On Wed, Nov 6, 2013 at 4:35 PM, David Rientjes <rientjes@xxxxxxxxxx> wrote: > On Wed, 6 Nov 2013, Sameer Nanda wrote: > >> David -- I think we can make the duration that the tasklist_lock is >> held smaller by consolidating the process selection logic that is >> currently split across select_bad_process and oom_kill_process into >> one place in select_bad_process. The tasklist_lock would then need to >> be held only when the thread lists are being traversed. Would you be >> ok with that? I can re-spin the patch if that sounds like a workable >> option. >> > > No, this caused hundreds of machines to hit soft lockups for Google > because there's no synchronization that prevents dozens of cpus to take > tasklist_lock in the oom killer during parallel memcg oom conditions and > never allow the write_lock_irq() on fork() or exit() to make progress. We > absolutely must hold tasklist_lock for as little time as possible in the > oom killer. > > That said, I've never actually seen your reported bug manifest in our > production environment so let's see if Oleg has any ideas. Is the path you are referring to mem_cgroup_out_of_memory calling oom_kill_process? If so, then that path doesn't appear to suffer from the two step select_bad_process, oom_kill_process race since mem_cgroup_out_of_memory directly calls oom_kill_process without going through select_bad_process. This also means that the patch I sent is incorrect since it removes the existing tasklist_lock protection in oom_kill_process. Respinning patch to take care of this case. -- Sameer -- 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>