On Wed, 11 Sep 2013, Michal Hocko wrote:
> On Wed 11-09-13 08:40:57, Anton Vorontsov wrote:
> > On Mon, Sep 09, 2013 at 01:08:47PM +0200, Michal Hocko wrote:
> > > On Fri 06-09-13 22:59:16, Hugh Dickins wrote:
> > > > Hit divide-by-0 in vmpressure_work_fn(): checking vmpr->scanned before
> > > > taking the lock is not enough, we must check scanned afterwards too.
> > >
> > > As vmpressure_work_fn seems the be the only place where we set scanned
> > > to 0 (except for the rare occasion when scanned overflows which
> > > would be really surprising) then the only possible way would be two
> > > vmpressure_work_fn racing over the same work item. system_wq is
> > > !WQ_NON_REENTRANT so one work item might be processed by multiple
> > > workers on different CPUs. This means that the vmpr->scanned check in
> > > the beginning of vmpressure_work_fn is inherently racy.
> > >
> > > Hugh's patch fixes the issue obviously but doesn't it make more sense to
> > > move the initial vmpr->scanned check under the lock instead?
> > >
> > > Anton, what was the initial motivation for the out of the lock
> > > check? Does it really optimize anything?
> >
> > Thanks a lot for the explanation.
> >
> > Answering your question: the idea was to minimize the lock section, but the
> > section is quite small anyway so I doubt that it makes any difference
> > (during
> > development I could not measure any effect of vmpressure() calls in my
> > system,
> > though the system itself was quite small).
> >
> > I am happy with moving the check under the lock
>
> The patch below. I find it little bit nicer than Hugh's original one
> because having the two checks sounds more confusing.
> What do you think Hugh, Anton?
>
> > or moving the work into its own WQ_NON_REENTRANT queue.
>
> That sounds like an overkill.
>
> ---
> From 888745909da34f8aee8a208a82d467236b828d0d Mon Sep 17 00:00:00 2001
> From: Michal Hocko <mho...@suse.cz>
> Date: Wed, 11 Sep 2013 17:48:10 +0200
> Subject: [PATCH] vmpressure: fix divide-by-0 in vmpressure_work_fn
>
> Hugh Dickins has reported a division by 0 when a vmpressure event is
> processed. The reason for the exception is that a single vmpressure
> work item (which is per memcg) might be processed by multiple CPUs
> because it is enqueued on system_wq which is !WQ_NON_REENTRANT.
> This means that the out of lock vmpr->scanned check in
> vmpressure_work_fn is inherently racy and the racing workers will see
> already zeroed scanned value after they manage to take the spin lock.
>
> The patch simply moves the vmp->scanned check inside the sr_lock to fix
> the race.
>
> The issue was there since the very beginning but "vmpressure: change
> vmpressure::sr_lock to spinlock" might have made it more visible as the
> racing workers would sleep on the mutex and give it more time to see
> updated value. The issue was still there, though.
>
> Reported-by: Hugh Dickins <hu...@google.com>
> Signed-off-by: Michal Hocko <mho...@suse.cz>
> Cc: sta...@vger.kernel.org
Nack! But equally Nack to my original.
Many thanks for looking into how this might have happened, Michal,
and for mentioning the WQ_NON_REENTRANT flag: which I knew nothing
about, but have now followed up.
I owe you all an abject apology: what I didn't mention in my patch
was that actually I hit the problem on a v3.3-based kernel to which
vmpressure had been backported.
I have not yet seen the problem on v3.11 or v3.10, and now believe
that it cannot happen there - which would explain why I was the
first to hit it.
When I looked up WQ_NON_REENTRANT in the latest tree, I found
WQ_NON_REENTRANT = 1 << 0, /* DEPRECATED */
and git blame on that line leads to Tejun explaining
dbf2576e37 ("workqueue: make all workqueues non-reentrant") made
WQ_NON_REENTRANT no-op but the following patches didn't remove the
flag or update the documentation. Let's mark the flag deprecated and
update the documentation accordingly.
dbf2576e37 went into v3.7, so I now believe this divide-by-0 could
only happen on a backport of vmpressure to an earlier kernel than that.
Tejun made that change precisely to guard against this kind of subtle
unsafe issue; but it does provide a good illustration of the danger of
backporting something to a kernel where primitives behave less safely.
Sorry for wasting all your time.
As to your code change itself, Michal: I don't really mind one way or
the other - it now seems unnecessary. On the one hand I liked Anton's
minor optimization, on the other hand your way is more proof against
future change.
My Nack is really to your comment (and the Cc stable): we cannot
explain in terms of WQ_NON_REENTRANT when that is a no-op!
Hugh
> ---
> mm/vmpressure.c | 17 +++++++++--------
> 1 file changed, 9 insertions(+), 8 deletions(-)
>
> diff --git a/mm/vmpressure.c b/mm/vmpressure.c
> index e0f6283..ad679a0 100644
> --- a/mm/vmpressure.c
> +++ b/mm/vmpressure.c
> @@ -164,18 +164,19 @@ static void vmpressure_work_fn(struct work_struct *work)
> unsigned long scanned;
> unsigned long reclaimed;
>
> + spin_lock(&vmpr->sr_lock);
> +
> /*
> - * Several contexts might be calling vmpressure(), so it is
> - * possible that the work was rescheduled again before the old
> - * work context cleared the counters. In that case we will run
> - * just after the old work returns, but then scanned might be zero
> - * here. No need for any locks here since we don't care if
> - * vmpr->reclaimed is in sync.
> + * Several contexts might be calling vmpressure() and the work
> + * item is sitting on !WQ_NON_REENTRANT workqueue so different
> + * CPUs might execute it concurrently. Bail out if the scanned
> + * counter is already 0 because all the work has been done already.
> */
> - if (!vmpr->scanned)
> + if (!vmpr->scanned) {
> + spin_unlock(&vmpr->sr_lock);
> return;
> + }
>
> - spin_lock(&vmpr->sr_lock);
> scanned = vmpr->scanned;
> reclaimed = vmpr->reclaimed;
> vmpr->scanned = 0;
> --
> 1.7.10.4
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
