[PATCH v15 6/8] nptl: Move the rseq area to the 'extra TLS' block

[Michael Jeanson]

On 2025-01-08 12:06, Florian Weimer wrote:
> * Michael Jeanson:
> 
>> On 2025-01-08 04:40, Florian Weimer wrote:
>>> * Michael Jeanson:
>>>
>>>> +#if TLS_TCB_AT_TP
>>>> +  /* The rseq area block should come before the thread pointer and be at least
>>>> +     32 bytes. */
>>>> +  TEST_VERIFY (__rseq_offset <= -RSEQ_AREA_SIZE_INITIAL);
>>>> +#elif TLS_DTV_AT_TP
>>>> +  /* The rseq area block should come after the thread pointer. */
>>>> +  TEST_VERIFY (__rseq_offset + TLS_TP_OFFSET >= 0);
>>>
>>> I think the TLS_TP_OFFSET is redudant here.  It's supposed have already
>>> been folded into __rseq_offset.  It certainly was before because a
>>> pointer difference was used to compute it.
>>
>> I should have updated the comment here, I'm testing that the rseq area
>> comes after the TCB on TLS_DTV_AT_TP architectures, my first iteration
>> looked like this:
>>
>>   TEST_VERIFY (__rseq_offset >= -TLS_TP_OFFSET);
>>
>> but I felt it was not very readable.
>>
>> Would this be better?
>>
>>  +  /* The rseq area block should come after the TCB, add the TLS block
>>  +     offset to the rseq offset to get a value relative to the TCB and
>>  +     test that it's positive.  */
>>  +  TEST_VERIFY (__rseq_offset + TLS_TP_OFFSET >= 0);
> 
> The comment helps, but it should say non-negative to match the actual
> comparison?

Ack.

> 
> In addition you could define a variable
> 
> static __thread char thread_var __attribute__ ((tls_model ("initial-exec")));
> 
> and check that the effective rseq address is greater than &thread_var.

I like this, I'll add this test.

Thanks,

Michael