On Thu, Dec 30, 2021 at 12:24 AM Michal Hocko <mhocko@xxxxxxxx> wrote:
>
> On Wed 29-12-21 21:59:55, Suren Baghdasaryan wrote:
> [...]
> > After some more digging I think there are two acceptable options:
> >
> > 1. Call unlock_range() under mmap_write_lock and then downgrade it to
> > read lock so that both exit_mmap() and __oom_reap_task_mm() can unmap
> > vmas in parallel like this:
> >
> > if (mm->locked_vm) {
> > mmap_write_lock(mm);
> > unlock_range(mm->mmap, ULONG_MAX);
> > mmap_write_downgrade(mm);
> > } else
> > mmap_read_lock(mm);
> > ...
> > unmap_vmas(&tlb, vma, 0, -1);
> > mmap_read_unlock(mm);
> > mmap_write_lock(mm);
> > free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
> > ...
> > mm->mmap = NULL;
> > mmap_write_unlock(mm);
> >
> > This way exit_mmap() might block __oom_reap_task_mm() but for a much
> > shorter time during unlock_range() call.
>
> IIRC unlock_range depends on page lock at some stage and that can mean
> this will block for a long time or for ever when the holder of the lock
> depends on a memory allocation. This was the primary problem why the oom
> reaper skips over mlocked vmas.
Oh, I missed that detail. I thought __oom_reap_task_mm() skips locked
vmas only to avoid destroying pgds from under follow_page().
>
> > 2. Introduce another vm_flag mask similar to VM_LOCKED which is set
> > before munlock_vma_pages_range() clears VM_LOCKED so that
> > __oom_reap_task_mm() can identify vmas being unlocked and skip them.
> >
> > Option 1 seems cleaner to me because it keeps the locking pattern
> > around unlock_range() in exit_mmap() consistent with all other places
> > it is used (in mremap() and munmap()) with mmap_write_lock taken.
> > WDYT?
>
> It would be really great to make unlock_range oom reaper aware IMHO.
What exactly do you envision? Say unlock_range() knows that it's
racing with __oom_reap_task_mm() and that calling follow_page() is
unsafe without locking, what should it do?
>
> You do not quote your change in the full length so it is not really
> clear whether you are planning to drop __oom_reap_task_mm from exit_mmap
> as well.
Yes, that was the plan.
> If yes then 1) could push oom reaper to timeout while the
> unlock_range could be dropped on something so that wouldn't be an
> improvement. 2) sounds like a workaround to me as it doesn't really
> address the underlying problem.
With (1) potentially blocking due to allocation I can see why this is a problem.
Agree about (2).
>
> I have to say that I am not really a great fan of __oom_reap_task_mm in
> exit_mmap but I would rather see it in place than making the surrounding
> code more complex/tricky.
Agree. So far I could not find a cleaner solution. I thought (1) would
be a good one but the point you made renders it invalid. If you
clarify your comment about making unlock_range oom reaper aware maybe
that will open a new line of investigation?
Thanks,
Suren.
>
> --
> Michal Hocko
> SUSE Labs
