Re: [PATCH] private futex support

[Rich Felker <dalias@libc.org>]

Just a quick explanation for some unexpected behavior you saw:

On Fri, Aug 08, 2014 at 11:38:57AM +0300, Timo Teras wrote:
> diff --git a/src/thread/pthread_cond_broadcast.c b/src/thread/pthread_cond_broadcast.c
> index 0901daf..4327a0e 100644
> --- a/src/thread/pthread_cond_broadcast.c
> +++ b/src/thread/pthread_cond_broadcast.c
> @@ -27,7 +27,7 @@ int pthread_cond_broadcast(pthread_cond_t *c)
>  
>  	/* Perform the futex requeue, waking one waiter unless we know
>  	 * that the calling thread holds the mutex. */
> -	__syscall(SYS_futex, &c->_c_seq, FUTEX_REQUEUE,
> +	__syscall(SYS_futex, &c->_c_seq, 128 | FUTEX_REQUEUE,
>  		!m->_m_type || (m->_m_lock&INT_MAX)!=__pthread_self()->tid,
>  		INT_MAX, &m->_m_lock);
>  
> diff --git a/src/thread/pthread_cond_timedwait.c b/src/thread/pthread_cond_timedwait.c
> index 99d62cc..73c1781 100644
> --- a/src/thread/pthread_cond_timedwait.c
> +++ b/src/thread/pthread_cond_timedwait.c
> @@ -64,7 +64,7 @@ int pthread_cond_timedwait(pthread_cond_t *restrict c, pthread_mutex_t *restrict
>  
>  	pthread_mutex_unlock(m);
>  
> -	do e = __timedwait(&c->_c_seq, seq, c->_c_clock, ts, cleanup, &cm, 0);
> +	do e = __timedwait(&c->_c_seq, seq, c->_c_clock, ts, cleanup, &cm, 1);
>  	while (c->_c_seq == seq && (!e || e==EINTR));
>  	if (e == EINTR) e = 0;
>  
> It seems that things worked even without these changes, but everything
> was a lot slower. Not sure why. But the code sounds buggy anyway,

Both the wait and wake (requeue) operations for this futex were
wrongly shared rather than private, so since they still matched it
makes sense that it didn't fail. The requeue was wrongly requeuing as
shared on a mutex that would only receive a wake as private, but since
there was only one waiter and the requeue requested one wake, I guess
no requeue ever happened.

Rich