Net driver using virtio
[linux-2.6] / drivers / net / virtio_net.c
1 /* A simple network driver using virtio.
2  *
3  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License as published by
7  * the Free Software Foundation; either version 2 of the License, or
8  * (at your option) any later version.
9  *
10  * This program is distributed in the hope that it will be useful,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13  * GNU General Public License for more details.
14  *
15  * You should have received a copy of the GNU General Public License
16  * along with this program; if not, write to the Free Software
17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
18  */
19 //#define DEBUG
20 #include <linux/netdevice.h>
21 #include <linux/etherdevice.h>
22 #include <linux/module.h>
23 #include <linux/virtio.h>
24 #include <linux/virtio_net.h>
25 #include <linux/scatterlist.h>
26
27 /* FIXME: MTU in config. */
28 #define MAX_PACKET_LEN (ETH_HLEN+ETH_DATA_LEN)
29
30 struct virtnet_info
31 {
32         struct virtio_device *vdev;
33         struct virtqueue *rvq, *svq;
34         struct net_device *dev;
35         struct napi_struct napi;
36
37         /* Number of input buffers, and max we've ever had. */
38         unsigned int num, max;
39
40         /* Receive & send queues. */
41         struct sk_buff_head recv;
42         struct sk_buff_head send;
43 };
44
45 static inline struct virtio_net_hdr *skb_vnet_hdr(struct sk_buff *skb)
46 {
47         return (struct virtio_net_hdr *)skb->cb;
48 }
49
50 static inline void vnet_hdr_to_sg(struct scatterlist *sg, struct sk_buff *skb)
51 {
52         sg_init_one(sg, skb_vnet_hdr(skb), sizeof(struct virtio_net_hdr));
53 }
54
55 static bool skb_xmit_done(struct virtqueue *rvq)
56 {
57         struct virtnet_info *vi = rvq->vdev->priv;
58
59         /* In case we were waiting for output buffers. */
60         netif_wake_queue(vi->dev);
61         return true;
62 }
63
64 static void receive_skb(struct net_device *dev, struct sk_buff *skb,
65                         unsigned len)
66 {
67         struct virtio_net_hdr *hdr = skb_vnet_hdr(skb);
68
69         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
70                 pr_debug("%s: short packet %i\n", dev->name, len);
71                 dev->stats.rx_length_errors++;
72                 goto drop;
73         }
74         len -= sizeof(struct virtio_net_hdr);
75         BUG_ON(len > MAX_PACKET_LEN);
76
77         skb_trim(skb, len);
78         skb->protocol = eth_type_trans(skb, dev);
79         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
80                  ntohs(skb->protocol), skb->len, skb->pkt_type);
81         dev->stats.rx_bytes += skb->len;
82         dev->stats.rx_packets++;
83
84         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
85                 pr_debug("Needs csum!\n");
86                 skb->ip_summed = CHECKSUM_PARTIAL;
87                 skb->csum_start = hdr->csum_start;
88                 skb->csum_offset = hdr->csum_offset;
89                 if (skb->csum_start > skb->len - 2
90                     || skb->csum_offset > skb->len - 2) {
91                         if (net_ratelimit())
92                                 printk(KERN_WARNING "%s: csum=%u/%u len=%u\n",
93                                        dev->name, skb->csum_start,
94                                        skb->csum_offset, skb->len);
95                         goto frame_err;
96                 }
97         }
98
99         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
100                 pr_debug("GSO!\n");
101                 switch (hdr->gso_type) {
102                 case VIRTIO_NET_HDR_GSO_TCPV4:
103                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
104                         break;
105                 case VIRTIO_NET_HDR_GSO_TCPV4_ECN:
106                         skb_shinfo(skb)->gso_type = SKB_GSO_TCP_ECN;
107                         break;
108                 case VIRTIO_NET_HDR_GSO_UDP:
109                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
110                         break;
111                 case VIRTIO_NET_HDR_GSO_TCPV6:
112                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
113                         break;
114                 default:
115                         if (net_ratelimit())
116                                 printk(KERN_WARNING "%s: bad gso type %u.\n",
117                                        dev->name, hdr->gso_type);
118                         goto frame_err;
119                 }
120
121                 skb_shinfo(skb)->gso_size = hdr->gso_size;
122                 if (skb_shinfo(skb)->gso_size == 0) {
123                         if (net_ratelimit())
124                                 printk(KERN_WARNING "%s: zero gso size.\n",
125                                        dev->name);
126                         goto frame_err;
127                 }
128
129                 /* Header must be checked, and gso_segs computed. */
130                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
131                 skb_shinfo(skb)->gso_segs = 0;
132         }
133
134         netif_receive_skb(skb);
135         return;
136
137 frame_err:
138         dev->stats.rx_frame_errors++;
139 drop:
140         dev_kfree_skb(skb);
141 }
142
143 static void try_fill_recv(struct virtnet_info *vi)
144 {
145         struct sk_buff *skb;
146         struct scatterlist sg[1+MAX_SKB_FRAGS];
147         int num, err;
148
149         for (;;) {
150                 skb = netdev_alloc_skb(vi->dev, MAX_PACKET_LEN);
151                 if (unlikely(!skb))
152                         break;
153
154                 skb_put(skb, MAX_PACKET_LEN);
155                 vnet_hdr_to_sg(sg, skb);
156                 num = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
157                 skb_queue_head(&vi->recv, skb);
158
159                 err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, num, skb);
160                 if (err) {
161                         skb_unlink(skb, &vi->recv);
162                         kfree_skb(skb);
163                         break;
164                 }
165                 vi->num++;
166         }
167         if (unlikely(vi->num > vi->max))
168                 vi->max = vi->num;
169         vi->rvq->vq_ops->kick(vi->rvq);
170 }
171
172 static bool skb_recv_done(struct virtqueue *rvq)
173 {
174         struct virtnet_info *vi = rvq->vdev->priv;
175         netif_rx_schedule(vi->dev, &vi->napi);
176         /* Suppress further interrupts. */
177         return false;
178 }
179
180 static int virtnet_poll(struct napi_struct *napi, int budget)
181 {
182         struct virtnet_info *vi = container_of(napi, struct virtnet_info, napi);
183         struct sk_buff *skb = NULL;
184         unsigned int len, received = 0;
185
186 again:
187         while (received < budget &&
188                (skb = vi->rvq->vq_ops->get_buf(vi->rvq, &len)) != NULL) {
189                 __skb_unlink(skb, &vi->recv);
190                 receive_skb(vi->dev, skb, len);
191                 vi->num--;
192                 received++;
193         }
194
195         /* FIXME: If we oom and completely run out of inbufs, we need
196          * to start a timer trying to fill more. */
197         if (vi->num < vi->max / 2)
198                 try_fill_recv(vi);
199
200         /* All done? */
201         if (!skb) {
202                 netif_rx_complete(vi->dev, napi);
203                 if (unlikely(!vi->rvq->vq_ops->restart(vi->rvq))
204                     && netif_rx_reschedule(vi->dev, napi))
205                         goto again;
206         }
207
208         return received;
209 }
210
211 static void free_old_xmit_skbs(struct virtnet_info *vi)
212 {
213         struct sk_buff *skb;
214         unsigned int len;
215
216         while ((skb = vi->svq->vq_ops->get_buf(vi->svq, &len)) != NULL) {
217                 pr_debug("Sent skb %p\n", skb);
218                 __skb_unlink(skb, &vi->send);
219                 vi->dev->stats.tx_bytes += len;
220                 vi->dev->stats.tx_packets++;
221                 kfree_skb(skb);
222         }
223 }
224
225 static int start_xmit(struct sk_buff *skb, struct net_device *dev)
226 {
227         struct virtnet_info *vi = netdev_priv(dev);
228         int num, err;
229         struct scatterlist sg[1+MAX_SKB_FRAGS];
230         struct virtio_net_hdr *hdr;
231         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
232         DECLARE_MAC_BUF(mac);
233
234         pr_debug("%s: xmit %p %s\n", dev->name, skb, print_mac(mac, dest));
235
236         free_old_xmit_skbs(vi);
237
238         /* Encode metadata header at front. */
239         hdr = skb_vnet_hdr(skb);
240         if (skb->ip_summed == CHECKSUM_PARTIAL) {
241                 hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
242                 hdr->csum_start = skb->csum_start - skb_headroom(skb);
243                 hdr->csum_offset = skb->csum_offset;
244         } else {
245                 hdr->flags = 0;
246                 hdr->csum_offset = hdr->csum_start = 0;
247         }
248
249         if (skb_is_gso(skb)) {
250                 hdr->gso_size = skb_shinfo(skb)->gso_size;
251                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
252                         hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4_ECN;
253                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
254                         hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
255                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
256                         hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
257                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
258                         hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
259                 else
260                         BUG();
261         } else {
262                 hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
263                 hdr->gso_size = 0;
264         }
265
266         vnet_hdr_to_sg(sg, skb);
267         num = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
268         __skb_queue_head(&vi->send, skb);
269         err = vi->svq->vq_ops->add_buf(vi->svq, sg, num, 0, skb);
270         if (err) {
271                 pr_debug("%s: virtio not prepared to send\n", dev->name);
272                 skb_unlink(skb, &vi->send);
273                 netif_stop_queue(dev);
274                 return NETDEV_TX_BUSY;
275         }
276         vi->svq->vq_ops->kick(vi->svq);
277
278         return 0;
279 }
280
281 static int virtnet_open(struct net_device *dev)
282 {
283         struct virtnet_info *vi = netdev_priv(dev);
284
285         try_fill_recv(vi);
286
287         /* If we didn't even get one input buffer, we're useless. */
288         if (vi->num == 0)
289                 return -ENOMEM;
290
291         napi_enable(&vi->napi);
292         return 0;
293 }
294
295 static int virtnet_close(struct net_device *dev)
296 {
297         struct virtnet_info *vi = netdev_priv(dev);
298         struct sk_buff *skb;
299
300         napi_disable(&vi->napi);
301
302         /* networking core has neutered skb_xmit_done/skb_recv_done, so don't
303          * worry about races vs. get(). */
304         vi->rvq->vq_ops->shutdown(vi->rvq);
305         while ((skb = __skb_dequeue(&vi->recv)) != NULL) {
306                 kfree_skb(skb);
307                 vi->num--;
308         }
309         vi->svq->vq_ops->shutdown(vi->svq);
310         while ((skb = __skb_dequeue(&vi->send)) != NULL)
311                 kfree_skb(skb);
312
313         BUG_ON(vi->num != 0);
314         return 0;
315 }
316
317 static int virtnet_probe(struct virtio_device *vdev)
318 {
319         int err;
320         unsigned int len;
321         struct net_device *dev;
322         struct virtnet_info *vi;
323         void *token;
324
325         /* Allocate ourselves a network device with room for our info */
326         dev = alloc_etherdev(sizeof(struct virtnet_info));
327         if (!dev)
328                 return -ENOMEM;
329
330         /* Set up network device as normal. */
331         ether_setup(dev);
332         dev->open = virtnet_open;
333         dev->stop = virtnet_close;
334         dev->hard_start_xmit = start_xmit;
335         dev->features = NETIF_F_HIGHDMA;
336         SET_NETDEV_DEV(dev, &vdev->dev);
337
338         /* Do we support "hardware" checksums? */
339         token = vdev->config->find(vdev, VIRTIO_CONFIG_NET_F, &len);
340         if (virtio_use_bit(vdev, token, len, VIRTIO_NET_F_NO_CSUM)) {
341                 /* This opens up the world of extra features. */
342                 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
343                 if (virtio_use_bit(vdev, token, len, VIRTIO_NET_F_TSO4))
344                         dev->features |= NETIF_F_TSO;
345                 if (virtio_use_bit(vdev, token, len, VIRTIO_NET_F_UFO))
346                         dev->features |= NETIF_F_UFO;
347                 if (virtio_use_bit(vdev, token, len, VIRTIO_NET_F_TSO4_ECN))
348                         dev->features |= NETIF_F_TSO_ECN;
349                 if (virtio_use_bit(vdev, token, len, VIRTIO_NET_F_TSO6))
350                         dev->features |= NETIF_F_TSO6;
351         }
352
353         /* Configuration may specify what MAC to use.  Otherwise random. */
354         token = vdev->config->find(vdev, VIRTIO_CONFIG_NET_MAC_F, &len);
355         if (token) {
356                 dev->addr_len = len;
357                 vdev->config->get(vdev, token, dev->dev_addr, len);
358         } else
359                 random_ether_addr(dev->dev_addr);
360
361         /* Set up our device-specific information */
362         vi = netdev_priv(dev);
363         netif_napi_add(dev, &vi->napi, virtnet_poll, 16);
364         vi->dev = dev;
365         vi->vdev = vdev;
366
367         /* We expect two virtqueues, receive then send. */
368         vi->rvq = vdev->config->find_vq(vdev, skb_recv_done);
369         if (IS_ERR(vi->rvq)) {
370                 err = PTR_ERR(vi->rvq);
371                 goto free;
372         }
373
374         vi->svq = vdev->config->find_vq(vdev, skb_xmit_done);
375         if (IS_ERR(vi->svq)) {
376                 err = PTR_ERR(vi->svq);
377                 goto free_recv;
378         }
379
380         /* Initialize our empty receive and send queues. */
381         skb_queue_head_init(&vi->recv);
382         skb_queue_head_init(&vi->send);
383
384         err = register_netdev(dev);
385         if (err) {
386                 pr_debug("virtio_net: registering device failed\n");
387                 goto free_send;
388         }
389         pr_debug("virtnet: registered device %s\n", dev->name);
390         vdev->priv = vi;
391         return 0;
392
393 free_send:
394         vdev->config->del_vq(vi->svq);
395 free_recv:
396         vdev->config->del_vq(vi->rvq);
397 free:
398         free_netdev(dev);
399         return err;
400 }
401
402 static void virtnet_remove(struct virtio_device *vdev)
403 {
404         unregister_netdev(vdev->priv);
405         free_netdev(vdev->priv);
406 }
407
408 static struct virtio_device_id id_table[] = {
409         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
410         { 0 },
411 };
412
413 static struct virtio_driver virtio_net = {
414         .driver.name =  KBUILD_MODNAME,
415         .driver.owner = THIS_MODULE,
416         .id_table =     id_table,
417         .probe =        virtnet_probe,
418         .remove =       __devexit_p(virtnet_remove),
419 };
420
421 static int __init init(void)
422 {
423         return register_virtio_driver(&virtio_net);
424 }
425
426 static void __exit fini(void)
427 {
428         unregister_virtio_driver(&virtio_net);
429 }
430 module_init(init);
431 module_exit(fini);
432
433 MODULE_DEVICE_TABLE(virtio, id_table);
434 MODULE_DESCRIPTION("Virtio network driver");
435 MODULE_LICENSE("GPL");