On Thu, Jul 11, 2019 at 09:12:45AM +0200, Michal Hocko wrote: > On Wed 10-07-19 16:36:58, Mike Kravetz wrote: > > On 7/10/19 12:44 PM, Michal Hocko wrote: > > > On Wed 10-07-19 11:42:40, Mike Kravetz wrote: > > > [...] > > >> As Michal suggested, I'm going to do some testing to see what impact > > >> dropping the __GFP_RETRY_MAYFAIL flag for these huge page allocations > > >> will have on the number of pages allocated. > > > > > > Just to clarify. I didn't mean to drop __GFP_RETRY_MAYFAIL from the > > > allocation request. I meant to drop the special casing of the flag in > > > should_continue_reclaim. I really have hard time to argue for this > > > special casing TBH. The flag is meant to retry harder but that shouldn't > > > be reduced to a single reclaim attempt because that alone doesn't really > > > help much with the high order allocation. It is more about compaction to > > > be retried harder. > > > > Thanks Michal. That is indeed what you suggested earlier. I remembered > > incorrectly. Sorry. > > > > Removing the special casing for __GFP_RETRY_MAYFAIL in should_continue_reclaim > > implies that it will return false if nothing was reclaimed (nr_reclaimed == 0) > > in the previous pass. > > > > When I make such a modification and test, I see long stalls as a result > > of should_compact_retry returning true too often. On a system I am currently > > testing, should_compact_retry has returned true 36000000 times. My guess > > is that this may stall forever. Vlastmil previously asked about this behavior, > > so I am capturing the reason. Like before [1], should_compact_retry is > > returning true mostly because compaction_withdrawn() returns COMPACT_DEFERRED. > > This smells like a problem to me. But somebody more familiar with > compaction should comment. > Examine in should_compact_retry if it's retrying because compaction_zonelist_suitable is true. Looking at it now, it would not necessarily do the right thing because any non-skipped zone would make it eligible which is too strong a condition as COMPACT_SKIPPED is not reliably set. If that function is the case, it would be reasonable remove "ret = compaction_zonelist_suitable(ac, order, alloc_flags);" and the implementation of compaction_zonelist_suitable entirely as part of your fix. -- Mel Gorman SUSE Labs
