On Thu 02-11-17 00:08:59, Tetsuo Handa wrote: > Michal Hocko wrote: [...] > > I am not sure about this part though. If the oom_reaper cannot take the > > mmap_sem then it retries for 1s. Have you ever seen the race to be that > > large? > > Like shown in [2], khugepaged can prevent oom_reaper from taking the mmap_sem > for 1 second. Also, it won't be impossible for OOM victims to spend 1 second > between post __gfp_pfmemalloc_flags(gfp_mask) and pre mutex_trylock(&oom_lock) > (in other words, the race window (1-2) above). Therefore, non artificial > workloads could hit the same result.
but this is a speculation so I wouldn't mention it in the changelog. It might confuse readers. -- Michal Hocko SUSE Labs
