| Message ID | 20170405192120.1009-1-d-gerlach@ti.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Russell, On 04/05/2017 02:21 PM, Dave Gerlach wrote: > Currently the sram-exec functionality, which allows allocation of > executable memory and provides an API to move code to it, is only > selected in configs for the ARM architecture. Based on commit > 5756e9dd0de6 ("ARM: 6640/1: Thumb-2: Symbol manipulation macros for > function body copying") simply copying a C function pointer address > using memcpy without consideration of alignment and Thumb is unsafe on > ARM platforms. > > The aforementioned patch introduces the fncpy macro which is a safe way > to copy executable code on ARM platforms, so let's make use of that here > rather than the unsafe plain memcpy that was previously used by > sram_exec_copy. > > In the future, architectures hoping to make use of the sram-exec > functionality must define an fncpy macro just as ARM has done to > guarantee or check for safe copying to executable memory before allowing > the arch to select CONFIG_SRAM_EXEC. > > Signed-off-by: Dave Gerlach <d-gerlach@ti.com> > --- > drivers/misc/sram-exec.c | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/drivers/misc/sram-exec.c b/drivers/misc/sram-exec.c > index ac522417c462..0057eabe5c03 100644 > --- a/drivers/misc/sram-exec.c > +++ b/drivers/misc/sram-exec.c > @@ -19,6 +19,7 @@ > #include <linux/sram.h> > > #include <asm/cacheflush.h> > +#include <asm/fncpy.h> > > #include "sram.h" > > @@ -93,7 +94,7 @@ int sram_exec_copy(struct gen_pool *pool, void *dst, void *src, > set_memory_nx((unsigned long)base, pages); > set_memory_rw((unsigned long)base, pages); > > - memcpy(dst, src, size); > + fncpy(dst, src, size); > > set_memory_ro((unsigned long)base, pages); > set_memory_x((unsigned long)base, pages); > Does this address your concerns from here [1]? Because the only user of this code is ARM right now I already only build the sram-exec code in if CONFIG_ARM is selected. I originally split the sram-exec code into its own file as it already depends on the changes you made to set_memory_* APIs for ARM which we have a hard dependency on here, and not all platforms support this. So this allowed me to constrain the sram_exec code to platforms with the proper set_memory_* APIs defined, but also now this lets us directly use the fncpy macro in this driver. For future platforms that want to make use of sram_exec we set the constraint that an arch must: * Support the required set_memory_* APIs * Define a fncpy macro that guarantees safe movement of a function. This seems reasonable to me and gives support for ARM right away with a path forward for additional architectures to support sram_exec. Regards, Dave [1] https://www.spinics.net/lists/arm-kernel/msg574481.html
On Wed, Apr 05, 2017 at 02:22:33PM -0500, Dave Gerlach wrote: > Russell, > On 04/05/2017 02:21 PM, Dave Gerlach wrote: > >Currently the sram-exec functionality, which allows allocation of > >executable memory and provides an API to move code to it, is only > >selected in configs for the ARM architecture. Based on commit > >5756e9dd0de6 ("ARM: 6640/1: Thumb-2: Symbol manipulation macros for > >function body copying") simply copying a C function pointer address > >using memcpy without consideration of alignment and Thumb is unsafe on > >ARM platforms. > > > >The aforementioned patch introduces the fncpy macro which is a safe way > >to copy executable code on ARM platforms, so let's make use of that here > >rather than the unsafe plain memcpy that was previously used by > >sram_exec_copy. > > > >In the future, architectures hoping to make use of the sram-exec > >functionality must define an fncpy macro just as ARM has done to > >guarantee or check for safe copying to executable memory before allowing > >the arch to select CONFIG_SRAM_EXEC. > > > >Signed-off-by: Dave Gerlach <d-gerlach@ti.com> > >--- > > drivers/misc/sram-exec.c | 3 ++- > > 1 file changed, 2 insertions(+), 1 deletion(-) > > > >diff --git a/drivers/misc/sram-exec.c b/drivers/misc/sram-exec.c > >index ac522417c462..0057eabe5c03 100644 > >--- a/drivers/misc/sram-exec.c > >+++ b/drivers/misc/sram-exec.c > >@@ -19,6 +19,7 @@ > > #include <linux/sram.h> > > > > #include <asm/cacheflush.h> > >+#include <asm/fncpy.h> > > > > #include "sram.h" > > > >@@ -93,7 +94,7 @@ int sram_exec_copy(struct gen_pool *pool, void *dst, void *src, > > set_memory_nx((unsigned long)base, pages); > > set_memory_rw((unsigned long)base, pages); > > > >- memcpy(dst, src, size); > >+ fncpy(dst, src, size); > > > > set_memory_ro((unsigned long)base, pages); > > set_memory_x((unsigned long)base, pages); > > > > Does this address your concerns from here [1]? Because the only user of this > code is ARM right now I already only build the sram-exec code in if > CONFIG_ARM is selected. Sorry, it does not. Please read the comments in asm/fncpy.h. Deviating from the proscribed usage means your code is, quite simply, buggy. There's no two ways about that.
On 04/06/2017 02:07 PM, Russell King - ARM Linux wrote: > On Wed, Apr 05, 2017 at 02:22:33PM -0500, Dave Gerlach wrote: >> Russell, >> On 04/05/2017 02:21 PM, Dave Gerlach wrote: >>> Currently the sram-exec functionality, which allows allocation of >>> executable memory and provides an API to move code to it, is only >>> selected in configs for the ARM architecture. Based on commit >>> 5756e9dd0de6 ("ARM: 6640/1: Thumb-2: Symbol manipulation macros for >>> function body copying") simply copying a C function pointer address >>> using memcpy without consideration of alignment and Thumb is unsafe on >>> ARM platforms. >>> >>> The aforementioned patch introduces the fncpy macro which is a safe way >>> to copy executable code on ARM platforms, so let's make use of that here >>> rather than the unsafe plain memcpy that was previously used by >>> sram_exec_copy. >>> >>> In the future, architectures hoping to make use of the sram-exec >>> functionality must define an fncpy macro just as ARM has done to >>> guarantee or check for safe copying to executable memory before allowing >>> the arch to select CONFIG_SRAM_EXEC. >>> >>> Signed-off-by: Dave Gerlach <d-gerlach@ti.com> >>> --- >>> drivers/misc/sram-exec.c | 3 ++- >>> 1 file changed, 2 insertions(+), 1 deletion(-) >>> >>> diff --git a/drivers/misc/sram-exec.c b/drivers/misc/sram-exec.c >>> index ac522417c462..0057eabe5c03 100644 >>> --- a/drivers/misc/sram-exec.c >>> +++ b/drivers/misc/sram-exec.c >>> @@ -19,6 +19,7 @@ >>> #include <linux/sram.h> >>> >>> #include <asm/cacheflush.h> >>> +#include <asm/fncpy.h> >>> >>> #include "sram.h" >>> >>> @@ -93,7 +94,7 @@ int sram_exec_copy(struct gen_pool *pool, void *dst, void *src, >>> set_memory_nx((unsigned long)base, pages); >>> set_memory_rw((unsigned long)base, pages); >>> >>> - memcpy(dst, src, size); >>> + fncpy(dst, src, size); >>> >>> set_memory_ro((unsigned long)base, pages); >>> set_memory_x((unsigned long)base, pages); >>> >> >> Does this address your concerns from here [1]? Because the only user of this >> code is ARM right now I already only build the sram-exec code in if >> CONFIG_ARM is selected. > > Sorry, it does not. Please read the comments in asm/fncpy.h. > > Deviating from the proscribed usage means your code is, quite simply, > buggy. There's no two ways about that. > I understand there are many constraints to using fncpy, as this is what we used before to copy our executable code. Apart from users being aware of what these constraints are (8-byte aligned, position independent) and making sure the code they are moving meets them, are you saying we need some sort of additional strict enforcement of them? Because fncpy today will throw a bug if you fail to align src and dst properly, so adding another check will just double the messages to the user. Regards, Dave
On Thu, Apr 06, 2017 at 02:14:12PM -0500, Dave Gerlach wrote: > On 04/06/2017 02:07 PM, Russell King - ARM Linux wrote: > >On Wed, Apr 05, 2017 at 02:22:33PM -0500, Dave Gerlach wrote: > >>Russell, > >>On 04/05/2017 02:21 PM, Dave Gerlach wrote: > >>>Currently the sram-exec functionality, which allows allocation of > >>>executable memory and provides an API to move code to it, is only > >>>selected in configs for the ARM architecture. Based on commit > >>>5756e9dd0de6 ("ARM: 6640/1: Thumb-2: Symbol manipulation macros for > >>>function body copying") simply copying a C function pointer address > >>>using memcpy without consideration of alignment and Thumb is unsafe on > >>>ARM platforms. > >>> > >>>The aforementioned patch introduces the fncpy macro which is a safe way > >>>to copy executable code on ARM platforms, so let's make use of that here > >>>rather than the unsafe plain memcpy that was previously used by > >>>sram_exec_copy. > >>> > >>>In the future, architectures hoping to make use of the sram-exec > >>>functionality must define an fncpy macro just as ARM has done to > >>>guarantee or check for safe copying to executable memory before allowing > >>>the arch to select CONFIG_SRAM_EXEC. > >>> > >>>Signed-off-by: Dave Gerlach <d-gerlach@ti.com> > >>>--- > >>>drivers/misc/sram-exec.c | 3 ++- > >>>1 file changed, 2 insertions(+), 1 deletion(-) > >>> > >>>diff --git a/drivers/misc/sram-exec.c b/drivers/misc/sram-exec.c > >>>index ac522417c462..0057eabe5c03 100644 > >>>--- a/drivers/misc/sram-exec.c > >>>+++ b/drivers/misc/sram-exec.c > >>>@@ -19,6 +19,7 @@ > >>>#include <linux/sram.h> > >>> > >>>#include <asm/cacheflush.h> > >>>+#include <asm/fncpy.h> > >>> > >>>#include "sram.h" > >>> > >>>@@ -93,7 +94,7 @@ int sram_exec_copy(struct gen_pool *pool, void *dst, void *src, > >>> set_memory_nx((unsigned long)base, pages); > >>> set_memory_rw((unsigned long)base, pages); > >>> > >>>- memcpy(dst, src, size); > >>>+ fncpy(dst, src, size); > >>> > >>> set_memory_ro((unsigned long)base, pages); > >>> set_memory_x((unsigned long)base, pages); > >>> > >> > >>Does this address your concerns from here [1]? Because the only user of this > >>code is ARM right now I already only build the sram-exec code in if > >>CONFIG_ARM is selected. > > > >Sorry, it does not. Please read the comments in asm/fncpy.h. > > > >Deviating from the proscribed usage means your code is, quite simply, > >buggy. There's no two ways about that. > > > > I understand there are many constraints to using fncpy, as this is what we > used before to copy our executable code. Apart from users being aware of > what these constraints are (8-byte aligned, position independent) and making > sure the code they are moving meets them, are you saying we need some sort > of additional strict enforcement of them? Because fncpy today will throw a > bug if you fail to align src and dst properly, so adding another check will > just double the messages to the user. Yes, fncpy() will throw a bug, but as I've already explained: sram = alloc(); sram_func = fncpy(sram, func, func_size); sram_func(); is the _only_ valid usage. You must not do: sram = alloc(); fncpy(sram, func, func_size); sram(); because that will not work with Thumb code. The only permitted usage is as per the first example above, everything else is buggy.
On 04/06/2017 02:29 PM, Russell King - ARM Linux wrote: > On Thu, Apr 06, 2017 at 02:14:12PM -0500, Dave Gerlach wrote: >> On 04/06/2017 02:07 PM, Russell King - ARM Linux wrote: >>> On Wed, Apr 05, 2017 at 02:22:33PM -0500, Dave Gerlach wrote: >>>> Russell, >>>> On 04/05/2017 02:21 PM, Dave Gerlach wrote: >>>>> Currently the sram-exec functionality, which allows allocation of >>>>> executable memory and provides an API to move code to it, is only >>>>> selected in configs for the ARM architecture. Based on commit >>>>> 5756e9dd0de6 ("ARM: 6640/1: Thumb-2: Symbol manipulation macros for >>>>> function body copying") simply copying a C function pointer address >>>>> using memcpy without consideration of alignment and Thumb is unsafe on >>>>> ARM platforms. >>>>> >>>>> The aforementioned patch introduces the fncpy macro which is a safe way >>>>> to copy executable code on ARM platforms, so let's make use of that here >>>>> rather than the unsafe plain memcpy that was previously used by >>>>> sram_exec_copy. >>>>> >>>>> In the future, architectures hoping to make use of the sram-exec >>>>> functionality must define an fncpy macro just as ARM has done to >>>>> guarantee or check for safe copying to executable memory before allowing >>>>> the arch to select CONFIG_SRAM_EXEC. >>>>> >>>>> Signed-off-by: Dave Gerlach <d-gerlach@ti.com> >>>>> --- >>>>> drivers/misc/sram-exec.c | 3 ++- >>>>> 1 file changed, 2 insertions(+), 1 deletion(-) >>>>> >>>>> diff --git a/drivers/misc/sram-exec.c b/drivers/misc/sram-exec.c >>>>> index ac522417c462..0057eabe5c03 100644 >>>>> --- a/drivers/misc/sram-exec.c >>>>> +++ b/drivers/misc/sram-exec.c >>>>> @@ -19,6 +19,7 @@ >>>>> #include <linux/sram.h> >>>>> >>>>> #include <asm/cacheflush.h> >>>>> +#include <asm/fncpy.h> >>>>> >>>>> #include "sram.h" >>>>> >>>>> @@ -93,7 +94,7 @@ int sram_exec_copy(struct gen_pool *pool, void *dst, void *src, >>>>> set_memory_nx((unsigned long)base, pages); >>>>> set_memory_rw((unsigned long)base, pages); >>>>> >>>>> - memcpy(dst, src, size); >>>>> + fncpy(dst, src, size); >>>>> >>>>> set_memory_ro((unsigned long)base, pages); >>>>> set_memory_x((unsigned long)base, pages); >>>>> >>>> >>>> Does this address your concerns from here [1]? Because the only user of this >>>> code is ARM right now I already only build the sram-exec code in if >>>> CONFIG_ARM is selected. >>> >>> Sorry, it does not. Please read the comments in asm/fncpy.h. >>> >>> Deviating from the proscribed usage means your code is, quite simply, >>> buggy. There's no two ways about that. >>> >> >> I understand there are many constraints to using fncpy, as this is what we >> used before to copy our executable code. Apart from users being aware of >> what these constraints are (8-byte aligned, position independent) and making >> sure the code they are moving meets them, are you saying we need some sort >> of additional strict enforcement of them? Because fncpy today will throw a >> bug if you fail to align src and dst properly, so adding another check will >> just double the messages to the user. > > Yes, fncpy() will throw a bug, but as I've already explained: > > sram = alloc(); > > sram_func = fncpy(sram, func, func_size); > > sram_func(); > > is the _only_ valid usage. > > You must not do: > > sram = alloc(); > > fncpy(sram, func, func_size); > > sram(); > > because that will not work with Thumb code. The only permitted usage > is as per the first example above, everything else is buggy. > I see exactly what you mean now. I missed that before, thank you for clarifying. Will update this patch and send a new version. Regards, Dave
diff --git a/drivers/misc/sram-exec.c b/drivers/misc/sram-exec.c index ac522417c462..0057eabe5c03 100644 --- a/drivers/misc/sram-exec.c +++ b/drivers/misc/sram-exec.c @@ -19,6 +19,7 @@ #include <linux/sram.h> #include <asm/cacheflush.h> +#include <asm/fncpy.h> #include "sram.h" @@ -93,7 +94,7 @@ int sram_exec_copy(struct gen_pool *pool, void *dst, void *src, set_memory_nx((unsigned long)base, pages); set_memory_rw((unsigned long)base, pages); - memcpy(dst, src, size); + fncpy(dst, src, size); set_memory_ro((unsigned long)base, pages); set_memory_x((unsigned long)base, pages);
Currently the sram-exec functionality, which allows allocation of executable memory and provides an API to move code to it, is only selected in configs for the ARM architecture. Based on commit 5756e9dd0de6 ("ARM: 6640/1: Thumb-2: Symbol manipulation macros for function body copying") simply copying a C function pointer address using memcpy without consideration of alignment and Thumb is unsafe on ARM platforms. The aforementioned patch introduces the fncpy macro which is a safe way to copy executable code on ARM platforms, so let's make use of that here rather than the unsafe plain memcpy that was previously used by sram_exec_copy. In the future, architectures hoping to make use of the sram-exec functionality must define an fncpy macro just as ARM has done to guarantee or check for safe copying to executable memory before allowing the arch to select CONFIG_SRAM_EXEC. Signed-off-by: Dave Gerlach <d-gerlach@ti.com> --- drivers/misc/sram-exec.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)