On Tue, Aug 16, 2016 at 04:34:08PM +0200, Oleg Nesterov wrote:
> On 08/16, Oleg Nesterov wrote:
> >
> > On 08/16, Oleg Nesterov wrote:
> > >
> > > On 08/15, Brenden Blanco wrote:
> > > >
> > > > Hi folks,
> > > >
> > > > I think I have come across a memory leak in uprobes, which is fairly
> > > > easy to
> > > > reproduce.
> > >
> > > At first glance this looks as a problem in memcg, add CC's...
> > >
> > > put_page(old_page) looks properly balanced, and I assume we do not need
> > > the additional "uncharge", we can rely on __page_cache_release().
> > >
> > > And I do not see any leak if I try to reproduce with CONFIG_MEMCG=n.
> >
> > Heh. it seems that mem_cgroup_*() logic was always wrong in
> > __replace_page().
>
> Yes, it seems this was broken by 00501b53 "mm: memcontrol: rewrite charge
> API".
>
> > Could you try the patch below?
>
> Please see v2 below. We don't need "cancel_charge" under "unlock:" at all.
>
> Johannes, could you review?
>
> Oleg.
> ---
> --- x/kernel/events/uprobes.c
> +++ x/kernel/events/uprobes.c
> @@ -172,8 +172,10 @@ static int __replace_page(struct vm_area
> mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
> err = -EAGAIN;
> ptep = page_check_address(page, mm, addr, &ptl, 0);
> - if (!ptep)
> + if (!ptep) {
> + mem_cgroup_cancel_charge(kpage, memcg, false);
> goto unlock;
> + }
>
> get_page(kpage);
> page_add_new_anon_rmap(kpage, vma, addr, false);
> @@ -200,7 +202,6 @@ static int __replace_page(struct vm_area
>
> err = 0;
> unlock:
> - mem_cgroup_cancel_charge(kpage, memcg, false);
> mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
> unlock_page(page);
> return err;
>
This passes my tests, thanks!
Please note that I applied this to 4.4.15+ubuntu-patches kernel, since that
was what I had most handy, therefore I had to adjust the patch to remove the
unavailable 'compound' bool parameter in 4.4 kernels.
