Re: [PATCH] AArch64 pauth: Indicate addresses in backtrace for kernel

[Luis Machado via Gdb-patches <gdb-patches@sourceware.org>]

On 10/25/21 8:47 AM, Kuan-Ying Lee via Gdb-patches wrote:
> Armv8.3-a Pointer Authentication cause the function return address to
> be changed. GDB need to use address bit[55] to know which mode is active
> and mask/unmask the link register in order to get backtrace.
> 
> If address is in kernel mode, we mask the address. If address is in user mode,
> we need to unmask the address.
> ---
>   gdb/aarch64-tdep.c | 5 ++++-
>   1 file changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/gdb/aarch64-tdep.c b/gdb/aarch64-tdep.c
> index 4b5af4616af..d4bb4305cea 100644
> --- a/gdb/aarch64-tdep.c
> +++ b/gdb/aarch64-tdep.c
> @@ -257,7 +257,10 @@ aarch64_frame_unmask_lr (struct gdbarch_tdep *tdep,
>       {
>         int cmask_num = AARCH64_PAUTH_CMASK_REGNUM (tdep->pauth_reg_base);
>         CORE_ADDR cmask = frame_unwind_register_unsigned (this_frame, cmask_num);
> -      addr = addr & ~cmask;
> +      if (addr & 0x0080000000000000ULL)
> +        addr = addr | cmask;
> +      else
> +        addr = addr & ~cmask;
>   
>         /* Record in the frame that the link register required unmasking.  */
>         set_frame_previous_pc_masked (this_frame);
> 

Could you please share more information about this problem? Why is it 
GDB needs to do things differently for a kernel mode and user mode 
address? What is the test setup?

If we entered the above conditional block, that means DWARF has told GDB 
that LR is masked (ra_state_regnum), and so it needs to be unmasked.

Given this is generic AArch64 code, we don't want to risk breaking 
existing use cases, and I'd like to understand what is not being handled 
properly.

We have another gdbarch method that handles sign-extending kernel mode 
addresses (gdbarch_significant_addr_bit), but it is not clear if that 
could be used here without some examples.

Thanks,
Luis