rseq/arm32: choosing rseq code signature

  ----- On Apr 9, 2019, at 3:32 PM, Mathieu Desnoyers mathieu.desnoyers@efficios.com wrote:

> Hi Will,
> 
> We are about to include the code signature required prior to restartable
> sequences abort handlers into glibc, which will make this ABI choice final.
> We need architecture maintainer input on that signature value.
> 
> That code signature is placed before each abort handler, so the kernel can
> validate that it is indeed jumping to an abort handler (and not some
> arbitrary attacker-chosen code). The signature is never executed.
> 
> The current discussion thread on the glibc mailing list leads us towards
> using a trap with uncommon immediate operand, which simplifies integration
> with disassemblers, emulators, makes it easier to debug if the control
> flow gets redirected there by mistake, and is nicer for some architecture's
> speculative execution.
> 
> We can have different signatures for each sub-architecture, as long as they
> don't have to co-exist within the same process. We can special-case with
> #ifdef for each sub-architecture and endianness if need be. If the architecture
> has instruction set extensions that can co-exist with the architecture
> instruction set within the same process (e.g. thumb for arm), we need to take
> into account to which instruction the chosen signature value would map (and
> possibly decide if we need to extend rseq to support many signatures).
> 
> Here is an example of rseq signature definition template:
> 
> /*
> * TODO: document trap instruction objdump output on each sub-architecture
> * instruction sets, as well as instruction set extensions.
> */
> #define RSEQ_SIG 0x########
> 
> Ideally we'd need a patch on top of the Linux kernel
> tools/testing/selftests/rseq/rseq-arm.h file that updates
> the signature value, so I can then pick it up for the glibc
> patchset.

Would the following diff work for you ? If so, can I get your
acked-by ?



> 
> Thanks!
> 
> Mathieu
> 
> --
> Mathieu Desnoyers
> EfficiOS Inc.
> http://www.efficios.com
  

Hi Mathieu,

On Wed, Apr 10, 2019 at 04:29:19PM -0400, Mathieu Desnoyers wrote:
> ----- On Apr 9, 2019, at 3:32 PM, Mathieu Desnoyers mathieu.desnoyers@efficios.com wrote:
> > We are about to include the code signature required prior to restartable
> > sequences abort handlers into glibc, which will make this ABI choice final.
> > We need architecture maintainer input on that signature value.
> > 
> > That code signature is placed before each abort handler, so the kernel can
> > validate that it is indeed jumping to an abort handler (and not some
> > arbitrary attacker-chosen code). The signature is never executed.
> > 
> > The current discussion thread on the glibc mailing list leads us towards
> > using a trap with uncommon immediate operand, which simplifies integration
> > with disassemblers, emulators, makes it easier to debug if the control
> > flow gets redirected there by mistake, and is nicer for some architecture's
> > speculative execution.
> > 
> > We can have different signatures for each sub-architecture, as long as they
> > don't have to co-exist within the same process. We can special-case with
> > #ifdef for each sub-architecture and endianness if need be. If the architecture
> > has instruction set extensions that can co-exist with the architecture
> > instruction set within the same process (e.g. thumb for arm), we need to take
> > into account to which instruction the chosen signature value would map (and
> > possibly decide if we need to extend rseq to support many signatures).
> > 
> > Here is an example of rseq signature definition template:
> > 
> > /*
> > * TODO: document trap instruction objdump output on each sub-architecture
> > * instruction sets, as well as instruction set extensions.
> > */
> > #define RSEQ_SIG 0x########
> > 
> > Ideally we'd need a patch on top of the Linux kernel
> > tools/testing/selftests/rseq/rseq-arm.h file that updates
> > the signature value, so I can then pick it up for the glibc
> > patchset.
> 
> Would the following diff work for you ? If so, can I get your
> acked-by ?

I had a quick chat with Richard and Peter (CC'd), since they're much more
familiar with the A32 instruction set than I am and also have a better view
of what might already be in use.

Peter suggests that anything of the form 0xe7fxdefx should trap in both A32
and T32, although it does assemble to UDF; B <imm11> in T16. I'm not sure we
should get too obsessed with trying to encode a signature that universally
decodes to a trap.

Whatever you choose, it would be worth checking that it doesn't clash with
other allocations such as software breakpoints in GDB.

Will

> diff --git a/tools/testing/selftests/rseq/rseq-arm.h b/tools/testing/selftests/rseq/rseq-arm.h
> index 5f262c54364f..1f261ad2ac1b 100644
> --- a/tools/testing/selftests/rseq/rseq-arm.h
> +++ b/tools/testing/selftests/rseq/rseq-arm.h
> @@ -5,7 +5,17 @@
>   * (C) Copyright 2016-2018 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
>   */
>  
> -#define RSEQ_SIG       0x53053053
> +/*
> + * RSEQ_SIG uses the udf A32 instruction with an uncommon immediate operand
> + * value 0x5305. This traps if user-space reaches this instruction by mistake,
> + * and the uncommon operand ensures the kernel does not move the instruction
> + * pointer to attacker-controlled code on rseq abort.
> + *
> + * The instruction pattern is:
> + *
> + * e7f530f5    udf    #21253    ; 0x5305
> + */
> +#define RSEQ_SIG       0xe7f530f5
>  
>  #define rseq_smp_mb()  __asm__ __volatile__ ("dmb" ::: "memory", "cc")
>  #define rseq_smp_rmb() __asm__ __volatile__ ("dmb" ::: "memory", "cc")
> @@ -78,7 +88,8 @@ do {                                                                  \
>                 __rseq_str(table_label) ":\n\t"                         \
>                 ".word " __rseq_str(version) ", " __rseq_str(flags) "\n\t" \
>                 ".word " __rseq_str(start_ip) ", 0x0, " __rseq_str(post_commit_offset) ", 0x0, " __rseq_str(abort_ip) ", 0x0\n\t" \
> -               ".word " __rseq_str(RSEQ_SIG) "\n\t"                    \
> +               ".arm\n\t"                                              \
> +               ".inst " __rseq_str(RSEQ_SIG) "\n\t"                    \
>                 __rseq_str(label) ":\n\t"                               \
>                 teardown                                                \
>                 "b %l[" __rseq_str(abort_label) "]\n\t"
  

----- On Apr 11, 2019, at 12:42 PM, Will Deacon will.deacon@arm.com wrote:

> Hi Mathieu,
> 
> On Wed, Apr 10, 2019 at 04:29:19PM -0400, Mathieu Desnoyers wrote:
>> ----- On Apr 9, 2019, at 3:32 PM, Mathieu Desnoyers
>> mathieu.desnoyers@efficios.com wrote:
>> > We are about to include the code signature required prior to restartable
>> > sequences abort handlers into glibc, which will make this ABI choice final.
>> > We need architecture maintainer input on that signature value.
>> > 
>> > That code signature is placed before each abort handler, so the kernel can
>> > validate that it is indeed jumping to an abort handler (and not some
>> > arbitrary attacker-chosen code). The signature is never executed.
>> > 
>> > The current discussion thread on the glibc mailing list leads us towards
>> > using a trap with uncommon immediate operand, which simplifies integration
>> > with disassemblers, emulators, makes it easier to debug if the control
>> > flow gets redirected there by mistake, and is nicer for some architecture's
>> > speculative execution.
>> > 
>> > We can have different signatures for each sub-architecture, as long as they
>> > don't have to co-exist within the same process. We can special-case with
>> > #ifdef for each sub-architecture and endianness if need be. If the architecture
>> > has instruction set extensions that can co-exist with the architecture
>> > instruction set within the same process (e.g. thumb for arm), we need to take
>> > into account to which instruction the chosen signature value would map (and
>> > possibly decide if we need to extend rseq to support many signatures).
>> > 
>> > Here is an example of rseq signature definition template:
>> > 
>> > /*
>> > * TODO: document trap instruction objdump output on each sub-architecture
>> > * instruction sets, as well as instruction set extensions.
>> > */
>> > #define RSEQ_SIG 0x########
>> > 
>> > Ideally we'd need a patch on top of the Linux kernel
>> > tools/testing/selftests/rseq/rseq-arm.h file that updates
>> > the signature value, so I can then pick it up for the glibc
>> > patchset.
>> 
>> Would the following diff work for you ? If so, can I get your
>> acked-by ?
> 
> I had a quick chat with Richard and Peter (CC'd), since they're much more
> familiar with the A32 instruction set than I am and also have a better view
> of what might already be in use.
> 
> Peter suggests that anything of the form 0xe7fxdefx should trap in both A32
> and T32, although it does assemble to UDF; B <imm11> in T16. I'm not sure we
> should get too obsessed with trying to encode a signature that universally
> decodes to a trap.

That's a nice trick.

> 
> Whatever you choose, it would be worth checking that it doesn't clash with
> other allocations such as software breakpoints in GDB.

GDB seems to have [1] :

#define ARM_LE_BREAKPOINT {0xFE,0xDE,0xFF,0xE7}
#define ARM_BE_BREAKPOINT {0xE7,0xFF,0xDE,0xFE}
#define THUMB_LE_BREAKPOINT {0xbe,0xbe}
#define THUMB_BE_BREAKPOINT {0xbe,0xbe}

None of which match the value you hint at.

So I could pick "0xe7f5def3", which would map to the following comment:

/*
 * RSEQ_SIG uses the udf A32 instruction with an uncommon immediate operand
 * value 0x5de3. This traps if user-space reaches this instruction by mistake,
 * and the uncommon operand ensures the kernel does not move the instruction
 * pointer to attacker-controlled code on rseq abort.
 *
 * The instruction pattern in the A32 instruction set is:
 *
 * e7f5def3    udf    #24035    ; 0x5de3
 *
 * This translates to the following instruction pattern in the T16 instruction
 * set:
 *
 * little endian:
 * def3        udf    #243      ; 0xf3
 * e7f5        b.n    <7f5>
 *
 * big endian:
 * e7f5        b.n    <7f5>
 * def3        udf    #243      ; 0xf3
 */
#define RSEQ_SIG        0xe7f5def3

Thoughts ?

Thanks!

Mathieu

[1] https://github.com/bminor/binutils-gdb/blob/master/gdb/arm-tdep.c#L7705-L7742
  

On Thu, 11 Apr 2019 at 18:51, Mathieu Desnoyers
<mathieu.desnoyers@efficios.com> wrote:
> ----- On Apr 11, 2019, at 12:42 PM, Will Deacon will.deacon@arm.com wrote:
> > Peter suggests that anything of the form 0xe7fxdefx should trap in both A32
> > and T32, although it does assemble to UDF; B <imm11> in T16. I'm not sure we
> > should get too obsessed with trying to encode a signature that universally
> > decodes to a trap.
>
> That's a nice trick.
>
> >
> > Whatever you choose, it would be worth checking that it doesn't clash with
> > other allocations such as software breakpoints in GDB.
>
> GDB seems to have [1] :
>
> #define ARM_LE_BREAKPOINT {0xFE,0xDE,0xFF,0xE7}
> #define ARM_BE_BREAKPOINT {0xE7,0xFF,0xDE,0xFE}
> #define THUMB_LE_BREAKPOINT {0xbe,0xbe}
> #define THUMB_BE_BREAKPOINT {0xbe,0xbe}
>
> None of which match the value you hint at.

Hmm? The ARM BPs match 0xe7fxdefx when considered with
the appropriate endianness (clearly somebody has
been down this line of thought before). Still, as long as
we pick different values for the 8 bits of freedom we
have it should be fine.

> /*
>  * RSEQ_SIG uses the udf A32 instruction with an uncommon immediate operand
>  * value 0x5de3. This traps if user-space reaches this instruction by mistake,
>  * and the uncommon operand ensures the kernel does not move the instruction
>  * pointer to attacker-controlled code on rseq abort.
>  *
>  * The instruction pattern in the A32 instruction set is:
>  *
>  * e7f5def3    udf    #24035    ; 0x5de3
>  *
>  * This translates to the following instruction pattern in the T16 instruction
>  * set:
>  *
>  * little endian:
>  * def3        udf    #243      ; 0xf3
>  * e7f5        b.n    <7f5>
>  *
>  * big endian:
>  * e7f5        b.n    <7f5>
>  * def3        udf    #243      ; 0xf3

Do we really care about big-endian instruction-ordering for Thumb?
It requires (AIUI) either an ARMv7R CPU which implements and sets
SCTLR.IE to 1, or a v6-or-earlier CPU using BE32, and it's going to
be even rarer than normal BE8 big-endian...

thanks
-- PMM
  

----- On Apr 11, 2019, at 3:55 PM, peter maydell peter.maydell@linaro.org wrote:

> On Thu, 11 Apr 2019 at 18:51, Mathieu Desnoyers
> <mathieu.desnoyers@efficios.com> wrote:
>> ----- On Apr 11, 2019, at 12:42 PM, Will Deacon will.deacon@arm.com wrote:
>> > Peter suggests that anything of the form 0xe7fxdefx should trap in both A32
>> > and T32, although it does assemble to UDF; B <imm11> in T16. I'm not sure we
>> > should get too obsessed with trying to encode a signature that universally
>> > decodes to a trap.
>>
>> That's a nice trick.
>>
>> >
>> > Whatever you choose, it would be worth checking that it doesn't clash with
>> > other allocations such as software breakpoints in GDB.
>>
>> GDB seems to have [1] :
>>
>> #define ARM_LE_BREAKPOINT {0xFE,0xDE,0xFF,0xE7}
>> #define ARM_BE_BREAKPOINT {0xE7,0xFF,0xDE,0xFE}
>> #define THUMB_LE_BREAKPOINT {0xbe,0xbe}
>> #define THUMB_BE_BREAKPOINT {0xbe,0xbe}
>>
>> None of which match the value you hint at.
> 
> Hmm? The ARM BPs match 0xe7fxdefx when considered with
> the appropriate endianness (clearly somebody has
> been down this line of thought before). Still, as long as
> we pick different values for the 8 bits of freedom we
> have it should be fine.

Right. I selected 0xe7f5def3, which should ensure we are distinct
from gdb's choice.

> 
>> /*
>>  * RSEQ_SIG uses the udf A32 instruction with an uncommon immediate operand
>>  * value 0x5de3. This traps if user-space reaches this instruction by mistake,
>>  * and the uncommon operand ensures the kernel does not move the instruction
>>  * pointer to attacker-controlled code on rseq abort.
>>  *
>>  * The instruction pattern in the A32 instruction set is:
>>  *
>>  * e7f5def3    udf    #24035    ; 0x5de3
>>  *
>>  * This translates to the following instruction pattern in the T16 instruction
>>  * set:
>>  *
>>  * little endian:
>>  * def3        udf    #243      ; 0xf3
>>  * e7f5        b.n    <7f5>
>>  *
>>  * big endian:
>>  * e7f5        b.n    <7f5>
>>  * def3        udf    #243      ; 0xf3
> 
> Do we really care about big-endian instruction-ordering for Thumb?
> It requires (AIUI) either an ARMv7R CPU which implements and sets
> SCTLR.IE to 1, or a v6-or-earlier CPU using BE32, and it's going to
> be even rarer than normal BE8 big-endian...

I don't think we care enough about it to look for a trick to
turn the branch into something else (which would not branch away from the
udf instruction), but considering this signature will be ABI, it's good to
be thorough documentation-wise and cover all existing cases.

Thoughts ?

Thanks,

Mathieu
  

On Mon, 15 Apr 2019 at 14:11, Mathieu Desnoyers
<mathieu.desnoyers@efficios.com> wrote:
>
> ----- On Apr 11, 2019, at 3:55 PM, peter maydell peter.maydell@linaro.org wrote:
>
> > On Thu, 11 Apr 2019 at 18:51, Mathieu Desnoyers
> > <mathieu.desnoyers@efficios.com> wrote:
> >>  * This translates to the following instruction pattern in the T16 instruction
> >>  * set:
> >>  *
> >>  * little endian:
> >>  * def3        udf    #243      ; 0xf3
> >>  * e7f5        b.n    <7f5>
> >>  *
> >>  * big endian:
> >>  * e7f5        b.n    <7f5>
> >>  * def3        udf    #243      ; 0xf3
> >
> > Do we really care about big-endian instruction-ordering for Thumb?
> > It requires (AIUI) either an ARMv7R CPU which implements and sets
> > SCTLR.IE to 1, or a v6-or-earlier CPU using BE32, and it's going to
> > be even rarer than normal BE8 big-endian...
>
> I don't think we care enough about it to look for a trick to
> turn the branch into something else (which would not branch away from the
> udf instruction), but considering this signature will be ABI, it's good to
> be thorough documentation-wise and cover all existing cases.

I think if you want to document it it would be helpful to
readers to make it clear that this is the ultra-rare
big-endian-instruction-order "big endian Thumb", not the only
moderately-rare little-endian-instructions-big-endian-data
"big endian Thumb".

thanks
-- PMM
  

----- On Apr 15, 2019, at 9:30 AM, peter maydell peter.maydell@linaro.org wrote:

> On Mon, 15 Apr 2019 at 14:11, Mathieu Desnoyers
> <mathieu.desnoyers@efficios.com> wrote:
>>
>> ----- On Apr 11, 2019, at 3:55 PM, peter maydell peter.maydell@linaro.org wrote:
>>
>> > On Thu, 11 Apr 2019 at 18:51, Mathieu Desnoyers
>> > <mathieu.desnoyers@efficios.com> wrote:
>> >>  * This translates to the following instruction pattern in the T16 instruction
>> >>  * set:
>> >>  *
>> >>  * little endian:
>> >>  * def3        udf    #243      ; 0xf3
>> >>  * e7f5        b.n    <7f5>
>> >>  *
>> >>  * big endian:
>> >>  * e7f5        b.n    <7f5>
>> >>  * def3        udf    #243      ; 0xf3
>> >
>> > Do we really care about big-endian instruction-ordering for Thumb?
>> > It requires (AIUI) either an ARMv7R CPU which implements and sets
>> > SCTLR.IE to 1, or a v6-or-earlier CPU using BE32, and it's going to
>> > be even rarer than normal BE8 big-endian...
>>
>> I don't think we care enough about it to look for a trick to
>> turn the branch into something else (which would not branch away from the
>> udf instruction), but considering this signature will be ABI, it's good to
>> be thorough documentation-wise and cover all existing cases.
> 
> I think if you want to document it it would be helpful to
> readers to make it clear that this is the ultra-rare
> big-endian-instruction-order "big endian Thumb", not the only
> moderately-rare little-endian-instructions-big-endian-data
> "big endian Thumb".

I'm actually very much concerned about environments with big endian
data and little endian code. Which gcc compiler flags do I need to
use to test it ?

I'm concerned about a signature mismatch between what is passed to
the rseq system call ("data-endian signature") and what is generated
in the code ("instruction-endian signature").

Thanks,

Mathieu
  

----- On Apr 15, 2019, at 9:37 AM, Mathieu Desnoyers mathieu.desnoyers@efficios.com wrote:

> ----- On Apr 15, 2019, at 9:30 AM, peter maydell peter.maydell@linaro.org wrote:
> 
>> On Mon, 15 Apr 2019 at 14:11, Mathieu Desnoyers
>> <mathieu.desnoyers@efficios.com> wrote:
>>>
>>> ----- On Apr 11, 2019, at 3:55 PM, peter maydell peter.maydell@linaro.org wrote:
>>>
>>> > On Thu, 11 Apr 2019 at 18:51, Mathieu Desnoyers
>>> > <mathieu.desnoyers@efficios.com> wrote:
>>> >>  * This translates to the following instruction pattern in the T16 instruction
>>> >>  * set:
>>> >>  *
>>> >>  * little endian:
>>> >>  * def3        udf    #243      ; 0xf3
>>> >>  * e7f5        b.n    <7f5>
>>> >>  *
>>> >>  * big endian:
>>> >>  * e7f5        b.n    <7f5>
>>> >>  * def3        udf    #243      ; 0xf3
>>> >
>>> > Do we really care about big-endian instruction-ordering for Thumb?
>>> > It requires (AIUI) either an ARMv7R CPU which implements and sets
>>> > SCTLR.IE to 1, or a v6-or-earlier CPU using BE32, and it's going to
>>> > be even rarer than normal BE8 big-endian...
>>>
>>> I don't think we care enough about it to look for a trick to
>>> turn the branch into something else (which would not branch away from the
>>> udf instruction), but considering this signature will be ABI, it's good to
>>> be thorough documentation-wise and cover all existing cases.
>> 
>> I think if you want to document it it would be helpful to
>> readers to make it clear that this is the ultra-rare
>> big-endian-instruction-order "big endian Thumb", not the only
>> moderately-rare little-endian-instructions-big-endian-data
>> "big endian Thumb".
> 
> I'm actually very much concerned about environments with big endian
> data and little endian code. Which gcc compiler flags do I need to
> use to test it ?
> 
> I'm concerned about a signature mismatch between what is passed to
> the rseq system call ("data-endian signature") and what is generated
> in the code ("instruction-endian signature").

Based on this page: http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0360f/CDFBBCHB.html

My understanding is that the situation is as follows (please confirm):

- Prior to ARMv6, you could build and run code that is either big or little endian,
  given you had a matching Linux kernel endianness. Code and data endianness needed
  to match,
- Starting from ARMv6, only little endian code is supported. The endianness for data
  access can be changed through bit [9], the E bit, of the Program Status Register,
  (mixed endianness)

Looking at ARM build options for gcc, it seems you can select either big or little
endian (-mbig-endian or -mlittle-endian (default)) which affects both instruction and
data endianness. So I suspect the -mbig-endian option is really only useful for
pre-ARMv6.

For ARMv6+ mixed-endianness, it seems to be a mode that temporarily swap endianness
of load/store instructions for specific memory accesses communicating with DMA devices,
so I don't see any scenario where we can generate a binary that has little endian code
and big endian data. If that is true, then it should be fine to declare the signature
with ".arm .inst" and expect the data endianness to be the same as code endianness.

Am I missing something ?

Thanks,

Mathieu
  

On 16/04/2019 14:39, Mathieu Desnoyers wrote:
> ----- On Apr 15, 2019, at 9:37 AM, Mathieu Desnoyers mathieu.desnoyers@efficios.com wrote:
> 
>> ----- On Apr 15, 2019, at 9:30 AM, peter maydell peter.maydell@linaro.org wrote:
>>
>>> On Mon, 15 Apr 2019 at 14:11, Mathieu Desnoyers
>>> <mathieu.desnoyers@efficios.com> wrote:
>>>>
>>>> ----- On Apr 11, 2019, at 3:55 PM, peter maydell peter.maydell@linaro.org wrote:
>>>>
>>>>> On Thu, 11 Apr 2019 at 18:51, Mathieu Desnoyers
>>>>> <mathieu.desnoyers@efficios.com> wrote:
>>>>>>  * This translates to the following instruction pattern in the T16 instruction
>>>>>>  * set:
>>>>>>  *
>>>>>>  * little endian:
>>>>>>  * def3        udf    #243      ; 0xf3
>>>>>>  * e7f5        b.n    <7f5>
>>>>>>  *
>>>>>>  * big endian:
>>>>>>  * e7f5        b.n    <7f5>
>>>>>>  * def3        udf    #243      ; 0xf3
>>>>>
>>>>> Do we really care about big-endian instruction-ordering for Thumb?
>>>>> It requires (AIUI) either an ARMv7R CPU which implements and sets
>>>>> SCTLR.IE to 1, or a v6-or-earlier CPU using BE32, and it's going to
>>>>> be even rarer than normal BE8 big-endian...
>>>>
>>>> I don't think we care enough about it to look for a trick to
>>>> turn the branch into something else (which would not branch away from the
>>>> udf instruction), but considering this signature will be ABI, it's good to
>>>> be thorough documentation-wise and cover all existing cases.
>>>
>>> I think if you want to document it it would be helpful to
>>> readers to make it clear that this is the ultra-rare
>>> big-endian-instruction-order "big endian Thumb", not the only
>>> moderately-rare little-endian-instructions-big-endian-data
>>> "big endian Thumb".
>>
>> I'm actually very much concerned about environments with big endian
>> data and little endian code. Which gcc compiler flags do I need to
>> use to test it ?
>>
>> I'm concerned about a signature mismatch between what is passed to
>> the rseq system call ("data-endian signature") and what is generated
>> in the code ("instruction-endian signature").
> 
> Based on this page: http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0360f/CDFBBCHB.html
> 
> My understanding is that the situation is as follows (please confirm):
> 
> - Prior to ARMv6, you could build and run code that is either big or little endian,
>   given you had a matching Linux kernel endianness. Code and data endianness needed
>   to match,
> - Starting from ARMv6, only little endian code is supported. The endianness for data
>   access can be changed through bit [9], the E bit, of the Program Status Register,
>   (mixed endianness)
> 
> Looking at ARM build options for gcc, it seems you can select either big or little
> endian (-mbig-endian or -mlittle-endian (default)) which affects both instruction and
> data endianness. So I suspect the -mbig-endian option is really only useful for
> pre-ARMv6.

-mbig-endian is still correct, even on later architectures.  The linker
gets involved, however, and (using the mapping symbol information) swaps
the code segments to little-endian form (this is why you have to use
.inst rather than .word when inserting instructions, so that the correct
mapping symbols are inserted).

> 
> For ARMv6+ mixed-endianness, it seems to be a mode that temporarily swap endianness
> of load/store instructions for specific memory accesses communicating with DMA devices,
> so I don't see any scenario where we can generate a binary that has little endian code
> and big endian data. If that is true, then it should be fine to declare the signature
> with ".arm .inst" and expect the data endianness to be the same as code endianness.
> 
> Am I missing something ?
> 
> Thanks,
> 
> Mathieu
>
  

----- On Apr 17, 2019, at 6:37 AM, richard earnshaw Richard.Earnshaw@arm.com wrote:

> On 16/04/2019 14:39, Mathieu Desnoyers wrote:
>> ----- On Apr 15, 2019, at 9:37 AM, Mathieu Desnoyers
>> mathieu.desnoyers@efficios.com wrote:
>> 
>>> ----- On Apr 15, 2019, at 9:30 AM, peter maydell peter.maydell@linaro.org wrote:
>>>
>>>> On Mon, 15 Apr 2019 at 14:11, Mathieu Desnoyers
>>>> <mathieu.desnoyers@efficios.com> wrote:
>>>>>
>>>>> ----- On Apr 11, 2019, at 3:55 PM, peter maydell peter.maydell@linaro.org wrote:
>>>>>
>>>>>> On Thu, 11 Apr 2019 at 18:51, Mathieu Desnoyers
>>>>>> <mathieu.desnoyers@efficios.com> wrote:
>>>>>>>  * This translates to the following instruction pattern in the T16 instruction
>>>>>>>  * set:
>>>>>>>  *
>>>>>>>  * little endian:
>>>>>>>  * def3        udf    #243      ; 0xf3
>>>>>>>  * e7f5        b.n    <7f5>
>>>>>>>  *
>>>>>>>  * big endian:
>>>>>>>  * e7f5        b.n    <7f5>
>>>>>>>  * def3        udf    #243      ; 0xf3
>>>>>>
>>>>>> Do we really care about big-endian instruction-ordering for Thumb?
>>>>>> It requires (AIUI) either an ARMv7R CPU which implements and sets
>>>>>> SCTLR.IE to 1, or a v6-or-earlier CPU using BE32, and it's going to
>>>>>> be even rarer than normal BE8 big-endian...
>>>>>
>>>>> I don't think we care enough about it to look for a trick to
>>>>> turn the branch into something else (which would not branch away from the
>>>>> udf instruction), but considering this signature will be ABI, it's good to
>>>>> be thorough documentation-wise and cover all existing cases.
>>>>
>>>> I think if you want to document it it would be helpful to
>>>> readers to make it clear that this is the ultra-rare
>>>> big-endian-instruction-order "big endian Thumb", not the only
>>>> moderately-rare little-endian-instructions-big-endian-data
>>>> "big endian Thumb".
>>>
>>> I'm actually very much concerned about environments with big endian
>>> data and little endian code. Which gcc compiler flags do I need to
>>> use to test it ?
>>>
>>> I'm concerned about a signature mismatch between what is passed to
>>> the rseq system call ("data-endian signature") and what is generated
>>> in the code ("instruction-endian signature").
>> 
>> Based on this page:
>> http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0360f/CDFBBCHB.html
>> 
>> My understanding is that the situation is as follows (please confirm):
>> 
>> - Prior to ARMv6, you could build and run code that is either big or little
>> endian,
>>   given you had a matching Linux kernel endianness. Code and data endianness
>>   needed
>>   to match,
>> - Starting from ARMv6, only little endian code is supported. The endianness for
>> data
>>   access can be changed through bit [9], the E bit, of the Program Status
>>   Register,
>>   (mixed endianness)
>> 
>> Looking at ARM build options for gcc, it seems you can select either big or
>> little
>> endian (-mbig-endian or -mlittle-endian (default)) which affects both
>> instruction and
>> data endianness. So I suspect the -mbig-endian option is really only useful for
>> pre-ARMv6.
> 
> -mbig-endian is still correct, even on later architectures.  The linker
> gets involved, however, and (using the mapping symbol information) swaps
> the code segments to little-endian form (this is why you have to use
> .inst rather than .word when inserting instructions, so that the correct
> mapping symbols are inserted).

So what you're saying is that if I have:

void main()
{
        asm volatile (
                        ".arm\n\t"
                        ".inst 0xe7f5def3\n\t"
                        ".long 0xe7f5def3\n\t");
}

and compile it with:

arm-linux-gnueabihf-gcc -mbig-endian -march=armv6 -c -o arm-big-endianv6.o arm-test-endian.c

It's expected that the generated .o will have big endian instructions, matching
the endianness of the data, e.g.:

hexdump arm-big-endianv6.o

[...]
0000030 0a00 0900 80b5 00af f5e7 f3de f5e7 f3de

But it's then at the linking stage that the linker will
reverse the endianness of the ".inst" (but not .long).

Let's see:

arm-linux-gnueabihf-gcc -nodefaultlibs -nostdlib -mbig-endian -march=armv6 -o arm-big-endianv6 arm-big-endianv6.o 
/usr/lib/gcc-cross/arm-linux-gnueabihf/7/../../../../arm-linux-gnueabihf/bin/ld: warning: cannot find entry symbol _start; defaulting to 00000000000001b0

hexdump gives me:
[...]
00001b0 80b5 00af f5e7 f3de f5e7 f3de c046 bd46

So it has not reversed the instruction endianness.

What am I doing wrong ?

I'm using:

gcc version 7.3.0 (Ubuntu/Linaro 7.3.0-27ubuntu1~18.04)
GNU ld (GNU Binutils for Ubuntu) 2.30

Thanks,

Mathieu

> 
>> 
>> For ARMv6+ mixed-endianness, it seems to be a mode that temporarily swap
>> endianness
>> of load/store instructions for specific memory accesses communicating with DMA
>> devices,
>> so I don't see any scenario where we can generate a binary that has little
>> endian code
>> and big endian data. If that is true, then it should be fine to declare the
>> signature
>> with ".arm .inst" and expect the data endianness to be the same as code
>> endianness.
>> 
>> Am I missing something ?
>> 
>> Thanks,
>> 
>> Mathieu
  

----- On Apr 17, 2019, at 10:43 AM, Mathieu Desnoyers mathieu.desnoyers@efficios.com wrote:

> ----- On Apr 17, 2019, at 6:37 AM, richard earnshaw Richard.Earnshaw@arm.com
> wrote:
> 
>> On 16/04/2019 14:39, Mathieu Desnoyers wrote:
>>> ----- On Apr 15, 2019, at 9:37 AM, Mathieu Desnoyers
>>> mathieu.desnoyers@efficios.com wrote:
>>> 
>>>> ----- On Apr 15, 2019, at 9:30 AM, peter maydell peter.maydell@linaro.org wrote:
>>>>
>>>>> On Mon, 15 Apr 2019 at 14:11, Mathieu Desnoyers
>>>>> <mathieu.desnoyers@efficios.com> wrote:
>>>>>>
>>>>>> ----- On Apr 11, 2019, at 3:55 PM, peter maydell peter.maydell@linaro.org wrote:
>>>>>>
>>>>>>> On Thu, 11 Apr 2019 at 18:51, Mathieu Desnoyers
>>>>>>> <mathieu.desnoyers@efficios.com> wrote:
>>>>>>>>  * This translates to the following instruction pattern in the T16 instruction
>>>>>>>>  * set:
>>>>>>>>  *
>>>>>>>>  * little endian:
>>>>>>>>  * def3        udf    #243      ; 0xf3
>>>>>>>>  * e7f5        b.n    <7f5>
>>>>>>>>  *
>>>>>>>>  * big endian:
>>>>>>>>  * e7f5        b.n    <7f5>
>>>>>>>>  * def3        udf    #243      ; 0xf3
>>>>>>>
>>>>>>> Do we really care about big-endian instruction-ordering for Thumb?
>>>>>>> It requires (AIUI) either an ARMv7R CPU which implements and sets
>>>>>>> SCTLR.IE to 1, or a v6-or-earlier CPU using BE32, and it's going to
>>>>>>> be even rarer than normal BE8 big-endian...
>>>>>>
>>>>>> I don't think we care enough about it to look for a trick to
>>>>>> turn the branch into something else (which would not branch away from the
>>>>>> udf instruction), but considering this signature will be ABI, it's good to
>>>>>> be thorough documentation-wise and cover all existing cases.
>>>>>
>>>>> I think if you want to document it it would be helpful to
>>>>> readers to make it clear that this is the ultra-rare
>>>>> big-endian-instruction-order "big endian Thumb", not the only
>>>>> moderately-rare little-endian-instructions-big-endian-data
>>>>> "big endian Thumb".
>>>>
>>>> I'm actually very much concerned about environments with big endian
>>>> data and little endian code. Which gcc compiler flags do I need to
>>>> use to test it ?
>>>>
>>>> I'm concerned about a signature mismatch between what is passed to
>>>> the rseq system call ("data-endian signature") and what is generated
>>>> in the code ("instruction-endian signature").
>>> 
>>> Based on this page:
>>> http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0360f/CDFBBCHB.html
>>> 
>>> My understanding is that the situation is as follows (please confirm):
>>> 
>>> - Prior to ARMv6, you could build and run code that is either big or little
>>> endian,
>>>   given you had a matching Linux kernel endianness. Code and data endianness
>>>   needed
>>>   to match,
>>> - Starting from ARMv6, only little endian code is supported. The endianness for
>>> data
>>>   access can be changed through bit [9], the E bit, of the Program Status
>>>   Register,
>>>   (mixed endianness)
>>> 
>>> Looking at ARM build options for gcc, it seems you can select either big or
>>> little
>>> endian (-mbig-endian or -mlittle-endian (default)) which affects both
>>> instruction and
>>> data endianness. So I suspect the -mbig-endian option is really only useful for
>>> pre-ARMv6.
>> 
>> -mbig-endian is still correct, even on later architectures.  The linker
>> gets involved, however, and (using the mapping symbol information) swaps
>> the code segments to little-endian form (this is why you have to use
>> .inst rather than .word when inserting instructions, so that the correct
>> mapping symbols are inserted).
> 
> So what you're saying is that if I have:
> 
> void main()
> {
>        asm volatile (
>                        ".arm\n\t"
>                        ".inst 0xe7f5def3\n\t"
>                        ".long 0xe7f5def3\n\t");
> }
> 
> and compile it with:
> 
> arm-linux-gnueabihf-gcc -mbig-endian -march=armv6 -c -o arm-big-endianv6.o
> arm-test-endian.c
> 
> It's expected that the generated .o will have big endian instructions, matching
> the endianness of the data, e.g.:
> 
> hexdump arm-big-endianv6.o
> 
> [...]
> 0000030 0a00 0900 80b5 00af f5e7 f3de f5e7 f3de
> 
> But it's then at the linking stage that the linker will
> reverse the endianness of the ".inst" (but not .long).
> 
> Let's see:
> 
> arm-linux-gnueabihf-gcc -nodefaultlibs -nostdlib -mbig-endian -march=armv6 -o
> arm-big-endianv6 arm-big-endianv6.o
> /usr/lib/gcc-cross/arm-linux-gnueabihf/7/../../../../arm-linux-gnueabihf/bin/ld:
> warning: cannot find entry symbol _start; defaulting to 00000000000001b0
> 
> hexdump gives me:
> [...]
> 00001b0 80b5 00af f5e7 f3de f5e7 f3de c046 bd46
> 
> So it has not reversed the instruction endianness.
> 
> What am I doing wrong ?

It seems to be specific to using armv6 and armv7* with gcc 7.
gcc 8 seems to indeed reverse the code vs data endianness.

So we need to figure out whether .inst is the right things to
do to declare a signature, or if it's better to use ".long"
which would probably generate an invalid instruction on BE...

Thanks,

Mathieu

> 
> I'm using:
> 
> gcc version 7.3.0 (Ubuntu/Linaro 7.3.0-27ubuntu1~18.04)
> GNU ld (GNU Binutils for Ubuntu) 2.30
> 
> Thanks,
> 
> Mathieu
> 
>> 
>>> 
>>> For ARMv6+ mixed-endianness, it seems to be a mode that temporarily swap
>>> endianness
>>> of load/store instructions for specific memory accesses communicating with DMA
>>> devices,
>>> so I don't see any scenario where we can generate a binary that has little
>>> endian code
>>> and big endian data. If that is true, then it should be fine to declare the
>>> signature
>>> with ".arm .inst" and expect the data endianness to be the same as code
>>> endianness.
>>> 
>>> Am I missing something ?
>>> 
>>> Thanks,
>>> 
>>> Mathieu
> 
> --
> Mathieu Desnoyers
> EfficiOS Inc.
> http://www.efficios.com
  

On 17/04/2019 16:30, Mathieu Desnoyers wrote:
> ----- On Apr 17, 2019, at 10:43 AM, Mathieu Desnoyers mathieu.desnoyers@efficios.com wrote:
> 
>> ----- On Apr 17, 2019, at 6:37 AM, richard earnshaw Richard.Earnshaw@arm.com
>> wrote:
>>
>>> On 16/04/2019 14:39, Mathieu Desnoyers wrote:
>>>> ----- On Apr 15, 2019, at 9:37 AM, Mathieu Desnoyers
>>>> mathieu.desnoyers@efficios.com wrote:
>>>>
>>>>> ----- On Apr 15, 2019, at 9:30 AM, peter maydell peter.maydell@linaro.org wrote:
>>>>>
>>>>>> On Mon, 15 Apr 2019 at 14:11, Mathieu Desnoyers
>>>>>> <mathieu.desnoyers@efficios.com> wrote:
>>>>>>>
>>>>>>> ----- On Apr 11, 2019, at 3:55 PM, peter maydell peter.maydell@linaro.org wrote:
>>>>>>>
>>>>>>>> On Thu, 11 Apr 2019 at 18:51, Mathieu Desnoyers
>>>>>>>> <mathieu.desnoyers@efficios.com> wrote:
>>>>>>>>>  * This translates to the following instruction pattern in the T16 instruction
>>>>>>>>>  * set:
>>>>>>>>>  *
>>>>>>>>>  * little endian:
>>>>>>>>>  * def3        udf    #243      ; 0xf3
>>>>>>>>>  * e7f5        b.n    <7f5>
>>>>>>>>>  *
>>>>>>>>>  * big endian:
>>>>>>>>>  * e7f5        b.n    <7f5>
>>>>>>>>>  * def3        udf    #243      ; 0xf3
>>>>>>>>
>>>>>>>> Do we really care about big-endian instruction-ordering for Thumb?
>>>>>>>> It requires (AIUI) either an ARMv7R CPU which implements and sets
>>>>>>>> SCTLR.IE to 1, or a v6-or-earlier CPU using BE32, and it's going to
>>>>>>>> be even rarer than normal BE8 big-endian...
>>>>>>>
>>>>>>> I don't think we care enough about it to look for a trick to
>>>>>>> turn the branch into something else (which would not branch away from the
>>>>>>> udf instruction), but considering this signature will be ABI, it's good to
>>>>>>> be thorough documentation-wise and cover all existing cases.
>>>>>>
>>>>>> I think if you want to document it it would be helpful to
>>>>>> readers to make it clear that this is the ultra-rare
>>>>>> big-endian-instruction-order "big endian Thumb", not the only
>>>>>> moderately-rare little-endian-instructions-big-endian-data
>>>>>> "big endian Thumb".
>>>>>
>>>>> I'm actually very much concerned about environments with big endian
>>>>> data and little endian code. Which gcc compiler flags do I need to
>>>>> use to test it ?
>>>>>
>>>>> I'm concerned about a signature mismatch between what is passed to
>>>>> the rseq system call ("data-endian signature") and what is generated
>>>>> in the code ("instruction-endian signature").
>>>>
>>>> Based on this page:
>>>> http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0360f/CDFBBCHB.html
>>>>
>>>> My understanding is that the situation is as follows (please confirm):
>>>>
>>>> - Prior to ARMv6, you could build and run code that is either big or little
>>>> endian,
>>>>   given you had a matching Linux kernel endianness. Code and data endianness
>>>>   needed
>>>>   to match,
>>>> - Starting from ARMv6, only little endian code is supported. The endianness for
>>>> data
>>>>   access can be changed through bit [9], the E bit, of the Program Status
>>>>   Register,
>>>>   (mixed endianness)
>>>>
>>>> Looking at ARM build options for gcc, it seems you can select either big or
>>>> little
>>>> endian (-mbig-endian or -mlittle-endian (default)) which affects both
>>>> instruction and
>>>> data endianness. So I suspect the -mbig-endian option is really only useful for
>>>> pre-ARMv6.
>>>
>>> -mbig-endian is still correct, even on later architectures.  The linker
>>> gets involved, however, and (using the mapping symbol information) swaps
>>> the code segments to little-endian form (this is why you have to use
>>> .inst rather than .word when inserting instructions, so that the correct
>>> mapping symbols are inserted).
>>
>> So what you're saying is that if I have:
>>
>> void main()
>> {
>>        asm volatile (
>>                        ".arm\n\t"
>>                        ".inst 0xe7f5def3\n\t"
>>                        ".long 0xe7f5def3\n\t");
>> }
>>
>> and compile it with:
>>
>> arm-linux-gnueabihf-gcc -mbig-endian -march=armv6 -c -o arm-big-endianv6.o
>> arm-test-endian.c
>>
>> It's expected that the generated .o will have big endian instructions, matching
>> the endianness of the data, e.g.:
>>
>> hexdump arm-big-endianv6.o
>>
>> [...]
>> 0000030 0a00 0900 80b5 00af f5e7 f3de f5e7 f3de
>>
>> But it's then at the linking stage that the linker will
>> reverse the endianness of the ".inst" (but not .long).
>>
>> Let's see:
>>
>> arm-linux-gnueabihf-gcc -nodefaultlibs -nostdlib -mbig-endian -march=armv6 -o
>> arm-big-endianv6 arm-big-endianv6.o
>> /usr/lib/gcc-cross/arm-linux-gnueabihf/7/../../../../arm-linux-gnueabihf/bin/ld:
>> warning: cannot find entry symbol _start; defaulting to 00000000000001b0
>>
>> hexdump gives me:
>> [...]
>> 00001b0 80b5 00af f5e7 f3de f5e7 f3de c046 bd46
>>
>> So it has not reversed the instruction endianness.
>>
>> What am I doing wrong ?
> 
> It seems to be specific to using armv6 and armv7* with gcc 7.
> gcc 8 seems to indeed reverse the code vs data endianness.
> 
> So we need to figure out whether .inst is the right things to
> do to declare a signature, or if it's better to use ".long"
> which would probably generate an invalid instruction on BE...

If you used .long the value would then be left as data, but tools that
scanned for data in the text segment (some environments forbid this)
would then fault the object as 'impure'.

So using an instruction would be preferable, but on a BE8 image you'll
need to byte-swap the 'value' before comparing it.  Such images will
have EF_ARM_BE8 set in the flags field of the image header.

R.

> 
> Thanks,
> 
> Mathieu
> 
>>
>> I'm using:
>>
>> gcc version 7.3.0 (Ubuntu/Linaro 7.3.0-27ubuntu1~18.04)
>> GNU ld (GNU Binutils for Ubuntu) 2.30
>>
>> Thanks,
>>
>> Mathieu
>>
>>>
>>>>
>>>> For ARMv6+ mixed-endianness, it seems to be a mode that temporarily swap
>>>> endianness
>>>> of load/store instructions for specific memory accesses communicating with DMA
>>>> devices,
>>>> so I don't see any scenario where we can generate a binary that has little
>>>> endian code
>>>> and big endian data. If that is true, then it should be fine to declare the
>>>> signature
>>>> with ".arm .inst" and expect the data endianness to be the same as code
>>>> endianness.
>>>>
>>>> Am I missing something ?
>>>>
>>>> Thanks,
>>>>
>>>> Mathieu
>>
>> --
>> Mathieu Desnoyers
>> EfficiOS Inc.
>> http://www.efficios.com
>