On Tue 20-03-18 22:50:21, Tetsuo Handa wrote: > Michal Hocko wrote: > > On Tue 20-03-18 22:30:16, Tetsuo Handa wrote: > > > Michal Hocko wrote: > > > > On Tue 20-03-18 21:52:33, Tetsuo Handa wrote: > > > > > Michal Hocko wrote: > > > > > > A single stack trace in the changelog would be sufficient IMHO. > > > > > > Appart from that. What do you expect users will do about this trace? > > > > > > Sure they will see a path which holds mmap_sem, we will see a bug report > > > > > > but we can hardly do anything about that. We simply cannot drop the lock > > > > > > from that path in 99% of situations. So _why_ do we want to add more > > > > > > information to the log? > > > > > > > > > > This case is blocked at i_mmap_lock_write(). > > > > > > > > But why does i_mmap_lock_write matter for oom_reaping. We are not > > > > touching hugetlb mappings. dup_mmap holds mmap_sem for write which is > > > > the most probable source of the backoff. > > > > > > If i_mmap_lock_write can bail out upon SIGKILL, the OOM victim will be able to > > > release mmap_sem held for write, which helps the OOM reaper not to back off. > > > > There are so many other blocking calls (including allocations) in > > dup_mmap > > Yes. But > > > that I do not really think i_mmap_lock_write is the biggest > > problem. That will be likely the case for other mmap_sem write lockers. > > i_mmap_lock_write() is one of the problems which we could afford fixing. > 8 out of 11 "oom_reaper: unable to reap" messages are blocked at i_mmap_lock_write(). > [...] > > Really I am not sure dumping more information is beneficial here. > > Converting to use killable where we can afford is beneficial. I am no questioning that. I am questioning the additional information because we won't be able to do anything about mmap_sem holder most of the time. Because they tend block on allocations... -- Michal Hocko SUSE Labs
