On Thu 10-08-17 21:10:30, Tetsuo Handa wrote:
> Michal Hocko wrote:
> > On Tue 08-08-17 11:14:50, Tetsuo Handa wrote:
> > > Michal Hocko wrote:
> > > > On Sat 05-08-17 10:02:55, Tetsuo Handa wrote:
> > > > > Michal Hocko wrote:
> > > > > > On Wed 26-07-17 20:33:21, Tetsuo Handa wrote:
> > > > > > > My question is, how can users know it if somebody was OOM-killed
> > > > > > > needlessly
> > > > > > > by allowing MMF_OOM_SKIP to race.
> > > > > >
> > > > > > Is it really important to know that the race is due to MMF_OOM_SKIP?
> > > > >
> > > > > Yes, it is really important. Needlessly selecting even one OOM victim
> > > > > is
> > > > > a pain which is difficult to explain to and persuade some of
> > > > > customers.
> > > >
> > > > How is this any different from a race with a task exiting an releasing
> > > > some memory after we have crossed the point of no return and will kill
> > > > something?
> > >
> > > I'm not complaining about an exiting task releasing some memory after we
> > > have
> > > crossed the point of no return.
> > >
> > > What I'm saying is that we can postpone "the point of no return" if we
> > > ignore
> > > MMF_OOM_SKIP for once (both this "oom_reaper: close race without using
> > > oom_lock"
> > > thread and "mm, oom: task_will_free_mem(current) should ignore
> > > MMF_OOM_SKIP for
> > > once." thread). These are race conditions we can avoid without crystal
> > > ball.
> > If those races are really that common than we can handle them even
> > without "try once more" tricks. Really this is just an ugly hack. If you
> > really care then make sure that we always try to allocate from memory
> > reserves before going down the oom path. In other words, try to find a
> > robust solution rather than tweaks around a problem.
> Since your "mm, oom: allow oom reaper to race with exit_mmap" patch removes
> oom_lock serialization from the OOM reaper, possibility of calling
> due to successful mutex_trylock(&oom_lock) would increase when the OOM reaper
> MMF_OOM_SKIP quickly.
> What if task_is_oom_victim(current) became true and MMF_OOM_SKIP was set
> on current->mm between after __gfp_pfmemalloc_flags() returned 0 and before
> out_of_memory() is called (due to successful mutex_trylock(&oom_lock)) ?
> Excuse me? Are you suggesting to try memory reserves before
> task_is_oom_victim(current) becomes true?
No what I've tried to say is that if this really is a real problem,
which I am not sure about, then the proper way to handle that is to
attempt to allocate from memory reserves for an oom victim. I would be
even willing to take the oom_lock back into the oom reaper path if the
former turnes out to be awkward to implement. But all this assumes this
is a _real_ problem.
> > [...]
> > > > Yes that is possible. Once you are in the shrinker land then you have to
> > > > count with everything. And if you want to imply that
> > > > get_page_from_freelist inside __alloc_pages_may_oom may lockup while
> > > > holding the oom_lock then you might be right but I haven't checked that
> > > > too deeply. It might be very well possible that the node reclaim bails
> > > > out early when we are under OOM.
> > >
> > > Yes, I worry that get_page_from_freelist() with oom_lock held might
> > > lockup.
> > >
> > > If we are about to invoke the OOM killer for the first time, it is likely
> > > that
> > > __node_reclaim() finds nothing to reclaim and will bail out immediately.
> > > But if
> > > we are about to invoke the OOM killer again, it is possible that small
> > > amount of
> > > memory was reclaimed by the OOM killer/reaper, and all reclaimed memory
> > > was assigned
> > > to things which __node_reclaim() will find and try to reclaim, and any
> > > thread which
> > > took oom_lock will call __node_reclaim() and __node_reclaim() find
> > > something
> > > reclaimable if __GFP_DIRECT_RECLAIM && !__GFP_NORETRY memory allocation
> > > is involved.
> > >
> > > We should consider such situation volatile (i.e. should not make
> > > assumption that
> > > get_page_from_freelist() with oom_lock held shall bail out immediately)
> > > if shrinkers
> > > which (directly or indirectly) involve __GFP_DIRECT_RECLAIM &&
> > > !__GFP_NORETRY memory
> > > allocation are permitted.
> > Well, I think you are so focused on details that you most probably miss
> > a large picture here. Just think about the purpose of the node reclaim.
> > It is there to _prefer_ local allocations than go to a distant NUMA
> > node. So rather than speculating about details maybe it makes sense to
> > consider whether it actually makes any sense to even try to node reclaim
> > when we are OOM. In other words why to do an additional reclaim when we
> > just found out that all reclaim attempts have failed...
> Below is what I will propose if there is possibility of lockup.
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index be5bd60..718b2e7 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -3271,9 +3271,11 @@ void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask,
> const char *fmt, ...)
> * Go through the zonelist yet one more time, keep very high watermark
> * here, this is only to catch a parallel oom killing, we must fail if
> - * we're still under heavy pressure.
> + * we're still under heavy pressure. But make sure that this reclaim
> + * attempt shall not involve __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
> + * allocation which will never fail due to oom_lock already held.
> - page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
> + page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
> ~__GFP_DIRECT_RECLAIM, order,
> ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
> if (page)
> goto out;
get_page_from_freelist doesn't check __GFP_DIRECT_RECLAIM. I was think
something like ALLOC_OOM which would skip node reclaim.