[PATCH v4 19/19] sched: Comment affine_move_task()

From: Peter Zijlstra <peterz@infradead.org>
To: tglx@linutronix.de, mingo@kernel.org
Cc: linux-kernel@vger.kernel.org, bigeasy@linutronix.de,
	qais.yousef@arm.com, swood@redhat.com, peterz@infradead.org,
	valentin.schneider@arm.com, juri.lelli@redhat.com,
	vincent.guittot@linaro.org, dietmar.eggemann@arm.com,
	rostedt@goodmis.org, bsegall@google.com, mgorman@suse.de,
	bristot@redhat.com, vincent.donnefort@arm.com, tj@kernel.org,
	ouwen210@hotmail.com
Subject: [PATCH v4 19/19] sched: Comment affine_move_task()
Date: Fri, 23 Oct 2020 12:12:17 +0200	[thread overview]
Message-ID: <20201023102347.866412704@infradead.org> (raw)
In-Reply-To: 20201023101158.088940906@infradead.org

From: Valentin Schneider <valentin.schneider@arm.com>


Signed-off-by: Valentin Schneider <valentin.schneider@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20201013140116.26651-2-valentin.schneider@arm.com
---
 kernel/sched/core.c |   81 ++++++++++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 79 insertions(+), 2 deletions(-)

--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -2082,7 +2082,75 @@ void do_set_cpus_allowed(struct task_str
 }
 
 /*
- * This function is wildly self concurrent, consider at least 3 times.
+ * This function is wildly self concurrent; here be dragons.
+ *
+ *
+ * When given a valid mask, __set_cpus_allowed_ptr() must block until the
+ * designated task is enqueued on an allowed CPU. If that task is currently
+ * running, we have to kick it out using the CPU stopper.
+ *
+ * Migrate-Disable comes along and tramples all over our nice sandcastle.
+ * Consider:
+ *
+ *     Initial conditions: P0->cpus_mask = [0, 1]
+ *
+ *     P0@CPU0                  P1
+ *
+ *     migrate_disable();
+ *     <preempted>
+ *                              set_cpus_allowed_ptr(P0, [1]);
+ *
+ * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
+ * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
+ * This means we need the following scheme:
+ *
+ *     P0@CPU0                  P1
+ *
+ *     migrate_disable();
+ *     <preempted>
+ *                              set_cpus_allowed_ptr(P0, [1]);
+ *                                <blocks>
+ *     <resumes>
+ *     migrate_enable();
+ *       __set_cpus_allowed_ptr();
+ *       <wakes local stopper>
+ *                         `--> <woken on migration completion>
+ *
+ * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
+ * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
+ * task p are serialized by p->pi_lock, which we can leverage: the one that
+ * should come into effect at the end of the Migrate-Disable region is the last
+ * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
+ * but we still need to properly signal those waiting tasks at the appropriate
+ * moment.
+ *
+ * This is implemented using struct set_affinity_pending. The first
+ * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
+ * setup an instance of that struct and install it on the targeted task_struct.
+ * Any and all further callers will reuse that instance. Those then wait for
+ * a completion signaled at the tail of the CPU stopper callback (1), triggered
+ * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
+ *
+ *
+ * (1) In the cases covered above. There is one more where the completion is
+ * signaled within affine_move_task() itself: when a subsequent affinity request
+ * cancels the need for an active migration. Consider:
+ *
+ *     Initial conditions: P0->cpus_mask = [0, 1]
+ *
+ *     P0@CPU0            P1                             P2
+ *
+ *     migrate_disable();
+ *     <preempted>
+ *                        set_cpus_allowed_ptr(P0, [1]);
+ *                          <blocks>
+ *                                                       set_cpus_allowed_ptr(P0, [0, 1]);
+ *                                                         <signal completion>
+ *                          <awakes>
+ *
+ * Note that the above is safe vs a concurrent migrate_enable(), as any
+ * pending affinity completion is preceded an uninstallion of
+ * p->migration_pending done with p->pi_lock held.
  */
 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
 			    int dest_cpu, unsigned int flags)
@@ -2126,6 +2194,7 @@ static int affine_move_task(struct rq *r
 	if (!(flags & SCA_MIGRATE_ENABLE)) {
 		/* serialized by p->pi_lock */
 		if (!p->migration_pending) {
+			/* Install the request */
 			refcount_set(&my_pending.refs, 1);
 			init_completion(&my_pending.done);
 			p->migration_pending = &my_pending;
@@ -2169,7 +2238,11 @@ static int affine_move_task(struct rq *r
 	}
 
 	if (task_running(rq, p) || p->state == TASK_WAKING) {
-
+		/*
+		 * Lessen races (and headaches) by delegating
+		 * is_migration_disabled(p) checks to the stopper, which will
+		 * run on the same CPU as said p.
+		 */
 		task_rq_unlock(rq, p, rf);
 		stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
 
@@ -2194,6 +2267,10 @@ static int affine_move_task(struct rq *r
 	if (refcount_dec_and_test(&pending->refs))
 		wake_up_var(&pending->refs);
 
+	/*
+	 * Block the original owner of &pending until all subsequent callers
+	 * have seen the completion and decremented the refcount
+	 */
 	wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
 
 	return 0;


