xref: /openbmc/linux/drivers/net/virtio_net.c (revision cdfce539)
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 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->curr_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_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
714 }
715 
716 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
717 {
718 	struct virtnet_info *vi = netdev_priv(dev);
719 	int qnum = skb_get_queue_mapping(skb);
720 	struct send_queue *sq = &vi->sq[qnum];
721 	int err;
722 
723 	/* Free up any pending old buffers before queueing new ones. */
724 	free_old_xmit_skbs(sq);
725 
726 	/* Try to transmit */
727 	err = xmit_skb(sq, skb);
728 
729 	/* This should not happen! */
730 	if (unlikely(err)) {
731 		dev->stats.tx_fifo_errors++;
732 		if (net_ratelimit())
733 			dev_warn(&dev->dev,
734 				 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
735 		dev->stats.tx_dropped++;
736 		kfree_skb(skb);
737 		return NETDEV_TX_OK;
738 	}
739 	virtqueue_kick(sq->vq);
740 
741 	/* Don't wait up for transmitted skbs to be freed. */
742 	skb_orphan(skb);
743 	nf_reset(skb);
744 
745 	/* Apparently nice girls don't return TX_BUSY; stop the queue
746 	 * before it gets out of hand.  Naturally, this wastes entries. */
747 	if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
748 		netif_stop_subqueue(dev, qnum);
749 		if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
750 			/* More just got used, free them then recheck. */
751 			free_old_xmit_skbs(sq);
752 			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
753 				netif_start_subqueue(dev, qnum);
754 				virtqueue_disable_cb(sq->vq);
755 			}
756 		}
757 	}
758 
759 	return NETDEV_TX_OK;
760 }
761 
762 /*
763  * Send command via the control virtqueue and check status.  Commands
764  * supported by the hypervisor, as indicated by feature bits, should
765  * never fail unless improperly formated.
766  */
767 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
768 				 struct scatterlist *out,
769 				 struct scatterlist *in)
770 {
771 	struct scatterlist *sgs[4], hdr, stat;
772 	struct virtio_net_ctrl_hdr ctrl;
773 	virtio_net_ctrl_ack status = ~0;
774 	unsigned out_num = 0, in_num = 0, tmp;
775 
776 	/* Caller should know better */
777 	BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
778 
779 	ctrl.class = class;
780 	ctrl.cmd = cmd;
781 	/* Add header */
782 	sg_init_one(&hdr, &ctrl, sizeof(ctrl));
783 	sgs[out_num++] = &hdr;
784 
785 	if (out)
786 		sgs[out_num++] = out;
787 	if (in)
788 		sgs[out_num + in_num++] = in;
789 
790 	/* Add return status. */
791 	sg_init_one(&stat, &status, sizeof(status));
792 	sgs[out_num + in_num++] = &stat;
793 
794 	BUG_ON(out_num + in_num > ARRAY_SIZE(sgs));
795 	BUG_ON(virtqueue_add_sgs(vi->cvq, sgs, out_num, in_num, vi, GFP_ATOMIC)
796 	       < 0);
797 
798 	virtqueue_kick(vi->cvq);
799 
800 	/* Spin for a response, the kick causes an ioport write, trapping
801 	 * into the hypervisor, so the request should be handled immediately.
802 	 */
803 	while (!virtqueue_get_buf(vi->cvq, &tmp))
804 		cpu_relax();
805 
806 	return status == VIRTIO_NET_OK;
807 }
808 
809 static int virtnet_set_mac_address(struct net_device *dev, void *p)
810 {
811 	struct virtnet_info *vi = netdev_priv(dev);
812 	struct virtio_device *vdev = vi->vdev;
813 	int ret;
814 	struct sockaddr *addr = p;
815 	struct scatterlist sg;
816 
817 	ret = eth_prepare_mac_addr_change(dev, p);
818 	if (ret)
819 		return ret;
820 
821 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
822 		sg_init_one(&sg, addr->sa_data, dev->addr_len);
823 		if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
824 					  VIRTIO_NET_CTRL_MAC_ADDR_SET,
825 					  &sg, NULL)) {
826 			dev_warn(&vdev->dev,
827 				 "Failed to set mac address by vq command.\n");
828 			return -EINVAL;
829 		}
830 	} else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
831 		vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
832 				  addr->sa_data, dev->addr_len);
833 	}
834 
835 	eth_commit_mac_addr_change(dev, p);
836 
837 	return 0;
838 }
839 
840 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
841 					       struct rtnl_link_stats64 *tot)
842 {
843 	struct virtnet_info *vi = netdev_priv(dev);
844 	int cpu;
845 	unsigned int start;
846 
847 	for_each_possible_cpu(cpu) {
848 		struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
849 		u64 tpackets, tbytes, rpackets, rbytes;
850 
851 		do {
852 			start = u64_stats_fetch_begin_bh(&stats->tx_syncp);
853 			tpackets = stats->tx_packets;
854 			tbytes   = stats->tx_bytes;
855 		} while (u64_stats_fetch_retry_bh(&stats->tx_syncp, start));
856 
857 		do {
858 			start = u64_stats_fetch_begin_bh(&stats->rx_syncp);
859 			rpackets = stats->rx_packets;
860 			rbytes   = stats->rx_bytes;
861 		} while (u64_stats_fetch_retry_bh(&stats->rx_syncp, start));
862 
863 		tot->rx_packets += rpackets;
864 		tot->tx_packets += tpackets;
865 		tot->rx_bytes   += rbytes;
866 		tot->tx_bytes   += tbytes;
867 	}
868 
869 	tot->tx_dropped = dev->stats.tx_dropped;
870 	tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
871 	tot->rx_dropped = dev->stats.rx_dropped;
872 	tot->rx_length_errors = dev->stats.rx_length_errors;
873 	tot->rx_frame_errors = dev->stats.rx_frame_errors;
874 
875 	return tot;
876 }
877 
878 #ifdef CONFIG_NET_POLL_CONTROLLER
879 static void virtnet_netpoll(struct net_device *dev)
880 {
881 	struct virtnet_info *vi = netdev_priv(dev);
882 	int i;
883 
884 	for (i = 0; i < vi->curr_queue_pairs; i++)
885 		napi_schedule(&vi->rq[i].napi);
886 }
887 #endif
888 
889 static void virtnet_ack_link_announce(struct virtnet_info *vi)
890 {
891 	rtnl_lock();
892 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
893 				  VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL, NULL))
894 		dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
895 	rtnl_unlock();
896 }
897 
898 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
899 {
900 	struct scatterlist sg;
901 	struct virtio_net_ctrl_mq s;
902 	struct net_device *dev = vi->dev;
903 	int i;
904 
905 	if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
906 		return 0;
907 
908 	s.virtqueue_pairs = queue_pairs;
909 	sg_init_one(&sg, &s, sizeof(s));
910 
911 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
912 				  VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg, NULL)) {
913 		dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
914 			 queue_pairs);
915 		return -EINVAL;
916 	} else {
917 		for (i = vi->curr_queue_pairs; i < queue_pairs; i++)
918 			if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
919 				schedule_delayed_work(&vi->refill, 0);
920 		vi->curr_queue_pairs = queue_pairs;
921 	}
922 
923 	return 0;
924 }
925 
926 static int virtnet_close(struct net_device *dev)
927 {
928 	struct virtnet_info *vi = netdev_priv(dev);
929 	int i;
930 
931 	/* Make sure refill_work doesn't re-enable napi! */
932 	cancel_delayed_work_sync(&vi->refill);
933 
934 	for (i = 0; i < vi->max_queue_pairs; i++)
935 		napi_disable(&vi->rq[i].napi);
936 
937 	return 0;
938 }
939 
940 static void virtnet_set_rx_mode(struct net_device *dev)
941 {
942 	struct virtnet_info *vi = netdev_priv(dev);
943 	struct scatterlist sg[2];
944 	u8 promisc, allmulti;
945 	struct virtio_net_ctrl_mac *mac_data;
946 	struct netdev_hw_addr *ha;
947 	int uc_count;
948 	int mc_count;
949 	void *buf;
950 	int i;
951 
952 	/* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
953 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
954 		return;
955 
956 	promisc = ((dev->flags & IFF_PROMISC) != 0);
957 	allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
958 
959 	sg_init_one(sg, &promisc, sizeof(promisc));
960 
961 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
962 				  VIRTIO_NET_CTRL_RX_PROMISC,
963 				  sg, NULL))
964 		dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
965 			 promisc ? "en" : "dis");
966 
967 	sg_init_one(sg, &allmulti, sizeof(allmulti));
968 
969 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
970 				  VIRTIO_NET_CTRL_RX_ALLMULTI,
971 				  sg, NULL))
972 		dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
973 			 allmulti ? "en" : "dis");
974 
975 	uc_count = netdev_uc_count(dev);
976 	mc_count = netdev_mc_count(dev);
977 	/* MAC filter - use one buffer for both lists */
978 	buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
979 		      (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
980 	mac_data = buf;
981 	if (!buf)
982 		return;
983 
984 	sg_init_table(sg, 2);
985 
986 	/* Store the unicast list and count in the front of the buffer */
987 	mac_data->entries = uc_count;
988 	i = 0;
989 	netdev_for_each_uc_addr(ha, dev)
990 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
991 
992 	sg_set_buf(&sg[0], mac_data,
993 		   sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
994 
995 	/* multicast list and count fill the end */
996 	mac_data = (void *)&mac_data->macs[uc_count][0];
997 
998 	mac_data->entries = mc_count;
999 	i = 0;
1000 	netdev_for_each_mc_addr(ha, dev)
1001 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1002 
1003 	sg_set_buf(&sg[1], mac_data,
1004 		   sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1005 
1006 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1007 				  VIRTIO_NET_CTRL_MAC_TABLE_SET,
1008 				  sg, NULL))
1009 		dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
1010 
1011 	kfree(buf);
1012 }
1013 
1014 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1015 				   __be16 proto, u16 vid)
1016 {
1017 	struct virtnet_info *vi = netdev_priv(dev);
1018 	struct scatterlist sg;
1019 
1020 	sg_init_one(&sg, &vid, sizeof(vid));
1021 
1022 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1023 				  VIRTIO_NET_CTRL_VLAN_ADD, &sg, NULL))
1024 		dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1025 	return 0;
1026 }
1027 
1028 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1029 				    __be16 proto, u16 vid)
1030 {
1031 	struct virtnet_info *vi = netdev_priv(dev);
1032 	struct scatterlist sg;
1033 
1034 	sg_init_one(&sg, &vid, sizeof(vid));
1035 
1036 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1037 				  VIRTIO_NET_CTRL_VLAN_DEL, &sg, NULL))
1038 		dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1039 	return 0;
1040 }
1041 
1042 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1043 {
1044 	int i;
1045 	int cpu;
1046 
1047 	if (vi->affinity_hint_set) {
1048 		for (i = 0; i < vi->max_queue_pairs; i++) {
1049 			virtqueue_set_affinity(vi->rq[i].vq, -1);
1050 			virtqueue_set_affinity(vi->sq[i].vq, -1);
1051 		}
1052 
1053 		vi->affinity_hint_set = false;
1054 	}
1055 
1056 	i = 0;
1057 	for_each_online_cpu(cpu) {
1058 		if (cpu == hcpu) {
1059 			*per_cpu_ptr(vi->vq_index, cpu) = -1;
1060 		} else {
1061 			*per_cpu_ptr(vi->vq_index, cpu) =
1062 				++i % vi->curr_queue_pairs;
1063 		}
1064 	}
1065 }
1066 
1067 static void virtnet_set_affinity(struct virtnet_info *vi)
1068 {
1069 	int i;
1070 	int cpu;
1071 
1072 	/* In multiqueue mode, when the number of cpu is equal to the number of
1073 	 * queue pairs, we let the queue pairs to be private to one cpu by
1074 	 * setting the affinity hint to eliminate the contention.
1075 	 */
1076 	if (vi->curr_queue_pairs == 1 ||
1077 	    vi->max_queue_pairs != num_online_cpus()) {
1078 		virtnet_clean_affinity(vi, -1);
1079 		return;
1080 	}
1081 
1082 	i = 0;
1083 	for_each_online_cpu(cpu) {
1084 		virtqueue_set_affinity(vi->rq[i].vq, cpu);
1085 		virtqueue_set_affinity(vi->sq[i].vq, cpu);
1086 		*per_cpu_ptr(vi->vq_index, cpu) = i;
1087 		i++;
1088 	}
1089 
1090 	vi->affinity_hint_set = true;
1091 }
1092 
1093 static int virtnet_cpu_callback(struct notifier_block *nfb,
1094 			        unsigned long action, void *hcpu)
1095 {
1096 	struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1097 
1098 	switch(action & ~CPU_TASKS_FROZEN) {
1099 	case CPU_ONLINE:
1100 	case CPU_DOWN_FAILED:
1101 	case CPU_DEAD:
1102 		virtnet_set_affinity(vi);
1103 		break;
1104 	case CPU_DOWN_PREPARE:
1105 		virtnet_clean_affinity(vi, (long)hcpu);
1106 		break;
1107 	default:
1108 		break;
1109 	}
1110 	return NOTIFY_OK;
1111 }
1112 
1113 static void virtnet_get_ringparam(struct net_device *dev,
1114 				struct ethtool_ringparam *ring)
1115 {
1116 	struct virtnet_info *vi = netdev_priv(dev);
1117 
1118 	ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1119 	ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1120 	ring->rx_pending = ring->rx_max_pending;
1121 	ring->tx_pending = ring->tx_max_pending;
1122 }
1123 
1124 
1125 static void virtnet_get_drvinfo(struct net_device *dev,
1126 				struct ethtool_drvinfo *info)
1127 {
1128 	struct virtnet_info *vi = netdev_priv(dev);
1129 	struct virtio_device *vdev = vi->vdev;
1130 
1131 	strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1132 	strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1133 	strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1134 
1135 }
1136 
1137 /* TODO: Eliminate OOO packets during switching */
1138 static int virtnet_set_channels(struct net_device *dev,
1139 				struct ethtool_channels *channels)
1140 {
1141 	struct virtnet_info *vi = netdev_priv(dev);
1142 	u16 queue_pairs = channels->combined_count;
1143 	int err;
1144 
1145 	/* We don't support separate rx/tx channels.
1146 	 * We don't allow setting 'other' channels.
1147 	 */
1148 	if (channels->rx_count || channels->tx_count || channels->other_count)
1149 		return -EINVAL;
1150 
1151 	if (queue_pairs > vi->max_queue_pairs)
1152 		return -EINVAL;
1153 
1154 	get_online_cpus();
1155 	err = virtnet_set_queues(vi, queue_pairs);
1156 	if (!err) {
1157 		netif_set_real_num_tx_queues(dev, queue_pairs);
1158 		netif_set_real_num_rx_queues(dev, queue_pairs);
1159 
1160 		virtnet_set_affinity(vi);
1161 	}
1162 	put_online_cpus();
1163 
1164 	return err;
1165 }
1166 
1167 static void virtnet_get_channels(struct net_device *dev,
1168 				 struct ethtool_channels *channels)
1169 {
1170 	struct virtnet_info *vi = netdev_priv(dev);
1171 
1172 	channels->combined_count = vi->curr_queue_pairs;
1173 	channels->max_combined = vi->max_queue_pairs;
1174 	channels->max_other = 0;
1175 	channels->rx_count = 0;
1176 	channels->tx_count = 0;
1177 	channels->other_count = 0;
1178 }
1179 
1180 static const struct ethtool_ops virtnet_ethtool_ops = {
1181 	.get_drvinfo = virtnet_get_drvinfo,
1182 	.get_link = ethtool_op_get_link,
1183 	.get_ringparam = virtnet_get_ringparam,
1184 	.set_channels = virtnet_set_channels,
1185 	.get_channels = virtnet_get_channels,
1186 };
1187 
1188 #define MIN_MTU 68
1189 #define MAX_MTU 65535
1190 
1191 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1192 {
1193 	if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1194 		return -EINVAL;
1195 	dev->mtu = new_mtu;
1196 	return 0;
1197 }
1198 
1199 /* To avoid contending a lock hold by a vcpu who would exit to host, select the
1200  * txq based on the processor id.
1201  */
1202 static u16 virtnet_select_queue(struct net_device *dev, struct sk_buff *skb)
1203 {
1204 	int txq;
1205 	struct virtnet_info *vi = netdev_priv(dev);
1206 
1207 	if (skb_rx_queue_recorded(skb)) {
1208 		txq = skb_get_rx_queue(skb);
1209 	} else {
1210 		txq = *__this_cpu_ptr(vi->vq_index);
1211 		if (txq == -1)
1212 			txq = 0;
1213 	}
1214 
1215 	while (unlikely(txq >= dev->real_num_tx_queues))
1216 		txq -= dev->real_num_tx_queues;
1217 
1218 	return txq;
1219 }
1220 
1221 static const struct net_device_ops virtnet_netdev = {
1222 	.ndo_open            = virtnet_open,
1223 	.ndo_stop   	     = virtnet_close,
1224 	.ndo_start_xmit      = start_xmit,
1225 	.ndo_validate_addr   = eth_validate_addr,
1226 	.ndo_set_mac_address = virtnet_set_mac_address,
1227 	.ndo_set_rx_mode     = virtnet_set_rx_mode,
1228 	.ndo_change_mtu	     = virtnet_change_mtu,
1229 	.ndo_get_stats64     = virtnet_stats,
1230 	.ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1231 	.ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1232 	.ndo_select_queue     = virtnet_select_queue,
1233 #ifdef CONFIG_NET_POLL_CONTROLLER
1234 	.ndo_poll_controller = virtnet_netpoll,
1235 #endif
1236 };
1237 
1238 static void virtnet_config_changed_work(struct work_struct *work)
1239 {
1240 	struct virtnet_info *vi =
1241 		container_of(work, struct virtnet_info, config_work);
1242 	u16 v;
1243 
1244 	mutex_lock(&vi->config_lock);
1245 	if (!vi->config_enable)
1246 		goto done;
1247 
1248 	if (virtio_config_val(vi->vdev, VIRTIO_NET_F_STATUS,
1249 			      offsetof(struct virtio_net_config, status),
1250 			      &v) < 0)
1251 		goto done;
1252 
1253 	if (v & VIRTIO_NET_S_ANNOUNCE) {
1254 		netdev_notify_peers(vi->dev);
1255 		virtnet_ack_link_announce(vi);
1256 	}
1257 
1258 	/* Ignore unknown (future) status bits */
1259 	v &= VIRTIO_NET_S_LINK_UP;
1260 
1261 	if (vi->status == v)
1262 		goto done;
1263 
1264 	vi->status = v;
1265 
1266 	if (vi->status & VIRTIO_NET_S_LINK_UP) {
1267 		netif_carrier_on(vi->dev);
1268 		netif_tx_wake_all_queues(vi->dev);
1269 	} else {
1270 		netif_carrier_off(vi->dev);
1271 		netif_tx_stop_all_queues(vi->dev);
1272 	}
1273 done:
1274 	mutex_unlock(&vi->config_lock);
1275 }
1276 
1277 static void virtnet_config_changed(struct virtio_device *vdev)
1278 {
1279 	struct virtnet_info *vi = vdev->priv;
1280 
1281 	schedule_work(&vi->config_work);
1282 }
1283 
1284 static void virtnet_free_queues(struct virtnet_info *vi)
1285 {
1286 	kfree(vi->rq);
1287 	kfree(vi->sq);
1288 }
1289 
1290 static void free_receive_bufs(struct virtnet_info *vi)
1291 {
1292 	int i;
1293 
1294 	for (i = 0; i < vi->max_queue_pairs; i++) {
1295 		while (vi->rq[i].pages)
1296 			__free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1297 	}
1298 }
1299 
1300 static void free_unused_bufs(struct virtnet_info *vi)
1301 {
1302 	void *buf;
1303 	int i;
1304 
1305 	for (i = 0; i < vi->max_queue_pairs; i++) {
1306 		struct virtqueue *vq = vi->sq[i].vq;
1307 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1308 			dev_kfree_skb(buf);
1309 	}
1310 
1311 	for (i = 0; i < vi->max_queue_pairs; i++) {
1312 		struct virtqueue *vq = vi->rq[i].vq;
1313 
1314 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1315 			if (vi->mergeable_rx_bufs || vi->big_packets)
1316 				give_pages(&vi->rq[i], buf);
1317 			else
1318 				dev_kfree_skb(buf);
1319 			--vi->rq[i].num;
1320 		}
1321 		BUG_ON(vi->rq[i].num != 0);
1322 	}
1323 }
1324 
1325 static void virtnet_del_vqs(struct virtnet_info *vi)
1326 {
1327 	struct virtio_device *vdev = vi->vdev;
1328 
1329 	virtnet_clean_affinity(vi, -1);
1330 
1331 	vdev->config->del_vqs(vdev);
1332 
1333 	virtnet_free_queues(vi);
1334 }
1335 
1336 static int virtnet_find_vqs(struct virtnet_info *vi)
1337 {
1338 	vq_callback_t **callbacks;
1339 	struct virtqueue **vqs;
1340 	int ret = -ENOMEM;
1341 	int i, total_vqs;
1342 	const char **names;
1343 
1344 	/* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1345 	 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1346 	 * possible control vq.
1347 	 */
1348 	total_vqs = vi->max_queue_pairs * 2 +
1349 		    virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1350 
1351 	/* Allocate space for find_vqs parameters */
1352 	vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1353 	if (!vqs)
1354 		goto err_vq;
1355 	callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1356 	if (!callbacks)
1357 		goto err_callback;
1358 	names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1359 	if (!names)
1360 		goto err_names;
1361 
1362 	/* Parameters for control virtqueue, if any */
1363 	if (vi->has_cvq) {
1364 		callbacks[total_vqs - 1] = NULL;
1365 		names[total_vqs - 1] = "control";
1366 	}
1367 
1368 	/* Allocate/initialize parameters for send/receive virtqueues */
1369 	for (i = 0; i < vi->max_queue_pairs; i++) {
1370 		callbacks[rxq2vq(i)] = skb_recv_done;
1371 		callbacks[txq2vq(i)] = skb_xmit_done;
1372 		sprintf(vi->rq[i].name, "input.%d", i);
1373 		sprintf(vi->sq[i].name, "output.%d", i);
1374 		names[rxq2vq(i)] = vi->rq[i].name;
1375 		names[txq2vq(i)] = vi->sq[i].name;
1376 	}
1377 
1378 	ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1379 					 names);
1380 	if (ret)
1381 		goto err_find;
1382 
1383 	if (vi->has_cvq) {
1384 		vi->cvq = vqs[total_vqs - 1];
1385 		if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1386 			vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1387 	}
1388 
1389 	for (i = 0; i < vi->max_queue_pairs; i++) {
1390 		vi->rq[i].vq = vqs[rxq2vq(i)];
1391 		vi->sq[i].vq = vqs[txq2vq(i)];
1392 	}
1393 
1394 	kfree(names);
1395 	kfree(callbacks);
1396 	kfree(vqs);
1397 
1398 	return 0;
1399 
1400 err_find:
1401 	kfree(names);
1402 err_names:
1403 	kfree(callbacks);
1404 err_callback:
1405 	kfree(vqs);
1406 err_vq:
1407 	return ret;
1408 }
1409 
1410 static int virtnet_alloc_queues(struct virtnet_info *vi)
1411 {
1412 	int i;
1413 
1414 	vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1415 	if (!vi->sq)
1416 		goto err_sq;
1417 	vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1418 	if (!vi->rq)
1419 		goto err_rq;
1420 
1421 	INIT_DELAYED_WORK(&vi->refill, refill_work);
1422 	for (i = 0; i < vi->max_queue_pairs; i++) {
1423 		vi->rq[i].pages = NULL;
1424 		netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1425 			       napi_weight);
1426 
1427 		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1428 		sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1429 	}
1430 
1431 	return 0;
1432 
1433 err_rq:
1434 	kfree(vi->sq);
1435 err_sq:
1436 	return -ENOMEM;
1437 }
1438 
1439 static int init_vqs(struct virtnet_info *vi)
1440 {
1441 	int ret;
1442 
1443 	/* Allocate send & receive queues */
1444 	ret = virtnet_alloc_queues(vi);
1445 	if (ret)
1446 		goto err;
1447 
1448 	ret = virtnet_find_vqs(vi);
1449 	if (ret)
1450 		goto err_free;
1451 
1452 	get_online_cpus();
1453 	virtnet_set_affinity(vi);
1454 	put_online_cpus();
1455 
1456 	return 0;
1457 
1458 err_free:
1459 	virtnet_free_queues(vi);
1460 err:
1461 	return ret;
1462 }
1463 
1464 static int virtnet_probe(struct virtio_device *vdev)
1465 {
1466 	int i, err;
1467 	struct net_device *dev;
1468 	struct virtnet_info *vi;
1469 	u16 max_queue_pairs;
1470 
1471 	/* Find if host supports multiqueue virtio_net device */
1472 	err = virtio_config_val(vdev, VIRTIO_NET_F_MQ,
1473 				offsetof(struct virtio_net_config,
1474 				max_virtqueue_pairs), &max_queue_pairs);
1475 
1476 	/* We need at least 2 queue's */
1477 	if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1478 	    max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1479 	    !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1480 		max_queue_pairs = 1;
1481 
1482 	/* Allocate ourselves a network device with room for our info */
1483 	dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1484 	if (!dev)
1485 		return -ENOMEM;
1486 
1487 	/* Set up network device as normal. */
1488 	dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1489 	dev->netdev_ops = &virtnet_netdev;
1490 	dev->features = NETIF_F_HIGHDMA;
1491 
1492 	SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1493 	SET_NETDEV_DEV(dev, &vdev->dev);
1494 
1495 	/* Do we support "hardware" checksums? */
1496 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1497 		/* This opens up the world of extra features. */
1498 		dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1499 		if (csum)
1500 			dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1501 
1502 		if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1503 			dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1504 				| NETIF_F_TSO_ECN | NETIF_F_TSO6;
1505 		}
1506 		/* Individual feature bits: what can host handle? */
1507 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1508 			dev->hw_features |= NETIF_F_TSO;
1509 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1510 			dev->hw_features |= NETIF_F_TSO6;
1511 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1512 			dev->hw_features |= NETIF_F_TSO_ECN;
1513 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1514 			dev->hw_features |= NETIF_F_UFO;
1515 
1516 		if (gso)
1517 			dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1518 		/* (!csum && gso) case will be fixed by register_netdev() */
1519 	}
1520 
1521 	dev->vlan_features = dev->features;
1522 
1523 	/* Configuration may specify what MAC to use.  Otherwise random. */
1524 	if (virtio_config_val_len(vdev, VIRTIO_NET_F_MAC,
1525 				  offsetof(struct virtio_net_config, mac),
1526 				  dev->dev_addr, dev->addr_len) < 0)
1527 		eth_hw_addr_random(dev);
1528 
1529 	/* Set up our device-specific information */
1530 	vi = netdev_priv(dev);
1531 	vi->dev = dev;
1532 	vi->vdev = vdev;
1533 	vdev->priv = vi;
1534 	vi->stats = alloc_percpu(struct virtnet_stats);
1535 	err = -ENOMEM;
1536 	if (vi->stats == NULL)
1537 		goto free;
1538 
1539 	vi->vq_index = alloc_percpu(int);
1540 	if (vi->vq_index == NULL)
1541 		goto free_stats;
1542 
1543 	mutex_init(&vi->config_lock);
1544 	vi->config_enable = true;
1545 	INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1546 
1547 	/* If we can receive ANY GSO packets, we must allocate large ones. */
1548 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1549 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1550 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1551 		vi->big_packets = true;
1552 
1553 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1554 		vi->mergeable_rx_bufs = true;
1555 
1556 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1557 		vi->has_cvq = true;
1558 
1559 	/* Use single tx/rx queue pair as default */
1560 	vi->curr_queue_pairs = 1;
1561 	vi->max_queue_pairs = max_queue_pairs;
1562 
1563 	/* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1564 	err = init_vqs(vi);
1565 	if (err)
1566 		goto free_index;
1567 
1568 	netif_set_real_num_tx_queues(dev, 1);
1569 	netif_set_real_num_rx_queues(dev, 1);
1570 
1571 	err = register_netdev(dev);
1572 	if (err) {
1573 		pr_debug("virtio_net: registering device failed\n");
1574 		goto free_vqs;
1575 	}
1576 
1577 	/* Last of all, set up some receive buffers. */
1578 	for (i = 0; i < vi->curr_queue_pairs; i++) {
1579 		try_fill_recv(&vi->rq[i], GFP_KERNEL);
1580 
1581 		/* If we didn't even get one input buffer, we're useless. */
1582 		if (vi->rq[i].num == 0) {
1583 			free_unused_bufs(vi);
1584 			err = -ENOMEM;
1585 			goto free_recv_bufs;
1586 		}
1587 	}
1588 
1589 	vi->nb.notifier_call = &virtnet_cpu_callback;
1590 	err = register_hotcpu_notifier(&vi->nb);
1591 	if (err) {
1592 		pr_debug("virtio_net: registering cpu notifier failed\n");
1593 		goto free_recv_bufs;
1594 	}
1595 
1596 	/* Assume link up if device can't report link status,
1597 	   otherwise get link status from config. */
1598 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1599 		netif_carrier_off(dev);
1600 		schedule_work(&vi->config_work);
1601 	} else {
1602 		vi->status = VIRTIO_NET_S_LINK_UP;
1603 		netif_carrier_on(dev);
1604 	}
1605 
1606 	pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1607 		 dev->name, max_queue_pairs);
1608 
1609 	return 0;
1610 
1611 free_recv_bufs:
1612 	free_receive_bufs(vi);
1613 	unregister_netdev(dev);
1614 free_vqs:
1615 	cancel_delayed_work_sync(&vi->refill);
1616 	virtnet_del_vqs(vi);
1617 free_index:
1618 	free_percpu(vi->vq_index);
1619 free_stats:
1620 	free_percpu(vi->stats);
1621 free:
1622 	free_netdev(dev);
1623 	return err;
1624 }
1625 
1626 static void remove_vq_common(struct virtnet_info *vi)
1627 {
1628 	vi->vdev->config->reset(vi->vdev);
1629 
1630 	/* Free unused buffers in both send and recv, if any. */
1631 	free_unused_bufs(vi);
1632 
1633 	free_receive_bufs(vi);
1634 
1635 	virtnet_del_vqs(vi);
1636 }
1637 
1638 static void virtnet_remove(struct virtio_device *vdev)
1639 {
1640 	struct virtnet_info *vi = vdev->priv;
1641 
1642 	unregister_hotcpu_notifier(&vi->nb);
1643 
1644 	/* Prevent config work handler from accessing the device. */
1645 	mutex_lock(&vi->config_lock);
1646 	vi->config_enable = false;
1647 	mutex_unlock(&vi->config_lock);
1648 
1649 	unregister_netdev(vi->dev);
1650 
1651 	remove_vq_common(vi);
1652 
1653 	flush_work(&vi->config_work);
1654 
1655 	free_percpu(vi->vq_index);
1656 	free_percpu(vi->stats);
1657 	free_netdev(vi->dev);
1658 }
1659 
1660 #ifdef CONFIG_PM
1661 static int virtnet_freeze(struct virtio_device *vdev)
1662 {
1663 	struct virtnet_info *vi = vdev->priv;
1664 	int i;
1665 
1666 	/* Prevent config work handler from accessing the device */
1667 	mutex_lock(&vi->config_lock);
1668 	vi->config_enable = false;
1669 	mutex_unlock(&vi->config_lock);
1670 
1671 	netif_device_detach(vi->dev);
1672 	cancel_delayed_work_sync(&vi->refill);
1673 
1674 	if (netif_running(vi->dev))
1675 		for (i = 0; i < vi->max_queue_pairs; i++) {
1676 			napi_disable(&vi->rq[i].napi);
1677 			netif_napi_del(&vi->rq[i].napi);
1678 		}
1679 
1680 	remove_vq_common(vi);
1681 
1682 	flush_work(&vi->config_work);
1683 
1684 	return 0;
1685 }
1686 
1687 static int virtnet_restore(struct virtio_device *vdev)
1688 {
1689 	struct virtnet_info *vi = vdev->priv;
1690 	int err, i;
1691 
1692 	err = init_vqs(vi);
1693 	if (err)
1694 		return err;
1695 
1696 	if (netif_running(vi->dev))
1697 		for (i = 0; i < vi->max_queue_pairs; i++)
1698 			virtnet_napi_enable(&vi->rq[i]);
1699 
1700 	netif_device_attach(vi->dev);
1701 
1702 	for (i = 0; i < vi->curr_queue_pairs; i++)
1703 		if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1704 			schedule_delayed_work(&vi->refill, 0);
1705 
1706 	mutex_lock(&vi->config_lock);
1707 	vi->config_enable = true;
1708 	mutex_unlock(&vi->config_lock);
1709 
1710 	virtnet_set_queues(vi, vi->curr_queue_pairs);
1711 
1712 	return 0;
1713 }
1714 #endif
1715 
1716 static struct virtio_device_id id_table[] = {
1717 	{ VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1718 	{ 0 },
1719 };
1720 
1721 static unsigned int features[] = {
1722 	VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1723 	VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1724 	VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1725 	VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1726 	VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1727 	VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1728 	VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1729 	VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1730 	VIRTIO_NET_F_CTRL_MAC_ADDR,
1731 };
1732 
1733 static struct virtio_driver virtio_net_driver = {
1734 	.feature_table = features,
1735 	.feature_table_size = ARRAY_SIZE(features),
1736 	.driver.name =	KBUILD_MODNAME,
1737 	.driver.owner =	THIS_MODULE,
1738 	.id_table =	id_table,
1739 	.probe =	virtnet_probe,
1740 	.remove =	virtnet_remove,
1741 	.config_changed = virtnet_config_changed,
1742 #ifdef CONFIG_PM
1743 	.freeze =	virtnet_freeze,
1744 	.restore =	virtnet_restore,
1745 #endif
1746 };
1747 
1748 module_virtio_driver(virtio_net_driver);
1749 
1750 MODULE_DEVICE_TABLE(virtio, id_table);
1751 MODULE_DESCRIPTION("Virtio network driver");
1752 MODULE_LICENSE("GPL");
1753