Re: Unending loop in __alloc_pages_slowpath following OOM-kill; rfc: patch.

[Minchan Kim <minchan.kim@gmail.com>]

On Thu, May 19, 2011 at 07:29:01AM +0900, Minchan Kim wrote:
> Hi Andrew,
> 
> On Wed, May 18, 2011 at 12:49 AM, Andrew Barry <abarry@cray.com> wrote:
> > On 05/17/2011 05:34 AM, Minchan Kim wrote:
> >> On Sat, May 14, 2011 at 6:31 AM, Andrew Barry <abarry@cray.com> wrote:
> >>> I believe I found a problem in __alloc_pages_slowpath, which allows a process to
> >>> get stuck endlessly looping, even when lots of memory is available.
> >>>
> >>> Running an I/O and memory intensive stress-test I see a 0-order page allocation
> >>> with __GFP_IO and __GFP_WAIT, running on a system with very little free memory.
> >>> Right about the same time that the stress-test gets killed by the OOM-killer,
> >>> the utility trying to allocate memory gets stuck in __alloc_pages_slowpath even
> >>> though most of the systems memory was freed by the oom-kill of the stress-test.
> >>>
> >>> The utility ends up looping from the rebalance label down through the
> >>> wait_iff_congested continiously. Because order=0, __alloc_pages_direct_compact
> >>> skips the call to get_page_from_freelist. Because all of the reclaimable memory
> >>> on the system has already been reclaimed, __alloc_pages_direct_reclaim skips the
> >>> call to get_page_from_freelist. Since there is no __GFP_FS flag, the block with
> >>> __alloc_pages_may_oom is skipped. The loop hits the wait_iff_congested, then
> >>> jumps back to rebalance without ever trying to get_page_from_freelist. This loop
> >>> repeats infinitely.
> >>>
> >>> Is there a reason that this loop is set up this way for 0 order allocations? I
> >>> applied the below patch, and the problem corrects itself. Does anyone have any
> >>> thoughts on the patch, or on a better way to address this situation?
> >>>
> >>> The test case is pretty pathological. Running a mix of I/O stress-tests that do
> >>> a lot of fork() and consume all of the system memory, I can pretty reliably hit
> >>> this on 600 nodes, in about 12 hours. 32GB/node.
> >>>
> >>
> >> It's amazing.
> >> I think it's _very_ rare but it's possible if test program killed by
> >> oom has only lots of anonymous pages and allocation tasks try to
> >> allocate order-0 page with GFP_NOFS.
> >
> > Unfortunately very rare is a subjective thing. We have been hitting it a couple
> > times a week in our test lab.
> 
> Okay.
> 
> >
> >> When the [in]active lists are empty suddenly(But I am not sure how
> >> come the situation happens.) and we are reclaiming order-0 page,
> >> compaction and __alloc_pages_direct_reclaim doesn't work. compaction
> >> doesn't work as it's order-0 page reclaiming.  In case of
> >> __alloc_pages_direct_reclaim, it would work only if we have lru pages
> >> in [in]active list. But unfortunately we don't have any pages in lru
> >> list.
> >> So, last resort is following codes in do_try_to_free_pages.
> >>
> >>         /* top priority shrink_zones still had more to do? don't OOM, then */
> >>         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
> >>                 return 1;
> >>
> >> But it has a problem, too. all_unreclaimable checks zone->all_unreclaimable.
> >> zone->all_unreclaimable is set by below condition.
> >>
> >> zone->pages_scanned < zone_reclaimable_pages(zone) * 6
> >>
> >> If lru list is completely empty, shrink_zone doesn't work so
> >> zone->pages_scanned would be zero. But as we know, zone_page_state
> >> isn't exact by per_cpu_pageset. So it might be positive value. After
> >> all, zone_reclaimable always return true. It means kswapd never set
> >> zone->all_unreclaimable.  So last resort become nop.
> >>
> >> In this case, current allocation doesn't have a chance to call
> >> get_page_from_freelist as Andrew Barry said.
> >>
> >> Does it make sense?
> >> If it is, how about this?
> >>
> >> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> >> index ebc7faa..4f64355 100644
> >> --- a/mm/page_alloc.c
> >> +++ b/mm/page_alloc.c
> >> @@ -2105,6 +2105,7 @@ restart:
> >>                 first_zones_zonelist(zonelist, high_zoneidx, NULL,
> >>                                         &preferred_zone);
> >>
> >> +rebalance:
> >>         /* This is the last chance, in general, before the goto nopage. */
> >>         page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
> >>                         high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
> >> @@ -2112,7 +2113,6 @@ restart:
> >>         if (page)
> >>                 goto got_pg;
> >>
> >> -rebalance:
> >>         /* Allocate without watermarks if the context allows */
> >>         if (alloc_flags & ALLOC_NO_WATERMARKS) {
> >>                 page = __alloc_pages_high_priority(gfp_mask, order,
> >
> > I think your solution is simpler than my patch.
> > Thanks very much.
> 
> You find the problem and it's harder than fix, I think.
> So I think you have to get a credit.
> 
> Could you send the patch to akpm with Cced Mel and me?
> (Maybe it's the subject to send stable).
> You can get my Reviewed-by.
> 
> Thanks for the good bug reporting.