Re: [PATCH -mm 2/2] memcg: fix css reference leak and endless loopin mem_cgroup_iter

From: Michal Hocko
Date: Wed Jan 22 2014 - 03:27:32 EST

On Tue 21-01-14 13:18:42, Hugh Dickins wrote:
> We do have a confusing situation. The hang goes back to 3.10 but takes
> two different forms, because of intervening changes: in 3.10 and 3.11
> mem_cgroup_iter repeatedly returns root memcg to its caller, in 3.12 and
> 3.13 mem_cgroup_iter repeatedly gets NULL memcg from mem_cgroup_iter_next
> and cannot return to its caller.
> Patch 1/2 is what's needed to fix 3.10 and 3.11 (and applies correctly
> to 3.11, but will have to be rediffed for 3.10 because of rearrangement
> in between).

I will backport it when it reaches stable queue.

> Patch 2/2 is what's needed to fix 3.12 and 3.13 (but applies
> correctly to neither of them because it's diffed on top of my CSS_ONLINE
> fix). Patch 1/2 is correct but unnecessary in 3.12 and 3.13: I'm unclear
> whether Michal is claiming that it would also fix the hang in 3.12 and
> 3.13 if we didn't have 2/2: I doubt that, and haven't tested that.

Actually both patches are needed. If we had only 2/2 then we wouldn't
endless loop inside mem_cgroup_iter but we could still return root to
caller all the time because mem_cgroup_iter_load would return NULL on
css_tryget failure on the cached root. Or am I missing something that
would prevent that?

> Given how Michal has diffed this patch on top of my CSS_ONLINE one
> (mm-memcg-iteration-skip-memcgs-not-yet-fully-initialized.patch),
> it would be helpful if you could mark that one also for stable 3.12+,
> to save us from having to rediff this one for stable. We don't have
> a concrete example of a problem it solves in the vanilla kernel, but
> it makes more sense to include it than to exclude it.

Yes, I think it makes sense to queue it for 3.12+ as well because it is
non intrusive and potential issues would be really subtle.

> (You would be right to point out that the CSS_ONLINE one fixes
> something that goes back to 3.10: I'm saying 3.12+ because I'm not
> motivated to rediff it for 3.10 and 3.11 when there's nothing to
> go on top; but that's not a very good reason to lie - overrule me.)
> Hugh

Michal Hocko
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at