[IA64] Minimize per_cpu reservations.
[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 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         /* Number of input buffers, and max we've ever had. */
45         unsigned int num, max;
46
47         /* Receive & send queues. */
48         struct sk_buff_head recv;
49         struct sk_buff_head send;
50 };
51
52 static inline struct virtio_net_hdr *skb_vnet_hdr(struct sk_buff *skb)
53 {
54         return (struct virtio_net_hdr *)skb->cb;
55 }
56
57 static inline void vnet_hdr_to_sg(struct scatterlist *sg, struct sk_buff *skb)
58 {
59         sg_init_one(sg, skb_vnet_hdr(skb), sizeof(struct virtio_net_hdr));
60 }
61
62 static void skb_xmit_done(struct virtqueue *svq)
63 {
64         struct virtnet_info *vi = svq->vdev->priv;
65
66         /* Suppress further interrupts. */
67         svq->vq_ops->disable_cb(svq);
68         /* We were waiting for more output buffers. */
69         netif_wake_queue(vi->dev);
70 }
71
72 static void receive_skb(struct net_device *dev, struct sk_buff *skb,
73                         unsigned len)
74 {
75         struct virtio_net_hdr *hdr = skb_vnet_hdr(skb);
76
77         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
78                 pr_debug("%s: short packet %i\n", dev->name, len);
79                 dev->stats.rx_length_errors++;
80                 goto drop;
81         }
82         len -= sizeof(struct virtio_net_hdr);
83         BUG_ON(len > MAX_PACKET_LEN);
84
85         skb_trim(skb, len);
86         skb->protocol = eth_type_trans(skb, dev);
87         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
88                  ntohs(skb->protocol), skb->len, skb->pkt_type);
89         dev->stats.rx_bytes += skb->len;
90         dev->stats.rx_packets++;
91
92         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
93                 pr_debug("Needs csum!\n");
94                 if (!skb_partial_csum_set(skb,hdr->csum_start,hdr->csum_offset))
95                         goto frame_err;
96         }
97
98         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
99                 pr_debug("GSO!\n");
100                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
101                 case VIRTIO_NET_HDR_GSO_TCPV4:
102                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
103                         break;
104                 case VIRTIO_NET_HDR_GSO_UDP:
105                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
106                         break;
107                 case VIRTIO_NET_HDR_GSO_TCPV6:
108                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
109                         break;
110                 default:
111                         if (net_ratelimit())
112                                 printk(KERN_WARNING "%s: bad gso type %u.\n",
113                                        dev->name, hdr->gso_type);
114                         goto frame_err;
115                 }
116
117                 if (hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN)
118                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
119
120                 skb_shinfo(skb)->gso_size = hdr->gso_size;
121                 if (skb_shinfo(skb)->gso_size == 0) {
122                         if (net_ratelimit())
123                                 printk(KERN_WARNING "%s: zero gso size.\n",
124                                        dev->name);
125                         goto frame_err;
126                 }
127
128                 /* Header must be checked, and gso_segs computed. */
129                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
130                 skb_shinfo(skb)->gso_segs = 0;
131         }
132
133         netif_receive_skb(skb);
134         return;
135
136 frame_err:
137         dev->stats.rx_frame_errors++;
138 drop:
139         dev_kfree_skb(skb);
140 }
141
142 static void try_fill_recv(struct virtnet_info *vi)
143 {
144         struct sk_buff *skb;
145         struct scatterlist sg[1+MAX_SKB_FRAGS];
146         int num, err;
147
148         sg_init_table(sg, 1+MAX_SKB_FRAGS);
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 void skb_recv_done(struct virtqueue *rvq)
173 {
174         struct virtnet_info *vi = rvq->vdev->priv;
175         /* Schedule NAPI, Suppress further interrupts if successful. */
176         if (netif_rx_schedule_prep(vi->dev, &vi->napi)) {
177                 rvq->vq_ops->disable_cb(rvq);
178                 __netif_rx_schedule(vi->dev, &vi->napi);
179         }
180 }
181
182 static int virtnet_poll(struct napi_struct *napi, int budget)
183 {
184         struct virtnet_info *vi = container_of(napi, struct virtnet_info, napi);
185         struct sk_buff *skb = NULL;
186         unsigned int len, received = 0;
187
188 again:
189         while (received < budget &&
190                (skb = vi->rvq->vq_ops->get_buf(vi->rvq, &len)) != NULL) {
191                 __skb_unlink(skb, &vi->recv);
192                 receive_skb(vi->dev, skb, len);
193                 vi->num--;
194                 received++;
195         }
196
197         /* FIXME: If we oom and completely run out of inbufs, we need
198          * to start a timer trying to fill more. */
199         if (vi->num < vi->max / 2)
200                 try_fill_recv(vi);
201
202         /* Out of packets? */
203         if (received < budget) {
204                 netif_rx_complete(vi->dev, napi);
205                 if (unlikely(!vi->rvq->vq_ops->enable_cb(vi->rvq))
206                     && napi_schedule_prep(napi)) {
207                         vi->rvq->vq_ops->disable_cb(vi->rvq);
208                         __netif_rx_schedule(vi->dev, napi);
209                         goto again;
210                 }
211         }
212
213         return received;
214 }
215
216 static void free_old_xmit_skbs(struct virtnet_info *vi)
217 {
218         struct sk_buff *skb;
219         unsigned int len;
220
221         while ((skb = vi->svq->vq_ops->get_buf(vi->svq, &len)) != NULL) {
222                 pr_debug("Sent skb %p\n", skb);
223                 __skb_unlink(skb, &vi->send);
224                 vi->dev->stats.tx_bytes += len;
225                 vi->dev->stats.tx_packets++;
226                 kfree_skb(skb);
227         }
228 }
229
230 static int start_xmit(struct sk_buff *skb, struct net_device *dev)
231 {
232         struct virtnet_info *vi = netdev_priv(dev);
233         int num, err;
234         struct scatterlist sg[1+MAX_SKB_FRAGS];
235         struct virtio_net_hdr *hdr;
236         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
237         DECLARE_MAC_BUF(mac);
238
239         sg_init_table(sg, 1+MAX_SKB_FRAGS);
240
241         pr_debug("%s: xmit %p %s\n", dev->name, skb, print_mac(mac, dest));
242
243         /* Encode metadata header at front. */
244         hdr = skb_vnet_hdr(skb);
245         if (skb->ip_summed == CHECKSUM_PARTIAL) {
246                 hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
247                 hdr->csum_start = skb->csum_start - skb_headroom(skb);
248                 hdr->csum_offset = skb->csum_offset;
249         } else {
250                 hdr->flags = 0;
251                 hdr->csum_offset = hdr->csum_start = 0;
252         }
253
254         if (skb_is_gso(skb)) {
255                 hdr->hdr_len = skb_transport_header(skb) - skb->data;
256                 hdr->gso_size = skb_shinfo(skb)->gso_size;
257                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
258                         hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
259                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
260                         hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
261                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
262                         hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
263                 else
264                         BUG();
265                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
266                         hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
267         } else {
268                 hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
269                 hdr->gso_size = hdr->hdr_len = 0;
270         }
271
272         vnet_hdr_to_sg(sg, skb);
273         num = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
274         __skb_queue_head(&vi->send, skb);
275
276 again:
277         /* Free up any pending old buffers before queueing new ones. */
278         free_old_xmit_skbs(vi);
279         err = vi->svq->vq_ops->add_buf(vi->svq, sg, num, 0, skb);
280         if (err) {
281                 pr_debug("%s: virtio not prepared to send\n", dev->name);
282                 netif_stop_queue(dev);
283
284                 /* Activate callback for using skbs: if this returns false it
285                  * means some were used in the meantime. */
286                 if (unlikely(!vi->svq->vq_ops->enable_cb(vi->svq))) {
287                         vi->svq->vq_ops->disable_cb(vi->svq);
288                         netif_start_queue(dev);
289                         goto again;
290                 }
291                 __skb_unlink(skb, &vi->send);
292
293                 return NETDEV_TX_BUSY;
294         }
295         vi->svq->vq_ops->kick(vi->svq);
296
297         return 0;
298 }
299
300 #ifdef CONFIG_NET_POLL_CONTROLLER
301 static void virtnet_netpoll(struct net_device *dev)
302 {
303         struct virtnet_info *vi = netdev_priv(dev);
304
305         napi_schedule(&vi->napi);
306 }
307 #endif
308
309 static int virtnet_open(struct net_device *dev)
310 {
311         struct virtnet_info *vi = netdev_priv(dev);
312
313         napi_enable(&vi->napi);
314
315         /* If all buffers were filled by other side before we napi_enabled, we
316          * won't get another interrupt, so process any outstanding packets
317          * now.  virtnet_poll wants re-enable the queue, so we disable here.
318          * We synchronize against interrupts via NAPI_STATE_SCHED */
319         if (netif_rx_schedule_prep(dev, &vi->napi)) {
320                 vi->rvq->vq_ops->disable_cb(vi->rvq);
321                 __netif_rx_schedule(dev, &vi->napi);
322         }
323         return 0;
324 }
325
326 static int virtnet_close(struct net_device *dev)
327 {
328         struct virtnet_info *vi = netdev_priv(dev);
329
330         napi_disable(&vi->napi);
331
332         return 0;
333 }
334
335 static int virtnet_probe(struct virtio_device *vdev)
336 {
337         int err;
338         struct net_device *dev;
339         struct virtnet_info *vi;
340
341         /* Allocate ourselves a network device with room for our info */
342         dev = alloc_etherdev(sizeof(struct virtnet_info));
343         if (!dev)
344                 return -ENOMEM;
345
346         /* Set up network device as normal. */
347         dev->open = virtnet_open;
348         dev->stop = virtnet_close;
349         dev->hard_start_xmit = start_xmit;
350         dev->features = NETIF_F_HIGHDMA;
351 #ifdef CONFIG_NET_POLL_CONTROLLER
352         dev->poll_controller = virtnet_netpoll;
353 #endif
354         SET_NETDEV_DEV(dev, &vdev->dev);
355
356         /* Do we support "hardware" checksums? */
357         if (csum && vdev->config->feature(vdev, VIRTIO_NET_F_CSUM)) {
358                 /* This opens up the world of extra features. */
359                 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
360                 if (gso && vdev->config->feature(vdev, VIRTIO_NET_F_GSO)) {
361                         dev->features |= NETIF_F_TSO | NETIF_F_UFO
362                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
363                 }
364         }
365
366         /* Configuration may specify what MAC to use.  Otherwise random. */
367         if (vdev->config->feature(vdev, VIRTIO_NET_F_MAC)) {
368                 vdev->config->get(vdev,
369                                   offsetof(struct virtio_net_config, mac),
370                                   dev->dev_addr, dev->addr_len);
371         } else
372                 random_ether_addr(dev->dev_addr);
373
374         /* Set up our device-specific information */
375         vi = netdev_priv(dev);
376         netif_napi_add(dev, &vi->napi, virtnet_poll, napi_weight);
377         vi->dev = dev;
378         vi->vdev = vdev;
379         vdev->priv = vi;
380
381         /* We expect two virtqueues, receive then send. */
382         vi->rvq = vdev->config->find_vq(vdev, 0, skb_recv_done);
383         if (IS_ERR(vi->rvq)) {
384                 err = PTR_ERR(vi->rvq);
385                 goto free;
386         }
387
388         vi->svq = vdev->config->find_vq(vdev, 1, skb_xmit_done);
389         if (IS_ERR(vi->svq)) {
390                 err = PTR_ERR(vi->svq);
391                 goto free_recv;
392         }
393
394         /* Initialize our empty receive and send queues. */
395         skb_queue_head_init(&vi->recv);
396         skb_queue_head_init(&vi->send);
397
398         err = register_netdev(dev);
399         if (err) {
400                 pr_debug("virtio_net: registering device failed\n");
401                 goto free_send;
402         }
403
404         /* Last of all, set up some receive buffers. */
405         try_fill_recv(vi);
406
407         /* If we didn't even get one input buffer, we're useless. */
408         if (vi->num == 0) {
409                 err = -ENOMEM;
410                 goto unregister;
411         }
412
413         pr_debug("virtnet: registered device %s\n", dev->name);
414         return 0;
415
416 unregister:
417         unregister_netdev(dev);
418 free_send:
419         vdev->config->del_vq(vi->svq);
420 free_recv:
421         vdev->config->del_vq(vi->rvq);
422 free:
423         free_netdev(dev);
424         return err;
425 }
426
427 static void virtnet_remove(struct virtio_device *vdev)
428 {
429         struct virtnet_info *vi = vdev->priv;
430         struct sk_buff *skb;
431
432         /* Stop all the virtqueues. */
433         vdev->config->reset(vdev);
434
435         /* Free our skbs in send and recv queues, if any. */
436         while ((skb = __skb_dequeue(&vi->recv)) != NULL) {
437                 kfree_skb(skb);
438                 vi->num--;
439         }
440         while ((skb = __skb_dequeue(&vi->send)) != NULL)
441                 kfree_skb(skb);
442
443         BUG_ON(vi->num != 0);
444
445         vdev->config->del_vq(vi->svq);
446         vdev->config->del_vq(vi->rvq);
447         unregister_netdev(vi->dev);
448         free_netdev(vi->dev);
449 }
450
451 static struct virtio_device_id id_table[] = {
452         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
453         { 0 },
454 };
455
456 static struct virtio_driver virtio_net = {
457         .driver.name =  KBUILD_MODNAME,
458         .driver.owner = THIS_MODULE,
459         .id_table =     id_table,
460         .probe =        virtnet_probe,
461         .remove =       __devexit_p(virtnet_remove),
462 };
463
464 static int __init init(void)
465 {
466         return register_virtio_driver(&virtio_net);
467 }
468
469 static void __exit fini(void)
470 {
471         unregister_virtio_driver(&virtio_net);
472 }
473 module_init(init);
474 module_exit(fini);
475
476 MODULE_DEVICE_TABLE(virtio, id_table);
477 MODULE_DESCRIPTION("Virtio network driver");
478 MODULE_LICENSE("GPL");