virtio_net: fix race in RX VQ processing
[linux-3.10.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 = NAPI_POLL_WEIGHT;
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_DRIVER_VERSION "1.0.0"
43
44 struct virtnet_stats {
45         struct u64_stats_sync tx_syncp;
46         struct u64_stats_sync rx_syncp;
47         u64 tx_bytes;
48         u64 tx_packets;
49
50         u64 rx_bytes;
51         u64 rx_packets;
52 };
53
54 /* Internal representation of a send virtqueue */
55 struct send_queue {
56         /* Virtqueue associated with this send _queue */
57         struct virtqueue *vq;
58
59         /* TX: fragments + linear part + virtio header */
60         struct scatterlist sg[MAX_SKB_FRAGS + 2];
61
62         /* Name of the send queue: output.$index */
63         char name[40];
64 };
65
66 /* Internal representation of a receive virtqueue */
67 struct receive_queue {
68         /* Virtqueue associated with this receive_queue */
69         struct virtqueue *vq;
70
71         struct napi_struct napi;
72
73         /* Number of input buffers, and max we've ever had. */
74         unsigned int num, max;
75
76         /* Chain pages by the private ptr. */
77         struct page *pages;
78
79         /* RX: fragments + linear part + virtio header */
80         struct scatterlist sg[MAX_SKB_FRAGS + 2];
81
82         /* Name of this receive queue: input.$index */
83         char name[40];
84 };
85
86 struct virtnet_info {
87         struct virtio_device *vdev;
88         struct virtqueue *cvq;
89         struct net_device *dev;
90         struct send_queue *sq;
91         struct receive_queue *rq;
92         unsigned int status;
93
94         /* Max # of queue pairs supported by the device */
95         u16 max_queue_pairs;
96
97         /* # of queue pairs currently used by the driver */
98         u16 curr_queue_pairs;
99
100         /* I like... big packets and I cannot lie! */
101         bool big_packets;
102
103         /* Host will merge rx buffers for big packets (shake it! shake it!) */
104         bool mergeable_rx_bufs;
105
106         /* Has control virtqueue */
107         bool has_cvq;
108
109         /* enable config space updates */
110         bool config_enable;
111
112         /* Active statistics */
113         struct virtnet_stats __percpu *stats;
114
115         /* Work struct for refilling if we run low on memory. */
116         struct delayed_work refill;
117
118         /* Work struct for config space updates */
119         struct work_struct config_work;
120
121         /* Lock for config space updates */
122         struct mutex config_lock;
123
124         /* Does the affinity hint is set for virtqueues? */
125         bool affinity_hint_set;
126
127         /* Per-cpu variable to show the mapping from CPU to virtqueue */
128         int __percpu *vq_index;
129
130         /* CPU hot plug notifier */
131         struct notifier_block nb;
132 };
133
134 struct skb_vnet_hdr {
135         union {
136                 struct virtio_net_hdr hdr;
137                 struct virtio_net_hdr_mrg_rxbuf mhdr;
138         };
139 };
140
141 struct padded_vnet_hdr {
142         struct virtio_net_hdr hdr;
143         /*
144          * virtio_net_hdr should be in a separated sg buffer because of a
145          * QEMU bug, and data sg buffer shares same page with this header sg.
146          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
147          */
148         char padding[6];
149 };
150
151 /* Converting between virtqueue no. and kernel tx/rx queue no.
152  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
153  */
154 static int vq2txq(struct virtqueue *vq)
155 {
156         return (vq->index - 1) / 2;
157 }
158
159 static int txq2vq(int txq)
160 {
161         return txq * 2 + 1;
162 }
163
164 static int vq2rxq(struct virtqueue *vq)
165 {
166         return vq->index / 2;
167 }
168
169 static int rxq2vq(int rxq)
170 {
171         return rxq * 2;
172 }
173
174 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
175 {
176         return (struct skb_vnet_hdr *)skb->cb;
177 }
178
179 /*
180  * private is used to chain pages for big packets, put the whole
181  * most recent used list in the beginning for reuse
182  */
183 static void give_pages(struct receive_queue *rq, struct page *page)
184 {
185         struct page *end;
186
187         /* Find end of list, sew whole thing into vi->rq.pages. */
188         for (end = page; end->private; end = (struct page *)end->private);
189         end->private = (unsigned long)rq->pages;
190         rq->pages = page;
191 }
192
193 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
194 {
195         struct page *p = rq->pages;
196
197         if (p) {
198                 rq->pages = (struct page *)p->private;
199                 /* clear private here, it is used to chain pages */
200                 p->private = 0;
201         } else
202                 p = alloc_page(gfp_mask);
203         return p;
204 }
205
206 static void skb_xmit_done(struct virtqueue *vq)
207 {
208         struct virtnet_info *vi = vq->vdev->priv;
209
210         /* Suppress further interrupts. */
211         virtqueue_disable_cb(vq);
212
213         /* We were probably waiting for more output buffers. */
214         netif_wake_subqueue(vi->dev, vq2txq(vq));
215 }
216
217 static void set_skb_frag(struct sk_buff *skb, struct page *page,
218                          unsigned int offset, unsigned int *len)
219 {
220         int size = min((unsigned)PAGE_SIZE - offset, *len);
221         int i = skb_shinfo(skb)->nr_frags;
222
223         __skb_fill_page_desc(skb, i, page, offset, size);
224
225         skb->data_len += size;
226         skb->len += size;
227         skb->truesize += PAGE_SIZE;
228         skb_shinfo(skb)->nr_frags++;
229         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
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_inbuf(rq->vq, rq->sg, 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_inbuf(rq->vq, rq->sg, 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_inbuf(rq->vq, rq->sg, 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->curr_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 r, 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                 r = virtqueue_enable_cb_prepare(rq->vq);
623                 napi_complete(napi);
624                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
625                     napi_schedule_prep(napi)) {
626                         virtqueue_disable_cb(rq->vq);
627                         __napi_schedule(napi);
628                         goto again;
629                 }
630         }
631
632         return received;
633 }
634
635 static int virtnet_open(struct net_device *dev)
636 {
637         struct virtnet_info *vi = netdev_priv(dev);
638         int i;
639
640         for (i = 0; i < vi->max_queue_pairs; i++) {
641                 if (i < vi->curr_queue_pairs)
642                         /* Make sure we have some buffers: if oom use wq. */
643                         if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
644                                 schedule_delayed_work(&vi->refill, 0);
645                 virtnet_napi_enable(&vi->rq[i]);
646         }
647
648         return 0;
649 }
650
651 static void free_old_xmit_skbs(struct send_queue *sq)
652 {
653         struct sk_buff *skb;
654         unsigned int len;
655         struct virtnet_info *vi = sq->vq->vdev->priv;
656         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
657
658         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
659                 pr_debug("Sent skb %p\n", skb);
660
661                 u64_stats_update_begin(&stats->tx_syncp);
662                 stats->tx_bytes += skb->len;
663                 stats->tx_packets++;
664                 u64_stats_update_end(&stats->tx_syncp);
665
666                 dev_kfree_skb_any(skb);
667         }
668 }
669
670 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
671 {
672         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
673         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
674         struct virtnet_info *vi = sq->vq->vdev->priv;
675         unsigned num_sg;
676
677         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
678
679         if (skb->ip_summed == CHECKSUM_PARTIAL) {
680                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
681                 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
682                 hdr->hdr.csum_offset = skb->csum_offset;
683         } else {
684                 hdr->hdr.flags = 0;
685                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
686         }
687
688         if (skb_is_gso(skb)) {
689                 hdr->hdr.hdr_len = skb_headlen(skb);
690                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
691                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
692                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
693                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
694                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
695                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
696                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
697                 else
698                         BUG();
699                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
700                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
701         } else {
702                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
703                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
704         }
705
706         hdr->mhdr.num_buffers = 0;
707
708         /* Encode metadata header at front. */
709         if (vi->mergeable_rx_bufs)
710                 sg_set_buf(sq->sg, &hdr->mhdr, sizeof hdr->mhdr);
711         else
712                 sg_set_buf(sq->sg, &hdr->hdr, sizeof hdr->hdr);
713
714         num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
715         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
716 }
717
718 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
719 {
720         struct virtnet_info *vi = netdev_priv(dev);
721         int qnum = skb_get_queue_mapping(skb);
722         struct send_queue *sq = &vi->sq[qnum];
723         int err;
724
725         /* Free up any pending old buffers before queueing new ones. */
726         free_old_xmit_skbs(sq);
727
728         /* Try to transmit */
729         err = xmit_skb(sq, skb);
730
731         /* This should not happen! */
732         if (unlikely(err)) {
733                 dev->stats.tx_fifo_errors++;
734                 if (net_ratelimit())
735                         dev_warn(&dev->dev,
736                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
737                 dev->stats.tx_dropped++;
738                 kfree_skb(skb);
739                 return NETDEV_TX_OK;
740         }
741         virtqueue_kick(sq->vq);
742
743         /* Don't wait up for transmitted skbs to be freed. */
744         skb_orphan(skb);
745         nf_reset(skb);
746
747         /* Apparently nice girls don't return TX_BUSY; stop the queue
748          * before it gets out of hand.  Naturally, this wastes entries. */
749         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
750                 netif_stop_subqueue(dev, qnum);
751                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
752                         /* More just got used, free them then recheck. */
753                         free_old_xmit_skbs(sq);
754                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
755                                 netif_start_subqueue(dev, qnum);
756                                 virtqueue_disable_cb(sq->vq);
757                         }
758                 }
759         }
760
761         return NETDEV_TX_OK;
762 }
763
764 /*
765  * Send command via the control virtqueue and check status.  Commands
766  * supported by the hypervisor, as indicated by feature bits, should
767  * never fail unless improperly formated.
768  */
769 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
770                                  struct scatterlist *out,
771                                  struct scatterlist *in)
772 {
773         struct scatterlist *sgs[4], hdr, stat;
774         struct virtio_net_ctrl_hdr ctrl;
775         virtio_net_ctrl_ack status = ~0;
776         unsigned out_num = 0, in_num = 0, tmp;
777
778         /* Caller should know better */
779         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
780
781         ctrl.class = class;
782         ctrl.cmd = cmd;
783         /* Add header */
784         sg_init_one(&hdr, &ctrl, sizeof(ctrl));
785         sgs[out_num++] = &hdr;
786
787         if (out)
788                 sgs[out_num++] = out;
789         if (in)
790                 sgs[out_num + in_num++] = in;
791
792         /* Add return status. */
793         sg_init_one(&stat, &status, sizeof(status));
794         sgs[out_num + in_num++] = &stat;
795
796         BUG_ON(out_num + in_num > ARRAY_SIZE(sgs));
797         BUG_ON(virtqueue_add_sgs(vi->cvq, sgs, out_num, in_num, vi, GFP_ATOMIC)
798                < 0);
799
800         virtqueue_kick(vi->cvq);
801
802         /* Spin for a response, the kick causes an ioport write, trapping
803          * into the hypervisor, so the request should be handled immediately.
804          */
805         while (!virtqueue_get_buf(vi->cvq, &tmp))
806                 cpu_relax();
807
808         return status == VIRTIO_NET_OK;
809 }
810
811 static int virtnet_set_mac_address(struct net_device *dev, void *p)
812 {
813         struct virtnet_info *vi = netdev_priv(dev);
814         struct virtio_device *vdev = vi->vdev;
815         int ret;
816         struct sockaddr *addr = p;
817         struct scatterlist sg;
818
819         ret = eth_prepare_mac_addr_change(dev, p);
820         if (ret)
821                 return ret;
822
823         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
824                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
825                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
826                                           VIRTIO_NET_CTRL_MAC_ADDR_SET,
827                                           &sg, NULL)) {
828                         dev_warn(&vdev->dev,
829                                  "Failed to set mac address by vq command.\n");
830                         return -EINVAL;
831                 }
832         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
833                 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
834                                   addr->sa_data, dev->addr_len);
835         }
836
837         eth_commit_mac_addr_change(dev, p);
838
839         return 0;
840 }
841
842 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
843                                                struct rtnl_link_stats64 *tot)
844 {
845         struct virtnet_info *vi = netdev_priv(dev);
846         int cpu;
847         unsigned int start;
848
849         for_each_possible_cpu(cpu) {
850                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
851                 u64 tpackets, tbytes, rpackets, rbytes;
852
853                 do {
854                         start = u64_stats_fetch_begin_bh(&stats->tx_syncp);
855                         tpackets = stats->tx_packets;
856                         tbytes   = stats->tx_bytes;
857                 } while (u64_stats_fetch_retry_bh(&stats->tx_syncp, start));
858
859                 do {
860                         start = u64_stats_fetch_begin_bh(&stats->rx_syncp);
861                         rpackets = stats->rx_packets;
862                         rbytes   = stats->rx_bytes;
863                 } while (u64_stats_fetch_retry_bh(&stats->rx_syncp, start));
864
865                 tot->rx_packets += rpackets;
866                 tot->tx_packets += tpackets;
867                 tot->rx_bytes   += rbytes;
868                 tot->tx_bytes   += tbytes;
869         }
870
871         tot->tx_dropped = dev->stats.tx_dropped;
872         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
873         tot->rx_dropped = dev->stats.rx_dropped;
874         tot->rx_length_errors = dev->stats.rx_length_errors;
875         tot->rx_frame_errors = dev->stats.rx_frame_errors;
876
877         return tot;
878 }
879
880 #ifdef CONFIG_NET_POLL_CONTROLLER
881 static void virtnet_netpoll(struct net_device *dev)
882 {
883         struct virtnet_info *vi = netdev_priv(dev);
884         int i;
885
886         for (i = 0; i < vi->curr_queue_pairs; i++)
887                 napi_schedule(&vi->rq[i].napi);
888 }
889 #endif
890
891 static void virtnet_ack_link_announce(struct virtnet_info *vi)
892 {
893         rtnl_lock();
894         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
895                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL, NULL))
896                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
897         rtnl_unlock();
898 }
899
900 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
901 {
902         struct scatterlist sg;
903         struct virtio_net_ctrl_mq s;
904         struct net_device *dev = vi->dev;
905         int i;
906
907         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
908                 return 0;
909
910         s.virtqueue_pairs = queue_pairs;
911         sg_init_one(&sg, &s, sizeof(s));
912
913         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
914                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg, NULL)) {
915                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
916                          queue_pairs);
917                 return -EINVAL;
918         } else {
919                 for (i = vi->curr_queue_pairs; i < queue_pairs; i++)
920                         if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
921                                 schedule_delayed_work(&vi->refill, 0);
922                 vi->curr_queue_pairs = queue_pairs;
923         }
924
925         return 0;
926 }
927
928 static int virtnet_close(struct net_device *dev)
929 {
930         struct virtnet_info *vi = netdev_priv(dev);
931         int i;
932
933         /* Make sure refill_work doesn't re-enable napi! */
934         cancel_delayed_work_sync(&vi->refill);
935
936         for (i = 0; i < vi->max_queue_pairs; i++)
937                 napi_disable(&vi->rq[i].napi);
938
939         return 0;
940 }
941
942 static void virtnet_set_rx_mode(struct net_device *dev)
943 {
944         struct virtnet_info *vi = netdev_priv(dev);
945         struct scatterlist sg[2];
946         u8 promisc, allmulti;
947         struct virtio_net_ctrl_mac *mac_data;
948         struct netdev_hw_addr *ha;
949         int uc_count;
950         int mc_count;
951         void *buf;
952         int i;
953
954         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
955         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
956                 return;
957
958         promisc = ((dev->flags & IFF_PROMISC) != 0);
959         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
960
961         sg_init_one(sg, &promisc, sizeof(promisc));
962
963         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
964                                   VIRTIO_NET_CTRL_RX_PROMISC,
965                                   sg, NULL))
966                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
967                          promisc ? "en" : "dis");
968
969         sg_init_one(sg, &allmulti, sizeof(allmulti));
970
971         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
972                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
973                                   sg, NULL))
974                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
975                          allmulti ? "en" : "dis");
976
977         uc_count = netdev_uc_count(dev);
978         mc_count = netdev_mc_count(dev);
979         /* MAC filter - use one buffer for both lists */
980         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
981                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
982         mac_data = buf;
983         if (!buf)
984                 return;
985
986         sg_init_table(sg, 2);
987
988         /* Store the unicast list and count in the front of the buffer */
989         mac_data->entries = uc_count;
990         i = 0;
991         netdev_for_each_uc_addr(ha, dev)
992                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
993
994         sg_set_buf(&sg[0], mac_data,
995                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
996
997         /* multicast list and count fill the end */
998         mac_data = (void *)&mac_data->macs[uc_count][0];
999
1000         mac_data->entries = mc_count;
1001         i = 0;
1002         netdev_for_each_mc_addr(ha, dev)
1003                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1004
1005         sg_set_buf(&sg[1], mac_data,
1006                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1007
1008         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1009                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
1010                                   sg, NULL))
1011                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
1012
1013         kfree(buf);
1014 }
1015
1016 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1017                                    __be16 proto, u16 vid)
1018 {
1019         struct virtnet_info *vi = netdev_priv(dev);
1020         struct scatterlist sg;
1021
1022         sg_init_one(&sg, &vid, sizeof(vid));
1023
1024         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1025                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, NULL))
1026                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1027         return 0;
1028 }
1029
1030 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1031                                     __be16 proto, u16 vid)
1032 {
1033         struct virtnet_info *vi = netdev_priv(dev);
1034         struct scatterlist sg;
1035
1036         sg_init_one(&sg, &vid, sizeof(vid));
1037
1038         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1039                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, NULL))
1040                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1041         return 0;
1042 }
1043
1044 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1045 {
1046         int i;
1047         int cpu;
1048
1049         if (vi->affinity_hint_set) {
1050                 for (i = 0; i < vi->max_queue_pairs; i++) {
1051                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1052                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1053                 }
1054
1055                 vi->affinity_hint_set = false;
1056         }
1057
1058         i = 0;
1059         for_each_online_cpu(cpu) {
1060                 if (cpu == hcpu) {
1061                         *per_cpu_ptr(vi->vq_index, cpu) = -1;
1062                 } else {
1063                         *per_cpu_ptr(vi->vq_index, cpu) =
1064                                 ++i % vi->curr_queue_pairs;
1065                 }
1066         }
1067 }
1068
1069 static void virtnet_set_affinity(struct virtnet_info *vi)
1070 {
1071         int i;
1072         int cpu;
1073
1074         /* In multiqueue mode, when the number of cpu is equal to the number of
1075          * queue pairs, we let the queue pairs to be private to one cpu by
1076          * setting the affinity hint to eliminate the contention.
1077          */
1078         if (vi->curr_queue_pairs == 1 ||
1079             vi->max_queue_pairs != num_online_cpus()) {
1080                 virtnet_clean_affinity(vi, -1);
1081                 return;
1082         }
1083
1084         i = 0;
1085         for_each_online_cpu(cpu) {
1086                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1087                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1088                 *per_cpu_ptr(vi->vq_index, cpu) = i;
1089                 i++;
1090         }
1091
1092         vi->affinity_hint_set = true;
1093 }
1094
1095 static int virtnet_cpu_callback(struct notifier_block *nfb,
1096                                 unsigned long action, void *hcpu)
1097 {
1098         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1099
1100         switch(action & ~CPU_TASKS_FROZEN) {
1101         case CPU_ONLINE:
1102         case CPU_DOWN_FAILED:
1103         case CPU_DEAD:
1104                 virtnet_set_affinity(vi);
1105                 break;
1106         case CPU_DOWN_PREPARE:
1107                 virtnet_clean_affinity(vi, (long)hcpu);
1108                 break;
1109         default:
1110                 break;
1111         }
1112         return NOTIFY_OK;
1113 }
1114
1115 static void virtnet_get_ringparam(struct net_device *dev,
1116                                 struct ethtool_ringparam *ring)
1117 {
1118         struct virtnet_info *vi = netdev_priv(dev);
1119
1120         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1121         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1122         ring->rx_pending = ring->rx_max_pending;
1123         ring->tx_pending = ring->tx_max_pending;
1124 }
1125
1126
1127 static void virtnet_get_drvinfo(struct net_device *dev,
1128                                 struct ethtool_drvinfo *info)
1129 {
1130         struct virtnet_info *vi = netdev_priv(dev);
1131         struct virtio_device *vdev = vi->vdev;
1132
1133         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1134         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1135         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1136
1137 }
1138
1139 /* TODO: Eliminate OOO packets during switching */
1140 static int virtnet_set_channels(struct net_device *dev,
1141                                 struct ethtool_channels *channels)
1142 {
1143         struct virtnet_info *vi = netdev_priv(dev);
1144         u16 queue_pairs = channels->combined_count;
1145         int err;
1146
1147         /* We don't support separate rx/tx channels.
1148          * We don't allow setting 'other' channels.
1149          */
1150         if (channels->rx_count || channels->tx_count || channels->other_count)
1151                 return -EINVAL;
1152
1153         if (queue_pairs > vi->max_queue_pairs)
1154                 return -EINVAL;
1155
1156         get_online_cpus();
1157         err = virtnet_set_queues(vi, queue_pairs);
1158         if (!err) {
1159                 netif_set_real_num_tx_queues(dev, queue_pairs);
1160                 netif_set_real_num_rx_queues(dev, queue_pairs);
1161
1162                 virtnet_set_affinity(vi);
1163         }
1164         put_online_cpus();
1165
1166         return err;
1167 }
1168
1169 static void virtnet_get_channels(struct net_device *dev,
1170                                  struct ethtool_channels *channels)
1171 {
1172         struct virtnet_info *vi = netdev_priv(dev);
1173
1174         channels->combined_count = vi->curr_queue_pairs;
1175         channels->max_combined = vi->max_queue_pairs;
1176         channels->max_other = 0;
1177         channels->rx_count = 0;
1178         channels->tx_count = 0;
1179         channels->other_count = 0;
1180 }
1181
1182 static const struct ethtool_ops virtnet_ethtool_ops = {
1183         .get_drvinfo = virtnet_get_drvinfo,
1184         .get_link = ethtool_op_get_link,
1185         .get_ringparam = virtnet_get_ringparam,
1186         .set_channels = virtnet_set_channels,
1187         .get_channels = virtnet_get_channels,
1188 };
1189
1190 #define MIN_MTU 68
1191 #define MAX_MTU 65535
1192
1193 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1194 {
1195         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1196                 return -EINVAL;
1197         dev->mtu = new_mtu;
1198         return 0;
1199 }
1200
1201 /* To avoid contending a lock hold by a vcpu who would exit to host, select the
1202  * txq based on the processor id.
1203  */
1204 static u16 virtnet_select_queue(struct net_device *dev, struct sk_buff *skb)
1205 {
1206         int txq;
1207         struct virtnet_info *vi = netdev_priv(dev);
1208
1209         if (skb_rx_queue_recorded(skb)) {
1210                 txq = skb_get_rx_queue(skb);
1211         } else {
1212                 txq = *__this_cpu_ptr(vi->vq_index);
1213                 if (txq == -1)
1214                         txq = 0;
1215         }
1216
1217         while (unlikely(txq >= dev->real_num_tx_queues))
1218                 txq -= dev->real_num_tx_queues;
1219
1220         return txq;
1221 }
1222
1223 static const struct net_device_ops virtnet_netdev = {
1224         .ndo_open            = virtnet_open,
1225         .ndo_stop            = virtnet_close,
1226         .ndo_start_xmit      = start_xmit,
1227         .ndo_validate_addr   = eth_validate_addr,
1228         .ndo_set_mac_address = virtnet_set_mac_address,
1229         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1230         .ndo_change_mtu      = virtnet_change_mtu,
1231         .ndo_get_stats64     = virtnet_stats,
1232         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1233         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1234         .ndo_select_queue     = virtnet_select_queue,
1235 #ifdef CONFIG_NET_POLL_CONTROLLER
1236         .ndo_poll_controller = virtnet_netpoll,
1237 #endif
1238 };
1239
1240 static void virtnet_config_changed_work(struct work_struct *work)
1241 {
1242         struct virtnet_info *vi =
1243                 container_of(work, struct virtnet_info, config_work);
1244         u16 v;
1245
1246         mutex_lock(&vi->config_lock);
1247         if (!vi->config_enable)
1248                 goto done;
1249
1250         if (virtio_config_val(vi->vdev, VIRTIO_NET_F_STATUS,
1251                               offsetof(struct virtio_net_config, status),
1252                               &v) < 0)
1253                 goto done;
1254
1255         if (v & VIRTIO_NET_S_ANNOUNCE) {
1256                 netdev_notify_peers(vi->dev);
1257                 virtnet_ack_link_announce(vi);
1258         }
1259
1260         /* Ignore unknown (future) status bits */
1261         v &= VIRTIO_NET_S_LINK_UP;
1262
1263         if (vi->status == v)
1264                 goto done;
1265
1266         vi->status = v;
1267
1268         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1269                 netif_carrier_on(vi->dev);
1270                 netif_tx_wake_all_queues(vi->dev);
1271         } else {
1272                 netif_carrier_off(vi->dev);
1273                 netif_tx_stop_all_queues(vi->dev);
1274         }
1275 done:
1276         mutex_unlock(&vi->config_lock);
1277 }
1278
1279 static void virtnet_config_changed(struct virtio_device *vdev)
1280 {
1281         struct virtnet_info *vi = vdev->priv;
1282
1283         schedule_work(&vi->config_work);
1284 }
1285
1286 static void virtnet_free_queues(struct virtnet_info *vi)
1287 {
1288         kfree(vi->rq);
1289         kfree(vi->sq);
1290 }
1291
1292 static void free_receive_bufs(struct virtnet_info *vi)
1293 {
1294         int i;
1295
1296         for (i = 0; i < vi->max_queue_pairs; i++) {
1297                 while (vi->rq[i].pages)
1298                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1299         }
1300 }
1301
1302 static void free_unused_bufs(struct virtnet_info *vi)
1303 {
1304         void *buf;
1305         int i;
1306
1307         for (i = 0; i < vi->max_queue_pairs; i++) {
1308                 struct virtqueue *vq = vi->sq[i].vq;
1309                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1310                         dev_kfree_skb(buf);
1311         }
1312
1313         for (i = 0; i < vi->max_queue_pairs; i++) {
1314                 struct virtqueue *vq = vi->rq[i].vq;
1315
1316                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1317                         if (vi->mergeable_rx_bufs || vi->big_packets)
1318                                 give_pages(&vi->rq[i], buf);
1319                         else
1320                                 dev_kfree_skb(buf);
1321                         --vi->rq[i].num;
1322                 }
1323                 BUG_ON(vi->rq[i].num != 0);
1324         }
1325 }
1326
1327 static void virtnet_del_vqs(struct virtnet_info *vi)
1328 {
1329         struct virtio_device *vdev = vi->vdev;
1330
1331         virtnet_clean_affinity(vi, -1);
1332
1333         vdev->config->del_vqs(vdev);
1334
1335         virtnet_free_queues(vi);
1336 }
1337
1338 static int virtnet_find_vqs(struct virtnet_info *vi)
1339 {
1340         vq_callback_t **callbacks;
1341         struct virtqueue **vqs;
1342         int ret = -ENOMEM;
1343         int i, total_vqs;
1344         const char **names;
1345
1346         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1347          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1348          * possible control vq.
1349          */
1350         total_vqs = vi->max_queue_pairs * 2 +
1351                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1352
1353         /* Allocate space for find_vqs parameters */
1354         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1355         if (!vqs)
1356                 goto err_vq;
1357         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1358         if (!callbacks)
1359                 goto err_callback;
1360         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1361         if (!names)
1362                 goto err_names;
1363
1364         /* Parameters for control virtqueue, if any */
1365         if (vi->has_cvq) {
1366                 callbacks[total_vqs - 1] = NULL;
1367                 names[total_vqs - 1] = "control";
1368         }
1369
1370         /* Allocate/initialize parameters for send/receive virtqueues */
1371         for (i = 0; i < vi->max_queue_pairs; i++) {
1372                 callbacks[rxq2vq(i)] = skb_recv_done;
1373                 callbacks[txq2vq(i)] = skb_xmit_done;
1374                 sprintf(vi->rq[i].name, "input.%d", i);
1375                 sprintf(vi->sq[i].name, "output.%d", i);
1376                 names[rxq2vq(i)] = vi->rq[i].name;
1377                 names[txq2vq(i)] = vi->sq[i].name;
1378         }
1379
1380         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1381                                          names);
1382         if (ret)
1383                 goto err_find;
1384
1385         if (vi->has_cvq) {
1386                 vi->cvq = vqs[total_vqs - 1];
1387                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1388                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1389         }
1390
1391         for (i = 0; i < vi->max_queue_pairs; i++) {
1392                 vi->rq[i].vq = vqs[rxq2vq(i)];
1393                 vi->sq[i].vq = vqs[txq2vq(i)];
1394         }
1395
1396         kfree(names);
1397         kfree(callbacks);
1398         kfree(vqs);
1399
1400         return 0;
1401
1402 err_find:
1403         kfree(names);
1404 err_names:
1405         kfree(callbacks);
1406 err_callback:
1407         kfree(vqs);
1408 err_vq:
1409         return ret;
1410 }
1411
1412 static int virtnet_alloc_queues(struct virtnet_info *vi)
1413 {
1414         int i;
1415
1416         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1417         if (!vi->sq)
1418                 goto err_sq;
1419         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1420         if (!vi->rq)
1421                 goto err_rq;
1422
1423         INIT_DELAYED_WORK(&vi->refill, refill_work);
1424         for (i = 0; i < vi->max_queue_pairs; i++) {
1425                 vi->rq[i].pages = NULL;
1426                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1427                                napi_weight);
1428
1429                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1430                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1431         }
1432
1433         return 0;
1434
1435 err_rq:
1436         kfree(vi->sq);
1437 err_sq:
1438         return -ENOMEM;
1439 }
1440
1441 static int init_vqs(struct virtnet_info *vi)
1442 {
1443         int ret;
1444
1445         /* Allocate send & receive queues */
1446         ret = virtnet_alloc_queues(vi);
1447         if (ret)
1448                 goto err;
1449
1450         ret = virtnet_find_vqs(vi);
1451         if (ret)
1452                 goto err_free;
1453
1454         get_online_cpus();
1455         virtnet_set_affinity(vi);
1456         put_online_cpus();
1457
1458         return 0;
1459
1460 err_free:
1461         virtnet_free_queues(vi);
1462 err:
1463         return ret;
1464 }
1465
1466 static int virtnet_probe(struct virtio_device *vdev)
1467 {
1468         int i, err;
1469         struct net_device *dev;
1470         struct virtnet_info *vi;
1471         u16 max_queue_pairs;
1472
1473         /* Find if host supports multiqueue virtio_net device */
1474         err = virtio_config_val(vdev, VIRTIO_NET_F_MQ,
1475                                 offsetof(struct virtio_net_config,
1476                                 max_virtqueue_pairs), &max_queue_pairs);
1477
1478         /* We need at least 2 queue's */
1479         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1480             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1481             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1482                 max_queue_pairs = 1;
1483
1484         /* Allocate ourselves a network device with room for our info */
1485         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1486         if (!dev)
1487                 return -ENOMEM;
1488
1489         /* Set up network device as normal. */
1490         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1491         dev->netdev_ops = &virtnet_netdev;
1492         dev->features = NETIF_F_HIGHDMA;
1493
1494         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1495         SET_NETDEV_DEV(dev, &vdev->dev);
1496
1497         /* Do we support "hardware" checksums? */
1498         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1499                 /* This opens up the world of extra features. */
1500                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1501                 if (csum)
1502                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1503
1504                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1505                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1506                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1507                 }
1508                 /* Individual feature bits: what can host handle? */
1509                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1510                         dev->hw_features |= NETIF_F_TSO;
1511                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1512                         dev->hw_features |= NETIF_F_TSO6;
1513                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1514                         dev->hw_features |= NETIF_F_TSO_ECN;
1515                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1516                         dev->hw_features |= NETIF_F_UFO;
1517
1518                 if (gso)
1519                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1520                 /* (!csum && gso) case will be fixed by register_netdev() */
1521         }
1522
1523         dev->vlan_features = dev->features;
1524
1525         /* Configuration may specify what MAC to use.  Otherwise random. */
1526         if (virtio_config_val_len(vdev, VIRTIO_NET_F_MAC,
1527                                   offsetof(struct virtio_net_config, mac),
1528                                   dev->dev_addr, dev->addr_len) < 0)
1529                 eth_hw_addr_random(dev);
1530
1531         /* Set up our device-specific information */
1532         vi = netdev_priv(dev);
1533         vi->dev = dev;
1534         vi->vdev = vdev;
1535         vdev->priv = vi;
1536         vi->stats = alloc_percpu(struct virtnet_stats);
1537         err = -ENOMEM;
1538         if (vi->stats == NULL)
1539                 goto free;
1540
1541         vi->vq_index = alloc_percpu(int);
1542         if (vi->vq_index == NULL)
1543                 goto free_stats;
1544
1545         mutex_init(&vi->config_lock);
1546         vi->config_enable = true;
1547         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1548
1549         /* If we can receive ANY GSO packets, we must allocate large ones. */
1550         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1551             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1552             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1553                 vi->big_packets = true;
1554
1555         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1556                 vi->mergeable_rx_bufs = true;
1557
1558         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1559                 vi->has_cvq = true;
1560
1561         /* Use single tx/rx queue pair as default */
1562         vi->curr_queue_pairs = 1;
1563         vi->max_queue_pairs = max_queue_pairs;
1564
1565         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1566         err = init_vqs(vi);
1567         if (err)
1568                 goto free_index;
1569
1570         netif_set_real_num_tx_queues(dev, 1);
1571         netif_set_real_num_rx_queues(dev, 1);
1572
1573         err = register_netdev(dev);
1574         if (err) {
1575                 pr_debug("virtio_net: registering device failed\n");
1576                 goto free_vqs;
1577         }
1578
1579         /* Last of all, set up some receive buffers. */
1580         for (i = 0; i < vi->curr_queue_pairs; i++) {
1581                 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1582
1583                 /* If we didn't even get one input buffer, we're useless. */
1584                 if (vi->rq[i].num == 0) {
1585                         free_unused_bufs(vi);
1586                         err = -ENOMEM;
1587                         goto free_recv_bufs;
1588                 }
1589         }
1590
1591         vi->nb.notifier_call = &virtnet_cpu_callback;
1592         err = register_hotcpu_notifier(&vi->nb);
1593         if (err) {
1594                 pr_debug("virtio_net: registering cpu notifier failed\n");
1595                 goto free_recv_bufs;
1596         }
1597
1598         /* Assume link up if device can't report link status,
1599            otherwise get link status from config. */
1600         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1601                 netif_carrier_off(dev);
1602                 schedule_work(&vi->config_work);
1603         } else {
1604                 vi->status = VIRTIO_NET_S_LINK_UP;
1605                 netif_carrier_on(dev);
1606         }
1607
1608         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1609                  dev->name, max_queue_pairs);
1610
1611         return 0;
1612
1613 free_recv_bufs:
1614         free_receive_bufs(vi);
1615         unregister_netdev(dev);
1616 free_vqs:
1617         cancel_delayed_work_sync(&vi->refill);
1618         virtnet_del_vqs(vi);
1619 free_index:
1620         free_percpu(vi->vq_index);
1621 free_stats:
1622         free_percpu(vi->stats);
1623 free:
1624         free_netdev(dev);
1625         return err;
1626 }
1627
1628 static void remove_vq_common(struct virtnet_info *vi)
1629 {
1630         vi->vdev->config->reset(vi->vdev);
1631
1632         /* Free unused buffers in both send and recv, if any. */
1633         free_unused_bufs(vi);
1634
1635         free_receive_bufs(vi);
1636
1637         virtnet_del_vqs(vi);
1638 }
1639
1640 static void virtnet_remove(struct virtio_device *vdev)
1641 {
1642         struct virtnet_info *vi = vdev->priv;
1643
1644         unregister_hotcpu_notifier(&vi->nb);
1645
1646         /* Prevent config work handler from accessing the device. */
1647         mutex_lock(&vi->config_lock);
1648         vi->config_enable = false;
1649         mutex_unlock(&vi->config_lock);
1650
1651         unregister_netdev(vi->dev);
1652
1653         remove_vq_common(vi);
1654
1655         flush_work(&vi->config_work);
1656
1657         free_percpu(vi->vq_index);
1658         free_percpu(vi->stats);
1659         free_netdev(vi->dev);
1660 }
1661
1662 #ifdef CONFIG_PM
1663 static int virtnet_freeze(struct virtio_device *vdev)
1664 {
1665         struct virtnet_info *vi = vdev->priv;
1666         int i;
1667
1668         /* Prevent config work handler from accessing the device */
1669         mutex_lock(&vi->config_lock);
1670         vi->config_enable = false;
1671         mutex_unlock(&vi->config_lock);
1672
1673         netif_device_detach(vi->dev);
1674         cancel_delayed_work_sync(&vi->refill);
1675
1676         if (netif_running(vi->dev))
1677                 for (i = 0; i < vi->max_queue_pairs; i++) {
1678                         napi_disable(&vi->rq[i].napi);
1679                         netif_napi_del(&vi->rq[i].napi);
1680                 }
1681
1682         remove_vq_common(vi);
1683
1684         flush_work(&vi->config_work);
1685
1686         return 0;
1687 }
1688
1689 static int virtnet_restore(struct virtio_device *vdev)
1690 {
1691         struct virtnet_info *vi = vdev->priv;
1692         int err, i;
1693
1694         err = init_vqs(vi);
1695         if (err)
1696                 return err;
1697
1698         if (netif_running(vi->dev))
1699                 for (i = 0; i < vi->max_queue_pairs; i++)
1700                         virtnet_napi_enable(&vi->rq[i]);
1701
1702         netif_device_attach(vi->dev);
1703
1704         for (i = 0; i < vi->curr_queue_pairs; i++)
1705                 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1706                         schedule_delayed_work(&vi->refill, 0);
1707
1708         mutex_lock(&vi->config_lock);
1709         vi->config_enable = true;
1710         mutex_unlock(&vi->config_lock);
1711
1712         virtnet_set_queues(vi, vi->curr_queue_pairs);
1713
1714         return 0;
1715 }
1716 #endif
1717
1718 static struct virtio_device_id id_table[] = {
1719         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1720         { 0 },
1721 };
1722
1723 static unsigned int features[] = {
1724         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1725         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1726         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1727         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1728         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1729         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1730         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1731         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1732         VIRTIO_NET_F_CTRL_MAC_ADDR,
1733 };
1734
1735 static struct virtio_driver virtio_net_driver = {
1736         .feature_table = features,
1737         .feature_table_size = ARRAY_SIZE(features),
1738         .driver.name =  KBUILD_MODNAME,
1739         .driver.owner = THIS_MODULE,
1740         .id_table =     id_table,
1741         .probe =        virtnet_probe,
1742         .remove =       virtnet_remove,
1743         .config_changed = virtnet_config_changed,
1744 #ifdef CONFIG_PM
1745         .freeze =       virtnet_freeze,
1746         .restore =      virtnet_restore,
1747 #endif
1748 };
1749
1750 module_virtio_driver(virtio_net_driver);
1751
1752 MODULE_DEVICE_TABLE(virtio, id_table);
1753 MODULE_DESCRIPTION("Virtio network driver");
1754 MODULE_LICENSE("GPL");