[PATCH] x86: Increase allowable TSX abort rate to 6%.

[Carlos O'Donell]

On 11/13/25 7:48 PM, H.J. Lu wrote:
> On Thu, Nov 13, 2025 at 11:01 PM Carlos O'Donell <carlos@redhat.com> wrote:
>>
>> In pre-commit CI on an E5-2698 v4 we sometimes see ~5% aborts.
>>
>> Set the trip point to 6%.
>> ---
>>   sysdeps/x86/tst-string-rtm.h | 20 +++++++++++++++++---
>>   1 file changed, 17 insertions(+), 3 deletions(-)
>>
>> diff --git a/sysdeps/x86/tst-string-rtm.h b/sysdeps/x86/tst-string-rtm.h
>> index 1bf863063f..699c52227b 100644
>> --- a/sysdeps/x86/tst-string-rtm.h
>> +++ b/sysdeps/x86/tst-string-rtm.h
>> @@ -36,6 +36,18 @@ do_test_1 (const char *name, unsigned int loop, int (*prepare) (void),
>>     unsigned int i;
>>     unsigned int naborts = 0;
>>     unsigned int failed = 0;
>> +
>> +  /* Iterate LOOP number of times, calling the same function over
>> +     and over again in an attempt to detect the following failure
>> +     modes:
>> +
>> +     1. Aborting instruction on the hot path.
>> +     2. Number of instructions executed exceeds maximum.
>> +     3. Memory reads or writes exceed maximum.
>> +
>> +     The function getting bigger may increase the abort rate due
>> +     to external events, despite not triggering #2 above, and may
>> +     need adjusting the abort rate failure test below.  */
>>     for (i = 0; i < loop; i++)
>>       {
>>         failed |= function ();
>> @@ -56,10 +68,12 @@ do_test_1 (const char *name, unsigned int loop, int (*prepare) (void),
>>
>>     if (naborts)
>>       {
>> -      /* NB: Low single digit (<= 5%) noise-level aborts are normal for
>> -        TSX.  */
>> +      /* Low single digit (<= 5%) noise-level aborts are normal for TSX
>> +        for reasons that are outside of the library's control.
>> +        In pre-commit CI on an E5-2698 v4 we sometimes see ~5% aborts.
>> +        Set the trip point to 6%.  */
>>         double rate = 100 * ((double) naborts) / ((double) loop);
>> -      if (rate > 5)
>> +      if (rate > 6)
>>          FAIL_EXIT1 ("TSX abort rate: %.2f%% (%d out of %d)",
>>                      rate, naborts, loop);
>>       }
>> --
>> 2.51.1
>>
> 
> LGTM.
> 
> Reviewed-by: H.J. Lu <hjl.tools@gmail.com>

Thanks HJ! Pushed. This should stop the false positives on some of the
pre-commit CI server we're using.

If we trigger this again we can raise the limit or review. I'm not ready
to remove the test because it does a good job identifying when we break
the string and memory operations for *users* of TSX, even if glibc
doesn't make extensive use.

-- 
Cheers,
Carlos.