Re: [PATCH] mm: sync rss-counters at the end of exit_mm()

[Konstantin Khlebnikov]

Hugh Dickins wrote:
> On Mon, 9 Apr 2012, KOSAKI Motohiro wrote:
> > On Mon, Apr 9, 2012 at 4:03 PM, Konstantin Khlebnikov
> > <khlebnikov@xxxxxxxxxx>  wrote:
> > > On task's exit do_exit() calls sync_mm_rss() but this is not enough,
> > > there can be page-faults after this point, for example exit_mm() ->
> > > mm_release() ->  put_user() (for processing tsk->clear_child_tid).
> > > Thus there may be some rss-counters delta in current->rss_stat.
> >
> > Seems reasonable.
>
> Yes, I think Konstantin has probably caught it;
> but I'd like to hear confirmation from Markus.

There is another bug in exec_mmap()

--- a/fs/exec.c
+++ b/fs/exec.c
@@ -823,8 +823,8 @@ static int exec_mmap(struct mm_struct *mm)
        /* Notify parent that we're no longer interested in the old VM */
        tsk = current;
        old_mm = current->mm;
-       sync_mm_rss(old_mm);
        mm_release(tsk, old_mm);
+       sync_mm_rss(old_mm);

        if (old_mm) {
                /*

> > but I have another question. Do we have any reason to
> > keep sync_mm_rss() in do_exit()? I havn't seen any reason that thread exiting
> > makes rss consistency.
>
> IIRC it's all about the hiwater_rss/maxrss stuff: we want to sync the
> maximum rss into mm->hiwater_rss before it's transferred to signal->maxrss,
> and later made visible to the user though getrusage(RUSAGE_CHILDREN,) -
> does your reading confirm that?
>
> Konstantin now finds the child_tid and futex stuff can trigger faults
> raising rss beyond that point, but usually it won't go higher than when
> it was captured for maxrss there.
>
> The sync_mm_rss() added by this patch (after "tsk->mm = NULL" so
> *_mm_counter_fast() cannot store any more into the tsk even if there
> were more faults) is solely to satisfy Konstantin's check_mm(), and
> it is irritating to have that duplicated on the exit path.

It was quick fix after the midnight. =) Now I think we can move mm_release()
from exit_mm() to do_exit(), and place it before sync_mm_rss(). Other stuff
there shouldn't trigger page-faults. Thus here will be only one sync_mm_rss():
at the end of mm_release()

> I'd be happy to see the new one put under CONFIG_DEBUG_VM along with
> check_mm(), once it's had a few -rcs of exposure without.
>
> Hugh
>
> > > Signed-off-by: Konstantin Khlebnikov<khlebnikov@xxxxxxxxxx>
> > > Reported-by: Markus Trippelsdorf<markus@xxxxxxxxxxxxxxx>
> > > Cc: Hugh Dickins<hughd@xxxxxxxxxx>
> > > Cc: KAMEZAWA Hiroyuki<kamezawa.hiroyu@xxxxxxxxxxxxxx>
> > > ---
> > >   kernel/exit.c |    1 +
> > >   1 file changed, 1 insertion(+)
> > >
> > > diff --git a/kernel/exit.c b/kernel/exit.c
> > > index d8bd3b42..8e09dbe 100644
> > > --- a/kernel/exit.c
> > > +++ b/kernel/exit.c
> > > @@ -683,6 +683,7 @@ static void exit_mm(struct task_struct * tsk)
> > >         enter_lazy_tlb(mm, current);
> > >         task_unlock(tsk);
> > >         mm_update_next_owner(mm);
> > > +       sync_mm_rss(mm);
> > >         mmput(mm);
> > >   }
> > >

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/