On Fri 19-12-14 21:28:15, Vladimir Davydov wrote:
> Hi,
>
> On Fri, Dec 19, 2014 at 04:57:47PM +0100, Michal Hocko wrote:
> > On Fri 19-12-14 14:01:55, Vlastimil Babka wrote:
> > > Charles Shirron and Paul Cassella from Cray Inc have reported kswapd stuck
> > > in a busy loop with nothing left to balance, but kswapd_try_to_sleep() failing
> > > to sleep. Their analysis found the cause to be a combination of several
> > > factors:
> > >
> > > 1. A process is waiting in throttle_direct_reclaim() on pgdat->pfmemalloc_wait
> > >
> > > 2. The process has been killed (by OOM in this case), but has not yet been
> > > scheduled to remove itself from the waitqueue and die.
> >
> > pfmemalloc_wait is used as wait_event and that one uses
> > autoremove_wake_function for wake ups so the task shouldn't stay on the
> > queue if it was woken up. Moreover pfmemalloc_wait sleeps are killable
> > by the OOM killer AFAICS.
> >
> > $ git grep "wait_event.*pfmemalloc_wait"
> > mm/vmscan.c:
> > wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
> > mm/vmscan.c: wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,))
> >
> > So OOM killer would wake it up already and kswapd shouldn't see this
> > task on the waitqueue anymore.
>
> OOM killer will wake up the process, but it won't remove it from the
> pfmemalloc_wait queue. Therefore, if kswapd gets scheduled before the
> dying process, it will see the wait queue being still active, but won't
> be able to wake anyone up, because the waiting process has already been
> woken by SIGKILL. I think this is what Vlastimil means.
OK, I see the point now. I didn't realize that autoremove_wake_function
doesn't remove the waiter from the queue if the state doesn't change.
> So AFAIU the problem does exist. However, I think it could be fixed by
> simply waking up all processes waiting on pfmemalloc_wait before putting
> kswapd to sleep:
I think that a simple cond_resched() in kswapd_try_to_sleep should be
sufficient and less risky fix, so basically what Vlastimil was proposing
in the beginning.
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index 744e2b491527..2a123634c220 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -2984,6 +2984,9 @@ static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
> if (remaining)
> return false;
>
> + if (!pgdat_balanced(pgdat, order, classzone_idx))
> + return false;
> +
What would be consequences of not waking up pfmemalloc waiters while the
node is not balanced?
> /*
> * There is a potential race between when kswapd checks its watermarks
> * and a process gets throttled. There is also a potential race if
> @@ -2993,12 +2996,9 @@ static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
> * so wake them now if necessary. If necessary, processes will wake
> * kswapd and get throttled again
> */
> - if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
> - wake_up(&pgdat->pfmemalloc_wait);
> - return false;
> - }
> + wake_up_all(&pgdat->pfmemalloc_wait);
>
> - return pgdat_balanced(pgdat, order, classzone_idx);
> + return true;
> }
>
> /*
--
Michal Hocko
SUSE Labs
--
To unsubscribe from this list: send the line "unsubscribe stable" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html