Re: [PATCH v2 1/2] mm: drop oom code from exit_mmap

[Suren Baghdasaryan <surenb@google.com>]

On Fri, May 20, 2022 at 12:21 AM Michal Hocko <mhocko@suse.com> wrote:
>
> On Thu 19-05-22 14:33:03, Suren Baghdasaryan wrote:
> > On Thu, May 19, 2022 at 1:22 PM Liam Howlett <liam.howlett@oracle.com> wrote:
> [...]
> > > arch_exit_mmap() was called under the write lock before, is it safe to
> > > call it under the read lock?
> >
> > Ah, good catch. I missed at least one call chain which I believe would
> > require arch_exit_mmap() to be called under write lock:
> >
> > arch_exit_mmap
> >     ldt_arch_exit_mmap
> >         free_ldt_pgtables
> >             free_pgd_range
>
> Why would be this a problem? This is LDT mapped into page tables but as
> far as I know oom_reaper cannot really ever see that range because it is
> not really reachable from any VMA.

Ah, thanks! I didn't realize these page tables are not reachable from
VMAs. The only other call that I'm not sure is ok without mmap write
lock is xen_hvm_exit_mmap:

arch_exit_mmap
    paravirt_arch_exit_mmap
        xen_hvm_exit_mmap

I'll look closer today but if someone can confirm it's safe then my
current patch should be fine as is.
Thanks,
Suren.

>
> > I'll need to check whether arch_exit_mmap() has to be called before
> > unmap_vmas(). If not, we could move it further down when we hold the
> > write lock.
> > Andrew, please remove this patchset from your tree for now until I fix this.
> >
> > >
> > > >
> > > >       vma = mm->mmap;
> > > >       if (!vma) {
> > > >               /* Can happen if dup_mmap() received an OOM */
> > > > -             mmap_write_unlock(mm);
> > > > +             mmap_read_unlock(mm);
> > > >               return;
> > > >       }
> > > >
> > > > @@ -3138,6 +3121,16 @@ void exit_mmap(struct mm_struct *mm)
> > > >       /* update_hiwater_rss(mm) here? but nobody should be looking */
> > > >       /* Use -1 here to ensure all VMAs in the mm are unmapped */
> > > >       unmap_vmas(&tlb, vma, 0, -1);
> > > > +     mmap_read_unlock(mm);
> > > > +
> > > > +     /*
> > > > +      * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
> > > > +      * because the memory has been already freed. Do not bother checking
> > > > +      * mm_is_oom_victim because setting a bit unconditionally is cheaper.
> > > > +      */
> > > > +     set_bit(MMF_OOM_SKIP, &mm->flags);
> > > > +
> > > > +     mmap_write_lock(mm);
> > >
> > > Is there a race here?  We had a VMA but after the read lock was dropped,
> > > could the oom killer cause the VMA to be invalidated?
>
> Nope, the oom killer itself doesn't do much beyond sending SIGKILL and
> scheduling the victim for the oom_reaper. dup_mmap is holding exclusive
> mmap_lock throughout the whole process.
>
> > > I don't think so
> > > but the comment above about dup_mmap() receiving an OOM makes me
> > > question it.  The code before kept the write lock from when the VMA was
> > > found until the end of the mm edits - and it had the check for !vma
> > > within the block itself.  We are also hiding it from the oom killer
> > > outside the read lock so it is possible for oom to find it in that
> > > window, right?
>
> The oom killer's victim selection doesn't really depend on the
> mmap_lock. If there is a race and MMF_OOM_SKIP is not set yet then it
> will consider the task and very likely bail out anyway because the
> address space has already been unampped so oom_badness() would consider
> this task boring.
>
> oom_reaper on the other hand would just try to unmap in parallel but
> that is fine regardless of MMF_OOM_SKIP. Seeing the flag would allow to
> bail out early rather than just trying to unmap something that is no
> longer there. The only problem for the oom_reaper is to see page tables
> of the address space disappearing from udner its feet. That is excluded
> by the the exlusive lock and as Suren mentions mm->mmap == NULL check
> if the exit_mmap wins the race.
> --
> Michal Hocko
> SUSE Labs