On Tue, Jul 26, 2011 at 10:35:17AM +0100, Mel Gorman wrote: > On Mon, Jul 25, 2011 at 10:38:41PM +0200, Johannes Weiner wrote: > > With defrag mode enabled per default, huge page allocations pass > > __GFP_WAIT and may drop compaction into sync-mode where they wait for > > pages under writeback. > > > > I observe applications hang for several minutes(!) when they fault in > > huge pages and compaction starts to wait on in-"flight" USB stick IO. > > > > This patch disables defrag mode for page fault allocations unless the > > VMA is madvised explicitely. Khugepaged will continue to allocate > > with __GFP_WAIT per default, but stalls are not a problem of > > application responsiveness there. > > > > Seems drastic. You could just avoid sync migration for transparent > hugepage allocations with something like the patch below? There still > is a stall as some order-0 pages will be reclaimed before compaction > is tried again but it will nothing like a sync migration. > > === CUT HERE === > diff --git a/mm/page_alloc.c b/mm/page_alloc.c > index 1fac154..40f2a9b 100644 > --- a/mm/page_alloc.c > +++ b/mm/page_alloc.c > @@ -2174,7 +2174,14 @@ rebalance: > sync_migration); > if (page) > goto got_pg; > - sync_migration = true; > + > + /* > + * Do not use sync migration for transparent hugepage allocations as > + * it could stall writing back pages which is far worse than simply > + * failing to promote a page. > + */ > + if (!(gfp_mask & __GFP_NO_KSWAPD)) > + sync_migration = true; For khugepaged it probably makes sense to enter sync migration. But it's less important and could be fixed with an extra GFP flag later, maybe? > As this is USB, the rate of pages getting written back may mean that > too much clean memory is reclaimed in direct reclaim while compaction > still fails due to dirty pages. If this is the case, it can be mitigated > with something like this before calling direct reclaim; > > if ((gfp_mask & __GFP_NO_KSWAPD) && compaction_deferred(preferred_zone)) > goto nopage; Ah, that looks sensible. Thanks, I'll add those hunks to my tree and see how they improve behaviour and keep an eye on the THP statistics. -- 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/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>