On Sun, Sep 18, 2016 at 1:26 PM, Lorenzo Stoakes <lstoakes@xxxxxxxxx> wrote: > > I encountered this even after applying the patch discussed in the > original thread at https://lkml.org/lkml/2016/8/22/184. It's not easily > reproducible but it is happening enough that I could probably check some > specific state when it next occurs or test out a patch to see if it > stops it if that'd be useful. Since you can at least try to recreate it, how about the series in -mm by Vlastimil? The series was called "reintroduce compaction feedback for OOM decisions", and is in -mm right now: Vlastimil Babka (4): Revert "mm, oom: prevent premature OOM killer invocation for high order request" mm, compaction: more reliably increase direct compaction priority mm, compaction: restrict full priority to non-costly orders mm, compaction: make full priority ignore pageblock suitability I'm not sure if Andrew has any other ones pending that are relevant to oom. A lot of the oom discussion seemed to be about the task stack allocation (order-2), but kmalloc() really can and does trigger those order-3 allocations even for small allocations. Just as an example, these are the slab entries for me that are order-3: bio-1, UDPv6, TCPv6, kcopyd_job, dm_uevent, mqueue_inode_cache, ext4_inode_cache, pid_namespace, PING, UDP, TCP, request_queue, net_namespace, bdev_cache, mm_struct, signal_cache, sighand_cache, task_struct, idr_layer_cache, dma-kmalloc-8192, dma-kmalloc-4096, dma-kmalloc-2048, dma-kmalloc-1024, kmalloc-8192, kmalloc-4096, kmalloc-2048, kmalloc-1024 and most of those are 1-2kB in size. Of course, any slab allocation failure is harder to trigger just because slab itself ends up often having empty cache entries, so only a small percentage makes it to the page allocator itself. But the page allocator failure case really needs to treat PAGE_ALLOC_COSTLY_ORDER specially. Which implies that if compaction is magical for page allocation success, then compaction needs to do so too. Linus -- 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>