rcu: Break call_rcu() deadlock involving scheduler and perf
authorPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Fri, 4 Oct 2013 21:33:34 +0000 (14:33 -0700)
committerPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Tue, 3 Dec 2013 18:10:18 +0000 (10:10 -0800)
commit96d3fd0d315a949e30adc80f086031c5cdf070d1
tree0fe7013d59b4d69a91bf031c0a53e8d279413e4a
parent78e4bc34e5d966cfd95f1238565afc399d56225c
rcu: Break call_rcu() deadlock involving scheduler and perf

Dave Jones got the following lockdep splat:

>  ======================================================
>  [ INFO: possible circular locking dependency detected ]
>  3.12.0-rc3+ #92 Not tainted
>  -------------------------------------------------------
>  trinity-child2/15191 is trying to acquire lock:
>   (&rdp->nocb_wq){......}, at: [<ffffffff8108ff43>] __wake_up+0x23/0x50
>
> but task is already holding lock:
>   (&ctx->lock){-.-...}, at: [<ffffffff81154c19>] perf_event_exit_task+0x109/0x230
>
> which lock already depends on the new lock.
>
>
> the existing dependency chain (in reverse order) is:
>
> -> #3 (&ctx->lock){-.-...}:
>         [<ffffffff810cc243>] lock_acquire+0x93/0x200
>         [<ffffffff81733f90>] _raw_spin_lock+0x40/0x80
>         [<ffffffff811500ff>] __perf_event_task_sched_out+0x2df/0x5e0
>         [<ffffffff81091b83>] perf_event_task_sched_out+0x93/0xa0
>         [<ffffffff81732052>] __schedule+0x1d2/0xa20
>         [<ffffffff81732f30>] preempt_schedule_irq+0x50/0xb0
>         [<ffffffff817352b6>] retint_kernel+0x26/0x30
>         [<ffffffff813eed04>] tty_flip_buffer_push+0x34/0x50
>         [<ffffffff813f0504>] pty_write+0x54/0x60
>         [<ffffffff813e900d>] n_tty_write+0x32d/0x4e0
>         [<ffffffff813e5838>] tty_write+0x158/0x2d0
>         [<ffffffff811c4850>] vfs_write+0xc0/0x1f0
>         [<ffffffff811c52cc>] SyS_write+0x4c/0xa0
>         [<ffffffff8173d4e4>] tracesys+0xdd/0xe2
>
> -> #2 (&rq->lock){-.-.-.}:
>         [<ffffffff810cc243>] lock_acquire+0x93/0x200
>         [<ffffffff81733f90>] _raw_spin_lock+0x40/0x80
>         [<ffffffff810980b2>] wake_up_new_task+0xc2/0x2e0
>         [<ffffffff81054336>] do_fork+0x126/0x460
>         [<ffffffff81054696>] kernel_thread+0x26/0x30
>         [<ffffffff8171ff93>] rest_init+0x23/0x140
>         [<ffffffff81ee1e4b>] start_kernel+0x3f6/0x403
>         [<ffffffff81ee1571>] x86_64_start_reservations+0x2a/0x2c
>         [<ffffffff81ee1664>] x86_64_start_kernel+0xf1/0xf4
>
> -> #1 (&p->pi_lock){-.-.-.}:
>         [<ffffffff810cc243>] lock_acquire+0x93/0x200
>         [<ffffffff8173419b>] _raw_spin_lock_irqsave+0x4b/0x90
>         [<ffffffff810979d1>] try_to_wake_up+0x31/0x350
>         [<ffffffff81097d62>] default_wake_function+0x12/0x20
>         [<ffffffff81084af8>] autoremove_wake_function+0x18/0x40
>         [<ffffffff8108ea38>] __wake_up_common+0x58/0x90
>         [<ffffffff8108ff59>] __wake_up+0x39/0x50
>         [<ffffffff8110d4f8>] __call_rcu_nocb_enqueue+0xa8/0xc0
>         [<ffffffff81111450>] __call_rcu+0x140/0x820
>         [<ffffffff81111b8d>] call_rcu+0x1d/0x20
>         [<ffffffff81093697>] cpu_attach_domain+0x287/0x360
>         [<ffffffff81099d7e>] build_sched_domains+0xe5e/0x10a0
>         [<ffffffff81efa7fc>] sched_init_smp+0x3b7/0x47a
>         [<ffffffff81ee1f4e>] kernel_init_freeable+0xf6/0x202
>         [<ffffffff817200be>] kernel_init+0xe/0x190
>         [<ffffffff8173d22c>] ret_from_fork+0x7c/0xb0
>
> -> #0 (&rdp->nocb_wq){......}:
>         [<ffffffff810cb7ca>] __lock_acquire+0x191a/0x1be0
>         [<ffffffff810cc243>] lock_acquire+0x93/0x200
>         [<ffffffff8173419b>] _raw_spin_lock_irqsave+0x4b/0x90
>         [<ffffffff8108ff43>] __wake_up+0x23/0x50
>         [<ffffffff8110d4f8>] __call_rcu_nocb_enqueue+0xa8/0xc0
>         [<ffffffff81111450>] __call_rcu+0x140/0x820
>         [<ffffffff81111bb0>] kfree_call_rcu+0x20/0x30
>         [<ffffffff81149abf>] put_ctx+0x4f/0x70
>         [<ffffffff81154c3e>] perf_event_exit_task+0x12e/0x230
>         [<ffffffff81056b8d>] do_exit+0x30d/0xcc0
>         [<ffffffff8105893c>] do_group_exit+0x4c/0xc0
>         [<ffffffff810589c4>] SyS_exit_group+0x14/0x20
>         [<ffffffff8173d4e4>] tracesys+0xdd/0xe2
>
> other info that might help us debug this:
>
> Chain exists of:
>   &rdp->nocb_wq --> &rq->lock --> &ctx->lock
>
>   Possible unsafe locking scenario:
>
>         CPU0                    CPU1
>         ----                    ----
>    lock(&ctx->lock);
>                                 lock(&rq->lock);
>                                 lock(&ctx->lock);
>    lock(&rdp->nocb_wq);
>
>  *** DEADLOCK ***
>
> 1 lock held by trinity-child2/15191:
>  #0:  (&ctx->lock){-.-...}, at: [<ffffffff81154c19>] perf_event_exit_task+0x109/0x230
>
> stack backtrace:
> CPU: 2 PID: 15191 Comm: trinity-child2 Not tainted 3.12.0-rc3+ #92
>  ffffffff82565b70 ffff880070c2dbf8 ffffffff8172a363 ffffffff824edf40
>  ffff880070c2dc38 ffffffff81726741 ffff880070c2dc90 ffff88022383b1c0
>  ffff88022383aac0 0000000000000000 ffff88022383b188 ffff88022383b1c0
> Call Trace:
>  [<ffffffff8172a363>] dump_stack+0x4e/0x82
>  [<ffffffff81726741>] print_circular_bug+0x200/0x20f
>  [<ffffffff810cb7ca>] __lock_acquire+0x191a/0x1be0
>  [<ffffffff810c6439>] ? get_lock_stats+0x19/0x60
>  [<ffffffff8100b2f4>] ? native_sched_clock+0x24/0x80
>  [<ffffffff810cc243>] lock_acquire+0x93/0x200
>  [<ffffffff8108ff43>] ? __wake_up+0x23/0x50
>  [<ffffffff8173419b>] _raw_spin_lock_irqsave+0x4b/0x90
>  [<ffffffff8108ff43>] ? __wake_up+0x23/0x50
>  [<ffffffff8108ff43>] __wake_up+0x23/0x50
>  [<ffffffff8110d4f8>] __call_rcu_nocb_enqueue+0xa8/0xc0
>  [<ffffffff81111450>] __call_rcu+0x140/0x820
>  [<ffffffff8109bc8f>] ? local_clock+0x3f/0x50
>  [<ffffffff81111bb0>] kfree_call_rcu+0x20/0x30
>  [<ffffffff81149abf>] put_ctx+0x4f/0x70
>  [<ffffffff81154c3e>] perf_event_exit_task+0x12e/0x230
>  [<ffffffff81056b8d>] do_exit+0x30d/0xcc0
>  [<ffffffff810c9af5>] ? trace_hardirqs_on_caller+0x115/0x1e0
>  [<ffffffff810c9bcd>] ? trace_hardirqs_on+0xd/0x10
>  [<ffffffff8105893c>] do_group_exit+0x4c/0xc0
>  [<ffffffff810589c4>] SyS_exit_group+0x14/0x20
>  [<ffffffff8173d4e4>] tracesys+0xdd/0xe2

The underlying problem is that perf is invoking call_rcu() with the
scheduler locks held, but in NOCB mode, call_rcu() will with high
probability invoke the scheduler -- which just might want to use its
locks.  The reason that call_rcu() needs to invoke the scheduler is
to wake up the corresponding rcuo callback-offload kthread, which
does the job of starting up a grace period and invoking the callbacks
afterwards.

One solution (championed on a related problem by Lai Jiangshan) is to
simply defer the wakeup to some point where scheduler locks are no longer
held.  Since we don't want to unnecessarily incur the cost of such
deferral, the task before us is threefold:

1. Determine when it is likely that a relevant scheduler lock is held.

2. Defer the wakeup in such cases.

3. Ensure that all deferred wakeups eventually happen, preferably
sooner rather than later.

We use irqs_disabled_flags() as a proxy for relevant scheduler locks
being held.  This works because the relevant locks are always acquired
with interrupts disabled.  We may defer more often than needed, but that
is at least safe.

The wakeup deferral is tracked via a new field in the per-CPU and
per-RCU-flavor rcu_data structure, namely ->nocb_defer_wakeup.

This flag is checked by the RCU core processing.  The __rcu_pending()
function now checks this flag, which causes rcu_check_callbacks()
to initiate RCU core processing at each scheduling-clock interrupt
where this flag is set.  Of course this is not sufficient because
scheduling-clock interrupts are often turned off (the things we used to
be able to count on!).  So the flags are also checked on entry to any
state that RCU considers to be idle, which includes both NO_HZ_IDLE idle
state and NO_HZ_FULL user-mode-execution state.

This approach should allow call_rcu() to be invoked regardless of what
locks you might be holding, the key word being "should".

Reported-by: Dave Jones <davej@redhat.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Documentation/RCU/trace.txt
kernel/rcu/tree.c
kernel/rcu/tree.h
kernel/rcu/tree_plugin.h
kernel/rcu/tree_trace.c