Re: [PATCH RFC v2] mm: use per_vma lock for MADV_DONTNEED

[Pedro Falcato]

On Tue, Jun 03, 2025 at 07:43:04PM +0100, Lorenzo Stoakes wrote:
> Hi Barry,
> 
> As promised, I enclose a patch to give a sense of how I think we might
> thread state through this operation.
> 
> There's a todo on the untagged stuff so you can figure that out. This is
> based on the v1 so it might not encompass everything you addressed in the
> v2.
> 
> Passing in madv_behavior to madvise_walk_vmas() twice kinda sucks, I
> _despise_ the void *arg function ptr stuff there added just for the anon
> vma name stuff (ughhh) so might be the only sensible way of threading
> state.
> 
> I don't need any attribution, so please use this patch as you see
> fit/adapt/delete/do whatever with it, just an easier way for me to show the
> idea!
> 
> I did some very basic testing and it seems to work, but nothing deeper.
> 
> Cheers, Lorenzo
> 
> ----8<----
> >From ff4ba0115cb31a0630b6f8c02c68f11b3fb71f7a Mon Sep 17 00:00:00 2001
> From: Lorenzo Stoakes <lorenzo.stoakes@xxxxxxxxxx>
> Date: Tue, 3 Jun 2025 18:22:55 +0100
> Subject: [PATCH] mm/madvise: support VMA read locks for MADV_DONTNEED[_LOCKED]
> 
> Refactor the madvise() code to retain state about the locking mode utilised
> for traversing VMAs.
> 
> Then use this mechanism to permit VMA locking to be done later in the
> madvise() logic and also to allow altering of the locking mode to permit
> falling back to an mmap read lock if required.
> 

Just as a quick drive-by comment: I was playing around with using per-vma locks
for GUP and mm_populate a few weeks ago. I never actually finished the work (and I
still plan on getting around doing it Eventually(tm)), but my final concept of an
approach was to simply read-lock every VMA in a range (if that fails, go back
to the mmap_lock).

I *think* it works, and doesn't have the same limitation for single VMAs.

I understand this is a super handwavy suggestion, but I know this discussion has
been happening and I just wanted to get this idea out of obscure IRC logs :)

-- 
Pedro