[0/4] Add AARCH64 pointer authentication support

  Hi,

I've included a set of patches in order to demangle return addresses in
aarch64 platforms with pointer authentication.

Besides adding the implementation of the negate_ra_state opcode, there
is a new function in the libdwfl.h header to feed the PAC masks to the
library.

Let me know if there are any concerns with the current version.

Thanks,
German

German Gomez (4):
  aarch64: Create definitions for AARCH64_RA_SIGN_STATE register
  libdw,aarch64: Implement DW_CFA_AARCH64_negate_ra_state CFI
    instruction
  libdwfl,aarch64: Demangle return addresses using a PAC mask
  libdwfl,eu-stack,aarch64: Add API for setting AARCH64 PAC mask.

 backends/aarch64_init.c    |  6 +++---
 backends/aarch64_initreg.c |  2 ++
 backends/aarch64_regs.c    |  5 ++++-
 libdw/cfi.c                | 14 +++++++++++++-
 libdw/dwarf.h              |  5 +++++
 libdw/libdw.map            |  5 +++++
 libdwfl/dwfl_frame.c       |  3 +++
 libdwfl/dwfl_frame_regs.c  | 10 ++++++++++
 libdwfl/frame_unwind.c     | 14 +++++++++++++-
 libdwfl/libdwfl.h          |  6 ++++++
 libdwfl/libdwflP.h         |  7 +++++++
 libdwfl/linux-pid-attach.c | 34 ++++++++++++++++++++++++++++++++--
 tests/run-addrcfi.sh       |  1 +
 tests/run-allregs.sh       |  1 +
 14 files changed, 105 insertions(+), 8 deletions(-)
  

Hi German,

On Mon, Apr 25, 2022 at 02:03:07PM +0000, German Gomez via Elfutils-devel wrote:
> I've included a set of patches in order to demangle return addresses in
> aarch64 platforms with pointer authentication.
> 
> Besides adding the implementation of the negate_ra_state opcode, there
> is a new function in the libdwfl.h header to feed the PAC masks to the
> library.
> 
> Let me know if there are any concerns with the current version.

Thanks a lot for this. Last time I looked at this didn't have any
means to test this, so I skipped implementing it. How did you test? Do
distributions now enable PAC by default and is there hardware (qemu?)
support?

I haven't been able to look at the actual patches yet. And I am on
vacation this week. But I'll review next week after I am back.

A quick scan shows we need a aarch64 special public function, which
would be slightly ugly imho. I had hoped it could be a variant of the
func_addr_mask. But maybe this is too different to make more generic.

Cheers,

Mark
  

Hi Mark, thanks for looking, and sorry for the delay

On 28/04/2022 20:56, Mark Wielaard wrote:
> Hi German,
>
> On Mon, Apr 25, 2022 at 02:03:07PM +0000, German Gomez via Elfutils-devel wrote:
>> I've included a set of patches in order to demangle return addresses in
>> aarch64 platforms with pointer authentication.
>>
>> Besides adding the implementation of the negate_ra_state opcode, there
>> is a new function in the libdwfl.h header to feed the PAC masks to the
>> library.
>>
>> Let me know if there are any concerns with the current version.
> Thanks a lot for this. Last time I looked at this didn't have any
> means to test this, so I skipped implementing it. How did you test? Do
> distributions now enable PAC by default and is there hardware (qemu?)
> support?

So far I've been testing on Graviton3 cores (running linux), which seem
to implement the PAC extension, and it came enabled by default.

https://www.kernel.org/doc/html/latest/arm64/pointer-authentication.html

> I haven't been able to look at the actual patches yet. And I am on
> vacation this week. But I'll review next week after I am back.

Thanks a lot for looking.

>
> A quick scan shows we need a aarch64 special public function, which
> would be slightly ugly imho. I had hoped it could be a variant of the
> func_addr_mask. But maybe this is too different to make more generic.

I did consider func_addr_mask initially, but when I wrote the patch it
wasn't exposed as a perf-thread value. Currently PAC masks are constant
but might be different from thread to thread in the future. So I placed
it in the Thread struct.

I agree the arch-specific naming is not pretty. I think I can certainly
rework it into a more generic feature. But I think I would need to make
sure that the masks can be supplied to the Thread struct before the   
unwind.

Thanks,
German

> Cheers,
>
> Mark
>
  

On 19/05/2022 14:30, German Gomez via Elfutils-devel wrote:
> Hi Mark, thanks for looking, and sorry for the delay
>
> On 28/04/2022 20:56, Mark Wielaard wrote:
>> Hi German,
>>
>> On Mon, Apr 25, 2022 at 02:03:07PM +0000, German Gomez via Elfutils-devel wrote:
>>> [...]
>> Thanks a lot for this. Last time I looked at this didn't have any
>> means to test this, so I skipped implementing it. How did you test? Do
>> distributions now enable PAC by default and is there hardware (qemu?)
>> support?
> So far I've been testing on Graviton3 cores (running linux), which seem
> to implement the PAC extension, and it came enabled by default.
>

Regarding qemu support, I haven't used it myself so I can't really speak
for it, but according to the docs FEAT_PAuth is supported.

https://www.qemu.org/docs/master/system/arm/emulation.html