High Performance UML Vector Network Driver

1. Provides infrastructure for vector IO using recvmmsg/sendmmsg.
    1.1. Multi-message read.
    1.2. Multi-message write.
    1.3. Optimized queue support for multi-packet enqueue/dequeue.
    1.4. BQL/DQL support.
2. Implements transports for several transports as well support
for direct wiring of PWEs to NIC. Allows direct connection of VMs
to host, other VMs and network devices with no switch in use.
    2.1. Raw socket >4 times higher PPS and 10 times higher tcp RX
    than existing pcap based transport (> 4Gbit)
    2.2. New tap transport using socket RX and tap xmit. Similar
    performance improvements (>4Gbit)
    2.3. GRE transport - direct wiring to GRE PWE
    2.4. L2TPv3 transport - direct wiring to L2TPv3 PWE
3. Tuning, performance and offload related setting support via ethtool.
4. Initial BPF support - used in tap/raw to avoid software looping
5. Scatter Gather support.
6. VNET and checksum offload support for raw socket transport.
7. TSO/GSO support where applicable or available
8. Migrates all error messages to netdevice_*() and rate limits
them where needed.

Signed-off-by: Anton Ivanov <anton.ivanov@cambridgegreys.com>
Signed-off-by: Richard Weinberger <richard@nod.at>

10 files changed, 2932 insertions, 3 deletions

diff --git a/arch/um/Kconfig.net b/arch/um/Kconfig.net
index e871af24d9cd..c390f3deb0dc 100644
--- a/arch/um/Kconfig.net
+++ b/arch/um/Kconfig.net
@@ -109,6 +109,17 @@ config UML_NET_DAEMON
         more than one without conflict.  If you don't need UML networking,
         say N.
 
+config UML_NET_VECTOR
+	bool "Vector I/O high performance network devices"
+	depends on UML_NET
+	help
+	This User-Mode Linux network driver uses multi-message send
+	and receive functions. The host running the UML guest must have
+	a linux kernel version above 3.0 and a libc version > 2.13.
+	This driver provides tap, raw, gre and l2tpv3 network transports
+	with up to 4 times higher network throughput than the UML network
+	drivers.
+
 config UML_NET_VDE
 	bool "VDE transport"
 	depends on UML_NET
diff --git a/arch/um/drivers/Makefile b/arch/um/drivers/Makefile
index e7582e1d248c..16b3cebddafb 100644
--- a/arch/um/drivers/Makefile
+++ b/arch/um/drivers/Makefile
@@ -9,6 +9,7 @@
 slip-objs := slip_kern.o slip_user.o
 slirp-objs := slirp_kern.o slirp_user.o
 daemon-objs := daemon_kern.o daemon_user.o
+vector-objs := vector_kern.o vector_user.o vector_transports.o
 umcast-objs := umcast_kern.o umcast_user.o
 net-objs := net_kern.o net_user.o
 mconsole-objs := mconsole_kern.o mconsole_user.o
@@ -43,6 +44,7 @@ obj-$(CONFIG_STDERR_CONSOLE) += stderr_console.o
 obj-$(CONFIG_UML_NET_SLIP) += slip.o slip_common.o
 obj-$(CONFIG_UML_NET_SLIRP) += slirp.o slip_common.o
 obj-$(CONFIG_UML_NET_DAEMON) += daemon.o 
+obj-$(CONFIG_UML_NET_VECTOR) += vector.o
 obj-$(CONFIG_UML_NET_VDE) += vde.o
 obj-$(CONFIG_UML_NET_MCAST) += umcast.o
 obj-$(CONFIG_UML_NET_PCAP) += pcap.o
@@ -61,7 +63,7 @@ obj-$(CONFIG_BLK_DEV_COW_COMMON) += cow_user.o
 obj-$(CONFIG_UML_RANDOM) += random.o
 
 # pcap_user.o must be added explicitly.
-USER_OBJS := fd.o null.o pty.o tty.o xterm.o slip_common.o pcap_user.o vde_user.o
+USER_OBJS := fd.o null.o pty.o tty.o xterm.o slip_common.o pcap_user.o vde_user.o vector_user.o
 CFLAGS_null.o = -DDEV_NULL=$(DEV_NULL_PATH)
 
 include arch/um/scripts/Makefile.rules
diff --git a/arch/um/drivers/net_kern.c b/arch/um/drivers/net_kern.c
index b305f8247909..3ef1b48e064a 100644
--- a/arch/um/drivers/net_kern.c
+++ b/arch/um/drivers/net_kern.c
@@ -288,7 +288,7 @@ static void uml_net_user_timer_expire(struct timer_list *t)
 #endif
 }
 
-static void setup_etheraddr(struct net_device *dev, char *str)
+void uml_net_setup_etheraddr(struct net_device *dev, char *str)
 {
 	unsigned char *addr = dev->dev_addr;
 	char *end;
@@ -412,7 +412,7 @@ static void eth_configure(int n, void *init, char *mac,
 	 */
 	snprintf(dev->name, sizeof(dev->name), "eth%d", n);
 
-	setup_etheraddr(dev, mac);
+	uml_net_setup_etheraddr(dev, mac);
 
 	printk(KERN_INFO "Netdevice %d (%pM) : ", n, dev->dev_addr);
 
diff --git a/arch/um/drivers/vector_kern.c b/arch/um/drivers/vector_kern.c
new file mode 100644
index 000000000000..d1d53015d572
--- /dev/null
+++ b/arch/um/drivers/vector_kern.c
@@ -0,0 +1,1630 @@
+/*
+ * Copyright (C) 2017 - Cambridge Greys Limited
+ * Copyright (C) 2011 - 2014 Cisco Systems Inc
+ * Copyright (C) 2001 - 2007 Jeff Dike (jdike@{addtoit,linux.intel}.com)
+ * Copyright (C) 2001 Lennert Buytenhek (buytenh@gnu.org) and
+ * James Leu (jleu@mindspring.net).
+ * Copyright (C) 2001 by various other people who didn't put their name here.
+ * Licensed under the GPL.
+ */
+
+#include <linux/version.h>
+#include <linux/bootmem.h>
+#include <linux/etherdevice.h>
+#include <linux/ethtool.h>
+#include <linux/inetdevice.h>
+#include <linux/init.h>
+#include <linux/list.h>
+#include <linux/netdevice.h>
+#include <linux/platform_device.h>
+#include <linux/rtnetlink.h>
+#include <linux/skbuff.h>
+#include <linux/slab.h>
+#include <linux/interrupt.h>
+#include <init.h>
+#include <irq_kern.h>
+#include <irq_user.h>
+#include <net_kern.h>
+#include <os.h>
+#include "mconsole_kern.h"
+#include "vector_user.h"
+#include "vector_kern.h"
+
+/*
+ * Adapted from network devices with the following major changes:
+ * All transports are static - simplifies the code significantly
+ * Multiple FDs/IRQs per device
+ * Vector IO optionally used for read/write, falling back to legacy
+ * based on configuration and/or availability
+ * Configuration is no longer positional - L2TPv3 and GRE require up to
+ * 10 parameters, passing this as positional is not fit for purpose.
+ * Only socket transports are supported
+ */
+
+
+#define DRIVER_NAME "uml-vector"
+#define DRIVER_VERSION "01"
+struct vector_cmd_line_arg {
+	struct list_head list;
+	int unit;
+	char *arguments;
+};
+
+struct vector_device {
+	struct list_head list;
+	struct net_device *dev;
+	struct platform_device pdev;
+	int unit;
+	int opened;
+};
+
+static LIST_HEAD(vec_cmd_line);
+
+static DEFINE_SPINLOCK(vector_devices_lock);
+static LIST_HEAD(vector_devices);
+
+static int driver_registered;
+
+static void vector_eth_configure(int n, struct arglist *def);
+
+/* Argument accessors to set variables (and/or set default values)
+ * mtu, buffer sizing, default headroom, etc
+ */
+
+#define DEFAULT_HEADROOM 2
+#define SAFETY_MARGIN 32
+#define DEFAULT_VECTOR_SIZE 64
+#define TX_SMALL_PACKET 128
+#define MAX_IOV_SIZE (MAX_SKB_FRAGS + 1)
+
+static const struct {
+	const char string[ETH_GSTRING_LEN];
+} ethtool_stats_keys[] = {
+	{ "rx_queue_max" },
+	{ "rx_queue_running_average" },
+	{ "tx_queue_max" },
+	{ "tx_queue_running_average" },
+	{ "rx_encaps_errors" },
+	{ "tx_timeout_count" },
+	{ "tx_restart_queue" },
+	{ "tx_kicks" },
+	{ "tx_flow_control_xon" },
+	{ "tx_flow_control_xoff" },
+	{ "rx_csum_offload_good" },
+	{ "rx_csum_offload_errors"},
+	{ "sg_ok"},
+	{ "sg_linearized"},
+};
+
+#define VECTOR_NUM_STATS	ARRAY_SIZE(ethtool_stats_keys)
+
+static void vector_reset_stats(struct vector_private *vp)
+{
+	vp->estats.rx_queue_max = 0;
+	vp->estats.rx_queue_running_average = 0;
+	vp->estats.tx_queue_max = 0;
+	vp->estats.tx_queue_running_average = 0;
+	vp->estats.rx_encaps_errors = 0;
+	vp->estats.tx_timeout_count = 0;
+	vp->estats.tx_restart_queue = 0;
+	vp->estats.tx_kicks = 0;
+	vp->estats.tx_flow_control_xon = 0;
+	vp->estats.tx_flow_control_xoff = 0;
+	vp->estats.sg_ok = 0;
+	vp->estats.sg_linearized = 0;
+}
+
+static int get_mtu(struct arglist *def)
+{
+	char *mtu = uml_vector_fetch_arg(def, "mtu");
+	long result;
+
+	if (mtu != NULL) {
+		if (kstrtoul(mtu, 10, &result) == 0)
+			return result;
+	}
+	return ETH_MAX_PACKET;
+}
+
+static int get_depth(struct arglist *def)
+{
+	char *mtu = uml_vector_fetch_arg(def, "depth");
+	long result;
+
+	if (mtu != NULL) {
+		if (kstrtoul(mtu, 10, &result) == 0)
+			return result;
+	}
+	return DEFAULT_VECTOR_SIZE;
+}
+
+static int get_headroom(struct arglist *def)
+{
+	char *mtu = uml_vector_fetch_arg(def, "headroom");
+	long result;
+
+	if (mtu != NULL) {
+		if (kstrtoul(mtu, 10, &result) == 0)
+			return result;
+	}
+	return DEFAULT_HEADROOM;
+}
+
+static int get_req_size(struct arglist *def)
+{
+	char *gro = uml_vector_fetch_arg(def, "gro");
+	long result;
+
+	if (gro != NULL) {
+		if (kstrtoul(gro, 10, &result) == 0) {
+			if (result > 0)
+				return 65536;
+		}
+	}
+	return get_mtu(def) + ETH_HEADER_OTHER +
+		get_headroom(def) + SAFETY_MARGIN;
+}
+
+
+static int get_transport_options(struct arglist *def)
+{
+	char *transport = uml_vector_fetch_arg(def, "transport");
+	char *vector = uml_vector_fetch_arg(def, "vec");
+
+	int vec_rx = VECTOR_RX;
+	int vec_tx = VECTOR_TX;
+	long parsed;
+
+	if (vector != NULL) {
+		if (kstrtoul(vector, 10, &parsed) == 0) {
+			if (parsed == 0) {
+				vec_rx = 0;
+				vec_tx = 0;
+			}
+		}
+	}
+
+
+	if (strncmp(transport, TRANS_TAP, TRANS_TAP_LEN) == 0)
+		return (vec_rx | VECTOR_BPF);
+	if (strncmp(transport, TRANS_RAW, TRANS_RAW_LEN) == 0)
+		return (vec_rx | vec_tx | VECTOR_BPF);
+	return (vec_rx | vec_tx);
+}
+
+
+/* A mini-buffer for packet drop read
+ * All of our supported transports are datagram oriented and we always
+ * read using recvmsg or recvmmsg. If we pass a buffer which is smaller
+ * than the packet size it still counts as full packet read and will
+ * clean the incoming stream to keep sigio/epoll happy
+ */
+
+#define DROP_BUFFER_SIZE 32
+
+static char *drop_buffer;
+
+/* Array backed queues optimized for bulk enqueue/dequeue and
+ * 1:N (small values of N) or 1:1 enqueuer/dequeuer ratios.
+ * For more details and full design rationale see
+ * http://foswiki.cambridgegreys.com/Main/EatYourTailAndEnjoyIt
+ */
+
+
+/*
+ * Advance the mmsg queue head by n = advance. Resets the queue to
+ * maximum enqueue/dequeue-at-once capacity if possible. Called by
+ * dequeuers. Caller must hold the head_lock!
+ */
+
+static int vector_advancehead(struct vector_queue *qi, int advance)
+{
+	int queue_depth;
+
+	qi->head =
+		(qi->head + advance)
+			% qi->max_depth;
+
+
+	spin_lock(&qi->tail_lock);
+	qi->queue_depth -= advance;
+
+	/* we are at 0, use this to
+	 * reset head and tail so we can use max size vectors
+	 */
+
+	if (qi->queue_depth == 0) {
+		qi->head = 0;
+		qi->tail = 0;
+	}
+	queue_depth = qi->queue_depth;
+	spin_unlock(&qi->tail_lock);
+	return queue_depth;
+}
+
+/*	Advance the queue tail by n = advance.
+ *	This is called by enqueuers which should hold the
+ *	head lock already
+ */
+
+static int vector_advancetail(struct vector_queue *qi, int advance)
+{
+	int queue_depth;
+
+	qi->tail =
+		(qi->tail + advance)
+			% qi->max_depth;
+	spin_lock(&qi->head_lock);
+	qi->queue_depth += advance;
+	queue_depth = qi->queue_depth;
+	spin_unlock(&qi->head_lock);
+	return queue_depth;
+}
+
+static int prep_msg(struct vector_private *vp,
+	struct sk_buff *skb,
+	struct iovec *iov)
+{
+	int iov_index = 0;
+	int nr_frags, frag;
+	skb_frag_t *skb_frag;
+
+	nr_frags = skb_shinfo(skb)->nr_frags;
+	if (nr_frags > MAX_IOV_SIZE) {
+		if (skb_linearize(skb) != 0)
+			goto drop;
+	}
+	if (vp->header_size > 0) {
+		iov[iov_index].iov_len = vp->header_size;
+		vp->form_header(iov[iov_index].iov_base, skb, vp);
+		iov_index++;
+	}
+	iov[iov_index].iov_base = skb->data;
+	if (nr_frags > 0) {
+		iov[iov_index].iov_len = skb->len - skb->data_len;
+		vp->estats.sg_ok++;
+	} else
+		iov[iov_index].iov_len = skb->len;
+	iov_index++;
+	for (frag = 0; frag < nr_frags; frag++) {
+		skb_frag = &skb_shinfo(skb)->frags[frag];
+		iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
+		iov[iov_index].iov_len = skb_frag_size(skb_frag);
+		iov_index++;
+	}
+	return iov_index;
+drop:
+	return -1;
+}
+/*
+ * Generic vector enqueue with support for forming headers using transport
+ * specific callback. Allows GRE, L2TPv3, RAW and other transports
+ * to use a common enqueue procedure in vector mode
+ */
+
+static int vector_enqueue(struct vector_queue *qi, struct sk_buff *skb)
+{
+	struct vector_private *vp = netdev_priv(qi->dev);
+	int queue_depth;
+	int packet_len;
+	struct mmsghdr *mmsg_vector = qi->mmsg_vector;
+	int iov_count;
+
+	spin_lock(&qi->tail_lock);
+	spin_lock(&qi->head_lock);
+	queue_depth = qi->queue_depth;
+	spin_unlock(&qi->head_lock);
+
+	if (skb)
+		packet_len = skb->len;
+
+	if (queue_depth < qi->max_depth) {
+
+		*(qi->skbuff_vector + qi->tail) = skb;
+		mmsg_vector += qi->tail;
+		iov_count = prep_msg(
+			vp,
+			skb,
+			mmsg_vector->msg_hdr.msg_iov
+		);
+		if (iov_count < 1)
+			goto drop;
+		mmsg_vector->msg_hdr.msg_iovlen = iov_count;
+		mmsg_vector->msg_hdr.msg_name = vp->fds->remote_addr;
+		mmsg_vector->msg_hdr.msg_namelen = vp->fds->remote_addr_size;
+		queue_depth = vector_advancetail(qi, 1);
+	} else
+		goto drop;
+	spin_unlock(&qi->tail_lock);
+	return queue_depth;
+drop:
+	qi->dev->stats.tx_dropped++;
+	if (skb != NULL) {
+		packet_len = skb->len;
+		dev_consume_skb_any(skb);
+		netdev_completed_queue(qi->dev, 1, packet_len);
+	}
+	spin_unlock(&qi->tail_lock);
+	return queue_depth;
+}
+
+static int consume_vector_skbs(struct vector_queue *qi, int count)
+{
+	struct sk_buff *skb;
+	int skb_index;
+	int bytes_compl = 0;
+
+	for (skb_index = qi->head; skb_index < qi->head + count; skb_index++) {
+		skb = *(qi->skbuff_vector + skb_index);
+		/* mark as empty to ensure correct destruction if
+		 * needed
+		 */
+		bytes_compl += skb->len;
+		*(qi->skbuff_vector + skb_index) = NULL;
+		dev_consume_skb_any(skb);
+	}
+	qi->dev->stats.tx_bytes += bytes_compl;
+	qi->dev->stats.tx_packets += count;
+	netdev_completed_queue(qi->dev, count, bytes_compl);
+	return vector_advancehead(qi, count);
+}
+
+/*
+ * Generic vector deque via sendmmsg with support for forming headers
+ * using transport specific callback. Allows GRE, L2TPv3, RAW and
+ * other transports to use a common dequeue procedure in vector mode
+ */
+
+
+static int vector_send(struct vector_queue *qi)
+{
+	struct vector_private *vp = netdev_priv(qi->dev);
+	struct mmsghdr *send_from;
+	int result = 0, send_len, queue_depth = qi->max_depth;
+
+	if (spin_trylock(&qi->head_lock)) {
+		if (spin_trylock(&qi->tail_lock)) {
+			/* update queue_depth to current value */
+			queue_depth = qi->queue_depth;
+			spin_unlock(&qi->tail_lock);
+			while (queue_depth > 0) {
+				/* Calculate the start of the vector */
+				send_len = queue_depth;
+				send_from = qi->mmsg_vector;
+				send_from += qi->head;
+				/* Adjust vector size if wraparound */
+				if (send_len + qi->head > qi->max_depth)
+					send_len = qi->max_depth - qi->head;
+				/* Try to TX as many packets as possible */
+				if (send_len > 0) {
+					result = uml_vector_sendmmsg(
+						 vp->fds->tx_fd,
+						 send_from,
+						 send_len,
+						 0
+					);
+					vp->in_write_poll =
+						(result != send_len);
+				}
+				/* For some of the sendmmsg error scenarios
+				 * we may end being unsure in the TX success
+				 * for all packets. It is safer to declare
+				 * them all TX-ed and blame the network.
+				 */
+				if (result < 0) {
+					if (net_ratelimit())
+						netdev_err(vp->dev, "sendmmsg err=%i\n",
+							result);
+					result = send_len;
+				}
+				if (result > 0) {
+					queue_depth =
+						consume_vector_skbs(qi, result);
+					/* This is equivalent to an TX IRQ.
+					 * Restart the upper layers to feed us
+					 * more packets.
+					 */
+					if (result > vp->estats.tx_queue_max)
+						vp->estats.tx_queue_max = result;
+					vp->estats.tx_queue_running_average =
+						(vp->estats.tx_queue_running_average + result) >> 1;
+				}
+				netif_trans_update(qi->dev);
+				netif_wake_queue(qi->dev);
+				/* if TX is busy, break out of the send loop,
+				 *  poll write IRQ will reschedule xmit for us
+				 */
+				if (result != send_len) {
+					vp->estats.tx_restart_queue++;
+					break;
+				}
+			}
+		}
+		spin_unlock(&qi->head_lock);
+	} else {
+		tasklet_schedule(&vp->tx_poll);
+	}
+	return queue_depth;
+}
+
+/* Queue destructor. Deliberately stateless so we can use
+ * it in queue cleanup if initialization fails.
+ */
+
+static void destroy_queue(struct vector_queue *qi)
+{
+	int i;
+	struct iovec *iov;
+	struct vector_private *vp = netdev_priv(qi->dev);
+	struct mmsghdr *mmsg_vector;
+
+	if (qi == NULL)
+		return;
+	/* deallocate any skbuffs - we rely on any unused to be
+	 * set to NULL.
+	 */
+	if (qi->skbuff_vector != NULL) {
+		for (i = 0; i < qi->max_depth; i++) {
+			if (*(qi->skbuff_vector + i) != NULL)
+				dev_kfree_skb_any(*(qi->skbuff_vector + i));
+		}
+		kfree(qi->skbuff_vector);
+	}
+	/* deallocate matching IOV structures including header buffs */
+	if (qi->mmsg_vector != NULL) {
+		mmsg_vector = qi->mmsg_vector;
+		for (i = 0; i < qi->max_depth; i++) {
+			iov = mmsg_vector->msg_hdr.msg_iov;
+			if (iov != NULL) {
+				if ((vp->header_size > 0) &&
+					(iov->iov_base != NULL))
+					kfree(iov->iov_base);
+				kfree(iov);
+			}
+			mmsg_vector++;
+		}
+		kfree(qi->mmsg_vector);
+	}
+	kfree(qi);
+}
+
+/*
+ * Queue constructor. Create a queue with a given side.
+ */
+static struct vector_queue *create_queue(
+	struct vector_private *vp,
+	int max_size,
+	int header_size,
+	int num_extra_frags)
+{
+	struct vector_queue *result;
+	int i;
+	struct iovec *iov;
+	struct mmsghdr *mmsg_vector;
+
+	result = kmalloc(sizeof(struct vector_queue), GFP_KERNEL);
+	if (result == NULL)
+		goto out_fail;
+	result->max_depth = max_size;
+	result->dev = vp->dev;
+	result->mmsg_vector = kmalloc(
+		(sizeof(struct mmsghdr) * max_size), GFP_KERNEL);
+	result->skbuff_vector = kmalloc(
+		(sizeof(void *) * max_size), GFP_KERNEL);
+	if (result->mmsg_vector == NULL || result->skbuff_vector == NULL)
+		goto out_fail;
+
+	mmsg_vector = result->mmsg_vector;
+	for (i = 0; i < max_size; i++) {
+		/* Clear all pointers - we use non-NULL as marking on
+		 * what to free on destruction
+		 */
+		*(result->skbuff_vector + i) = NULL;
+		mmsg_vector->msg_hdr.msg_iov = NULL;
+		mmsg_vector++;
+	}
+	mmsg_vector = result->mmsg_vector;
+	result->max_iov_frags = num_extra_frags;
+	for (i = 0; i < max_size; i++) {
+		if (vp->header_size > 0)
+			iov = kmalloc(
+				sizeof(struct iovec) * (3 + num_extra_frags),
+				GFP_KERNEL
+			);
+		else
+			iov = kmalloc(
+				sizeof(struct iovec) * (2 + num_extra_frags),
+				GFP_KERNEL
+			);
+		if (iov == NULL)
+			goto out_fail;
+		mmsg_vector->msg_hdr.msg_iov = iov;
+		mmsg_vector->msg_hdr.msg_iovlen = 1;
+		mmsg_vector->msg_hdr.msg_control = NULL;
+		mmsg_vector->msg_hdr.msg_controllen = 0;
+		mmsg_vector->msg_hdr.msg_flags = MSG_DONTWAIT;
+		mmsg_vector->msg_hdr.msg_name = NULL;
+		mmsg_vector->msg_hdr.msg_namelen = 0;
+		if (vp->header_size > 0) {
+			iov->iov_base = kmalloc(header_size, GFP_KERNEL);
+			if (iov->iov_base == NULL)
+				goto out_fail;
+			iov->iov_len = header_size;
+			mmsg_vector->msg_hdr.msg_iovlen = 2;
+			iov++;
+		}
+		iov->iov_base = NULL;
+		iov->iov_len = 0;
+		mmsg_vector++;
+	}
+	spin_lock_init(&result->head_lock);
+	spin_lock_init(&result->tail_lock);
+	result->queue_depth = 0;
+	result->head = 0;
+	result->tail = 0;
+	return result;
+out_fail:
+	destroy_queue(result);
+	return NULL;
+}
+
+/*
+ * We do not use the RX queue as a proper wraparound queue for now
+ * This is not necessary because the consumption via netif_rx()
+ * happens in-line. While we can try using the return code of
+ * netif_rx() for flow control there are no drivers doing this today.
+ * For this RX specific use we ignore the tail/head locks and
+ * just read into a prepared queue filled with skbuffs.
+ */
+
+static struct sk_buff *prep_skb(
+	struct vector_private *vp,
+	struct user_msghdr *msg)
+{
+	int linear = vp->max_packet + vp->headroom + SAFETY_MARGIN;
+	struct sk_buff *result;
+	int iov_index = 0, len;
+	struct iovec *iov = msg->msg_iov;
+	int err, nr_frags, frag;
+	skb_frag_t *skb_frag;
+
+	if (vp->req_size <= linear)
+		len = linear;
+	else
+		len = vp->req_size;
+	result = alloc_skb_with_frags(
+		linear,
+		len - vp->max_packet,
+		3,
+		&err,
+		GFP_ATOMIC
+	);
+	if (vp->header_size > 0)
+		iov_index++;
+	if (result == NULL) {
+		iov[iov_index].iov_base = NULL;
+		iov[iov_index].iov_len = 0;
+		goto done;
+	}
+	skb_reserve(result, vp->headroom);
+	result->dev = vp->dev;
+	skb_put(result, vp->max_packet);
+	result->data_len = len - vp->max_packet;
+	result->len += len - vp->max_packet;
+	skb_reset_mac_header(result);
+	result->ip_summed = CHECKSUM_NONE;
+	iov[iov_index].iov_base = result->data;
+	iov[iov_index].iov_len = vp->max_packet;
+	iov_index++;
+
+	nr_frags = skb_shinfo(result)->nr_frags;
+	for (frag = 0; frag < nr_frags; frag++) {
+		skb_frag = &skb_shinfo(result)->frags[frag];
+		iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
+		if (iov[iov_index].iov_base != NULL)
+			iov[iov_index].iov_len = skb_frag_size(skb_frag);
+		else
+			iov[iov_index].iov_len = 0;
+		iov_index++;
+	}
+done:
+	msg->msg_iovlen = iov_index;
+	return result;
+}
+
+
+/* Prepare queue for recvmmsg one-shot rx - fill with fresh sk_buffs*/
+
+static void prep_queue_for_rx(struct vector_queue *qi)
+{
+	struct vector_private *vp = netdev_priv(qi->dev);
+	struct mmsghdr *mmsg_vector = qi->mmsg_vector;
+	void **skbuff_vector = qi->skbuff_vector;
+	int i;
+
+	if (qi->queue_depth == 0)
+		return;
+	for (i = 0; i < qi->queue_depth; i++) {
+		/* it is OK if allocation fails - recvmmsg with NULL data in
+		 * iov argument still performs an RX, just drops the packet
+		 * This allows us stop faffing around with a "drop buffer"
+		 */
+
+		*skbuff_vector = prep_skb(vp, &mmsg_vector->msg_hdr);
+		skbuff_vector++;
+		mmsg_vector++;
+	}
+	qi->queue_depth = 0;
+}
+
+static struct vector_device *find_device(int n)
+{
+	struct vector_device *device;
+	struct list_head *ele;
+
+	spin_lock(&vector_devices_lock);
+	list_for_each(ele, &vector_devices) {
+		device = list_entry(ele, struct vector_device, list);
+		if (device->unit == n)
+			goto out;
+	}
+	device = NULL;
+ out:
+	spin_unlock(&vector_devices_lock);
+	return device;
+}
+
+static int vector_parse(char *str, int *index_out, char **str_out,
+			char **error_out)
+{
+	int n, len, err = -EINVAL;
+	char *start = str;
+
+	len = strlen(str);
+
+	while ((*str != ':') && (strlen(str) > 1))
+		str++;
+	if (*str != ':') {
+		*error_out = "Expected ':' after device number";
+		return err;
+	}
+	*str = '\0';
+
+	err = kstrtouint(start, 0, &n);
+	if (err < 0) {
+		*error_out = "Bad device number";
+		return err;
+	}
+
+	str++;
+	if (find_device(n)) {
+		*error_out = "Device already configured";
+		return err;
+	}
+
+	*index_out = n;
+	*str_out = str;
+	return 0;
+}
+
+static int vector_config(char *str, char **error_out)
+{
+	int err, n;
+	char *params;
+	struct arglist *parsed;
+
+	err = vector_parse(str, &n, &params, error_out);
+	if (err != 0)
+		return err;
+
+	/* This string is broken up and the pieces used by the underlying
+	 * driver. We should copy it to make sure things do not go wrong
+	 * later.
+	 */
+
+	params = kstrdup(params, GFP_KERNEL);
+	if (str == NULL) {
+		*error_out = "vector_config failed to strdup string";
+		return -ENOMEM;
+	}
+
+	parsed = uml_parse_vector_ifspec(params);
+
+	if (parsed == NULL) {
+		*error_out = "vector_config failed to parse parameters";
+		return -EINVAL;
+	}
+
+	vector_eth_configure(n, parsed);
+	return 0;
+}
+
+static int vector_id(char **str, int *start_out, int *end_out)
+{
+	char *end;
+	int n;
+
+	n = simple_strtoul(*str, &end, 0);
+	if ((*end != '\0') || (end == *str))
+		return -1;
+
+	*start_out = n;
+	*end_out = n;
+	*str = end;
+	return n;
+}
+
+static int vector_remove(int n, char **error_out)
+{
+	struct vector_device *vec_d;
+	struct net_device *dev;
+	struct vector_private *vp;
+
+	vec_d = find_device(n);
+	if (vec_d == NULL)
+		return -ENODEV;
+	dev = vec_d->dev;
+	vp = netdev_priv(dev);
+	if (vp->fds != NULL)
+		return -EBUSY;
+	unregister_netdev(dev);
+	platform_device_unregister(&vec_d->pdev);
+	return 0;
+}
+
+/*
+ * There is no shared per-transport initialization code, so
+ * we will just initialize each interface one by one and
+ * add them to a list
+ */
+
+static struct platform_driver uml_net_driver = {
+	.driver = {
+		.name = DRIVER_NAME,
+	},
+};
+
+
+static void vector_device_release(struct device *dev)
+{
+	struct vector_device *device = dev_get_drvdata(dev);
+	struct net_device *netdev = device->dev;
+
+	list_del(&device->list);
+	kfree(device);
+	free_netdev(netdev);
+}
+
+/* Bog standard recv using recvmsg - not used normally unless the user
+ * explicitly specifies not to use recvmmsg vector RX.
+ */
+
+static int vector_legacy_rx(struct vector_private *vp)
+{
+	int pkt_len;
+	struct user_msghdr hdr;
+	struct iovec iov[2 + MAX_IOV_SIZE]; /* header + data use case only */
+	int iovpos = 0;
+	struct sk_buff *skb;
+	int header_check;
+
+	hdr.msg_name = NULL;
+	hdr.msg_namelen = 0;
+	hdr.msg_iov = (struct iovec *) &iov;
+	hdr.msg_control = NULL;
+	hdr.msg_controllen = 0;
+	hdr.msg_flags = 0;
+
+	if (vp->header_size > 0) {
+		iov[0].iov_base = vp->header_rxbuffer;
+		iov[0].iov_len = vp->header_size;
+	}
+
+	skb = prep_skb(vp, &hdr);
+
+	if (skb == NULL) {
+		/* Read a packet into drop_buffer and don't do
+		 * anything with it.
+		 */
+		iov[iovpos].iov_base = drop_buffer;
+		iov[iovpos].iov_len = DROP_BUFFER_SIZE;
+		hdr.msg_iovlen = 1;
+		vp->dev->stats.rx_dropped++;
+	}
+
+	pkt_len = uml_vector_recvmsg(vp->fds->rx_fd, &hdr, 0);
+
+	if (skb != NULL) {
+		if (pkt_len > vp->header_size) {
+			if (vp->header_size > 0) {
+				header_check = vp->verify_header(
+					vp->header_rxbuffer, skb, vp);
+				if (header_check < 0) {
+					dev_kfree_skb_irq(skb);
+					vp->dev->stats.rx_dropped++;
+					vp->estats.rx_encaps_errors++;
+					return 0;
+				}
+				if (header_check > 0) {
+					vp->estats.rx_csum_offload_good++;
+					skb->ip_summed = CHECKSUM_UNNECESSARY;
+				}
+			}
+			pskb_trim(skb, pkt_len - vp->rx_header_size);
+			skb->protocol = eth_type_trans(skb, skb->dev);
+			vp->dev->stats.rx_bytes += skb->len;
+			vp->dev->stats.rx_packets++;
+			netif_rx(skb);
+		} else {
+			dev_kfree_skb_irq(skb);
+		}
+	}
+	return pkt_len;
+}
+
+/*
+ * Packet at a time TX which falls back to vector TX if the
+ * underlying transport is busy.
+ */
+
+
+
+static int writev_tx(struct vector_private *vp, struct sk_buff *skb)
+{
+	struct iovec iov[3 + MAX_IOV_SIZE];
+	int iov_count, pkt_len = 0;
+
+	iov[0].iov_base = vp->header_txbuffer;
+	iov_count = prep_msg(vp, skb, (struct iovec *) &iov);
+
+	if (iov_count < 1)
+		goto drop;
+	pkt_len = uml_vector_writev(
+		vp->fds->tx_fd,
+		(struct iovec *) &iov,
+		iov_count
+	);
+
+	netif_trans_update(vp->dev);
+	netif_wake_queue(vp->dev);
+
+	if (pkt_len > 0) {
+		vp->dev->stats.tx_bytes += skb->len;
+		vp->dev->stats.tx_packets++;
+	} else {
+		vp->dev->stats.tx_dropped++;
+	}
+	consume_skb(skb);
+	return pkt_len;
+drop:
+	vp->dev->stats.tx_dropped++;
+	consume_skb(skb);
+	return pkt_len;
+}
+
+/*
+ * Receive as many messages as we can in one call using the special
+ * mmsg vector matched to an skb vector which we prepared earlier.
+ */
+
+static int vector_mmsg_rx(struct vector_private *vp)
+{
+	int packet_count, i;
+	struct vector_queue *qi = vp->rx_queue;
+	struct sk_buff *skb;
+	struct mmsghdr *mmsg_vector = qi->mmsg_vector;
+	void **skbuff_vector = qi->skbuff_vector;
+	int header_check;
+
+	/* Refresh the vector and make sure it is with new skbs and the
+	 * iovs are updated to point to them.
+	 */
+
+	prep_queue_for_rx(qi);
+
+	/* Fire the Lazy Gun - get as many packets as we can in one go. */
+
+	packet_count = uml_vector_recvmmsg(
+		vp->fds->rx_fd, qi->mmsg_vector, qi->max_depth, 0);
+
+	if (packet_count <= 0)
+		return packet_count;
+
+	/* We treat packet processing as enqueue, buffer refresh as dequeue
+	 * The queue_depth tells us how many buffers have been used and how
+	 * many do we need to prep the next time prep_queue_for_rx() is called.
+	 */
+
+	qi->queue_depth = packet_count;
+
+	for (i = 0; i < packet_count; i++) {
+		skb = (*skbuff_vector);
+		if (mmsg_vector->msg_len > vp->header_size) {
+			if (vp->header_size > 0) {
+				header_check = vp->verify_header(
+					mmsg_vector->msg_hdr.msg_iov->iov_base,
+					skb,
+					vp
+				);
+				if (header_check < 0) {
+				/* Overlay header failed to verify - discard.
+				 * We can actually keep this skb and reuse it,
+				 * but that will make the prep logic too
+				 * complex.
+				 */
+					dev_kfree_skb_irq(skb);
+					vp->estats.rx_encaps_errors++;
+					continue;
+				}
+				if (header_check > 0) {
+					vp->estats.rx_csum_offload_good++;
+					skb->ip_summed = CHECKSUM_UNNECESSARY;
+				}
+			}
+			pskb_trim(skb,
+				mmsg_vector->msg_len - vp->rx_header_size);
+			skb->protocol = eth_type_trans(skb, skb->dev);
+			/*
+			 * We do not need to lock on updating stats here
+			 * The interrupt loop is non-reentrant.
+			 */
+			vp->dev->stats.rx_bytes += skb->len;
+			vp->dev->stats.rx_packets++;
+			netif_rx(skb);
+		} else {
+			/* Overlay header too short to do anything - discard.
+			 * We can actually keep this skb and reuse it,
+			 * but that will make the prep logic too complex.
+			 */
+			if (skb != NULL)
+				dev_kfree_skb_irq(skb);
+		}
+		(*skbuff_vector) = NULL;
+		/* Move to the next buffer element */
+		mmsg_vector++;
+		skbuff_vector++;
+	}
+	if (packet_count > 0) {
+		if (vp->estats.rx_queue_max < packet_count)
+			vp->estats.rx_queue_max = packet_count;
+		vp->estats.rx_queue_running_average =
+			(vp->estats.rx_queue_running_average + packet_count) >> 1;
+	}
+	return packet_count;
+}
+
+static void vector_rx(struct vector_private *vp)
+{
+	int err;
+
+	if ((vp->options & VECTOR_RX) > 0)
+		while ((err = vector_mmsg_rx(vp)) > 0)
+			;
+	else
+		while ((err = vector_legacy_rx(vp)) > 0)
+			;
+	if ((err != 0) && net_ratelimit())
+		netdev_err(vp->dev, "vector_rx: error(%d)\n", err);
+}
+
+static int vector_net_start_xmit(struct sk_buff *skb, struct net_device *dev)
+{
+	struct vector_private *vp = netdev_priv(dev);
+	int queue_depth = 0;
+
+	if ((vp->options & VECTOR_TX) == 0) {
+		writev_tx(vp, skb);
+		return NETDEV_TX_OK;
+	}
+
+	/* We do BQL only in the vector path, no point doing it in
+	 * packet at a time mode as there is no device queue
+	 */