On Wed 22-06-16 08:40:15, Michal Hocko wrote: > On Wed 22-06-16 06:47:48, Tetsuo Handa wrote: > > Michal Hocko wrote: > > > On Wed 22-06-16 00:32:29, Tetsuo Handa wrote: > > > > Michal Hocko wrote: > > > [...] > > > > > Hmm, what about the following instead. It is rather a workaround than > > > > > a > > > > > full flaged fix but it seems much more easier and shouldn't introduce > > > > > new issues. > > > > > > > > Yes, I think that will work. But I think below patch (marking > > > > signal_struct > > > > to ignore TIF_MEMDIE instead of clearing TIF_MEMDIE from task_struct) > > > > on top of > > > > current linux.git will implement no-lockup requirement. No race is > > > > possible unlike > > > > "[PATCH 10/10] mm, oom: hide mm which is shared with kthread or global > > > > init". > > > > > > Not really. Because without the exit_oom_victim from oom_reaper you have > > > no guarantee that the oom_killer_disable will ever return. I have > > > mentioned that in the changelog. There is simply no guarantee the oom > > > victim will ever reach exit_mm->exit_oom_victim. > > > > Why? Since any allocation after setting oom_killer_disabled = true will be > > forced to fail, nobody will be blocked on waiting for memory allocation. > > Thus, > > the TIF_MEMDIE tasks will eventually reach exit_mm->exit_oom_victim, won't > > it? > > What if it gets blocked waiting for an operation which cannot make any > forward progress because it cannot proceed with an allocation (e.g. > an open coded allocation retry loop - not that uncommon when sending > a bio)? I mean if we want to guarantee a forward progress then there has > to be something to clear the flag no matter in what state the oom victim > is or give up on oom_killer_disable.
That being said I guess the patch to try_to_freeze_tasks after oom_killer_disable should be simple enough to go for now and stable trees and we can come up with something less hackish later. I do not like the fact that oom_killer_disable doesn't act as a full "barrier" anymore. What do you think? -- Michal Hocko SUSE Labs
