lkml.org 
[lkml]   [2012]   [Feb]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 3.3] memcg: fix deadlock by inverting lrucare nesting
On Tue, Feb 28, 2012 at 09:25:02PM -0800, Hugh Dickins wrote:
> We have forgotten the rules of lock nesting: the irq-safe ones must be
> taken inside the non-irq-safe ones, otherwise we are open to deadlock:
>
> CPU0 CPU1
> ---- ----
> lock(&(&pc->lock)->rlock);
> local_irq_disable();
> lock(&(&zone->lru_lock)->rlock);
> lock(&(&pc->lock)->rlock);
> <Interrupt>
> lock(&(&zone->lru_lock)->rlock);
>
> To check a different locking issue, I happened to add a spin_lock to
> memcg's bit_spin_lock in lock_page_cgroup(), and lockdep very quickly
> complained about __mem_cgroup_commit_charge_lrucare() (on CPU1 above).
>
> So delete __mem_cgroup_commit_charge_lrucare(), passing a bool lrucare
> to __mem_cgroup_commit_charge() instead, taking zone->lru_lock under
> lock_page_cgroup() in the lrucare case.
>
> The original was using spin_lock_irqsave, but we'd be in more trouble
> if it were ever called at interrupt time: unconditional _irq is enough.
> And ClearPageLRU before del from lru, SetPageLRU before add to lru: no
> strong reason, but that is the ordering used consistently elsewhere.
>
> Signed-off-by: Hugh Dickins <hughd@google.com>

Acked-by: Johannes Weiner <hannes@cmpxchg.org>


\
 
 \ /
  Last update: 2012-02-29 20:03    [from the cache]
©2003-2011 Jasper Spaans