| Message ID | 20200623020134.16655-1-pcc@google.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [v8] arm64: Expose FAR_EL1 tag bits in siginfo | expand |
Peter Collingbourne <pcc@google.com> writes: > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > index 47f651df781c..a8380a2b6361 100644 > --- a/arch/arm64/kernel/traps.c > +++ b/arch/arm64/kernel/traps.c > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > } > > void arm64_force_sig_fault(int signo, int code, void __user *addr, > + unsigned long far, unsigned char far_tb_mask, > const char *str) > { > arm64_show_signal(signo, str); > - if (signo == SIGKILL) > + if (signo == SIGKILL) { > force_sig(SIGKILL); > - else > - force_sig_fault(signo, code, addr); > + } else { > + struct kernel_siginfo info; > + clear_siginfo(&info); > + info.si_signo = signo; > + info.si_errno = 0; > + info.si_code = code; > + info.si_addr = addr; > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > + info.si_addr_top_byte_mask = far_tb_mask; > + force_sig_info(&info); > + } > } > > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > - const char *str) > + unsigned long far, const char *str) > { > + struct kernel_siginfo info; > + > arm64_show_signal(SIGBUS, str); > - force_sig_mceerr(code, addr, lsb); > + > + clear_siginfo(&info); > + info.si_signo = SIGBUS; > + info.si_errno = 0; > + info.si_code = code; > + info.si_addr = addr; > + info.si_addr_lsb = lsb; > + info.si_addr_top_byte = far >> 56; > + info.si_addr_top_byte_mask = 0xff; > + force_sig_info(&info); > } I have a real problem with this construction. force_sig_info is not an interface that should be used for anything except to define a wrapper that takes it's parameters. It is not clear to me that if you have adapted siginfo_layout. > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > index cb3d6c267181..6dd82373eb2d 100644 > --- a/include/uapi/asm-generic/siginfo.h > +++ b/include/uapi/asm-generic/siginfo.h > @@ -91,6 +91,14 @@ union __sifields { > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > __u32 _pkey; > } _addr_pkey; > +#ifdef __aarch64__ > + /* used with all si_codes */ > + struct { > + short _dummy_top_byte; > + unsigned char _top_byte; > + unsigned char _top_byte_mask; > + } _addr_top_byte; > +#endif > }; > } _sigfault; > Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". Please remove the "#ifdef __aarch64__". If at all possible we want to design this so any other architecture who has this challenge can use the code. The kind of code does not get enough attention/maintenance if it is built for a single architecture. > @@ -148,6 +156,10 @@ typedef struct siginfo { > #define si_int _sifields._rt._sigval.sival_int > #define si_ptr _sifields._rt._sigval.sival_ptr > #define si_addr _sifields._sigfault._addr > +#ifdef __aarch64__ > +#define si_addr_top_byte _sifields._sigfault._addr_top_byte._top_byte > +#define si_addr_top_byte_mask _sifields._sigfault._addr_top_byte._top_byte_mask > +#endif > #ifdef __ARCH_SI_TRAPNO > #define si_trapno _sifields._sigfault._trapno > #endif The details asside I think this set of changes is making progress. Eric
On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > Peter Collingbourne <pcc@google.com> writes: > > > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > > index 47f651df781c..a8380a2b6361 100644 > > --- a/arch/arm64/kernel/traps.c > > +++ b/arch/arm64/kernel/traps.c > > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > > } > > > > void arm64_force_sig_fault(int signo, int code, void __user *addr, > > + unsigned long far, unsigned char far_tb_mask, > > const char *str) > > { > > arm64_show_signal(signo, str); > > - if (signo == SIGKILL) > > + if (signo == SIGKILL) { > > force_sig(SIGKILL); > > - else > > - force_sig_fault(signo, code, addr); > > + } else { > > + struct kernel_siginfo info; > > + clear_siginfo(&info); > > + info.si_signo = signo; > > + info.si_errno = 0; > > + info.si_code = code; > > + info.si_addr = addr; > > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > > + info.si_addr_top_byte_mask = far_tb_mask; > > + force_sig_info(&info); > > + } > > } > > > > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > > - const char *str) > > + unsigned long far, const char *str) > > { > > + struct kernel_siginfo info; > > + > > arm64_show_signal(SIGBUS, str); > > - force_sig_mceerr(code, addr, lsb); > > + > > + clear_siginfo(&info); > > + info.si_signo = SIGBUS; > > + info.si_errno = 0; > > + info.si_code = code; > > + info.si_addr = addr; > > + info.si_addr_lsb = lsb; > > + info.si_addr_top_byte = far >> 56; > > + info.si_addr_top_byte_mask = 0xff; > > + force_sig_info(&info); > > } > > I have a real problem with this construction. force_sig_info is not an > interface that should be used for anything except to define a wrapper > that takes it's parameters. Can you elaborate? How would you do this king of thing. AIUI we absolutely need a forced signal here, we need to supply metadata, and we don't have to open-code all that at every relevant signal generation site... > It is not clear to me that if you have adapted siginfo_layout. Garbled sentence? > > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > index cb3d6c267181..6dd82373eb2d 100644 > > --- a/include/uapi/asm-generic/siginfo.h > > +++ b/include/uapi/asm-generic/siginfo.h > > @@ -91,6 +91,14 @@ union __sifields { > > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > __u32 _pkey; > > } _addr_pkey; > > +#ifdef __aarch64__ > > + /* used with all si_codes */ > > + struct { > > + short _dummy_top_byte; ^ What's this for? I don't have Eric's insight here. > > + unsigned char _top_byte; > > + unsigned char _top_byte_mask; > > + } _addr_top_byte; > > +#endif > > }; > > } _sigfault; > > > > Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > > Please remove the "#ifdef __aarch64__". If at all possible we want to > design this so any other architecture who has this challenge can use the > code. The kind of code does not get enough attention/maintenance if it > is built for a single architecture. Does this belong in the user-facing siginfo? It seems a bit strange, when other closely-related information such as esr_context is in the arch-specific signal frame. If trying to make this reusable, I wonder if we should have some sort of "address attributes" field. An alternative approach would be to add some opaque "arch_data" field, that the arch code can go look at when delivering the signal. I think that's all we were trying to achieve here: tack some arch private data onto the signal, to avoid having to stash the same info in thread_info and pray that it doesn't get clobbered in between signal generation and delivery. At signal delivery time, the arch signal delivery code could inspect this data and emit it into the signal frame as appropriate for the arch. [...] Cheers ---Dave
On Mon, Jun 22, 2020 at 07:01:34PM -0700, Peter Collingbourne wrote: > The kernel currently clears the tag bits (i.e. bits 56-63) in the fault > address exposed via siginfo.si_addr and sigcontext.fault_address. However, > the tag bits may be needed by tools in order to accurately diagnose > memory errors, such as HWASan [1] or future tools based on the Memory > Tagging Extension (MTE). [...] Minor nit: Can you stop making each version of this series in-reply-to the previous series please? Most people don't do that, and it's giving me weird threading in my mailbox... Cheers ---Dave
Dave Martin <Dave.Martin@arm.com> writes: > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: >> Peter Collingbourne <pcc@google.com> writes: >> >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c >> > index 47f651df781c..a8380a2b6361 100644 >> > --- a/arch/arm64/kernel/traps.c >> > +++ b/arch/arm64/kernel/traps.c >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) >> > } >> > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, >> > + unsigned long far, unsigned char far_tb_mask, >> > const char *str) >> > { >> > arm64_show_signal(signo, str); >> > - if (signo == SIGKILL) >> > + if (signo == SIGKILL) { >> > force_sig(SIGKILL); >> > - else >> > - force_sig_fault(signo, code, addr); >> > + } else { >> > + struct kernel_siginfo info; >> > + clear_siginfo(&info); >> > + info.si_signo = signo; >> > + info.si_errno = 0; >> > + info.si_code = code; >> > + info.si_addr = addr; >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; >> > + info.si_addr_top_byte_mask = far_tb_mask; >> > + force_sig_info(&info); >> > + } >> > } >> > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, >> > - const char *str) >> > + unsigned long far, const char *str) >> > { >> > + struct kernel_siginfo info; >> > + >> > arm64_show_signal(SIGBUS, str); >> > - force_sig_mceerr(code, addr, lsb); >> > + >> > + clear_siginfo(&info); >> > + info.si_signo = SIGBUS; >> > + info.si_errno = 0; >> > + info.si_code = code; >> > + info.si_addr = addr; >> > + info.si_addr_lsb = lsb; >> > + info.si_addr_top_byte = far >> 56; >> > + info.si_addr_top_byte_mask = 0xff; >> > + force_sig_info(&info); >> > } >> >> I have a real problem with this construction. force_sig_info is not an >> interface that should be used for anything except to define a wrapper >> that takes it's parameters. > > Can you elaborate? How would you do this king of thing. There are no other uses of force_sig_info in architecture code. I just removed them _all_ because they were almost all broken. In fact your mcerr case is broken because it uses two different union members simultantiously. So I am looking for something like force_sig_mcerr or force_sig_fault that includes your new information that then calls force_sig_info. I know of no other way to safely use the siginfo struct. > AIUI we absolutely need a forced signal here, we need to supply > metadata, and we don't have to open-code all that at every relevant > signal generation site... > >> It is not clear to me that if you have adapted siginfo_layout. > > Garbled sentence? Looks like. One of the pieces of code that needs to change when siginfo gets updated is siginfo_layout so that the structure can be properly decoded and made sense of. I am not seeing anything like that. >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h >> > index cb3d6c267181..6dd82373eb2d 100644 >> > --- a/include/uapi/asm-generic/siginfo.h >> > +++ b/include/uapi/asm-generic/siginfo.h >> > @@ -91,6 +91,14 @@ union __sifields { >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; >> > __u32 _pkey; >> > } _addr_pkey; >> > +#ifdef __aarch64__ >> > + /* used with all si_codes */ >> > + struct { >> > + short _dummy_top_byte; > > ^ What's this for? I don't have Eric's insight here. > >> > + unsigned char _top_byte; >> > + unsigned char _top_byte_mask; >> > + } _addr_top_byte; >> > +#endif >> > }; >> > } _sigfault; >> > >> >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". >> >> Please remove the "#ifdef __aarch64__". If at all possible we want to >> design this so any other architecture who has this challenge can use the >> code. The kind of code does not get enough attention/maintenance if it >> is built for a single architecture. > > Does this belong in the user-facing siginfo? It seems a bit strange, > when other closely-related information such as esr_context is in the > arch-specific signal frame. > > > If trying to make this reusable, I wonder if we should have some sort of > "address attributes" field. > > An alternative approach would be to add some opaque "arch_data" field, > that the arch code can go look at when delivering the signal. My point is arch specific hacks don't get looked at, and wind up being broken. So I am not encouraging anything that doesn't get looked at, and winds up being broken. > I think that's all we were trying to achieve here: tack some arch > private data onto the signal, to avoid having to stash the same info in > thread_info and pray that it doesn't get clobbered in between signal > generation and delivery. What makes it arch private data? Why isn't it just data that your arch happens to have that other architectures don't yet. > At signal delivery time, the arch signal delivery code could inspect > this data and emit it into the signal frame as appropriate for the > arch. Sorry this probably isn't what you mean but when I read that description I get the feeling that you are asking for code that won't be reviewed or looked at by anyone else. So inevitably that code will be broken. Frankly it is bad enough finding people to review and maintain the generic code of the kernel. With that said, and your desire for this data to go into the sigframe (despite it sounding a lot like generic data that only aarch64 has implemented yet) can you remind me why siginfo comes into the equation at all? Last I remember the discussion there were some issues and the plan was to simply solve the problem generically and use siginfo, and there would not need to be any sigframe changes. But if you want to deliver via sigframe force_sig_info and all it's variants will be delivered when the kernel returns back to userspace. So there should be no need to touch siginfo or anything else in that scenario. Eric
On Tue, Jun 23, 2020 at 10:52 AM Eric W. Biederman <ebiederm@xmission.com> wrote: > > Dave Martin <Dave.Martin@arm.com> writes: > > > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > >> Peter Collingbourne <pcc@google.com> writes: > >> > >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > >> > index 47f651df781c..a8380a2b6361 100644 > >> > --- a/arch/arm64/kernel/traps.c > >> > +++ b/arch/arm64/kernel/traps.c > >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > >> > } > >> > > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, > >> > + unsigned long far, unsigned char far_tb_mask, > >> > const char *str) > >> > { > >> > arm64_show_signal(signo, str); > >> > - if (signo == SIGKILL) > >> > + if (signo == SIGKILL) { > >> > force_sig(SIGKILL); > >> > - else > >> > - force_sig_fault(signo, code, addr); > >> > + } else { > >> > + struct kernel_siginfo info; > >> > + clear_siginfo(&info); > >> > + info.si_signo = signo; > >> > + info.si_errno = 0; > >> > + info.si_code = code; > >> > + info.si_addr = addr; > >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > >> > + info.si_addr_top_byte_mask = far_tb_mask; > >> > + force_sig_info(&info); > >> > + } > >> > } > >> > > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > >> > - const char *str) > >> > + unsigned long far, const char *str) > >> > { > >> > + struct kernel_siginfo info; > >> > + > >> > arm64_show_signal(SIGBUS, str); > >> > - force_sig_mceerr(code, addr, lsb); > >> > + > >> > + clear_siginfo(&info); > >> > + info.si_signo = SIGBUS; > >> > + info.si_errno = 0; > >> > + info.si_code = code; > >> > + info.si_addr = addr; > >> > + info.si_addr_lsb = lsb; > >> > + info.si_addr_top_byte = far >> 56; > >> > + info.si_addr_top_byte_mask = 0xff; > >> > + force_sig_info(&info); > >> > } > >> > >> I have a real problem with this construction. force_sig_info is not an > >> interface that should be used for anything except to define a wrapper > >> that takes it's parameters. > > > > Can you elaborate? How would you do this king of thing. > > There are no other uses of force_sig_info in architecture code. > > I just removed them _all_ because they were almost all broken. > In fact your mcerr case is broken because it uses two different > union members simultantiously. Is that really broken? I thought that the Linux kernel deliberately didn't care about strict aliasing rules (the top-level Makefile passes -fno-strict-aliasing) so I thought that it was valid in "Linux kernel C" even though from a standards point of view it is invalid. (That being said, this is probably moot with my proposed changes below though.) > So I am looking for something like force_sig_mcerr or force_sig_fault > that includes your new information that then calls force_sig_info. > > I know of no other way to safely use the siginfo struct. So you want something like: int force_sig_fault_with_ignored_bits(int signo, int code, void __user *addr, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); int force_sig_mceerr_with_ignored_bits(int code, void __user *addr, short lsb, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); in kernel/signal.c and the code in arch/arm64 would call that? > > AIUI we absolutely need a forced signal here, we need to supply > > metadata, and we don't have to open-code all that at every relevant > > signal generation site... > > > >> It is not clear to me that if you have adapted siginfo_layout. > > > > Garbled sentence? > > Looks like. One of the pieces of code that needs to change > when siginfo gets updated is siginfo_layout so that the structure > can be properly decoded and made sense of. > > I am not seeing anything like that. Okay, this has to do with copying between the compat and non-compat versions of the struct? Sure, I can update that, although the code would be basically non-functional on arm64 because TBI isn't supported on 32-bit ARM. > >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > >> > index cb3d6c267181..6dd82373eb2d 100644 > >> > --- a/include/uapi/asm-generic/siginfo.h > >> > +++ b/include/uapi/asm-generic/siginfo.h > >> > @@ -91,6 +91,14 @@ union __sifields { > >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > >> > __u32 _pkey; > >> > } _addr_pkey; > >> > +#ifdef __aarch64__ > >> > + /* used with all si_codes */ > >> > + struct { > >> > + short _dummy_top_byte; > > > > ^ What's this for? I don't have Eric's insight here. We would need a short's worth of padding in order to prevent the fields from occupying the same address as si_addr_lsb. > > > >> > + unsigned char _top_byte; > >> > + unsigned char _top_byte_mask; > >> > + } _addr_top_byte; > >> > +#endif > >> > }; > >> > } _sigfault; > >> > > >> > >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > >> > >> Please remove the "#ifdef __aarch64__". If at all possible we want to > >> design this so any other architecture who has this challenge can use the > >> code. The kind of code does not get enough attention/maintenance if it > >> is built for a single architecture. Seems reasonable. I was recently made aware that RISC-V was considering a similar feature: https://lists.riscv.org/g/tech-tee/topic/risc_v_tbi_proposal/72855478 I would have opted to expand this to other architectures on an as-needed basis, but I'd also be fine with having it on all architectures from the start. If we make this arch-independent, we have an additional concern, which is "what if some future architecture wants more than one byte here?" For example, an architecture may have a "top-two-bytes-ignore" feature, which would imply two-byte (misnamed) "si_addr_top_byte" and "si_addr_top_byte_mask" fields. And the RISC-V proposal potentially implies many more ignored bits (see slide 13 of the presentation). The maximum size that these fields can possibly be is the size of a pointer, and with that there wouldn't be enough room in the padding at this point to accommodate the new fields. That basically implies your earlier suggestion of adding a union member here to accommodate future expansion of the union, and adding the new fields after the union. I'm happy to make that change, with the fields renamed "si_addr_ignored" and "si_addr_ignored_mask". > > > > Does this belong in the user-facing siginfo? It seems a bit strange, > > when other closely-related information such as esr_context is in the > > arch-specific signal frame. > > > > > > If trying to make this reusable, I wonder if we should have some sort of > > "address attributes" field. > > > > An alternative approach would be to add some opaque "arch_data" field, > > that the arch code can go look at when delivering the signal. > > My point is arch specific hacks don't get looked at, and wind up being > broken. So I am not encouraging anything that doesn't get looked at, > and winds up being broken. > > > I think that's all we were trying to achieve here: tack some arch > > private data onto the signal, to avoid having to stash the same info in > > thread_info and pray that it doesn't get clobbered in between signal > > generation and delivery. > > What makes it arch private data? Why isn't it just data that your arch > happens to have that other architectures don't yet. > > > At signal delivery time, the arch signal delivery code could inspect > > this data and emit it into the signal frame as appropriate for the > > arch. > > Sorry this probably isn't what you mean but when I read that description > I get the feeling that you are asking for code that won't be reviewed or > looked at by anyone else. So inevitably that code will be broken. > Frankly it is bad enough finding people to review and maintain the > generic code of the kernel. > > > With that said, and your desire for this data to go into the sigframe > (despite it sounding a lot like generic data that only aarch64 has > implemented yet) can you remind me why siginfo comes into the equation > at all? > > Last I remember the discussion there were some issues and the plan was > to simply solve the problem generically and use siginfo, and there would > not need to be any sigframe changes. > > But if you want to deliver via sigframe force_sig_info and all it's > variants will be delivered when the kernel returns back to userspace. > So there should be no need to touch siginfo or anything else in that > scenario. My understanding is that siginfo should contain information about the signal itself, while sigcontext should contain any information about the machine state at the point when the signal was delivered that is needed in order to restore the state after returning from a signal handler. The fault address isn't really part of the restorable machine state (despite the existence of a "fault_address" field in sigcontext), so any information relating to it belongs (at least morally) in siginfo. Peter
On Tue, Jun 23, 2020 at 05:40:08PM -0700, Peter Collingbourne wrote: > On Tue, Jun 23, 2020 at 10:52 AM Eric W. Biederman > <ebiederm@xmission.com> wrote: > > > > Dave Martin <Dave.Martin@arm.com> writes: > > > > > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > > >> Peter Collingbourne <pcc@google.com> writes: > > >> > > >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > > >> > index 47f651df781c..a8380a2b6361 100644 > > >> > --- a/arch/arm64/kernel/traps.c > > >> > +++ b/arch/arm64/kernel/traps.c > > >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > > >> > } > > >> > > > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, > > >> > + unsigned long far, unsigned char far_tb_mask, > > >> > const char *str) > > >> > { > > >> > arm64_show_signal(signo, str); > > >> > - if (signo == SIGKILL) > > >> > + if (signo == SIGKILL) { > > >> > force_sig(SIGKILL); > > >> > - else > > >> > - force_sig_fault(signo, code, addr); > > >> > + } else { > > >> > + struct kernel_siginfo info; > > >> > + clear_siginfo(&info); > > >> > + info.si_signo = signo; > > >> > + info.si_errno = 0; > > >> > + info.si_code = code; > > >> > + info.si_addr = addr; > > >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > > >> > + info.si_addr_top_byte_mask = far_tb_mask; > > >> > + force_sig_info(&info); > > >> > + } > > >> > } > > >> > > > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > > >> > - const char *str) > > >> > + unsigned long far, const char *str) > > >> > { > > >> > + struct kernel_siginfo info; > > >> > + > > >> > arm64_show_signal(SIGBUS, str); > > >> > - force_sig_mceerr(code, addr, lsb); > > >> > + > > >> > + clear_siginfo(&info); > > >> > + info.si_signo = SIGBUS; > > >> > + info.si_errno = 0; > > >> > + info.si_code = code; > > >> > + info.si_addr = addr; > > >> > + info.si_addr_lsb = lsb; > > >> > + info.si_addr_top_byte = far >> 56; > > >> > + info.si_addr_top_byte_mask = 0xff; > > >> > + force_sig_info(&info); > > >> > } > > >> > > >> I have a real problem with this construction. force_sig_info is not an > > >> interface that should be used for anything except to define a wrapper > > >> that takes it's parameters. > > > > > > Can you elaborate? How would you do this king of thing. > > > > There are no other uses of force_sig_info in architecture code. > > > > I just removed them _all_ because they were almost all broken. > > In fact your mcerr case is broken because it uses two different > > union members simultantiously. > > Is that really broken? I thought that the Linux kernel deliberately > didn't care about strict aliasing rules (the top-level Makefile passes > -fno-strict-aliasing) so I thought that it was valid in "Linux kernel > C" even though from a standards point of view it is invalid. (That > being said, this is probably moot with my proposed changes below > though.) I have a feeling that -fno-strict-aliasing only allows you to _read_ a different union member from the one previously written. Writing a different member from the last one written can still splatter on the other members IIUC. It would be better to keep things separate rather than risk incorrectness just to save a few bytes. IMHO -fno-strict-aliasing is no excuse for gratuitous type-punning. > > So I am looking for something like force_sig_mcerr or force_sig_fault > > that includes your new information that then calls force_sig_info. > > > > I know of no other way to safely use the siginfo struct. > > So you want something like: > > int force_sig_fault_with_ignored_bits(int signo, int code, void __user > *addr, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > int force_sig_mceerr_with_ignored_bits(int code, void __user *addr, > short lsb, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > in kernel/signal.c and the code in arch/arm64 would call that? > > > > AIUI we absolutely need a forced signal here, we need to supply > > > metadata, and we don't have to open-code all that at every relevant > > > signal generation site... > > > > > >> It is not clear to me that if you have adapted siginfo_layout. > > > > > > Garbled sentence? > > > > Looks like. One of the pieces of code that needs to change > > when siginfo gets updated is siginfo_layout so that the structure > > can be properly decoded and made sense of. > > > > I am not seeing anything like that. > > Okay, this has to do with copying between the compat and non-compat > versions of the struct? Sure, I can update that, although the code > would be basically non-functional on arm64 because TBI isn't supported > on 32-bit ARM. > > > >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > >> > index cb3d6c267181..6dd82373eb2d 100644 > > >> > --- a/include/uapi/asm-generic/siginfo.h > > >> > +++ b/include/uapi/asm-generic/siginfo.h > > >> > @@ -91,6 +91,14 @@ union __sifields { > > >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > >> > __u32 _pkey; > > >> > } _addr_pkey; > > >> > +#ifdef __aarch64__ > > >> > + /* used with all si_codes */ > > >> > + struct { > > >> > + short _dummy_top_byte; > > > > > > ^ What's this for? I don't have Eric's insight here. > > We would need a short's worth of padding in order to prevent the > fields from occupying the same address as si_addr_lsb. > > > > > > >> > + unsigned char _top_byte; > > >> > + unsigned char _top_byte_mask; > > >> > + } _addr_top_byte; > > >> > +#endif > > >> > }; > > >> > } _sigfault; > > >> > > > >> > > >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > > >> > > >> Please remove the "#ifdef __aarch64__". If at all possible we want to > > >> design this so any other architecture who has this challenge can use the > > >> code. The kind of code does not get enough attention/maintenance if it > > >> is built for a single architecture. > > Seems reasonable. I was recently made aware that RISC-V was > considering a similar feature: > https://lists.riscv.org/g/tech-tee/topic/risc_v_tbi_proposal/72855478 > I would have opted to expand this to other architectures on an > as-needed basis, but I'd also be fine with having it on all > architectures from the start. > > If we make this arch-independent, we have an additional concern, which > is "what if some future architecture wants more than one byte here?" > For example, an architecture may have a "top-two-bytes-ignore" > feature, which would imply two-byte (misnamed) "si_addr_top_byte" and > "si_addr_top_byte_mask" fields. And the RISC-V proposal potentially > implies many more ignored bits (see slide 13 of the presentation). The > maximum size that these fields can possibly be is the size of a > pointer, and with that there wouldn't be enough room in the padding at > this point to accommodate the new fields. > > That basically implies your earlier suggestion of adding a union > member here to accommodate future expansion of the union, and adding > the new fields after the union. I'm happy to make that change, with > the fields renamed "si_addr_ignored" and "si_addr_ignored_mask". I think what we need here is basically a flags word. So long as we keep a flag spare to indicate the existence of a further flags word, we can extend as needed. How the existence of the first flags words is detected is another problem. If it only applies for newly-defined si_code values, then I guess si_code may be sufficient. > > > > > > Does this belong in the user-facing siginfo? It seems a bit strange, > > > when other closely-related information such as esr_context is in the > > > arch-specific signal frame. > > > > > > > > > If trying to make this reusable, I wonder if we should have some sort of > > > "address attributes" field. > > > > > > An alternative approach would be to add some opaque "arch_data" field, > > > that the arch code can go look at when delivering the signal. > > > > My point is arch specific hacks don't get looked at, and wind up being > > broken. So I am not encouraging anything that doesn't get looked at, > > and winds up being broken. Arch code will get looked at, and is automatically inherently broken. Nor is the core code always perfect... I agree that generic is best, both for getting more eyes on it and for coming up with a clean design, but there's also a risk of pointless over-abstraction for things that just aren't generic enough. Part of the issue is that each arch necessarily has its own way of dumping its register state, while siginfo contains abstract diagnostic information. The boundary between these two is not clear-cut: for example, arm64 dumps its exception syndrome register which contains (among other things) imformation about whether a faulted access was a read or write. Is this generic information, or arch-specific information? A side problem is that siginfo_t as originally designed is quite hard to extend. AFAICT, any extension needs a new si_code, otherwise there is no way to detect that the extension fields are present. This is fine for defining entirely new signal types, but seems to make it hard to add supplementary information for existing signals. Have I missed something here? Say we wanted to add extra data for SIGSEGV to indicate the size of access and whether it was a read or write. If we try to add a new si_code for this, then all software that inspects si_code at all for SIGSEGV now has no idea what to do with this new si_code. Reading between the lines, I wonder whether this is part of the reason arches tend to go their own way: such information can't be added generically precisely because it _is_ generic -- too generic to justify a new si_code. If so, this problem is going to crop up again and again... > > > I think that's all we were trying to achieve here: tack some arch > > > private data onto the signal, to avoid having to stash the same info in > > > thread_info and pray that it doesn't get clobbered in between signal > > > generation and delivery. > > > > What makes it arch private data? Why isn't it just data that your arch > > happens to have that other architectures don't yet. I didn't mean it must be private, just that it can be. > > > At signal delivery time, the arch signal delivery code could inspect > > > this data and emit it into the signal frame as appropriate for the > > > arch. > > > > Sorry this probably isn't what you mean but when I read that description > > I get the feeling that you are asking for code that won't be reviewed or > > looked at by anyone else. So inevitably that code will be broken. > > Frankly it is bad enough finding people to review and maintain the > > generic code of the kernel. Does this need flag up to the arch maintainers? Signal code has been heavily arch-specific for ages, and that's where the force of gravity seems to point. I a lot of work has gone into cleaning this up, but it sounds like arch maintainers might need to push back harder on anything that _could_ be done in the common code. > > With that said, and your desire for this data to go into the sigframe > > (despite it sounding a lot like generic data that only aarch64 has > > implemented yet) can you remind me why siginfo comes into the equation > > at all? > > > > Last I remember the discussion there were some issues and the plan was > > to simply solve the problem generically and use siginfo, and there would > > not need to be any sigframe changes. > > > > But if you want to deliver via sigframe force_sig_info and all it's > > variants will be delivered when the kernel returns back to userspace. > > So there should be no need to touch siginfo or anything else in that > > scenario. > > My understanding is that siginfo should contain information about the > signal itself, while sigcontext should contain any information about > the machine state at the point when the signal was delivered that is > needed in order to restore the state after returning from a signal > handler. The fault address isn't really part of the restorable machine > state (despite the existence of a "fault_address" field in > sigcontext), so any information relating to it belongs (at least > morally) in siginfo. I think this is more than just a principle. Diagnostic information that is supposed to accompany a signal needs to be captured at the time the signal is generated, otherwise it may have been overwritten by another signal-generating event by the time the first signal is actually delivered. Currently this isn't handled properly in the arm64 code, so it looks like some diagnostic fields in the arm64 signal frame can be wrong in some situations. (I know, that's your "non-generic code that hardly anyone relies on will be broken" argument. But the need to keep diagnostic information with the signal instance it relates to feels like a generic problem.) I have no objection to finding a generic way to report the address tag information, but "address tag" is not the most generic concept in the world, even if there are a few arches with something analogous. Cheers ---Dave
On Wed, Jun 24, 2020 at 2:28 AM Dave Martin <Dave.Martin@arm.com> wrote: > > On Tue, Jun 23, 2020 at 05:40:08PM -0700, Peter Collingbourne wrote: > > On Tue, Jun 23, 2020 at 10:52 AM Eric W. Biederman > > <ebiederm@xmission.com> wrote: > > > > > > Dave Martin <Dave.Martin@arm.com> writes: > > > > > > > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > > > >> Peter Collingbourne <pcc@google.com> writes: > > > >> > > > >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > > > >> > index 47f651df781c..a8380a2b6361 100644 > > > >> > --- a/arch/arm64/kernel/traps.c > > > >> > +++ b/arch/arm64/kernel/traps.c > > > >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > > > >> > } > > > >> > > > > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, > > > >> > + unsigned long far, unsigned char far_tb_mask, > > > >> > const char *str) > > > >> > { > > > >> > arm64_show_signal(signo, str); > > > >> > - if (signo == SIGKILL) > > > >> > + if (signo == SIGKILL) { > > > >> > force_sig(SIGKILL); > > > >> > - else > > > >> > - force_sig_fault(signo, code, addr); > > > >> > + } else { > > > >> > + struct kernel_siginfo info; > > > >> > + clear_siginfo(&info); > > > >> > + info.si_signo = signo; > > > >> > + info.si_errno = 0; > > > >> > + info.si_code = code; > > > >> > + info.si_addr = addr; > > > >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > > > >> > + info.si_addr_top_byte_mask = far_tb_mask; > > > >> > + force_sig_info(&info); > > > >> > + } > > > >> > } > > > >> > > > > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > > > >> > - const char *str) > > > >> > + unsigned long far, const char *str) > > > >> > { > > > >> > + struct kernel_siginfo info; > > > >> > + > > > >> > arm64_show_signal(SIGBUS, str); > > > >> > - force_sig_mceerr(code, addr, lsb); > > > >> > + > > > >> > + clear_siginfo(&info); > > > >> > + info.si_signo = SIGBUS; > > > >> > + info.si_errno = 0; > > > >> > + info.si_code = code; > > > >> > + info.si_addr = addr; > > > >> > + info.si_addr_lsb = lsb; > > > >> > + info.si_addr_top_byte = far >> 56; > > > >> > + info.si_addr_top_byte_mask = 0xff; > > > >> > + force_sig_info(&info); > > > >> > } > > > >> > > > >> I have a real problem with this construction. force_sig_info is not an > > > >> interface that should be used for anything except to define a wrapper > > > >> that takes it's parameters. > > > > > > > > Can you elaborate? How would you do this king of thing. > > > > > > There are no other uses of force_sig_info in architecture code. > > > > > > I just removed them _all_ because they were almost all broken. > > > In fact your mcerr case is broken because it uses two different > > > union members simultantiously. > > > > Is that really broken? I thought that the Linux kernel deliberately > > didn't care about strict aliasing rules (the top-level Makefile passes > > -fno-strict-aliasing) so I thought that it was valid in "Linux kernel > > C" even though from a standards point of view it is invalid. (That > > being said, this is probably moot with my proposed changes below > > though.) > > I have a feeling that -fno-strict-aliasing only allows you to _read_ a > different union member from the one previously written. > > Writing a different member from the last one written can still splatter > on the other members IIUC. > > It would be better to keep things separate rather than risk > incorrectness just to save a few bytes. > > IMHO -fno-strict-aliasing is no excuse for gratuitous type-punning. > > > > So I am looking for something like force_sig_mcerr or force_sig_fault > > > that includes your new information that then calls force_sig_info. > > > > > > I know of no other way to safely use the siginfo struct. > > > > So you want something like: > > > > int force_sig_fault_with_ignored_bits(int signo, int code, void __user > > *addr, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > int force_sig_mceerr_with_ignored_bits(int code, void __user *addr, > > short lsb, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > in kernel/signal.c and the code in arch/arm64 would call that? > > > > > > AIUI we absolutely need a forced signal here, we need to supply > > > > metadata, and we don't have to open-code all that at every relevant > > > > signal generation site... > > > > > > > >> It is not clear to me that if you have adapted siginfo_layout. > > > > > > > > Garbled sentence? > > > > > > Looks like. One of the pieces of code that needs to change > > > when siginfo gets updated is siginfo_layout so that the structure > > > can be properly decoded and made sense of. > > > > > > I am not seeing anything like that. > > > > Okay, this has to do with copying between the compat and non-compat > > versions of the struct? Sure, I can update that, although the code > > would be basically non-functional on arm64 because TBI isn't supported > > on 32-bit ARM. > > > > > >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > > >> > index cb3d6c267181..6dd82373eb2d 100644 > > > >> > --- a/include/uapi/asm-generic/siginfo.h > > > >> > +++ b/include/uapi/asm-generic/siginfo.h > > > >> > @@ -91,6 +91,14 @@ union __sifields { > > > >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > >> > __u32 _pkey; > > > >> > } _addr_pkey; > > > >> > +#ifdef __aarch64__ > > > >> > + /* used with all si_codes */ > > > >> > + struct { > > > >> > + short _dummy_top_byte; > > > > > > > > ^ What's this for? I don't have Eric's insight here. > > > > We would need a short's worth of padding in order to prevent the > > fields from occupying the same address as si_addr_lsb. > > > > > > > > > >> > + unsigned char _top_byte; > > > >> > + unsigned char _top_byte_mask; > > > >> > + } _addr_top_byte; > > > >> > +#endif > > > >> > }; > > > >> > } _sigfault; > > > >> > > > > >> > > > >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > > > >> > > > >> Please remove the "#ifdef __aarch64__". If at all possible we want to > > > >> design this so any other architecture who has this challenge can use the > > > >> code. The kind of code does not get enough attention/maintenance if it > > > >> is built for a single architecture. > > > > Seems reasonable. I was recently made aware that RISC-V was > > considering a similar feature: > > https://lists.riscv.org/g/tech-tee/topic/risc_v_tbi_proposal/72855478 > > I would have opted to expand this to other architectures on an > > as-needed basis, but I'd also be fine with having it on all > > architectures from the start. > > > > If we make this arch-independent, we have an additional concern, which > > is "what if some future architecture wants more than one byte here?" > > For example, an architecture may have a "top-two-bytes-ignore" > > feature, which would imply two-byte (misnamed) "si_addr_top_byte" and > > "si_addr_top_byte_mask" fields. And the RISC-V proposal potentially > > implies many more ignored bits (see slide 13 of the presentation). The > > maximum size that these fields can possibly be is the size of a > > pointer, and with that there wouldn't be enough room in the padding at > > this point to accommodate the new fields. > > > > That basically implies your earlier suggestion of adding a union > > member here to accommodate future expansion of the union, and adding > > the new fields after the union. I'm happy to make that change, with > > the fields renamed "si_addr_ignored" and "si_addr_ignored_mask". > > I think what we need here is basically a flags word. > > So long as we keep a flag spare to indicate the existence of a further > flags word, we can extend as needed. > > How the existence of the first flags words is detected is another > problem. If it only applies for newly-defined si_code values, then > I guess si_code may be sufficient. Existing kernels will zero-initialize unused regions of the siginfo data structure. The zero-initialization of the padding at the end of the struct is done by the clear_user call here: https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/kernel/signal.c#L3193 and the zero-initialization of the padding between fields and unused union members is done by the clear_siginfo function which the kernel calls when initializing the data structure: https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/include/linux/signal.h#L20 Therefore, a flag word value of 0 may be used to detect a lack of support for flagged fields. That being said, in this particular case, we do not need a flag word. We can just take advantage of this zero-initialization behavior in existing kernels to set si_addr_ignored_mask to 0, which indicates that none of the bits in si_addr_ignored are valid. Peter > > > > > > > > Does this belong in the user-facing siginfo? It seems a bit strange, > > > > when other closely-related information such as esr_context is in the > > > > arch-specific signal frame. > > > > > > > > > > > > If trying to make this reusable, I wonder if we should have some sort of > > > > "address attributes" field. > > > > > > > > An alternative approach would be to add some opaque "arch_data" field, > > > > that the arch code can go look at when delivering the signal. > > > > > > My point is arch specific hacks don't get looked at, and wind up being > > > broken. So I am not encouraging anything that doesn't get looked at, > > > and winds up being broken. > > Arch code will get looked at, and is automatically inherently broken. > Nor is the core code always perfect... > > I agree that generic is best, both for getting more eyes on it and for > coming up with a clean design, but there's also a risk of pointless > over-abstraction for things that just aren't generic enough. > > Part of the issue is that each arch necessarily has its own way of > dumping its register state, while siginfo contains abstract diagnostic > information. The boundary between these two is not clear-cut: for > example, arm64 dumps its exception syndrome register which contains > (among other things) imformation about whether a faulted access was a > read or write. Is this generic information, or arch-specific > information? > > > A side problem is that siginfo_t as originally designed is quite hard to > extend. > > AFAICT, any extension needs a new si_code, otherwise there is no way > to detect that the extension fields are present. This is fine for > defining entirely new signal types, but seems to make it hard to add > supplementary information for existing signals. Have I missed something > here? > > Say we wanted to add extra data for SIGSEGV to indicate the size of > access and whether it was a read or write. If we try to add a new > si_code for this, then all software that inspects si_code at all for > SIGSEGV now has no idea what to do with this new si_code. > > Reading between the lines, I wonder whether this is part of the reason > arches tend to go their own way: such information can't be added > generically precisely because it _is_ generic -- too generic to justify > a new si_code. If so, this problem is going to crop up again and > again... > > > > > I think that's all we were trying to achieve here: tack some arch > > > > private data onto the signal, to avoid having to stash the same info in > > > > thread_info and pray that it doesn't get clobbered in between signal > > > > generation and delivery. > > > > > > What makes it arch private data? Why isn't it just data that your arch > > > happens to have that other architectures don't yet. > > I didn't mean it must be private, just that it can be. > > > > > At signal delivery time, the arch signal delivery code could inspect > > > > this data and emit it into the signal frame as appropriate for the > > > > arch. > > > > > > Sorry this probably isn't what you mean but when I read that description > > > I get the feeling that you are asking for code that won't be reviewed or > > > looked at by anyone else. So inevitably that code will be broken. > > > Frankly it is bad enough finding people to review and maintain the > > > generic code of the kernel. > > Does this need flag up to the arch maintainers? Signal code has been > heavily arch-specific for ages, and that's where the force of gravity > seems to point. I a lot of work has gone into cleaning this up, but it > sounds like arch maintainers might need to push back harder on anything > that _could_ be done in the common code. > > > > With that said, and your desire for this data to go into the sigframe > > > (despite it sounding a lot like generic data that only aarch64 has > > > implemented yet) can you remind me why siginfo comes into the equation > > > at all? > > > > > > Last I remember the discussion there were some issues and the plan was > > > to simply solve the problem generically and use siginfo, and there would > > > not need to be any sigframe changes. > > > > > > But if you want to deliver via sigframe force_sig_info and all it's > > > variants will be delivered when the kernel returns back to userspace. > > > So there should be no need to touch siginfo or anything else in that > > > scenario. > > > > My understanding is that siginfo should contain information about the > > signal itself, while sigcontext should contain any information about > > the machine state at the point when the signal was delivered that is > > needed in order to restore the state after returning from a signal > > handler. The fault address isn't really part of the restorable machine > > state (despite the existence of a "fault_address" field in > > sigcontext), so any information relating to it belongs (at least > > morally) in siginfo. > > I think this is more than just a principle. > > Diagnostic information that is supposed to accompany a signal needs to > be captured at the time the signal is generated, otherwise it may have > been overwritten by another signal-generating event by the time the > first signal is actually delivered. > > Currently this isn't handled properly in the arm64 code, so it looks > like some diagnostic fields in the arm64 signal frame can be wrong in > some situations. (I know, that's your "non-generic code that hardly > anyone relies on will be broken" argument. But the need to keep > diagnostic information with the signal instance it relates to feels like > a generic problem.) > > I have no objection to finding a generic way to report the address tag > information, but "address tag" is not the most generic concept in the > world, even if there are a few arches with something analogous. > > Cheers > ---Dave
On Wed, Jun 24, 2020 at 09:51:49AM -0700, Peter Collingbourne wrote: > On Wed, Jun 24, 2020 at 2:28 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > On Tue, Jun 23, 2020 at 05:40:08PM -0700, Peter Collingbourne wrote: > > > On Tue, Jun 23, 2020 at 10:52 AM Eric W. Biederman > > > <ebiederm@xmission.com> wrote: > > > > > > > > Dave Martin <Dave.Martin@arm.com> writes: > > > > > > > > > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > > > > >> Peter Collingbourne <pcc@google.com> writes: > > > > >> > > > > >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > > > > >> > index 47f651df781c..a8380a2b6361 100644 > > > > >> > --- a/arch/arm64/kernel/traps.c > > > > >> > +++ b/arch/arm64/kernel/traps.c > > > > >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > > > > >> > } > > > > >> > > > > > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, > > > > >> > + unsigned long far, unsigned char far_tb_mask, > > > > >> > const char *str) > > > > >> > { > > > > >> > arm64_show_signal(signo, str); > > > > >> > - if (signo == SIGKILL) > > > > >> > + if (signo == SIGKILL) { > > > > >> > force_sig(SIGKILL); > > > > >> > - else > > > > >> > - force_sig_fault(signo, code, addr); > > > > >> > + } else { > > > > >> > + struct kernel_siginfo info; > > > > >> > + clear_siginfo(&info); > > > > >> > + info.si_signo = signo; > > > > >> > + info.si_errno = 0; > > > > >> > + info.si_code = code; > > > > >> > + info.si_addr = addr; > > > > >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > > > > >> > + info.si_addr_top_byte_mask = far_tb_mask; > > > > >> > + force_sig_info(&info); > > > > >> > + } > > > > >> > } > > > > >> > > > > > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > > > > >> > - const char *str) > > > > >> > + unsigned long far, const char *str) > > > > >> > { > > > > >> > + struct kernel_siginfo info; > > > > >> > + > > > > >> > arm64_show_signal(SIGBUS, str); > > > > >> > - force_sig_mceerr(code, addr, lsb); > > > > >> > + > > > > >> > + clear_siginfo(&info); > > > > >> > + info.si_signo = SIGBUS; > > > > >> > + info.si_errno = 0; > > > > >> > + info.si_code = code; > > > > >> > + info.si_addr = addr; > > > > >> > + info.si_addr_lsb = lsb; > > > > >> > + info.si_addr_top_byte = far >> 56; > > > > >> > + info.si_addr_top_byte_mask = 0xff; > > > > >> > + force_sig_info(&info); > > > > >> > } > > > > >> > > > > >> I have a real problem with this construction. force_sig_info is not an > > > > >> interface that should be used for anything except to define a wrapper > > > > >> that takes it's parameters. > > > > > > > > > > Can you elaborate? How would you do this king of thing. > > > > > > > > There are no other uses of force_sig_info in architecture code. > > > > > > > > I just removed them _all_ because they were almost all broken. > > > > In fact your mcerr case is broken because it uses two different > > > > union members simultantiously. > > > > > > Is that really broken? I thought that the Linux kernel deliberately > > > didn't care about strict aliasing rules (the top-level Makefile passes > > > -fno-strict-aliasing) so I thought that it was valid in "Linux kernel > > > C" even though from a standards point of view it is invalid. (That > > > being said, this is probably moot with my proposed changes below > > > though.) > > > > I have a feeling that -fno-strict-aliasing only allows you to _read_ a > > different union member from the one previously written. > > > > Writing a different member from the last one written can still splatter > > on the other members IIUC. > > > > It would be better to keep things separate rather than risk > > incorrectness just to save a few bytes. > > > > IMHO -fno-strict-aliasing is no excuse for gratuitous type-punning. > > > > > > So I am looking for something like force_sig_mcerr or force_sig_fault > > > > that includes your new information that then calls force_sig_info. > > > > > > > > I know of no other way to safely use the siginfo struct. > > > > > > So you want something like: > > > > > > int force_sig_fault_with_ignored_bits(int signo, int code, void __user > > > *addr, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > int force_sig_mceerr_with_ignored_bits(int code, void __user *addr, > > > short lsb, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > in kernel/signal.c and the code in arch/arm64 would call that? > > > > > > > > AIUI we absolutely need a forced signal here, we need to supply > > > > > metadata, and we don't have to open-code all that at every relevant > > > > > signal generation site... > > > > > > > > > >> It is not clear to me that if you have adapted siginfo_layout. > > > > > > > > > > Garbled sentence? > > > > > > > > Looks like. One of the pieces of code that needs to change > > > > when siginfo gets updated is siginfo_layout so that the structure > > > > can be properly decoded and made sense of. > > > > > > > > I am not seeing anything like that. > > > > > > Okay, this has to do with copying between the compat and non-compat > > > versions of the struct? Sure, I can update that, although the code > > > would be basically non-functional on arm64 because TBI isn't supported > > > on 32-bit ARM. > > > > > > > >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > > > >> > index cb3d6c267181..6dd82373eb2d 100644 > > > > >> > --- a/include/uapi/asm-generic/siginfo.h > > > > >> > +++ b/include/uapi/asm-generic/siginfo.h > > > > >> > @@ -91,6 +91,14 @@ union __sifields { > > > > >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > >> > __u32 _pkey; > > > > >> > } _addr_pkey; > > > > >> > +#ifdef __aarch64__ > > > > >> > + /* used with all si_codes */ > > > > >> > + struct { > > > > >> > + short _dummy_top_byte; > > > > > > > > > > ^ What's this for? I don't have Eric's insight here. > > > > > > We would need a short's worth of padding in order to prevent the > > > fields from occupying the same address as si_addr_lsb. > > > > > > > > > > > > >> > + unsigned char _top_byte; > > > > >> > + unsigned char _top_byte_mask; > > > > >> > + } _addr_top_byte; > > > > >> > +#endif > > > > >> > }; > > > > >> > } _sigfault; > > > > >> > > > > > >> > > > > >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > > > > >> > > > > >> Please remove the "#ifdef __aarch64__". If at all possible we want to > > > > >> design this so any other architecture who has this challenge can use the > > > > >> code. The kind of code does not get enough attention/maintenance if it > > > > >> is built for a single architecture. > > > > > > Seems reasonable. I was recently made aware that RISC-V was > > > considering a similar feature: > > > https://lists.riscv.org/g/tech-tee/topic/risc_v_tbi_proposal/72855478 > > > I would have opted to expand this to other architectures on an > > > as-needed basis, but I'd also be fine with having it on all > > > architectures from the start. > > > > > > If we make this arch-independent, we have an additional concern, which > > > is "what if some future architecture wants more than one byte here?" > > > For example, an architecture may have a "top-two-bytes-ignore" > > > feature, which would imply two-byte (misnamed) "si_addr_top_byte" and > > > "si_addr_top_byte_mask" fields. And the RISC-V proposal potentially > > > implies many more ignored bits (see slide 13 of the presentation). The > > > maximum size that these fields can possibly be is the size of a > > > pointer, and with that there wouldn't be enough room in the padding at > > > this point to accommodate the new fields. > > > > > > That basically implies your earlier suggestion of adding a union > > > member here to accommodate future expansion of the union, and adding > > > the new fields after the union. I'm happy to make that change, with > > > the fields renamed "si_addr_ignored" and "si_addr_ignored_mask". > > > > I think what we need here is basically a flags word. > > > > So long as we keep a flag spare to indicate the existence of a further > > flags word, we can extend as needed. > > > > How the existence of the first flags words is detected is another > > problem. If it only applies for newly-defined si_code values, then > > I guess si_code may be sufficient. > > Existing kernels will zero-initialize unused regions of the siginfo > data structure. The zero-initialization of the padding at the end of > the struct is done by the clear_user call here: > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/kernel/signal.c#L3193 > > and the zero-initialization of the padding between fields and unused > union members is done by the clear_siginfo function which the kernel > calls when initializing the data structure: > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/include/linux/signal.h#L20 > > Therefore, a flag word value of 0 may be used to detect a lack of > support for flagged fields. It's not enough that we do this today. We would have had to do it back to the dawn of time (though in the arm64 case I guess we just need to go back to when the arch/arm64 was merged). v2.6.12:kernel/signal.c:copy_siginfo_to_user() suggests that this wasn't always the case, so unused parts of siginfo could be full of old junk from the user stack, if the kernel is sufficiently old. If we're trying to do something generic that makes sense on all arches, this matters. I may have misunderstood something about the code though. > That being said, in this particular case, we do not need a flag word. > We can just take advantage of this zero-initialization behavior in > existing kernels to set si_addr_ignored_mask to 0, which indicates > that none of the bits in si_addr_ignored are valid. [...] Cheers ---Dave
On Wed, Jun 24, 2020 at 10:12 AM Dave Martin <Dave.Martin@arm.com> wrote: > > On Wed, Jun 24, 2020 at 09:51:49AM -0700, Peter Collingbourne wrote: > > On Wed, Jun 24, 2020 at 2:28 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > On Tue, Jun 23, 2020 at 05:40:08PM -0700, Peter Collingbourne wrote: > > > > On Tue, Jun 23, 2020 at 10:52 AM Eric W. Biederman > > > > <ebiederm@xmission.com> wrote: > > > > > > > > > > Dave Martin <Dave.Martin@arm.com> writes: > > > > > > > > > > > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > > > > > >> Peter Collingbourne <pcc@google.com> writes: > > > > > >> > > > > > >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > > > > > >> > index 47f651df781c..a8380a2b6361 100644 > > > > > >> > --- a/arch/arm64/kernel/traps.c > > > > > >> > +++ b/arch/arm64/kernel/traps.c > > > > > >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > > > > > >> > } > > > > > >> > > > > > > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, > > > > > >> > + unsigned long far, unsigned char far_tb_mask, > > > > > >> > const char *str) > > > > > >> > { > > > > > >> > arm64_show_signal(signo, str); > > > > > >> > - if (signo == SIGKILL) > > > > > >> > + if (signo == SIGKILL) { > > > > > >> > force_sig(SIGKILL); > > > > > >> > - else > > > > > >> > - force_sig_fault(signo, code, addr); > > > > > >> > + } else { > > > > > >> > + struct kernel_siginfo info; > > > > > >> > + clear_siginfo(&info); > > > > > >> > + info.si_signo = signo; > > > > > >> > + info.si_errno = 0; > > > > > >> > + info.si_code = code; > > > > > >> > + info.si_addr = addr; > > > > > >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > > > > > >> > + info.si_addr_top_byte_mask = far_tb_mask; > > > > > >> > + force_sig_info(&info); > > > > > >> > + } > > > > > >> > } > > > > > >> > > > > > > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > > > > > >> > - const char *str) > > > > > >> > + unsigned long far, const char *str) > > > > > >> > { > > > > > >> > + struct kernel_siginfo info; > > > > > >> > + > > > > > >> > arm64_show_signal(SIGBUS, str); > > > > > >> > - force_sig_mceerr(code, addr, lsb); > > > > > >> > + > > > > > >> > + clear_siginfo(&info); > > > > > >> > + info.si_signo = SIGBUS; > > > > > >> > + info.si_errno = 0; > > > > > >> > + info.si_code = code; > > > > > >> > + info.si_addr = addr; > > > > > >> > + info.si_addr_lsb = lsb; > > > > > >> > + info.si_addr_top_byte = far >> 56; > > > > > >> > + info.si_addr_top_byte_mask = 0xff; > > > > > >> > + force_sig_info(&info); > > > > > >> > } > > > > > >> > > > > > >> I have a real problem with this construction. force_sig_info is not an > > > > > >> interface that should be used for anything except to define a wrapper > > > > > >> that takes it's parameters. > > > > > > > > > > > > Can you elaborate? How would you do this king of thing. > > > > > > > > > > There are no other uses of force_sig_info in architecture code. > > > > > > > > > > I just removed them _all_ because they were almost all broken. > > > > > In fact your mcerr case is broken because it uses two different > > > > > union members simultantiously. > > > > > > > > Is that really broken? I thought that the Linux kernel deliberately > > > > didn't care about strict aliasing rules (the top-level Makefile passes > > > > -fno-strict-aliasing) so I thought that it was valid in "Linux kernel > > > > C" even though from a standards point of view it is invalid. (That > > > > being said, this is probably moot with my proposed changes below > > > > though.) > > > > > > I have a feeling that -fno-strict-aliasing only allows you to _read_ a > > > different union member from the one previously written. > > > > > > Writing a different member from the last one written can still splatter > > > on the other members IIUC. > > > > > > It would be better to keep things separate rather than risk > > > incorrectness just to save a few bytes. > > > > > > IMHO -fno-strict-aliasing is no excuse for gratuitous type-punning. > > > > > > > > So I am looking for something like force_sig_mcerr or force_sig_fault > > > > > that includes your new information that then calls force_sig_info. > > > > > > > > > > I know of no other way to safely use the siginfo struct. > > > > > > > > So you want something like: > > > > > > > > int force_sig_fault_with_ignored_bits(int signo, int code, void __user > > > > *addr, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > int force_sig_mceerr_with_ignored_bits(int code, void __user *addr, > > > > short lsb, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > in kernel/signal.c and the code in arch/arm64 would call that? > > > > > > > > > > AIUI we absolutely need a forced signal here, we need to supply > > > > > > metadata, and we don't have to open-code all that at every relevant > > > > > > signal generation site... > > > > > > > > > > > >> It is not clear to me that if you have adapted siginfo_layout. > > > > > > > > > > > > Garbled sentence? > > > > > > > > > > Looks like. One of the pieces of code that needs to change > > > > > when siginfo gets updated is siginfo_layout so that the structure > > > > > can be properly decoded and made sense of. > > > > > > > > > > I am not seeing anything like that. > > > > > > > > Okay, this has to do with copying between the compat and non-compat > > > > versions of the struct? Sure, I can update that, although the code > > > > would be basically non-functional on arm64 because TBI isn't supported > > > > on 32-bit ARM. > > > > > > > > > >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > > > > >> > index cb3d6c267181..6dd82373eb2d 100644 > > > > > >> > --- a/include/uapi/asm-generic/siginfo.h > > > > > >> > +++ b/include/uapi/asm-generic/siginfo.h > > > > > >> > @@ -91,6 +91,14 @@ union __sifields { > > > > > >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > > >> > __u32 _pkey; > > > > > >> > } _addr_pkey; > > > > > >> > +#ifdef __aarch64__ > > > > > >> > + /* used with all si_codes */ > > > > > >> > + struct { > > > > > >> > + short _dummy_top_byte; > > > > > > > > > > > > ^ What's this for? I don't have Eric's insight here. > > > > > > > > We would need a short's worth of padding in order to prevent the > > > > fields from occupying the same address as si_addr_lsb. > > > > > > > > > > > > > > > >> > + unsigned char _top_byte; > > > > > >> > + unsigned char _top_byte_mask; > > > > > >> > + } _addr_top_byte; > > > > > >> > +#endif > > > > > >> > }; > > > > > >> > } _sigfault; > > > > > >> > > > > > > >> > > > > > >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > > > > > >> > > > > > >> Please remove the "#ifdef __aarch64__". If at all possible we want to > > > > > >> design this so any other architecture who has this challenge can use the > > > > > >> code. The kind of code does not get enough attention/maintenance if it > > > > > >> is built for a single architecture. > > > > > > > > Seems reasonable. I was recently made aware that RISC-V was > > > > considering a similar feature: > > > > https://lists.riscv.org/g/tech-tee/topic/risc_v_tbi_proposal/72855478 > > > > I would have opted to expand this to other architectures on an > > > > as-needed basis, but I'd also be fine with having it on all > > > > architectures from the start. > > > > > > > > If we make this arch-independent, we have an additional concern, which > > > > is "what if some future architecture wants more than one byte here?" > > > > For example, an architecture may have a "top-two-bytes-ignore" > > > > feature, which would imply two-byte (misnamed) "si_addr_top_byte" and > > > > "si_addr_top_byte_mask" fields. And the RISC-V proposal potentially > > > > implies many more ignored bits (see slide 13 of the presentation). The > > > > maximum size that these fields can possibly be is the size of a > > > > pointer, and with that there wouldn't be enough room in the padding at > > > > this point to accommodate the new fields. > > > > > > > > That basically implies your earlier suggestion of adding a union > > > > member here to accommodate future expansion of the union, and adding > > > > the new fields after the union. I'm happy to make that change, with > > > > the fields renamed "si_addr_ignored" and "si_addr_ignored_mask". > > > > > > I think what we need here is basically a flags word. > > > > > > So long as we keep a flag spare to indicate the existence of a further > > > flags word, we can extend as needed. > > > > > > How the existence of the first flags words is detected is another > > > problem. If it only applies for newly-defined si_code values, then > > > I guess si_code may be sufficient. > > > > Existing kernels will zero-initialize unused regions of the siginfo > > data structure. The zero-initialization of the padding at the end of > > the struct is done by the clear_user call here: > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/kernel/signal.c#L3193 > > > > and the zero-initialization of the padding between fields and unused > > union members is done by the clear_siginfo function which the kernel > > calls when initializing the data structure: > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/include/linux/signal.h#L20 > > > > Therefore, a flag word value of 0 may be used to detect a lack of > > support for flagged fields. > > It's not enough that we do this today. We would have had to do it back > to the dawn of time (though in the arm64 case I guess we just need to go > back to when the arch/arm64 was merged). > > v2.6.12:kernel/signal.c:copy_siginfo_to_user() suggests that this wasn't > always the case, so unused parts of siginfo could be full of old junk > from the user stack, if the kernel is sufficiently old. > > If we're trying to do something generic that makes sense on all arches, > this matters. I may have misunderstood something about the code though. Hmm, I think you're right. The current behavior was introduced by commit c999b933faa5e281e3af2e110eccaf91698b0a81 which was first released in 4.18. So if an application wants to be compatible with pre-4.18 kernels then there would need to be some other way to indicate that the fields are valid. Probably the simplest way would be to have the application issue a uname(2) syscall and check the kernel version before reading these fields. I have a couple of other ideas that don't rely on version detection, if we'd prefer to avoid that. (They are somewhat ugly, but our hand is forced by backwards compatibility.) One idea is to re-purpose the si_errno field as a flags field for certain signal numbers. I checked a few kernel releases going back to 2.6.18 and it looks like the field is set to 0 except in the following circumstances: - sending a hardware breakpoint (SIGTRAP/TRAP_HWBKPT) - seccomp failures (SIGSYS/SYS_SECCOMP) - user-defined signal via kill_pid_usb_asyncio - SIGSWI in 3.18 and before (code since removed) It is also set to EFAULT for certain SIGSEGV/SEGV_MAPERR signals on powerpc since commit c96c4436aba4c12f1f48369f2f90bc43e12fe36c, which is currently unreleased. So if we wanted to go this route for SIGSEGV we would need to stop the kernel from setting si_errno to EFAULT for this signal before the 5.8 release. Another idea was to have userspace set a flag in sa_flags when registering a signal handler meaning "this signal handler requires unknown siginfo fields to be zeroed", and have existing kernels reject the syscall due to an unknown flag being set, but unfortunately this won't work because existing kernels do not reject sigaction syscalls with unknown flags set in sa_flags. A perhaps more radical idea in this vein would be to claim some of the upper bits of the signal number as flags that will cause the syscall to be rejected if set and unknown to the kernel. Existing kernels (going back to at least 2.6.18) contain this code in do_sigaction: if (!valid_signal(sig) || sig < 1 || (act && sig_kernel_only(sig))) return -EINVAL; and vald_signal is defined as: static inline int valid_signal(unsigned long sig) { return sig <= _NSIG ? 1 : 0; } All architectures define _NSIG as a value <= 128, so they will reject a signal number with any of bits 8-31 set. This means that we can use any of those bits for mandatory flags. Most likely we could use bit 30 (expanding down as necessary), as it keeps the signal number positive and permits future expansion of the signal number range. Peter
On Wed, Jun 24, 2020 at 12:51:43PM -0700, Peter Collingbourne wrote: > On Wed, Jun 24, 2020 at 10:12 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > On Wed, Jun 24, 2020 at 09:51:49AM -0700, Peter Collingbourne wrote: > > > On Wed, Jun 24, 2020 at 2:28 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > On Tue, Jun 23, 2020 at 05:40:08PM -0700, Peter Collingbourne wrote: > > > > > On Tue, Jun 23, 2020 at 10:52 AM Eric W. Biederman > > > > > <ebiederm@xmission.com> wrote: > > > > > > > > > > > > Dave Martin <Dave.Martin@arm.com> writes: > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > > > > > > >> Peter Collingbourne <pcc@google.com> writes: > > > > > > >> > > > > > > >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > > > > > > >> > index 47f651df781c..a8380a2b6361 100644 > > > > > > >> > --- a/arch/arm64/kernel/traps.c > > > > > > >> > +++ b/arch/arm64/kernel/traps.c > > > > > > >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > > > > > > >> > } > > > > > > >> > > > > > > > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, > > > > > > >> > + unsigned long far, unsigned char far_tb_mask, > > > > > > >> > const char *str) > > > > > > >> > { > > > > > > >> > arm64_show_signal(signo, str); > > > > > > >> > - if (signo == SIGKILL) > > > > > > >> > + if (signo == SIGKILL) { > > > > > > >> > force_sig(SIGKILL); > > > > > > >> > - else > > > > > > >> > - force_sig_fault(signo, code, addr); > > > > > > >> > + } else { > > > > > > >> > + struct kernel_siginfo info; > > > > > > >> > + clear_siginfo(&info); > > > > > > >> > + info.si_signo = signo; > > > > > > >> > + info.si_errno = 0; > > > > > > >> > + info.si_code = code; > > > > > > >> > + info.si_addr = addr; > > > > > > >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > > > > > > >> > + info.si_addr_top_byte_mask = far_tb_mask; > > > > > > >> > + force_sig_info(&info); > > > > > > >> > + } > > > > > > >> > } > > > > > > >> > > > > > > > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > > > > > > >> > - const char *str) > > > > > > >> > + unsigned long far, const char *str) > > > > > > >> > { > > > > > > >> > + struct kernel_siginfo info; > > > > > > >> > + > > > > > > >> > arm64_show_signal(SIGBUS, str); > > > > > > >> > - force_sig_mceerr(code, addr, lsb); > > > > > > >> > + > > > > > > >> > + clear_siginfo(&info); > > > > > > >> > + info.si_signo = SIGBUS; > > > > > > >> > + info.si_errno = 0; > > > > > > >> > + info.si_code = code; > > > > > > >> > + info.si_addr = addr; > > > > > > >> > + info.si_addr_lsb = lsb; > > > > > > >> > + info.si_addr_top_byte = far >> 56; > > > > > > >> > + info.si_addr_top_byte_mask = 0xff; > > > > > > >> > + force_sig_info(&info); > > > > > > >> > } > > > > > > >> > > > > > > >> I have a real problem with this construction. force_sig_info is not an > > > > > > >> interface that should be used for anything except to define a wrapper > > > > > > >> that takes it's parameters. > > > > > > > > > > > > > > Can you elaborate? How would you do this king of thing. > > > > > > > > > > > > There are no other uses of force_sig_info in architecture code. > > > > > > > > > > > > I just removed them _all_ because they were almost all broken. > > > > > > In fact your mcerr case is broken because it uses two different > > > > > > union members simultantiously. > > > > > > > > > > Is that really broken? I thought that the Linux kernel deliberately > > > > > didn't care about strict aliasing rules (the top-level Makefile passes > > > > > -fno-strict-aliasing) so I thought that it was valid in "Linux kernel > > > > > C" even though from a standards point of view it is invalid. (That > > > > > being said, this is probably moot with my proposed changes below > > > > > though.) > > > > > > > > I have a feeling that -fno-strict-aliasing only allows you to _read_ a > > > > different union member from the one previously written. > > > > > > > > Writing a different member from the last one written can still splatter > > > > on the other members IIUC. > > > > > > > > It would be better to keep things separate rather than risk > > > > incorrectness just to save a few bytes. > > > > > > > > IMHO -fno-strict-aliasing is no excuse for gratuitous type-punning. > > > > > > > > > > So I am looking for something like force_sig_mcerr or force_sig_fault > > > > > > that includes your new information that then calls force_sig_info. > > > > > > > > > > > > I know of no other way to safely use the siginfo struct. > > > > > > > > > > So you want something like: > > > > > > > > > > int force_sig_fault_with_ignored_bits(int signo, int code, void __user > > > > > *addr, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > int force_sig_mceerr_with_ignored_bits(int code, void __user *addr, > > > > > short lsb, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > in kernel/signal.c and the code in arch/arm64 would call that? > > > > > > > > > > > > AIUI we absolutely need a forced signal here, we need to supply > > > > > > > metadata, and we don't have to open-code all that at every relevant > > > > > > > signal generation site... > > > > > > > > > > > > > >> It is not clear to me that if you have adapted siginfo_layout. > > > > > > > > > > > > > > Garbled sentence? > > > > > > > > > > > > Looks like. One of the pieces of code that needs to change > > > > > > when siginfo gets updated is siginfo_layout so that the structure > > > > > > can be properly decoded and made sense of. > > > > > > > > > > > > I am not seeing anything like that. > > > > > > > > > > Okay, this has to do with copying between the compat and non-compat > > > > > versions of the struct? Sure, I can update that, although the code > > > > > would be basically non-functional on arm64 because TBI isn't supported > > > > > on 32-bit ARM. > > > > > > > > > > > >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > > > > > >> > index cb3d6c267181..6dd82373eb2d 100644 > > > > > > >> > --- a/include/uapi/asm-generic/siginfo.h > > > > > > >> > +++ b/include/uapi/asm-generic/siginfo.h > > > > > > >> > @@ -91,6 +91,14 @@ union __sifields { > > > > > > >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > > > >> > __u32 _pkey; > > > > > > >> > } _addr_pkey; > > > > > > >> > +#ifdef __aarch64__ > > > > > > >> > + /* used with all si_codes */ > > > > > > >> > + struct { > > > > > > >> > + short _dummy_top_byte; > > > > > > > > > > > > > > ^ What's this for? I don't have Eric's insight here. > > > > > > > > > > We would need a short's worth of padding in order to prevent the > > > > > fields from occupying the same address as si_addr_lsb. > > > > > > > > > > > > > > > > > > >> > + unsigned char _top_byte; > > > > > > >> > + unsigned char _top_byte_mask; > > > > > > >> > + } _addr_top_byte; > > > > > > >> > +#endif > > > > > > >> > }; > > > > > > >> > } _sigfault; > > > > > > >> > > > > > > > >> > > > > > > >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > > > > > > >> > > > > > > >> Please remove the "#ifdef __aarch64__". If at all possible we want to > > > > > > >> design this so any other architecture who has this challenge can use the > > > > > > >> code. The kind of code does not get enough attention/maintenance if it > > > > > > >> is built for a single architecture. > > > > > > > > > > Seems reasonable. I was recently made aware that RISC-V was > > > > > considering a similar feature: > > > > > https://lists.riscv.org/g/tech-tee/topic/risc_v_tbi_proposal/72855478 > > > > > I would have opted to expand this to other architectures on an > > > > > as-needed basis, but I'd also be fine with having it on all > > > > > architectures from the start. > > > > > > > > > > If we make this arch-independent, we have an additional concern, which > > > > > is "what if some future architecture wants more than one byte here?" > > > > > For example, an architecture may have a "top-two-bytes-ignore" > > > > > feature, which would imply two-byte (misnamed) "si_addr_top_byte" and > > > > > "si_addr_top_byte_mask" fields. And the RISC-V proposal potentially > > > > > implies many more ignored bits (see slide 13 of the presentation). The > > > > > maximum size that these fields can possibly be is the size of a > > > > > pointer, and with that there wouldn't be enough room in the padding at > > > > > this point to accommodate the new fields. > > > > > > > > > > That basically implies your earlier suggestion of adding a union > > > > > member here to accommodate future expansion of the union, and adding > > > > > the new fields after the union. I'm happy to make that change, with > > > > > the fields renamed "si_addr_ignored" and "si_addr_ignored_mask". > > > > > > > > I think what we need here is basically a flags word. > > > > > > > > So long as we keep a flag spare to indicate the existence of a further > > > > flags word, we can extend as needed. > > > > > > > > How the existence of the first flags words is detected is another > > > > problem. If it only applies for newly-defined si_code values, then > > > > I guess si_code may be sufficient. > > > > > > Existing kernels will zero-initialize unused regions of the siginfo > > > data structure. The zero-initialization of the padding at the end of > > > the struct is done by the clear_user call here: > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/kernel/signal.c#L3193 > > > > > > and the zero-initialization of the padding between fields and unused > > > union members is done by the clear_siginfo function which the kernel > > > calls when initializing the data structure: > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/include/linux/signal.h#L20 > > > > > > Therefore, a flag word value of 0 may be used to detect a lack of > > > support for flagged fields. > > > > It's not enough that we do this today. We would have had to do it back > > to the dawn of time (though in the arm64 case I guess we just need to go > > back to when the arch/arm64 was merged). > > > > v2.6.12:kernel/signal.c:copy_siginfo_to_user() suggests that this wasn't > > always the case, so unused parts of siginfo could be full of old junk > > from the user stack, if the kernel is sufficiently old. > > > > If we're trying to do something generic that makes sense on all arches, > > this matters. I may have misunderstood something about the code though. > > Hmm, I think you're right. The current behavior was introduced by > commit c999b933faa5e281e3af2e110eccaf91698b0a81 which was first > released in 4.18. So if an application wants to be compatible with > pre-4.18 kernels then there would need to be some other way to > indicate that the fields are valid. Probably the simplest way would be > to have the application issue a uname(2) syscall and check the kernel > version before reading these fields. I have a couple of other ideas > that don't rely on version detection, if we'd prefer to avoid that. > (They are somewhat ugly, but our hand is forced by backwards > compatibility.) > > One idea is to re-purpose the si_errno field as a flags field for > certain signal numbers. I checked a few kernel releases going back to > 2.6.18 and it looks like the field is set to 0 except in the following > circumstances: > - sending a hardware breakpoint (SIGTRAP/TRAP_HWBKPT) > - seccomp failures (SIGSYS/SYS_SECCOMP) > - user-defined signal via kill_pid_usb_asyncio > - SIGSWI in 3.18 and before (code since removed) > > It is also set to EFAULT for certain SIGSEGV/SEGV_MAPERR signals on > powerpc since commit c96c4436aba4c12f1f48369f2f90bc43e12fe36c, which > is currently unreleased. So if we wanted to go this route for SIGSEGV > we would need to stop the kernel from setting si_errno to EFAULT for > this signal before the 5.8 release. > > Another idea was to have userspace set a flag in sa_flags when > registering a signal handler meaning "this signal handler requires > unknown siginfo fields to be zeroed", and have existing kernels reject > the syscall due to an unknown flag being set, but unfortunately this > won't work because existing kernels do not reject sigaction syscalls > with unknown flags set in sa_flags. A perhaps more radical idea in > this vein would be to claim some of the upper bits of the signal > number as flags that will cause the syscall to be rejected if set and > unknown to the kernel. Existing kernels (going back to at least > 2.6.18) contain this code in do_sigaction: > > if (!valid_signal(sig) || sig < 1 || (act && sig_kernel_only(sig))) > return -EINVAL; > > and vald_signal is defined as: > > static inline int valid_signal(unsigned long sig) > { > return sig <= _NSIG ? 1 : 0; > } > > All architectures define _NSIG as a value <= 128, so they will reject > a signal number with any of bits 8-31 set. This means that we can use > any of those bits for mandatory flags. Most likely we could use bit 30 > (expanding down as necessary), as it keeps the signal number positive > and permits future expansion of the signal number range. Does the signal core code actually gurantee to zero the unused fields? Unless the fields are poked in by hand this is fraught with subtlelies, especially when unions are involved. (I'm sure the code tries to do it, but I've not eyeballed it in detail...) Using unused bits in the signal number to turn on new functionality feels risky. As currently specified, this is just a number. Since today a successful sigaction(n ...) guarantees that n is a valid signal number, reasonable code like the following would trigger a buffer overrun if we start trying to encode anything else in there: struct sigaction actions[NSIG]; int do_something( ... ) { ... if (!sigaction(n, sa, ...)) { actions[n] = *sa; return 0; } ... } I think it would be cleaner for to add a single flag field that can be used for detecting other extensions, and request it via a new sa_flags bit. This removes the need for sematically useless zeroing of unused fields (though for hygiene and backwards compatibility reasons we would probably want to carry on zeroing them anyway). I can see no simpler way to add supplementary siginfo fields for existing si_codes. For si_codes that didn't exist before the zeroing came in we could still detect optional si_code-specific fields via zeroing, but it seems messary to have two ways of detecting extensions. Cheers ---Dave
On Mon, Jul 6, 2020 at 9:41 AM Dave Martin <Dave.Martin@arm.com> wrote: > > On Wed, Jun 24, 2020 at 12:51:43PM -0700, Peter Collingbourne wrote: > > On Wed, Jun 24, 2020 at 10:12 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > On Wed, Jun 24, 2020 at 09:51:49AM -0700, Peter Collingbourne wrote: > > > > On Wed, Jun 24, 2020 at 2:28 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > On Tue, Jun 23, 2020 at 05:40:08PM -0700, Peter Collingbourne wrote: > > > > > > On Tue, Jun 23, 2020 at 10:52 AM Eric W. Biederman > > > > > > <ebiederm@xmission.com> wrote: > > > > > > > > > > > > > > Dave Martin <Dave.Martin@arm.com> writes: > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > > > > > > > >> Peter Collingbourne <pcc@google.com> writes: > > > > > > > >> > > > > > > > >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > > > > > > > >> > index 47f651df781c..a8380a2b6361 100644 > > > > > > > >> > --- a/arch/arm64/kernel/traps.c > > > > > > > >> > +++ b/arch/arm64/kernel/traps.c > > > > > > > >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > > > > > > > >> > } > > > > > > > >> > > > > > > > > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, > > > > > > > >> > + unsigned long far, unsigned char far_tb_mask, > > > > > > > >> > const char *str) > > > > > > > >> > { > > > > > > > >> > arm64_show_signal(signo, str); > > > > > > > >> > - if (signo == SIGKILL) > > > > > > > >> > + if (signo == SIGKILL) { > > > > > > > >> > force_sig(SIGKILL); > > > > > > > >> > - else > > > > > > > >> > - force_sig_fault(signo, code, addr); > > > > > > > >> > + } else { > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > >> > + clear_siginfo(&info); > > > > > > > >> > + info.si_signo = signo; > > > > > > > >> > + info.si_errno = 0; > > > > > > > >> > + info.si_code = code; > > > > > > > >> > + info.si_addr = addr; > > > > > > > >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > > > > > > > >> > + info.si_addr_top_byte_mask = far_tb_mask; > > > > > > > >> > + force_sig_info(&info); > > > > > > > >> > + } > > > > > > > >> > } > > > > > > > >> > > > > > > > > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > > > > > > > >> > - const char *str) > > > > > > > >> > + unsigned long far, const char *str) > > > > > > > >> > { > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > >> > + > > > > > > > >> > arm64_show_signal(SIGBUS, str); > > > > > > > >> > - force_sig_mceerr(code, addr, lsb); > > > > > > > >> > + > > > > > > > >> > + clear_siginfo(&info); > > > > > > > >> > + info.si_signo = SIGBUS; > > > > > > > >> > + info.si_errno = 0; > > > > > > > >> > + info.si_code = code; > > > > > > > >> > + info.si_addr = addr; > > > > > > > >> > + info.si_addr_lsb = lsb; > > > > > > > >> > + info.si_addr_top_byte = far >> 56; > > > > > > > >> > + info.si_addr_top_byte_mask = 0xff; > > > > > > > >> > + force_sig_info(&info); > > > > > > > >> > } > > > > > > > >> > > > > > > > >> I have a real problem with this construction. force_sig_info is not an > > > > > > > >> interface that should be used for anything except to define a wrapper > > > > > > > >> that takes it's parameters. > > > > > > > > > > > > > > > > Can you elaborate? How would you do this king of thing. > > > > > > > > > > > > > > There are no other uses of force_sig_info in architecture code. > > > > > > > > > > > > > > I just removed them _all_ because they were almost all broken. > > > > > > > In fact your mcerr case is broken because it uses two different > > > > > > > union members simultantiously. > > > > > > > > > > > > Is that really broken? I thought that the Linux kernel deliberately > > > > > > didn't care about strict aliasing rules (the top-level Makefile passes > > > > > > -fno-strict-aliasing) so I thought that it was valid in "Linux kernel > > > > > > C" even though from a standards point of view it is invalid. (That > > > > > > being said, this is probably moot with my proposed changes below > > > > > > though.) > > > > > > > > > > I have a feeling that -fno-strict-aliasing only allows you to _read_ a > > > > > different union member from the one previously written. > > > > > > > > > > Writing a different member from the last one written can still splatter > > > > > on the other members IIUC. > > > > > > > > > > It would be better to keep things separate rather than risk > > > > > incorrectness just to save a few bytes. > > > > > > > > > > IMHO -fno-strict-aliasing is no excuse for gratuitous type-punning. > > > > > > > > > > > > So I am looking for something like force_sig_mcerr or force_sig_fault > > > > > > > that includes your new information that then calls force_sig_info. > > > > > > > > > > > > > > I know of no other way to safely use the siginfo struct. > > > > > > > > > > > > So you want something like: > > > > > > > > > > > > int force_sig_fault_with_ignored_bits(int signo, int code, void __user > > > > > > *addr, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > int force_sig_mceerr_with_ignored_bits(int code, void __user *addr, > > > > > > short lsb, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > > > in kernel/signal.c and the code in arch/arm64 would call that? > > > > > > > > > > > > > > AIUI we absolutely need a forced signal here, we need to supply > > > > > > > > metadata, and we don't have to open-code all that at every relevant > > > > > > > > signal generation site... > > > > > > > > > > > > > > > >> It is not clear to me that if you have adapted siginfo_layout. > > > > > > > > > > > > > > > > Garbled sentence? > > > > > > > > > > > > > > Looks like. One of the pieces of code that needs to change > > > > > > > when siginfo gets updated is siginfo_layout so that the structure > > > > > > > can be properly decoded and made sense of. > > > > > > > > > > > > > > I am not seeing anything like that. > > > > > > > > > > > > Okay, this has to do with copying between the compat and non-compat > > > > > > versions of the struct? Sure, I can update that, although the code > > > > > > would be basically non-functional on arm64 because TBI isn't supported > > > > > > on 32-bit ARM. > > > > > > > > > > > > > >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > > > > > > >> > index cb3d6c267181..6dd82373eb2d 100644 > > > > > > > >> > --- a/include/uapi/asm-generic/siginfo.h > > > > > > > >> > +++ b/include/uapi/asm-generic/siginfo.h > > > > > > > >> > @@ -91,6 +91,14 @@ union __sifields { > > > > > > > >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > > > > >> > __u32 _pkey; > > > > > > > >> > } _addr_pkey; > > > > > > > >> > +#ifdef __aarch64__ > > > > > > > >> > + /* used with all si_codes */ > > > > > > > >> > + struct { > > > > > > > >> > + short _dummy_top_byte; > > > > > > > > > > > > > > > > ^ What's this for? I don't have Eric's insight here. > > > > > > > > > > > > We would need a short's worth of padding in order to prevent the > > > > > > fields from occupying the same address as si_addr_lsb. > > > > > > > > > > > > > > > > > > > > > >> > + unsigned char _top_byte; > > > > > > > >> > + unsigned char _top_byte_mask; > > > > > > > >> > + } _addr_top_byte; > > > > > > > >> > +#endif > > > > > > > >> > }; > > > > > > > >> > } _sigfault; > > > > > > > >> > > > > > > > > >> > > > > > > > >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > > > > > > > >> > > > > > > > >> Please remove the "#ifdef __aarch64__". If at all possible we want to > > > > > > > >> design this so any other architecture who has this challenge can use the > > > > > > > >> code. The kind of code does not get enough attention/maintenance if it > > > > > > > >> is built for a single architecture. > > > > > > > > > > > > Seems reasonable. I was recently made aware that RISC-V was > > > > > > considering a similar feature: > > > > > > https://lists.riscv.org/g/tech-tee/topic/risc_v_tbi_proposal/72855478 > > > > > > I would have opted to expand this to other architectures on an > > > > > > as-needed basis, but I'd also be fine with having it on all > > > > > > architectures from the start. > > > > > > > > > > > > If we make this arch-independent, we have an additional concern, which > > > > > > is "what if some future architecture wants more than one byte here?" > > > > > > For example, an architecture may have a "top-two-bytes-ignore" > > > > > > feature, which would imply two-byte (misnamed) "si_addr_top_byte" and > > > > > > "si_addr_top_byte_mask" fields. And the RISC-V proposal potentially > > > > > > implies many more ignored bits (see slide 13 of the presentation). The > > > > > > maximum size that these fields can possibly be is the size of a > > > > > > pointer, and with that there wouldn't be enough room in the padding at > > > > > > this point to accommodate the new fields. > > > > > > > > > > > > That basically implies your earlier suggestion of adding a union > > > > > > member here to accommodate future expansion of the union, and adding > > > > > > the new fields after the union. I'm happy to make that change, with > > > > > > the fields renamed "si_addr_ignored" and "si_addr_ignored_mask". > > > > > > > > > > I think what we need here is basically a flags word. > > > > > > > > > > So long as we keep a flag spare to indicate the existence of a further > > > > > flags word, we can extend as needed. > > > > > > > > > > How the existence of the first flags words is detected is another > > > > > problem. If it only applies for newly-defined si_code values, then > > > > > I guess si_code may be sufficient. > > > > > > > > Existing kernels will zero-initialize unused regions of the siginfo > > > > data structure. The zero-initialization of the padding at the end of > > > > the struct is done by the clear_user call here: > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/kernel/signal.c#L3193 > > > > > > > > and the zero-initialization of the padding between fields and unused > > > > union members is done by the clear_siginfo function which the kernel > > > > calls when initializing the data structure: > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/include/linux/signal.h#L20 > > > > > > > > Therefore, a flag word value of 0 may be used to detect a lack of > > > > support for flagged fields. > > > > > > It's not enough that we do this today. We would have had to do it back > > > to the dawn of time (though in the arm64 case I guess we just need to go > > > back to when the arch/arm64 was merged). > > > > > > v2.6.12:kernel/signal.c:copy_siginfo_to_user() suggests that this wasn't > > > always the case, so unused parts of siginfo could be full of old junk > > > from the user stack, if the kernel is sufficiently old. > > > > > > If we're trying to do something generic that makes sense on all arches, > > > this matters. I may have misunderstood something about the code though. > > > > Hmm, I think you're right. The current behavior was introduced by > > commit c999b933faa5e281e3af2e110eccaf91698b0a81 which was first > > released in 4.18. So if an application wants to be compatible with > > pre-4.18 kernels then there would need to be some other way to > > indicate that the fields are valid. Probably the simplest way would be > > to have the application issue a uname(2) syscall and check the kernel > > version before reading these fields. I have a couple of other ideas > > that don't rely on version detection, if we'd prefer to avoid that. > > (They are somewhat ugly, but our hand is forced by backwards > > compatibility.) > > > > One idea is to re-purpose the si_errno field as a flags field for > > certain signal numbers. I checked a few kernel releases going back to > > 2.6.18 and it looks like the field is set to 0 except in the following > > circumstances: > > - sending a hardware breakpoint (SIGTRAP/TRAP_HWBKPT) > > - seccomp failures (SIGSYS/SYS_SECCOMP) > > - user-defined signal via kill_pid_usb_asyncio > > - SIGSWI in 3.18 and before (code since removed) > > > > It is also set to EFAULT for certain SIGSEGV/SEGV_MAPERR signals on > > powerpc since commit c96c4436aba4c12f1f48369f2f90bc43e12fe36c, which > > is currently unreleased. So if we wanted to go this route for SIGSEGV > > we would need to stop the kernel from setting si_errno to EFAULT for > > this signal before the 5.8 release. > > > > Another idea was to have userspace set a flag in sa_flags when > > registering a signal handler meaning "this signal handler requires > > unknown siginfo fields to be zeroed", and have existing kernels reject > > the syscall due to an unknown flag being set, but unfortunately this > > won't work because existing kernels do not reject sigaction syscalls > > with unknown flags set in sa_flags. A perhaps more radical idea in > > this vein would be to claim some of the upper bits of the signal > > number as flags that will cause the syscall to be rejected if set and > > unknown to the kernel. Existing kernels (going back to at least > > 2.6.18) contain this code in do_sigaction: > > > > if (!valid_signal(sig) || sig < 1 || (act && sig_kernel_only(sig))) > > return -EINVAL; > > > > and vald_signal is defined as: > > > > static inline int valid_signal(unsigned long sig) > > { > > return sig <= _NSIG ? 1 : 0; > > } > > > > All architectures define _NSIG as a value <= 128, so they will reject > > a signal number with any of bits 8-31 set. This means that we can use > > any of those bits for mandatory flags. Most likely we could use bit 30 > > (expanding down as necessary), as it keeps the signal number positive > > and permits future expansion of the signal number range. > > Does the signal core code actually gurantee to zero the unused fields? > Unless the fields are poked in by hand this is fraught with subtlelies, > especially when unions are involved. (I'm sure the code tries to do it, > but I've not eyeballed it in detail...) It memsets the siginfo structure before setting the fields and sending the signal (grep for clear_siginfo which is just a memset; you should find a call before all callers of force_sig_info). Memset is the right approach here since unlike setting fields by hand it clears padding which could lead to information leaks from the kernel. IIUC this is the reason why Eric wants all of the signals to be raised via wrappers in kernel/signal.c instead of via force_sig_info directly (to make this aspect easier to audit). > Using unused bits in the signal number to turn on new functionality > feels risky. As currently specified, this is just a number. Since > today a successful sigaction(n ...) guarantees that n is a valid signal > number, reasonable code like the following would trigger a buffer > overrun if we start trying to encode anything else in there: > > struct sigaction actions[NSIG]; > > int do_something( ... ) > { > ... > > if (!sigaction(n, sa, ...)) { > actions[n] = *sa; > return 0; > } > > ... > } I imagine the bit in the signal number being set by the direct caller to sigaction, and we could specifically recommend that calling pattern. In that case, your "n" wouldn't have the bit set in it. It could only appear in newly-written code that doesn't follow our recommendations, and there are already plenty of much more likely ways to cause buffer overflows in C code that doesn't follow recommendations anyway. (And even if such a buffer overflow occurred, it would very likely be caught early in development by the MMU due to the magnitude of the number 1<<30.) > I think it would be cleaner for to add a single flag field that can be > used for detecting other extensions, and request it via a new sa_flags > bit. This removes the need for sematically useless zeroing of unused > fields (though for hygiene and backwards compatibility reasons we would > probably want to carry on zeroing them anyway). > > I can see no simpler way to add supplementary siginfo fields for > existing si_codes. For si_codes that didn't exist before the zeroing > came in we could still detect optional si_code-specific fields via > zeroing, but it seems messary to have two ways of detecting extensions. That would certainly be cleaner if it worked, but that would only be the case if old kernels rejected unknown bits in sa_flags, and unfortunately they don't. With the bit in the signal number, the "old kernels reject" behavior admits relatively straightforward usage code: void set_segv_handler(void) { struct sigaction sa; sa.sa_sigaction = handle_segv; sa.sa_flags = SA_SIGINFO; if (sigaction(SIGSEGV | SF_CLEAR_UNKNOWN_FIELDS, &sa, 0) < 0) { // succeeds in new kernels, fails in old kernels sa.sa_sigaction = clear_fields_and_handle_segv; if (sigaction(SIGSEGV, &sa, 0) < 0) // succeeds in old kernels perror("sigaction"); } } void clear_fields_and_handle_segv(int signum, siginfo_t *sa, void *ctx) { sa->si_future_field = 0; handle_segv(signum, sa, ctx); } void handle_segv(int signum, siginfo_t *sa, void *ctx) { // At this point, si_future_field will have the value 0 in old kernels and the kernel-supplied value in new kernels. } Imagine if we moved the flag SF_CLEAR_UNKNOWN_FIELDS from the signal number to sa_flags. In that case, the first sigaction would succeed in old kernels so handle_segv wouldn't know whether it can safely read from si_future_field. With the sa_flags approach, you would need kernel version number checking via uname before setting the flag in sa_flags, and at that point why even have the flag in sa_flags at all since you could just have the signal handler conditionally read from si_future_field based on the uname? Note that the same applies to a flag indicating the availability of a si_flags field in sigaction (just s/SF_CLEAR_UNKNOWN_FIELDS/SF_HAS_SI_FLAGS/ and s/si_future_field/si_flags/ in the usage code above). In terms of SF_CLEAR_UNKNOWN_FIELDS versus SF_HAS_SI_FLAGS I'd be fine either way. Another thought that occurred to me is that we may consider generalizing this a step further and introducing a single flag bit in the signal number that means "reject unknown flags in sa_flags". This would mean that we wouldn't need to add any more flag bits to the signal number in the future, thus limiting this signal number hack to a single bit; all future mandatory behavior changes could just be put behind a flag in sa_flags and userspace code would easily be able to detect missing support for a flag and fall back if necessary. In our case, this would imply usage code like this: void set_segv_handler(void) { struct sigaction sa; sa.sa_sigaction = handle_segv; sa.sa_flags = SA_SIGINFO | SA_CLEAR_UNKNOWN_FIELDS; // Succeeds in kernels with SA_CLEAR_UNKNOWN_FIELDS support. // Fails in kernels with SF_CHECK_SA_FLAGS support but no SA_CLEAR_UNKNOWN_FIELDS support (because of the unknown flags check). // Fails in kernels without SF_CHECK_SA_FLAGS support (because of the bounds check on the signal number). if (sigaction(SIGSEGV | SF_CHECK_SA_FLAGS, &sa, 0) < 0) { sa.sa_sigaction = clear_fields_and_handle_segv; sa.sa_flags = SA_SIGINFO; // Succeeds in old kernels, no need to use SF_CHECK_SA_FLAGS since we're using sa_flags from the beginning of time. if (sigaction(SIGSEGV, &sa, 0) < 0) perror("sigaction"); } } Peter
On Mon, Jul 06, 2020 at 12:20:33PM -0700, Peter Collingbourne wrote: > On Mon, Jul 6, 2020 at 9:41 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > On Wed, Jun 24, 2020 at 12:51:43PM -0700, Peter Collingbourne wrote: > > > On Wed, Jun 24, 2020 at 10:12 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > On Wed, Jun 24, 2020 at 09:51:49AM -0700, Peter Collingbourne wrote: > > > > > On Wed, Jun 24, 2020 at 2:28 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > On Tue, Jun 23, 2020 at 05:40:08PM -0700, Peter Collingbourne wrote: > > > > > > > On Tue, Jun 23, 2020 at 10:52 AM Eric W. Biederman > > > > > > > <ebiederm@xmission.com> wrote: > > > > > > > > > > > > > > > > Dave Martin <Dave.Martin@arm.com> writes: > > > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > > > > > > > > >> Peter Collingbourne <pcc@google.com> writes: > > > > > > > > >> > > > > > > > > >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > > > > > > > > >> > index 47f651df781c..a8380a2b6361 100644 > > > > > > > > >> > --- a/arch/arm64/kernel/traps.c > > > > > > > > >> > +++ b/arch/arm64/kernel/traps.c > > > > > > > > >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > > > > > > > > >> > } > > > > > > > > >> > > > > > > > > > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, > > > > > > > > >> > + unsigned long far, unsigned char far_tb_mask, > > > > > > > > >> > const char *str) > > > > > > > > >> > { > > > > > > > > >> > arm64_show_signal(signo, str); > > > > > > > > >> > - if (signo == SIGKILL) > > > > > > > > >> > + if (signo == SIGKILL) { > > > > > > > > >> > force_sig(SIGKILL); > > > > > > > > >> > - else > > > > > > > > >> > - force_sig_fault(signo, code, addr); > > > > > > > > >> > + } else { > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > >> > + info.si_signo = signo; > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > >> > + info.si_code = code; > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > > > > > > > > >> > + info.si_addr_top_byte_mask = far_tb_mask; > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > >> > + } > > > > > > > > >> > } > > > > > > > > >> > > > > > > > > > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > > > > > > > > >> > - const char *str) > > > > > > > > >> > + unsigned long far, const char *str) > > > > > > > > >> > { > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > >> > + > > > > > > > > >> > arm64_show_signal(SIGBUS, str); > > > > > > > > >> > - force_sig_mceerr(code, addr, lsb); > > > > > > > > >> > + > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > >> > + info.si_signo = SIGBUS; > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > >> > + info.si_code = code; > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > >> > + info.si_addr_lsb = lsb; > > > > > > > > >> > + info.si_addr_top_byte = far >> 56; > > > > > > > > >> > + info.si_addr_top_byte_mask = 0xff; > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > >> > } > > > > > > > > >> > > > > > > > > >> I have a real problem with this construction. force_sig_info is not an > > > > > > > > >> interface that should be used for anything except to define a wrapper > > > > > > > > >> that takes it's parameters. > > > > > > > > > > > > > > > > > > Can you elaborate? How would you do this king of thing. > > > > > > > > > > > > > > > > There are no other uses of force_sig_info in architecture code. > > > > > > > > > > > > > > > > I just removed them _all_ because they were almost all broken. > > > > > > > > In fact your mcerr case is broken because it uses two different > > > > > > > > union members simultantiously. > > > > > > > > > > > > > > Is that really broken? I thought that the Linux kernel deliberately > > > > > > > didn't care about strict aliasing rules (the top-level Makefile passes > > > > > > > -fno-strict-aliasing) so I thought that it was valid in "Linux kernel > > > > > > > C" even though from a standards point of view it is invalid. (That > > > > > > > being said, this is probably moot with my proposed changes below > > > > > > > though.) > > > > > > > > > > > > I have a feeling that -fno-strict-aliasing only allows you to _read_ a > > > > > > different union member from the one previously written. > > > > > > > > > > > > Writing a different member from the last one written can still splatter > > > > > > on the other members IIUC. > > > > > > > > > > > > It would be better to keep things separate rather than risk > > > > > > incorrectness just to save a few bytes. > > > > > > > > > > > > IMHO -fno-strict-aliasing is no excuse for gratuitous type-punning. > > > > > > > > > > > > > > So I am looking for something like force_sig_mcerr or force_sig_fault > > > > > > > > that includes your new information that then calls force_sig_info. > > > > > > > > > > > > > > > > I know of no other way to safely use the siginfo struct. > > > > > > > > > > > > > > So you want something like: > > > > > > > > > > > > > > int force_sig_fault_with_ignored_bits(int signo, int code, void __user > > > > > > > *addr, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > int force_sig_mceerr_with_ignored_bits(int code, void __user *addr, > > > > > > > short lsb, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > > > > > in kernel/signal.c and the code in arch/arm64 would call that? > > > > > > > > > > > > > > > > AIUI we absolutely need a forced signal here, we need to supply > > > > > > > > > metadata, and we don't have to open-code all that at every relevant > > > > > > > > > signal generation site... > > > > > > > > > > > > > > > > > >> It is not clear to me that if you have adapted siginfo_layout. > > > > > > > > > > > > > > > > > > Garbled sentence? > > > > > > > > > > > > > > > > Looks like. One of the pieces of code that needs to change > > > > > > > > when siginfo gets updated is siginfo_layout so that the structure > > > > > > > > can be properly decoded and made sense of. > > > > > > > > > > > > > > > > I am not seeing anything like that. > > > > > > > > > > > > > > Okay, this has to do with copying between the compat and non-compat > > > > > > > versions of the struct? Sure, I can update that, although the code > > > > > > > would be basically non-functional on arm64 because TBI isn't supported > > > > > > > on 32-bit ARM. > > > > > > > > > > > > > > > >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > > > > > > > >> > index cb3d6c267181..6dd82373eb2d 100644 > > > > > > > > >> > --- a/include/uapi/asm-generic/siginfo.h > > > > > > > > >> > +++ b/include/uapi/asm-generic/siginfo.h > > > > > > > > >> > @@ -91,6 +91,14 @@ union __sifields { > > > > > > > > >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > > > > > >> > __u32 _pkey; > > > > > > > > >> > } _addr_pkey; > > > > > > > > >> > +#ifdef __aarch64__ > > > > > > > > >> > + /* used with all si_codes */ > > > > > > > > >> > + struct { > > > > > > > > >> > + short _dummy_top_byte; > > > > > > > > > > > > > > > > > > ^ What's this for? I don't have Eric's insight here. > > > > > > > > > > > > > > We would need a short's worth of padding in order to prevent the > > > > > > > fields from occupying the same address as si_addr_lsb. > > > > > > > > > > > > > > > > > > > > > > > > >> > + unsigned char _top_byte; > > > > > > > > >> > + unsigned char _top_byte_mask; > > > > > > > > >> > + } _addr_top_byte; > > > > > > > > >> > +#endif > > > > > > > > >> > }; > > > > > > > > >> > } _sigfault; > > > > > > > > >> > > > > > > > > > >> > > > > > > > > >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > > > > > > > > >> > > > > > > > > >> Please remove the "#ifdef __aarch64__". If at all possible we want to > > > > > > > > >> design this so any other architecture who has this challenge can use the > > > > > > > > >> code. The kind of code does not get enough attention/maintenance if it > > > > > > > > >> is built for a single architecture. > > > > > > > > > > > > > > Seems reasonable. I was recently made aware that RISC-V was > > > > > > > considering a similar feature: > > > > > > > https://lists.riscv.org/g/tech-tee/topic/risc_v_tbi_proposal/72855478 > > > > > > > I would have opted to expand this to other architectures on an > > > > > > > as-needed basis, but I'd also be fine with having it on all > > > > > > > architectures from the start. > > > > > > > > > > > > > > If we make this arch-independent, we have an additional concern, which > > > > > > > is "what if some future architecture wants more than one byte here?" > > > > > > > For example, an architecture may have a "top-two-bytes-ignore" > > > > > > > feature, which would imply two-byte (misnamed) "si_addr_top_byte" and > > > > > > > "si_addr_top_byte_mask" fields. And the RISC-V proposal potentially > > > > > > > implies many more ignored bits (see slide 13 of the presentation). The > > > > > > > maximum size that these fields can possibly be is the size of a > > > > > > > pointer, and with that there wouldn't be enough room in the padding at > > > > > > > this point to accommodate the new fields. > > > > > > > > > > > > > > That basically implies your earlier suggestion of adding a union > > > > > > > member here to accommodate future expansion of the union, and adding > > > > > > > the new fields after the union. I'm happy to make that change, with > > > > > > > the fields renamed "si_addr_ignored" and "si_addr_ignored_mask". > > > > > > > > > > > > I think what we need here is basically a flags word. > > > > > > > > > > > > So long as we keep a flag spare to indicate the existence of a further > > > > > > flags word, we can extend as needed. > > > > > > > > > > > > How the existence of the first flags words is detected is another > > > > > > problem. If it only applies for newly-defined si_code values, then > > > > > > I guess si_code may be sufficient. > > > > > > > > > > Existing kernels will zero-initialize unused regions of the siginfo > > > > > data structure. The zero-initialization of the padding at the end of > > > > > the struct is done by the clear_user call here: > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/kernel/signal.c#L3193 > > > > > > > > > > and the zero-initialization of the padding between fields and unused > > > > > union members is done by the clear_siginfo function which the kernel > > > > > calls when initializing the data structure: > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/include/linux/signal.h#L20 > > > > > > > > > > Therefore, a flag word value of 0 may be used to detect a lack of > > > > > support for flagged fields. > > > > > > > > It's not enough that we do this today. We would have had to do it back > > > > to the dawn of time (though in the arm64 case I guess we just need to go > > > > back to when the arch/arm64 was merged). > > > > > > > > v2.6.12:kernel/signal.c:copy_siginfo_to_user() suggests that this wasn't > > > > always the case, so unused parts of siginfo could be full of old junk > > > > from the user stack, if the kernel is sufficiently old. > > > > > > > > If we're trying to do something generic that makes sense on all arches, > > > > this matters. I may have misunderstood something about the code though. > > > > > > Hmm, I think you're right. The current behavior was introduced by > > > commit c999b933faa5e281e3af2e110eccaf91698b0a81 which was first > > > released in 4.18. So if an application wants to be compatible with > > > pre-4.18 kernels then there would need to be some other way to > > > indicate that the fields are valid. Probably the simplest way would be > > > to have the application issue a uname(2) syscall and check the kernel > > > version before reading these fields. I have a couple of other ideas > > > that don't rely on version detection, if we'd prefer to avoid that. > > > (They are somewhat ugly, but our hand is forced by backwards > > > compatibility.) > > > > > > One idea is to re-purpose the si_errno field as a flags field for > > > certain signal numbers. I checked a few kernel releases going back to > > > 2.6.18 and it looks like the field is set to 0 except in the following > > > circumstances: > > > - sending a hardware breakpoint (SIGTRAP/TRAP_HWBKPT) > > > - seccomp failures (SIGSYS/SYS_SECCOMP) > > > - user-defined signal via kill_pid_usb_asyncio > > > - SIGSWI in 3.18 and before (code since removed) > > > > > > It is also set to EFAULT for certain SIGSEGV/SEGV_MAPERR signals on > > > powerpc since commit c96c4436aba4c12f1f48369f2f90bc43e12fe36c, which > > > is currently unreleased. So if we wanted to go this route for SIGSEGV > > > we would need to stop the kernel from setting si_errno to EFAULT for > > > this signal before the 5.8 release. > > > > > > Another idea was to have userspace set a flag in sa_flags when > > > registering a signal handler meaning "this signal handler requires > > > unknown siginfo fields to be zeroed", and have existing kernels reject > > > the syscall due to an unknown flag being set, but unfortunately this > > > won't work because existing kernels do not reject sigaction syscalls > > > with unknown flags set in sa_flags. A perhaps more radical idea in > > > this vein would be to claim some of the upper bits of the signal > > > number as flags that will cause the syscall to be rejected if set and > > > unknown to the kernel. Existing kernels (going back to at least > > > 2.6.18) contain this code in do_sigaction: > > > > > > if (!valid_signal(sig) || sig < 1 || (act && sig_kernel_only(sig))) > > > return -EINVAL; > > > > > > and vald_signal is defined as: > > > > > > static inline int valid_signal(unsigned long sig) > > > { > > > return sig <= _NSIG ? 1 : 0; > > > } > > > > > > All architectures define _NSIG as a value <= 128, so they will reject > > > a signal number with any of bits 8-31 set. This means that we can use > > > any of those bits for mandatory flags. Most likely we could use bit 30 > > > (expanding down as necessary), as it keeps the signal number positive > > > and permits future expansion of the signal number range. > > > > Does the signal core code actually gurantee to zero the unused fields? > > Unless the fields are poked in by hand this is fraught with subtlelies, > > especially when unions are involved. (I'm sure the code tries to do it, > > but I've not eyeballed it in detail...) > > It memsets the siginfo structure before setting the fields and sending > the signal (grep for clear_siginfo which is just a memset; you should > find a call before all callers of force_sig_info). Memset is the right > approach here since unlike setting fields by hand it clears padding > which could lead to information leaks from the kernel. IIUC this is > the reason why Eric wants all of the signals to be raised via wrappers > in kernel/signal.c instead of via force_sig_info directly (to make > this aspect easier to audit). My impression was that the reason for this model is partly to ensure that siginfo fields are populated more consistently. When it was all down to the individual callers, inconsistencies creeped in. With regard to memset(), this is not a complete defence against data leakage. Assigning to a struct member can set any or all padding in the struct to random garbage (consider write-combining of neighboring member writes into a single larger accesses in asm for example). The only way to avoid this is to ensure that the struct is 100% padding-free, and that each member of a union is the same size. A quick clance at <uapi/asm-generic/siginfo.h> confirms that this is not the case. This might need to be looked at separately. But it does mean, strictly speaking, that we can't reliably add new fields anywhere that there was previously padding: assigning to neighboring members can still fill those with garbage after the memset(). > > Using unused bits in the signal number to turn on new functionality > > feels risky. As currently specified, this is just a number. Since > > today a successful sigaction(n ...) guarantees that n is a valid signal > > number, reasonable code like the following would trigger a buffer > > overrun if we start trying to encode anything else in there: > > > > struct sigaction actions[NSIG]; > > > > int do_something( ... ) > > { > > ... > > > > if (!sigaction(n, sa, ...)) { > > actions[n] = *sa; > > return 0; > > } > > > > ... > > } > > I imagine the bit in the signal number being set by the direct caller > to sigaction, and we could specifically recommend that calling > pattern. In that case, your "n" wouldn't have the bit set in it. It I can imagine this too, but that doesn't mean that software does it. If the above kind of thing exists in a framework or library somewhere, we could get problems. Similarly, a pre-existing LD_PRELOAD framework that provides a wrapper for sigaction may now go wrong even if your pattern is followed -- i.e., the caller thinks it's calling sigaction directly but in fact it isn't. > could only appear in newly-written code that doesn't follow our > recommendations, and there are already plenty of much more likely ways > to cause buffer overflows in C code that doesn't follow > recommendations anyway. (And even if such a buffer overflow occurred, > it would very likely be caught early in development by the MMU due to > the magnitude of the number 1<<30.) Choosing the bit value is hard. If shitfing it overflows, this can trigger random undefined behaviour in the compiler in addition to (or perhaps instead of) an out-of-bounds access or segfault. If shifting it doesn't overflow, we might still fall into a valid mapping, though I'd agree a segfault is more likely. > > > I think it would be cleaner for to add a single flag field that can be > > used for detecting other extensions, and request it via a new sa_flags > > bit. This removes the need for sematically useless zeroing of unused > > fields (though for hygiene and backwards compatibility reasons we would > > probably want to carry on zeroing them anyway). > > > > I can see no simpler way to add supplementary siginfo fields for > > existing si_codes. For si_codes that didn't exist before the zeroing > > came in we could still detect optional si_code-specific fields via > > zeroing, but it seems messary to have two ways of detecting extensions. > > That would certainly be cleaner if it worked, but that would only be > the case if old kernels rejected unknown bits in sa_flags, and > unfortunately they don't. With the bit in the signal number, the "old Hmm, that is a problem I wasn't aware of. > kernels reject" behavior admits relatively straightforward usage code: > > void set_segv_handler(void) { > struct sigaction sa; > sa.sa_sigaction = handle_segv; > sa.sa_flags = SA_SIGINFO; > if (sigaction(SIGSEGV | SF_CLEAR_UNKNOWN_FIELDS, &sa, 0) < 0) { // > succeeds in new kernels, fails in old kernels > sa.sa_sigaction = clear_fields_and_handle_segv; > if (sigaction(SIGSEGV, &sa, 0) < 0) // succeeds in old kernels > perror("sigaction"); > } > } > > void clear_fields_and_handle_segv(int signum, siginfo_t *sa, void *ctx) { > sa->si_future_field = 0; > handle_segv(signum, sa, ctx); > } > > void handle_segv(int signum, siginfo_t *sa, void *ctx) { > // At this point, si_future_field will have the value 0 in old > kernels and the kernel-supplied value in new kernels. > } > > Imagine if we moved the flag SF_CLEAR_UNKNOWN_FIELDS from the signal > number to sa_flags. In that case, the first sigaction would succeed in > old kernels so handle_segv wouldn't know whether it can safely read > from si_future_field. With the sa_flags approach, you would need > kernel version number checking via uname before setting the flag in > sa_flags, and at that point why even have the flag in sa_flags at all > since you could just have the signal handler conditionally read from > si_future_field based on the uname? Software setting SA_SIFLAGS (or whatever) is new by definition, since it would be using a new #define. So it might be reasonable to put the burden on that software to verify that the flag was really accepted by the kernel, by reading it back. Unfortunately, even relatively recent kernels blindly store sa_flags in the kernel without validating it, and so it looks like duff flags can be read back out via a sigaction() call. Dang. Perhaps a new frontend syscall could be added. A new libc that knows about this "sigaction2" could use it and mask off problem bits from sa_flags in its sigaction() wrapper before calling sigaction2. An old libc would call the old sigaction syscall, where we would ignore these new sa_flags bits as before. This may not be a popular approach though, and software wouldn't be able to use our new features until libc is updated to match. If we go down this route, it may provide additional opportunities to fix annoying defects in the old interface. > Note that the same applies to a flag indicating the availability of a > si_flags field in sigaction (just > s/SF_CLEAR_UNKNOWN_FIELDS/SF_HAS_SI_FLAGS/ and > s/si_future_field/si_flags/ in the usage code above). In terms of > SF_CLEAR_UNKNOWN_FIELDS versus SF_HAS_SI_FLAGS I'd be fine either way. > > Another thought that occurred to me is that we may consider > generalizing this a step further and introducing a single flag bit in > the signal number that means "reject unknown flags in sa_flags". This > would mean that we wouldn't need to add any more flag bits to the > signal number in the future, thus limiting this signal number hack to > a single bit; all future mandatory behavior changes could just be put > behind a flag in sa_flags and userspace code would easily be able to > detect missing support for a flag and fall back if necessary. In our > case, this would imply usage code like this: > > void set_segv_handler(void) { > struct sigaction sa; > sa.sa_sigaction = handle_segv; > sa.sa_flags = SA_SIGINFO | SA_CLEAR_UNKNOWN_FIELDS; > // Succeeds in kernels with SA_CLEAR_UNKNOWN_FIELDS support. > // Fails in kernels with SF_CHECK_SA_FLAGS support but no > SA_CLEAR_UNKNOWN_FIELDS support (because of the unknown flags check). > // Fails in kernels without SF_CHECK_SA_FLAGS support (because of > the bounds check on the signal number). > if (sigaction(SIGSEGV | SF_CHECK_SA_FLAGS, &sa, 0) < 0) { > sa.sa_sigaction = clear_fields_and_handle_segv; > sa.sa_flags = SA_SIGINFO; > // Succeeds in old kernels, no need to use SF_CHECK_SA_FLAGS since > we're using sa_flags from the beginning of time. > if (sigaction(SIGSEGV, &sa, 0) < 0) > perror("sigaction"); > } > } As with the other options this could work, but looks like it could break the ABI due to violating the original semantics for the signal number argument. Perhaps I'm being too paranoid. Cheers ---Dave
On Tue, Jul 7, 2020 at 7:19 AM Dave Martin <Dave.Martin@arm.com> wrote: > > On Mon, Jul 06, 2020 at 12:20:33PM -0700, Peter Collingbourne wrote: > > On Mon, Jul 6, 2020 at 9:41 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > On Wed, Jun 24, 2020 at 12:51:43PM -0700, Peter Collingbourne wrote: > > > > On Wed, Jun 24, 2020 at 10:12 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > On Wed, Jun 24, 2020 at 09:51:49AM -0700, Peter Collingbourne wrote: > > > > > > On Wed, Jun 24, 2020 at 2:28 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 05:40:08PM -0700, Peter Collingbourne wrote: > > > > > > > > On Tue, Jun 23, 2020 at 10:52 AM Eric W. Biederman > > > > > > > > <ebiederm@xmission.com> wrote: > > > > > > > > > > > > > > > > > > Dave Martin <Dave.Martin@arm.com> writes: > > > > > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > > > > > > > > > >> Peter Collingbourne <pcc@google.com> writes: > > > > > > > > > >> > > > > > > > > > >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > > > > > > > > > >> > index 47f651df781c..a8380a2b6361 100644 > > > > > > > > > >> > --- a/arch/arm64/kernel/traps.c > > > > > > > > > >> > +++ b/arch/arm64/kernel/traps.c > > > > > > > > > >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > > > > > > > > > >> > } > > > > > > > > > >> > > > > > > > > > > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, > > > > > > > > > >> > + unsigned long far, unsigned char far_tb_mask, > > > > > > > > > >> > const char *str) > > > > > > > > > >> > { > > > > > > > > > >> > arm64_show_signal(signo, str); > > > > > > > > > >> > - if (signo == SIGKILL) > > > > > > > > > >> > + if (signo == SIGKILL) { > > > > > > > > > >> > force_sig(SIGKILL); > > > > > > > > > >> > - else > > > > > > > > > >> > - force_sig_fault(signo, code, addr); > > > > > > > > > >> > + } else { > > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > > >> > + info.si_signo = signo; > > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > > >> > + info.si_code = code; > > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > > >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > > > > > > > > > >> > + info.si_addr_top_byte_mask = far_tb_mask; > > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > > >> > + } > > > > > > > > > >> > } > > > > > > > > > >> > > > > > > > > > > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > > > > > > > > > >> > - const char *str) > > > > > > > > > >> > + unsigned long far, const char *str) > > > > > > > > > >> > { > > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > > >> > + > > > > > > > > > >> > arm64_show_signal(SIGBUS, str); > > > > > > > > > >> > - force_sig_mceerr(code, addr, lsb); > > > > > > > > > >> > + > > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > > >> > + info.si_signo = SIGBUS; > > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > > >> > + info.si_code = code; > > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > > >> > + info.si_addr_lsb = lsb; > > > > > > > > > >> > + info.si_addr_top_byte = far >> 56; > > > > > > > > > >> > + info.si_addr_top_byte_mask = 0xff; > > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > > >> > } > > > > > > > > > >> > > > > > > > > > >> I have a real problem with this construction. force_sig_info is not an > > > > > > > > > >> interface that should be used for anything except to define a wrapper > > > > > > > > > >> that takes it's parameters. > > > > > > > > > > > > > > > > > > > > Can you elaborate? How would you do this king of thing. > > > > > > > > > > > > > > > > > > There are no other uses of force_sig_info in architecture code. > > > > > > > > > > > > > > > > > > I just removed them _all_ because they were almost all broken. > > > > > > > > > In fact your mcerr case is broken because it uses two different > > > > > > > > > union members simultantiously. > > > > > > > > > > > > > > > > Is that really broken? I thought that the Linux kernel deliberately > > > > > > > > didn't care about strict aliasing rules (the top-level Makefile passes > > > > > > > > -fno-strict-aliasing) so I thought that it was valid in "Linux kernel > > > > > > > > C" even though from a standards point of view it is invalid. (That > > > > > > > > being said, this is probably moot with my proposed changes below > > > > > > > > though.) > > > > > > > > > > > > > > I have a feeling that -fno-strict-aliasing only allows you to _read_ a > > > > > > > different union member from the one previously written. > > > > > > > > > > > > > > Writing a different member from the last one written can still splatter > > > > > > > on the other members IIUC. > > > > > > > > > > > > > > It would be better to keep things separate rather than risk > > > > > > > incorrectness just to save a few bytes. > > > > > > > > > > > > > > IMHO -fno-strict-aliasing is no excuse for gratuitous type-punning. > > > > > > > > > > > > > > > > So I am looking for something like force_sig_mcerr or force_sig_fault > > > > > > > > > that includes your new information that then calls force_sig_info. > > > > > > > > > > > > > > > > > > I know of no other way to safely use the siginfo struct. > > > > > > > > > > > > > > > > So you want something like: > > > > > > > > > > > > > > > > int force_sig_fault_with_ignored_bits(int signo, int code, void __user > > > > > > > > *addr, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > int force_sig_mceerr_with_ignored_bits(int code, void __user *addr, > > > > > > > > short lsb, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > > > > > > > in kernel/signal.c and the code in arch/arm64 would call that? > > > > > > > > > > > > > > > > > > AIUI we absolutely need a forced signal here, we need to supply > > > > > > > > > > metadata, and we don't have to open-code all that at every relevant > > > > > > > > > > signal generation site... > > > > > > > > > > > > > > > > > > > >> It is not clear to me that if you have adapted siginfo_layout. > > > > > > > > > > > > > > > > > > > > Garbled sentence? > > > > > > > > > > > > > > > > > > Looks like. One of the pieces of code that needs to change > > > > > > > > > when siginfo gets updated is siginfo_layout so that the structure > > > > > > > > > can be properly decoded and made sense of. > > > > > > > > > > > > > > > > > > I am not seeing anything like that. > > > > > > > > > > > > > > > > Okay, this has to do with copying between the compat and non-compat > > > > > > > > versions of the struct? Sure, I can update that, although the code > > > > > > > > would be basically non-functional on arm64 because TBI isn't supported > > > > > > > > on 32-bit ARM. > > > > > > > > > > > > > > > > > >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > > > > > > > > >> > index cb3d6c267181..6dd82373eb2d 100644 > > > > > > > > > >> > --- a/include/uapi/asm-generic/siginfo.h > > > > > > > > > >> > +++ b/include/uapi/asm-generic/siginfo.h > > > > > > > > > >> > @@ -91,6 +91,14 @@ union __sifields { > > > > > > > > > >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > > > > > > >> > __u32 _pkey; > > > > > > > > > >> > } _addr_pkey; > > > > > > > > > >> > +#ifdef __aarch64__ > > > > > > > > > >> > + /* used with all si_codes */ > > > > > > > > > >> > + struct { > > > > > > > > > >> > + short _dummy_top_byte; > > > > > > > > > > > > > > > > > > > > ^ What's this for? I don't have Eric's insight here. > > > > > > > > > > > > > > > > We would need a short's worth of padding in order to prevent the > > > > > > > > fields from occupying the same address as si_addr_lsb. > > > > > > > > > > > > > > > > > > > > > > > > > > > >> > + unsigned char _top_byte; > > > > > > > > > >> > + unsigned char _top_byte_mask; > > > > > > > > > >> > + } _addr_top_byte; > > > > > > > > > >> > +#endif > > > > > > > > > >> > }; > > > > > > > > > >> > } _sigfault; > > > > > > > > > >> > > > > > > > > > > >> > > > > > > > > > >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > > > > > > > > > >> > > > > > > > > > >> Please remove the "#ifdef __aarch64__". If at all possible we want to > > > > > > > > > >> design this so any other architecture who has this challenge can use the > > > > > > > > > >> code. The kind of code does not get enough attention/maintenance if it > > > > > > > > > >> is built for a single architecture. > > > > > > > > > > > > > > > > Seems reasonable. I was recently made aware that RISC-V was > > > > > > > > considering a similar feature: > > > > > > > > https://lists.riscv.org/g/tech-tee/topic/risc_v_tbi_proposal/72855478 > > > > > > > > I would have opted to expand this to other architectures on an > > > > > > > > as-needed basis, but I'd also be fine with having it on all > > > > > > > > architectures from the start. > > > > > > > > > > > > > > > > If we make this arch-independent, we have an additional concern, which > > > > > > > > is "what if some future architecture wants more than one byte here?" > > > > > > > > For example, an architecture may have a "top-two-bytes-ignore" > > > > > > > > feature, which would imply two-byte (misnamed) "si_addr_top_byte" and > > > > > > > > "si_addr_top_byte_mask" fields. And the RISC-V proposal potentially > > > > > > > > implies many more ignored bits (see slide 13 of the presentation). The > > > > > > > > maximum size that these fields can possibly be is the size of a > > > > > > > > pointer, and with that there wouldn't be enough room in the padding at > > > > > > > > this point to accommodate the new fields. > > > > > > > > > > > > > > > > That basically implies your earlier suggestion of adding a union > > > > > > > > member here to accommodate future expansion of the union, and adding > > > > > > > > the new fields after the union. I'm happy to make that change, with > > > > > > > > the fields renamed "si_addr_ignored" and "si_addr_ignored_mask". > > > > > > > > > > > > > > I think what we need here is basically a flags word. > > > > > > > > > > > > > > So long as we keep a flag spare to indicate the existence of a further > > > > > > > flags word, we can extend as needed. > > > > > > > > > > > > > > How the existence of the first flags words is detected is another > > > > > > > problem. If it only applies for newly-defined si_code values, then > > > > > > > I guess si_code may be sufficient. > > > > > > > > > > > > Existing kernels will zero-initialize unused regions of the siginfo > > > > > > data structure. The zero-initialization of the padding at the end of > > > > > > the struct is done by the clear_user call here: > > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/kernel/signal.c#L3193 > > > > > > > > > > > > and the zero-initialization of the padding between fields and unused > > > > > > union members is done by the clear_siginfo function which the kernel > > > > > > calls when initializing the data structure: > > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/include/linux/signal.h#L20 > > > > > > > > > > > > Therefore, a flag word value of 0 may be used to detect a lack of > > > > > > support for flagged fields. > > > > > > > > > > It's not enough that we do this today. We would have had to do it back > > > > > to the dawn of time (though in the arm64 case I guess we just need to go > > > > > back to when the arch/arm64 was merged). > > > > > > > > > > v2.6.12:kernel/signal.c:copy_siginfo_to_user() suggests that this wasn't > > > > > always the case, so unused parts of siginfo could be full of old junk > > > > > from the user stack, if the kernel is sufficiently old. > > > > > > > > > > If we're trying to do something generic that makes sense on all arches, > > > > > this matters. I may have misunderstood something about the code though. > > > > > > > > Hmm, I think you're right. The current behavior was introduced by > > > > commit c999b933faa5e281e3af2e110eccaf91698b0a81 which was first > > > > released in 4.18. So if an application wants to be compatible with > > > > pre-4.18 kernels then there would need to be some other way to > > > > indicate that the fields are valid. Probably the simplest way would be > > > > to have the application issue a uname(2) syscall and check the kernel > > > > version before reading these fields. I have a couple of other ideas > > > > that don't rely on version detection, if we'd prefer to avoid that. > > > > (They are somewhat ugly, but our hand is forced by backwards > > > > compatibility.) > > > > > > > > One idea is to re-purpose the si_errno field as a flags field for > > > > certain signal numbers. I checked a few kernel releases going back to > > > > 2.6.18 and it looks like the field is set to 0 except in the following > > > > circumstances: > > > > - sending a hardware breakpoint (SIGTRAP/TRAP_HWBKPT) > > > > - seccomp failures (SIGSYS/SYS_SECCOMP) > > > > - user-defined signal via kill_pid_usb_asyncio > > > > - SIGSWI in 3.18 and before (code since removed) > > > > > > > > It is also set to EFAULT for certain SIGSEGV/SEGV_MAPERR signals on > > > > powerpc since commit c96c4436aba4c12f1f48369f2f90bc43e12fe36c, which > > > > is currently unreleased. So if we wanted to go this route for SIGSEGV > > > > we would need to stop the kernel from setting si_errno to EFAULT for > > > > this signal before the 5.8 release. > > > > > > > > Another idea was to have userspace set a flag in sa_flags when > > > > registering a signal handler meaning "this signal handler requires > > > > unknown siginfo fields to be zeroed", and have existing kernels reject > > > > the syscall due to an unknown flag being set, but unfortunately this > > > > won't work because existing kernels do not reject sigaction syscalls > > > > with unknown flags set in sa_flags. A perhaps more radical idea in > > > > this vein would be to claim some of the upper bits of the signal > > > > number as flags that will cause the syscall to be rejected if set and > > > > unknown to the kernel. Existing kernels (going back to at least > > > > 2.6.18) contain this code in do_sigaction: > > > > > > > > if (!valid_signal(sig) || sig < 1 || (act && sig_kernel_only(sig))) > > > > return -EINVAL; > > > > > > > > and vald_signal is defined as: > > > > > > > > static inline int valid_signal(unsigned long sig) > > > > { > > > > return sig <= _NSIG ? 1 : 0; > > > > } > > > > > > > > All architectures define _NSIG as a value <= 128, so they will reject > > > > a signal number with any of bits 8-31 set. This means that we can use > > > > any of those bits for mandatory flags. Most likely we could use bit 30 > > > > (expanding down as necessary), as it keeps the signal number positive > > > > and permits future expansion of the signal number range. > > > > > > Does the signal core code actually gurantee to zero the unused fields? > > > Unless the fields are poked in by hand this is fraught with subtlelies, > > > especially when unions are involved. (I'm sure the code tries to do it, > > > but I've not eyeballed it in detail...) > > > > It memsets the siginfo structure before setting the fields and sending > > the signal (grep for clear_siginfo which is just a memset; you should > > find a call before all callers of force_sig_info). Memset is the right > > approach here since unlike setting fields by hand it clears padding > > which could lead to information leaks from the kernel. IIUC this is > > the reason why Eric wants all of the signals to be raised via wrappers > > in kernel/signal.c instead of via force_sig_info directly (to make > > this aspect easier to audit). > > My impression was that the reason for this model is partly to ensure > that siginfo fields are populated more consistently. When it was all > down to the individual callers, inconsistencies creeped in. > > With regard to memset(), this is not a complete defence against data > leakage. Assigning to a struct member can set any or all padding in > the struct to random garbage (consider write-combining of neighboring > member writes into a single larger accesses in asm for example). The I don't believe that LLVM will store to padding like this. I don't know about GCC, though, but I wouldn't be surprised if this is something that the kernel would want to turn off in "kernel C" (like it turns off strict aliasing) specifically because of the information leak issue. > only way to avoid this is to ensure that the struct is 100% > padding-free, and that each member of a union is the same size. A > quick clance at <uapi/asm-generic/siginfo.h> confirms that this is not > the case. > > This might need to be looked at separately. > > But it does mean, strictly speaking, that we can't reliably add new > fields anywhere that there was previously padding: assigning to > neighboring members can still fill those with garbage after the > memset(). ...but this is largely moot because I'm not proposing to add new fields in the padding any more (because the fields needed to become larger in order to accommodate future hypothetical architectures which might want to use the fields, and thus they wouldn't fit in the padding). The siginfo.h diff would be something like: diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h index cb3d6c267181..4a2fe257415d 100644 --- a/include/uapi/asm-generic/siginfo.h +++ b/include/uapi/asm-generic/siginfo.h @@ -91,7 +91,10 @@ union __sifields { char _dummy_pkey[__ADDR_BND_PKEY_PAD]; __u32 _pkey; } _addr_pkey; + void *_pad[6]; }; + uintptr_t _ignored_bits; + uintptr_t _ignored_bits_mask; } _sigfault; /* SIGPOLL */ or with a "uintptr_t _flags" added in before _ignored_bits if we go with that. > > > Using unused bits in the signal number to turn on new functionality > > > feels risky. As currently specified, this is just a number. Since > > > today a successful sigaction(n ...) guarantees that n is a valid signal > > > number, reasonable code like the following would trigger a buffer > > > overrun if we start trying to encode anything else in there: > > > > > > struct sigaction actions[NSIG]; > > > > > > int do_something( ... ) > > > { > > > ... > > > > > > if (!sigaction(n, sa, ...)) { > > > actions[n] = *sa; > > > return 0; > > > } > > > > > > ... > > > } > > > > I imagine the bit in the signal number being set by the direct caller > > to sigaction, and we could specifically recommend that calling > > pattern. In that case, your "n" wouldn't have the bit set in it. It > > I can imagine this too, but that doesn't mean that software does it. > > If the above kind of thing exists in a framework or library somewhere, > we could get problems. Similarly, a pre-existing LD_PRELOAD framework > that provides a wrapper for sigaction may now go wrong even if your > pattern is followed -- i.e., the caller thinks it's calling sigaction > directly but in fact it isn't. I'm aware of one library like that. It's called libsigchain, and it has an early bounds check: https://cs.android.com/android/platform/superproject/+/master:art/sigchainlib/sigchain.cc;l=371 Until the library is changed to recognize the flag, calling code would see the return value of -1 as if the kernel failed the syscall, and would fall back to the code for old kernels. In general I think that any library like this with independent tracking of the kernel's purported signal handler state would need to be very sensitive to which syscalls are capable of setting signal handlers, what their semantics are, and so on. This applies to any change that we might make to the signal handler interface. So for example, if we introduced a new syscall as you propose below, and the library hasn't been updated to recognize the new syscall, it will silently miss changes in signal handler state caused by the new syscall. At the end of this argument lies "we can never change anything about how signal handlers work because it could break some interposing library somewhere" -- replace "signal handlers" with any kernel feature whose behavior may be modified by an interposing library if you like -- and I don't think we want to go that far. As far as I know, this isn't really the kernel's business anyway -- the kernel's stable ABI contract starts and ends with the syscall interface and not some library on top. That being said, we should perhaps try to define our interface so that something reasonable will probably happen if there is such a library and it hasn't been updated. With the new syscall, the library will sometimes silently fail to work in some non-local fashion. With the flag bit in the signal number, the library will either cause the caller to fall back to the old kernel code path (if there is a bounds check) or likely crash loudly (if there is no bounds check). To me, the "flag bit in the signal number" behavior seems more reasonable, since either something correct or something easy to debug will probably happen at runtime. > > could only appear in newly-written code that doesn't follow our > > recommendations, and there are already plenty of much more likely ways > > to cause buffer overflows in C code that doesn't follow > > recommendations anyway. (And even if such a buffer overflow occurred, > > it would very likely be caught early in development by the MMU due to > > the magnitude of the number 1<<30.) > > Choosing the bit value is hard. If shitfing it overflows, this can > trigger random undefined behaviour in the compiler in addition to (or > perhaps instead of) an out-of-bounds access or segfault. It wouldn't overflow on a 64-bit architecture assuming normal array indexing (the index would be promoted to pointer width before being scaled to the array element size), and to begin with the users of this would be 64-bit. > If shifting it doesn't overflow, we might still fall into a valid > mapping, though I'd agree a segfault is more likely. > > > > > > I think it would be cleaner for to add a single flag field that can be > > > used for detecting other extensions, and request it via a new sa_flags > > > bit. This removes the need for sematically useless zeroing of unused > > > fields (though for hygiene and backwards compatibility reasons we would > > > probably want to carry on zeroing them anyway). > > > > > > I can see no simpler way to add supplementary siginfo fields for > > > existing si_codes. For si_codes that didn't exist before the zeroing > > > came in we could still detect optional si_code-specific fields via > > > zeroing, but it seems messary to have two ways of detecting extensions. > > > > That would certainly be cleaner if it worked, but that would only be > > the case if old kernels rejected unknown bits in sa_flags, and > > unfortunately they don't. With the bit in the signal number, the "old > > Hmm, that is a problem I wasn't aware of. > > > kernels reject" behavior admits relatively straightforward usage code: > > > > void set_segv_handler(void) { > > struct sigaction sa; > > sa.sa_sigaction = handle_segv; > > sa.sa_flags = SA_SIGINFO; > > if (sigaction(SIGSEGV | SF_CLEAR_UNKNOWN_FIELDS, &sa, 0) < 0) { // > > succeeds in new kernels, fails in old kernels > > sa.sa_sigaction = clear_fields_and_handle_segv; > > if (sigaction(SIGSEGV, &sa, 0) < 0) // succeeds in old kernels > > perror("sigaction"); > > } > > } > > > > void clear_fields_and_handle_segv(int signum, siginfo_t *sa, void *ctx) { > > sa->si_future_field = 0; > > handle_segv(signum, sa, ctx); > > } > > > > void handle_segv(int signum, siginfo_t *sa, void *ctx) { > > // At this point, si_future_field will have the value 0 in old > > kernels and the kernel-supplied value in new kernels. > > } > > > > Imagine if we moved the flag SF_CLEAR_UNKNOWN_FIELDS from the signal > > number to sa_flags. In that case, the first sigaction would succeed in > > old kernels so handle_segv wouldn't know whether it can safely read > > from si_future_field. With the sa_flags approach, you would need > > kernel version number checking via uname before setting the flag in > > sa_flags, and at that point why even have the flag in sa_flags at all > > since you could just have the signal handler conditionally read from > > si_future_field based on the uname? > > Software setting SA_SIFLAGS (or whatever) is new by definition, since > it would be using a new #define. So it might be reasonable to put the > burden on that software to verify that the flag was really accepted by > the kernel, by reading it back. That doesn't seem like a good idea even if it worked, because it could lead to race conditions. If the si_flags-reading signal handler were invoked in response to a signal between when you set it and when you ended up replacing it with the fallback signal handler for old kernels, the handler may end up reading garbage data from si_flags. > Unfortunately, even relatively recent kernels blindly store sa_flags > in the kernel without validating it, and so it looks like duff flags > can be read back out via a sigaction() call. Dang. > > > Perhaps a new frontend syscall could be added. A new libc that knows > about this "sigaction2" could use it and mask off problem bits from > sa_flags in its sigaction() wrapper before calling sigaction2. An old > libc would call the old sigaction syscall, where we would ignore these > new sa_flags bits as before. I'm not currently in favor of the new syscall but if we do this I would keep sigaction and sigaction2 separate. That is, libc sigaction should always use the sigaction syscall, and libc sigaction2 should always use the sigaction2 syscall. We should avoid libc's sigaction having different behavior based on the libc version and kernel version, as that would make it harder to reason about its behavior. Calling code would need to check for presence of sigaction2 in both libc and the kernel, e.g. __attribute__((weak)) decltype(sigaction2) sigaction2; void set_segv_handler(void) { struct sigaction sa; sa.sa_sigaction = handle_segv; sa.sa_flags = SA_SIGINFO | SA_SIFLAGS; if (!sigaction2 || sigaction2(SIGSEGV, &sa, 0) < 0) { sa.sa_sigaction = clear_fields_and_handle_segv; sa.sa_flags = SA_SIGINFO; if (sigaction(SIGSEGV, &sa, 0) < 0) perror("sigaction"); } } > This may not be a popular approach though, and software wouldn't be able > to use our new features until libc is updated to match. > > If we go down this route, it may provide additional opportunities to fix > annoying defects in the old interface. > > > > Note that the same applies to a flag indicating the availability of a > > si_flags field in sigaction (just > > s/SF_CLEAR_UNKNOWN_FIELDS/SF_HAS_SI_FLAGS/ and > > s/si_future_field/si_flags/ in the usage code above). In terms of > > SF_CLEAR_UNKNOWN_FIELDS versus SF_HAS_SI_FLAGS I'd be fine either way. > > > > Another thought that occurred to me is that we may consider > > generalizing this a step further and introducing a single flag bit in > > the signal number that means "reject unknown flags in sa_flags". This > > would mean that we wouldn't need to add any more flag bits to the > > signal number in the future, thus limiting this signal number hack to > > a single bit; all future mandatory behavior changes could just be put > > behind a flag in sa_flags and userspace code would easily be able to > > detect missing support for a flag and fall back if necessary. In our > > case, this would imply usage code like this: > > > > void set_segv_handler(void) { > > struct sigaction sa; > > sa.sa_sigaction = handle_segv; > > sa.sa_flags = SA_SIGINFO | SA_CLEAR_UNKNOWN_FIELDS; > > // Succeeds in kernels with SA_CLEAR_UNKNOWN_FIELDS support. > > // Fails in kernels with SF_CHECK_SA_FLAGS support but no > > SA_CLEAR_UNKNOWN_FIELDS support (because of the unknown flags check). > > // Fails in kernels without SF_CHECK_SA_FLAGS support (because of > > the bounds check on the signal number). > > if (sigaction(SIGSEGV | SF_CHECK_SA_FLAGS, &sa, 0) < 0) { > > sa.sa_sigaction = clear_fields_and_handle_segv; > > sa.sa_flags = SA_SIGINFO; > > // Succeeds in old kernels, no need to use SF_CHECK_SA_FLAGS since > > we're using sa_flags from the beginning of time. > > if (sigaction(SIGSEGV, &sa, 0) < 0) > > perror("sigaction"); > > } > > } > > As with the other options this could work, but looks like it could > break the ABI due to violating the original semantics for the signal > number argument. Perhaps I'm being too paranoid. There's no ABI being broken here, as long as we consider syscalls to be the stable ABI layer. Old kernels are simply rejecting arguments that they don't know about yet. By that argument, any introduction of a new syscall is an ABI break because it changes the semantics of a previously-unallocated syscall number. Peter
On Tue, Jul 07, 2020 at 12:07:09PM -0700, Peter Collingbourne wrote: > On Tue, Jul 7, 2020 at 7:19 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > On Mon, Jul 06, 2020 at 12:20:33PM -0700, Peter Collingbourne wrote: > > > On Mon, Jul 6, 2020 at 9:41 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > On Wed, Jun 24, 2020 at 12:51:43PM -0700, Peter Collingbourne wrote: > > > > > On Wed, Jun 24, 2020 at 10:12 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > On Wed, Jun 24, 2020 at 09:51:49AM -0700, Peter Collingbourne wrote: > > > > > > > On Wed, Jun 24, 2020 at 2:28 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 05:40:08PM -0700, Peter Collingbourne wrote: > > > > > > > > > On Tue, Jun 23, 2020 at 10:52 AM Eric W. Biederman > > > > > > > > > <ebiederm@xmission.com> wrote: > > > > > > > > > > > > > > > > > > > > Dave Martin <Dave.Martin@arm.com> writes: > > > > > > > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > > > > > > > > > > >> Peter Collingbourne <pcc@google.com> writes: > > > > > > > > > > >> > > > > > > > > > > >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > > > > > > > > > > >> > index 47f651df781c..a8380a2b6361 100644 > > > > > > > > > > >> > --- a/arch/arm64/kernel/traps.c > > > > > > > > > > >> > +++ b/arch/arm64/kernel/traps.c > > > > > > > > > > >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > > > > > > > > > > >> > } > > > > > > > > > > >> > > > > > > > > > > > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, > > > > > > > > > > >> > + unsigned long far, unsigned char far_tb_mask, > > > > > > > > > > >> > const char *str) > > > > > > > > > > >> > { > > > > > > > > > > >> > arm64_show_signal(signo, str); > > > > > > > > > > >> > - if (signo == SIGKILL) > > > > > > > > > > >> > + if (signo == SIGKILL) { > > > > > > > > > > >> > force_sig(SIGKILL); > > > > > > > > > > >> > - else > > > > > > > > > > >> > - force_sig_fault(signo, code, addr); > > > > > > > > > > >> > + } else { > > > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > > > >> > + info.si_signo = signo; > > > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > > > >> > + info.si_code = code; > > > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > > > >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > > > > > > > > > > >> > + info.si_addr_top_byte_mask = far_tb_mask; > > > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > > > >> > + } > > > > > > > > > > >> > } > > > > > > > > > > >> > > > > > > > > > > > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > > > > > > > > > > >> > - const char *str) > > > > > > > > > > >> > + unsigned long far, const char *str) > > > > > > > > > > >> > { > > > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > > > >> > + > > > > > > > > > > >> > arm64_show_signal(SIGBUS, str); > > > > > > > > > > >> > - force_sig_mceerr(code, addr, lsb); > > > > > > > > > > >> > + > > > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > > > >> > + info.si_signo = SIGBUS; > > > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > > > >> > + info.si_code = code; > > > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > > > >> > + info.si_addr_lsb = lsb; > > > > > > > > > > >> > + info.si_addr_top_byte = far >> 56; > > > > > > > > > > >> > + info.si_addr_top_byte_mask = 0xff; > > > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > > > >> > } > > > > > > > > > > >> > > > > > > > > > > >> I have a real problem with this construction. force_sig_info is not an > > > > > > > > > > >> interface that should be used for anything except to define a wrapper > > > > > > > > > > >> that takes it's parameters. > > > > > > > > > > > > > > > > > > > > > > Can you elaborate? How would you do this king of thing. > > > > > > > > > > > > > > > > > > > > There are no other uses of force_sig_info in architecture code. > > > > > > > > > > > > > > > > > > > > I just removed them _all_ because they were almost all broken. > > > > > > > > > > In fact your mcerr case is broken because it uses two different > > > > > > > > > > union members simultantiously. > > > > > > > > > > > > > > > > > > Is that really broken? I thought that the Linux kernel deliberately > > > > > > > > > didn't care about strict aliasing rules (the top-level Makefile passes > > > > > > > > > -fno-strict-aliasing) so I thought that it was valid in "Linux kernel > > > > > > > > > C" even though from a standards point of view it is invalid. (That > > > > > > > > > being said, this is probably moot with my proposed changes below > > > > > > > > > though.) > > > > > > > > > > > > > > > > I have a feeling that -fno-strict-aliasing only allows you to _read_ a > > > > > > > > different union member from the one previously written. > > > > > > > > > > > > > > > > Writing a different member from the last one written can still splatter > > > > > > > > on the other members IIUC. > > > > > > > > > > > > > > > > It would be better to keep things separate rather than risk > > > > > > > > incorrectness just to save a few bytes. > > > > > > > > > > > > > > > > IMHO -fno-strict-aliasing is no excuse for gratuitous type-punning. > > > > > > > > > > > > > > > > > > So I am looking for something like force_sig_mcerr or force_sig_fault > > > > > > > > > > that includes your new information that then calls force_sig_info. > > > > > > > > > > > > > > > > > > > > I know of no other way to safely use the siginfo struct. > > > > > > > > > > > > > > > > > > So you want something like: > > > > > > > > > > > > > > > > > > int force_sig_fault_with_ignored_bits(int signo, int code, void __user > > > > > > > > > *addr, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > int force_sig_mceerr_with_ignored_bits(int code, void __user *addr, > > > > > > > > > short lsb, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > > > > > > > > > in kernel/signal.c and the code in arch/arm64 would call that? > > > > > > > > > > > > > > > > > > > > AIUI we absolutely need a forced signal here, we need to supply > > > > > > > > > > > metadata, and we don't have to open-code all that at every relevant > > > > > > > > > > > signal generation site... > > > > > > > > > > > > > > > > > > > > > >> It is not clear to me that if you have adapted siginfo_layout. > > > > > > > > > > > > > > > > > > > > > > Garbled sentence? > > > > > > > > > > > > > > > > > > > > Looks like. One of the pieces of code that needs to change > > > > > > > > > > when siginfo gets updated is siginfo_layout so that the structure > > > > > > > > > > can be properly decoded and made sense of. > > > > > > > > > > > > > > > > > > > > I am not seeing anything like that. > > > > > > > > > > > > > > > > > > Okay, this has to do with copying between the compat and non-compat > > > > > > > > > versions of the struct? Sure, I can update that, although the code > > > > > > > > > would be basically non-functional on arm64 because TBI isn't supported > > > > > > > > > on 32-bit ARM. > > > > > > > > > > > > > > > > > > > >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > > > > > > > > > >> > index cb3d6c267181..6dd82373eb2d 100644 > > > > > > > > > > >> > --- a/include/uapi/asm-generic/siginfo.h > > > > > > > > > > >> > +++ b/include/uapi/asm-generic/siginfo.h > > > > > > > > > > >> > @@ -91,6 +91,14 @@ union __sifields { > > > > > > > > > > >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > > > > > > > >> > __u32 _pkey; > > > > > > > > > > >> > } _addr_pkey; > > > > > > > > > > >> > +#ifdef __aarch64__ > > > > > > > > > > >> > + /* used with all si_codes */ > > > > > > > > > > >> > + struct { > > > > > > > > > > >> > + short _dummy_top_byte; > > > > > > > > > > > > > > > > > > > > > > ^ What's this for? I don't have Eric's insight here. > > > > > > > > > > > > > > > > > > We would need a short's worth of padding in order to prevent the > > > > > > > > > fields from occupying the same address as si_addr_lsb. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >> > + unsigned char _top_byte; > > > > > > > > > > >> > + unsigned char _top_byte_mask; > > > > > > > > > > >> > + } _addr_top_byte; > > > > > > > > > > >> > +#endif > > > > > > > > > > >> > }; > > > > > > > > > > >> > } _sigfault; > > > > > > > > > > >> > > > > > > > > > > > >> > > > > > > > > > > >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > > > > > > > > > > >> > > > > > > > > > > >> Please remove the "#ifdef __aarch64__". If at all possible we want to > > > > > > > > > > >> design this so any other architecture who has this challenge can use the > > > > > > > > > > >> code. The kind of code does not get enough attention/maintenance if it > > > > > > > > > > >> is built for a single architecture. > > > > > > > > > > > > > > > > > > Seems reasonable. I was recently made aware that RISC-V was > > > > > > > > > considering a similar feature: > > > > > > > > > https://lists.riscv.org/g/tech-tee/topic/risc_v_tbi_proposal/72855478 > > > > > > > > > I would have opted to expand this to other architectures on an > > > > > > > > > as-needed basis, but I'd also be fine with having it on all > > > > > > > > > architectures from the start. > > > > > > > > > > > > > > > > > > If we make this arch-independent, we have an additional concern, which > > > > > > > > > is "what if some future architecture wants more than one byte here?" > > > > > > > > > For example, an architecture may have a "top-two-bytes-ignore" > > > > > > > > > feature, which would imply two-byte (misnamed) "si_addr_top_byte" and > > > > > > > > > "si_addr_top_byte_mask" fields. And the RISC-V proposal potentially > > > > > > > > > implies many more ignored bits (see slide 13 of the presentation). The > > > > > > > > > maximum size that these fields can possibly be is the size of a > > > > > > > > > pointer, and with that there wouldn't be enough room in the padding at > > > > > > > > > this point to accommodate the new fields. > > > > > > > > > > > > > > > > > > That basically implies your earlier suggestion of adding a union > > > > > > > > > member here to accommodate future expansion of the union, and adding > > > > > > > > > the new fields after the union. I'm happy to make that change, with > > > > > > > > > the fields renamed "si_addr_ignored" and "si_addr_ignored_mask". > > > > > > > > > > > > > > > > I think what we need here is basically a flags word. > > > > > > > > > > > > > > > > So long as we keep a flag spare to indicate the existence of a further > > > > > > > > flags word, we can extend as needed. > > > > > > > > > > > > > > > > How the existence of the first flags words is detected is another > > > > > > > > problem. If it only applies for newly-defined si_code values, then > > > > > > > > I guess si_code may be sufficient. > > > > > > > > > > > > > > Existing kernels will zero-initialize unused regions of the siginfo > > > > > > > data structure. The zero-initialization of the padding at the end of > > > > > > > the struct is done by the clear_user call here: > > > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/kernel/signal.c#L3193 > > > > > > > > > > > > > > and the zero-initialization of the padding between fields and unused > > > > > > > union members is done by the clear_siginfo function which the kernel > > > > > > > calls when initializing the data structure: > > > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/include/linux/signal.h#L20 > > > > > > > > > > > > > > Therefore, a flag word value of 0 may be used to detect a lack of > > > > > > > support for flagged fields. > > > > > > > > > > > > It's not enough that we do this today. We would have had to do it back > > > > > > to the dawn of time (though in the arm64 case I guess we just need to go > > > > > > back to when the arch/arm64 was merged). > > > > > > > > > > > > v2.6.12:kernel/signal.c:copy_siginfo_to_user() suggests that this wasn't > > > > > > always the case, so unused parts of siginfo could be full of old junk > > > > > > from the user stack, if the kernel is sufficiently old. > > > > > > > > > > > > If we're trying to do something generic that makes sense on all arches, > > > > > > this matters. I may have misunderstood something about the code though. > > > > > > > > > > Hmm, I think you're right. The current behavior was introduced by > > > > > commit c999b933faa5e281e3af2e110eccaf91698b0a81 which was first > > > > > released in 4.18. So if an application wants to be compatible with > > > > > pre-4.18 kernels then there would need to be some other way to > > > > > indicate that the fields are valid. Probably the simplest way would be > > > > > to have the application issue a uname(2) syscall and check the kernel > > > > > version before reading these fields. I have a couple of other ideas > > > > > that don't rely on version detection, if we'd prefer to avoid that. > > > > > (They are somewhat ugly, but our hand is forced by backwards > > > > > compatibility.) > > > > > > > > > > One idea is to re-purpose the si_errno field as a flags field for > > > > > certain signal numbers. I checked a few kernel releases going back to > > > > > 2.6.18 and it looks like the field is set to 0 except in the following > > > > > circumstances: > > > > > - sending a hardware breakpoint (SIGTRAP/TRAP_HWBKPT) > > > > > - seccomp failures (SIGSYS/SYS_SECCOMP) > > > > > - user-defined signal via kill_pid_usb_asyncio > > > > > - SIGSWI in 3.18 and before (code since removed) > > > > > > > > > > It is also set to EFAULT for certain SIGSEGV/SEGV_MAPERR signals on > > > > > powerpc since commit c96c4436aba4c12f1f48369f2f90bc43e12fe36c, which > > > > > is currently unreleased. So if we wanted to go this route for SIGSEGV > > > > > we would need to stop the kernel from setting si_errno to EFAULT for > > > > > this signal before the 5.8 release. > > > > > > > > > > Another idea was to have userspace set a flag in sa_flags when > > > > > registering a signal handler meaning "this signal handler requires > > > > > unknown siginfo fields to be zeroed", and have existing kernels reject > > > > > the syscall due to an unknown flag being set, but unfortunately this > > > > > won't work because existing kernels do not reject sigaction syscalls > > > > > with unknown flags set in sa_flags. A perhaps more radical idea in > > > > > this vein would be to claim some of the upper bits of the signal > > > > > number as flags that will cause the syscall to be rejected if set and > > > > > unknown to the kernel. Existing kernels (going back to at least > > > > > 2.6.18) contain this code in do_sigaction: > > > > > > > > > > if (!valid_signal(sig) || sig < 1 || (act && sig_kernel_only(sig))) > > > > > return -EINVAL; > > > > > > > > > > and vald_signal is defined as: > > > > > > > > > > static inline int valid_signal(unsigned long sig) > > > > > { > > > > > return sig <= _NSIG ? 1 : 0; > > > > > } > > > > > > > > > > All architectures define _NSIG as a value <= 128, so they will reject > > > > > a signal number with any of bits 8-31 set. This means that we can use > > > > > any of those bits for mandatory flags. Most likely we could use bit 30 > > > > > (expanding down as necessary), as it keeps the signal number positive > > > > > and permits future expansion of the signal number range. > > > > > > > > Does the signal core code actually gurantee to zero the unused fields? > > > > Unless the fields are poked in by hand this is fraught with subtlelies, > > > > especially when unions are involved. (I'm sure the code tries to do it, > > > > but I've not eyeballed it in detail...) > > > > > > It memsets the siginfo structure before setting the fields and sending > > > the signal (grep for clear_siginfo which is just a memset; you should > > > find a call before all callers of force_sig_info). Memset is the right > > > approach here since unlike setting fields by hand it clears padding > > > which could lead to information leaks from the kernel. IIUC this is > > > the reason why Eric wants all of the signals to be raised via wrappers > > > in kernel/signal.c instead of via force_sig_info directly (to make > > > this aspect easier to audit). > > > > My impression was that the reason for this model is partly to ensure > > that siginfo fields are populated more consistently. When it was all > > down to the individual callers, inconsistencies creeped in. > > > > With regard to memset(), this is not a complete defence against data > > leakage. Assigning to a struct member can set any or all padding in > > the struct to random garbage (consider write-combining of neighboring > > member writes into a single larger accesses in asm for example). The > > I don't believe that LLVM will store to padding like this. I don't > know about GCC, though, but I wouldn't be surprised if this is > something that the kernel would want to turn off in "kernel C" (like > it turns off strict aliasing) specifically because of the information > leak issue. Again, the issue is not future kernel builds -- we can always find a way to fix the behaviour for those -- but past kernel builds. > > only way to avoid this is to ensure that the struct is 100% > > padding-free, and that each member of a union is the same size. A > > quick clance at <uapi/asm-generic/siginfo.h> confirms that this is not > > the case. > > > > This might need to be looked at separately. > > > > But it does mean, strictly speaking, that we can't reliably add new > > fields anywhere that there was previously padding: assigning to > > neighboring members can still fill those with garbage after the > > memset(). > > ...but this is largely moot because I'm not proposing to add new > fields in the padding any more (because the fields needed to become > larger in order to accommodate future hypothetical architectures which > might want to use the fields, and thus they wouldn't fit in the > padding). The siginfo.h diff would be something like: > > diff --git a/include/uapi/asm-generic/siginfo.h > b/include/uapi/asm-generic/siginfo.h > index cb3d6c267181..4a2fe257415d 100644 > --- a/include/uapi/asm-generic/siginfo.h > +++ b/include/uapi/asm-generic/siginfo.h > @@ -91,7 +91,10 @@ union __sifields { > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > __u32 _pkey; > } _addr_pkey; > + void *_pad[6]; > }; > + uintptr_t _ignored_bits; > + uintptr_t _ignored_bits_mask; This _is_ in padding: the tail-padding of the (previously smaller) _sigfault. Again, the compiler was allowed to populate this area with junk before these fields were added. I agree that it seems fairly unlikely that the compiler would have been overwriting this in normal circumstances, but that's not a guarantee. My worry is that if this goes wrong, it will go wrong silently and unpredictably. > } _sigfault; > > /* SIGPOLL */ > > or with a "uintptr_t _flags" added in before _ignored_bits if we go with that. > > > > > Using unused bits in the signal number to turn on new functionality > > > > feels risky. As currently specified, this is just a number. Since > > > > today a successful sigaction(n ...) guarantees that n is a valid signal > > > > number, reasonable code like the following would trigger a buffer > > > > overrun if we start trying to encode anything else in there: > > > > > > > > struct sigaction actions[NSIG]; > > > > > > > > int do_something( ... ) > > > > { > > > > ... > > > > > > > > if (!sigaction(n, sa, ...)) { > > > > actions[n] = *sa; > > > > return 0; > > > > } > > > > > > > > ... > > > > } > > > > > > I imagine the bit in the signal number being set by the direct caller > > > to sigaction, and we could specifically recommend that calling > > > pattern. In that case, your "n" wouldn't have the bit set in it. It > > > > I can imagine this too, but that doesn't mean that software does it. > > > > If the above kind of thing exists in a framework or library somewhere, > > we could get problems. Similarly, a pre-existing LD_PRELOAD framework > > that provides a wrapper for sigaction may now go wrong even if your > > pattern is followed -- i.e., the caller thinks it's calling sigaction > > directly but in fact it isn't. > > I'm aware of one library like that. It's called libsigchain, and it > has an early bounds check: > https://cs.android.com/android/platform/superproject/+/master:art/sigchainlib/sigchain.cc;l=371 > > Until the library is changed to recognize the flag, calling code would > see the return value of -1 as if the kernel failed the syscall, and > would fall back to the code for old kernels. But only after some bad dereferences. If these were writes, this means that memory _may_ be silently corrupted (I don't say it't likely in a given case, and we cannot pick a flag bit that makes this impossible). So, _even though the user program is correct_, our change may trigger the corruption of arbitrary user memory. This what I mean by an ABI break. The fact that the corruption is not done by the syscall itself is no excuse. We also fail to notice failures in sigaddset() etc., though in this code it looks like that should not matter. > In general I think that any library like this with independent > tracking of the kernel's purported signal handler state would need to > be very sensitive to which syscalls are capable of setting signal > handlers, what their semantics are, and so on. This applies to any > change that we might make to the signal handler interface. So for > example, if we introduced a new syscall as you propose below, and the > library hasn't been updated to recognize the new syscall, it will > silently miss changes in signal handler state caused by the new > syscall. > > At the end of this argument lies "we can never change anything about > how signal handlers work because it could break some interposing > library somewhere" -- replace "signal handlers" with any kernel > feature whose behavior may be modified by an interposing library if > you like -- and I don't think we want to go that far. As far as I > know, this isn't really the kernel's business anyway -- the kernel's > stable ABI contract starts and ends with the syscall interface and not > some library on top. > > That being said, we should perhaps try to define our interface so that > something reasonable will probably happen if there is such a library > and it hasn't been updated. With the new syscall, the library will > sometimes silently fail to work in some non-local fashion. With the > flag bit in the signal number, the library will either cause the > caller to fall back to the old kernel code path (if there is a bounds > check) or likely crash loudly (if there is no bounds check). To me, > the "flag bit in the signal number" behavior seems more reasonable, > since either something correct or something easy to debug will > probably happen at runtime. > > > > could only appear in newly-written code that doesn't follow our > > > recommendations, and there are already plenty of much more likely ways > > > to cause buffer overflows in C code that doesn't follow > > > recommendations anyway. (And even if such a buffer overflow occurred, > > > it would very likely be caught early in development by the MMU due to > > > the magnitude of the number 1<<30.) > > > > Choosing the bit value is hard. If shitfing it overflows, this can > > trigger random undefined behaviour in the compiler in addition to (or > > perhaps instead of) an out-of-bounds access or segfault. > > It wouldn't overflow on a 64-bit architecture assuming normal array > indexing (the index would be promoted to pointer width before being > scaled to the array element size), and to begin with the users of this > would be 64-bit. Unless we don't offer this feature for 32-bit at all (possible, if ugly) we can't stop people using it. > > If shifting it doesn't overflow, we might still fall into a valid > > mapping, though I'd agree a segfault is more likely. > > > > > > > > > I think it would be cleaner for to add a single flag field that can be > > > > used for detecting other extensions, and request it via a new sa_flags > > > > bit. This removes the need for sematically useless zeroing of unused > > > > fields (though for hygiene and backwards compatibility reasons we would > > > > probably want to carry on zeroing them anyway). > > > > > > > > I can see no simpler way to add supplementary siginfo fields for > > > > existing si_codes. For si_codes that didn't exist before the zeroing > > > > came in we could still detect optional si_code-specific fields via > > > > zeroing, but it seems messary to have two ways of detecting extensions. > > > > > > That would certainly be cleaner if it worked, but that would only be > > > the case if old kernels rejected unknown bits in sa_flags, and > > > unfortunately they don't. With the bit in the signal number, the "old > > > > Hmm, that is a problem I wasn't aware of. > > > > > kernels reject" behavior admits relatively straightforward usage code: > > > > > > void set_segv_handler(void) { > > > struct sigaction sa; > > > sa.sa_sigaction = handle_segv; > > > sa.sa_flags = SA_SIGINFO; > > > if (sigaction(SIGSEGV | SF_CLEAR_UNKNOWN_FIELDS, &sa, 0) < 0) { // > > > succeeds in new kernels, fails in old kernels > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > if (sigaction(SIGSEGV, &sa, 0) < 0) // succeeds in old kernels > > > perror("sigaction"); > > > } > > > } > > > > > > void clear_fields_and_handle_segv(int signum, siginfo_t *sa, void *ctx) { > > > sa->si_future_field = 0; > > > handle_segv(signum, sa, ctx); > > > } > > > > > > void handle_segv(int signum, siginfo_t *sa, void *ctx) { > > > // At this point, si_future_field will have the value 0 in old > > > kernels and the kernel-supplied value in new kernels. > > > } > > > > > > Imagine if we moved the flag SF_CLEAR_UNKNOWN_FIELDS from the signal > > > number to sa_flags. In that case, the first sigaction would succeed in > > > old kernels so handle_segv wouldn't know whether it can safely read > > > from si_future_field. With the sa_flags approach, you would need > > > kernel version number checking via uname before setting the flag in > > > sa_flags, and at that point why even have the flag in sa_flags at all > > > since you could just have the signal handler conditionally read from > > > si_future_field based on the uname? > > > > Software setting SA_SIFLAGS (or whatever) is new by definition, since > > it would be using a new #define. So it might be reasonable to put the > > burden on that software to verify that the flag was really accepted by > > the kernel, by reading it back. > > That doesn't seem like a good idea even if it worked, because it could > lead to race conditions. If the si_flags-reading signal handler were > invoked in response to a signal between when you set it and when you > ended up replacing it with the fallback signal handler for old > kernels, the handler may end up reading garbage data from si_flags. Not really. My example may have this problem, but the signal handler can be written to support both scenarios, based on testing a flag that the main program sets after verifying that the flag could be set. Or the signal could be blocked around establishment (often a good idea for other reasons). But I agree it's a bit gross, and anyway doesn't work due to the fact that the kernel doesn't filter out unrecognised flags anyway. > > Unfortunately, even relatively recent kernels blindly store sa_flags > > in the kernel without validating it, and so it looks like duff flags > > can be read back out via a sigaction() call. Dang. > > > > > > Perhaps a new frontend syscall could be added. A new libc that knows > > about this "sigaction2" could use it and mask off problem bits from > > sa_flags in its sigaction() wrapper before calling sigaction2. An old > > libc would call the old sigaction syscall, where we would ignore these > > new sa_flags bits as before. > > I'm not currently in favor of the new syscall but if we do this I > would keep sigaction and sigaction2 separate. That is, libc sigaction > should always use the sigaction syscall, and libc sigaction2 should > always use the sigaction2 syscall. We should avoid libc's sigaction > having different behavior based on the libc version and kernel > version, as that would make it harder to reason about its behavior. > Calling code would need to check for presence of sigaction2 in both > libc and the kernel, e.g. > > __attribute__((weak)) decltype(sigaction2) sigaction2; > > void set_segv_handler(void) { > struct sigaction sa; > sa.sa_sigaction = handle_segv; > sa.sa_flags = SA_SIGINFO | SA_SIFLAGS; > if (!sigaction2 || sigaction2(SIGSEGV, &sa, 0) < 0) { > sa.sa_sigaction = clear_fields_and_handle_segv; > sa.sa_flags = SA_SIGINFO; > if (sigaction(SIGSEGV, &sa, 0) < 0) > perror("sigaction"); > } > } I guess. But I share your distaste for adding a new syscall. > > > This may not be a popular approach though, and software wouldn't be able > > to use our new features until libc is updated to match. > > > > If we go down this route, it may provide additional opportunities to fix > > annoying defects in the old interface. > > > > > > > Note that the same applies to a flag indicating the availability of a > > > si_flags field in sigaction (just > > > s/SF_CLEAR_UNKNOWN_FIELDS/SF_HAS_SI_FLAGS/ and > > > s/si_future_field/si_flags/ in the usage code above). In terms of > > > SF_CLEAR_UNKNOWN_FIELDS versus SF_HAS_SI_FLAGS I'd be fine either way. > > > > > > Another thought that occurred to me is that we may consider > > > generalizing this a step further and introducing a single flag bit in > > > the signal number that means "reject unknown flags in sa_flags". This > > > would mean that we wouldn't need to add any more flag bits to the > > > signal number in the future, thus limiting this signal number hack to > > > a single bit; all future mandatory behavior changes could just be put > > > behind a flag in sa_flags and userspace code would easily be able to > > > detect missing support for a flag and fall back if necessary. In our > > > case, this would imply usage code like this: > > > > > > void set_segv_handler(void) { > > > struct sigaction sa; > > > sa.sa_sigaction = handle_segv; > > > sa.sa_flags = SA_SIGINFO | SA_CLEAR_UNKNOWN_FIELDS; > > > // Succeeds in kernels with SA_CLEAR_UNKNOWN_FIELDS support. > > > // Fails in kernels with SF_CHECK_SA_FLAGS support but no > > > SA_CLEAR_UNKNOWN_FIELDS support (because of the unknown flags check). > > > // Fails in kernels without SF_CHECK_SA_FLAGS support (because of > > > the bounds check on the signal number). > > > if (sigaction(SIGSEGV | SF_CHECK_SA_FLAGS, &sa, 0) < 0) { > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > sa.sa_flags = SA_SIGINFO; > > > // Succeeds in old kernels, no need to use SF_CHECK_SA_FLAGS since > > > we're using sa_flags from the beginning of time. > > > if (sigaction(SIGSEGV, &sa, 0) < 0) > > > perror("sigaction"); > > > } > > > } > > > > As with the other options this could work, but looks like it could > > break the ABI due to violating the original semantics for the signal > > number argument. Perhaps I'm being too paranoid. > > There's no ABI being broken here, as long as we consider syscalls to > be the stable ABI layer. Old kernels are simply rejecting arguments > that they don't know about yet. By that argument, any introduction of > a new syscall is an ABI break because it changes the semantics of a > previously-unallocated syscall number. As argued above, I think this is an invalid argument. Although any addition will change behaviour (so is a break in some sense), the key is not to make "surprising" changes. Having something random happen when setting a previously reserved flag bit, or when issuing a syscall when an unknown syscall number, or not surprising at all. Making fundamental changes to the encoding of an existing argument is highly surprising, on the other hand: as your example shows, it is reasonable to index an array using a signal number. I agree that this doesn't get us closer to a practical solution though. But we do seem to need some mechanism in addition to (or instead of) sa_flags. Here's another thought: Since si_flags would be either always present or always absent, it could make sense to have a global property to report this, rather than an sa_flags or signal number bit to request it per-signal. Requiring software to parse uname() might be reasonable for that, if cumbersome (did you suggest this previously?). If we're concerned that the awkwardness of this would encourage people not to bother (or encourage people to do it wrong) then we might opt for something simpler like an AT_FLAGS bit. Ultimately libc could provide a more portable interface for discovery, such as via sysconf(). Thoughts? ---Dave
On Wed, Jul 08, 2020 at 12:00:22PM +0100, Dave Martin wrote: > On Tue, Jul 07, 2020 at 12:07:09PM -0700, Peter Collingbourne wrote: > > On Tue, Jul 7, 2020 at 7:19 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > On Mon, Jul 06, 2020 at 12:20:33PM -0700, Peter Collingbourne wrote: > > > > On Mon, Jul 6, 2020 at 9:41 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > On Wed, Jun 24, 2020 at 12:51:43PM -0700, Peter Collingbourne wrote: > > > > > > On Wed, Jun 24, 2020 at 10:12 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > On Wed, Jun 24, 2020 at 09:51:49AM -0700, Peter Collingbourne wrote: > > > > > > > > On Wed, Jun 24, 2020 at 2:28 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 05:40:08PM -0700, Peter Collingbourne wrote: > > > > > > > > > > On Tue, Jun 23, 2020 at 10:52 AM Eric W. Biederman > > > > > > > > > > <ebiederm@xmission.com> wrote: > > > > > > > > > > > > > > > > > > > > > > Dave Martin <Dave.Martin@arm.com> writes: > > > > > > > > > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > > > > > > > > > > > >> Peter Collingbourne <pcc@google.com> writes: > > > > > > > > > > > >> > > > > > > > > > > > >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > > > > > > > > > > > >> > index 47f651df781c..a8380a2b6361 100644 > > > > > > > > > > > >> > --- a/arch/arm64/kernel/traps.c > > > > > > > > > > > >> > +++ b/arch/arm64/kernel/traps.c > > > > > > > > > > > >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > > > > > > > > > > > >> > } > > > > > > > > > > > >> > > > > > > > > > > > > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, > > > > > > > > > > > >> > + unsigned long far, unsigned char far_tb_mask, > > > > > > > > > > > >> > const char *str) > > > > > > > > > > > >> > { > > > > > > > > > > > >> > arm64_show_signal(signo, str); > > > > > > > > > > > >> > - if (signo == SIGKILL) > > > > > > > > > > > >> > + if (signo == SIGKILL) { > > > > > > > > > > > >> > force_sig(SIGKILL); > > > > > > > > > > > >> > - else > > > > > > > > > > > >> > - force_sig_fault(signo, code, addr); > > > > > > > > > > > >> > + } else { > > > > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > > > > >> > + info.si_signo = signo; > > > > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > > > > >> > + info.si_code = code; > > > > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > > > > >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > > > > > > > > > > > >> > + info.si_addr_top_byte_mask = far_tb_mask; > > > > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > > > > >> > + } > > > > > > > > > > > >> > } > > > > > > > > > > > >> > > > > > > > > > > > > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > > > > > > > > > > > >> > - const char *str) > > > > > > > > > > > >> > + unsigned long far, const char *str) > > > > > > > > > > > >> > { > > > > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > > > > >> > + > > > > > > > > > > > >> > arm64_show_signal(SIGBUS, str); > > > > > > > > > > > >> > - force_sig_mceerr(code, addr, lsb); > > > > > > > > > > > >> > + > > > > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > > > > >> > + info.si_signo = SIGBUS; > > > > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > > > > >> > + info.si_code = code; > > > > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > > > > >> > + info.si_addr_lsb = lsb; > > > > > > > > > > > >> > + info.si_addr_top_byte = far >> 56; > > > > > > > > > > > >> > + info.si_addr_top_byte_mask = 0xff; > > > > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > > > > >> > } > > > > > > > > > > > >> > > > > > > > > > > > >> I have a real problem with this construction. force_sig_info is not an > > > > > > > > > > > >> interface that should be used for anything except to define a wrapper > > > > > > > > > > > >> that takes it's parameters. > > > > > > > > > > > > > > > > > > > > > > > > Can you elaborate? How would you do this king of thing. > > > > > > > > > > > > > > > > > > > > > > There are no other uses of force_sig_info in architecture code. > > > > > > > > > > > > > > > > > > > > > > I just removed them _all_ because they were almost all broken. > > > > > > > > > > > In fact your mcerr case is broken because it uses two different > > > > > > > > > > > union members simultantiously. > > > > > > > > > > > > > > > > > > > > Is that really broken? I thought that the Linux kernel deliberately > > > > > > > > > > didn't care about strict aliasing rules (the top-level Makefile passes > > > > > > > > > > -fno-strict-aliasing) so I thought that it was valid in "Linux kernel > > > > > > > > > > C" even though from a standards point of view it is invalid. (That > > > > > > > > > > being said, this is probably moot with my proposed changes below > > > > > > > > > > though.) > > > > > > > > > > > > > > > > > > I have a feeling that -fno-strict-aliasing only allows you to _read_ a > > > > > > > > > different union member from the one previously written. > > > > > > > > > > > > > > > > > > Writing a different member from the last one written can still splatter > > > > > > > > > on the other members IIUC. > > > > > > > > > > > > > > > > > > It would be better to keep things separate rather than risk > > > > > > > > > incorrectness just to save a few bytes. > > > > > > > > > > > > > > > > > > IMHO -fno-strict-aliasing is no excuse for gratuitous type-punning. > > > > > > > > > > > > > > > > > > > > So I am looking for something like force_sig_mcerr or force_sig_fault > > > > > > > > > > > that includes your new information that then calls force_sig_info. > > > > > > > > > > > > > > > > > > > > > > I know of no other way to safely use the siginfo struct. > > > > > > > > > > > > > > > > > > > > So you want something like: > > > > > > > > > > > > > > > > > > > > int force_sig_fault_with_ignored_bits(int signo, int code, void __user > > > > > > > > > > *addr, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > int force_sig_mceerr_with_ignored_bits(int code, void __user *addr, > > > > > > > > > > short lsb, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > > > > > > > > > > > in kernel/signal.c and the code in arch/arm64 would call that? > > > > > > > > > > > > > > > > > > > > > > AIUI we absolutely need a forced signal here, we need to supply > > > > > > > > > > > > metadata, and we don't have to open-code all that at every relevant > > > > > > > > > > > > signal generation site... > > > > > > > > > > > > > > > > > > > > > > > >> It is not clear to me that if you have adapted siginfo_layout. > > > > > > > > > > > > > > > > > > > > > > > > Garbled sentence? > > > > > > > > > > > > > > > > > > > > > > Looks like. One of the pieces of code that needs to change > > > > > > > > > > > when siginfo gets updated is siginfo_layout so that the structure > > > > > > > > > > > can be properly decoded and made sense of. > > > > > > > > > > > > > > > > > > > > > > I am not seeing anything like that. > > > > > > > > > > > > > > > > > > > > Okay, this has to do with copying between the compat and non-compat > > > > > > > > > > versions of the struct? Sure, I can update that, although the code > > > > > > > > > > would be basically non-functional on arm64 because TBI isn't supported > > > > > > > > > > on 32-bit ARM. > > > > > > > > > > > > > > > > > > > > > >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > >> > index cb3d6c267181..6dd82373eb2d 100644 > > > > > > > > > > > >> > --- a/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > >> > +++ b/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > >> > @@ -91,6 +91,14 @@ union __sifields { > > > > > > > > > > > >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > > > > > > > > >> > __u32 _pkey; > > > > > > > > > > > >> > } _addr_pkey; > > > > > > > > > > > >> > +#ifdef __aarch64__ > > > > > > > > > > > >> > + /* used with all si_codes */ > > > > > > > > > > > >> > + struct { > > > > > > > > > > > >> > + short _dummy_top_byte; > > > > > > > > > > > > > > > > > > > > > > > > ^ What's this for? I don't have Eric's insight here. > > > > > > > > > > > > > > > > > > > > We would need a short's worth of padding in order to prevent the > > > > > > > > > > fields from occupying the same address as si_addr_lsb. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >> > + unsigned char _top_byte; > > > > > > > > > > > >> > + unsigned char _top_byte_mask; > > > > > > > > > > > >> > + } _addr_top_byte; > > > > > > > > > > > >> > +#endif > > > > > > > > > > > >> > }; > > > > > > > > > > > >> > } _sigfault; > > > > > > > > > > > >> > > > > > > > > > > > > >> > > > > > > > > > > > >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > > > > > > > > > > > >> > > > > > > > > > > > >> Please remove the "#ifdef __aarch64__". If at all possible we want to > > > > > > > > > > > >> design this so any other architecture who has this challenge can use the > > > > > > > > > > > >> code. The kind of code does not get enough attention/maintenance if it > > > > > > > > > > > >> is built for a single architecture. > > > > > > > > > > > > > > > > > > > > Seems reasonable. I was recently made aware that RISC-V was > > > > > > > > > > considering a similar feature: > > > > > > > > > > https://lists.riscv.org/g/tech-tee/topic/risc_v_tbi_proposal/72855478 > > > > > > > > > > I would have opted to expand this to other architectures on an > > > > > > > > > > as-needed basis, but I'd also be fine with having it on all > > > > > > > > > > architectures from the start. > > > > > > > > > > > > > > > > > > > > If we make this arch-independent, we have an additional concern, which > > > > > > > > > > is "what if some future architecture wants more than one byte here?" > > > > > > > > > > For example, an architecture may have a "top-two-bytes-ignore" > > > > > > > > > > feature, which would imply two-byte (misnamed) "si_addr_top_byte" and > > > > > > > > > > "si_addr_top_byte_mask" fields. And the RISC-V proposal potentially > > > > > > > > > > implies many more ignored bits (see slide 13 of the presentation). The > > > > > > > > > > maximum size that these fields can possibly be is the size of a > > > > > > > > > > pointer, and with that there wouldn't be enough room in the padding at > > > > > > > > > > this point to accommodate the new fields. > > > > > > > > > > > > > > > > > > > > That basically implies your earlier suggestion of adding a union > > > > > > > > > > member here to accommodate future expansion of the union, and adding > > > > > > > > > > the new fields after the union. I'm happy to make that change, with > > > > > > > > > > the fields renamed "si_addr_ignored" and "si_addr_ignored_mask". > > > > > > > > > > > > > > > > > > I think what we need here is basically a flags word. > > > > > > > > > > > > > > > > > > So long as we keep a flag spare to indicate the existence of a further > > > > > > > > > flags word, we can extend as needed. > > > > > > > > > > > > > > > > > > How the existence of the first flags words is detected is another > > > > > > > > > problem. If it only applies for newly-defined si_code values, then > > > > > > > > > I guess si_code may be sufficient. > > > > > > > > > > > > > > > > Existing kernels will zero-initialize unused regions of the siginfo > > > > > > > > data structure. The zero-initialization of the padding at the end of > > > > > > > > the struct is done by the clear_user call here: > > > > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/kernel/signal.c#L3193 > > > > > > > > > > > > > > > > and the zero-initialization of the padding between fields and unused > > > > > > > > union members is done by the clear_siginfo function which the kernel > > > > > > > > calls when initializing the data structure: > > > > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/include/linux/signal.h#L20 > > > > > > > > > > > > > > > > Therefore, a flag word value of 0 may be used to detect a lack of > > > > > > > > support for flagged fields. > > > > > > > > > > > > > > It's not enough that we do this today. We would have had to do it back > > > > > > > to the dawn of time (though in the arm64 case I guess we just need to go > > > > > > > back to when the arch/arm64 was merged). > > > > > > > > > > > > > > v2.6.12:kernel/signal.c:copy_siginfo_to_user() suggests that this wasn't > > > > > > > always the case, so unused parts of siginfo could be full of old junk > > > > > > > from the user stack, if the kernel is sufficiently old. > > > > > > > > > > > > > > If we're trying to do something generic that makes sense on all arches, > > > > > > > this matters. I may have misunderstood something about the code though. > > > > > > > > > > > > Hmm, I think you're right. The current behavior was introduced by > > > > > > commit c999b933faa5e281e3af2e110eccaf91698b0a81 which was first > > > > > > released in 4.18. So if an application wants to be compatible with > > > > > > pre-4.18 kernels then there would need to be some other way to > > > > > > indicate that the fields are valid. Probably the simplest way would be > > > > > > to have the application issue a uname(2) syscall and check the kernel > > > > > > version before reading these fields. I have a couple of other ideas > > > > > > that don't rely on version detection, if we'd prefer to avoid that. > > > > > > (They are somewhat ugly, but our hand is forced by backwards > > > > > > compatibility.) > > > > > > > > > > > > One idea is to re-purpose the si_errno field as a flags field for > > > > > > certain signal numbers. I checked a few kernel releases going back to > > > > > > 2.6.18 and it looks like the field is set to 0 except in the following > > > > > > circumstances: > > > > > > - sending a hardware breakpoint (SIGTRAP/TRAP_HWBKPT) > > > > > > - seccomp failures (SIGSYS/SYS_SECCOMP) > > > > > > - user-defined signal via kill_pid_usb_asyncio > > > > > > - SIGSWI in 3.18 and before (code since removed) > > > > > > > > > > > > It is also set to EFAULT for certain SIGSEGV/SEGV_MAPERR signals on > > > > > > powerpc since commit c96c4436aba4c12f1f48369f2f90bc43e12fe36c, which > > > > > > is currently unreleased. So if we wanted to go this route for SIGSEGV > > > > > > we would need to stop the kernel from setting si_errno to EFAULT for > > > > > > this signal before the 5.8 release. > > > > > > > > > > > > Another idea was to have userspace set a flag in sa_flags when > > > > > > registering a signal handler meaning "this signal handler requires > > > > > > unknown siginfo fields to be zeroed", and have existing kernels reject > > > > > > the syscall due to an unknown flag being set, but unfortunately this > > > > > > won't work because existing kernels do not reject sigaction syscalls > > > > > > with unknown flags set in sa_flags. A perhaps more radical idea in > > > > > > this vein would be to claim some of the upper bits of the signal > > > > > > number as flags that will cause the syscall to be rejected if set and > > > > > > unknown to the kernel. Existing kernels (going back to at least > > > > > > 2.6.18) contain this code in do_sigaction: > > > > > > > > > > > > if (!valid_signal(sig) || sig < 1 || (act && sig_kernel_only(sig))) > > > > > > return -EINVAL; > > > > > > > > > > > > and vald_signal is defined as: > > > > > > > > > > > > static inline int valid_signal(unsigned long sig) > > > > > > { > > > > > > return sig <= _NSIG ? 1 : 0; > > > > > > } > > > > > > > > > > > > All architectures define _NSIG as a value <= 128, so they will reject > > > > > > a signal number with any of bits 8-31 set. This means that we can use > > > > > > any of those bits for mandatory flags. Most likely we could use bit 30 > > > > > > (expanding down as necessary), as it keeps the signal number positive > > > > > > and permits future expansion of the signal number range. > > > > > > > > > > Does the signal core code actually gurantee to zero the unused fields? > > > > > Unless the fields are poked in by hand this is fraught with subtlelies, > > > > > especially when unions are involved. (I'm sure the code tries to do it, > > > > > but I've not eyeballed it in detail...) > > > > > > > > It memsets the siginfo structure before setting the fields and sending > > > > the signal (grep for clear_siginfo which is just a memset; you should > > > > find a call before all callers of force_sig_info). Memset is the right > > > > approach here since unlike setting fields by hand it clears padding > > > > which could lead to information leaks from the kernel. IIUC this is > > > > the reason why Eric wants all of the signals to be raised via wrappers > > > > in kernel/signal.c instead of via force_sig_info directly (to make > > > > this aspect easier to audit). > > > > > > My impression was that the reason for this model is partly to ensure > > > that siginfo fields are populated more consistently. When it was all > > > down to the individual callers, inconsistencies creeped in. > > > > > > With regard to memset(), this is not a complete defence against data > > > leakage. Assigning to a struct member can set any or all padding in > > > the struct to random garbage (consider write-combining of neighboring > > > member writes into a single larger accesses in asm for example). The > > > > I don't believe that LLVM will store to padding like this. I don't > > know about GCC, though, but I wouldn't be surprised if this is > > something that the kernel would want to turn off in "kernel C" (like > > it turns off strict aliasing) specifically because of the information > > leak issue. > > Again, the issue is not future kernel builds -- we can always find a way > to fix the behaviour for those -- but past kernel builds. > > > > only way to avoid this is to ensure that the struct is 100% > > > padding-free, and that each member of a union is the same size. A > > > quick clance at <uapi/asm-generic/siginfo.h> confirms that this is not > > > the case. > > > > > > This might need to be looked at separately. > > > > > > But it does mean, strictly speaking, that we can't reliably add new > > > fields anywhere that there was previously padding: assigning to > > > neighboring members can still fill those with garbage after the > > > memset(). > > > > ...but this is largely moot because I'm not proposing to add new > > fields in the padding any more (because the fields needed to become > > larger in order to accommodate future hypothetical architectures which > > might want to use the fields, and thus they wouldn't fit in the > > padding). The siginfo.h diff would be something like: > > > > diff --git a/include/uapi/asm-generic/siginfo.h > > b/include/uapi/asm-generic/siginfo.h > > index cb3d6c267181..4a2fe257415d 100644 > > --- a/include/uapi/asm-generic/siginfo.h > > +++ b/include/uapi/asm-generic/siginfo.h > > @@ -91,7 +91,10 @@ union __sifields { > > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > __u32 _pkey; > > } _addr_pkey; > > + void *_pad[6]; > > }; > > + uintptr_t _ignored_bits; > > + uintptr_t _ignored_bits_mask; > > This _is_ in padding: the tail-padding of the (previously smaller) > _sigfault. Again, the compiler was allowed to populate this area with > junk before these fields were added. > > I agree that it seems fairly unlikely that the compiler would have been > overwriting this in normal circumstances, but that's not a guarantee. > My worry is that if this goes wrong, it will go wrong silently and > unpredictably. > > > } _sigfault; > > > > /* SIGPOLL */ > > > > or with a "uintptr_t _flags" added in before _ignored_bits if we go with that. > > > > > > > Using unused bits in the signal number to turn on new functionality > > > > > feels risky. As currently specified, this is just a number. Since > > > > > today a successful sigaction(n ...) guarantees that n is a valid signal > > > > > number, reasonable code like the following would trigger a buffer > > > > > overrun if we start trying to encode anything else in there: > > > > > > > > > > struct sigaction actions[NSIG]; > > > > > > > > > > int do_something( ... ) > > > > > { > > > > > ... > > > > > > > > > > if (!sigaction(n, sa, ...)) { > > > > > actions[n] = *sa; > > > > > return 0; > > > > > } > > > > > > > > > > ... > > > > > } > > > > > > > > I imagine the bit in the signal number being set by the direct caller > > > > to sigaction, and we could specifically recommend that calling > > > > pattern. In that case, your "n" wouldn't have the bit set in it. It > > > > > > I can imagine this too, but that doesn't mean that software does it. > > > > > > If the above kind of thing exists in a framework or library somewhere, > > > we could get problems. Similarly, a pre-existing LD_PRELOAD framework > > > that provides a wrapper for sigaction may now go wrong even if your > > > pattern is followed -- i.e., the caller thinks it's calling sigaction > > > directly but in fact it isn't. > > > > I'm aware of one library like that. It's called libsigchain, and it > > has an early bounds check: > > https://cs.android.com/android/platform/superproject/+/master:art/sigchainlib/sigchain.cc;l=371 > > > > Until the library is changed to recognize the flag, calling code would > > see the return value of -1 as if the kernel failed the syscall, and > > would fall back to the code for old kernels. > > But only after some bad dereferences. If these were writes, this means > that memory _may_ be silently corrupted (I don't say it't likely in a > given case, and we cannot pick a flag bit that makes this impossible). > > So, _even though the user program is correct_, our change may trigger > the corruption of arbitrary user memory. This what I mean by an ABI > break. The fact that the corruption is not done by the syscall itself > is no excuse. > > We also fail to notice failures in sigaddset() etc., though in this code > it looks like that should not matter. > > > In general I think that any library like this with independent > > tracking of the kernel's purported signal handler state would need to > > be very sensitive to which syscalls are capable of setting signal > > handlers, what their semantics are, and so on. This applies to any > > change that we might make to the signal handler interface. So for > > example, if we introduced a new syscall as you propose below, and the > > library hasn't been updated to recognize the new syscall, it will > > silently miss changes in signal handler state caused by the new > > syscall. > > > > At the end of this argument lies "we can never change anything about > > how signal handlers work because it could break some interposing > > library somewhere" -- replace "signal handlers" with any kernel > > feature whose behavior may be modified by an interposing library if > > you like -- and I don't think we want to go that far. As far as I > > know, this isn't really the kernel's business anyway -- the kernel's > > stable ABI contract starts and ends with the syscall interface and not > > some library on top. > > > > That being said, we should perhaps try to define our interface so that > > something reasonable will probably happen if there is such a library > > and it hasn't been updated. With the new syscall, the library will > > sometimes silently fail to work in some non-local fashion. With the > > flag bit in the signal number, the library will either cause the > > caller to fall back to the old kernel code path (if there is a bounds > > check) or likely crash loudly (if there is no bounds check). To me, > > the "flag bit in the signal number" behavior seems more reasonable, > > since either something correct or something easy to debug will > > probably happen at runtime. > > > > > > could only appear in newly-written code that doesn't follow our > > > > recommendations, and there are already plenty of much more likely ways > > > > to cause buffer overflows in C code that doesn't follow > > > > recommendations anyway. (And even if such a buffer overflow occurred, > > > > it would very likely be caught early in development by the MMU due to > > > > the magnitude of the number 1<<30.) > > > > > > Choosing the bit value is hard. If shitfing it overflows, this can > > > trigger random undefined behaviour in the compiler in addition to (or > > > perhaps instead of) an out-of-bounds access or segfault. > > > > It wouldn't overflow on a 64-bit architecture assuming normal array > > indexing (the index would be promoted to pointer width before being > > scaled to the array element size), and to begin with the users of this > > would be 64-bit. > > Unless we don't offer this feature for 32-bit at all (possible, if ugly) > we can't stop people using it. > > > > If shifting it doesn't overflow, we might still fall into a valid > > > mapping, though I'd agree a segfault is more likely. > > > > > > > > > > > > I think it would be cleaner for to add a single flag field that can be > > > > > used for detecting other extensions, and request it via a new sa_flags > > > > > bit. This removes the need for sematically useless zeroing of unused > > > > > fields (though for hygiene and backwards compatibility reasons we would > > > > > probably want to carry on zeroing them anyway). > > > > > > > > > > I can see no simpler way to add supplementary siginfo fields for > > > > > existing si_codes. For si_codes that didn't exist before the zeroing > > > > > came in we could still detect optional si_code-specific fields via > > > > > zeroing, but it seems messary to have two ways of detecting extensions. > > > > > > > > That would certainly be cleaner if it worked, but that would only be > > > > the case if old kernels rejected unknown bits in sa_flags, and > > > > unfortunately they don't. With the bit in the signal number, the "old > > > > > > Hmm, that is a problem I wasn't aware of. > > > > > > > kernels reject" behavior admits relatively straightforward usage code: > > > > > > > > void set_segv_handler(void) { > > > > struct sigaction sa; > > > > sa.sa_sigaction = handle_segv; > > > > sa.sa_flags = SA_SIGINFO; > > > > if (sigaction(SIGSEGV | SF_CLEAR_UNKNOWN_FIELDS, &sa, 0) < 0) { // > > > > succeeds in new kernels, fails in old kernels > > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > > if (sigaction(SIGSEGV, &sa, 0) < 0) // succeeds in old kernels > > > > perror("sigaction"); > > > > } > > > > } > > > > > > > > void clear_fields_and_handle_segv(int signum, siginfo_t *sa, void *ctx) { > > > > sa->si_future_field = 0; > > > > handle_segv(signum, sa, ctx); > > > > } > > > > > > > > void handle_segv(int signum, siginfo_t *sa, void *ctx) { > > > > // At this point, si_future_field will have the value 0 in old > > > > kernels and the kernel-supplied value in new kernels. > > > > } > > > > > > > > Imagine if we moved the flag SF_CLEAR_UNKNOWN_FIELDS from the signal > > > > number to sa_flags. In that case, the first sigaction would succeed in > > > > old kernels so handle_segv wouldn't know whether it can safely read > > > > from si_future_field. With the sa_flags approach, you would need > > > > kernel version number checking via uname before setting the flag in > > > > sa_flags, and at that point why even have the flag in sa_flags at all > > > > since you could just have the signal handler conditionally read from > > > > si_future_field based on the uname? > > > > > > Software setting SA_SIFLAGS (or whatever) is new by definition, since > > > it would be using a new #define. So it might be reasonable to put the > > > burden on that software to verify that the flag was really accepted by > > > the kernel, by reading it back. > > > > That doesn't seem like a good idea even if it worked, because it could > > lead to race conditions. If the si_flags-reading signal handler were > > invoked in response to a signal between when you set it and when you > > ended up replacing it with the fallback signal handler for old > > kernels, the handler may end up reading garbage data from si_flags. > > Not really. My example may have this problem, but the signal handler > can be written to support both scenarios, based on testing a flag that > the main program sets after verifying that the flag could be set. Or > the signal could be blocked around establishment (often a good idea for > other reasons). > > But I agree it's a bit gross, and anyway doesn't work due to the fact > that the kernel doesn't filter out unrecognised flags anyway. > > > > Unfortunately, even relatively recent kernels blindly store sa_flags > > > in the kernel without validating it, and so it looks like duff flags > > > can be read back out via a sigaction() call. Dang. > > > > > > > > > Perhaps a new frontend syscall could be added. A new libc that knows > > > about this "sigaction2" could use it and mask off problem bits from > > > sa_flags in its sigaction() wrapper before calling sigaction2. An old > > > libc would call the old sigaction syscall, where we would ignore these > > > new sa_flags bits as before. > > > > I'm not currently in favor of the new syscall but if we do this I > > would keep sigaction and sigaction2 separate. That is, libc sigaction > > should always use the sigaction syscall, and libc sigaction2 should > > always use the sigaction2 syscall. We should avoid libc's sigaction > > having different behavior based on the libc version and kernel > > version, as that would make it harder to reason about its behavior. > > Calling code would need to check for presence of sigaction2 in both > > libc and the kernel, e.g. > > > > __attribute__((weak)) decltype(sigaction2) sigaction2; > > > > void set_segv_handler(void) { > > struct sigaction sa; > > sa.sa_sigaction = handle_segv; > > sa.sa_flags = SA_SIGINFO | SA_SIFLAGS; > > if (!sigaction2 || sigaction2(SIGSEGV, &sa, 0) < 0) { > > sa.sa_sigaction = clear_fields_and_handle_segv; > > sa.sa_flags = SA_SIGINFO; > > if (sigaction(SIGSEGV, &sa, 0) < 0) > > perror("sigaction"); > > } > > } > > I guess. But I share your distaste for adding a new syscall. > > > > > > This may not be a popular approach though, and software wouldn't be able > > > to use our new features until libc is updated to match. > > > > > > If we go down this route, it may provide additional opportunities to fix > > > annoying defects in the old interface. > > > > > > > > > > Note that the same applies to a flag indicating the availability of a > > > > si_flags field in sigaction (just > > > > s/SF_CLEAR_UNKNOWN_FIELDS/SF_HAS_SI_FLAGS/ and > > > > s/si_future_field/si_flags/ in the usage code above). In terms of > > > > SF_CLEAR_UNKNOWN_FIELDS versus SF_HAS_SI_FLAGS I'd be fine either way. > > > > > > > > Another thought that occurred to me is that we may consider > > > > generalizing this a step further and introducing a single flag bit in > > > > the signal number that means "reject unknown flags in sa_flags". This > > > > would mean that we wouldn't need to add any more flag bits to the > > > > signal number in the future, thus limiting this signal number hack to > > > > a single bit; all future mandatory behavior changes could just be put > > > > behind a flag in sa_flags and userspace code would easily be able to > > > > detect missing support for a flag and fall back if necessary. In our > > > > case, this would imply usage code like this: > > > > > > > > void set_segv_handler(void) { > > > > struct sigaction sa; > > > > sa.sa_sigaction = handle_segv; > > > > sa.sa_flags = SA_SIGINFO | SA_CLEAR_UNKNOWN_FIELDS; > > > > // Succeeds in kernels with SA_CLEAR_UNKNOWN_FIELDS support. > > > > // Fails in kernels with SF_CHECK_SA_FLAGS support but no > > > > SA_CLEAR_UNKNOWN_FIELDS support (because of the unknown flags check). > > > > // Fails in kernels without SF_CHECK_SA_FLAGS support (because of > > > > the bounds check on the signal number). > > > > if (sigaction(SIGSEGV | SF_CHECK_SA_FLAGS, &sa, 0) < 0) { > > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > > sa.sa_flags = SA_SIGINFO; > > > > // Succeeds in old kernels, no need to use SF_CHECK_SA_FLAGS since > > > > we're using sa_flags from the beginning of time. > > > > if (sigaction(SIGSEGV, &sa, 0) < 0) > > > > perror("sigaction"); > > > > } > > > > } > > > > > > As with the other options this could work, but looks like it could > > > break the ABI due to violating the original semantics for the signal > > > number argument. Perhaps I'm being too paranoid. > > > > There's no ABI being broken here, as long as we consider syscalls to > > be the stable ABI layer. Old kernels are simply rejecting arguments > > that they don't know about yet. By that argument, any introduction of > > a new syscall is an ABI break because it changes the semantics of a > > previously-unallocated syscall number. > > As argued above, I think this is an invalid argument. > > Although any addition will change behaviour (so is a break in some > sense), the key is not to make "surprising" changes. > > Having something random happen when setting a previously reserved flag > bit, or when issuing a syscall when an unknown syscall number, or not > surprising at all. > > Making fundamental changes to the encoding of an existing argument is > highly surprising, on the other hand: as your example shows, it is > reasonable to index an array using a signal number. > > I agree that this doesn't get us closer to a practical solution though. > > > But we do seem to need some mechanism in addition to (or instead of) > sa_flags. > > Here's another thought: > > Since si_flags would be either always present or always absent, it > could make sense to have a global property to report this, rather than > an sa_flags or signal number bit to request it per-signal. > > Requiring software to parse uname() might be reasonable for that, if > cumbersome (did you suggest this previously?). If we're concerned that > the awkwardness of this would encourage people not to bother (or > encourage people to do it wrong) then we might opt for something simpler > like an AT_FLAGS bit. > > Ultimately libc could provide a more portable interface for discovery, > such as via sysconf(). > > Thoughts? While you're thinking about that, here's another idea: It occurs to me that there are spare bits in si_code. si_code is an enum, but unlike the signal number there are no specific bounds for this value, so we may have an easier time recycling bits here. The high bits of si_code are usually sign-extension and so not always 0, but we can XOR flags into them provided we don't forget the real sign. Software that isn't expecting twiddled bits would get confused, so we need a new SA_ flag to enable this. But this flag (SA_CODEX) below is now just a request. If the kernel doesn't understand it (or without SA_SIGINFO) then no flags would be reported in si_code, which is backwards-compatible. A handler would now do void handler(int n, siginfo_t *si, ...) { int flags = SI_FLAGS(si->si_code); int code = SI_CODE(si->si_code); if (!(flags & SIF_CODEX) { /* flags not supported */ /* Careful assignment of flag meanings may make this check unnecessary, but it's probably useful for developers for testing their code. */ } /* Handle signal based on n, code and flags */ } If the kernel doesn't report any flags (perhaps because it's too old) then SI_FLAGS() will yield 0 and SI_CODE() will just return si_code unchanged. This means that even non-SA_CODEX handlers can use these macros, which may ease migration. Cheers ---Dave --8<-- diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h index cb3d6c2..4e77c71 100644 --- a/include/uapi/asm-generic/siginfo.h +++ b/include/uapi/asm-generic/siginfo.h @@ -176,6 +176,18 @@ typedef struct siginfo { #define SI_DETHREAD -7 /* sent by execve() killing subsidiary threads */ #define SI_ASYNCNL -60 /* sent by glibc async name lookup completion */ +#define __SI_FLAGS 0x7ffff000 /* optional code extension flags */ +#define SIF_CODEX 0x40000000 /* code extension flags supported */ + +/* + * Extract value and extension flags from si_code. + * These are only required in handlers registered with SA_CODEX. + */ +#define SI_CODE(sicode) + ((sicode) >= 0 ? (sicode) & ~__SI_FLAGS : (sicode) | __SI_FLAGS) +#define SI_FLAGS(sicode) \ + (((sicode) >= 0 ? (sicode) : ~(sicode)) & __SI_FLAGS) + #define SI_FROMUSER(siptr) ((siptr)->si_code <= 0) #define SI_FROMKERNEL(siptr) ((siptr)->si_code > 0) diff --git a/include/uapi/asm-generic/signal.h b/include/uapi/asm-generic/signal.h index 5c716a9..c20f5f61 100644 --- a/include/uapi/asm-generic/signal.h +++ b/include/uapi/asm-generic/signal.h @@ -61,6 +61,7 @@ * SA_RESETHAND clears the handler when the signal is delivered. * SA_NOCLDWAIT flag on SIGCHLD to inhibit zombies. * SA_NODEFER prevents the current signal from being masked in the handler. + * SA_CODEX allows extension flag reporting in si_code. * * SA_ONESHOT and SA_NOMASK are the historical Linux names for the Single * Unix names RESETHAND and NODEFER respectively. @@ -68,6 +69,7 @@ #define SA_NOCLDSTOP 0x00000001 #define SA_NOCLDWAIT 0x00000002 #define SA_SIGINFO 0x00000004 +#define SA_CODEX 0x04000000 #define SA_ONSTACK 0x08000000 #define SA_RESTART 0x10000000 #define SA_NODEFER 0x40000000 diff --git a/kernel/signal.c b/kernel/signal.c index ee22ec7..8e8550a 100644 --- a/kernel/signal.c +++ b/kernel/signal.c
On Wed, Jul 8, 2020 at 6:58 AM Dave Martin <Dave.Martin@arm.com> wrote: > > On Wed, Jul 08, 2020 at 12:00:22PM +0100, Dave Martin wrote: > > On Tue, Jul 07, 2020 at 12:07:09PM -0700, Peter Collingbourne wrote: > > > On Tue, Jul 7, 2020 at 7:19 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > On Mon, Jul 06, 2020 at 12:20:33PM -0700, Peter Collingbourne wrote: > > > > > On Mon, Jul 6, 2020 at 9:41 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > On Wed, Jun 24, 2020 at 12:51:43PM -0700, Peter Collingbourne wrote: > > > > > > > On Wed, Jun 24, 2020 at 10:12 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > > > On Wed, Jun 24, 2020 at 09:51:49AM -0700, Peter Collingbourne wrote: > > > > > > > > > On Wed, Jun 24, 2020 at 2:28 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 05:40:08PM -0700, Peter Collingbourne wrote: > > > > > > > > > > > On Tue, Jun 23, 2020 at 10:52 AM Eric W. Biederman > > > > > > > > > > > <ebiederm@xmission.com> wrote: > > > > > > > > > > > > > > > > > > > > > > > > Dave Martin <Dave.Martin@arm.com> writes: > > > > > > > > > > > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > > > > > > > > > > > > >> Peter Collingbourne <pcc@google.com> writes: > > > > > > > > > > > > >> > > > > > > > > > > > > >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > > > > > > > > > > > > >> > index 47f651df781c..a8380a2b6361 100644 > > > > > > > > > > > > >> > --- a/arch/arm64/kernel/traps.c > > > > > > > > > > > > >> > +++ b/arch/arm64/kernel/traps.c > > > > > > > > > > > > >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > > > > > > > > > > > > >> > } > > > > > > > > > > > > >> > > > > > > > > > > > > > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, > > > > > > > > > > > > >> > + unsigned long far, unsigned char far_tb_mask, > > > > > > > > > > > > >> > const char *str) > > > > > > > > > > > > >> > { > > > > > > > > > > > > >> > arm64_show_signal(signo, str); > > > > > > > > > > > > >> > - if (signo == SIGKILL) > > > > > > > > > > > > >> > + if (signo == SIGKILL) { > > > > > > > > > > > > >> > force_sig(SIGKILL); > > > > > > > > > > > > >> > - else > > > > > > > > > > > > >> > - force_sig_fault(signo, code, addr); > > > > > > > > > > > > >> > + } else { > > > > > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > > > > > >> > + info.si_signo = signo; > > > > > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > > > > > >> > + info.si_code = code; > > > > > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > > > > > >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > > > > > > > > > > > > >> > + info.si_addr_top_byte_mask = far_tb_mask; > > > > > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > > > > > >> > + } > > > > > > > > > > > > >> > } > > > > > > > > > > > > >> > > > > > > > > > > > > > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > > > > > > > > > > > > >> > - const char *str) > > > > > > > > > > > > >> > + unsigned long far, const char *str) > > > > > > > > > > > > >> > { > > > > > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > > > > > >> > + > > > > > > > > > > > > >> > arm64_show_signal(SIGBUS, str); > > > > > > > > > > > > >> > - force_sig_mceerr(code, addr, lsb); > > > > > > > > > > > > >> > + > > > > > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > > > > > >> > + info.si_signo = SIGBUS; > > > > > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > > > > > >> > + info.si_code = code; > > > > > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > > > > > >> > + info.si_addr_lsb = lsb; > > > > > > > > > > > > >> > + info.si_addr_top_byte = far >> 56; > > > > > > > > > > > > >> > + info.si_addr_top_byte_mask = 0xff; > > > > > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > > > > > >> > } > > > > > > > > > > > > >> > > > > > > > > > > > > >> I have a real problem with this construction. force_sig_info is not an > > > > > > > > > > > > >> interface that should be used for anything except to define a wrapper > > > > > > > > > > > > >> that takes it's parameters. > > > > > > > > > > > > > > > > > > > > > > > > > > Can you elaborate? How would you do this king of thing. > > > > > > > > > > > > > > > > > > > > > > > > There are no other uses of force_sig_info in architecture code. > > > > > > > > > > > > > > > > > > > > > > > > I just removed them _all_ because they were almost all broken. > > > > > > > > > > > > In fact your mcerr case is broken because it uses two different > > > > > > > > > > > > union members simultantiously. > > > > > > > > > > > > > > > > > > > > > > Is that really broken? I thought that the Linux kernel deliberately > > > > > > > > > > > didn't care about strict aliasing rules (the top-level Makefile passes > > > > > > > > > > > -fno-strict-aliasing) so I thought that it was valid in "Linux kernel > > > > > > > > > > > C" even though from a standards point of view it is invalid. (That > > > > > > > > > > > being said, this is probably moot with my proposed changes below > > > > > > > > > > > though.) > > > > > > > > > > > > > > > > > > > > I have a feeling that -fno-strict-aliasing only allows you to _read_ a > > > > > > > > > > different union member from the one previously written. > > > > > > > > > > > > > > > > > > > > Writing a different member from the last one written can still splatter > > > > > > > > > > on the other members IIUC. > > > > > > > > > > > > > > > > > > > > It would be better to keep things separate rather than risk > > > > > > > > > > incorrectness just to save a few bytes. > > > > > > > > > > > > > > > > > > > > IMHO -fno-strict-aliasing is no excuse for gratuitous type-punning. > > > > > > > > > > > > > > > > > > > > > > So I am looking for something like force_sig_mcerr or force_sig_fault > > > > > > > > > > > > that includes your new information that then calls force_sig_info. > > > > > > > > > > > > > > > > > > > > > > > > I know of no other way to safely use the siginfo struct. > > > > > > > > > > > > > > > > > > > > > > So you want something like: > > > > > > > > > > > > > > > > > > > > > > int force_sig_fault_with_ignored_bits(int signo, int code, void __user > > > > > > > > > > > *addr, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > > int force_sig_mceerr_with_ignored_bits(int code, void __user *addr, > > > > > > > > > > > short lsb, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > > > > > > > > > > > > > in kernel/signal.c and the code in arch/arm64 would call that? > > > > > > > > > > > > > > > > > > > > > > > > AIUI we absolutely need a forced signal here, we need to supply > > > > > > > > > > > > > metadata, and we don't have to open-code all that at every relevant > > > > > > > > > > > > > signal generation site... > > > > > > > > > > > > > > > > > > > > > > > > > >> It is not clear to me that if you have adapted siginfo_layout. > > > > > > > > > > > > > > > > > > > > > > > > > > Garbled sentence? > > > > > > > > > > > > > > > > > > > > > > > > Looks like. One of the pieces of code that needs to change > > > > > > > > > > > > when siginfo gets updated is siginfo_layout so that the structure > > > > > > > > > > > > can be properly decoded and made sense of. > > > > > > > > > > > > > > > > > > > > > > > > I am not seeing anything like that. > > > > > > > > > > > > > > > > > > > > > > Okay, this has to do with copying between the compat and non-compat > > > > > > > > > > > versions of the struct? Sure, I can update that, although the code > > > > > > > > > > > would be basically non-functional on arm64 because TBI isn't supported > > > > > > > > > > > on 32-bit ARM. > > > > > > > > > > > > > > > > > > > > > > > >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > > >> > index cb3d6c267181..6dd82373eb2d 100644 > > > > > > > > > > > > >> > --- a/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > > >> > +++ b/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > > >> > @@ -91,6 +91,14 @@ union __sifields { > > > > > > > > > > > > >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > > > > > > > > > >> > __u32 _pkey; > > > > > > > > > > > > >> > } _addr_pkey; > > > > > > > > > > > > >> > +#ifdef __aarch64__ > > > > > > > > > > > > >> > + /* used with all si_codes */ > > > > > > > > > > > > >> > + struct { > > > > > > > > > > > > >> > + short _dummy_top_byte; > > > > > > > > > > > > > > > > > > > > > > > > > > ^ What's this for? I don't have Eric's insight here. > > > > > > > > > > > > > > > > > > > > > > We would need a short's worth of padding in order to prevent the > > > > > > > > > > > fields from occupying the same address as si_addr_lsb. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >> > + unsigned char _top_byte; > > > > > > > > > > > > >> > + unsigned char _top_byte_mask; > > > > > > > > > > > > >> > + } _addr_top_byte; > > > > > > > > > > > > >> > +#endif > > > > > > > > > > > > >> > }; > > > > > > > > > > > > >> > } _sigfault; > > > > > > > > > > > > >> > > > > > > > > > > > > > >> > > > > > > > > > > > > >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > > > > > > > > > > > > >> > > > > > > > > > > > > >> Please remove the "#ifdef __aarch64__". If at all possible we want to > > > > > > > > > > > > >> design this so any other architecture who has this challenge can use the > > > > > > > > > > > > >> code. The kind of code does not get enough attention/maintenance if it > > > > > > > > > > > > >> is built for a single architecture. > > > > > > > > > > > > > > > > > > > > > > Seems reasonable. I was recently made aware that RISC-V was > > > > > > > > > > > considering a similar feature: > > > > > > > > > > > https://lists.riscv.org/g/tech-tee/topic/risc_v_tbi_proposal/72855478 > > > > > > > > > > > I would have opted to expand this to other architectures on an > > > > > > > > > > > as-needed basis, but I'd also be fine with having it on all > > > > > > > > > > > architectures from the start. > > > > > > > > > > > > > > > > > > > > > > If we make this arch-independent, we have an additional concern, which > > > > > > > > > > > is "what if some future architecture wants more than one byte here?" > > > > > > > > > > > For example, an architecture may have a "top-two-bytes-ignore" > > > > > > > > > > > feature, which would imply two-byte (misnamed) "si_addr_top_byte" and > > > > > > > > > > > "si_addr_top_byte_mask" fields. And the RISC-V proposal potentially > > > > > > > > > > > implies many more ignored bits (see slide 13 of the presentation). The > > > > > > > > > > > maximum size that these fields can possibly be is the size of a > > > > > > > > > > > pointer, and with that there wouldn't be enough room in the padding at > > > > > > > > > > > this point to accommodate the new fields. > > > > > > > > > > > > > > > > > > > > > > That basically implies your earlier suggestion of adding a union > > > > > > > > > > > member here to accommodate future expansion of the union, and adding > > > > > > > > > > > the new fields after the union. I'm happy to make that change, with > > > > > > > > > > > the fields renamed "si_addr_ignored" and "si_addr_ignored_mask". > > > > > > > > > > > > > > > > > > > > I think what we need here is basically a flags word. > > > > > > > > > > > > > > > > > > > > So long as we keep a flag spare to indicate the existence of a further > > > > > > > > > > flags word, we can extend as needed. > > > > > > > > > > > > > > > > > > > > How the existence of the first flags words is detected is another > > > > > > > > > > problem. If it only applies for newly-defined si_code values, then > > > > > > > > > > I guess si_code may be sufficient. > > > > > > > > > > > > > > > > > > Existing kernels will zero-initialize unused regions of the siginfo > > > > > > > > > data structure. The zero-initialization of the padding at the end of > > > > > > > > > the struct is done by the clear_user call here: > > > > > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/kernel/signal.c#L3193 > > > > > > > > > > > > > > > > > > and the zero-initialization of the padding between fields and unused > > > > > > > > > union members is done by the clear_siginfo function which the kernel > > > > > > > > > calls when initializing the data structure: > > > > > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/include/linux/signal.h#L20 > > > > > > > > > > > > > > > > > > Therefore, a flag word value of 0 may be used to detect a lack of > > > > > > > > > support for flagged fields. > > > > > > > > > > > > > > > > It's not enough that we do this today. We would have had to do it back > > > > > > > > to the dawn of time (though in the arm64 case I guess we just need to go > > > > > > > > back to when the arch/arm64 was merged). > > > > > > > > > > > > > > > > v2.6.12:kernel/signal.c:copy_siginfo_to_user() suggests that this wasn't > > > > > > > > always the case, so unused parts of siginfo could be full of old junk > > > > > > > > from the user stack, if the kernel is sufficiently old. > > > > > > > > > > > > > > > > If we're trying to do something generic that makes sense on all arches, > > > > > > > > this matters. I may have misunderstood something about the code though. > > > > > > > > > > > > > > Hmm, I think you're right. The current behavior was introduced by > > > > > > > commit c999b933faa5e281e3af2e110eccaf91698b0a81 which was first > > > > > > > released in 4.18. So if an application wants to be compatible with > > > > > > > pre-4.18 kernels then there would need to be some other way to > > > > > > > indicate that the fields are valid. Probably the simplest way would be > > > > > > > to have the application issue a uname(2) syscall and check the kernel > > > > > > > version before reading these fields. I have a couple of other ideas > > > > > > > that don't rely on version detection, if we'd prefer to avoid that. > > > > > > > (They are somewhat ugly, but our hand is forced by backwards > > > > > > > compatibility.) > > > > > > > > > > > > > > One idea is to re-purpose the si_errno field as a flags field for > > > > > > > certain signal numbers. I checked a few kernel releases going back to > > > > > > > 2.6.18 and it looks like the field is set to 0 except in the following > > > > > > > circumstances: > > > > > > > - sending a hardware breakpoint (SIGTRAP/TRAP_HWBKPT) > > > > > > > - seccomp failures (SIGSYS/SYS_SECCOMP) > > > > > > > - user-defined signal via kill_pid_usb_asyncio > > > > > > > - SIGSWI in 3.18 and before (code since removed) > > > > > > > > > > > > > > It is also set to EFAULT for certain SIGSEGV/SEGV_MAPERR signals on > > > > > > > powerpc since commit c96c4436aba4c12f1f48369f2f90bc43e12fe36c, which > > > > > > > is currently unreleased. So if we wanted to go this route for SIGSEGV > > > > > > > we would need to stop the kernel from setting si_errno to EFAULT for > > > > > > > this signal before the 5.8 release. > > > > > > > > > > > > > > Another idea was to have userspace set a flag in sa_flags when > > > > > > > registering a signal handler meaning "this signal handler requires > > > > > > > unknown siginfo fields to be zeroed", and have existing kernels reject > > > > > > > the syscall due to an unknown flag being set, but unfortunately this > > > > > > > won't work because existing kernels do not reject sigaction syscalls > > > > > > > with unknown flags set in sa_flags. A perhaps more radical idea in > > > > > > > this vein would be to claim some of the upper bits of the signal > > > > > > > number as flags that will cause the syscall to be rejected if set and > > > > > > > unknown to the kernel. Existing kernels (going back to at least > > > > > > > 2.6.18) contain this code in do_sigaction: > > > > > > > > > > > > > > if (!valid_signal(sig) || sig < 1 || (act && sig_kernel_only(sig))) > > > > > > > return -EINVAL; > > > > > > > > > > > > > > and vald_signal is defined as: > > > > > > > > > > > > > > static inline int valid_signal(unsigned long sig) > > > > > > > { > > > > > > > return sig <= _NSIG ? 1 : 0; > > > > > > > } > > > > > > > > > > > > > > All architectures define _NSIG as a value <= 128, so they will reject > > > > > > > a signal number with any of bits 8-31 set. This means that we can use > > > > > > > any of those bits for mandatory flags. Most likely we could use bit 30 > > > > > > > (expanding down as necessary), as it keeps the signal number positive > > > > > > > and permits future expansion of the signal number range. > > > > > > > > > > > > Does the signal core code actually gurantee to zero the unused fields? > > > > > > Unless the fields are poked in by hand this is fraught with subtlelies, > > > > > > especially when unions are involved. (I'm sure the code tries to do it, > > > > > > but I've not eyeballed it in detail...) > > > > > > > > > > It memsets the siginfo structure before setting the fields and sending > > > > > the signal (grep for clear_siginfo which is just a memset; you should > > > > > find a call before all callers of force_sig_info). Memset is the right > > > > > approach here since unlike setting fields by hand it clears padding > > > > > which could lead to information leaks from the kernel. IIUC this is > > > > > the reason why Eric wants all of the signals to be raised via wrappers > > > > > in kernel/signal.c instead of via force_sig_info directly (to make > > > > > this aspect easier to audit). > > > > > > > > My impression was that the reason for this model is partly to ensure > > > > that siginfo fields are populated more consistently. When it was all > > > > down to the individual callers, inconsistencies creeped in. > > > > > > > > With regard to memset(), this is not a complete defence against data > > > > leakage. Assigning to a struct member can set any or all padding in > > > > the struct to random garbage (consider write-combining of neighboring > > > > member writes into a single larger accesses in asm for example). The > > > > > > I don't believe that LLVM will store to padding like this. I don't > > > know about GCC, though, but I wouldn't be surprised if this is > > > something that the kernel would want to turn off in "kernel C" (like > > > it turns off strict aliasing) specifically because of the information > > > leak issue. > > > > Again, the issue is not future kernel builds -- we can always find a way > > to fix the behaviour for those -- but past kernel builds. I thought that the whole point of the "bit in the signal number" (or SI_CODEX or whatever) was that we didn't need to worry about the behavior of past kernel builds? > > > > > > only way to avoid this is to ensure that the struct is 100% > > > > padding-free, and that each member of a union is the same size. A > > > > quick clance at <uapi/asm-generic/siginfo.h> confirms that this is not > > > > the case. > > > > > > > > This might need to be looked at separately. > > > > > > > > But it does mean, strictly speaking, that we can't reliably add new > > > > fields anywhere that there was previously padding: assigning to > > > > neighboring members can still fill those with garbage after the > > > > memset(). > > > > > > ...but this is largely moot because I'm not proposing to add new > > > fields in the padding any more (because the fields needed to become > > > larger in order to accommodate future hypothetical architectures which > > > might want to use the fields, and thus they wouldn't fit in the > > > padding). The siginfo.h diff would be something like: > > > > > > diff --git a/include/uapi/asm-generic/siginfo.h > > > b/include/uapi/asm-generic/siginfo.h > > > index cb3d6c267181..4a2fe257415d 100644 > > > --- a/include/uapi/asm-generic/siginfo.h > > > +++ b/include/uapi/asm-generic/siginfo.h > > > @@ -91,7 +91,10 @@ union __sifields { > > > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > __u32 _pkey; > > > } _addr_pkey; > > > + void *_pad[6]; > > > }; > > > + uintptr_t _ignored_bits; > > > + uintptr_t _ignored_bits_mask; > > > > This _is_ in padding: the tail-padding of the (previously smaller) > > _sigfault. Again, the compiler was allowed to populate this area with > > junk before these fields were added. > > > > I agree that it seems fairly unlikely that the compiler would have been > > overwriting this in normal circumstances, but that's not a guarantee. > > My worry is that if this goes wrong, it will go wrong silently and > > unpredictably. > > > > > } _sigfault; > > > > > > /* SIGPOLL */ > > > > > > or with a "uintptr_t _flags" added in before _ignored_bits if we go with that. > > > > > > > > > Using unused bits in the signal number to turn on new functionality > > > > > > feels risky. As currently specified, this is just a number. Since > > > > > > today a successful sigaction(n ...) guarantees that n is a valid signal > > > > > > number, reasonable code like the following would trigger a buffer > > > > > > overrun if we start trying to encode anything else in there: > > > > > > > > > > > > struct sigaction actions[NSIG]; > > > > > > > > > > > > int do_something( ... ) > > > > > > { > > > > > > ... > > > > > > > > > > > > if (!sigaction(n, sa, ...)) { > > > > > > actions[n] = *sa; > > > > > > return 0; > > > > > > } > > > > > > > > > > > > ... > > > > > > } > > > > > > > > > > I imagine the bit in the signal number being set by the direct caller > > > > > to sigaction, and we could specifically recommend that calling > > > > > pattern. In that case, your "n" wouldn't have the bit set in it. It > > > > > > > > I can imagine this too, but that doesn't mean that software does it. > > > > > > > > If the above kind of thing exists in a framework or library somewhere, > > > > we could get problems. Similarly, a pre-existing LD_PRELOAD framework > > > > that provides a wrapper for sigaction may now go wrong even if your > > > > pattern is followed -- i.e., the caller thinks it's calling sigaction > > > > directly but in fact it isn't. > > > > > > I'm aware of one library like that. It's called libsigchain, and it > > > has an early bounds check: > > > https://cs.android.com/android/platform/superproject/+/master:art/sigchainlib/sigchain.cc;l=371 > > > > > > Until the library is changed to recognize the flag, calling code would > > > see the return value of -1 as if the kernel failed the syscall, and > > > would fall back to the code for old kernels. > > > > But only after some bad dereferences. If these were writes, this means > > that memory _may_ be silently corrupted (I don't say it't likely in a > > given case, and we cannot pick a flag bit that makes this impossible). You're talking about libsigchain, right? I don't see any bad references, the function returns after noticing the bounds check failure. > > So, _even though the user program is correct_, our change may trigger Let's say that you were talking about some other library and not libsigchain. Such an interceptor wouldn't be correct though, it failed to account for our change to the syscall semantics. If the accesses were before the syscall (or the bounds check), then the interceptor would not have been correct in the first place because POSIX requires returning -1 with errno=EINVAL (and not crashing) if the signal number is invalid. > > the corruption of arbitrary user memory. This what I mean by an ABI > > break. The fact that the corruption is not done by the syscall itself > > is no excuse. At some point, though, accommodating interceptors becomes pretty much tantamount to saying "we can never change anything". Even just adding a field to __sifields (which is pretty much required for what we need to do) could break things in the presence of some interceptors because the interceptor could be copying the fields manually to a new data structure before calling the user's signal handler (e.g. because it wants to defer the signal until later) and miss our new field. I think most of the other ideas we're discussing fail to meet this bar as well and I'll go into more details later on. > > We also fail to notice failures in sigaddset() etc., though in this code > > it looks like that should not matter. Maybe you're looking at the handler ("SignalChain::Handler")? The bit wouldn't be set in the signo argument to the handler. I'm talking about line 371 of the code I linked, in the sigaction interceptor "__sigaction" (it looks like sometimes the link doesn't take you to the correct line for some reason). > > > > > In general I think that any library like this with independent > > > tracking of the kernel's purported signal handler state would need to > > > be very sensitive to which syscalls are capable of setting signal > > > handlers, what their semantics are, and so on. This applies to any > > > change that we might make to the signal handler interface. So for > > > example, if we introduced a new syscall as you propose below, and the > > > library hasn't been updated to recognize the new syscall, it will > > > silently miss changes in signal handler state caused by the new > > > syscall. > > > > > > At the end of this argument lies "we can never change anything about > > > how signal handlers work because it could break some interposing > > > library somewhere" -- replace "signal handlers" with any kernel > > > feature whose behavior may be modified by an interposing library if > > > you like -- and I don't think we want to go that far. As far as I > > > know, this isn't really the kernel's business anyway -- the kernel's > > > stable ABI contract starts and ends with the syscall interface and not > > > some library on top. > > > > > > That being said, we should perhaps try to define our interface so that > > > something reasonable will probably happen if there is such a library > > > and it hasn't been updated. With the new syscall, the library will > > > sometimes silently fail to work in some non-local fashion. With the > > > flag bit in the signal number, the library will either cause the > > > caller to fall back to the old kernel code path (if there is a bounds > > > check) or likely crash loudly (if there is no bounds check). To me, > > > the "flag bit in the signal number" behavior seems more reasonable, > > > since either something correct or something easy to debug will > > > probably happen at runtime. > > > > > > > > could only appear in newly-written code that doesn't follow our > > > > > recommendations, and there are already plenty of much more likely ways > > > > > to cause buffer overflows in C code that doesn't follow > > > > > recommendations anyway. (And even if such a buffer overflow occurred, > > > > > it would very likely be caught early in development by the MMU due to > > > > > the magnitude of the number 1<<30.) > > > > > > > > Choosing the bit value is hard. If shitfing it overflows, this can > > > > trigger random undefined behaviour in the compiler in addition to (or > > > > perhaps instead of) an out-of-bounds access or segfault. > > > > > > It wouldn't overflow on a 64-bit architecture assuming normal array > > > indexing (the index would be promoted to pointer width before being > > > scaled to the array element size), and to begin with the users of this > > > would be 64-bit. > > > > Unless we don't offer this feature for 32-bit at all (possible, if ugly) > > we can't stop people using it. My point is that the problem in the interceptor library would probably be noticed on 64-bit (since that's what most people use these days), which would probably result in it being fixed by the time it reaches 32-bit users. > > > > > > If shifting it doesn't overflow, we might still fall into a valid > > > > mapping, though I'd agree a segfault is more likely. > > > > > > > > > > > > > > > I think it would be cleaner for to add a single flag field that can be > > > > > > used for detecting other extensions, and request it via a new sa_flags > > > > > > bit. This removes the need for sematically useless zeroing of unused > > > > > > fields (though for hygiene and backwards compatibility reasons we would > > > > > > probably want to carry on zeroing them anyway). > > > > > > > > > > > > I can see no simpler way to add supplementary siginfo fields for > > > > > > existing si_codes. For si_codes that didn't exist before the zeroing > > > > > > came in we could still detect optional si_code-specific fields via > > > > > > zeroing, but it seems messary to have two ways of detecting extensions. > > > > > > > > > > That would certainly be cleaner if it worked, but that would only be > > > > > the case if old kernels rejected unknown bits in sa_flags, and > > > > > unfortunately they don't. With the bit in the signal number, the "old > > > > > > > > Hmm, that is a problem I wasn't aware of. > > > > > > > > > kernels reject" behavior admits relatively straightforward usage code: > > > > > > > > > > void set_segv_handler(void) { > > > > > struct sigaction sa; > > > > > sa.sa_sigaction = handle_segv; > > > > > sa.sa_flags = SA_SIGINFO; > > > > > if (sigaction(SIGSEGV | SF_CLEAR_UNKNOWN_FIELDS, &sa, 0) < 0) { // > > > > > succeeds in new kernels, fails in old kernels > > > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > > > if (sigaction(SIGSEGV, &sa, 0) < 0) // succeeds in old kernels > > > > > perror("sigaction"); > > > > > } > > > > > } > > > > > > > > > > void clear_fields_and_handle_segv(int signum, siginfo_t *sa, void *ctx) { > > > > > sa->si_future_field = 0; > > > > > handle_segv(signum, sa, ctx); > > > > > } > > > > > > > > > > void handle_segv(int signum, siginfo_t *sa, void *ctx) { > > > > > // At this point, si_future_field will have the value 0 in old > > > > > kernels and the kernel-supplied value in new kernels. > > > > > } > > > > > > > > > > Imagine if we moved the flag SF_CLEAR_UNKNOWN_FIELDS from the signal > > > > > number to sa_flags. In that case, the first sigaction would succeed in > > > > > old kernels so handle_segv wouldn't know whether it can safely read > > > > > from si_future_field. With the sa_flags approach, you would need > > > > > kernel version number checking via uname before setting the flag in > > > > > sa_flags, and at that point why even have the flag in sa_flags at all > > > > > since you could just have the signal handler conditionally read from > > > > > si_future_field based on the uname? > > > > > > > > Software setting SA_SIFLAGS (or whatever) is new by definition, since > > > > it would be using a new #define. So it might be reasonable to put the > > > > burden on that software to verify that the flag was really accepted by > > > > the kernel, by reading it back. > > > > > > That doesn't seem like a good idea even if it worked, because it could > > > lead to race conditions. If the si_flags-reading signal handler were > > > invoked in response to a signal between when you set it and when you > > > ended up replacing it with the fallback signal handler for old > > > kernels, the handler may end up reading garbage data from si_flags. > > > > Not really. My example may have this problem, but the signal handler > > can be written to support both scenarios, based on testing a flag that > > the main program sets after verifying that the flag could be set. Or > > the signal could be blocked around establishment (often a good idea for > > other reasons). > > > > But I agree it's a bit gross, and anyway doesn't work due to the fact > > that the kernel doesn't filter out unrecognised flags anyway. > > > > > > Unfortunately, even relatively recent kernels blindly store sa_flags > > > > in the kernel without validating it, and so it looks like duff flags > > > > can be read back out via a sigaction() call. Dang. > > > > > > > > > > > > Perhaps a new frontend syscall could be added. A new libc that knows > > > > about this "sigaction2" could use it and mask off problem bits from > > > > sa_flags in its sigaction() wrapper before calling sigaction2. An old > > > > libc would call the old sigaction syscall, where we would ignore these > > > > new sa_flags bits as before. > > > > > > I'm not currently in favor of the new syscall but if we do this I > > > would keep sigaction and sigaction2 separate. That is, libc sigaction > > > should always use the sigaction syscall, and libc sigaction2 should > > > always use the sigaction2 syscall. We should avoid libc's sigaction > > > having different behavior based on the libc version and kernel > > > version, as that would make it harder to reason about its behavior. > > > Calling code would need to check for presence of sigaction2 in both > > > libc and the kernel, e.g. > > > > > > __attribute__((weak)) decltype(sigaction2) sigaction2; > > > > > > void set_segv_handler(void) { > > > struct sigaction sa; > > > sa.sa_sigaction = handle_segv; > > > sa.sa_flags = SA_SIGINFO | SA_SIFLAGS; > > > if (!sigaction2 || sigaction2(SIGSEGV, &sa, 0) < 0) { > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > sa.sa_flags = SA_SIGINFO; > > > if (sigaction(SIGSEGV, &sa, 0) < 0) > > > perror("sigaction"); > > > } > > > } > > > > I guess. But I share your distaste for adding a new syscall. > > > > > > > > > This may not be a popular approach though, and software wouldn't be able > > > > to use our new features until libc is updated to match. > > > > > > > > If we go down this route, it may provide additional opportunities to fix > > > > annoying defects in the old interface. > > > > > > > > > > > > > Note that the same applies to a flag indicating the availability of a > > > > > si_flags field in sigaction (just > > > > > s/SF_CLEAR_UNKNOWN_FIELDS/SF_HAS_SI_FLAGS/ and > > > > > s/si_future_field/si_flags/ in the usage code above). In terms of > > > > > SF_CLEAR_UNKNOWN_FIELDS versus SF_HAS_SI_FLAGS I'd be fine either way. > > > > > > > > > > Another thought that occurred to me is that we may consider > > > > > generalizing this a step further and introducing a single flag bit in > > > > > the signal number that means "reject unknown flags in sa_flags". This > > > > > would mean that we wouldn't need to add any more flag bits to the > > > > > signal number in the future, thus limiting this signal number hack to > > > > > a single bit; all future mandatory behavior changes could just be put > > > > > behind a flag in sa_flags and userspace code would easily be able to > > > > > detect missing support for a flag and fall back if necessary. In our > > > > > case, this would imply usage code like this: > > > > > > > > > > void set_segv_handler(void) { > > > > > struct sigaction sa; > > > > > sa.sa_sigaction = handle_segv; > > > > > sa.sa_flags = SA_SIGINFO | SA_CLEAR_UNKNOWN_FIELDS; > > > > > // Succeeds in kernels with SA_CLEAR_UNKNOWN_FIELDS support. > > > > > // Fails in kernels with SF_CHECK_SA_FLAGS support but no > > > > > SA_CLEAR_UNKNOWN_FIELDS support (because of the unknown flags check). > > > > > // Fails in kernels without SF_CHECK_SA_FLAGS support (because of > > > > > the bounds check on the signal number). > > > > > if (sigaction(SIGSEGV | SF_CHECK_SA_FLAGS, &sa, 0) < 0) { > > > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > > > sa.sa_flags = SA_SIGINFO; > > > > > // Succeeds in old kernels, no need to use SF_CHECK_SA_FLAGS since > > > > > we're using sa_flags from the beginning of time. > > > > > if (sigaction(SIGSEGV, &sa, 0) < 0) > > > > > perror("sigaction"); > > > > > } > > > > > } > > > > > > > > As with the other options this could work, but looks like it could > > > > break the ABI due to violating the original semantics for the signal > > > > number argument. Perhaps I'm being too paranoid. > > > > > > There's no ABI being broken here, as long as we consider syscalls to > > > be the stable ABI layer. Old kernels are simply rejecting arguments > > > that they don't know about yet. By that argument, any introduction of > > > a new syscall is an ABI break because it changes the semantics of a > > > previously-unallocated syscall number. > > > > As argued above, I think this is an invalid argument. > > > > Although any addition will change behaviour (so is a break in some > > sense), the key is not to make "surprising" changes. If we care about interceptors then I don't think "surprising" comes into it. It's more a question of "does the anticipated behavior of the interceptor match our desired behavior", where "desired" means "most likely to avoid silent breakage". We would need to get into the head of a potential interceptor author and think about how they would have handled the signal number argument, as well as other arguments like sa_flags if we want to go that route, and see whether that behavior would lead to the desired result. In this case, I think we exactly want the interceptor author to have thought "oh, it's just a number, I'll (possibly do a bounds check and then) use the number as an index into an array". This will lead to one of two outcomes: crashing (yes, yes, it won't always crash, but if the alternative is that it never crashes and we get silently incorrect behavior all of the time, I'll take sometimes crashing) or fail the bounds check and pretend to be an old kernel (the latter is anticipated by POSIX which requires returning -1/EINVAL for an invalid signal number). Each of these behaviors are desirable, as they are observable failures, which are more likely to result in fixes than silent ones. > > Having something random happen when setting a previously reserved flag > > bit, or when issuing a syscall when an unknown syscall number, or not > > surprising at all. Introducing a new syscall is right out in this model. The interceptor author wouldn't have anticipated our introducing a new syscall, so the new syscall wouldn't be intercepted and calls to the new syscall would silently bypass the interceptor. For example, adding sigaction2 could result in signal handlers being set without the interceptor's knowledge. Regarding a sa_flags bit, let's get inside the head of the interceptor author again. How would they handle a flag bit that they don't recognize when replacing the signal handler? It wouldn't be correct to just pass it through to the kernel, or drop the flag on the floor, as it might be semantically meaningful (and thus could change the calling convention as SA_SIGINFO does, or change the meaning of fields in siginfo, as SA_CODEX would do). A correctly written sigaction interceptor should probably abort the program upon encountering an unknown flag (thus giving a human a chance to update the interceptor), but chances are that they don't. Ignoring all but a few flags (and passing a fixed set of flags to the kernel) seems to be what libsigchain does, and in the case of SA_CODEX it would seem to result in desirable behavior (but I suspect that it isn't handling the other flags correctly), but I could also see an interceptor author just passing it unchanged to the kernel without checking it (perhaps because they didn't think about these issues, and because that didn't matter until now, with the exception of from-the-beginning-of-time flags like SA_SIGINFO). And with SA_CODEX that could lead to silent misreading of si_code in the interceptor's signal handler, if it hasn't been updated to use the new macros. > > Making fundamental changes to the encoding of an existing argument is > > highly surprising, on the other hand: as your example shows, it is > > reasonable to index an array using a signal number. > > > > I agree that this doesn't get us closer to a practical solution though. > > > > > > But we do seem to need some mechanism in addition to (or instead of) > > sa_flags. > > > > Here's another thought: > > > > Since si_flags would be either always present or always absent, it > > could make sense to have a global property to report this, rather than > > an sa_flags or signal number bit to request it per-signal. > > > > Requiring software to parse uname() might be reasonable for that, if > > cumbersome (did you suggest this previously?). If we're concerned that > > the awkwardness of this would encourage people not to bother (or > > encourage people to do it wrong) then we might opt for something simpler > > like an AT_FLAGS bit. > > > > Ultimately libc could provide a more portable interface for discovery, > > such as via sysconf(). > > > > Thoughts? Yes, this was all to avoid the userspace code needing to contain a version check (or equivalent). Maybe the version check would be better than the alternatives though (although it's still vulnerable to non-updated interceptors not copying our new fields). AT_FLAGS sounds good to me. > While you're thinking about that, here's another idea: > > It occurs to me that there are spare bits in si_code. si_code is an > enum, but unlike the signal number there are no specific bounds for > this value, so we may have an easier time recycling bits here. > > The high bits of si_code are usually sign-extension and so not always > 0, but we can XOR flags into them provided we don't forget the real sign. > > Software that isn't expecting twiddled bits would get confused, so we > need a new SA_ flag to enable this. But this flag (SA_CODEX) below is > now just a request. If the kernel doesn't understand it (or without > SA_SIGINFO) then no flags would be reported in si_code, which is > backwards-compatible. > > A handler would now do > > void handler(int n, siginfo_t *si, ...) > { > int flags = SI_FLAGS(si->si_code); > int code = SI_CODE(si->si_code); > > if (!(flags & SIF_CODEX) { > /* flags not supported */ > /* Careful assignment of flag meanings may make this > check unnecessary, but it's probably useful for > developers for testing their code. */ > } > > /* Handle signal based on n, code and flags */ > } > > > If the kernel doesn't report any flags (perhaps because it's too old) > then SI_FLAGS() will yield 0 and SI_CODE() will just return si_code > unchanged. This means that even non-SA_CODEX handlers can use these > macros, which may ease migration. Thanks, this seems more appealing to me than the bit in the signal number idea. It uses the sa_flags field as intended and doesn't abuse the fields of the siginfo data structure too much. I don't think we should put as much weight into interceptor concerns as you do, so you can consider my above argumentation to be from a devil's advocate perspective. > > Cheers > ---Dave > > --8<-- > > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > index cb3d6c2..4e77c71 100644 > --- a/include/uapi/asm-generic/siginfo.h > +++ b/include/uapi/asm-generic/siginfo.h > @@ -176,6 +176,18 @@ typedef struct siginfo { > #define SI_DETHREAD -7 /* sent by execve() killing subsidiary threads */ > #define SI_ASYNCNL -60 /* sent by glibc async name lookup completion */ > > +#define __SI_FLAGS 0x7ffff000 /* optional code extension flags */ > +#define SIF_CODEX 0x40000000 /* code extension flags supported */ I don't think we would need this flag because even old kernels "support" the extension flags under this scheme (they just don't report any features). We would only need to introduce flags for actual features. Peter > + > +/* > + * Extract value and extension flags from si_code. > + * These are only required in handlers registered with SA_CODEX. > + */ > +#define SI_CODE(sicode) > + ((sicode) >= 0 ? (sicode) & ~__SI_FLAGS : (sicode) | __SI_FLAGS) > +#define SI_FLAGS(sicode) \ > + (((sicode) >= 0 ? (sicode) : ~(sicode)) & __SI_FLAGS) > + > #define SI_FROMUSER(siptr) ((siptr)->si_code <= 0) > #define SI_FROMKERNEL(siptr) ((siptr)->si_code > 0) > > diff --git a/include/uapi/asm-generic/signal.h b/include/uapi/asm-generic/signal.h > index 5c716a9..c20f5f61 100644 > --- a/include/uapi/asm-generic/signal.h > +++ b/include/uapi/asm-generic/signal.h > @@ -61,6 +61,7 @@ > * SA_RESETHAND clears the handler when the signal is delivered. > * SA_NOCLDWAIT flag on SIGCHLD to inhibit zombies. > * SA_NODEFER prevents the current signal from being masked in the handler. > + * SA_CODEX allows extension flag reporting in si_code. > * > * SA_ONESHOT and SA_NOMASK are the historical Linux names for the Single > * Unix names RESETHAND and NODEFER respectively. > @@ -68,6 +69,7 @@ > #define SA_NOCLDSTOP 0x00000001 > #define SA_NOCLDWAIT 0x00000002 > #define SA_SIGINFO 0x00000004 > +#define SA_CODEX 0x04000000 > #define SA_ONSTACK 0x08000000 > #define SA_RESTART 0x10000000 > #define SA_NODEFER 0x40000000 > diff --git a/kernel/signal.c b/kernel/signal.c > index ee22ec7..8e8550a 100644 > --- a/kernel/signal.c > +++ b/kernel/signal.c
On Wed, Jul 08, 2020 at 03:21:13PM -0700, Peter Collingbourne wrote: > On Wed, Jul 8, 2020 at 6:58 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > On Wed, Jul 08, 2020 at 12:00:22PM +0100, Dave Martin wrote: > > > On Tue, Jul 07, 2020 at 12:07:09PM -0700, Peter Collingbourne wrote: > > > > On Tue, Jul 7, 2020 at 7:19 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > On Mon, Jul 06, 2020 at 12:20:33PM -0700, Peter Collingbourne wrote: > > > > > > On Mon, Jul 6, 2020 at 9:41 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > On Wed, Jun 24, 2020 at 12:51:43PM -0700, Peter Collingbourne wrote: > > > > > > > > On Wed, Jun 24, 2020 at 10:12 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > > > > > On Wed, Jun 24, 2020 at 09:51:49AM -0700, Peter Collingbourne wrote: > > > > > > > > > > On Wed, Jun 24, 2020 at 2:28 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 05:40:08PM -0700, Peter Collingbourne wrote: > > > > > > > > > > > > On Tue, Jun 23, 2020 at 10:52 AM Eric W. Biederman > > > > > > > > > > > > <ebiederm@xmission.com> wrote: > > > > > > > > > > > > > > > > > > > > > > > > > > Dave Martin <Dave.Martin@arm.com> writes: > > > > > > > > > > > > > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > > > > > > > > > > > > > >> Peter Collingbourne <pcc@google.com> writes: > > > > > > > > > > > > > >> > > > > > > > > > > > > > >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > > > > > > > > > > > > > >> > index 47f651df781c..a8380a2b6361 100644 > > > > > > > > > > > > > >> > --- a/arch/arm64/kernel/traps.c > > > > > > > > > > > > > >> > +++ b/arch/arm64/kernel/traps.c > > > > > > > > > > > > > >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > > > > > > > > > > > > > >> > } > > > > > > > > > > > > > >> > > > > > > > > > > > > > > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, > > > > > > > > > > > > > >> > + unsigned long far, unsigned char far_tb_mask, > > > > > > > > > > > > > >> > const char *str) > > > > > > > > > > > > > >> > { > > > > > > > > > > > > > >> > arm64_show_signal(signo, str); > > > > > > > > > > > > > >> > - if (signo == SIGKILL) > > > > > > > > > > > > > >> > + if (signo == SIGKILL) { > > > > > > > > > > > > > >> > force_sig(SIGKILL); > > > > > > > > > > > > > >> > - else > > > > > > > > > > > > > >> > - force_sig_fault(signo, code, addr); > > > > > > > > > > > > > >> > + } else { > > > > > > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > > > > > > >> > + info.si_signo = signo; > > > > > > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > > > > > > >> > + info.si_code = code; > > > > > > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > > > > > > >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > > > > > > > > > > > > > >> > + info.si_addr_top_byte_mask = far_tb_mask; > > > > > > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > > > > > > >> > + } > > > > > > > > > > > > > >> > } > > > > > > > > > > > > > >> > > > > > > > > > > > > > > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > > > > > > > > > > > > > >> > - const char *str) > > > > > > > > > > > > > >> > + unsigned long far, const char *str) > > > > > > > > > > > > > >> > { > > > > > > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > > > > > > >> > + > > > > > > > > > > > > > >> > arm64_show_signal(SIGBUS, str); > > > > > > > > > > > > > >> > - force_sig_mceerr(code, addr, lsb); > > > > > > > > > > > > > >> > + > > > > > > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > > > > > > >> > + info.si_signo = SIGBUS; > > > > > > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > > > > > > >> > + info.si_code = code; > > > > > > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > > > > > > >> > + info.si_addr_lsb = lsb; > > > > > > > > > > > > > >> > + info.si_addr_top_byte = far >> 56; > > > > > > > > > > > > > >> > + info.si_addr_top_byte_mask = 0xff; > > > > > > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > > > > > > >> > } > > > > > > > > > > > > > >> > > > > > > > > > > > > > >> I have a real problem with this construction. force_sig_info is not an > > > > > > > > > > > > > >> interface that should be used for anything except to define a wrapper > > > > > > > > > > > > > >> that takes it's parameters. > > > > > > > > > > > > > > > > > > > > > > > > > > > > Can you elaborate? How would you do this king of thing. > > > > > > > > > > > > > > > > > > > > > > > > > > There are no other uses of force_sig_info in architecture code. > > > > > > > > > > > > > > > > > > > > > > > > > > I just removed them _all_ because they were almost all broken. > > > > > > > > > > > > > In fact your mcerr case is broken because it uses two different > > > > > > > > > > > > > union members simultantiously. > > > > > > > > > > > > > > > > > > > > > > > > Is that really broken? I thought that the Linux kernel deliberately > > > > > > > > > > > > didn't care about strict aliasing rules (the top-level Makefile passes > > > > > > > > > > > > -fno-strict-aliasing) so I thought that it was valid in "Linux kernel > > > > > > > > > > > > C" even though from a standards point of view it is invalid. (That > > > > > > > > > > > > being said, this is probably moot with my proposed changes below > > > > > > > > > > > > though.) > > > > > > > > > > > > > > > > > > > > > > I have a feeling that -fno-strict-aliasing only allows you to _read_ a > > > > > > > > > > > different union member from the one previously written. > > > > > > > > > > > > > > > > > > > > > > Writing a different member from the last one written can still splatter > > > > > > > > > > > on the other members IIUC. > > > > > > > > > > > > > > > > > > > > > > It would be better to keep things separate rather than risk > > > > > > > > > > > incorrectness just to save a few bytes. > > > > > > > > > > > > > > > > > > > > > > IMHO -fno-strict-aliasing is no excuse for gratuitous type-punning. > > > > > > > > > > > > > > > > > > > > > > > > So I am looking for something like force_sig_mcerr or force_sig_fault > > > > > > > > > > > > > that includes your new information that then calls force_sig_info. > > > > > > > > > > > > > > > > > > > > > > > > > > I know of no other way to safely use the siginfo struct. > > > > > > > > > > > > > > > > > > > > > > > > So you want something like: > > > > > > > > > > > > > > > > > > > > > > > > int force_sig_fault_with_ignored_bits(int signo, int code, void __user > > > > > > > > > > > > *addr, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > > > int force_sig_mceerr_with_ignored_bits(int code, void __user *addr, > > > > > > > > > > > > short lsb, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > > > > > > > > > > > > > > > in kernel/signal.c and the code in arch/arm64 would call that? > > > > > > > > > > > > > > > > > > > > > > > > > > AIUI we absolutely need a forced signal here, we need to supply > > > > > > > > > > > > > > metadata, and we don't have to open-code all that at every relevant > > > > > > > > > > > > > > signal generation site... > > > > > > > > > > > > > > > > > > > > > > > > > > > >> It is not clear to me that if you have adapted siginfo_layout. > > > > > > > > > > > > > > > > > > > > > > > > > > > > Garbled sentence? > > > > > > > > > > > > > > > > > > > > > > > > > > Looks like. One of the pieces of code that needs to change > > > > > > > > > > > > > when siginfo gets updated is siginfo_layout so that the structure > > > > > > > > > > > > > can be properly decoded and made sense of. > > > > > > > > > > > > > > > > > > > > > > > > > > I am not seeing anything like that. > > > > > > > > > > > > > > > > > > > > > > > > Okay, this has to do with copying between the compat and non-compat > > > > > > > > > > > > versions of the struct? Sure, I can update that, although the code > > > > > > > > > > > > would be basically non-functional on arm64 because TBI isn't supported > > > > > > > > > > > > on 32-bit ARM. > > > > > > > > > > > > > > > > > > > > > > > > > >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > > > >> > index cb3d6c267181..6dd82373eb2d 100644 > > > > > > > > > > > > > >> > --- a/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > > > >> > +++ b/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > > > >> > @@ -91,6 +91,14 @@ union __sifields { > > > > > > > > > > > > > >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > > > > > > > > > > >> > __u32 _pkey; > > > > > > > > > > > > > >> > } _addr_pkey; > > > > > > > > > > > > > >> > +#ifdef __aarch64__ > > > > > > > > > > > > > >> > + /* used with all si_codes */ > > > > > > > > > > > > > >> > + struct { > > > > > > > > > > > > > >> > + short _dummy_top_byte; > > > > > > > > > > > > > > > > > > > > > > > > > > > > ^ What's this for? I don't have Eric's insight here. > > > > > > > > > > > > > > > > > > > > > > > > We would need a short's worth of padding in order to prevent the > > > > > > > > > > > > fields from occupying the same address as si_addr_lsb. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >> > + unsigned char _top_byte; > > > > > > > > > > > > > >> > + unsigned char _top_byte_mask; > > > > > > > > > > > > > >> > + } _addr_top_byte; > > > > > > > > > > > > > >> > +#endif > > > > > > > > > > > > > >> > }; > > > > > > > > > > > > > >> > } _sigfault; > > > > > > > > > > > > > >> > > > > > > > > > > > > > > >> > > > > > > > > > > > > > >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > > > > > > > > > > > > > >> > > > > > > > > > > > > > >> Please remove the "#ifdef __aarch64__". If at all possible we want to > > > > > > > > > > > > > >> design this so any other architecture who has this challenge can use the > > > > > > > > > > > > > >> code. The kind of code does not get enough attention/maintenance if it > > > > > > > > > > > > > >> is built for a single architecture. > > > > > > > > > > > > > > > > > > > > > > > > Seems reasonable. I was recently made aware that RISC-V was > > > > > > > > > > > > considering a similar feature: > > > > > > > > > > > > https://lists.riscv.org/g/tech-tee/topic/risc_v_tbi_proposal/72855478 > > > > > > > > > > > > I would have opted to expand this to other architectures on an > > > > > > > > > > > > as-needed basis, but I'd also be fine with having it on all > > > > > > > > > > > > architectures from the start. > > > > > > > > > > > > > > > > > > > > > > > > If we make this arch-independent, we have an additional concern, which > > > > > > > > > > > > is "what if some future architecture wants more than one byte here?" > > > > > > > > > > > > For example, an architecture may have a "top-two-bytes-ignore" > > > > > > > > > > > > feature, which would imply two-byte (misnamed) "si_addr_top_byte" and > > > > > > > > > > > > "si_addr_top_byte_mask" fields. And the RISC-V proposal potentially > > > > > > > > > > > > implies many more ignored bits (see slide 13 of the presentation). The > > > > > > > > > > > > maximum size that these fields can possibly be is the size of a > > > > > > > > > > > > pointer, and with that there wouldn't be enough room in the padding at > > > > > > > > > > > > this point to accommodate the new fields. > > > > > > > > > > > > > > > > > > > > > > > > That basically implies your earlier suggestion of adding a union > > > > > > > > > > > > member here to accommodate future expansion of the union, and adding > > > > > > > > > > > > the new fields after the union. I'm happy to make that change, with > > > > > > > > > > > > the fields renamed "si_addr_ignored" and "si_addr_ignored_mask". > > > > > > > > > > > > > > > > > > > > > > I think what we need here is basically a flags word. > > > > > > > > > > > > > > > > > > > > > > So long as we keep a flag spare to indicate the existence of a further > > > > > > > > > > > flags word, we can extend as needed. > > > > > > > > > > > > > > > > > > > > > > How the existence of the first flags words is detected is another > > > > > > > > > > > problem. If it only applies for newly-defined si_code values, then > > > > > > > > > > > I guess si_code may be sufficient. > > > > > > > > > > > > > > > > > > > > Existing kernels will zero-initialize unused regions of the siginfo > > > > > > > > > > data structure. The zero-initialization of the padding at the end of > > > > > > > > > > the struct is done by the clear_user call here: > > > > > > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/kernel/signal.c#L3193 > > > > > > > > > > > > > > > > > > > > and the zero-initialization of the padding between fields and unused > > > > > > > > > > union members is done by the clear_siginfo function which the kernel > > > > > > > > > > calls when initializing the data structure: > > > > > > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/include/linux/signal.h#L20 > > > > > > > > > > > > > > > > > > > > Therefore, a flag word value of 0 may be used to detect a lack of > > > > > > > > > > support for flagged fields. > > > > > > > > > > > > > > > > > > It's not enough that we do this today. We would have had to do it back > > > > > > > > > to the dawn of time (though in the arm64 case I guess we just need to go > > > > > > > > > back to when the arch/arm64 was merged). > > > > > > > > > > > > > > > > > > v2.6.12:kernel/signal.c:copy_siginfo_to_user() suggests that this wasn't > > > > > > > > > always the case, so unused parts of siginfo could be full of old junk > > > > > > > > > from the user stack, if the kernel is sufficiently old. > > > > > > > > > > > > > > > > > > If we're trying to do something generic that makes sense on all arches, > > > > > > > > > this matters. I may have misunderstood something about the code though. > > > > > > > > > > > > > > > > Hmm, I think you're right. The current behavior was introduced by > > > > > > > > commit c999b933faa5e281e3af2e110eccaf91698b0a81 which was first > > > > > > > > released in 4.18. So if an application wants to be compatible with > > > > > > > > pre-4.18 kernels then there would need to be some other way to > > > > > > > > indicate that the fields are valid. Probably the simplest way would be > > > > > > > > to have the application issue a uname(2) syscall and check the kernel > > > > > > > > version before reading these fields. I have a couple of other ideas > > > > > > > > that don't rely on version detection, if we'd prefer to avoid that. > > > > > > > > (They are somewhat ugly, but our hand is forced by backwards > > > > > > > > compatibility.) > > > > > > > > > > > > > > > > One idea is to re-purpose the si_errno field as a flags field for > > > > > > > > certain signal numbers. I checked a few kernel releases going back to > > > > > > > > 2.6.18 and it looks like the field is set to 0 except in the following > > > > > > > > circumstances: > > > > > > > > - sending a hardware breakpoint (SIGTRAP/TRAP_HWBKPT) > > > > > > > > - seccomp failures (SIGSYS/SYS_SECCOMP) > > > > > > > > - user-defined signal via kill_pid_usb_asyncio > > > > > > > > - SIGSWI in 3.18 and before (code since removed) > > > > > > > > > > > > > > > > It is also set to EFAULT for certain SIGSEGV/SEGV_MAPERR signals on > > > > > > > > powerpc since commit c96c4436aba4c12f1f48369f2f90bc43e12fe36c, which > > > > > > > > is currently unreleased. So if we wanted to go this route for SIGSEGV > > > > > > > > we would need to stop the kernel from setting si_errno to EFAULT for > > > > > > > > this signal before the 5.8 release. > > > > > > > > > > > > > > > > Another idea was to have userspace set a flag in sa_flags when > > > > > > > > registering a signal handler meaning "this signal handler requires > > > > > > > > unknown siginfo fields to be zeroed", and have existing kernels reject > > > > > > > > the syscall due to an unknown flag being set, but unfortunately this > > > > > > > > won't work because existing kernels do not reject sigaction syscalls > > > > > > > > with unknown flags set in sa_flags. A perhaps more radical idea in > > > > > > > > this vein would be to claim some of the upper bits of the signal > > > > > > > > number as flags that will cause the syscall to be rejected if set and > > > > > > > > unknown to the kernel. Existing kernels (going back to at least > > > > > > > > 2.6.18) contain this code in do_sigaction: > > > > > > > > > > > > > > > > if (!valid_signal(sig) || sig < 1 || (act && sig_kernel_only(sig))) > > > > > > > > return -EINVAL; > > > > > > > > > > > > > > > > and vald_signal is defined as: > > > > > > > > > > > > > > > > static inline int valid_signal(unsigned long sig) > > > > > > > > { > > > > > > > > return sig <= _NSIG ? 1 : 0; > > > > > > > > } > > > > > > > > > > > > > > > > All architectures define _NSIG as a value <= 128, so they will reject > > > > > > > > a signal number with any of bits 8-31 set. This means that we can use > > > > > > > > any of those bits for mandatory flags. Most likely we could use bit 30 > > > > > > > > (expanding down as necessary), as it keeps the signal number positive > > > > > > > > and permits future expansion of the signal number range. > > > > > > > > > > > > > > Does the signal core code actually gurantee to zero the unused fields? > > > > > > > Unless the fields are poked in by hand this is fraught with subtlelies, > > > > > > > especially when unions are involved. (I'm sure the code tries to do it, > > > > > > > but I've not eyeballed it in detail...) > > > > > > > > > > > > It memsets the siginfo structure before setting the fields and sending > > > > > > the signal (grep for clear_siginfo which is just a memset; you should > > > > > > find a call before all callers of force_sig_info). Memset is the right > > > > > > approach here since unlike setting fields by hand it clears padding > > > > > > which could lead to information leaks from the kernel. IIUC this is > > > > > > the reason why Eric wants all of the signals to be raised via wrappers > > > > > > in kernel/signal.c instead of via force_sig_info directly (to make > > > > > > this aspect easier to audit). > > > > > > > > > > My impression was that the reason for this model is partly to ensure > > > > > that siginfo fields are populated more consistently. When it was all > > > > > down to the individual callers, inconsistencies creeped in. > > > > > > > > > > With regard to memset(), this is not a complete defence against data > > > > > leakage. Assigning to a struct member can set any or all padding in > > > > > the struct to random garbage (consider write-combining of neighboring > > > > > member writes into a single larger accesses in asm for example). The > > > > > > > > I don't believe that LLVM will store to padding like this. I don't > > > > know about GCC, though, but I wouldn't be surprised if this is > > > > something that the kernel would want to turn off in "kernel C" (like > > > > it turns off strict aliasing) specifically because of the information > > > > leak issue. > > > > > > Again, the issue is not future kernel builds -- we can always find a way > > > to fix the behaviour for those -- but past kernel builds. > > I thought that the whole point of the "bit in the signal number" (or > SI_CODEX or whatever) was that we didn't need to worry about the > behavior of past kernel builds? It depends on what we use the new flag(s) for. If the flag means just that unused padding is safely zeroed, that could work -- but we'd want high confidence that it really is zeroed even in wacky configurations. > > > > > only way to avoid this is to ensure that the struct is 100% > > > > > padding-free, and that each member of a union is the same size. A > > > > > quick clance at <uapi/asm-generic/siginfo.h> confirms that this is not > > > > > the case. > > > > > > > > > > This might need to be looked at separately. > > > > > > > > > > But it does mean, strictly speaking, that we can't reliably add new > > > > > fields anywhere that there was previously padding: assigning to > > > > > neighboring members can still fill those with garbage after the > > > > > memset(). > > > > > > > > ...but this is largely moot because I'm not proposing to add new > > > > fields in the padding any more (because the fields needed to become > > > > larger in order to accommodate future hypothetical architectures which > > > > might want to use the fields, and thus they wouldn't fit in the > > > > padding). The siginfo.h diff would be something like: > > > > > > > > diff --git a/include/uapi/asm-generic/siginfo.h > > > > b/include/uapi/asm-generic/siginfo.h > > > > index cb3d6c267181..4a2fe257415d 100644 > > > > --- a/include/uapi/asm-generic/siginfo.h > > > > +++ b/include/uapi/asm-generic/siginfo.h > > > > @@ -91,7 +91,10 @@ union __sifields { > > > > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > __u32 _pkey; > > > > } _addr_pkey; > > > > + void *_pad[6]; > > > > }; > > > > + uintptr_t _ignored_bits; > > > > + uintptr_t _ignored_bits_mask; > > > > > > This _is_ in padding: the tail-padding of the (previously smaller) > > > _sigfault. Again, the compiler was allowed to populate this area with > > > junk before these fields were added. > > > > > > I agree that it seems fairly unlikely that the compiler would have been > > > overwriting this in normal circumstances, but that's not a guarantee. > > > My worry is that if this goes wrong, it will go wrong silently and > > > unpredictably. > > > > > > > } _sigfault; > > > > > > > > /* SIGPOLL */ > > > > > > > > or with a "uintptr_t _flags" added in before _ignored_bits if we go with that. > > > > > > > > > > > Using unused bits in the signal number to turn on new functionality > > > > > > > feels risky. As currently specified, this is just a number. Since > > > > > > > today a successful sigaction(n ...) guarantees that n is a valid signal > > > > > > > number, reasonable code like the following would trigger a buffer > > > > > > > overrun if we start trying to encode anything else in there: > > > > > > > > > > > > > > struct sigaction actions[NSIG]; > > > > > > > > > > > > > > int do_something( ... ) > > > > > > > { > > > > > > > ... > > > > > > > > > > > > > > if (!sigaction(n, sa, ...)) { > > > > > > > actions[n] = *sa; > > > > > > > return 0; > > > > > > > } > > > > > > > > > > > > > > ... > > > > > > > } > > > > > > > > > > > > I imagine the bit in the signal number being set by the direct caller > > > > > > to sigaction, and we could specifically recommend that calling > > > > > > pattern. In that case, your "n" wouldn't have the bit set in it. It > > > > > > > > > > I can imagine this too, but that doesn't mean that software does it. > > > > > > > > > > If the above kind of thing exists in a framework or library somewhere, > > > > > we could get problems. Similarly, a pre-existing LD_PRELOAD framework > > > > > that provides a wrapper for sigaction may now go wrong even if your > > > > > pattern is followed -- i.e., the caller thinks it's calling sigaction > > > > > directly but in fact it isn't. > > > > > > > > I'm aware of one library like that. It's called libsigchain, and it > > > > has an early bounds check: > > > > https://cs.android.com/android/platform/superproject/+/master:art/sigchainlib/sigchain.cc;l=371 > > > > > > > > Until the library is changed to recognize the flag, calling code would > > > > see the return value of -1 as if the kernel failed the syscall, and > > > > would fall back to the code for old kernels. > > > > > > But only after some bad dereferences. If these were writes, this means > > > that memory _may_ be silently corrupted (I don't say it't likely in a > > > given case, and we cannot pick a flag bit that makes this impossible). > > You're talking about libsigchain, right? I don't see any bad > references, the function returns after noticing the bounds check > failure. Yes, I confused myself by reading Handler() out of context. The kernel will invoke this with signo to a real signal number (without any flags). The sigaction wrapper does the bounds check before doing anything else, just as you say -- so that looks fine. (Side question: is all this thread-safe? Is there some implicit locking somewhere?) > > > So, _even though the user program is correct_, our change may trigger > > Let's say that you were talking about some other library and not > libsigchain. Such an interceptor wouldn't be correct though, it failed > to account for our change to the syscall semantics. If the accesses > were before the syscall (or the bounds check), then the interceptor > would not have been correct in the first place because POSIX requires > returning -1 with errno=EINVAL (and not crashing) if the signal number > is invalid. > > > > the corruption of arbitrary user memory. This what I mean by an ABI > > > break. The fact that the corruption is not done by the syscall itself > > > is no excuse. > > At some point, though, accommodating interceptors becomes pretty much > tantamount to saying "we can never change anything". Even just adding > a field to __sifields (which is pretty much required for what we need > to do) could break things in the presence of some interceptors because > the interceptor could be copying the fields manually to a new data > structure before calling the user's signal handler (e.g. because it > wants to defer the signal until later) and miss our new field. I think > most of the other ideas we're discussing fail to meet this bar as well > and I'll go into more details later on. I agree we cannot always avoid breaking such things. But we should do our best to avoid it. > > > We also fail to notice failures in sigaddset() etc., though in this code > > > it looks like that should not matter. > > Maybe you're looking at the handler ("SignalChain::Handler")? The bit > wouldn't be set in the signo argument to the handler. I'm talking > about line 371 of the code I linked, in the sigaction interceptor > "__sigaction" (it looks like sometimes the link doesn't take you to > the correct line for some reason). Ack, I confused myself. > > > > In general I think that any library like this with independent > > > > tracking of the kernel's purported signal handler state would need to > > > > be very sensitive to which syscalls are capable of setting signal > > > > handlers, what their semantics are, and so on. This applies to any > > > > change that we might make to the signal handler interface. So for > > > > example, if we introduced a new syscall as you propose below, and the > > > > library hasn't been updated to recognize the new syscall, it will > > > > silently miss changes in signal handler state caused by the new > > > > syscall. > > > > > > > > At the end of this argument lies "we can never change anything about > > > > how signal handlers work because it could break some interposing > > > > library somewhere" -- replace "signal handlers" with any kernel > > > > feature whose behavior may be modified by an interposing library if > > > > you like -- and I don't think we want to go that far. As far as I > > > > know, this isn't really the kernel's business anyway -- the kernel's > > > > stable ABI contract starts and ends with the syscall interface and not > > > > some library on top. > > > > > > > > That being said, we should perhaps try to define our interface so that > > > > something reasonable will probably happen if there is such a library > > > > and it hasn't been updated. With the new syscall, the library will > > > > sometimes silently fail to work in some non-local fashion. With the > > > > flag bit in the signal number, the library will either cause the > > > > caller to fall back to the old kernel code path (if there is a bounds > > > > check) or likely crash loudly (if there is no bounds check). To me, > > > > the "flag bit in the signal number" behavior seems more reasonable, > > > > since either something correct or something easy to debug will > > > > probably happen at runtime. > > > > > > > > > > could only appear in newly-written code that doesn't follow our > > > > > > recommendations, and there are already plenty of much more likely ways > > > > > > to cause buffer overflows in C code that doesn't follow > > > > > > recommendations anyway. (And even if such a buffer overflow occurred, > > > > > > it would very likely be caught early in development by the MMU due to > > > > > > the magnitude of the number 1<<30.) > > > > > > > > > > Choosing the bit value is hard. If shitfing it overflows, this can > > > > > trigger random undefined behaviour in the compiler in addition to (or > > > > > perhaps instead of) an out-of-bounds access or segfault. > > > > > > > > It wouldn't overflow on a 64-bit architecture assuming normal array > > > > indexing (the index would be promoted to pointer width before being > > > > scaled to the array element size), and to begin with the users of this > > > > would be 64-bit. > > > > > > Unless we don't offer this feature for 32-bit at all (possible, if ugly) > > > we can't stop people using it. > > My point is that the problem in the interceptor library would probably > be noticed on 64-bit (since that's what most people use these days), > which would probably result in it being fixed by the time it reaches > 32-bit users. Agreed. But we shouldn't take such bets unless we really have to. > > > > > If shifting it doesn't overflow, we might still fall into a valid > > > > > mapping, though I'd agree a segfault is more likely. > > > > > > > > > > > > > > > > > > I think it would be cleaner for to add a single flag field that can be > > > > > > > used for detecting other extensions, and request it via a new sa_flags > > > > > > > bit. This removes the need for sematically useless zeroing of unused > > > > > > > fields (though for hygiene and backwards compatibility reasons we would > > > > > > > probably want to carry on zeroing them anyway). > > > > > > > > > > > > > > I can see no simpler way to add supplementary siginfo fields for > > > > > > > existing si_codes. For si_codes that didn't exist before the zeroing > > > > > > > came in we could still detect optional si_code-specific fields via > > > > > > > zeroing, but it seems messary to have two ways of detecting extensions. > > > > > > > > > > > > That would certainly be cleaner if it worked, but that would only be > > > > > > the case if old kernels rejected unknown bits in sa_flags, and > > > > > > unfortunately they don't. With the bit in the signal number, the "old > > > > > > > > > > Hmm, that is a problem I wasn't aware of. > > > > > > > > > > > kernels reject" behavior admits relatively straightforward usage code: > > > > > > > > > > > > void set_segv_handler(void) { > > > > > > struct sigaction sa; > > > > > > sa.sa_sigaction = handle_segv; > > > > > > sa.sa_flags = SA_SIGINFO; > > > > > > if (sigaction(SIGSEGV | SF_CLEAR_UNKNOWN_FIELDS, &sa, 0) < 0) { // > > > > > > succeeds in new kernels, fails in old kernels > > > > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > > > > if (sigaction(SIGSEGV, &sa, 0) < 0) // succeeds in old kernels > > > > > > perror("sigaction"); > > > > > > } > > > > > > } > > > > > > > > > > > > void clear_fields_and_handle_segv(int signum, siginfo_t *sa, void *ctx) { > > > > > > sa->si_future_field = 0; > > > > > > handle_segv(signum, sa, ctx); > > > > > > } > > > > > > > > > > > > void handle_segv(int signum, siginfo_t *sa, void *ctx) { > > > > > > // At this point, si_future_field will have the value 0 in old > > > > > > kernels and the kernel-supplied value in new kernels. > > > > > > } > > > > > > > > > > > > Imagine if we moved the flag SF_CLEAR_UNKNOWN_FIELDS from the signal > > > > > > number to sa_flags. In that case, the first sigaction would succeed in > > > > > > old kernels so handle_segv wouldn't know whether it can safely read > > > > > > from si_future_field. With the sa_flags approach, you would need > > > > > > kernel version number checking via uname before setting the flag in > > > > > > sa_flags, and at that point why even have the flag in sa_flags at all > > > > > > since you could just have the signal handler conditionally read from > > > > > > si_future_field based on the uname? > > > > > > > > > > Software setting SA_SIFLAGS (or whatever) is new by definition, since > > > > > it would be using a new #define. So it might be reasonable to put the > > > > > burden on that software to verify that the flag was really accepted by > > > > > the kernel, by reading it back. > > > > > > > > That doesn't seem like a good idea even if it worked, because it could > > > > lead to race conditions. If the si_flags-reading signal handler were > > > > invoked in response to a signal between when you set it and when you > > > > ended up replacing it with the fallback signal handler for old > > > > kernels, the handler may end up reading garbage data from si_flags. > > > > > > Not really. My example may have this problem, but the signal handler > > > can be written to support both scenarios, based on testing a flag that > > > the main program sets after verifying that the flag could be set. Or > > > the signal could be blocked around establishment (often a good idea for > > > other reasons). > > > > > > But I agree it's a bit gross, and anyway doesn't work due to the fact > > > that the kernel doesn't filter out unrecognised flags anyway. > > > > > > > > Unfortunately, even relatively recent kernels blindly store sa_flags > > > > > in the kernel without validating it, and so it looks like duff flags > > > > > can be read back out via a sigaction() call. Dang. > > > > > > > > > > > > > > > Perhaps a new frontend syscall could be added. A new libc that knows > > > > > about this "sigaction2" could use it and mask off problem bits from > > > > > sa_flags in its sigaction() wrapper before calling sigaction2. An old > > > > > libc would call the old sigaction syscall, where we would ignore these > > > > > new sa_flags bits as before. > > > > > > > > I'm not currently in favor of the new syscall but if we do this I > > > > would keep sigaction and sigaction2 separate. That is, libc sigaction > > > > should always use the sigaction syscall, and libc sigaction2 should > > > > always use the sigaction2 syscall. We should avoid libc's sigaction > > > > having different behavior based on the libc version and kernel > > > > version, as that would make it harder to reason about its behavior. > > > > Calling code would need to check for presence of sigaction2 in both > > > > libc and the kernel, e.g. > > > > > > > > __attribute__((weak)) decltype(sigaction2) sigaction2; > > > > > > > > void set_segv_handler(void) { > > > > struct sigaction sa; > > > > sa.sa_sigaction = handle_segv; > > > > sa.sa_flags = SA_SIGINFO | SA_SIFLAGS; > > > > if (!sigaction2 || sigaction2(SIGSEGV, &sa, 0) < 0) { > > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > > sa.sa_flags = SA_SIGINFO; > > > > if (sigaction(SIGSEGV, &sa, 0) < 0) > > > > perror("sigaction"); > > > > } > > > > } > > > > > > I guess. But I share your distaste for adding a new syscall. > > > > > > > > > > > > This may not be a popular approach though, and software wouldn't be able > > > > > to use our new features until libc is updated to match. > > > > > > > > > > If we go down this route, it may provide additional opportunities to fix > > > > > annoying defects in the old interface. > > > > > > > > > > > > > > > > Note that the same applies to a flag indicating the availability of a > > > > > > si_flags field in sigaction (just > > > > > > s/SF_CLEAR_UNKNOWN_FIELDS/SF_HAS_SI_FLAGS/ and > > > > > > s/si_future_field/si_flags/ in the usage code above). In terms of > > > > > > SF_CLEAR_UNKNOWN_FIELDS versus SF_HAS_SI_FLAGS I'd be fine either way. > > > > > > > > > > > > Another thought that occurred to me is that we may consider > > > > > > generalizing this a step further and introducing a single flag bit in > > > > > > the signal number that means "reject unknown flags in sa_flags". This > > > > > > would mean that we wouldn't need to add any more flag bits to the > > > > > > signal number in the future, thus limiting this signal number hack to > > > > > > a single bit; all future mandatory behavior changes could just be put > > > > > > behind a flag in sa_flags and userspace code would easily be able to > > > > > > detect missing support for a flag and fall back if necessary. In our > > > > > > case, this would imply usage code like this: > > > > > > > > > > > > void set_segv_handler(void) { > > > > > > struct sigaction sa; > > > > > > sa.sa_sigaction = handle_segv; > > > > > > sa.sa_flags = SA_SIGINFO | SA_CLEAR_UNKNOWN_FIELDS; > > > > > > // Succeeds in kernels with SA_CLEAR_UNKNOWN_FIELDS support. > > > > > > // Fails in kernels with SF_CHECK_SA_FLAGS support but no > > > > > > SA_CLEAR_UNKNOWN_FIELDS support (because of the unknown flags check). > > > > > > // Fails in kernels without SF_CHECK_SA_FLAGS support (because of > > > > > > the bounds check on the signal number). > > > > > > if (sigaction(SIGSEGV | SF_CHECK_SA_FLAGS, &sa, 0) < 0) { > > > > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > > > > sa.sa_flags = SA_SIGINFO; > > > > > > // Succeeds in old kernels, no need to use SF_CHECK_SA_FLAGS since > > > > > > we're using sa_flags from the beginning of time. > > > > > > if (sigaction(SIGSEGV, &sa, 0) < 0) > > > > > > perror("sigaction"); > > > > > > } > > > > > > } > > > > > > > > > > As with the other options this could work, but looks like it could > > > > > break the ABI due to violating the original semantics for the signal > > > > > number argument. Perhaps I'm being too paranoid. > > > > > > > > There's no ABI being broken here, as long as we consider syscalls to > > > > be the stable ABI layer. Old kernels are simply rejecting arguments > > > > that they don't know about yet. By that argument, any introduction of > > > > a new syscall is an ABI break because it changes the semantics of a > > > > previously-unallocated syscall number. > > > > > > As argued above, I think this is an invalid argument. > > > > > > Although any addition will change behaviour (so is a break in some > > > sense), the key is not to make "surprising" changes. > > If we care about interceptors then I don't think "surprising" comes > into it. It's more a question of "does the anticipated behavior of the > interceptor match our desired behavior", where "desired" means "most > likely to avoid silent breakage". We would need to get into the head > of a potential interceptor author and think about how they would have > handled the signal number argument, as well as other arguments like > sa_flags if we want to go that route, and see whether that behavior > would lead to the desired result. That's exactly what I mean by "surprising". However, not every interceptor author will be making the same assumptions, and not every bit of software affected will be an interceptor. So some judgement needs to be applied. > In this case, I think we exactly want the interceptor author to have > thought "oh, it's just a number, I'll (possibly do a bounds check and > then) use the number as an index into an array". This will lead to one > of two outcomes: crashing (yes, yes, it won't always crash, but if the > alternative is that it never crashes and we get silently incorrect > behavior all of the time, I'll take sometimes crashing) or fail the > bounds check and pretend to be an old kernel (the latter is > anticipated by POSIX which requires returning -1/EINVAL for an invalid > signal number). Each of these behaviors are desirable, as they are > observable failures, which are more likely to result in fixes than > silent ones. Agreed, except wanting the author to have thought something doesn't ensure that they actually did think that. > > > Having something random happen when setting a previously reserved flag > > > bit, or when issuing a syscall when an unknown syscall number, or not > > > surprising at all. > > Introducing a new syscall is right out in this model. The interceptor > author wouldn't have anticipated our introducing a new syscall, so the > new syscall wouldn't be intercepted and calls to the new syscall would > silently bypass the interceptor. For example, adding sigaction2 could > result in signal handlers being set without the interceptor's > knowledge. Agreed. My sentence was a bit mangled: I mean to say "Having something random happen when [...] issuing a syscall *with* an unknown syscall number *is* not surprising at all." I agree that adding a new syscall is problematic if we want to avoid breaking existing interceptors in particular. Other types of code are much less likely to be affected by the addition of new syscalls. > Regarding a sa_flags bit, let's get inside the head of the interceptor > author again. How would they handle a flag bit that they don't > recognize when replacing the signal handler? It wouldn't be correct to > just pass it through to the kernel, or drop the flag on the floor, as > it might be semantically meaningful (and thus could change the calling > convention as SA_SIGINFO does, or change the meaning of fields in > siginfo, as SA_CODEX would do). A correctly written sigaction > interceptor should probably abort the program upon encountering an > unknown flag (thus giving a human a chance to update the interceptor), > but chances are that they don't. Ignoring all but a few flags (and > passing a fixed set of flags to the kernel) seems to be what > libsigchain does, and in the case of SA_CODEX it would seem to result > in desirable behavior (but I suspect that it isn't handling the other > flags correctly), but I could also see an interceptor author just > passing it unchanged to the kernel without checking it (perhaps > because they didn't think about these issues, and because that didn't > matter until now, with the exception of from-the-beginning-of-time > flags like SA_SIGINFO). And with SA_CODEX that could lead to silent > misreading of si_code in the interceptor's signal handler, if it > hasn't been updated to use the new macros. Agreed. I've tried to implement things rather like this in the past, and how to interpret the flags is a tricky issue. Some of the flags are impossible to emulate even when you know what they mean, in particular SA_NODEFER and SA_RESTART. Making new flags safe to ignore and harmless to set of you don't know what they mean is the safest approach, but not always possible (I think I managed this with by suggestion below, though). Ideally, a flags field should be specified with rules that say exactly what to do with flags you don't recognise. Sadly this is usually not thought about until it's too late. > > > Making fundamental changes to the encoding of an existing argument is > > > highly surprising, on the other hand: as your example shows, it is > > > reasonable to index an array using a signal number. > > > > > > I agree that this doesn't get us closer to a practical solution though. > > > > > > > > > But we do seem to need some mechanism in addition to (or instead of) > > > sa_flags. > > > > > > Here's another thought: > > > > > > Since si_flags would be either always present or always absent, it > > > could make sense to have a global property to report this, rather than > > > an sa_flags or signal number bit to request it per-signal. > > > > > > Requiring software to parse uname() might be reasonable for that, if > > > cumbersome (did you suggest this previously?). If we're concerned that > > > the awkwardness of this would encourage people not to bother (or > > > encourage people to do it wrong) then we might opt for something simpler > > > like an AT_FLAGS bit. > > > > > > Ultimately libc could provide a more portable interface for discovery, > > > such as via sysconf(). > > > > > > Thoughts? > > Yes, this was all to avoid the userspace code needing to contain a > version check (or equivalent). Maybe the version check would be better > than the alternatives though (although it's still vulnerable to > non-updated interceptors not copying our new fields). AT_FLAGS sounds > good to me. Ack (though not needed if we use the approach outlined below). > > > While you're thinking about that, here's another idea: > > > > It occurs to me that there are spare bits in si_code. si_code is an > > enum, but unlike the signal number there are no specific bounds for > > this value, so we may have an easier time recycling bits here. > > > > The high bits of si_code are usually sign-extension and so not always > > 0, but we can XOR flags into them provided we don't forget the real sign. > > > > Software that isn't expecting twiddled bits would get confused, so we > > need a new SA_ flag to enable this. But this flag (SA_CODEX) below is > > now just a request. If the kernel doesn't understand it (or without > > SA_SIGINFO) then no flags would be reported in si_code, which is > > backwards-compatible. > > > > A handler would now do > > > > void handler(int n, siginfo_t *si, ...) > > { > > int flags = SI_FLAGS(si->si_code); > > int code = SI_CODE(si->si_code); > > > > if (!(flags & SIF_CODEX) { > > /* flags not supported */ > > /* Careful assignment of flag meanings may make this > > check unnecessary, but it's probably useful for > > developers for testing their code. */ > > } > > > > /* Handle signal based on n, code and flags */ > > } > > > > > > If the kernel doesn't report any flags (perhaps because it's too old) > > then SI_FLAGS() will yield 0 and SI_CODE() will just return si_code > > unchanged. This means that even non-SA_CODEX handlers can use these > > macros, which may ease migration. > > Thanks, this seems more appealing to me than the bit in the signal > number idea. It uses the sa_flags field as intended and doesn't abuse > the fields of the siginfo data structure too much. I don't think we > should put as much weight into interceptor concerns as you do, so you > can consider my above argumentation to be from a devil's advocate > perspective. Ditto: we shouldn't break forwards compatibility wherever we can avoid doing so, but with APIs that are not well designed for forwards compatibility (the signal API included) we have to be realistic. The interceptor case is just one example of something that could break, but I'll admit I couldn't come up with a more convincing example so far. > > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > index cb3d6c2..4e77c71 100644 > > --- a/include/uapi/asm-generic/siginfo.h > > +++ b/include/uapi/asm-generic/siginfo.h > > @@ -176,6 +176,18 @@ typedef struct siginfo { > > #define SI_DETHREAD -7 /* sent by execve() killing subsidiary threads */ > > #define SI_ASYNCNL -60 /* sent by glibc async name lookup completion */ > > > > +#define __SI_FLAGS 0x7ffff000 /* optional code extension flags */ > > +#define SIF_CODEX 0x40000000 /* code extension flags supported */ > > I don't think we would need this flag because even old kernels > "support" the extension flags under this scheme (they just don't > report any features). We would only need to introduce flags for actual > features. Agreed, it's not strictly required. It occurred to me that someone writing code in userspace would want a way to verify that they'd set the flag correctly and that the kernel supported it. But there ought to be other ways to do that once real si_code flags get added. Cheers ---Dave
On Mon, Jul 13, 2020 at 6:24 AM Dave Martin <Dave.Martin@arm.com> wrote: > > On Wed, Jul 08, 2020 at 03:21:13PM -0700, Peter Collingbourne wrote: > > On Wed, Jul 8, 2020 at 6:58 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > On Wed, Jul 08, 2020 at 12:00:22PM +0100, Dave Martin wrote: > > > > On Tue, Jul 07, 2020 at 12:07:09PM -0700, Peter Collingbourne wrote: > > > > > On Tue, Jul 7, 2020 at 7:19 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > On Mon, Jul 06, 2020 at 12:20:33PM -0700, Peter Collingbourne wrote: > > > > > > > On Mon, Jul 6, 2020 at 9:41 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > > > On Wed, Jun 24, 2020 at 12:51:43PM -0700, Peter Collingbourne wrote: > > > > > > > > > On Wed, Jun 24, 2020 at 10:12 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > > > > > > > On Wed, Jun 24, 2020 at 09:51:49AM -0700, Peter Collingbourne wrote: > > > > > > > > > > > On Wed, Jun 24, 2020 at 2:28 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 05:40:08PM -0700, Peter Collingbourne wrote: > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 10:52 AM Eric W. Biederman > > > > > > > > > > > > > <ebiederm@xmission.com> wrote: > > > > > > > > > > > > > > > > > > > > > > > > > > > > Dave Martin <Dave.Martin@arm.com> writes: > > > > > > > > > > > > > > > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > > > > > > > > > > > > > > >> Peter Collingbourne <pcc@google.com> writes: > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > > > > > > > > > > > > > > >> > index 47f651df781c..a8380a2b6361 100644 > > > > > > > > > > > > > > >> > --- a/arch/arm64/kernel/traps.c > > > > > > > > > > > > > > >> > +++ b/arch/arm64/kernel/traps.c > > > > > > > > > > > > > > >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > > > > > > > > > > > > > > >> > } > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, > > > > > > > > > > > > > > >> > + unsigned long far, unsigned char far_tb_mask, > > > > > > > > > > > > > > >> > const char *str) > > > > > > > > > > > > > > >> > { > > > > > > > > > > > > > > >> > arm64_show_signal(signo, str); > > > > > > > > > > > > > > >> > - if (signo == SIGKILL) > > > > > > > > > > > > > > >> > + if (signo == SIGKILL) { > > > > > > > > > > > > > > >> > force_sig(SIGKILL); > > > > > > > > > > > > > > >> > - else > > > > > > > > > > > > > > >> > - force_sig_fault(signo, code, addr); > > > > > > > > > > > > > > >> > + } else { > > > > > > > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > > > > > > > >> > + info.si_signo = signo; > > > > > > > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > > > > > > > >> > + info.si_code = code; > > > > > > > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > > > > > > > >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > > > > > > > > > > > > > > >> > + info.si_addr_top_byte_mask = far_tb_mask; > > > > > > > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > > > > > > > >> > + } > > > > > > > > > > > > > > >> > } > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > > > > > > > > > > > > > > >> > - const char *str) > > > > > > > > > > > > > > >> > + unsigned long far, const char *str) > > > > > > > > > > > > > > >> > { > > > > > > > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > > > > > > > >> > + > > > > > > > > > > > > > > >> > arm64_show_signal(SIGBUS, str); > > > > > > > > > > > > > > >> > - force_sig_mceerr(code, addr, lsb); > > > > > > > > > > > > > > >> > + > > > > > > > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > > > > > > > >> > + info.si_signo = SIGBUS; > > > > > > > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > > > > > > > >> > + info.si_code = code; > > > > > > > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > > > > > > > >> > + info.si_addr_lsb = lsb; > > > > > > > > > > > > > > >> > + info.si_addr_top_byte = far >> 56; > > > > > > > > > > > > > > >> > + info.si_addr_top_byte_mask = 0xff; > > > > > > > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > > > > > > > >> > } > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > >> I have a real problem with this construction. force_sig_info is not an > > > > > > > > > > > > > > >> interface that should be used for anything except to define a wrapper > > > > > > > > > > > > > > >> that takes it's parameters. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Can you elaborate? How would you do this king of thing. > > > > > > > > > > > > > > > > > > > > > > > > > > > > There are no other uses of force_sig_info in architecture code. > > > > > > > > > > > > > > > > > > > > > > > > > > > > I just removed them _all_ because they were almost all broken. > > > > > > > > > > > > > > In fact your mcerr case is broken because it uses two different > > > > > > > > > > > > > > union members simultantiously. > > > > > > > > > > > > > > > > > > > > > > > > > > Is that really broken? I thought that the Linux kernel deliberately > > > > > > > > > > > > > didn't care about strict aliasing rules (the top-level Makefile passes > > > > > > > > > > > > > -fno-strict-aliasing) so I thought that it was valid in "Linux kernel > > > > > > > > > > > > > C" even though from a standards point of view it is invalid. (That > > > > > > > > > > > > > being said, this is probably moot with my proposed changes below > > > > > > > > > > > > > though.) > > > > > > > > > > > > > > > > > > > > > > > > I have a feeling that -fno-strict-aliasing only allows you to _read_ a > > > > > > > > > > > > different union member from the one previously written. > > > > > > > > > > > > > > > > > > > > > > > > Writing a different member from the last one written can still splatter > > > > > > > > > > > > on the other members IIUC. > > > > > > > > > > > > > > > > > > > > > > > > It would be better to keep things separate rather than risk > > > > > > > > > > > > incorrectness just to save a few bytes. > > > > > > > > > > > > > > > > > > > > > > > > IMHO -fno-strict-aliasing is no excuse for gratuitous type-punning. > > > > > > > > > > > > > > > > > > > > > > > > > > So I am looking for something like force_sig_mcerr or force_sig_fault > > > > > > > > > > > > > > that includes your new information that then calls force_sig_info. > > > > > > > > > > > > > > > > > > > > > > > > > > > > I know of no other way to safely use the siginfo struct. > > > > > > > > > > > > > > > > > > > > > > > > > > So you want something like: > > > > > > > > > > > > > > > > > > > > > > > > > > int force_sig_fault_with_ignored_bits(int signo, int code, void __user > > > > > > > > > > > > > *addr, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > > > > int force_sig_mceerr_with_ignored_bits(int code, void __user *addr, > > > > > > > > > > > > > short lsb, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > > > > > > > > > > > > > > > > > in kernel/signal.c and the code in arch/arm64 would call that? > > > > > > > > > > > > > > > > > > > > > > > > > > > > AIUI we absolutely need a forced signal here, we need to supply > > > > > > > > > > > > > > > metadata, and we don't have to open-code all that at every relevant > > > > > > > > > > > > > > > signal generation site... > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >> It is not clear to me that if you have adapted siginfo_layout. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Garbled sentence? > > > > > > > > > > > > > > > > > > > > > > > > > > > > Looks like. One of the pieces of code that needs to change > > > > > > > > > > > > > > when siginfo gets updated is siginfo_layout so that the structure > > > > > > > > > > > > > > can be properly decoded and made sense of. > > > > > > > > > > > > > > > > > > > > > > > > > > > > I am not seeing anything like that. > > > > > > > > > > > > > > > > > > > > > > > > > > Okay, this has to do with copying between the compat and non-compat > > > > > > > > > > > > > versions of the struct? Sure, I can update that, although the code > > > > > > > > > > > > > would be basically non-functional on arm64 because TBI isn't supported > > > > > > > > > > > > > on 32-bit ARM. > > > > > > > > > > > > > > > > > > > > > > > > > > > >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > > > > >> > index cb3d6c267181..6dd82373eb2d 100644 > > > > > > > > > > > > > > >> > --- a/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > > > > >> > +++ b/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > > > > >> > @@ -91,6 +91,14 @@ union __sifields { > > > > > > > > > > > > > > >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > > > > > > > > > > > >> > __u32 _pkey; > > > > > > > > > > > > > > >> > } _addr_pkey; > > > > > > > > > > > > > > >> > +#ifdef __aarch64__ > > > > > > > > > > > > > > >> > + /* used with all si_codes */ > > > > > > > > > > > > > > >> > + struct { > > > > > > > > > > > > > > >> > + short _dummy_top_byte; > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > ^ What's this for? I don't have Eric's insight here. > > > > > > > > > > > > > > > > > > > > > > > > > > We would need a short's worth of padding in order to prevent the > > > > > > > > > > > > > fields from occupying the same address as si_addr_lsb. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >> > + unsigned char _top_byte; > > > > > > > > > > > > > > >> > + unsigned char _top_byte_mask; > > > > > > > > > > > > > > >> > + } _addr_top_byte; > > > > > > > > > > > > > > >> > +#endif > > > > > > > > > > > > > > >> > }; > > > > > > > > > > > > > > >> > } _sigfault; > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > >> Please remove the "#ifdef __aarch64__". If at all possible we want to > > > > > > > > > > > > > > >> design this so any other architecture who has this challenge can use the > > > > > > > > > > > > > > >> code. The kind of code does not get enough attention/maintenance if it > > > > > > > > > > > > > > >> is built for a single architecture. > > > > > > > > > > > > > > > > > > > > > > > > > > Seems reasonable. I was recently made aware that RISC-V was > > > > > > > > > > > > > considering a similar feature: > > > > > > > > > > > > > https://lists.riscv.org/g/tech-tee/topic/risc_v_tbi_proposal/72855478 > > > > > > > > > > > > > I would have opted to expand this to other architectures on an > > > > > > > > > > > > > as-needed basis, but I'd also be fine with having it on all > > > > > > > > > > > > > architectures from the start. > > > > > > > > > > > > > > > > > > > > > > > > > > If we make this arch-independent, we have an additional concern, which > > > > > > > > > > > > > is "what if some future architecture wants more than one byte here?" > > > > > > > > > > > > > For example, an architecture may have a "top-two-bytes-ignore" > > > > > > > > > > > > > feature, which would imply two-byte (misnamed) "si_addr_top_byte" and > > > > > > > > > > > > > "si_addr_top_byte_mask" fields. And the RISC-V proposal potentially > > > > > > > > > > > > > implies many more ignored bits (see slide 13 of the presentation). The > > > > > > > > > > > > > maximum size that these fields can possibly be is the size of a > > > > > > > > > > > > > pointer, and with that there wouldn't be enough room in the padding at > > > > > > > > > > > > > this point to accommodate the new fields. > > > > > > > > > > > > > > > > > > > > > > > > > > That basically implies your earlier suggestion of adding a union > > > > > > > > > > > > > member here to accommodate future expansion of the union, and adding > > > > > > > > > > > > > the new fields after the union. I'm happy to make that change, with > > > > > > > > > > > > > the fields renamed "si_addr_ignored" and "si_addr_ignored_mask". > > > > > > > > > > > > > > > > > > > > > > > > I think what we need here is basically a flags word. > > > > > > > > > > > > > > > > > > > > > > > > So long as we keep a flag spare to indicate the existence of a further > > > > > > > > > > > > flags word, we can extend as needed. > > > > > > > > > > > > > > > > > > > > > > > > How the existence of the first flags words is detected is another > > > > > > > > > > > > problem. If it only applies for newly-defined si_code values, then > > > > > > > > > > > > I guess si_code may be sufficient. > > > > > > > > > > > > > > > > > > > > > > Existing kernels will zero-initialize unused regions of the siginfo > > > > > > > > > > > data structure. The zero-initialization of the padding at the end of > > > > > > > > > > > the struct is done by the clear_user call here: > > > > > > > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/kernel/signal.c#L3193 > > > > > > > > > > > > > > > > > > > > > > and the zero-initialization of the padding between fields and unused > > > > > > > > > > > union members is done by the clear_siginfo function which the kernel > > > > > > > > > > > calls when initializing the data structure: > > > > > > > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/include/linux/signal.h#L20 > > > > > > > > > > > > > > > > > > > > > > Therefore, a flag word value of 0 may be used to detect a lack of > > > > > > > > > > > support for flagged fields. > > > > > > > > > > > > > > > > > > > > It's not enough that we do this today. We would have had to do it back > > > > > > > > > > to the dawn of time (though in the arm64 case I guess we just need to go > > > > > > > > > > back to when the arch/arm64 was merged). > > > > > > > > > > > > > > > > > > > > v2.6.12:kernel/signal.c:copy_siginfo_to_user() suggests that this wasn't > > > > > > > > > > always the case, so unused parts of siginfo could be full of old junk > > > > > > > > > > from the user stack, if the kernel is sufficiently old. > > > > > > > > > > > > > > > > > > > > If we're trying to do something generic that makes sense on all arches, > > > > > > > > > > this matters. I may have misunderstood something about the code though. > > > > > > > > > > > > > > > > > > Hmm, I think you're right. The current behavior was introduced by > > > > > > > > > commit c999b933faa5e281e3af2e110eccaf91698b0a81 which was first > > > > > > > > > released in 4.18. So if an application wants to be compatible with > > > > > > > > > pre-4.18 kernels then there would need to be some other way to > > > > > > > > > indicate that the fields are valid. Probably the simplest way would be > > > > > > > > > to have the application issue a uname(2) syscall and check the kernel > > > > > > > > > version before reading these fields. I have a couple of other ideas > > > > > > > > > that don't rely on version detection, if we'd prefer to avoid that. > > > > > > > > > (They are somewhat ugly, but our hand is forced by backwards > > > > > > > > > compatibility.) > > > > > > > > > > > > > > > > > > One idea is to re-purpose the si_errno field as a flags field for > > > > > > > > > certain signal numbers. I checked a few kernel releases going back to > > > > > > > > > 2.6.18 and it looks like the field is set to 0 except in the following > > > > > > > > > circumstances: > > > > > > > > > - sending a hardware breakpoint (SIGTRAP/TRAP_HWBKPT) > > > > > > > > > - seccomp failures (SIGSYS/SYS_SECCOMP) > > > > > > > > > - user-defined signal via kill_pid_usb_asyncio > > > > > > > > > - SIGSWI in 3.18 and before (code since removed) > > > > > > > > > > > > > > > > > > It is also set to EFAULT for certain SIGSEGV/SEGV_MAPERR signals on > > > > > > > > > powerpc since commit c96c4436aba4c12f1f48369f2f90bc43e12fe36c, which > > > > > > > > > is currently unreleased. So if we wanted to go this route for SIGSEGV > > > > > > > > > we would need to stop the kernel from setting si_errno to EFAULT for > > > > > > > > > this signal before the 5.8 release. > > > > > > > > > > > > > > > > > > Another idea was to have userspace set a flag in sa_flags when > > > > > > > > > registering a signal handler meaning "this signal handler requires > > > > > > > > > unknown siginfo fields to be zeroed", and have existing kernels reject > > > > > > > > > the syscall due to an unknown flag being set, but unfortunately this > > > > > > > > > won't work because existing kernels do not reject sigaction syscalls > > > > > > > > > with unknown flags set in sa_flags. A perhaps more radical idea in > > > > > > > > > this vein would be to claim some of the upper bits of the signal > > > > > > > > > number as flags that will cause the syscall to be rejected if set and > > > > > > > > > unknown to the kernel. Existing kernels (going back to at least > > > > > > > > > 2.6.18) contain this code in do_sigaction: > > > > > > > > > > > > > > > > > > if (!valid_signal(sig) || sig < 1 || (act && sig_kernel_only(sig))) > > > > > > > > > return -EINVAL; > > > > > > > > > > > > > > > > > > and vald_signal is defined as: > > > > > > > > > > > > > > > > > > static inline int valid_signal(unsigned long sig) > > > > > > > > > { > > > > > > > > > return sig <= _NSIG ? 1 : 0; > > > > > > > > > } > > > > > > > > > > > > > > > > > > All architectures define _NSIG as a value <= 128, so they will reject > > > > > > > > > a signal number with any of bits 8-31 set. This means that we can use > > > > > > > > > any of those bits for mandatory flags. Most likely we could use bit 30 > > > > > > > > > (expanding down as necessary), as it keeps the signal number positive > > > > > > > > > and permits future expansion of the signal number range. > > > > > > > > > > > > > > > > Does the signal core code actually gurantee to zero the unused fields? > > > > > > > > Unless the fields are poked in by hand this is fraught with subtlelies, > > > > > > > > especially when unions are involved. (I'm sure the code tries to do it, > > > > > > > > but I've not eyeballed it in detail...) > > > > > > > > > > > > > > It memsets the siginfo structure before setting the fields and sending > > > > > > > the signal (grep for clear_siginfo which is just a memset; you should > > > > > > > find a call before all callers of force_sig_info). Memset is the right > > > > > > > approach here since unlike setting fields by hand it clears padding > > > > > > > which could lead to information leaks from the kernel. IIUC this is > > > > > > > the reason why Eric wants all of the signals to be raised via wrappers > > > > > > > in kernel/signal.c instead of via force_sig_info directly (to make > > > > > > > this aspect easier to audit). > > > > > > > > > > > > My impression was that the reason for this model is partly to ensure > > > > > > that siginfo fields are populated more consistently. When it was all > > > > > > down to the individual callers, inconsistencies creeped in. > > > > > > > > > > > > With regard to memset(), this is not a complete defence against data > > > > > > leakage. Assigning to a struct member can set any or all padding in > > > > > > the struct to random garbage (consider write-combining of neighboring > > > > > > member writes into a single larger accesses in asm for example). The > > > > > > > > > > I don't believe that LLVM will store to padding like this. I don't > > > > > know about GCC, though, but I wouldn't be surprised if this is > > > > > something that the kernel would want to turn off in "kernel C" (like > > > > > it turns off strict aliasing) specifically because of the information > > > > > leak issue. > > > > > > > > Again, the issue is not future kernel builds -- we can always find a way > > > > to fix the behaviour for those -- but past kernel builds. > > > > I thought that the whole point of the "bit in the signal number" (or > > SI_CODEX or whatever) was that we didn't need to worry about the > > behavior of past kernel builds? > > It depends on what we use the new flag(s) for. > > If the flag means just that unused padding is safely zeroed, that could > work -- but we'd want high confidence that it really is zeroed even in > wacky configurations. > > > > > > > only way to avoid this is to ensure that the struct is 100% > > > > > > padding-free, and that each member of a union is the same size. A > > > > > > quick clance at <uapi/asm-generic/siginfo.h> confirms that this is not > > > > > > the case. > > > > > > > > > > > > This might need to be looked at separately. > > > > > > > > > > > > But it does mean, strictly speaking, that we can't reliably add new > > > > > > fields anywhere that there was previously padding: assigning to > > > > > > neighboring members can still fill those with garbage after the > > > > > > memset(). > > > > > > > > > > ...but this is largely moot because I'm not proposing to add new > > > > > fields in the padding any more (because the fields needed to become > > > > > larger in order to accommodate future hypothetical architectures which > > > > > might want to use the fields, and thus they wouldn't fit in the > > > > > padding). The siginfo.h diff would be something like: > > > > > > > > > > diff --git a/include/uapi/asm-generic/siginfo.h > > > > > b/include/uapi/asm-generic/siginfo.h > > > > > index cb3d6c267181..4a2fe257415d 100644 > > > > > --- a/include/uapi/asm-generic/siginfo.h > > > > > +++ b/include/uapi/asm-generic/siginfo.h > > > > > @@ -91,7 +91,10 @@ union __sifields { > > > > > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > > __u32 _pkey; > > > > > } _addr_pkey; > > > > > + void *_pad[6]; > > > > > }; > > > > > + uintptr_t _ignored_bits; > > > > > + uintptr_t _ignored_bits_mask; > > > > > > > > This _is_ in padding: the tail-padding of the (previously smaller) > > > > _sigfault. Again, the compiler was allowed to populate this area with > > > > junk before these fields were added. > > > > > > > > I agree that it seems fairly unlikely that the compiler would have been > > > > overwriting this in normal circumstances, but that's not a guarantee. > > > > My worry is that if this goes wrong, it will go wrong silently and > > > > unpredictably. > > > > > > > > > } _sigfault; > > > > > > > > > > /* SIGPOLL */ > > > > > > > > > > or with a "uintptr_t _flags" added in before _ignored_bits if we go with that. > > > > > > > > > > > > > Using unused bits in the signal number to turn on new functionality > > > > > > > > feels risky. As currently specified, this is just a number. Since > > > > > > > > today a successful sigaction(n ...) guarantees that n is a valid signal > > > > > > > > number, reasonable code like the following would trigger a buffer > > > > > > > > overrun if we start trying to encode anything else in there: > > > > > > > > > > > > > > > > struct sigaction actions[NSIG]; > > > > > > > > > > > > > > > > int do_something( ... ) > > > > > > > > { > > > > > > > > ... > > > > > > > > > > > > > > > > if (!sigaction(n, sa, ...)) { > > > > > > > > actions[n] = *sa; > > > > > > > > return 0; > > > > > > > > } > > > > > > > > > > > > > > > > ... > > > > > > > > } > > > > > > > > > > > > > > I imagine the bit in the signal number being set by the direct caller > > > > > > > to sigaction, and we could specifically recommend that calling > > > > > > > pattern. In that case, your "n" wouldn't have the bit set in it. It > > > > > > > > > > > > I can imagine this too, but that doesn't mean that software does it. > > > > > > > > > > > > If the above kind of thing exists in a framework or library somewhere, > > > > > > we could get problems. Similarly, a pre-existing LD_PRELOAD framework > > > > > > that provides a wrapper for sigaction may now go wrong even if your > > > > > > pattern is followed -- i.e., the caller thinks it's calling sigaction > > > > > > directly but in fact it isn't. > > > > > > > > > > I'm aware of one library like that. It's called libsigchain, and it > > > > > has an early bounds check: > > > > > https://cs.android.com/android/platform/superproject/+/master:art/sigchainlib/sigchain.cc;l=371 > > > > > > > > > > Until the library is changed to recognize the flag, calling code would > > > > > see the return value of -1 as if the kernel failed the syscall, and > > > > > would fall back to the code for old kernels. > > > > > > > > But only after some bad dereferences. If these were writes, this means > > > > that memory _may_ be silently corrupted (I don't say it't likely in a > > > > given case, and we cannot pick a flag bit that makes this impossible). > > > > You're talking about libsigchain, right? I don't see any bad > > references, the function returns after noticing the bounds check > > failure. > > Yes, I confused myself by reading Handler() out of context. The kernel > will invoke this with signo to a real signal number (without any flags). > > The sigaction wrapper does the bounds check before doing anything else, > just as you say -- so that looks fine. > > (Side question: is all this thread-safe? Is there some implicit locking > somewhere?) I think maybe it isn't? There seem to be possible races on the handler_ field. One possibility is that the function could race with itself on another thread, which could be fixed via locking, but it would also need to handle races between itself and the signal handler, most likely by blocking the signal while setting it. > > > > So, _even though the user program is correct_, our change may trigger > > > > Let's say that you were talking about some other library and not > > libsigchain. Such an interceptor wouldn't be correct though, it failed > > to account for our change to the syscall semantics. If the accesses > > were before the syscall (or the bounds check), then the interceptor > > would not have been correct in the first place because POSIX requires > > returning -1 with errno=EINVAL (and not crashing) if the signal number > > is invalid. > > > > > > the corruption of arbitrary user memory. This what I mean by an ABI > > > > break. The fact that the corruption is not done by the syscall itself > > > > is no excuse. > > > > At some point, though, accommodating interceptors becomes pretty much > > tantamount to saying "we can never change anything". Even just adding > > a field to __sifields (which is pretty much required for what we need > > to do) could break things in the presence of some interceptors because > > the interceptor could be copying the fields manually to a new data > > structure before calling the user's signal handler (e.g. because it > > wants to defer the signal until later) and miss our new field. I think > > most of the other ideas we're discussing fail to meet this bar as well > > and I'll go into more details later on. > > I agree we cannot always avoid breaking such things. But we should do > our best to avoid it. I think that given the hand that we've been dealt, no matter what we do, we can't really avoid risking breaking something. The relevant questions are "what are we going to risk breaking", "how much risk is there", "will it be easily noticed/fixable", and "once we're on the other side of the potential breakage, will we find ourselves in a position where changing things involves less breakage risk". > > > > We also fail to notice failures in sigaddset() etc., though in this code > > > > it looks like that should not matter. > > > > Maybe you're looking at the handler ("SignalChain::Handler")? The bit > > wouldn't be set in the signo argument to the handler. I'm talking > > about line 371 of the code I linked, in the sigaction interceptor > > "__sigaction" (it looks like sometimes the link doesn't take you to > > the correct line for some reason). > > Ack, I confused myself. > > > > > > In general I think that any library like this with independent > > > > > tracking of the kernel's purported signal handler state would need to > > > > > be very sensitive to which syscalls are capable of setting signal > > > > > handlers, what their semantics are, and so on. This applies to any > > > > > change that we might make to the signal handler interface. So for > > > > > example, if we introduced a new syscall as you propose below, and the > > > > > library hasn't been updated to recognize the new syscall, it will > > > > > silently miss changes in signal handler state caused by the new > > > > > syscall. > > > > > > > > > > At the end of this argument lies "we can never change anything about > > > > > how signal handlers work because it could break some interposing > > > > > library somewhere" -- replace "signal handlers" with any kernel > > > > > feature whose behavior may be modified by an interposing library if > > > > > you like -- and I don't think we want to go that far. As far as I > > > > > know, this isn't really the kernel's business anyway -- the kernel's > > > > > stable ABI contract starts and ends with the syscall interface and not > > > > > some library on top. > > > > > > > > > > That being said, we should perhaps try to define our interface so that > > > > > something reasonable will probably happen if there is such a library > > > > > and it hasn't been updated. With the new syscall, the library will > > > > > sometimes silently fail to work in some non-local fashion. With the > > > > > flag bit in the signal number, the library will either cause the > > > > > caller to fall back to the old kernel code path (if there is a bounds > > > > > check) or likely crash loudly (if there is no bounds check). To me, > > > > > the "flag bit in the signal number" behavior seems more reasonable, > > > > > since either something correct or something easy to debug will > > > > > probably happen at runtime. > > > > > > > > > > > > could only appear in newly-written code that doesn't follow our > > > > > > > recommendations, and there are already plenty of much more likely ways > > > > > > > to cause buffer overflows in C code that doesn't follow > > > > > > > recommendations anyway. (And even if such a buffer overflow occurred, > > > > > > > it would very likely be caught early in development by the MMU due to > > > > > > > the magnitude of the number 1<<30.) > > > > > > > > > > > > Choosing the bit value is hard. If shitfing it overflows, this can > > > > > > trigger random undefined behaviour in the compiler in addition to (or > > > > > > perhaps instead of) an out-of-bounds access or segfault. > > > > > > > > > > It wouldn't overflow on a 64-bit architecture assuming normal array > > > > > indexing (the index would be promoted to pointer width before being > > > > > scaled to the array element size), and to begin with the users of this > > > > > would be 64-bit. > > > > > > > > Unless we don't offer this feature for 32-bit at all (possible, if ugly) > > > > we can't stop people using it. > > > > My point is that the problem in the interceptor library would probably > > be noticed on 64-bit (since that's what most people use these days), > > which would probably result in it being fixed by the time it reaches > > 32-bit users. > > Agreed. But we shouldn't take such bets unless we really have to. > > > > > > > If shifting it doesn't overflow, we might still fall into a valid > > > > > > mapping, though I'd agree a segfault is more likely. > > > > > > > > > > > > > > > > > > > > > I think it would be cleaner for to add a single flag field that can be > > > > > > > > used for detecting other extensions, and request it via a new sa_flags > > > > > > > > bit. This removes the need for sematically useless zeroing of unused > > > > > > > > fields (though for hygiene and backwards compatibility reasons we would > > > > > > > > probably want to carry on zeroing them anyway). > > > > > > > > > > > > > > > > I can see no simpler way to add supplementary siginfo fields for > > > > > > > > existing si_codes. For si_codes that didn't exist before the zeroing > > > > > > > > came in we could still detect optional si_code-specific fields via > > > > > > > > zeroing, but it seems messary to have two ways of detecting extensions. > > > > > > > > > > > > > > That would certainly be cleaner if it worked, but that would only be > > > > > > > the case if old kernels rejected unknown bits in sa_flags, and > > > > > > > unfortunately they don't. With the bit in the signal number, the "old > > > > > > > > > > > > Hmm, that is a problem I wasn't aware of. > > > > > > > > > > > > > kernels reject" behavior admits relatively straightforward usage code: > > > > > > > > > > > > > > void set_segv_handler(void) { > > > > > > > struct sigaction sa; > > > > > > > sa.sa_sigaction = handle_segv; > > > > > > > sa.sa_flags = SA_SIGINFO; > > > > > > > if (sigaction(SIGSEGV | SF_CLEAR_UNKNOWN_FIELDS, &sa, 0) < 0) { // > > > > > > > succeeds in new kernels, fails in old kernels > > > > > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > > > > > if (sigaction(SIGSEGV, &sa, 0) < 0) // succeeds in old kernels > > > > > > > perror("sigaction"); > > > > > > > } > > > > > > > } > > > > > > > > > > > > > > void clear_fields_and_handle_segv(int signum, siginfo_t *sa, void *ctx) { > > > > > > > sa->si_future_field = 0; > > > > > > > handle_segv(signum, sa, ctx); > > > > > > > } > > > > > > > > > > > > > > void handle_segv(int signum, siginfo_t *sa, void *ctx) { > > > > > > > // At this point, si_future_field will have the value 0 in old > > > > > > > kernels and the kernel-supplied value in new kernels. > > > > > > > } > > > > > > > > > > > > > > Imagine if we moved the flag SF_CLEAR_UNKNOWN_FIELDS from the signal > > > > > > > number to sa_flags. In that case, the first sigaction would succeed in > > > > > > > old kernels so handle_segv wouldn't know whether it can safely read > > > > > > > from si_future_field. With the sa_flags approach, you would need > > > > > > > kernel version number checking via uname before setting the flag in > > > > > > > sa_flags, and at that point why even have the flag in sa_flags at all > > > > > > > since you could just have the signal handler conditionally read from > > > > > > > si_future_field based on the uname? > > > > > > > > > > > > Software setting SA_SIFLAGS (or whatever) is new by definition, since > > > > > > it would be using a new #define. So it might be reasonable to put the > > > > > > burden on that software to verify that the flag was really accepted by > > > > > > the kernel, by reading it back. > > > > > > > > > > That doesn't seem like a good idea even if it worked, because it could > > > > > lead to race conditions. If the si_flags-reading signal handler were > > > > > invoked in response to a signal between when you set it and when you > > > > > ended up replacing it with the fallback signal handler for old > > > > > kernels, the handler may end up reading garbage data from si_flags. > > > > > > > > Not really. My example may have this problem, but the signal handler > > > > can be written to support both scenarios, based on testing a flag that > > > > the main program sets after verifying that the flag could be set. Or > > > > the signal could be blocked around establishment (often a good idea for > > > > other reasons). > > > > > > > > But I agree it's a bit gross, and anyway doesn't work due to the fact > > > > that the kernel doesn't filter out unrecognised flags anyway. > > > > > > > > > > Unfortunately, even relatively recent kernels blindly store sa_flags > > > > > > in the kernel without validating it, and so it looks like duff flags > > > > > > can be read back out via a sigaction() call. Dang. > > > > > > > > > > > > > > > > > > Perhaps a new frontend syscall could be added. A new libc that knows > > > > > > about this "sigaction2" could use it and mask off problem bits from > > > > > > sa_flags in its sigaction() wrapper before calling sigaction2. An old > > > > > > libc would call the old sigaction syscall, where we would ignore these > > > > > > new sa_flags bits as before. > > > > > > > > > > I'm not currently in favor of the new syscall but if we do this I > > > > > would keep sigaction and sigaction2 separate. That is, libc sigaction > > > > > should always use the sigaction syscall, and libc sigaction2 should > > > > > always use the sigaction2 syscall. We should avoid libc's sigaction > > > > > having different behavior based on the libc version and kernel > > > > > version, as that would make it harder to reason about its behavior. > > > > > Calling code would need to check for presence of sigaction2 in both > > > > > libc and the kernel, e.g. > > > > > > > > > > __attribute__((weak)) decltype(sigaction2) sigaction2; > > > > > > > > > > void set_segv_handler(void) { > > > > > struct sigaction sa; > > > > > sa.sa_sigaction = handle_segv; > > > > > sa.sa_flags = SA_SIGINFO | SA_SIFLAGS; > > > > > if (!sigaction2 || sigaction2(SIGSEGV, &sa, 0) < 0) { > > > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > > > sa.sa_flags = SA_SIGINFO; > > > > > if (sigaction(SIGSEGV, &sa, 0) < 0) > > > > > perror("sigaction"); > > > > > } > > > > > } > > > > > > > > I guess. But I share your distaste for adding a new syscall. > > > > > > > > > > > > > > > This may not be a popular approach though, and software wouldn't be able > > > > > > to use our new features until libc is updated to match. > > > > > > > > > > > > If we go down this route, it may provide additional opportunities to fix > > > > > > annoying defects in the old interface. > > > > > > > > > > > > > > > > > > > Note that the same applies to a flag indicating the availability of a > > > > > > > si_flags field in sigaction (just > > > > > > > s/SF_CLEAR_UNKNOWN_FIELDS/SF_HAS_SI_FLAGS/ and > > > > > > > s/si_future_field/si_flags/ in the usage code above). In terms of > > > > > > > SF_CLEAR_UNKNOWN_FIELDS versus SF_HAS_SI_FLAGS I'd be fine either way. > > > > > > > > > > > > > > Another thought that occurred to me is that we may consider > > > > > > > generalizing this a step further and introducing a single flag bit in > > > > > > > the signal number that means "reject unknown flags in sa_flags". This > > > > > > > would mean that we wouldn't need to add any more flag bits to the > > > > > > > signal number in the future, thus limiting this signal number hack to > > > > > > > a single bit; all future mandatory behavior changes could just be put > > > > > > > behind a flag in sa_flags and userspace code would easily be able to > > > > > > > detect missing support for a flag and fall back if necessary. In our > > > > > > > case, this would imply usage code like this: > > > > > > > > > > > > > > void set_segv_handler(void) { > > > > > > > struct sigaction sa; > > > > > > > sa.sa_sigaction = handle_segv; > > > > > > > sa.sa_flags = SA_SIGINFO | SA_CLEAR_UNKNOWN_FIELDS; > > > > > > > // Succeeds in kernels with SA_CLEAR_UNKNOWN_FIELDS support. > > > > > > > // Fails in kernels with SF_CHECK_SA_FLAGS support but no > > > > > > > SA_CLEAR_UNKNOWN_FIELDS support (because of the unknown flags check). > > > > > > > // Fails in kernels without SF_CHECK_SA_FLAGS support (because of > > > > > > > the bounds check on the signal number). > > > > > > > if (sigaction(SIGSEGV | SF_CHECK_SA_FLAGS, &sa, 0) < 0) { > > > > > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > > > > > sa.sa_flags = SA_SIGINFO; > > > > > > > // Succeeds in old kernels, no need to use SF_CHECK_SA_FLAGS since > > > > > > > we're using sa_flags from the beginning of time. > > > > > > > if (sigaction(SIGSEGV, &sa, 0) < 0) > > > > > > > perror("sigaction"); > > > > > > > } > > > > > > > } > > > > > > > > > > > > As with the other options this could work, but looks like it could > > > > > > break the ABI due to violating the original semantics for the signal > > > > > > number argument. Perhaps I'm being too paranoid. > > > > > > > > > > There's no ABI being broken here, as long as we consider syscalls to > > > > > be the stable ABI layer. Old kernels are simply rejecting arguments > > > > > that they don't know about yet. By that argument, any introduction of > > > > > a new syscall is an ABI break because it changes the semantics of a > > > > > previously-unallocated syscall number. > > > > > > > > As argued above, I think this is an invalid argument. > > > > > > > > Although any addition will change behaviour (so is a break in some > > > > sense), the key is not to make "surprising" changes. > > > > If we care about interceptors then I don't think "surprising" comes > > into it. It's more a question of "does the anticipated behavior of the > > interceptor match our desired behavior", where "desired" means "most > > likely to avoid silent breakage". We would need to get into the head > > of a potential interceptor author and think about how they would have > > handled the signal number argument, as well as other arguments like > > sa_flags if we want to go that route, and see whether that behavior > > would lead to the desired result. > > That's exactly what I mean by "surprising". Not quite, see below. > However, not every > interceptor author will be making the same assumptions, and not every > bit of software affected will be an interceptor. I can see a couple of ways in which non-interceptor software could be affected: - It's doing something like "call sigaction on every possible signal number in the 31-bit range and end up failing if the syscall succeeded" (e.g. with an OOB write). Perhaps software could be doing something like this in a loop to collect all currently registered signal handlers. That being said, this program: #include <limits.h> #include <signal.h> int main() { struct sigaction act; for (int i = 1; i != INT_MAX; ++i) { sigaction(i, 0, &act); } } takes around 5 seconds to run on my relatively-fast machine, so I would expect any such code to be noticed as a performance issue and either be changed to be bounded on _NSIG or break on EINVAL. This is probably the largest potential flaw that I can currently see in the "bit in the signal number" idea, since it could conceivably result in userspace code being broken without having first required it to have been changed to make use of the new feature. I'm not convinced that it would be an ABI break though, because the code seems unlikely to exist in this form in the wild because of the performance issue, and you could anyway make the argument that the code is incorrect because, in order to contain a loop like this, it would need to be able to handle large, previously-unknown signal numbers somehow. If we accept that the code is incorrect, a similar line of argument applies as for interceptors (i.e. likely to result in an OOB access which will fail loudly and be easily debugged and fixed). - If we do something that involves introducing a new flag in sa_flags, the flag may be exposed to unaware software via the oldact argument to sigaction, and I suppose that it's conceivable that exposing a previously-unknown flag like this could somehow break something. But this seems like an unreasonable restriction because it would mean that we can never add a flag to sa_flags no matter what. > So some judgement > needs to be applied. Of course. We need to agree *how* to apply the judgement though. > > In this case, I think we exactly want the interceptor author to have > > thought "oh, it's just a number, I'll (possibly do a bounds check and > > then) use the number as an index into an array". This will lead to one > > of two outcomes: crashing (yes, yes, it won't always crash, but if the > > alternative is that it never crashes and we get silently incorrect > > behavior all of the time, I'll take sometimes crashing) or fail the > > bounds check and pretend to be an old kernel (the latter is > > anticipated by POSIX which requires returning -1/EINVAL for an invalid > > signal number). Each of these behaviors are desirable, as they are > > observable failures, which are more likely to result in fixes than > > silent ones. > > Agreed, except wanting the author to have thought something doesn't > ensure that they actually did think that. True, but if our goal is only to accommodate reasonably written interceptors, we don't actually need to ensure anything here. > > > > Having something random happen when setting a previously reserved flag > > > > bit, or when issuing a syscall when an unknown syscall number, or not > > > > surprising at all. > > > > Introducing a new syscall is right out in this model. The interceptor > > author wouldn't have anticipated our introducing a new syscall, so the > > new syscall wouldn't be intercepted and calls to the new syscall would > > silently bypass the interceptor. For example, adding sigaction2 could > > result in signal handlers being set without the interceptor's > > knowledge. > > Agreed. My sentence was a bit mangled: I mean to say "Having something > random happen when [...] issuing a syscall *with* an unknown syscall > number *is* not surprising at all." > > I agree that adding a new syscall is problematic if we want to avoid > breaking existing interceptors in particular. Other types of code are > much less likely to be affected by the addition of new syscalls. Right, and this to me is a case in point for why I would say that "surprising" isn't the right frame of analysis here. My analysis seems to generally be that "anticipated interceptor behavior matches desired behavior" is positively correlated with "surprising" (i.e. the interceptor viewpoint is the dual of the user viewpoint), so if we care about interceptors we may end up making a "surprising" change even though it doesn't intuitively seem like the right thing to do. > > Regarding a sa_flags bit, let's get inside the head of the interceptor > > author again. How would they handle a flag bit that they don't > > recognize when replacing the signal handler? It wouldn't be correct to > > just pass it through to the kernel, or drop the flag on the floor, as > > it might be semantically meaningful (and thus could change the calling > > convention as SA_SIGINFO does, or change the meaning of fields in > > siginfo, as SA_CODEX would do). A correctly written sigaction > > interceptor should probably abort the program upon encountering an > > unknown flag (thus giving a human a chance to update the interceptor), > > but chances are that they don't. Ignoring all but a few flags (and > > passing a fixed set of flags to the kernel) seems to be what > > libsigchain does, and in the case of SA_CODEX it would seem to result > > in desirable behavior (but I suspect that it isn't handling the other > > flags correctly), but I could also see an interceptor author just > > passing it unchanged to the kernel without checking it (perhaps > > because they didn't think about these issues, and because that didn't > > matter until now, with the exception of from-the-beginning-of-time > > flags like SA_SIGINFO). And with SA_CODEX that could lead to silent > > misreading of si_code in the interceptor's signal handler, if it > > hasn't been updated to use the new macros. > > Agreed. I've tried to implement things rather like this in the past, > and how to interpret the flags is a tricky issue. Some of the flags are > impossible to emulate even when you know what they mean, in particular > SA_NODEFER and SA_RESTART. > > Making new flags safe to ignore and harmless to set of you don't know > what they mean is the safest approach, but not always possible (I think > I managed this with by suggestion below, though). Again, this suggestion could lead to silent failures in an interceptor, if: - the interceptor passes the sa_flags through to the kernel unchanged (or otherwise doesn't touch SA_CODEX) - the interceptor replaces the user's sa_sigaction - the interceptor's replacement sa_sigaction tests the provided si_code. Maybe you're not concerned about that, though? At least to me it seems in the same ballpark of likelihood as the ways in which things could go wrong with the signal number bit. > Ideally, a flags field should be specified with rules that say exactly > what to do with flags you don't recognise. Sadly this is usually not > thought about until it's too late. It perhaps isn't too late to introduce such rules for sigaction if we adopt the signal number bit and we make it mean "reject unknown flags". Peter
On Mon, Jul 13, 2020 at 01:50:30PM -0700, Peter Collingbourne wrote: > On Mon, Jul 13, 2020 at 6:24 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > On Wed, Jul 08, 2020 at 03:21:13PM -0700, Peter Collingbourne wrote: > > > On Wed, Jul 8, 2020 at 6:58 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > On Wed, Jul 08, 2020 at 12:00:22PM +0100, Dave Martin wrote: > > > > > On Tue, Jul 07, 2020 at 12:07:09PM -0700, Peter Collingbourne wrote: > > > > > > On Tue, Jul 7, 2020 at 7:19 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > On Mon, Jul 06, 2020 at 12:20:33PM -0700, Peter Collingbourne wrote: > > > > > > > > On Mon, Jul 6, 2020 at 9:41 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > > > > > On Wed, Jun 24, 2020 at 12:51:43PM -0700, Peter Collingbourne wrote: > > > > > > > > > > On Wed, Jun 24, 2020 at 10:12 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > > > > > > > > > On Wed, Jun 24, 2020 at 09:51:49AM -0700, Peter Collingbourne wrote: > > > > > > > > > > > > On Wed, Jun 24, 2020 at 2:28 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 05:40:08PM -0700, Peter Collingbourne wrote: > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 10:52 AM Eric W. Biederman > > > > > > > > > > > > > > <ebiederm@xmission.com> wrote: > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Dave Martin <Dave.Martin@arm.com> writes: > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > > > > > > > > > > > > > > > >> Peter Collingbourne <pcc@google.com> writes: > > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > > > > > > > > > > > > > > > >> > index 47f651df781c..a8380a2b6361 100644 > > > > > > > > > > > > > > > >> > --- a/arch/arm64/kernel/traps.c > > > > > > > > > > > > > > > >> > +++ b/arch/arm64/kernel/traps.c > > > > > > > > > > > > > > > >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > > > > > > > > > > > > > > > >> > } > > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, > > > > > > > > > > > > > > > >> > + unsigned long far, unsigned char far_tb_mask, > > > > > > > > > > > > > > > >> > const char *str) > > > > > > > > > > > > > > > >> > { > > > > > > > > > > > > > > > >> > arm64_show_signal(signo, str); > > > > > > > > > > > > > > > >> > - if (signo == SIGKILL) > > > > > > > > > > > > > > > >> > + if (signo == SIGKILL) { > > > > > > > > > > > > > > > >> > force_sig(SIGKILL); > > > > > > > > > > > > > > > >> > - else > > > > > > > > > > > > > > > >> > - force_sig_fault(signo, code, addr); > > > > > > > > > > > > > > > >> > + } else { > > > > > > > > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > > > > > > > > >> > + info.si_signo = signo; > > > > > > > > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > > > > > > > > >> > + info.si_code = code; > > > > > > > > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > > > > > > > > >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > > > > > > > > > > > > > > > >> > + info.si_addr_top_byte_mask = far_tb_mask; > > > > > > > > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > > > > > > > > >> > + } > > > > > > > > > > > > > > > >> > } > > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > > > > > > > > > > > > > > > >> > - const char *str) > > > > > > > > > > > > > > > >> > + unsigned long far, const char *str) > > > > > > > > > > > > > > > >> > { > > > > > > > > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > > > > > > > > >> > + > > > > > > > > > > > > > > > >> > arm64_show_signal(SIGBUS, str); > > > > > > > > > > > > > > > >> > - force_sig_mceerr(code, addr, lsb); > > > > > > > > > > > > > > > >> > + > > > > > > > > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > > > > > > > > >> > + info.si_signo = SIGBUS; > > > > > > > > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > > > > > > > > >> > + info.si_code = code; > > > > > > > > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > > > > > > > > >> > + info.si_addr_lsb = lsb; > > > > > > > > > > > > > > > >> > + info.si_addr_top_byte = far >> 56; > > > > > > > > > > > > > > > >> > + info.si_addr_top_byte_mask = 0xff; > > > > > > > > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > > > > > > > > >> > } > > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > >> I have a real problem with this construction. force_sig_info is not an > > > > > > > > > > > > > > > >> interface that should be used for anything except to define a wrapper > > > > > > > > > > > > > > > >> that takes it's parameters. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Can you elaborate? How would you do this king of thing. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > There are no other uses of force_sig_info in architecture code. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > I just removed them _all_ because they were almost all broken. > > > > > > > > > > > > > > > In fact your mcerr case is broken because it uses two different > > > > > > > > > > > > > > > union members simultantiously. > > > > > > > > > > > > > > > > > > > > > > > > > > > > Is that really broken? I thought that the Linux kernel deliberately > > > > > > > > > > > > > > didn't care about strict aliasing rules (the top-level Makefile passes > > > > > > > > > > > > > > -fno-strict-aliasing) so I thought that it was valid in "Linux kernel > > > > > > > > > > > > > > C" even though from a standards point of view it is invalid. (That > > > > > > > > > > > > > > being said, this is probably moot with my proposed changes below > > > > > > > > > > > > > > though.) > > > > > > > > > > > > > > > > > > > > > > > > > > I have a feeling that -fno-strict-aliasing only allows you to _read_ a > > > > > > > > > > > > > different union member from the one previously written. > > > > > > > > > > > > > > > > > > > > > > > > > > Writing a different member from the last one written can still splatter > > > > > > > > > > > > > on the other members IIUC. > > > > > > > > > > > > > > > > > > > > > > > > > > It would be better to keep things separate rather than risk > > > > > > > > > > > > > incorrectness just to save a few bytes. > > > > > > > > > > > > > > > > > > > > > > > > > > IMHO -fno-strict-aliasing is no excuse for gratuitous type-punning. > > > > > > > > > > > > > > > > > > > > > > > > > > > > So I am looking for something like force_sig_mcerr or force_sig_fault > > > > > > > > > > > > > > > that includes your new information that then calls force_sig_info. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > I know of no other way to safely use the siginfo struct. > > > > > > > > > > > > > > > > > > > > > > > > > > > > So you want something like: > > > > > > > > > > > > > > > > > > > > > > > > > > > > int force_sig_fault_with_ignored_bits(int signo, int code, void __user > > > > > > > > > > > > > > *addr, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > > > > > int force_sig_mceerr_with_ignored_bits(int code, void __user *addr, > > > > > > > > > > > > > > short lsb, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > > > > > > > > > > > > > > > > > > > in kernel/signal.c and the code in arch/arm64 would call that? > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > AIUI we absolutely need a forced signal here, we need to supply > > > > > > > > > > > > > > > > metadata, and we don't have to open-code all that at every relevant > > > > > > > > > > > > > > > > signal generation site... > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >> It is not clear to me that if you have adapted siginfo_layout. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Garbled sentence? > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Looks like. One of the pieces of code that needs to change > > > > > > > > > > > > > > > when siginfo gets updated is siginfo_layout so that the structure > > > > > > > > > > > > > > > can be properly decoded and made sense of. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > I am not seeing anything like that. > > > > > > > > > > > > > > > > > > > > > > > > > > > > Okay, this has to do with copying between the compat and non-compat > > > > > > > > > > > > > > versions of the struct? Sure, I can update that, although the code > > > > > > > > > > > > > > would be basically non-functional on arm64 because TBI isn't supported > > > > > > > > > > > > > > on 32-bit ARM. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > > > > > >> > index cb3d6c267181..6dd82373eb2d 100644 > > > > > > > > > > > > > > > >> > --- a/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > > > > > >> > +++ b/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > > > > > >> > @@ -91,6 +91,14 @@ union __sifields { > > > > > > > > > > > > > > > >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > > > > > > > > > > > > >> > __u32 _pkey; > > > > > > > > > > > > > > > >> > } _addr_pkey; > > > > > > > > > > > > > > > >> > +#ifdef __aarch64__ > > > > > > > > > > > > > > > >> > + /* used with all si_codes */ > > > > > > > > > > > > > > > >> > + struct { > > > > > > > > > > > > > > > >> > + short _dummy_top_byte; > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > ^ What's this for? I don't have Eric's insight here. > > > > > > > > > > > > > > > > > > > > > > > > > > > > We would need a short's worth of padding in order to prevent the > > > > > > > > > > > > > > fields from occupying the same address as si_addr_lsb. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >> > + unsigned char _top_byte; > > > > > > > > > > > > > > > >> > + unsigned char _top_byte_mask; > > > > > > > > > > > > > > > >> > + } _addr_top_byte; > > > > > > > > > > > > > > > >> > +#endif > > > > > > > > > > > > > > > >> > }; > > > > > > > > > > > > > > > >> > } _sigfault; > > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > >> Please remove the "#ifdef __aarch64__". If at all possible we want to > > > > > > > > > > > > > > > >> design this so any other architecture who has this challenge can use the > > > > > > > > > > > > > > > >> code. The kind of code does not get enough attention/maintenance if it > > > > > > > > > > > > > > > >> is built for a single architecture. > > > > > > > > > > > > > > > > > > > > > > > > > > > > Seems reasonable. I was recently made aware that RISC-V was > > > > > > > > > > > > > > considering a similar feature: > > > > > > > > > > > > > > https://lists.riscv.org/g/tech-tee/topic/risc_v_tbi_proposal/72855478 > > > > > > > > > > > > > > I would have opted to expand this to other architectures on an > > > > > > > > > > > > > > as-needed basis, but I'd also be fine with having it on all > > > > > > > > > > > > > > architectures from the start. > > > > > > > > > > > > > > > > > > > > > > > > > > > > If we make this arch-independent, we have an additional concern, which > > > > > > > > > > > > > > is "what if some future architecture wants more than one byte here?" > > > > > > > > > > > > > > For example, an architecture may have a "top-two-bytes-ignore" > > > > > > > > > > > > > > feature, which would imply two-byte (misnamed) "si_addr_top_byte" and > > > > > > > > > > > > > > "si_addr_top_byte_mask" fields. And the RISC-V proposal potentially > > > > > > > > > > > > > > implies many more ignored bits (see slide 13 of the presentation). The > > > > > > > > > > > > > > maximum size that these fields can possibly be is the size of a > > > > > > > > > > > > > > pointer, and with that there wouldn't be enough room in the padding at > > > > > > > > > > > > > > this point to accommodate the new fields. > > > > > > > > > > > > > > > > > > > > > > > > > > > > That basically implies your earlier suggestion of adding a union > > > > > > > > > > > > > > member here to accommodate future expansion of the union, and adding > > > > > > > > > > > > > > the new fields after the union. I'm happy to make that change, with > > > > > > > > > > > > > > the fields renamed "si_addr_ignored" and "si_addr_ignored_mask". > > > > > > > > > > > > > > > > > > > > > > > > > > I think what we need here is basically a flags word. > > > > > > > > > > > > > > > > > > > > > > > > > > So long as we keep a flag spare to indicate the existence of a further > > > > > > > > > > > > > flags word, we can extend as needed. > > > > > > > > > > > > > > > > > > > > > > > > > > How the existence of the first flags words is detected is another > > > > > > > > > > > > > problem. If it only applies for newly-defined si_code values, then > > > > > > > > > > > > > I guess si_code may be sufficient. > > > > > > > > > > > > > > > > > > > > > > > > Existing kernels will zero-initialize unused regions of the siginfo > > > > > > > > > > > > data structure. The zero-initialization of the padding at the end of > > > > > > > > > > > > the struct is done by the clear_user call here: > > > > > > > > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/kernel/signal.c#L3193 > > > > > > > > > > > > > > > > > > > > > > > > and the zero-initialization of the padding between fields and unused > > > > > > > > > > > > union members is done by the clear_siginfo function which the kernel > > > > > > > > > > > > calls when initializing the data structure: > > > > > > > > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/include/linux/signal.h#L20 > > > > > > > > > > > > > > > > > > > > > > > > Therefore, a flag word value of 0 may be used to detect a lack of > > > > > > > > > > > > support for flagged fields. > > > > > > > > > > > > > > > > > > > > > > It's not enough that we do this today. We would have had to do it back > > > > > > > > > > > to the dawn of time (though in the arm64 case I guess we just need to go > > > > > > > > > > > back to when the arch/arm64 was merged). > > > > > > > > > > > > > > > > > > > > > > v2.6.12:kernel/signal.c:copy_siginfo_to_user() suggests that this wasn't > > > > > > > > > > > always the case, so unused parts of siginfo could be full of old junk > > > > > > > > > > > from the user stack, if the kernel is sufficiently old. > > > > > > > > > > > > > > > > > > > > > > If we're trying to do something generic that makes sense on all arches, > > > > > > > > > > > this matters. I may have misunderstood something about the code though. > > > > > > > > > > > > > > > > > > > > Hmm, I think you're right. The current behavior was introduced by > > > > > > > > > > commit c999b933faa5e281e3af2e110eccaf91698b0a81 which was first > > > > > > > > > > released in 4.18. So if an application wants to be compatible with > > > > > > > > > > pre-4.18 kernels then there would need to be some other way to > > > > > > > > > > indicate that the fields are valid. Probably the simplest way would be > > > > > > > > > > to have the application issue a uname(2) syscall and check the kernel > > > > > > > > > > version before reading these fields. I have a couple of other ideas > > > > > > > > > > that don't rely on version detection, if we'd prefer to avoid that. > > > > > > > > > > (They are somewhat ugly, but our hand is forced by backwards > > > > > > > > > > compatibility.) > > > > > > > > > > > > > > > > > > > > One idea is to re-purpose the si_errno field as a flags field for > > > > > > > > > > certain signal numbers. I checked a few kernel releases going back to > > > > > > > > > > 2.6.18 and it looks like the field is set to 0 except in the following > > > > > > > > > > circumstances: > > > > > > > > > > - sending a hardware breakpoint (SIGTRAP/TRAP_HWBKPT) > > > > > > > > > > - seccomp failures (SIGSYS/SYS_SECCOMP) > > > > > > > > > > - user-defined signal via kill_pid_usb_asyncio > > > > > > > > > > - SIGSWI in 3.18 and before (code since removed) > > > > > > > > > > > > > > > > > > > > It is also set to EFAULT for certain SIGSEGV/SEGV_MAPERR signals on > > > > > > > > > > powerpc since commit c96c4436aba4c12f1f48369f2f90bc43e12fe36c, which > > > > > > > > > > is currently unreleased. So if we wanted to go this route for SIGSEGV > > > > > > > > > > we would need to stop the kernel from setting si_errno to EFAULT for > > > > > > > > > > this signal before the 5.8 release. > > > > > > > > > > > > > > > > > > > > Another idea was to have userspace set a flag in sa_flags when > > > > > > > > > > registering a signal handler meaning "this signal handler requires > > > > > > > > > > unknown siginfo fields to be zeroed", and have existing kernels reject > > > > > > > > > > the syscall due to an unknown flag being set, but unfortunately this > > > > > > > > > > won't work because existing kernels do not reject sigaction syscalls > > > > > > > > > > with unknown flags set in sa_flags. A perhaps more radical idea in > > > > > > > > > > this vein would be to claim some of the upper bits of the signal > > > > > > > > > > number as flags that will cause the syscall to be rejected if set and > > > > > > > > > > unknown to the kernel. Existing kernels (going back to at least > > > > > > > > > > 2.6.18) contain this code in do_sigaction: > > > > > > > > > > > > > > > > > > > > if (!valid_signal(sig) || sig < 1 || (act && sig_kernel_only(sig))) > > > > > > > > > > return -EINVAL; > > > > > > > > > > > > > > > > > > > > and vald_signal is defined as: > > > > > > > > > > > > > > > > > > > > static inline int valid_signal(unsigned long sig) > > > > > > > > > > { > > > > > > > > > > return sig <= _NSIG ? 1 : 0; > > > > > > > > > > } > > > > > > > > > > > > > > > > > > > > All architectures define _NSIG as a value <= 128, so they will reject > > > > > > > > > > a signal number with any of bits 8-31 set. This means that we can use > > > > > > > > > > any of those bits for mandatory flags. Most likely we could use bit 30 > > > > > > > > > > (expanding down as necessary), as it keeps the signal number positive > > > > > > > > > > and permits future expansion of the signal number range. > > > > > > > > > > > > > > > > > > Does the signal core code actually gurantee to zero the unused fields? > > > > > > > > > Unless the fields are poked in by hand this is fraught with subtlelies, > > > > > > > > > especially when unions are involved. (I'm sure the code tries to do it, > > > > > > > > > but I've not eyeballed it in detail...) > > > > > > > > > > > > > > > > It memsets the siginfo structure before setting the fields and sending > > > > > > > > the signal (grep for clear_siginfo which is just a memset; you should > > > > > > > > find a call before all callers of force_sig_info). Memset is the right > > > > > > > > approach here since unlike setting fields by hand it clears padding > > > > > > > > which could lead to information leaks from the kernel. IIUC this is > > > > > > > > the reason why Eric wants all of the signals to be raised via wrappers > > > > > > > > in kernel/signal.c instead of via force_sig_info directly (to make > > > > > > > > this aspect easier to audit). > > > > > > > > > > > > > > My impression was that the reason for this model is partly to ensure > > > > > > > that siginfo fields are populated more consistently. When it was all > > > > > > > down to the individual callers, inconsistencies creeped in. > > > > > > > > > > > > > > With regard to memset(), this is not a complete defence against data > > > > > > > leakage. Assigning to a struct member can set any or all padding in > > > > > > > the struct to random garbage (consider write-combining of neighboring > > > > > > > member writes into a single larger accesses in asm for example). The > > > > > > > > > > > > I don't believe that LLVM will store to padding like this. I don't > > > > > > know about GCC, though, but I wouldn't be surprised if this is > > > > > > something that the kernel would want to turn off in "kernel C" (like > > > > > > it turns off strict aliasing) specifically because of the information > > > > > > leak issue. > > > > > > > > > > Again, the issue is not future kernel builds -- we can always find a way > > > > > to fix the behaviour for those -- but past kernel builds. > > > > > > I thought that the whole point of the "bit in the signal number" (or > > > SI_CODEX or whatever) was that we didn't need to worry about the > > > behavior of past kernel builds? > > > > It depends on what we use the new flag(s) for. > > > > If the flag means just that unused padding is safely zeroed, that could > > work -- but we'd want high confidence that it really is zeroed even in > > wacky configurations. > > > > > > > > > only way to avoid this is to ensure that the struct is 100% > > > > > > > padding-free, and that each member of a union is the same size. A > > > > > > > quick clance at <uapi/asm-generic/siginfo.h> confirms that this is not > > > > > > > the case. > > > > > > > > > > > > > > This might need to be looked at separately. > > > > > > > > > > > > > > But it does mean, strictly speaking, that we can't reliably add new > > > > > > > fields anywhere that there was previously padding: assigning to > > > > > > > neighboring members can still fill those with garbage after the > > > > > > > memset(). > > > > > > > > > > > > ...but this is largely moot because I'm not proposing to add new > > > > > > fields in the padding any more (because the fields needed to become > > > > > > larger in order to accommodate future hypothetical architectures which > > > > > > might want to use the fields, and thus they wouldn't fit in the > > > > > > padding). The siginfo.h diff would be something like: > > > > > > > > > > > > diff --git a/include/uapi/asm-generic/siginfo.h > > > > > > b/include/uapi/asm-generic/siginfo.h > > > > > > index cb3d6c267181..4a2fe257415d 100644 > > > > > > --- a/include/uapi/asm-generic/siginfo.h > > > > > > +++ b/include/uapi/asm-generic/siginfo.h > > > > > > @@ -91,7 +91,10 @@ union __sifields { > > > > > > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > > > __u32 _pkey; > > > > > > } _addr_pkey; > > > > > > + void *_pad[6]; > > > > > > }; > > > > > > + uintptr_t _ignored_bits; > > > > > > + uintptr_t _ignored_bits_mask; > > > > > > > > > > This _is_ in padding: the tail-padding of the (previously smaller) > > > > > _sigfault. Again, the compiler was allowed to populate this area with > > > > > junk before these fields were added. > > > > > > > > > > I agree that it seems fairly unlikely that the compiler would have been > > > > > overwriting this in normal circumstances, but that's not a guarantee. > > > > > My worry is that if this goes wrong, it will go wrong silently and > > > > > unpredictably. > > > > > > > > > > > } _sigfault; > > > > > > > > > > > > /* SIGPOLL */ > > > > > > > > > > > > or with a "uintptr_t _flags" added in before _ignored_bits if we go with that. > > > > > > > > > > > > > > > Using unused bits in the signal number to turn on new functionality > > > > > > > > > feels risky. As currently specified, this is just a number. Since > > > > > > > > > today a successful sigaction(n ...) guarantees that n is a valid signal > > > > > > > > > number, reasonable code like the following would trigger a buffer > > > > > > > > > overrun if we start trying to encode anything else in there: > > > > > > > > > > > > > > > > > > struct sigaction actions[NSIG]; > > > > > > > > > > > > > > > > > > int do_something( ... ) > > > > > > > > > { > > > > > > > > > ... > > > > > > > > > > > > > > > > > > if (!sigaction(n, sa, ...)) { > > > > > > > > > actions[n] = *sa; > > > > > > > > > return 0; > > > > > > > > > } > > > > > > > > > > > > > > > > > > ... > > > > > > > > > } > > > > > > > > > > > > > > > > I imagine the bit in the signal number being set by the direct caller > > > > > > > > to sigaction, and we could specifically recommend that calling > > > > > > > > pattern. In that case, your "n" wouldn't have the bit set in it. It > > > > > > > > > > > > > > I can imagine this too, but that doesn't mean that software does it. > > > > > > > > > > > > > > If the above kind of thing exists in a framework or library somewhere, > > > > > > > we could get problems. Similarly, a pre-existing LD_PRELOAD framework > > > > > > > that provides a wrapper for sigaction may now go wrong even if your > > > > > > > pattern is followed -- i.e., the caller thinks it's calling sigaction > > > > > > > directly but in fact it isn't. > > > > > > > > > > > > I'm aware of one library like that. It's called libsigchain, and it > > > > > > has an early bounds check: > > > > > > https://cs.android.com/android/platform/superproject/+/master:art/sigchainlib/sigchain.cc;l=371 > > > > > > > > > > > > Until the library is changed to recognize the flag, calling code would > > > > > > see the return value of -1 as if the kernel failed the syscall, and > > > > > > would fall back to the code for old kernels. > > > > > > > > > > But only after some bad dereferences. If these were writes, this means > > > > > that memory _may_ be silently corrupted (I don't say it't likely in a > > > > > given case, and we cannot pick a flag bit that makes this impossible). > > > > > > You're talking about libsigchain, right? I don't see any bad > > > references, the function returns after noticing the bounds check > > > failure. > > > > Yes, I confused myself by reading Handler() out of context. The kernel > > will invoke this with signo to a real signal number (without any flags). > > > > The sigaction wrapper does the bounds check before doing anything else, > > just as you say -- so that looks fine. > > > > (Side question: is all this thread-safe? Is there some implicit locking > > somewhere?) > > I think maybe it isn't? There seem to be possible races on the > handler_ field. One possibility is that the function could race with > itself on another thread, which could be fixed via locking, but it > would also need to handle races between itself and the signal handler, > most likely by blocking the signal while setting it. Hmmm, tricky... anyway, that's not my problem ;) > > > > > So, _even though the user program is correct_, our change may trigger > > > > > > Let's say that you were talking about some other library and not > > > libsigchain. Such an interceptor wouldn't be correct though, it failed > > > to account for our change to the syscall semantics. If the accesses > > > were before the syscall (or the bounds check), then the interceptor > > > would not have been correct in the first place because POSIX requires > > > returning -1 with errno=EINVAL (and not crashing) if the signal number > > > is invalid. > > > > > > > > the corruption of arbitrary user memory. This what I mean by an ABI > > > > > break. The fact that the corruption is not done by the syscall itself > > > > > is no excuse. > > > > > > At some point, though, accommodating interceptors becomes pretty much > > > tantamount to saying "we can never change anything". Even just adding > > > a field to __sifields (which is pretty much required for what we need > > > to do) could break things in the presence of some interceptors because > > > the interceptor could be copying the fields manually to a new data > > > structure before calling the user's signal handler (e.g. because it > > > wants to defer the signal until later) and miss our new field. I think > > > most of the other ideas we're discussing fail to meet this bar as well > > > and I'll go into more details later on. > > > > I agree we cannot always avoid breaking such things. But we should do > > our best to avoid it. > > I think that given the hand that we've been dealt, no matter what we > do, we can't really avoid risking breaking something. The relevant > questions are "what are we going to risk breaking", "how much risk is > there", "will it be easily noticed/fixable", and "once we're on the > other side of the potential breakage, will we find ourselves in a > position where changing things involves less breakage risk". > > > > > > We also fail to notice failures in sigaddset() etc., though in this code > > > > > it looks like that should not matter. > > > > > > Maybe you're looking at the handler ("SignalChain::Handler")? The bit > > > wouldn't be set in the signo argument to the handler. I'm talking > > > about line 371 of the code I linked, in the sigaction interceptor > > > "__sigaction" (it looks like sometimes the link doesn't take you to > > > the correct line for some reason). > > > > Ack, I confused myself. > > > > > > > > In general I think that any library like this with independent > > > > > > tracking of the kernel's purported signal handler state would need to > > > > > > be very sensitive to which syscalls are capable of setting signal > > > > > > handlers, what their semantics are, and so on. This applies to any > > > > > > change that we might make to the signal handler interface. So for > > > > > > example, if we introduced a new syscall as you propose below, and the > > > > > > library hasn't been updated to recognize the new syscall, it will > > > > > > silently miss changes in signal handler state caused by the new > > > > > > syscall. > > > > > > > > > > > > At the end of this argument lies "we can never change anything about > > > > > > how signal handlers work because it could break some interposing > > > > > > library somewhere" -- replace "signal handlers" with any kernel > > > > > > feature whose behavior may be modified by an interposing library if > > > > > > you like -- and I don't think we want to go that far. As far as I > > > > > > know, this isn't really the kernel's business anyway -- the kernel's > > > > > > stable ABI contract starts and ends with the syscall interface and not > > > > > > some library on top. > > > > > > > > > > > > That being said, we should perhaps try to define our interface so that > > > > > > something reasonable will probably happen if there is such a library > > > > > > and it hasn't been updated. With the new syscall, the library will > > > > > > sometimes silently fail to work in some non-local fashion. With the > > > > > > flag bit in the signal number, the library will either cause the > > > > > > caller to fall back to the old kernel code path (if there is a bounds > > > > > > check) or likely crash loudly (if there is no bounds check). To me, > > > > > > the "flag bit in the signal number" behavior seems more reasonable, > > > > > > since either something correct or something easy to debug will > > > > > > probably happen at runtime. > > > > > > > > > > > > > > could only appear in newly-written code that doesn't follow our > > > > > > > > recommendations, and there are already plenty of much more likely ways > > > > > > > > to cause buffer overflows in C code that doesn't follow > > > > > > > > recommendations anyway. (And even if such a buffer overflow occurred, > > > > > > > > it would very likely be caught early in development by the MMU due to > > > > > > > > the magnitude of the number 1<<30.) > > > > > > > > > > > > > > Choosing the bit value is hard. If shitfing it overflows, this can > > > > > > > trigger random undefined behaviour in the compiler in addition to (or > > > > > > > perhaps instead of) an out-of-bounds access or segfault. > > > > > > > > > > > > It wouldn't overflow on a 64-bit architecture assuming normal array > > > > > > indexing (the index would be promoted to pointer width before being > > > > > > scaled to the array element size), and to begin with the users of this > > > > > > would be 64-bit. > > > > > > > > > > Unless we don't offer this feature for 32-bit at all (possible, if ugly) > > > > > we can't stop people using it. > > > > > > My point is that the problem in the interceptor library would probably > > > be noticed on 64-bit (since that's what most people use these days), > > > which would probably result in it being fixed by the time it reaches > > > 32-bit users. > > > > Agreed. But we shouldn't take such bets unless we really have to. > > > > > > > > > If shifting it doesn't overflow, we might still fall into a valid > > > > > > > mapping, though I'd agree a segfault is more likely. > > > > > > > > > > > > > > > > > > > > > > > > I think it would be cleaner for to add a single flag field that can be > > > > > > > > > used for detecting other extensions, and request it via a new sa_flags > > > > > > > > > bit. This removes the need for sematically useless zeroing of unused > > > > > > > > > fields (though for hygiene and backwards compatibility reasons we would > > > > > > > > > probably want to carry on zeroing them anyway). > > > > > > > > > > > > > > > > > > I can see no simpler way to add supplementary siginfo fields for > > > > > > > > > existing si_codes. For si_codes that didn't exist before the zeroing > > > > > > > > > came in we could still detect optional si_code-specific fields via > > > > > > > > > zeroing, but it seems messary to have two ways of detecting extensions. > > > > > > > > > > > > > > > > That would certainly be cleaner if it worked, but that would only be > > > > > > > > the case if old kernels rejected unknown bits in sa_flags, and > > > > > > > > unfortunately they don't. With the bit in the signal number, the "old > > > > > > > > > > > > > > Hmm, that is a problem I wasn't aware of. > > > > > > > > > > > > > > > kernels reject" behavior admits relatively straightforward usage code: > > > > > > > > > > > > > > > > void set_segv_handler(void) { > > > > > > > > struct sigaction sa; > > > > > > > > sa.sa_sigaction = handle_segv; > > > > > > > > sa.sa_flags = SA_SIGINFO; > > > > > > > > if (sigaction(SIGSEGV | SF_CLEAR_UNKNOWN_FIELDS, &sa, 0) < 0) { // > > > > > > > > succeeds in new kernels, fails in old kernels > > > > > > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > > > > > > if (sigaction(SIGSEGV, &sa, 0) < 0) // succeeds in old kernels > > > > > > > > perror("sigaction"); > > > > > > > > } > > > > > > > > } > > > > > > > > > > > > > > > > void clear_fields_and_handle_segv(int signum, siginfo_t *sa, void *ctx) { > > > > > > > > sa->si_future_field = 0; > > > > > > > > handle_segv(signum, sa, ctx); > > > > > > > > } > > > > > > > > > > > > > > > > void handle_segv(int signum, siginfo_t *sa, void *ctx) { > > > > > > > > // At this point, si_future_field will have the value 0 in old > > > > > > > > kernels and the kernel-supplied value in new kernels. > > > > > > > > } > > > > > > > > > > > > > > > > Imagine if we moved the flag SF_CLEAR_UNKNOWN_FIELDS from the signal > > > > > > > > number to sa_flags. In that case, the first sigaction would succeed in > > > > > > > > old kernels so handle_segv wouldn't know whether it can safely read > > > > > > > > from si_future_field. With the sa_flags approach, you would need > > > > > > > > kernel version number checking via uname before setting the flag in > > > > > > > > sa_flags, and at that point why even have the flag in sa_flags at all > > > > > > > > since you could just have the signal handler conditionally read from > > > > > > > > si_future_field based on the uname? > > > > > > > > > > > > > > Software setting SA_SIFLAGS (or whatever) is new by definition, since > > > > > > > it would be using a new #define. So it might be reasonable to put the > > > > > > > burden on that software to verify that the flag was really accepted by > > > > > > > the kernel, by reading it back. > > > > > > > > > > > > That doesn't seem like a good idea even if it worked, because it could > > > > > > lead to race conditions. If the si_flags-reading signal handler were > > > > > > invoked in response to a signal between when you set it and when you > > > > > > ended up replacing it with the fallback signal handler for old > > > > > > kernels, the handler may end up reading garbage data from si_flags. > > > > > > > > > > Not really. My example may have this problem, but the signal handler > > > > > can be written to support both scenarios, based on testing a flag that > > > > > the main program sets after verifying that the flag could be set. Or > > > > > the signal could be blocked around establishment (often a good idea for > > > > > other reasons). > > > > > > > > > > But I agree it's a bit gross, and anyway doesn't work due to the fact > > > > > that the kernel doesn't filter out unrecognised flags anyway. > > > > > > > > > > > > Unfortunately, even relatively recent kernels blindly store sa_flags > > > > > > > in the kernel without validating it, and so it looks like duff flags > > > > > > > can be read back out via a sigaction() call. Dang. > > > > > > > > > > > > > > > > > > > > > Perhaps a new frontend syscall could be added. A new libc that knows > > > > > > > about this "sigaction2" could use it and mask off problem bits from > > > > > > > sa_flags in its sigaction() wrapper before calling sigaction2. An old > > > > > > > libc would call the old sigaction syscall, where we would ignore these > > > > > > > new sa_flags bits as before. > > > > > > > > > > > > I'm not currently in favor of the new syscall but if we do this I > > > > > > would keep sigaction and sigaction2 separate. That is, libc sigaction > > > > > > should always use the sigaction syscall, and libc sigaction2 should > > > > > > always use the sigaction2 syscall. We should avoid libc's sigaction > > > > > > having different behavior based on the libc version and kernel > > > > > > version, as that would make it harder to reason about its behavior. > > > > > > Calling code would need to check for presence of sigaction2 in both > > > > > > libc and the kernel, e.g. > > > > > > > > > > > > __attribute__((weak)) decltype(sigaction2) sigaction2; > > > > > > > > > > > > void set_segv_handler(void) { > > > > > > struct sigaction sa; > > > > > > sa.sa_sigaction = handle_segv; > > > > > > sa.sa_flags = SA_SIGINFO | SA_SIFLAGS; > > > > > > if (!sigaction2 || sigaction2(SIGSEGV, &sa, 0) < 0) { > > > > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > > > > sa.sa_flags = SA_SIGINFO; > > > > > > if (sigaction(SIGSEGV, &sa, 0) < 0) > > > > > > perror("sigaction"); > > > > > > } > > > > > > } > > > > > > > > > > I guess. But I share your distaste for adding a new syscall. > > > > > > > > > > > > > > > > > > This may not be a popular approach though, and software wouldn't be able > > > > > > > to use our new features until libc is updated to match. > > > > > > > > > > > > > > If we go down this route, it may provide additional opportunities to fix > > > > > > > annoying defects in the old interface. > > > > > > > > > > > > > > > > > > > > > > Note that the same applies to a flag indicating the availability of a > > > > > > > > si_flags field in sigaction (just > > > > > > > > s/SF_CLEAR_UNKNOWN_FIELDS/SF_HAS_SI_FLAGS/ and > > > > > > > > s/si_future_field/si_flags/ in the usage code above). In terms of > > > > > > > > SF_CLEAR_UNKNOWN_FIELDS versus SF_HAS_SI_FLAGS I'd be fine either way. > > > > > > > > > > > > > > > > Another thought that occurred to me is that we may consider > > > > > > > > generalizing this a step further and introducing a single flag bit in > > > > > > > > the signal number that means "reject unknown flags in sa_flags". This > > > > > > > > would mean that we wouldn't need to add any more flag bits to the > > > > > > > > signal number in the future, thus limiting this signal number hack to > > > > > > > > a single bit; all future mandatory behavior changes could just be put > > > > > > > > behind a flag in sa_flags and userspace code would easily be able to > > > > > > > > detect missing support for a flag and fall back if necessary. In our > > > > > > > > case, this would imply usage code like this: > > > > > > > > > > > > > > > > void set_segv_handler(void) { > > > > > > > > struct sigaction sa; > > > > > > > > sa.sa_sigaction = handle_segv; > > > > > > > > sa.sa_flags = SA_SIGINFO | SA_CLEAR_UNKNOWN_FIELDS; > > > > > > > > // Succeeds in kernels with SA_CLEAR_UNKNOWN_FIELDS support. > > > > > > > > // Fails in kernels with SF_CHECK_SA_FLAGS support but no > > > > > > > > SA_CLEAR_UNKNOWN_FIELDS support (because of the unknown flags check). > > > > > > > > // Fails in kernels without SF_CHECK_SA_FLAGS support (because of > > > > > > > > the bounds check on the signal number). > > > > > > > > if (sigaction(SIGSEGV | SF_CHECK_SA_FLAGS, &sa, 0) < 0) { > > > > > > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > > > > > > sa.sa_flags = SA_SIGINFO; > > > > > > > > // Succeeds in old kernels, no need to use SF_CHECK_SA_FLAGS since > > > > > > > > we're using sa_flags from the beginning of time. > > > > > > > > if (sigaction(SIGSEGV, &sa, 0) < 0) > > > > > > > > perror("sigaction"); > > > > > > > > } > > > > > > > > } > > > > > > > > > > > > > > As with the other options this could work, but looks like it could > > > > > > > break the ABI due to violating the original semantics for the signal > > > > > > > number argument. Perhaps I'm being too paranoid. > > > > > > > > > > > > There's no ABI being broken here, as long as we consider syscalls to > > > > > > be the stable ABI layer. Old kernels are simply rejecting arguments > > > > > > that they don't know about yet. By that argument, any introduction of > > > > > > a new syscall is an ABI break because it changes the semantics of a > > > > > > previously-unallocated syscall number. > > > > > > > > > > As argued above, I think this is an invalid argument. > > > > > > > > > > Although any addition will change behaviour (so is a break in some > > > > > sense), the key is not to make "surprising" changes. > > > > > > If we care about interceptors then I don't think "surprising" comes > > > into it. It's more a question of "does the anticipated behavior of the > > > interceptor match our desired behavior", where "desired" means "most > > > likely to avoid silent breakage". We would need to get into the head > > > of a potential interceptor author and think about how they would have > > > handled the signal number argument, as well as other arguments like > > > sa_flags if we want to go that route, and see whether that behavior > > > would lead to the desired result. > > > > That's exactly what I mean by "surprising". > > Not quite, see below. > > > However, not every > > interceptor author will be making the same assumptions, and not every > > bit of software affected will be an interceptor. > > I can see a couple of ways in which non-interceptor software could be affected: > > - It's doing something like "call sigaction on every possible signal > number in the 31-bit range and end up failing if the syscall > succeeded" (e.g. with an OOB write). Perhaps software could be doing > something like this in a loop to collect all currently registered > signal handlers. That being said, this program: > > #include <limits.h> > #include <signal.h> > > int main() { > struct sigaction act; > for (int i = 1; i != INT_MAX; ++i) { > sigaction(i, 0, &act); > } > } > > takes around 5 seconds to run on my relatively-fast machine, so I > would expect any such code to be noticed as a performance issue and > either be changed to be bounded on _NSIG or break on EINVAL. > > This is probably the largest potential flaw that I can currently see > in the "bit in the signal number" idea, since it could conceivably > result in userspace code being broken without having first required it > to have been changed to make use of the new feature. I'm not convinced > that it would be an ABI break though, because the code seems unlikely > to exist in this form in the wild because of the performance issue, > and you could anyway make the argument that the code is incorrect > because, in order to contain a loop like this, it would need to be > able to handle large, previously-unknown signal numbers somehow. If we > accept that the code is incorrect, a similar line of argument applies > as for interceptors (i.e. likely to result in an OOB access which will > fail loudly and be easily debugged and fixed). > > - If we do something that involves introducing a new flag in sa_flags, > the flag may be exposed to unaware software via the oldact argument to > sigaction, and I suppose that it's conceivable that exposing a > previously-unknown flag like this could somehow break something. But > this seems like an unreasonable restriction because it would mean that > we can never add a flag to sa_flags no matter what. > > > So some judgement > > needs to be applied. > > Of course. We need to agree *how* to apply the judgement though. > > > > In this case, I think we exactly want the interceptor author to have > > > thought "oh, it's just a number, I'll (possibly do a bounds check and > > > then) use the number as an index into an array". This will lead to one > > > of two outcomes: crashing (yes, yes, it won't always crash, but if the > > > alternative is that it never crashes and we get silently incorrect > > > behavior all of the time, I'll take sometimes crashing) or fail the > > > bounds check and pretend to be an old kernel (the latter is > > > anticipated by POSIX which requires returning -1/EINVAL for an invalid > > > signal number). Each of these behaviors are desirable, as they are > > > observable failures, which are more likely to result in fixes than > > > silent ones. > > > > Agreed, except wanting the author to have thought something doesn't > > ensure that they actually did think that. > > True, but if our goal is only to accommodate reasonably written > interceptors, we don't actually need to ensure anything here. > > > > > > Having something random happen when setting a previously reserved flag > > > > > bit, or when issuing a syscall when an unknown syscall number, or not > > > > > surprising at all. > > > > > > Introducing a new syscall is right out in this model. The interceptor > > > author wouldn't have anticipated our introducing a new syscall, so the > > > new syscall wouldn't be intercepted and calls to the new syscall would > > > silently bypass the interceptor. For example, adding sigaction2 could > > > result in signal handlers being set without the interceptor's > > > knowledge. > > > > Agreed. My sentence was a bit mangled: I mean to say "Having something > > random happen when [...] issuing a syscall *with* an unknown syscall > > number *is* not surprising at all." > > > > I agree that adding a new syscall is problematic if we want to avoid > > breaking existing interceptors in particular. Other types of code are > > much less likely to be affected by the addition of new syscalls. > > Right, and this to me is a case in point for why I would say that > "surprising" isn't the right frame of analysis here. My analysis seems > to generally be that "anticipated interceptor behavior matches desired > behavior" is positively correlated with "surprising" (i.e. the > interceptor viewpoint is the dual of the user viewpoint), so if we > care about interceptors we may end up making a "surprising" change > even though it doesn't intuitively seem like the right thing to do. You're right that interceptors are different from normal callers. I'm not sure I follow your argument, but an alternative way of looking at it might be to say that an interceptor is both an implementation of an interface and a caller of the same interface. Since API specs are rarely complete enough to cover the corner cases that arise from this, full portability is hard to achieve on top of an evolving kernel. However, I think this interceptor thing is a bit of a red herring. I just intended that as an illustration of the kind of code that might fall foul. This doesn't mean that it's 100% certain that no other software can be affected. The starting points for this discussion were: "is it reasonable for a caller to pass an unvalidated signal number to sigaction(), and rely on sigaction() to validate it?" and "is it reasonable to assume that a signal number accepted by sigaction() fits the POSIX specification of a valid signal number?" I think yes; you aren't (or weren't) convinced. The mere fact that it's hard to agree suggests to me that the specification is too weak to extend safely in this area. Unfortunately, it's rather weak for sa_flags too, although a non-full flags argument does at least suggest that future extensions might appear. > > > Regarding a sa_flags bit, let's get inside the head of the interceptor > > > author again. How would they handle a flag bit that they don't > > > recognize when replacing the signal handler? It wouldn't be correct to > > > just pass it through to the kernel, or drop the flag on the floor, as > > > it might be semantically meaningful (and thus could change the calling > > > convention as SA_SIGINFO does, or change the meaning of fields in > > > siginfo, as SA_CODEX would do). A correctly written sigaction > > > interceptor should probably abort the program upon encountering an > > > unknown flag (thus giving a human a chance to update the interceptor), > > > but chances are that they don't. Ignoring all but a few flags (and > > > passing a fixed set of flags to the kernel) seems to be what > > > libsigchain does, and in the case of SA_CODEX it would seem to result > > > in desirable behavior (but I suspect that it isn't handling the other > > > flags correctly), but I could also see an interceptor author just > > > passing it unchanged to the kernel without checking it (perhaps > > > because they didn't think about these issues, and because that didn't > > > matter until now, with the exception of from-the-beginning-of-time > > > flags like SA_SIGINFO). And with SA_CODEX that could lead to silent > > > misreading of si_code in the interceptor's signal handler, if it > > > hasn't been updated to use the new macros. > > > > Agreed. I've tried to implement things rather like this in the past, > > and how to interpret the flags is a tricky issue. Some of the flags are > > impossible to emulate even when you know what they mean, in particular > > SA_NODEFER and SA_RESTART. > > > > Making new flags safe to ignore and harmless to set of you don't know > > what they mean is the safest approach, but not always possible (I think > > I managed this with by suggestion below, though). > > Again, this suggestion could lead to silent failures in an interceptor, if: > - the interceptor passes the sa_flags through to the kernel unchanged > (or otherwise doesn't touch SA_CODEX) > - the interceptor replaces the user's sa_sigaction > - the interceptor's replacement sa_sigaction tests the provided si_code. > > Maybe you're not concerned about that, though? At least to me it seems > in the same ballpark of likelihood as the ways in which things could > go wrong with the signal number bit. I agree this is a concern, and perhaps a bit nastier in practice than side-effects of setting random bits in the signal number. Personally I do tend to be paranoid about flags arguments and try to police them in any code that isn't a trivial pass-through, but this doesn't mean that all code out there does it. (Including the kernel's sigaction()!) > > Ideally, a flags field should be specified with rules that say exactly > > what to do with flags you don't recognise. Sadly this is usually not > > thought about until it's too late. > > It perhaps isn't too late to introduce such rules for sigaction if we > adopt the signal number bit and we make it mean "reject unknown > flags". If Eric likes the idea then fair enough, but as I've tried to argue this may still just be moving the problem around rather than solving it. As a final random idea to add to the mix, we could add two or more flags in sa_flags, and require the kernel to transform them in a specific way, say: #define SA_WANT_FLAGS 0x00c700000 #define SA_HAVE_FLAGS 0x009200000 #define SA_FLAGS_MASK 0x00ff00000 volatile sig_atomic_t have_flags = 0; sa.sa_flags |= SA_WANT_FLAGS; if (sigaction(n, &sa, NULL)) if (!sigaction(n, NULL, &sa) && (sa.sa_flags & SA_FLAGS_MASK) == SA_HAVE_FLAGS) have_flags = 1; This is at least proof against "dumb readback". Provided that the handler can cope with the have_flags == 0 case and just reads the flag once per call, I don't think we would need to worry about races. Of course, an interceptor that doesn't understand this mechanism and munges or manufactures its own siginfo might still fail to properly initialise our new field before passing it on to a signal handler that is expecting it. But that's already broken: such an interceptor might also not understand new si_codes that the client code absolutely relies on. And new si_codes _do_ get added (that's another extensibility fail in the existing signal API). So... overall, maybe a bit in the signal number isn't a lot worse, and perhaps it _will_ lead to cleaner failures. Really, I don't see a way to solve it properly without a new API. In the meantime, can I suggest: (1) Come up with an extensible way of encoding supplementary information in siginfo. If the consensus is that zeroing unused fields is sufficient and that the kernel and compiler will reliably do it, then great. Otherwise, we might need explicit flags fields or something. (2) Hack up any simple mechanism (such as your signal number flag) for requesting/detecting the extra information. Along with an illustration of a application of the mechanism (i.e., reporting address tag bits), this should at least provide a basis for further review. We can then try to swap in a different mechanism for (2) if people have still have concerns (or it not, keep it). Cheers ---Dave
On Tue, Jul 14, 2020 at 10:36 AM Dave Martin <Dave.Martin@arm.com> wrote: > > On Mon, Jul 13, 2020 at 01:50:30PM -0700, Peter Collingbourne wrote: > > On Mon, Jul 13, 2020 at 6:24 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > On Wed, Jul 08, 2020 at 03:21:13PM -0700, Peter Collingbourne wrote: > > > > On Wed, Jul 8, 2020 at 6:58 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > On Wed, Jul 08, 2020 at 12:00:22PM +0100, Dave Martin wrote: > > > > > > On Tue, Jul 07, 2020 at 12:07:09PM -0700, Peter Collingbourne wrote: > > > > > > > On Tue, Jul 7, 2020 at 7:19 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > > > On Mon, Jul 06, 2020 at 12:20:33PM -0700, Peter Collingbourne wrote: > > > > > > > > > On Mon, Jul 6, 2020 at 9:41 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > > > > > > > On Wed, Jun 24, 2020 at 12:51:43PM -0700, Peter Collingbourne wrote: > > > > > > > > > > > On Wed, Jun 24, 2020 at 10:12 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > > > > > > > > > > > On Wed, Jun 24, 2020 at 09:51:49AM -0700, Peter Collingbourne wrote: > > > > > > > > > > > > > On Wed, Jun 24, 2020 at 2:28 AM Dave Martin <Dave.Martin@arm.com> wrote: > > > > > > > > > > > > > > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 05:40:08PM -0700, Peter Collingbourne wrote: > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 10:52 AM Eric W. Biederman > > > > > > > > > > > > > > > <ebiederm@xmission.com> wrote: > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Dave Martin <Dave.Martin@arm.com> writes: > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > On Tue, Jun 23, 2020 at 07:54:59AM -0500, Eric W. Biederman wrote: > > > > > > > > > > > > > > > > >> Peter Collingbourne <pcc@google.com> writes: > > > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > > >> > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c > > > > > > > > > > > > > > > > >> > index 47f651df781c..a8380a2b6361 100644 > > > > > > > > > > > > > > > > >> > --- a/arch/arm64/kernel/traps.c > > > > > > > > > > > > > > > > >> > +++ b/arch/arm64/kernel/traps.c > > > > > > > > > > > > > > > > >> > @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) > > > > > > > > > > > > > > > > >> > } > > > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > > > >> > void arm64_force_sig_fault(int signo, int code, void __user *addr, > > > > > > > > > > > > > > > > >> > + unsigned long far, unsigned char far_tb_mask, > > > > > > > > > > > > > > > > >> > const char *str) > > > > > > > > > > > > > > > > >> > { > > > > > > > > > > > > > > > > >> > arm64_show_signal(signo, str); > > > > > > > > > > > > > > > > >> > - if (signo == SIGKILL) > > > > > > > > > > > > > > > > >> > + if (signo == SIGKILL) { > > > > > > > > > > > > > > > > >> > force_sig(SIGKILL); > > > > > > > > > > > > > > > > >> > - else > > > > > > > > > > > > > > > > >> > - force_sig_fault(signo, code, addr); > > > > > > > > > > > > > > > > >> > + } else { > > > > > > > > > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > > > > > > > > > >> > + info.si_signo = signo; > > > > > > > > > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > > > > > > > > > >> > + info.si_code = code; > > > > > > > > > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > > > > > > > > > >> > + info.si_addr_top_byte = (far >> 56) & far_tb_mask; > > > > > > > > > > > > > > > > >> > + info.si_addr_top_byte_mask = far_tb_mask; > > > > > > > > > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > > > > > > > > > >> > + } > > > > > > > > > > > > > > > > >> > } > > > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > > > >> > void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, > > > > > > > > > > > > > > > > >> > - const char *str) > > > > > > > > > > > > > > > > >> > + unsigned long far, const char *str) > > > > > > > > > > > > > > > > >> > { > > > > > > > > > > > > > > > > >> > + struct kernel_siginfo info; > > > > > > > > > > > > > > > > >> > + > > > > > > > > > > > > > > > > >> > arm64_show_signal(SIGBUS, str); > > > > > > > > > > > > > > > > >> > - force_sig_mceerr(code, addr, lsb); > > > > > > > > > > > > > > > > >> > + > > > > > > > > > > > > > > > > >> > + clear_siginfo(&info); > > > > > > > > > > > > > > > > >> > + info.si_signo = SIGBUS; > > > > > > > > > > > > > > > > >> > + info.si_errno = 0; > > > > > > > > > > > > > > > > >> > + info.si_code = code; > > > > > > > > > > > > > > > > >> > + info.si_addr = addr; > > > > > > > > > > > > > > > > >> > + info.si_addr_lsb = lsb; > > > > > > > > > > > > > > > > >> > + info.si_addr_top_byte = far >> 56; > > > > > > > > > > > > > > > > >> > + info.si_addr_top_byte_mask = 0xff; > > > > > > > > > > > > > > > > >> > + force_sig_info(&info); > > > > > > > > > > > > > > > > >> > } > > > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > > >> I have a real problem with this construction. force_sig_info is not an > > > > > > > > > > > > > > > > >> interface that should be used for anything except to define a wrapper > > > > > > > > > > > > > > > > >> that takes it's parameters. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Can you elaborate? How would you do this king of thing. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > There are no other uses of force_sig_info in architecture code. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > I just removed them _all_ because they were almost all broken. > > > > > > > > > > > > > > > > In fact your mcerr case is broken because it uses two different > > > > > > > > > > > > > > > > union members simultantiously. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Is that really broken? I thought that the Linux kernel deliberately > > > > > > > > > > > > > > > didn't care about strict aliasing rules (the top-level Makefile passes > > > > > > > > > > > > > > > -fno-strict-aliasing) so I thought that it was valid in "Linux kernel > > > > > > > > > > > > > > > C" even though from a standards point of view it is invalid. (That > > > > > > > > > > > > > > > being said, this is probably moot with my proposed changes below > > > > > > > > > > > > > > > though.) > > > > > > > > > > > > > > > > > > > > > > > > > > > > I have a feeling that -fno-strict-aliasing only allows you to _read_ a > > > > > > > > > > > > > > different union member from the one previously written. > > > > > > > > > > > > > > > > > > > > > > > > > > > > Writing a different member from the last one written can still splatter > > > > > > > > > > > > > > on the other members IIUC. > > > > > > > > > > > > > > > > > > > > > > > > > > > > It would be better to keep things separate rather than risk > > > > > > > > > > > > > > incorrectness just to save a few bytes. > > > > > > > > > > > > > > > > > > > > > > > > > > > > IMHO -fno-strict-aliasing is no excuse for gratuitous type-punning. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > So I am looking for something like force_sig_mcerr or force_sig_fault > > > > > > > > > > > > > > > > that includes your new information that then calls force_sig_info. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > I know of no other way to safely use the siginfo struct. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > So you want something like: > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > int force_sig_fault_with_ignored_bits(int signo, int code, void __user > > > > > > > > > > > > > > > *addr, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > > > > > > int force_sig_mceerr_with_ignored_bits(int code, void __user *addr, > > > > > > > > > > > > > > > short lsb, uintptr_t addr_ignored, uintptr_t addr_ignored_mask); > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > in kernel/signal.c and the code in arch/arm64 would call that? > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > AIUI we absolutely need a forced signal here, we need to supply > > > > > > > > > > > > > > > > > metadata, and we don't have to open-code all that at every relevant > > > > > > > > > > > > > > > > > signal generation site... > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >> It is not clear to me that if you have adapted siginfo_layout. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Garbled sentence? > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Looks like. One of the pieces of code that needs to change > > > > > > > > > > > > > > > > when siginfo gets updated is siginfo_layout so that the structure > > > > > > > > > > > > > > > > can be properly decoded and made sense of. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > I am not seeing anything like that. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Okay, this has to do with copying between the compat and non-compat > > > > > > > > > > > > > > > versions of the struct? Sure, I can update that, although the code > > > > > > > > > > > > > > > would be basically non-functional on arm64 because TBI isn't supported > > > > > > > > > > > > > > > on 32-bit ARM. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > > > > > > >> > index cb3d6c267181..6dd82373eb2d 100644 > > > > > > > > > > > > > > > > >> > --- a/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > > > > > > >> > +++ b/include/uapi/asm-generic/siginfo.h > > > > > > > > > > > > > > > > >> > @@ -91,6 +91,14 @@ union __sifields { > > > > > > > > > > > > > > > > >> > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > > > > > > > > > > > > > >> > __u32 _pkey; > > > > > > > > > > > > > > > > >> > } _addr_pkey; > > > > > > > > > > > > > > > > >> > +#ifdef __aarch64__ > > > > > > > > > > > > > > > > >> > + /* used with all si_codes */ > > > > > > > > > > > > > > > > >> > + struct { > > > > > > > > > > > > > > > > >> > + short _dummy_top_byte; > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > ^ What's this for? I don't have Eric's insight here. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > We would need a short's worth of padding in order to prevent the > > > > > > > > > > > > > > > fields from occupying the same address as si_addr_lsb. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >> > + unsigned char _top_byte; > > > > > > > > > > > > > > > > >> > + unsigned char _top_byte_mask; > > > > > > > > > > > > > > > > >> > + } _addr_top_byte; > > > > > > > > > > > > > > > > >> > +#endif > > > > > > > > > > > > > > > > >> > }; > > > > > > > > > > > > > > > > >> > } _sigfault; > > > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > > >> Why the _dummy_top_byte? Oh I see it should be spelled "short _addr_lsb;". > > > > > > > > > > > > > > > > >> > > > > > > > > > > > > > > > > >> Please remove the "#ifdef __aarch64__". If at all possible we want to > > > > > > > > > > > > > > > > >> design this so any other architecture who has this challenge can use the > > > > > > > > > > > > > > > > >> code. The kind of code does not get enough attention/maintenance if it > > > > > > > > > > > > > > > > >> is built for a single architecture. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Seems reasonable. I was recently made aware that RISC-V was > > > > > > > > > > > > > > > considering a similar feature: > > > > > > > > > > > > > > > https://lists.riscv.org/g/tech-tee/topic/risc_v_tbi_proposal/72855478 > > > > > > > > > > > > > > > I would have opted to expand this to other architectures on an > > > > > > > > > > > > > > > as-needed basis, but I'd also be fine with having it on all > > > > > > > > > > > > > > > architectures from the start. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > If we make this arch-independent, we have an additional concern, which > > > > > > > > > > > > > > > is "what if some future architecture wants more than one byte here?" > > > > > > > > > > > > > > > For example, an architecture may have a "top-two-bytes-ignore" > > > > > > > > > > > > > > > feature, which would imply two-byte (misnamed) "si_addr_top_byte" and > > > > > > > > > > > > > > > "si_addr_top_byte_mask" fields. And the RISC-V proposal potentially > > > > > > > > > > > > > > > implies many more ignored bits (see slide 13 of the presentation). The > > > > > > > > > > > > > > > maximum size that these fields can possibly be is the size of a > > > > > > > > > > > > > > > pointer, and with that there wouldn't be enough room in the padding at > > > > > > > > > > > > > > > this point to accommodate the new fields. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > That basically implies your earlier suggestion of adding a union > > > > > > > > > > > > > > > member here to accommodate future expansion of the union, and adding > > > > > > > > > > > > > > > the new fields after the union. I'm happy to make that change, with > > > > > > > > > > > > > > > the fields renamed "si_addr_ignored" and "si_addr_ignored_mask". > > > > > > > > > > > > > > > > > > > > > > > > > > > > I think what we need here is basically a flags word. > > > > > > > > > > > > > > > > > > > > > > > > > > > > So long as we keep a flag spare to indicate the existence of a further > > > > > > > > > > > > > > flags word, we can extend as needed. > > > > > > > > > > > > > > > > > > > > > > > > > > > > How the existence of the first flags words is detected is another > > > > > > > > > > > > > > problem. If it only applies for newly-defined si_code values, then > > > > > > > > > > > > > > I guess si_code may be sufficient. > > > > > > > > > > > > > > > > > > > > > > > > > > Existing kernels will zero-initialize unused regions of the siginfo > > > > > > > > > > > > > data structure. The zero-initialization of the padding at the end of > > > > > > > > > > > > > the struct is done by the clear_user call here: > > > > > > > > > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/kernel/signal.c#L3193 > > > > > > > > > > > > > > > > > > > > > > > > > > and the zero-initialization of the padding between fields and unused > > > > > > > > > > > > > union members is done by the clear_siginfo function which the kernel > > > > > > > > > > > > > calls when initializing the data structure: > > > > > > > > > > > > > https://github.com/torvalds/linux/blob/3e08a95294a4fb3702bb3d35ed08028433c37fe6/include/linux/signal.h#L20 > > > > > > > > > > > > > > > > > > > > > > > > > > Therefore, a flag word value of 0 may be used to detect a lack of > > > > > > > > > > > > > support for flagged fields. > > > > > > > > > > > > > > > > > > > > > > > > It's not enough that we do this today. We would have had to do it back > > > > > > > > > > > > to the dawn of time (though in the arm64 case I guess we just need to go > > > > > > > > > > > > back to when the arch/arm64 was merged). > > > > > > > > > > > > > > > > > > > > > > > > v2.6.12:kernel/signal.c:copy_siginfo_to_user() suggests that this wasn't > > > > > > > > > > > > always the case, so unused parts of siginfo could be full of old junk > > > > > > > > > > > > from the user stack, if the kernel is sufficiently old. > > > > > > > > > > > > > > > > > > > > > > > > If we're trying to do something generic that makes sense on all arches, > > > > > > > > > > > > this matters. I may have misunderstood something about the code though. > > > > > > > > > > > > > > > > > > > > > > Hmm, I think you're right. The current behavior was introduced by > > > > > > > > > > > commit c999b933faa5e281e3af2e110eccaf91698b0a81 which was first > > > > > > > > > > > released in 4.18. So if an application wants to be compatible with > > > > > > > > > > > pre-4.18 kernels then there would need to be some other way to > > > > > > > > > > > indicate that the fields are valid. Probably the simplest way would be > > > > > > > > > > > to have the application issue a uname(2) syscall and check the kernel > > > > > > > > > > > version before reading these fields. I have a couple of other ideas > > > > > > > > > > > that don't rely on version detection, if we'd prefer to avoid that. > > > > > > > > > > > (They are somewhat ugly, but our hand is forced by backwards > > > > > > > > > > > compatibility.) > > > > > > > > > > > > > > > > > > > > > > One idea is to re-purpose the si_errno field as a flags field for > > > > > > > > > > > certain signal numbers. I checked a few kernel releases going back to > > > > > > > > > > > 2.6.18 and it looks like the field is set to 0 except in the following > > > > > > > > > > > circumstances: > > > > > > > > > > > - sending a hardware breakpoint (SIGTRAP/TRAP_HWBKPT) > > > > > > > > > > > - seccomp failures (SIGSYS/SYS_SECCOMP) > > > > > > > > > > > - user-defined signal via kill_pid_usb_asyncio > > > > > > > > > > > - SIGSWI in 3.18 and before (code since removed) > > > > > > > > > > > > > > > > > > > > > > It is also set to EFAULT for certain SIGSEGV/SEGV_MAPERR signals on > > > > > > > > > > > powerpc since commit c96c4436aba4c12f1f48369f2f90bc43e12fe36c, which > > > > > > > > > > > is currently unreleased. So if we wanted to go this route for SIGSEGV > > > > > > > > > > > we would need to stop the kernel from setting si_errno to EFAULT for > > > > > > > > > > > this signal before the 5.8 release. > > > > > > > > > > > > > > > > > > > > > > Another idea was to have userspace set a flag in sa_flags when > > > > > > > > > > > registering a signal handler meaning "this signal handler requires > > > > > > > > > > > unknown siginfo fields to be zeroed", and have existing kernels reject > > > > > > > > > > > the syscall due to an unknown flag being set, but unfortunately this > > > > > > > > > > > won't work because existing kernels do not reject sigaction syscalls > > > > > > > > > > > with unknown flags set in sa_flags. A perhaps more radical idea in > > > > > > > > > > > this vein would be to claim some of the upper bits of the signal > > > > > > > > > > > number as flags that will cause the syscall to be rejected if set and > > > > > > > > > > > unknown to the kernel. Existing kernels (going back to at least > > > > > > > > > > > 2.6.18) contain this code in do_sigaction: > > > > > > > > > > > > > > > > > > > > > > if (!valid_signal(sig) || sig < 1 || (act && sig_kernel_only(sig))) > > > > > > > > > > > return -EINVAL; > > > > > > > > > > > > > > > > > > > > > > and vald_signal is defined as: > > > > > > > > > > > > > > > > > > > > > > static inline int valid_signal(unsigned long sig) > > > > > > > > > > > { > > > > > > > > > > > return sig <= _NSIG ? 1 : 0; > > > > > > > > > > > } > > > > > > > > > > > > > > > > > > > > > > All architectures define _NSIG as a value <= 128, so they will reject > > > > > > > > > > > a signal number with any of bits 8-31 set. This means that we can use > > > > > > > > > > > any of those bits for mandatory flags. Most likely we could use bit 30 > > > > > > > > > > > (expanding down as necessary), as it keeps the signal number positive > > > > > > > > > > > and permits future expansion of the signal number range. > > > > > > > > > > > > > > > > > > > > Does the signal core code actually gurantee to zero the unused fields? > > > > > > > > > > Unless the fields are poked in by hand this is fraught with subtlelies, > > > > > > > > > > especially when unions are involved. (I'm sure the code tries to do it, > > > > > > > > > > but I've not eyeballed it in detail...) > > > > > > > > > > > > > > > > > > It memsets the siginfo structure before setting the fields and sending > > > > > > > > > the signal (grep for clear_siginfo which is just a memset; you should > > > > > > > > > find a call before all callers of force_sig_info). Memset is the right > > > > > > > > > approach here since unlike setting fields by hand it clears padding > > > > > > > > > which could lead to information leaks from the kernel. IIUC this is > > > > > > > > > the reason why Eric wants all of the signals to be raised via wrappers > > > > > > > > > in kernel/signal.c instead of via force_sig_info directly (to make > > > > > > > > > this aspect easier to audit). > > > > > > > > > > > > > > > > My impression was that the reason for this model is partly to ensure > > > > > > > > that siginfo fields are populated more consistently. When it was all > > > > > > > > down to the individual callers, inconsistencies creeped in. > > > > > > > > > > > > > > > > With regard to memset(), this is not a complete defence against data > > > > > > > > leakage. Assigning to a struct member can set any or all padding in > > > > > > > > the struct to random garbage (consider write-combining of neighboring > > > > > > > > member writes into a single larger accesses in asm for example). The > > > > > > > > > > > > > > I don't believe that LLVM will store to padding like this. I don't > > > > > > > know about GCC, though, but I wouldn't be surprised if this is > > > > > > > something that the kernel would want to turn off in "kernel C" (like > > > > > > > it turns off strict aliasing) specifically because of the information > > > > > > > leak issue. > > > > > > > > > > > > Again, the issue is not future kernel builds -- we can always find a way > > > > > > to fix the behaviour for those -- but past kernel builds. > > > > > > > > I thought that the whole point of the "bit in the signal number" (or > > > > SI_CODEX or whatever) was that we didn't need to worry about the > > > > behavior of past kernel builds? > > > > > > It depends on what we use the new flag(s) for. > > > > > > If the flag means just that unused padding is safely zeroed, that could > > > work -- but we'd want high confidence that it really is zeroed even in > > > wacky configurations. > > > > > > > > > > > only way to avoid this is to ensure that the struct is 100% > > > > > > > > padding-free, and that each member of a union is the same size. A > > > > > > > > quick clance at <uapi/asm-generic/siginfo.h> confirms that this is not > > > > > > > > the case. > > > > > > > > > > > > > > > > This might need to be looked at separately. > > > > > > > > > > > > > > > > But it does mean, strictly speaking, that we can't reliably add new > > > > > > > > fields anywhere that there was previously padding: assigning to > > > > > > > > neighboring members can still fill those with garbage after the > > > > > > > > memset(). > > > > > > > > > > > > > > ...but this is largely moot because I'm not proposing to add new > > > > > > > fields in the padding any more (because the fields needed to become > > > > > > > larger in order to accommodate future hypothetical architectures which > > > > > > > might want to use the fields, and thus they wouldn't fit in the > > > > > > > padding). The siginfo.h diff would be something like: > > > > > > > > > > > > > > diff --git a/include/uapi/asm-generic/siginfo.h > > > > > > > b/include/uapi/asm-generic/siginfo.h > > > > > > > index cb3d6c267181..4a2fe257415d 100644 > > > > > > > --- a/include/uapi/asm-generic/siginfo.h > > > > > > > +++ b/include/uapi/asm-generic/siginfo.h > > > > > > > @@ -91,7 +91,10 @@ union __sifields { > > > > > > > char _dummy_pkey[__ADDR_BND_PKEY_PAD]; > > > > > > > __u32 _pkey; > > > > > > > } _addr_pkey; > > > > > > > + void *_pad[6]; > > > > > > > }; > > > > > > > + uintptr_t _ignored_bits; > > > > > > > + uintptr_t _ignored_bits_mask; > > > > > > > > > > > > This _is_ in padding: the tail-padding of the (previously smaller) > > > > > > _sigfault. Again, the compiler was allowed to populate this area with > > > > > > junk before these fields were added. > > > > > > > > > > > > I agree that it seems fairly unlikely that the compiler would have been > > > > > > overwriting this in normal circumstances, but that's not a guarantee. > > > > > > My worry is that if this goes wrong, it will go wrong silently and > > > > > > unpredictably. > > > > > > > > > > > > > } _sigfault; > > > > > > > > > > > > > > /* SIGPOLL */ > > > > > > > > > > > > > > or with a "uintptr_t _flags" added in before _ignored_bits if we go with that. > > > > > > > > > > > > > > > > > Using unused bits in the signal number to turn on new functionality > > > > > > > > > > feels risky. As currently specified, this is just a number. Since > > > > > > > > > > today a successful sigaction(n ...) guarantees that n is a valid signal > > > > > > > > > > number, reasonable code like the following would trigger a buffer > > > > > > > > > > overrun if we start trying to encode anything else in there: > > > > > > > > > > > > > > > > > > > > struct sigaction actions[NSIG]; > > > > > > > > > > > > > > > > > > > > int do_something( ... ) > > > > > > > > > > { > > > > > > > > > > ... > > > > > > > > > > > > > > > > > > > > if (!sigaction(n, sa, ...)) { > > > > > > > > > > actions[n] = *sa; > > > > > > > > > > return 0; > > > > > > > > > > } > > > > > > > > > > > > > > > > > > > > ... > > > > > > > > > > } > > > > > > > > > > > > > > > > > > I imagine the bit in the signal number being set by the direct caller > > > > > > > > > to sigaction, and we could specifically recommend that calling > > > > > > > > > pattern. In that case, your "n" wouldn't have the bit set in it. It > > > > > > > > > > > > > > > > I can imagine this too, but that doesn't mean that software does it. > > > > > > > > > > > > > > > > If the above kind of thing exists in a framework or library somewhere, > > > > > > > > we could get problems. Similarly, a pre-existing LD_PRELOAD framework > > > > > > > > that provides a wrapper for sigaction may now go wrong even if your > > > > > > > > pattern is followed -- i.e., the caller thinks it's calling sigaction > > > > > > > > directly but in fact it isn't. > > > > > > > > > > > > > > I'm aware of one library like that. It's called libsigchain, and it > > > > > > > has an early bounds check: > > > > > > > https://cs.android.com/android/platform/superproject/+/master:art/sigchainlib/sigchain.cc;l=371 > > > > > > > > > > > > > > Until the library is changed to recognize the flag, calling code would > > > > > > > see the return value of -1 as if the kernel failed the syscall, and > > > > > > > would fall back to the code for old kernels. > > > > > > > > > > > > But only after some bad dereferences. If these were writes, this means > > > > > > that memory _may_ be silently corrupted (I don't say it't likely in a > > > > > > given case, and we cannot pick a flag bit that makes this impossible). > > > > > > > > You're talking about libsigchain, right? I don't see any bad > > > > references, the function returns after noticing the bounds check > > > > failure. > > > > > > Yes, I confused myself by reading Handler() out of context. The kernel > > > will invoke this with signo to a real signal number (without any flags). > > > > > > The sigaction wrapper does the bounds check before doing anything else, > > > just as you say -- so that looks fine. > > > > > > (Side question: is all this thread-safe? Is there some implicit locking > > > somewhere?) > > > > I think maybe it isn't? There seem to be possible races on the > > handler_ field. One possibility is that the function could race with > > itself on another thread, which could be fixed via locking, but it > > would also need to handle races between itself and the signal handler, > > most likely by blocking the signal while setting it. > > Hmmm, tricky... anyway, that's not my problem ;) > > > > > > > So, _even though the user program is correct_, our change may trigger > > > > > > > > Let's say that you were talking about some other library and not > > > > libsigchain. Such an interceptor wouldn't be correct though, it failed > > > > to account for our change to the syscall semantics. If the accesses > > > > were before the syscall (or the bounds check), then the interceptor > > > > would not have been correct in the first place because POSIX requires > > > > returning -1 with errno=EINVAL (and not crashing) if the signal number > > > > is invalid. > > > > > > > > > > the corruption of arbitrary user memory. This what I mean by an ABI > > > > > > break. The fact that the corruption is not done by the syscall itself > > > > > > is no excuse. > > > > > > > > At some point, though, accommodating interceptors becomes pretty much > > > > tantamount to saying "we can never change anything". Even just adding > > > > a field to __sifields (which is pretty much required for what we need > > > > to do) could break things in the presence of some interceptors because > > > > the interceptor could be copying the fields manually to a new data > > > > structure before calling the user's signal handler (e.g. because it > > > > wants to defer the signal until later) and miss our new field. I think > > > > most of the other ideas we're discussing fail to meet this bar as well > > > > and I'll go into more details later on. > > > > > > I agree we cannot always avoid breaking such things. But we should do > > > our best to avoid it. > > > > I think that given the hand that we've been dealt, no matter what we > > do, we can't really avoid risking breaking something. The relevant > > questions are "what are we going to risk breaking", "how much risk is > > there", "will it be easily noticed/fixable", and "once we're on the > > other side of the potential breakage, will we find ourselves in a > > position where changing things involves less breakage risk". > > > > > > > > We also fail to notice failures in sigaddset() etc., though in this code > > > > > > it looks like that should not matter. > > > > > > > > Maybe you're looking at the handler ("SignalChain::Handler")? The bit > > > > wouldn't be set in the signo argument to the handler. I'm talking > > > > about line 371 of the code I linked, in the sigaction interceptor > > > > "__sigaction" (it looks like sometimes the link doesn't take you to > > > > the correct line for some reason). > > > > > > Ack, I confused myself. > > > > > > > > > > In general I think that any library like this with independent > > > > > > > tracking of the kernel's purported signal handler state would need to > > > > > > > be very sensitive to which syscalls are capable of setting signal > > > > > > > handlers, what their semantics are, and so on. This applies to any > > > > > > > change that we might make to the signal handler interface. So for > > > > > > > example, if we introduced a new syscall as you propose below, and the > > > > > > > library hasn't been updated to recognize the new syscall, it will > > > > > > > silently miss changes in signal handler state caused by the new > > > > > > > syscall. > > > > > > > > > > > > > > At the end of this argument lies "we can never change anything about > > > > > > > how signal handlers work because it could break some interposing > > > > > > > library somewhere" -- replace "signal handlers" with any kernel > > > > > > > feature whose behavior may be modified by an interposing library if > > > > > > > you like -- and I don't think we want to go that far. As far as I > > > > > > > know, this isn't really the kernel's business anyway -- the kernel's > > > > > > > stable ABI contract starts and ends with the syscall interface and not > > > > > > > some library on top. > > > > > > > > > > > > > > That being said, we should perhaps try to define our interface so that > > > > > > > something reasonable will probably happen if there is such a library > > > > > > > and it hasn't been updated. With the new syscall, the library will > > > > > > > sometimes silently fail to work in some non-local fashion. With the > > > > > > > flag bit in the signal number, the library will either cause the > > > > > > > caller to fall back to the old kernel code path (if there is a bounds > > > > > > > check) or likely crash loudly (if there is no bounds check). To me, > > > > > > > the "flag bit in the signal number" behavior seems more reasonable, > > > > > > > since either something correct or something easy to debug will > > > > > > > probably happen at runtime. > > > > > > > > > > > > > > > > could only appear in newly-written code that doesn't follow our > > > > > > > > > recommendations, and there are already plenty of much more likely ways > > > > > > > > > to cause buffer overflows in C code that doesn't follow > > > > > > > > > recommendations anyway. (And even if such a buffer overflow occurred, > > > > > > > > > it would very likely be caught early in development by the MMU due to > > > > > > > > > the magnitude of the number 1<<30.) > > > > > > > > > > > > > > > > Choosing the bit value is hard. If shitfing it overflows, this can > > > > > > > > trigger random undefined behaviour in the compiler in addition to (or > > > > > > > > perhaps instead of) an out-of-bounds access or segfault. > > > > > > > > > > > > > > It wouldn't overflow on a 64-bit architecture assuming normal array > > > > > > > indexing (the index would be promoted to pointer width before being > > > > > > > scaled to the array element size), and to begin with the users of this > > > > > > > would be 64-bit. > > > > > > > > > > > > Unless we don't offer this feature for 32-bit at all (possible, if ugly) > > > > > > we can't stop people using it. > > > > > > > > My point is that the problem in the interceptor library would probably > > > > be noticed on 64-bit (since that's what most people use these days), > > > > which would probably result in it being fixed by the time it reaches > > > > 32-bit users. > > > > > > Agreed. But we shouldn't take such bets unless we really have to. > > > > > > > > > > > If shifting it doesn't overflow, we might still fall into a valid > > > > > > > > mapping, though I'd agree a segfault is more likely. > > > > > > > > > > > > > > > > > > > > > > > > > > > I think it would be cleaner for to add a single flag field that can be > > > > > > > > > > used for detecting other extensions, and request it via a new sa_flags > > > > > > > > > > bit. This removes the need for sematically useless zeroing of unused > > > > > > > > > > fields (though for hygiene and backwards compatibility reasons we would > > > > > > > > > > probably want to carry on zeroing them anyway). > > > > > > > > > > > > > > > > > > > > I can see no simpler way to add supplementary siginfo fields for > > > > > > > > > > existing si_codes. For si_codes that didn't exist before the zeroing > > > > > > > > > > came in we could still detect optional si_code-specific fields via > > > > > > > > > > zeroing, but it seems messary to have two ways of detecting extensions. > > > > > > > > > > > > > > > > > > That would certainly be cleaner if it worked, but that would only be > > > > > > > > > the case if old kernels rejected unknown bits in sa_flags, and > > > > > > > > > unfortunately they don't. With the bit in the signal number, the "old > > > > > > > > > > > > > > > > Hmm, that is a problem I wasn't aware of. > > > > > > > > > > > > > > > > > kernels reject" behavior admits relatively straightforward usage code: > > > > > > > > > > > > > > > > > > void set_segv_handler(void) { > > > > > > > > > struct sigaction sa; > > > > > > > > > sa.sa_sigaction = handle_segv; > > > > > > > > > sa.sa_flags = SA_SIGINFO; > > > > > > > > > if (sigaction(SIGSEGV | SF_CLEAR_UNKNOWN_FIELDS, &sa, 0) < 0) { // > > > > > > > > > succeeds in new kernels, fails in old kernels > > > > > > > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > > > > > > > if (sigaction(SIGSEGV, &sa, 0) < 0) // succeeds in old kernels > > > > > > > > > perror("sigaction"); > > > > > > > > > } > > > > > > > > > } > > > > > > > > > > > > > > > > > > void clear_fields_and_handle_segv(int signum, siginfo_t *sa, void *ctx) { > > > > > > > > > sa->si_future_field = 0; > > > > > > > > > handle_segv(signum, sa, ctx); > > > > > > > > > } > > > > > > > > > > > > > > > > > > void handle_segv(int signum, siginfo_t *sa, void *ctx) { > > > > > > > > > // At this point, si_future_field will have the value 0 in old > > > > > > > > > kernels and the kernel-supplied value in new kernels. > > > > > > > > > } > > > > > > > > > > > > > > > > > > Imagine if we moved the flag SF_CLEAR_UNKNOWN_FIELDS from the signal > > > > > > > > > number to sa_flags. In that case, the first sigaction would succeed in > > > > > > > > > old kernels so handle_segv wouldn't know whether it can safely read > > > > > > > > > from si_future_field. With the sa_flags approach, you would need > > > > > > > > > kernel version number checking via uname before setting the flag in > > > > > > > > > sa_flags, and at that point why even have the flag in sa_flags at all > > > > > > > > > since you could just have the signal handler conditionally read from > > > > > > > > > si_future_field based on the uname? > > > > > > > > > > > > > > > > Software setting SA_SIFLAGS (or whatever) is new by definition, since > > > > > > > > it would be using a new #define. So it might be reasonable to put the > > > > > > > > burden on that software to verify that the flag was really accepted by > > > > > > > > the kernel, by reading it back. > > > > > > > > > > > > > > That doesn't seem like a good idea even if it worked, because it could > > > > > > > lead to race conditions. If the si_flags-reading signal handler were > > > > > > > invoked in response to a signal between when you set it and when you > > > > > > > ended up replacing it with the fallback signal handler for old > > > > > > > kernels, the handler may end up reading garbage data from si_flags. > > > > > > > > > > > > Not really. My example may have this problem, but the signal handler > > > > > > can be written to support both scenarios, based on testing a flag that > > > > > > the main program sets after verifying that the flag could be set. Or > > > > > > the signal could be blocked around establishment (often a good idea for > > > > > > other reasons). > > > > > > > > > > > > But I agree it's a bit gross, and anyway doesn't work due to the fact > > > > > > that the kernel doesn't filter out unrecognised flags anyway. > > > > > > > > > > > > > > Unfortunately, even relatively recent kernels blindly store sa_flags > > > > > > > > in the kernel without validating it, and so it looks like duff flags > > > > > > > > can be read back out via a sigaction() call. Dang. > > > > > > > > > > > > > > > > > > > > > > > > Perhaps a new frontend syscall could be added. A new libc that knows > > > > > > > > about this "sigaction2" could use it and mask off problem bits from > > > > > > > > sa_flags in its sigaction() wrapper before calling sigaction2. An old > > > > > > > > libc would call the old sigaction syscall, where we would ignore these > > > > > > > > new sa_flags bits as before. > > > > > > > > > > > > > > I'm not currently in favor of the new syscall but if we do this I > > > > > > > would keep sigaction and sigaction2 separate. That is, libc sigaction > > > > > > > should always use the sigaction syscall, and libc sigaction2 should > > > > > > > always use the sigaction2 syscall. We should avoid libc's sigaction > > > > > > > having different behavior based on the libc version and kernel > > > > > > > version, as that would make it harder to reason about its behavior. > > > > > > > Calling code would need to check for presence of sigaction2 in both > > > > > > > libc and the kernel, e.g. > > > > > > > > > > > > > > __attribute__((weak)) decltype(sigaction2) sigaction2; > > > > > > > > > > > > > > void set_segv_handler(void) { > > > > > > > struct sigaction sa; > > > > > > > sa.sa_sigaction = handle_segv; > > > > > > > sa.sa_flags = SA_SIGINFO | SA_SIFLAGS; > > > > > > > if (!sigaction2 || sigaction2(SIGSEGV, &sa, 0) < 0) { > > > > > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > > > > > sa.sa_flags = SA_SIGINFO; > > > > > > > if (sigaction(SIGSEGV, &sa, 0) < 0) > > > > > > > perror("sigaction"); > > > > > > > } > > > > > > > } > > > > > > > > > > > > I guess. But I share your distaste for adding a new syscall. > > > > > > > > > > > > > > > > > > > > > This may not be a popular approach though, and software wouldn't be able > > > > > > > > to use our new features until libc is updated to match. > > > > > > > > > > > > > > > > If we go down this route, it may provide additional opportunities to fix > > > > > > > > annoying defects in the old interface. > > > > > > > > > > > > > > > > > > > > > > > > > Note that the same applies to a flag indicating the availability of a > > > > > > > > > si_flags field in sigaction (just > > > > > > > > > s/SF_CLEAR_UNKNOWN_FIELDS/SF_HAS_SI_FLAGS/ and > > > > > > > > > s/si_future_field/si_flags/ in the usage code above). In terms of > > > > > > > > > SF_CLEAR_UNKNOWN_FIELDS versus SF_HAS_SI_FLAGS I'd be fine either way. > > > > > > > > > > > > > > > > > > Another thought that occurred to me is that we may consider > > > > > > > > > generalizing this a step further and introducing a single flag bit in > > > > > > > > > the signal number that means "reject unknown flags in sa_flags". This > > > > > > > > > would mean that we wouldn't need to add any more flag bits to the > > > > > > > > > signal number in the future, thus limiting this signal number hack to > > > > > > > > > a single bit; all future mandatory behavior changes could just be put > > > > > > > > > behind a flag in sa_flags and userspace code would easily be able to > > > > > > > > > detect missing support for a flag and fall back if necessary. In our > > > > > > > > > case, this would imply usage code like this: > > > > > > > > > > > > > > > > > > void set_segv_handler(void) { > > > > > > > > > struct sigaction sa; > > > > > > > > > sa.sa_sigaction = handle_segv; > > > > > > > > > sa.sa_flags = SA_SIGINFO | SA_CLEAR_UNKNOWN_FIELDS; > > > > > > > > > // Succeeds in kernels with SA_CLEAR_UNKNOWN_FIELDS support. > > > > > > > > > // Fails in kernels with SF_CHECK_SA_FLAGS support but no > > > > > > > > > SA_CLEAR_UNKNOWN_FIELDS support (because of the unknown flags check). > > > > > > > > > // Fails in kernels without SF_CHECK_SA_FLAGS support (because of > > > > > > > > > the bounds check on the signal number). > > > > > > > > > if (sigaction(SIGSEGV | SF_CHECK_SA_FLAGS, &sa, 0) < 0) { > > > > > > > > > sa.sa_sigaction = clear_fields_and_handle_segv; > > > > > > > > > sa.sa_flags = SA_SIGINFO; > > > > > > > > > // Succeeds in old kernels, no need to use SF_CHECK_SA_FLAGS since > > > > > > > > > we're using sa_flags from the beginning of time. > > > > > > > > > if (sigaction(SIGSEGV, &sa, 0) < 0) > > > > > > > > > perror("sigaction"); > > > > > > > > > } > > > > > > > > > } > > > > > > > > > > > > > > > > As with the other options this could work, but looks like it could > > > > > > > > break the ABI due to violating the original semantics for the signal > > > > > > > > number argument. Perhaps I'm being too paranoid. > > > > > > > > > > > > > > There's no ABI being broken here, as long as we consider syscalls to > > > > > > > be the stable ABI layer. Old kernels are simply rejecting arguments > > > > > > > that they don't know about yet. By that argument, any introduction of > > > > > > > a new syscall is an ABI break because it changes the semantics of a > > > > > > > previously-unallocated syscall number. > > > > > > > > > > > > As argued above, I think this is an invalid argument. > > > > > > > > > > > > Although any addition will change behaviour (so is a break in some > > > > > > sense), the key is not to make "surprising" changes. > > > > > > > > If we care about interceptors then I don't think "surprising" comes > > > > into it. It's more a question of "does the anticipated behavior of the > > > > interceptor match our desired behavior", where "desired" means "most > > > > likely to avoid silent breakage". We would need to get into the head > > > > of a potential interceptor author and think about how they would have > > > > handled the signal number argument, as well as other arguments like > > > > sa_flags if we want to go that route, and see whether that behavior > > > > would lead to the desired result. > > > > > > That's exactly what I mean by "surprising". > > > > Not quite, see below. > > > > > However, not every > > > interceptor author will be making the same assumptions, and not every > > > bit of software affected will be an interceptor. > > > > I can see a couple of ways in which non-interceptor software could be affected: > > > > - It's doing something like "call sigaction on every possible signal > > number in the 31-bit range and end up failing if the syscall > > succeeded" (e.g. with an OOB write). Perhaps software could be doing > > something like this in a loop to collect all currently registered > > signal handlers. That being said, this program: > > > > #include <limits.h> > > #include <signal.h> > > > > int main() { > > struct sigaction act; > > for (int i = 1; i != INT_MAX; ++i) { > > sigaction(i, 0, &act); > > } > > } > > > > takes around 5 seconds to run on my relatively-fast machine, so I > > would expect any such code to be noticed as a performance issue and > > either be changed to be bounded on _NSIG or break on EINVAL. > > > > This is probably the largest potential flaw that I can currently see > > in the "bit in the signal number" idea, since it could conceivably > > result in userspace code being broken without having first required it > > to have been changed to make use of the new feature. I'm not convinced > > that it would be an ABI break though, because the code seems unlikely > > to exist in this form in the wild because of the performance issue, > > and you could anyway make the argument that the code is incorrect > > because, in order to contain a loop like this, it would need to be > > able to handle large, previously-unknown signal numbers somehow. If we > > accept that the code is incorrect, a similar line of argument applies > > as for interceptors (i.e. likely to result in an OOB access which will > > fail loudly and be easily debugged and fixed). > > > > - If we do something that involves introducing a new flag in sa_flags, > > the flag may be exposed to unaware software via the oldact argument to > > sigaction, and I suppose that it's conceivable that exposing a > > previously-unknown flag like this could somehow break something. But > > this seems like an unreasonable restriction because it would mean that > > we can never add a flag to sa_flags no matter what. > > > > > So some judgement > > > needs to be applied. > > > > Of course. We need to agree *how* to apply the judgement though. > > > > > > In this case, I think we exactly want the interceptor author to have > > > > thought "oh, it's just a number, I'll (possibly do a bounds check and > > > > then) use the number as an index into an array". This will lead to one > > > > of two outcomes: crashing (yes, yes, it won't always crash, but if the > > > > alternative is that it never crashes and we get silently incorrect > > > > behavior all of the time, I'll take sometimes crashing) or fail the > > > > bounds check and pretend to be an old kernel (the latter is > > > > anticipated by POSIX which requires returning -1/EINVAL for an invalid > > > > signal number). Each of these behaviors are desirable, as they are > > > > observable failures, which are more likely to result in fixes than > > > > silent ones. > > > > > > Agreed, except wanting the author to have thought something doesn't > > > ensure that they actually did think that. > > > > True, but if our goal is only to accommodate reasonably written > > interceptors, we don't actually need to ensure anything here. > > > > > > > > Having something random happen when setting a previously reserved flag > > > > > > bit, or when issuing a syscall when an unknown syscall number, or not > > > > > > surprising at all. > > > > > > > > Introducing a new syscall is right out in this model. The interceptor > > > > author wouldn't have anticipated our introducing a new syscall, so the > > > > new syscall wouldn't be intercepted and calls to the new syscall would > > > > silently bypass the interceptor. For example, adding sigaction2 could > > > > result in signal handlers being set without the interceptor's > > > > knowledge. > > > > > > Agreed. My sentence was a bit mangled: I mean to say "Having something > > > random happen when [...] issuing a syscall *with* an unknown syscall > > > number *is* not surprising at all." > > > > > > I agree that adding a new syscall is problematic if we want to avoid > > > breaking existing interceptors in particular. Other types of code are > > > much less likely to be affected by the addition of new syscalls. > > > > Right, and this to me is a case in point for why I would say that > > "surprising" isn't the right frame of analysis here. My analysis seems > > to generally be that "anticipated interceptor behavior matches desired > > behavior" is positively correlated with "surprising" (i.e. the > > interceptor viewpoint is the dual of the user viewpoint), so if we > > care about interceptors we may end up making a "surprising" change > > even though it doesn't intuitively seem like the right thing to do. > > You're right that interceptors are different from normal callers. I'm > not sure I follow your argument, but an alternative way of looking at > it might be to say that an interceptor is both an implementation of an > interface and a caller of the same interface. Since API specs are > rarely complete enough to cover the corner cases that arise from this, > full portability is hard to achieve on top of an evolving kernel. > > However, I think this interceptor thing is a bit of a red herring. I > just intended that as an illustration of the kind of code that might > fall foul. This doesn't mean that it's 100% certain that no other > software can be affected. > > The starting points for this discussion were: "is it reasonable for a > caller to pass an unvalidated signal number to sigaction(), and rely on > sigaction() to validate it?" and "is it reasonable to assume that a > signal number accepted by sigaction() fits the POSIX specification of > a valid signal number?" > > I think yes; you aren't (or weren't) convinced. > > The mere fact that it's hard to agree suggests to me that the > specification is too weak to extend safely in this area. Unfortunately, > it's rather weak for sa_flags too, although a non-full flags argument > does at least suggest that future extensions might appear. > > > > > Regarding a sa_flags bit, let's get inside the head of the interceptor > > > > author again. How would they handle a flag bit that they don't > > > > recognize when replacing the signal handler? It wouldn't be correct to > > > > just pass it through to the kernel, or drop the flag on the floor, as > > > > it might be semantically meaningful (and thus could change the calling > > > > convention as SA_SIGINFO does, or change the meaning of fields in > > > > siginfo, as SA_CODEX would do). A correctly written sigaction > > > > interceptor should probably abort the program upon encountering an > > > > unknown flag (thus giving a human a chance to update the interceptor), > > > > but chances are that they don't. Ignoring all but a few flags (and > > > > passing a fixed set of flags to the kernel) seems to be what > > > > libsigchain does, and in the case of SA_CODEX it would seem to result > > > > in desirable behavior (but I suspect that it isn't handling the other > > > > flags correctly), but I could also see an interceptor author just > > > > passing it unchanged to the kernel without checking it (perhaps > > > > because they didn't think about these issues, and because that didn't > > > > matter until now, with the exception of from-the-beginning-of-time > > > > flags like SA_SIGINFO). And with SA_CODEX that could lead to silent > > > > misreading of si_code in the interceptor's signal handler, if it > > > > hasn't been updated to use the new macros. > > > > > > Agreed. I've tried to implement things rather like this in the past, > > > and how to interpret the flags is a tricky issue. Some of the flags are > > > impossible to emulate even when you know what they mean, in particular > > > SA_NODEFER and SA_RESTART. > > > > > > Making new flags safe to ignore and harmless to set of you don't know > > > what they mean is the safest approach, but not always possible (I think > > > I managed this with by suggestion below, though). > > > > Again, this suggestion could lead to silent failures in an interceptor, if: > > - the interceptor passes the sa_flags through to the kernel unchanged > > (or otherwise doesn't touch SA_CODEX) > > - the interceptor replaces the user's sa_sigaction > > - the interceptor's replacement sa_sigaction tests the provided si_code. > > > > Maybe you're not concerned about that, though? At least to me it seems > > in the same ballpark of likelihood as the ways in which things could > > go wrong with the signal number bit. > > I agree this is a concern, and perhaps a bit nastier in practice than > side-effects of setting random bits in the signal number. > > Personally I do tend to be paranoid about flags arguments and try to > police them in any code that isn't a trivial pass-through, but this > doesn't mean that all code out there does it. (Including the kernel's > sigaction()!) > > > > Ideally, a flags field should be specified with rules that say exactly > > > what to do with flags you don't recognise. Sadly this is usually not > > > thought about until it's too late. > > > > It perhaps isn't too late to introduce such rules for sigaction if we > > adopt the signal number bit and we make it mean "reject unknown > > flags". > > If Eric likes the idea then fair enough, but as I've tried to argue this > may still just be moving the problem around rather than solving it. > > > As a final random idea to add to the mix, we could add two or more > flags in sa_flags, and require the kernel to transform them in a > specific way, say: > > #define SA_WANT_FLAGS 0x00c700000 > #define SA_HAVE_FLAGS 0x009200000 > #define SA_FLAGS_MASK 0x00ff00000 > > volatile sig_atomic_t have_flags = 0; > > sa.sa_flags |= SA_WANT_FLAGS; > if (sigaction(n, &sa, NULL)) > if (!sigaction(n, NULL, &sa) && > (sa.sa_flags & SA_FLAGS_MASK) == SA_HAVE_FLAGS) > have_flags = 1; > > This is at least proof against "dumb readback". > > Provided that the handler can cope with the have_flags == 0 case and > just reads the flag once per call, I don't think we would need to worry > about races. > > Of course, an interceptor that doesn't understand this mechanism and > munges or manufactures its own siginfo might still fail to properly > initialise our new field before passing it on to a signal handler that > is expecting it. But that's already broken: such an interceptor might > also not understand new si_codes that the client code absolutely relies > on. And new si_codes _do_ get added (that's another extensibility fail > in the existing signal API). > > > So... overall, maybe a bit in the signal number isn't a lot worse, and > perhaps it _will_ lead to cleaner failures. > > Really, I don't see a way to solve it properly without a new API. I started on implementing my signal number bit idea, and in the process of doing so came up with another idea that may be better from the "don't abuse existing arguments" perspective. It involves a sigaction protocol similar to the one that you describe above, but it only requires one new bit (plus one bit per new flag) so it is less wasteful of sa_flags bits. The idea is twofold: 1. Require the kernel to clear unknown flag bits in sa_flags when passing them back in oldact. I suppose that this is technically a behavior change for sigaction, but critically, this change in behavior only applies to unallocated flags, which we are free to change the meaning of. We can simply define each existing unallocated flag bit to mean "clear this bit in oldact (unless the bit becomes supported in the future)". There was already code doing something similar in a limited fashion on x86, which we can remove by using this approach. 2. Define a flag bit SA_UNSUPPORTED which will never be supported by the kernel. Now userspace can use the fact that the bit has been cleared to mean that it can trust that other unsupported bits have also been cleared. Now we may have code like this: #define SA_UNSUPPORTED 0x400 #define SA_XFLAGS 0x800 volatile sig_atomic_t have_xflags = 0; sa.sa_flags |= SA_UNSUPPORTED | SA_XFLAGS; if (sigaction(n, &sa, NULL)) if (!sigaction(n, NULL, &sa) && !(sa.sa_flags & SA_UNSUPPORTED) && (sa.sa_flags & SA_XFLAGS)) have_xflags = 1; > In the meantime, can I suggest: > > (1) Come up with an extensible way of encoding supplementary > information in siginfo. If the consensus is that zeroing unused > fields is sufficient and that the kernel and compiler will > reliably do it, then great. Otherwise, we might need explicit > flags fields or something. I thought about this for a while and concluded that we probably want a flags field anyway. si_addr_ignored_bits is something of a special case in the sense that we can define the zero value to mean "unknown" by taking advantage of the mask field (which I suppose is something of a flags field), but we can't necessarily say that the same is true for any fields that we may add in the future. For example, if we wanted to communicate whether the failing access is a read or a write, we would need a tristate: read, write and "unknown" (and arrange for old kernels' behavior to be interpreted as "unknown"). If we rely on zeroing then we may implement this by adding a field like: char si_access_type; // 0 = unknown, 1 = read, 2 = write But that's really just a (slightly wasteful, because we use the entire byte) flags field, so we may as well define an actual flags field to begin with and let people add their flags there. Unfortunately we can't name it sa_flags because ia64 got there first. We may consider making the ia64 field generic though (ia64 only uses one bit of their field, so we would have 31 free bits). In the meantime, I added a separate field, sa_xflags. > (2) Hack up any simple mechanism (such as your signal number flag) for > requesting/detecting the extra information. > > Along with an illustration of a application of the mechanism (i.e., > reporting address tag bits), this should at least provide a basis for > further review. > > We can then try to swap in a different mechanism for (2) if people have > still have concerns (or it not, keep it). Sounds good. Apologies for not replying sooner, I was hoping that Eric would chime in so that I would get a sense of which approach he would prefer (so that I wouldn't spend as much time implementing in an undesired direction), then this fell off my radar. I decided to go with the SA_UNSUPPORTED approach that I mentioned above for now, and I'll send a v9 with that implemented shortly. Most of the change is about letting the architecture-independent code know which bits are supported, so it should be easy to replace the detection mechanism with another idea like the signal number bit. Peter
On Mon, Aug 17, 2020 at 08:16:42PM -0700, Peter Collingbourne wrote: > On Tue, Jul 14, 2020 at 10:36 AM Dave Martin <Dave.Martin@arm.com> wrote: [...] > > As a final random idea to add to the mix, we could add two or more > > flags in sa_flags, and require the kernel to transform them in a > > specific way, say: > > > > #define SA_WANT_FLAGS 0x00c700000 > > #define SA_HAVE_FLAGS 0x009200000 > > #define SA_FLAGS_MASK 0x00ff00000 > > > > volatile sig_atomic_t have_flags = 0; > > > > sa.sa_flags |= SA_WANT_FLAGS; > > if (sigaction(n, &sa, NULL)) > > if (!sigaction(n, NULL, &sa) && > > (sa.sa_flags & SA_FLAGS_MASK) == SA_HAVE_FLAGS) > > have_flags = 1; > > > > This is at least proof against "dumb readback". > > > > Provided that the handler can cope with the have_flags == 0 case and > > just reads the flag once per call, I don't think we would need to worry > > about races. > > > > Of course, an interceptor that doesn't understand this mechanism and > > munges or manufactures its own siginfo might still fail to properly > > initialise our new field before passing it on to a signal handler that > > is expecting it. But that's already broken: such an interceptor might > > also not understand new si_codes that the client code absolutely relies > > on. And new si_codes _do_ get added (that's another extensibility fail > > in the existing signal API). > > > > > > So... overall, maybe a bit in the signal number isn't a lot worse, and > > perhaps it _will_ lead to cleaner failures. > > > > Really, I don't see a way to solve it properly without a new API. > > I started on implementing my signal number bit idea, and in the > process of doing so came up with another idea that may be better from > the "don't abuse existing arguments" perspective. It involves a > sigaction protocol similar to the one that you describe above, but it > only requires one new bit (plus one bit per new flag) so it is less > wasteful of sa_flags bits. > > The idea is twofold: > > 1. Require the kernel to clear unknown flag bits in sa_flags when > passing them back in oldact. I suppose that this is technically a > behavior change for sigaction, but critically, this change in behavior > only applies to unallocated flags, which we are free to change the > meaning of. We can simply define each existing unallocated flag bit to > mean "clear this bit in oldact (unless the bit becomes supported in > the future)". There was already code doing something similar in a > limited fashion on x86, which we can remove by using this approach. Sounds reasonable. It's quite hard to imagine how software could accidentally rely on unallocated sa_flags bits reading back out unmodified through sigaction(). Software that deliberately relies on this for some bit that is never allocated would be obviously non POSIX compliant. > 2. Define a flag bit SA_UNSUPPORTED which will never be supported by > the kernel. Now userspace can use the fact that the bit has been > cleared to mean that it can trust that other unsupported bits have > also been cleared. > > Now we may have code like this: > > #define SA_UNSUPPORTED 0x400 > #define SA_XFLAGS 0x800 > > volatile sig_atomic_t have_xflags = 0; > > sa.sa_flags |= SA_UNSUPPORTED | SA_XFLAGS; > if (sigaction(n, &sa, NULL)) > if (!sigaction(n, NULL, &sa) && > !(sa.sa_flags & SA_UNSUPPORTED) && > (sa.sa_flags & SA_XFLAGS)) > have_xflags = 1; OK, so I think the novelty here is that we detect support by requiring the kernel to clear one bit while preserving another. That does indeed seem more robust: an old kernel (unless bizarrely buggy) would either clear all unsupported bits or preserve them all. Other OSes that extend their sigaction() in line with POSIX would also be highly unlikely to exhibit this behaviour by accident IMHO, making it easier for this to coexist with other people's extensions. Interceptors still may not transparently work with this approach, but I think that's a reasonable price to pay. > > > In the meantime, can I suggest: > > > > (1) Come up with an extensible way of encoding supplementary > > information in siginfo. If the consensus is that zeroing unused > > fields is sufficient and that the kernel and compiler will > > reliably do it, then great. Otherwise, we might need explicit > > flags fields or something. > > I thought about this for a while and concluded that we probably want a > flags field anyway. si_addr_ignored_bits is something of a special > case in the sense that we can define the zero value to mean > "unknown" by taking advantage of the mask field (which I suppose is > something of a flags field), but we can't necessarily say that the > same is true for any fields that we may add in the future. For > example, if we wanted to communicate whether the failing access is a > read or a write, we would need a tristate: read, write and "unknown" > (and arrange for old kernels' behavior to be interpreted as > "unknown"). If we rely on zeroing then we may implement this by adding > a field like: > > char si_access_type; // 0 = unknown, 1 = read, 2 = write > > But that's really just a (slightly wasteful, because we use the entire > byte) flags field, so we may as well define an actual flags field to > begin with and let people add their flags there. > > Unfortunately we can't name it sa_flags because ia64 got there first. > We may consider making the ia64 field generic though (ia64 only uses > one bit of their field, so we would have 31 free bits). In the > meantime, I added a separate field, sa_xflags. Seems fair enough to me. We still have the option to use zeroing for detection of fields where it works. Would SA_XFLAGS imply zeroing of unallocated siginfo fields? This may just be a matter of documentation, if the kernel already does the zeroing today. > > > (2) Hack up any simple mechanism (such as your signal number flag) for > > requesting/detecting the extra information. > > > > Along with an illustration of a application of the mechanism (i.e., > > reporting address tag bits), this should at least provide a basis for > > further review. > > > > We can then try to swap in a different mechanism for (2) if people have > > still have concerns (or it not, keep it). > > Sounds good. Apologies for not replying sooner, I was hoping that Eric > would chime in so that I would get a sense of which approach he would > prefer (so that I wouldn't spend as much time implementing in an > undesired direction), then this fell off my radar. I decided to go > with the SA_UNSUPPORTED approach that I mentioned above for now, and > I'll send a v9 with that implemented shortly. Most of the change is > about letting the architecture-independent code know which bits are > supported, so it should be easy to replace the detection mechanism > with another idea like the signal number bit. No worries, I think various people have had distractions (I certainly have ... but I digress). I'll take a look at your v9. Cheers ---Dave
diff --git a/Documentation/arm64/tagged-pointers.rst b/Documentation/arm64/tagged-pointers.rst index eab4323609b9..f1880ed5cdf2 100644 --- a/Documentation/arm64/tagged-pointers.rst +++ b/Documentation/arm64/tagged-pointers.rst @@ -53,12 +53,17 @@ visibility. Preserving tags --------------- -Non-zero tags are not preserved when delivering signals. This means that -signal handlers in applications making use of tags cannot rely on the -tag information for user virtual addresses being maintained for fields -inside siginfo_t. One exception to this rule is for signals raised in -response to watchpoint debug exceptions, where the tag information will -be preserved. +Non-zero tags are not preserved in the fault address fields +siginfo.si_addr or sigcontext.fault_address when delivering +signals. This means that signal handlers in applications making use +of tags cannot rely on the tag information for user virtual addresses +being maintained in these fields. One exception to this rule is for +signals raised in response to watchpoint debug exceptions, where the +tag information will be preserved. + +The fault address tag is preserved in the si_addr_top_byte field of +siginfo, which is set for signals raised in response to data aborts +and instruction aborts. The architecture prevents the use of a tagged PC, so the upper byte will be set to a sign-extension of bit 55 on exception return. diff --git a/arch/arm64/include/asm/exception.h b/arch/arm64/include/asm/exception.h index 7577a754d443..950d55dae948 100644 --- a/arch/arm64/include/asm/exception.h +++ b/arch/arm64/include/asm/exception.h @@ -32,7 +32,7 @@ static inline u32 disr_to_esr(u64 disr) } asmlinkage void enter_from_user_mode(void); -void do_mem_abort(unsigned long addr, unsigned int esr, struct pt_regs *regs); +void do_mem_abort(unsigned long far, unsigned int esr, struct pt_regs *regs); void do_undefinstr(struct pt_regs *regs); void do_bti(struct pt_regs *regs); asmlinkage void bad_mode(struct pt_regs *regs, int reason, unsigned int esr); diff --git a/arch/arm64/include/asm/traps.h b/arch/arm64/include/asm/traps.h index cee5928e1b7d..8e4f6c5b97af 100644 --- a/arch/arm64/include/asm/traps.h +++ b/arch/arm64/include/asm/traps.h @@ -26,8 +26,11 @@ void register_undef_hook(struct undef_hook *hook); void unregister_undef_hook(struct undef_hook *hook); void force_signal_inject(int signal, int code, unsigned long address); void arm64_notify_segfault(unsigned long addr); -void arm64_force_sig_fault(int signo, int code, void __user *addr, const char *str); -void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, const char *str); +void arm64_force_sig_fault(int signo, int code, void __user *addr, + unsigned long far, unsigned char far_tb_mask, + const char *str); +void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, + unsigned long far, const char *str); void arm64_force_sig_ptrace_errno_trap(int errno, void __user *addr, const char *str); /* diff --git a/arch/arm64/kernel/debug-monitors.c b/arch/arm64/kernel/debug-monitors.c index 5df49366e9ab..41776df532b4 100644 --- a/arch/arm64/kernel/debug-monitors.c +++ b/arch/arm64/kernel/debug-monitors.c @@ -232,8 +232,8 @@ static void send_user_sigtrap(int si_code) local_irq_enable(); arm64_force_sig_fault(SIGTRAP, si_code, - (void __user *)instruction_pointer(regs), - "User debug trap"); + (void __user *)instruction_pointer(regs), 0, 0, + "User debug trap"); } static int single_step_handler(unsigned long unused, unsigned int esr, diff --git a/arch/arm64/kernel/entry-common.c b/arch/arm64/kernel/entry-common.c index 3dbdf9752b11..59bbfe084043 100644 --- a/arch/arm64/kernel/entry-common.c +++ b/arch/arm64/kernel/entry-common.c @@ -22,7 +22,6 @@ static void notrace el1_abort(struct pt_regs *regs, unsigned long esr) unsigned long far = read_sysreg(far_el1); local_daif_inherit(regs); - far = untagged_addr(far); do_mem_abort(far, esr, regs); } NOKPROBE_SYMBOL(el1_abort); @@ -104,7 +103,6 @@ static void notrace el0_da(struct pt_regs *regs, unsigned long esr) user_exit_irqoff(); local_daif_restore(DAIF_PROCCTX); - far = untagged_addr(far); do_mem_abort(far, esr, regs); } NOKPROBE_SYMBOL(el0_da); diff --git a/arch/arm64/kernel/ptrace.c b/arch/arm64/kernel/ptrace.c index 68b7f34a08f5..f3bf39859e84 100644 --- a/arch/arm64/kernel/ptrace.c +++ b/arch/arm64/kernel/ptrace.c @@ -197,7 +197,7 @@ static void ptrace_hbptriggered(struct perf_event *bp, } #endif arm64_force_sig_fault(SIGTRAP, TRAP_HWBKPT, - (void __user *)(bkpt->trigger), + (void __user *)(bkpt->trigger), 0, 0, desc); } diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c index 47f651df781c..a8380a2b6361 100644 --- a/arch/arm64/kernel/traps.c +++ b/arch/arm64/kernel/traps.c @@ -235,20 +235,41 @@ static void arm64_show_signal(int signo, const char *str) } void arm64_force_sig_fault(int signo, int code, void __user *addr, + unsigned long far, unsigned char far_tb_mask, const char *str) { arm64_show_signal(signo, str); - if (signo == SIGKILL) + if (signo == SIGKILL) { force_sig(SIGKILL); - else - force_sig_fault(signo, code, addr); + } else { + struct kernel_siginfo info; + clear_siginfo(&info); + info.si_signo = signo; + info.si_errno = 0; + info.si_code = code; + info.si_addr = addr; + info.si_addr_top_byte = (far >> 56) & far_tb_mask; + info.si_addr_top_byte_mask = far_tb_mask; + force_sig_info(&info); + } } void arm64_force_sig_mceerr(int code, void __user *addr, short lsb, - const char *str) + unsigned long far, const char *str) { + struct kernel_siginfo info; + arm64_show_signal(SIGBUS, str); - force_sig_mceerr(code, addr, lsb); + + clear_siginfo(&info); + info.si_signo = SIGBUS; + info.si_errno = 0; + info.si_code = code; + info.si_addr = addr; + info.si_addr_lsb = lsb; + info.si_addr_top_byte = far >> 56; + info.si_addr_top_byte_mask = 0xff; + force_sig_info(&info); } void arm64_force_sig_ptrace_errno_trap(int errno, void __user *addr, @@ -267,7 +288,7 @@ void arm64_notify_die(const char *str, struct pt_regs *regs, current->thread.fault_address = 0; current->thread.fault_code = err; - arm64_force_sig_fault(signo, sicode, addr, str); + arm64_force_sig_fault(signo, sicode, addr, 0, 0, str); } else { die(str, regs, err); } @@ -829,7 +850,7 @@ void bad_el0_sync(struct pt_regs *regs, int reason, unsigned int esr) current->thread.fault_address = 0; current->thread.fault_code = esr; - arm64_force_sig_fault(SIGILL, ILL_ILLOPC, pc, + arm64_force_sig_fault(SIGILL, ILL_ILLOPC, pc, 0, 0, "Bad EL0 synchronous exception"); } diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c index 8afb238ff335..73748e419aa7 100644 --- a/arch/arm64/mm/fault.c +++ b/arch/arm64/mm/fault.c @@ -40,7 +40,7 @@ #include <asm/traps.h> struct fault_info { - int (*fn)(unsigned long addr, unsigned int esr, + int (*fn)(unsigned long far, unsigned int esr, struct pt_regs *regs); int sig; int code; @@ -383,8 +383,11 @@ static void set_thread_esr(unsigned long address, unsigned int esr) current->thread.fault_code = esr; } -static void do_bad_area(unsigned long addr, unsigned int esr, struct pt_regs *regs) +static void do_bad_area(unsigned long far, unsigned int esr, + struct pt_regs *regs) { + unsigned long addr = untagged_addr(far); + /* * If we are in kernel mode at this point, we have no context to * handle this fault with. @@ -394,7 +397,7 @@ static void do_bad_area(unsigned long addr, unsigned int esr, struct pt_regs *re set_thread_esr(addr, esr); arm64_force_sig_fault(inf->sig, inf->code, (void __user *)addr, - inf->name); + far, 0xff, inf->name); } else { __do_kernel_fault(addr, esr, regs); } @@ -445,7 +448,7 @@ static bool is_write_abort(unsigned int esr) return (esr & ESR_ELx_WNR) && !(esr & ESR_ELx_CM); } -static int __kprobes do_page_fault(unsigned long addr, unsigned int esr, +static int __kprobes do_page_fault(unsigned long far, unsigned int esr, struct pt_regs *regs) { const struct fault_info *inf; @@ -453,6 +456,7 @@ static int __kprobes do_page_fault(unsigned long addr, unsigned int esr, vm_fault_t fault, major = 0; unsigned long vm_flags = VM_ACCESS_FLAGS; unsigned int mm_flags = FAULT_FLAG_DEFAULT; + unsigned long addr = untagged_addr(far); if (kprobe_page_fault(regs, esr)) return 0; @@ -583,7 +587,7 @@ static int __kprobes do_page_fault(unsigned long addr, unsigned int esr, * this page fault. */ arm64_force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)addr, - inf->name); + far, 0xff, inf->name); } else if (fault & (VM_FAULT_HWPOISON_LARGE | VM_FAULT_HWPOISON)) { unsigned int lsb; @@ -592,7 +596,7 @@ static int __kprobes do_page_fault(unsigned long addr, unsigned int esr, lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault)); arm64_force_sig_mceerr(BUS_MCEERR_AR, (void __user *)addr, lsb, - inf->name); + far, inf->name); } else { /* * Something tried to access memory that isn't in our memory @@ -600,8 +604,7 @@ static int __kprobes do_page_fault(unsigned long addr, unsigned int esr, */ arm64_force_sig_fault(SIGSEGV, fault == VM_FAULT_BADACCESS ? SEGV_ACCERR : SEGV_MAPERR, - (void __user *)addr, - inf->name); + (void __user *)addr, far, 0xff, inf->name); } return 0; @@ -611,30 +614,32 @@ static int __kprobes do_page_fault(unsigned long addr, unsigned int esr, return 0; } -static int __kprobes do_translation_fault(unsigned long addr, +static int __kprobes do_translation_fault(unsigned long far, unsigned int esr, struct pt_regs *regs) { + unsigned long addr = untagged_addr(far); + if (is_ttbr0_addr(addr)) - return do_page_fault(addr, esr, regs); + return do_page_fault(far, esr, regs); - do_bad_area(addr, esr, regs); + do_bad_area(far, esr, regs); return 0; } -static int do_alignment_fault(unsigned long addr, unsigned int esr, +static int do_alignment_fault(unsigned long far, unsigned int esr, struct pt_regs *regs) { - do_bad_area(addr, esr, regs); + do_bad_area(far, esr, regs); return 0; } -static int do_bad(unsigned long addr, unsigned int esr, struct pt_regs *regs) +static int do_bad(unsigned long far, unsigned int esr, struct pt_regs *regs) { return 1; /* "fault" */ } -static int do_sea(unsigned long addr, unsigned int esr, struct pt_regs *regs) +static int do_sea(unsigned long far, unsigned int esr, struct pt_regs *regs) { const struct fault_info *inf; void __user *siaddr; @@ -652,7 +657,7 @@ static int do_sea(unsigned long addr, unsigned int esr, struct pt_regs *regs) if (esr & ESR_ELx_FnV) siaddr = NULL; else - siaddr = (void __user *)addr; + siaddr = (void __user *)untagged_addr(far); arm64_notify_die(inf->name, regs, inf->sig, inf->code, siaddr, esr); return 0; @@ -725,11 +730,12 @@ static const struct fault_info fault_info[] = { { do_bad, SIGKILL, SI_KERNEL, "unknown 63" }, }; -void do_mem_abort(unsigned long addr, unsigned int esr, struct pt_regs *regs) +void do_mem_abort(unsigned long far, unsigned int esr, struct pt_regs *regs) { const struct fault_info *inf = esr_to_fault_info(esr); + unsigned long addr = untagged_addr(far); - if (!inf->fn(addr, esr, regs)) + if (!inf->fn(far, esr, regs)) return; if (!user_mode(regs)) { @@ -738,8 +744,8 @@ void do_mem_abort(unsigned long addr, unsigned int esr, struct pt_regs *regs) show_pte(addr); } - arm64_notify_die(inf->name, regs, - inf->sig, inf->code, (void __user *)addr, esr); + arm64_notify_die(inf->name, regs, inf->sig, inf->code, + (void __user *)addr, esr); } NOKPROBE_SYMBOL(do_mem_abort); @@ -752,8 +758,8 @@ NOKPROBE_SYMBOL(do_el0_irq_bp_hardening); void do_sp_pc_abort(unsigned long addr, unsigned int esr, struct pt_regs *regs) { - arm64_notify_die("SP/PC alignment exception", regs, - SIGBUS, BUS_ADRALN, (void __user *)addr, esr); + arm64_notify_die("SP/PC alignment exception", regs, SIGBUS, BUS_ADRALN, + (void __user *)addr, esr); } NOKPROBE_SYMBOL(do_sp_pc_abort); @@ -879,8 +885,8 @@ void do_debug_exception(unsigned long addr_if_watchpoint, unsigned int esr, arm64_apply_bp_hardening(); if (inf->fn(addr_if_watchpoint, esr, regs)) { - arm64_notify_die(inf->name, regs, - inf->sig, inf->code, (void __user *)pc, esr); + arm64_notify_die(inf->name, regs, inf->sig, inf->code, + (void __user *)pc, esr); } debug_exception_exit(regs); diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h index cb3d6c267181..6dd82373eb2d 100644 --- a/include/uapi/asm-generic/siginfo.h +++ b/include/uapi/asm-generic/siginfo.h @@ -91,6 +91,14 @@ union __sifields { char _dummy_pkey[__ADDR_BND_PKEY_PAD]; __u32 _pkey; } _addr_pkey; +#ifdef __aarch64__ + /* used with all si_codes */ + struct { + short _dummy_top_byte; + unsigned char _top_byte; + unsigned char _top_byte_mask; + } _addr_top_byte; +#endif }; } _sigfault; @@ -148,6 +156,10 @@ typedef struct siginfo { #define si_int _sifields._rt._sigval.sival_int #define si_ptr _sifields._rt._sigval.sival_ptr #define si_addr _sifields._sigfault._addr +#ifdef __aarch64__ +#define si_addr_top_byte _sifields._sigfault._addr_top_byte._top_byte +#define si_addr_top_byte_mask _sifields._sigfault._addr_top_byte._top_byte_mask +#endif #ifdef __ARCH_SI_TRAPNO #define si_trapno _sifields._sigfault._trapno #endif
The kernel currently clears the tag bits (i.e. bits 56-63) in the fault address exposed via siginfo.si_addr and sigcontext.fault_address. However, the tag bits may be needed by tools in order to accurately diagnose memory errors, such as HWASan [1] or future tools based on the Memory Tagging Extension (MTE). We should not stop clearing these bits in the existing fault address fields, because there may be existing userspace applications that are expecting the tag bits to be cleared. Instead, create a new aarch64-specific union field in siginfo, and store the tag bits of FAR_EL1 there, together with a mask specifying which bits are valid. The new fields are laid out in a part of siginfo that is currently unused due to having previously been used for padding between si_addr_lsb and the union. Existing kernels will zero-initialize the padding, setting both fields to 0, which is a valid value for the fields. [1] http://clang.llvm.org/docs/HardwareAssistedAddressSanitizerDesign.html Signed-off-by: Peter Collingbourne <pcc@google.com> --- v8: - rebase onto 5.8rc2 v7: - switch to a new siginfo field instead of using sigcontext - merge the patch back into one since the other patches are now unnecessary v6: - move fault address and fault code into the kernel_siginfo data structure - split the patch in three since it was getting large and now has generic and arch-specific parts v5: - add padding to fault_addr_top_byte_context in order to ensure the correct size and preserve sp alignment v4: - expose only the tag bits in the context instead of the entire FAR_EL1 - remove mention of the new context from the sigcontext.__reserved[] note v3: - add documentation to tagged-pointers.rst - update comments in sigcontext.h v2: - revert changes to hw_breakpoint.c - rename set_thread_esr to set_thread_far_esr Documentation/arm64/tagged-pointers.rst | 17 +++++--- arch/arm64/include/asm/exception.h | 2 +- arch/arm64/include/asm/traps.h | 7 +++- arch/arm64/kernel/debug-monitors.c | 4 +- arch/arm64/kernel/entry-common.c | 2 - arch/arm64/kernel/ptrace.c | 2 +- arch/arm64/kernel/traps.c | 35 ++++++++++++---- arch/arm64/mm/fault.c | 54 ++++++++++++++----------- include/uapi/asm-generic/siginfo.h | 12 ++++++ 9 files changed, 90 insertions(+), 45 deletions(-)