[PATCH v2 0/5] mm: reduce mmap_lock contention and improve page fault performance

[Lorenzo Stoakes]

On Tue, May 19, 2026 at 07:07:37PM +0800, Barry Song wrote:
> On Tue, May 19, 2026 at 5:21 AM Yang Shi <shy828301 at gmail.com> wrote:
> >
> > On Sun, May 17, 2026 at 1:45 AM Barry Song <baohua at kernel.org> wrote:
> > >
> > > On Sat, May 2, 2026 at 1:58 AM Matthew Wilcox <willy at infradead.org> wrote:
> > > >
> > > > On Sat, May 02, 2026 at 01:44:34AM +0800, Barry Song wrote:
> > > > > On Fri, May 1, 2026 at 10:57 PM Matthew Wilcox <willy at infradead.org> wrote:
> > > > > >
> > > > > > On Fri, May 01, 2026 at 06:49:58AM +0800, Barry Song wrote:
> > > > > > > 1. There is no deterministic latency for I/O completion. It depends on
> > > > > > > both the hardware and the software stack (bio/request queues and the
> > > > > > > block scheduler). Sometimes the latency is short; at other times it can
> > > > > > > be quite long. In such cases, a high-priority thread performing operations
> > > > > > > such as mprotect, unmap, prctl_set_vma, or madvise may be forced to wait
> > > > > > > for an unpredictable amount of time.
> > > > > >
> > > > > > But does that actually happen?  I find it hard to believe that thread A
> > > > > > unmaps a VMA while thread B is in the middle of taking a page fault in
> > > > > > that same VMA.  mprotect() and madvise() are more likely to happen, but
> > > > > > it still seems really unlikely to me.
> > > > >
> > > > > It doesn’t have to involve unmapping or applying mprotect to
> > > > > the entire VMA—just a portion of it is sufficient.
> > > >
> > > > Yes, but that still fails to answer "does this actually happen".  How much
> > > > performance is all this complexity in the page fault handler buying us?
> > > > If you don't answer this question, I'm just going to go in and rip it
> > > > all out.
> > > >
> > >
> > > Hi Matthew (and Lorenzo, Jan, and anyone else who may be
> > > waiting for answers),
> > >
> > > As promised during LSF/MM/BPF, we conducted thorough
> > > testing on Android phones to determine whether performing
> > > I/O in `filemap_fault()` can block `vma_start_write()`.
> > > I wanted to give a quick update on this question.
> > >
> > > Nanzhe at Xiaomi created tracing scripts and ran various
> > > applications on Android devices with I/O performed under
> > > the VMA lock in `filemap_fault()`. We found that:
> > >
> > > 1. There are very few cases where unmap() is blocked by
> > >    page faults. I assume this is due to buggy user code
> > >    or poor synchronization between reads and unmap().
> > > So I assume it is not a problem.
> > >
> > > 2. We observed many cases where `vma_start_write()`
> > >    is blocked by page-fault I/O in some applications.
> > >    The blocking occurs in the `dup_mmap()` path during
> > >    fork().
> > >
> > > With Suren's commit fb49c455323ff ("fork: lock VMAs of
> > > the parent process when forking"), we now always hold
> > > `vma_write_lock()` for each VMA. Note that the
> > > `mmap_lock` write lock is also held, which could lead to
> > > chained waiting if page-fault I/O is performed without
> > > releasing the VMA lock.
> > >
> > > My gut feeling is that Suren's commit may be overshooting,
> > > so my rough idea is that we might want to do something like
> > > the following (we haven't tested it yet and it might be
> > > wrong):
> > >
> > > diff --git a/mm/mmap.c b/mm/mmap.c
> > > index 2311ae7c2ff4..5ddaf297f31a 100644
> > > --- a/mm/mmap.c
> > > +++ b/mm/mmap.c
> > > @@ -1762,7 +1762,13 @@ __latent_entropy int dup_mmap(struct mm_struct
> > > *mm, struct mm_struct *oldmm)
> > >         for_each_vma(vmi, mpnt) {
> > >                 struct file *file;
> > >
> > > -               retval = vma_start_write_killable(mpnt);
> > > +               /*
> > > +                * For anonymous or writable private VMAs, prevent
> > > +                * concurrent CoW faults.
> > > +                */
> > > +               if (!mpnt->vm_file || (!(mpnt->vm_flags & VM_SHARED) &&
> > > +                                       (mpnt->vm_flags & VM_WRITE)))
> > > +                       retval = vma_start_write_killable(mpnt);
> > >                 if (retval < 0)
> > >                         goto loop_out;
> > >                 if (mpnt->vm_flags & VM_DONTCOPY) {
> >
> > Maybe a little bit off topic. This is an interesting idea. It seems
> > possible we don't have to take vma write lock unconditionally. IIUC
> > the write lock is mainly used to serialize against page fault and
> > madvise, right? I got a crazy idea off the top of my head. We may be
> > able to just take vma write lock iff vma->anon_vma is not NULL.
> >
> > First of all, write mmap_lock is held, so the vma can't go or be
> > changed under us.
> >
> > Secondly, if vma->anon_vma is NULL, it basically means either no page
> > fault happened or no cow happened, so there is no page table to copy,
> > this is also what copy_page_range() does currently. So we can shrink
> > the critical section to:
> >
> > if (vma->anon_vma) {
> >     vma_start_write_killable(src_vma);
> >     anon_vma_fork(dst_vma, src_vma);
> >     copy_page_range(dst_vma, src_vma);
> > }
> >
> > But page fault can happen before write mmap_lock is taken, when we
> > check vma->anon_vma, it is possible it has not been set up yet. But it
> > seems to be equivalent to page fault after fork and won't break the
> > semantic.
>
> Re-reading Suren's commit log for fb49c455323ff8
> ("fork: lock VMAs of the parent process when forking"),
> it seems that vm_start_write() is used to protect
> against a race where anon_vma changes from NULL to
> non-NULL during fork. In that scenario, we hold the
> mmap_lock write lock, but not vma_start_write(), so a
> concurrent anon_vma_prepare() could still install an
> anon_vma.
>
> "    A concurrent page fault on a page newly marked read-only by the page
>     copy might trigger wp_page_copy() and a anon_vma_prepare(vma) on the
>     source vma, defeating the anon_vma_clone() that wasn't done because the
>     parent vma originally didn't have an anon_vma, but we now might end up
>     copying a pte entry for a page that has one.
> "
>
> If that is the case, then your change does not work.
>
> Nowadays, nobody calls anon_vma_prepare(vma) directly.

I see callers? Am I imagining them? :)
https://elixir.bootlin.com/linux/v7.0.9/A/ident/anon_vma_prepare

> Instead, vmf_anon_prepare() is used, and we always
> require the mmap_lock read lock before calling
> __anon_vma_prepare(). As a result, anon_vma cannot
> transition from NULL to non-NULL during fork.

Right, yes the mmap read lock is required for that.

>
> So the original race condition has effectively
> disappeared.

Err the page tables? All the other cases which require page table copying?

Concurrent faults mean that copy_page_range() and faulting with vma->anon_vma
_or_ any of the multiple cases mentioned elsewhere.

And who knows what else serialises on that.

>
> You also mentioned the madvise() case. If I understand
> correctly, madvise() should take mmap_lock before
> modifying anon_vma. Only some parts of madvise() can
> support per-VMA locking. Therefore, we probably do not
> need:
>
> if (vma->anon_vma) {
> vma_start_write_killable(src_vma);
> ...
> }

I like how you hand wave the VMA lock operations in madvise() :)

(Maybe) guard regions being present cause page tables to be copied, they're
installed under VMA (read) lock, and can race now.

And it sets traps for future changes - introducing more horrible edge case race
conditions in fork is just a big nope nope nope.

This isn't an area to play around in.

>
> >
> > Anyway, just a crazy idea, I may miss some corner cases.
>
> To me, it seems that we could remove vma_start_write()
> entirely now. Or is that an even crazier idea?

As above that'd be totally broken. NAK.

>
> Thanks
> Barry

Thanks, Lorenzo