On Thu 14-06-18 13:42:59, David Rientjes wrote: > The oom reaper ensures forward progress by setting MMF_OOM_SKIP itself if > it cannot reap an mm. This can happen for a variety of reasons, > including: > > - the inability to grab mm->mmap_sem in a sufficient amount of time, > > - when the mm has blockable mmu notifiers that could cause the oom reaper > to stall indefinitely, > > but we can also add a third when the oom reaper can "reap" an mm but doing > so is unlikely to free any amount of memory: > > - when the mm's memory is fully mlocked. > > When all memory is mlocked, the oom reaper will not be able to free any > substantial amount of memory. It sets MMF_OOM_SKIP before the victim can > unmap and free its memory in exit_mmap() and subsequent oom victims are > chosen unnecessarily. This is trivial to reproduce if all eligible > processes on the system have mlocked their memory: the oom killer calls > panic() even though forward progress can be made. > > This is the same issue where the exit path sets MMF_OOM_SKIP before > unmapping memory and additional processes can be chosen unnecessarily > because the oom killer is racing with exit_mmap(). > > We can't simply defer setting MMF_OOM_SKIP, however, because if there is > a true oom livelock in progress, it never gets set and no additional > killing is possible. > > To fix this, this patch introduces a per-mm reaping timeout, initially set > at 10s. It requires that the oom reaper's list becomes a properly linked > list so that other mm's may be reaped while waiting for an mm's timeout to > expire. > > This replaces the current timeouts in the oom reaper: (1) when trying to > grab mm->mmap_sem 10 times in a row with HZ/10 sleeps in between and (2) > a HZ sleep if there are blockable mmu notifiers. It extends it with > timeout to allow an oom victim to reach exit_mmap() before choosing > additional processes unnecessarily. > > The exit path will now set MMF_OOM_SKIP only after all memory has been > freed, so additional oom killing is justified, and rely on MMF_UNSTABLE to > determine when it can race with the oom reaper. > > The oom reaper will now set MMF_OOM_SKIP only after the reap timeout has > lapsed because it can no longer guarantee forward progress. > > The reaping timeout is intentionally set for a substantial amount of time > since oom livelock is a very rare occurrence and it's better to optimize > for preventing additional (unnecessary) oom killing than a scenario that > is much more unlikely. > > Signed-off-by: David Rientjes <rientjes@xxxxxxxxxx> Nacked-by: Michal Hocko <mhocko@xxxxxxxx> as already explained elsewhere in this email thread. > --- > Note: I understand there is an objection based on timeout based delays. > This is currently the only possible way to avoid oom killing important > processes completely unnecessarily. If the oom reaper can someday free > all memory, including mlocked memory and those mm's with blockable mmu > notifiers, and is guaranteed to always be able to grab mm->mmap_sem, > this can be removed. I do not believe any such guarantee is possible > and consider the massive killing of additional processes unnecessarily > to be a regression introduced by the oom reaper and its very quick > setting of MMF_OOM_SKIP to allow additional processes to be oom killed. If you find oom reaper more harmful than useful I would be willing to ack a comman line option to disable it. Especially when you keep claiming that the lockups are not really happening in your environment. Other than that I've already pointed to a more robust solution. If you are reluctant to try it out I will do, but introducing a timeout is just papering over the real problem. Maybe we will not reach the state that _all_ the memory is reapable but we definitely should try to make as much as possible to be reapable and I do not see any fundamental problems in that direction. -- Michal Hocko SUSE Labs
