Re: [RFC PATCH 0/3] rework mmap-exit vs. oom_reaper handover

[Michal Hocko <mhocko@kernel.org>]

On Wed 12-09-18 16:58:53, Tetsuo Handa wrote:
> Michal Hocko wrote:
> > OK, I will fold the following to the patch
> 
> OK. But at that point, my patch which tries to wait for reclaimed memory
> to be re-allocatable addresses a different problem which you are refusing.

I am trying to address a real world example of when the excessive amount
of memory is in page tables. As David pointed, this can happen with some
userspace allocators.

> By the way, is it guaranteed that vma->vm_ops->close(vma) in remove_vma() never
> sleeps? Since remove_vma() has might_sleep() since 2005, and that might_sleep()
> predates the git history, I don't know what that ->close() would do.

Hmm, I am afraid we cannot assume anything so we have to consider it
unsafe. A cursory look at some callers shows that they are taking locks.
E.g. drm_gem_object_put_unlocked might take a mutex. So MMF_OOM_SKIP
would have to set right after releasing page tables.

> Anyway, please fix free_pgd_range() crash in this patchset.

I will try to get to this later today.
-- 
Michal Hocko
SUSE Labs