hrtimers: Push pending hrtimers away from outgoing CPU earlier

2b8272ff4a70 ("cpu/hotplug: Prevent self deadlock on CPU hot-unplug")
solved the straight forward CPU hotplug deadlock vs. the scheduler
bandwidth timer. Yu discovered a more involved variant where a task which
has a bandwidth timer started on the outgoing CPU holds a lock and then
gets throttled. If the lock required by one of the CPU hotplug callbacks
the hotplug operation deadlocks because the unthrottling timer event is not
handled on the dying CPU and can only be recovered once the control CPU
reaches the hotplug state which pulls the pending hrtimers from the dead
CPU.

Solve this by pushing the hrtimers away from the dying CPU in the dying
callbacks. Nothing can queue a hrtimer on the dying CPU at that point because
all other CPUs spin in stop_machine() with interrupts disabled and once the
operation is finished the CPU is marked offline.

Reported-by: Yu Liao <liaoyu15@huawei.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Liu Tie <liutie4@huawei.com>
Link: https://lore.kernel.org/r/87a5rphara.ffs@tglx

2 files changed, 19 insertions, 22 deletions

diff --git a/kernel/cpu.c b/kernel/cpu.c
index 6de7c6bb74ee..2e69a1deaa31 100644
--- a/kernel/cpu.c
+++ b/kernel/cpu.c
@@ -2098,7 +2098,7 @@ static struct cpuhp_step cpuhp_hp_states[] = {
 	[CPUHP_HRTIMERS_PREPARE] = {
 		.name			= "hrtimers:prepare",
 		.startup.single		= hrtimers_prepare_cpu,
-		.teardown.single	= hrtimers_dead_cpu,
+		.teardown.single	= NULL,
 	},
 	[CPUHP_SMPCFD_PREPARE] = {
 		.name			= "smpcfd:prepare",
@@ -2190,6 +2190,12 @@ static struct cpuhp_step cpuhp_hp_states[] = {
 		.startup.single		= NULL,
 		.teardown.single	= smpcfd_dying_cpu,
 	},
+	[CPUHP_AP_HRTIMERS_DYING] = {
+		.name			= "hrtimers:dying",
+		.startup.single		= NULL,
+		.teardown.single	= hrtimers_cpu_dying,
+	},
+
 	/* Entry state on starting. Interrupts enabled from here on. Transient
 	 * state for synchronsization */
 	[CPUHP_AP_ONLINE] = {
diff --git a/kernel/time/hrtimer.c b/kernel/time/hrtimer.c
index 238262e4aba7..760793998cdd 100644
--- a/kernel/time/hrtimer.c
+++ b/kernel/time/hrtimer.c
@@ -2219,29 +2219,22 @@ static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
 	}
 }
 
-int hrtimers_dead_cpu(unsigned int scpu)
+int hrtimers_cpu_dying(unsigned int dying_cpu)
 {
 	struct hrtimer_cpu_base *old_base, *new_base;
-	int i;
+	int i, ncpu = cpumask_first(cpu_active_mask);
 
-	BUG_ON(cpu_online(scpu));
-	tick_cancel_sched_timer(scpu);
+	tick_cancel_sched_timer(dying_cpu);
+
+	old_base = this_cpu_ptr(&hrtimer_bases);
+	new_base = &per_cpu(hrtimer_bases, ncpu);
 
-	/*
-	 * this BH disable ensures that raise_softirq_irqoff() does
-	 * not wakeup ksoftirqd (and acquire the pi-lock) while
-	 * holding the cpu_base lock
-	 */
-	local_bh_disable();
-	local_irq_disable();
-	old_base = &per_cpu(hrtimer_bases, scpu);
-	new_base = this_cpu_ptr(&hrtimer_bases);
 	/*
 	 * The caller is globally serialized and nobody else
 	 * takes two locks at once, deadlock is not possible.
 	 */
-	raw_spin_lock(&new_base->lock);
-	raw_spin_lock_nested(&old_base->lock, SINGLE_DEPTH_NESTING);
+	raw_spin_lock(&old_base->lock);
+	raw_spin_lock_nested(&new_base->lock, SINGLE_DEPTH_NESTING);
 
 	for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
 		migrate_hrtimer_list(&old_base->clock_base[i],
@@ -2252,15 +2245,13 @@ int hrtimers_dead_cpu(unsigned int scpu)
 	 * The migration might have changed the first expiring softirq
 	 * timer on this CPU. Update it.
 	 */
-	hrtimer_update_softirq_timer(new_base, false);
+	__hrtimer_get_next_event(new_base, HRTIMER_ACTIVE_SOFT);
+	/* Tell the other CPU to retrigger the next event */
+	smp_call_function_single(ncpu, retrigger_next_event, NULL, 0);
 
-	raw_spin_unlock(&old_base->lock);
 	raw_spin_unlock(&new_base->lock);
+	raw_spin_unlock(&old_base->lock);
 
-	/* Check, if we got expired work to do */
-	__hrtimer_peek_ahead_timers();
-	local_irq_enable();
-	local_bh_enable();
 	return 0;
 }