Re: [PATCH,v2] [aarch64] Fix removal of non-address bits for PAuth

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

On 8/19/22 00:47, Thiago Jung Bauermann wrote:
> 
> Hello Luis,
> 
> The patch looks great to me, FWIW.
> I have just a few small comments.
> 
> Luis Machado via Gdb-patches <gdb-patches@sourceware.org> writes:
> 
>> While at it, make GDB and GDBServer share some more code for aarch64
> 
> Nice!
> 
>> +/* AArch64 implementation of the remove_non_address_bits gdbarch hook.  Remove
>> +   non address bits from a pointer value.  */
>> +
>> +static CORE_ADDR
>> +aarch64_remove_non_address_bits (struct gdbarch *gdbarch, CORE_ADDR pointer)
>> +{
>> +  aarch64_gdbarch_tdep *tdep
>> +    = (aarch64_gdbarch_tdep *) gdbarch_tdep (gdbarch);
>> +
>> +  /* By default, we assume TBI and discard the top 8 bits plus the VA range
>> +     select bit (55).  */
>> +  CORE_ADDR mask = (((CORE_ADDR) 1) << (64 - 9)) - 1;
>> +  mask = ~mask;
> 
> One suggestion is to put the above in gdb/arch/aarch64.h as:
> 
> /* By default, we assume TBI and discard the top 8 bits plus the VA range
>     select bit (55).  */
> #define AARCH64_TOP_BITS_MASK (~((((CORE_ADDR) 1) << (64 - 9)) - 1))
> 
> and then use it here and in gdbserver's version of this function so that
> there's a little bit more code shared between them.
> 

That makes sense, but I wonder if it will fit gdb/arch/aarch64.h. The VA bit is
a generic ISA thing, but the TBI is OS-specific I think.

Maybe a gdb/arch/aarch64-pauth.h header that gets shared between gdb and gdbserver
then?

>> +
>> +  if (tdep->has_pauth ())
>> +    {
>> +      /* Fetch the PAC masks.  These masks are per-process, so we can just
>> +	 fetch data from whatever thread we have at the moment.
>> +
>> +	 Also, we have both a code mask and a data mask.  For now they are the
>> +	 same, but this may change in the future.  */
>> +      struct regcache *regs = get_current_regcache ();
>> +      CORE_ADDR cmask, dmask;
>> +
>> +      if (regs->cooked_read (tdep->pauth_reg_base, &dmask) != REG_VALID)
>> +	dmask = mask;
>> +
>> +      if (regs->cooked_read (tdep->pauth_reg_base + 1, &cmask) != REG_VALID)
>> +	cmask = mask;
>> +
>> +      if (dmask != cmask)
>> +	{
>> +	  /* Warn if the masks are different.  */
>> +	  warning (_("C mask and D mask differ"));
> 
> This warning is a bit hard to read if one isn't familiar enough with
> PAuth. Perhaps something like “Pointer authentication masks for code and
> data differ”?
> 
> Same comment applies to the warning message in the gdbserver version of
> this function.
> 

You're right. Let me adjust those.

>> diff --git a/gdb/gdbarch-components.py b/gdb/gdbarch-components.py
>> index fc10e8600ba..c6836b63c50 100644
>> --- a/gdb/gdbarch-components.py
>> +++ b/gdb/gdbarch-components.py
>> @@ -1116,15 +1116,21 @@ possible it should be in TARGET_READ_PC instead).
>>       invalid=False,
>>   )
>>   
>> -Value(
>> +Method(
>>       comment="""
>> -On some machines, not all bits of an address word are significant.
>> -For example, on AArch64, the top bits of an address known as the "tag"
>> -are ignored by the kernel, the hardware, etc. and can be regarded as
>> -additional data associated with the address.
>> +On some architectures, not all bits of a pointer are significant.
>> +On AArch64, for example, the top bits of a pointer may carry a "tag", which
>> +can be ignored by the kernel and the hardware. The "tag" can be regarded as
>> +additional data associated with the pointer, but it is not part of the address.
>> +
>> +Given a pointer for the architecture, this hook removes all the
>> +non-significant bits and sign-extends things as needed. It is used mostly
>> +for data pointers, as opposed to code pointers.
> 
> Is the last sentence just FYI or does it imply some consequence for the
> GDB user or even GDB itself?
> 

More a FYI really. This was documented somewhere in the code, and I wanted to
bring it to the manual. I'll clarify it.