linux-kernel - Re: [patch] slab: Fix a cpu hotplug race condition while tuning slab cpu caches

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

Date:	Mon, 16 Oct 2006 12:48:08 -0700
From:	Andrew Morton <akpm@...l.org>
To:	Ravikiran G Thirumalai <kiran@...lex86.org>
Cc:	linux-kernel@...r.kernel.org,
	Christoph Lameter <clameter@...r.sgi.com>,
	Alok Kataria <alok.kataria@...softinc.com>,
	"Shai Fultheim (Shai@...lex86.org)" <shai@...lex86.org>,
	"Benzi Galili (Benzi@...leMP.com)" <benzi@...lemp.com>
Subject: Re: [patch] slab: Fix a cpu hotplug race condition while tuning
 slab cpu caches

On Mon, 16 Oct 2006 12:26:15 -0700
Ravikiran G Thirumalai <kiran@...lex86.org> wrote:

> On Mon, Oct 16, 2006 at 11:15:11AM -0700, Andrew Morton wrote:
> > On Mon, 16 Oct 2006 01:54:39 -0700
> > Ravikiran G Thirumalai <kiran@...lex86.org> wrote:
> > 
> > The problem is obvious: we have some data (the array caches) and we have a
> > data structure which is used to look up that data (cpu_online_map).  But
> > we're releasing the lock while these two things are in an inconsistent
> > state.
> > 
> > So you could have fixed this by taking cache_chain_mutex in CPU_UP_PREPARE
> > and releasing it in CPU_ONLINE and CPU_UP_CANCELED.
> 
> Hmm, yes. I suppose so. Maybe we can do away with other uses of
> lock_cpu_hotplug() in slab.c as well then!

Yes, we can.

>  Will give it a shot. Slab
> locking might look uglier than what it already is though no?

I suspect it will improve - all the calls to lock_cpu_hotplug() go away.

But it gets conceptually more complex: we're now holding cache_chain_mutex
across some or all of the _other_ callbacks which are registered on the cpu
notifier list.  And they could be anything at all, and in any order.

But as long as things are well-designed and layered and interactions are
minimised (all things we like to achieve) then we're OK.  But it's a worry.
Fortunately lockdep helps a lot here.

> > 
> > >  	list_for_each_entry(cachep, &cache_chain, next) {
> > > @@ -4087,6 +4088,7 @@ ssize_t slabinfo_write(struct file *file
> > >  		}
> > >  	}
> > >  	mutex_unlock(&cache_chain_mutex);
> > > +	unlock_cpu_hotplug();
> > >  	if (res >= 0)
> > >  		res = count;
> > >  	return res;
> > 
> > Given that this lock_cpu_hotplug() happens at a high level I guess it'll
> > avoid the usual lock_cpu_hotplug() horrors and we can live with it.  I
> > assume lockdep was enabled when you were testing this?
> 
> Not when I tested it.  I just retested with lockdep on and things seemed 
> fine on a SMP.

Great, thanks.  Please ensure that lockdep is used when testing the kernel.
 Also preempt, DEBUG_SLAB, DEBUG_SPINLOCK_SLEEP and various other things. 
I guess the list in Documentation/SubmitChecklist is the definitive one.
-
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/