[PATCH v2 10/19] nptl: Use tidlock when accessing TID on pthread_getaffinity_np

[Florian Weimer]

* Adhemerval Zanella:

>> The pthread_cancel reference looks like a cut-and-paste-bug.
>
> It is, but I think it is applicable. Maybe:
>
>   /* Block all signal, since the lock is not recursive and used on                                                                                             
>      async-signal-safe functions. */

Just mention async-signal-safe, please.  I do not think recursive locks
can achieve async-signal-safety.

>>> +  sigset_t oldmask;
>>> +  __libc_signal_block_all (&oldmask);
>>> +  lll_lock (pd->tidlock, LLL_PRIVATE);
>>> +
>>> +  int res = pd->tid == 0
>>> +	    ? -ESRCH
>>> +	    : INTERNAL_SYSCALL_CALL (sched_getaffinity, pd->tid,
>>> +				     MIN (INT_MAX, cpusetsize), cpuset);
>>> +
>>> +  lll_unlock (pd->tidlock, LLL_PRIVATE);
>>> +  __libc_signal_restore_set (&oldmask);
>>> +
>>> +  if (res < 0)
>>> +    return -res;
>> 
>> ESRCH doesn't look like the right error code here.  Should we return an
>> affinity mask without any bits set?
>
> Why not? Returning anything but an error does not improve things here
> since the information won't make much sense.  Also it follows what other
> symbols is already doing (such as pthread_cancel()).

POSIX reserves ESRCH for using a pthread_t thread ID outside of the
thread lifetime.  In glibc, this is always undefined.  In contrast, the
case above involves a valid thread ID of a thread that has exited.
Returning ESRCH suggests that the thread lifetime has ended, which is
not true.

Thanks,
Florian