Re: [PATCH v4 04/10] mm/vma: add+use vma lockdep acquire/release defines

[Suren Baghdasaryan]

On Wed, Jan 28, 2026 at 3:48 AM Lorenzo Stoakes
<lorenzo.stoakes@xxxxxxxxxx> wrote:
>
> On Wed, Jan 28, 2026 at 12:37:49PM +0100, Sebastian Andrzej Siewior wrote:
> > On 2026-01-23 20:12:14 [+0000], Lorenzo Stoakes wrote:
> > > --- a/include/linux/mmap_lock.h
> > > +++ b/include/linux/mmap_lock.h
> > > @@ -78,6 +78,37 @@ static inline void mmap_assert_write_locked(const struct mm_struct *mm)
> > …
> > > +/* Only meaningful if CONFIG_LOCK_STAT is defined. */
> > > +#define __vma_lockdep_stat_mark_acquired(vma) \
> > > +   lock_acquired(&vma->vmlock_dep_map, _RET_IP_)
> > > +
> >
> > After going through the remaining series, I don't think I found a
> > matching lock_contended(). So perf/ tracing just give you a few
> > lock-acquired events. Wouldn't it make sense to also some
> > lock_contended() events where the caller had to wait before it could
> > acquire the lock?
>
> Yeah I did wonder about this actually. The series really just abstracts this
> part, so I think doing something with that should be a follow-up.
>
> Suren - what was your intent with this? I did wonder what we actually really
> accomplished with this.
>
> VMA locks are always try-locks.
>
> Write locks can't be contended against one another since VMAs are always a
> per-process entity and not obtained remotely, so either a VMA is write-locked by
> us or not write-locked, never write-locked by anybody else.

Correct, all paths which call __vma_enter_locked() to either
write-lock a vma or to detach it have to already hold the mmap write
lock on vma->vm_mm. So, lock_contended() will never be triggered for
vma write locking paths.

>
> Read locks immediately give up if the VMA is write locked.
>
> Would we want to record a lock_contended() event in that case I guess then?

Yes, we could have it in vma_start_read_locked_nested() if refcount
increment fails and if the refcount != 0 (if it's 0 then the vma is
detached and the refcount can't change concurrently).

>
> I don't think we'd want to do that if the VMA were detached, only if it were
> write-locked?

Yes, that's why we would need an additional check for 0 in there.

I can add this logic after your patchset lands. I don't think we need
to block it for this.
Thanks,
Suren.

>
> >
> > Sebastian
>
> Cheers, Lorenzo