Re: [PATCH 1/6] mm: Allow per-VMA locks on file-backed VMAs

[Suren Baghdasaryan <surenb@google.com>]

On Fri, Apr 7, 2023 at 1:12 PM Matthew Wilcox <willy@infradead.org> wrote:
>
> On Fri, Apr 07, 2023 at 10:54:00AM -0700, Suren Baghdasaryan wrote:
> > On Tue, Apr 4, 2023 at 6:59 AM Matthew Wilcox (Oracle)
> > <willy@infradead.org> wrote:
> > >
> > > The fault path will immediately fail in handle_mm_fault(), so this
> > > is the minimal step which allows the per-VMA lock to be taken on
> > > file-backed VMAs.  There may be a small performance reduction as a
> > > little unnecessary work will be done on each page fault.  See later
> > > patches for the improvement.
> > >
> > > Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
> > > ---
> > >  mm/memory.c | 9 ++++-----
> > >  1 file changed, 4 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/mm/memory.c b/mm/memory.c
> > > index fdaec7772fff..f726f85f0081 100644
> > > --- a/mm/memory.c
> > > +++ b/mm/memory.c
> > > @@ -5223,6 +5223,9 @@ vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
> > >                                             flags & FAULT_FLAG_REMOTE))
> > >                 return VM_FAULT_SIGSEGV;
> > >
> > > +       if ((flags & FAULT_FLAG_VMA_LOCK) && !vma_is_anonymous(vma))
> > > +               return VM_FAULT_RETRY;
> > > +
> >
> > There are count_vm_event(PGFAULT) and count_memcg_event_mm(vma->vm_mm,
> > PGFAULT) earlier in this function. Returning here and retrying I think
> > will double-count this page fault. Returning before this accounting
> > should fix this issue.
>
> You're right, but this will be an issue with later patches in the series
> anyway as we move the check further and further down the call-chain.
> For that matter, it's an issue in do_swap_page() right now, isn't it?
> I suppose we don't care too much because it's the rare case where we go
> into do_swap_page() and so the stats are "correct enough".

True. do_swap_page() has the same issue. Can we move these
count_vm_event() calls to the end of handle_mm_fault():

vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
   unsigned int flags, struct pt_regs *regs)
{
       vm_fault_t ret;

       __set_current_state(TASK_RUNNING);

-       count_vm_event(PGFAULT);
-       count_memcg_event_mm(vma->vm_mm, PGFAULT);

       ret = sanitize_fault_flags(vma, &flags);
       if (ret)
-              return ret;
-              goto out;
       ...
       mm_account_fault(regs, address, flags, ret);
+out:
+       if (ret != VM_FAULT_RETRY) {
+              count_vm_event(PGFAULT);
+              count_memcg_event_mm(vma->vm_mm, PGFAULT);
+       }

       return ret;
}

?