hrtimer: removing all ur callback modes, fix hotplug
Peter Zijlstra [Thu, 4 Dec 2008 10:17:10 +0000 (11:17 +0100)]
Impact: fix hrtimer locking (reported by lockdep) in the CPU hotplug case

This addition fixes the hotplug locking issue on my machine

Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

kernel/hrtimer.c

index efd6f41..b09c7a2 100644 (file)
@@ -1496,7 +1496,7 @@ static void __cpuinit init_hrtimers_cpu(int cpu)
 #ifdef CONFIG_HOTPLUG_CPU
 
 static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
-                                struct hrtimer_clock_base *new_base, int dcpu)
+                               struct hrtimer_clock_base *new_base)
 {
        struct hrtimer *timer;
        struct rb_node *node;
@@ -1514,40 +1514,34 @@ static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
                __remove_hrtimer(timer, old_base, HRTIMER_STATE_MIGRATE, 0);
                timer->base = new_base;
                /*
-                * Enqueue the timer. Allow reprogramming of the event device
+                * Enqueue the timers on the new cpu, but do not reprogram 
+                * the timer as that would enable a deadlock between
+                * hrtimer_enqueue_reprogramm() running the timer and us still
+                * holding a nested base lock.
+                *
+                * Instead we tickle the hrtimer interrupt after the migration
+                * is done, which will run all expired timers and re-programm
+                * the timer device.
                 */
-               enqueue_hrtimer(timer, new_base, 1);
+               enqueue_hrtimer(timer, new_base, 0);
 
-#ifdef CONFIG_HIGH_RES_TIMERS
-               /*
-                * Happens with high res enabled when the timer was
-                * already expired and the callback mode is
-                * HRTIMER_CB_IRQSAFE_UNLOCKED (hrtimer_sleeper). The
-                * enqueue code does not move them to the soft irq
-                * pending list for performance/latency reasons, but
-                * in the migration state, we need to do that
-                * otherwise we end up with a stale timer.
-                */
-               if (timer->state == HRTIMER_STATE_MIGRATE) {
-                       /* XXX: running on offline cpu */
-                       __run_hrtimer(timer);
-               }
-#endif
                /* Clear the migration state bit */
                timer->state &= ~HRTIMER_STATE_MIGRATE;
        }
 }
 
-static void migrate_hrtimers(int cpu)
+static int migrate_hrtimers(int scpu)
 {
        struct hrtimer_cpu_base *old_base, *new_base;
-       int i;
+       int dcpu, i;
 
-       BUG_ON(cpu_online(cpu));
-       old_base = &per_cpu(hrtimer_bases, cpu);
+       BUG_ON(cpu_online(scpu));
+       old_base = &per_cpu(hrtimer_bases, scpu);
        new_base = &get_cpu_var(hrtimer_bases);
 
-       tick_cancel_sched_timer(cpu);
+       dcpu = smp_processor_id();
+
+       tick_cancel_sched_timer(scpu);
        /*
         * The caller is globally serialized and nobody else
         * takes two locks at once, deadlock is not possible.
@@ -1557,32 +1551,47 @@ static void migrate_hrtimers(int cpu)
 
        for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
                migrate_hrtimer_list(&old_base->clock_base[i],
-                                    &new_base->clock_base[i], cpu);
+                                    &new_base->clock_base[i]);
        }
 
        spin_unlock(&old_base->lock);
        spin_unlock_irq(&new_base->lock);
        put_cpu_var(hrtimer_bases);
+
+       return dcpu;
+}
+
+static void tickle_timers(void *arg)
+{
+       hrtimer_peek_ahead_timers();
 }
+
 #endif /* CONFIG_HOTPLUG_CPU */
 
 static int __cpuinit hrtimer_cpu_notify(struct notifier_block *self,
                                        unsigned long action, void *hcpu)
 {
-       unsigned int cpu = (long)hcpu;
+       int dcpu = -1, scpu = (long)hcpu;
 
        switch (action) {
 
        case CPU_UP_PREPARE:
        case CPU_UP_PREPARE_FROZEN:
-               init_hrtimers_cpu(cpu);
+               init_hrtimers_cpu(scpu);
                break;
 
 #ifdef CONFIG_HOTPLUG_CPU
        case CPU_DEAD:
        case CPU_DEAD_FROZEN:
-               clockevents_notify(CLOCK_EVT_NOTIFY_CPU_DEAD, &cpu);
-               migrate_hrtimers(cpu);
+               clockevents_notify(CLOCK_EVT_NOTIFY_CPU_DEAD, &scpu);
+               dcpu = migrate_hrtimers(scpu);
+               break;
+
+       case CPU_POST_DEAD:
+               if (dcpu == -1)
+                       break;
+
+               smp_call_function_single(dcpu, tickle_timers, NULL, 0);
                break;
 #endif