mm, slab: release slab_mutex earlier in kmem_cache_destroy()
authorJiri Kosina <jkosina@suse.cz>
Mon, 8 Oct 2012 07:26:01 +0000 (09:26 +0200)
committerPekka Enberg <penberg@kernel.org>
Wed, 10 Oct 2012 06:25:08 +0000 (09:25 +0300)
commit210ed9defffca13b909f040d7338d8062e5594a3
treeb939e0be42022f20bcdc73ed7a669ad611d30595
parente2087be35a8ed101c1e748ef688c889419c69829
mm, slab: release slab_mutex earlier in kmem_cache_destroy()

Commit 1331e7a1bbe1 ("rcu: Remove _rcu_barrier() dependency on
__stop_machine()") introduced slab_mutex -> cpu_hotplug.lock dependency
through kmem_cache_destroy() -> rcu_barrier() -> _rcu_barrier() ->
get_online_cpus().

Lockdep thinks that this might actually result in ABBA deadlock,
and reports it as below:

=== [ cut here ] ===
 ======================================================
 [ INFO: possible circular locking dependency detected ]
 3.6.0-rc5-00004-g0d8ee37 #143 Not tainted
 -------------------------------------------------------
 kworker/u:2/40 is trying to acquire lock:
  (rcu_sched_state.barrier_mutex){+.+...}, at: [<ffffffff810f2126>] _rcu_barrier+0x26/0x1e0

 but task is already holding lock:
  (slab_mutex){+.+.+.}, at: [<ffffffff81176e15>] kmem_cache_destroy+0x45/0xe0

 which lock already depends on the new lock.

 the existing dependency chain (in reverse order) is:

 -> #2 (slab_mutex){+.+.+.}:
        [<ffffffff810ae1e2>] validate_chain+0x632/0x720
        [<ffffffff810ae5d9>] __lock_acquire+0x309/0x530
        [<ffffffff810ae921>] lock_acquire+0x121/0x190
        [<ffffffff8155d4cc>] __mutex_lock_common+0x5c/0x450
        [<ffffffff8155d9ee>] mutex_lock_nested+0x3e/0x50
        [<ffffffff81558cb5>] cpuup_callback+0x2f/0xbe
        [<ffffffff81564b83>] notifier_call_chain+0x93/0x140
        [<ffffffff81076f89>] __raw_notifier_call_chain+0x9/0x10
        [<ffffffff8155719d>] _cpu_up+0xba/0x14e
        [<ffffffff815572ed>] cpu_up+0xbc/0x117
        [<ffffffff81ae05e3>] smp_init+0x6b/0x9f
        [<ffffffff81ac47d6>] kernel_init+0x147/0x1dc
        [<ffffffff8156ab44>] kernel_thread_helper+0x4/0x10

 -> #1 (cpu_hotplug.lock){+.+.+.}:
        [<ffffffff810ae1e2>] validate_chain+0x632/0x720
        [<ffffffff810ae5d9>] __lock_acquire+0x309/0x530
        [<ffffffff810ae921>] lock_acquire+0x121/0x190
        [<ffffffff8155d4cc>] __mutex_lock_common+0x5c/0x450
        [<ffffffff8155d9ee>] mutex_lock_nested+0x3e/0x50
        [<ffffffff81049197>] get_online_cpus+0x37/0x50
        [<ffffffff810f21bb>] _rcu_barrier+0xbb/0x1e0
        [<ffffffff810f22f0>] rcu_barrier_sched+0x10/0x20
        [<ffffffff810f2309>] rcu_barrier+0x9/0x10
        [<ffffffff8118c129>] deactivate_locked_super+0x49/0x90
        [<ffffffff8118cc01>] deactivate_super+0x61/0x70
        [<ffffffff811aaaa7>] mntput_no_expire+0x127/0x180
        [<ffffffff811ab49e>] sys_umount+0x6e/0xd0
        [<ffffffff81569979>] system_call_fastpath+0x16/0x1b

 -> #0 (rcu_sched_state.barrier_mutex){+.+...}:
        [<ffffffff810adb4e>] check_prev_add+0x3de/0x440
        [<ffffffff810ae1e2>] validate_chain+0x632/0x720
        [<ffffffff810ae5d9>] __lock_acquire+0x309/0x530
        [<ffffffff810ae921>] lock_acquire+0x121/0x190
        [<ffffffff8155d4cc>] __mutex_lock_common+0x5c/0x450
        [<ffffffff8155d9ee>] mutex_lock_nested+0x3e/0x50
        [<ffffffff810f2126>] _rcu_barrier+0x26/0x1e0
        [<ffffffff810f22f0>] rcu_barrier_sched+0x10/0x20
        [<ffffffff810f2309>] rcu_barrier+0x9/0x10
        [<ffffffff81176ea1>] kmem_cache_destroy+0xd1/0xe0
        [<ffffffffa04c3154>] nf_conntrack_cleanup_net+0xe4/0x110 [nf_conntrack]
        [<ffffffffa04c31aa>] nf_conntrack_cleanup+0x2a/0x70 [nf_conntrack]
        [<ffffffffa04c42ce>] nf_conntrack_net_exit+0x5e/0x80 [nf_conntrack]
        [<ffffffff81454b79>] ops_exit_list+0x39/0x60
        [<ffffffff814551ab>] cleanup_net+0xfb/0x1b0
        [<ffffffff8106917b>] process_one_work+0x26b/0x4c0
        [<ffffffff81069f3e>] worker_thread+0x12e/0x320
        [<ffffffff8106f73e>] kthread+0x9e/0xb0
        [<ffffffff8156ab44>] kernel_thread_helper+0x4/0x10

 other info that might help us debug this:

 Chain exists of:
   rcu_sched_state.barrier_mutex --> cpu_hotplug.lock --> slab_mutex

  Possible unsafe locking scenario:

        CPU0                    CPU1
        ----                    ----
   lock(slab_mutex);
                                lock(cpu_hotplug.lock);
                                lock(slab_mutex);
   lock(rcu_sched_state.barrier_mutex);

  *** DEADLOCK ***
=== [ cut here ] ===

This is actually a false positive. Lockdep has no way of knowing the fact
that the ABBA can actually never happen, because of special semantics of
cpu_hotplug.refcount and its handling in cpu_hotplug_begin(); the mutual
exclusion there is not achieved through mutex, but through
cpu_hotplug.refcount.

The "neither cpu_up() nor cpu_down() will proceed past cpu_hotplug_begin()
until everyone who called get_online_cpus() will call put_online_cpus()"
semantics is totally invisible to lockdep.

This patch therefore moves the unlock of slab_mutex so that rcu_barrier()
is being called with it unlocked. It has two advantages:

- it slightly reduces hold time of slab_mutex; as it's used to protect
  the cachep list, it's not necessary to hold it over kmem_cache_free()
  call any more
- it silences the lockdep false positive warning, as it avoids lockdep ever
  learning about slab_mutex -> cpu_hotplug.lock dependency

Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Jiri Kosina <jkosina@suse.cz>
Signed-off-by: Pekka Enberg <penberg@kernel.org>
mm/slab_common.c