Re: [PATCH v5 6/6] mm: handle userfaults under VMA lock

[Suren Baghdasaryan]

On Thu, Jun 29, 2023 at 9:39 AM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
>
> On Thu, Jun 29, 2023 at 9:33 AM Peter Xu <peterx@xxxxxxxxxx> wrote:
> >
> > On Wed, Jun 28, 2023 at 05:19:31PM -0700, Suren Baghdasaryan wrote:
> > > On Wed, Jun 28, 2023 at 10:32 AM Peter Xu <peterx@xxxxxxxxxx> wrote:
> > > >
> > > > On Wed, Jun 28, 2023 at 10:25:29AM -0700, Suren Baghdasaryan wrote:
> > > > > Enable handle_userfault to operate under VMA lock by releasing VMA lock
> > > > > instead of mmap_lock and retrying. Note that FAULT_FLAG_RETRY_NOWAIT
> > > > > should never be used when handling faults under per-VMA lock protection
> > > > > because that would break the assumption that lock is dropped on retry.
> > > > >
> > > > > Signed-off-by: Suren Baghdasaryan <surenb@xxxxxxxxxx>
> > > >
> > > > Maybe the sanitize_fault_flags() changes suite more in patch 3, but not a
> > > > big deal I guess.
> > >
> > > IIUC FAULT_FLAG_RETRY_NOWAIT comes into play in this patchset only in
> > > the context of uffds, therefore that check seems to be needed when we
> > > enable per-VMA lock uffd support, which is this patch. Does that make
> > > sense?
> >
> > I don't see why uffd is special in this regard, as e.g. swap also checks
> > NOWAIT when folio_lock_or_retry() so I assume it's also used there.
> >
> > IMHO the "NOWAIT should never apply with VMA_LOCK so far" assumption starts
> > from patch 3 where it conditionally releases the vma lock when
> > !(RETRY|COMPLETE); that is the real place where it can start to go wrong if
> > anyone breaks the assumption.
>
> Um, yes, you are right as usual. It was clear to me from the code that
> NOWAIT is not used with swap under VMA_LOCK, that's why I didn't
> consider this check earlier. Yeah, patch 3 seems like a more
> appropriate place for it. I'll move it and post a new patchset later
> today or tomorrow morning with your Acks.

Posted v6 at https://lore.kernel.org/all/20230630020436.1066016-1-surenb@xxxxxxxxxx/

> Thanks,
> Suren.
>
> >
> > Thanks,
> >
> > --
> > Peter Xu
> >