On Sat, Nov 08, 2014 at 11:18:37PM +0100, Vlastimil Babka wrote: > On 11/08/2014 02:11 PM, P. Christeas wrote: > > On Thursday 06 November 2014, Vlastimil Babka wrote: > >>> On Wednesday 05 November 2014, Vlastimil Babka wrote: > >>>> Can you please try the following patch? > >>>> - compaction_defer_reset(zone, order, false); > >> Oh and did I ask in this thread for /proc/zoneinfo yet? :) > > > > Using that same kernel[1], got again into a race, gathered a few more data. > > > > This time, I had 1x "urpmq" process [2] hung at 100% CPU , when "kwin" got > > apparently blocked (100% CPU, too) trying to resize a GUI window. I suppose > > the resizing operation would mean heavy memory alloc/free. > > > > The rest of the system was responsive, I could easily get a console, login, > > gather the files.. Then, I have *killed* -9 the "urpmq" process, which solved > > the race and my system is still alive! "kwin" is still running, returned to > > regular CPU load. > > > > Attached is traces from SysRq+l (pressed a few times, wanted to "snapshot" the > > stack) and /proc/zoneinfo + /proc/vmstat > > > > Bisection is not yet meaningful, IMHO, because I cannot be sure that "good" > > points are really free from this issue. I'd estimate that each test would take > > +3days, unless I really find a deterministic way to reproduce the issue . > > Hi, > > I think I finally found the cause by staring into the code... CCing > people from all 4 separate threads I know about this issue. > The problem with finding the cause was that the first report I got from > Markus was about isolate_freepages_block() overhead, and later Norbert > reported that reverting a patch for isolate_freepages* helped. But the > problem seems to be that although the loop in isolate_migratepages exits > because the scanners almost meet (they are within same pageblock), they > don't truly meet, therefore compact_finished() decides to continue, but > isolate_migratepages() exits immediately... boom! But indeed e14c720efdd7 > made this situation possible, as free scaner pfn can now point to a > middle of pageblock. Indeed. > > So I hope the attached patch will fix the soft-lockup issues in > compact_zone. Please apply on 3.18-rc3 or later without any other reverts, > and test. It probably won't help Markus and his isolate_freepages_block() > overhead though... Yes, I found this bug too, but, it can't explain isolate_freepages_block() overhead. Anyway, I can't find another bug related to isolate_freepages_block(). :/ > Thanks, > Vlastimil > > ------8<------ > >From fbf8eb0bcd2897090312e23da6a31bad9cc6b337 Mon Sep 17 00:00:00 2001 > From: Vlastimil Babka <vbabka@xxxxxxx> > Date: Sat, 8 Nov 2014 22:20:43 +0100 > Subject: [PATCH] mm, compaction: prevent endless loop in migrate scanner > > --- > mm/compaction.c | 8 ++++++-- > 1 file changed, 6 insertions(+), 2 deletions(-) > > diff --git a/mm/compaction.c b/mm/compaction.c > index ec74cf0..1b7a1be 100644 > --- a/mm/compaction.c > +++ b/mm/compaction.c > @@ -1029,8 +1029,12 @@ static isolate_migrate_t isolate_migratepages(struct zone *zone, > } > > acct_isolated(zone, cc); > - /* Record where migration scanner will be restarted */ > - cc->migrate_pfn = low_pfn; > + /* > + * Record where migration scanner will be restarted. If we end up in > + * the same pageblock as the free scanner, make the scanners fully > + * meet so that compact_finished() terminates compaction. > + */ > + cc->migrate_pfn = (end_pfn <= cc->free_pfn) ? low_pfn : cc->free_pfn; > > return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE; > } IMHO, proper fix is not to change this logic, but, to change decision logic in compact_finished() and in compact_zone(). Maybe helper function would be good for readability. Thanks. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>