virtio-net: move sg off stack
[linux-2.6.git] / drivers / net / virtio_net.c
1 /* A 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/ethtool.h>
23 #include <linux/module.h>
24 #include <linux/virtio.h>
25 #include <linux/virtio_net.h>
26 #include <linux/scatterlist.h>
27 #include <linux/if_vlan.h>
28
29 static int napi_weight = 128;
30 module_param(napi_weight, int, 0444);
31
32 static int csum = 1, gso = 1;
33 module_param(csum, bool, 0444);
34 module_param(gso, bool, 0444);
35
36 /* FIXME: MTU in config. */
37 #define MAX_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
38 #define GOOD_COPY_LEN   128
39
40 #define VIRTNET_SEND_COMMAND_SG_MAX    2
41
42 struct virtnet_info {
43         struct virtio_device *vdev;
44         struct virtqueue *rvq, *svq, *cvq;
45         struct net_device *dev;
46         struct napi_struct napi;
47         unsigned int status;
48
49         /* Number of input buffers, and max we've ever had. */
50         unsigned int num, max;
51
52         /* I like... big packets and I cannot lie! */
53         bool big_packets;
54
55         /* Host will merge rx buffers for big packets (shake it! shake it!) */
56         bool mergeable_rx_bufs;
57
58         /* Work struct for refilling if we run low on memory. */
59         struct delayed_work refill;
60
61         /* Chain pages by the private ptr. */
62         struct page *pages;
63
64         /* fragments + linear part + virtio header */
65         struct scatterlist rx_sg[MAX_SKB_FRAGS + 2];
66         struct scatterlist tx_sg[MAX_SKB_FRAGS + 2];
67 };
68
69 struct skb_vnet_hdr {
70         union {
71                 struct virtio_net_hdr hdr;
72                 struct virtio_net_hdr_mrg_rxbuf mhdr;
73         };
74         unsigned int num_sg;
75 };
76
77 struct padded_vnet_hdr {
78         struct virtio_net_hdr hdr;
79         /*
80          * virtio_net_hdr should be in a separated sg buffer because of a
81          * QEMU bug, and data sg buffer shares same page with this header sg.
82          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
83          */
84         char padding[6];
85 };
86
87 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
88 {
89         return (struct skb_vnet_hdr *)skb->cb;
90 }
91
92 /*
93  * private is used to chain pages for big packets, put the whole
94  * most recent used list in the beginning for reuse
95  */
96 static void give_pages(struct virtnet_info *vi, struct page *page)
97 {
98         struct page *end;
99
100         /* Find end of list, sew whole thing into vi->pages. */
101         for (end = page; end->private; end = (struct page *)end->private);
102         end->private = (unsigned long)vi->pages;
103         vi->pages = page;
104 }
105
106 static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
107 {
108         struct page *p = vi->pages;
109
110         if (p) {
111                 vi->pages = (struct page *)p->private;
112                 /* clear private here, it is used to chain pages */
113                 p->private = 0;
114         } else
115                 p = alloc_page(gfp_mask);
116         return p;
117 }
118
119 static void skb_xmit_done(struct virtqueue *svq)
120 {
121         struct virtnet_info *vi = svq->vdev->priv;
122
123         /* Suppress further interrupts. */
124         svq->vq_ops->disable_cb(svq);
125
126         /* We were probably waiting for more output buffers. */
127         netif_wake_queue(vi->dev);
128 }
129
130 static void set_skb_frag(struct sk_buff *skb, struct page *page,
131                          unsigned int offset, unsigned int *len)
132 {
133         int i = skb_shinfo(skb)->nr_frags;
134         skb_frag_t *f;
135
136         f = &skb_shinfo(skb)->frags[i];
137         f->size = min((unsigned)PAGE_SIZE - offset, *len);
138         f->page_offset = offset;
139         f->page = page;
140
141         skb->data_len += f->size;
142         skb->len += f->size;
143         skb_shinfo(skb)->nr_frags++;
144         *len -= f->size;
145 }
146
147 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
148                                    struct page *page, unsigned int len)
149 {
150         struct sk_buff *skb;
151         struct skb_vnet_hdr *hdr;
152         unsigned int copy, hdr_len, offset;
153         char *p;
154
155         p = page_address(page);
156
157         /* copy small packet so we can reuse these pages for small data */
158         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
159         if (unlikely(!skb))
160                 return NULL;
161
162         hdr = skb_vnet_hdr(skb);
163
164         if (vi->mergeable_rx_bufs) {
165                 hdr_len = sizeof hdr->mhdr;
166                 offset = hdr_len;
167         } else {
168                 hdr_len = sizeof hdr->hdr;
169                 offset = sizeof(struct padded_vnet_hdr);
170         }
171
172         memcpy(hdr, p, hdr_len);
173
174         len -= hdr_len;
175         p += offset;
176
177         copy = len;
178         if (copy > skb_tailroom(skb))
179                 copy = skb_tailroom(skb);
180         memcpy(skb_put(skb, copy), p, copy);
181
182         len -= copy;
183         offset += copy;
184
185         while (len) {
186                 set_skb_frag(skb, page, offset, &len);
187                 page = (struct page *)page->private;
188                 offset = 0;
189         }
190
191         if (page)
192                 give_pages(vi, page);
193
194         return skb;
195 }
196
197 static int receive_mergeable(struct virtnet_info *vi, struct sk_buff *skb)
198 {
199         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
200         struct page *page;
201         int num_buf, i, len;
202
203         num_buf = hdr->mhdr.num_buffers;
204         while (--num_buf) {
205                 i = skb_shinfo(skb)->nr_frags;
206                 if (i >= MAX_SKB_FRAGS) {
207                         pr_debug("%s: packet too long\n", skb->dev->name);
208                         skb->dev->stats.rx_length_errors++;
209                         return -EINVAL;
210                 }
211
212                 page = vi->rvq->vq_ops->get_buf(vi->rvq, &len);
213                 if (!page) {
214                         pr_debug("%s: rx error: %d buffers missing\n",
215                                  skb->dev->name, hdr->mhdr.num_buffers);
216                         skb->dev->stats.rx_length_errors++;
217                         return -EINVAL;
218                 }
219                 if (len > PAGE_SIZE)
220                         len = PAGE_SIZE;
221
222                 set_skb_frag(skb, page, 0, &len);
223
224                 --vi->num;
225         }
226         return 0;
227 }
228
229 static void receive_buf(struct net_device *dev, void *buf, unsigned int len)
230 {
231         struct virtnet_info *vi = netdev_priv(dev);
232         struct sk_buff *skb;
233         struct page *page;
234         struct skb_vnet_hdr *hdr;
235
236         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
237                 pr_debug("%s: short packet %i\n", dev->name, len);
238                 dev->stats.rx_length_errors++;
239                 if (vi->mergeable_rx_bufs || vi->big_packets)
240                         give_pages(vi, buf);
241                 else
242                         dev_kfree_skb(buf);
243                 return;
244         }
245
246         if (!vi->mergeable_rx_bufs && !vi->big_packets) {
247                 skb = buf;
248                 len -= sizeof(struct virtio_net_hdr);
249                 skb_trim(skb, len);
250         } else {
251                 page = buf;
252                 skb = page_to_skb(vi, page, len);
253                 if (unlikely(!skb)) {
254                         dev->stats.rx_dropped++;
255                         give_pages(vi, page);
256                         return;
257                 }
258                 if (vi->mergeable_rx_bufs)
259                         if (receive_mergeable(vi, skb)) {
260                                 dev_kfree_skb(skb);
261                                 return;
262                         }
263         }
264
265         hdr = skb_vnet_hdr(skb);
266         skb->truesize += skb->data_len;
267         dev->stats.rx_bytes += skb->len;
268         dev->stats.rx_packets++;
269
270         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
271                 pr_debug("Needs csum!\n");
272                 if (!skb_partial_csum_set(skb,
273                                           hdr->hdr.csum_start,
274                                           hdr->hdr.csum_offset))
275                         goto frame_err;
276         }
277
278         skb->protocol = eth_type_trans(skb, dev);
279         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
280                  ntohs(skb->protocol), skb->len, skb->pkt_type);
281
282         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
283                 pr_debug("GSO!\n");
284                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
285                 case VIRTIO_NET_HDR_GSO_TCPV4:
286                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
287                         break;
288                 case VIRTIO_NET_HDR_GSO_UDP:
289                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
290                         break;
291                 case VIRTIO_NET_HDR_GSO_TCPV6:
292                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
293                         break;
294                 default:
295                         if (net_ratelimit())
296                                 printk(KERN_WARNING "%s: bad gso type %u.\n",
297                                        dev->name, hdr->hdr.gso_type);
298                         goto frame_err;
299                 }
300
301                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
302                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
303
304                 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
305                 if (skb_shinfo(skb)->gso_size == 0) {
306                         if (net_ratelimit())
307                                 printk(KERN_WARNING "%s: zero gso size.\n",
308                                        dev->name);
309                         goto frame_err;
310                 }
311
312                 /* Header must be checked, and gso_segs computed. */
313                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
314                 skb_shinfo(skb)->gso_segs = 0;
315         }
316
317         netif_receive_skb(skb);
318         return;
319
320 frame_err:
321         dev->stats.rx_frame_errors++;
322         dev_kfree_skb(skb);
323 }
324
325 static int add_recvbuf_small(struct virtnet_info *vi, gfp_t gfp)
326 {
327         struct sk_buff *skb;
328         struct skb_vnet_hdr *hdr;
329         int err;
330
331         skb = netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN);
332         if (unlikely(!skb))
333                 return -ENOMEM;
334
335         skb_put(skb, MAX_PACKET_LEN);
336
337         hdr = skb_vnet_hdr(skb);
338         sg_set_buf(vi->rx_sg, &hdr->hdr, sizeof hdr->hdr);
339
340         skb_to_sgvec(skb, vi->rx_sg + 1, 0, skb->len);
341
342         err = vi->rvq->vq_ops->add_buf(vi->rvq, vi->rx_sg, 0, 2, skb);
343         if (err < 0)
344                 dev_kfree_skb(skb);
345
346         return err;
347 }
348
349 static int add_recvbuf_big(struct virtnet_info *vi, gfp_t gfp)
350 {
351         struct page *first, *list = NULL;
352         char *p;
353         int i, err, offset;
354
355         /* page in vi->rx_sg[MAX_SKB_FRAGS + 1] is list tail */
356         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
357                 first = get_a_page(vi, gfp);
358                 if (!first) {
359                         if (list)
360                                 give_pages(vi, list);
361                         return -ENOMEM;
362                 }
363                 sg_set_buf(&vi->rx_sg[i], page_address(first), PAGE_SIZE);
364
365                 /* chain new page in list head to match sg */
366                 first->private = (unsigned long)list;
367                 list = first;
368         }
369
370         first = get_a_page(vi, gfp);
371         if (!first) {
372                 give_pages(vi, list);
373                 return -ENOMEM;
374         }
375         p = page_address(first);
376
377         /* vi->rx_sg[0], vi->rx_sg[1] share the same page */
378         /* a separated vi->rx_sg[0] for virtio_net_hdr only due to QEMU bug */
379         sg_set_buf(&vi->rx_sg[0], p, sizeof(struct virtio_net_hdr));
380
381         /* vi->rx_sg[1] for data packet, from offset */
382         offset = sizeof(struct padded_vnet_hdr);
383         sg_set_buf(&vi->rx_sg[1], p + offset, PAGE_SIZE - offset);
384
385         /* chain first in list head */
386         first->private = (unsigned long)list;
387         err = vi->rvq->vq_ops->add_buf(vi->rvq, vi->rx_sg, 0, MAX_SKB_FRAGS + 2,
388                                        first);
389         if (err < 0)
390                 give_pages(vi, first);
391
392         return err;
393 }
394
395 static int add_recvbuf_mergeable(struct virtnet_info *vi, gfp_t gfp)
396 {
397         struct page *page;
398         int err;
399
400         page = get_a_page(vi, gfp);
401         if (!page)
402                 return -ENOMEM;
403
404         sg_init_one(vi->rx_sg, page_address(page), PAGE_SIZE);
405
406         err = vi->rvq->vq_ops->add_buf(vi->rvq, vi->rx_sg, 0, 1, page);
407         if (err < 0)
408                 give_pages(vi, page);
409
410         return err;
411 }
412
413 /* Returns false if we couldn't fill entirely (OOM). */
414 static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
415 {
416         int err;
417         bool oom = false;
418
419         do {
420                 if (vi->mergeable_rx_bufs)
421                         err = add_recvbuf_mergeable(vi, gfp);
422                 else if (vi->big_packets)
423                         err = add_recvbuf_big(vi, gfp);
424                 else
425                         err = add_recvbuf_small(vi, gfp);
426
427                 if (err < 0) {
428                         oom = true;
429                         break;
430                 }
431                 ++vi->num;
432         } while (err > 0);
433         if (unlikely(vi->num > vi->max))
434                 vi->max = vi->num;
435         vi->rvq->vq_ops->kick(vi->rvq);
436         return !oom;
437 }
438
439 static void skb_recv_done(struct virtqueue *rvq)
440 {
441         struct virtnet_info *vi = rvq->vdev->priv;
442         /* Schedule NAPI, Suppress further interrupts if successful. */
443         if (napi_schedule_prep(&vi->napi)) {
444                 rvq->vq_ops->disable_cb(rvq);
445                 __napi_schedule(&vi->napi);
446         }
447 }
448
449 static void refill_work(struct work_struct *work)
450 {
451         struct virtnet_info *vi;
452         bool still_empty;
453
454         vi = container_of(work, struct virtnet_info, refill.work);
455         napi_disable(&vi->napi);
456         still_empty = !try_fill_recv(vi, GFP_KERNEL);
457         napi_enable(&vi->napi);
458
459         /* In theory, this can happen: if we don't get any buffers in
460          * we will *never* try to fill again. */
461         if (still_empty)
462                 schedule_delayed_work(&vi->refill, HZ/2);
463 }
464
465 static int virtnet_poll(struct napi_struct *napi, int budget)
466 {
467         struct virtnet_info *vi = container_of(napi, struct virtnet_info, napi);
468         void *buf;
469         unsigned int len, received = 0;
470
471 again:
472         while (received < budget &&
473                (buf = vi->rvq->vq_ops->get_buf(vi->rvq, &len)) != NULL) {
474                 receive_buf(vi->dev, buf, len);
475                 --vi->num;
476                 received++;
477         }
478
479         if (vi->num < vi->max / 2) {
480                 if (!try_fill_recv(vi, GFP_ATOMIC))
481                         schedule_delayed_work(&vi->refill, 0);
482         }
483
484         /* Out of packets? */
485         if (received < budget) {
486                 napi_complete(napi);
487                 if (unlikely(!vi->rvq->vq_ops->enable_cb(vi->rvq)) &&
488                     napi_schedule_prep(napi)) {
489                         vi->rvq->vq_ops->disable_cb(vi->rvq);
490                         __napi_schedule(napi);
491                         goto again;
492                 }
493         }
494
495         return received;
496 }
497
498 static unsigned int free_old_xmit_skbs(struct virtnet_info *vi)
499 {
500         struct sk_buff *skb;
501         unsigned int len, tot_sgs = 0;
502
503         while ((skb = vi->svq->vq_ops->get_buf(vi->svq, &len)) != NULL) {
504                 pr_debug("Sent skb %p\n", skb);
505                 vi->dev->stats.tx_bytes += skb->len;
506                 vi->dev->stats.tx_packets++;
507                 tot_sgs += skb_vnet_hdr(skb)->num_sg;
508                 dev_kfree_skb_any(skb);
509         }
510         return tot_sgs;
511 }
512
513 static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
514 {
515         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
516         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
517
518         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
519
520         if (skb->ip_summed == CHECKSUM_PARTIAL) {
521                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
522                 hdr->hdr.csum_start = skb->csum_start - skb_headroom(skb);
523                 hdr->hdr.csum_offset = skb->csum_offset;
524         } else {
525                 hdr->hdr.flags = 0;
526                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
527         }
528
529         if (skb_is_gso(skb)) {
530                 hdr->hdr.hdr_len = skb_headlen(skb);
531                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
532                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
533                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
534                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
535                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
536                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
537                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
538                 else
539                         BUG();
540                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
541                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
542         } else {
543                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
544                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
545         }
546
547         hdr->mhdr.num_buffers = 0;
548
549         /* Encode metadata header at front. */
550         if (vi->mergeable_rx_bufs)
551                 sg_set_buf(vi->tx_sg, &hdr->mhdr, sizeof hdr->mhdr);
552         else
553                 sg_set_buf(vi->tx_sg, &hdr->hdr, sizeof hdr->hdr);
554
555         hdr->num_sg = skb_to_sgvec(skb, vi->tx_sg + 1, 0, skb->len) + 1;
556         return vi->svq->vq_ops->add_buf(vi->svq, vi->tx_sg, hdr->num_sg,
557                                         0, skb);
558 }
559
560 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
561 {
562         struct virtnet_info *vi = netdev_priv(dev);
563         int capacity;
564
565 again:
566         /* Free up any pending old buffers before queueing new ones. */
567         free_old_xmit_skbs(vi);
568
569         /* Try to transmit */
570         capacity = xmit_skb(vi, skb);
571
572         /* This can happen with OOM and indirect buffers. */
573         if (unlikely(capacity < 0)) {
574                 netif_stop_queue(dev);
575                 dev_warn(&dev->dev, "Unexpected full queue\n");
576                 if (unlikely(!vi->svq->vq_ops->enable_cb(vi->svq))) {
577                         vi->svq->vq_ops->disable_cb(vi->svq);
578                         netif_start_queue(dev);
579                         goto again;
580                 }
581                 return NETDEV_TX_BUSY;
582         }
583         vi->svq->vq_ops->kick(vi->svq);
584
585         /* Don't wait up for transmitted skbs to be freed. */
586         skb_orphan(skb);
587         nf_reset(skb);
588
589         /* Apparently nice girls don't return TX_BUSY; stop the queue
590          * before it gets out of hand.  Naturally, this wastes entries. */
591         if (capacity < 2+MAX_SKB_FRAGS) {
592                 netif_stop_queue(dev);
593                 if (unlikely(!vi->svq->vq_ops->enable_cb(vi->svq))) {
594                         /* More just got used, free them then recheck. */
595                         capacity += free_old_xmit_skbs(vi);
596                         if (capacity >= 2+MAX_SKB_FRAGS) {
597                                 netif_start_queue(dev);
598                                 vi->svq->vq_ops->disable_cb(vi->svq);
599                         }
600                 }
601         }
602
603         return NETDEV_TX_OK;
604 }
605
606 static int virtnet_set_mac_address(struct net_device *dev, void *p)
607 {
608         struct virtnet_info *vi = netdev_priv(dev);
609         struct virtio_device *vdev = vi->vdev;
610         int ret;
611
612         ret = eth_mac_addr(dev, p);
613         if (ret)
614                 return ret;
615
616         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
617                 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
618                                   dev->dev_addr, dev->addr_len);
619
620         return 0;
621 }
622
623 #ifdef CONFIG_NET_POLL_CONTROLLER
624 static void virtnet_netpoll(struct net_device *dev)
625 {
626         struct virtnet_info *vi = netdev_priv(dev);
627
628         napi_schedule(&vi->napi);
629 }
630 #endif
631
632 static int virtnet_open(struct net_device *dev)
633 {
634         struct virtnet_info *vi = netdev_priv(dev);
635
636         napi_enable(&vi->napi);
637
638         /* If all buffers were filled by other side before we napi_enabled, we
639          * won't get another interrupt, so process any outstanding packets
640          * now.  virtnet_poll wants re-enable the queue, so we disable here.
641          * We synchronize against interrupts via NAPI_STATE_SCHED */
642         if (napi_schedule_prep(&vi->napi)) {
643                 vi->rvq->vq_ops->disable_cb(vi->rvq);
644                 __napi_schedule(&vi->napi);
645         }
646         return 0;
647 }
648
649 /*
650  * Send command via the control virtqueue and check status.  Commands
651  * supported by the hypervisor, as indicated by feature bits, should
652  * never fail unless improperly formated.
653  */
654 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
655                                  struct scatterlist *data, int out, int in)
656 {
657         struct scatterlist *s, sg[VIRTNET_SEND_COMMAND_SG_MAX + 2];
658         struct virtio_net_ctrl_hdr ctrl;
659         virtio_net_ctrl_ack status = ~0;
660         unsigned int tmp;
661         int i;
662
663         /* Caller should know better */
664         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ||
665                 (out + in > VIRTNET_SEND_COMMAND_SG_MAX));
666
667         out++; /* Add header */
668         in++; /* Add return status */
669
670         ctrl.class = class;
671         ctrl.cmd = cmd;
672
673         sg_init_table(sg, out + in);
674
675         sg_set_buf(&sg[0], &ctrl, sizeof(ctrl));
676         for_each_sg(data, s, out + in - 2, i)
677                 sg_set_buf(&sg[i + 1], sg_virt(s), s->length);
678         sg_set_buf(&sg[out + in - 1], &status, sizeof(status));
679
680         BUG_ON(vi->cvq->vq_ops->add_buf(vi->cvq, sg, out, in, vi) < 0);
681
682         vi->cvq->vq_ops->kick(vi->cvq);
683
684         /*
685          * Spin for a response, the kick causes an ioport write, trapping
686          * into the hypervisor, so the request should be handled immediately.
687          */
688         while (!vi->cvq->vq_ops->get_buf(vi->cvq, &tmp))
689                 cpu_relax();
690
691         return status == VIRTIO_NET_OK;
692 }
693
694 static int virtnet_close(struct net_device *dev)
695 {
696         struct virtnet_info *vi = netdev_priv(dev);
697
698         napi_disable(&vi->napi);
699
700         return 0;
701 }
702
703 static int virtnet_set_tx_csum(struct net_device *dev, u32 data)
704 {
705         struct virtnet_info *vi = netdev_priv(dev);
706         struct virtio_device *vdev = vi->vdev;
707
708         if (data && !virtio_has_feature(vdev, VIRTIO_NET_F_CSUM))
709                 return -ENOSYS;
710
711         return ethtool_op_set_tx_hw_csum(dev, data);
712 }
713
714 static void virtnet_set_rx_mode(struct net_device *dev)
715 {
716         struct virtnet_info *vi = netdev_priv(dev);
717         struct scatterlist sg[2];
718         u8 promisc, allmulti;
719         struct virtio_net_ctrl_mac *mac_data;
720         struct netdev_hw_addr *ha;
721         int uc_count;
722         int mc_count;
723         void *buf;
724         int i;
725
726         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
727         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
728                 return;
729
730         promisc = ((dev->flags & IFF_PROMISC) != 0);
731         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
732
733         sg_init_one(sg, &promisc, sizeof(promisc));
734
735         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
736                                   VIRTIO_NET_CTRL_RX_PROMISC,
737                                   sg, 1, 0))
738                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
739                          promisc ? "en" : "dis");
740
741         sg_init_one(sg, &allmulti, sizeof(allmulti));
742
743         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
744                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
745                                   sg, 1, 0))
746                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
747                          allmulti ? "en" : "dis");
748
749         uc_count = netdev_uc_count(dev);
750         mc_count = netdev_mc_count(dev);
751         /* MAC filter - use one buffer for both lists */
752         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
753                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
754         mac_data = buf;
755         if (!buf) {
756                 dev_warn(&dev->dev, "No memory for MAC address buffer\n");
757                 return;
758         }
759
760         sg_init_table(sg, 2);
761
762         /* Store the unicast list and count in the front of the buffer */
763         mac_data->entries = uc_count;
764         i = 0;
765         netdev_for_each_uc_addr(ha, dev)
766                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
767
768         sg_set_buf(&sg[0], mac_data,
769                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
770
771         /* multicast list and count fill the end */
772         mac_data = (void *)&mac_data->macs[uc_count][0];
773
774         mac_data->entries = mc_count;
775         i = 0;
776         netdev_for_each_mc_addr(ha, dev)
777                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
778
779         sg_set_buf(&sg[1], mac_data,
780                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
781
782         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
783                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
784                                   sg, 2, 0))
785                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
786
787         kfree(buf);
788 }
789
790 static void virtnet_vlan_rx_add_vid(struct net_device *dev, u16 vid)
791 {
792         struct virtnet_info *vi = netdev_priv(dev);
793         struct scatterlist sg;
794
795         sg_init_one(&sg, &vid, sizeof(vid));
796
797         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
798                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, 1, 0))
799                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
800 }
801
802 static void virtnet_vlan_rx_kill_vid(struct net_device *dev, u16 vid)
803 {
804         struct virtnet_info *vi = netdev_priv(dev);
805         struct scatterlist sg;
806
807         sg_init_one(&sg, &vid, sizeof(vid));
808
809         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
810                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, 1, 0))
811                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
812 }
813
814 static const struct ethtool_ops virtnet_ethtool_ops = {
815         .set_tx_csum = virtnet_set_tx_csum,
816         .set_sg = ethtool_op_set_sg,
817         .set_tso = ethtool_op_set_tso,
818         .set_ufo = ethtool_op_set_ufo,
819         .get_link = ethtool_op_get_link,
820 };
821
822 #define MIN_MTU 68
823 #define MAX_MTU 65535
824
825 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
826 {
827         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
828                 return -EINVAL;
829         dev->mtu = new_mtu;
830         return 0;
831 }
832
833 static const struct net_device_ops virtnet_netdev = {
834         .ndo_open            = virtnet_open,
835         .ndo_stop            = virtnet_close,
836         .ndo_start_xmit      = start_xmit,
837         .ndo_validate_addr   = eth_validate_addr,
838         .ndo_set_mac_address = virtnet_set_mac_address,
839         .ndo_set_rx_mode     = virtnet_set_rx_mode,
840         .ndo_change_mtu      = virtnet_change_mtu,
841         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
842         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
843 #ifdef CONFIG_NET_POLL_CONTROLLER
844         .ndo_poll_controller = virtnet_netpoll,
845 #endif
846 };
847
848 static void virtnet_update_status(struct virtnet_info *vi)
849 {
850         u16 v;
851
852         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS))
853                 return;
854
855         vi->vdev->config->get(vi->vdev,
856                               offsetof(struct virtio_net_config, status),
857                               &v, sizeof(v));
858
859         /* Ignore unknown (future) status bits */
860         v &= VIRTIO_NET_S_LINK_UP;
861
862         if (vi->status == v)
863                 return;
864
865         vi->status = v;
866
867         if (vi->status & VIRTIO_NET_S_LINK_UP) {
868                 netif_carrier_on(vi->dev);
869                 netif_wake_queue(vi->dev);
870         } else {
871                 netif_carrier_off(vi->dev);
872                 netif_stop_queue(vi->dev);
873         }
874 }
875
876 static void virtnet_config_changed(struct virtio_device *vdev)
877 {
878         struct virtnet_info *vi = vdev->priv;
879
880         virtnet_update_status(vi);
881 }
882
883 static int virtnet_probe(struct virtio_device *vdev)
884 {
885         int err;
886         struct net_device *dev;
887         struct virtnet_info *vi;
888         struct virtqueue *vqs[3];
889         vq_callback_t *callbacks[] = { skb_recv_done, skb_xmit_done, NULL};
890         const char *names[] = { "input", "output", "control" };
891         int nvqs;
892
893         /* Allocate ourselves a network device with room for our info */
894         dev = alloc_etherdev(sizeof(struct virtnet_info));
895         if (!dev)
896                 return -ENOMEM;
897
898         /* Set up network device as normal. */
899         dev->netdev_ops = &virtnet_netdev;
900         dev->features = NETIF_F_HIGHDMA;
901         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
902         SET_NETDEV_DEV(dev, &vdev->dev);
903
904         /* Do we support "hardware" checksums? */
905         if (csum && virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
906                 /* This opens up the world of extra features. */
907                 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
908                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
909                         dev->features |= NETIF_F_TSO | NETIF_F_UFO
910                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
911                 }
912                 /* Individual feature bits: what can host handle? */
913                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
914                         dev->features |= NETIF_F_TSO;
915                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
916                         dev->features |= NETIF_F_TSO6;
917                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
918                         dev->features |= NETIF_F_TSO_ECN;
919                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
920                         dev->features |= NETIF_F_UFO;
921         }
922
923         /* Configuration may specify what MAC to use.  Otherwise random. */
924         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
925                 vdev->config->get(vdev,
926                                   offsetof(struct virtio_net_config, mac),
927                                   dev->dev_addr, dev->addr_len);
928         } else
929                 random_ether_addr(dev->dev_addr);
930
931         /* Set up our device-specific information */
932         vi = netdev_priv(dev);
933         netif_napi_add(dev, &vi->napi, virtnet_poll, napi_weight);
934         vi->dev = dev;
935         vi->vdev = vdev;
936         vdev->priv = vi;
937         vi->pages = NULL;
938         INIT_DELAYED_WORK(&vi->refill, refill_work);
939         sg_init_table(vi->rx_sg, ARRAY_SIZE(vi->rx_sg));
940         sg_init_table(vi->tx_sg, ARRAY_SIZE(vi->tx_sg));
941
942         /* If we can receive ANY GSO packets, we must allocate large ones. */
943         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
944             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
945             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
946                 vi->big_packets = true;
947
948         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
949                 vi->mergeable_rx_bufs = true;
950
951         /* We expect two virtqueues, receive then send,
952          * and optionally control. */
953         nvqs = virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ? 3 : 2;
954
955         err = vdev->config->find_vqs(vdev, nvqs, vqs, callbacks, names);
956         if (err)
957                 goto free;
958
959         vi->rvq = vqs[0];
960         vi->svq = vqs[1];
961
962         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)) {
963                 vi->cvq = vqs[2];
964
965                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
966                         dev->features |= NETIF_F_HW_VLAN_FILTER;
967         }
968
969         err = register_netdev(dev);
970         if (err) {
971                 pr_debug("virtio_net: registering device failed\n");
972                 goto free_vqs;
973         }
974
975         /* Last of all, set up some receive buffers. */
976         try_fill_recv(vi, GFP_KERNEL);
977
978         /* If we didn't even get one input buffer, we're useless. */
979         if (vi->num == 0) {
980                 err = -ENOMEM;
981                 goto unregister;
982         }
983
984         vi->status = VIRTIO_NET_S_LINK_UP;
985         virtnet_update_status(vi);
986         netif_carrier_on(dev);
987
988         pr_debug("virtnet: registered device %s\n", dev->name);
989         return 0;
990
991 unregister:
992         unregister_netdev(dev);
993         cancel_delayed_work_sync(&vi->refill);
994 free_vqs:
995         vdev->config->del_vqs(vdev);
996 free:
997         free_netdev(dev);
998         return err;
999 }
1000
1001 static void free_unused_bufs(struct virtnet_info *vi)
1002 {
1003         void *buf;
1004         while (1) {
1005                 buf = vi->svq->vq_ops->detach_unused_buf(vi->svq);
1006                 if (!buf)
1007                         break;
1008                 dev_kfree_skb(buf);
1009         }
1010         while (1) {
1011                 buf = vi->rvq->vq_ops->detach_unused_buf(vi->rvq);
1012                 if (!buf)
1013                         break;
1014                 if (vi->mergeable_rx_bufs || vi->big_packets)
1015                         give_pages(vi, buf);
1016                 else
1017                         dev_kfree_skb(buf);
1018                 --vi->num;
1019         }
1020         BUG_ON(vi->num != 0);
1021 }
1022
1023 static void __devexit virtnet_remove(struct virtio_device *vdev)
1024 {
1025         struct virtnet_info *vi = vdev->priv;
1026
1027         /* Stop all the virtqueues. */
1028         vdev->config->reset(vdev);
1029
1030
1031         unregister_netdev(vi->dev);
1032         cancel_delayed_work_sync(&vi->refill);
1033
1034         /* Free unused buffers in both send and recv, if any. */
1035         free_unused_bufs(vi);
1036
1037         vdev->config->del_vqs(vi->vdev);
1038
1039         while (vi->pages)
1040                 __free_pages(get_a_page(vi, GFP_KERNEL), 0);
1041
1042         free_netdev(vi->dev);
1043 }
1044
1045 static struct virtio_device_id id_table[] = {
1046         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1047         { 0 },
1048 };
1049
1050 static unsigned int features[] = {
1051         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1052         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1053         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1054         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1055         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1056         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1057         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1058 };
1059
1060 static struct virtio_driver virtio_net_driver = {
1061         .feature_table = features,
1062         .feature_table_size = ARRAY_SIZE(features),
1063         .driver.name =  KBUILD_MODNAME,
1064         .driver.owner = THIS_MODULE,
1065         .id_table =     id_table,
1066         .probe =        virtnet_probe,
1067         .remove =       __devexit_p(virtnet_remove),
1068         .config_changed = virtnet_config_changed,
1069 };
1070
1071 static int __init init(void)
1072 {
1073         return register_virtio_driver(&virtio_net_driver);
1074 }
1075
1076 static void __exit fini(void)
1077 {
1078         unregister_virtio_driver(&virtio_net_driver);
1079 }
1080 module_init(init);
1081 module_exit(fini);
1082
1083 MODULE_DEVICE_TABLE(virtio, id_table);
1084 MODULE_DESCRIPTION("Virtio network driver");
1085 MODULE_LICENSE("GPL");