[PATCH 12/19] kthread: Default affine kthread to its preferred NUMA node

[Frederic Weisbecker <frederic@kernel.org>]

Kthreads attached to a preferred NUMA node for their task structure
allocation can also be assumed to run preferrably within that same node.

A more precise affinity is usually notified by calling
kthread_create_on_cpu() or kthread_bind[_mask]() before the first wakeup.

For the others, a default affinity to the node is desired and sometimes
implemented with more or less success when it comes to deal with hotplug
events and nohz_full / CPU Isolation interactions:

- kcompactd is affine to its node and handles hotplug but not CPU Isolation
- kswapd is affine to its node and ignores hotplug and CPU Isolation
- A bunch of drivers create their kthreads on a specific node and
  don't take care about affining further.

Handle that default node affinity preference at the generic level
instead, provided a kthread is created on an actual node and doesn't
apply any specific affinity such as a given CPU or a custom cpumask to
bind to before its first wake-up.

This generic handling is aware of CPU hotplug events and CPU isolation
such that:

* When a housekeeping CPU goes up that is part of the node of a given
  kthread, the related task is re-affined to that own node if it was
  previously running on the default last resort online housekeeping set
  from other nodes.

* When a housekeeping CPU goes down while it was part of the node of a
  kthread, the running task is migrated (or the sleeping task is woken
  up) automatically by the scheduler to other housekeepers within the
  same node or, as a last resort, to all housekeepers from other nodes.

Acked-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
---
 include/linux/cpuhotplug.h |   1 +
 kernel/kthread.c           | 106 ++++++++++++++++++++++++++++++++++++-
 2 files changed, 106 insertions(+), 1 deletion(-)

diff --git a/include/linux/cpuhotplug.h b/include/linux/cpuhotplug.h
index a04b73c40173..6cc5e484547c 100644
--- a/include/linux/cpuhotplug.h
+++ b/include/linux/cpuhotplug.h
@@ -240,6 +240,7 @@ enum cpuhp_state {
 	CPUHP_AP_WORKQUEUE_ONLINE,
 	CPUHP_AP_RANDOM_ONLINE,
 	CPUHP_AP_RCUTREE_ONLINE,
+	CPUHP_AP_KTHREADS_ONLINE,
 	CPUHP_AP_BASE_CACHEINFO_ONLINE,
 	CPUHP_AP_ONLINE_DYN,
 	CPUHP_AP_ONLINE_DYN_END		= CPUHP_AP_ONLINE_DYN + 40,
diff --git a/kernel/kthread.c b/kernel/kthread.c
index b6f9ce475a4f..3394ff024a5a 100644
--- a/kernel/kthread.c
+++ b/kernel/kthread.c
@@ -35,6 +35,9 @@ static DEFINE_SPINLOCK(kthread_create_lock);
 static LIST_HEAD(kthread_create_list);
 struct task_struct *kthreadd_task;
 
+static LIST_HEAD(kthreads_hotplug);
+static DEFINE_MUTEX(kthreads_hotplug_lock);
+
 struct kthread_create_info
 {
 	/* Information passed to kthread() from kthreadd. */
@@ -53,6 +56,7 @@ struct kthread_create_info
 struct kthread {
 	unsigned long flags;
 	unsigned int cpu;
+	unsigned int node;
 	int started;
 	int result;
 	int (*threadfn)(void *);
@@ -64,6 +68,8 @@ struct kthread {
 #endif
 	/* To store the full name if task comm is truncated. */
 	char *full_name;
+	struct task_struct *task;
+	struct list_head hotplug_node;
 };
 
 enum KTHREAD_BITS {
@@ -122,8 +128,11 @@ bool set_kthread_struct(struct task_struct *p)
 
 	init_completion(&kthread->exited);
 	init_completion(&kthread->parked);
+	INIT_LIST_HEAD(&kthread->hotplug_node);
 	p->vfork_done = &kthread->exited;
 
+	kthread->task = p;
+	kthread->node = tsk_fork_get_node(current);
 	p->worker_private = kthread;
 	return true;
 }
@@ -314,6 +323,11 @@ void __noreturn kthread_exit(long result)
 {
 	struct kthread *kthread = to_kthread(current);
 	kthread->result = result;
+	if (!list_empty(&kthread->hotplug_node)) {
+		mutex_lock(&kthreads_hotplug_lock);
+		list_del(&kthread->hotplug_node);
+		mutex_unlock(&kthreads_hotplug_lock);
+	}
 	do_exit(0);
 }
 EXPORT_SYMBOL(kthread_exit);
@@ -339,6 +353,48 @@ void __noreturn kthread_complete_and_exit(struct completion *comp, long code)
 }
 EXPORT_SYMBOL(kthread_complete_and_exit);
 
+static void kthread_fetch_affinity(struct kthread *kthread, struct cpumask *cpumask)
+{
+	cpumask_and(cpumask, cpumask_of_node(kthread->node),
+		    housekeeping_cpumask(HK_TYPE_KTHREAD));
+
+	if (cpumask_empty(cpumask))
+		cpumask_copy(cpumask, housekeeping_cpumask(HK_TYPE_KTHREAD));
+}
+
+static void kthread_affine_node(void)
+{
+	struct kthread *kthread = to_kthread(current);
+	cpumask_var_t affinity;
+
+	WARN_ON_ONCE(kthread_is_per_cpu(current));
+
+	if (kthread->node == NUMA_NO_NODE) {
+		housekeeping_affine(current, HK_TYPE_KTHREAD);
+	} else {
+		if (!zalloc_cpumask_var(&affinity, GFP_KERNEL)) {
+			WARN_ON_ONCE(1);
+			return;
+		}
+
+		mutex_lock(&kthreads_hotplug_lock);
+		WARN_ON_ONCE(!list_empty(&kthread->hotplug_node));
+		list_add_tail(&kthread->hotplug_node, &kthreads_hotplug);
+		/*
+		 * The node cpumask is racy when read from kthread() but:
+		 * - a racing CPU going down will either fail on the subsequent
+		 *   call to set_cpus_allowed_ptr() or be migrated to housekeepers
+		 *   afterwards by the scheduler.
+		 * - a racing CPU going up will be handled by kthreads_online_cpu()
+		 */
+		kthread_fetch_affinity(kthread, affinity);
+		set_cpus_allowed_ptr(current, affinity);
+		mutex_unlock(&kthreads_hotplug_lock);
+
+		free_cpumask_var(affinity);
+	}
+}
+
 static int kthread(void *_create)
 {
 	static const struct sched_param param = { .sched_priority = 0 };
@@ -369,7 +425,6 @@ static int kthread(void *_create)
 	 * back to default in case they have been changed.
 	 */
 	sched_setscheduler_nocheck(current, SCHED_NORMAL, &param);
-	set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_KTHREAD));
 
 	/* OK, tell user we're spawned, wait for stop or wakeup */
 	__set_current_state(TASK_UNINTERRUPTIBLE);
@@ -385,6 +440,9 @@ static int kthread(void *_create)
 
 	self->started = 1;
 
+	if (!(current->flags & PF_NO_SETAFFINITY))
+		kthread_affine_node();
+
 	ret = -EINTR;
 	if (!test_bit(KTHREAD_SHOULD_STOP, &self->flags)) {
 		cgroup_kthread_ready();
@@ -781,6 +839,52 @@ int kthreadd(void *unused)
 	return 0;
 }
 
+/*
+ * Re-affine kthreads according to their preferences
+ * and the newly online CPU. The CPU down part is handled
+ * by select_fallback_rq() which default re-affines to
+ * housekeepers in case the preferred affinity doesn't
+ * apply anymore.
+ */
+static int kthreads_online_cpu(unsigned int cpu)
+{
+	cpumask_var_t affinity;
+	struct kthread *k;
+	int ret;
+
+	guard(mutex)(&kthreads_hotplug_lock);
+
+	if (list_empty(&kthreads_hotplug))
+		return 0;
+
+	if (!zalloc_cpumask_var(&affinity, GFP_KERNEL))
+		return -ENOMEM;
+
+	ret = 0;
+
+	list_for_each_entry(k, &kthreads_hotplug, hotplug_node) {
+		if (WARN_ON_ONCE((k->task->flags & PF_NO_SETAFFINITY) ||
+				 kthread_is_per_cpu(k->task) ||
+				 k->node == NUMA_NO_NODE)) {
+			ret = -EINVAL;
+			continue;
+		}
+		kthread_fetch_affinity(k, affinity);
+		set_cpus_allowed_ptr(k->task, affinity);
+	}
+
+	free_cpumask_var(affinity);
+
+	return ret;
+}
+
+static int kthreads_init(void)
+{
+	return cpuhp_setup_state(CPUHP_AP_KTHREADS_ONLINE, "kthreads:online",
+				kthreads_online_cpu, NULL);
+}
+early_initcall(kthreads_init);
+
 void __kthread_init_worker(struct kthread_worker *worker,
 				const char *name,
 				struct lock_class_key *key)
-- 
2.46.0