diff --git a/drivers/net/wireguard/queueing.c b/drivers/net/wireguard/queueing.c
index 8084e7408c0a..26d235d15235 100644
--- a/drivers/net/wireguard/queueing.c
+++ b/drivers/net/wireguard/queueing.c
@@ -28,6 +28,7 @@ int wg_packet_queue_init(struct crypt_queue *queue, work_func_t function,
 	int ret;
 
 	memset(queue, 0, sizeof(*queue));
+	queue->last_cpu = -1;
 	ret = ptr_ring_init(&queue->ring, len, GFP_KERNEL);
 	if (ret)
 		return ret;
diff --git a/drivers/net/wireguard/queueing.h b/drivers/net/wireguard/queueing.h
index 03850c43ebaf..4c81cc5bce59 100644
--- a/drivers/net/wireguard/queueing.h
+++ b/drivers/net/wireguard/queueing.h
@@ -122,20 +122,17 @@ static inline int wg_cpumask_choose_online(int *stored_cpu, unsigned int id)
 	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(*last_cpu, cpu_online_mask);
+	if (cpu >= nr_cpu_ids)
+		cpu = cpumask_first(cpu_online_mask);
+	*last_cpu = cpu;
 	return cpu;
 }
 
@@ -164,7 +161,7 @@ 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;
 
@@ -178,7 +175,7 @@ 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);
diff --git a/drivers/net/wireguard/receive.c b/drivers/net/wireguard/receive.c
index 214889edb48e..0955a2e735d9 100644
--- a/drivers/net/wireguard/receive.c
+++ b/drivers/net/wireguard/receive.c
@@ -540,7 +540,7 @@ static void wg_packet_consume_data(struct wg_device *wg, struct sk_buff *skb)
 		goto err;
 
 	ret = wg_queue_enqueue_per_device_and_peer(&wg->decrypt_queue, &peer->rx_queue, skb,
-						   wg->packet_crypt_wq, &wg->decrypt_queue.last_cpu);
+						   wg->packet_crypt_wq);
 	if (unlikely(ret == -EPIPE))
 		wg_queue_enqueue_per_peer_rx(skb, PACKET_STATE_DEAD);
 	if (likely(!ret || ret == -EPIPE)) {
diff --git a/drivers/net/wireguard/send.c b/drivers/net/wireguard/send.c
index 55bb0c9313d7..79e759ee4f2b 100644
--- a/drivers/net/wireguard/send.c
+++ b/drivers/net/wireguard/send.c
@@ -325,7 +325,7 @@ static void wg_packet_create_data(struct wg_peer *peer, struct sk_buff *first)
 		goto err;
 
 	ret = wg_queue_enqueue_per_device_and_peer(&wg->encrypt_queue, &peer->tx_queue, first,
-						   wg->packet_crypt_wq, &wg->encrypt_queue.last_cpu);
+						   wg->packet_crypt_wq);
 	if (unlikely(ret == -EPIPE))
 		wg_queue_enqueue_per_peer_tx(first, PACKET_STATE_DEAD);
 err: