Re: [PATCH 3/4] mm: replace rw_semaphore with atomic_t in vma_lock

[Suren Baghdasaryan <surenb@google.com>]

On Mon, Nov 11, 2024 at 4:35 PM Davidlohr Bueso <dave@stgolabs.net> wrote:
>
> On Mon, 11 Nov 2024, Suren Baghdasaryan wrote:
>
> >@@ -787,7 +893,10 @@ static inline void vma_start_write(struct vm_area_struct *vma)
> >        * we should use WRITE_ONCE() for cleanliness and to keep KCSAN happy.
> >        */
> >       WRITE_ONCE(vma->vm_lock_seq, mm_lock_seq);
> >-      up_write(&vma->vm_lock.lock);
> >+      /* Write barrier to ensure vm_lock_seq change is visible before count */
> >+      smp_wmb();
> >+      rwsem_release(&vma->vm_lock.dep_map, _RET_IP_);
> >+      atomic_set(&vma->vm_lock.count, VMA_LOCK_UNLOCKED);
>
> Too many barriers here. Just do atomic_set_release and remove that
> smp_wmb(). And what you are doing is really ensuring nothing leaks out
> of the critical region, so that comment should also be more generic.

Uh, yes. I missed that.

>
> I would expect regression testing to catch this sort of thing.

Well, it's in vma_start_write(), which is in the write-locking path.
Maybe that's why it's not very visible.

>
> ...
>
> > #ifdef CONFIG_PER_VMA_LOCK
> >+              struct wait_queue_head vma_writer_wait;
>
> You might want to use rcuwait here instead, which is much more
> optimized for the single waiter requirement vmas have.

Thanks for the suggestion! I'll give it a try.

>
> Thanks,
> Davidlohr