Date:   Fri, 14 Apr 2023 15:26:03 -0700
From:   Suren Baghdasaryan <surenb@...gle.com>
To:     Peter Xu <peterx@...hat.com>
Cc:     akpm@...ux-foundation.org, willy@...radead.org, hannes@...xchg.org,
        mhocko@...e.com, josef@...icpanda.com, jack@...e.cz,
        ldufour@...ux.ibm.com, laurent.dufour@...ibm.com,
        michel@...pinasse.org, liam.howlett@...cle.com, jglisse@...gle.com,
        vbabka@...e.cz, minchan@...gle.com, dave@...olabs.net,
        punit.agrawal@...edance.com, lstoakes@...il.com,
        linux-mm@...ck.org, linux-kernel@...r.kernel.org,
        kernel-team@...roid.com
Subject: Re: [PATCH 1/1] mm: do not increment pgfault stats when page fault
 handler retries

On Fri, Apr 14, 2023 at 3:14 PM Suren Baghdasaryan <surenb@...gle.com> wrote:
>
> On Fri, Apr 14, 2023 at 2:47 PM Peter Xu <peterx@...hat.com> wrote:
> >
> > Hi, Suren,
>
> Hi Peter,
>
> >
> > On Fri, Apr 14, 2023 at 10:54:44AM -0700, Suren Baghdasaryan wrote:
> > > If the page fault handler requests a retry, we will count the fault
> > > multiple times.  This is a relatively harmless problem as the retry paths
> > > are not often requested, and the only user-visible problem is that the
> > > fault counter will be slightly higher than it should be.  Nevertheless,
> > > userspace only took one fault, and should not see the fact that the
> > > kernel had to retry the fault multiple times.
> > >
> > > Fixes: 6b4c9f446981 ("filemap: drop the mmap_sem for all blocking operations")
> > > Signed-off-by: Suren Baghdasaryan <surenb@...gle.com>
> > > Reviewed-by: Matthew Wilcox (Oracle) <willy@...radead.org>
> > > ---
> > > Patch applies cleanly over linux-next and mm-unstable
> > >
> > >  mm/memory.c | 16 ++++++++++------
> > >  1 file changed, 10 insertions(+), 6 deletions(-)
> > >
> > > diff --git a/mm/memory.c b/mm/memory.c
> > > index 1c5b231fe6e3..d88f370eacd1 100644
> > > --- a/mm/memory.c
> > > +++ b/mm/memory.c
> > > @@ -5212,17 +5212,16 @@ vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
> > >
> > >       __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;
> > >
> > >       if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
> > >                                           flags & FAULT_FLAG_INSTRUCTION,
> > > -                                         flags & FAULT_FLAG_REMOTE))
> > > -             return VM_FAULT_SIGSEGV;
> > > +                                         flags & FAULT_FLAG_REMOTE)) {
> > > +             ret = VM_FAULT_SIGSEGV;
> > > +             goto out;
> > > +     }
> > >
> > >       /*
> > >        * Enable the memcg OOM handling for faults triggered in user
> > > @@ -5253,6 +5252,11 @@ vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
> > >       }
> > >
> > >       mm_account_fault(regs, address, flags, ret);
> >
> > Here is the mm_account_fault() function taking care of some other
> > accountings.  Perhaps good to put things into it?
>
> That seems appropriate. Let me take a closer look.
>
> >
> > It also already ignores invalid faults:
> >
> >         if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
> >                 return;
>
> Can there be a case of (!VM_FAULT_ERROR && VM_FAULT_RETRY) - basically
> we need to retry but no errors happened? If so then this condition
> would double-count pagefaults in such cases. If such return code is
> impossible then it's the same as checking for VM_FAULT_RETRY.
>
> >
> > I see that you may also want to account for sigbus, however I really don't
> > know why.  Explanations would be great when it would matter.  So far it
> > makes sense to me if we skip both RETRY or ERROR cases.
>
> Accounting in case of a sigbus is not affected by this patch I think.
> We account for sigbus or any other error cases because there was a
> pagefault and we need to account for it. Whether we failed to handle
> it or not should not affect the count. We skip the retry case because
> we know the same fault will be retried. If we don't skip then we will
> double-count this fault.

mm_account_fault() has a nice comment explaining why it skips errors
and that now makes sense to me. Let me move the accounting there and
see if others agree that's the right place.

>
> >
> > > +out:
> > > +     if (!(ret & VM_FAULT_RETRY)) {
> > > +             count_vm_event(PGFAULT);
> > > +             count_memcg_event_mm(vma->vm_mm, PGFAULT);
> >
> > There is one thing worth noticing is here vma may or may not be valid
> > depending on the retval of the fault.
> >
> > RETRY is exactly one of the cases that accessing vma may be unsafe due to
> > releasing of mmap read lock.  The other one is the recently added
> > VM_FAULT_COMPLETE.  So if we want to move this chunk (or any vma reference)
> > to be later we need to consider a valid vma / mm being there first, or
> > we're prone to accessing a vma that has already been released, I think.
>
> Good catch! I think you are right and I should have stored vma->vm_mm
> in the beginning and used it when calling count_memcg_event_mm().
> I'll prepare a new patch which handles this correctly.
> Thanks,
> Suren.
>
> >
> > > +     }
> > >
> > >       return ret;
> > >  }
> > > --
> > > 2.40.0.634.g4ca3ef3211-goog
> > >
> > >
> >
> > Thanks,
> >
> > --
> > Peter Xu
> >
> > --
> > To unsubscribe from this group and stop receiving emails from it, send an email to kernel-team+unsubscribe@...roid.com.
> >

Powered by blists - more mailing lists