On Sun, Sep 18, 2016 at 1:26 PM, Lorenzo Stoakes [off-list ref] 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@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>