linux-kernel - Re: [PATCH mmotm/next] memcg-mm-introduce-lowlimit-reclaim-fix2.patch

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Date:	Wed, 28 May 2014 10:01:37 +0200
From:	Michal Hocko <mhocko@...e.cz>
To:	Hugh Dickins <hughd@...gle.com>
Cc:	Andrew Morton <akpm@...ux-foundation.org>,
	Johannes Weiner <hannes@...xchg.org>, linux-mm@...ck.org,
	linux-kernel@...r.kernel.org
Subject: Re: [PATCH mmotm/next] memcg-mm-introduce-lowlimit-reclaim-fix2.patch

On Tue 27-05-14 16:05:36, Hugh Dickins wrote:
> On Tue, 27 May 2014, Andrew Morton wrote:
> > On Tue, 27 May 2014 14:36:04 -0700 (PDT) Hugh Dickins <hughd@...gle.com> wrote:
> > 
> > > mem_cgroup_within_guarantee() oopses in _raw_spin_lock_irqsave() when
> > > booted with cgroup_disable=memory.  Fix that in the obvious inelegant
> > > way for now - though I hope we are moving towards a world in which
> > > almost all of the mem_cgroup_disabled() tests will vanish, with a
> > > root_mem_cgroup which can handle the basics even when disabled.
> > > 
> > > I bet there's a neater way of doing this, rearranging the loop (and we
> > > shall want to avoid spinlocking on root_mem_cgroup when we reach that
> > > new world), but that's the kind of thing I'd get wrong in a hurry!
> > > 
> > > ...
> > >
> > > @@ -2793,6 +2793,9 @@ static struct mem_cgroup *mem_cgroup_loo
> > >  bool mem_cgroup_within_guarantee(struct mem_cgroup *memcg,
> > >  		struct mem_cgroup *root)
> > >  {
> > > +	if (mem_cgroup_disabled())
> > > +		return false;
> > > +
> > >  	do {
> > >  		if (!res_counter_low_limit_excess(&memcg->res))
> > >  			return true;
> > 
> > This seems to be an awfully late and deep place at which to be noticing
> > mem_cgroup_disabled().  Should mem_cgroup_within_guarantee() even be called
> > in this state?
> 
> I think it's a natural consequence of our preferring to use a single
> path for memcg and non-memcg, outside of memcontrol.c itself.  So in
> vmscan.c there are loops iterating through a subtree of memcgs, which
> in the non-memcg case can only ever encounter root_mem_cgroup (or NULL).
> 
> In doing so, it's not surprising that __shrink_zone() should want to
> check mem_cgroup_within_guarantee().  Now, __shrink_zone() does have an
> honor_memcg_guarantee arg passed in, and I did consider initializing
> that according to !mem_cgroup_disabled(): which would be not so late
> and not so deep.  But then noticed mem_cgroup_all_within_guarantee(),
> which is called without condition on honor_guarantee, so backed away:
> we could very easily change that, I suppose, but...

I think that hiding the check inside mem_cgroup_all_within_guarantee
makes more sense than playing games with mem_cgroup_disabled in the
shrinking code. We do not want to convolute the generic mm code more
than necessary.

> I'm sure there is a better way of dealing with this than sprinkling
> mem_cgroup_disabled() tests all over, and IIUC Hannes is moving us
> towards that by making root_mem_cgroup more of a first-class citizen
> (following on from earlier per-cpu-ification of memcg's most expensive
> fields).

That is definitely the future direction.

> My attitude is that for now we just chuck in a !mem_cgroup_disabled()
> wherever it stops a crash, as before; but in future aim to give the
> cgroup_disabled=memory root_mem_cgroup all it needs to handle this
> seamlessly.  Ideally just a !mem_cgroup_disabled() test at the point
> of memcg creation, and everything else fall out naturally (but maybe
> some more lookup_page_cgroup() NULL tests).  In practice we may identify
> other places, where it's useful to add a special test to avoid expense;
> though usually that would be expense worth avoiding at the root, even
> when !mem_cgroup_disabled().

Yes, I would like to move mem_cgroup_disabled to jump labels at some
point and disable the possible runtime overhead.

> And probably a static dummy root_mem_cgroup even when !CONFIG_MEMCG.
> 
> (Not that I'm expecting to do any of this work myself!)
> 
> Hugh

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