Loading...
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#include <linux/slab.h>
29
30static int napi_weight = 128;
31module_param(napi_weight, int, 0444);
32
33static int csum = 1, gso = 1;
34module_param(csum, bool, 0444);
35module_param(gso, bool, 0444);
36
37/* FIXME: MTU in config. */
38#define MAX_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
39#define GOOD_COPY_LEN 128
40
41#define VIRTNET_SEND_COMMAND_SG_MAX 2
42
43struct virtnet_stats {
44 struct u64_stats_sync syncp;
45 u64 tx_bytes;
46 u64 tx_packets;
47
48 u64 rx_bytes;
49 u64 rx_packets;
50};
51
52struct virtnet_info {
53 struct virtio_device *vdev;
54 struct virtqueue *rvq, *svq, *cvq;
55 struct net_device *dev;
56 struct napi_struct napi;
57 unsigned int status;
58
59 /* Number of input buffers, and max we've ever had. */
60 unsigned int num, max;
61
62 /* I like... big packets and I cannot lie! */
63 bool big_packets;
64
65 /* Host will merge rx buffers for big packets (shake it! shake it!) */
66 bool mergeable_rx_bufs;
67
68 /* Active statistics */
69 struct virtnet_stats __percpu *stats;
70
71 /* Work struct for refilling if we run low on memory. */
72 struct delayed_work refill;
73
74 /* Chain pages by the private ptr. */
75 struct page *pages;
76
77 /* fragments + linear part + virtio header */
78 struct scatterlist rx_sg[MAX_SKB_FRAGS + 2];
79 struct scatterlist tx_sg[MAX_SKB_FRAGS + 2];
80};
81
82struct skb_vnet_hdr {
83 union {
84 struct virtio_net_hdr hdr;
85 struct virtio_net_hdr_mrg_rxbuf mhdr;
86 };
87 unsigned int num_sg;
88};
89
90struct padded_vnet_hdr {
91 struct virtio_net_hdr hdr;
92 /*
93 * virtio_net_hdr should be in a separated sg buffer because of a
94 * QEMU bug, and data sg buffer shares same page with this header sg.
95 * This padding makes next sg 16 byte aligned after virtio_net_hdr.
96 */
97 char padding[6];
98};
99
100static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
101{
102 return (struct skb_vnet_hdr *)skb->cb;
103}
104
105/*
106 * private is used to chain pages for big packets, put the whole
107 * most recent used list in the beginning for reuse
108 */
109static void give_pages(struct virtnet_info *vi, struct page *page)
110{
111 struct page *end;
112
113 /* Find end of list, sew whole thing into vi->pages. */
114 for (end = page; end->private; end = (struct page *)end->private);
115 end->private = (unsigned long)vi->pages;
116 vi->pages = page;
117}
118
119static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
120{
121 struct page *p = vi->pages;
122
123 if (p) {
124 vi->pages = (struct page *)p->private;
125 /* clear private here, it is used to chain pages */
126 p->private = 0;
127 } else
128 p = alloc_page(gfp_mask);
129 return p;
130}
131
132static void skb_xmit_done(struct virtqueue *svq)
133{
134 struct virtnet_info *vi = svq->vdev->priv;
135
136 /* Suppress further interrupts. */
137 virtqueue_disable_cb(svq);
138
139 /* We were probably waiting for more output buffers. */
140 netif_wake_queue(vi->dev);
141}
142
143static void set_skb_frag(struct sk_buff *skb, struct page *page,
144 unsigned int offset, unsigned int *len)
145{
146 int i = skb_shinfo(skb)->nr_frags;
147 skb_frag_t *f;
148
149 f = &skb_shinfo(skb)->frags[i];
150 f->size = min((unsigned)PAGE_SIZE - offset, *len);
151 f->page_offset = offset;
152 f->page = page;
153
154 skb->data_len += f->size;
155 skb->len += f->size;
156 skb_shinfo(skb)->nr_frags++;
157 *len -= f->size;
158}
159
160static struct sk_buff *page_to_skb(struct virtnet_info *vi,
161 struct page *page, unsigned int len)
162{
163 struct sk_buff *skb;
164 struct skb_vnet_hdr *hdr;
165 unsigned int copy, hdr_len, offset;
166 char *p;
167
168 p = page_address(page);
169
170 /* copy small packet so we can reuse these pages for small data */
171 skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
172 if (unlikely(!skb))
173 return NULL;
174
175 hdr = skb_vnet_hdr(skb);
176
177 if (vi->mergeable_rx_bufs) {
178 hdr_len = sizeof hdr->mhdr;
179 offset = hdr_len;
180 } else {
181 hdr_len = sizeof hdr->hdr;
182 offset = sizeof(struct padded_vnet_hdr);
183 }
184
185 memcpy(hdr, p, hdr_len);
186
187 len -= hdr_len;
188 p += offset;
189
190 copy = len;
191 if (copy > skb_tailroom(skb))
192 copy = skb_tailroom(skb);
193 memcpy(skb_put(skb, copy), p, copy);
194
195 len -= copy;
196 offset += copy;
197
198 while (len) {
199 set_skb_frag(skb, page, offset, &len);
200 page = (struct page *)page->private;
201 offset = 0;
202 }
203
204 if (page)
205 give_pages(vi, page);
206
207 return skb;
208}
209
210static int receive_mergeable(struct virtnet_info *vi, struct sk_buff *skb)
211{
212 struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
213 struct page *page;
214 int num_buf, i, len;
215
216 num_buf = hdr->mhdr.num_buffers;
217 while (--num_buf) {
218 i = skb_shinfo(skb)->nr_frags;
219 if (i >= MAX_SKB_FRAGS) {
220 pr_debug("%s: packet too long\n", skb->dev->name);
221 skb->dev->stats.rx_length_errors++;
222 return -EINVAL;
223 }
224 page = virtqueue_get_buf(vi->rvq, &len);
225 if (!page) {
226 pr_debug("%s: rx error: %d buffers missing\n",
227 skb->dev->name, hdr->mhdr.num_buffers);
228 skb->dev->stats.rx_length_errors++;
229 return -EINVAL;
230 }
231
232 if (len > PAGE_SIZE)
233 len = PAGE_SIZE;
234
235 set_skb_frag(skb, page, 0, &len);
236
237 --vi->num;
238 }
239 return 0;
240}
241
242static void receive_buf(struct net_device *dev, void *buf, unsigned int len)
243{
244 struct virtnet_info *vi = netdev_priv(dev);
245 struct virtnet_stats __percpu *stats = this_cpu_ptr(vi->stats);
246 struct sk_buff *skb;
247 struct page *page;
248 struct skb_vnet_hdr *hdr;
249
250 if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
251 pr_debug("%s: short packet %i\n", dev->name, len);
252 dev->stats.rx_length_errors++;
253 if (vi->mergeable_rx_bufs || vi->big_packets)
254 give_pages(vi, buf);
255 else
256 dev_kfree_skb(buf);
257 return;
258 }
259
260 if (!vi->mergeable_rx_bufs && !vi->big_packets) {
261 skb = buf;
262 len -= sizeof(struct virtio_net_hdr);
263 skb_trim(skb, len);
264 } else {
265 page = buf;
266 skb = page_to_skb(vi, page, len);
267 if (unlikely(!skb)) {
268 dev->stats.rx_dropped++;
269 give_pages(vi, page);
270 return;
271 }
272 if (vi->mergeable_rx_bufs)
273 if (receive_mergeable(vi, skb)) {
274 dev_kfree_skb(skb);
275 return;
276 }
277 }
278
279 hdr = skb_vnet_hdr(skb);
280 skb->truesize += skb->data_len;
281
282 u64_stats_update_begin(&stats->syncp);
283 stats->rx_bytes += skb->len;
284 stats->rx_packets++;
285 u64_stats_update_end(&stats->syncp);
286
287 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
288 pr_debug("Needs csum!\n");
289 if (!skb_partial_csum_set(skb,
290 hdr->hdr.csum_start,
291 hdr->hdr.csum_offset))
292 goto frame_err;
293 } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
294 skb->ip_summed = CHECKSUM_UNNECESSARY;
295 }
296
297 skb->protocol = eth_type_trans(skb, dev);
298 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
299 ntohs(skb->protocol), skb->len, skb->pkt_type);
300
301 if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
302 pr_debug("GSO!\n");
303 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
304 case VIRTIO_NET_HDR_GSO_TCPV4:
305 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
306 break;
307 case VIRTIO_NET_HDR_GSO_UDP:
308 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
309 break;
310 case VIRTIO_NET_HDR_GSO_TCPV6:
311 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
312 break;
313 default:
314 if (net_ratelimit())
315 printk(KERN_WARNING "%s: bad gso type %u.\n",
316 dev->name, hdr->hdr.gso_type);
317 goto frame_err;
318 }
319
320 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
321 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
322
323 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
324 if (skb_shinfo(skb)->gso_size == 0) {
325 if (net_ratelimit())
326 printk(KERN_WARNING "%s: zero gso size.\n",
327 dev->name);
328 goto frame_err;
329 }
330
331 /* Header must be checked, and gso_segs computed. */
332 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
333 skb_shinfo(skb)->gso_segs = 0;
334 }
335
336 netif_receive_skb(skb);
337 return;
338
339frame_err:
340 dev->stats.rx_frame_errors++;
341 dev_kfree_skb(skb);
342}
343
344static int add_recvbuf_small(struct virtnet_info *vi, gfp_t gfp)
345{
346 struct sk_buff *skb;
347 struct skb_vnet_hdr *hdr;
348 int err;
349
350 skb = netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN);
351 if (unlikely(!skb))
352 return -ENOMEM;
353
354 skb_put(skb, MAX_PACKET_LEN);
355
356 hdr = skb_vnet_hdr(skb);
357 sg_set_buf(vi->rx_sg, &hdr->hdr, sizeof hdr->hdr);
358
359 skb_to_sgvec(skb, vi->rx_sg + 1, 0, skb->len);
360
361 err = virtqueue_add_buf_gfp(vi->rvq, vi->rx_sg, 0, 2, skb, gfp);
362 if (err < 0)
363 dev_kfree_skb(skb);
364
365 return err;
366}
367
368static int add_recvbuf_big(struct virtnet_info *vi, gfp_t gfp)
369{
370 struct page *first, *list = NULL;
371 char *p;
372 int i, err, offset;
373
374 /* page in vi->rx_sg[MAX_SKB_FRAGS + 1] is list tail */
375 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
376 first = get_a_page(vi, gfp);
377 if (!first) {
378 if (list)
379 give_pages(vi, list);
380 return -ENOMEM;
381 }
382 sg_set_buf(&vi->rx_sg[i], page_address(first), PAGE_SIZE);
383
384 /* chain new page in list head to match sg */
385 first->private = (unsigned long)list;
386 list = first;
387 }
388
389 first = get_a_page(vi, gfp);
390 if (!first) {
391 give_pages(vi, list);
392 return -ENOMEM;
393 }
394 p = page_address(first);
395
396 /* vi->rx_sg[0], vi->rx_sg[1] share the same page */
397 /* a separated vi->rx_sg[0] for virtio_net_hdr only due to QEMU bug */
398 sg_set_buf(&vi->rx_sg[0], p, sizeof(struct virtio_net_hdr));
399
400 /* vi->rx_sg[1] for data packet, from offset */
401 offset = sizeof(struct padded_vnet_hdr);
402 sg_set_buf(&vi->rx_sg[1], p + offset, PAGE_SIZE - offset);
403
404 /* chain first in list head */
405 first->private = (unsigned long)list;
406 err = virtqueue_add_buf_gfp(vi->rvq, vi->rx_sg, 0, MAX_SKB_FRAGS + 2,
407 first, gfp);
408 if (err < 0)
409 give_pages(vi, first);
410
411 return err;
412}
413
414static int add_recvbuf_mergeable(struct virtnet_info *vi, gfp_t gfp)
415{
416 struct page *page;
417 int err;
418
419 page = get_a_page(vi, gfp);
420 if (!page)
421 return -ENOMEM;
422
423 sg_init_one(vi->rx_sg, page_address(page), PAGE_SIZE);
424
425 err = virtqueue_add_buf_gfp(vi->rvq, vi->rx_sg, 0, 1, page, gfp);
426 if (err < 0)
427 give_pages(vi, page);
428
429 return err;
430}
431
432/* Returns false if we couldn't fill entirely (OOM). */
433static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
434{
435 int err;
436 bool oom;
437
438 do {
439 if (vi->mergeable_rx_bufs)
440 err = add_recvbuf_mergeable(vi, gfp);
441 else if (vi->big_packets)
442 err = add_recvbuf_big(vi, gfp);
443 else
444 err = add_recvbuf_small(vi, gfp);
445
446 oom = err == -ENOMEM;
447 if (err < 0)
448 break;
449 ++vi->num;
450 } while (err > 0);
451 if (unlikely(vi->num > vi->max))
452 vi->max = vi->num;
453 virtqueue_kick(vi->rvq);
454 return !oom;
455}
456
457static void skb_recv_done(struct virtqueue *rvq)
458{
459 struct virtnet_info *vi = rvq->vdev->priv;
460 /* Schedule NAPI, Suppress further interrupts if successful. */
461 if (napi_schedule_prep(&vi->napi)) {
462 virtqueue_disable_cb(rvq);
463 __napi_schedule(&vi->napi);
464 }
465}
466
467static void virtnet_napi_enable(struct virtnet_info *vi)
468{
469 napi_enable(&vi->napi);
470
471 /* If all buffers were filled by other side before we napi_enabled, we
472 * won't get another interrupt, so process any outstanding packets
473 * now. virtnet_poll wants re-enable the queue, so we disable here.
474 * We synchronize against interrupts via NAPI_STATE_SCHED */
475 if (napi_schedule_prep(&vi->napi)) {
476 virtqueue_disable_cb(vi->rvq);
477 __napi_schedule(&vi->napi);
478 }
479}
480
481static void refill_work(struct work_struct *work)
482{
483 struct virtnet_info *vi;
484 bool still_empty;
485
486 vi = container_of(work, struct virtnet_info, refill.work);
487 napi_disable(&vi->napi);
488 still_empty = !try_fill_recv(vi, GFP_KERNEL);
489 virtnet_napi_enable(vi);
490
491 /* In theory, this can happen: if we don't get any buffers in
492 * we will *never* try to fill again. */
493 if (still_empty)
494 schedule_delayed_work(&vi->refill, HZ/2);
495}
496
497static int virtnet_poll(struct napi_struct *napi, int budget)
498{
499 struct virtnet_info *vi = container_of(napi, struct virtnet_info, napi);
500 void *buf;
501 unsigned int len, received = 0;
502
503again:
504 while (received < budget &&
505 (buf = virtqueue_get_buf(vi->rvq, &len)) != NULL) {
506 receive_buf(vi->dev, buf, len);
507 --vi->num;
508 received++;
509 }
510
511 if (vi->num < vi->max / 2) {
512 if (!try_fill_recv(vi, GFP_ATOMIC))
513 schedule_delayed_work(&vi->refill, 0);
514 }
515
516 /* Out of packets? */
517 if (received < budget) {
518 napi_complete(napi);
519 if (unlikely(!virtqueue_enable_cb(vi->rvq)) &&
520 napi_schedule_prep(napi)) {
521 virtqueue_disable_cb(vi->rvq);
522 __napi_schedule(napi);
523 goto again;
524 }
525 }
526
527 return received;
528}
529
530static unsigned int free_old_xmit_skbs(struct virtnet_info *vi)
531{
532 struct sk_buff *skb;
533 unsigned int len, tot_sgs = 0;
534 struct virtnet_stats __percpu *stats = this_cpu_ptr(vi->stats);
535
536 while ((skb = virtqueue_get_buf(vi->svq, &len)) != NULL) {
537 pr_debug("Sent skb %p\n", skb);
538
539 u64_stats_update_begin(&stats->syncp);
540 stats->tx_bytes += skb->len;
541 stats->tx_packets++;
542 u64_stats_update_end(&stats->syncp);
543
544 tot_sgs += skb_vnet_hdr(skb)->num_sg;
545 dev_kfree_skb_any(skb);
546 }
547 return tot_sgs;
548}
549
550static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
551{
552 struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
553 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
554
555 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
556
557 if (skb->ip_summed == CHECKSUM_PARTIAL) {
558 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
559 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
560 hdr->hdr.csum_offset = skb->csum_offset;
561 } else {
562 hdr->hdr.flags = 0;
563 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
564 }
565
566 if (skb_is_gso(skb)) {
567 hdr->hdr.hdr_len = skb_headlen(skb);
568 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
569 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
570 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
571 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
572 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
573 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
574 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
575 else
576 BUG();
577 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
578 hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
579 } else {
580 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
581 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
582 }
583
584 hdr->mhdr.num_buffers = 0;
585
586 /* Encode metadata header at front. */
587 if (vi->mergeable_rx_bufs)
588 sg_set_buf(vi->tx_sg, &hdr->mhdr, sizeof hdr->mhdr);
589 else
590 sg_set_buf(vi->tx_sg, &hdr->hdr, sizeof hdr->hdr);
591
592 hdr->num_sg = skb_to_sgvec(skb, vi->tx_sg + 1, 0, skb->len) + 1;
593 return virtqueue_add_buf(vi->svq, vi->tx_sg, hdr->num_sg,
594 0, skb);
595}
596
597static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
598{
599 struct virtnet_info *vi = netdev_priv(dev);
600 int capacity;
601
602 /* Free up any pending old buffers before queueing new ones. */
603 free_old_xmit_skbs(vi);
604
605 /* Try to transmit */
606 capacity = xmit_skb(vi, skb);
607
608 /* This can happen with OOM and indirect buffers. */
609 if (unlikely(capacity < 0)) {
610 if (net_ratelimit()) {
611 if (likely(capacity == -ENOMEM)) {
612 dev_warn(&dev->dev,
613 "TX queue failure: out of memory\n");
614 } else {
615 dev->stats.tx_fifo_errors++;
616 dev_warn(&dev->dev,
617 "Unexpected TX queue failure: %d\n",
618 capacity);
619 }
620 }
621 dev->stats.tx_dropped++;
622 kfree_skb(skb);
623 return NETDEV_TX_OK;
624 }
625 virtqueue_kick(vi->svq);
626
627 /* Don't wait up for transmitted skbs to be freed. */
628 skb_orphan(skb);
629 nf_reset(skb);
630
631 /* Apparently nice girls don't return TX_BUSY; stop the queue
632 * before it gets out of hand. Naturally, this wastes entries. */
633 if (capacity < 2+MAX_SKB_FRAGS) {
634 netif_stop_queue(dev);
635 if (unlikely(!virtqueue_enable_cb_delayed(vi->svq))) {
636 /* More just got used, free them then recheck. */
637 capacity += free_old_xmit_skbs(vi);
638 if (capacity >= 2+MAX_SKB_FRAGS) {
639 netif_start_queue(dev);
640 virtqueue_disable_cb(vi->svq);
641 }
642 }
643 }
644
645 return NETDEV_TX_OK;
646}
647
648static int virtnet_set_mac_address(struct net_device *dev, void *p)
649{
650 struct virtnet_info *vi = netdev_priv(dev);
651 struct virtio_device *vdev = vi->vdev;
652 int ret;
653
654 ret = eth_mac_addr(dev, p);
655 if (ret)
656 return ret;
657
658 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
659 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
660 dev->dev_addr, dev->addr_len);
661
662 return 0;
663}
664
665static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
666 struct rtnl_link_stats64 *tot)
667{
668 struct virtnet_info *vi = netdev_priv(dev);
669 int cpu;
670 unsigned int start;
671
672 for_each_possible_cpu(cpu) {
673 struct virtnet_stats __percpu *stats
674 = per_cpu_ptr(vi->stats, cpu);
675 u64 tpackets, tbytes, rpackets, rbytes;
676
677 do {
678 start = u64_stats_fetch_begin(&stats->syncp);
679 tpackets = stats->tx_packets;
680 tbytes = stats->tx_bytes;
681 rpackets = stats->rx_packets;
682 rbytes = stats->rx_bytes;
683 } while (u64_stats_fetch_retry(&stats->syncp, start));
684
685 tot->rx_packets += rpackets;
686 tot->tx_packets += tpackets;
687 tot->rx_bytes += rbytes;
688 tot->tx_bytes += tbytes;
689 }
690
691 tot->tx_dropped = dev->stats.tx_dropped;
692 tot->rx_dropped = dev->stats.rx_dropped;
693 tot->rx_length_errors = dev->stats.rx_length_errors;
694 tot->rx_frame_errors = dev->stats.rx_frame_errors;
695
696 return tot;
697}
698
699#ifdef CONFIG_NET_POLL_CONTROLLER
700static void virtnet_netpoll(struct net_device *dev)
701{
702 struct virtnet_info *vi = netdev_priv(dev);
703
704 napi_schedule(&vi->napi);
705}
706#endif
707
708static int virtnet_open(struct net_device *dev)
709{
710 struct virtnet_info *vi = netdev_priv(dev);
711
712 virtnet_napi_enable(vi);
713 return 0;
714}
715
716/*
717 * Send command via the control virtqueue and check status. Commands
718 * supported by the hypervisor, as indicated by feature bits, should
719 * never fail unless improperly formated.
720 */
721static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
722 struct scatterlist *data, int out, int in)
723{
724 struct scatterlist *s, sg[VIRTNET_SEND_COMMAND_SG_MAX + 2];
725 struct virtio_net_ctrl_hdr ctrl;
726 virtio_net_ctrl_ack status = ~0;
727 unsigned int tmp;
728 int i;
729
730 /* Caller should know better */
731 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ||
732 (out + in > VIRTNET_SEND_COMMAND_SG_MAX));
733
734 out++; /* Add header */
735 in++; /* Add return status */
736
737 ctrl.class = class;
738 ctrl.cmd = cmd;
739
740 sg_init_table(sg, out + in);
741
742 sg_set_buf(&sg[0], &ctrl, sizeof(ctrl));
743 for_each_sg(data, s, out + in - 2, i)
744 sg_set_buf(&sg[i + 1], sg_virt(s), s->length);
745 sg_set_buf(&sg[out + in - 1], &status, sizeof(status));
746
747 BUG_ON(virtqueue_add_buf(vi->cvq, sg, out, in, vi) < 0);
748
749 virtqueue_kick(vi->cvq);
750
751 /*
752 * Spin for a response, the kick causes an ioport write, trapping
753 * into the hypervisor, so the request should be handled immediately.
754 */
755 while (!virtqueue_get_buf(vi->cvq, &tmp))
756 cpu_relax();
757
758 return status == VIRTIO_NET_OK;
759}
760
761static int virtnet_close(struct net_device *dev)
762{
763 struct virtnet_info *vi = netdev_priv(dev);
764
765 napi_disable(&vi->napi);
766
767 return 0;
768}
769
770static void virtnet_set_rx_mode(struct net_device *dev)
771{
772 struct virtnet_info *vi = netdev_priv(dev);
773 struct scatterlist sg[2];
774 u8 promisc, allmulti;
775 struct virtio_net_ctrl_mac *mac_data;
776 struct netdev_hw_addr *ha;
777 int uc_count;
778 int mc_count;
779 void *buf;
780 int i;
781
782 /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
783 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
784 return;
785
786 promisc = ((dev->flags & IFF_PROMISC) != 0);
787 allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
788
789 sg_init_one(sg, &promisc, sizeof(promisc));
790
791 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
792 VIRTIO_NET_CTRL_RX_PROMISC,
793 sg, 1, 0))
794 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
795 promisc ? "en" : "dis");
796
797 sg_init_one(sg, &allmulti, sizeof(allmulti));
798
799 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
800 VIRTIO_NET_CTRL_RX_ALLMULTI,
801 sg, 1, 0))
802 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
803 allmulti ? "en" : "dis");
804
805 uc_count = netdev_uc_count(dev);
806 mc_count = netdev_mc_count(dev);
807 /* MAC filter - use one buffer for both lists */
808 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
809 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
810 mac_data = buf;
811 if (!buf) {
812 dev_warn(&dev->dev, "No memory for MAC address buffer\n");
813 return;
814 }
815
816 sg_init_table(sg, 2);
817
818 /* Store the unicast list and count in the front of the buffer */
819 mac_data->entries = uc_count;
820 i = 0;
821 netdev_for_each_uc_addr(ha, dev)
822 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
823
824 sg_set_buf(&sg[0], mac_data,
825 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
826
827 /* multicast list and count fill the end */
828 mac_data = (void *)&mac_data->macs[uc_count][0];
829
830 mac_data->entries = mc_count;
831 i = 0;
832 netdev_for_each_mc_addr(ha, dev)
833 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
834
835 sg_set_buf(&sg[1], mac_data,
836 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
837
838 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
839 VIRTIO_NET_CTRL_MAC_TABLE_SET,
840 sg, 2, 0))
841 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
842
843 kfree(buf);
844}
845
846static void virtnet_vlan_rx_add_vid(struct net_device *dev, u16 vid)
847{
848 struct virtnet_info *vi = netdev_priv(dev);
849 struct scatterlist sg;
850
851 sg_init_one(&sg, &vid, sizeof(vid));
852
853 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
854 VIRTIO_NET_CTRL_VLAN_ADD, &sg, 1, 0))
855 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
856}
857
858static void virtnet_vlan_rx_kill_vid(struct net_device *dev, u16 vid)
859{
860 struct virtnet_info *vi = netdev_priv(dev);
861 struct scatterlist sg;
862
863 sg_init_one(&sg, &vid, sizeof(vid));
864
865 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
866 VIRTIO_NET_CTRL_VLAN_DEL, &sg, 1, 0))
867 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
868}
869
870static const struct ethtool_ops virtnet_ethtool_ops = {
871 .get_link = ethtool_op_get_link,
872};
873
874#define MIN_MTU 68
875#define MAX_MTU 65535
876
877static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
878{
879 if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
880 return -EINVAL;
881 dev->mtu = new_mtu;
882 return 0;
883}
884
885static const struct net_device_ops virtnet_netdev = {
886 .ndo_open = virtnet_open,
887 .ndo_stop = virtnet_close,
888 .ndo_start_xmit = start_xmit,
889 .ndo_validate_addr = eth_validate_addr,
890 .ndo_set_mac_address = virtnet_set_mac_address,
891 .ndo_set_rx_mode = virtnet_set_rx_mode,
892 .ndo_change_mtu = virtnet_change_mtu,
893 .ndo_get_stats64 = virtnet_stats,
894 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
895 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
896#ifdef CONFIG_NET_POLL_CONTROLLER
897 .ndo_poll_controller = virtnet_netpoll,
898#endif
899};
900
901static void virtnet_update_status(struct virtnet_info *vi)
902{
903 u16 v;
904
905 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS))
906 return;
907
908 vi->vdev->config->get(vi->vdev,
909 offsetof(struct virtio_net_config, status),
910 &v, sizeof(v));
911
912 /* Ignore unknown (future) status bits */
913 v &= VIRTIO_NET_S_LINK_UP;
914
915 if (vi->status == v)
916 return;
917
918 vi->status = v;
919
920 if (vi->status & VIRTIO_NET_S_LINK_UP) {
921 netif_carrier_on(vi->dev);
922 netif_wake_queue(vi->dev);
923 } else {
924 netif_carrier_off(vi->dev);
925 netif_stop_queue(vi->dev);
926 }
927}
928
929static void virtnet_config_changed(struct virtio_device *vdev)
930{
931 struct virtnet_info *vi = vdev->priv;
932
933 virtnet_update_status(vi);
934}
935
936static int virtnet_probe(struct virtio_device *vdev)
937{
938 int err;
939 struct net_device *dev;
940 struct virtnet_info *vi;
941 struct virtqueue *vqs[3];
942 vq_callback_t *callbacks[] = { skb_recv_done, skb_xmit_done, NULL};
943 const char *names[] = { "input", "output", "control" };
944 int nvqs;
945
946 /* Allocate ourselves a network device with room for our info */
947 dev = alloc_etherdev(sizeof(struct virtnet_info));
948 if (!dev)
949 return -ENOMEM;
950
951 /* Set up network device as normal. */
952 dev->netdev_ops = &virtnet_netdev;
953 dev->features = NETIF_F_HIGHDMA;
954
955 SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
956 SET_NETDEV_DEV(dev, &vdev->dev);
957
958 /* Do we support "hardware" checksums? */
959 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
960 /* This opens up the world of extra features. */
961 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
962 if (csum)
963 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
964
965 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
966 dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
967 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
968 }
969 /* Individual feature bits: what can host handle? */
970 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
971 dev->hw_features |= NETIF_F_TSO;
972 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
973 dev->hw_features |= NETIF_F_TSO6;
974 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
975 dev->hw_features |= NETIF_F_TSO_ECN;
976 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
977 dev->hw_features |= NETIF_F_UFO;
978
979 if (gso)
980 dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
981 /* (!csum && gso) case will be fixed by register_netdev() */
982 }
983
984 /* Configuration may specify what MAC to use. Otherwise random. */
985 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
986 vdev->config->get(vdev,
987 offsetof(struct virtio_net_config, mac),
988 dev->dev_addr, dev->addr_len);
989 } else
990 random_ether_addr(dev->dev_addr);
991
992 /* Set up our device-specific information */
993 vi = netdev_priv(dev);
994 netif_napi_add(dev, &vi->napi, virtnet_poll, napi_weight);
995 vi->dev = dev;
996 vi->vdev = vdev;
997 vdev->priv = vi;
998 vi->pages = NULL;
999 vi->stats = alloc_percpu(struct virtnet_stats);
1000 err = -ENOMEM;
1001 if (vi->stats == NULL)
1002 goto free;
1003
1004 INIT_DELAYED_WORK(&vi->refill, refill_work);
1005 sg_init_table(vi->rx_sg, ARRAY_SIZE(vi->rx_sg));
1006 sg_init_table(vi->tx_sg, ARRAY_SIZE(vi->tx_sg));
1007
1008 /* If we can receive ANY GSO packets, we must allocate large ones. */
1009 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1010 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1011 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1012 vi->big_packets = true;
1013
1014 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1015 vi->mergeable_rx_bufs = true;
1016
1017 /* We expect two virtqueues, receive then send,
1018 * and optionally control. */
1019 nvqs = virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ? 3 : 2;
1020
1021 err = vdev->config->find_vqs(vdev, nvqs, vqs, callbacks, names);
1022 if (err)
1023 goto free_stats;
1024
1025 vi->rvq = vqs[0];
1026 vi->svq = vqs[1];
1027
1028 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)) {
1029 vi->cvq = vqs[2];
1030
1031 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1032 dev->features |= NETIF_F_HW_VLAN_FILTER;
1033 }
1034
1035 err = register_netdev(dev);
1036 if (err) {
1037 pr_debug("virtio_net: registering device failed\n");
1038 goto free_vqs;
1039 }
1040
1041 /* Last of all, set up some receive buffers. */
1042 try_fill_recv(vi, GFP_KERNEL);
1043
1044 /* If we didn't even get one input buffer, we're useless. */
1045 if (vi->num == 0) {
1046 err = -ENOMEM;
1047 goto unregister;
1048 }
1049
1050 /* Assume link up if device can't report link status,
1051 otherwise get link status from config. */
1052 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1053 netif_carrier_off(dev);
1054 virtnet_update_status(vi);
1055 } else {
1056 vi->status = VIRTIO_NET_S_LINK_UP;
1057 netif_carrier_on(dev);
1058 }
1059
1060 pr_debug("virtnet: registered device %s\n", dev->name);
1061 return 0;
1062
1063unregister:
1064 unregister_netdev(dev);
1065 cancel_delayed_work_sync(&vi->refill);
1066free_vqs:
1067 vdev->config->del_vqs(vdev);
1068free_stats:
1069 free_percpu(vi->stats);
1070free:
1071 free_netdev(dev);
1072 return err;
1073}
1074
1075static void free_unused_bufs(struct virtnet_info *vi)
1076{
1077 void *buf;
1078 while (1) {
1079 buf = virtqueue_detach_unused_buf(vi->svq);
1080 if (!buf)
1081 break;
1082 dev_kfree_skb(buf);
1083 }
1084 while (1) {
1085 buf = virtqueue_detach_unused_buf(vi->rvq);
1086 if (!buf)
1087 break;
1088 if (vi->mergeable_rx_bufs || vi->big_packets)
1089 give_pages(vi, buf);
1090 else
1091 dev_kfree_skb(buf);
1092 --vi->num;
1093 }
1094 BUG_ON(vi->num != 0);
1095}
1096
1097static void __devexit virtnet_remove(struct virtio_device *vdev)
1098{
1099 struct virtnet_info *vi = vdev->priv;
1100
1101 /* Stop all the virtqueues. */
1102 vdev->config->reset(vdev);
1103
1104
1105 unregister_netdev(vi->dev);
1106 cancel_delayed_work_sync(&vi->refill);
1107
1108 /* Free unused buffers in both send and recv, if any. */
1109 free_unused_bufs(vi);
1110
1111 vdev->config->del_vqs(vi->vdev);
1112
1113 while (vi->pages)
1114 __free_pages(get_a_page(vi, GFP_KERNEL), 0);
1115
1116 free_percpu(vi->stats);
1117 free_netdev(vi->dev);
1118}
1119
1120static struct virtio_device_id id_table[] = {
1121 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1122 { 0 },
1123};
1124
1125static unsigned int features[] = {
1126 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1127 VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1128 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1129 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1130 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1131 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1132 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1133};
1134
1135static struct virtio_driver virtio_net_driver = {
1136 .feature_table = features,
1137 .feature_table_size = ARRAY_SIZE(features),
1138 .driver.name = KBUILD_MODNAME,
1139 .driver.owner = THIS_MODULE,
1140 .id_table = id_table,
1141 .probe = virtnet_probe,
1142 .remove = __devexit_p(virtnet_remove),
1143 .config_changed = virtnet_config_changed,
1144};
1145
1146static int __init init(void)
1147{
1148 return register_virtio_driver(&virtio_net_driver);
1149}
1150
1151static void __exit fini(void)
1152{
1153 unregister_virtio_driver(&virtio_net_driver);
1154}
1155module_init(init);
1156module_exit(fini);
1157
1158MODULE_DEVICE_TABLE(virtio, id_table);
1159MODULE_DESCRIPTION("Virtio network driver");
1160MODULE_LICENSE("GPL");
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, see <http://www.gnu.org/licenses/>.
17 */
18//#define DEBUG
19#include <linux/netdevice.h>
20#include <linux/etherdevice.h>
21#include <linux/ethtool.h>
22#include <linux/module.h>
23#include <linux/virtio.h>
24#include <linux/virtio_net.h>
25#include <linux/scatterlist.h>
26#include <linux/if_vlan.h>
27#include <linux/slab.h>
28#include <linux/cpu.h>
29#include <linux/average.h>
30
31static int napi_weight = NAPI_POLL_WEIGHT;
32module_param(napi_weight, int, 0444);
33
34static bool csum = true, gso = true;
35module_param(csum, bool, 0444);
36module_param(gso, bool, 0444);
37
38/* FIXME: MTU in config. */
39#define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
40#define GOOD_COPY_LEN 128
41
42/* Weight used for the RX packet size EWMA. The average packet size is used to
43 * determine the packet buffer size when refilling RX rings. As the entire RX
44 * ring may be refilled at once, the weight is chosen so that the EWMA will be
45 * insensitive to short-term, transient changes in packet size.
46 */
47#define RECEIVE_AVG_WEIGHT 64
48
49/* Minimum alignment for mergeable packet buffers. */
50#define MERGEABLE_BUFFER_ALIGN max(L1_CACHE_BYTES, 256)
51
52#define VIRTNET_DRIVER_VERSION "1.0.0"
53
54struct virtnet_stats {
55 struct u64_stats_sync tx_syncp;
56 struct u64_stats_sync rx_syncp;
57 u64 tx_bytes;
58 u64 tx_packets;
59
60 u64 rx_bytes;
61 u64 rx_packets;
62};
63
64/* Internal representation of a send virtqueue */
65struct send_queue {
66 /* Virtqueue associated with this send _queue */
67 struct virtqueue *vq;
68
69 /* TX: fragments + linear part + virtio header */
70 struct scatterlist sg[MAX_SKB_FRAGS + 2];
71
72 /* Name of the send queue: output.$index */
73 char name[40];
74};
75
76/* Internal representation of a receive virtqueue */
77struct receive_queue {
78 /* Virtqueue associated with this receive_queue */
79 struct virtqueue *vq;
80
81 struct napi_struct napi;
82
83 /* Chain pages by the private ptr. */
84 struct page *pages;
85
86 /* Average packet length for mergeable receive buffers. */
87 struct ewma mrg_avg_pkt_len;
88
89 /* Page frag for packet buffer allocation. */
90 struct page_frag alloc_frag;
91
92 /* RX: fragments + linear part + virtio header */
93 struct scatterlist sg[MAX_SKB_FRAGS + 2];
94
95 /* Name of this receive queue: input.$index */
96 char name[40];
97};
98
99struct virtnet_info {
100 struct virtio_device *vdev;
101 struct virtqueue *cvq;
102 struct net_device *dev;
103 struct send_queue *sq;
104 struct receive_queue *rq;
105 unsigned int status;
106
107 /* Max # of queue pairs supported by the device */
108 u16 max_queue_pairs;
109
110 /* # of queue pairs currently used by the driver */
111 u16 curr_queue_pairs;
112
113 /* I like... big packets and I cannot lie! */
114 bool big_packets;
115
116 /* Host will merge rx buffers for big packets (shake it! shake it!) */
117 bool mergeable_rx_bufs;
118
119 /* Has control virtqueue */
120 bool has_cvq;
121
122 /* Host can handle any s/g split between our header and packet data */
123 bool any_header_sg;
124
125 /* enable config space updates */
126 bool config_enable;
127
128 /* Active statistics */
129 struct virtnet_stats __percpu *stats;
130
131 /* Work struct for refilling if we run low on memory. */
132 struct delayed_work refill;
133
134 /* Work struct for config space updates */
135 struct work_struct config_work;
136
137 /* Lock for config space updates */
138 struct mutex config_lock;
139
140 /* Does the affinity hint is set for virtqueues? */
141 bool affinity_hint_set;
142
143 /* CPU hot plug notifier */
144 struct notifier_block nb;
145};
146
147struct skb_vnet_hdr {
148 union {
149 struct virtio_net_hdr hdr;
150 struct virtio_net_hdr_mrg_rxbuf mhdr;
151 };
152};
153
154struct padded_vnet_hdr {
155 struct virtio_net_hdr hdr;
156 /*
157 * virtio_net_hdr should be in a separated sg buffer because of a
158 * QEMU bug, and data sg buffer shares same page with this header sg.
159 * This padding makes next sg 16 byte aligned after virtio_net_hdr.
160 */
161 char padding[6];
162};
163
164/* Converting between virtqueue no. and kernel tx/rx queue no.
165 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
166 */
167static int vq2txq(struct virtqueue *vq)
168{
169 return (vq->index - 1) / 2;
170}
171
172static int txq2vq(int txq)
173{
174 return txq * 2 + 1;
175}
176
177static int vq2rxq(struct virtqueue *vq)
178{
179 return vq->index / 2;
180}
181
182static int rxq2vq(int rxq)
183{
184 return rxq * 2;
185}
186
187static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
188{
189 return (struct skb_vnet_hdr *)skb->cb;
190}
191
192/*
193 * private is used to chain pages for big packets, put the whole
194 * most recent used list in the beginning for reuse
195 */
196static void give_pages(struct receive_queue *rq, struct page *page)
197{
198 struct page *end;
199
200 /* Find end of list, sew whole thing into vi->rq.pages. */
201 for (end = page; end->private; end = (struct page *)end->private);
202 end->private = (unsigned long)rq->pages;
203 rq->pages = page;
204}
205
206static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
207{
208 struct page *p = rq->pages;
209
210 if (p) {
211 rq->pages = (struct page *)p->private;
212 /* clear private here, it is used to chain pages */
213 p->private = 0;
214 } else
215 p = alloc_page(gfp_mask);
216 return p;
217}
218
219static void skb_xmit_done(struct virtqueue *vq)
220{
221 struct virtnet_info *vi = vq->vdev->priv;
222
223 /* Suppress further interrupts. */
224 virtqueue_disable_cb(vq);
225
226 /* We were probably waiting for more output buffers. */
227 netif_wake_subqueue(vi->dev, vq2txq(vq));
228}
229
230static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
231{
232 unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
233 return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
234}
235
236static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
237{
238 return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
239
240}
241
242static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
243{
244 unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
245 return (unsigned long)buf | (size - 1);
246}
247
248/* Called from bottom half context */
249static struct sk_buff *page_to_skb(struct receive_queue *rq,
250 struct page *page, unsigned int offset,
251 unsigned int len, unsigned int truesize)
252{
253 struct virtnet_info *vi = rq->vq->vdev->priv;
254 struct sk_buff *skb;
255 struct skb_vnet_hdr *hdr;
256 unsigned int copy, hdr_len, hdr_padded_len;
257 char *p;
258
259 p = page_address(page) + offset;
260
261 /* copy small packet so we can reuse these pages for small data */
262 skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
263 if (unlikely(!skb))
264 return NULL;
265
266 hdr = skb_vnet_hdr(skb);
267
268 if (vi->mergeable_rx_bufs) {
269 hdr_len = sizeof hdr->mhdr;
270 hdr_padded_len = sizeof hdr->mhdr;
271 } else {
272 hdr_len = sizeof hdr->hdr;
273 hdr_padded_len = sizeof(struct padded_vnet_hdr);
274 }
275
276 memcpy(hdr, p, hdr_len);
277
278 len -= hdr_len;
279 offset += hdr_padded_len;
280 p += hdr_padded_len;
281
282 copy = len;
283 if (copy > skb_tailroom(skb))
284 copy = skb_tailroom(skb);
285 memcpy(skb_put(skb, copy), p, copy);
286
287 len -= copy;
288 offset += copy;
289
290 if (vi->mergeable_rx_bufs) {
291 if (len)
292 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
293 else
294 put_page(page);
295 return skb;
296 }
297
298 /*
299 * Verify that we can indeed put this data into a skb.
300 * This is here to handle cases when the device erroneously
301 * tries to receive more than is possible. This is usually
302 * the case of a broken device.
303 */
304 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
305 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
306 dev_kfree_skb(skb);
307 return NULL;
308 }
309 BUG_ON(offset >= PAGE_SIZE);
310 while (len) {
311 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
312 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
313 frag_size, truesize);
314 len -= frag_size;
315 page = (struct page *)page->private;
316 offset = 0;
317 }
318
319 if (page)
320 give_pages(rq, page);
321
322 return skb;
323}
324
325static struct sk_buff *receive_small(void *buf, unsigned int len)
326{
327 struct sk_buff * skb = buf;
328
329 len -= sizeof(struct virtio_net_hdr);
330 skb_trim(skb, len);
331
332 return skb;
333}
334
335static struct sk_buff *receive_big(struct net_device *dev,
336 struct receive_queue *rq,
337 void *buf,
338 unsigned int len)
339{
340 struct page *page = buf;
341 struct sk_buff *skb = page_to_skb(rq, page, 0, len, PAGE_SIZE);
342
343 if (unlikely(!skb))
344 goto err;
345
346 return skb;
347
348err:
349 dev->stats.rx_dropped++;
350 give_pages(rq, page);
351 return NULL;
352}
353
354static struct sk_buff *receive_mergeable(struct net_device *dev,
355 struct receive_queue *rq,
356 unsigned long ctx,
357 unsigned int len)
358{
359 void *buf = mergeable_ctx_to_buf_address(ctx);
360 struct skb_vnet_hdr *hdr = buf;
361 int num_buf = hdr->mhdr.num_buffers;
362 struct page *page = virt_to_head_page(buf);
363 int offset = buf - page_address(page);
364 unsigned int truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
365
366 struct sk_buff *head_skb = page_to_skb(rq, page, offset, len, truesize);
367 struct sk_buff *curr_skb = head_skb;
368
369 if (unlikely(!curr_skb))
370 goto err_skb;
371 while (--num_buf) {
372 int num_skb_frags;
373
374 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
375 if (unlikely(!ctx)) {
376 pr_debug("%s: rx error: %d buffers out of %d missing\n",
377 dev->name, num_buf, hdr->mhdr.num_buffers);
378 dev->stats.rx_length_errors++;
379 goto err_buf;
380 }
381
382 buf = mergeable_ctx_to_buf_address(ctx);
383 page = virt_to_head_page(buf);
384
385 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
386 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
387 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
388
389 if (unlikely(!nskb))
390 goto err_skb;
391 if (curr_skb == head_skb)
392 skb_shinfo(curr_skb)->frag_list = nskb;
393 else
394 curr_skb->next = nskb;
395 curr_skb = nskb;
396 head_skb->truesize += nskb->truesize;
397 num_skb_frags = 0;
398 }
399 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
400 if (curr_skb != head_skb) {
401 head_skb->data_len += len;
402 head_skb->len += len;
403 head_skb->truesize += truesize;
404 }
405 offset = buf - page_address(page);
406 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
407 put_page(page);
408 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
409 len, truesize);
410 } else {
411 skb_add_rx_frag(curr_skb, num_skb_frags, page,
412 offset, len, truesize);
413 }
414 }
415
416 ewma_add(&rq->mrg_avg_pkt_len, head_skb->len);
417 return head_skb;
418
419err_skb:
420 put_page(page);
421 while (--num_buf) {
422 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
423 if (unlikely(!ctx)) {
424 pr_debug("%s: rx error: %d buffers missing\n",
425 dev->name, num_buf);
426 dev->stats.rx_length_errors++;
427 break;
428 }
429 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
430 put_page(page);
431 }
432err_buf:
433 dev->stats.rx_dropped++;
434 dev_kfree_skb(head_skb);
435 return NULL;
436}
437
438static void receive_buf(struct receive_queue *rq, void *buf, unsigned int len)
439{
440 struct virtnet_info *vi = rq->vq->vdev->priv;
441 struct net_device *dev = vi->dev;
442 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
443 struct sk_buff *skb;
444 struct skb_vnet_hdr *hdr;
445
446 if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
447 pr_debug("%s: short packet %i\n", dev->name, len);
448 dev->stats.rx_length_errors++;
449 if (vi->mergeable_rx_bufs) {
450 unsigned long ctx = (unsigned long)buf;
451 void *base = mergeable_ctx_to_buf_address(ctx);
452 put_page(virt_to_head_page(base));
453 } else if (vi->big_packets) {
454 give_pages(rq, buf);
455 } else {
456 dev_kfree_skb(buf);
457 }
458 return;
459 }
460
461 if (vi->mergeable_rx_bufs)
462 skb = receive_mergeable(dev, rq, (unsigned long)buf, len);
463 else if (vi->big_packets)
464 skb = receive_big(dev, rq, buf, len);
465 else
466 skb = receive_small(buf, len);
467
468 if (unlikely(!skb))
469 return;
470
471 hdr = skb_vnet_hdr(skb);
472
473 u64_stats_update_begin(&stats->rx_syncp);
474 stats->rx_bytes += skb->len;
475 stats->rx_packets++;
476 u64_stats_update_end(&stats->rx_syncp);
477
478 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
479 pr_debug("Needs csum!\n");
480 if (!skb_partial_csum_set(skb,
481 hdr->hdr.csum_start,
482 hdr->hdr.csum_offset))
483 goto frame_err;
484 } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
485 skb->ip_summed = CHECKSUM_UNNECESSARY;
486 }
487
488 skb->protocol = eth_type_trans(skb, dev);
489 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
490 ntohs(skb->protocol), skb->len, skb->pkt_type);
491
492 if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
493 pr_debug("GSO!\n");
494 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
495 case VIRTIO_NET_HDR_GSO_TCPV4:
496 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
497 break;
498 case VIRTIO_NET_HDR_GSO_UDP:
499 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
500 break;
501 case VIRTIO_NET_HDR_GSO_TCPV6:
502 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
503 break;
504 default:
505 net_warn_ratelimited("%s: bad gso type %u.\n",
506 dev->name, hdr->hdr.gso_type);
507 goto frame_err;
508 }
509
510 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
511 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
512
513 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
514 if (skb_shinfo(skb)->gso_size == 0) {
515 net_warn_ratelimited("%s: zero gso size.\n", dev->name);
516 goto frame_err;
517 }
518
519 /* Header must be checked, and gso_segs computed. */
520 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
521 skb_shinfo(skb)->gso_segs = 0;
522 }
523
524 netif_receive_skb(skb);
525 return;
526
527frame_err:
528 dev->stats.rx_frame_errors++;
529 dev_kfree_skb(skb);
530}
531
532static int add_recvbuf_small(struct receive_queue *rq, gfp_t gfp)
533{
534 struct virtnet_info *vi = rq->vq->vdev->priv;
535 struct sk_buff *skb;
536 struct skb_vnet_hdr *hdr;
537 int err;
538
539 skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
540 if (unlikely(!skb))
541 return -ENOMEM;
542
543 skb_put(skb, GOOD_PACKET_LEN);
544
545 hdr = skb_vnet_hdr(skb);
546 sg_set_buf(rq->sg, &hdr->hdr, sizeof hdr->hdr);
547
548 skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
549
550 err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
551 if (err < 0)
552 dev_kfree_skb(skb);
553
554 return err;
555}
556
557static int add_recvbuf_big(struct receive_queue *rq, gfp_t gfp)
558{
559 struct page *first, *list = NULL;
560 char *p;
561 int i, err, offset;
562
563 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
564 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
565 first = get_a_page(rq, gfp);
566 if (!first) {
567 if (list)
568 give_pages(rq, list);
569 return -ENOMEM;
570 }
571 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
572
573 /* chain new page in list head to match sg */
574 first->private = (unsigned long)list;
575 list = first;
576 }
577
578 first = get_a_page(rq, gfp);
579 if (!first) {
580 give_pages(rq, list);
581 return -ENOMEM;
582 }
583 p = page_address(first);
584
585 /* rq->sg[0], rq->sg[1] share the same page */
586 /* a separated rq->sg[0] for virtio_net_hdr only due to QEMU bug */
587 sg_set_buf(&rq->sg[0], p, sizeof(struct virtio_net_hdr));
588
589 /* rq->sg[1] for data packet, from offset */
590 offset = sizeof(struct padded_vnet_hdr);
591 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
592
593 /* chain first in list head */
594 first->private = (unsigned long)list;
595 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
596 first, gfp);
597 if (err < 0)
598 give_pages(rq, first);
599
600 return err;
601}
602
603static unsigned int get_mergeable_buf_len(struct ewma *avg_pkt_len)
604{
605 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
606 unsigned int len;
607
608 len = hdr_len + clamp_t(unsigned int, ewma_read(avg_pkt_len),
609 GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
610 return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
611}
612
613static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
614{
615 struct page_frag *alloc_frag = &rq->alloc_frag;
616 char *buf;
617 unsigned long ctx;
618 int err;
619 unsigned int len, hole;
620
621 len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
622 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
623 return -ENOMEM;
624
625 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
626 ctx = mergeable_buf_to_ctx(buf, len);
627 get_page(alloc_frag->page);
628 alloc_frag->offset += len;
629 hole = alloc_frag->size - alloc_frag->offset;
630 if (hole < len) {
631 /* To avoid internal fragmentation, if there is very likely not
632 * enough space for another buffer, add the remaining space to
633 * the current buffer. This extra space is not included in
634 * the truesize stored in ctx.
635 */
636 len += hole;
637 alloc_frag->offset += hole;
638 }
639
640 sg_init_one(rq->sg, buf, len);
641 err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
642 if (err < 0)
643 put_page(virt_to_head_page(buf));
644
645 return err;
646}
647
648/*
649 * Returns false if we couldn't fill entirely (OOM).
650 *
651 * Normally run in the receive path, but can also be run from ndo_open
652 * before we're receiving packets, or from refill_work which is
653 * careful to disable receiving (using napi_disable).
654 */
655static bool try_fill_recv(struct receive_queue *rq, gfp_t gfp)
656{
657 struct virtnet_info *vi = rq->vq->vdev->priv;
658 int err;
659 bool oom;
660
661 gfp |= __GFP_COLD;
662 do {
663 if (vi->mergeable_rx_bufs)
664 err = add_recvbuf_mergeable(rq, gfp);
665 else if (vi->big_packets)
666 err = add_recvbuf_big(rq, gfp);
667 else
668 err = add_recvbuf_small(rq, gfp);
669
670 oom = err == -ENOMEM;
671 if (err)
672 break;
673 } while (rq->vq->num_free);
674 virtqueue_kick(rq->vq);
675 return !oom;
676}
677
678static void skb_recv_done(struct virtqueue *rvq)
679{
680 struct virtnet_info *vi = rvq->vdev->priv;
681 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
682
683 /* Schedule NAPI, Suppress further interrupts if successful. */
684 if (napi_schedule_prep(&rq->napi)) {
685 virtqueue_disable_cb(rvq);
686 __napi_schedule(&rq->napi);
687 }
688}
689
690static void virtnet_napi_enable(struct receive_queue *rq)
691{
692 napi_enable(&rq->napi);
693
694 /* If all buffers were filled by other side before we napi_enabled, we
695 * won't get another interrupt, so process any outstanding packets
696 * now. virtnet_poll wants re-enable the queue, so we disable here.
697 * We synchronize against interrupts via NAPI_STATE_SCHED */
698 if (napi_schedule_prep(&rq->napi)) {
699 virtqueue_disable_cb(rq->vq);
700 local_bh_disable();
701 __napi_schedule(&rq->napi);
702 local_bh_enable();
703 }
704}
705
706static void refill_work(struct work_struct *work)
707{
708 struct virtnet_info *vi =
709 container_of(work, struct virtnet_info, refill.work);
710 bool still_empty;
711 int i;
712
713 for (i = 0; i < vi->curr_queue_pairs; i++) {
714 struct receive_queue *rq = &vi->rq[i];
715
716 napi_disable(&rq->napi);
717 still_empty = !try_fill_recv(rq, GFP_KERNEL);
718 virtnet_napi_enable(rq);
719
720 /* In theory, this can happen: if we don't get any buffers in
721 * we will *never* try to fill again.
722 */
723 if (still_empty)
724 schedule_delayed_work(&vi->refill, HZ/2);
725 }
726}
727
728static int virtnet_poll(struct napi_struct *napi, int budget)
729{
730 struct receive_queue *rq =
731 container_of(napi, struct receive_queue, napi);
732 struct virtnet_info *vi = rq->vq->vdev->priv;
733 void *buf;
734 unsigned int r, len, received = 0;
735
736again:
737 while (received < budget &&
738 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
739 receive_buf(rq, buf, len);
740 received++;
741 }
742
743 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
744 if (!try_fill_recv(rq, GFP_ATOMIC))
745 schedule_delayed_work(&vi->refill, 0);
746 }
747
748 /* Out of packets? */
749 if (received < budget) {
750 r = virtqueue_enable_cb_prepare(rq->vq);
751 napi_complete(napi);
752 if (unlikely(virtqueue_poll(rq->vq, r)) &&
753 napi_schedule_prep(napi)) {
754 virtqueue_disable_cb(rq->vq);
755 __napi_schedule(napi);
756 goto again;
757 }
758 }
759
760 return received;
761}
762
763static int virtnet_open(struct net_device *dev)
764{
765 struct virtnet_info *vi = netdev_priv(dev);
766 int i;
767
768 for (i = 0; i < vi->max_queue_pairs; i++) {
769 if (i < vi->curr_queue_pairs)
770 /* Make sure we have some buffers: if oom use wq. */
771 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
772 schedule_delayed_work(&vi->refill, 0);
773 virtnet_napi_enable(&vi->rq[i]);
774 }
775
776 return 0;
777}
778
779static void free_old_xmit_skbs(struct send_queue *sq)
780{
781 struct sk_buff *skb;
782 unsigned int len;
783 struct virtnet_info *vi = sq->vq->vdev->priv;
784 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
785
786 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
787 pr_debug("Sent skb %p\n", skb);
788
789 u64_stats_update_begin(&stats->tx_syncp);
790 stats->tx_bytes += skb->len;
791 stats->tx_packets++;
792 u64_stats_update_end(&stats->tx_syncp);
793
794 dev_kfree_skb_any(skb);
795 }
796}
797
798static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
799{
800 struct skb_vnet_hdr *hdr;
801 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
802 struct virtnet_info *vi = sq->vq->vdev->priv;
803 unsigned num_sg;
804 unsigned hdr_len;
805 bool can_push;
806
807 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
808 if (vi->mergeable_rx_bufs)
809 hdr_len = sizeof hdr->mhdr;
810 else
811 hdr_len = sizeof hdr->hdr;
812
813 can_push = vi->any_header_sg &&
814 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
815 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
816 /* Even if we can, don't push here yet as this would skew
817 * csum_start offset below. */
818 if (can_push)
819 hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
820 else
821 hdr = skb_vnet_hdr(skb);
822
823 if (skb->ip_summed == CHECKSUM_PARTIAL) {
824 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
825 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
826 hdr->hdr.csum_offset = skb->csum_offset;
827 } else {
828 hdr->hdr.flags = 0;
829 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
830 }
831
832 if (skb_is_gso(skb)) {
833 hdr->hdr.hdr_len = skb_headlen(skb);
834 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
835 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
836 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
837 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
838 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
839 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
840 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
841 else
842 BUG();
843 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
844 hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
845 } else {
846 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
847 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
848 }
849
850 if (vi->mergeable_rx_bufs)
851 hdr->mhdr.num_buffers = 0;
852
853 if (can_push) {
854 __skb_push(skb, hdr_len);
855 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
856 /* Pull header back to avoid skew in tx bytes calculations. */
857 __skb_pull(skb, hdr_len);
858 } else {
859 sg_set_buf(sq->sg, hdr, hdr_len);
860 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
861 }
862 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
863}
864
865static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
866{
867 struct virtnet_info *vi = netdev_priv(dev);
868 int qnum = skb_get_queue_mapping(skb);
869 struct send_queue *sq = &vi->sq[qnum];
870 int err;
871
872 /* Free up any pending old buffers before queueing new ones. */
873 free_old_xmit_skbs(sq);
874
875 /* Try to transmit */
876 err = xmit_skb(sq, skb);
877
878 /* This should not happen! */
879 if (unlikely(err)) {
880 dev->stats.tx_fifo_errors++;
881 if (net_ratelimit())
882 dev_warn(&dev->dev,
883 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
884 dev->stats.tx_dropped++;
885 dev_kfree_skb_any(skb);
886 return NETDEV_TX_OK;
887 }
888 virtqueue_kick(sq->vq);
889
890 /* Don't wait up for transmitted skbs to be freed. */
891 skb_orphan(skb);
892 nf_reset(skb);
893
894 /* Apparently nice girls don't return TX_BUSY; stop the queue
895 * before it gets out of hand. Naturally, this wastes entries. */
896 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
897 netif_stop_subqueue(dev, qnum);
898 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
899 /* More just got used, free them then recheck. */
900 free_old_xmit_skbs(sq);
901 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
902 netif_start_subqueue(dev, qnum);
903 virtqueue_disable_cb(sq->vq);
904 }
905 }
906 }
907
908 return NETDEV_TX_OK;
909}
910
911/*
912 * Send command via the control virtqueue and check status. Commands
913 * supported by the hypervisor, as indicated by feature bits, should
914 * never fail unless improperly formatted.
915 */
916static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
917 struct scatterlist *out)
918{
919 struct scatterlist *sgs[4], hdr, stat;
920 struct virtio_net_ctrl_hdr ctrl;
921 virtio_net_ctrl_ack status = ~0;
922 unsigned out_num = 0, tmp;
923
924 /* Caller should know better */
925 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
926
927 ctrl.class = class;
928 ctrl.cmd = cmd;
929 /* Add header */
930 sg_init_one(&hdr, &ctrl, sizeof(ctrl));
931 sgs[out_num++] = &hdr;
932
933 if (out)
934 sgs[out_num++] = out;
935
936 /* Add return status. */
937 sg_init_one(&stat, &status, sizeof(status));
938 sgs[out_num] = &stat;
939
940 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
941 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
942
943 if (unlikely(!virtqueue_kick(vi->cvq)))
944 return status == VIRTIO_NET_OK;
945
946 /* Spin for a response, the kick causes an ioport write, trapping
947 * into the hypervisor, so the request should be handled immediately.
948 */
949 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
950 !virtqueue_is_broken(vi->cvq))
951 cpu_relax();
952
953 return status == VIRTIO_NET_OK;
954}
955
956static int virtnet_set_mac_address(struct net_device *dev, void *p)
957{
958 struct virtnet_info *vi = netdev_priv(dev);
959 struct virtio_device *vdev = vi->vdev;
960 int ret;
961 struct sockaddr *addr = p;
962 struct scatterlist sg;
963
964 ret = eth_prepare_mac_addr_change(dev, p);
965 if (ret)
966 return ret;
967
968 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
969 sg_init_one(&sg, addr->sa_data, dev->addr_len);
970 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
971 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
972 dev_warn(&vdev->dev,
973 "Failed to set mac address by vq command.\n");
974 return -EINVAL;
975 }
976 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
977 unsigned int i;
978
979 /* Naturally, this has an atomicity problem. */
980 for (i = 0; i < dev->addr_len; i++)
981 virtio_cwrite8(vdev,
982 offsetof(struct virtio_net_config, mac) +
983 i, addr->sa_data[i]);
984 }
985
986 eth_commit_mac_addr_change(dev, p);
987
988 return 0;
989}
990
991static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
992 struct rtnl_link_stats64 *tot)
993{
994 struct virtnet_info *vi = netdev_priv(dev);
995 int cpu;
996 unsigned int start;
997
998 for_each_possible_cpu(cpu) {
999 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1000 u64 tpackets, tbytes, rpackets, rbytes;
1001
1002 do {
1003 start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1004 tpackets = stats->tx_packets;
1005 tbytes = stats->tx_bytes;
1006 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1007
1008 do {
1009 start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1010 rpackets = stats->rx_packets;
1011 rbytes = stats->rx_bytes;
1012 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1013
1014 tot->rx_packets += rpackets;
1015 tot->tx_packets += tpackets;
1016 tot->rx_bytes += rbytes;
1017 tot->tx_bytes += tbytes;
1018 }
1019
1020 tot->tx_dropped = dev->stats.tx_dropped;
1021 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1022 tot->rx_dropped = dev->stats.rx_dropped;
1023 tot->rx_length_errors = dev->stats.rx_length_errors;
1024 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1025
1026 return tot;
1027}
1028
1029#ifdef CONFIG_NET_POLL_CONTROLLER
1030static void virtnet_netpoll(struct net_device *dev)
1031{
1032 struct virtnet_info *vi = netdev_priv(dev);
1033 int i;
1034
1035 for (i = 0; i < vi->curr_queue_pairs; i++)
1036 napi_schedule(&vi->rq[i].napi);
1037}
1038#endif
1039
1040static void virtnet_ack_link_announce(struct virtnet_info *vi)
1041{
1042 rtnl_lock();
1043 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1044 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1045 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1046 rtnl_unlock();
1047}
1048
1049static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1050{
1051 struct scatterlist sg;
1052 struct virtio_net_ctrl_mq s;
1053 struct net_device *dev = vi->dev;
1054
1055 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1056 return 0;
1057
1058 s.virtqueue_pairs = queue_pairs;
1059 sg_init_one(&sg, &s, sizeof(s));
1060
1061 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1062 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1063 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1064 queue_pairs);
1065 return -EINVAL;
1066 } else {
1067 vi->curr_queue_pairs = queue_pairs;
1068 /* virtnet_open() will refill when device is going to up. */
1069 if (dev->flags & IFF_UP)
1070 schedule_delayed_work(&vi->refill, 0);
1071 }
1072
1073 return 0;
1074}
1075
1076static int virtnet_close(struct net_device *dev)
1077{
1078 struct virtnet_info *vi = netdev_priv(dev);
1079 int i;
1080
1081 /* Make sure refill_work doesn't re-enable napi! */
1082 cancel_delayed_work_sync(&vi->refill);
1083
1084 for (i = 0; i < vi->max_queue_pairs; i++)
1085 napi_disable(&vi->rq[i].napi);
1086
1087 return 0;
1088}
1089
1090static void virtnet_set_rx_mode(struct net_device *dev)
1091{
1092 struct virtnet_info *vi = netdev_priv(dev);
1093 struct scatterlist sg[2];
1094 u8 promisc, allmulti;
1095 struct virtio_net_ctrl_mac *mac_data;
1096 struct netdev_hw_addr *ha;
1097 int uc_count;
1098 int mc_count;
1099 void *buf;
1100 int i;
1101
1102 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1103 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1104 return;
1105
1106 promisc = ((dev->flags & IFF_PROMISC) != 0);
1107 allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1108
1109 sg_init_one(sg, &promisc, sizeof(promisc));
1110
1111 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1112 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1113 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1114 promisc ? "en" : "dis");
1115
1116 sg_init_one(sg, &allmulti, sizeof(allmulti));
1117
1118 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1119 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1120 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1121 allmulti ? "en" : "dis");
1122
1123 uc_count = netdev_uc_count(dev);
1124 mc_count = netdev_mc_count(dev);
1125 /* MAC filter - use one buffer for both lists */
1126 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1127 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1128 mac_data = buf;
1129 if (!buf)
1130 return;
1131
1132 sg_init_table(sg, 2);
1133
1134 /* Store the unicast list and count in the front of the buffer */
1135 mac_data->entries = uc_count;
1136 i = 0;
1137 netdev_for_each_uc_addr(ha, dev)
1138 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1139
1140 sg_set_buf(&sg[0], mac_data,
1141 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1142
1143 /* multicast list and count fill the end */
1144 mac_data = (void *)&mac_data->macs[uc_count][0];
1145
1146 mac_data->entries = mc_count;
1147 i = 0;
1148 netdev_for_each_mc_addr(ha, dev)
1149 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1150
1151 sg_set_buf(&sg[1], mac_data,
1152 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1153
1154 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1155 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1156 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1157
1158 kfree(buf);
1159}
1160
1161static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1162 __be16 proto, u16 vid)
1163{
1164 struct virtnet_info *vi = netdev_priv(dev);
1165 struct scatterlist sg;
1166
1167 sg_init_one(&sg, &vid, sizeof(vid));
1168
1169 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1170 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1171 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1172 return 0;
1173}
1174
1175static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1176 __be16 proto, u16 vid)
1177{
1178 struct virtnet_info *vi = netdev_priv(dev);
1179 struct scatterlist sg;
1180
1181 sg_init_one(&sg, &vid, sizeof(vid));
1182
1183 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1184 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1185 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1186 return 0;
1187}
1188
1189static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1190{
1191 int i;
1192
1193 if (vi->affinity_hint_set) {
1194 for (i = 0; i < vi->max_queue_pairs; i++) {
1195 virtqueue_set_affinity(vi->rq[i].vq, -1);
1196 virtqueue_set_affinity(vi->sq[i].vq, -1);
1197 }
1198
1199 vi->affinity_hint_set = false;
1200 }
1201}
1202
1203static void virtnet_set_affinity(struct virtnet_info *vi)
1204{
1205 int i;
1206 int cpu;
1207
1208 /* In multiqueue mode, when the number of cpu is equal to the number of
1209 * queue pairs, we let the queue pairs to be private to one cpu by
1210 * setting the affinity hint to eliminate the contention.
1211 */
1212 if (vi->curr_queue_pairs == 1 ||
1213 vi->max_queue_pairs != num_online_cpus()) {
1214 virtnet_clean_affinity(vi, -1);
1215 return;
1216 }
1217
1218 i = 0;
1219 for_each_online_cpu(cpu) {
1220 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1221 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1222 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1223 i++;
1224 }
1225
1226 vi->affinity_hint_set = true;
1227}
1228
1229static int virtnet_cpu_callback(struct notifier_block *nfb,
1230 unsigned long action, void *hcpu)
1231{
1232 struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1233
1234 switch(action & ~CPU_TASKS_FROZEN) {
1235 case CPU_ONLINE:
1236 case CPU_DOWN_FAILED:
1237 case CPU_DEAD:
1238 virtnet_set_affinity(vi);
1239 break;
1240 case CPU_DOWN_PREPARE:
1241 virtnet_clean_affinity(vi, (long)hcpu);
1242 break;
1243 default:
1244 break;
1245 }
1246
1247 return NOTIFY_OK;
1248}
1249
1250static void virtnet_get_ringparam(struct net_device *dev,
1251 struct ethtool_ringparam *ring)
1252{
1253 struct virtnet_info *vi = netdev_priv(dev);
1254
1255 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1256 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1257 ring->rx_pending = ring->rx_max_pending;
1258 ring->tx_pending = ring->tx_max_pending;
1259}
1260
1261
1262static void virtnet_get_drvinfo(struct net_device *dev,
1263 struct ethtool_drvinfo *info)
1264{
1265 struct virtnet_info *vi = netdev_priv(dev);
1266 struct virtio_device *vdev = vi->vdev;
1267
1268 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1269 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1270 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1271
1272}
1273
1274/* TODO: Eliminate OOO packets during switching */
1275static int virtnet_set_channels(struct net_device *dev,
1276 struct ethtool_channels *channels)
1277{
1278 struct virtnet_info *vi = netdev_priv(dev);
1279 u16 queue_pairs = channels->combined_count;
1280 int err;
1281
1282 /* We don't support separate rx/tx channels.
1283 * We don't allow setting 'other' channels.
1284 */
1285 if (channels->rx_count || channels->tx_count || channels->other_count)
1286 return -EINVAL;
1287
1288 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1289 return -EINVAL;
1290
1291 get_online_cpus();
1292 err = virtnet_set_queues(vi, queue_pairs);
1293 if (!err) {
1294 netif_set_real_num_tx_queues(dev, queue_pairs);
1295 netif_set_real_num_rx_queues(dev, queue_pairs);
1296
1297 virtnet_set_affinity(vi);
1298 }
1299 put_online_cpus();
1300
1301 return err;
1302}
1303
1304static void virtnet_get_channels(struct net_device *dev,
1305 struct ethtool_channels *channels)
1306{
1307 struct virtnet_info *vi = netdev_priv(dev);
1308
1309 channels->combined_count = vi->curr_queue_pairs;
1310 channels->max_combined = vi->max_queue_pairs;
1311 channels->max_other = 0;
1312 channels->rx_count = 0;
1313 channels->tx_count = 0;
1314 channels->other_count = 0;
1315}
1316
1317static const struct ethtool_ops virtnet_ethtool_ops = {
1318 .get_drvinfo = virtnet_get_drvinfo,
1319 .get_link = ethtool_op_get_link,
1320 .get_ringparam = virtnet_get_ringparam,
1321 .set_channels = virtnet_set_channels,
1322 .get_channels = virtnet_get_channels,
1323};
1324
1325#define MIN_MTU 68
1326#define MAX_MTU 65535
1327
1328static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1329{
1330 if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1331 return -EINVAL;
1332 dev->mtu = new_mtu;
1333 return 0;
1334}
1335
1336static const struct net_device_ops virtnet_netdev = {
1337 .ndo_open = virtnet_open,
1338 .ndo_stop = virtnet_close,
1339 .ndo_start_xmit = start_xmit,
1340 .ndo_validate_addr = eth_validate_addr,
1341 .ndo_set_mac_address = virtnet_set_mac_address,
1342 .ndo_set_rx_mode = virtnet_set_rx_mode,
1343 .ndo_change_mtu = virtnet_change_mtu,
1344 .ndo_get_stats64 = virtnet_stats,
1345 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1346 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1347#ifdef CONFIG_NET_POLL_CONTROLLER
1348 .ndo_poll_controller = virtnet_netpoll,
1349#endif
1350};
1351
1352static void virtnet_config_changed_work(struct work_struct *work)
1353{
1354 struct virtnet_info *vi =
1355 container_of(work, struct virtnet_info, config_work);
1356 u16 v;
1357
1358 mutex_lock(&vi->config_lock);
1359 if (!vi->config_enable)
1360 goto done;
1361
1362 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1363 struct virtio_net_config, status, &v) < 0)
1364 goto done;
1365
1366 if (v & VIRTIO_NET_S_ANNOUNCE) {
1367 netdev_notify_peers(vi->dev);
1368 virtnet_ack_link_announce(vi);
1369 }
1370
1371 /* Ignore unknown (future) status bits */
1372 v &= VIRTIO_NET_S_LINK_UP;
1373
1374 if (vi->status == v)
1375 goto done;
1376
1377 vi->status = v;
1378
1379 if (vi->status & VIRTIO_NET_S_LINK_UP) {
1380 netif_carrier_on(vi->dev);
1381 netif_tx_wake_all_queues(vi->dev);
1382 } else {
1383 netif_carrier_off(vi->dev);
1384 netif_tx_stop_all_queues(vi->dev);
1385 }
1386done:
1387 mutex_unlock(&vi->config_lock);
1388}
1389
1390static void virtnet_config_changed(struct virtio_device *vdev)
1391{
1392 struct virtnet_info *vi = vdev->priv;
1393
1394 schedule_work(&vi->config_work);
1395}
1396
1397static void virtnet_free_queues(struct virtnet_info *vi)
1398{
1399 int i;
1400
1401 for (i = 0; i < vi->max_queue_pairs; i++)
1402 netif_napi_del(&vi->rq[i].napi);
1403
1404 kfree(vi->rq);
1405 kfree(vi->sq);
1406}
1407
1408static void free_receive_bufs(struct virtnet_info *vi)
1409{
1410 int i;
1411
1412 for (i = 0; i < vi->max_queue_pairs; i++) {
1413 while (vi->rq[i].pages)
1414 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1415 }
1416}
1417
1418static void free_receive_page_frags(struct virtnet_info *vi)
1419{
1420 int i;
1421 for (i = 0; i < vi->max_queue_pairs; i++)
1422 if (vi->rq[i].alloc_frag.page)
1423 put_page(vi->rq[i].alloc_frag.page);
1424}
1425
1426static void free_unused_bufs(struct virtnet_info *vi)
1427{
1428 void *buf;
1429 int i;
1430
1431 for (i = 0; i < vi->max_queue_pairs; i++) {
1432 struct virtqueue *vq = vi->sq[i].vq;
1433 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1434 dev_kfree_skb(buf);
1435 }
1436
1437 for (i = 0; i < vi->max_queue_pairs; i++) {
1438 struct virtqueue *vq = vi->rq[i].vq;
1439
1440 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1441 if (vi->mergeable_rx_bufs) {
1442 unsigned long ctx = (unsigned long)buf;
1443 void *base = mergeable_ctx_to_buf_address(ctx);
1444 put_page(virt_to_head_page(base));
1445 } else if (vi->big_packets) {
1446 give_pages(&vi->rq[i], buf);
1447 } else {
1448 dev_kfree_skb(buf);
1449 }
1450 }
1451 }
1452}
1453
1454static void virtnet_del_vqs(struct virtnet_info *vi)
1455{
1456 struct virtio_device *vdev = vi->vdev;
1457
1458 virtnet_clean_affinity(vi, -1);
1459
1460 vdev->config->del_vqs(vdev);
1461
1462 virtnet_free_queues(vi);
1463}
1464
1465static int virtnet_find_vqs(struct virtnet_info *vi)
1466{
1467 vq_callback_t **callbacks;
1468 struct virtqueue **vqs;
1469 int ret = -ENOMEM;
1470 int i, total_vqs;
1471 const char **names;
1472
1473 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1474 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1475 * possible control vq.
1476 */
1477 total_vqs = vi->max_queue_pairs * 2 +
1478 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1479
1480 /* Allocate space for find_vqs parameters */
1481 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1482 if (!vqs)
1483 goto err_vq;
1484 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1485 if (!callbacks)
1486 goto err_callback;
1487 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1488 if (!names)
1489 goto err_names;
1490
1491 /* Parameters for control virtqueue, if any */
1492 if (vi->has_cvq) {
1493 callbacks[total_vqs - 1] = NULL;
1494 names[total_vqs - 1] = "control";
1495 }
1496
1497 /* Allocate/initialize parameters for send/receive virtqueues */
1498 for (i = 0; i < vi->max_queue_pairs; i++) {
1499 callbacks[rxq2vq(i)] = skb_recv_done;
1500 callbacks[txq2vq(i)] = skb_xmit_done;
1501 sprintf(vi->rq[i].name, "input.%d", i);
1502 sprintf(vi->sq[i].name, "output.%d", i);
1503 names[rxq2vq(i)] = vi->rq[i].name;
1504 names[txq2vq(i)] = vi->sq[i].name;
1505 }
1506
1507 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1508 names);
1509 if (ret)
1510 goto err_find;
1511
1512 if (vi->has_cvq) {
1513 vi->cvq = vqs[total_vqs - 1];
1514 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1515 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1516 }
1517
1518 for (i = 0; i < vi->max_queue_pairs; i++) {
1519 vi->rq[i].vq = vqs[rxq2vq(i)];
1520 vi->sq[i].vq = vqs[txq2vq(i)];
1521 }
1522
1523 kfree(names);
1524 kfree(callbacks);
1525 kfree(vqs);
1526
1527 return 0;
1528
1529err_find:
1530 kfree(names);
1531err_names:
1532 kfree(callbacks);
1533err_callback:
1534 kfree(vqs);
1535err_vq:
1536 return ret;
1537}
1538
1539static int virtnet_alloc_queues(struct virtnet_info *vi)
1540{
1541 int i;
1542
1543 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1544 if (!vi->sq)
1545 goto err_sq;
1546 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1547 if (!vi->rq)
1548 goto err_rq;
1549
1550 INIT_DELAYED_WORK(&vi->refill, refill_work);
1551 for (i = 0; i < vi->max_queue_pairs; i++) {
1552 vi->rq[i].pages = NULL;
1553 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1554 napi_weight);
1555
1556 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1557 ewma_init(&vi->rq[i].mrg_avg_pkt_len, 1, RECEIVE_AVG_WEIGHT);
1558 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1559 }
1560
1561 return 0;
1562
1563err_rq:
1564 kfree(vi->sq);
1565err_sq:
1566 return -ENOMEM;
1567}
1568
1569static int init_vqs(struct virtnet_info *vi)
1570{
1571 int ret;
1572
1573 /* Allocate send & receive queues */
1574 ret = virtnet_alloc_queues(vi);
1575 if (ret)
1576 goto err;
1577
1578 ret = virtnet_find_vqs(vi);
1579 if (ret)
1580 goto err_free;
1581
1582 get_online_cpus();
1583 virtnet_set_affinity(vi);
1584 put_online_cpus();
1585
1586 return 0;
1587
1588err_free:
1589 virtnet_free_queues(vi);
1590err:
1591 return ret;
1592}
1593
1594#ifdef CONFIG_SYSFS
1595static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1596 struct rx_queue_attribute *attribute, char *buf)
1597{
1598 struct virtnet_info *vi = netdev_priv(queue->dev);
1599 unsigned int queue_index = get_netdev_rx_queue_index(queue);
1600 struct ewma *avg;
1601
1602 BUG_ON(queue_index >= vi->max_queue_pairs);
1603 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1604 return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1605}
1606
1607static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1608 __ATTR_RO(mergeable_rx_buffer_size);
1609
1610static struct attribute *virtio_net_mrg_rx_attrs[] = {
1611 &mergeable_rx_buffer_size_attribute.attr,
1612 NULL
1613};
1614
1615static const struct attribute_group virtio_net_mrg_rx_group = {
1616 .name = "virtio_net",
1617 .attrs = virtio_net_mrg_rx_attrs
1618};
1619#endif
1620
1621static int virtnet_probe(struct virtio_device *vdev)
1622{
1623 int i, err;
1624 struct net_device *dev;
1625 struct virtnet_info *vi;
1626 u16 max_queue_pairs;
1627
1628 /* Find if host supports multiqueue virtio_net device */
1629 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1630 struct virtio_net_config,
1631 max_virtqueue_pairs, &max_queue_pairs);
1632
1633 /* We need at least 2 queue's */
1634 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1635 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1636 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1637 max_queue_pairs = 1;
1638
1639 /* Allocate ourselves a network device with room for our info */
1640 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1641 if (!dev)
1642 return -ENOMEM;
1643
1644 /* Set up network device as normal. */
1645 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1646 dev->netdev_ops = &virtnet_netdev;
1647 dev->features = NETIF_F_HIGHDMA;
1648
1649 SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1650 SET_NETDEV_DEV(dev, &vdev->dev);
1651
1652 /* Do we support "hardware" checksums? */
1653 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1654 /* This opens up the world of extra features. */
1655 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1656 if (csum)
1657 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1658
1659 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1660 dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1661 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1662 }
1663 /* Individual feature bits: what can host handle? */
1664 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1665 dev->hw_features |= NETIF_F_TSO;
1666 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1667 dev->hw_features |= NETIF_F_TSO6;
1668 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1669 dev->hw_features |= NETIF_F_TSO_ECN;
1670 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1671 dev->hw_features |= NETIF_F_UFO;
1672
1673 if (gso)
1674 dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1675 /* (!csum && gso) case will be fixed by register_netdev() */
1676 }
1677 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1678 dev->features |= NETIF_F_RXCSUM;
1679
1680 dev->vlan_features = dev->features;
1681
1682 /* Configuration may specify what MAC to use. Otherwise random. */
1683 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1684 virtio_cread_bytes(vdev,
1685 offsetof(struct virtio_net_config, mac),
1686 dev->dev_addr, dev->addr_len);
1687 else
1688 eth_hw_addr_random(dev);
1689
1690 /* Set up our device-specific information */
1691 vi = netdev_priv(dev);
1692 vi->dev = dev;
1693 vi->vdev = vdev;
1694 vdev->priv = vi;
1695 vi->stats = alloc_percpu(struct virtnet_stats);
1696 err = -ENOMEM;
1697 if (vi->stats == NULL)
1698 goto free;
1699
1700 for_each_possible_cpu(i) {
1701 struct virtnet_stats *virtnet_stats;
1702 virtnet_stats = per_cpu_ptr(vi->stats, i);
1703 u64_stats_init(&virtnet_stats->tx_syncp);
1704 u64_stats_init(&virtnet_stats->rx_syncp);
1705 }
1706
1707 mutex_init(&vi->config_lock);
1708 vi->config_enable = true;
1709 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1710
1711 /* If we can receive ANY GSO packets, we must allocate large ones. */
1712 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1713 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1714 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
1715 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
1716 vi->big_packets = true;
1717
1718 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1719 vi->mergeable_rx_bufs = true;
1720
1721 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1722 vi->any_header_sg = true;
1723
1724 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1725 vi->has_cvq = true;
1726
1727 /* Use single tx/rx queue pair as default */
1728 vi->curr_queue_pairs = 1;
1729 vi->max_queue_pairs = max_queue_pairs;
1730
1731 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1732 err = init_vqs(vi);
1733 if (err)
1734 goto free_stats;
1735
1736#ifdef CONFIG_SYSFS
1737 if (vi->mergeable_rx_bufs)
1738 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
1739#endif
1740 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1741 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1742
1743 err = register_netdev(dev);
1744 if (err) {
1745 pr_debug("virtio_net: registering device failed\n");
1746 goto free_vqs;
1747 }
1748
1749 /* Last of all, set up some receive buffers. */
1750 for (i = 0; i < vi->curr_queue_pairs; i++) {
1751 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1752
1753 /* If we didn't even get one input buffer, we're useless. */
1754 if (vi->rq[i].vq->num_free ==
1755 virtqueue_get_vring_size(vi->rq[i].vq)) {
1756 free_unused_bufs(vi);
1757 err = -ENOMEM;
1758 goto free_recv_bufs;
1759 }
1760 }
1761
1762 vi->nb.notifier_call = &virtnet_cpu_callback;
1763 err = register_hotcpu_notifier(&vi->nb);
1764 if (err) {
1765 pr_debug("virtio_net: registering cpu notifier failed\n");
1766 goto free_recv_bufs;
1767 }
1768
1769 /* Assume link up if device can't report link status,
1770 otherwise get link status from config. */
1771 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1772 netif_carrier_off(dev);
1773 schedule_work(&vi->config_work);
1774 } else {
1775 vi->status = VIRTIO_NET_S_LINK_UP;
1776 netif_carrier_on(dev);
1777 }
1778
1779 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1780 dev->name, max_queue_pairs);
1781
1782 return 0;
1783
1784free_recv_bufs:
1785 free_receive_bufs(vi);
1786 unregister_netdev(dev);
1787free_vqs:
1788 cancel_delayed_work_sync(&vi->refill);
1789 free_receive_page_frags(vi);
1790 virtnet_del_vqs(vi);
1791free_stats:
1792 free_percpu(vi->stats);
1793free:
1794 free_netdev(dev);
1795 return err;
1796}
1797
1798static void remove_vq_common(struct virtnet_info *vi)
1799{
1800 vi->vdev->config->reset(vi->vdev);
1801
1802 /* Free unused buffers in both send and recv, if any. */
1803 free_unused_bufs(vi);
1804
1805 free_receive_bufs(vi);
1806
1807 free_receive_page_frags(vi);
1808
1809 virtnet_del_vqs(vi);
1810}
1811
1812static void virtnet_remove(struct virtio_device *vdev)
1813{
1814 struct virtnet_info *vi = vdev->priv;
1815
1816 unregister_hotcpu_notifier(&vi->nb);
1817
1818 /* Prevent config work handler from accessing the device. */
1819 mutex_lock(&vi->config_lock);
1820 vi->config_enable = false;
1821 mutex_unlock(&vi->config_lock);
1822
1823 unregister_netdev(vi->dev);
1824
1825 remove_vq_common(vi);
1826
1827 flush_work(&vi->config_work);
1828
1829 free_percpu(vi->stats);
1830 free_netdev(vi->dev);
1831}
1832
1833#ifdef CONFIG_PM_SLEEP
1834static int virtnet_freeze(struct virtio_device *vdev)
1835{
1836 struct virtnet_info *vi = vdev->priv;
1837 int i;
1838
1839 unregister_hotcpu_notifier(&vi->nb);
1840
1841 /* Prevent config work handler from accessing the device */
1842 mutex_lock(&vi->config_lock);
1843 vi->config_enable = false;
1844 mutex_unlock(&vi->config_lock);
1845
1846 netif_device_detach(vi->dev);
1847 cancel_delayed_work_sync(&vi->refill);
1848
1849 if (netif_running(vi->dev))
1850 for (i = 0; i < vi->max_queue_pairs; i++) {
1851 napi_disable(&vi->rq[i].napi);
1852 netif_napi_del(&vi->rq[i].napi);
1853 }
1854
1855 remove_vq_common(vi);
1856
1857 flush_work(&vi->config_work);
1858
1859 return 0;
1860}
1861
1862static int virtnet_restore(struct virtio_device *vdev)
1863{
1864 struct virtnet_info *vi = vdev->priv;
1865 int err, i;
1866
1867 err = init_vqs(vi);
1868 if (err)
1869 return err;
1870
1871 if (netif_running(vi->dev)) {
1872 for (i = 0; i < vi->curr_queue_pairs; i++)
1873 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1874 schedule_delayed_work(&vi->refill, 0);
1875
1876 for (i = 0; i < vi->max_queue_pairs; i++)
1877 virtnet_napi_enable(&vi->rq[i]);
1878 }
1879
1880 netif_device_attach(vi->dev);
1881
1882 mutex_lock(&vi->config_lock);
1883 vi->config_enable = true;
1884 mutex_unlock(&vi->config_lock);
1885
1886 rtnl_lock();
1887 virtnet_set_queues(vi, vi->curr_queue_pairs);
1888 rtnl_unlock();
1889
1890 err = register_hotcpu_notifier(&vi->nb);
1891 if (err)
1892 return err;
1893
1894 return 0;
1895}
1896#endif
1897
1898static struct virtio_device_id id_table[] = {
1899 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1900 { 0 },
1901};
1902
1903static unsigned int features[] = {
1904 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1905 VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1906 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1907 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1908 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1909 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1910 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1911 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1912 VIRTIO_NET_F_CTRL_MAC_ADDR,
1913 VIRTIO_F_ANY_LAYOUT,
1914};
1915
1916static struct virtio_driver virtio_net_driver = {
1917 .feature_table = features,
1918 .feature_table_size = ARRAY_SIZE(features),
1919 .driver.name = KBUILD_MODNAME,
1920 .driver.owner = THIS_MODULE,
1921 .id_table = id_table,
1922 .probe = virtnet_probe,
1923 .remove = virtnet_remove,
1924 .config_changed = virtnet_config_changed,
1925#ifdef CONFIG_PM_SLEEP
1926 .freeze = virtnet_freeze,
1927 .restore = virtnet_restore,
1928#endif
1929};
1930
1931module_virtio_driver(virtio_net_driver);
1932
1933MODULE_DEVICE_TABLE(virtio, id_table);
1934MODULE_DESCRIPTION("Virtio network driver");
1935MODULE_LICENSE("GPL");