On Thu 30-04-15 02:27:44, Tetsuo Handa wrote: > Michal Hocko wrote: > > On Wed 29-04-15 08:55:06, Johannes Weiner wrote: > > > What we can do to mitigate this is tie the timeout to the setting of > > > TIF_MEMDIE so that the wait is not 5s from the point of calling > > > out_of_memory() but from the point of where TIF_MEMDIE was set. > > > Subsequent allocations will then go straight to the reserves. > > > > That would deplete the reserves very easily. Shouldn't we rather > > go other way around? Allow OOM killer context to dive into memory > > reserves some more (ALLOC_OOM on top of current ALLOC flags and > > __zone_watermark_ok would allow an additional 1/4 of the reserves) and > > start waiting for the victim after that reserve is depleted. We would > > still have some room for TIF_MEMDIE to allocate, the reserves consumption > > would be throttled somehow and the holders of resources would have some > > chance to release them and allow the victim to die. > > Does OOM killer context mean memory allocations which can call out_of_memory()? Yes, that was the idea, because others will not reclaim any memory. Even all those which invoke out_of_memory will not kill a new task but one killed task should compensate for the ALLOC_OOM part of the memory reserves. > If yes, there is no guarantee that such memory reserve is used by threads which > the OOM victim is waiting for, for they might do only !__GFP_FS allocations. OK, so we are back to GFP_NOFS. Right, those are your main pain point because you can see i_mutex deadlocks. But really, those allocations should simply fail because looping in the allocator and rely on others to make a progress is simply retarded. I thought that Dave was quite explicit that they do not strictly need nofail behavior of GFP_NOFS but rather a GFP flag which would allow to dive into reserves some more for specific contexts (http://marc.info/?l=linux-mm&m=142897087230385&w=2). I also do not remember him or anybody else saying that _every_ GFP_NOFS should get the access to reserves automatically. Dave, could you clarify/confirm, please? Because we are going back and forth about GFP_NOFS without any progress for a very long time already and it seems one class of issues could be handled by this change already. I mean we should eventually fail all the allocation types but GFP_NOFS is coming from _carefully_ handled code paths which is an easier starting point than a random code path in the kernel/drivers. So can we finally move at least in this direction? > Likewise, there is possibility that such memory reserve is used by threads > which the OOM victim is not waiting for, for malloc() + memset() causes > __GFP_FS allocations. We cannot be certain without complete dependency tracking. This is just a heuristic. -- Michal Hocko SUSE Labs -- 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>