Michal Hocko wrote:
> That being said, I will keep refusing other such tweaks unless you have
> a sound usecase behind. If you really _want_ to help out here then you
> can focus on the reaping of the mlock memory.
Not the reaping of the mlock'ed memory. Although Manish's report was mlock'ed
case, there are other cases (e.g. MAP_SHARED, mmu_notifier, mmap_sem held for
write) which can lead to this race condition. If we think about artificial case,
it would be possible to run 1024 threads not sharing signal_struct but consume
almost 0KB memory (i.e. written without using C library) and many of them are
running between __gfp_pfmemalloc_flags() and mutex_trylock() waiting for
ALLOC_OOM.
What the Manish's report revealed is the fact that we accidentally broke the
/*
* Kill all user processes sharing victim->mm in other thread groups, if
* any. They don't get access to memory reserves, though, to avoid
* depletion of all memory. This prevents mm->mmap_sem livelock when an
* oom killed thread cannot exit because it requires the semaphore and
* its contended by another thread trying to allocate memory itself.
* That thread will now get access to memory reserves since it has a
* pending fatal signal.
*/
assumption via 696453e66630ad45 ("mm, oom: task_will_free_mem should skip
oom_reaped tasks"), and we already wasted for 16 months. There is no need to
wait for fixing mlock'ed, MAP_SHARED, mmu_notifier and mmap_sem cases because
"OOM victims consuming almost 0KB memory" case cannot be solved.
The mlock'ed, MAP_SHARED, mmu_notifier and mmap_sem cases are a sort of alias
of "OOM victims consuming almost 0KB memory" case.
Anyway, since you introduced MMF_OOM_VICTIM flag, I will try a patch which
checks MMF_OOM_VICTIM instead of oom_reserves_allowed().
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
