Re: [PATCH 1/1] aarch64: fix a crash during maintenance print cooked-registers

[Luis Machado <luis.machado@arm.com>]

Hi,

On 2/3/25 13:52, Klaus Gerlicher wrote:
> Hi all, Luis,
> 
> I found this accidently while looking at Thiago's GDBserver improvements. Luis
> I think added the pseudo register in question, so maybe he can comment
> on this issue. I used a aarch64 qemu 9.2.

Thanks for catching this.

> 
> Thanks
> Klaus
> 
> On aarch64 with pauth enabled I can see a crash when
> using "maintenance print cooked-registers".
> 
> This happens because the register dump code tries to read
> a pseudo reg that is not handled here because it is supposedly
> only used in unwinding.

Indeed. ra_sign_state is only used to track which frames have PAC enabled or not.

I was wondering why I had not caught this before, and I just noticed my VM had pauth disabled. :-(

Enabling it, now I see the crash. In any case...

> 
> Fix this by returning a zero value typed as a built-in uint64.
> ---
>  gdb/aarch64-tdep.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/gdb/aarch64-tdep.c b/gdb/aarch64-tdep.c
> index 840f9877361..6e712b12b86 100644
> --- a/gdb/aarch64-tdep.c
> +++ b/gdb/aarch64-tdep.c
> @@ -3290,6 +3290,9 @@ aarch64_pseudo_read_value (gdbarch *gdbarch, const frame_info_ptr &next_frame,
>      return aarch64_pseudo_read_value_1 (next_frame, pseudo_reg_num,
>  					pseudo_offset - AARCH64_SVE_V0_REGNUM);
>  
> +  if (tdep->has_pauth () && pseudo_reg_num == tdep->ra_sign_state_regnum)
> +    return value::zero (builtin_type (gdbarch)->builtin_uint64, lval_register);
> +
>    gdb_assert_not_reached ("regnum out of bound");
>  }
>  

... the fix above looks OK to me. Though it makes me wonder why we starting hitting this.

Internally we don't assign a name to this register so it technically doesn't show up in any
listing. Maybe something changed and we now list even pseudo registers without a name?

Approved-By: Luis Machado <luis.machado@arm.com>