On Tue 28-01-20 00:30:44, Matthew Wilcox wrote: > On Tue, Jan 28, 2020 at 09:17:12AM +0100, Michal Hocko wrote: > > On Mon 27-01-20 11:06:53, Matthew Wilcox wrote: > > > On Mon, Jan 27, 2020 at 04:00:24PM +0100, Michal Hocko wrote: > > > > On Sun 26-01-20 15:39:35, Matthew Wilcox wrote: > > > > > On Sun, Jan 26, 2020 at 11:53:55AM -0800, Cong Wang wrote: > > > > > > I suspect the process gets stuck in the retry loop in try_charge(), as > > > > > > the _shortest_ stacktrace of the perf samples indicated: > > > > > > > > > > > > cycles:ppp: > > > > > > ffffffffa72963db mem_cgroup_iter > > > > > > ffffffffa72980ca mem_cgroup_oom_unlock > > > > > > ffffffffa7298c15 try_charge > > > > > > ffffffffa729a886 mem_cgroup_try_charge > > > > > > ffffffffa720ec03 __add_to_page_cache_locked > > > > > > ffffffffa720ee3a add_to_page_cache_lru > > > > > > ffffffffa7312ddb iomap_readpages_actor > > > > > > ffffffffa73133f7 iomap_apply > > > > > > ffffffffa73135da iomap_readpages > > > > > > ffffffffa722062e read_pages > > > > > > ffffffffa7220b3f __do_page_cache_readahead > > > > > > ffffffffa7210554 filemap_fault > > > > > > ffffffffc039e41f __xfs_filemap_fault > > > > > > ffffffffa724f5e7 __do_fault > > > > > > ffffffffa724c5f2 __handle_mm_fault > > > > > > ffffffffa724cbc6 handle_mm_fault > > > > > > ffffffffa70a313e __do_page_fault > > > > > > ffffffffa7a00dfe page_fault > > > > > > > > > > > > But I don't see how it could be, the only possible case is when > > > > > > mem_cgroup_oom() returns OOM_SUCCESS. However I can't > > > > > > find any clue in dmesg pointing to OOM. These processes in the > > > > > > same memcg are either running or sleeping (that is not exiting or > > > > > > coredump'ing), I don't see how and why they could be selected as > > > > > > a victim of OOM killer. I don't see any signal pending either from > > > > > > their /proc/X/status. > > > > > > > > > > I think this is a situation where we might end up with a genuine deadlock > > > > > if we're not trylocking the pages. readahead allocates a batch of > > > > > locked pages and adds them to the pagecache. If it has allocated, > > > > > say, 5 pages, successfully inserted the first three into i_pages, then > > > > > needs to allocate memory to insert the fourth one into i_pages, and > > > > > the process then attempts to migrate the pages which are still locked, > > > > > they will never come unlocked because they haven't yet been submitted > > > > > to the filesystem for reading. > > > > > > > > Just to make sure I understand. Do you mean this? > > > > lock_page(A) > > > > alloc_pages > > > > try_to_compact_pages > > > > compact_zone_order > > > > compact_zone(MIGRATE_SYNC_LIGHT) > > > > migrate_pages > > > > unmap_and_move > > > > __unmap_and_move > > > > lock_page(A) > > > > > > Yes. There's a little more to it than that, eg slab is involved, but > > > you have it in a nutshell. > > > > I am not deeply familiar with the readahead code. But is there really a > > high oerder allocation (order > 1) that would trigger compaction in the > > phase when pages are locked? > > Thanks to sl*b, yes: > > radix_tree_node 80890 102536 584 28 4 : tunables 0 0 0 : slabdata 3662 3662 0 > > so it's allocating 4 pages for an allocation of a 576 byte node. I am not really sure that we do sync migration for costly orders. > > Btw. the compaction rejects to consider file backed pages when __GFP_FS > > is not present AFAIR. > > Ah, that would save us. So the NOFS comes from the mapping GFP mask, right? That is something I was hoping to get rid of eventually :/ Anyway it would be better to have an explicit NOFS with a comment explaining why we need that. If for nothing else then for documentation. -- Michal Hocko SUSE Labs
