sched_ext: Avoid live-locking bypass mode switching

A poorly behaving BPF scheduler can live-lock the system by e.g. incessantly
banging on the same DSQ on a large NUMA system to the point where switching
to the bypass mode can take a long time. Turning on the bypass mode requires
dequeueing and re-enqueueing currently runnable tasks, if the DSQs that they
are on are live-locked, this can take tens of seconds cascading into other
failures. This was observed on 2 x Intel Sapphire Rapids machines with 224
logical CPUs.

Inject artifical delays while the bypass mode is switching to guarantee
timely completion.

While at it, move __scx_ops_bypass_lock into scx_ops_bypass() and rename it
to bypass_lock.

Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Valentin Andrei <vandrei@meta.com>
Reported-by: Patrick Lu <patlu@meta.com>

1 files changed, 52 insertions, 3 deletions

diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
index c074a64c20f0..2d41f1917464 100644
--- a/kernel/sched/ext.c
+++ b/kernel/sched/ext.c
@@ -867,8 +867,8 @@ static DEFINE_MUTEX(scx_ops_enable_mutex);
 DEFINE_STATIC_KEY_FALSE(__scx_ops_enabled);
 DEFINE_STATIC_PERCPU_RWSEM(scx_fork_rwsem);
 static atomic_t scx_ops_enable_state_var = ATOMIC_INIT(SCX_OPS_DISABLED);
+static atomic_t scx_ops_breather_depth = ATOMIC_INIT(0);
 static int scx_ops_bypass_depth;
-static DEFINE_RAW_SPINLOCK(__scx_ops_bypass_lock);
 static bool scx_ops_init_task_enabled;
 static bool scx_switching_all;
 DEFINE_STATIC_KEY_FALSE(__scx_switched_all);
@@ -2474,11 +2474,48 @@ static struct rq *move_task_between_dsqs(struct task_struct *p, u64 enq_flags,
 	return dst_rq;
 }
 
+/*
+ * A poorly behaving BPF scheduler can live-lock the system by e.g. incessantly
+ * banging on the same DSQ on a large NUMA system to the point where switching
+ * to the bypass mode can take a long time. Inject artifical delays while the
+ * bypass mode is switching to guarantee timely completion.
+ */
+static void scx_ops_breather(struct rq *rq)
+{
+	u64 until;
+
+	lockdep_assert_rq_held(rq);
+
+	if (likely(!atomic_read(&scx_ops_breather_depth)))
+		return;
+
+	raw_spin_rq_unlock(rq);
+
+	until = ktime_get_ns() + NSEC_PER_MSEC;
+
+	do {
+		int cnt = 1024;
+		while (atomic_read(&scx_ops_breather_depth) && --cnt)
+			cpu_relax();
+	} while (atomic_read(&scx_ops_breather_depth) &&
+		 time_before64(ktime_get_ns(), until));
+
+	raw_spin_rq_lock(rq);
+}
+
 static bool consume_dispatch_q(struct rq *rq, struct scx_dispatch_q *dsq)
 {
 	struct task_struct *p;
 retry:
 	/*
+	 * This retry loop can repeatedly race against scx_ops_bypass()
+	 * dequeueing tasks from @dsq trying to put the system into the bypass
+	 * mode. On some multi-socket machines (e.g. 2x Intel 8480c), this can
+	 * live-lock the machine into soft lockups. Give a breather.
+	 */
+	scx_ops_breather(rq);
+
+	/*
 	 * The caller can't expect to successfully consume a task if the task's
 	 * addition to @dsq isn't guaranteed to be visible somehow. Test
 	 * @dsq->list without locking and skip if it seems empty.
@@ -4609,10 +4646,11 @@ bool task_should_scx(struct task_struct *p)
  */
 static void scx_ops_bypass(bool bypass)
 {
+	static DEFINE_RAW_SPINLOCK(bypass_lock);
 	int cpu;
 	unsigned long flags;
 
-	raw_spin_lock_irqsave(&__scx_ops_bypass_lock, flags);
+	raw_spin_lock_irqsave(&bypass_lock, flags);
 	if (bypass) {
 		scx_ops_bypass_depth++;
 		WARN_ON_ONCE(scx_ops_bypass_depth <= 0);
@@ -4625,6 +4663,8 @@ static void scx_ops_bypass(bool bypass)
 			goto unlock;
 	}
 
+	atomic_inc(&scx_ops_breather_depth);
+
 	/*
 	 * No task property is changing. We just need to make sure all currently
 	 * queued tasks are re-queued according to the new scx_rq_bypassing()
@@ -4680,8 +4720,10 @@ static void scx_ops_bypass(bool bypass)
 		/* resched to restore ticks and idle state */
 		resched_cpu(cpu);
 	}
+
+	atomic_dec(&scx_ops_breather_depth);
 unlock:
-	raw_spin_unlock_irqrestore(&__scx_ops_bypass_lock, flags);
+	raw_spin_unlock_irqrestore(&bypass_lock, flags);
 }
 
 static void free_exit_info(struct scx_exit_info *ei)
@@ -6334,6 +6376,13 @@ static bool scx_dispatch_from_dsq(struct bpf_iter_scx_dsq_kern *kit,
 		raw_spin_rq_lock(src_rq);
 	}
 
+	/*
+	 * If the BPF scheduler keeps calling this function repeatedly, it can
+	 * cause similar live-lock conditions as consume_dispatch_q(). Insert a
+	 * breather if necessary.
+	 */
+	scx_ops_breather(src_rq);
+
 	locked_rq = src_rq;
 	raw_spin_lock(&src_dsq->lock);