cpu/rt: Rework cpu down for PREEMPT_RT

Bringing a CPU down is a pain with the PREEMPT_RT kernel because
tasks can be preempted in many more places than in non-RT. In
order to handle per_cpu variables, tasks may be pinned to a CPU
for a while, and even sleep. But these tasks need to be off the CPU
if that CPU is going down.

Several synchronization methods have been tried, but when stressed
they failed. This is a new approach.

A sync_tsk thread is still created and tasks may still block on a
lock when the CPU is going down, but how that works is a bit different.
When cpu_down() starts, it will create the sync_tsk and wait on it
to inform that current tasks that are pinned on the CPU are no longer
pinned. But new tasks that are about to be pinned will still be allowed
to do so at this time.

Then the notifiers are called. Several notifiers will bring down tasks
that will enter these locations. Some of these tasks will take locks
of other tasks that are on the CPU. If we don't let those other tasks
continue, but make them block until CPU down is done, the tasks that
the notifiers are waiting on will never complete as they are waiting
for the locks held by the tasks that are blocked.

Thus we still let the task pin the CPU until the notifiers are done.
After the notifiers run, we then make new tasks entering the pinned
CPU sections grab a mutex and wait. This mutex is now a per CPU mutex
in the hotplug_pcp descriptor.

To help things along, a new function in the scheduler code is created
called migrate_me(). This function will try to migrate the current task
off the CPU this is going down if possible. When the sync_tsk is created,
all tasks will then try to migrate off the CPU going down. There are
several cases that this wont work, but it helps in most cases.

After the notifiers are called and if a task can't migrate off but enters
the pin CPU sections, it will be forced to wait on the hotplug_pcp mutex
until the CPU down is complete. Then the scheduler will force the migration
anyway.

Also, I found that THREAD_BOUND need to also be accounted for in the
pinned CPU, and the migrate_disable no longer treats them special.
This helps fix issues with ksoftirqd and workqueue that unbind on CPU down.

Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>

3 files changed, 281 insertions, 44 deletions

diff --git a/include/linux/sched.h b/include/linux/sched.h
index 813d614d38dd71..2fa0caa557932d 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -2288,6 +2288,10 @@ extern void do_set_cpus_allowed(struct task_struct *p,
 
 extern int set_cpus_allowed_ptr(struct task_struct *p,
 				const struct cpumask *new_mask);
+int migrate_me(void);
+void tell_sched_cpu_down_begin(int cpu);
+void tell_sched_cpu_down_done(int cpu);
+
 #else
 static inline void do_set_cpus_allowed(struct task_struct *p,
 				      const struct cpumask *new_mask)
@@ -2300,6 +2304,9 @@ static inline int set_cpus_allowed_ptr(struct task_struct *p,
 		return -EINVAL;
 	return 0;
 }
+static inline int migrate_me(void) { return 0; }
+static inline void tell_sched_cpu_down_begin(int cpu) { }
+static inline void tell_sched_cpu_down_done(int cpu) { }
 #endif
 
 #ifdef CONFIG_NO_HZ_COMMON
diff --git a/kernel/cpu.c b/kernel/cpu.c
index 503f248f28291e..c48208b921601d 100644
--- a/kernel/cpu.c
+++ b/kernel/cpu.c
@@ -60,16 +60,10 @@ static int cpu_hotplug_disabled;
 
 static struct {
 	struct task_struct *active_writer;
-
 	/* wait queue to wake up the active_writer */
 	wait_queue_head_t wq;
-#ifdef CONFIG_PREEMPT_RT_FULL
-	/* Makes the lock keep the task's state */
-	spinlock_t lock;
-#else
 	/* verifies that no writer will get active while readers are active */
 	struct mutex lock;
-#endif
 	/*
 	 * Also blocks the new readers during
 	 * an ongoing cpu hotplug operation.
@@ -81,27 +75,13 @@ static struct {
 #endif
 } cpu_hotplug = {
 	.active_writer = NULL,
-	.wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
-#ifdef CONFIG_PREEMPT_RT_FULL
-	.lock = __SPIN_LOCK_UNLOCKED(cpu_hotplug.lock),
-#else
 	.lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
-#endif
+	.wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 	.dep_map = {.name = "cpu_hotplug.lock" },
 #endif
 };
 
-#ifdef CONFIG_PREEMPT_RT_FULL
-# define hotplug_lock()		rt_spin_lock(&cpu_hotplug.lock)
-# define hotplug_trylock()	rt_spin_trylock(&cpu_hotplug.lock)
-# define hotplug_unlock()	rt_spin_unlock(&cpu_hotplug.lock)
-#else
-# define hotplug_lock()		mutex_lock(&cpu_hotplug.lock)
-# define hotplug_trylock()	mutex_trylock(&cpu_hotplug.lock)
-# define hotplug_unlock()	mutex_unlock(&cpu_hotplug.lock)
-#endif
-
 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
 #define cpuhp_lock_acquire_tryread() \
@@ -109,12 +89,42 @@ static struct {
 #define cpuhp_lock_acquire()      lock_map_acquire(&cpu_hotplug.dep_map)
 #define cpuhp_lock_release()      lock_map_release(&cpu_hotplug.dep_map)
 
+/**
+ * hotplug_pcp	- per cpu hotplug descriptor
+ * @unplug:	set when pin_current_cpu() needs to sync tasks
+ * @sync_tsk:	the task that waits for tasks to finish pinned sections
+ * @refcount:	counter of tasks in pinned sections
+ * @grab_lock:	set when the tasks entering pinned sections should wait
+ * @synced:	notifier for @sync_tsk to tell cpu_down it's finished
+ * @mutex:	the mutex to make tasks wait (used when @grab_lock is true)
+ * @mutex_init:	zero if the mutex hasn't been initialized yet.
+ *
+ * Although @unplug and @sync_tsk may point to the same task, the @unplug
+ * is used as a flag and still exists after @sync_tsk has exited and
+ * @sync_tsk set to NULL.
+ */
 struct hotplug_pcp {
 	struct task_struct *unplug;
+	struct task_struct *sync_tsk;
 	int refcount;
+	int grab_lock;
 	struct completion synced;
+#ifdef CONFIG_PREEMPT_RT_FULL
+	spinlock_t lock;
+#else
+	struct mutex mutex;
+#endif
+	int mutex_init;
 };
 
+#ifdef CONFIG_PREEMPT_RT_FULL
+# define hotplug_lock(hp) rt_spin_lock(&(hp)->lock)
+# define hotplug_unlock(hp) rt_spin_unlock(&(hp)->lock)
+#else
+# define hotplug_lock(hp) mutex_lock(&(hp)->mutex)
+# define hotplug_unlock(hp) mutex_unlock(&(hp)->mutex)
+#endif
+
 static DEFINE_PER_CPU(struct hotplug_pcp, hotplug_pcp);
 
 /**
@@ -128,18 +138,39 @@ static DEFINE_PER_CPU(struct hotplug_pcp, hotplug_pcp);
 void pin_current_cpu(void)
 {
 	struct hotplug_pcp *hp;
+	int force = 0;
 
 retry:
 	hp = this_cpu_ptr(&hotplug_pcp);
 
-	if (!hp->unplug || hp->refcount || preempt_count() > 1 ||
+	if (!hp->unplug || hp->refcount || force || preempt_count() > 1 ||
 	    hp->unplug == current) {
 		hp->refcount++;
 		return;
 	}
-	preempt_enable();
-	hotplug_lock();
-	hotplug_unlock();
+	if (hp->grab_lock) {
+		preempt_enable();
+		hotplug_lock(hp);
+		hotplug_unlock(hp);
+	} else {
+		preempt_enable();
+		/*
+		 * Try to push this task off of this CPU.
+		 */
+		if (!migrate_me()) {
+			preempt_disable();
+			hp = this_cpu_ptr(&hotplug_pcp);
+			if (!hp->grab_lock) {
+				/*
+				 * Just let it continue it's already pinned
+				 * or about to sleep.
+				 */
+				force = 1;
+				goto retry;
+			}
+			preempt_enable();
+		}
+	}
 	preempt_disable();
 	goto retry;
 }
@@ -160,26 +191,84 @@ void unpin_current_cpu(void)
 		wake_up_process(hp->unplug);
 }
 
-/*
- * FIXME: Is this really correct under all circumstances ?
- */
+static void wait_for_pinned_cpus(struct hotplug_pcp *hp)
+{
+	set_current_state(TASK_UNINTERRUPTIBLE);
+	while (hp->refcount) {
+		schedule_preempt_disabled();
+		set_current_state(TASK_UNINTERRUPTIBLE);
+	}
+}
+
 static int sync_unplug_thread(void *data)
 {
 	struct hotplug_pcp *hp = data;
 
 	preempt_disable();
 	hp->unplug = current;
+	wait_for_pinned_cpus(hp);
+
+	/*
+	 * This thread will synchronize the cpu_down() with threads
+	 * that have pinned the CPU. When the pinned CPU count reaches
+	 * zero, we inform the cpu_down code to continue to the next step.
+	 */
 	set_current_state(TASK_UNINTERRUPTIBLE);
-	while (hp->refcount) {
-		schedule_preempt_disabled();
+	preempt_enable();
+	complete(&hp->synced);
+
+	/*
+	 * If all succeeds, the next step will need tasks to wait till
+	 * the CPU is offline before continuing. To do this, the grab_lock
+	 * is set and tasks going into pin_current_cpu() will block on the
+	 * mutex. But we still need to wait for those that are already in
+	 * pinned CPU sections. If the cpu_down() failed, the kthread_should_stop()
+	 * will kick this thread out.
+	 */
+	while (!hp->grab_lock && !kthread_should_stop()) {
+		schedule();
+		set_current_state(TASK_UNINTERRUPTIBLE);
+	}
+
+	/* Make sure grab_lock is seen before we see a stale completion */
+	smp_mb();
+
+	/*
+	 * Now just before cpu_down() enters stop machine, we need to make
+	 * sure all tasks that are in pinned CPU sections are out, and new
+	 * tasks will now grab the lock, keeping them from entering pinned
+	 * CPU sections.
+	 */
+	if (!kthread_should_stop()) {
+		preempt_disable();
+		wait_for_pinned_cpus(hp);
+		preempt_enable();
+		complete(&hp->synced);
+	}
+
+	set_current_state(TASK_UNINTERRUPTIBLE);
+	while (!kthread_should_stop()) {
+		schedule();
 		set_current_state(TASK_UNINTERRUPTIBLE);
 	}
 	set_current_state(TASK_RUNNING);
-	preempt_enable();
-	complete(&hp->synced);
+
+	/*
+	 * Force this thread off this CPU as it's going down and
+	 * we don't want any more work on this CPU.
+	 */
+	current->flags &= ~PF_NO_SETAFFINITY;
+	do_set_cpus_allowed(current, cpu_present_mask);
+	migrate_me();
 	return 0;
 }
 
+static void __cpu_unplug_sync(struct hotplug_pcp *hp)
+{
+	wake_up_process(hp->sync_tsk);
+	wait_for_completion(&hp->synced);
+}
+
 /*
  * Start the sync_unplug_thread on the target cpu and wait for it to
  * complete.
@@ -187,23 +276,83 @@ static int sync_unplug_thread(void *data)
 static int cpu_unplug_begin(unsigned int cpu)
 {
 	struct hotplug_pcp *hp = &per_cpu(hotplug_pcp, cpu);
-	struct task_struct *tsk;
+	int err;
+
+	/* Protected by cpu_hotplug.lock */
+	if (!hp->mutex_init) {
+#ifdef CONFIG_PREEMPT_RT_FULL
+		spin_lock_init(&hp->lock);
+#else
+		mutex_init(&hp->mutex);
+#endif
+		hp->mutex_init = 1;
+	}
+
+	/* Inform the scheduler to migrate tasks off this CPU */
+	tell_sched_cpu_down_begin(cpu);
 
 	init_completion(&hp->synced);
-	tsk = kthread_create(sync_unplug_thread, hp, "sync_unplug/%d", cpu);
-	if (IS_ERR(tsk))
-		return (PTR_ERR(tsk));
-	kthread_bind(tsk, cpu);
-	wake_up_process(tsk);
-	wait_for_completion(&hp->synced);
+
+	hp->sync_tsk = kthread_create(sync_unplug_thread, hp, "sync_unplug/%d", cpu);
+	if (IS_ERR(hp->sync_tsk)) {
+		err = PTR_ERR(hp->sync_tsk);
+		hp->sync_tsk = NULL;
+		return err;
+	}
+	kthread_bind(hp->sync_tsk, cpu);
+
+	/*
+	 * Wait for tasks to get out of the pinned sections,
+	 * it's still OK if new tasks enter. Some CPU notifiers will
+	 * wait for tasks that are going to enter these sections and
+	 * we must not have them block.
+	 */
+	__cpu_unplug_sync(hp);
+
 	return 0;
 }
 
+static void cpu_unplug_sync(unsigned int cpu)
+{
+	struct hotplug_pcp *hp = &per_cpu(hotplug_pcp, cpu);
+
+	init_completion(&hp->synced);
+	/* The completion needs to be initialzied before setting grab_lock */
+	smp_wmb();
+
+	/* Grab the mutex before setting grab_lock */
+	hotplug_lock(hp);
+	hp->grab_lock = 1;
+
+	/*
+	 * The CPU notifiers have been completed.
+	 * Wait for tasks to get out of pinned CPU sections and have new
+	 * tasks block until the CPU is completely down.
+	 */
+	__cpu_unplug_sync(hp);
+
+	/* All done with the sync thread */
+	kthread_stop(hp->sync_tsk);
+	hp->sync_tsk = NULL;
+}
+
 static void cpu_unplug_done(unsigned int cpu)
 {
 	struct hotplug_pcp *hp = &per_cpu(hotplug_pcp, cpu);
 
 	hp->unplug = NULL;
+	/* Let all tasks know cpu unplug is finished before cleaning up */
+	smp_wmb();
+
+	if (hp->sync_tsk)
+		kthread_stop(hp->sync_tsk);
+
+	if (hp->grab_lock) {
+		hotplug_unlock(hp);
+		/* protected by cpu_hotplug.lock */
+		hp->grab_lock = 0;
+	}
+	tell_sched_cpu_down_done(cpu);
 }
 
 void get_online_cpus(void)
@@ -212,9 +361,9 @@ void get_online_cpus(void)
 	if (cpu_hotplug.active_writer == current)
 		return;
 	cpuhp_lock_acquire_read();
-	hotplug_lock();
+	mutex_lock(&cpu_hotplug.lock);
 	atomic_inc(&cpu_hotplug.refcount);
-	hotplug_unlock();
+	mutex_unlock(&cpu_hotplug.lock);
 }
 EXPORT_SYMBOL_GPL(get_online_cpus);
 
@@ -267,11 +416,11 @@ void cpu_hotplug_begin(void)
 	cpuhp_lock_acquire();
 
 	for (;;) {
-		hotplug_lock();
+		mutex_lock(&cpu_hotplug.lock);
 		prepare_to_wait(&cpu_hotplug.wq, &wait, TASK_UNINTERRUPTIBLE);
 		if (likely(!atomic_read(&cpu_hotplug.refcount)))
 				break;
-		hotplug_unlock();
+		mutex_unlock(&cpu_hotplug.lock);
 		schedule();
 	}
 	finish_wait(&cpu_hotplug.wq, &wait);
@@ -280,7 +429,7 @@ void cpu_hotplug_begin(void)
 void cpu_hotplug_done(void)
 {
 	cpu_hotplug.active_writer = NULL;
-	hotplug_unlock();
+	mutex_unlock(&cpu_hotplug.lock);
 	cpuhp_lock_release();
 }
 
@@ -516,6 +665,9 @@ static int _cpu_down(unsigned int cpu, int tasks_frozen)
 
 	smpboot_park_threads(cpu);
 
+	/* Notifiers are done. Don't let any more tasks pin this CPU. */
+	cpu_unplug_sync(cpu);
+
 	/*
 	 * Prevent irq alloc/free while the dying cpu reorganizes the
 	 * interrupt affinities.
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 21cba30df6f693..50e5184f85efb9 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -1211,6 +1211,84 @@ void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
 		enqueue_task(rq, p, ENQUEUE_RESTORE);
 }
 
+static DEFINE_PER_CPU(struct cpumask, sched_cpumasks);
+static DEFINE_MUTEX(sched_down_mutex);
+static cpumask_t sched_down_cpumask;
+
+void tell_sched_cpu_down_begin(int cpu)
+{
+	mutex_lock(&sched_down_mutex);
+	cpumask_set_cpu(cpu, &sched_down_cpumask);
+	mutex_unlock(&sched_down_mutex);
+}
+
+void tell_sched_cpu_down_done(int cpu)
+{
+	mutex_lock(&sched_down_mutex);
+	cpumask_clear_cpu(cpu, &sched_down_cpumask);
+	mutex_unlock(&sched_down_mutex);
+}
+
+/**
+ * migrate_me - try to move the current task off this cpu
+ *
+ * Used by the pin_current_cpu() code to try to get tasks
+ * to move off the current CPU as it is going down.
+ * It will only move the task if the task isn't pinned to
+ * the CPU (with migrate_disable, affinity or NO_SETAFFINITY)
+ * and the task has to be in a RUNNING state. Otherwise the
+ * movement of the task will wake it up (change its state
+ * to running) when the task did not expect it.
+ *
+ * Returns 1 if it succeeded in moving the current task
+ *         0 otherwise.
+ */
+int migrate_me(void)
+{
+	struct task_struct *p = current;
+	struct migration_arg arg;
+	struct cpumask *cpumask;
+	struct cpumask *mask;
+	unsigned long flags;
+	unsigned int dest_cpu;
+	struct rq *rq;
+
+	/*
+	 * We can not migrate tasks bounded to a CPU or tasks not
+	 * running. The movement of the task will wake it up.
+	 */
+	if (p->flags & PF_NO_SETAFFINITY || p->state)
+		return 0;
+
+	mutex_lock(&sched_down_mutex);
+	rq = task_rq_lock(p, &flags);
+
+	cpumask = this_cpu_ptr(&sched_cpumasks);
+	mask = &p->cpus_allowed;
+
+	cpumask_andnot(cpumask, mask, &sched_down_cpumask);
+
+	if (!cpumask_weight(cpumask)) {
+		/* It's only on this CPU? */
+		task_rq_unlock(rq, p, &flags);
+		mutex_unlock(&sched_down_mutex);
+		return 0;
+	}
+
+	dest_cpu = cpumask_any_and(cpu_active_mask, cpumask);
+
+	arg.task = p;
+	arg.dest_cpu = dest_cpu;
+
+	task_rq_unlock(rq, p, &flags);
+
+	stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
+	tlb_migrate_finish(p->mm);
+	mutex_unlock(&sched_down_mutex);
+
+	return 1;
+}
+
 /*
  * Change a given task's CPU affinity. Migrate the thread to a
  * proper CPU and schedule it away if the CPU it's executing on