[tip:timers/urgent] tick/broadcast: Prevent deadlock on tick_broadcast_lock

[tip-bot for Mike Galbraith]

Commit-ID:  202461e2f3c15dbfb05825d29ace0d20cdf55fa4
Gitweb:     http://git.kernel.org/tip/202461e2f3c15dbfb05825d29ace0d20cdf55fa4
Author:     Mike Galbraith <efault@xxxxxx>
AuthorDate: Mon, 13 Feb 2017 03:31:55 +0100
Committer:  Thomas Gleixner <tglx@xxxxxxxxxxxxx>
CommitDate: Mon, 13 Feb 2017 09:49:31 +0100

tick/broadcast: Prevent deadlock on tick_broadcast_lock

tick_broadcast_lock is taken from interrupt context, but the following call
chain takes the lock without disabling interrupts:

[   12.703736]  _raw_spin_lock+0x3b/0x50
[   12.703738]  tick_broadcast_control+0x5a/0x1a0
[   12.703742]  intel_idle_cpu_online+0x22/0x100
[   12.703744]  cpuhp_invoke_callback+0x245/0x9d0
[   12.703752]  cpuhp_thread_fun+0x52/0x110
[   12.703754]  smpboot_thread_fn+0x276/0x320

So the following deadlock can happen:

   lock(tick_broadcast_lock);
   <Interrupt>
      lock(tick_broadcast_lock);

intel_idle_cpu_online() is the only place which violates the calling
convention of tick_broadcast_control(). This was caused by the removal of
the smp function call in course of the cpu hotplug rework.

Instead of slapping local_irq_disable/enable() at the call site, we can
relax the calling convention and handle it in the core code, which makes
the whole machinery more robust.

Fixes: 29d7bbada98e ("intel_idle: Remove superfluous SMP fuction call")
Reported-by: Gabriel C <nix.or.die@xxxxxxxxx>
Signed-off-by: Mike Galbraith <efault@xxxxxx>
Cc: Ruslan Ruslichenko <rruslich@xxxxxxxxx>
Cc: Jiri Slaby <jslaby@xxxxxxx>
Cc: Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx>
Cc: Borislav Petkov <bp@xxxxxxxxx>
Cc: lwn@xxxxxxx
Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
Cc: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
Cc: Anna-Maria Gleixner <anna-maria@xxxxxxxxxxxxx>
Cc: Sebastian Siewior <bigeasy@xxxxxxxxxxxxx>
Cc: stable <stable@xxxxxxxxxxxxxxx>
Link: http://lkml.kernel.org/r/1486953115.5912.4.camel@xxxxxx
Signed-off-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>

---
 kernel/time/tick-broadcast.c | 15 +++++++--------
 1 file changed, 7 insertions(+), 8 deletions(-)

diff --git a/kernel/time/tick-broadcast.c b/kernel/time/tick-broadcast.c
index 3109204..17ac99b 100644
--- a/kernel/time/tick-broadcast.c
+++ b/kernel/time/tick-broadcast.c
@@ -347,17 +347,16 @@ static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
  *
  * Called when the system enters a state where affected tick devices
  * might stop. Note: TICK_BROADCAST_FORCE cannot be undone.
- *
- * Called with interrupts disabled, so clockevents_lock is not
- * required here because the local clock event device cannot go away
- * under us.
  */
 void tick_broadcast_control(enum tick_broadcast_mode mode)
 {
 	struct clock_event_device *bc, *dev;
 	struct tick_device *td;
 	int cpu, bc_stopped;
+	unsigned long flags;
 
+	/* Protects also the local clockevent device. */
+	raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 	td = this_cpu_ptr(&tick_cpu_device);
 	dev = td->evtdev;
 
@@ -365,12 +364,11 @@ void tick_broadcast_control(enum tick_broadcast_mode mode)
 	 * Is the device not affected by the powerstate ?
 	 */
 	if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
-		return;
+		goto out;
 
 	if (!tick_device_is_functional(dev))
-		return;
+		goto out;
 
-	raw_spin_lock(&tick_broadcast_lock);
 	cpu = smp_processor_id();
 	bc = tick_broadcast_device.evtdev;
 	bc_stopped = cpumask_empty(tick_broadcast_mask);
@@ -420,7 +418,8 @@ void tick_broadcast_control(enum tick_broadcast_mode mode)
 				tick_broadcast_setup_oneshot(bc);
 		}
 	}
-	raw_spin_unlock(&tick_broadcast_lock);
+out:
+	raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 }
 EXPORT_SYMBOL_GPL(tick_broadcast_control);