xen: add virtual network device driver
authorJeremy Fitzhardinge <jeremy@xensource.com>
Wed, 18 Jul 2007 01:37:06 +0000 (18:37 -0700)
committerJeremy Fitzhardinge <jeremy@goop.org>
Wed, 18 Jul 2007 15:47:45 +0000 (08:47 -0700)
The network device frontend driver allows the kernel to access network
devices exported exported by a virtual machine containing a physical
network device driver.

Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Chris Wright <chrisw@sous-sol.org>
Acked-by: Jeff Garzik <jeff@garzik.org>
Cc: Ian Pratt <ian.pratt@xensource.com>
Cc: Christian Limpach <Christian.Limpach@cl.cam.ac.uk>
Cc: Stephen Hemminger <shemminger@linux-foundation.org>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Keir Fraser <Keir.Fraser@cl.cam.ac.uk>
Cc: netdev@vger.kernel.org
drivers/net/Kconfig
drivers/net/Makefile
drivers/net/xen-netfront.c [new file with mode: 0644]

index 43d03178064dfb96f8c7957c328e0d1f2754cea8..5fb659f8b20e537c0b22ff504b182811c59a47cd 100644 (file)
@@ -2486,6 +2486,18 @@ source "drivers/atm/Kconfig"
 
 source "drivers/s390/net/Kconfig"
 
+config XEN_NETDEV_FRONTEND
+       tristate "Xen network device frontend driver"
+       depends on XEN
+       default y
+       help
+         The network device frontend driver allows the kernel to
+         access network devices exported exported by a virtual
+         machine containing a physical network device driver. The
+         frontend driver is intended for unprivileged guest domains;
+         if you are compiling a kernel for a Xen guest, you almost
+         certainly want to enable this.
+
 config ISERIES_VETH
        tristate "iSeries Virtual Ethernet driver support"
        depends on PPC_ISERIES
index eb4167622a6abdacafaa9a042f2152be9aa36865..0e286ab8855a5202d925f684e092e9117ff1f6dd 100644 (file)
@@ -127,6 +127,8 @@ obj-$(CONFIG_PPPOL2TP) += pppox.o pppol2tp.o
 obj-$(CONFIG_SLIP) += slip.o
 obj-$(CONFIG_SLHC) += slhc.o
 
+obj-$(CONFIG_XEN_NETDEV_FRONTEND) += xen-netfront.o
+
 obj-$(CONFIG_DUMMY) += dummy.o
 obj-$(CONFIG_IFB) += ifb.o
 obj-$(CONFIG_MACVLAN) += macvlan.o
diff --git a/drivers/net/xen-netfront.c b/drivers/net/xen-netfront.c
new file mode 100644 (file)
index 0000000..489f69c
--- /dev/null
@@ -0,0 +1,1863 @@
+/*
+ * Virtual network driver for conversing with remote driver backends.
+ *
+ * Copyright (c) 2002-2005, K A Fraser
+ * Copyright (c) 2005, XenSource Ltd
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License version 2
+ * as published by the Free Software Foundation; or, when distributed
+ * separately from the Linux kernel or incorporated into other
+ * software packages, subject to the following license:
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this source file (the "Software"), to deal in the Software without
+ * restriction, including without limitation the rights to use, copy, modify,
+ * merge, publish, distribute, sublicense, and/or sell copies of the Software,
+ * and to permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <linux/module.h>
+#include <linux/kernel.h>
+#include <linux/netdevice.h>
+#include <linux/etherdevice.h>
+#include <linux/skbuff.h>
+#include <linux/ethtool.h>
+#include <linux/if_ether.h>
+#include <linux/tcp.h>
+#include <linux/udp.h>
+#include <linux/moduleparam.h>
+#include <linux/mm.h>
+#include <net/ip.h>
+
+#include <xen/xenbus.h>
+#include <xen/events.h>
+#include <xen/page.h>
+#include <xen/grant_table.h>
+
+#include <xen/interface/io/netif.h>
+#include <xen/interface/memory.h>
+#include <xen/interface/grant_table.h>
+
+static struct ethtool_ops xennet_ethtool_ops;
+
+struct netfront_cb {
+       struct page *page;
+       unsigned offset;
+};
+
+#define NETFRONT_SKB_CB(skb)   ((struct netfront_cb *)((skb)->cb))
+
+#define RX_COPY_THRESHOLD 256
+
+#define GRANT_INVALID_REF      0
+
+#define NET_TX_RING_SIZE __RING_SIZE((struct xen_netif_tx_sring *)0, PAGE_SIZE)
+#define NET_RX_RING_SIZE __RING_SIZE((struct xen_netif_rx_sring *)0, PAGE_SIZE)
+#define TX_MAX_TARGET min_t(int, NET_RX_RING_SIZE, 256)
+
+struct netfront_info {
+       struct list_head list;
+       struct net_device *netdev;
+
+       struct net_device_stats stats;
+
+       struct xen_netif_tx_front_ring tx;
+       struct xen_netif_rx_front_ring rx;
+
+       spinlock_t   tx_lock;
+       spinlock_t   rx_lock;
+
+       unsigned int evtchn;
+
+       /* Receive-ring batched refills. */
+#define RX_MIN_TARGET 8
+#define RX_DFL_MIN_TARGET 64
+#define RX_MAX_TARGET min_t(int, NET_RX_RING_SIZE, 256)
+       unsigned rx_min_target, rx_max_target, rx_target;
+       struct sk_buff_head rx_batch;
+
+       struct timer_list rx_refill_timer;
+
+       /*
+        * {tx,rx}_skbs store outstanding skbuffs. Free tx_skb entries
+        * are linked from tx_skb_freelist through skb_entry.link.
+        *
+        *  NB. Freelist index entries are always going to be less than
+        *  PAGE_OFFSET, whereas pointers to skbs will always be equal or
+        *  greater than PAGE_OFFSET: we use this property to distinguish
+        *  them.
+        */
+       union skb_entry {
+               struct sk_buff *skb;
+               unsigned link;
+       } tx_skbs[NET_TX_RING_SIZE];
+       grant_ref_t gref_tx_head;
+       grant_ref_t grant_tx_ref[NET_TX_RING_SIZE];
+       unsigned tx_skb_freelist;
+
+       struct sk_buff *rx_skbs[NET_RX_RING_SIZE];
+       grant_ref_t gref_rx_head;
+       grant_ref_t grant_rx_ref[NET_RX_RING_SIZE];
+
+       struct xenbus_device *xbdev;
+       int tx_ring_ref;
+       int rx_ring_ref;
+
+       unsigned long rx_pfn_array[NET_RX_RING_SIZE];
+       struct multicall_entry rx_mcl[NET_RX_RING_SIZE+1];
+       struct mmu_update rx_mmu[NET_RX_RING_SIZE];
+};
+
+struct netfront_rx_info {
+       struct xen_netif_rx_response rx;
+       struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1];
+};
+
+/*
+ * Access macros for acquiring freeing slots in tx_skbs[].
+ */
+
+static void add_id_to_freelist(unsigned *head, union skb_entry *list,
+                              unsigned short id)
+{
+       list[id].link = *head;
+       *head = id;
+}
+
+static unsigned short get_id_from_freelist(unsigned *head,
+                                          union skb_entry *list)
+{
+       unsigned int id = *head;
+       *head = list[id].link;
+       return id;
+}
+
+static int xennet_rxidx(RING_IDX idx)
+{
+       return idx & (NET_RX_RING_SIZE - 1);
+}
+
+static struct sk_buff *xennet_get_rx_skb(struct netfront_info *np,
+                                        RING_IDX ri)
+{
+       int i = xennet_rxidx(ri);
+       struct sk_buff *skb = np->rx_skbs[i];
+       np->rx_skbs[i] = NULL;
+       return skb;
+}
+
+static grant_ref_t xennet_get_rx_ref(struct netfront_info *np,
+                                           RING_IDX ri)
+{
+       int i = xennet_rxidx(ri);
+       grant_ref_t ref = np->grant_rx_ref[i];
+       np->grant_rx_ref[i] = GRANT_INVALID_REF;
+       return ref;
+}
+
+#ifdef CONFIG_SYSFS
+static int xennet_sysfs_addif(struct net_device *netdev);
+static void xennet_sysfs_delif(struct net_device *netdev);
+#else /* !CONFIG_SYSFS */
+#define xennet_sysfs_addif(dev) (0)
+#define xennet_sysfs_delif(dev) do { } while (0)
+#endif
+
+static int xennet_can_sg(struct net_device *dev)
+{
+       return dev->features & NETIF_F_SG;
+}
+
+
+static void rx_refill_timeout(unsigned long data)
+{
+       struct net_device *dev = (struct net_device *)data;
+       netif_rx_schedule(dev);
+}
+
+static int netfront_tx_slot_available(struct netfront_info *np)
+{
+       return ((np->tx.req_prod_pvt - np->tx.rsp_cons) <
+               (TX_MAX_TARGET - MAX_SKB_FRAGS - 2));
+}
+
+static void xennet_maybe_wake_tx(struct net_device *dev)
+{
+       struct netfront_info *np = netdev_priv(dev);
+
+       if (unlikely(netif_queue_stopped(dev)) &&
+           netfront_tx_slot_available(np) &&
+           likely(netif_running(dev)))
+               netif_wake_queue(dev);
+}
+
+static void xennet_alloc_rx_buffers(struct net_device *dev)
+{
+       unsigned short id;
+       struct netfront_info *np = netdev_priv(dev);
+       struct sk_buff *skb;
+       struct page *page;
+       int i, batch_target, notify;
+       RING_IDX req_prod = np->rx.req_prod_pvt;
+       struct xen_memory_reservation reservation;
+       grant_ref_t ref;
+       unsigned long pfn;
+       void *vaddr;
+       int nr_flips;
+       struct xen_netif_rx_request *req;
+
+       if (unlikely(!netif_carrier_ok(dev)))
+               return;
+
+       /*
+        * Allocate skbuffs greedily, even though we batch updates to the
+        * receive ring. This creates a less bursty demand on the memory
+        * allocator, so should reduce the chance of failed allocation requests
+        * both for ourself and for other kernel subsystems.
+        */
+       batch_target = np->rx_target - (req_prod - np->rx.rsp_cons);
+       for (i = skb_queue_len(&np->rx_batch); i < batch_target; i++) {
+               skb = __netdev_alloc_skb(dev, RX_COPY_THRESHOLD,
+                                        GFP_ATOMIC | __GFP_NOWARN);
+               if (unlikely(!skb))
+                       goto no_skb;
+
+               page = alloc_page(GFP_ATOMIC | __GFP_NOWARN);
+               if (!page) {
+                       kfree_skb(skb);
+no_skb:
+                       /* Any skbuffs queued for refill? Force them out. */
+                       if (i != 0)
+                               goto refill;
+                       /* Could not allocate any skbuffs. Try again later. */
+                       mod_timer(&np->rx_refill_timer,
+                                 jiffies + (HZ/10));
+                       break;
+               }
+
+               skb_shinfo(skb)->frags[0].page = page;
+               skb_shinfo(skb)->nr_frags = 1;
+               __skb_queue_tail(&np->rx_batch, skb);
+       }
+
+       /* Is the batch large enough to be worthwhile? */
+       if (i < (np->rx_target/2)) {
+               if (req_prod > np->rx.sring->req_prod)
+                       goto push;
+               return;
+       }
+
+       /* Adjust our fill target if we risked running out of buffers. */
+       if (((req_prod - np->rx.sring->rsp_prod) < (np->rx_target / 4)) &&
+           ((np->rx_target *= 2) > np->rx_max_target))
+               np->rx_target = np->rx_max_target;
+
+ refill:
+       for (nr_flips = i = 0; ; i++) {
+               skb = __skb_dequeue(&np->rx_batch);
+               if (skb == NULL)
+                       break;
+
+               skb->dev = dev;
+
+               id = xennet_rxidx(req_prod + i);
+
+               BUG_ON(np->rx_skbs[id]);
+               np->rx_skbs[id] = skb;
+
+               ref = gnttab_claim_grant_reference(&np->gref_rx_head);
+               BUG_ON((signed short)ref < 0);
+               np->grant_rx_ref[id] = ref;
+
+               pfn = page_to_pfn(skb_shinfo(skb)->frags[0].page);
+               vaddr = page_address(skb_shinfo(skb)->frags[0].page);
+
+               req = RING_GET_REQUEST(&np->rx, req_prod + i);
+               gnttab_grant_foreign_access_ref(ref,
+                                               np->xbdev->otherend_id,
+                                               pfn_to_mfn(pfn),
+                                               0);
+
+               req->id = id;
+               req->gref = ref;
+       }
+
+       if (nr_flips != 0) {
+               reservation.extent_start = np->rx_pfn_array;
+               reservation.nr_extents   = nr_flips;
+               reservation.extent_order = 0;
+               reservation.address_bits = 0;
+               reservation.domid        = DOMID_SELF;
+
+               if (!xen_feature(XENFEAT_auto_translated_physmap)) {
+                       /* After all PTEs have been zapped, flush the TLB. */
+                       np->rx_mcl[i-1].args[MULTI_UVMFLAGS_INDEX] =
+                               UVMF_TLB_FLUSH|UVMF_ALL;
+
+                       /* Give away a batch of pages. */
+                       np->rx_mcl[i].op = __HYPERVISOR_memory_op;
+                       np->rx_mcl[i].args[0] = XENMEM_decrease_reservation;
+                       np->rx_mcl[i].args[1] = (unsigned long)&reservation;
+
+                       /* Zap PTEs and give away pages in one big
+                        * multicall. */
+                       (void)HYPERVISOR_multicall(np->rx_mcl, i+1);
+
+                       /* Check return status of HYPERVISOR_memory_op(). */
+                       if (unlikely(np->rx_mcl[i].result != i))
+                               panic("Unable to reduce memory reservation\n");
+               } else {
+                       if (HYPERVISOR_memory_op(XENMEM_decrease_reservation,
+                                                &reservation) != i)
+                               panic("Unable to reduce memory reservation\n");
+               }
+       } else {
+               wmb();          /* barrier so backend seens requests */
+       }
+
+       /* Above is a suitable barrier to ensure backend will see requests. */
+       np->rx.req_prod_pvt = req_prod + i;
+ push:
+       RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&np->rx, notify);
+       if (notify)
+               notify_remote_via_irq(np->netdev->irq);
+}
+
+static int xennet_open(struct net_device *dev)
+{
+       struct netfront_info *np = netdev_priv(dev);
+
+       memset(&np->stats, 0, sizeof(np->stats));
+
+       spin_lock_bh(&np->rx_lock);
+       if (netif_carrier_ok(dev)) {
+               xennet_alloc_rx_buffers(dev);
+               np->rx.sring->rsp_event = np->rx.rsp_cons + 1;
+               if (RING_HAS_UNCONSUMED_RESPONSES(&np->rx))
+                       netif_rx_schedule(dev);
+       }
+       spin_unlock_bh(&np->rx_lock);
+
+       xennet_maybe_wake_tx(dev);
+
+       return 0;
+}
+
+static void xennet_tx_buf_gc(struct net_device *dev)
+{
+       RING_IDX cons, prod;
+       unsigned short id;
+       struct netfront_info *np = netdev_priv(dev);
+       struct sk_buff *skb;
+
+       BUG_ON(!netif_carrier_ok(dev));
+
+       do {
+               prod = np->tx.sring->rsp_prod;
+               rmb(); /* Ensure we see responses up to 'rp'. */
+
+               for (cons = np->tx.rsp_cons; cons != prod; cons++) {
+                       struct xen_netif_tx_response *txrsp;
+
+                       txrsp = RING_GET_RESPONSE(&np->tx, cons);
+                       if (txrsp->status == NETIF_RSP_NULL)
+                               continue;
+
+                       id  = txrsp->id;
+                       skb = np->tx_skbs[id].skb;
+                       if (unlikely(gnttab_query_foreign_access(
+                               np->grant_tx_ref[id]) != 0)) {
+                               printk(KERN_ALERT "xennet_tx_buf_gc: warning "
+                                      "-- grant still in use by backend "
+                                      "domain.\n");
+                               BUG();
+                       }
+                       gnttab_end_foreign_access_ref(
+                               np->grant_tx_ref[id], GNTMAP_readonly);
+                       gnttab_release_grant_reference(
+                               &np->gref_tx_head, np->grant_tx_ref[id]);
+                       np->grant_tx_ref[id] = GRANT_INVALID_REF;
+                       add_id_to_freelist(&np->tx_skb_freelist, np->tx_skbs, id);
+                       dev_kfree_skb_irq(skb);
+               }
+
+               np->tx.rsp_cons = prod;
+
+               /*
+                * Set a new event, then check for race with update of tx_cons.
+                * Note that it is essential to schedule a callback, no matter
+                * how few buffers are pending. Even if there is space in the
+                * transmit ring, higher layers may be blocked because too much
+                * data is outstanding: in such cases notification from Xen is
+                * likely to be the only kick that we'll get.
+                */
+               np->tx.sring->rsp_event =
+                       prod + ((np->tx.sring->req_prod - prod) >> 1) + 1;
+               mb();           /* update shared area */
+       } while ((cons == prod) && (prod != np->tx.sring->rsp_prod));
+
+       xennet_maybe_wake_tx(dev);
+}
+
+static void xennet_make_frags(struct sk_buff *skb, struct net_device *dev,
+                             struct xen_netif_tx_request *tx)
+{
+       struct netfront_info *np = netdev_priv(dev);
+       char *data = skb->data;
+       unsigned long mfn;
+       RING_IDX prod = np->tx.req_prod_pvt;
+       int frags = skb_shinfo(skb)->nr_frags;
+       unsigned int offset = offset_in_page(data);
+       unsigned int len = skb_headlen(skb);
+       unsigned int id;
+       grant_ref_t ref;
+       int i;
+
+       /* While the header overlaps a page boundary (including being
+          larger than a page), split it it into page-sized chunks. */
+       while (len > PAGE_SIZE - offset) {
+               tx->size = PAGE_SIZE - offset;
+               tx->flags |= NETTXF_more_data;
+               len -= tx->size;
+               data += tx->size;
+               offset = 0;
+
+               id = get_id_from_freelist(&np->tx_skb_freelist, np->tx_skbs);
+               np->tx_skbs[id].skb = skb_get(skb);
+               tx = RING_GET_REQUEST(&np->tx, prod++);
+               tx->id = id;
+               ref = gnttab_claim_grant_reference(&np->gref_tx_head);
+               BUG_ON((signed short)ref < 0);
+
+               mfn = virt_to_mfn(data);
+               gnttab_grant_foreign_access_ref(ref, np->xbdev->otherend_id,
+                                               mfn, GNTMAP_readonly);
+
+               tx->gref = np->grant_tx_ref[id] = ref;
+               tx->offset = offset;
+               tx->size = len;
+               tx->flags = 0;
+       }
+
+       /* Grant backend access to each skb fragment page. */
+       for (i = 0; i < frags; i++) {
+               skb_frag_t *frag = skb_shinfo(skb)->frags + i;
+
+               tx->flags |= NETTXF_more_data;
+
+               id = get_id_from_freelist(&np->tx_skb_freelist, np->tx_skbs);
+               np->tx_skbs[id].skb = skb_get(skb);
+               tx = RING_GET_REQUEST(&np->tx, prod++);
+               tx->id = id;
+               ref = gnttab_claim_grant_reference(&np->gref_tx_head);
+               BUG_ON((signed short)ref < 0);
+
+               mfn = pfn_to_mfn(page_to_pfn(frag->page));
+               gnttab_grant_foreign_access_ref(ref, np->xbdev->otherend_id,
+                                               mfn, GNTMAP_readonly);
+
+               tx->gref = np->grant_tx_ref[id] = ref;
+               tx->offset = frag->page_offset;
+               tx->size = frag->size;
+               tx->flags = 0;
+       }
+
+       np->tx.req_prod_pvt = prod;
+}
+
+static int xennet_start_xmit(struct sk_buff *skb, struct net_device *dev)
+{
+       unsigned short id;
+       struct netfront_info *np = netdev_priv(dev);
+       struct xen_netif_tx_request *tx;
+       struct xen_netif_extra_info *extra;
+       char *data = skb->data;
+       RING_IDX i;
+       grant_ref_t ref;
+       unsigned long mfn;
+       int notify;
+       int frags = skb_shinfo(skb)->nr_frags;
+       unsigned int offset = offset_in_page(data);
+       unsigned int len = skb_headlen(skb);
+
+       frags += (offset + len + PAGE_SIZE - 1) / PAGE_SIZE;
+       if (unlikely(frags > MAX_SKB_FRAGS + 1)) {
+               printk(KERN_ALERT "xennet: skb rides the rocket: %d frags\n",
+                      frags);
+               dump_stack();
+               goto drop;
+       }
+
+       spin_lock_irq(&np->tx_lock);
+
+       if (unlikely(!netif_carrier_ok(dev) ||
+                    (frags > 1 && !xennet_can_sg(dev)) ||
+                    netif_needs_gso(dev, skb))) {
+               spin_unlock_irq(&np->tx_lock);
+               goto drop;
+       }
+
+       i = np->tx.req_prod_pvt;
+
+       id = get_id_from_freelist(&np->tx_skb_freelist, np->tx_skbs);
+       np->tx_skbs[id].skb = skb;
+
+       tx = RING_GET_REQUEST(&np->tx, i);
+
+       tx->id   = id;
+       ref = gnttab_claim_grant_reference(&np->gref_tx_head);
+       BUG_ON((signed short)ref < 0);
+       mfn = virt_to_mfn(data);
+       gnttab_grant_foreign_access_ref(
+               ref, np->xbdev->otherend_id, mfn, GNTMAP_readonly);
+       tx->gref = np->grant_tx_ref[id] = ref;
+       tx->offset = offset;
+       tx->size = len;
+       extra = NULL;
+
+       tx->flags = 0;
+       if (skb->ip_summed == CHECKSUM_PARTIAL)
+               /* local packet? */
+               tx->flags |= NETTXF_csum_blank | NETTXF_data_validated;
+       else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
+               /* remote but checksummed. */
+               tx->flags |= NETTXF_data_validated;
+
+       if (skb_shinfo(skb)->gso_size) {
+               struct xen_netif_extra_info *gso;
+
+               gso = (struct xen_netif_extra_info *)
+                       RING_GET_REQUEST(&np->tx, ++i);
+
+               if (extra)
+                       extra->flags |= XEN_NETIF_EXTRA_FLAG_MORE;
+               else
+                       tx->flags |= NETTXF_extra_info;
+
+               gso->u.gso.size = skb_shinfo(skb)->gso_size;
+               gso->u.gso.type = XEN_NETIF_GSO_TYPE_TCPV4;
+               gso->u.gso.pad = 0;
+               gso->u.gso.features = 0;
+
+               gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
+               gso->flags = 0;
+               extra = gso;
+       }
+
+       np->tx.req_prod_pvt = i + 1;
+
+       xennet_make_frags(skb, dev, tx);
+       tx->size = skb->len;
+
+       RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&np->tx, notify);
+       if (notify)
+               notify_remote_via_irq(np->netdev->irq);
+
+       xennet_tx_buf_gc(dev);
+
+       if (!netfront_tx_slot_available(np))
+               netif_stop_queue(dev);
+
+       spin_unlock_irq(&np->tx_lock);
+
+       np->stats.tx_bytes += skb->len;
+       np->stats.tx_packets++;
+
+       return 0;
+
+ drop:
+       np->stats.tx_dropped++;
+       dev_kfree_skb(skb);
+       return 0;
+}
+
+static int xennet_close(struct net_device *dev)
+{
+       struct netfront_info *np = netdev_priv(dev);
+       netif_stop_queue(np->netdev);
+       return 0;
+}
+
+static struct net_device_stats *xennet_get_stats(struct net_device *dev)
+{
+       struct netfront_info *np = netdev_priv(dev);
+       return &np->stats;
+}
+
+static void xennet_move_rx_slot(struct netfront_info *np, struct sk_buff *skb,
+                               grant_ref_t ref)
+{
+       int new = xennet_rxidx(np->rx.req_prod_pvt);
+
+       BUG_ON(np->rx_skbs[new]);
+       np->rx_skbs[new] = skb;
+       np->grant_rx_ref[new] = ref;
+       RING_GET_REQUEST(&np->rx, np->rx.req_prod_pvt)->id = new;
+       RING_GET_REQUEST(&np->rx, np->rx.req_prod_pvt)->gref = ref;
+       np->rx.req_prod_pvt++;
+}
+
+static int xennet_get_extras(struct netfront_info *np,
+                            struct xen_netif_extra_info *extras,
+                            RING_IDX rp)
+
+{
+       struct xen_netif_extra_info *extra;
+       struct device *dev = &np->netdev->dev;
+       RING_IDX cons = np->rx.rsp_cons;
+       int err = 0;
+
+       do {
+               struct sk_buff *skb;
+               grant_ref_t ref;
+
+               if (unlikely(cons + 1 == rp)) {
+                       if (net_ratelimit())
+                               dev_warn(dev, "Missing extra info\n");
+                       err = -EBADR;
+                       break;
+               }
+
+               extra = (struct xen_netif_extra_info *)
+                       RING_GET_RESPONSE(&np->rx, ++cons);
+
+               if (unlikely(!extra->type ||
+                            extra->type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
+                       if (net_ratelimit())
+                               dev_warn(dev, "Invalid extra type: %d\n",
+                                       extra->type);
+                       err = -EINVAL;
+               } else {
+                       memcpy(&extras[extra->type - 1], extra,
+                              sizeof(*extra));
+               }
+
+               skb = xennet_get_rx_skb(np, cons);
+               ref = xennet_get_rx_ref(np, cons);
+               xennet_move_rx_slot(np, skb, ref);
+       } while (extra->flags & XEN_NETIF_EXTRA_FLAG_MORE);
+
+       np->rx.rsp_cons = cons;
+       return err;
+}
+
+static int xennet_get_responses(struct netfront_info *np,
+                               struct netfront_rx_info *rinfo, RING_IDX rp,
+                               struct sk_buff_head *list)
+{
+       struct xen_netif_rx_response *rx = &rinfo->rx;
+       struct xen_netif_extra_info *extras = rinfo->extras;
+       struct device *dev = &np->netdev->dev;
+       RING_IDX cons = np->rx.rsp_cons;
+       struct sk_buff *skb = xennet_get_rx_skb(np, cons);
+       grant_ref_t ref = xennet_get_rx_ref(np, cons);
+       int max = MAX_SKB_FRAGS + (rx->status <= RX_COPY_THRESHOLD);
+       int frags = 1;
+       int err = 0;
+       unsigned long ret;
+
+       if (rx->flags & NETRXF_extra_info) {
+               err = xennet_get_extras(np, extras, rp);
+               cons = np->rx.rsp_cons;
+       }
+
+       for (;;) {
+               if (unlikely(rx->status < 0 ||
+                            rx->offset + rx->status > PAGE_SIZE)) {
+                       if (net_ratelimit())
+                               dev_warn(dev, "rx->offset: %x, size: %u\n",
+                                        rx->offset, rx->status);
+                       xennet_move_rx_slot(np, skb, ref);
+                       err = -EINVAL;
+                       goto next;
+               }
+
+               /*
+                * This definitely indicates a bug, either in this driver or in
+                * the backend driver. In future this should flag the bad
+                * situation to the system controller to reboot the backed.
+                */
+               if (ref == GRANT_INVALID_REF) {
+                       if (net_ratelimit())
+                               dev_warn(dev, "Bad rx response id %d.\n",
+                                        rx->id);
+                       err = -EINVAL;
+                       goto next;
+               }
+
+               ret = gnttab_end_foreign_access_ref(ref, 0);
+               BUG_ON(!ret);
+
+               gnttab_release_grant_reference(&np->gref_rx_head, ref);
+
+               __skb_queue_tail(list, skb);
+
+next:
+               if (!(rx->flags & NETRXF_more_data))
+                       break;
+
+               if (cons + frags == rp) {
+                       if (net_ratelimit())
+                               dev_warn(dev, "Need more frags\n");
+                       err = -ENOENT;
+                       break;
+               }
+
+               rx = RING_GET_RESPONSE(&np->rx, cons + frags);
+               skb = xennet_get_rx_skb(np, cons + frags);
+               ref = xennet_get_rx_ref(np, cons + frags);
+               frags++;
+       }
+
+       if (unlikely(frags > max)) {
+               if (net_ratelimit())
+                       dev_warn(dev, "Too many frags\n");
+               err = -E2BIG;
+       }
+
+       if (unlikely(err))
+               np->rx.rsp_cons = cons + frags;
+
+       return err;
+}
+
+static int xennet_set_skb_gso(struct sk_buff *skb,
+                             struct xen_netif_extra_info *gso)
+{
+       if (!gso->u.gso.size) {
+               if (net_ratelimit())
+                       printk(KERN_WARNING "GSO size must not be zero.\n");
+               return -EINVAL;
+       }
+
+       /* Currently only TCPv4 S.O. is supported. */
+       if (gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV4) {
+               if (net_ratelimit())
+                       printk(KERN_WARNING "Bad GSO type %d.\n", gso->u.gso.type);
+               return -EINVAL;
+       }
+
+       skb_shinfo(skb)->gso_size = gso->u.gso.size;
+       skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
+
+       /* Header must be checked, and gso_segs computed. */
+       skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
+       skb_shinfo(skb)->gso_segs = 0;
+
+       return 0;
+}
+
+static RING_IDX xennet_fill_frags(struct netfront_info *np,
+                                 struct sk_buff *skb,
+                                 struct sk_buff_head *list)
+{
+       struct skb_shared_info *shinfo = skb_shinfo(skb);
+       int nr_frags = shinfo->nr_frags;
+       RING_IDX cons = np->rx.rsp_cons;
+       skb_frag_t *frag = shinfo->frags + nr_frags;
+       struct sk_buff *nskb;
+
+       while ((nskb = __skb_dequeue(list))) {
+               struct xen_netif_rx_response *rx =
+                       RING_GET_RESPONSE(&np->rx, ++cons);
+
+               frag->page = skb_shinfo(nskb)->frags[0].page;
+               frag->page_offset = rx->offset;
+               frag->size = rx->status;
+
+               skb->data_len += rx->status;
+
+               skb_shinfo(nskb)->nr_frags = 0;
+               kfree_skb(nskb);
+
+               frag++;
+               nr_frags++;
+       }
+
+       shinfo->nr_frags = nr_frags;
+       return cons;
+}
+
+static int skb_checksum_setup(struct sk_buff *skb)
+{
+       struct iphdr *iph;
+       unsigned char *th;
+       int err = -EPROTO;
+
+       if (skb->protocol != htons(ETH_P_IP))
+               goto out;
+
+       iph = (void *)skb->data;
+       th = skb->data + 4 * iph->ihl;
+       if (th >= skb_tail_pointer(skb))
+               goto out;
+
+       skb->csum_start = th - skb->head;
+       switch (iph->protocol) {
+       case IPPROTO_TCP:
+               skb->csum_offset = offsetof(struct tcphdr, check);
+               break;
+       case IPPROTO_UDP:
+               skb->csum_offset = offsetof(struct udphdr, check);
+               break;
+       default:
+               if (net_ratelimit())
+                       printk(KERN_ERR "Attempting to checksum a non-"
+                              "TCP/UDP packet, dropping a protocol"
+                              " %d packet", iph->protocol);
+               goto out;
+       }
+
+       if ((th + skb->csum_offset + 2) > skb_tail_pointer(skb))
+               goto out;
+
+       err = 0;
+
+out:
+       return err;
+}
+
+static int handle_incoming_queue(struct net_device *dev,
+                                 struct sk_buff_head *rxq)
+{
+       struct netfront_info *np = netdev_priv(dev);
+       int packets_dropped = 0;
+       struct sk_buff *skb;
+
+       while ((skb = __skb_dequeue(rxq)) != NULL) {
+               struct page *page = NETFRONT_SKB_CB(skb)->page;
+               void *vaddr = page_address(page);
+               unsigned offset = NETFRONT_SKB_CB(skb)->offset;
+
+               memcpy(skb->data, vaddr + offset,
+                      skb_headlen(skb));
+
+               if (page != skb_shinfo(skb)->frags[0].page)
+                       __free_page(page);
+
+               /* Ethernet work: Delayed to here as it peeks the header. */
+               skb->protocol = eth_type_trans(skb, dev);
+
+               if (skb->ip_summed == CHECKSUM_PARTIAL) {
+                       if (skb_checksum_setup(skb)) {
+                               kfree_skb(skb);
+                               packets_dropped++;
+                               np->stats.rx_errors++;
+                               continue;
+                       }
+               }
+
+               np->stats.rx_packets++;
+               np->stats.rx_bytes += skb->len;
+
+               /* Pass it up. */
+               netif_receive_skb(skb);
+               dev->last_rx = jiffies;
+       }
+
+       return packets_dropped;
+}
+
+static int xennet_poll(struct net_device *dev, int *pbudget)
+{
+       struct netfront_info *np = netdev_priv(dev);
+       struct sk_buff *skb;
+       struct netfront_rx_info rinfo;
+       struct xen_netif_rx_response *rx = &rinfo.rx;
+       struct xen_netif_extra_info *extras = rinfo.extras;
+       RING_IDX i, rp;
+       int work_done, budget, more_to_do = 1;
+       struct sk_buff_head rxq;
+       struct sk_buff_head errq;
+       struct sk_buff_head tmpq;
+       unsigned long flags;
+       unsigned int len;
+       int err;
+
+       spin_lock(&np->rx_lock);
+
+       if (unlikely(!netif_carrier_ok(dev))) {
+               spin_unlock(&np->rx_lock);
+               return 0;
+       }
+
+       skb_queue_head_init(&rxq);
+       skb_queue_head_init(&errq);
+       skb_queue_head_init(&tmpq);
+
+       budget = *pbudget;
+       if (budget > dev->quota)
+               budget = dev->quota;
+       rp = np->rx.sring->rsp_prod;
+       rmb(); /* Ensure we see queued responses up to 'rp'. */
+
+       i = np->rx.rsp_cons;
+       work_done = 0;
+       while ((i != rp) && (work_done < budget)) {
+               memcpy(rx, RING_GET_RESPONSE(&np->rx, i), sizeof(*rx));
+               memset(extras, 0, sizeof(rinfo.extras));
+
+               err = xennet_get_responses(np, &rinfo, rp, &tmpq);
+
+               if (unlikely(err)) {
+err:
+                       while ((skb = __skb_dequeue(&tmpq)))
+                               __skb_queue_tail(&errq, skb);
+                       np->stats.rx_errors++;
+                       i = np->rx.rsp_cons;
+                       continue;
+               }
+
+               skb = __skb_dequeue(&tmpq);
+
+               if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
+                       struct xen_netif_extra_info *gso;
+                       gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
+
+                       if (unlikely(xennet_set_skb_gso(skb, gso))) {
+                               __skb_queue_head(&tmpq, skb);
+                               np->rx.rsp_cons += skb_queue_len(&tmpq);
+                               goto err;
+                       }
+               }
+
+               NETFRONT_SKB_CB(skb)->page = skb_shinfo(skb)->frags[0].page;
+               NETFRONT_SKB_CB(skb)->offset = rx->offset;
+
+               len = rx->status;
+               if (len > RX_COPY_THRESHOLD)
+                       len = RX_COPY_THRESHOLD;
+               skb_put(skb, len);
+
+               if (rx->status > len) {
+                       skb_shinfo(skb)->frags[0].page_offset =
+                               rx->offset + len;
+                       skb_shinfo(skb)->frags[0].size = rx->status - len;
+                       skb->data_len = rx->status - len;
+               } else {
+                       skb_shinfo(skb)->frags[0].page = NULL;
+                       skb_shinfo(skb)->nr_frags = 0;
+               }
+
+               i = xennet_fill_frags(np, skb, &tmpq);
+
+               /*
+                * Truesize approximates the size of true data plus
+                * any supervisor overheads. Adding hypervisor
+                * overheads has been shown to significantly reduce
+                * achievable bandwidth with the default receive
+                * buffer size. It is therefore not wise to account
+                * for it here.
+                *
+                * After alloc_skb(RX_COPY_THRESHOLD), truesize is set
+                * to RX_COPY_THRESHOLD + the supervisor
+                * overheads. Here, we add the size of the data pulled
+                * in xennet_fill_frags().
+                *
+                * We also adjust for any unused space in the main
+                * data area by subtracting (RX_COPY_THRESHOLD -
+                * len). This is especially important with drivers
+                * which split incoming packets into header and data,
+                * using only 66 bytes of the main data area (see the
+                * e1000 driver for example.)  On such systems,
+                * without this last adjustement, our achievable
+                * receive throughout using the standard receive
+                * buffer size was cut by 25%(!!!).
+                */
+               skb->truesize += skb->data_len - (RX_COPY_THRESHOLD - len);
+               skb->len += skb->data_len;
+
+               if (rx->flags & NETRXF_csum_blank)
+                       skb->ip_summed = CHECKSUM_PARTIAL;
+               else if (rx->flags & NETRXF_data_validated)
+                       skb->ip_summed = CHECKSUM_UNNECESSARY;
+
+               __skb_queue_tail(&rxq, skb);
+
+               np->rx.rsp_cons = ++i;
+               work_done++;
+       }
+
+       while ((skb = __skb_dequeue(&errq)))
+               kfree_skb(skb);
+
+       work_done -= handle_incoming_queue(dev, &rxq);
+
+       /* If we get a callback with very few responses, reduce fill target. */
+       /* NB. Note exponential increase, linear decrease. */
+       if (((np->rx.req_prod_pvt - np->rx.sring->rsp_prod) >
+            ((3*np->rx_target) / 4)) &&
+           (--np->rx_target < np->rx_min_target))
+               np->rx_target = np->rx_min_target;
+
+       xennet_alloc_rx_buffers(dev);
+
+       *pbudget   -= work_done;
+       dev->quota -= work_done;
+
+       if (work_done < budget) {
+               local_irq_save(flags);
+
+               RING_FINAL_CHECK_FOR_RESPONSES(&np->rx, more_to_do);
+               if (!more_to_do)
+                       __netif_rx_complete(dev);
+
+               local_irq_restore(flags);
+       }
+
+       spin_unlock(&np->rx_lock);
+
+       return more_to_do;
+}
+
+static int xennet_change_mtu(struct net_device *dev, int mtu)
+{
+       int max = xennet_can_sg(dev) ? 65535 - ETH_HLEN : ETH_DATA_LEN;
+
+       if (mtu > max)
+               return -EINVAL;
+       dev->mtu = mtu;
+       return 0;
+}
+
+static void xennet_release_tx_bufs(struct netfront_info *np)
+{
+       struct sk_buff *skb;
+       int i;
+
+       for (i = 0; i < NET_TX_RING_SIZE; i++) {
+               /* Skip over entries which are actually freelist references */
+               if ((unsigned long)np->tx_skbs[i].skb < PAGE_OFFSET)
+                       continue;
+
+               skb = np->tx_skbs[i].skb;
+               gnttab_end_foreign_access_ref(np->grant_tx_ref[i],
+                                             GNTMAP_readonly);
+               gnttab_release_grant_reference(&np->gref_tx_head,
+                                              np->grant_tx_ref[i]);
+               np->grant_tx_ref[i] = GRANT_INVALID_REF;
+               add_id_to_freelist(&np->tx_skb_freelist, np->tx_skbs, i);
+               dev_kfree_skb_irq(skb);
+       }
+}
+
+static void xennet_release_rx_bufs(struct netfront_info *np)
+{
+       struct mmu_update      *mmu = np->rx_mmu;
+       struct multicall_entry *mcl = np->rx_mcl;
+       struct sk_buff_head free_list;
+       struct sk_buff *skb;
+       unsigned long mfn;
+       int xfer = 0, noxfer = 0, unused = 0;
+       int id, ref;
+
+       dev_warn(&np->netdev->dev, "%s: fix me for copying receiver.\n",
+                        __func__);
+       return;
+
+       skb_queue_head_init(&free_list);
+
+       spin_lock_bh(&np->rx_lock);
+
+       for (id = 0; id < NET_RX_RING_SIZE; id++) {
+               ref = np->grant_rx_ref[id];
+               if (ref == GRANT_INVALID_REF) {
+                       unused++;
+                       continue;
+               }
+
+               skb = np->rx_skbs[id];
+               mfn = gnttab_end_foreign_transfer_ref(ref);
+               gnttab_release_grant_reference(&np->gref_rx_head, ref);
+               np->grant_rx_ref[id] = GRANT_INVALID_REF;
+
+               if (0 == mfn) {
+                       skb_shinfo(skb)->nr_frags = 0;
+                       dev_kfree_skb(skb);
+                       noxfer++;
+                       continue;
+               }
+
+               if (!xen_feature(XENFEAT_auto_translated_physmap)) {
+                       /* Remap the page. */
+                       struct page *page = skb_shinfo(skb)->frags[0].page;
+                       unsigned long pfn = page_to_pfn(page);
+                       void *vaddr = page_address(page);
+
+                       MULTI_update_va_mapping(mcl, (unsigned long)vaddr,
+                                               mfn_pte(mfn, PAGE_KERNEL),
+                                               0);
+                       mcl++;
+                       mmu->ptr = ((u64)mfn << PAGE_SHIFT)
+                               | MMU_MACHPHYS_UPDATE;
+                       mmu->val = pfn;
+                       mmu++;
+
+                       set_phys_to_machine(pfn, mfn);
+               }
+               __skb_queue_tail(&free_list, skb);
+               xfer++;
+       }
+
+       dev_info(&np->netdev->dev, "%s: %d xfer, %d noxfer, %d unused\n",
+                __func__, xfer, noxfer, unused);
+
+       if (xfer) {
+               if (!xen_feature(XENFEAT_auto_translated_physmap)) {
+                       /* Do all the remapping work and M2P updates. */
+                       MULTI_mmu_update(mcl, np->rx_mmu, mmu - np->rx_mmu,
+                                        0, DOMID_SELF);
+                       mcl++;
+                       HYPERVISOR_multicall(np->rx_mcl, mcl - np->rx_mcl);
+               }
+       }
+
+       while ((skb = __skb_dequeue(&free_list)) != NULL)
+               dev_kfree_skb(skb);
+
+       spin_unlock_bh(&np->rx_lock);
+}
+
+static void xennet_uninit(struct net_device *dev)
+{
+       struct netfront_info *np = netdev_priv(dev);
+       xennet_release_tx_bufs(np);
+       xennet_release_rx_bufs(np);
+       gnttab_free_grant_references(np->gref_tx_head);
+       gnttab_free_grant_references(np->gref_rx_head);
+}
+
+static struct net_device * __devinit xennet_create_dev(struct xenbus_device *dev)
+{
+       int i, err;
+       struct net_device *netdev;
+       struct netfront_info *np;
+
+       netdev = alloc_etherdev(sizeof(struct netfront_info));
+       if (!netdev) {
+               printk(KERN_WARNING "%s> alloc_etherdev failed.\n",
+                      __func__);
+               return ERR_PTR(-ENOMEM);
+       }
+
+       np                   = netdev_priv(netdev);
+       np->xbdev            = dev;
+
+       spin_lock_init(&np->tx_lock);
+       spin_lock_init(&np->rx_lock);
+
+       skb_queue_head_init(&np->rx_batch);
+       np->rx_target     = RX_DFL_MIN_TARGET;
+       np->rx_min_target = RX_DFL_MIN_TARGET;
+       np->rx_max_target = RX_MAX_TARGET;
+
+       init_timer(&np->rx_refill_timer);
+       np->rx_refill_timer.data = (unsigned long)netdev;
+       np->rx_refill_timer.function = rx_refill_timeout;
+
+       /* Initialise tx_skbs as a free chain containing every entry. */
+       np->tx_skb_freelist = 0;
+       for (i = 0; i < NET_TX_RING_SIZE; i++) {
+               np->tx_skbs[i].link = i+1;
+               np->grant_tx_ref[i] = GRANT_INVALID_REF;
+       }
+
+       /* Clear out rx_skbs */
+       for (i = 0; i < NET_RX_RING_SIZE; i++) {
+               np->rx_skbs[i] = NULL;
+               np->grant_rx_ref[i] = GRANT_INVALID_REF;
+       }
+
+       /* A grant for every tx ring slot */
+       if (gnttab_alloc_grant_references(TX_MAX_TARGET,
+                                         &np->gref_tx_head) < 0) {
+               printk(KERN_ALERT "#### netfront can't alloc tx grant refs\n");
+               err = -ENOMEM;
+               goto exit;
+       }
+       /* A grant for every rx ring slot */
+       if (gnttab_alloc_grant_references(RX_MAX_TARGET,
+                                         &np->gref_rx_head) < 0) {
+               printk(KERN_ALERT "#### netfront can't alloc rx grant refs\n");
+               err = -ENOMEM;
+               goto exit_free_tx;
+       }
+
+       netdev->open            = xennet_open;
+       netdev->hard_start_xmit = xennet_start_xmit;
+       netdev->stop            = xennet_close;
+       netdev->get_stats       = xennet_get_stats;
+       netdev->poll            = xennet_poll;
+       netdev->uninit          = xennet_uninit;
+       netdev->change_mtu      = xennet_change_mtu;
+       netdev->weight          = 64;
+       netdev->features        = NETIF_F_IP_CSUM;
+
+       SET_ETHTOOL_OPS(netdev, &xennet_ethtool_ops);
+       SET_MODULE_OWNER(netdev);
+       SET_NETDEV_DEV(netdev, &dev->dev);
+
+       np->netdev = netdev;
+
+       netif_carrier_off(netdev);
+
+       return netdev;
+
+ exit_free_tx:
+       gnttab_free_grant_references(np->gref_tx_head);
+ exit:
+       free_netdev(netdev);
+       return ERR_PTR(err);
+}
+
+/**
+ * Entry point to this code when a new device is created.  Allocate the basic
+ * structures and the ring buffers for communication with the backend, and
+ * inform the backend of the appropriate details for those.
+ */
+static int __devinit netfront_probe(struct xenbus_device *dev,
+                                   const struct xenbus_device_id *id)
+{
+       int err;
+       struct net_device *netdev;
+       struct netfront_info *info;
+
+       netdev = xennet_create_dev(dev);
+       if (IS_ERR(netdev)) {
+               err = PTR_ERR(netdev);
+               xenbus_dev_fatal(dev, err, "creating netdev");
+               return err;
+       }
+
+       info = netdev_priv(netdev);
+       dev->dev.driver_data = info;
+
+       err = register_netdev(info->netdev);
+       if (err) {
+               printk(KERN_WARNING "%s: register_netdev err=%d\n",
+                      __func__, err);
+               goto fail;
+       }
+
+       err = xennet_sysfs_addif(info->netdev);
+       if (err) {
+               unregister_netdev(info->netdev);
+               printk(KERN_WARNING "%s: add sysfs failed err=%d\n",
+                      __func__, err);
+               goto fail;
+       }
+
+       return 0;
+
+ fail:
+       free_netdev(netdev);
+       dev->dev.driver_data = NULL;
+       return err;
+}
+
+static void xennet_end_access(int ref, void *page)
+{
+       /* This frees the page as a side-effect */
+       if (ref != GRANT_INVALID_REF)
+               gnttab_end_foreign_access(ref, 0, (unsigned long)page);
+}
+
+static void xennet_disconnect_backend(struct netfront_info *info)
+{
+       /* Stop old i/f to prevent errors whilst we rebuild the state. */
+       spin_lock_bh(&info->rx_lock);
+       spin_lock_irq(&info->tx_lock);
+       netif_carrier_off(info->netdev);
+       spin_unlock_irq(&info->tx_lock);
+       spin_unlock_bh(&info->rx_lock);
+
+       if (info->netdev->irq)
+               unbind_from_irqhandler(info->netdev->irq, info->netdev);
+       info->evtchn = info->netdev->irq = 0;
+
+       /* End access and free the pages */
+       xennet_end_access(info->tx_ring_ref, info->tx.sring);
+       xennet_end_access(info->rx_ring_ref, info->rx.sring);
+
+       info->tx_ring_ref = GRANT_INVALID_REF;
+       info->rx_ring_ref = GRANT_INVALID_REF;
+       info->tx.sring = NULL;
+       info->rx.sring = NULL;
+}
+
+/**
+ * We are reconnecting to the backend, due to a suspend/resume, or a backend
+ * driver restart.  We tear down our netif structure and recreate it, but
+ * leave the device-layer structures intact so that this is transparent to the
+ * rest of the kernel.
+ */
+static int netfront_resume(struct xenbus_device *dev)
+{
+       struct netfront_info *info = dev->dev.driver_data;
+
+       dev_dbg(&dev->dev, "%s\n", dev->nodename);
+
+       xennet_disconnect_backend(info);
+       return 0;
+}
+
+static int xen_net_read_mac(struct xenbus_device *dev, u8 mac[])
+{
+       char *s, *e, *macstr;
+       int i;
+
+       macstr = s = xenbus_read(XBT_NIL, dev->nodename, "mac", NULL);
+       if (IS_ERR(macstr))
+               return PTR_ERR(macstr);
+
+       for (i = 0; i < ETH_ALEN; i++) {
+               mac[i] = simple_strtoul(s, &e, 16);
+               if ((s == e) || (*e != ((i == ETH_ALEN-1) ? '\0' : ':'))) {
+                       kfree(macstr);
+                       return -ENOENT;
+               }
+               s = e+1;
+       }
+
+       kfree(macstr);
+       return 0;
+}
+
+static irqreturn_t xennet_interrupt(int irq, void *dev_id)
+{
+       struct net_device *dev = dev_id;
+       struct netfront_info *np = netdev_priv(dev);
+       unsigned long flags;
+
+       spin_lock_irqsave(&np->tx_lock, flags);
+
+       if (likely(netif_carrier_ok(dev))) {
+               xennet_tx_buf_gc(dev);
+               /* Under tx_lock: protects access to rx shared-ring indexes. */
+               if (RING_HAS_UNCONSUMED_RESPONSES(&np->rx))
+                       netif_rx_schedule(dev);
+       }
+
+       spin_unlock_irqrestore(&np->tx_lock, flags);
+
+       return IRQ_HANDLED;
+}
+
+static int setup_netfront(struct xenbus_device *dev, struct netfront_info *info)
+{
+       struct xen_netif_tx_sring *txs;
+       struct xen_netif_rx_sring *rxs;
+       int err;
+       struct net_device *netdev = info->netdev;
+
+       info->tx_ring_ref = GRANT_INVALID_REF;
+       info->rx_ring_ref = GRANT_INVALID_REF;
+       info->rx.sring = NULL;
+       info->tx.sring = NULL;
+       netdev->irq = 0;
+
+       err = xen_net_read_mac(dev, netdev->dev_addr);
+       if (err) {
+               xenbus_dev_fatal(dev, err, "parsing %s/mac", dev->nodename);
+               goto fail;
+       }
+
+       txs = (struct xen_netif_tx_sring *)get_zeroed_page(GFP_KERNEL);
+       if (!txs) {
+               err = -ENOMEM;
+               xenbus_dev_fatal(dev, err, "allocating tx ring page");
+               goto fail;
+       }
+       SHARED_RING_INIT(txs);
+       FRONT_RING_INIT(&info->tx, txs, PAGE_SIZE);
+
+       err = xenbus_grant_ring(dev, virt_to_mfn(txs));
+       if (err < 0) {
+               free_page((unsigned long)txs);
+               goto fail;
+       }
+
+       info->tx_ring_ref = err;
+       rxs = (struct xen_netif_rx_sring *)get_zeroed_page(GFP_KERNEL);
+       if (!rxs) {
+               err = -ENOMEM;
+               xenbus_dev_fatal(dev, err, "allocating rx ring page");
+               goto fail;
+       }
+       SHARED_RING_INIT(rxs);
+       FRONT_RING_INIT(&info->rx, rxs, PAGE_SIZE);
+
+       err = xenbus_grant_ring(dev, virt_to_mfn(rxs));
+       if (err < 0) {
+               free_page((unsigned long)rxs);
+               goto fail;
+       }
+       info->rx_ring_ref = err;
+
+       err = xenbus_alloc_evtchn(dev, &info->evtchn);
+       if (err)
+               goto fail;
+
+       err = bind_evtchn_to_irqhandler(info->evtchn, xennet_interrupt,
+                                       IRQF_SAMPLE_RANDOM, netdev->name,
+                                       netdev);
+       if (err < 0)
+               goto fail;
+       netdev->irq = err;
+       return 0;
+
+ fail:
+       return err;
+}
+
+/* Common code used when first setting up, and when resuming. */
+static int talk_to_backend(struct xenbus_device *dev,
+                          struct netfront_info *info)
+{
+       const char *message;
+       struct xenbus_transaction xbt;
+       int err;
+
+       /* Create shared ring, alloc event channel. */
+       err = setup_netfront(dev, info);
+       if (err)
+               goto out;
+
+again:
+       err = xenbus_transaction_start(&xbt);
+       if (err) {
+               xenbus_dev_fatal(dev, err, "starting transaction");
+               goto destroy_ring;
+       }
+
+       err = xenbus_printf(xbt, dev->nodename, "tx-ring-ref", "%u",
+                           info->tx_ring_ref);
+       if (err) {
+               message = "writing tx ring-ref";
+               goto abort_transaction;
+       }
+       err = xenbus_printf(xbt, dev->nodename, "rx-ring-ref", "%u",
+                           info->rx_ring_ref);
+       if (err) {
+               message = "writing rx ring-ref";
+               goto abort_transaction;
+       }
+       err = xenbus_printf(xbt, dev->nodename,
+                           "event-channel", "%u", info->evtchn);
+       if (err) {
+               message = "writing event-channel";
+               goto abort_transaction;
+       }
+
+       err = xenbus_printf(xbt, dev->nodename, "request-rx-copy", "%u",
+                           1);
+       if (err) {
+               message = "writing request-rx-copy";
+               goto abort_transaction;
+       }
+
+       err = xenbus_printf(xbt, dev->nodename, "feature-rx-notify", "%d", 1);
+       if (err) {
+               message = "writing feature-rx-notify";
+               goto abort_transaction;
+       }
+
+       err = xenbus_printf(xbt, dev->nodename, "feature-sg", "%d", 1);
+       if (err) {
+               message = "writing feature-sg";
+               goto abort_transaction;
+       }
+
+       err = xenbus_printf(xbt, dev->nodename, "feature-gso-tcpv4", "%d", 1);
+       if (err) {
+               message = "writing feature-gso-tcpv4";
+               goto abort_transaction;
+       }
+
+       err = xenbus_transaction_end(xbt, 0);
+       if (err) {
+               if (err == -EAGAIN)
+                       goto again;
+               xenbus_dev_fatal(dev, err, "completing transaction");
+               goto destroy_ring;
+       }
+
+       return 0;
+
+ abort_transaction:
+       xenbus_transaction_end(xbt, 1);
+       xenbus_dev_fatal(dev, err, "%s", message);
+ destroy_ring:
+       xennet_disconnect_backend(info);
+ out:
+       return err;
+}
+
+static int xennet_set_sg(struct net_device *dev, u32 data)
+{
+       if (data) {
+               struct netfront_info *np = netdev_priv(dev);
+               int val;
+
+               if (xenbus_scanf(XBT_NIL, np->xbdev->otherend, "feature-sg",
+                                "%d", &val) < 0)
+                       val = 0;
+               if (!val)
+                       return -ENOSYS;
+       } else if (dev->mtu > ETH_DATA_LEN)
+               dev->mtu = ETH_DATA_LEN;
+
+       return ethtool_op_set_sg(dev, data);
+}
+
+static int xennet_set_tso(struct net_device *dev, u32 data)
+{
+       if (data) {
+               struct netfront_info *np = netdev_priv(dev);
+               int val;
+
+               if (xenbus_scanf(XBT_NIL, np->xbdev->otherend,
+                                "feature-gso-tcpv4", "%d", &val) < 0)
+                       val = 0;
+               if (!val)
+                       return -ENOSYS;
+       }
+
+       return ethtool_op_set_tso(dev, data);
+}
+
+static void xennet_set_features(struct net_device *dev)
+{
+       /* Turn off all GSO bits except ROBUST. */
+       dev->features &= (1 << NETIF_F_GSO_SHIFT) - 1;
+       dev->features |= NETIF_F_GSO_ROBUST;
+       xennet_set_sg(dev, 0);
+
+       /* We need checksum offload to enable scatter/gather and TSO. */
+       if (!(dev->features & NETIF_F_IP_CSUM))
+               return;
+
+       if (!xennet_set_sg(dev, 1))
+               xennet_set_tso(dev, 1);
+}
+
+static int xennet_connect(struct net_device *dev)
+{
+       struct netfront_info *np = netdev_priv(dev);
+       int i, requeue_idx, err;
+       struct sk_buff *skb;
+       grant_ref_t ref;
+       struct xen_netif_rx_request *req;
+       unsigned int feature_rx_copy;
+
+       err = xenbus_scanf(XBT_NIL, np->xbdev->otherend,
+                          "feature-rx-copy", "%u", &feature_rx_copy);
+       if (err != 1)
+               feature_rx_copy = 0;
+
+       if (!feature_rx_copy) {
+               dev_info(&dev->dev,
+                        "backend does not support copying recieve path");
+               return -ENODEV;
+       }
+
+       err = talk_to_backend(np->xbdev, np);
+       if (err)
+               return err;
+
+       xennet_set_features(dev);
+
+       spin_lock_bh(&np->rx_lock);
+       spin_lock_irq(&np->tx_lock);
+
+       /* Step 1: Discard all pending TX packet fragments. */
+       xennet_release_tx_bufs(np);
+
+       /* Step 2: Rebuild the RX buffer freelist and the RX ring itself. */
+       for (requeue_idx = 0, i = 0; i < NET_RX_RING_SIZE; i++) {
+               if (!np->rx_skbs[i])
+                       continue;
+
+               skb = np->rx_skbs[requeue_idx] = xennet_get_rx_skb(np, i);
+               ref = np->grant_rx_ref[requeue_idx] = xennet_get_rx_ref(np, i);
+               req = RING_GET_REQUEST(&np->rx, requeue_idx);
+
+               gnttab_grant_foreign_access_ref(
+                       ref, np->xbdev->otherend_id,
+                       pfn_to_mfn(page_to_pfn(skb_shinfo(skb)->
+                                              frags->page)),
+                       0);
+               req->gref = ref;
+               req->id   = requeue_idx;
+
+               requeue_idx++;
+       }
+
+       np->rx.req_prod_pvt = requeue_idx;
+
+       /*
+        * Step 3: All public and private state should now be sane.  Get
+        * ready to start sending and receiving packets and give the driver
+        * domain a kick because we've probably just requeued some
+        * packets.
+        */
+       netif_carrier_on(np->netdev);
+       notify_remote_via_irq(np->netdev->irq);
+       xennet_tx_buf_gc(dev);
+       xennet_alloc_rx_buffers(dev);
+
+       spin_unlock_irq(&np->tx_lock);
+       spin_unlock_bh(&np->rx_lock);
+
+       return 0;
+}
+
+/**
+ * Callback received when the backend's state changes.
+ */
+static void backend_changed(struct xenbus_device *dev,
+                           enum xenbus_state backend_state)
+{
+       struct netfront_info *np = dev->dev.driver_data;
+       struct net_device *netdev = np->netdev;
+
+       dev_dbg(&dev->dev, "%s\n", xenbus_strstate(backend_state));
+
+       switch (backend_state) {
+       case XenbusStateInitialising:
+       case XenbusStateInitialised:
+       case XenbusStateConnected:
+       case XenbusStateUnknown:
+       case XenbusStateClosed:
+               break;
+
+       case XenbusStateInitWait:
+               if (dev->state != XenbusStateInitialising)
+                       break;
+               if (xennet_connect(netdev) != 0)
+                       break;
+               xenbus_switch_state(dev, XenbusStateConnected);
+               break;
+
+       case XenbusStateClosing:
+               xenbus_frontend_closed(dev);
+               break;
+       }
+}
+
+static struct ethtool_ops xennet_ethtool_ops =
+{
+       .get_tx_csum = ethtool_op_get_tx_csum,
+       .set_tx_csum = ethtool_op_set_tx_csum,
+       .get_sg = ethtool_op_get_sg,
+       .set_sg = xennet_set_sg,
+       .get_tso = ethtool_op_get_tso,
+       .set_tso = xennet_set_tso,
+       .get_link = ethtool_op_get_link,
+};
+
+#ifdef CONFIG_SYSFS
+static ssize_t show_rxbuf_min(struct device *dev,
+                             struct device_attribute *attr, char *buf)
+{
+       struct net_device *netdev = to_net_dev(dev);
+       struct netfront_info *info = netdev_priv(netdev);
+
+       return sprintf(buf, "%u\n", info->rx_min_target);
+}
+
+static ssize_t store_rxbuf_min(struct device *dev,
+                              struct device_attribute *attr,
+                              const char *buf, size_t len)
+{
+       struct net_device *netdev = to_net_dev(dev);
+       struct netfront_info *np = netdev_priv(netdev);
+       char *endp;
+       unsigned long target;
+
+       if (!capable(CAP_NET_ADMIN))
+               return -EPERM;
+
+       target = simple_strtoul(buf, &endp, 0);
+       if (endp == buf)
+               return -EBADMSG;
+
+       if (target < RX_MIN_TARGET)
+               target = RX_MIN_TARGET;
+       if (target > RX_MAX_TARGET)
+               target = RX_MAX_TARGET;
+
+       spin_lock_bh(&np->rx_lock);
+       if (target > np->rx_max_target)
+               np->rx_max_target = target;
+       np->rx_min_target = target;
+       if (target > np->rx_target)
+               np->rx_target = target;
+
+       xennet_alloc_rx_buffers(netdev);
+
+       spin_unlock_bh(&np->rx_lock);
+       return len;
+}
+
+static ssize_t show_rxbuf_max(struct device *dev,
+                             struct device_attribute *attr, char *buf)
+{
+       struct net_device *netdev = to_net_dev(dev);
+       struct netfront_info *info = netdev_priv(netdev);
+
+       return sprintf(buf, "%u\n", info->rx_max_target);
+}
+
+static ssize_t store_rxbuf_max(struct device *dev,
+                              struct device_attribute *attr,
+                              const char *buf, size_t len)
+{
+       struct net_device *netdev = to_net_dev(dev);
+       struct netfront_info *np = netdev_priv(netdev);
+       char *endp;
+       unsigned long target;
+
+       if (!capable(CAP_NET_ADMIN))
+               return -EPERM;
+
+       target = simple_strtoul(buf, &endp, 0);
+       if (endp == buf)
+               return -EBADMSG;
+
+       if (target < RX_MIN_TARGET)
+               target = RX_MIN_TARGET;
+       if (target > RX_MAX_TARGET)
+               target = RX_MAX_TARGET;
+
+       spin_lock_bh(&np->rx_lock);
+       if (target < np->rx_min_target)
+               np->rx_min_target = target;
+       np->rx_max_target = target;
+       if (target < np->rx_target)
+               np->rx_target = target;
+
+       xennet_alloc_rx_buffers(netdev);
+
+       spin_unlock_bh(&np->rx_lock);
+       return len;
+}
+
+static ssize_t show_rxbuf_cur(struct device *dev,
+                             struct device_attribute *attr, char *buf)
+{
+       struct net_device *netdev = to_net_dev(dev);
+       struct netfront_info *info = netdev_priv(netdev);
+
+       return sprintf(buf, "%u\n", info->rx_target);
+}
+
+static struct device_attribute xennet_attrs[] = {
+       __ATTR(rxbuf_min, S_IRUGO|S_IWUSR, show_rxbuf_min, store_rxbuf_min),
+       __ATTR(rxbuf_max, S_IRUGO|S_IWUSR, show_rxbuf_max, store_rxbuf_max),
+       __ATTR(rxbuf_cur, S_IRUGO, show_rxbuf_cur, NULL),
+};
+
+static int xennet_sysfs_addif(struct net_device *netdev)
+{
+       int i;
+       int err;
+
+       for (i = 0; i < ARRAY_SIZE(xennet_attrs); i++) {
+               err = device_create_file(&netdev->dev,
+                                          &xennet_attrs[i]);
+               if (err)
+                       goto fail;
+       }
+       return 0;
+
+ fail:
+       while (--i >= 0)
+               device_remove_file(&netdev->dev, &xennet_attrs[i]);
+       return err;
+}
+
+static void xennet_sysfs_delif(struct net_device *netdev)
+{
+       int i;
+
+       for (i = 0; i < ARRAY_SIZE(xennet_attrs); i++)
+               device_remove_file(&netdev->dev, &xennet_attrs[i]);
+}
+
+#endif /* CONFIG_SYSFS */
+
+static struct xenbus_device_id netfront_ids[] = {
+       { "vif" },
+       { "" }
+};
+
+
+static int __devexit xennet_remove(struct xenbus_device *dev)
+{
+       struct netfront_info *info = dev->dev.driver_data;
+
+       dev_dbg(&dev->dev, "%s\n", dev->nodename);
+
+       unregister_netdev(info->netdev);
+
+       xennet_disconnect_backend(info);
+
+       del_timer_sync(&info->rx_refill_timer);
+
+       xennet_sysfs_delif(info->netdev);
+
+       free_netdev(info->netdev);
+
+       return 0;
+}
+
+static struct xenbus_driver netfront = {
+       .name = "vif",
+       .owner = THIS_MODULE,
+       .ids = netfront_ids,
+       .probe = netfront_probe,
+       .remove = __devexit_p(xennet_remove),
+       .resume = netfront_resume,
+       .otherend_changed = backend_changed,
+};
+
+static int __init netif_init(void)
+{
+       if (!is_running_on_xen())
+               return -ENODEV;
+
+       if (is_initial_xendomain())
+               return 0;
+
+       printk(KERN_INFO "Initialising Xen virtual ethernet driver.\n");
+
+       return xenbus_register_frontend(&netfront);
+}
+module_init(netif_init);
+
+
+static void __exit netif_exit(void)
+{
+       if (is_initial_xendomain())
+               return;
+
+       return xenbus_unregister_driver(&netfront);
+}
+module_exit(netif_exit);
+
+MODULE_DESCRIPTION("Xen virtual network device frontend");
+MODULE_LICENSE("GPL");