diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c
index d5495d8..12f2f72 100644
--- a/net/packet/af_packet.c
+++ b/net/packet/af_packet.c
@@ -89,6 +89,7 @@
 #include <linux/errqueue.h>
 #include <linux/net_tstamp.h>
 #include <linux/reciprocal_div.h>
+#include <linux/percpu.h>
 #ifdef CONFIG_INET
 #include <net/inet_common.h>
 #endif
@@ -1168,6 +1169,47 @@
 	buff->head = buff->head != buff->frame_max ? buff->head+1 : 0;
 }
 
+static void packet_inc_pending(struct packet_ring_buffer *rb)
+{
+	this_cpu_inc(*rb->pending_refcnt);
+}
+
+static void packet_dec_pending(struct packet_ring_buffer *rb)
+{
+	this_cpu_dec(*rb->pending_refcnt);
+}
+
+static unsigned int packet_read_pending(const struct packet_ring_buffer *rb)
+{
+	unsigned int refcnt = 0;
+	int cpu;
+
+	/* We don't use pending refcount in rx_ring. */
+	if (rb->pending_refcnt == NULL)
+		return 0;
+
+	for_each_possible_cpu(cpu)
+		refcnt += *per_cpu_ptr(rb->pending_refcnt, cpu);
+
+	return refcnt;
+}
+
+static int packet_alloc_pending(struct packet_sock *po)
+{
+	po->rx_ring.pending_refcnt = NULL;
+
+	po->tx_ring.pending_refcnt = alloc_percpu(unsigned int);
+	if (unlikely(po->tx_ring.pending_refcnt == NULL))
+		return -ENOBUFS;
+
+	return 0;
+}
+
+static void packet_free_pending(struct packet_sock *po)
+{
+	free_percpu(po->tx_ring.pending_refcnt);
+}
+
 static bool packet_rcv_has_room(struct packet_sock *po, struct sk_buff *skb)
 {
 	struct sock *sk = &po->sk;
@@ -2014,8 +2056,7 @@
 		__u32 ts;
 
 		ph = skb_shinfo(skb)->destructor_arg;
-		BUG_ON(atomic_read(&po->tx_ring.pending) == 0);
-		atomic_dec(&po->tx_ring.pending);
+		packet_dec_pending(&po->tx_ring);
 
 		ts = __packet_set_timestamp(po, ph, skb);
 		__packet_set_status(po, ph, TP_STATUS_AVAILABLE | ts);
@@ -2236,7 +2277,7 @@
 		skb_set_queue_mapping(skb, packet_pick_tx_queue(dev));
 		skb->destructor = tpacket_destruct_skb;
 		__packet_set_status(po, ph, TP_STATUS_SENDING);
-		atomic_inc(&po->tx_ring.pending);
+		packet_inc_pending(&po->tx_ring);
 
 		status = TP_STATUS_SEND_REQUEST;
 		err = po->xmit(skb);
@@ -2256,8 +2297,14 @@
 		}
 		packet_increment_head(&po->tx_ring);
 		len_sum += tp_len;
-	} while (likely((ph != NULL) || (need_wait &&
-					 atomic_read(&po->tx_ring.pending))));
+	} while (likely((ph != NULL) ||
+		/* Note: packet_read_pending() might be slow if we have
+		 * to call it as it's per_cpu variable, but in fast-path
+		 * we already short-circuit the loop with the first
+		 * condition, and luckily don't have to go that path
+		 * anyway.
+		 */
+		 (need_wait && packet_read_pending(&po->tx_ring))));
 
 	err = len_sum;
 	goto out_put;
@@ -2556,6 +2603,7 @@
 	/* Purge queues */
 
 	skb_queue_purge(&sk->sk_receive_queue);
+	packet_free_pending(po);
 	sk_refcnt_debug_release(sk);
 
 	sock_put(sk);
@@ -2717,6 +2765,10 @@
 	po->num = proto;
 	po->xmit = dev_queue_xmit;
 
+	err = packet_alloc_pending(po);
+	if (err)
+		goto out2;
+
 	packet_cached_dev_reset(po);
 
 	sk->sk_destruct = packet_sock_destruct;
@@ -2749,6 +2801,8 @@
 	preempt_enable();
 
 	return 0;
+out2:
+	sk_free(sk);
 out:
 	return err;
 }
@@ -3676,7 +3730,7 @@
 	if (!closing) {
 		if (atomic_read(&po->mapped))
 			goto out;
-		if (atomic_read(&rb->pending))
+		if (packet_read_pending(rb))
 			goto out;
 	}
 
diff --git a/net/packet/diag.c b/net/packet/diag.c
index a9584a2..533ce4f 100644
--- a/net/packet/diag.c
+++ b/net/packet/diag.c
@@ -3,6 +3,7 @@
 #include <linux/net.h>
 #include <linux/netdevice.h>
 #include <linux/packet_diag.h>
+#include <linux/percpu.h>
 #include <net/net_namespace.h>
 #include <net/sock.h>
 
diff --git a/net/packet/internal.h b/net/packet/internal.h
index 0a87d7b..eb9580a 100644
--- a/net/packet/internal.h
+++ b/net/packet/internal.h
@@ -64,7 +64,7 @@
 	unsigned int		pg_vec_pages;
 	unsigned int		pg_vec_len;
 
-	atomic_t		pending;
+	unsigned int __percpu	*pending_refcnt;
 
 	struct tpacket_kbdq_core	prb_bdqc;
 };