Re: [PATCH 1/2] vmscan: fix initial shrinker size handling
From: Andrew Morton
Date: Wed Oct 05 2011 - 17:58:32 EST
On Wed, 5 Oct 2011 23:52:05 +0200
Johannes Weiner <jweiner@xxxxxxxxxx> wrote:
> On Tue, Sep 13, 2011 at 08:38:36PM +0200, Johannes Weiner wrote:
> > On Tue, Aug 23, 2011 at 10:00:54AM +1000, Dave Chinner wrote:
> > > On Mon, Aug 22, 2011 at 04:38:21PM -0700, Andrew Morton wrote:
> > > > On Tue, 23 Aug 2011 09:22:57 +1000
> > > > Dave Chinner <david@xxxxxxxxxxxxx> wrote:
> > > >
> > > > > On Mon, Aug 22, 2011 at 02:17:21PM +0300, Konstantin Khlebnikov wrote:
> > > > > > Shrinker function can returns -1, it means it cannot do anything without a risk of deadlock.
> > > > > > For example prune_super() do this if it cannot grab superblock refrence, even if nr_to_scan=0.
> > > > > > Currenly we interpret this like ULONG_MAX size shrinker, evaluate total_scan according this,
> > > > > > and next time this shrinker can get really big pressure. Let's skip such shrinkers instead.
> > > > > >
> > > > > > Also make total_scan signed, otherwise check (total_scan < 0) below never works.
> > > > >
> > > > > I've got a patch set I am going to post out today that makes this
> > > > > irrelevant.
> > > >
> > > > Well, how serious is the bug? If it's a non-issue then we can leave
> > > > the fix until 3.1. If it's a non-non-issue then we'd need a minimal
> > > > patch to fix up 3.1 and 3.0.x.
> > >
> > > I'm pretty sure it's a non-issue. I'm pretty sure all of the
> > > shrinkers return a count >= 0 rather than -1 when passed nr_to_scan
> > > == 0 (i.e. they skip the GFP_NOFS checking), so getting a max_pass
> > > of -1 isn't going to happen very often....
> > Except for the case which Konstantin laid out, grabbing the super
> > block reference. How likely is that? And why isn't once enough to
> > build up quite a high number?
> > > And with total_scan being unsigned, the negative check is followed
> > > by a "if (total_scan > max_pass * 2)" check, which will catch
> > > numbers that would have gone negative anyway because max_pass won't
> > > be negative....
> > total_scan = nr;
> > max_pass = do_shrinker_shrink(shrinker, shrink, 0);
> > delta = (4 * nr_pages_scanned) / shrinker->seeks;
> > delta *= max_pass;
> > do_div(delta, lru_pages + 1);
> > total_scan += delta;
> > max_pass, an unsigned long, is what the shrinker returned, so
> > ULONG_MAX. ULONG_MAX * 2 is ULONG_MAX - 1, still pretty big?
> > Even for high values of delta (lots of pages scanned, few lru pages
> > left), it won't come nowhere near max_pass such that the product of
> > the two is a reasonable number again.
I have a note against that patch that David was working on some
alternative. So ... ping, indeed.
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/