Re: [PATCH] hugetlb/cgroup: Simplify pre_destroy callback

[Kamezawa Hiroyuki]

(2012/07/19 18:41), Aneesh Kumar K.V wrote:
> Li Zefan <lizefan@xxxxxxxxxx> writes:
>
> > on 2012/7/19 10:55, Aneesh Kumar K.V wrote:
> >
> > > Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> writes:
> > >
> > > > On Wed, 18 Jul 2012 11:04:09 +0530
> > > > "Aneesh Kumar K.V" <aneesh.kumar@xxxxxxxxxxxxxxxxxx> wrote:
> > > >
> > > > > From: "Aneesh Kumar K.V" <aneesh.kumar@xxxxxxxxxxxxxxxxxx>
> > > > >
> > > > > Since we cannot fail in hugetlb_cgroup_move_parent, we don't really
> > > > > need to check whether cgroup have any change left after that. Also skip
> > > > > those hstates for which we don't have any charge in this cgroup.
> > > > >
> > > > > ...
> > > > >
> > > > > +	for_each_hstate(h) {
> > > > > +		/*
> > > > > +		 * if we don't have any charge, skip this hstate
> > > > > +		 */
> > > > > +		idx = hstate_index(h);
> > > > > +		if (res_counter_read_u64(&h_cg->hugepage[idx], RES_USAGE) == 0)
> > > > > +			continue;
> > > > > +		spin_lock(&hugetlb_lock);
> > > > > +		list_for_each_entry(page, &h->hugepage_activelist, lru)
> > > > > +			hugetlb_cgroup_move_parent(idx, cgroup, page);
> > > > > +		spin_unlock(&hugetlb_lock);
> > > > > +		VM_BUG_ON(res_counter_read_u64(&h_cg->hugepage[idx], RES_USAGE));
> > > > > +	}
> > > > >   out:
> > > > >   	return ret;
> > > > >   }
> > > >
> > > > This looks fishy.
> > > >
> > > > We test RES_USAGE before taking hugetlb_lock.  What prevents some other
> > > > thread from increasing RES_USAGE after that test?
> > > >
> > > > After walking the list we test RES_USAGE after dropping hugetlb_lock.
> > > > What prevents another thread from incrementing RES_USAGE before that
> > > > test, triggering the BUG?
> > >
> > > IIUC core cgroup will prevent a new task getting added to the cgroup
> > > when we are in pre_destroy. Since we already check that the cgroup doesn't
> > > have any task, the RES_USAGE cannot increase in pre_destroy.
> >
> >
> > You're wrong here. We release cgroup_lock before calling pre_destroy and retrieve
> > the lock after that, so a task can be attached to the cgroup in this interval.
>
> But that means rmdir can be racy right ? What happens if the task got
> added, allocated few pages and then moved out ? We still would have task
> count 0 but few pages, which we missed to to move to parent cgroup.

That's a problem even if it's verrrry unlikely.
I'd like to look into it and fix the race in cgroup layer.
But I'm sorry I'm a bit busy in these days...

Thanks,
-Kame


--
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/