Merge branch 'vhost_net-TX-batching'

Jason Wang says:

====================
vhost_net TX batching

This series tries to batch submitting packets to underlayer socket
through msg_control during sendmsg(). This is done by:

1) Doing userspace copy inside vhost_net
2) Build XDP buff
3) Batch at most 64 (VHOST_NET_BATCH) XDP buffs and submit them once
   through msg_control during sendmsg().
4) Underlayer sockets can use XDP buffs directly when XDP is enalbed,
   or build skb based on XDP buff.

For the packet that can not be built easily with XDP or for the case
that batch submission is hard (e.g sndbuf is limited). We will go for
the previous slow path, passing iov iterator to underlayer socket
through sendmsg() once per packet.

This can help to improve cache utilization and avoid lots of indirect
calls with sendmsg(). It can also co-operate with the batching support
of the underlayer sockets (e.g the case of XDP redirection through
maps).

Testpmd(txonly) in guest shows obvious improvements:

Test                /+pps%
XDP_DROP on TAP     /+44.8%
XDP_REDIRECT on TAP /+29%
macvtap (skb)       /+26%

Netperf TCP_STREAM TX from guest shows obvious improvements on small
packet:

    size/session/+thu%/+normalize%
       64/     1/   +2%/    0%
       64/     2/   +3%/   +1%
       64/     4/   +7%/   +5%
       64/     8/   +8%/   +6%
      256/     1/   +3%/    0%
      256/     2/  +10%/   +7%
      256/     4/  +26%/  +22%
      256/     8/  +27%/  +23%
      512/     1/   +3%/   +2%
      512/     2/  +19%/  +14%
      512/     4/  +43%/  +40%
      512/     8/  +45%/  +41%
     1024/     1/   +4%/    0%
     1024/     2/  +27%/  +21%
     1024/     4/  +38%/  +73%
     1024/     8/  +15%/  +24%
     2048/     1/  +10%/   +7%
     2048/     2/  +16%/  +12%
     2048/     4/    0%/   +2%
     2048/     8/    0%/   +2%
     4096/     1/  +36%/  +60%
     4096/     2/  -11%/  -26%
     4096/     4/    0%/  +14%
     4096/     8/    0%/   +4%
    16384/     1/   -1%/   +5%
    16384/     2/    0%/   +2%
    16384/     4/    0%/   -3%
    16384/     8/    0%/   +4%
    65535/     1/    0%/  +10%
    65535/     2/    0%/   +8%
    65535/     4/    0%/   +1%
    65535/     8/    0%/   +3%

Please review.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/tap.c

@@ -619,7 +619,7 @@ static inline struct sk_buff *tap_alloc_skb(struct sock *sk, size_t prepad,
 #define TAP_RESERVE HH_DATA_OFF(ETH_HLEN)
 
 /* Get packet from user space buffer */
-static ssize_t tap_get_user(struct tap_queue *q, struct msghdr *m,
+static ssize_t tap_get_user(struct tap_queue *q, void *msg_control,
 			    struct iov_iter *from, int noblock)
 {
 	int good_linear = SKB_MAX_HEAD(TAP_RESERVE);
@@ -663,7 +663,7 @@ static ssize_t tap_get_user(struct tap_queue *q, struct msghdr *m,
 	if (unlikely(len < ETH_HLEN))
 		goto err;
 
-	if (m && m->msg_control && sock_flag(&q->sk, SOCK_ZEROCOPY)) {
+	if (msg_control && sock_flag(&q->sk, SOCK_ZEROCOPY)) {
 		struct iov_iter i;
 
 		copylen = vnet_hdr.hdr_len ?
@@ -724,11 +724,11 @@ static ssize_t tap_get_user(struct tap_queue *q, struct msghdr *m,
 	tap = rcu_dereference(q->tap);
 	/* copy skb_ubuf_info for callback when skb has no error */
 	if (zerocopy) {
-		skb_shinfo(skb)->destructor_arg = m->msg_control;
+		skb_shinfo(skb)->destructor_arg = msg_control;
 		skb_shinfo(skb)->tx_flags |= SKBTX_DEV_ZEROCOPY;
 		skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
-	} else if (m && m->msg_control) {
-		struct ubuf_info *uarg = m->msg_control;
+	} else if (msg_control) {
+		struct ubuf_info *uarg = msg_control;
 		uarg->callback(uarg, false);
 	}
 
@@ -1146,11 +1146,87 @@ static const struct file_operations tap_fops = {
 #endif
 };
 
+static int tap_get_user_xdp(struct tap_queue *q, struct xdp_buff *xdp)
+{
+	struct tun_xdp_hdr *hdr = xdp->data_hard_start;
+	struct virtio_net_hdr *gso = &hdr->gso;
+	int buflen = hdr->buflen;
+	int vnet_hdr_len = 0;
+	struct tap_dev *tap;
+	struct sk_buff *skb;
+	int err, depth;
+
+	if (q->flags & IFF_VNET_HDR)
+		vnet_hdr_len = READ_ONCE(q->vnet_hdr_sz);
+
+	skb = build_skb(xdp->data_hard_start, buflen);
+	if (!skb) {
+		err = -ENOMEM;
+		goto err;
+	}
+
+	skb_reserve(skb, xdp->data - xdp->data_hard_start);
+	skb_put(skb, xdp->data_end - xdp->data);
+
+	skb_set_network_header(skb, ETH_HLEN);
+	skb_reset_mac_header(skb);
+	skb->protocol = eth_hdr(skb)->h_proto;
+
+	if (vnet_hdr_len) {
+		err = virtio_net_hdr_to_skb(skb, gso, tap_is_little_endian(q));
+		if (err)
+			goto err_kfree;
+	}
+
+	skb_probe_transport_header(skb, ETH_HLEN);
+
+	/* Move network header to the right position for VLAN tagged packets */
+	if ((skb->protocol == htons(ETH_P_8021Q) ||
+	     skb->protocol == htons(ETH_P_8021AD)) &&
+	    __vlan_get_protocol(skb, skb->protocol, &depth) != 0)
+		skb_set_network_header(skb, depth);
+
+	rcu_read_lock();
+	tap = rcu_dereference(q->tap);
+	if (tap) {
+		skb->dev = tap->dev;
+		dev_queue_xmit(skb);
+	} else {
+		kfree_skb(skb);
+	}
+	rcu_read_unlock();
+
+	return 0;
+
+err_kfree:
+	kfree_skb(skb);
+err:
+	rcu_read_lock();
+		tap = rcu_dereference(q->tap);
+	if (tap && tap->count_tx_dropped)
+		tap->count_tx_dropped(tap);
+	rcu_read_unlock();
+	return err;
+}
+
 static int tap_sendmsg(struct socket *sock, struct msghdr *m,
 		       size_t total_len)
 {
 	struct tap_queue *q = container_of(sock, struct tap_queue, sock);
-	return tap_get_user(q, m, &m->msg_iter, m->msg_flags & MSG_DONTWAIT);
+	struct tun_msg_ctl *ctl = m->msg_control;
+	struct xdp_buff *xdp;
+	int i;
+
+	if (ctl && (ctl->type == TUN_MSG_PTR)) {
+		for (i = 0; i < ctl->num; i++) {
+			xdp = &((struct xdp_buff *)ctl->ptr)[i];
+			tap_get_user_xdp(q, xdp);
+		}
+		return 0;
+	}
+
+	return tap_get_user(q, ctl ? ctl->ptr : NULL, &m->msg_iter,
+			    m->msg_flags & MSG_DONTWAIT);
 }
 
 static int tap_recvmsg(struct socket *sock, struct msghdr *m,

drivers/net/tun.c

@@ -113,7 +113,6 @@ do {								\
 } while (0)
 #endif
 
-#define TUN_HEADROOM 256
 #define TUN_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
 
 /* TUN device flags */
@@ -869,6 +868,9 @@ static int tun_attach(struct tun_struct *tun, struct file *file,
 		tun_napi_init(tun, tfile, napi);
 	}
 
+	if (rtnl_dereference(tun->xdp_prog))
+		sock_set_flag(&tfile->sk, SOCK_XDP);
+
 	tun_set_real_num_queues(tun);
 
 	/* device is allowed to go away first, so no need to hold extra
@@ -1241,13 +1243,29 @@ static int tun_xdp_set(struct net_device *dev, struct bpf_prog *prog,
 		       struct netlink_ext_ack *extack)
 {
 	struct tun_struct *tun = netdev_priv(dev);
+	struct tun_file *tfile;
 	struct bpf_prog *old_prog;
+	int i;
 
 	old_prog = rtnl_dereference(tun->xdp_prog);
 	rcu_assign_pointer(tun->xdp_prog, prog);
 	if (old_prog)
 		bpf_prog_put(old_prog);
 
+	for (i = 0; i < tun->numqueues; i++) {
+		tfile = rtnl_dereference(tun->tfiles[i]);
+		if (prog)
+			sock_set_flag(&tfile->sk, SOCK_XDP);
+		else
+			sock_reset_flag(&tfile->sk, SOCK_XDP);
+	}
+	list_for_each_entry(tfile, &tun->disabled, next) {
+		if (prog)
+			sock_set_flag(&tfile->sk, SOCK_XDP);
+		else
+			sock_reset_flag(&tfile->sk, SOCK_XDP);
+	}
+
 	return 0;
 }
 
@@ -1617,6 +1635,55 @@ static bool tun_can_build_skb(struct tun_struct *tun, struct tun_file *tfile,
 	return true;
 }
 
+static struct sk_buff *__tun_build_skb(struct page_frag *alloc_frag, char *buf,
+				       int buflen, int len, int pad)
+{
+	struct sk_buff *skb = build_skb(buf, buflen);
+
+	if (!skb)
+		return ERR_PTR(-ENOMEM);
+
+	skb_reserve(skb, pad);
+	skb_put(skb, len);
+
+	get_page(alloc_frag->page);
+	alloc_frag->offset += buflen;
+
+	return skb;
+}
+
+static int tun_xdp_act(struct tun_struct *tun, struct bpf_prog *xdp_prog,
+		       struct xdp_buff *xdp, u32 act)
+{
+	int err;
+
+	switch (act) {
+	case XDP_REDIRECT:
+		err = xdp_do_redirect(tun->dev, xdp, xdp_prog);
+		if (err)
+			return err;
+		break;
+	case XDP_TX:
+		err = tun_xdp_tx(tun->dev, xdp);
+		if (err < 0)
+			return err;
+		break;
+	case XDP_PASS:
+		break;
+	default:
+		bpf_warn_invalid_xdp_action(act);
+		/* fall through */
+	case XDP_ABORTED:
+		trace_xdp_exception(tun->dev, xdp_prog, act);
+		/* fall through */
+	case XDP_DROP:
+		this_cpu_inc(tun->pcpu_stats->rx_dropped);
+		break;
+	}
+
+	return act;
+}
+
 static struct sk_buff *tun_build_skb(struct tun_struct *tun,
 				     struct tun_file *tfile,
 				     struct iov_iter *from,
@@ -1624,18 +1691,17 @@ static struct sk_buff *tun_build_skb(struct tun_struct *tun,
 				     int len, int *skb_xdp)
 {
 	struct page_frag *alloc_frag = &current->task_frag;
-	struct sk_buff *skb;
 	struct bpf_prog *xdp_prog;
 	int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
-	unsigned int delta = 0;
 	char *buf;
 	size_t copied;
-	int err, pad = TUN_RX_PAD;
+	int pad = TUN_RX_PAD;
+	int err = 0;
 
 	rcu_read_lock();
 	xdp_prog = rcu_dereference(tun->xdp_prog);
 	if (xdp_prog)
-		pad += TUN_HEADROOM;
+		pad += XDP_PACKET_HEADROOM;
 	buflen += SKB_DATA_ALIGN(len + pad);
 	rcu_read_unlock();
 
@@ -1654,17 +1720,18 @@ static struct sk_buff *tun_build_skb(struct tun_struct *tun,
 	 * of xdp_prog above, this should be rare and for simplicity
 	 * we do XDP on skb in case the headroom is not enough.
 	 */
-	if (hdr->gso_type || !xdp_prog)
+	if (hdr->gso_type || !xdp_prog) {
 		*skb_xdp = 1;
-	else
-		*skb_xdp = 0;
+		return __tun_build_skb(alloc_frag, buf, buflen, len, pad);
+	}
+
+	*skb_xdp = 0;
 
 	local_bh_disable();
 	rcu_read_lock();
 	xdp_prog = rcu_dereference(tun->xdp_prog);
-	if (xdp_prog && !*skb_xdp) {
+	if (xdp_prog) {
 		struct xdp_buff xdp;
-		void *orig_data;
 		u32 act;
 
 		xdp.data_hard_start = buf;
@@ -1672,66 +1739,33 @@ static struct sk_buff *tun_build_skb(struct tun_struct *tun,
 		xdp_set_data_meta_invalid(&xdp);
 		xdp.data_end = xdp.data + len;
 		xdp.rxq = &tfile->xdp_rxq;
-		orig_data = xdp.data;
-		act = bpf_prog_run_xdp(xdp_prog, &xdp);
 
-		switch (act) {
-		case XDP_REDIRECT:
-			get_page(alloc_frag->page);
-			alloc_frag->offset += buflen;
-			err = xdp_do_redirect(tun->dev, &xdp, xdp_prog);
-			xdp_do_flush_map();
-			if (err)
-				goto err_redirect;
-			rcu_read_unlock();
-			local_bh_enable();
-			return NULL;
-		case XDP_TX:
+		act = bpf_prog_run_xdp(xdp_prog, &xdp);
+		if (act == XDP_REDIRECT || act == XDP_TX) {
 			get_page(alloc_frag->page);
 			alloc_frag->offset += buflen;
-			if (tun_xdp_tx(tun->dev, &xdp) < 0)
-				goto err_redirect;
-			rcu_read_unlock();
-			local_bh_enable();
-			return NULL;
-		case XDP_PASS:
-			delta = orig_data - xdp.data;
-			len = xdp.data_end - xdp.data;
-			break;
-		default:
-			bpf_warn_invalid_xdp_action(act);
-			/* fall through */
-		case XDP_ABORTED:
-			trace_xdp_exception(tun->dev, xdp_prog, act);
-			/* fall through */
-		case XDP_DROP:
-			goto err_xdp;
 		}
-	}
+		err = tun_xdp_act(tun, xdp_prog, &xdp, act);
+		if (err < 0)
+			goto err_xdp;
+		if (err == XDP_REDIRECT)
+			xdp_do_flush_map();
+		if (err != XDP_PASS)
+			goto out;
 
-	skb = build_skb(buf, buflen);
-	if (!skb) {
-		rcu_read_unlock();
-		local_bh_enable();
-		return ERR_PTR(-ENOMEM);
+		pad = xdp.data - xdp.data_hard_start;
+		len = xdp.data_end - xdp.data;
 	}
-
-	skb_reserve(skb, pad - delta);
-	skb_put(skb, len);
-	get_page(alloc_frag->page);
-	alloc_frag->offset += buflen;
-
 	rcu_read_unlock();
 	local_bh_enable();
 
-	return skb;
+	return __tun_build_skb(alloc_frag, buf, buflen, len, pad);
 
-err_redirect:
-	put_page(alloc_frag->page);
 err_xdp:
+	put_page(alloc_frag->page);
+out:
 	rcu_read_unlock();
 	local_bh_enable();
-	this_cpu_inc(tun->pcpu_stats->rx_dropped);
 	return NULL;
 }
 
@@ -2392,18 +2426,133 @@ static void tun_sock_write_space(struct sock *sk)
 	kill_fasync(&tfile->fasync, SIGIO, POLL_OUT);
 }
 
+static int tun_xdp_one(struct tun_struct *tun,
+		       struct tun_file *tfile,
+		       struct xdp_buff *xdp, int *flush)
+{
+	struct tun_xdp_hdr *hdr = xdp->data_hard_start;
+	struct virtio_net_hdr *gso = &hdr->gso;
+	struct tun_pcpu_stats *stats;
+	struct bpf_prog *xdp_prog;
+	struct sk_buff *skb = NULL;
+	u32 rxhash = 0, act;
+	int buflen = hdr->buflen;
+	int err = 0;
+	bool skb_xdp = false;
+
+	xdp_prog = rcu_dereference(tun->xdp_prog);
+	if (xdp_prog) {
+		if (gso->gso_type) {
+			skb_xdp = true;
+			goto build;
+		}
+		xdp_set_data_meta_invalid(xdp);
+		xdp->rxq = &tfile->xdp_rxq;
+
+		act = bpf_prog_run_xdp(xdp_prog, xdp);
+		err = tun_xdp_act(tun, xdp_prog, xdp, act);
+		if (err < 0) {
+			put_page(virt_to_head_page(xdp->data));
+			return err;
+		}
+
+		switch (err) {
+		case XDP_REDIRECT:
+			*flush = true;
+			/* fall through */
+		case XDP_TX:
+			return 0;
+		case XDP_PASS:
+			break;
+		default:
+			put_page(virt_to_head_page(xdp->data));
+			return 0;
+		}
+	}
+
+build:
+	skb = build_skb(xdp->data_hard_start, buflen);
+	if (!skb) {
+		err = -ENOMEM;
+		goto out;
+	}
+
+	skb_reserve(skb, xdp->data - xdp->data_hard_start);
+	skb_put(skb, xdp->data_end - xdp->data);
+
+	if (virtio_net_hdr_to_skb(skb, gso, tun_is_little_endian(tun))) {
+		this_cpu_inc(tun->pcpu_stats->rx_frame_errors);
+		kfree_skb(skb);
+		err = -EINVAL;
+		goto out;
+	}
+
+	skb->protocol = eth_type_trans(skb, tun->dev);
+	skb_reset_network_header(skb);
+	skb_probe_transport_header(skb, 0);
+
+	if (skb_xdp) {
+		err = do_xdp_generic(xdp_prog, skb);
+		if (err != XDP_PASS)
+			goto out;
+	}
+
+	if (!rcu_dereference(tun->steering_prog))
+		rxhash = __skb_get_hash_symmetric(skb);
+
+	netif_receive_skb(skb);
+
+	stats = get_cpu_ptr(tun->pcpu_stats);
+	u64_stats_update_begin(&stats->syncp);
+	stats->rx_packets++;
+	stats->rx_bytes += skb->len;
+	u64_stats_update_end(&stats->syncp);
+	put_cpu_ptr(stats);
+
+	if (rxhash)
+		tun_flow_update(tun, rxhash, tfile);
+
+out:
+	return err;
+}
+
 static int tun_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 {
-	int ret;
+	int ret, i;
 	struct tun_file *tfile = container_of(sock, struct tun_file, socket);
 	struct tun_struct *tun = tun_get(tfile);
+	struct tun_msg_ctl *ctl = m->msg_control;
+	struct xdp_buff *xdp;
 
 	if (!tun)
 		return -EBADFD;
 
-	ret = tun_get_user(tun, tfile, m->msg_control, &m->msg_iter,
+	if (ctl && (ctl->type == TUN_MSG_PTR)) {
+		int n = ctl->num;
+		int flush = 0;
+
+		local_bh_disable();
+		rcu_read_lock();
+
+		for (i = 0; i < n; i++) {
+			xdp = &((struct xdp_buff *)ctl->ptr)[i];
+			tun_xdp_one(tun, tfile, xdp, &flush);
+		}
+
+		if (flush)
+			xdp_do_flush_map();
+
+		rcu_read_unlock();
+		local_bh_enable();
+
+		ret = total_len;
+		goto out;
+	}
+
+	ret = tun_get_user(tun, tfile, ctl ? ctl->ptr : NULL, &m->msg_iter,
 			   m->msg_flags & MSG_DONTWAIT,
 			   m->msg_flags & MSG_MORE);
+out:
 	tun_put(tun);
 	return ret;
 }

drivers/vhost/net.c

@@ -116,6 +116,8 @@ struct vhost_net_virtqueue {
 	 * For RX, number of batched heads
 	 */
 	int done_idx;
+	/* Number of XDP frames batched */
+	int batched_xdp;
 	/* an array of userspace buffers info */
 	struct ubuf_info *ubuf_info;
 	/* Reference counting for outstanding ubufs.
@@ -123,6 +125,8 @@ struct vhost_net_virtqueue {
 	struct vhost_net_ubuf_ref *ubufs;
 	struct ptr_ring *rx_ring;
 	struct vhost_net_buf rxq;
+	/* Batched XDP buffs */
+	struct xdp_buff *xdp;
 };
 
 struct vhost_net {
@@ -338,6 +342,11 @@ static bool vhost_sock_zcopy(struct socket *sock)
 		sock_flag(sock->sk, SOCK_ZEROCOPY);
 }
 
+static bool vhost_sock_xdp(struct socket *sock)
+{
+	return sock_flag(sock->sk, SOCK_XDP);
+}
+
 /* In case of DMA done not in order in lower device driver for some reason.
  * upend_idx is used to track end of used idx, done_idx is used to track head
  * of used idx. Once lower device DMA done contiguously, we will signal KVM
@@ -444,10 +453,37 @@ static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq)
 	nvq->done_idx = 0;
 }
 
+static void vhost_tx_batch(struct vhost_net *net,
+			   struct vhost_net_virtqueue *nvq,
+			   struct socket *sock,
+			   struct msghdr *msghdr)
+{
+	struct tun_msg_ctl ctl = {
+		.type = TUN_MSG_PTR,
+		.num = nvq->batched_xdp,
+		.ptr = nvq->xdp,
+	};
+	int err;
+
+	if (nvq->batched_xdp == 0)
+		goto signal_used;
+
+	msghdr->msg_control = &ctl;
+	err = sock->ops->sendmsg(sock, msghdr, 0);
+	if (unlikely(err < 0)) {
+		vq_err(&nvq->vq, "Fail to batch sending packets\n");
+		return;
+	}
+
+signal_used:
+	vhost_net_signal_used(nvq);
+	nvq->batched_xdp = 0;
+}
+
 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 				    struct vhost_net_virtqueue *nvq,
 				    unsigned int *out_num, unsigned int *in_num,
-				    bool *busyloop_intr)
+				    struct msghdr *msghdr, bool *busyloop_intr)
 {
 	struct vhost_virtqueue *vq = &nvq->vq;
 	unsigned long uninitialized_var(endtime);
@@ -455,8 +491,9 @@ static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 				  out_num, in_num, NULL, NULL);
 
 	if (r == vq->num && vq->busyloop_timeout) {
+		/* Flush batched packets first */
 		if (!vhost_sock_zcopy(vq->private_data))
-			vhost_net_signal_used(nvq);
+			vhost_tx_batch(net, nvq, vq->private_data, msghdr);
 		preempt_disable();
 		endtime = busy_clock() + vq->busyloop_timeout;
 		while (vhost_can_busy_poll(endtime)) {
@@ -512,7 +549,7 @@ static int get_tx_bufs(struct vhost_net *net,
 	struct vhost_virtqueue *vq = &nvq->vq;
 	int ret;
 
-	ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, busyloop_intr);
+	ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg, busyloop_intr);
 
 	if (ret < 0 || ret == vq->num)
 		return ret;
@@ -540,6 +577,80 @@ static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
 	       !vhost_vq_avail_empty(vq->dev, vq);
 }
 
+#define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
+
+static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
+			       struct iov_iter *from)
+{
+	struct vhost_virtqueue *vq = &nvq->vq;
+	struct socket *sock = vq->private_data;
+	struct page_frag *alloc_frag = &current->task_frag;
+	struct virtio_net_hdr *gso;
+	struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
+	struct tun_xdp_hdr *hdr;
+	size_t len = iov_iter_count(from);
+	int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
+	int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
+	int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
+	int sock_hlen = nvq->sock_hlen;
+	void *buf;
+	int copied;
+
+	if (unlikely(len < nvq->sock_hlen))
+		return -EFAULT;
+
+	if (SKB_DATA_ALIGN(len + pad) +
+	    SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
+		return -ENOSPC;
+
+	buflen += SKB_DATA_ALIGN(len + pad);
+	alloc_frag->offset = ALIGN((u64)alloc_frag->offset, SMP_CACHE_BYTES);
+	if (unlikely(!skb_page_frag_refill(buflen, alloc_frag, GFP_KERNEL)))
+		return -ENOMEM;
+
+	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
+	copied = copy_page_from_iter(alloc_frag->page,
+				     alloc_frag->offset +
+				     offsetof(struct tun_xdp_hdr, gso),
+				     sock_hlen, from);
+	if (copied != sock_hlen)
+		return -EFAULT;
+
+	hdr = buf;
+	gso = &hdr->gso;
+
+	if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
+	    vhost16_to_cpu(vq, gso->csum_start) +
+	    vhost16_to_cpu(vq, gso->csum_offset) + 2 >
+	    vhost16_to_cpu(vq, gso->hdr_len)) {
+		gso->hdr_len = cpu_to_vhost16(vq,
+			       vhost16_to_cpu(vq, gso->csum_start) +
+			       vhost16_to_cpu(vq, gso->csum_offset) + 2);
+
+		if (vhost16_to_cpu(vq, gso->hdr_len) > len)
+			return -EINVAL;
+	}
+
+	len -= sock_hlen;
+	copied = copy_page_from_iter(alloc_frag->page,
+				     alloc_frag->offset + pad,
+				     len, from);
+	if (copied != len)
+		return -EFAULT;
+
+	xdp->data_hard_start = buf;
+	xdp->data = buf + pad;
+	xdp->data_end = xdp->data + len;
+	hdr->buflen = buflen;
+
+	get_page(alloc_frag->page);
+	alloc_frag->offset += buflen;
+
+	++nvq->batched_xdp;
+
+	return 0;
+}
+
 static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
 {
 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
@@ -556,10 +667,14 @@ static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
 	size_t len, total_len = 0;
 	int err;
 	int sent_pkts = 0;
+	bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
 
 	for (;;) {
 		bool busyloop_intr = false;
 
+		if (nvq->done_idx == VHOST_NET_BATCH)
+			vhost_tx_batch(net, nvq, sock, &msg);
+
 		head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 				   &busyloop_intr);
 		/* On error, stop handling until the next kick. */
@@ -577,14 +692,34 @@ static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
 			break;
 		}
 
-		vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
-		vq->heads[nvq->done_idx].len = 0;
-
 		total_len += len;
-		if (tx_can_batch(vq, total_len))
-			msg.msg_flags |= MSG_MORE;
-		else
-			msg.msg_flags &= ~MSG_MORE;
+
+		/* For simplicity, TX batching is only enabled if
+		 * sndbuf is unlimited.
+		 */
+		if (sock_can_batch) {
+			err = vhost_net_build_xdp(nvq, &msg.msg_iter);
+			if (!err) {
+				goto done;
+			} else if (unlikely(err != -ENOSPC)) {
+				vhost_tx_batch(net, nvq, sock, &msg);
+				vhost_discard_vq_desc(vq, 1);
+				vhost_net_enable_vq(net, vq);
+				break;
+			}
+
+			/* We can't build XDP buff, go for single
+			 * packet path but let's flush batched
+			 * packets.
+			 */
+			vhost_tx_batch(net, nvq, sock, &msg);
+			msg.msg_control = NULL;
+		} else {
+			if (tx_can_batch(vq, total_len))
+				msg.msg_flags |= MSG_MORE;
+			else
+				msg.msg_flags &= ~MSG_MORE;
+		}
 
 		/* TODO: Check specific error and bomb out unless ENOBUFS? */
 		err = sock->ops->sendmsg(sock, &msg, len);
@@ -596,15 +731,17 @@ static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
 		if (err != len)
 			pr_debug("Truncated TX packet: len %d != %zd\n",
 				 err, len);
-		if (++nvq->done_idx >= VHOST_NET_BATCH)
-			vhost_net_signal_used(nvq);
+done:
+		vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
+		vq->heads[nvq->done_idx].len = 0;
+		++nvq->done_idx;
 		if (vhost_exceeds_weight(++sent_pkts, total_len)) {
 			vhost_poll_queue(&vq->poll);
 			break;
 		}
 	}
 
-	vhost_net_signal_used(nvq);
+	vhost_tx_batch(net, nvq, sock, &msg);
 }
 
 static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
@@ -620,6 +757,7 @@ static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
 		.msg_controllen = 0,
 		.msg_flags = MSG_DONTWAIT,
 	};
+	struct tun_msg_ctl ctl;
 	size_t len, total_len = 0;
 	int err;
 	struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
@@ -664,8 +802,10 @@ static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
 			ubuf->ctx = nvq->ubufs;
 			ubuf->desc = nvq->upend_idx;
 			refcount_set(&ubuf->refcnt, 1);
-			msg.msg_control = ubuf;
-			msg.msg_controllen = sizeof(ubuf);
+			msg.msg_control = &ctl;
+			ctl.type = TUN_MSG_UBUF;
+			ctl.ptr = ubuf;
+			msg.msg_controllen = sizeof(ctl);
 			ubufs = nvq->ubufs;
 			atomic_inc(&ubufs->refcount);
 			nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
@@ -1078,6 +1218,7 @@ static int vhost_net_open(struct inode *inode, struct file *f)
 	struct vhost_dev *dev;
 	struct vhost_virtqueue **vqs;
 	void **queue;
+	struct xdp_buff *xdp;
 	int i;
 
 	n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
@@ -1098,6 +1239,14 @@ static int vhost_net_open(struct inode *inode, struct file *f)
 	}
 	n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
 
+	xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL);
+	if (!xdp) {
+		kfree(vqs);
+		kvfree(n);
+		kfree(queue);
+	}
+	n->vqs[VHOST_NET_VQ_TX].xdp = xdp;
+
 	dev = &n->dev;
 	vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 	vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
@@ -1108,6 +1257,7 @@ static int vhost_net_open(struct inode *inode, struct file *f)
 		n->vqs[i].ubuf_info = NULL;
 		n->vqs[i].upend_idx = 0;
 		n->vqs[i].done_idx = 0;
+		n->vqs[i].batched_xdp = 0;
 		n->vqs[i].vhost_hlen = 0;
 		n->vqs[i].sock_hlen = 0;
 		n->vqs[i].rx_ring = NULL;
@@ -1191,6 +1341,7 @@ static int vhost_net_release(struct inode *inode, struct file *f)
 	 * since jobs can re-queue themselves. */
 	vhost_net_flush(n);
 	kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
+	kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
 	kfree(n->dev.vqs);
 	kvfree(n);
 	return 0;

include/linux/if_tun.h

@@ -16,9 +16,23 @@
 #define __IF_TUN_H
 
 #include <uapi/linux/if_tun.h>
+#include <uapi/linux/virtio_net.h>
 
 #define TUN_XDP_FLAG 0x1UL
 
+#define TUN_MSG_UBUF 1
+#define TUN_MSG_PTR  2
+struct tun_msg_ctl {
+	unsigned short type;
+	unsigned short num;
+	void *ptr;
+};
+
+struct tun_xdp_hdr {
+	int buflen;
+	struct virtio_net_hdr gso;
+};
+
 #if defined(CONFIG_TUN) || defined(CONFIG_TUN_MODULE)
 struct socket *tun_get_socket(struct file *);
 struct ptr_ring *tun_get_tx_ring(struct file *file);

include/net/sock.h

@@ -800,6 +800,7 @@ enum sock_flags {
 	SOCK_SELECT_ERR_QUEUE, /* Wake select on error queue */
 	SOCK_RCU_FREE, /* wait rcu grace period in sk_destruct() */
 	SOCK_TXTIME,
+	SOCK_XDP, /* XDP is attached */
 };
 
 #define SK_FLAGS_TIMESTAMP ((1UL << SOCK_TIMESTAMP) | (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE))