perf: Fix lockdep warning on process exit

Sasha Levin reported:

> While fuzzing with trinity inside a KVM tools guest running the latest -next
> kernel I've stumbled on the following spew:
>
> ======================================================
> [ INFO: possible circular locking dependency detected ]
> 3.15.0-next-20140613-sasha-00026-g6dd125d-dirty #654 Not tainted
> -------------------------------------------------------
> trinity-c578/9725 is trying to acquire lock:
> (&(&pool->lock)->rlock){-.-...}, at: __queue_work (kernel/workqueue.c:1346)
>
> but task is already holding lock:
> (&ctx->lock){-.....}, at: perf_event_exit_task (kernel/events/core.c:7471 kernel/events/core.c:7533)
>
> which lock already depends on the new lock.

> 1 lock held by trinity-c578/9725:
> #0: (&ctx->lock){-.....}, at: perf_event_exit_task (kernel/events/core.c:7471 kernel/events/core.c:7533)
>
>  Call Trace:
>  dump_stack (lib/dump_stack.c:52)
>  print_circular_bug (kernel/locking/lockdep.c:1216)
>  __lock_acquire (kernel/locking/lockdep.c:1840 kernel/locking/lockdep.c:1945 kernel/locking/lockdep.c:2131 kernel/locking/lockdep.c:3182)
>  lock_acquire (./arch/x86/include/asm/current.h:14 kernel/locking/lockdep.c:3602)
>  _raw_spin_lock (include/linux/spinlock_api_smp.h:143 kernel/locking/spinlock.c:151)
>  __queue_work (kernel/workqueue.c:1346)
>  queue_work_on (kernel/workqueue.c:1424)
>  free_object (lib/debugobjects.c:209)
>  __debug_check_no_obj_freed (lib/debugobjects.c:715)
>  debug_check_no_obj_freed (lib/debugobjects.c:727)
>  kmem_cache_free (mm/slub.c:2683 mm/slub.c:2711)
>  free_task (kernel/fork.c:221)
>  __put_task_struct (kernel/fork.c:250)
>  put_ctx (include/linux/sched.h:1855 kernel/events/core.c:898)
>  perf_event_exit_task (kernel/events/core.c:907 kernel/events/core.c:7478 kernel/events/core.c:7533)
>  do_exit (kernel/exit.c:766)
>  do_group_exit (kernel/exit.c:884)
>  get_signal_to_deliver (kernel/signal.c:2347)
>  do_signal (arch/x86/kernel/signal.c:698)
>  do_notify_resume (arch/x86/kernel/signal.c:751)
>  int_signal (arch/x86/kernel/entry_64.S:600)

Urgh.. so the only way I can make that happen is through:

  perf_event_exit_task_context()
    raw_spin_lock(&child_ctx->lock);
    unclone_ctx(child_ctx)
      put_ctx(ctx->parent_ctx);
    raw_spin_unlock_irqrestore(&child_ctx->lock);

And we can avoid this by doing the change below.

I can't immediately see how this changed recently, but given that you
say it's easy to reproduce, lets fix this.

Reported-by: Sasha Levin <[email protected]>
Signed-off-by: Peter Zijlstra <[email protected]>
Cc: Tejun Heo <[email protected]>
Cc: Dave Jones <[email protected]>
Cc: Arnaldo Carvalho de Melo <[email protected]>
Cc: Linus Torvalds <[email protected]>
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Ingo Molnar <[email protected]>

kernel/events/core.c

@@ -7486,7 +7486,7 @@ __perf_event_exit_task(struct perf_event *child_event,
 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
 {
 	struct perf_event *child_event, *next;
-	struct perf_event_context *child_ctx;
+	struct perf_event_context *child_ctx, *parent_ctx;
 	unsigned long flags;
 
 	if (likely(!child->perf_event_ctxp[ctxn])) {
@@ -7511,6 +7511,15 @@ static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
 	raw_spin_lock(&child_ctx->lock);
 	task_ctx_sched_out(child_ctx);
 	child->perf_event_ctxp[ctxn] = NULL;
+
+	/*
+	 * In order to avoid freeing: child_ctx->parent_ctx->task
+	 * under perf_event_context::lock, grab another reference.
+	 */
+	parent_ctx = child_ctx->parent_ctx;
+	if (parent_ctx)
+		get_ctx(parent_ctx);
+
 	/*
 	 * If this context is a clone; unclone it so it can't get
 	 * swapped to another process while we're removing all
@@ -7520,6 +7529,13 @@ static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
 	update_context_time(child_ctx);
 	raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
 
+	/*
+	 * Now that we no longer hold perf_event_context::lock, drop
+	 * our extra child_ctx->parent_ctx reference.
+	 */
+	if (parent_ctx)
+		put_ctx(parent_ctx);
+
 	/*
 	 * Report the task dead after unscheduling the events so that we
 	 * won't get any samples after PERF_RECORD_EXIT. We can however still