Re: [RFC][RFT] memcg fix cgroup_mutex deadlock when cpuset reclaims memory

[KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>]

On Wed, 10 Dec 2008 16:41:26 +0900
Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> wrote:

> On Wed, 10 Dec 2008 15:19:48 +0900, Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> wrote:
> > On Wed, 10 Dec 2008 10:49:47 +0530, Balbir Singh <balbir@linux.vnet.ibm.com> wrote:
> > > Hi,
> > > 
> > > Here is a proposed fix for the memory controller cgroup_mutex deadlock
> > > reported. It is lightly tested and reviewed. I need help with review
> > > and test. Is the reported deadlock reproducible after this patch? A
> > > careful review of the cpuset impact will also be highly appreciated.
> > > 
> > > From: Balbir Singh <balbir@linux.vnet.ibm.com>
> > > 
> > > cpuset_migrate_mm() holds cgroup_mutex throughout the duration of
> > > do_migrate_pages(). The issue with that is that
> > > 
> > > 1. It can lead to deadlock with memcg, as do_migrate_pages()
> > >    enters reclaim
> > > 2. It can lead to long latencies, preventing users from creating/
> > >    destroying other cgroups anywhere else
> > > 
> > > The patch holds callback_mutex through the duration of cpuset_migrate_mm() and
> > > gives up cgroup_mutex while doing so.
> > > 
> > I agree changing cpuset_migrate_mm not to hold cgroup_mutex to fix the dead lock
> > is one choice, and it looks good to me at the first impression.
> > 
> > But I'm not sure it's good to change cpuset(other subsystem) code because of memcg.
> > 
> > Anyway, I'll test this patch and report the result tomorrow.
> > (Sorry, I don't have enough time today.)
> > 
> Unfortunately, this patch doesn't seem enough.
> 
> This patch can fix dead lock caused by "circular lock of cgroup_mutex",
> but cannot that of caused by "race between page reclaim and cpuset_attach(mpol_rebind_mm)".
> 
> (The dead lock I fixed in memcg-avoid-dead-lock-caused-by-race-between-oom-and-cpuset_attach.patch
> was caused by "race between memcg's oom and mpol_rebind_mm, and was independent of hierarchy.)
> 
> I attach logs I got in testing this patch.
> 
Hmm, ok then, what you  mention to is this race.
--
	cgroup_lock()
		-> cpuset_attach()
			-> down_write(&mm->mmap_sem);

	down_read()
		-> page fault
			-> reclaim in memcg
				-> cgroup_lock().
--
What this patch tries to fix is this recursive locks
--
	cgroup_lock()
		-> cpuset_attach()
			-> cpuset_migrate_mm()
				-> charge to migration
					-> go to reclaim and meet cgroup_lock.
--


Right ?

BTW, releasing cgroup_lock() while attach() is going on is finally safe ?
If not, can this lock for attach be replaced with (new) cgroup private mutex ?

a new mutex like this ?
--
struct cgroup {
	.....
	mutex_t		attach_mutex; /* for serializing attach() ops. 
					 while attach() is going on, rmdir() will fail */
}
--
Do we need the big lock of cgroup_lock for attach(), at last ?

-Kame

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>