virtio-net: reset virtqueue affinity when doing cpu hotplug
[~shefty/rdma-dev.git] / drivers / net / virtio_net.c
1 /* A network driver using virtio.
2  *
3  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License as published by
7  * the Free Software Foundation; either version 2 of the License, or
8  * (at your option) any later version.
9  *
10  * This program is distributed in the hope that it will be useful,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13  * GNU General Public License for more details.
14  *
15  * You should have received a copy of the GNU General Public License
16  * along with this program; if not, write to the Free Software
17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
18  */
19 //#define DEBUG
20 #include <linux/netdevice.h>
21 #include <linux/etherdevice.h>
22 #include <linux/ethtool.h>
23 #include <linux/module.h>
24 #include <linux/virtio.h>
25 #include <linux/virtio_net.h>
26 #include <linux/scatterlist.h>
27 #include <linux/if_vlan.h>
28 #include <linux/slab.h>
29 #include <linux/cpu.h>
30
31 static int napi_weight = 128;
32 module_param(napi_weight, int, 0444);
33
34 static bool csum = true, gso = true;
35 module_param(csum, bool, 0444);
36 module_param(gso, bool, 0444);
37
38 /* FIXME: MTU in config. */
39 #define MAX_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
40 #define GOOD_COPY_LEN   128
41
42 #define VIRTNET_SEND_COMMAND_SG_MAX    2
43 #define VIRTNET_DRIVER_VERSION "1.0.0"
44
45 struct virtnet_stats {
46         struct u64_stats_sync tx_syncp;
47         struct u64_stats_sync rx_syncp;
48         u64 tx_bytes;
49         u64 tx_packets;
50
51         u64 rx_bytes;
52         u64 rx_packets;
53 };
54
55 /* Internal representation of a send virtqueue */
56 struct send_queue {
57         /* Virtqueue associated with this send _queue */
58         struct virtqueue *vq;
59
60         /* TX: fragments + linear part + virtio header */
61         struct scatterlist sg[MAX_SKB_FRAGS + 2];
62
63         /* Name of the send queue: output.$index */
64         char name[40];
65 };
66
67 /* Internal representation of a receive virtqueue */
68 struct receive_queue {
69         /* Virtqueue associated with this receive_queue */
70         struct virtqueue *vq;
71
72         struct napi_struct napi;
73
74         /* Number of input buffers, and max we've ever had. */
75         unsigned int num, max;
76
77         /* Chain pages by the private ptr. */
78         struct page *pages;
79
80         /* RX: fragments + linear part + virtio header */
81         struct scatterlist sg[MAX_SKB_FRAGS + 2];
82
83         /* Name of this receive queue: input.$index */
84         char name[40];
85 };
86
87 struct virtnet_info {
88         struct virtio_device *vdev;
89         struct virtqueue *cvq;
90         struct net_device *dev;
91         struct send_queue *sq;
92         struct receive_queue *rq;
93         unsigned int status;
94
95         /* Max # of queue pairs supported by the device */
96         u16 max_queue_pairs;
97
98         /* # of queue pairs currently used by the driver */
99         u16 curr_queue_pairs;
100
101         /* I like... big packets and I cannot lie! */
102         bool big_packets;
103
104         /* Host will merge rx buffers for big packets (shake it! shake it!) */
105         bool mergeable_rx_bufs;
106
107         /* Has control virtqueue */
108         bool has_cvq;
109
110         /* enable config space updates */
111         bool config_enable;
112
113         /* Active statistics */
114         struct virtnet_stats __percpu *stats;
115
116         /* Work struct for refilling if we run low on memory. */
117         struct delayed_work refill;
118
119         /* Work struct for config space updates */
120         struct work_struct config_work;
121
122         /* Lock for config space updates */
123         struct mutex config_lock;
124
125         /* Does the affinity hint is set for virtqueues? */
126         bool affinity_hint_set;
127
128         /* Per-cpu variable to show the mapping from CPU to virtqueue */
129         int __percpu *vq_index;
130
131         /* CPU hot plug notifier */
132         struct notifier_block nb;
133 };
134
135 struct skb_vnet_hdr {
136         union {
137                 struct virtio_net_hdr hdr;
138                 struct virtio_net_hdr_mrg_rxbuf mhdr;
139         };
140 };
141
142 struct padded_vnet_hdr {
143         struct virtio_net_hdr hdr;
144         /*
145          * virtio_net_hdr should be in a separated sg buffer because of a
146          * QEMU bug, and data sg buffer shares same page with this header sg.
147          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
148          */
149         char padding[6];
150 };
151
152 /* Converting between virtqueue no. and kernel tx/rx queue no.
153  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
154  */
155 static int vq2txq(struct virtqueue *vq)
156 {
157         return (virtqueue_get_queue_index(vq) - 1) / 2;
158 }
159
160 static int txq2vq(int txq)
161 {
162         return txq * 2 + 1;
163 }
164
165 static int vq2rxq(struct virtqueue *vq)
166 {
167         return virtqueue_get_queue_index(vq) / 2;
168 }
169
170 static int rxq2vq(int rxq)
171 {
172         return rxq * 2;
173 }
174
175 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
176 {
177         return (struct skb_vnet_hdr *)skb->cb;
178 }
179
180 /*
181  * private is used to chain pages for big packets, put the whole
182  * most recent used list in the beginning for reuse
183  */
184 static void give_pages(struct receive_queue *rq, struct page *page)
185 {
186         struct page *end;
187
188         /* Find end of list, sew whole thing into vi->rq.pages. */
189         for (end = page; end->private; end = (struct page *)end->private);
190         end->private = (unsigned long)rq->pages;
191         rq->pages = page;
192 }
193
194 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
195 {
196         struct page *p = rq->pages;
197
198         if (p) {
199                 rq->pages = (struct page *)p->private;
200                 /* clear private here, it is used to chain pages */
201                 p->private = 0;
202         } else
203                 p = alloc_page(gfp_mask);
204         return p;
205 }
206
207 static void skb_xmit_done(struct virtqueue *vq)
208 {
209         struct virtnet_info *vi = vq->vdev->priv;
210
211         /* Suppress further interrupts. */
212         virtqueue_disable_cb(vq);
213
214         /* We were probably waiting for more output buffers. */
215         netif_wake_subqueue(vi->dev, vq2txq(vq));
216 }
217
218 static void set_skb_frag(struct sk_buff *skb, struct page *page,
219                          unsigned int offset, unsigned int *len)
220 {
221         int size = min((unsigned)PAGE_SIZE - offset, *len);
222         int i = skb_shinfo(skb)->nr_frags;
223
224         __skb_fill_page_desc(skb, i, page, offset, size);
225
226         skb->data_len += size;
227         skb->len += size;
228         skb->truesize += PAGE_SIZE;
229         skb_shinfo(skb)->nr_frags++;
230         *len -= size;
231 }
232
233 /* Called from bottom half context */
234 static struct sk_buff *page_to_skb(struct receive_queue *rq,
235                                    struct page *page, unsigned int len)
236 {
237         struct virtnet_info *vi = rq->vq->vdev->priv;
238         struct sk_buff *skb;
239         struct skb_vnet_hdr *hdr;
240         unsigned int copy, hdr_len, offset;
241         char *p;
242
243         p = page_address(page);
244
245         /* copy small packet so we can reuse these pages for small data */
246         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
247         if (unlikely(!skb))
248                 return NULL;
249
250         hdr = skb_vnet_hdr(skb);
251
252         if (vi->mergeable_rx_bufs) {
253                 hdr_len = sizeof hdr->mhdr;
254                 offset = hdr_len;
255         } else {
256                 hdr_len = sizeof hdr->hdr;
257                 offset = sizeof(struct padded_vnet_hdr);
258         }
259
260         memcpy(hdr, p, hdr_len);
261
262         len -= hdr_len;
263         p += offset;
264
265         copy = len;
266         if (copy > skb_tailroom(skb))
267                 copy = skb_tailroom(skb);
268         memcpy(skb_put(skb, copy), p, copy);
269
270         len -= copy;
271         offset += copy;
272
273         /*
274          * Verify that we can indeed put this data into a skb.
275          * This is here to handle cases when the device erroneously
276          * tries to receive more than is possible. This is usually
277          * the case of a broken device.
278          */
279         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
280                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
281                 dev_kfree_skb(skb);
282                 return NULL;
283         }
284
285         while (len) {
286                 set_skb_frag(skb, page, offset, &len);
287                 page = (struct page *)page->private;
288                 offset = 0;
289         }
290
291         if (page)
292                 give_pages(rq, page);
293
294         return skb;
295 }
296
297 static int receive_mergeable(struct receive_queue *rq, struct sk_buff *skb)
298 {
299         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
300         struct page *page;
301         int num_buf, i, len;
302
303         num_buf = hdr->mhdr.num_buffers;
304         while (--num_buf) {
305                 i = skb_shinfo(skb)->nr_frags;
306                 if (i >= MAX_SKB_FRAGS) {
307                         pr_debug("%s: packet too long\n", skb->dev->name);
308                         skb->dev->stats.rx_length_errors++;
309                         return -EINVAL;
310                 }
311                 page = virtqueue_get_buf(rq->vq, &len);
312                 if (!page) {
313                         pr_debug("%s: rx error: %d buffers missing\n",
314                                  skb->dev->name, hdr->mhdr.num_buffers);
315                         skb->dev->stats.rx_length_errors++;
316                         return -EINVAL;
317                 }
318
319                 if (len > PAGE_SIZE)
320                         len = PAGE_SIZE;
321
322                 set_skb_frag(skb, page, 0, &len);
323
324                 --rq->num;
325         }
326         return 0;
327 }
328
329 static void receive_buf(struct receive_queue *rq, void *buf, unsigned int len)
330 {
331         struct virtnet_info *vi = rq->vq->vdev->priv;
332         struct net_device *dev = vi->dev;
333         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
334         struct sk_buff *skb;
335         struct page *page;
336         struct skb_vnet_hdr *hdr;
337
338         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
339                 pr_debug("%s: short packet %i\n", dev->name, len);
340                 dev->stats.rx_length_errors++;
341                 if (vi->mergeable_rx_bufs || vi->big_packets)
342                         give_pages(rq, buf);
343                 else
344                         dev_kfree_skb(buf);
345                 return;
346         }
347
348         if (!vi->mergeable_rx_bufs && !vi->big_packets) {
349                 skb = buf;
350                 len -= sizeof(struct virtio_net_hdr);
351                 skb_trim(skb, len);
352         } else {
353                 page = buf;
354                 skb = page_to_skb(rq, page, len);
355                 if (unlikely(!skb)) {
356                         dev->stats.rx_dropped++;
357                         give_pages(rq, page);
358                         return;
359                 }
360                 if (vi->mergeable_rx_bufs)
361                         if (receive_mergeable(rq, skb)) {
362                                 dev_kfree_skb(skb);
363                                 return;
364                         }
365         }
366
367         hdr = skb_vnet_hdr(skb);
368
369         u64_stats_update_begin(&stats->rx_syncp);
370         stats->rx_bytes += skb->len;
371         stats->rx_packets++;
372         u64_stats_update_end(&stats->rx_syncp);
373
374         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
375                 pr_debug("Needs csum!\n");
376                 if (!skb_partial_csum_set(skb,
377                                           hdr->hdr.csum_start,
378                                           hdr->hdr.csum_offset))
379                         goto frame_err;
380         } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
381                 skb->ip_summed = CHECKSUM_UNNECESSARY;
382         }
383
384         skb->protocol = eth_type_trans(skb, dev);
385         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
386                  ntohs(skb->protocol), skb->len, skb->pkt_type);
387
388         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
389                 pr_debug("GSO!\n");
390                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
391                 case VIRTIO_NET_HDR_GSO_TCPV4:
392                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
393                         break;
394                 case VIRTIO_NET_HDR_GSO_UDP:
395                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
396                         break;
397                 case VIRTIO_NET_HDR_GSO_TCPV6:
398                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
399                         break;
400                 default:
401                         net_warn_ratelimited("%s: bad gso type %u.\n",
402                                              dev->name, hdr->hdr.gso_type);
403                         goto frame_err;
404                 }
405
406                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
407                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
408
409                 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
410                 if (skb_shinfo(skb)->gso_size == 0) {
411                         net_warn_ratelimited("%s: zero gso size.\n", dev->name);
412                         goto frame_err;
413                 }
414
415                 /* Header must be checked, and gso_segs computed. */
416                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
417                 skb_shinfo(skb)->gso_segs = 0;
418         }
419
420         netif_receive_skb(skb);
421         return;
422
423 frame_err:
424         dev->stats.rx_frame_errors++;
425         dev_kfree_skb(skb);
426 }
427
428 static int add_recvbuf_small(struct receive_queue *rq, gfp_t gfp)
429 {
430         struct virtnet_info *vi = rq->vq->vdev->priv;
431         struct sk_buff *skb;
432         struct skb_vnet_hdr *hdr;
433         int err;
434
435         skb = __netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN, gfp);
436         if (unlikely(!skb))
437                 return -ENOMEM;
438
439         skb_put(skb, MAX_PACKET_LEN);
440
441         hdr = skb_vnet_hdr(skb);
442         sg_set_buf(rq->sg, &hdr->hdr, sizeof hdr->hdr);
443
444         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
445
446         err = virtqueue_add_buf(rq->vq, rq->sg, 0, 2, skb, gfp);
447         if (err < 0)
448                 dev_kfree_skb(skb);
449
450         return err;
451 }
452
453 static int add_recvbuf_big(struct receive_queue *rq, gfp_t gfp)
454 {
455         struct page *first, *list = NULL;
456         char *p;
457         int i, err, offset;
458
459         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
460         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
461                 first = get_a_page(rq, gfp);
462                 if (!first) {
463                         if (list)
464                                 give_pages(rq, list);
465                         return -ENOMEM;
466                 }
467                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
468
469                 /* chain new page in list head to match sg */
470                 first->private = (unsigned long)list;
471                 list = first;
472         }
473
474         first = get_a_page(rq, gfp);
475         if (!first) {
476                 give_pages(rq, list);
477                 return -ENOMEM;
478         }
479         p = page_address(first);
480
481         /* rq->sg[0], rq->sg[1] share the same page */
482         /* a separated rq->sg[0] for virtio_net_hdr only due to QEMU bug */
483         sg_set_buf(&rq->sg[0], p, sizeof(struct virtio_net_hdr));
484
485         /* rq->sg[1] for data packet, from offset */
486         offset = sizeof(struct padded_vnet_hdr);
487         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
488
489         /* chain first in list head */
490         first->private = (unsigned long)list;
491         err = virtqueue_add_buf(rq->vq, rq->sg, 0, MAX_SKB_FRAGS + 2,
492                                 first, gfp);
493         if (err < 0)
494                 give_pages(rq, first);
495
496         return err;
497 }
498
499 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
500 {
501         struct page *page;
502         int err;
503
504         page = get_a_page(rq, gfp);
505         if (!page)
506                 return -ENOMEM;
507
508         sg_init_one(rq->sg, page_address(page), PAGE_SIZE);
509
510         err = virtqueue_add_buf(rq->vq, rq->sg, 0, 1, page, gfp);
511         if (err < 0)
512                 give_pages(rq, page);
513
514         return err;
515 }
516
517 /*
518  * Returns false if we couldn't fill entirely (OOM).
519  *
520  * Normally run in the receive path, but can also be run from ndo_open
521  * before we're receiving packets, or from refill_work which is
522  * careful to disable receiving (using napi_disable).
523  */
524 static bool try_fill_recv(struct receive_queue *rq, gfp_t gfp)
525 {
526         struct virtnet_info *vi = rq->vq->vdev->priv;
527         int err;
528         bool oom;
529
530         do {
531                 if (vi->mergeable_rx_bufs)
532                         err = add_recvbuf_mergeable(rq, gfp);
533                 else if (vi->big_packets)
534                         err = add_recvbuf_big(rq, gfp);
535                 else
536                         err = add_recvbuf_small(rq, gfp);
537
538                 oom = err == -ENOMEM;
539                 if (err)
540                         break;
541                 ++rq->num;
542         } while (rq->vq->num_free);
543         if (unlikely(rq->num > rq->max))
544                 rq->max = rq->num;
545         virtqueue_kick(rq->vq);
546         return !oom;
547 }
548
549 static void skb_recv_done(struct virtqueue *rvq)
550 {
551         struct virtnet_info *vi = rvq->vdev->priv;
552         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
553
554         /* Schedule NAPI, Suppress further interrupts if successful. */
555         if (napi_schedule_prep(&rq->napi)) {
556                 virtqueue_disable_cb(rvq);
557                 __napi_schedule(&rq->napi);
558         }
559 }
560
561 static void virtnet_napi_enable(struct receive_queue *rq)
562 {
563         napi_enable(&rq->napi);
564
565         /* If all buffers were filled by other side before we napi_enabled, we
566          * won't get another interrupt, so process any outstanding packets
567          * now.  virtnet_poll wants re-enable the queue, so we disable here.
568          * We synchronize against interrupts via NAPI_STATE_SCHED */
569         if (napi_schedule_prep(&rq->napi)) {
570                 virtqueue_disable_cb(rq->vq);
571                 local_bh_disable();
572                 __napi_schedule(&rq->napi);
573                 local_bh_enable();
574         }
575 }
576
577 static void refill_work(struct work_struct *work)
578 {
579         struct virtnet_info *vi =
580                 container_of(work, struct virtnet_info, refill.work);
581         bool still_empty;
582         int i;
583
584         for (i = 0; i < vi->max_queue_pairs; i++) {
585                 struct receive_queue *rq = &vi->rq[i];
586
587                 napi_disable(&rq->napi);
588                 still_empty = !try_fill_recv(rq, GFP_KERNEL);
589                 virtnet_napi_enable(rq);
590
591                 /* In theory, this can happen: if we don't get any buffers in
592                  * we will *never* try to fill again.
593                  */
594                 if (still_empty)
595                         schedule_delayed_work(&vi->refill, HZ/2);
596         }
597 }
598
599 static int virtnet_poll(struct napi_struct *napi, int budget)
600 {
601         struct receive_queue *rq =
602                 container_of(napi, struct receive_queue, napi);
603         struct virtnet_info *vi = rq->vq->vdev->priv;
604         void *buf;
605         unsigned int len, received = 0;
606
607 again:
608         while (received < budget &&
609                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
610                 receive_buf(rq, buf, len);
611                 --rq->num;
612                 received++;
613         }
614
615         if (rq->num < rq->max / 2) {
616                 if (!try_fill_recv(rq, GFP_ATOMIC))
617                         schedule_delayed_work(&vi->refill, 0);
618         }
619
620         /* Out of packets? */
621         if (received < budget) {
622                 napi_complete(napi);
623                 if (unlikely(!virtqueue_enable_cb(rq->vq)) &&
624                     napi_schedule_prep(napi)) {
625                         virtqueue_disable_cb(rq->vq);
626                         __napi_schedule(napi);
627                         goto again;
628                 }
629         }
630
631         return received;
632 }
633
634 static int virtnet_open(struct net_device *dev)
635 {
636         struct virtnet_info *vi = netdev_priv(dev);
637         int i;
638
639         for (i = 0; i < vi->max_queue_pairs; i++) {
640                 /* Make sure we have some buffers: if oom use wq. */
641                 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
642                         schedule_delayed_work(&vi->refill, 0);
643                 virtnet_napi_enable(&vi->rq[i]);
644         }
645
646         return 0;
647 }
648
649 static void free_old_xmit_skbs(struct send_queue *sq)
650 {
651         struct sk_buff *skb;
652         unsigned int len;
653         struct virtnet_info *vi = sq->vq->vdev->priv;
654         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
655
656         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
657                 pr_debug("Sent skb %p\n", skb);
658
659                 u64_stats_update_begin(&stats->tx_syncp);
660                 stats->tx_bytes += skb->len;
661                 stats->tx_packets++;
662                 u64_stats_update_end(&stats->tx_syncp);
663
664                 dev_kfree_skb_any(skb);
665         }
666 }
667
668 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
669 {
670         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
671         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
672         struct virtnet_info *vi = sq->vq->vdev->priv;
673         unsigned num_sg;
674
675         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
676
677         if (skb->ip_summed == CHECKSUM_PARTIAL) {
678                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
679                 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
680                 hdr->hdr.csum_offset = skb->csum_offset;
681         } else {
682                 hdr->hdr.flags = 0;
683                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
684         }
685
686         if (skb_is_gso(skb)) {
687                 hdr->hdr.hdr_len = skb_headlen(skb);
688                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
689                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
690                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
691                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
692                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
693                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
694                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
695                 else
696                         BUG();
697                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
698                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
699         } else {
700                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
701                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
702         }
703
704         hdr->mhdr.num_buffers = 0;
705
706         /* Encode metadata header at front. */
707         if (vi->mergeable_rx_bufs)
708                 sg_set_buf(sq->sg, &hdr->mhdr, sizeof hdr->mhdr);
709         else
710                 sg_set_buf(sq->sg, &hdr->hdr, sizeof hdr->hdr);
711
712         num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
713         return virtqueue_add_buf(sq->vq, sq->sg, num_sg,
714                                  0, skb, GFP_ATOMIC);
715 }
716
717 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
718 {
719         struct virtnet_info *vi = netdev_priv(dev);
720         int qnum = skb_get_queue_mapping(skb);
721         struct send_queue *sq = &vi->sq[qnum];
722         int err;
723
724         /* Free up any pending old buffers before queueing new ones. */
725         free_old_xmit_skbs(sq);
726
727         /* Try to transmit */
728         err = xmit_skb(sq, skb);
729
730         /* This should not happen! */
731         if (unlikely(err)) {
732                 dev->stats.tx_fifo_errors++;
733                 if (net_ratelimit())
734                         dev_warn(&dev->dev,
735                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
736                 dev->stats.tx_dropped++;
737                 kfree_skb(skb);
738                 return NETDEV_TX_OK;
739         }
740         virtqueue_kick(sq->vq);
741
742         /* Don't wait up for transmitted skbs to be freed. */
743         skb_orphan(skb);
744         nf_reset(skb);
745
746         /* Apparently nice girls don't return TX_BUSY; stop the queue
747          * before it gets out of hand.  Naturally, this wastes entries. */
748         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
749                 netif_stop_subqueue(dev, qnum);
750                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
751                         /* More just got used, free them then recheck. */
752                         free_old_xmit_skbs(sq);
753                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
754                                 netif_start_subqueue(dev, qnum);
755                                 virtqueue_disable_cb(sq->vq);
756                         }
757                 }
758         }
759
760         return NETDEV_TX_OK;
761 }
762
763 static int virtnet_set_mac_address(struct net_device *dev, void *p)
764 {
765         struct virtnet_info *vi = netdev_priv(dev);
766         struct virtio_device *vdev = vi->vdev;
767         int ret;
768
769         ret = eth_mac_addr(dev, p);
770         if (ret)
771                 return ret;
772
773         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
774                 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
775                                   dev->dev_addr, dev->addr_len);
776
777         return 0;
778 }
779
780 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
781                                                struct rtnl_link_stats64 *tot)
782 {
783         struct virtnet_info *vi = netdev_priv(dev);
784         int cpu;
785         unsigned int start;
786
787         for_each_possible_cpu(cpu) {
788                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
789                 u64 tpackets, tbytes, rpackets, rbytes;
790
791                 do {
792                         start = u64_stats_fetch_begin_bh(&stats->tx_syncp);
793                         tpackets = stats->tx_packets;
794                         tbytes   = stats->tx_bytes;
795                 } while (u64_stats_fetch_retry_bh(&stats->tx_syncp, start));
796
797                 do {
798                         start = u64_stats_fetch_begin_bh(&stats->rx_syncp);
799                         rpackets = stats->rx_packets;
800                         rbytes   = stats->rx_bytes;
801                 } while (u64_stats_fetch_retry_bh(&stats->rx_syncp, start));
802
803                 tot->rx_packets += rpackets;
804                 tot->tx_packets += tpackets;
805                 tot->rx_bytes   += rbytes;
806                 tot->tx_bytes   += tbytes;
807         }
808
809         tot->tx_dropped = dev->stats.tx_dropped;
810         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
811         tot->rx_dropped = dev->stats.rx_dropped;
812         tot->rx_length_errors = dev->stats.rx_length_errors;
813         tot->rx_frame_errors = dev->stats.rx_frame_errors;
814
815         return tot;
816 }
817
818 #ifdef CONFIG_NET_POLL_CONTROLLER
819 static void virtnet_netpoll(struct net_device *dev)
820 {
821         struct virtnet_info *vi = netdev_priv(dev);
822         int i;
823
824         for (i = 0; i < vi->curr_queue_pairs; i++)
825                 napi_schedule(&vi->rq[i].napi);
826 }
827 #endif
828
829 /*
830  * Send command via the control virtqueue and check status.  Commands
831  * supported by the hypervisor, as indicated by feature bits, should
832  * never fail unless improperly formated.
833  */
834 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
835                                  struct scatterlist *data, int out, int in)
836 {
837         struct scatterlist *s, sg[VIRTNET_SEND_COMMAND_SG_MAX + 2];
838         struct virtio_net_ctrl_hdr ctrl;
839         virtio_net_ctrl_ack status = ~0;
840         unsigned int tmp;
841         int i;
842
843         /* Caller should know better */
844         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ||
845                 (out + in > VIRTNET_SEND_COMMAND_SG_MAX));
846
847         out++; /* Add header */
848         in++; /* Add return status */
849
850         ctrl.class = class;
851         ctrl.cmd = cmd;
852
853         sg_init_table(sg, out + in);
854
855         sg_set_buf(&sg[0], &ctrl, sizeof(ctrl));
856         for_each_sg(data, s, out + in - 2, i)
857                 sg_set_buf(&sg[i + 1], sg_virt(s), s->length);
858         sg_set_buf(&sg[out + in - 1], &status, sizeof(status));
859
860         BUG_ON(virtqueue_add_buf(vi->cvq, sg, out, in, vi, GFP_ATOMIC) < 0);
861
862         virtqueue_kick(vi->cvq);
863
864         /*
865          * Spin for a response, the kick causes an ioport write, trapping
866          * into the hypervisor, so the request should be handled immediately.
867          */
868         while (!virtqueue_get_buf(vi->cvq, &tmp))
869                 cpu_relax();
870
871         return status == VIRTIO_NET_OK;
872 }
873
874 static void virtnet_ack_link_announce(struct virtnet_info *vi)
875 {
876         rtnl_lock();
877         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
878                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL,
879                                   0, 0))
880                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
881         rtnl_unlock();
882 }
883
884 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
885 {
886         struct scatterlist sg;
887         struct virtio_net_ctrl_mq s;
888         struct net_device *dev = vi->dev;
889
890         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
891                 return 0;
892
893         s.virtqueue_pairs = queue_pairs;
894         sg_init_one(&sg, &s, sizeof(s));
895
896         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
897                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg, 1, 0)){
898                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
899                          queue_pairs);
900                 return -EINVAL;
901         } else
902                 vi->curr_queue_pairs = queue_pairs;
903
904         return 0;
905 }
906
907 static int virtnet_close(struct net_device *dev)
908 {
909         struct virtnet_info *vi = netdev_priv(dev);
910         int i;
911
912         /* Make sure refill_work doesn't re-enable napi! */
913         cancel_delayed_work_sync(&vi->refill);
914
915         for (i = 0; i < vi->max_queue_pairs; i++)
916                 napi_disable(&vi->rq[i].napi);
917
918         return 0;
919 }
920
921 static void virtnet_set_rx_mode(struct net_device *dev)
922 {
923         struct virtnet_info *vi = netdev_priv(dev);
924         struct scatterlist sg[2];
925         u8 promisc, allmulti;
926         struct virtio_net_ctrl_mac *mac_data;
927         struct netdev_hw_addr *ha;
928         int uc_count;
929         int mc_count;
930         void *buf;
931         int i;
932
933         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
934         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
935                 return;
936
937         promisc = ((dev->flags & IFF_PROMISC) != 0);
938         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
939
940         sg_init_one(sg, &promisc, sizeof(promisc));
941
942         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
943                                   VIRTIO_NET_CTRL_RX_PROMISC,
944                                   sg, 1, 0))
945                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
946                          promisc ? "en" : "dis");
947
948         sg_init_one(sg, &allmulti, sizeof(allmulti));
949
950         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
951                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
952                                   sg, 1, 0))
953                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
954                          allmulti ? "en" : "dis");
955
956         uc_count = netdev_uc_count(dev);
957         mc_count = netdev_mc_count(dev);
958         /* MAC filter - use one buffer for both lists */
959         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
960                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
961         mac_data = buf;
962         if (!buf) {
963                 dev_warn(&dev->dev, "No memory for MAC address buffer\n");
964                 return;
965         }
966
967         sg_init_table(sg, 2);
968
969         /* Store the unicast list and count in the front of the buffer */
970         mac_data->entries = uc_count;
971         i = 0;
972         netdev_for_each_uc_addr(ha, dev)
973                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
974
975         sg_set_buf(&sg[0], mac_data,
976                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
977
978         /* multicast list and count fill the end */
979         mac_data = (void *)&mac_data->macs[uc_count][0];
980
981         mac_data->entries = mc_count;
982         i = 0;
983         netdev_for_each_mc_addr(ha, dev)
984                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
985
986         sg_set_buf(&sg[1], mac_data,
987                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
988
989         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
990                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
991                                   sg, 2, 0))
992                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
993
994         kfree(buf);
995 }
996
997 static int virtnet_vlan_rx_add_vid(struct net_device *dev, u16 vid)
998 {
999         struct virtnet_info *vi = netdev_priv(dev);
1000         struct scatterlist sg;
1001
1002         sg_init_one(&sg, &vid, sizeof(vid));
1003
1004         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1005                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, 1, 0))
1006                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1007         return 0;
1008 }
1009
1010 static int virtnet_vlan_rx_kill_vid(struct net_device *dev, u16 vid)
1011 {
1012         struct virtnet_info *vi = netdev_priv(dev);
1013         struct scatterlist sg;
1014
1015         sg_init_one(&sg, &vid, sizeof(vid));
1016
1017         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1018                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, 1, 0))
1019                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1020         return 0;
1021 }
1022
1023 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1024 {
1025         int i;
1026         int cpu;
1027
1028         if (vi->affinity_hint_set) {
1029                 for (i = 0; i < vi->max_queue_pairs; i++) {
1030                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1031                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1032                 }
1033
1034                 vi->affinity_hint_set = false;
1035         }
1036
1037         i = 0;
1038         for_each_online_cpu(cpu) {
1039                 if (cpu == hcpu) {
1040                         *per_cpu_ptr(vi->vq_index, cpu) = -1;
1041                 } else {
1042                         *per_cpu_ptr(vi->vq_index, cpu) =
1043                                 ++i % vi->curr_queue_pairs;
1044                 }
1045         }
1046 }
1047
1048 static void virtnet_set_affinity(struct virtnet_info *vi)
1049 {
1050         int i;
1051         int cpu;
1052
1053         /* In multiqueue mode, when the number of cpu is equal to the number of
1054          * queue pairs, we let the queue pairs to be private to one cpu by
1055          * setting the affinity hint to eliminate the contention.
1056          */
1057         if (vi->curr_queue_pairs == 1 ||
1058             vi->max_queue_pairs != num_online_cpus()) {
1059                 virtnet_clean_affinity(vi, -1);
1060                 return;
1061         }
1062
1063         i = 0;
1064         for_each_online_cpu(cpu) {
1065                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1066                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1067                 *per_cpu_ptr(vi->vq_index, cpu) = i;
1068                 i++;
1069         }
1070
1071         vi->affinity_hint_set = true;
1072 }
1073
1074 static int virtnet_cpu_callback(struct notifier_block *nfb,
1075                                 unsigned long action, void *hcpu)
1076 {
1077         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1078
1079         switch(action & ~CPU_TASKS_FROZEN) {
1080         case CPU_ONLINE:
1081         case CPU_DOWN_FAILED:
1082         case CPU_DEAD:
1083                 virtnet_set_affinity(vi);
1084                 break;
1085         case CPU_DOWN_PREPARE:
1086                 virtnet_clean_affinity(vi, (long)hcpu);
1087                 break;
1088         default:
1089                 break;
1090         }
1091         return NOTIFY_OK;
1092 }
1093
1094 static void virtnet_get_ringparam(struct net_device *dev,
1095                                 struct ethtool_ringparam *ring)
1096 {
1097         struct virtnet_info *vi = netdev_priv(dev);
1098
1099         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1100         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1101         ring->rx_pending = ring->rx_max_pending;
1102         ring->tx_pending = ring->tx_max_pending;
1103 }
1104
1105
1106 static void virtnet_get_drvinfo(struct net_device *dev,
1107                                 struct ethtool_drvinfo *info)
1108 {
1109         struct virtnet_info *vi = netdev_priv(dev);
1110         struct virtio_device *vdev = vi->vdev;
1111
1112         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1113         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1114         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1115
1116 }
1117
1118 /* TODO: Eliminate OOO packets during switching */
1119 static int virtnet_set_channels(struct net_device *dev,
1120                                 struct ethtool_channels *channels)
1121 {
1122         struct virtnet_info *vi = netdev_priv(dev);
1123         u16 queue_pairs = channels->combined_count;
1124         int err;
1125
1126         /* We don't support separate rx/tx channels.
1127          * We don't allow setting 'other' channels.
1128          */
1129         if (channels->rx_count || channels->tx_count || channels->other_count)
1130                 return -EINVAL;
1131
1132         if (queue_pairs > vi->max_queue_pairs)
1133                 return -EINVAL;
1134
1135         get_online_cpus();
1136         err = virtnet_set_queues(vi, queue_pairs);
1137         if (!err) {
1138                 netif_set_real_num_tx_queues(dev, queue_pairs);
1139                 netif_set_real_num_rx_queues(dev, queue_pairs);
1140
1141                 virtnet_set_affinity(vi);
1142         }
1143         put_online_cpus();
1144
1145         return err;
1146 }
1147
1148 static void virtnet_get_channels(struct net_device *dev,
1149                                  struct ethtool_channels *channels)
1150 {
1151         struct virtnet_info *vi = netdev_priv(dev);
1152
1153         channels->combined_count = vi->curr_queue_pairs;
1154         channels->max_combined = vi->max_queue_pairs;
1155         channels->max_other = 0;
1156         channels->rx_count = 0;
1157         channels->tx_count = 0;
1158         channels->other_count = 0;
1159 }
1160
1161 static const struct ethtool_ops virtnet_ethtool_ops = {
1162         .get_drvinfo = virtnet_get_drvinfo,
1163         .get_link = ethtool_op_get_link,
1164         .get_ringparam = virtnet_get_ringparam,
1165         .set_channels = virtnet_set_channels,
1166         .get_channels = virtnet_get_channels,
1167 };
1168
1169 #define MIN_MTU 68
1170 #define MAX_MTU 65535
1171
1172 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1173 {
1174         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1175                 return -EINVAL;
1176         dev->mtu = new_mtu;
1177         return 0;
1178 }
1179
1180 /* To avoid contending a lock hold by a vcpu who would exit to host, select the
1181  * txq based on the processor id.
1182  */
1183 static u16 virtnet_select_queue(struct net_device *dev, struct sk_buff *skb)
1184 {
1185         int txq;
1186         struct virtnet_info *vi = netdev_priv(dev);
1187
1188         if (skb_rx_queue_recorded(skb)) {
1189                 txq = skb_get_rx_queue(skb);
1190         } else {
1191                 txq = *__this_cpu_ptr(vi->vq_index);
1192                 if (txq == -1)
1193                         txq = 0;
1194         }
1195
1196         while (unlikely(txq >= dev->real_num_tx_queues))
1197                 txq -= dev->real_num_tx_queues;
1198
1199         return txq;
1200 }
1201
1202 static const struct net_device_ops virtnet_netdev = {
1203         .ndo_open            = virtnet_open,
1204         .ndo_stop            = virtnet_close,
1205         .ndo_start_xmit      = start_xmit,
1206         .ndo_validate_addr   = eth_validate_addr,
1207         .ndo_set_mac_address = virtnet_set_mac_address,
1208         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1209         .ndo_change_mtu      = virtnet_change_mtu,
1210         .ndo_get_stats64     = virtnet_stats,
1211         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1212         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1213         .ndo_select_queue     = virtnet_select_queue,
1214 #ifdef CONFIG_NET_POLL_CONTROLLER
1215         .ndo_poll_controller = virtnet_netpoll,
1216 #endif
1217 };
1218
1219 static void virtnet_config_changed_work(struct work_struct *work)
1220 {
1221         struct virtnet_info *vi =
1222                 container_of(work, struct virtnet_info, config_work);
1223         u16 v;
1224
1225         mutex_lock(&vi->config_lock);
1226         if (!vi->config_enable)
1227                 goto done;
1228
1229         if (virtio_config_val(vi->vdev, VIRTIO_NET_F_STATUS,
1230                               offsetof(struct virtio_net_config, status),
1231                               &v) < 0)
1232                 goto done;
1233
1234         if (v & VIRTIO_NET_S_ANNOUNCE) {
1235                 netdev_notify_peers(vi->dev);
1236                 virtnet_ack_link_announce(vi);
1237         }
1238
1239         /* Ignore unknown (future) status bits */
1240         v &= VIRTIO_NET_S_LINK_UP;
1241
1242         if (vi->status == v)
1243                 goto done;
1244
1245         vi->status = v;
1246
1247         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1248                 netif_carrier_on(vi->dev);
1249                 netif_tx_wake_all_queues(vi->dev);
1250         } else {
1251                 netif_carrier_off(vi->dev);
1252                 netif_tx_stop_all_queues(vi->dev);
1253         }
1254 done:
1255         mutex_unlock(&vi->config_lock);
1256 }
1257
1258 static void virtnet_config_changed(struct virtio_device *vdev)
1259 {
1260         struct virtnet_info *vi = vdev->priv;
1261
1262         schedule_work(&vi->config_work);
1263 }
1264
1265 static void virtnet_free_queues(struct virtnet_info *vi)
1266 {
1267         kfree(vi->rq);
1268         kfree(vi->sq);
1269 }
1270
1271 static void free_receive_bufs(struct virtnet_info *vi)
1272 {
1273         int i;
1274
1275         for (i = 0; i < vi->max_queue_pairs; i++) {
1276                 while (vi->rq[i].pages)
1277                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1278         }
1279 }
1280
1281 static void free_unused_bufs(struct virtnet_info *vi)
1282 {
1283         void *buf;
1284         int i;
1285
1286         for (i = 0; i < vi->max_queue_pairs; i++) {
1287                 struct virtqueue *vq = vi->sq[i].vq;
1288                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1289                         dev_kfree_skb(buf);
1290         }
1291
1292         for (i = 0; i < vi->max_queue_pairs; i++) {
1293                 struct virtqueue *vq = vi->rq[i].vq;
1294
1295                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1296                         if (vi->mergeable_rx_bufs || vi->big_packets)
1297                                 give_pages(&vi->rq[i], buf);
1298                         else
1299                                 dev_kfree_skb(buf);
1300                         --vi->rq[i].num;
1301                 }
1302                 BUG_ON(vi->rq[i].num != 0);
1303         }
1304 }
1305
1306 static void virtnet_del_vqs(struct virtnet_info *vi)
1307 {
1308         struct virtio_device *vdev = vi->vdev;
1309
1310         virtnet_clean_affinity(vi, -1);
1311
1312         vdev->config->del_vqs(vdev);
1313
1314         virtnet_free_queues(vi);
1315 }
1316
1317 static int virtnet_find_vqs(struct virtnet_info *vi)
1318 {
1319         vq_callback_t **callbacks;
1320         struct virtqueue **vqs;
1321         int ret = -ENOMEM;
1322         int i, total_vqs;
1323         const char **names;
1324
1325         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1326          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1327          * possible control vq.
1328          */
1329         total_vqs = vi->max_queue_pairs * 2 +
1330                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1331
1332         /* Allocate space for find_vqs parameters */
1333         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1334         if (!vqs)
1335                 goto err_vq;
1336         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1337         if (!callbacks)
1338                 goto err_callback;
1339         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1340         if (!names)
1341                 goto err_names;
1342
1343         /* Parameters for control virtqueue, if any */
1344         if (vi->has_cvq) {
1345                 callbacks[total_vqs - 1] = NULL;
1346                 names[total_vqs - 1] = "control";
1347         }
1348
1349         /* Allocate/initialize parameters for send/receive virtqueues */
1350         for (i = 0; i < vi->max_queue_pairs; i++) {
1351                 callbacks[rxq2vq(i)] = skb_recv_done;
1352                 callbacks[txq2vq(i)] = skb_xmit_done;
1353                 sprintf(vi->rq[i].name, "input.%d", i);
1354                 sprintf(vi->sq[i].name, "output.%d", i);
1355                 names[rxq2vq(i)] = vi->rq[i].name;
1356                 names[txq2vq(i)] = vi->sq[i].name;
1357         }
1358
1359         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1360                                          names);
1361         if (ret)
1362                 goto err_find;
1363
1364         if (vi->has_cvq) {
1365                 vi->cvq = vqs[total_vqs - 1];
1366                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1367                         vi->dev->features |= NETIF_F_HW_VLAN_FILTER;
1368         }
1369
1370         for (i = 0; i < vi->max_queue_pairs; i++) {
1371                 vi->rq[i].vq = vqs[rxq2vq(i)];
1372                 vi->sq[i].vq = vqs[txq2vq(i)];
1373         }
1374
1375         kfree(names);
1376         kfree(callbacks);
1377         kfree(vqs);
1378
1379         return 0;
1380
1381 err_find:
1382         kfree(names);
1383 err_names:
1384         kfree(callbacks);
1385 err_callback:
1386         kfree(vqs);
1387 err_vq:
1388         return ret;
1389 }
1390
1391 static int virtnet_alloc_queues(struct virtnet_info *vi)
1392 {
1393         int i;
1394
1395         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1396         if (!vi->sq)
1397                 goto err_sq;
1398         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1399         if (!vi->rq)
1400                 goto err_rq;
1401
1402         INIT_DELAYED_WORK(&vi->refill, refill_work);
1403         for (i = 0; i < vi->max_queue_pairs; i++) {
1404                 vi->rq[i].pages = NULL;
1405                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1406                                napi_weight);
1407
1408                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1409                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1410         }
1411
1412         return 0;
1413
1414 err_rq:
1415         kfree(vi->sq);
1416 err_sq:
1417         return -ENOMEM;
1418 }
1419
1420 static int init_vqs(struct virtnet_info *vi)
1421 {
1422         int ret;
1423
1424         /* Allocate send & receive queues */
1425         ret = virtnet_alloc_queues(vi);
1426         if (ret)
1427                 goto err;
1428
1429         ret = virtnet_find_vqs(vi);
1430         if (ret)
1431                 goto err_free;
1432
1433         get_online_cpus();
1434         virtnet_set_affinity(vi);
1435         put_online_cpus();
1436
1437         return 0;
1438
1439 err_free:
1440         virtnet_free_queues(vi);
1441 err:
1442         return ret;
1443 }
1444
1445 static int virtnet_probe(struct virtio_device *vdev)
1446 {
1447         int i, err;
1448         struct net_device *dev;
1449         struct virtnet_info *vi;
1450         u16 max_queue_pairs;
1451
1452         /* Find if host supports multiqueue virtio_net device */
1453         err = virtio_config_val(vdev, VIRTIO_NET_F_MQ,
1454                                 offsetof(struct virtio_net_config,
1455                                 max_virtqueue_pairs), &max_queue_pairs);
1456
1457         /* We need at least 2 queue's */
1458         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1459             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1460             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1461                 max_queue_pairs = 1;
1462
1463         /* Allocate ourselves a network device with room for our info */
1464         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1465         if (!dev)
1466                 return -ENOMEM;
1467
1468         /* Set up network device as normal. */
1469         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1470         dev->netdev_ops = &virtnet_netdev;
1471         dev->features = NETIF_F_HIGHDMA;
1472
1473         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1474         SET_NETDEV_DEV(dev, &vdev->dev);
1475
1476         /* Do we support "hardware" checksums? */
1477         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1478                 /* This opens up the world of extra features. */
1479                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1480                 if (csum)
1481                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1482
1483                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1484                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1485                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1486                 }
1487                 /* Individual feature bits: what can host handle? */
1488                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1489                         dev->hw_features |= NETIF_F_TSO;
1490                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1491                         dev->hw_features |= NETIF_F_TSO6;
1492                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1493                         dev->hw_features |= NETIF_F_TSO_ECN;
1494                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1495                         dev->hw_features |= NETIF_F_UFO;
1496
1497                 if (gso)
1498                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1499                 /* (!csum && gso) case will be fixed by register_netdev() */
1500         }
1501
1502         /* Configuration may specify what MAC to use.  Otherwise random. */
1503         if (virtio_config_val_len(vdev, VIRTIO_NET_F_MAC,
1504                                   offsetof(struct virtio_net_config, mac),
1505                                   dev->dev_addr, dev->addr_len) < 0)
1506                 eth_hw_addr_random(dev);
1507
1508         /* Set up our device-specific information */
1509         vi = netdev_priv(dev);
1510         vi->dev = dev;
1511         vi->vdev = vdev;
1512         vdev->priv = vi;
1513         vi->stats = alloc_percpu(struct virtnet_stats);
1514         err = -ENOMEM;
1515         if (vi->stats == NULL)
1516                 goto free;
1517
1518         vi->vq_index = alloc_percpu(int);
1519         if (vi->vq_index == NULL)
1520                 goto free_stats;
1521
1522         mutex_init(&vi->config_lock);
1523         vi->config_enable = true;
1524         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1525
1526         /* If we can receive ANY GSO packets, we must allocate large ones. */
1527         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1528             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1529             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1530                 vi->big_packets = true;
1531
1532         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1533                 vi->mergeable_rx_bufs = true;
1534
1535         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1536                 vi->has_cvq = true;
1537
1538         /* Use single tx/rx queue pair as default */
1539         vi->curr_queue_pairs = 1;
1540         vi->max_queue_pairs = max_queue_pairs;
1541
1542         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1543         err = init_vqs(vi);
1544         if (err)
1545                 goto free_index;
1546
1547         netif_set_real_num_tx_queues(dev, 1);
1548         netif_set_real_num_rx_queues(dev, 1);
1549
1550         err = register_netdev(dev);
1551         if (err) {
1552                 pr_debug("virtio_net: registering device failed\n");
1553                 goto free_vqs;
1554         }
1555
1556         /* Last of all, set up some receive buffers. */
1557         for (i = 0; i < vi->max_queue_pairs; i++) {
1558                 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1559
1560                 /* If we didn't even get one input buffer, we're useless. */
1561                 if (vi->rq[i].num == 0) {
1562                         free_unused_bufs(vi);
1563                         err = -ENOMEM;
1564                         goto free_recv_bufs;
1565                 }
1566         }
1567
1568         vi->nb.notifier_call = &virtnet_cpu_callback;
1569         err = register_hotcpu_notifier(&vi->nb);
1570         if (err) {
1571                 pr_debug("virtio_net: registering cpu notifier failed\n");
1572                 goto free_recv_bufs;
1573         }
1574
1575         /* Assume link up if device can't report link status,
1576            otherwise get link status from config. */
1577         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1578                 netif_carrier_off(dev);
1579                 schedule_work(&vi->config_work);
1580         } else {
1581                 vi->status = VIRTIO_NET_S_LINK_UP;
1582                 netif_carrier_on(dev);
1583         }
1584
1585         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1586                  dev->name, max_queue_pairs);
1587
1588         return 0;
1589
1590 free_recv_bufs:
1591         free_receive_bufs(vi);
1592         unregister_netdev(dev);
1593 free_vqs:
1594         cancel_delayed_work_sync(&vi->refill);
1595         virtnet_del_vqs(vi);
1596 free_index:
1597         free_percpu(vi->vq_index);
1598 free_stats:
1599         free_percpu(vi->stats);
1600 free:
1601         free_netdev(dev);
1602         return err;
1603 }
1604
1605 static void remove_vq_common(struct virtnet_info *vi)
1606 {
1607         vi->vdev->config->reset(vi->vdev);
1608
1609         /* Free unused buffers in both send and recv, if any. */
1610         free_unused_bufs(vi);
1611
1612         free_receive_bufs(vi);
1613
1614         virtnet_del_vqs(vi);
1615 }
1616
1617 static void virtnet_remove(struct virtio_device *vdev)
1618 {
1619         struct virtnet_info *vi = vdev->priv;
1620
1621         unregister_hotcpu_notifier(&vi->nb);
1622
1623         /* Prevent config work handler from accessing the device. */
1624         mutex_lock(&vi->config_lock);
1625         vi->config_enable = false;
1626         mutex_unlock(&vi->config_lock);
1627
1628         unregister_netdev(vi->dev);
1629
1630         remove_vq_common(vi);
1631
1632         flush_work(&vi->config_work);
1633
1634         free_percpu(vi->vq_index);
1635         free_percpu(vi->stats);
1636         free_netdev(vi->dev);
1637 }
1638
1639 #ifdef CONFIG_PM
1640 static int virtnet_freeze(struct virtio_device *vdev)
1641 {
1642         struct virtnet_info *vi = vdev->priv;
1643         int i;
1644
1645         /* Prevent config work handler from accessing the device */
1646         mutex_lock(&vi->config_lock);
1647         vi->config_enable = false;
1648         mutex_unlock(&vi->config_lock);
1649
1650         netif_device_detach(vi->dev);
1651         cancel_delayed_work_sync(&vi->refill);
1652
1653         if (netif_running(vi->dev))
1654                 for (i = 0; i < vi->max_queue_pairs; i++) {
1655                         napi_disable(&vi->rq[i].napi);
1656                         netif_napi_del(&vi->rq[i].napi);
1657                 }
1658
1659         remove_vq_common(vi);
1660
1661         flush_work(&vi->config_work);
1662
1663         return 0;
1664 }
1665
1666 static int virtnet_restore(struct virtio_device *vdev)
1667 {
1668         struct virtnet_info *vi = vdev->priv;
1669         int err, i;
1670
1671         err = init_vqs(vi);
1672         if (err)
1673                 return err;
1674
1675         if (netif_running(vi->dev))
1676                 for (i = 0; i < vi->max_queue_pairs; i++)
1677                         virtnet_napi_enable(&vi->rq[i]);
1678
1679         netif_device_attach(vi->dev);
1680
1681         for (i = 0; i < vi->max_queue_pairs; i++)
1682                 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1683                         schedule_delayed_work(&vi->refill, 0);
1684
1685         mutex_lock(&vi->config_lock);
1686         vi->config_enable = true;
1687         mutex_unlock(&vi->config_lock);
1688
1689         virtnet_set_queues(vi, vi->curr_queue_pairs);
1690
1691         return 0;
1692 }
1693 #endif
1694
1695 static struct virtio_device_id id_table[] = {
1696         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1697         { 0 },
1698 };
1699
1700 static unsigned int features[] = {
1701         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1702         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1703         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1704         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1705         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1706         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1707         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1708         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1709 };
1710
1711 static struct virtio_driver virtio_net_driver = {
1712         .feature_table = features,
1713         .feature_table_size = ARRAY_SIZE(features),
1714         .driver.name =  KBUILD_MODNAME,
1715         .driver.owner = THIS_MODULE,
1716         .id_table =     id_table,
1717         .probe =        virtnet_probe,
1718         .remove =       virtnet_remove,
1719         .config_changed = virtnet_config_changed,
1720 #ifdef CONFIG_PM
1721         .freeze =       virtnet_freeze,
1722         .restore =      virtnet_restore,
1723 #endif
1724 };
1725
1726 static int __init init(void)
1727 {
1728         return register_virtio_driver(&virtio_net_driver);
1729 }
1730
1731 static void __exit fini(void)
1732 {
1733         unregister_virtio_driver(&virtio_net_driver);
1734 }
1735 module_init(init);
1736 module_exit(fini);
1737
1738 MODULE_DEVICE_TABLE(virtio, id_table);
1739 MODULE_DESCRIPTION("Virtio network driver");
1740 MODULE_LICENSE("GPL");