Re: Dead code in pthread_cond_wait() for spin-wait

[Florian Weimer <fw@deneb.enyo.de>]

* Olivier Dion:

> On Wed, 11 Oct 2023, Adhemerval Zanella Netto <adhemerval.zanella@linaro.org> wrote:
>> On 11/10/23 15:30, Olivier Dion via Libc-help wrote:
>
> [...]
>
>>> However, looking at the following snippet in nptl/pthread_cond_wait.c:
>>> 
>>>  static __always_inline int
>>>  __pthread_cond_wait_common (pthread_cond_t *cond, pthread_mutex_t *mutex,
>>>      clockid_t clockid, const struct __timespec64 *abstime)
>>>  {
>>>    const int maxspin = 0;
>>>    [...]
>>>            /* Spin-wait first. [...] */
>>> 	   unsigned int spin = maxspin;
>>> 	   while (signals == 0 && spin > 0) {
>>>              [...]
>>>            }
>>>    [...]
>>>  }
>>> 
>>
>> Afaik it was added a placeholder for future extension, so fell free to
>> send a patch to delete it.
>
> [...]
>
> What about making it work instead?  Would this be something to consider?
> IMHO it would be worth it.  Having a fast-path that spins instead of
> taking the futex directly can greatly impact performance in some cases.

Changes to the condvar code are very difficult for us.  There's an
open bug with patches that have been under review for years:

  pthread_cond_signal failed to wake up pthread_cond_wait due to a bug
  in undoing stealing
  <https://sourceware.org/bugzilla/show_bug.cgi?id=25847>