Re: [PATCH v3] mm: use per_vma lock for MADV_DONTNEED

[Barry Song]

On Sat, Jun 7, 2025 at 4:05 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> On Sat, Jun 07, 2025 at 12:46:23PM +1200, Barry Song wrote:
> > To simplify handling, the implementation falls back to the standard
> > mmap_lock if userfaultfd is enabled on the VMA, avoiding the complexity of
> > userfaultfd_remove().
>
> This feels too complex to me.  Why do we defer grabbing the vma lock
> so late, instead of grabbing it at the start like the fault handler does?

Hi Matthew,

It looks like you missed the spot where your comment should have gone:
https://lore.kernel.org/all/0b96ce61-a52c-4036-b5b6-5c50783db51f@lucifer.local/
So I believe Lorenzo is the best person to respond to your concern.

In both v1 and v2 [1][2], we did try to fall back as early as possible:

[1] https://lore.kernel.org/linux-mm/20250527044145.13153-1-21cnbao@xxxxxxxxx/
[2] https://lore.kernel.org/all/20250530104439.64841-1-21cnbao@xxxxxxxxx/

But that approach had its own problems:
* It's not extensible to other madvise operations.
* It's not easy to adapt to vector_madvise.

I also initially found the new approach too complex and tried a few
alternatives, but each had its own problems. In the end, Lorenzo's
solution still seems to be the cleanest among them.

I even forgot to move the code below back to visit() from
madvise_vma_behavior(). I had changed it while exploring an
alternative and should have reverted it.

+       if (madv_behavior && madv_behavior->lock_mode ==
MADVISE_VMA_READ_LOCK) {
+               vma = try_vma_read_lock(mm, madv_behavior, start, end);
+               if (vma) {
+                       error = madvise_vma_behavior(vma, &prev, start, end,
+                               madv_behavior);  /* better to be visit() */
+                       vma_end_read(vma);
+                       return error;
+               }
+       }

Thanks
Barry