1 files changed, 15 insertions, 23 deletions

diff --git a/drivers/net/wireguard/queueing.h b/drivers/net/wireguard/queueing.h
index 583adb37ee1e..79b6d70de236 100644
--- a/drivers/net/wireguard/queueing.h
+++ b/drivers/net/wireguard/queueing.h
@@ -104,33 +104,25 @@ static inline void wg_reset_packet(struct sk_buff *skb, bool encapsulating)
 
 static inline int wg_cpumask_choose_online(int *stored_cpu, unsigned int id)
 {
-	unsigned int cpu = *stored_cpu, cpu_index, i;
+	unsigned int cpu = *stored_cpu;
+
+	while (unlikely(cpu >= nr_cpu_ids || !cpu_online(cpu)))
+		cpu = *stored_cpu = cpumask_nth(id % num_online_cpus(), cpu_online_mask);
 
-	if (unlikely(cpu == nr_cpumask_bits ||
-		     !cpumask_test_cpu(cpu, cpu_online_mask))) {
-		cpu_index = id % cpumask_weight(cpu_online_mask);
-		cpu = cpumask_first(cpu_online_mask);
-		for (i = 0; i < cpu_index; ++i)
-			cpu = cpumask_next(cpu, cpu_online_mask);
-		*stored_cpu = cpu;
-	}
 	return cpu;
 }
 
-/* This function is racy, in the sense that next is unlocked, so it could return
- * the same CPU twice. A race-free version of this would be to instead store an
- * atomic sequence number, do an increment-and-return, and then iterate through
- * every possible CPU until we get to that index -- choose_cpu. However that's
- * a bit slower, and it doesn't seem like this potential race actually
- * introduces any performance loss, so we live with it.
+/* This function is racy, in the sense that it's called while last_cpu is
+ * unlocked, so it could return the same CPU twice. Adding locking or using
+ * atomic sequence numbers is slower though, and the consequences of racing are
+ * harmless, so live with it.
  */
-static inline int wg_cpumask_next_online(int *next)
+static inline int wg_cpumask_next_online(int *last_cpu)
 {
-	int cpu = *next;
-
-	while (unlikely(!cpumask_test_cpu(cpu, cpu_online_mask)))
-		cpu = cpumask_next(cpu, cpu_online_mask) % nr_cpumask_bits;
-	*next = cpumask_next(cpu, cpu_online_mask) % nr_cpumask_bits;
+	int cpu = cpumask_next(READ_ONCE(*last_cpu), cpu_online_mask);
+	if (cpu >= nr_cpu_ids)
+		cpu = cpumask_first(cpu_online_mask);
+	WRITE_ONCE(*last_cpu, cpu);
 	return cpu;
 }
 
@@ -159,7 +151,7 @@ static inline void wg_prev_queue_drop_peeked(struct prev_queue *queue)
 
 static inline int wg_queue_enqueue_per_device_and_peer(
 	struct crypt_queue *device_queue, struct prev_queue *peer_queue,
-	struct sk_buff *skb, struct workqueue_struct *wq, int *next_cpu)
+	struct sk_buff *skb, struct workqueue_struct *wq)
 {
 	int cpu;
 
@@ -173,7 +165,7 @@ static inline int wg_queue_enqueue_per_device_and_peer(
 	/* Then we queue it up in the device queue, which consumes the
 	 * packet as soon as it can.
 	 */
-	cpu = wg_cpumask_next_online(next_cpu);
+	cpu = wg_cpumask_next_online(&device_queue->last_cpu);
 	if (unlikely(ptr_ring_produce_bh(&device_queue->ring, skb)))
 		return -EPIPE;
 	queue_work_on(cpu, wq, &per_cpu_ptr(device_queue->worker, cpu)->work);