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, 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");
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#include <net/busy_poll.h>
31
32static int napi_weight = NAPI_POLL_WEIGHT;
33module_param(napi_weight, int, 0444);
34
35static bool csum = true, gso = true;
36module_param(csum, bool, 0444);
37module_param(gso, bool, 0444);
38
39/* FIXME: MTU in config. */
40#define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
41#define GOOD_COPY_LEN 128
42
43/* RX packet size EWMA. The average packet size is used to determine the packet
44 * buffer size when refilling RX rings. As the entire RX ring may be refilled
45 * at once, the weight is chosen so that the EWMA will be insensitive to short-
46 * term, transient changes in packet size.
47 */
48DECLARE_EWMA(pkt_len, 1, 64)
49
50/* Minimum alignment for mergeable packet buffers. */
51#define MERGEABLE_BUFFER_ALIGN max(L1_CACHE_BYTES, 256)
52
53#define VIRTNET_DRIVER_VERSION "1.0.0"
54
55struct virtnet_stats {
56 struct u64_stats_sync tx_syncp;
57 struct u64_stats_sync rx_syncp;
58 u64 tx_bytes;
59 u64 tx_packets;
60
61 u64 rx_bytes;
62 u64 rx_packets;
63};
64
65/* Internal representation of a send virtqueue */
66struct send_queue {
67 /* Virtqueue associated with this send _queue */
68 struct virtqueue *vq;
69
70 /* TX: fragments + linear part + virtio header */
71 struct scatterlist sg[MAX_SKB_FRAGS + 2];
72
73 /* Name of the send queue: output.$index */
74 char name[40];
75};
76
77/* Internal representation of a receive virtqueue */
78struct receive_queue {
79 /* Virtqueue associated with this receive_queue */
80 struct virtqueue *vq;
81
82 struct napi_struct napi;
83
84 /* Chain pages by the private ptr. */
85 struct page *pages;
86
87 /* Average packet length for mergeable receive buffers. */
88 struct ewma_pkt_len mrg_avg_pkt_len;
89
90 /* Page frag for packet buffer allocation. */
91 struct page_frag alloc_frag;
92
93 /* RX: fragments + linear part + virtio header */
94 struct scatterlist sg[MAX_SKB_FRAGS + 2];
95
96 /* Name of this receive queue: input.$index */
97 char name[40];
98};
99
100struct virtnet_info {
101 struct virtio_device *vdev;
102 struct virtqueue *cvq;
103 struct net_device *dev;
104 struct send_queue *sq;
105 struct receive_queue *rq;
106 unsigned int status;
107
108 /* Max # of queue pairs supported by the device */
109 u16 max_queue_pairs;
110
111 /* # of queue pairs currently used by the driver */
112 u16 curr_queue_pairs;
113
114 /* I like... big packets and I cannot lie! */
115 bool big_packets;
116
117 /* Host will merge rx buffers for big packets (shake it! shake it!) */
118 bool mergeable_rx_bufs;
119
120 /* Has control virtqueue */
121 bool has_cvq;
122
123 /* Host can handle any s/g split between our header and packet data */
124 bool any_header_sg;
125
126 /* Packet virtio header size */
127 u8 hdr_len;
128
129 /* Active statistics */
130 struct virtnet_stats __percpu *stats;
131
132 /* Work struct for refilling if we run low on memory. */
133 struct delayed_work refill;
134
135 /* Work struct for config space updates */
136 struct work_struct config_work;
137
138 /* Does the affinity hint is set for virtqueues? */
139 bool affinity_hint_set;
140
141 /* CPU hot plug notifier */
142 struct notifier_block nb;
143
144 /* Control VQ buffers: protected by the rtnl lock */
145 struct virtio_net_ctrl_hdr ctrl_hdr;
146 virtio_net_ctrl_ack ctrl_status;
147 u8 ctrl_promisc;
148 u8 ctrl_allmulti;
149
150 /* Ethtool settings */
151 u8 duplex;
152 u32 speed;
153};
154
155struct padded_vnet_hdr {
156 struct virtio_net_hdr_mrg_rxbuf hdr;
157 /*
158 * hdr is in a separate sg buffer, and data sg buffer shares same page
159 * with this header sg. This padding makes next sg 16 byte aligned
160 * after the header.
161 */
162 char padding[4];
163};
164
165/* Converting between virtqueue no. and kernel tx/rx queue no.
166 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
167 */
168static int vq2txq(struct virtqueue *vq)
169{
170 return (vq->index - 1) / 2;
171}
172
173static int txq2vq(int txq)
174{
175 return txq * 2 + 1;
176}
177
178static int vq2rxq(struct virtqueue *vq)
179{
180 return vq->index / 2;
181}
182
183static int rxq2vq(int rxq)
184{
185 return rxq * 2;
186}
187
188static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
189{
190 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
191}
192
193/*
194 * private is used to chain pages for big packets, put the whole
195 * most recent used list in the beginning for reuse
196 */
197static void give_pages(struct receive_queue *rq, struct page *page)
198{
199 struct page *end;
200
201 /* Find end of list, sew whole thing into vi->rq.pages. */
202 for (end = page; end->private; end = (struct page *)end->private);
203 end->private = (unsigned long)rq->pages;
204 rq->pages = page;
205}
206
207static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
208{
209 struct page *p = rq->pages;
210
211 if (p) {
212 rq->pages = (struct page *)p->private;
213 /* clear private here, it is used to chain pages */
214 p->private = 0;
215 } else
216 p = alloc_page(gfp_mask);
217 return p;
218}
219
220static void skb_xmit_done(struct virtqueue *vq)
221{
222 struct virtnet_info *vi = vq->vdev->priv;
223
224 /* Suppress further interrupts. */
225 virtqueue_disable_cb(vq);
226
227 /* We were probably waiting for more output buffers. */
228 netif_wake_subqueue(vi->dev, vq2txq(vq));
229}
230
231static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
232{
233 unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
234 return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
235}
236
237static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
238{
239 return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
240
241}
242
243static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
244{
245 unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
246 return (unsigned long)buf | (size - 1);
247}
248
249/* Called from bottom half context */
250static struct sk_buff *page_to_skb(struct virtnet_info *vi,
251 struct receive_queue *rq,
252 struct page *page, unsigned int offset,
253 unsigned int len, unsigned int truesize)
254{
255 struct sk_buff *skb;
256 struct virtio_net_hdr_mrg_rxbuf *hdr;
257 unsigned int copy, hdr_len, hdr_padded_len;
258 char *p;
259
260 p = page_address(page) + offset;
261
262 /* copy small packet so we can reuse these pages for small data */
263 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
264 if (unlikely(!skb))
265 return NULL;
266
267 hdr = skb_vnet_hdr(skb);
268
269 hdr_len = vi->hdr_len;
270 if (vi->mergeable_rx_bufs)
271 hdr_padded_len = sizeof *hdr;
272 else
273 hdr_padded_len = sizeof(struct padded_vnet_hdr);
274
275 memcpy(hdr, p, hdr_len);
276
277 len -= hdr_len;
278 offset += hdr_padded_len;
279 p += hdr_padded_len;
280
281 copy = len;
282 if (copy > skb_tailroom(skb))
283 copy = skb_tailroom(skb);
284 memcpy(skb_put(skb, copy), p, copy);
285
286 len -= copy;
287 offset += copy;
288
289 if (vi->mergeable_rx_bufs) {
290 if (len)
291 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
292 else
293 put_page(page);
294 return skb;
295 }
296
297 /*
298 * Verify that we can indeed put this data into a skb.
299 * This is here to handle cases when the device erroneously
300 * tries to receive more than is possible. This is usually
301 * the case of a broken device.
302 */
303 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
304 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
305 dev_kfree_skb(skb);
306 return NULL;
307 }
308 BUG_ON(offset >= PAGE_SIZE);
309 while (len) {
310 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
311 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
312 frag_size, truesize);
313 len -= frag_size;
314 page = (struct page *)page->private;
315 offset = 0;
316 }
317
318 if (page)
319 give_pages(rq, page);
320
321 return skb;
322}
323
324static struct sk_buff *receive_small(struct virtnet_info *vi, void *buf, unsigned int len)
325{
326 struct sk_buff * skb = buf;
327
328 len -= vi->hdr_len;
329 skb_trim(skb, len);
330
331 return skb;
332}
333
334static struct sk_buff *receive_big(struct net_device *dev,
335 struct virtnet_info *vi,
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(vi, 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 virtnet_info *vi,
356 struct receive_queue *rq,
357 unsigned long ctx,
358 unsigned int len)
359{
360 void *buf = mergeable_ctx_to_buf_address(ctx);
361 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
362 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
363 struct page *page = virt_to_head_page(buf);
364 int offset = buf - page_address(page);
365 unsigned int truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
366
367 struct sk_buff *head_skb = page_to_skb(vi, rq, page, offset, len,
368 truesize);
369 struct sk_buff *curr_skb = head_skb;
370
371 if (unlikely(!curr_skb))
372 goto err_skb;
373 while (--num_buf) {
374 int num_skb_frags;
375
376 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
377 if (unlikely(!ctx)) {
378 pr_debug("%s: rx error: %d buffers out of %d missing\n",
379 dev->name, num_buf,
380 virtio16_to_cpu(vi->vdev,
381 hdr->num_buffers));
382 dev->stats.rx_length_errors++;
383 goto err_buf;
384 }
385
386 buf = mergeable_ctx_to_buf_address(ctx);
387 page = virt_to_head_page(buf);
388
389 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
390 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
391 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
392
393 if (unlikely(!nskb))
394 goto err_skb;
395 if (curr_skb == head_skb)
396 skb_shinfo(curr_skb)->frag_list = nskb;
397 else
398 curr_skb->next = nskb;
399 curr_skb = nskb;
400 head_skb->truesize += nskb->truesize;
401 num_skb_frags = 0;
402 }
403 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
404 if (curr_skb != head_skb) {
405 head_skb->data_len += len;
406 head_skb->len += len;
407 head_skb->truesize += truesize;
408 }
409 offset = buf - page_address(page);
410 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
411 put_page(page);
412 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
413 len, truesize);
414 } else {
415 skb_add_rx_frag(curr_skb, num_skb_frags, page,
416 offset, len, truesize);
417 }
418 }
419
420 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
421 return head_skb;
422
423err_skb:
424 put_page(page);
425 while (--num_buf) {
426 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
427 if (unlikely(!ctx)) {
428 pr_debug("%s: rx error: %d buffers missing\n",
429 dev->name, num_buf);
430 dev->stats.rx_length_errors++;
431 break;
432 }
433 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
434 put_page(page);
435 }
436err_buf:
437 dev->stats.rx_dropped++;
438 dev_kfree_skb(head_skb);
439 return NULL;
440}
441
442static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
443 void *buf, unsigned int len)
444{
445 struct net_device *dev = vi->dev;
446 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
447 struct sk_buff *skb;
448 struct virtio_net_hdr_mrg_rxbuf *hdr;
449
450 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
451 pr_debug("%s: short packet %i\n", dev->name, len);
452 dev->stats.rx_length_errors++;
453 if (vi->mergeable_rx_bufs) {
454 unsigned long ctx = (unsigned long)buf;
455 void *base = mergeable_ctx_to_buf_address(ctx);
456 put_page(virt_to_head_page(base));
457 } else if (vi->big_packets) {
458 give_pages(rq, buf);
459 } else {
460 dev_kfree_skb(buf);
461 }
462 return;
463 }
464
465 if (vi->mergeable_rx_bufs)
466 skb = receive_mergeable(dev, vi, rq, (unsigned long)buf, len);
467 else if (vi->big_packets)
468 skb = receive_big(dev, vi, rq, buf, len);
469 else
470 skb = receive_small(vi, buf, len);
471
472 if (unlikely(!skb))
473 return;
474
475 hdr = skb_vnet_hdr(skb);
476
477 u64_stats_update_begin(&stats->rx_syncp);
478 stats->rx_bytes += skb->len;
479 stats->rx_packets++;
480 u64_stats_update_end(&stats->rx_syncp);
481
482 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
483 pr_debug("Needs csum!\n");
484 if (!skb_partial_csum_set(skb,
485 virtio16_to_cpu(vi->vdev, hdr->hdr.csum_start),
486 virtio16_to_cpu(vi->vdev, hdr->hdr.csum_offset)))
487 goto frame_err;
488 } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
489 skb->ip_summed = CHECKSUM_UNNECESSARY;
490 }
491
492 skb->protocol = eth_type_trans(skb, dev);
493 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
494 ntohs(skb->protocol), skb->len, skb->pkt_type);
495
496 if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
497 pr_debug("GSO!\n");
498 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
499 case VIRTIO_NET_HDR_GSO_TCPV4:
500 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
501 break;
502 case VIRTIO_NET_HDR_GSO_UDP:
503 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
504 break;
505 case VIRTIO_NET_HDR_GSO_TCPV6:
506 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
507 break;
508 default:
509 net_warn_ratelimited("%s: bad gso type %u.\n",
510 dev->name, hdr->hdr.gso_type);
511 goto frame_err;
512 }
513
514 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
515 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
516
517 skb_shinfo(skb)->gso_size = virtio16_to_cpu(vi->vdev,
518 hdr->hdr.gso_size);
519 if (skb_shinfo(skb)->gso_size == 0) {
520 net_warn_ratelimited("%s: zero gso size.\n", dev->name);
521 goto frame_err;
522 }
523
524 /* Header must be checked, and gso_segs computed. */
525 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
526 skb_shinfo(skb)->gso_segs = 0;
527 }
528
529 napi_gro_receive(&rq->napi, skb);
530 return;
531
532frame_err:
533 dev->stats.rx_frame_errors++;
534 dev_kfree_skb(skb);
535}
536
537static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
538 gfp_t gfp)
539{
540 struct sk_buff *skb;
541 struct virtio_net_hdr_mrg_rxbuf *hdr;
542 int err;
543
544 skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
545 if (unlikely(!skb))
546 return -ENOMEM;
547
548 skb_put(skb, GOOD_PACKET_LEN);
549
550 hdr = skb_vnet_hdr(skb);
551 sg_init_table(rq->sg, 2);
552 sg_set_buf(rq->sg, hdr, vi->hdr_len);
553 skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
554
555 err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
556 if (err < 0)
557 dev_kfree_skb(skb);
558
559 return err;
560}
561
562static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
563 gfp_t gfp)
564{
565 struct page *first, *list = NULL;
566 char *p;
567 int i, err, offset;
568
569 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
570
571 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
572 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
573 first = get_a_page(rq, gfp);
574 if (!first) {
575 if (list)
576 give_pages(rq, list);
577 return -ENOMEM;
578 }
579 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
580
581 /* chain new page in list head to match sg */
582 first->private = (unsigned long)list;
583 list = first;
584 }
585
586 first = get_a_page(rq, gfp);
587 if (!first) {
588 give_pages(rq, list);
589 return -ENOMEM;
590 }
591 p = page_address(first);
592
593 /* rq->sg[0], rq->sg[1] share the same page */
594 /* a separated rq->sg[0] for header - required in case !any_header_sg */
595 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
596
597 /* rq->sg[1] for data packet, from offset */
598 offset = sizeof(struct padded_vnet_hdr);
599 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
600
601 /* chain first in list head */
602 first->private = (unsigned long)list;
603 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
604 first, gfp);
605 if (err < 0)
606 give_pages(rq, first);
607
608 return err;
609}
610
611static unsigned int get_mergeable_buf_len(struct ewma_pkt_len *avg_pkt_len)
612{
613 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
614 unsigned int len;
615
616 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
617 GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
618 return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
619}
620
621static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
622{
623 struct page_frag *alloc_frag = &rq->alloc_frag;
624 char *buf;
625 unsigned long ctx;
626 int err;
627 unsigned int len, hole;
628
629 len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
630 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
631 return -ENOMEM;
632
633 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
634 ctx = mergeable_buf_to_ctx(buf, len);
635 get_page(alloc_frag->page);
636 alloc_frag->offset += len;
637 hole = alloc_frag->size - alloc_frag->offset;
638 if (hole < len) {
639 /* To avoid internal fragmentation, if there is very likely not
640 * enough space for another buffer, add the remaining space to
641 * the current buffer. This extra space is not included in
642 * the truesize stored in ctx.
643 */
644 len += hole;
645 alloc_frag->offset += hole;
646 }
647
648 sg_init_one(rq->sg, buf, len);
649 err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
650 if (err < 0)
651 put_page(virt_to_head_page(buf));
652
653 return err;
654}
655
656/*
657 * Returns false if we couldn't fill entirely (OOM).
658 *
659 * Normally run in the receive path, but can also be run from ndo_open
660 * before we're receiving packets, or from refill_work which is
661 * careful to disable receiving (using napi_disable).
662 */
663static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
664 gfp_t gfp)
665{
666 int err;
667 bool oom;
668
669 gfp |= __GFP_COLD;
670 do {
671 if (vi->mergeable_rx_bufs)
672 err = add_recvbuf_mergeable(rq, gfp);
673 else if (vi->big_packets)
674 err = add_recvbuf_big(vi, rq, gfp);
675 else
676 err = add_recvbuf_small(vi, rq, gfp);
677
678 oom = err == -ENOMEM;
679 if (err)
680 break;
681 } while (rq->vq->num_free);
682 virtqueue_kick(rq->vq);
683 return !oom;
684}
685
686static void skb_recv_done(struct virtqueue *rvq)
687{
688 struct virtnet_info *vi = rvq->vdev->priv;
689 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
690
691 /* Schedule NAPI, Suppress further interrupts if successful. */
692 if (napi_schedule_prep(&rq->napi)) {
693 virtqueue_disable_cb(rvq);
694 __napi_schedule(&rq->napi);
695 }
696}
697
698static void virtnet_napi_enable(struct receive_queue *rq)
699{
700 napi_enable(&rq->napi);
701
702 /* If all buffers were filled by other side before we napi_enabled, we
703 * won't get another interrupt, so process any outstanding packets
704 * now. virtnet_poll wants re-enable the queue, so we disable here.
705 * We synchronize against interrupts via NAPI_STATE_SCHED */
706 if (napi_schedule_prep(&rq->napi)) {
707 virtqueue_disable_cb(rq->vq);
708 local_bh_disable();
709 __napi_schedule(&rq->napi);
710 local_bh_enable();
711 }
712}
713
714static void refill_work(struct work_struct *work)
715{
716 struct virtnet_info *vi =
717 container_of(work, struct virtnet_info, refill.work);
718 bool still_empty;
719 int i;
720
721 for (i = 0; i < vi->curr_queue_pairs; i++) {
722 struct receive_queue *rq = &vi->rq[i];
723
724 napi_disable(&rq->napi);
725 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
726 virtnet_napi_enable(rq);
727
728 /* In theory, this can happen: if we don't get any buffers in
729 * we will *never* try to fill again.
730 */
731 if (still_empty)
732 schedule_delayed_work(&vi->refill, HZ/2);
733 }
734}
735
736static int virtnet_receive(struct receive_queue *rq, int budget)
737{
738 struct virtnet_info *vi = rq->vq->vdev->priv;
739 unsigned int len, received = 0;
740 void *buf;
741
742 while (received < budget &&
743 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
744 receive_buf(vi, rq, buf, len);
745 received++;
746 }
747
748 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
749 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
750 schedule_delayed_work(&vi->refill, 0);
751 }
752
753 return received;
754}
755
756static int virtnet_poll(struct napi_struct *napi, int budget)
757{
758 struct receive_queue *rq =
759 container_of(napi, struct receive_queue, napi);
760 unsigned int r, received;
761
762 received = virtnet_receive(rq, budget);
763
764 /* Out of packets? */
765 if (received < budget) {
766 r = virtqueue_enable_cb_prepare(rq->vq);
767 napi_complete_done(napi, received);
768 if (unlikely(virtqueue_poll(rq->vq, r)) &&
769 napi_schedule_prep(napi)) {
770 virtqueue_disable_cb(rq->vq);
771 __napi_schedule(napi);
772 }
773 }
774
775 return received;
776}
777
778#ifdef CONFIG_NET_RX_BUSY_POLL
779/* must be called with local_bh_disable()d */
780static int virtnet_busy_poll(struct napi_struct *napi)
781{
782 struct receive_queue *rq =
783 container_of(napi, struct receive_queue, napi);
784 struct virtnet_info *vi = rq->vq->vdev->priv;
785 int r, received = 0, budget = 4;
786
787 if (!(vi->status & VIRTIO_NET_S_LINK_UP))
788 return LL_FLUSH_FAILED;
789
790 if (!napi_schedule_prep(napi))
791 return LL_FLUSH_BUSY;
792
793 virtqueue_disable_cb(rq->vq);
794
795again:
796 received += virtnet_receive(rq, budget);
797
798 r = virtqueue_enable_cb_prepare(rq->vq);
799 clear_bit(NAPI_STATE_SCHED, &napi->state);
800 if (unlikely(virtqueue_poll(rq->vq, r)) &&
801 napi_schedule_prep(napi)) {
802 virtqueue_disable_cb(rq->vq);
803 if (received < budget) {
804 budget -= received;
805 goto again;
806 } else {
807 __napi_schedule(napi);
808 }
809 }
810
811 return received;
812}
813#endif /* CONFIG_NET_RX_BUSY_POLL */
814
815static int virtnet_open(struct net_device *dev)
816{
817 struct virtnet_info *vi = netdev_priv(dev);
818 int i;
819
820 for (i = 0; i < vi->max_queue_pairs; i++) {
821 if (i < vi->curr_queue_pairs)
822 /* Make sure we have some buffers: if oom use wq. */
823 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
824 schedule_delayed_work(&vi->refill, 0);
825 virtnet_napi_enable(&vi->rq[i]);
826 }
827
828 return 0;
829}
830
831static void free_old_xmit_skbs(struct send_queue *sq)
832{
833 struct sk_buff *skb;
834 unsigned int len;
835 struct virtnet_info *vi = sq->vq->vdev->priv;
836 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
837
838 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
839 pr_debug("Sent skb %p\n", skb);
840
841 u64_stats_update_begin(&stats->tx_syncp);
842 stats->tx_bytes += skb->len;
843 stats->tx_packets++;
844 u64_stats_update_end(&stats->tx_syncp);
845
846 dev_kfree_skb_any(skb);
847 }
848}
849
850static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
851{
852 struct virtio_net_hdr_mrg_rxbuf *hdr;
853 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
854 struct virtnet_info *vi = sq->vq->vdev->priv;
855 unsigned num_sg;
856 unsigned hdr_len = vi->hdr_len;
857 bool can_push;
858
859 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
860
861 can_push = vi->any_header_sg &&
862 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
863 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
864 /* Even if we can, don't push here yet as this would skew
865 * csum_start offset below. */
866 if (can_push)
867 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
868 else
869 hdr = skb_vnet_hdr(skb);
870
871 if (skb->ip_summed == CHECKSUM_PARTIAL) {
872 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
873 hdr->hdr.csum_start = cpu_to_virtio16(vi->vdev,
874 skb_checksum_start_offset(skb));
875 hdr->hdr.csum_offset = cpu_to_virtio16(vi->vdev,
876 skb->csum_offset);
877 } else {
878 hdr->hdr.flags = 0;
879 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
880 }
881
882 if (skb_is_gso(skb)) {
883 hdr->hdr.hdr_len = cpu_to_virtio16(vi->vdev, skb_headlen(skb));
884 hdr->hdr.gso_size = cpu_to_virtio16(vi->vdev,
885 skb_shinfo(skb)->gso_size);
886 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
887 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
888 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
889 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
890 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
891 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
892 else
893 BUG();
894 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
895 hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
896 } else {
897 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
898 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
899 }
900
901 if (vi->mergeable_rx_bufs)
902 hdr->num_buffers = 0;
903
904 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
905 if (can_push) {
906 __skb_push(skb, hdr_len);
907 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
908 /* Pull header back to avoid skew in tx bytes calculations. */
909 __skb_pull(skb, hdr_len);
910 } else {
911 sg_set_buf(sq->sg, hdr, hdr_len);
912 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
913 }
914 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
915}
916
917static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
918{
919 struct virtnet_info *vi = netdev_priv(dev);
920 int qnum = skb_get_queue_mapping(skb);
921 struct send_queue *sq = &vi->sq[qnum];
922 int err;
923 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
924 bool kick = !skb->xmit_more;
925
926 /* Free up any pending old buffers before queueing new ones. */
927 free_old_xmit_skbs(sq);
928
929 /* timestamp packet in software */
930 skb_tx_timestamp(skb);
931
932 /* Try to transmit */
933 err = xmit_skb(sq, skb);
934
935 /* This should not happen! */
936 if (unlikely(err)) {
937 dev->stats.tx_fifo_errors++;
938 if (net_ratelimit())
939 dev_warn(&dev->dev,
940 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
941 dev->stats.tx_dropped++;
942 dev_kfree_skb_any(skb);
943 return NETDEV_TX_OK;
944 }
945
946 /* Don't wait up for transmitted skbs to be freed. */
947 skb_orphan(skb);
948 nf_reset(skb);
949
950 /* If running out of space, stop queue to avoid getting packets that we
951 * are then unable to transmit.
952 * An alternative would be to force queuing layer to requeue the skb by
953 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
954 * returned in a normal path of operation: it means that driver is not
955 * maintaining the TX queue stop/start state properly, and causes
956 * the stack to do a non-trivial amount of useless work.
957 * Since most packets only take 1 or 2 ring slots, stopping the queue
958 * early means 16 slots are typically wasted.
959 */
960 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
961 netif_stop_subqueue(dev, qnum);
962 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
963 /* More just got used, free them then recheck. */
964 free_old_xmit_skbs(sq);
965 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
966 netif_start_subqueue(dev, qnum);
967 virtqueue_disable_cb(sq->vq);
968 }
969 }
970 }
971
972 if (kick || netif_xmit_stopped(txq))
973 virtqueue_kick(sq->vq);
974
975 return NETDEV_TX_OK;
976}
977
978/*
979 * Send command via the control virtqueue and check status. Commands
980 * supported by the hypervisor, as indicated by feature bits, should
981 * never fail unless improperly formatted.
982 */
983static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
984 struct scatterlist *out)
985{
986 struct scatterlist *sgs[4], hdr, stat;
987 unsigned out_num = 0, tmp;
988
989 /* Caller should know better */
990 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
991
992 vi->ctrl_status = ~0;
993 vi->ctrl_hdr.class = class;
994 vi->ctrl_hdr.cmd = cmd;
995 /* Add header */
996 sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
997 sgs[out_num++] = &hdr;
998
999 if (out)
1000 sgs[out_num++] = out;
1001
1002 /* Add return status. */
1003 sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1004 sgs[out_num] = &stat;
1005
1006 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1007 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1008
1009 if (unlikely(!virtqueue_kick(vi->cvq)))
1010 return vi->ctrl_status == VIRTIO_NET_OK;
1011
1012 /* Spin for a response, the kick causes an ioport write, trapping
1013 * into the hypervisor, so the request should be handled immediately.
1014 */
1015 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1016 !virtqueue_is_broken(vi->cvq))
1017 cpu_relax();
1018
1019 return vi->ctrl_status == VIRTIO_NET_OK;
1020}
1021
1022static int virtnet_set_mac_address(struct net_device *dev, void *p)
1023{
1024 struct virtnet_info *vi = netdev_priv(dev);
1025 struct virtio_device *vdev = vi->vdev;
1026 int ret;
1027 struct sockaddr *addr = p;
1028 struct scatterlist sg;
1029
1030 ret = eth_prepare_mac_addr_change(dev, p);
1031 if (ret)
1032 return ret;
1033
1034 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1035 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1036 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1037 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1038 dev_warn(&vdev->dev,
1039 "Failed to set mac address by vq command.\n");
1040 return -EINVAL;
1041 }
1042 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1043 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1044 unsigned int i;
1045
1046 /* Naturally, this has an atomicity problem. */
1047 for (i = 0; i < dev->addr_len; i++)
1048 virtio_cwrite8(vdev,
1049 offsetof(struct virtio_net_config, mac) +
1050 i, addr->sa_data[i]);
1051 }
1052
1053 eth_commit_mac_addr_change(dev, p);
1054
1055 return 0;
1056}
1057
1058static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1059 struct rtnl_link_stats64 *tot)
1060{
1061 struct virtnet_info *vi = netdev_priv(dev);
1062 int cpu;
1063 unsigned int start;
1064
1065 for_each_possible_cpu(cpu) {
1066 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1067 u64 tpackets, tbytes, rpackets, rbytes;
1068
1069 do {
1070 start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1071 tpackets = stats->tx_packets;
1072 tbytes = stats->tx_bytes;
1073 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1074
1075 do {
1076 start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1077 rpackets = stats->rx_packets;
1078 rbytes = stats->rx_bytes;
1079 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1080
1081 tot->rx_packets += rpackets;
1082 tot->tx_packets += tpackets;
1083 tot->rx_bytes += rbytes;
1084 tot->tx_bytes += tbytes;
1085 }
1086
1087 tot->tx_dropped = dev->stats.tx_dropped;
1088 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1089 tot->rx_dropped = dev->stats.rx_dropped;
1090 tot->rx_length_errors = dev->stats.rx_length_errors;
1091 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1092
1093 return tot;
1094}
1095
1096#ifdef CONFIG_NET_POLL_CONTROLLER
1097static void virtnet_netpoll(struct net_device *dev)
1098{
1099 struct virtnet_info *vi = netdev_priv(dev);
1100 int i;
1101
1102 for (i = 0; i < vi->curr_queue_pairs; i++)
1103 napi_schedule(&vi->rq[i].napi);
1104}
1105#endif
1106
1107static void virtnet_ack_link_announce(struct virtnet_info *vi)
1108{
1109 rtnl_lock();
1110 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1111 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1112 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1113 rtnl_unlock();
1114}
1115
1116static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1117{
1118 struct scatterlist sg;
1119 struct virtio_net_ctrl_mq s;
1120 struct net_device *dev = vi->dev;
1121
1122 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1123 return 0;
1124
1125 s.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1126 sg_init_one(&sg, &s, sizeof(s));
1127
1128 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1129 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1130 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1131 queue_pairs);
1132 return -EINVAL;
1133 } else {
1134 vi->curr_queue_pairs = queue_pairs;
1135 /* virtnet_open() will refill when device is going to up. */
1136 if (dev->flags & IFF_UP)
1137 schedule_delayed_work(&vi->refill, 0);
1138 }
1139
1140 return 0;
1141}
1142
1143static int virtnet_close(struct net_device *dev)
1144{
1145 struct virtnet_info *vi = netdev_priv(dev);
1146 int i;
1147
1148 /* Make sure refill_work doesn't re-enable napi! */
1149 cancel_delayed_work_sync(&vi->refill);
1150
1151 for (i = 0; i < vi->max_queue_pairs; i++)
1152 napi_disable(&vi->rq[i].napi);
1153
1154 return 0;
1155}
1156
1157static void virtnet_set_rx_mode(struct net_device *dev)
1158{
1159 struct virtnet_info *vi = netdev_priv(dev);
1160 struct scatterlist sg[2];
1161 struct virtio_net_ctrl_mac *mac_data;
1162 struct netdev_hw_addr *ha;
1163 int uc_count;
1164 int mc_count;
1165 void *buf;
1166 int i;
1167
1168 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1169 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1170 return;
1171
1172 vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1173 vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1174
1175 sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1176
1177 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1178 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1179 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1180 vi->ctrl_promisc ? "en" : "dis");
1181
1182 sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1183
1184 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1185 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1186 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1187 vi->ctrl_allmulti ? "en" : "dis");
1188
1189 uc_count = netdev_uc_count(dev);
1190 mc_count = netdev_mc_count(dev);
1191 /* MAC filter - use one buffer for both lists */
1192 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1193 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1194 mac_data = buf;
1195 if (!buf)
1196 return;
1197
1198 sg_init_table(sg, 2);
1199
1200 /* Store the unicast list and count in the front of the buffer */
1201 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1202 i = 0;
1203 netdev_for_each_uc_addr(ha, dev)
1204 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1205
1206 sg_set_buf(&sg[0], mac_data,
1207 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1208
1209 /* multicast list and count fill the end */
1210 mac_data = (void *)&mac_data->macs[uc_count][0];
1211
1212 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1213 i = 0;
1214 netdev_for_each_mc_addr(ha, dev)
1215 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1216
1217 sg_set_buf(&sg[1], mac_data,
1218 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1219
1220 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1221 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1222 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1223
1224 kfree(buf);
1225}
1226
1227static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1228 __be16 proto, u16 vid)
1229{
1230 struct virtnet_info *vi = netdev_priv(dev);
1231 struct scatterlist sg;
1232
1233 sg_init_one(&sg, &vid, sizeof(vid));
1234
1235 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1236 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1237 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1238 return 0;
1239}
1240
1241static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1242 __be16 proto, u16 vid)
1243{
1244 struct virtnet_info *vi = netdev_priv(dev);
1245 struct scatterlist sg;
1246
1247 sg_init_one(&sg, &vid, sizeof(vid));
1248
1249 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1250 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1251 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1252 return 0;
1253}
1254
1255static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1256{
1257 int i;
1258
1259 if (vi->affinity_hint_set) {
1260 for (i = 0; i < vi->max_queue_pairs; i++) {
1261 virtqueue_set_affinity(vi->rq[i].vq, -1);
1262 virtqueue_set_affinity(vi->sq[i].vq, -1);
1263 }
1264
1265 vi->affinity_hint_set = false;
1266 }
1267}
1268
1269static void virtnet_set_affinity(struct virtnet_info *vi)
1270{
1271 int i;
1272 int cpu;
1273
1274 /* In multiqueue mode, when the number of cpu is equal to the number of
1275 * queue pairs, we let the queue pairs to be private to one cpu by
1276 * setting the affinity hint to eliminate the contention.
1277 */
1278 if (vi->curr_queue_pairs == 1 ||
1279 vi->max_queue_pairs != num_online_cpus()) {
1280 virtnet_clean_affinity(vi, -1);
1281 return;
1282 }
1283
1284 i = 0;
1285 for_each_online_cpu(cpu) {
1286 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1287 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1288 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1289 i++;
1290 }
1291
1292 vi->affinity_hint_set = true;
1293}
1294
1295static int virtnet_cpu_callback(struct notifier_block *nfb,
1296 unsigned long action, void *hcpu)
1297{
1298 struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1299
1300 switch(action & ~CPU_TASKS_FROZEN) {
1301 case CPU_ONLINE:
1302 case CPU_DOWN_FAILED:
1303 case CPU_DEAD:
1304 virtnet_set_affinity(vi);
1305 break;
1306 case CPU_DOWN_PREPARE:
1307 virtnet_clean_affinity(vi, (long)hcpu);
1308 break;
1309 default:
1310 break;
1311 }
1312
1313 return NOTIFY_OK;
1314}
1315
1316static void virtnet_get_ringparam(struct net_device *dev,
1317 struct ethtool_ringparam *ring)
1318{
1319 struct virtnet_info *vi = netdev_priv(dev);
1320
1321 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1322 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1323 ring->rx_pending = ring->rx_max_pending;
1324 ring->tx_pending = ring->tx_max_pending;
1325}
1326
1327
1328static void virtnet_get_drvinfo(struct net_device *dev,
1329 struct ethtool_drvinfo *info)
1330{
1331 struct virtnet_info *vi = netdev_priv(dev);
1332 struct virtio_device *vdev = vi->vdev;
1333
1334 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1335 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1336 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1337
1338}
1339
1340/* TODO: Eliminate OOO packets during switching */
1341static int virtnet_set_channels(struct net_device *dev,
1342 struct ethtool_channels *channels)
1343{
1344 struct virtnet_info *vi = netdev_priv(dev);
1345 u16 queue_pairs = channels->combined_count;
1346 int err;
1347
1348 /* We don't support separate rx/tx channels.
1349 * We don't allow setting 'other' channels.
1350 */
1351 if (channels->rx_count || channels->tx_count || channels->other_count)
1352 return -EINVAL;
1353
1354 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1355 return -EINVAL;
1356
1357 get_online_cpus();
1358 err = virtnet_set_queues(vi, queue_pairs);
1359 if (!err) {
1360 netif_set_real_num_tx_queues(dev, queue_pairs);
1361 netif_set_real_num_rx_queues(dev, queue_pairs);
1362
1363 virtnet_set_affinity(vi);
1364 }
1365 put_online_cpus();
1366
1367 return err;
1368}
1369
1370static void virtnet_get_channels(struct net_device *dev,
1371 struct ethtool_channels *channels)
1372{
1373 struct virtnet_info *vi = netdev_priv(dev);
1374
1375 channels->combined_count = vi->curr_queue_pairs;
1376 channels->max_combined = vi->max_queue_pairs;
1377 channels->max_other = 0;
1378 channels->rx_count = 0;
1379 channels->tx_count = 0;
1380 channels->other_count = 0;
1381}
1382
1383/* Check if the user is trying to change anything besides speed/duplex */
1384static bool virtnet_validate_ethtool_cmd(const struct ethtool_cmd *cmd)
1385{
1386 struct ethtool_cmd diff1 = *cmd;
1387 struct ethtool_cmd diff2 = {};
1388
1389 /* cmd is always set so we need to clear it, validate the port type
1390 * and also without autonegotiation we can ignore advertising
1391 */
1392 ethtool_cmd_speed_set(&diff1, 0);
1393 diff2.port = PORT_OTHER;
1394 diff1.advertising = 0;
1395 diff1.duplex = 0;
1396 diff1.cmd = 0;
1397
1398 return !memcmp(&diff1, &diff2, sizeof(diff1));
1399}
1400
1401static int virtnet_set_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1402{
1403 struct virtnet_info *vi = netdev_priv(dev);
1404 u32 speed;
1405
1406 speed = ethtool_cmd_speed(cmd);
1407 /* don't allow custom speed and duplex */
1408 if (!ethtool_validate_speed(speed) ||
1409 !ethtool_validate_duplex(cmd->duplex) ||
1410 !virtnet_validate_ethtool_cmd(cmd))
1411 return -EINVAL;
1412 vi->speed = speed;
1413 vi->duplex = cmd->duplex;
1414
1415 return 0;
1416}
1417
1418static int virtnet_get_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1419{
1420 struct virtnet_info *vi = netdev_priv(dev);
1421
1422 ethtool_cmd_speed_set(cmd, vi->speed);
1423 cmd->duplex = vi->duplex;
1424 cmd->port = PORT_OTHER;
1425
1426 return 0;
1427}
1428
1429static void virtnet_init_settings(struct net_device *dev)
1430{
1431 struct virtnet_info *vi = netdev_priv(dev);
1432
1433 vi->speed = SPEED_UNKNOWN;
1434 vi->duplex = DUPLEX_UNKNOWN;
1435}
1436
1437static const struct ethtool_ops virtnet_ethtool_ops = {
1438 .get_drvinfo = virtnet_get_drvinfo,
1439 .get_link = ethtool_op_get_link,
1440 .get_ringparam = virtnet_get_ringparam,
1441 .set_channels = virtnet_set_channels,
1442 .get_channels = virtnet_get_channels,
1443 .get_ts_info = ethtool_op_get_ts_info,
1444 .get_settings = virtnet_get_settings,
1445 .set_settings = virtnet_set_settings,
1446};
1447
1448#define MIN_MTU 68
1449#define MAX_MTU 65535
1450
1451static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1452{
1453 if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1454 return -EINVAL;
1455 dev->mtu = new_mtu;
1456 return 0;
1457}
1458
1459static const struct net_device_ops virtnet_netdev = {
1460 .ndo_open = virtnet_open,
1461 .ndo_stop = virtnet_close,
1462 .ndo_start_xmit = start_xmit,
1463 .ndo_validate_addr = eth_validate_addr,
1464 .ndo_set_mac_address = virtnet_set_mac_address,
1465 .ndo_set_rx_mode = virtnet_set_rx_mode,
1466 .ndo_change_mtu = virtnet_change_mtu,
1467 .ndo_get_stats64 = virtnet_stats,
1468 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1469 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1470#ifdef CONFIG_NET_POLL_CONTROLLER
1471 .ndo_poll_controller = virtnet_netpoll,
1472#endif
1473#ifdef CONFIG_NET_RX_BUSY_POLL
1474 .ndo_busy_poll = virtnet_busy_poll,
1475#endif
1476};
1477
1478static void virtnet_config_changed_work(struct work_struct *work)
1479{
1480 struct virtnet_info *vi =
1481 container_of(work, struct virtnet_info, config_work);
1482 u16 v;
1483
1484 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1485 struct virtio_net_config, status, &v) < 0)
1486 return;
1487
1488 if (v & VIRTIO_NET_S_ANNOUNCE) {
1489 netdev_notify_peers(vi->dev);
1490 virtnet_ack_link_announce(vi);
1491 }
1492
1493 /* Ignore unknown (future) status bits */
1494 v &= VIRTIO_NET_S_LINK_UP;
1495
1496 if (vi->status == v)
1497 return;
1498
1499 vi->status = v;
1500
1501 if (vi->status & VIRTIO_NET_S_LINK_UP) {
1502 netif_carrier_on(vi->dev);
1503 netif_tx_wake_all_queues(vi->dev);
1504 } else {
1505 netif_carrier_off(vi->dev);
1506 netif_tx_stop_all_queues(vi->dev);
1507 }
1508}
1509
1510static void virtnet_config_changed(struct virtio_device *vdev)
1511{
1512 struct virtnet_info *vi = vdev->priv;
1513
1514 schedule_work(&vi->config_work);
1515}
1516
1517static void virtnet_free_queues(struct virtnet_info *vi)
1518{
1519 int i;
1520
1521 for (i = 0; i < vi->max_queue_pairs; i++) {
1522 napi_hash_del(&vi->rq[i].napi);
1523 netif_napi_del(&vi->rq[i].napi);
1524 }
1525
1526 kfree(vi->rq);
1527 kfree(vi->sq);
1528}
1529
1530static void free_receive_bufs(struct virtnet_info *vi)
1531{
1532 int i;
1533
1534 for (i = 0; i < vi->max_queue_pairs; i++) {
1535 while (vi->rq[i].pages)
1536 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1537 }
1538}
1539
1540static void free_receive_page_frags(struct virtnet_info *vi)
1541{
1542 int i;
1543 for (i = 0; i < vi->max_queue_pairs; i++)
1544 if (vi->rq[i].alloc_frag.page)
1545 put_page(vi->rq[i].alloc_frag.page);
1546}
1547
1548static void free_unused_bufs(struct virtnet_info *vi)
1549{
1550 void *buf;
1551 int i;
1552
1553 for (i = 0; i < vi->max_queue_pairs; i++) {
1554 struct virtqueue *vq = vi->sq[i].vq;
1555 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1556 dev_kfree_skb(buf);
1557 }
1558
1559 for (i = 0; i < vi->max_queue_pairs; i++) {
1560 struct virtqueue *vq = vi->rq[i].vq;
1561
1562 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1563 if (vi->mergeable_rx_bufs) {
1564 unsigned long ctx = (unsigned long)buf;
1565 void *base = mergeable_ctx_to_buf_address(ctx);
1566 put_page(virt_to_head_page(base));
1567 } else if (vi->big_packets) {
1568 give_pages(&vi->rq[i], buf);
1569 } else {
1570 dev_kfree_skb(buf);
1571 }
1572 }
1573 }
1574}
1575
1576static void virtnet_del_vqs(struct virtnet_info *vi)
1577{
1578 struct virtio_device *vdev = vi->vdev;
1579
1580 virtnet_clean_affinity(vi, -1);
1581
1582 vdev->config->del_vqs(vdev);
1583
1584 virtnet_free_queues(vi);
1585}
1586
1587static int virtnet_find_vqs(struct virtnet_info *vi)
1588{
1589 vq_callback_t **callbacks;
1590 struct virtqueue **vqs;
1591 int ret = -ENOMEM;
1592 int i, total_vqs;
1593 const char **names;
1594
1595 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1596 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1597 * possible control vq.
1598 */
1599 total_vqs = vi->max_queue_pairs * 2 +
1600 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1601
1602 /* Allocate space for find_vqs parameters */
1603 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1604 if (!vqs)
1605 goto err_vq;
1606 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1607 if (!callbacks)
1608 goto err_callback;
1609 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1610 if (!names)
1611 goto err_names;
1612
1613 /* Parameters for control virtqueue, if any */
1614 if (vi->has_cvq) {
1615 callbacks[total_vqs - 1] = NULL;
1616 names[total_vqs - 1] = "control";
1617 }
1618
1619 /* Allocate/initialize parameters for send/receive virtqueues */
1620 for (i = 0; i < vi->max_queue_pairs; i++) {
1621 callbacks[rxq2vq(i)] = skb_recv_done;
1622 callbacks[txq2vq(i)] = skb_xmit_done;
1623 sprintf(vi->rq[i].name, "input.%d", i);
1624 sprintf(vi->sq[i].name, "output.%d", i);
1625 names[rxq2vq(i)] = vi->rq[i].name;
1626 names[txq2vq(i)] = vi->sq[i].name;
1627 }
1628
1629 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1630 names);
1631 if (ret)
1632 goto err_find;
1633
1634 if (vi->has_cvq) {
1635 vi->cvq = vqs[total_vqs - 1];
1636 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1637 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1638 }
1639
1640 for (i = 0; i < vi->max_queue_pairs; i++) {
1641 vi->rq[i].vq = vqs[rxq2vq(i)];
1642 vi->sq[i].vq = vqs[txq2vq(i)];
1643 }
1644
1645 kfree(names);
1646 kfree(callbacks);
1647 kfree(vqs);
1648
1649 return 0;
1650
1651err_find:
1652 kfree(names);
1653err_names:
1654 kfree(callbacks);
1655err_callback:
1656 kfree(vqs);
1657err_vq:
1658 return ret;
1659}
1660
1661static int virtnet_alloc_queues(struct virtnet_info *vi)
1662{
1663 int i;
1664
1665 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1666 if (!vi->sq)
1667 goto err_sq;
1668 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1669 if (!vi->rq)
1670 goto err_rq;
1671
1672 INIT_DELAYED_WORK(&vi->refill, refill_work);
1673 for (i = 0; i < vi->max_queue_pairs; i++) {
1674 vi->rq[i].pages = NULL;
1675 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1676 napi_weight);
1677
1678 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1679 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
1680 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1681 }
1682
1683 return 0;
1684
1685err_rq:
1686 kfree(vi->sq);
1687err_sq:
1688 return -ENOMEM;
1689}
1690
1691static int init_vqs(struct virtnet_info *vi)
1692{
1693 int ret;
1694
1695 /* Allocate send & receive queues */
1696 ret = virtnet_alloc_queues(vi);
1697 if (ret)
1698 goto err;
1699
1700 ret = virtnet_find_vqs(vi);
1701 if (ret)
1702 goto err_free;
1703
1704 get_online_cpus();
1705 virtnet_set_affinity(vi);
1706 put_online_cpus();
1707
1708 return 0;
1709
1710err_free:
1711 virtnet_free_queues(vi);
1712err:
1713 return ret;
1714}
1715
1716#ifdef CONFIG_SYSFS
1717static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1718 struct rx_queue_attribute *attribute, char *buf)
1719{
1720 struct virtnet_info *vi = netdev_priv(queue->dev);
1721 unsigned int queue_index = get_netdev_rx_queue_index(queue);
1722 struct ewma_pkt_len *avg;
1723
1724 BUG_ON(queue_index >= vi->max_queue_pairs);
1725 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1726 return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1727}
1728
1729static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1730 __ATTR_RO(mergeable_rx_buffer_size);
1731
1732static struct attribute *virtio_net_mrg_rx_attrs[] = {
1733 &mergeable_rx_buffer_size_attribute.attr,
1734 NULL
1735};
1736
1737static const struct attribute_group virtio_net_mrg_rx_group = {
1738 .name = "virtio_net",
1739 .attrs = virtio_net_mrg_rx_attrs
1740};
1741#endif
1742
1743static bool virtnet_fail_on_feature(struct virtio_device *vdev,
1744 unsigned int fbit,
1745 const char *fname, const char *dname)
1746{
1747 if (!virtio_has_feature(vdev, fbit))
1748 return false;
1749
1750 dev_err(&vdev->dev, "device advertises feature %s but not %s",
1751 fname, dname);
1752
1753 return true;
1754}
1755
1756#define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
1757 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
1758
1759static bool virtnet_validate_features(struct virtio_device *vdev)
1760{
1761 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
1762 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
1763 "VIRTIO_NET_F_CTRL_VQ") ||
1764 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
1765 "VIRTIO_NET_F_CTRL_VQ") ||
1766 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
1767 "VIRTIO_NET_F_CTRL_VQ") ||
1768 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
1769 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
1770 "VIRTIO_NET_F_CTRL_VQ"))) {
1771 return false;
1772 }
1773
1774 return true;
1775}
1776
1777static int virtnet_probe(struct virtio_device *vdev)
1778{
1779 int i, err;
1780 struct net_device *dev;
1781 struct virtnet_info *vi;
1782 u16 max_queue_pairs;
1783
1784 if (!vdev->config->get) {
1785 dev_err(&vdev->dev, "%s failure: config access disabled\n",
1786 __func__);
1787 return -EINVAL;
1788 }
1789
1790 if (!virtnet_validate_features(vdev))
1791 return -EINVAL;
1792
1793 /* Find if host supports multiqueue virtio_net device */
1794 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1795 struct virtio_net_config,
1796 max_virtqueue_pairs, &max_queue_pairs);
1797
1798 /* We need at least 2 queue's */
1799 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1800 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1801 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1802 max_queue_pairs = 1;
1803
1804 /* Allocate ourselves a network device with room for our info */
1805 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1806 if (!dev)
1807 return -ENOMEM;
1808
1809 /* Set up network device as normal. */
1810 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1811 dev->netdev_ops = &virtnet_netdev;
1812 dev->features = NETIF_F_HIGHDMA;
1813
1814 dev->ethtool_ops = &virtnet_ethtool_ops;
1815 SET_NETDEV_DEV(dev, &vdev->dev);
1816
1817 /* Do we support "hardware" checksums? */
1818 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1819 /* This opens up the world of extra features. */
1820 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
1821 if (csum)
1822 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
1823
1824 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1825 dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1826 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1827 }
1828 /* Individual feature bits: what can host handle? */
1829 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1830 dev->hw_features |= NETIF_F_TSO;
1831 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1832 dev->hw_features |= NETIF_F_TSO6;
1833 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1834 dev->hw_features |= NETIF_F_TSO_ECN;
1835 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1836 dev->hw_features |= NETIF_F_UFO;
1837
1838 dev->features |= NETIF_F_GSO_ROBUST;
1839
1840 if (gso)
1841 dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1842 /* (!csum && gso) case will be fixed by register_netdev() */
1843 }
1844 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1845 dev->features |= NETIF_F_RXCSUM;
1846
1847 dev->vlan_features = dev->features;
1848
1849 /* Configuration may specify what MAC to use. Otherwise random. */
1850 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1851 virtio_cread_bytes(vdev,
1852 offsetof(struct virtio_net_config, mac),
1853 dev->dev_addr, dev->addr_len);
1854 else
1855 eth_hw_addr_random(dev);
1856
1857 /* Set up our device-specific information */
1858 vi = netdev_priv(dev);
1859 vi->dev = dev;
1860 vi->vdev = vdev;
1861 vdev->priv = vi;
1862 vi->stats = alloc_percpu(struct virtnet_stats);
1863 err = -ENOMEM;
1864 if (vi->stats == NULL)
1865 goto free;
1866
1867 for_each_possible_cpu(i) {
1868 struct virtnet_stats *virtnet_stats;
1869 virtnet_stats = per_cpu_ptr(vi->stats, i);
1870 u64_stats_init(&virtnet_stats->tx_syncp);
1871 u64_stats_init(&virtnet_stats->rx_syncp);
1872 }
1873
1874 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1875
1876 /* If we can receive ANY GSO packets, we must allocate large ones. */
1877 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1878 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1879 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
1880 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
1881 vi->big_packets = true;
1882
1883 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1884 vi->mergeable_rx_bufs = true;
1885
1886 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
1887 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1888 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1889 else
1890 vi->hdr_len = sizeof(struct virtio_net_hdr);
1891
1892 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
1893 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1894 vi->any_header_sg = true;
1895
1896 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1897 vi->has_cvq = true;
1898
1899 if (vi->any_header_sg)
1900 dev->needed_headroom = vi->hdr_len;
1901
1902 /* Use single tx/rx queue pair as default */
1903 vi->curr_queue_pairs = 1;
1904 vi->max_queue_pairs = max_queue_pairs;
1905
1906 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1907 err = init_vqs(vi);
1908 if (err)
1909 goto free_stats;
1910
1911#ifdef CONFIG_SYSFS
1912 if (vi->mergeable_rx_bufs)
1913 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
1914#endif
1915 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1916 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1917
1918 virtnet_init_settings(dev);
1919
1920 err = register_netdev(dev);
1921 if (err) {
1922 pr_debug("virtio_net: registering device failed\n");
1923 goto free_vqs;
1924 }
1925
1926 virtio_device_ready(vdev);
1927
1928 /* Last of all, set up some receive buffers. */
1929 for (i = 0; i < vi->curr_queue_pairs; i++) {
1930 try_fill_recv(vi, &vi->rq[i], GFP_KERNEL);
1931
1932 /* If we didn't even get one input buffer, we're useless. */
1933 if (vi->rq[i].vq->num_free ==
1934 virtqueue_get_vring_size(vi->rq[i].vq)) {
1935 free_unused_bufs(vi);
1936 err = -ENOMEM;
1937 goto free_recv_bufs;
1938 }
1939 }
1940
1941 vi->nb.notifier_call = &virtnet_cpu_callback;
1942 err = register_hotcpu_notifier(&vi->nb);
1943 if (err) {
1944 pr_debug("virtio_net: registering cpu notifier failed\n");
1945 goto free_recv_bufs;
1946 }
1947
1948 /* Assume link up if device can't report link status,
1949 otherwise get link status from config. */
1950 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1951 netif_carrier_off(dev);
1952 schedule_work(&vi->config_work);
1953 } else {
1954 vi->status = VIRTIO_NET_S_LINK_UP;
1955 netif_carrier_on(dev);
1956 }
1957
1958 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1959 dev->name, max_queue_pairs);
1960
1961 return 0;
1962
1963free_recv_bufs:
1964 vi->vdev->config->reset(vdev);
1965
1966 free_receive_bufs(vi);
1967 unregister_netdev(dev);
1968free_vqs:
1969 cancel_delayed_work_sync(&vi->refill);
1970 free_receive_page_frags(vi);
1971 virtnet_del_vqs(vi);
1972free_stats:
1973 free_percpu(vi->stats);
1974free:
1975 free_netdev(dev);
1976 return err;
1977}
1978
1979static void remove_vq_common(struct virtnet_info *vi)
1980{
1981 vi->vdev->config->reset(vi->vdev);
1982
1983 /* Free unused buffers in both send and recv, if any. */
1984 free_unused_bufs(vi);
1985
1986 free_receive_bufs(vi);
1987
1988 free_receive_page_frags(vi);
1989
1990 virtnet_del_vqs(vi);
1991}
1992
1993static void virtnet_remove(struct virtio_device *vdev)
1994{
1995 struct virtnet_info *vi = vdev->priv;
1996
1997 unregister_hotcpu_notifier(&vi->nb);
1998
1999 /* Make sure no work handler is accessing the device. */
2000 flush_work(&vi->config_work);
2001
2002 unregister_netdev(vi->dev);
2003
2004 remove_vq_common(vi);
2005
2006 free_percpu(vi->stats);
2007 free_netdev(vi->dev);
2008}
2009
2010#ifdef CONFIG_PM_SLEEP
2011static int virtnet_freeze(struct virtio_device *vdev)
2012{
2013 struct virtnet_info *vi = vdev->priv;
2014 int i;
2015
2016 unregister_hotcpu_notifier(&vi->nb);
2017
2018 /* Make sure no work handler is accessing the device */
2019 flush_work(&vi->config_work);
2020
2021 netif_device_detach(vi->dev);
2022 cancel_delayed_work_sync(&vi->refill);
2023
2024 if (netif_running(vi->dev)) {
2025 for (i = 0; i < vi->max_queue_pairs; i++)
2026 napi_disable(&vi->rq[i].napi);
2027 }
2028
2029 remove_vq_common(vi);
2030
2031 return 0;
2032}
2033
2034static int virtnet_restore(struct virtio_device *vdev)
2035{
2036 struct virtnet_info *vi = vdev->priv;
2037 int err, i;
2038
2039 err = init_vqs(vi);
2040 if (err)
2041 return err;
2042
2043 virtio_device_ready(vdev);
2044
2045 if (netif_running(vi->dev)) {
2046 for (i = 0; i < vi->curr_queue_pairs; i++)
2047 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2048 schedule_delayed_work(&vi->refill, 0);
2049
2050 for (i = 0; i < vi->max_queue_pairs; i++)
2051 virtnet_napi_enable(&vi->rq[i]);
2052 }
2053
2054 netif_device_attach(vi->dev);
2055
2056 rtnl_lock();
2057 virtnet_set_queues(vi, vi->curr_queue_pairs);
2058 rtnl_unlock();
2059
2060 err = register_hotcpu_notifier(&vi->nb);
2061 if (err)
2062 return err;
2063
2064 return 0;
2065}
2066#endif
2067
2068static struct virtio_device_id id_table[] = {
2069 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2070 { 0 },
2071};
2072
2073static unsigned int features[] = {
2074 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
2075 VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
2076 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
2077 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
2078 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
2079 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
2080 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
2081 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
2082 VIRTIO_NET_F_CTRL_MAC_ADDR,
2083 VIRTIO_F_ANY_LAYOUT,
2084};
2085
2086static struct virtio_driver virtio_net_driver = {
2087 .feature_table = features,
2088 .feature_table_size = ARRAY_SIZE(features),
2089 .driver.name = KBUILD_MODNAME,
2090 .driver.owner = THIS_MODULE,
2091 .id_table = id_table,
2092 .probe = virtnet_probe,
2093 .remove = virtnet_remove,
2094 .config_changed = virtnet_config_changed,
2095#ifdef CONFIG_PM_SLEEP
2096 .freeze = virtnet_freeze,
2097 .restore = virtnet_restore,
2098#endif
2099};
2100
2101module_virtio_driver(virtio_net_driver);
2102
2103MODULE_DEVICE_TABLE(virtio, id_table);
2104MODULE_DESCRIPTION("Virtio network driver");
2105MODULE_LICENSE("GPL");