Re: [PATCH v4] mm, slab: release slab_mutex earlier inkmem_cache_destroy()

From: David Rientjes
Date: Wed Oct 03 2012 - 14:54:39 EST


On Wed, 3 Oct 2012, Jiri Kosina wrote:

> 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 itss 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
>
> Signed-off-by: Jiri Kosina <jkosina@xxxxxxx>

Acked-by: David Rientjes <rientjes@xxxxxxxxxx>
--
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/