rcu: Don't call wakeup() with rcu_node structure ->lock held
authorSteven Rostedt <rostedt@goodmis.org>
Tue, 28 May 2013 21:32:53 +0000 (17:32 -0400)
committerPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Mon, 10 Jun 2013 20:37:11 +0000 (13:37 -0700)
This commit fixes a lockdep-detected deadlock by moving a wake_up()
call out from a rnp->lock critical section.  Please see below for
the long version of this story.

On Tue, 2013-05-28 at 16:13 -0400, Dave Jones wrote:

> [12572.705832] ======================================================
> [12572.750317] [ INFO: possible circular locking dependency detected ]
> [12572.796978] 3.10.0-rc3+ #39 Not tainted
> [12572.833381] -------------------------------------------------------
> [12572.862233] trinity-child17/31341 is trying to acquire lock:
> [12572.870390]  (rcu_node_0){..-.-.}, at: [<ffffffff811054ff>] rcu_read_unlock_special+0x9f/0x4c0
> [12572.878859]
> but task is already holding lock:
> [12572.894894]  (&ctx->lock){-.-...}, at: [<ffffffff811390ed>] perf_lock_task_context+0x7d/0x2d0
> [12572.903381]
> which lock already depends on the new lock.
>
> [12572.927541]
> the existing dependency chain (in reverse order) is:
> [12572.943736]
> -> #4 (&ctx->lock){-.-...}:
> [12572.960032]        [<ffffffff810b9851>] lock_acquire+0x91/0x1f0
> [12572.968337]        [<ffffffff816ebc90>] _raw_spin_lock+0x40/0x80
> [12572.976633]        [<ffffffff8113c987>] __perf_event_task_sched_out+0x2e7/0x5e0
> [12572.984969]        [<ffffffff81088953>] perf_event_task_sched_out+0x93/0xa0
> [12572.993326]        [<ffffffff816ea0bf>] __schedule+0x2cf/0x9c0
> [12573.001652]        [<ffffffff816eacfe>] schedule_user+0x2e/0x70
> [12573.009998]        [<ffffffff816ecd64>] retint_careful+0x12/0x2e
> [12573.018321]
> -> #3 (&rq->lock){-.-.-.}:
> [12573.034628]        [<ffffffff810b9851>] lock_acquire+0x91/0x1f0
> [12573.042930]        [<ffffffff816ebc90>] _raw_spin_lock+0x40/0x80
> [12573.051248]        [<ffffffff8108e6a7>] wake_up_new_task+0xb7/0x260
> [12573.059579]        [<ffffffff810492f5>] do_fork+0x105/0x470
> [12573.067880]        [<ffffffff81049686>] kernel_thread+0x26/0x30
> [12573.076202]        [<ffffffff816cee63>] rest_init+0x23/0x140
> [12573.084508]        [<ffffffff81ed8e1f>] start_kernel+0x3f1/0x3fe
> [12573.092852]        [<ffffffff81ed856f>] x86_64_start_reservations+0x2a/0x2c
> [12573.101233]        [<ffffffff81ed863d>] x86_64_start_kernel+0xcc/0xcf
> [12573.109528]
> -> #2 (&p->pi_lock){-.-.-.}:
> [12573.125675]        [<ffffffff810b9851>] lock_acquire+0x91/0x1f0
> [12573.133829]        [<ffffffff816ebe9b>] _raw_spin_lock_irqsave+0x4b/0x90
> [12573.141964]        [<ffffffff8108e881>] try_to_wake_up+0x31/0x320
> [12573.150065]        [<ffffffff8108ebe2>] default_wake_function+0x12/0x20
> [12573.158151]        [<ffffffff8107bbf8>] autoremove_wake_function+0x18/0x40
> [12573.166195]        [<ffffffff81085398>] __wake_up_common+0x58/0x90
> [12573.174215]        [<ffffffff81086909>] __wake_up+0x39/0x50
> [12573.182146]        [<ffffffff810fc3da>] rcu_start_gp_advanced.isra.11+0x4a/0x50
> [12573.190119]        [<ffffffff810fdb09>] rcu_start_future_gp+0x1c9/0x1f0
> [12573.198023]        [<ffffffff810fe2c4>] rcu_nocb_kthread+0x114/0x930
> [12573.205860]        [<ffffffff8107a91d>] kthread+0xed/0x100
> [12573.213656]        [<ffffffff816f4b1c>] ret_from_fork+0x7c/0xb0
> [12573.221379]
> -> #1 (&rsp->gp_wq){..-.-.}:
> [12573.236329]        [<ffffffff810b9851>] lock_acquire+0x91/0x1f0
> [12573.243783]        [<ffffffff816ebe9b>] _raw_spin_lock_irqsave+0x4b/0x90
> [12573.251178]        [<ffffffff810868f3>] __wake_up+0x23/0x50
> [12573.258505]        [<ffffffff810fc3da>] rcu_start_gp_advanced.isra.11+0x4a/0x50
> [12573.265891]        [<ffffffff810fdb09>] rcu_start_future_gp+0x1c9/0x1f0
> [12573.273248]        [<ffffffff810fe2c4>] rcu_nocb_kthread+0x114/0x930
> [12573.280564]        [<ffffffff8107a91d>] kthread+0xed/0x100
> [12573.287807]        [<ffffffff816f4b1c>] ret_from_fork+0x7c/0xb0

Notice the above call chain.

rcu_start_future_gp() is called with the rnp->lock held. Then it calls
rcu_start_gp_advance, which does a wakeup.

You can't do wakeups while holding the rnp->lock, as that would mean
that you could not do a rcu_read_unlock() while holding the rq lock, or
any lock that was taken while holding the rq lock. This is because...
(See below).

> [12573.295067]
> -> #0 (rcu_node_0){..-.-.}:
> [12573.309293]        [<ffffffff810b8d36>] __lock_acquire+0x1786/0x1af0
> [12573.316568]        [<ffffffff810b9851>] lock_acquire+0x91/0x1f0
> [12573.323825]        [<ffffffff816ebc90>] _raw_spin_lock+0x40/0x80
> [12573.331081]        [<ffffffff811054ff>] rcu_read_unlock_special+0x9f/0x4c0
> [12573.338377]        [<ffffffff810760a6>] __rcu_read_unlock+0x96/0xa0
> [12573.345648]        [<ffffffff811391b3>] perf_lock_task_context+0x143/0x2d0
> [12573.352942]        [<ffffffff8113938e>] find_get_context+0x4e/0x1f0
> [12573.360211]        [<ffffffff811403f4>] SYSC_perf_event_open+0x514/0xbd0
> [12573.367514]        [<ffffffff81140e49>] SyS_perf_event_open+0x9/0x10
> [12573.374816]        [<ffffffff816f4dd4>] tracesys+0xdd/0xe2

Notice the above trace.

perf took its own ctx->lock, which can be taken while holding the rq
lock. While holding this lock, it did a rcu_read_unlock(). The
perf_lock_task_context() basically looks like:

rcu_read_lock();
raw_spin_lock(ctx->lock);
rcu_read_unlock();

Now, what looks to have happened, is that we scheduled after taking that
first rcu_read_lock() but before taking the spin lock. When we scheduled
back in and took the ctx->lock, the following rcu_read_unlock()
triggered the "special" code.

The rcu_read_unlock_special() takes the rnp->lock, which gives us a
possible deadlock scenario.

CPU0 CPU1 CPU2
---- ---- ----

     rcu_nocb_kthread()
    lock(rq->lock);
    lock(ctx->lock);
     lock(rnp->lock);

     wake_up();

     lock(rq->lock);

    rcu_read_unlock();

    rcu_read_unlock_special();

    lock(rnp->lock);
    lock(ctx->lock);

**** DEADLOCK ****

> [12573.382068]
> other info that might help us debug this:
>
> [12573.403229] Chain exists of:
>   rcu_node_0 --> &rq->lock --> &ctx->lock
>
> [12573.424471]  Possible unsafe locking scenario:
>
> [12573.438499]        CPU0                    CPU1
> [12573.445599]        ----                    ----
> [12573.452691]   lock(&ctx->lock);
> [12573.459799]                                lock(&rq->lock);
> [12573.467010]                                lock(&ctx->lock);
> [12573.474192]   lock(rcu_node_0);
> [12573.481262]
>  *** DEADLOCK ***
>
> [12573.501931] 1 lock held by trinity-child17/31341:
> [12573.508990]  #0:  (&ctx->lock){-.-...}, at: [<ffffffff811390ed>] perf_lock_task_context+0x7d/0x2d0
> [12573.516475]
> stack backtrace:
> [12573.530395] CPU: 1 PID: 31341 Comm: trinity-child17 Not tainted 3.10.0-rc3+ #39
> [12573.545357]  ffffffff825b4f90 ffff880219f1dbc0 ffffffff816e375b ffff880219f1dc00
> [12573.552868]  ffffffff816dfa5d ffff880219f1dc50 ffff88023ce4d1f8 ffff88023ce4ca40
> [12573.560353]  0000000000000001 0000000000000001 ffff88023ce4d1f8 ffff880219f1dcc0
> [12573.567856] Call Trace:
> [12573.575011]  [<ffffffff816e375b>] dump_stack+0x19/0x1b
> [12573.582284]  [<ffffffff816dfa5d>] print_circular_bug+0x200/0x20f
> [12573.589637]  [<ffffffff810b8d36>] __lock_acquire+0x1786/0x1af0
> [12573.596982]  [<ffffffff810918f5>] ? sched_clock_cpu+0xb5/0x100
> [12573.604344]  [<ffffffff810b9851>] lock_acquire+0x91/0x1f0
> [12573.611652]  [<ffffffff811054ff>] ? rcu_read_unlock_special+0x9f/0x4c0
> [12573.619030]  [<ffffffff816ebc90>] _raw_spin_lock+0x40/0x80
> [12573.626331]  [<ffffffff811054ff>] ? rcu_read_unlock_special+0x9f/0x4c0
> [12573.633671]  [<ffffffff811054ff>] rcu_read_unlock_special+0x9f/0x4c0
> [12573.640992]  [<ffffffff811390ed>] ? perf_lock_task_context+0x7d/0x2d0
> [12573.648330]  [<ffffffff810b429e>] ? put_lock_stats.isra.29+0xe/0x40
> [12573.655662]  [<ffffffff813095a0>] ? delay_tsc+0x90/0xe0
> [12573.662964]  [<ffffffff810760a6>] __rcu_read_unlock+0x96/0xa0
> [12573.670276]  [<ffffffff811391b3>] perf_lock_task_context+0x143/0x2d0
> [12573.677622]  [<ffffffff81139070>] ? __perf_event_enable+0x370/0x370
> [12573.684981]  [<ffffffff8113938e>] find_get_context+0x4e/0x1f0
> [12573.692358]  [<ffffffff811403f4>] SYSC_perf_event_open+0x514/0xbd0
> [12573.699753]  [<ffffffff8108cd9d>] ? get_parent_ip+0xd/0x50
> [12573.707135]  [<ffffffff810b71fd>] ? trace_hardirqs_on_caller+0xfd/0x1c0
> [12573.714599]  [<ffffffff81140e49>] SyS_perf_event_open+0x9/0x10
> [12573.721996]  [<ffffffff816f4dd4>] tracesys+0xdd/0xe2

This commit delays the wakeup via irq_work(), which is what
perf and ftrace use to perform wakeups in critical sections.

Reported-by: Dave Jones <davej@redhat.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
init/Kconfig
kernel/rcutree.c
kernel/rcutree.h

index 9d3a7887a6d3e0b7a8323865b658e1d088dcea85..2d9b83104dcf715197f1a84d34b892783ff79fe2 100644 (file)
@@ -431,6 +431,7 @@ choice
 config TREE_RCU
        bool "Tree-based hierarchical RCU"
        depends on !PREEMPT && SMP
+       select IRQ_WORK
        help
          This option selects the RCU implementation that is
          designed for very large SMP system with hundreds or
index 16ea67925015f19e7f693ed7a4557d898942beea..b61d20c5ee7b3e90aa385f2ce40a31ec2ab132eb 100644 (file)
@@ -1613,6 +1613,14 @@ static int __noreturn rcu_gp_kthread(void *arg)
        }
 }
 
+static void rsp_wakeup(struct irq_work *work)
+{
+       struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
+
+       /* Wake up rcu_gp_kthread() to start the grace period. */
+       wake_up(&rsp->gp_wq);
+}
+
 /*
  * Start a new RCU grace period if warranted, re-initializing the hierarchy
  * in preparation for detecting the next grace period.  The caller must hold
@@ -1637,8 +1645,12 @@ rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
        }
        rsp->gp_flags = RCU_GP_FLAG_INIT;
 
-       /* Wake up rcu_gp_kthread() to start the grace period. */
-       wake_up(&rsp->gp_wq);
+       /*
+        * We can't do wakeups while holding the rnp->lock, as that
+        * could cause possible deadlocks with the rq->lock. Deter
+        * the wakeup to interrupt context.
+        */
+       irq_work_queue(&rsp->wakeup_work);
 }
 
 /*
@@ -3235,6 +3247,7 @@ static void __init rcu_init_one(struct rcu_state *rsp,
 
        rsp->rda = rda;
        init_waitqueue_head(&rsp->gp_wq);
+       init_irq_work(&rsp->wakeup_work, rsp_wakeup);
        rnp = rsp->level[rcu_num_lvls - 1];
        for_each_possible_cpu(i) {
                while (i > rnp->grphi)
index da77a8f57ff95f80c7546684eba2ac293fdde08c..4df503470e420a24420edaf5d836d6aae508a155 100644 (file)
@@ -27,6 +27,7 @@
 #include <linux/threads.h>
 #include <linux/cpumask.h>
 #include <linux/seqlock.h>
+#include <linux/irq_work.h>
 
 /*
  * Define shape of hierarchy based on NR_CPUS, CONFIG_RCU_FANOUT, and
@@ -442,6 +443,7 @@ struct rcu_state {
        char *name;                             /* Name of structure. */
        char abbr;                              /* Abbreviated name. */
        struct list_head flavors;               /* List of RCU flavors. */
+       struct irq_work wakeup_work;            /* Postponed wakeups */
 };
 
 /* Values for rcu_state structure's gp_flags field. */