virtio_net: invoke softirqs after __napi_schedule
[linux-flexiantxendom0-3.2.10.git] / drivers / net / virtio_net.c
index 78eb319..cbefe67 100644 (file)
@@ -30,7 +30,7 @@
 static int napi_weight = 128;
 module_param(napi_weight, int, 0444);
 
-static int csum = 1, gso = 1;
+static bool csum = true, gso = true;
 module_param(csum, bool, 0444);
 module_param(gso, bool, 0444);
 
@@ -39,6 +39,16 @@ module_param(gso, bool, 0444);
 #define GOOD_COPY_LEN  128
 
 #define VIRTNET_SEND_COMMAND_SG_MAX    2
+#define VIRTNET_DRIVER_VERSION "1.0.0"
+
+struct virtnet_stats {
+       struct u64_stats_sync syncp;
+       u64 tx_bytes;
+       u64 tx_packets;
+
+       u64 rx_bytes;
+       u64 rx_packets;
+};
 
 struct virtnet_info {
        struct virtio_device *vdev;
@@ -56,6 +66,9 @@ struct virtnet_info {
        /* Host will merge rx buffers for big packets (shake it! shake it!) */
        bool mergeable_rx_bufs;
 
+       /* Active statistics */
+       struct virtnet_stats __percpu *stats;
+
        /* Work struct for refilling if we run low on memory. */
        struct delayed_work refill;
 
@@ -131,20 +144,19 @@ static void skb_xmit_done(struct virtqueue *svq)
 static void set_skb_frag(struct sk_buff *skb, struct page *page,
                         unsigned int offset, unsigned int *len)
 {
+       int size = min((unsigned)PAGE_SIZE - offset, *len);
        int i = skb_shinfo(skb)->nr_frags;
-       skb_frag_t *f;
 
-       f = &skb_shinfo(skb)->frags[i];
-       f->size = min((unsigned)PAGE_SIZE - offset, *len);
-       f->page_offset = offset;
-       f->page = page;
+       __skb_fill_page_desc(skb, i, page, offset, size);
 
-       skb->data_len += f->size;
-       skb->len += f->size;
+       skb->data_len += size;
+       skb->len += size;
+       skb->truesize += PAGE_SIZE;
        skb_shinfo(skb)->nr_frags++;
-       *len -= f->size;
+       *len -= size;
 }
 
+/* Called from bottom half context */
 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
                                   struct page *page, unsigned int len)
 {
@@ -183,6 +195,19 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
        len -= copy;
        offset += copy;
 
+       /*
+        * Verify that we can indeed put this data into a skb.
+        * This is here to handle cases when the device erroneously
+        * tries to receive more than is possible. This is usually
+        * the case of a broken device.
+        */
+       if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
+               if (net_ratelimit())
+                       pr_debug("%s: too much data\n", skb->dev->name);
+               dev_kfree_skb(skb);
+               return NULL;
+       }
+
        while (len) {
                set_skb_frag(skb, page, offset, &len);
                page = (struct page *)page->private;
@@ -209,7 +234,6 @@ static int receive_mergeable(struct virtnet_info *vi, struct sk_buff *skb)
                        skb->dev->stats.rx_length_errors++;
                        return -EINVAL;
                }
-
                page = virtqueue_get_buf(vi->rvq, &len);
                if (!page) {
                        pr_debug("%s: rx error: %d buffers missing\n",
@@ -217,6 +241,7 @@ static int receive_mergeable(struct virtnet_info *vi, struct sk_buff *skb)
                        skb->dev->stats.rx_length_errors++;
                        return -EINVAL;
                }
+
                if (len > PAGE_SIZE)
                        len = PAGE_SIZE;
 
@@ -230,6 +255,7 @@ static int receive_mergeable(struct virtnet_info *vi, struct sk_buff *skb)
 static void receive_buf(struct net_device *dev, void *buf, unsigned int len)
 {
        struct virtnet_info *vi = netdev_priv(dev);
+       struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
        struct sk_buff *skb;
        struct page *page;
        struct skb_vnet_hdr *hdr;
@@ -264,9 +290,11 @@ static void receive_buf(struct net_device *dev, void *buf, unsigned int len)
        }
 
        hdr = skb_vnet_hdr(skb);
-       skb->truesize += skb->data_len;
-       dev->stats.rx_bytes += skb->len;
-       dev->stats.rx_packets++;
+
+       u64_stats_update_begin(&stats->syncp);
+       stats->rx_bytes += skb->len;
+       stats->rx_packets++;
+       u64_stats_update_end(&stats->syncp);
 
        if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
                pr_debug("Needs csum!\n");
@@ -274,6 +302,8 @@ static void receive_buf(struct net_device *dev, void *buf, unsigned int len)
                                          hdr->hdr.csum_start,
                                          hdr->hdr.csum_offset))
                        goto frame_err;
+       } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
+               skb->ip_summed = CHECKSUM_UNNECESSARY;
        }
 
        skb->protocol = eth_type_trans(skb, dev);
@@ -329,7 +359,7 @@ static int add_recvbuf_small(struct virtnet_info *vi, gfp_t gfp)
        struct skb_vnet_hdr *hdr;
        int err;
 
-       skb = netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN);
+       skb = __netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN, gfp);
        if (unlikely(!skb))
                return -ENOMEM;
 
@@ -340,7 +370,7 @@ static int add_recvbuf_small(struct virtnet_info *vi, gfp_t gfp)
 
        skb_to_sgvec(skb, vi->rx_sg + 1, 0, skb->len);
 
-       err = virtqueue_add_buf(vi->rvq, vi->rx_sg, 0, 2, skb);
+       err = virtqueue_add_buf(vi->rvq, vi->rx_sg, 0, 2, skb, gfp);
        if (err < 0)
                dev_kfree_skb(skb);
 
@@ -386,7 +416,7 @@ static int add_recvbuf_big(struct virtnet_info *vi, gfp_t gfp)
        /* chain first in list head */
        first->private = (unsigned long)list;
        err = virtqueue_add_buf(vi->rvq, vi->rx_sg, 0, MAX_SKB_FRAGS + 2,
-                                      first);
+                               first, gfp);
        if (err < 0)
                give_pages(vi, first);
 
@@ -404,18 +434,24 @@ static int add_recvbuf_mergeable(struct virtnet_info *vi, gfp_t gfp)
 
        sg_init_one(vi->rx_sg, page_address(page), PAGE_SIZE);
 
-       err = virtqueue_add_buf(vi->rvq, vi->rx_sg, 0, 1, page);
+       err = virtqueue_add_buf(vi->rvq, vi->rx_sg, 0, 1, page, gfp);
        if (err < 0)
                give_pages(vi, page);
 
        return err;
 }
 
-/* Returns false if we couldn't fill entirely (OOM). */
+/*
+ * Returns false if we couldn't fill entirely (OOM).
+ *
+ * Normally run in the receive path, but can also be run from ndo_open
+ * before we're receiving packets, or from refill_work which is
+ * careful to disable receiving (using napi_disable).
+ */
 static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
 {
        int err;
-       bool oom = false;
+       bool oom;
 
        do {
                if (vi->mergeable_rx_bufs)
@@ -425,10 +461,9 @@ static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
                else
                        err = add_recvbuf_small(vi, gfp);
 
-               if (err < 0) {
-                       oom = true;
+               oom = err == -ENOMEM;
+               if (err < 0)
                        break;
-               }
                ++vi->num;
        } while (err > 0);
        if (unlikely(vi->num > vi->max))
@@ -447,6 +482,22 @@ static void skb_recv_done(struct virtqueue *rvq)
        }
 }
 
+static void virtnet_napi_enable(struct virtnet_info *vi)
+{
+       napi_enable(&vi->napi);
+
+       /* If all buffers were filled by other side before we napi_enabled, we
+        * won't get another interrupt, so process any outstanding packets
+        * now.  virtnet_poll wants re-enable the queue, so we disable here.
+        * We synchronize against interrupts via NAPI_STATE_SCHED */
+       if (napi_schedule_prep(&vi->napi)) {
+               virtqueue_disable_cb(vi->rvq);
+               local_bh_disable();
+               __napi_schedule(&vi->napi);
+               local_bh_enable();
+       }
+}
+
 static void refill_work(struct work_struct *work)
 {
        struct virtnet_info *vi;
@@ -455,12 +506,12 @@ static void refill_work(struct work_struct *work)
        vi = container_of(work, struct virtnet_info, refill.work);
        napi_disable(&vi->napi);
        still_empty = !try_fill_recv(vi, GFP_KERNEL);
-       napi_enable(&vi->napi);
+       virtnet_napi_enable(vi);
 
        /* In theory, this can happen: if we don't get any buffers in
         * we will *never* try to fill again. */
        if (still_empty)
-               schedule_delayed_work(&vi->refill, HZ/2);
+               queue_delayed_work(system_nrt_wq, &vi->refill, HZ/2);
 }
 
 static int virtnet_poll(struct napi_struct *napi, int budget)
@@ -479,7 +530,7 @@ again:
 
        if (vi->num < vi->max / 2) {
                if (!try_fill_recv(vi, GFP_ATOMIC))
-                       schedule_delayed_work(&vi->refill, 0);
+                       queue_delayed_work(system_nrt_wq, &vi->refill, 0);
        }
 
        /* Out of packets? */
@@ -500,11 +551,16 @@ static unsigned int free_old_xmit_skbs(struct virtnet_info *vi)
 {
        struct sk_buff *skb;
        unsigned int len, tot_sgs = 0;
+       struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 
        while ((skb = virtqueue_get_buf(vi->svq, &len)) != NULL) {
                pr_debug("Sent skb %p\n", skb);
-               vi->dev->stats.tx_bytes += skb->len;
-               vi->dev->stats.tx_packets++;
+
+               u64_stats_update_begin(&stats->syncp);
+               stats->tx_bytes += skb->len;
+               stats->tx_packets++;
+               u64_stats_update_end(&stats->syncp);
+
                tot_sgs += skb_vnet_hdr(skb)->num_sg;
                dev_kfree_skb_any(skb);
        }
@@ -520,7 +576,7 @@ static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
 
        if (skb->ip_summed == CHECKSUM_PARTIAL) {
                hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
-               hdr->hdr.csum_start = skb->csum_start - skb_headroom(skb);
+               hdr->hdr.csum_start = skb_checksum_start_offset(skb);
                hdr->hdr.csum_offset = skb->csum_offset;
        } else {
                hdr->hdr.flags = 0;
@@ -555,7 +611,7 @@ static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
 
        hdr->num_sg = skb_to_sgvec(skb, vi->tx_sg + 1, 0, skb->len) + 1;
        return virtqueue_add_buf(vi->svq, vi->tx_sg, hdr->num_sg,
-                                       0, skb);
+                                0, skb, GFP_ATOMIC);
 }
 
 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
@@ -563,7 +619,6 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
        struct virtnet_info *vi = netdev_priv(dev);
        int capacity;
 
-again:
        /* Free up any pending old buffers before queueing new ones. */
        free_old_xmit_skbs(vi);
 
@@ -572,14 +627,20 @@ again:
 
        /* This can happen with OOM and indirect buffers. */
        if (unlikely(capacity < 0)) {
-               netif_stop_queue(dev);
-               dev_warn(&dev->dev, "Unexpected full queue\n");
-               if (unlikely(!virtqueue_enable_cb(vi->svq))) {
-                       virtqueue_disable_cb(vi->svq);
-                       netif_start_queue(dev);
-                       goto again;
+               if (likely(capacity == -ENOMEM)) {
+                       if (net_ratelimit())
+                               dev_warn(&dev->dev,
+                                        "TX queue failure: out of memory\n");
+               } else {
+                       dev->stats.tx_fifo_errors++;
+                       if (net_ratelimit())
+                               dev_warn(&dev->dev,
+                                        "Unexpected TX queue failure: %d\n",
+                                        capacity);
                }
-               return NETDEV_TX_BUSY;
+               dev->stats.tx_dropped++;
+               kfree_skb(skb);
+               return NETDEV_TX_OK;
        }
        virtqueue_kick(vi->svq);
 
@@ -591,7 +652,7 @@ again:
         * before it gets out of hand.  Naturally, this wastes entries. */
        if (capacity < 2+MAX_SKB_FRAGS) {
                netif_stop_queue(dev);
-               if (unlikely(!virtqueue_enable_cb(vi->svq))) {
+               if (unlikely(!virtqueue_enable_cb_delayed(vi->svq))) {
                        /* More just got used, free them then recheck. */
                        capacity += free_old_xmit_skbs(vi);
                        if (capacity >= 2+MAX_SKB_FRAGS) {
@@ -621,6 +682,40 @@ static int virtnet_set_mac_address(struct net_device *dev, void *p)
        return 0;
 }
 
+static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
+                                              struct rtnl_link_stats64 *tot)
+{
+       struct virtnet_info *vi = netdev_priv(dev);
+       int cpu;
+       unsigned int start;
+
+       for_each_possible_cpu(cpu) {
+               struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
+               u64 tpackets, tbytes, rpackets, rbytes;
+
+               do {
+                       start = u64_stats_fetch_begin(&stats->syncp);
+                       tpackets = stats->tx_packets;
+                       tbytes   = stats->tx_bytes;
+                       rpackets = stats->rx_packets;
+                       rbytes   = stats->rx_bytes;
+               } while (u64_stats_fetch_retry(&stats->syncp, start));
+
+               tot->rx_packets += rpackets;
+               tot->tx_packets += tpackets;
+               tot->rx_bytes   += rbytes;
+               tot->tx_bytes   += tbytes;
+       }
+
+       tot->tx_dropped = dev->stats.tx_dropped;
+       tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
+       tot->rx_dropped = dev->stats.rx_dropped;
+       tot->rx_length_errors = dev->stats.rx_length_errors;
+       tot->rx_frame_errors = dev->stats.rx_frame_errors;
+
+       return tot;
+}
+
 #ifdef CONFIG_NET_POLL_CONTROLLER
 static void virtnet_netpoll(struct net_device *dev)
 {
@@ -634,16 +729,11 @@ static int virtnet_open(struct net_device *dev)
 {
        struct virtnet_info *vi = netdev_priv(dev);
 
-       napi_enable(&vi->napi);
+       /* Make sure we have some buffers: if oom use wq. */
+       if (!try_fill_recv(vi, GFP_KERNEL))
+               queue_delayed_work(system_nrt_wq, &vi->refill, 0);
 
-       /* If all buffers were filled by other side before we napi_enabled, we
-        * won't get another interrupt, so process any outstanding packets
-        * now.  virtnet_poll wants re-enable the queue, so we disable here.
-        * We synchronize against interrupts via NAPI_STATE_SCHED */
-       if (napi_schedule_prep(&vi->napi)) {
-               virtqueue_disable_cb(vi->rvq);
-               __napi_schedule(&vi->napi);
-       }
+       virtnet_napi_enable(vi);
        return 0;
 }
 
@@ -678,7 +768,7 @@ static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
                sg_set_buf(&sg[i + 1], sg_virt(s), s->length);
        sg_set_buf(&sg[out + in - 1], &status, sizeof(status));
 
-       BUG_ON(virtqueue_add_buf(vi->cvq, sg, out, in, vi) < 0);
+       BUG_ON(virtqueue_add_buf(vi->cvq, sg, out, in, vi, GFP_ATOMIC) < 0);
 
        virtqueue_kick(vi->cvq);
 
@@ -696,22 +786,13 @@ static int virtnet_close(struct net_device *dev)
 {
        struct virtnet_info *vi = netdev_priv(dev);
 
+       /* Make sure refill_work doesn't re-enable napi! */
+       cancel_delayed_work_sync(&vi->refill);
        napi_disable(&vi->napi);
 
        return 0;
 }
 
-static int virtnet_set_tx_csum(struct net_device *dev, u32 data)
-{
-       struct virtnet_info *vi = netdev_priv(dev);
-       struct virtio_device *vdev = vi->vdev;
-
-       if (data && !virtio_has_feature(vdev, VIRTIO_NET_F_CSUM))
-               return -ENOSYS;
-
-       return ethtool_op_set_tx_hw_csum(dev, data);
-}
-
 static void virtnet_set_rx_mode(struct net_device *dev)
 {
        struct virtnet_info *vi = netdev_priv(dev);
@@ -788,7 +869,7 @@ static void virtnet_set_rx_mode(struct net_device *dev)
        kfree(buf);
 }
 
-static void virtnet_vlan_rx_add_vid(struct net_device *dev, u16 vid)
+static int virtnet_vlan_rx_add_vid(struct net_device *dev, u16 vid)
 {
        struct virtnet_info *vi = netdev_priv(dev);
        struct scatterlist sg;
@@ -798,9 +879,10 @@ static void virtnet_vlan_rx_add_vid(struct net_device *dev, u16 vid)
        if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
                                  VIRTIO_NET_CTRL_VLAN_ADD, &sg, 1, 0))
                dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
+       return 0;
 }
 
-static void virtnet_vlan_rx_kill_vid(struct net_device *dev, u16 vid)
+static int virtnet_vlan_rx_kill_vid(struct net_device *dev, u16 vid)
 {
        struct virtnet_info *vi = netdev_priv(dev);
        struct scatterlist sg;
@@ -810,14 +892,38 @@ static void virtnet_vlan_rx_kill_vid(struct net_device *dev, u16 vid)
        if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
                                  VIRTIO_NET_CTRL_VLAN_DEL, &sg, 1, 0))
                dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
+       return 0;
+}
+
+static void virtnet_get_ringparam(struct net_device *dev,
+                               struct ethtool_ringparam *ring)
+{
+       struct virtnet_info *vi = netdev_priv(dev);
+
+       ring->rx_max_pending = virtqueue_get_vring_size(vi->rvq);
+       ring->tx_max_pending = virtqueue_get_vring_size(vi->svq);
+       ring->rx_pending = ring->rx_max_pending;
+       ring->tx_pending = ring->tx_max_pending;
+
+}
+
+
+static void virtnet_get_drvinfo(struct net_device *dev,
+                               struct ethtool_drvinfo *info)
+{
+       struct virtnet_info *vi = netdev_priv(dev);
+       struct virtio_device *vdev = vi->vdev;
+
+       strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
+       strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
+       strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
+
 }
 
 static const struct ethtool_ops virtnet_ethtool_ops = {
-       .set_tx_csum = virtnet_set_tx_csum,
-       .set_sg = ethtool_op_set_sg,
-       .set_tso = ethtool_op_set_tso,
-       .set_ufo = ethtool_op_set_ufo,
+       .get_drvinfo = virtnet_get_drvinfo,
        .get_link = ethtool_op_get_link,
+       .get_ringparam = virtnet_get_ringparam,
 };
 
 #define MIN_MTU 68
@@ -839,6 +945,7 @@ static const struct net_device_ops virtnet_netdev = {
        .ndo_set_mac_address = virtnet_set_mac_address,
        .ndo_set_rx_mode     = virtnet_set_rx_mode,
        .ndo_change_mtu      = virtnet_change_mtu,
+       .ndo_get_stats64     = virtnet_stats,
        .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
        .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
 #ifdef CONFIG_NET_POLL_CONTROLLER
@@ -850,12 +957,10 @@ static void virtnet_update_status(struct virtnet_info *vi)
 {
        u16 v;
 
-       if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS))
-               return;
-
-       vi->vdev->config->get(vi->vdev,
+       if (virtio_config_val(vi->vdev, VIRTIO_NET_F_STATUS,
                              offsetof(struct virtio_net_config, status),
-                             &v, sizeof(v));
+                             &v) < 0)
+               return;
 
        /* Ignore unknown (future) status bits */
        v &= VIRTIO_NET_S_LINK_UP;
@@ -881,15 +986,38 @@ static void virtnet_config_changed(struct virtio_device *vdev)
        virtnet_update_status(vi);
 }
 
+static int init_vqs(struct virtnet_info *vi)
+{
+       struct virtqueue *vqs[3];
+       vq_callback_t *callbacks[] = { skb_recv_done, skb_xmit_done, NULL};
+       const char *names[] = { "input", "output", "control" };
+       int nvqs, err;
+
+       /* We expect two virtqueues, receive then send,
+        * and optionally control. */
+       nvqs = virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ? 3 : 2;
+
+       err = vi->vdev->config->find_vqs(vi->vdev, nvqs, vqs, callbacks, names);
+       if (err)
+               return err;
+
+       vi->rvq = vqs[0];
+       vi->svq = vqs[1];
+
+       if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)) {
+               vi->cvq = vqs[2];
+
+               if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
+                       vi->dev->features |= NETIF_F_HW_VLAN_FILTER;
+       }
+       return 0;
+}
+
 static int virtnet_probe(struct virtio_device *vdev)
 {
        int err;
        struct net_device *dev;
        struct virtnet_info *vi;
-       struct virtqueue *vqs[3];
-       vq_callback_t *callbacks[] = { skb_recv_done, skb_xmit_done, NULL};
-       const char *names[] = { "input", "output", "control" };
-       int nvqs;
 
        /* Allocate ourselves a network device with room for our info */
        dev = alloc_etherdev(sizeof(struct virtnet_info));
@@ -897,37 +1025,44 @@ static int virtnet_probe(struct virtio_device *vdev)
                return -ENOMEM;
 
        /* Set up network device as normal. */
+       dev->priv_flags |= IFF_UNICAST_FLT;
        dev->netdev_ops = &virtnet_netdev;
        dev->features = NETIF_F_HIGHDMA;
+
        SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
        SET_NETDEV_DEV(dev, &vdev->dev);
 
        /* Do we support "hardware" checksums? */
-       if (csum && virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
+       if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
                /* This opens up the world of extra features. */
-               dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
-               if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
-                       dev->features |= NETIF_F_TSO | NETIF_F_UFO
+               dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
+               if (csum)
+                       dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
+
+               if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
+                       dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
                                | NETIF_F_TSO_ECN | NETIF_F_TSO6;
                }
                /* Individual feature bits: what can host handle? */
-               if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
-                       dev->features |= NETIF_F_TSO;
-               if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
-                       dev->features |= NETIF_F_TSO6;
-               if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
-                       dev->features |= NETIF_F_TSO_ECN;
-               if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
-                       dev->features |= NETIF_F_UFO;
+               if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
+                       dev->hw_features |= NETIF_F_TSO;
+               if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
+                       dev->hw_features |= NETIF_F_TSO6;
+               if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
+                       dev->hw_features |= NETIF_F_TSO_ECN;
+               if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
+                       dev->hw_features |= NETIF_F_UFO;
+
+               if (gso)
+                       dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
+               /* (!csum && gso) case will be fixed by register_netdev() */
        }
 
        /* Configuration may specify what MAC to use.  Otherwise random. */
-       if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
-               vdev->config->get(vdev,
+       if (virtio_config_val_len(vdev, VIRTIO_NET_F_MAC,
                                  offsetof(struct virtio_net_config, mac),
-                                 dev->dev_addr, dev->addr_len);
-       } else
-               random_ether_addr(dev->dev_addr);
+                                 dev->dev_addr, dev->addr_len) < 0)
+               eth_hw_addr_random(dev);
 
        /* Set up our device-specific information */
        vi = netdev_priv(dev);
@@ -936,6 +1071,11 @@ static int virtnet_probe(struct virtio_device *vdev)
        vi->vdev = vdev;
        vdev->priv = vi;
        vi->pages = NULL;
+       vi->stats = alloc_percpu(struct virtnet_stats);
+       err = -ENOMEM;
+       if (vi->stats == NULL)
+               goto free;
+
        INIT_DELAYED_WORK(&vi->refill, refill_work);
        sg_init_table(vi->rx_sg, ARRAY_SIZE(vi->rx_sg));
        sg_init_table(vi->tx_sg, ARRAY_SIZE(vi->tx_sg));
@@ -949,23 +1089,9 @@ static int virtnet_probe(struct virtio_device *vdev)
        if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
                vi->mergeable_rx_bufs = true;
 
-       /* We expect two virtqueues, receive then send,
-        * and optionally control. */
-       nvqs = virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ? 3 : 2;
-
-       err = vdev->config->find_vqs(vdev, nvqs, vqs, callbacks, names);
+       err = init_vqs(vi);
        if (err)
-               goto free;
-
-       vi->rvq = vqs[0];
-       vi->svq = vqs[1];
-
-       if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)) {
-               vi->cvq = vqs[2];
-
-               if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
-                       dev->features |= NETIF_F_HW_VLAN_FILTER;
-       }
+               goto free_stats;
 
        err = register_netdev(dev);
        if (err) {
@@ -982,18 +1108,25 @@ static int virtnet_probe(struct virtio_device *vdev)
                goto unregister;
        }
 
-       vi->status = VIRTIO_NET_S_LINK_UP;
-       virtnet_update_status(vi);
-       netif_carrier_on(dev);
+       /* Assume link up if device can't report link status,
+          otherwise get link status from config. */
+       if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
+               netif_carrier_off(dev);
+               virtnet_update_status(vi);
+       } else {
+               vi->status = VIRTIO_NET_S_LINK_UP;
+               netif_carrier_on(dev);
+       }
 
        pr_debug("virtnet: registered device %s\n", dev->name);
        return 0;
 
 unregister:
        unregister_netdev(dev);
-       cancel_delayed_work_sync(&vi->refill);
 free_vqs:
        vdev->config->del_vqs(vdev);
+free_stats:
+       free_percpu(vi->stats);
 free:
        free_netdev(dev);
        return err;
@@ -1021,27 +1154,72 @@ static void free_unused_bufs(struct virtnet_info *vi)
        BUG_ON(vi->num != 0);
 }
 
+static void remove_vq_common(struct virtnet_info *vi)
+{
+       vi->vdev->config->reset(vi->vdev);
+
+       /* Free unused buffers in both send and recv, if any. */
+       free_unused_bufs(vi);
+
+       vi->vdev->config->del_vqs(vi->vdev);
+
+       while (vi->pages)
+               __free_pages(get_a_page(vi, GFP_KERNEL), 0);
+}
+
 static void __devexit virtnet_remove(struct virtio_device *vdev)
 {
        struct virtnet_info *vi = vdev->priv;
 
-       /* Stop all the virtqueues. */
-       vdev->config->reset(vdev);
+       unregister_netdev(vi->dev);
 
+       remove_vq_common(vi);
 
-       unregister_netdev(vi->dev);
+       free_percpu(vi->stats);
+       free_netdev(vi->dev);
+}
+
+#ifdef CONFIG_PM
+static int virtnet_freeze(struct virtio_device *vdev)
+{
+       struct virtnet_info *vi = vdev->priv;
+
+       virtqueue_disable_cb(vi->rvq);
+       virtqueue_disable_cb(vi->svq);
+       if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ))
+               virtqueue_disable_cb(vi->cvq);
+
+       netif_device_detach(vi->dev);
        cancel_delayed_work_sync(&vi->refill);
 
-       /* Free unused buffers in both send and recv, if any. */
-       free_unused_bufs(vi);
+       if (netif_running(vi->dev))
+               napi_disable(&vi->napi);
 
-       vdev->config->del_vqs(vi->vdev);
+       remove_vq_common(vi);
 
-       while (vi->pages)
-               __free_pages(get_a_page(vi, GFP_KERNEL), 0);
+       return 0;
+}
 
-       free_netdev(vi->dev);
+static int virtnet_restore(struct virtio_device *vdev)
+{
+       struct virtnet_info *vi = vdev->priv;
+       int err;
+
+       err = init_vqs(vi);
+       if (err)
+               return err;
+
+       if (netif_running(vi->dev))
+               virtnet_napi_enable(vi);
+
+       netif_device_attach(vi->dev);
+
+       if (!try_fill_recv(vi, GFP_KERNEL))
+               queue_delayed_work(system_nrt_wq, &vi->refill, 0);
+
+       return 0;
 }
+#endif
 
 static struct virtio_device_id id_table[] = {
        { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
@@ -1067,6 +1245,10 @@ static struct virtio_driver virtio_net_driver = {
        .probe =        virtnet_probe,
        .remove =       __devexit_p(virtnet_remove),
        .config_changed = virtnet_config_changed,
+#ifdef CONFIG_PM
+       .freeze =       virtnet_freeze,
+       .restore =      virtnet_restore,
+#endif
 };
 
 static int __init init(void)