Message ID | 20210416131820.2566571-1-aaronlewis@google.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | [1/2] kvm: x86: Allow userspace to handle emulation errors | expand |
Thanks for sending the patches. On Friday, 2021-04-16 at 06:18:19 -07, Aaron Lewis wrote: > Add a fallback mechanism to the in-kernel instruction emulator that > allows userspace the opportunity to process an instruction the emulator > was unable to. When the in-kernel instruction emulator fails to process > an instruction it will either inject a #UD into the guest or exit to > userspace with exit reason KVM_INTERNAL_ERROR. This is because it does > not know how to proceed in an appropriate manner. This feature lets > userspace get involved to see if it can figure out a better path > forward. Given that you are intending to try and handle the instruction in user-space, it seems a little odd to overload the KVM_EXIT_INTERNAL_ERROR/KVM_INTERNAL_ERROR_EMULATION exit reason/sub error. Why not add a new exit reason, particularly given that the caller has to enable the capability to get the relevant data? (It would also remove the need for the flag field and any mechanism for packing multiple bits of detail into the structure.) > > Signed-off-by: Aaron Lewis <aaronlewis@google.com> > Change-Id: If9876bc73d26f6c3ff9a8bce177c2fc6f160e629 > --- > Documentation/virt/kvm/api.rst | 16 ++++++++++++++++ > arch/x86/include/asm/kvm_host.h | 6 ++++++ > arch/x86/kvm/x86.c | 33 +++++++++++++++++++++++++++++---- > include/uapi/linux/kvm.h | 20 ++++++++++++++++++++ > tools/include/uapi/linux/kvm.h | 20 ++++++++++++++++++++ > 5 files changed, 91 insertions(+), 4 deletions(-) > > diff --git a/Documentation/virt/kvm/api.rst b/Documentation/virt/kvm/api.rst > index 307f2fcf1b02..f8278e893fbe 100644 > --- a/Documentation/virt/kvm/api.rst > +++ b/Documentation/virt/kvm/api.rst > @@ -6233,6 +6233,22 @@ KVM_RUN_BUS_LOCK flag is used to distinguish between them. > This capability can be used to check / enable 2nd DAWR feature provided > by POWER10 processor. > > +7.24 KVM_CAP_EXIT_ON_EMULATION_FAILURE > +-------------------------------------- > + > +:Architectures: x86 > +:Parameters: args[0] whether the feature should be enabled or not > + > +With this capability enabled, the in-kernel instruction emulator packs the exit > +struct of KVM_INTERNAL_ERROR with the instruction length and instruction bytes > +when an error occurs while emulating an instruction. This allows userspace to > +then take a look at the instruction and see if it is able to handle it more > +gracefully than the in-kernel emulator. > + > +When this capability is enabled use the emulation_failure struct instead of the > +internal struct for the exit struct. They have the same layout, but the > +emulation_failure struct matches the content better. > + > 8. Other capabilities. > ====================== > > diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h > index 3768819693e5..07235d08e976 100644 > --- a/arch/x86/include/asm/kvm_host.h > +++ b/arch/x86/include/asm/kvm_host.h > @@ -1049,6 +1049,12 @@ struct kvm_arch { > bool exception_payload_enabled; > > bool bus_lock_detection_enabled; > + /* > + * If exit_on_emulation_error is set, and the in-kernel instruction > + * emulator fails to emulate an instruction, allow userspace > + * the opportunity to look at it. > + */ > + bool exit_on_emulation_error; > > /* Deflect RDMSR and WRMSR to user space when they trigger a #GP */ > u32 user_space_msr_mask; > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c > index eca63625aee4..f9a207f815fb 100644 > --- a/arch/x86/kvm/x86.c > +++ b/arch/x86/kvm/x86.c > @@ -3771,6 +3771,7 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext) > case KVM_CAP_X86_USER_SPACE_MSR: > case KVM_CAP_X86_MSR_FILTER: > case KVM_CAP_ENFORCE_PV_FEATURE_CPUID: > + case KVM_CAP_EXIT_ON_EMULATION_FAILURE: > r = 1; > break; > #ifdef CONFIG_KVM_XEN > @@ -5357,6 +5358,10 @@ int kvm_vm_ioctl_enable_cap(struct kvm *kvm, > kvm->arch.bus_lock_detection_enabled = true; > r = 0; > break; > + case KVM_CAP_EXIT_ON_EMULATION_FAILURE: > + kvm->arch.exit_on_emulation_error = cap->args[0]; > + r = 0; > + break; > default: > r = -EINVAL; > break; > @@ -7119,8 +7124,29 @@ void kvm_inject_realmode_interrupt(struct kvm_vcpu *vcpu, int irq, int inc_eip) > } > EXPORT_SYMBOL_GPL(kvm_inject_realmode_interrupt); > > +static void prepare_emulation_failure_exit(struct kvm_vcpu *vcpu) > +{ > + struct x86_emulate_ctxt *ctxt = vcpu->arch.emulate_ctxt; > + u64 insn_size = ctxt->fetch.end - ctxt->fetch.data; > + struct kvm *kvm = vcpu->kvm; > + > + vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR; > + vcpu->run->emulation_failure.suberror = KVM_INTERNAL_ERROR_EMULATION; > + vcpu->run->emulation_failure.ndata = 0; > + if (kvm->arch.exit_on_emulation_error && insn_size > 0) { > + vcpu->run->emulation_failure.ndata = 3; > + vcpu->run->emulation_failure.flags = > + KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES; > + vcpu->run->emulation_failure.insn_size = insn_size; > + memcpy(vcpu->run->emulation_failure.insn_bytes, > + ctxt->fetch.data, sizeof(ctxt->fetch.data)); > + } > +} > + > static int handle_emulation_failure(struct kvm_vcpu *vcpu, int emulation_type) > { > + struct kvm *kvm = vcpu->kvm; > + > ++vcpu->stat.insn_emulation_fail; > trace_kvm_emulate_insn_failed(vcpu); > > @@ -7129,10 +7155,9 @@ static int handle_emulation_failure(struct kvm_vcpu *vcpu, int emulation_type) > return 1; > } > > - if (emulation_type & EMULTYPE_SKIP) { > - vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR; > - vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION; > - vcpu->run->internal.ndata = 0; > + if (kvm->arch.exit_on_emulation_error || > + (emulation_type & EMULTYPE_SKIP)) { > + prepare_emulation_failure_exit(vcpu); > return 0; > } > > diff --git a/include/uapi/linux/kvm.h b/include/uapi/linux/kvm.h > index f6afee209620..7c77099235b2 100644 > --- a/include/uapi/linux/kvm.h > +++ b/include/uapi/linux/kvm.h > @@ -279,6 +279,17 @@ struct kvm_xen_exit { > /* Encounter unexpected vm-exit reason */ > #define KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON 4 > > +/* > + * When using the suberror KVM_INTERNAL_ERROR_EMULATION, these flags are used > + * to describe what is contained in the exit struct. The flags are used to > + * describe it's contents, and the contents should be in ascending numerical > + * order of the flag values. For example, if the flag > + * KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES is set, the instruction > + * length and instruction bytes would be expected to show up first because this > + * flag has the lowest numerical value (1) of all the other flags. > + */ > +#define KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES (1ULL << 0) > + > /* for KVM_RUN, returned by mmap(vcpu_fd, offset=0) */ > struct kvm_run { > /* in */ > @@ -382,6 +393,14 @@ struct kvm_run { > __u32 ndata; > __u64 data[16]; > } internal; > + /* KVM_EXIT_INTERNAL_ERROR, too (not 2) */ > + struct { > + __u32 suberror; > + __u32 ndata; > + __u64 flags; > + __u8 insn_size; > + __u8 insn_bytes[15]; > + } emulation_failure; > /* KVM_EXIT_OSI */ > struct { > __u64 gprs[32]; > @@ -1078,6 +1097,7 @@ struct kvm_ppc_resize_hpt { > #define KVM_CAP_DIRTY_LOG_RING 192 > #define KVM_CAP_X86_BUS_LOCK_EXIT 193 > #define KVM_CAP_PPC_DAWR1 194 > +#define KVM_CAP_EXIT_ON_EMULATION_FAILURE 195 > > #ifdef KVM_CAP_IRQ_ROUTING > > diff --git a/tools/include/uapi/linux/kvm.h b/tools/include/uapi/linux/kvm.h > index f6afee209620..7c77099235b2 100644 > --- a/tools/include/uapi/linux/kvm.h > +++ b/tools/include/uapi/linux/kvm.h > @@ -279,6 +279,17 @@ struct kvm_xen_exit { > /* Encounter unexpected vm-exit reason */ > #define KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON 4 > > +/* > + * When using the suberror KVM_INTERNAL_ERROR_EMULATION, these flags are used > + * to describe what is contained in the exit struct. The flags are used to > + * describe it's contents, and the contents should be in ascending numerical > + * order of the flag values. For example, if the flag > + * KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES is set, the instruction > + * length and instruction bytes would be expected to show up first because this > + * flag has the lowest numerical value (1) of all the other flags. When adding a new flag, do I steal bytes from insn_bytes[] for my associated payload? If so, how many do I have to leave? > + */ > +#define KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES (1ULL << 0) > + > /* for KVM_RUN, returned by mmap(vcpu_fd, offset=0) */ > struct kvm_run { > /* in */ > @@ -382,6 +393,14 @@ struct kvm_run { > __u32 ndata; > __u64 data[16]; > } internal; > + /* KVM_EXIT_INTERNAL_ERROR, too (not 2) */ > + struct { > + __u32 suberror; > + __u32 ndata; > + __u64 flags; > + __u8 insn_size; > + __u8 insn_bytes[15]; > + } emulation_failure; > /* KVM_EXIT_OSI */ > struct { > __u64 gprs[32]; > @@ -1078,6 +1097,7 @@ struct kvm_ppc_resize_hpt { > #define KVM_CAP_DIRTY_LOG_RING 192 > #define KVM_CAP_X86_BUS_LOCK_EXIT 193 > #define KVM_CAP_PPC_DAWR1 194 > +#define KVM_CAP_EXIT_ON_EMULATION_FAILURE 195 > > #ifdef KVM_CAP_IRQ_ROUTING > > -- > 2.31.1.368.gbe11c130af-goog dme.
> > Add a fallback mechanism to the in-kernel instruction emulator that > > allows userspace the opportunity to process an instruction the emulator > > was unable to. When the in-kernel instruction emulator fails to process > > an instruction it will either inject a #UD into the guest or exit to > > userspace with exit reason KVM_INTERNAL_ERROR. This is because it does > > not know how to proceed in an appropriate manner. This feature lets > > userspace get involved to see if it can figure out a better path > > forward. > > Given that you are intending to try and handle the instruction in > user-space, it seems a little odd to overload the > KVM_EXIT_INTERNAL_ERROR/KVM_INTERNAL_ERROR_EMULATION exit reason/sub > error. > > Why not add a new exit reason, particularly given that the caller has to > enable the capability to get the relevant data? (It would also remove > the need for the flag field and any mechanism for packing multiple bits > of detail into the structure.) I considered that, but I opted for the extensibility of the exiting KVM_EXIT_INTERNAL_ERROR instead. To me it was six of one or half a dozen of the other. With either strategy I still wanted to provide for future extensibility, and had a flags field in place. That way we can add to this in the future if we find something that is missing (ie: potentially wanting a way to mark dirty pages, possibly passing a fault address, etc...) > > +/* > > + * When using the suberror KVM_INTERNAL_ERROR_EMULATION, these flags are used > > + * to describe what is contained in the exit struct. The flags are used to > > + * describe it's contents, and the contents should be in ascending numerical > > + * order of the flag values. For example, if the flag > > + * KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES is set, the instruction > > + * length and instruction bytes would be expected to show up first because this > > + * flag has the lowest numerical value (1) of all the other flags. > > + */ > > +#define KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES (1ULL << 0) > > + > > /* for KVM_RUN, returned by mmap(vcpu_fd, offset=0) */ > > struct kvm_run { > > /* in */ > > @@ -382,6 +393,14 @@ struct kvm_run { > > __u32 ndata; > > __u64 data[16]; > > } internal; > > + /* KVM_EXIT_INTERNAL_ERROR, too (not 2) */ > > + struct { > > + __u32 suberror; > > + __u32 ndata; > > + __u64 flags; > > + __u8 insn_size; > > + __u8 insn_bytes[15]; > > + } emulation_failure; > > +/* > > + * When using the suberror KVM_INTERNAL_ERROR_EMULATION, these flags are used > > + * to describe what is contained in the exit struct. The flags are used to > > + * describe it's contents, and the contents should be in ascending numerical > > + * order of the flag values. For example, if the flag > > + * KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES is set, the instruction > > + * length and instruction bytes would be expected to show up first because this > > + * flag has the lowest numerical value (1) of all the other flags. > > When adding a new flag, do I steal bytes from insn_bytes[] for my > associated payload? If so, how many do I have to leave? > The emulation_failure struct mirrors the internal struct, so if you are just adding to what I have, you can safely add up to 16 __u64's. I'm currently using the size equivalent to 3 of them (flags, insn_size, insn_bytes), so there should be plenty of space left for you to add what you need to the end. Just add the fields you need to the end of emulation_failure struct, increase 'ndata' to the new count, add a new flag to 'flags' so we know its contents.
On Monday, 2021-04-19 at 09:47:19 -07, Aaron Lewis wrote: >> > Add a fallback mechanism to the in-kernel instruction emulator that >> > allows userspace the opportunity to process an instruction the emulator >> > was unable to. When the in-kernel instruction emulator fails to process >> > an instruction it will either inject a #UD into the guest or exit to >> > userspace with exit reason KVM_INTERNAL_ERROR. This is because it does >> > not know how to proceed in an appropriate manner. This feature lets >> > userspace get involved to see if it can figure out a better path >> > forward. >> >> Given that you are intending to try and handle the instruction in >> user-space, it seems a little odd to overload the >> KVM_EXIT_INTERNAL_ERROR/KVM_INTERNAL_ERROR_EMULATION exit reason/sub >> error. >> >> Why not add a new exit reason, particularly given that the caller has to >> enable the capability to get the relevant data? (It would also remove >> the need for the flag field and any mechanism for packing multiple bits >> of detail into the structure.) > > I considered that, but I opted for the extensibility of the exiting > KVM_EXIT_INTERNAL_ERROR instead. To me it was six of one or half a > dozen of the other. With either strategy I still wanted to provide > for future extensibility, and had a flags field in place. That way we > can add to this in the future if we find something that is missing > (ie: potentially wanting a way to mark dirty pages, possibly passing a > fault address, etc...) How many of the flag based optional fields do you anticipate needing for any one particular exit scenario? If it's one, then using the flags to disambiguate the emulation failure cases after choosing to stuff all of the cases into KVM_EXIT_INTERNAL_ERROR / KVM_INTERNAL_ERROR_EMULATION would be odd. (I'm presuming that it's not one, but don't understand the use case.) >> > +/* >> > + * When using the suberror KVM_INTERNAL_ERROR_EMULATION, these flags are used >> > + * to describe what is contained in the exit struct. The flags are used to >> > + * describe it's contents, and the contents should be in ascending numerical >> > + * order of the flag values. For example, if the flag >> > + * KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES is set, the instruction >> > + * length and instruction bytes would be expected to show up first because this >> > + * flag has the lowest numerical value (1) of all the other flags. >> > + */ >> > +#define KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES (1ULL << 0) >> > + >> > /* for KVM_RUN, returned by mmap(vcpu_fd, offset=0) */ >> > struct kvm_run { >> > /* in */ >> > @@ -382,6 +393,14 @@ struct kvm_run { >> > __u32 ndata; >> > __u64 data[16]; >> > } internal; >> > + /* KVM_EXIT_INTERNAL_ERROR, too (not 2) */ >> > + struct { >> > + __u32 suberror; >> > + __u32 ndata; >> > + __u64 flags; >> > + __u8 insn_size; >> > + __u8 insn_bytes[15]; >> > + } emulation_failure; >> > +/* >> > + * When using the suberror KVM_INTERNAL_ERROR_EMULATION, these flags are used >> > + * to describe what is contained in the exit struct. The flags are used to >> > + * describe it's contents, and the contents should be in ascending numerical >> > + * order of the flag values. For example, if the flag >> > + * KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES is set, the instruction >> > + * length and instruction bytes would be expected to show up first because this >> > + * flag has the lowest numerical value (1) of all the other flags. >> >> When adding a new flag, do I steal bytes from insn_bytes[] for my >> associated payload? If so, how many do I have to leave? >> > > The emulation_failure struct mirrors the internal struct, so if you > are just adding to what I have, you can safely add up to 16 __u64's. > I'm currently using the size equivalent to 3 of them (flags, > insn_size, insn_bytes), so there should be plenty of space left for > you to add what you need to the end. Just add the fields you need to > the end of emulation_failure struct, increase 'ndata' to the new > count, add a new flag to 'flags' so we know its contents. My apologies, I mis-read the u8 as u64, so figured that you'd eaten all of the remaining space. dme.
On Tue, Apr 20, 2021 at 12:21 AM David Edmondson <dme@dme.org> wrote: > > On Monday, 2021-04-19 at 09:47:19 -07, Aaron Lewis wrote: > > >> > Add a fallback mechanism to the in-kernel instruction emulator that > >> > allows userspace the opportunity to process an instruction the emulator > >> > was unable to. When the in-kernel instruction emulator fails to process > >> > an instruction it will either inject a #UD into the guest or exit to > >> > userspace with exit reason KVM_INTERNAL_ERROR. This is because it does > >> > not know how to proceed in an appropriate manner. This feature lets > >> > userspace get involved to see if it can figure out a better path > >> > forward. > >> > >> Given that you are intending to try and handle the instruction in > >> user-space, it seems a little odd to overload the > >> KVM_EXIT_INTERNAL_ERROR/KVM_INTERNAL_ERROR_EMULATION exit reason/sub > >> error. > >> > >> Why not add a new exit reason, particularly given that the caller has to > >> enable the capability to get the relevant data? (It would also remove > >> the need for the flag field and any mechanism for packing multiple bits > >> of detail into the structure.) > > > > I considered that, but I opted for the extensibility of the exiting > > KVM_EXIT_INTERNAL_ERROR instead. To me it was six of one or half a > > dozen of the other. With either strategy I still wanted to provide > > for future extensibility, and had a flags field in place. That way we > > can add to this in the future if we find something that is missing > > (ie: potentially wanting a way to mark dirty pages, possibly passing a > > fault address, etc...) > > How many of the flag based optional fields do you anticipate needing for > any one particular exit scenario? > > If it's one, then using the flags to disambiguate the emulation failure > cases after choosing to stuff all of the cases into > KVM_EXIT_INTERNAL_ERROR / KVM_INTERNAL_ERROR_EMULATION would be odd. > > (I'm presuming that it's not one, but don't understand the use case.) > The motivation was to allow for maximum flexibility in the future, and not be tied down to something we potentially missed now. I agree the flags aren't needed if we are only adding to what's currently there, but they are needed if we want to remove something or pack something differently. I didn't see how I could achieve that without adding a flags field. Seemed like low overhead to be more future proof. > >> > +/* > >> > + * When using the suberror KVM_INTERNAL_ERROR_EMULATION, these flags are used > >> > + * to describe what is contained in the exit struct. The flags are used to > >> > + * describe it's contents, and the contents should be in ascending numerical > >> > + * order of the flag values. For example, if the flag > >> > + * KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES is set, the instruction > >> > + * length and instruction bytes would be expected to show up first because this > >> > + * flag has the lowest numerical value (1) of all the other flags. > >> > + */ > >> > +#define KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES (1ULL << 0) > >> > + > >> > /* for KVM_RUN, returned by mmap(vcpu_fd, offset=0) */ > >> > struct kvm_run { > >> > /* in */ > >> > @@ -382,6 +393,14 @@ struct kvm_run { > >> > __u32 ndata; > >> > __u64 data[16]; > >> > } internal; > >> > + /* KVM_EXIT_INTERNAL_ERROR, too (not 2) */ > >> > + struct { > >> > + __u32 suberror; > >> > + __u32 ndata; > >> > + __u64 flags; > >> > + __u8 insn_size; > >> > + __u8 insn_bytes[15]; > >> > + } emulation_failure; > >> > +/* > >> > + * When using the suberror KVM_INTERNAL_ERROR_EMULATION, these flags are used > >> > + * to describe what is contained in the exit struct. The flags are used to > >> > + * describe it's contents, and the contents should be in ascending numerical > >> > + * order of the flag values. For example, if the flag > >> > + * KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES is set, the instruction > >> > + * length and instruction bytes would be expected to show up first because this > >> > + * flag has the lowest numerical value (1) of all the other flags. > >> > >> When adding a new flag, do I steal bytes from insn_bytes[] for my > >> associated payload? If so, how many do I have to leave? > >> > > > > The emulation_failure struct mirrors the internal struct, so if you > > are just adding to what I have, you can safely add up to 16 __u64's. > > I'm currently using the size equivalent to 3 of them (flags, > > insn_size, insn_bytes), so there should be plenty of space left for > > you to add what you need to the end. Just add the fields you need to > > the end of emulation_failure struct, increase 'ndata' to the new > > count, add a new flag to 'flags' so we know its contents. > > My apologies, I mis-read the u8 as u64, so figured that you'd eaten all > of the remaining space. > > dme. > -- > I walk like a building, I never get wet.
On Tuesday, 2021-04-20 at 07:57:27 -07, Aaron Lewis wrote: >> >> Why not add a new exit reason, particularly given that the caller has to >> >> enable the capability to get the relevant data? (It would also remove >> >> the need for the flag field and any mechanism for packing multiple bits >> >> of detail into the structure.) >> > >> > I considered that, but I opted for the extensibility of the exiting >> > KVM_EXIT_INTERNAL_ERROR instead. To me it was six of one or half a >> > dozen of the other. With either strategy I still wanted to provide >> > for future extensibility, and had a flags field in place. That way we >> > can add to this in the future if we find something that is missing >> > (ie: potentially wanting a way to mark dirty pages, possibly passing a >> > fault address, etc...) >> >> How many of the flag based optional fields do you anticipate needing for >> any one particular exit scenario? >> >> If it's one, then using the flags to disambiguate the emulation failure >> cases after choosing to stuff all of the cases into >> KVM_EXIT_INTERNAL_ERROR / KVM_INTERNAL_ERROR_EMULATION would be odd. >> >> (I'm presuming that it's not one, but don't understand the use case.) >> > > The motivation was to allow for maximum flexibility in the future, and > not be tied down to something we potentially missed now. I agree the > flags aren't needed if we are only adding to what's currently there, > but they are needed if we want to remove something or pack something > differently. I didn't see how I could achieve that without adding a > flags field. Seemed like low overhead to be more future proof. With what you have now, the ndata field seems unnecessary - I should be able to determine the contents of the rest of the structure based on the flags. That also suggests to me that using something other than KVM_INTERNAL_ERROR_EMULATION would make sense. This comment: >> >> > + * When using the suberror KVM_INTERNAL_ERROR_EMULATION, these flags are used >> >> > + * to describe what is contained in the exit struct. The flags are used to >> >> > + * describe it's contents, and the contents should be in ascending numerical >> >> > + * order of the flag values. For example, if the flag >> >> > + * KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES is set, the instruction >> >> > + * length and instruction bytes would be expected to show up first because this >> >> > + * flag has the lowest numerical value (1) of all the other flags. originally made me think that the flag-indicated elements were going to be packed into the remaining space of the structure at a position depending on which flags are set. For example, if I add a new flag KVM_INTERNAL_ERROR_EMULATION_FLAG_EXIT_CODE, value 2, and then want to pass back an exit code but *not* instruction bytes, the comment appears to suggest that the exit code will appear immediately after the flags. This is contradicted by your other reply: >> > Just add the fields you need to >> > the end of emulation_failure struct, increase 'ndata' to the new >> > count, add a new flag to 'flags' so we know its contents. Given this, the ordering of flag values does not seem significant - the structure elements corresponding to a flag value will always be present, just not filled with relevant data. dme.
On Tue, Apr 20, 2021, David Edmondson wrote: > On Tuesday, 2021-04-20 at 07:57:27 -07, Aaron Lewis wrote: > > >> >> Why not add a new exit reason, particularly given that the caller has to > >> >> enable the capability to get the relevant data? (It would also remove > >> >> the need for the flag field and any mechanism for packing multiple bits > >> >> of detail into the structure.) > >> > > >> > I considered that, but I opted for the extensibility of the exiting > >> > KVM_EXIT_INTERNAL_ERROR instead. To me it was six of one or half a > >> > dozen of the other. With either strategy I still wanted to provide > >> > for future extensibility, and had a flags field in place. That way we > >> > can add to this in the future if we find something that is missing > >> > (ie: potentially wanting a way to mark dirty pages, possibly passing a > >> > fault address, etc...) > >> > >> How many of the flag based optional fields do you anticipate needing for > >> any one particular exit scenario? > >> > >> If it's one, then using the flags to disambiguate the emulation failure > >> cases after choosing to stuff all of the cases into > >> KVM_EXIT_INTERNAL_ERROR / KVM_INTERNAL_ERROR_EMULATION would be odd. > >> > >> (I'm presuming that it's not one, but don't understand the use case.) > > > > The motivation was to allow for maximum flexibility in the future, and > > not be tied down to something we potentially missed now. I agree the > > flags aren't needed if we are only adding to what's currently there, > > but they are needed if we want to remove something or pack something > > differently. I didn't see how I could achieve that without adding a > > flags field. Seemed like low overhead to be more future proof. > > With what you have now, the ndata field seems unnecessary - I should be > able to determine the contents of the rest of the structure based on the > flags. Keeping ndata is necessary if we piggyback KVM_INTERNAL_ERROR_EMULATION, otherwise we'll break for VMMs that are not aware of the new format. E.g. if ndata gets stuffed with a large number, KVM could cause a buffer overrun in an old VMM. > That also suggests to me that using something other than > KVM_INTERNAL_ERROR_EMULATION would make sense. Like Aaron, I'm on the fence as to whether or not a new exit reason is in order. On one hand, it would be slightly cleaner. On the other hand, the existing "KVM_INTERNAL_ERROR_EMULATION" really is the best name. It implies nothing about the userspace VMM, only that KVM attempted to emulate an instruction and failed. The other motivation is that KVM can opportunistically start dumping extra info for old VMMs, though this patch does not do that; feedback imminent. :-) > This comment: > > >> >> > + * When using the suberror KVM_INTERNAL_ERROR_EMULATION, these flags are used > >> >> > + * to describe what is contained in the exit struct. The flags are used to > >> >> > + * describe it's contents, and the contents should be in ascending numerical > >> >> > + * order of the flag values. For example, if the flag > >> >> > + * KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES is set, the instruction > >> >> > + * length and instruction bytes would be expected to show up first because this > >> >> > + * flag has the lowest numerical value (1) of all the other flags. > > originally made me think that the flag-indicated elements were going to > be packed into the remaining space of the structure at a position > depending on which flags are set. > > For example, if I add a new flag > KVM_INTERNAL_ERROR_EMULATION_FLAG_EXIT_CODE, value 2, and then want to > pass back an exit code but *not* instruction bytes, the comment appears > to suggest that the exit code will appear immediately after the flags. > > This is contradicted by your other reply: > > >> > Just add the fields you need to > >> > the end of emulation_failure struct, increase 'ndata' to the new > >> > count, add a new flag to 'flags' so we know its contents. > > Given this, the ordering of flag values does not seem significant - the > structure elements corresponding to a flag value will always be present, > just not filled with relevant data. I think what Aaron is trying to say is that the order in the aliased data[] is associated with the lowest _defined_ flag value, not the lowest _set_ flag. That said, I would just omit the "ascending numerical" stuff entirely, e.g. I think for the #defines, this will suffice: /* Flags that describe what fields in emulation_failure hold valid data */ As for not breaking userspace if/when additional fields are added, we can instead document the new struct (and drop my snarky comment :-D), e.g.: /* * KVM_INTERNAL_ERROR_EMULATION * * "struct emulation_failure" is an overlay of "struct internal" * that is used for the KVM_INTERNAL_ERROR_EMULATION sub-type of * KVM_EXIT_INTERNAL_ERROR. Note, unlike other internal error * sub-types, this struct is ABI! It also needs to be backwards * compabile with "struct internal". Take special care that * "ndata" is correct, that new fields are enumerated in "flags", * and that each flag enumerates fields that are 64-bit aligned * and sized (so that ndata+internal.data[] is valid/accurate). */ struct { __u32 suberror; __u32 ndata; __u64 flags; __u8 insn_size; __u8 insn_bytes[15]; } emulation_failure;
On Fri, Apr 16, 2021, Aaron Lewis wrote: > +7.24 KVM_CAP_EXIT_ON_EMULATION_FAILURE > +-------------------------------------- > + > +:Architectures: x86 > +:Parameters: args[0] whether the feature should be enabled or not > + > +With this capability enabled, the in-kernel instruction emulator packs the exit > +struct of KVM_INTERNAL_ERROR with the instruction length and instruction bytes > +when an error occurs while emulating an instruction. This allows userspace to > +then take a look at the instruction and see if it is able to handle it more > +gracefully than the in-kernel emulator. As alluded to later in the thread, I don't think we should condition the extra information on this capability. By crafting the struct overlay to be backwards compatibile, KVM can safely dump all the new information, even for old VMMs. An old VMM may not programmatically use the data, but I suspect most VMMs at least dump all info, e.g. QEMU does: if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) { int i; for (i = 0; i < run->internal.ndata; ++i) { fprintf(stderr, "extra data[%d]: %"PRIx64"\n", i, (uint64_t)run->internal.data[i]); } } This would be a way to feed more info to the poor sod that has to debug emulation failures :-) > + > +When this capability is enabled use the emulation_failure struct instead of the > +internal struct for the exit struct. They have the same layout, but the > +emulation_failure struct matches the content better. This documentation misses the arguably more important details of what exactly "EXIT_ON_EMULATION_FAILURE" means. E.g. it should call out the KVM still exits on certain types (skip) even if this capability is not enabled, and that KVM will _never_ exit if VMware #GP emulation fails. > + > 8. Other capabilities. > ====================== > @@ -7119,8 +7124,29 @@ void kvm_inject_realmode_interrupt(struct kvm_vcpu *vcpu, int irq, int inc_eip) ... > } > EXPORT_SYMBOL_GPL(kvm_inject_realmode_interrupt); > > +static void prepare_emulation_failure_exit(struct kvm_vcpu *vcpu) > +{ > + struct x86_emulate_ctxt *ctxt = vcpu->arch.emulate_ctxt; > + u64 insn_size = ctxt->fetch.end - ctxt->fetch.data; > + struct kvm *kvm = vcpu->kvm; > + > + vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR; Grab vcpu->run in a local variable. > + vcpu->run->emulation_failure.suberror = KVM_INTERNAL_ERROR_EMULATION; > + vcpu->run->emulation_failure.ndata = 0; > + if (kvm->arch.exit_on_emulation_error && insn_size > 0) { I definitely think this should not be conditioned on exit_on_emulation_error. No need for "> 0", it's an unsigned value. > + vcpu->run->emulation_failure.ndata = 3; > + vcpu->run->emulation_failure.flags = Flags needs to be zeroed when insn_size==0. And use |= for new flags so that, if we add new flags, the existing code doesn't need to be modified. > + KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES; > + vcpu->run->emulation_failure.insn_size = insn_size; > + memcpy(vcpu->run->emulation_failure.insn_bytes, > + ctxt->fetch.data, sizeof(ctxt->fetch.data)); Doesn't truly matter, but I think it's less confusing to copy over insn_size bytes. > + } ... struct kvm_run *kvm_run = vcpu->run; kvm_run->exit_reason = KVM_EXIT_INTERNAL_ERROR; kvm_run->emulation_failure.suberror = KVM_INTERNAL_ERROR_EMULATION; kvm_run->emulation_failure.ndata = 1; kvm_run->emulation_failure.flags = 0; if (insn_size) { kvm_run->emulation_failure.ndata = 3; kvm_run->emulation_failure.flags |= KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES; kvm_run->emulation_failure.insn_size = insn_size; memcpy(kvm_run->emulation_failure.insn_bytes, ctxt->fetch.data, insn_size); } > +}
On Tuesday, 2021-04-20 at 18:21:42 GMT, Sean Christopherson wrote: > On Tue, Apr 20, 2021, David Edmondson wrote: >> >> With what you have now, the ndata field seems unnecessary - I should be >> able to determine the contents of the rest of the structure based on the >> flags. > > Keeping ndata is necessary if we piggyback KVM_INTERNAL_ERROR_EMULATION, > otherwise we'll break for VMMs that are not aware of the new format. E.g. if > ndata gets stuffed with a large number, KVM could cause a buffer overrun in an > old VMM. Agreed. >> That also suggests to me that using something other than >> KVM_INTERNAL_ERROR_EMULATION would make sense. > > Like Aaron, I'm on the fence as to whether or not a new exit reason is in order. > On one hand, it would be slightly cleaner. On the other hand, the existing > "KVM_INTERNAL_ERROR_EMULATION" really is the best name. It implies nothing > about the userspace VMM, only that KVM attempted to emulate an instruction and > failed. > > The other motivation is that KVM can opportunistically start dumping extra info > for old VMMs, though this patch does not do that; feedback imminent. :-) It's nothing more than that the interface ends up feeling a little strange. With several flags added and some of the earlier flags unused, ndata ends up indicating the largest extent of the flag-indicated data, but the earlier elements of the structure are unused. Hence the question about how many flags we anticipate using simultaneously. (I'm not really arguing that we should be packing the stuff in and having to decode it, as that is also unpleasant.) >> This comment: >> >> >> >> > + * When using the suberror KVM_INTERNAL_ERROR_EMULATION, these flags are used >> >> >> > + * to describe what is contained in the exit struct. The flags are used to >> >> >> > + * describe it's contents, and the contents should be in ascending numerical >> >> >> > + * order of the flag values. For example, if the flag >> >> >> > + * KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES is set, the instruction >> >> >> > + * length and instruction bytes would be expected to show up first because this >> >> >> > + * flag has the lowest numerical value (1) of all the other flags. >> >> originally made me think that the flag-indicated elements were going to >> be packed into the remaining space of the structure at a position >> depending on which flags are set. >> >> For example, if I add a new flag >> KVM_INTERNAL_ERROR_EMULATION_FLAG_EXIT_CODE, value 2, and then want to >> pass back an exit code but *not* instruction bytes, the comment appears >> to suggest that the exit code will appear immediately after the flags. >> >> This is contradicted by your other reply: >> >> >> > Just add the fields you need to >> >> > the end of emulation_failure struct, increase 'ndata' to the new >> >> > count, add a new flag to 'flags' so we know its contents. >> >> Given this, the ordering of flag values does not seem significant - the >> structure elements corresponding to a flag value will always be present, >> just not filled with relevant data. > > I think what Aaron is trying to say is that the order in the aliased data[] is > associated with the lowest _defined_ flag value, not the lowest _set_ flag. > > That said, I would just omit the "ascending numerical" stuff entirely, e.g. I > think for the #defines, this will suffice: > > /* Flags that describe what fields in emulation_failure hold valid data */ Agreed. > As for not breaking userspace if/when additional fields are added, we can instead > document the new struct (and drop my snarky comment :-D), e.g.: > > /* > * KVM_INTERNAL_ERROR_EMULATION > * > * "struct emulation_failure" is an overlay of "struct internal" > * that is used for the KVM_INTERNAL_ERROR_EMULATION sub-type of > * KVM_EXIT_INTERNAL_ERROR. Note, unlike other internal error > * sub-types, this struct is ABI! It also needs to be backwards > * compabile with "struct internal". Take special care that > * "ndata" is correct, that new fields are enumerated in "flags", > * and that each flag enumerates fields that are 64-bit aligned > * and sized (so that ndata+internal.data[] is valid/accurate). > */ > struct { > __u32 suberror; > __u32 ndata; > __u64 flags; > __u8 insn_size; > __u8 insn_bytes[15]; > } emulation_failure; Looks good (even with the snark). dme.
On Tuesday, 2021-04-20 at 18:34:48 UTC, Sean Christopherson wrote: > On Fri, Apr 16, 2021, Aaron Lewis wrote: >> + KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES; >> + vcpu->run->emulation_failure.insn_size = insn_size; >> + memcpy(vcpu->run->emulation_failure.insn_bytes, >> + ctxt->fetch.data, sizeof(ctxt->fetch.data)); > > Doesn't truly matter, but I think it's less confusing to copy over insn_size > bytes. And zero out the rest? dme.
>> The other motivation is that KVM can opportunistically start dumping extra info >> for old VMMs, though this patch does not do that; feedback imminent. :-) > > It's nothing more than that the interface ends up feeling a little > strange. With several flags added and some of the earlier flags unused, > ndata ends up indicating the largest extent of the flag-indicated data, > but the earlier elements of the structure are unused. Hence the question > about how many flags we anticipate using simultaneously. > > (I'm not really arguing that we should be packing the stuff in and > having to decode it, as that is also unpleasant.) It's hard to say how many flags will be used, but I suspect it will be a small set (I mentioned previously a possibility of 2 more), so I like Sean's suggestion to keep it simple and just use the flags to indicate which fields are used, instead of trying to pack everything in tightly. On Wed, Apr 21, 2021 at 1:39 AM David Edmondson <dme@dme.org> wrote: > > On Tuesday, 2021-04-20 at 18:34:48 UTC, Sean Christopherson wrote: > > > On Fri, Apr 16, 2021, Aaron Lewis wrote: > >> + KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES; > >> + vcpu->run->emulation_failure.insn_size = insn_size; > >> + memcpy(vcpu->run->emulation_failure.insn_bytes, > >> + ctxt->fetch.data, sizeof(ctxt->fetch.data)); > > > > Doesn't truly matter, but I think it's less confusing to copy over insn_size > > bytes. > > And zero out the rest? > I left the memcpy alone as this seems like a more concervative / safer approach. I've had and have seen too many memcpy fails that my preference is to be conservative here. I could add a comment to clarify any confusion that this may bring by not using insn_size if that will help. Cheers, Aaron
On Wed, Apr 21, 2021 at 1:39 AM David Edmondson <dme@dme.org> wrote: > > On Tuesday, 2021-04-20 at 18:34:48 UTC, Sean Christopherson wrote: > > > On Fri, Apr 16, 2021, Aaron Lewis wrote: > >> + KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES; > >> + vcpu->run->emulation_failure.insn_size = insn_size; > >> + memcpy(vcpu->run->emulation_failure.insn_bytes, > >> + ctxt->fetch.data, sizeof(ctxt->fetch.data)); > > > > Doesn't truly matter, but I think it's less confusing to copy over insn_size > > bytes. > And zero out the rest? Why zero? Since we're talking about an instruction stream, wouldn't 0x90 make more sense than zero?
On Tue, Apr 20, 2021 at 11:34 AM Sean Christopherson <seanjc@google.com> wrote: > > On Fri, Apr 16, 2021, Aaron Lewis wrote: ... > > + vcpu->run->emulation_failure.suberror = KVM_INTERNAL_ERROR_EMULATION; > > + vcpu->run->emulation_failure.ndata = 0; > > + if (kvm->arch.exit_on_emulation_error && insn_size > 0) { > > I definitely think this should not be conditioned on exit_on_emulation_error. > > No need for "> 0", it's an unsigned value. Unsigned doesn't imply non-zero. If insn_size is 0, then something is wrong. ... > > + KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES; > > + vcpu->run->emulation_failure.insn_size = insn_size; > > + memcpy(vcpu->run->emulation_failure.insn_bytes, > > + ctxt->fetch.data, sizeof(ctxt->fetch.data)); > > Doesn't truly matter, but I think it's less confusing to copy over insn_size > bytes. Are you convinced that insn_size is always less than or equal to sizeof(ctxt->fetch.data)? I'm not. It shouldn't be, of course, but perhaps we should confirm that before copying insn_size bytes.
On Wednesday, 2021-04-21 at 09:26:34 -07, Jim Mattson wrote: > On Wed, Apr 21, 2021 at 1:39 AM David Edmondson <dme@dme.org> wrote: >> >> On Tuesday, 2021-04-20 at 18:34:48 UTC, Sean Christopherson wrote: >> >> > On Fri, Apr 16, 2021, Aaron Lewis wrote: >> >> + KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES; >> >> + vcpu->run->emulation_failure.insn_size = insn_size; >> >> + memcpy(vcpu->run->emulation_failure.insn_bytes, >> >> + ctxt->fetch.data, sizeof(ctxt->fetch.data)); >> > >> > Doesn't truly matter, but I think it's less confusing to copy over insn_size >> > bytes. > >> And zero out the rest? > > Why zero? Since we're talking about an instruction stream, wouldn't > 0x90 make more sense than zero? I'm not sure if you are serious or not. Zero-ing out the rest was intended to be to avoid leaking any previous emulated instruction stream. If the user-level code wants to start looking for instructions after insn_bytes[insn_size], they get what they deserve. dme.
On Wed, Apr 21, 2021 at 10:01 AM David Edmondson <dme@dme.org> wrote: > > On Wednesday, 2021-04-21 at 09:26:34 -07, Jim Mattson wrote: > > > On Wed, Apr 21, 2021 at 1:39 AM David Edmondson <dme@dme.org> wrote: > >> > >> On Tuesday, 2021-04-20 at 18:34:48 UTC, Sean Christopherson wrote: > >> > >> > On Fri, Apr 16, 2021, Aaron Lewis wrote: > >> >> + KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES; > >> >> + vcpu->run->emulation_failure.insn_size = insn_size; > >> >> + memcpy(vcpu->run->emulation_failure.insn_bytes, > >> >> + ctxt->fetch.data, sizeof(ctxt->fetch.data)); > >> > > >> > Doesn't truly matter, but I think it's less confusing to copy over insn_size > >> > bytes. > > > >> And zero out the rest? > > > > Why zero? Since we're talking about an instruction stream, wouldn't > > 0x90 make more sense than zero? > > I'm not sure if you are serious or not. > > Zero-ing out the rest was intended to be to avoid leaking any previous > emulated instruction stream. If the user-level code wants to start > looking for instructions after insn_bytes[insn_size], they get what they > deserve. If we limit the copy to insn_size bytes, the only leak is what was already present in the vcpu->run structure before this, which may or may not be a previous instruction stream, and which has been readable by userspace all along.
diff --git a/Documentation/virt/kvm/api.rst b/Documentation/virt/kvm/api.rst index 307f2fcf1b02..f8278e893fbe 100644 --- a/Documentation/virt/kvm/api.rst +++ b/Documentation/virt/kvm/api.rst @@ -6233,6 +6233,22 @@ KVM_RUN_BUS_LOCK flag is used to distinguish between them. This capability can be used to check / enable 2nd DAWR feature provided by POWER10 processor. +7.24 KVM_CAP_EXIT_ON_EMULATION_FAILURE +-------------------------------------- + +:Architectures: x86 +:Parameters: args[0] whether the feature should be enabled or not + +With this capability enabled, the in-kernel instruction emulator packs the exit +struct of KVM_INTERNAL_ERROR with the instruction length and instruction bytes +when an error occurs while emulating an instruction. This allows userspace to +then take a look at the instruction and see if it is able to handle it more +gracefully than the in-kernel emulator. + +When this capability is enabled use the emulation_failure struct instead of the +internal struct for the exit struct. They have the same layout, but the +emulation_failure struct matches the content better. + 8. Other capabilities. ====================== diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h index 3768819693e5..07235d08e976 100644 --- a/arch/x86/include/asm/kvm_host.h +++ b/arch/x86/include/asm/kvm_host.h @@ -1049,6 +1049,12 @@ struct kvm_arch { bool exception_payload_enabled; bool bus_lock_detection_enabled; + /* + * If exit_on_emulation_error is set, and the in-kernel instruction + * emulator fails to emulate an instruction, allow userspace + * the opportunity to look at it. + */ + bool exit_on_emulation_error; /* Deflect RDMSR and WRMSR to user space when they trigger a #GP */ u32 user_space_msr_mask; diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c index eca63625aee4..f9a207f815fb 100644 --- a/arch/x86/kvm/x86.c +++ b/arch/x86/kvm/x86.c @@ -3771,6 +3771,7 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext) case KVM_CAP_X86_USER_SPACE_MSR: case KVM_CAP_X86_MSR_FILTER: case KVM_CAP_ENFORCE_PV_FEATURE_CPUID: + case KVM_CAP_EXIT_ON_EMULATION_FAILURE: r = 1; break; #ifdef CONFIG_KVM_XEN @@ -5357,6 +5358,10 @@ int kvm_vm_ioctl_enable_cap(struct kvm *kvm, kvm->arch.bus_lock_detection_enabled = true; r = 0; break; + case KVM_CAP_EXIT_ON_EMULATION_FAILURE: + kvm->arch.exit_on_emulation_error = cap->args[0]; + r = 0; + break; default: r = -EINVAL; break; @@ -7119,8 +7124,29 @@ void kvm_inject_realmode_interrupt(struct kvm_vcpu *vcpu, int irq, int inc_eip) } EXPORT_SYMBOL_GPL(kvm_inject_realmode_interrupt); +static void prepare_emulation_failure_exit(struct kvm_vcpu *vcpu) +{ + struct x86_emulate_ctxt *ctxt = vcpu->arch.emulate_ctxt; + u64 insn_size = ctxt->fetch.end - ctxt->fetch.data; + struct kvm *kvm = vcpu->kvm; + + vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR; + vcpu->run->emulation_failure.suberror = KVM_INTERNAL_ERROR_EMULATION; + vcpu->run->emulation_failure.ndata = 0; + if (kvm->arch.exit_on_emulation_error && insn_size > 0) { + vcpu->run->emulation_failure.ndata = 3; + vcpu->run->emulation_failure.flags = + KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES; + vcpu->run->emulation_failure.insn_size = insn_size; + memcpy(vcpu->run->emulation_failure.insn_bytes, + ctxt->fetch.data, sizeof(ctxt->fetch.data)); + } +} + static int handle_emulation_failure(struct kvm_vcpu *vcpu, int emulation_type) { + struct kvm *kvm = vcpu->kvm; + ++vcpu->stat.insn_emulation_fail; trace_kvm_emulate_insn_failed(vcpu); @@ -7129,10 +7155,9 @@ static int handle_emulation_failure(struct kvm_vcpu *vcpu, int emulation_type) return 1; } - if (emulation_type & EMULTYPE_SKIP) { - vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR; - vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION; - vcpu->run->internal.ndata = 0; + if (kvm->arch.exit_on_emulation_error || + (emulation_type & EMULTYPE_SKIP)) { + prepare_emulation_failure_exit(vcpu); return 0; } diff --git a/include/uapi/linux/kvm.h b/include/uapi/linux/kvm.h index f6afee209620..7c77099235b2 100644 --- a/include/uapi/linux/kvm.h +++ b/include/uapi/linux/kvm.h @@ -279,6 +279,17 @@ struct kvm_xen_exit { /* Encounter unexpected vm-exit reason */ #define KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON 4 +/* + * When using the suberror KVM_INTERNAL_ERROR_EMULATION, these flags are used + * to describe what is contained in the exit struct. The flags are used to + * describe it's contents, and the contents should be in ascending numerical + * order of the flag values. For example, if the flag + * KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES is set, the instruction + * length and instruction bytes would be expected to show up first because this + * flag has the lowest numerical value (1) of all the other flags. + */ +#define KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES (1ULL << 0) + /* for KVM_RUN, returned by mmap(vcpu_fd, offset=0) */ struct kvm_run { /* in */ @@ -382,6 +393,14 @@ struct kvm_run { __u32 ndata; __u64 data[16]; } internal; + /* KVM_EXIT_INTERNAL_ERROR, too (not 2) */ + struct { + __u32 suberror; + __u32 ndata; + __u64 flags; + __u8 insn_size; + __u8 insn_bytes[15]; + } emulation_failure; /* KVM_EXIT_OSI */ struct { __u64 gprs[32]; @@ -1078,6 +1097,7 @@ struct kvm_ppc_resize_hpt { #define KVM_CAP_DIRTY_LOG_RING 192 #define KVM_CAP_X86_BUS_LOCK_EXIT 193 #define KVM_CAP_PPC_DAWR1 194 +#define KVM_CAP_EXIT_ON_EMULATION_FAILURE 195 #ifdef KVM_CAP_IRQ_ROUTING diff --git a/tools/include/uapi/linux/kvm.h b/tools/include/uapi/linux/kvm.h index f6afee209620..7c77099235b2 100644 --- a/tools/include/uapi/linux/kvm.h +++ b/tools/include/uapi/linux/kvm.h @@ -279,6 +279,17 @@ struct kvm_xen_exit { /* Encounter unexpected vm-exit reason */ #define KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON 4 +/* + * When using the suberror KVM_INTERNAL_ERROR_EMULATION, these flags are used + * to describe what is contained in the exit struct. The flags are used to + * describe it's contents, and the contents should be in ascending numerical + * order of the flag values. For example, if the flag + * KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES is set, the instruction + * length and instruction bytes would be expected to show up first because this + * flag has the lowest numerical value (1) of all the other flags. + */ +#define KVM_INTERNAL_ERROR_EMULATION_FLAG_INSTRUCTION_BYTES (1ULL << 0) + /* for KVM_RUN, returned by mmap(vcpu_fd, offset=0) */ struct kvm_run { /* in */ @@ -382,6 +393,14 @@ struct kvm_run { __u32 ndata; __u64 data[16]; } internal; + /* KVM_EXIT_INTERNAL_ERROR, too (not 2) */ + struct { + __u32 suberror; + __u32 ndata; + __u64 flags; + __u8 insn_size; + __u8 insn_bytes[15]; + } emulation_failure; /* KVM_EXIT_OSI */ struct { __u64 gprs[32]; @@ -1078,6 +1097,7 @@ struct kvm_ppc_resize_hpt { #define KVM_CAP_DIRTY_LOG_RING 192 #define KVM_CAP_X86_BUS_LOCK_EXIT 193 #define KVM_CAP_PPC_DAWR1 194 +#define KVM_CAP_EXIT_ON_EMULATION_FAILURE 195 #ifdef KVM_CAP_IRQ_ROUTING
Add a fallback mechanism to the in-kernel instruction emulator that allows userspace the opportunity to process an instruction the emulator was unable to. When the in-kernel instruction emulator fails to process an instruction it will either inject a #UD into the guest or exit to userspace with exit reason KVM_INTERNAL_ERROR. This is because it does not know how to proceed in an appropriate manner. This feature lets userspace get involved to see if it can figure out a better path forward. Signed-off-by: Aaron Lewis <aaronlewis@google.com> Change-Id: If9876bc73d26f6c3ff9a8bce177c2fc6f160e629 --- Documentation/virt/kvm/api.rst | 16 ++++++++++++++++ arch/x86/include/asm/kvm_host.h | 6 ++++++ arch/x86/kvm/x86.c | 33 +++++++++++++++++++++++++++++---- include/uapi/linux/kvm.h | 20 ++++++++++++++++++++ tools/include/uapi/linux/kvm.h | 20 ++++++++++++++++++++ 5 files changed, 91 insertions(+), 4 deletions(-)