net/packet: remove po->xmit

Use PACKET_SOCK_QDISC_BYPASS atomic bit instead of a pointer.

This removes one indirect call in fast path,
and READ_ONCE()/WRITE_ONCE() annotations as well.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Suggested-by: Willem de Bruijn <willemb@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/packet/af_packet.c

@@ -270,8 +270,11 @@ static noinline struct sk_buff *nf_hook_direct_egress(struct sk_buff *skb)
 }
 #endif
 
-static int packet_direct_xmit(struct sk_buff *skb)
+static int packet_xmit(const struct packet_sock *po, struct sk_buff *skb)
 {
+	if (!packet_sock_flag(po, PACKET_SOCK_QDISC_BYPASS))
+		return dev_queue_xmit(skb);
+
 #ifdef CONFIG_NETFILTER_EGRESS
 	if (nf_hook_egress_active()) {
 		skb = nf_hook_direct_egress(skb);
@@ -305,12 +308,6 @@ static void packet_cached_dev_reset(struct packet_sock *po)
 	RCU_INIT_POINTER(po->cached_dev, NULL);
 }
 
-static bool packet_use_direct_xmit(const struct packet_sock *po)
-{
-	/* Paired with WRITE_ONCE() in packet_setsockopt() */
-	return READ_ONCE(po->xmit) == packet_direct_xmit;
-}
-
 static u16 packet_pick_tx_queue(struct sk_buff *skb)
 {
 	struct net_device *dev = skb->dev;
@@ -2872,8 +2869,7 @@ tpacket_error:
 		packet_inc_pending(&po->tx_ring);
 
 		status = TP_STATUS_SEND_REQUEST;
-		/* Paired with WRITE_ONCE() in packet_setsockopt() */
+		err = packet_xmit(po, skb);
-		err = READ_ONCE(po->xmit)(skb);
 		if (unlikely(err != 0)) {
 			if (err > 0)
 				err = net_xmit_errno(err);
@@ -3076,8 +3072,8 @@ static int packet_snd(struct socket *sock, struct msghdr *msg, size_t len)
 		virtio_net_hdr_set_proto(skb, &vnet_hdr);
 	}
 
-	/* Paired with WRITE_ONCE() in packet_setsockopt() */
+	err = packet_xmit(po, skb);
-	err = READ_ONCE(po->xmit)(skb);
+
 	if (unlikely(err != 0)) {
 		if (err > 0)
 			err = net_xmit_errno(err);
@@ -3359,7 +3355,6 @@ static int packet_create(struct net *net, struct socket *sock, int protocol,
 	init_completion(&po->skb_completion);
 	sk->sk_family = PF_PACKET;
 	po->num = proto;
-	po->xmit = dev_queue_xmit;
 
 	err = packet_alloc_pending(po);
 	if (err)
@@ -4010,8 +4005,7 @@ packet_setsockopt(struct socket *sock, int level, int optname, sockptr_t optval,
 		if (copy_from_sockptr(&val, optval, sizeof(val)))
 			return -EFAULT;
 
-		/* Paired with all lockless reads of po->xmit */
+		packet_sock_flag_set(po, PACKET_SOCK_QDISC_BYPASS, val);
-		WRITE_ONCE(po->xmit, val ? packet_direct_xmit : dev_queue_xmit);
 		return 0;
 	}
 	default:
@@ -4126,7 +4120,7 @@ static int packet_getsockopt(struct socket *sock, int level, int optname,
 		val = packet_sock_flag(po, PACKET_SOCK_TX_HAS_OFF);
 		break;
 	case PACKET_QDISC_BYPASS:
-		val = packet_use_direct_xmit(po);
+		val = packet_sock_flag(po, PACKET_SOCK_QDISC_BYPASS);
 		break;
 	default:
 		return -ENOPROTOOPT;

net/packet/internal.h

@@ -128,7 +128,6 @@ struct packet_sock {
 	unsigned int		tp_tstamp;
 	struct completion	skb_completion;
 	struct net_device __rcu	*cached_dev;
-	int			(*xmit)(struct sk_buff *skb);
 	struct packet_type	prot_hook ____cacheline_aligned_in_smp;
 	atomic_t		tp_drops ____cacheline_aligned_in_smp;
 };
@@ -143,6 +142,7 @@ enum packet_sock_flags {
 	PACKET_SOCK_HAS_VNET_HDR,
 	PACKET_SOCK_RUNNING,
 	PACKET_SOCK_PRESSURE,
+	PACKET_SOCK_QDISC_BYPASS,
 };
 
 static inline void packet_sock_flag_set(struct packet_sock *po,