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/bpf.h>
26#include <linux/scatterlist.h>
27#include <linux/if_vlan.h>
28#include <linux/slab.h>
29#include <linux/cpu.h>
30#include <linux/average.h>
31#include <net/busy_poll.h>
32
33static int napi_weight = NAPI_POLL_WEIGHT;
34module_param(napi_weight, int, 0444);
35
36static bool csum = true, gso = true;
37module_param(csum, bool, 0444);
38module_param(gso, bool, 0444);
39
40/* FIXME: MTU in config. */
41#define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
42#define GOOD_COPY_LEN 128
43
44/* RX packet size EWMA. The average packet size is used to determine the packet
45 * buffer size when refilling RX rings. As the entire RX ring may be refilled
46 * at once, the weight is chosen so that the EWMA will be insensitive to short-
47 * term, transient changes in packet size.
48 */
49DECLARE_EWMA(pkt_len, 1, 64)
50
51/* With mergeable buffers we align buffer address and use the low bits to
52 * encode its true size. Buffer size is up to 1 page so we need to align to
53 * square root of page size to ensure we reserve enough bits to encode the true
54 * size.
55 */
56#define MERGEABLE_BUFFER_MIN_ALIGN_SHIFT ((PAGE_SHIFT + 1) / 2)
57
58/* Minimum alignment for mergeable packet buffers. */
59#define MERGEABLE_BUFFER_ALIGN max(L1_CACHE_BYTES, \
60 1 << MERGEABLE_BUFFER_MIN_ALIGN_SHIFT)
61
62#define VIRTNET_DRIVER_VERSION "1.0.0"
63
64struct virtnet_stats {
65 struct u64_stats_sync tx_syncp;
66 struct u64_stats_sync rx_syncp;
67 u64 tx_bytes;
68 u64 tx_packets;
69
70 u64 rx_bytes;
71 u64 rx_packets;
72};
73
74/* Internal representation of a send virtqueue */
75struct send_queue {
76 /* Virtqueue associated with this send _queue */
77 struct virtqueue *vq;
78
79 /* TX: fragments + linear part + virtio header */
80 struct scatterlist sg[MAX_SKB_FRAGS + 2];
81
82 /* Name of the send queue: output.$index */
83 char name[40];
84};
85
86/* Internal representation of a receive virtqueue */
87struct receive_queue {
88 /* Virtqueue associated with this receive_queue */
89 struct virtqueue *vq;
90
91 struct napi_struct napi;
92
93 struct bpf_prog __rcu *xdp_prog;
94
95 /* Chain pages by the private ptr. */
96 struct page *pages;
97
98 /* Average packet length for mergeable receive buffers. */
99 struct ewma_pkt_len mrg_avg_pkt_len;
100
101 /* Page frag for packet buffer allocation. */
102 struct page_frag alloc_frag;
103
104 /* RX: fragments + linear part + virtio header */
105 struct scatterlist sg[MAX_SKB_FRAGS + 2];
106
107 /* Name of this receive queue: input.$index */
108 char name[40];
109};
110
111struct virtnet_info {
112 struct virtio_device *vdev;
113 struct virtqueue *cvq;
114 struct net_device *dev;
115 struct send_queue *sq;
116 struct receive_queue *rq;
117 unsigned int status;
118
119 /* Max # of queue pairs supported by the device */
120 u16 max_queue_pairs;
121
122 /* # of queue pairs currently used by the driver */
123 u16 curr_queue_pairs;
124
125 /* # of XDP queue pairs currently used by the driver */
126 u16 xdp_queue_pairs;
127
128 /* I like... big packets and I cannot lie! */
129 bool big_packets;
130
131 /* Host will merge rx buffers for big packets (shake it! shake it!) */
132 bool mergeable_rx_bufs;
133
134 /* Has control virtqueue */
135 bool has_cvq;
136
137 /* Host can handle any s/g split between our header and packet data */
138 bool any_header_sg;
139
140 /* Packet virtio header size */
141 u8 hdr_len;
142
143 /* Active statistics */
144 struct virtnet_stats __percpu *stats;
145
146 /* Work struct for refilling if we run low on memory. */
147 struct delayed_work refill;
148
149 /* Work struct for config space updates */
150 struct work_struct config_work;
151
152 /* Does the affinity hint is set for virtqueues? */
153 bool affinity_hint_set;
154
155 /* CPU hotplug instances for online & dead */
156 struct hlist_node node;
157 struct hlist_node node_dead;
158
159 /* Control VQ buffers: protected by the rtnl lock */
160 struct virtio_net_ctrl_hdr ctrl_hdr;
161 virtio_net_ctrl_ack ctrl_status;
162 struct virtio_net_ctrl_mq ctrl_mq;
163 u8 ctrl_promisc;
164 u8 ctrl_allmulti;
165 u16 ctrl_vid;
166
167 /* Ethtool settings */
168 u8 duplex;
169 u32 speed;
170};
171
172struct padded_vnet_hdr {
173 struct virtio_net_hdr_mrg_rxbuf hdr;
174 /*
175 * hdr is in a separate sg buffer, and data sg buffer shares same page
176 * with this header sg. This padding makes next sg 16 byte aligned
177 * after the header.
178 */
179 char padding[4];
180};
181
182/* Converting between virtqueue no. and kernel tx/rx queue no.
183 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
184 */
185static int vq2txq(struct virtqueue *vq)
186{
187 return (vq->index - 1) / 2;
188}
189
190static int txq2vq(int txq)
191{
192 return txq * 2 + 1;
193}
194
195static int vq2rxq(struct virtqueue *vq)
196{
197 return vq->index / 2;
198}
199
200static int rxq2vq(int rxq)
201{
202 return rxq * 2;
203}
204
205static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
206{
207 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
208}
209
210/*
211 * private is used to chain pages for big packets, put the whole
212 * most recent used list in the beginning for reuse
213 */
214static void give_pages(struct receive_queue *rq, struct page *page)
215{
216 struct page *end;
217
218 /* Find end of list, sew whole thing into vi->rq.pages. */
219 for (end = page; end->private; end = (struct page *)end->private);
220 end->private = (unsigned long)rq->pages;
221 rq->pages = page;
222}
223
224static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
225{
226 struct page *p = rq->pages;
227
228 if (p) {
229 rq->pages = (struct page *)p->private;
230 /* clear private here, it is used to chain pages */
231 p->private = 0;
232 } else
233 p = alloc_page(gfp_mask);
234 return p;
235}
236
237static void skb_xmit_done(struct virtqueue *vq)
238{
239 struct virtnet_info *vi = vq->vdev->priv;
240
241 /* Suppress further interrupts. */
242 virtqueue_disable_cb(vq);
243
244 /* We were probably waiting for more output buffers. */
245 netif_wake_subqueue(vi->dev, vq2txq(vq));
246}
247
248static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
249{
250 unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
251 return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
252}
253
254static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
255{
256 return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
257
258}
259
260static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
261{
262 unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
263 return (unsigned long)buf | (size - 1);
264}
265
266/* Called from bottom half context */
267static struct sk_buff *page_to_skb(struct virtnet_info *vi,
268 struct receive_queue *rq,
269 struct page *page, unsigned int offset,
270 unsigned int len, unsigned int truesize)
271{
272 struct sk_buff *skb;
273 struct virtio_net_hdr_mrg_rxbuf *hdr;
274 unsigned int copy, hdr_len, hdr_padded_len;
275 char *p;
276
277 p = page_address(page) + offset;
278
279 /* copy small packet so we can reuse these pages for small data */
280 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
281 if (unlikely(!skb))
282 return NULL;
283
284 hdr = skb_vnet_hdr(skb);
285
286 hdr_len = vi->hdr_len;
287 if (vi->mergeable_rx_bufs)
288 hdr_padded_len = sizeof *hdr;
289 else
290 hdr_padded_len = sizeof(struct padded_vnet_hdr);
291
292 memcpy(hdr, p, hdr_len);
293
294 len -= hdr_len;
295 offset += hdr_padded_len;
296 p += hdr_padded_len;
297
298 copy = len;
299 if (copy > skb_tailroom(skb))
300 copy = skb_tailroom(skb);
301 memcpy(skb_put(skb, copy), p, copy);
302
303 len -= copy;
304 offset += copy;
305
306 if (vi->mergeable_rx_bufs) {
307 if (len)
308 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
309 else
310 put_page(page);
311 return skb;
312 }
313
314 /*
315 * Verify that we can indeed put this data into a skb.
316 * This is here to handle cases when the device erroneously
317 * tries to receive more than is possible. This is usually
318 * the case of a broken device.
319 */
320 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
321 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
322 dev_kfree_skb(skb);
323 return NULL;
324 }
325 BUG_ON(offset >= PAGE_SIZE);
326 while (len) {
327 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
328 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
329 frag_size, truesize);
330 len -= frag_size;
331 page = (struct page *)page->private;
332 offset = 0;
333 }
334
335 if (page)
336 give_pages(rq, page);
337
338 return skb;
339}
340
341static void virtnet_xdp_xmit(struct virtnet_info *vi,
342 struct receive_queue *rq,
343 struct send_queue *sq,
344 struct xdp_buff *xdp,
345 void *data)
346{
347 struct virtio_net_hdr_mrg_rxbuf *hdr;
348 unsigned int num_sg, len;
349 void *xdp_sent;
350 int err;
351
352 /* Free up any pending old buffers before queueing new ones. */
353 while ((xdp_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) {
354 if (vi->mergeable_rx_bufs) {
355 struct page *sent_page = virt_to_head_page(xdp_sent);
356
357 put_page(sent_page);
358 } else { /* small buffer */
359 struct sk_buff *skb = xdp_sent;
360
361 kfree_skb(skb);
362 }
363 }
364
365 if (vi->mergeable_rx_bufs) {
366 /* Zero header and leave csum up to XDP layers */
367 hdr = xdp->data;
368 memset(hdr, 0, vi->hdr_len);
369
370 num_sg = 1;
371 sg_init_one(sq->sg, xdp->data, xdp->data_end - xdp->data);
372 } else { /* small buffer */
373 struct sk_buff *skb = data;
374
375 /* Zero header and leave csum up to XDP layers */
376 hdr = skb_vnet_hdr(skb);
377 memset(hdr, 0, vi->hdr_len);
378
379 num_sg = 2;
380 sg_init_table(sq->sg, 2);
381 sg_set_buf(sq->sg, hdr, vi->hdr_len);
382 skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
383 }
384 err = virtqueue_add_outbuf(sq->vq, sq->sg, num_sg,
385 data, GFP_ATOMIC);
386 if (unlikely(err)) {
387 if (vi->mergeable_rx_bufs) {
388 struct page *page = virt_to_head_page(xdp->data);
389
390 put_page(page);
391 } else /* small buffer */
392 kfree_skb(data);
393 return; // On error abort to avoid unnecessary kick
394 }
395
396 virtqueue_kick(sq->vq);
397}
398
399static u32 do_xdp_prog(struct virtnet_info *vi,
400 struct receive_queue *rq,
401 struct bpf_prog *xdp_prog,
402 void *data, int len)
403{
404 int hdr_padded_len;
405 struct xdp_buff xdp;
406 void *buf;
407 unsigned int qp;
408 u32 act;
409
410 if (vi->mergeable_rx_bufs) {
411 hdr_padded_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
412 xdp.data = data + hdr_padded_len;
413 xdp.data_end = xdp.data + (len - vi->hdr_len);
414 buf = data;
415 } else { /* small buffers */
416 struct sk_buff *skb = data;
417
418 xdp.data = skb->data;
419 xdp.data_end = xdp.data + len;
420 buf = skb->data;
421 }
422
423 act = bpf_prog_run_xdp(xdp_prog, &xdp);
424 switch (act) {
425 case XDP_PASS:
426 return XDP_PASS;
427 case XDP_TX:
428 qp = vi->curr_queue_pairs -
429 vi->xdp_queue_pairs +
430 smp_processor_id();
431 xdp.data = buf;
432 virtnet_xdp_xmit(vi, rq, &vi->sq[qp], &xdp, data);
433 return XDP_TX;
434 default:
435 bpf_warn_invalid_xdp_action(act);
436 case XDP_ABORTED:
437 case XDP_DROP:
438 return XDP_DROP;
439 }
440}
441
442static struct sk_buff *receive_small(struct net_device *dev,
443 struct virtnet_info *vi,
444 struct receive_queue *rq,
445 void *buf, unsigned int len)
446{
447 struct sk_buff * skb = buf;
448 struct bpf_prog *xdp_prog;
449
450 len -= vi->hdr_len;
451 skb_trim(skb, len);
452
453 rcu_read_lock();
454 xdp_prog = rcu_dereference(rq->xdp_prog);
455 if (xdp_prog) {
456 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
457 u32 act;
458
459 if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags))
460 goto err_xdp;
461 act = do_xdp_prog(vi, rq, xdp_prog, skb, len);
462 switch (act) {
463 case XDP_PASS:
464 break;
465 case XDP_TX:
466 rcu_read_unlock();
467 goto xdp_xmit;
468 case XDP_DROP:
469 default:
470 goto err_xdp;
471 }
472 }
473 rcu_read_unlock();
474
475 return skb;
476
477err_xdp:
478 rcu_read_unlock();
479 dev->stats.rx_dropped++;
480 kfree_skb(skb);
481xdp_xmit:
482 return NULL;
483}
484
485static struct sk_buff *receive_big(struct net_device *dev,
486 struct virtnet_info *vi,
487 struct receive_queue *rq,
488 void *buf,
489 unsigned int len)
490{
491 struct page *page = buf;
492 struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
493
494 if (unlikely(!skb))
495 goto err;
496
497 return skb;
498
499err:
500 dev->stats.rx_dropped++;
501 give_pages(rq, page);
502 return NULL;
503}
504
505/* The conditions to enable XDP should preclude the underlying device from
506 * sending packets across multiple buffers (num_buf > 1). However per spec
507 * it does not appear to be illegal to do so but rather just against convention.
508 * So in order to avoid making a system unresponsive the packets are pushed
509 * into a page and the XDP program is run. This will be extremely slow and we
510 * push a warning to the user to fix this as soon as possible. Fixing this may
511 * require resolving the underlying hardware to determine why multiple buffers
512 * are being received or simply loading the XDP program in the ingress stack
513 * after the skb is built because there is no advantage to running it here
514 * anymore.
515 */
516static struct page *xdp_linearize_page(struct receive_queue *rq,
517 u16 *num_buf,
518 struct page *p,
519 int offset,
520 unsigned int *len)
521{
522 struct page *page = alloc_page(GFP_ATOMIC);
523 unsigned int page_off = 0;
524
525 if (!page)
526 return NULL;
527
528 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
529 page_off += *len;
530
531 while (--*num_buf) {
532 unsigned int buflen;
533 unsigned long ctx;
534 void *buf;
535 int off;
536
537 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &buflen);
538 if (unlikely(!ctx))
539 goto err_buf;
540
541 buf = mergeable_ctx_to_buf_address(ctx);
542 p = virt_to_head_page(buf);
543 off = buf - page_address(p);
544
545 /* guard against a misconfigured or uncooperative backend that
546 * is sending packet larger than the MTU.
547 */
548 if ((page_off + buflen) > PAGE_SIZE) {
549 put_page(p);
550 goto err_buf;
551 }
552
553 memcpy(page_address(page) + page_off,
554 page_address(p) + off, buflen);
555 page_off += buflen;
556 put_page(p);
557 }
558
559 *len = page_off;
560 return page;
561err_buf:
562 __free_pages(page, 0);
563 return NULL;
564}
565
566static struct sk_buff *receive_mergeable(struct net_device *dev,
567 struct virtnet_info *vi,
568 struct receive_queue *rq,
569 unsigned long ctx,
570 unsigned int len)
571{
572 void *buf = mergeable_ctx_to_buf_address(ctx);
573 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
574 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
575 struct page *page = virt_to_head_page(buf);
576 int offset = buf - page_address(page);
577 struct sk_buff *head_skb, *curr_skb;
578 struct bpf_prog *xdp_prog;
579 unsigned int truesize;
580
581 head_skb = NULL;
582
583 rcu_read_lock();
584 xdp_prog = rcu_dereference(rq->xdp_prog);
585 if (xdp_prog) {
586 struct page *xdp_page;
587 u32 act;
588
589 /* This happens when rx buffer size is underestimated */
590 if (unlikely(num_buf > 1)) {
591 /* linearize data for XDP */
592 xdp_page = xdp_linearize_page(rq, &num_buf,
593 page, offset, &len);
594 if (!xdp_page)
595 goto err_xdp;
596 offset = 0;
597 } else {
598 xdp_page = page;
599 }
600
601 /* Transient failure which in theory could occur if
602 * in-flight packets from before XDP was enabled reach
603 * the receive path after XDP is loaded. In practice I
604 * was not able to create this condition.
605 */
606 if (unlikely(hdr->hdr.gso_type))
607 goto err_xdp;
608
609 act = do_xdp_prog(vi, rq, xdp_prog,
610 page_address(xdp_page) + offset, len);
611 switch (act) {
612 case XDP_PASS:
613 /* We can only create skb based on xdp_page. */
614 if (unlikely(xdp_page != page)) {
615 rcu_read_unlock();
616 put_page(page);
617 head_skb = page_to_skb(vi, rq, xdp_page,
618 0, len, PAGE_SIZE);
619 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
620 return head_skb;
621 }
622 break;
623 case XDP_TX:
624 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
625 if (unlikely(xdp_page != page))
626 goto err_xdp;
627 rcu_read_unlock();
628 goto xdp_xmit;
629 case XDP_DROP:
630 default:
631 if (unlikely(xdp_page != page))
632 __free_pages(xdp_page, 0);
633 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
634 goto err_xdp;
635 }
636 }
637 rcu_read_unlock();
638
639 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
640 head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
641 curr_skb = head_skb;
642
643 if (unlikely(!curr_skb))
644 goto err_skb;
645 while (--num_buf) {
646 int num_skb_frags;
647
648 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
649 if (unlikely(!ctx)) {
650 pr_debug("%s: rx error: %d buffers out of %d missing\n",
651 dev->name, num_buf,
652 virtio16_to_cpu(vi->vdev,
653 hdr->num_buffers));
654 dev->stats.rx_length_errors++;
655 goto err_buf;
656 }
657
658 buf = mergeable_ctx_to_buf_address(ctx);
659 page = virt_to_head_page(buf);
660
661 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
662 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
663 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
664
665 if (unlikely(!nskb))
666 goto err_skb;
667 if (curr_skb == head_skb)
668 skb_shinfo(curr_skb)->frag_list = nskb;
669 else
670 curr_skb->next = nskb;
671 curr_skb = nskb;
672 head_skb->truesize += nskb->truesize;
673 num_skb_frags = 0;
674 }
675 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
676 if (curr_skb != head_skb) {
677 head_skb->data_len += len;
678 head_skb->len += len;
679 head_skb->truesize += truesize;
680 }
681 offset = buf - page_address(page);
682 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
683 put_page(page);
684 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
685 len, truesize);
686 } else {
687 skb_add_rx_frag(curr_skb, num_skb_frags, page,
688 offset, len, truesize);
689 }
690 }
691
692 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
693 return head_skb;
694
695err_xdp:
696 rcu_read_unlock();
697err_skb:
698 put_page(page);
699 while (--num_buf) {
700 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
701 if (unlikely(!ctx)) {
702 pr_debug("%s: rx error: %d buffers missing\n",
703 dev->name, num_buf);
704 dev->stats.rx_length_errors++;
705 break;
706 }
707 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
708 put_page(page);
709 }
710err_buf:
711 dev->stats.rx_dropped++;
712 dev_kfree_skb(head_skb);
713xdp_xmit:
714 return NULL;
715}
716
717static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
718 void *buf, unsigned int len)
719{
720 struct net_device *dev = vi->dev;
721 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
722 struct sk_buff *skb;
723 struct virtio_net_hdr_mrg_rxbuf *hdr;
724
725 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
726 pr_debug("%s: short packet %i\n", dev->name, len);
727 dev->stats.rx_length_errors++;
728 if (vi->mergeable_rx_bufs) {
729 unsigned long ctx = (unsigned long)buf;
730 void *base = mergeable_ctx_to_buf_address(ctx);
731 put_page(virt_to_head_page(base));
732 } else if (vi->big_packets) {
733 give_pages(rq, buf);
734 } else {
735 dev_kfree_skb(buf);
736 }
737 return;
738 }
739
740 if (vi->mergeable_rx_bufs)
741 skb = receive_mergeable(dev, vi, rq, (unsigned long)buf, len);
742 else if (vi->big_packets)
743 skb = receive_big(dev, vi, rq, buf, len);
744 else
745 skb = receive_small(dev, vi, rq, buf, len);
746
747 if (unlikely(!skb))
748 return;
749
750 hdr = skb_vnet_hdr(skb);
751
752 u64_stats_update_begin(&stats->rx_syncp);
753 stats->rx_bytes += skb->len;
754 stats->rx_packets++;
755 u64_stats_update_end(&stats->rx_syncp);
756
757 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
758 skb->ip_summed = CHECKSUM_UNNECESSARY;
759
760 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
761 virtio_is_little_endian(vi->vdev))) {
762 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
763 dev->name, hdr->hdr.gso_type,
764 hdr->hdr.gso_size);
765 goto frame_err;
766 }
767
768 skb->protocol = eth_type_trans(skb, dev);
769 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
770 ntohs(skb->protocol), skb->len, skb->pkt_type);
771
772 napi_gro_receive(&rq->napi, skb);
773 return;
774
775frame_err:
776 dev->stats.rx_frame_errors++;
777 dev_kfree_skb(skb);
778}
779
780static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
781 gfp_t gfp)
782{
783 struct sk_buff *skb;
784 struct virtio_net_hdr_mrg_rxbuf *hdr;
785 int err;
786
787 skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
788 if (unlikely(!skb))
789 return -ENOMEM;
790
791 skb_put(skb, GOOD_PACKET_LEN);
792
793 hdr = skb_vnet_hdr(skb);
794 sg_init_table(rq->sg, 2);
795 sg_set_buf(rq->sg, hdr, vi->hdr_len);
796 skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
797
798 err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
799 if (err < 0)
800 dev_kfree_skb(skb);
801
802 return err;
803}
804
805static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
806 gfp_t gfp)
807{
808 struct page *first, *list = NULL;
809 char *p;
810 int i, err, offset;
811
812 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
813
814 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
815 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
816 first = get_a_page(rq, gfp);
817 if (!first) {
818 if (list)
819 give_pages(rq, list);
820 return -ENOMEM;
821 }
822 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
823
824 /* chain new page in list head to match sg */
825 first->private = (unsigned long)list;
826 list = first;
827 }
828
829 first = get_a_page(rq, gfp);
830 if (!first) {
831 give_pages(rq, list);
832 return -ENOMEM;
833 }
834 p = page_address(first);
835
836 /* rq->sg[0], rq->sg[1] share the same page */
837 /* a separated rq->sg[0] for header - required in case !any_header_sg */
838 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
839
840 /* rq->sg[1] for data packet, from offset */
841 offset = sizeof(struct padded_vnet_hdr);
842 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
843
844 /* chain first in list head */
845 first->private = (unsigned long)list;
846 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
847 first, gfp);
848 if (err < 0)
849 give_pages(rq, first);
850
851 return err;
852}
853
854static unsigned int get_mergeable_buf_len(struct ewma_pkt_len *avg_pkt_len)
855{
856 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
857 unsigned int len;
858
859 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
860 GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
861 return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
862}
863
864static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
865{
866 struct page_frag *alloc_frag = &rq->alloc_frag;
867 char *buf;
868 unsigned long ctx;
869 int err;
870 unsigned int len, hole;
871
872 len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
873 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
874 return -ENOMEM;
875
876 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
877 ctx = mergeable_buf_to_ctx(buf, len);
878 get_page(alloc_frag->page);
879 alloc_frag->offset += len;
880 hole = alloc_frag->size - alloc_frag->offset;
881 if (hole < len) {
882 /* To avoid internal fragmentation, if there is very likely not
883 * enough space for another buffer, add the remaining space to
884 * the current buffer. This extra space is not included in
885 * the truesize stored in ctx.
886 */
887 len += hole;
888 alloc_frag->offset += hole;
889 }
890
891 sg_init_one(rq->sg, buf, len);
892 err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
893 if (err < 0)
894 put_page(virt_to_head_page(buf));
895
896 return err;
897}
898
899/*
900 * Returns false if we couldn't fill entirely (OOM).
901 *
902 * Normally run in the receive path, but can also be run from ndo_open
903 * before we're receiving packets, or from refill_work which is
904 * careful to disable receiving (using napi_disable).
905 */
906static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
907 gfp_t gfp)
908{
909 int err;
910 bool oom;
911
912 gfp |= __GFP_COLD;
913 do {
914 if (vi->mergeable_rx_bufs)
915 err = add_recvbuf_mergeable(rq, gfp);
916 else if (vi->big_packets)
917 err = add_recvbuf_big(vi, rq, gfp);
918 else
919 err = add_recvbuf_small(vi, rq, gfp);
920
921 oom = err == -ENOMEM;
922 if (err)
923 break;
924 } while (rq->vq->num_free);
925 virtqueue_kick(rq->vq);
926 return !oom;
927}
928
929static void skb_recv_done(struct virtqueue *rvq)
930{
931 struct virtnet_info *vi = rvq->vdev->priv;
932 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
933
934 /* Schedule NAPI, Suppress further interrupts if successful. */
935 if (napi_schedule_prep(&rq->napi)) {
936 virtqueue_disable_cb(rvq);
937 __napi_schedule(&rq->napi);
938 }
939}
940
941static void virtnet_napi_enable(struct receive_queue *rq)
942{
943 napi_enable(&rq->napi);
944
945 /* If all buffers were filled by other side before we napi_enabled, we
946 * won't get another interrupt, so process any outstanding packets
947 * now. virtnet_poll wants re-enable the queue, so we disable here.
948 * We synchronize against interrupts via NAPI_STATE_SCHED */
949 if (napi_schedule_prep(&rq->napi)) {
950 virtqueue_disable_cb(rq->vq);
951 local_bh_disable();
952 __napi_schedule(&rq->napi);
953 local_bh_enable();
954 }
955}
956
957static void refill_work(struct work_struct *work)
958{
959 struct virtnet_info *vi =
960 container_of(work, struct virtnet_info, refill.work);
961 bool still_empty;
962 int i;
963
964 for (i = 0; i < vi->curr_queue_pairs; i++) {
965 struct receive_queue *rq = &vi->rq[i];
966
967 napi_disable(&rq->napi);
968 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
969 virtnet_napi_enable(rq);
970
971 /* In theory, this can happen: if we don't get any buffers in
972 * we will *never* try to fill again.
973 */
974 if (still_empty)
975 schedule_delayed_work(&vi->refill, HZ/2);
976 }
977}
978
979static int virtnet_receive(struct receive_queue *rq, int budget)
980{
981 struct virtnet_info *vi = rq->vq->vdev->priv;
982 unsigned int len, received = 0;
983 void *buf;
984
985 while (received < budget &&
986 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
987 receive_buf(vi, rq, buf, len);
988 received++;
989 }
990
991 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
992 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
993 schedule_delayed_work(&vi->refill, 0);
994 }
995
996 return received;
997}
998
999static int virtnet_poll(struct napi_struct *napi, int budget)
1000{
1001 struct receive_queue *rq =
1002 container_of(napi, struct receive_queue, napi);
1003 unsigned int r, received;
1004
1005 received = virtnet_receive(rq, budget);
1006
1007 /* Out of packets? */
1008 if (received < budget) {
1009 r = virtqueue_enable_cb_prepare(rq->vq);
1010 napi_complete_done(napi, received);
1011 if (unlikely(virtqueue_poll(rq->vq, r)) &&
1012 napi_schedule_prep(napi)) {
1013 virtqueue_disable_cb(rq->vq);
1014 __napi_schedule(napi);
1015 }
1016 }
1017
1018 return received;
1019}
1020
1021#ifdef CONFIG_NET_RX_BUSY_POLL
1022/* must be called with local_bh_disable()d */
1023static int virtnet_busy_poll(struct napi_struct *napi)
1024{
1025 struct receive_queue *rq =
1026 container_of(napi, struct receive_queue, napi);
1027 struct virtnet_info *vi = rq->vq->vdev->priv;
1028 int r, received = 0, budget = 4;
1029
1030 if (!(vi->status & VIRTIO_NET_S_LINK_UP))
1031 return LL_FLUSH_FAILED;
1032
1033 if (!napi_schedule_prep(napi))
1034 return LL_FLUSH_BUSY;
1035
1036 virtqueue_disable_cb(rq->vq);
1037
1038again:
1039 received += virtnet_receive(rq, budget);
1040
1041 r = virtqueue_enable_cb_prepare(rq->vq);
1042 clear_bit(NAPI_STATE_SCHED, &napi->state);
1043 if (unlikely(virtqueue_poll(rq->vq, r)) &&
1044 napi_schedule_prep(napi)) {
1045 virtqueue_disable_cb(rq->vq);
1046 if (received < budget) {
1047 budget -= received;
1048 goto again;
1049 } else {
1050 __napi_schedule(napi);
1051 }
1052 }
1053
1054 return received;
1055}
1056#endif /* CONFIG_NET_RX_BUSY_POLL */
1057
1058static int virtnet_open(struct net_device *dev)
1059{
1060 struct virtnet_info *vi = netdev_priv(dev);
1061 int i;
1062
1063 for (i = 0; i < vi->max_queue_pairs; i++) {
1064 if (i < vi->curr_queue_pairs)
1065 /* Make sure we have some buffers: if oom use wq. */
1066 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1067 schedule_delayed_work(&vi->refill, 0);
1068 virtnet_napi_enable(&vi->rq[i]);
1069 }
1070
1071 return 0;
1072}
1073
1074static void free_old_xmit_skbs(struct send_queue *sq)
1075{
1076 struct sk_buff *skb;
1077 unsigned int len;
1078 struct virtnet_info *vi = sq->vq->vdev->priv;
1079 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1080
1081 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1082 pr_debug("Sent skb %p\n", skb);
1083
1084 u64_stats_update_begin(&stats->tx_syncp);
1085 stats->tx_bytes += skb->len;
1086 stats->tx_packets++;
1087 u64_stats_update_end(&stats->tx_syncp);
1088
1089 dev_kfree_skb_any(skb);
1090 }
1091}
1092
1093static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1094{
1095 struct virtio_net_hdr_mrg_rxbuf *hdr;
1096 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1097 struct virtnet_info *vi = sq->vq->vdev->priv;
1098 unsigned num_sg;
1099 unsigned hdr_len = vi->hdr_len;
1100 bool can_push;
1101
1102 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1103
1104 can_push = vi->any_header_sg &&
1105 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1106 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1107 /* Even if we can, don't push here yet as this would skew
1108 * csum_start offset below. */
1109 if (can_push)
1110 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1111 else
1112 hdr = skb_vnet_hdr(skb);
1113
1114 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1115 virtio_is_little_endian(vi->vdev), false))
1116 BUG();
1117
1118 if (vi->mergeable_rx_bufs)
1119 hdr->num_buffers = 0;
1120
1121 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1122 if (can_push) {
1123 __skb_push(skb, hdr_len);
1124 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1125 /* Pull header back to avoid skew in tx bytes calculations. */
1126 __skb_pull(skb, hdr_len);
1127 } else {
1128 sg_set_buf(sq->sg, hdr, hdr_len);
1129 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
1130 }
1131 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1132}
1133
1134static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1135{
1136 struct virtnet_info *vi = netdev_priv(dev);
1137 int qnum = skb_get_queue_mapping(skb);
1138 struct send_queue *sq = &vi->sq[qnum];
1139 int err;
1140 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1141 bool kick = !skb->xmit_more;
1142
1143 /* Free up any pending old buffers before queueing new ones. */
1144 free_old_xmit_skbs(sq);
1145
1146 /* timestamp packet in software */
1147 skb_tx_timestamp(skb);
1148
1149 /* Try to transmit */
1150 err = xmit_skb(sq, skb);
1151
1152 /* This should not happen! */
1153 if (unlikely(err)) {
1154 dev->stats.tx_fifo_errors++;
1155 if (net_ratelimit())
1156 dev_warn(&dev->dev,
1157 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1158 dev->stats.tx_dropped++;
1159 dev_kfree_skb_any(skb);
1160 return NETDEV_TX_OK;
1161 }
1162
1163 /* Don't wait up for transmitted skbs to be freed. */
1164 skb_orphan(skb);
1165 nf_reset(skb);
1166
1167 /* If running out of space, stop queue to avoid getting packets that we
1168 * are then unable to transmit.
1169 * An alternative would be to force queuing layer to requeue the skb by
1170 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1171 * returned in a normal path of operation: it means that driver is not
1172 * maintaining the TX queue stop/start state properly, and causes
1173 * the stack to do a non-trivial amount of useless work.
1174 * Since most packets only take 1 or 2 ring slots, stopping the queue
1175 * early means 16 slots are typically wasted.
1176 */
1177 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1178 netif_stop_subqueue(dev, qnum);
1179 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1180 /* More just got used, free them then recheck. */
1181 free_old_xmit_skbs(sq);
1182 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1183 netif_start_subqueue(dev, qnum);
1184 virtqueue_disable_cb(sq->vq);
1185 }
1186 }
1187 }
1188
1189 if (kick || netif_xmit_stopped(txq))
1190 virtqueue_kick(sq->vq);
1191
1192 return NETDEV_TX_OK;
1193}
1194
1195/*
1196 * Send command via the control virtqueue and check status. Commands
1197 * supported by the hypervisor, as indicated by feature bits, should
1198 * never fail unless improperly formatted.
1199 */
1200static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1201 struct scatterlist *out)
1202{
1203 struct scatterlist *sgs[4], hdr, stat;
1204 unsigned out_num = 0, tmp;
1205
1206 /* Caller should know better */
1207 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1208
1209 vi->ctrl_status = ~0;
1210 vi->ctrl_hdr.class = class;
1211 vi->ctrl_hdr.cmd = cmd;
1212 /* Add header */
1213 sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
1214 sgs[out_num++] = &hdr;
1215
1216 if (out)
1217 sgs[out_num++] = out;
1218
1219 /* Add return status. */
1220 sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1221 sgs[out_num] = &stat;
1222
1223 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1224 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1225
1226 if (unlikely(!virtqueue_kick(vi->cvq)))
1227 return vi->ctrl_status == VIRTIO_NET_OK;
1228
1229 /* Spin for a response, the kick causes an ioport write, trapping
1230 * into the hypervisor, so the request should be handled immediately.
1231 */
1232 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1233 !virtqueue_is_broken(vi->cvq))
1234 cpu_relax();
1235
1236 return vi->ctrl_status == VIRTIO_NET_OK;
1237}
1238
1239static int virtnet_set_mac_address(struct net_device *dev, void *p)
1240{
1241 struct virtnet_info *vi = netdev_priv(dev);
1242 struct virtio_device *vdev = vi->vdev;
1243 int ret;
1244 struct sockaddr *addr;
1245 struct scatterlist sg;
1246
1247 addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1248 if (!addr)
1249 return -ENOMEM;
1250 memcpy(addr, p, sizeof(*addr));
1251
1252 ret = eth_prepare_mac_addr_change(dev, addr);
1253 if (ret)
1254 goto out;
1255
1256 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1257 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1258 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1259 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1260 dev_warn(&vdev->dev,
1261 "Failed to set mac address by vq command.\n");
1262 ret = -EINVAL;
1263 goto out;
1264 }
1265 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1266 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1267 unsigned int i;
1268
1269 /* Naturally, this has an atomicity problem. */
1270 for (i = 0; i < dev->addr_len; i++)
1271 virtio_cwrite8(vdev,
1272 offsetof(struct virtio_net_config, mac) +
1273 i, addr->sa_data[i]);
1274 }
1275
1276 eth_commit_mac_addr_change(dev, p);
1277 ret = 0;
1278
1279out:
1280 kfree(addr);
1281 return ret;
1282}
1283
1284static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1285 struct rtnl_link_stats64 *tot)
1286{
1287 struct virtnet_info *vi = netdev_priv(dev);
1288 int cpu;
1289 unsigned int start;
1290
1291 for_each_possible_cpu(cpu) {
1292 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1293 u64 tpackets, tbytes, rpackets, rbytes;
1294
1295 do {
1296 start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1297 tpackets = stats->tx_packets;
1298 tbytes = stats->tx_bytes;
1299 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1300
1301 do {
1302 start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1303 rpackets = stats->rx_packets;
1304 rbytes = stats->rx_bytes;
1305 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1306
1307 tot->rx_packets += rpackets;
1308 tot->tx_packets += tpackets;
1309 tot->rx_bytes += rbytes;
1310 tot->tx_bytes += tbytes;
1311 }
1312
1313 tot->tx_dropped = dev->stats.tx_dropped;
1314 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1315 tot->rx_dropped = dev->stats.rx_dropped;
1316 tot->rx_length_errors = dev->stats.rx_length_errors;
1317 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1318
1319 return tot;
1320}
1321
1322#ifdef CONFIG_NET_POLL_CONTROLLER
1323static void virtnet_netpoll(struct net_device *dev)
1324{
1325 struct virtnet_info *vi = netdev_priv(dev);
1326 int i;
1327
1328 for (i = 0; i < vi->curr_queue_pairs; i++)
1329 napi_schedule(&vi->rq[i].napi);
1330}
1331#endif
1332
1333static void virtnet_ack_link_announce(struct virtnet_info *vi)
1334{
1335 rtnl_lock();
1336 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1337 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1338 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1339 rtnl_unlock();
1340}
1341
1342static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1343{
1344 struct scatterlist sg;
1345 struct net_device *dev = vi->dev;
1346
1347 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1348 return 0;
1349
1350 vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1351 sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1352
1353 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1354 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1355 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1356 queue_pairs);
1357 return -EINVAL;
1358 } else {
1359 vi->curr_queue_pairs = queue_pairs;
1360 /* virtnet_open() will refill when device is going to up. */
1361 if (dev->flags & IFF_UP)
1362 schedule_delayed_work(&vi->refill, 0);
1363 }
1364
1365 return 0;
1366}
1367
1368static int virtnet_close(struct net_device *dev)
1369{
1370 struct virtnet_info *vi = netdev_priv(dev);
1371 int i;
1372
1373 /* Make sure refill_work doesn't re-enable napi! */
1374 cancel_delayed_work_sync(&vi->refill);
1375
1376 for (i = 0; i < vi->max_queue_pairs; i++)
1377 napi_disable(&vi->rq[i].napi);
1378
1379 return 0;
1380}
1381
1382static void virtnet_set_rx_mode(struct net_device *dev)
1383{
1384 struct virtnet_info *vi = netdev_priv(dev);
1385 struct scatterlist sg[2];
1386 struct virtio_net_ctrl_mac *mac_data;
1387 struct netdev_hw_addr *ha;
1388 int uc_count;
1389 int mc_count;
1390 void *buf;
1391 int i;
1392
1393 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1394 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1395 return;
1396
1397 vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1398 vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1399
1400 sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1401
1402 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1403 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1404 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1405 vi->ctrl_promisc ? "en" : "dis");
1406
1407 sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1408
1409 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1410 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1411 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1412 vi->ctrl_allmulti ? "en" : "dis");
1413
1414 uc_count = netdev_uc_count(dev);
1415 mc_count = netdev_mc_count(dev);
1416 /* MAC filter - use one buffer for both lists */
1417 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1418 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1419 mac_data = buf;
1420 if (!buf)
1421 return;
1422
1423 sg_init_table(sg, 2);
1424
1425 /* Store the unicast list and count in the front of the buffer */
1426 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1427 i = 0;
1428 netdev_for_each_uc_addr(ha, dev)
1429 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1430
1431 sg_set_buf(&sg[0], mac_data,
1432 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1433
1434 /* multicast list and count fill the end */
1435 mac_data = (void *)&mac_data->macs[uc_count][0];
1436
1437 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1438 i = 0;
1439 netdev_for_each_mc_addr(ha, dev)
1440 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1441
1442 sg_set_buf(&sg[1], mac_data,
1443 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1444
1445 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1446 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1447 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1448
1449 kfree(buf);
1450}
1451
1452static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1453 __be16 proto, u16 vid)
1454{
1455 struct virtnet_info *vi = netdev_priv(dev);
1456 struct scatterlist sg;
1457
1458 vi->ctrl_vid = vid;
1459 sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1460
1461 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1462 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1463 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1464 return 0;
1465}
1466
1467static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1468 __be16 proto, u16 vid)
1469{
1470 struct virtnet_info *vi = netdev_priv(dev);
1471 struct scatterlist sg;
1472
1473 vi->ctrl_vid = vid;
1474 sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1475
1476 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1477 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1478 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1479 return 0;
1480}
1481
1482static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1483{
1484 int i;
1485
1486 if (vi->affinity_hint_set) {
1487 for (i = 0; i < vi->max_queue_pairs; i++) {
1488 virtqueue_set_affinity(vi->rq[i].vq, -1);
1489 virtqueue_set_affinity(vi->sq[i].vq, -1);
1490 }
1491
1492 vi->affinity_hint_set = false;
1493 }
1494}
1495
1496static void virtnet_set_affinity(struct virtnet_info *vi)
1497{
1498 int i;
1499 int cpu;
1500
1501 /* In multiqueue mode, when the number of cpu is equal to the number of
1502 * queue pairs, we let the queue pairs to be private to one cpu by
1503 * setting the affinity hint to eliminate the contention.
1504 */
1505 if (vi->curr_queue_pairs == 1 ||
1506 vi->max_queue_pairs != num_online_cpus()) {
1507 virtnet_clean_affinity(vi, -1);
1508 return;
1509 }
1510
1511 i = 0;
1512 for_each_online_cpu(cpu) {
1513 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1514 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1515 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1516 i++;
1517 }
1518
1519 vi->affinity_hint_set = true;
1520}
1521
1522static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1523{
1524 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1525 node);
1526 virtnet_set_affinity(vi);
1527 return 0;
1528}
1529
1530static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1531{
1532 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1533 node_dead);
1534 virtnet_set_affinity(vi);
1535 return 0;
1536}
1537
1538static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1539{
1540 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1541 node);
1542
1543 virtnet_clean_affinity(vi, cpu);
1544 return 0;
1545}
1546
1547static enum cpuhp_state virtionet_online;
1548
1549static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1550{
1551 int ret;
1552
1553 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1554 if (ret)
1555 return ret;
1556 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1557 &vi->node_dead);
1558 if (!ret)
1559 return ret;
1560 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1561 return ret;
1562}
1563
1564static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1565{
1566 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1567 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1568 &vi->node_dead);
1569}
1570
1571static void virtnet_get_ringparam(struct net_device *dev,
1572 struct ethtool_ringparam *ring)
1573{
1574 struct virtnet_info *vi = netdev_priv(dev);
1575
1576 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1577 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1578 ring->rx_pending = ring->rx_max_pending;
1579 ring->tx_pending = ring->tx_max_pending;
1580}
1581
1582
1583static void virtnet_get_drvinfo(struct net_device *dev,
1584 struct ethtool_drvinfo *info)
1585{
1586 struct virtnet_info *vi = netdev_priv(dev);
1587 struct virtio_device *vdev = vi->vdev;
1588
1589 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1590 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1591 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1592
1593}
1594
1595/* TODO: Eliminate OOO packets during switching */
1596static int virtnet_set_channels(struct net_device *dev,
1597 struct ethtool_channels *channels)
1598{
1599 struct virtnet_info *vi = netdev_priv(dev);
1600 u16 queue_pairs = channels->combined_count;
1601 int err;
1602
1603 /* We don't support separate rx/tx channels.
1604 * We don't allow setting 'other' channels.
1605 */
1606 if (channels->rx_count || channels->tx_count || channels->other_count)
1607 return -EINVAL;
1608
1609 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1610 return -EINVAL;
1611
1612 /* For now we don't support modifying channels while XDP is loaded
1613 * also when XDP is loaded all RX queues have XDP programs so we only
1614 * need to check a single RX queue.
1615 */
1616 if (vi->rq[0].xdp_prog)
1617 return -EINVAL;
1618
1619 get_online_cpus();
1620 err = virtnet_set_queues(vi, queue_pairs);
1621 if (!err) {
1622 netif_set_real_num_tx_queues(dev, queue_pairs);
1623 netif_set_real_num_rx_queues(dev, queue_pairs);
1624
1625 virtnet_set_affinity(vi);
1626 }
1627 put_online_cpus();
1628
1629 return err;
1630}
1631
1632static void virtnet_get_channels(struct net_device *dev,
1633 struct ethtool_channels *channels)
1634{
1635 struct virtnet_info *vi = netdev_priv(dev);
1636
1637 channels->combined_count = vi->curr_queue_pairs;
1638 channels->max_combined = vi->max_queue_pairs;
1639 channels->max_other = 0;
1640 channels->rx_count = 0;
1641 channels->tx_count = 0;
1642 channels->other_count = 0;
1643}
1644
1645/* Check if the user is trying to change anything besides speed/duplex */
1646static bool virtnet_validate_ethtool_cmd(const struct ethtool_cmd *cmd)
1647{
1648 struct ethtool_cmd diff1 = *cmd;
1649 struct ethtool_cmd diff2 = {};
1650
1651 /* cmd is always set so we need to clear it, validate the port type
1652 * and also without autonegotiation we can ignore advertising
1653 */
1654 ethtool_cmd_speed_set(&diff1, 0);
1655 diff2.port = PORT_OTHER;
1656 diff1.advertising = 0;
1657 diff1.duplex = 0;
1658 diff1.cmd = 0;
1659
1660 return !memcmp(&diff1, &diff2, sizeof(diff1));
1661}
1662
1663static int virtnet_set_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1664{
1665 struct virtnet_info *vi = netdev_priv(dev);
1666 u32 speed;
1667
1668 speed = ethtool_cmd_speed(cmd);
1669 /* don't allow custom speed and duplex */
1670 if (!ethtool_validate_speed(speed) ||
1671 !ethtool_validate_duplex(cmd->duplex) ||
1672 !virtnet_validate_ethtool_cmd(cmd))
1673 return -EINVAL;
1674 vi->speed = speed;
1675 vi->duplex = cmd->duplex;
1676
1677 return 0;
1678}
1679
1680static int virtnet_get_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1681{
1682 struct virtnet_info *vi = netdev_priv(dev);
1683
1684 ethtool_cmd_speed_set(cmd, vi->speed);
1685 cmd->duplex = vi->duplex;
1686 cmd->port = PORT_OTHER;
1687
1688 return 0;
1689}
1690
1691static void virtnet_init_settings(struct net_device *dev)
1692{
1693 struct virtnet_info *vi = netdev_priv(dev);
1694
1695 vi->speed = SPEED_UNKNOWN;
1696 vi->duplex = DUPLEX_UNKNOWN;
1697}
1698
1699static const struct ethtool_ops virtnet_ethtool_ops = {
1700 .get_drvinfo = virtnet_get_drvinfo,
1701 .get_link = ethtool_op_get_link,
1702 .get_ringparam = virtnet_get_ringparam,
1703 .set_channels = virtnet_set_channels,
1704 .get_channels = virtnet_get_channels,
1705 .get_ts_info = ethtool_op_get_ts_info,
1706 .get_settings = virtnet_get_settings,
1707 .set_settings = virtnet_set_settings,
1708};
1709
1710static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog)
1711{
1712 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
1713 struct virtnet_info *vi = netdev_priv(dev);
1714 struct bpf_prog *old_prog;
1715 u16 xdp_qp = 0, curr_qp;
1716 int i, err;
1717
1718 if (prog && prog->xdp_adjust_head) {
1719 netdev_warn(dev, "Does not support bpf_xdp_adjust_head()\n");
1720 return -EOPNOTSUPP;
1721 }
1722
1723 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1724 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1725 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
1726 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO)) {
1727 netdev_warn(dev, "can't set XDP while host is implementing LRO, disable LRO first\n");
1728 return -EOPNOTSUPP;
1729 }
1730
1731 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
1732 netdev_warn(dev, "XDP expects header/data in single page, any_header_sg required\n");
1733 return -EINVAL;
1734 }
1735
1736 if (dev->mtu > max_sz) {
1737 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
1738 return -EINVAL;
1739 }
1740
1741 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
1742 if (prog)
1743 xdp_qp = nr_cpu_ids;
1744
1745 /* XDP requires extra queues for XDP_TX */
1746 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
1747 netdev_warn(dev, "request %i queues but max is %i\n",
1748 curr_qp + xdp_qp, vi->max_queue_pairs);
1749 return -ENOMEM;
1750 }
1751
1752 err = virtnet_set_queues(vi, curr_qp + xdp_qp);
1753 if (err) {
1754 dev_warn(&dev->dev, "XDP Device queue allocation failure.\n");
1755 return err;
1756 }
1757
1758 if (prog) {
1759 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
1760 if (IS_ERR(prog)) {
1761 virtnet_set_queues(vi, curr_qp);
1762 return PTR_ERR(prog);
1763 }
1764 }
1765
1766 vi->xdp_queue_pairs = xdp_qp;
1767 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
1768
1769 for (i = 0; i < vi->max_queue_pairs; i++) {
1770 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1771 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
1772 if (old_prog)
1773 bpf_prog_put(old_prog);
1774 }
1775
1776 return 0;
1777}
1778
1779static bool virtnet_xdp_query(struct net_device *dev)
1780{
1781 struct virtnet_info *vi = netdev_priv(dev);
1782 int i;
1783
1784 for (i = 0; i < vi->max_queue_pairs; i++) {
1785 if (vi->rq[i].xdp_prog)
1786 return true;
1787 }
1788 return false;
1789}
1790
1791static int virtnet_xdp(struct net_device *dev, struct netdev_xdp *xdp)
1792{
1793 switch (xdp->command) {
1794 case XDP_SETUP_PROG:
1795 return virtnet_xdp_set(dev, xdp->prog);
1796 case XDP_QUERY_PROG:
1797 xdp->prog_attached = virtnet_xdp_query(dev);
1798 return 0;
1799 default:
1800 return -EINVAL;
1801 }
1802}
1803
1804static const struct net_device_ops virtnet_netdev = {
1805 .ndo_open = virtnet_open,
1806 .ndo_stop = virtnet_close,
1807 .ndo_start_xmit = start_xmit,
1808 .ndo_validate_addr = eth_validate_addr,
1809 .ndo_set_mac_address = virtnet_set_mac_address,
1810 .ndo_set_rx_mode = virtnet_set_rx_mode,
1811 .ndo_get_stats64 = virtnet_stats,
1812 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1813 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1814#ifdef CONFIG_NET_POLL_CONTROLLER
1815 .ndo_poll_controller = virtnet_netpoll,
1816#endif
1817#ifdef CONFIG_NET_RX_BUSY_POLL
1818 .ndo_busy_poll = virtnet_busy_poll,
1819#endif
1820 .ndo_xdp = virtnet_xdp,
1821};
1822
1823static void virtnet_config_changed_work(struct work_struct *work)
1824{
1825 struct virtnet_info *vi =
1826 container_of(work, struct virtnet_info, config_work);
1827 u16 v;
1828
1829 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1830 struct virtio_net_config, status, &v) < 0)
1831 return;
1832
1833 if (v & VIRTIO_NET_S_ANNOUNCE) {
1834 netdev_notify_peers(vi->dev);
1835 virtnet_ack_link_announce(vi);
1836 }
1837
1838 /* Ignore unknown (future) status bits */
1839 v &= VIRTIO_NET_S_LINK_UP;
1840
1841 if (vi->status == v)
1842 return;
1843
1844 vi->status = v;
1845
1846 if (vi->status & VIRTIO_NET_S_LINK_UP) {
1847 netif_carrier_on(vi->dev);
1848 netif_tx_wake_all_queues(vi->dev);
1849 } else {
1850 netif_carrier_off(vi->dev);
1851 netif_tx_stop_all_queues(vi->dev);
1852 }
1853}
1854
1855static void virtnet_config_changed(struct virtio_device *vdev)
1856{
1857 struct virtnet_info *vi = vdev->priv;
1858
1859 schedule_work(&vi->config_work);
1860}
1861
1862static void virtnet_free_queues(struct virtnet_info *vi)
1863{
1864 int i;
1865
1866 for (i = 0; i < vi->max_queue_pairs; i++) {
1867 napi_hash_del(&vi->rq[i].napi);
1868 netif_napi_del(&vi->rq[i].napi);
1869 }
1870
1871 /* We called napi_hash_del() before netif_napi_del(),
1872 * we need to respect an RCU grace period before freeing vi->rq
1873 */
1874 synchronize_net();
1875
1876 kfree(vi->rq);
1877 kfree(vi->sq);
1878}
1879
1880static void free_receive_bufs(struct virtnet_info *vi)
1881{
1882 struct bpf_prog *old_prog;
1883 int i;
1884
1885 rtnl_lock();
1886 for (i = 0; i < vi->max_queue_pairs; i++) {
1887 while (vi->rq[i].pages)
1888 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1889
1890 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1891 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
1892 if (old_prog)
1893 bpf_prog_put(old_prog);
1894 }
1895 rtnl_unlock();
1896}
1897
1898static void free_receive_page_frags(struct virtnet_info *vi)
1899{
1900 int i;
1901 for (i = 0; i < vi->max_queue_pairs; i++)
1902 if (vi->rq[i].alloc_frag.page)
1903 put_page(vi->rq[i].alloc_frag.page);
1904}
1905
1906static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1907{
1908 /* For small receive mode always use kfree_skb variants */
1909 if (!vi->mergeable_rx_bufs)
1910 return false;
1911
1912 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1913 return false;
1914 else if (q < vi->curr_queue_pairs)
1915 return true;
1916 else
1917 return false;
1918}
1919
1920static void free_unused_bufs(struct virtnet_info *vi)
1921{
1922 void *buf;
1923 int i;
1924
1925 for (i = 0; i < vi->max_queue_pairs; i++) {
1926 struct virtqueue *vq = vi->sq[i].vq;
1927 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1928 if (!is_xdp_raw_buffer_queue(vi, i))
1929 dev_kfree_skb(buf);
1930 else
1931 put_page(virt_to_head_page(buf));
1932 }
1933 }
1934
1935 for (i = 0; i < vi->max_queue_pairs; i++) {
1936 struct virtqueue *vq = vi->rq[i].vq;
1937
1938 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1939 if (vi->mergeable_rx_bufs) {
1940 unsigned long ctx = (unsigned long)buf;
1941 void *base = mergeable_ctx_to_buf_address(ctx);
1942 put_page(virt_to_head_page(base));
1943 } else if (vi->big_packets) {
1944 give_pages(&vi->rq[i], buf);
1945 } else {
1946 dev_kfree_skb(buf);
1947 }
1948 }
1949 }
1950}
1951
1952static void virtnet_del_vqs(struct virtnet_info *vi)
1953{
1954 struct virtio_device *vdev = vi->vdev;
1955
1956 virtnet_clean_affinity(vi, -1);
1957
1958 vdev->config->del_vqs(vdev);
1959
1960 virtnet_free_queues(vi);
1961}
1962
1963static int virtnet_find_vqs(struct virtnet_info *vi)
1964{
1965 vq_callback_t **callbacks;
1966 struct virtqueue **vqs;
1967 int ret = -ENOMEM;
1968 int i, total_vqs;
1969 const char **names;
1970
1971 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1972 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1973 * possible control vq.
1974 */
1975 total_vqs = vi->max_queue_pairs * 2 +
1976 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1977
1978 /* Allocate space for find_vqs parameters */
1979 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1980 if (!vqs)
1981 goto err_vq;
1982 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1983 if (!callbacks)
1984 goto err_callback;
1985 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1986 if (!names)
1987 goto err_names;
1988
1989 /* Parameters for control virtqueue, if any */
1990 if (vi->has_cvq) {
1991 callbacks[total_vqs - 1] = NULL;
1992 names[total_vqs - 1] = "control";
1993 }
1994
1995 /* Allocate/initialize parameters for send/receive virtqueues */
1996 for (i = 0; i < vi->max_queue_pairs; i++) {
1997 callbacks[rxq2vq(i)] = skb_recv_done;
1998 callbacks[txq2vq(i)] = skb_xmit_done;
1999 sprintf(vi->rq[i].name, "input.%d", i);
2000 sprintf(vi->sq[i].name, "output.%d", i);
2001 names[rxq2vq(i)] = vi->rq[i].name;
2002 names[txq2vq(i)] = vi->sq[i].name;
2003 }
2004
2005 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2006 names);
2007 if (ret)
2008 goto err_find;
2009
2010 if (vi->has_cvq) {
2011 vi->cvq = vqs[total_vqs - 1];
2012 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2013 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2014 }
2015
2016 for (i = 0; i < vi->max_queue_pairs; i++) {
2017 vi->rq[i].vq = vqs[rxq2vq(i)];
2018 vi->sq[i].vq = vqs[txq2vq(i)];
2019 }
2020
2021 kfree(names);
2022 kfree(callbacks);
2023 kfree(vqs);
2024
2025 return 0;
2026
2027err_find:
2028 kfree(names);
2029err_names:
2030 kfree(callbacks);
2031err_callback:
2032 kfree(vqs);
2033err_vq:
2034 return ret;
2035}
2036
2037static int virtnet_alloc_queues(struct virtnet_info *vi)
2038{
2039 int i;
2040
2041 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2042 if (!vi->sq)
2043 goto err_sq;
2044 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2045 if (!vi->rq)
2046 goto err_rq;
2047
2048 INIT_DELAYED_WORK(&vi->refill, refill_work);
2049 for (i = 0; i < vi->max_queue_pairs; i++) {
2050 vi->rq[i].pages = NULL;
2051 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2052 napi_weight);
2053
2054 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2055 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2056 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2057 }
2058
2059 return 0;
2060
2061err_rq:
2062 kfree(vi->sq);
2063err_sq:
2064 return -ENOMEM;
2065}
2066
2067static int init_vqs(struct virtnet_info *vi)
2068{
2069 int ret;
2070
2071 /* Allocate send & receive queues */
2072 ret = virtnet_alloc_queues(vi);
2073 if (ret)
2074 goto err;
2075
2076 ret = virtnet_find_vqs(vi);
2077 if (ret)
2078 goto err_free;
2079
2080 get_online_cpus();
2081 virtnet_set_affinity(vi);
2082 put_online_cpus();
2083
2084 return 0;
2085
2086err_free:
2087 virtnet_free_queues(vi);
2088err:
2089 return ret;
2090}
2091
2092#ifdef CONFIG_SYSFS
2093static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2094 struct rx_queue_attribute *attribute, char *buf)
2095{
2096 struct virtnet_info *vi = netdev_priv(queue->dev);
2097 unsigned int queue_index = get_netdev_rx_queue_index(queue);
2098 struct ewma_pkt_len *avg;
2099
2100 BUG_ON(queue_index >= vi->max_queue_pairs);
2101 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2102 return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
2103}
2104
2105static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2106 __ATTR_RO(mergeable_rx_buffer_size);
2107
2108static struct attribute *virtio_net_mrg_rx_attrs[] = {
2109 &mergeable_rx_buffer_size_attribute.attr,
2110 NULL
2111};
2112
2113static const struct attribute_group virtio_net_mrg_rx_group = {
2114 .name = "virtio_net",
2115 .attrs = virtio_net_mrg_rx_attrs
2116};
2117#endif
2118
2119static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2120 unsigned int fbit,
2121 const char *fname, const char *dname)
2122{
2123 if (!virtio_has_feature(vdev, fbit))
2124 return false;
2125
2126 dev_err(&vdev->dev, "device advertises feature %s but not %s",
2127 fname, dname);
2128
2129 return true;
2130}
2131
2132#define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2133 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2134
2135static bool virtnet_validate_features(struct virtio_device *vdev)
2136{
2137 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2138 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2139 "VIRTIO_NET_F_CTRL_VQ") ||
2140 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2141 "VIRTIO_NET_F_CTRL_VQ") ||
2142 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2143 "VIRTIO_NET_F_CTRL_VQ") ||
2144 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2145 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2146 "VIRTIO_NET_F_CTRL_VQ"))) {
2147 return false;
2148 }
2149
2150 return true;
2151}
2152
2153#define MIN_MTU ETH_MIN_MTU
2154#define MAX_MTU ETH_MAX_MTU
2155
2156static int virtnet_probe(struct virtio_device *vdev)
2157{
2158 int i, err;
2159 struct net_device *dev;
2160 struct virtnet_info *vi;
2161 u16 max_queue_pairs;
2162 int mtu;
2163
2164 if (!vdev->config->get) {
2165 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2166 __func__);
2167 return -EINVAL;
2168 }
2169
2170 if (!virtnet_validate_features(vdev))
2171 return -EINVAL;
2172
2173 /* Find if host supports multiqueue virtio_net device */
2174 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2175 struct virtio_net_config,
2176 max_virtqueue_pairs, &max_queue_pairs);
2177
2178 /* We need at least 2 queue's */
2179 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2180 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2181 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2182 max_queue_pairs = 1;
2183
2184 /* Allocate ourselves a network device with room for our info */
2185 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2186 if (!dev)
2187 return -ENOMEM;
2188
2189 /* Set up network device as normal. */
2190 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2191 dev->netdev_ops = &virtnet_netdev;
2192 dev->features = NETIF_F_HIGHDMA;
2193
2194 dev->ethtool_ops = &virtnet_ethtool_ops;
2195 SET_NETDEV_DEV(dev, &vdev->dev);
2196
2197 /* Do we support "hardware" checksums? */
2198 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2199 /* This opens up the world of extra features. */
2200 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2201 if (csum)
2202 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2203
2204 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2205 dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
2206 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2207 }
2208 /* Individual feature bits: what can host handle? */
2209 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2210 dev->hw_features |= NETIF_F_TSO;
2211 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2212 dev->hw_features |= NETIF_F_TSO6;
2213 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2214 dev->hw_features |= NETIF_F_TSO_ECN;
2215 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
2216 dev->hw_features |= NETIF_F_UFO;
2217
2218 dev->features |= NETIF_F_GSO_ROBUST;
2219
2220 if (gso)
2221 dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
2222 /* (!csum && gso) case will be fixed by register_netdev() */
2223 }
2224 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2225 dev->features |= NETIF_F_RXCSUM;
2226
2227 dev->vlan_features = dev->features;
2228
2229 /* MTU range: 68 - 65535 */
2230 dev->min_mtu = MIN_MTU;
2231 dev->max_mtu = MAX_MTU;
2232
2233 /* Configuration may specify what MAC to use. Otherwise random. */
2234 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2235 virtio_cread_bytes(vdev,
2236 offsetof(struct virtio_net_config, mac),
2237 dev->dev_addr, dev->addr_len);
2238 else
2239 eth_hw_addr_random(dev);
2240
2241 /* Set up our device-specific information */
2242 vi = netdev_priv(dev);
2243 vi->dev = dev;
2244 vi->vdev = vdev;
2245 vdev->priv = vi;
2246 vi->stats = alloc_percpu(struct virtnet_stats);
2247 err = -ENOMEM;
2248 if (vi->stats == NULL)
2249 goto free;
2250
2251 for_each_possible_cpu(i) {
2252 struct virtnet_stats *virtnet_stats;
2253 virtnet_stats = per_cpu_ptr(vi->stats, i);
2254 u64_stats_init(&virtnet_stats->tx_syncp);
2255 u64_stats_init(&virtnet_stats->rx_syncp);
2256 }
2257
2258 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2259
2260 /* If we can receive ANY GSO packets, we must allocate large ones. */
2261 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2262 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2263 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2264 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2265 vi->big_packets = true;
2266
2267 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2268 vi->mergeable_rx_bufs = true;
2269
2270 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2271 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2272 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2273 else
2274 vi->hdr_len = sizeof(struct virtio_net_hdr);
2275
2276 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2277 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2278 vi->any_header_sg = true;
2279
2280 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2281 vi->has_cvq = true;
2282
2283 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2284 mtu = virtio_cread16(vdev,
2285 offsetof(struct virtio_net_config,
2286 mtu));
2287 if (mtu < dev->min_mtu) {
2288 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2289 } else {
2290 dev->mtu = mtu;
2291 dev->max_mtu = mtu;
2292 }
2293 }
2294
2295 if (vi->any_header_sg)
2296 dev->needed_headroom = vi->hdr_len;
2297
2298 /* Enable multiqueue by default */
2299 if (num_online_cpus() >= max_queue_pairs)
2300 vi->curr_queue_pairs = max_queue_pairs;
2301 else
2302 vi->curr_queue_pairs = num_online_cpus();
2303 vi->max_queue_pairs = max_queue_pairs;
2304
2305 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2306 err = init_vqs(vi);
2307 if (err)
2308 goto free_stats;
2309
2310#ifdef CONFIG_SYSFS
2311 if (vi->mergeable_rx_bufs)
2312 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2313#endif
2314 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2315 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2316
2317 virtnet_init_settings(dev);
2318
2319 err = register_netdev(dev);
2320 if (err) {
2321 pr_debug("virtio_net: registering device failed\n");
2322 goto free_vqs;
2323 }
2324
2325 virtio_device_ready(vdev);
2326
2327 err = virtnet_cpu_notif_add(vi);
2328 if (err) {
2329 pr_debug("virtio_net: registering cpu notifier failed\n");
2330 goto free_unregister_netdev;
2331 }
2332
2333 rtnl_lock();
2334 virtnet_set_queues(vi, vi->curr_queue_pairs);
2335 rtnl_unlock();
2336
2337 /* Assume link up if device can't report link status,
2338 otherwise get link status from config. */
2339 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2340 netif_carrier_off(dev);
2341 schedule_work(&vi->config_work);
2342 } else {
2343 vi->status = VIRTIO_NET_S_LINK_UP;
2344 netif_carrier_on(dev);
2345 }
2346
2347 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2348 dev->name, max_queue_pairs);
2349
2350 return 0;
2351
2352free_unregister_netdev:
2353 vi->vdev->config->reset(vdev);
2354
2355 unregister_netdev(dev);
2356free_vqs:
2357 cancel_delayed_work_sync(&vi->refill);
2358 free_receive_page_frags(vi);
2359 virtnet_del_vqs(vi);
2360free_stats:
2361 free_percpu(vi->stats);
2362free:
2363 free_netdev(dev);
2364 return err;
2365}
2366
2367static void remove_vq_common(struct virtnet_info *vi)
2368{
2369 vi->vdev->config->reset(vi->vdev);
2370
2371 /* Free unused buffers in both send and recv, if any. */
2372 free_unused_bufs(vi);
2373
2374 free_receive_bufs(vi);
2375
2376 free_receive_page_frags(vi);
2377
2378 virtnet_del_vqs(vi);
2379}
2380
2381static void virtnet_remove(struct virtio_device *vdev)
2382{
2383 struct virtnet_info *vi = vdev->priv;
2384
2385 virtnet_cpu_notif_remove(vi);
2386
2387 /* Make sure no work handler is accessing the device. */
2388 flush_work(&vi->config_work);
2389
2390 unregister_netdev(vi->dev);
2391
2392 remove_vq_common(vi);
2393
2394 free_percpu(vi->stats);
2395 free_netdev(vi->dev);
2396}
2397
2398#ifdef CONFIG_PM_SLEEP
2399static int virtnet_freeze(struct virtio_device *vdev)
2400{
2401 struct virtnet_info *vi = vdev->priv;
2402 int i;
2403
2404 virtnet_cpu_notif_remove(vi);
2405
2406 /* Make sure no work handler is accessing the device */
2407 flush_work(&vi->config_work);
2408
2409 netif_device_detach(vi->dev);
2410 cancel_delayed_work_sync(&vi->refill);
2411
2412 if (netif_running(vi->dev)) {
2413 for (i = 0; i < vi->max_queue_pairs; i++)
2414 napi_disable(&vi->rq[i].napi);
2415 }
2416
2417 remove_vq_common(vi);
2418
2419 return 0;
2420}
2421
2422static int virtnet_restore(struct virtio_device *vdev)
2423{
2424 struct virtnet_info *vi = vdev->priv;
2425 int err, i;
2426
2427 err = init_vqs(vi);
2428 if (err)
2429 return err;
2430
2431 virtio_device_ready(vdev);
2432
2433 if (netif_running(vi->dev)) {
2434 for (i = 0; i < vi->curr_queue_pairs; i++)
2435 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2436 schedule_delayed_work(&vi->refill, 0);
2437
2438 for (i = 0; i < vi->max_queue_pairs; i++)
2439 virtnet_napi_enable(&vi->rq[i]);
2440 }
2441
2442 netif_device_attach(vi->dev);
2443
2444 rtnl_lock();
2445 virtnet_set_queues(vi, vi->curr_queue_pairs);
2446 rtnl_unlock();
2447
2448 err = virtnet_cpu_notif_add(vi);
2449 if (err)
2450 return err;
2451
2452 return 0;
2453}
2454#endif
2455
2456static struct virtio_device_id id_table[] = {
2457 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2458 { 0 },
2459};
2460
2461#define VIRTNET_FEATURES \
2462 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2463 VIRTIO_NET_F_MAC, \
2464 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2465 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2466 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2467 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2468 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2469 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2470 VIRTIO_NET_F_CTRL_MAC_ADDR, \
2471 VIRTIO_NET_F_MTU
2472
2473static unsigned int features[] = {
2474 VIRTNET_FEATURES,
2475};
2476
2477static unsigned int features_legacy[] = {
2478 VIRTNET_FEATURES,
2479 VIRTIO_NET_F_GSO,
2480 VIRTIO_F_ANY_LAYOUT,
2481};
2482
2483static struct virtio_driver virtio_net_driver = {
2484 .feature_table = features,
2485 .feature_table_size = ARRAY_SIZE(features),
2486 .feature_table_legacy = features_legacy,
2487 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2488 .driver.name = KBUILD_MODNAME,
2489 .driver.owner = THIS_MODULE,
2490 .id_table = id_table,
2491 .probe = virtnet_probe,
2492 .remove = virtnet_remove,
2493 .config_changed = virtnet_config_changed,
2494#ifdef CONFIG_PM_SLEEP
2495 .freeze = virtnet_freeze,
2496 .restore = virtnet_restore,
2497#endif
2498};
2499
2500static __init int virtio_net_driver_init(void)
2501{
2502 int ret;
2503
2504 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
2505 virtnet_cpu_online,
2506 virtnet_cpu_down_prep);
2507 if (ret < 0)
2508 goto out;
2509 virtionet_online = ret;
2510 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
2511 NULL, virtnet_cpu_dead);
2512 if (ret)
2513 goto err_dead;
2514
2515 ret = register_virtio_driver(&virtio_net_driver);
2516 if (ret)
2517 goto err_virtio;
2518 return 0;
2519err_virtio:
2520 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2521err_dead:
2522 cpuhp_remove_multi_state(virtionet_online);
2523out:
2524 return ret;
2525}
2526module_init(virtio_net_driver_init);
2527
2528static __exit void virtio_net_driver_exit(void)
2529{
2530 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2531 cpuhp_remove_multi_state(virtionet_online);
2532 unregister_virtio_driver(&virtio_net_driver);
2533}
2534module_exit(virtio_net_driver_exit);
2535
2536MODULE_DEVICE_TABLE(virtio, id_table);
2537MODULE_DESCRIPTION("Virtio network driver");
2538MODULE_LICENSE("GPL");