Fix PR51298, libgomp barrier failure

On Mon, Nov 28, 2011 at 08:09:01PM +1030, Alan Modra wrote:
> On Mon, Nov 28, 2011 at 09:13:27AM +0100, Jakub Jelinek wrote:
> > On Mon, Nov 28, 2011 at 10:12:09AM +1030, Alan Modra wrote:
> > > --- libgomp/config/linux/bar.c	(revision 181718)
> > > +++ libgomp/config/linux/bar.c	(working copy)
> > > @@ -36,18 +36,15 @@ gomp_barrier_wait_end (gomp_barrier_t *b
> > >    if (__builtin_expect ((state & 1) != 0, 0))
> > >      {
> > >        /* Next time we'll be awaiting TOTAL threads again.  */
> > > -      bar->awaited = bar->total;
> > > -      atomic_write_barrier ();
> > > -      bar->generation += 4;
> > > +      __atomic_store_4 (&bar->awaited, bar->total, MEMMODEL_RELEASE);
> > > +      __atomic_add_fetch (&bar->generation, 4, MEMMODEL_ACQ_REL);
> > >        futex_wake ((int *) &bar->generation, INT_MAX);
> > >      }
> > 
> > The above is unfortunate, bar->generation should be only modified from a
> > single thread at this point, but the __atomic_add_fetch will enforce there
> > an atomic instruction on it rather than just some kind of barrier.
> 
> I will try without the atomic add and see what happens.

Seems to be fine.  I took out the extra write barriers too, so we now
just have two MEMMODEL_RELEASE atomic store barriers replacing the two
old atomic_write_barriers, and a number of new acquire barriers.

	PR libgomp/51298
	* config/linux/bar.h: Use atomic rather than sync builtins.
	* config/linux/bar.c: Likewise.  Add missing acquire
	synchronisation on generation field.
	* task.c (gomp_barrier_handle_tasks): Regain lock so as to not
	double unlock.

On 11/28/2011 06:02 AM, Alan Modra wrote:
> -  unsigned int ret = bar->generation & ~3;
> -  /* Do we need any barrier here or is __sync_add_and_fetch acting
> -     as the needed LoadLoad barrier already?  */
> -  ret += __sync_add_and_fetch (&bar->awaited, -1) == 0;
> +  unsigned int ret = __atomic_load_4 (&bar->generation, MEMMODEL_ACQUIRE) & ~3;
> +  ret += __atomic_add_fetch (&bar->awaited, -1, MEMMODEL_ACQ_REL) == 0;

Given that the read from bar->generation is ACQ, we don't need a duplicate 
barrier from the REL on the atomic add.  I believe both can be MEMMODEL_ACQUIRE
both in order to force the ordering of these two memops, as well as force these
to happen before anything subsequent.

The s/_4/_n/ change needs doing.

Otherwise ok.


r~

On Mon, Nov 28, 2011 at 05:42:15PM -0800, Richard Henderson wrote:
> On 11/28/2011 06:02 AM, Alan Modra wrote:
> > -  unsigned int ret = bar->generation & ~3;
> > -  /* Do we need any barrier here or is __sync_add_and_fetch acting
> > -     as the needed LoadLoad barrier already?  */
> > -  ret += __sync_add_and_fetch (&bar->awaited, -1) == 0;
> > +  unsigned int ret = __atomic_load_4 (&bar->generation, MEMMODEL_ACQUIRE) & ~3;
> > +  ret += __atomic_add_fetch (&bar->awaited, -1, MEMMODEL_ACQ_REL) == 0;
> 
> Given that the read from bar->generation is ACQ, we don't need a duplicate 
> barrier from the REL on the atomic add.  I believe both can be MEMMODEL_ACQUIRE
> both in order to force the ordering of these two memops, as well as force these
> to happen before anything subsequent.

I tried with MEMMODEL_ACQUIRE and ran force-parallel-6.exe, the test
that seems most sensitive to barrier problems, many times, and it hangs
occasionally in futex_wait called via gomp_team_barrier_wait_end.

I believe that threads can't be allowed to exit from
gomp_{,team_}barrier_wait without hitting a release barrier, and
perhaps from gomp_barrier_wait_last too.  gomp_barrier_wait_start is a
convenient point to insert the barrier, and a minimal change from the
old code using __sync_add_and_fetch.  I can add a comment. ;-)

On Tue, Nov 29, 2011 at 04:15:36PM +1030, Alan Modra wrote:
> On Mon, Nov 28, 2011 at 05:42:15PM -0800, Richard Henderson wrote:
> > On 11/28/2011 06:02 AM, Alan Modra wrote:
> > > -  unsigned int ret = bar->generation & ~3;
> > > -  /* Do we need any barrier here or is __sync_add_and_fetch acting
> > > -     as the needed LoadLoad barrier already?  */
> > > -  ret += __sync_add_and_fetch (&bar->awaited, -1) == 0;
> > > +  unsigned int ret = __atomic_load_4 (&bar->generation, MEMMODEL_ACQUIRE) & ~3;
> > > +  ret += __atomic_add_fetch (&bar->awaited, -1, MEMMODEL_ACQ_REL) == 0;
> > 
> > Given that the read from bar->generation is ACQ, we don't need a duplicate 
> > barrier from the REL on the atomic add.  I believe both can be MEMMODEL_ACQUIRE
> > both in order to force the ordering of these two memops, as well as force these
> > to happen before anything subsequent.
> 
> I tried with MEMMODEL_ACQUIRE and ran force-parallel-6.exe, the test
> that seems most sensitive to barrier problems, many times, and it hangs
> occasionally in futex_wait called via gomp_team_barrier_wait_end.
> 
> I believe that threads can't be allowed to exit from
> gomp_{,team_}barrier_wait without hitting a release barrier, and
> perhaps from gomp_barrier_wait_last too.  gomp_barrier_wait_start is a
> convenient point to insert the barrier, and a minimal change from the
> old code using __sync_add_and_fetch.  I can add a comment. ;-)

Committed rev 181833.