virtio: fix delayed xmit of packet and freeing of old packets.
[linux-2.6.git] / 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 static int napi_weight = 128;
28 module_param(napi_weight, int, 0444);
29
30 static int csum = 1, gso = 1;
31 module_param(csum, bool, 0444);
32 module_param(gso, bool, 0444);
33
34 /* FIXME: MTU in config. */
35 #define MAX_PACKET_LEN (ETH_HLEN+ETH_DATA_LEN)
36
37 struct virtnet_info
38 {
39         struct virtio_device *vdev;
40         struct virtqueue *rvq, *svq;
41         struct net_device *dev;
42         struct napi_struct napi;
43
44         /* The skb we couldn't send because buffers were full. */
45         struct sk_buff *last_xmit_skb;
46
47         /* Number of input buffers, and max we've ever had. */
48         unsigned int num, max;
49
50         /* For cleaning up after transmission. */
51         struct tasklet_struct tasklet;
52
53         /* Receive & send queues. */
54         struct sk_buff_head recv;
55         struct sk_buff_head send;
56 };
57
58 static inline struct virtio_net_hdr *skb_vnet_hdr(struct sk_buff *skb)
59 {
60         return (struct virtio_net_hdr *)skb->cb;
61 }
62
63 static inline void vnet_hdr_to_sg(struct scatterlist *sg, struct sk_buff *skb)
64 {
65         sg_init_one(sg, skb_vnet_hdr(skb), sizeof(struct virtio_net_hdr));
66 }
67
68 static void skb_xmit_done(struct virtqueue *svq)
69 {
70         struct virtnet_info *vi = svq->vdev->priv;
71
72         /* Suppress further interrupts. */
73         svq->vq_ops->disable_cb(svq);
74
75         /* We were waiting for more output buffers. */
76         netif_wake_queue(vi->dev);
77
78         /* Make sure we re-xmit last_xmit_skb: if there are no more packets
79          * queued, start_xmit won't be called. */
80         tasklet_schedule(&vi->tasklet);
81 }
82
83 static void receive_skb(struct net_device *dev, struct sk_buff *skb,
84                         unsigned len)
85 {
86         struct virtio_net_hdr *hdr = skb_vnet_hdr(skb);
87
88         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
89                 pr_debug("%s: short packet %i\n", dev->name, len);
90                 dev->stats.rx_length_errors++;
91                 goto drop;
92         }
93         len -= sizeof(struct virtio_net_hdr);
94         BUG_ON(len > MAX_PACKET_LEN);
95
96         skb_trim(skb, len);
97         skb->protocol = eth_type_trans(skb, dev);
98         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
99                  ntohs(skb->protocol), skb->len, skb->pkt_type);
100         dev->stats.rx_bytes += skb->len;
101         dev->stats.rx_packets++;
102
103         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
104                 pr_debug("Needs csum!\n");
105                 if (!skb_partial_csum_set(skb,hdr->csum_start,hdr->csum_offset))
106                         goto frame_err;
107         }
108
109         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
110                 pr_debug("GSO!\n");
111                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
112                 case VIRTIO_NET_HDR_GSO_TCPV4:
113                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
114                         break;
115                 case VIRTIO_NET_HDR_GSO_UDP:
116                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
117                         break;
118                 case VIRTIO_NET_HDR_GSO_TCPV6:
119                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
120                         break;
121                 default:
122                         if (net_ratelimit())
123                                 printk(KERN_WARNING "%s: bad gso type %u.\n",
124                                        dev->name, hdr->gso_type);
125                         goto frame_err;
126                 }
127
128                 if (hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN)
129                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
130
131                 skb_shinfo(skb)->gso_size = hdr->gso_size;
132                 if (skb_shinfo(skb)->gso_size == 0) {
133                         if (net_ratelimit())
134                                 printk(KERN_WARNING "%s: zero gso size.\n",
135                                        dev->name);
136                         goto frame_err;
137                 }
138
139                 /* Header must be checked, and gso_segs computed. */
140                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
141                 skb_shinfo(skb)->gso_segs = 0;
142         }
143
144         netif_receive_skb(skb);
145         return;
146
147 frame_err:
148         dev->stats.rx_frame_errors++;
149 drop:
150         dev_kfree_skb(skb);
151 }
152
153 static void try_fill_recv(struct virtnet_info *vi)
154 {
155         struct sk_buff *skb;
156         struct scatterlist sg[2+MAX_SKB_FRAGS];
157         int num, err;
158
159         sg_init_table(sg, 2+MAX_SKB_FRAGS);
160         for (;;) {
161                 skb = netdev_alloc_skb(vi->dev, MAX_PACKET_LEN);
162                 if (unlikely(!skb))
163                         break;
164
165                 skb_put(skb, MAX_PACKET_LEN);
166                 vnet_hdr_to_sg(sg, skb);
167                 num = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
168                 skb_queue_head(&vi->recv, skb);
169
170                 err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, num, skb);
171                 if (err) {
172                         skb_unlink(skb, &vi->recv);
173                         kfree_skb(skb);
174                         break;
175                 }
176                 vi->num++;
177         }
178         if (unlikely(vi->num > vi->max))
179                 vi->max = vi->num;
180         vi->rvq->vq_ops->kick(vi->rvq);
181 }
182
183 static void skb_recv_done(struct virtqueue *rvq)
184 {
185         struct virtnet_info *vi = rvq->vdev->priv;
186         /* Schedule NAPI, Suppress further interrupts if successful. */
187         if (netif_rx_schedule_prep(vi->dev, &vi->napi)) {
188                 rvq->vq_ops->disable_cb(rvq);
189                 __netif_rx_schedule(vi->dev, &vi->napi);
190         }
191 }
192
193 static int virtnet_poll(struct napi_struct *napi, int budget)
194 {
195         struct virtnet_info *vi = container_of(napi, struct virtnet_info, napi);
196         struct sk_buff *skb = NULL;
197         unsigned int len, received = 0;
198
199 again:
200         while (received < budget &&
201                (skb = vi->rvq->vq_ops->get_buf(vi->rvq, &len)) != NULL) {
202                 __skb_unlink(skb, &vi->recv);
203                 receive_skb(vi->dev, skb, len);
204                 vi->num--;
205                 received++;
206         }
207
208         /* FIXME: If we oom and completely run out of inbufs, we need
209          * to start a timer trying to fill more. */
210         if (vi->num < vi->max / 2)
211                 try_fill_recv(vi);
212
213         /* Out of packets? */
214         if (received < budget) {
215                 netif_rx_complete(vi->dev, napi);
216                 if (unlikely(!vi->rvq->vq_ops->enable_cb(vi->rvq))
217                     && napi_schedule_prep(napi)) {
218                         vi->rvq->vq_ops->disable_cb(vi->rvq);
219                         __netif_rx_schedule(vi->dev, napi);
220                         goto again;
221                 }
222         }
223
224         return received;
225 }
226
227 static void free_old_xmit_skbs(struct virtnet_info *vi)
228 {
229         struct sk_buff *skb;
230         unsigned int len;
231
232         while ((skb = vi->svq->vq_ops->get_buf(vi->svq, &len)) != NULL) {
233                 pr_debug("Sent skb %p\n", skb);
234                 __skb_unlink(skb, &vi->send);
235                 vi->dev->stats.tx_bytes += skb->len;
236                 vi->dev->stats.tx_packets++;
237                 kfree_skb(skb);
238         }
239 }
240
241 static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
242 {
243         int num;
244         struct scatterlist sg[2+MAX_SKB_FRAGS];
245         struct virtio_net_hdr *hdr;
246         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
247
248         sg_init_table(sg, 2+MAX_SKB_FRAGS);
249
250         pr_debug("%s: xmit %p " MAC_FMT "\n", vi->dev->name, skb,
251                  dest[0], dest[1], dest[2],
252                  dest[3], dest[4], dest[5]);
253
254         /* Encode metadata header at front. */
255         hdr = skb_vnet_hdr(skb);
256         if (skb->ip_summed == CHECKSUM_PARTIAL) {
257                 hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
258                 hdr->csum_start = skb->csum_start - skb_headroom(skb);
259                 hdr->csum_offset = skb->csum_offset;
260         } else {
261                 hdr->flags = 0;
262                 hdr->csum_offset = hdr->csum_start = 0;
263         }
264
265         if (skb_is_gso(skb)) {
266                 hdr->hdr_len = skb_transport_header(skb) - skb->data;
267                 hdr->gso_size = skb_shinfo(skb)->gso_size;
268                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
269                         hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
270                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
271                         hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
272                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
273                         hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
274                 else
275                         BUG();
276                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
277                         hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
278         } else {
279                 hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
280                 hdr->gso_size = hdr->hdr_len = 0;
281         }
282
283         vnet_hdr_to_sg(sg, skb);
284         num = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
285
286         return vi->svq->vq_ops->add_buf(vi->svq, sg, num, 0, skb);
287 }
288
289 static void xmit_tasklet(unsigned long data)
290 {
291         struct virtnet_info *vi = (void *)data;
292
293         netif_tx_lock_bh(vi->dev);
294         if (vi->last_xmit_skb && xmit_skb(vi, vi->last_xmit_skb) == 0) {
295                 vi->svq->vq_ops->kick(vi->svq);
296                 vi->last_xmit_skb = NULL;
297         }
298         netif_tx_unlock_bh(vi->dev);
299 }
300
301 static int start_xmit(struct sk_buff *skb, struct net_device *dev)
302 {
303         struct virtnet_info *vi = netdev_priv(dev);
304
305 again:
306         /* Free up any pending old buffers before queueing new ones. */
307         free_old_xmit_skbs(vi);
308
309         /* If we has a buffer left over from last time, send it now. */
310         if (unlikely(vi->last_xmit_skb)) {
311                 if (xmit_skb(vi, vi->last_xmit_skb) != 0) {
312                         /* Drop this skb: we only queue one. */
313                         vi->dev->stats.tx_dropped++;
314                         kfree_skb(skb);
315                         skb = NULL;
316                         goto stop_queue;
317                 }
318                 vi->last_xmit_skb = NULL;
319         }
320
321         /* Put new one in send queue and do transmit */
322         if (likely(skb)) {
323                 __skb_queue_head(&vi->send, skb);
324                 if (xmit_skb(vi, skb) != 0) {
325                         vi->last_xmit_skb = skb;
326                         skb = NULL;
327                         goto stop_queue;
328                 }
329         }
330 done:
331         vi->svq->vq_ops->kick(vi->svq);
332         return NETDEV_TX_OK;
333
334 stop_queue:
335         pr_debug("%s: virtio not prepared to send\n", dev->name);
336         netif_stop_queue(dev);
337
338         /* Activate callback for using skbs: if this returns false it
339          * means some were used in the meantime. */
340         if (unlikely(!vi->svq->vq_ops->enable_cb(vi->svq))) {
341                 vi->svq->vq_ops->disable_cb(vi->svq);
342                 netif_start_queue(dev);
343                 goto again;
344         }
345         goto done;
346 }
347
348 #ifdef CONFIG_NET_POLL_CONTROLLER
349 static void virtnet_netpoll(struct net_device *dev)
350 {
351         struct virtnet_info *vi = netdev_priv(dev);
352
353         napi_schedule(&vi->napi);
354 }
355 #endif
356
357 static int virtnet_open(struct net_device *dev)
358 {
359         struct virtnet_info *vi = netdev_priv(dev);
360
361         napi_enable(&vi->napi);
362
363         /* If all buffers were filled by other side before we napi_enabled, we
364          * won't get another interrupt, so process any outstanding packets
365          * now.  virtnet_poll wants re-enable the queue, so we disable here.
366          * We synchronize against interrupts via NAPI_STATE_SCHED */
367         if (netif_rx_schedule_prep(dev, &vi->napi)) {
368                 vi->rvq->vq_ops->disable_cb(vi->rvq);
369                 __netif_rx_schedule(dev, &vi->napi);
370         }
371         return 0;
372 }
373
374 static int virtnet_close(struct net_device *dev)
375 {
376         struct virtnet_info *vi = netdev_priv(dev);
377
378         napi_disable(&vi->napi);
379
380         return 0;
381 }
382
383 static int virtnet_probe(struct virtio_device *vdev)
384 {
385         int err;
386         struct net_device *dev;
387         struct virtnet_info *vi;
388
389         /* Allocate ourselves a network device with room for our info */
390         dev = alloc_etherdev(sizeof(struct virtnet_info));
391         if (!dev)
392                 return -ENOMEM;
393
394         /* Set up network device as normal. */
395         dev->open = virtnet_open;
396         dev->stop = virtnet_close;
397         dev->hard_start_xmit = start_xmit;
398         dev->features = NETIF_F_HIGHDMA;
399 #ifdef CONFIG_NET_POLL_CONTROLLER
400         dev->poll_controller = virtnet_netpoll;
401 #endif
402         SET_NETDEV_DEV(dev, &vdev->dev);
403
404         /* Do we support "hardware" checksums? */
405         if (csum && virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
406                 /* This opens up the world of extra features. */
407                 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
408                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
409                         dev->features |= NETIF_F_TSO | NETIF_F_UFO
410                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
411                 }
412                 /* Individual feature bits: what can host handle? */
413                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
414                         dev->features |= NETIF_F_TSO;
415                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
416                         dev->features |= NETIF_F_TSO6;
417                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
418                         dev->features |= NETIF_F_TSO_ECN;
419                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
420                         dev->features |= NETIF_F_UFO;
421         }
422
423         /* Configuration may specify what MAC to use.  Otherwise random. */
424         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
425                 vdev->config->get(vdev,
426                                   offsetof(struct virtio_net_config, mac),
427                                   dev->dev_addr, dev->addr_len);
428         } else
429                 random_ether_addr(dev->dev_addr);
430
431         /* Set up our device-specific information */
432         vi = netdev_priv(dev);
433         netif_napi_add(dev, &vi->napi, virtnet_poll, napi_weight);
434         vi->dev = dev;
435         vi->vdev = vdev;
436         vdev->priv = vi;
437
438         /* We expect two virtqueues, receive then send. */
439         vi->rvq = vdev->config->find_vq(vdev, 0, skb_recv_done);
440         if (IS_ERR(vi->rvq)) {
441                 err = PTR_ERR(vi->rvq);
442                 goto free;
443         }
444
445         vi->svq = vdev->config->find_vq(vdev, 1, skb_xmit_done);
446         if (IS_ERR(vi->svq)) {
447                 err = PTR_ERR(vi->svq);
448                 goto free_recv;
449         }
450
451         /* Initialize our empty receive and send queues. */
452         skb_queue_head_init(&vi->recv);
453         skb_queue_head_init(&vi->send);
454
455         tasklet_init(&vi->tasklet, xmit_tasklet, (unsigned long)vi);
456
457         err = register_netdev(dev);
458         if (err) {
459                 pr_debug("virtio_net: registering device failed\n");
460                 goto free_send;
461         }
462
463         /* Last of all, set up some receive buffers. */
464         try_fill_recv(vi);
465
466         /* If we didn't even get one input buffer, we're useless. */
467         if (vi->num == 0) {
468                 err = -ENOMEM;
469                 goto unregister;
470         }
471
472         pr_debug("virtnet: registered device %s\n", dev->name);
473         return 0;
474
475 unregister:
476         unregister_netdev(dev);
477 free_send:
478         vdev->config->del_vq(vi->svq);
479 free_recv:
480         vdev->config->del_vq(vi->rvq);
481 free:
482         free_netdev(dev);
483         return err;
484 }
485
486 static void virtnet_remove(struct virtio_device *vdev)
487 {
488         struct virtnet_info *vi = vdev->priv;
489         struct sk_buff *skb;
490
491         /* Stop all the virtqueues. */
492         vdev->config->reset(vdev);
493
494         /* Free our skbs in send and recv queues, if any. */
495         while ((skb = __skb_dequeue(&vi->recv)) != NULL) {
496                 kfree_skb(skb);
497                 vi->num--;
498         }
499         __skb_queue_purge(&vi->send);
500
501         BUG_ON(vi->num != 0);
502
503         vdev->config->del_vq(vi->svq);
504         vdev->config->del_vq(vi->rvq);
505         unregister_netdev(vi->dev);
506         free_netdev(vi->dev);
507 }
508
509 static struct virtio_device_id id_table[] = {
510         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
511         { 0 },
512 };
513
514 static unsigned int features[] = {
515         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
516         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
517         VIRTIO_NET_F_HOST_ECN,
518 };
519
520 static struct virtio_driver virtio_net = {
521         .feature_table = features,
522         .feature_table_size = ARRAY_SIZE(features),
523         .driver.name =  KBUILD_MODNAME,
524         .driver.owner = THIS_MODULE,
525         .id_table =     id_table,
526         .probe =        virtnet_probe,
527         .remove =       __devexit_p(virtnet_remove),
528 };
529
530 static int __init init(void)
531 {
532         return register_virtio_driver(&virtio_net);
533 }
534
535 static void __exit fini(void)
536 {
537         unregister_virtio_driver(&virtio_net);
538 }
539 module_init(init);
540 module_exit(fini);
541
542 MODULE_DEVICE_TABLE(virtio, id_table);
543 MODULE_DESCRIPTION("Virtio network driver");
544 MODULE_LICENSE("GPL");