[PATCH v4 4/4] workqueue: Unbind workers before sending them to exit()

[Valentin Schneider]

It has been reported that isolated CPUs can suffer from interference due to
per-CPU kworkers waking up just to die.

A surge of workqueue activity during initial setup of a latency-sensitive
application (refresh_vm_stats() being one of the culprits) can cause extra
per-CPU kworkers to be spawned. Then, said latency-sensitive task can be
running merrily on an isolated CPU only to be interrupted sometime later by
a kworker marked for death (cf. IDLE_WORKER_TIMEOUT, 5 minutes after last
kworker activity).

Prevent this by affining kworkers to the wq_unbound_cpumask (which doesn't
contain isolated CPUs, cf. HK_TYPE_WQ) before waking them up after marking
them with WORKER_DIE.

Changing the affinity does require a sleepable context, leverage the newly
introduced pool->idle_reaper_work to get that.
Remove dying workers from pool->workers and keep track of them in a
separate list. This intentionally prevents for_each_loop_worker() from
iterating over workers that are marked for death.

Signed-off-by: Valentin Schneider <vschneid@xxxxxxxxxx>
---
 kernel/workqueue.c | 80 ++++++++++++++++++++++++++++++++++++++++------
 1 file changed, 70 insertions(+), 10 deletions(-)

diff --git a/kernel/workqueue.c b/kernel/workqueue.c
index 436b1dbdf9ff..714db7df7105 100644
--- a/kernel/workqueue.c
+++ b/kernel/workqueue.c
@@ -177,6 +177,7 @@ struct worker_pool {
 
 	struct worker		*manager;	/* L: purely informational */
 	struct list_head	workers;	/* A: attached workers */
+	struct list_head        dying_workers;  /* A: workers about to die */
 	struct completion	*detach_completion; /* all workers detached */
 
 	struct ida		worker_ida;	/* worker IDs for task name */
@@ -1902,7 +1903,7 @@ static void worker_detach_from_pool(struct worker *worker)
 	list_del(&worker->node);
 	worker->pool = NULL;
 
-	if (list_empty(&pool->workers))
+	if (list_empty(&pool->workers) && list_empty(&pool->dying_workers))
 		detach_completion = pool->detach_completion;
 	mutex_unlock(&wq_pool_attach_mutex);
 
@@ -1991,9 +1992,31 @@ static void rebind_worker(struct worker *worker, struct worker_pool *pool)
 	WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask) < 0);
 }
 
+static void reap_workers(struct list_head *reaplist)
+{
+	struct worker *worker, *tmp;
+
+	list_for_each_entry_safe(worker, tmp, reaplist, entry) {
+		list_del_init(&worker->entry);
+		unbind_worker(worker);
+		/*
+		 * If the worker was somehow already running, then it had to be
+		 * in pool->idle_list when destroy_worker() happened or we
+		 * wouldn't have gotten here.
+		 *
+		 * Thus, the worker must either have observed the WORKER_DIE
+		 * flag, or have set its state to TASK_IDLE. Either way, the
+		 * below will be observed by the worker and is safe to do
+		 * outside of pool->lock.
+		 */
+		wake_up_process(worker->task);
+	}
+}
+
 /**
  * destroy_worker - destroy a workqueue worker
  * @worker: worker to be destroyed
+ * @list: transfer worker away from its pool->idle_list and into list
  *
  * Destroy @worker and adjust @pool stats accordingly.  The worker should
  * be idle.
@@ -2001,11 +2024,12 @@ static void rebind_worker(struct worker *worker, struct worker_pool *pool)
  * CONTEXT:
  * raw_spin_lock_irq(pool->lock).
  */
-static void destroy_worker(struct worker *worker)
+static void destroy_worker(struct worker *worker, struct list_head *list)
 {
 	struct worker_pool *pool = worker->pool;
 
 	lockdep_assert_held(&pool->lock);
+	lockdep_assert_held(&wq_pool_attach_mutex);
 
 	/* sanity check frenzy */
 	if (WARN_ON(worker->current_work) ||
@@ -2016,21 +2040,50 @@ static void destroy_worker(struct worker *worker)
 	pool->nr_workers--;
 	pool->nr_idle--;
 
-	list_del_init(&worker->entry);
 	worker->flags |= WORKER_DIE;
-	wake_up_process(worker->task);
+
+	list_move(&worker->entry, list);
+	list_move(&worker->node, &pool->dying_workers);
 }
 
 /*
  * idle_reaper_fn - reap workers that have been idle for too long.
  *
+ * Unbinding marked-for-destruction workers requires a sleepable context, as
+ * changing a task's affinity is not an atomic operation, and we don't want
+ * to disturb isolated CPUs IDLE_WORKER_TIMEOUT in the future just for a kworker
+ * to do_exit().
+ *
+ * Percpu kworkers should meet the conditions for the affinity change to not
+ * block (not migration-disabled and not running), but there is no *hard*
+ * guarantee that they are not running when we get here.
+ *
  * The delayed_work is only ever modified under raw_spin_lock_irq(pool->lock).
  */
 static void idle_reaper_fn(struct work_struct *work)
 {
 	struct delayed_work *dwork = to_delayed_work(work);
 	struct worker_pool *pool = container_of(dwork, struct worker_pool, idle_reaper_work);
+	struct list_head reaplist;
 
+	INIT_LIST_HEAD(&reaplist);
+
+	/*
+	 * Unlikely as it may be, a to-be-reaped worker could run after
+	 * idle_reaper_fn()::destroy_worker() has happened but before
+	 * idle_reaper_fn()::reap_workers() (consider a worker that stays
+	 * preempted after setting itself in the idle list, or before removing
+	 * itself from it).
+	 *
+	 * WORKER_DIE would be set in worker->flags, so it would be able to
+	 * kfree(worker) and head out to do_exit(), which wouldn't be nice to
+	 * the idle reaper.
+	 *
+	 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
+	 * won't go beyond worker_detach_from_pool() in its self-destruct path
+	 * (WORKER_DIE is set with wq_pool_attach_mutex set).
+	 */
+	mutex_lock(&wq_pool_attach_mutex);
 	raw_spin_lock_irq(&pool->lock);
 
 	while (too_many_workers(pool)) {
@@ -2055,10 +2108,11 @@ static void idle_reaper_fn(struct work_struct *work)
 			break;
 		}
 
-		destroy_worker(worker);
+		destroy_worker(worker, &reaplist);
 	}
-
 	raw_spin_unlock_irq(&pool->lock);
+	reap_workers(&reaplist);
+	mutex_unlock(&wq_pool_attach_mutex);
 }
 
 static void send_mayday(struct work_struct *work)
@@ -2422,12 +2476,12 @@ static int worker_thread(void *__worker)
 	/* am I supposed to die? */
 	if (unlikely(worker->flags & WORKER_DIE)) {
 		raw_spin_unlock_irq(&pool->lock);
-		WARN_ON_ONCE(!list_empty(&worker->entry));
 		set_pf_worker(false);
 
 		set_task_comm(worker->task, "kworker/dying");
 		ida_free(&pool->worker_ida, worker->id);
 		worker_detach_from_pool(worker);
+		WARN_ON_ONCE(!list_empty(&worker->entry));
 		kfree(worker);
 		return 0;
 	}
@@ -3500,6 +3554,7 @@ static int init_worker_pool(struct worker_pool *pool)
 	timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
 
 	INIT_LIST_HEAD(&pool->workers);
+	INIT_LIST_HEAD(&pool->dying_workers);
 
 	ida_init(&pool->worker_ida);
 	INIT_HLIST_NODE(&pool->hash_node);
@@ -3600,8 +3655,11 @@ static bool wq_manager_inactive(struct worker_pool *pool)
 static void put_unbound_pool(struct worker_pool *pool)
 {
 	DECLARE_COMPLETION_ONSTACK(detach_completion);
+	struct list_head reaplist;
 	struct worker *worker;
 
+	INIT_LIST_HEAD(&reaplist);
+
 	lockdep_assert_held(&wq_pool_mutex);
 
 	if (--pool->refcnt)
@@ -3624,17 +3682,19 @@ static void put_unbound_pool(struct worker_pool *pool)
 	 * Because of how wq_manager_inactive() works, we will hold the
 	 * spinlock after a successful wait.
 	 */
+	mutex_lock(&wq_pool_attach_mutex);
 	rcuwait_wait_event(&manager_wait, wq_manager_inactive(pool),
 			   TASK_UNINTERRUPTIBLE);
 	pool->flags |= POOL_MANAGER_ACTIVE;
 
 	while ((worker = first_idle_worker(pool)))
-		destroy_worker(worker);
+		destroy_worker(worker, &reaplist);
 	WARN_ON(pool->nr_workers || pool->nr_idle);
 	raw_spin_unlock_irq(&pool->lock);
 
-	mutex_lock(&wq_pool_attach_mutex);
-	if (!list_empty(&pool->workers))
+	reap_workers(&reaplist);
+
+	if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers))
 		pool->detach_completion = &detach_completion;
 	mutex_unlock(&wq_pool_attach_mutex);
 
-- 
2.31.1