Re: [PATCH] oom: handle overflow in mem_cgroup_out_of_memory()

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

On Wed, 26 Jan 2011 14:29:09 -0800
Andrew Morton <akpm@linux-foundation.org> wrote:

> On Wed, 26 Jan 2011 12:32:04 -0800
> Greg Thelen <gthelen@google.com> wrote:
> 
> > > That being said, does this have any practical impact at all?  I mean,
> > > this code runs when the cgroup limit is breached.  But if the number
> > > of allowed pages (not bytes!) can not fit into 32 bits, it means you
> > > have a group of processes using more than 16T.  On a 32-bit machine.
> > 
> > The value of this patch is up for debate.  I do not have an example
> > situation where this truncation causes the wrong thing to happen.  I
> > suppose it might be possible for a racing update to
> > memory.limit_in_bytes which grows the limit from a reasonable (example:
> > 100M) limit to a large limit (example 1<<45) could benefit from this
> > patch.  I admit that this case seems pathological and may not be likely
> > or even worth bothering over.  If neither the memcg nor the oom
> > maintainers want the patch, then feel free to drop it.  I just noticed
> > the issue and thought it might be worth addressing.
> 
> Ah.  I was scratching my head over that.
> 
> In zillions of places the kernel assumes that a 32-bit kernel has less
> than 2^32 pages of memory, so the code as it stands is, umm, idiomatic.
> 

I think we can assume that. 

> But afaict the only way the patch makes a real-world difference is if
> res_counter_read_u64() is busted?
> 
> And, as you point out, res_counter_read_u64() is indeed busted on
> 32-bit machines.  It has 25 callsites in mm/memcontrol.c - has anyone
> looked at the implications of this?  What happens in all those
> callsites if the counter is read during a count rollover?
> 

I'll review. Against the roll-over, I think we just need to take lock.
So, res_counter_read_u64() implementation was wrong. It should take lock.
Please give me time.


THanks,
-Kame