Re: [PATCH v3 1/5] mm: Make per-VMA locks available universally

On Tue, Aug 4, 2026 at 1:56 AM Lorenzo Stoakes (ARM) <[email protected]> wrote:
>
> On Mon, Aug 03, 2026 at 11:12:01PM +0200, Jann Horn wrote:
> > On Mon, Aug 3, 2026 at 6:08 PM Lorenzo Stoakes (ARM) <[email protected]> wrote:
> > > On Mon, Aug 03, 2026 at 08:24:44AM -0700, Suren Baghdasaryan wrote:
> > > > On Sun, Aug 2, 2026 at 2:55 PM Suren Baghdasaryan <[email protected]> wrote:
> > > > > -#endif /* CONFIG_PER_VMA_LOCK */
> > > >
> > > > Now that I'm looking closer into this, I think we would break NOMMU
> > > > case because nommu.c does not take VMA write locks at all. So,
> > > > lock_vma_under_rcu() for example would always succeed.
> > >
> > > I don't think anything's broken actually.
> > >
> > > Per-VMA locks was gated on CONFIG_MMU so nothing there assumes per-VMA flags,
> > > but now you have stuff that happens that didn't before but:
> > >
> > > * vm_area_free() -> vma_assert_detached() - fine - it's always detached in nommu.
> >
> > I don't think that's true, AFAICS vma_mark_attached() is called from
> > vma_iter_store_new() which is called from the nommu version of
> > do_mmap(). But I think you're right that one blunt workaround for this
> > would be to add a hack that prevents ever marking VMAs in nommu as
> > attached...
>
> Thanks Jann, though also ugh god at that :)
>
> (Really the ideal solution for nommu would be a 'virtual' MMU with noops for
> everything IMO.)
>
> Anyway, I think we could fix this with a:
>
> static inline void vma_mark_attached(struct vm_area_struct *vma)
> {
> +       if (!IS_ENBLED(CONFIG_MMU))
> +               return;
>
>         vma_assert_write_locked(vma);
>         vma_assert_detached(vma);
>         refcount_set_release(&vma->vm_refcnt, 1);
> }
>
> Here?

Yes, I think so.
I also need to make sure mmap_lock fallbacks we are removing are not
used in NOMMU, otherwise it will fail to lock_vma_under_rcu() and will
have no fallbacks.
TCP zerocopy and binder are gated on CONFIG_MMU, so that's fine. Need
to check the BPF part...


>
> >
> > > * vm_area_dup() -> vma_lock_init() - no asserts, just sets refcount to 0 (correct).
>
> --
> Cheers, Lorenzo