xref: /openbmc/linux/drivers/net/virtio_net.c (revision 161f4089)
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 	/* Host can handle any s/g split between our header and packet data */
110 	bool any_header_sg;
111 
112 	/* enable config space updates */
113 	bool config_enable;
114 
115 	/* Active statistics */
116 	struct virtnet_stats __percpu *stats;
117 
118 	/* Work struct for refilling if we run low on memory. */
119 	struct delayed_work refill;
120 
121 	/* Work struct for config space updates */
122 	struct work_struct config_work;
123 
124 	/* Lock for config space updates */
125 	struct mutex config_lock;
126 
127 	/* Does the affinity hint is set for virtqueues? */
128 	bool affinity_hint_set;
129 
130 	/* Per-cpu variable to show the mapping from CPU to virtqueue */
131 	int __percpu *vq_index;
132 
133 	/* CPU hot plug notifier */
134 	struct notifier_block nb;
135 };
136 
137 struct skb_vnet_hdr {
138 	union {
139 		struct virtio_net_hdr hdr;
140 		struct virtio_net_hdr_mrg_rxbuf mhdr;
141 	};
142 };
143 
144 struct padded_vnet_hdr {
145 	struct virtio_net_hdr hdr;
146 	/*
147 	 * virtio_net_hdr should be in a separated sg buffer because of a
148 	 * QEMU bug, and data sg buffer shares same page with this header sg.
149 	 * This padding makes next sg 16 byte aligned after virtio_net_hdr.
150 	 */
151 	char padding[6];
152 };
153 
154 /* Converting between virtqueue no. and kernel tx/rx queue no.
155  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
156  */
157 static int vq2txq(struct virtqueue *vq)
158 {
159 	return (vq->index - 1) / 2;
160 }
161 
162 static int txq2vq(int txq)
163 {
164 	return txq * 2 + 1;
165 }
166 
167 static int vq2rxq(struct virtqueue *vq)
168 {
169 	return vq->index / 2;
170 }
171 
172 static int rxq2vq(int rxq)
173 {
174 	return rxq * 2;
175 }
176 
177 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
178 {
179 	return (struct skb_vnet_hdr *)skb->cb;
180 }
181 
182 /*
183  * private is used to chain pages for big packets, put the whole
184  * most recent used list in the beginning for reuse
185  */
186 static void give_pages(struct receive_queue *rq, struct page *page)
187 {
188 	struct page *end;
189 
190 	/* Find end of list, sew whole thing into vi->rq.pages. */
191 	for (end = page; end->private; end = (struct page *)end->private);
192 	end->private = (unsigned long)rq->pages;
193 	rq->pages = page;
194 }
195 
196 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
197 {
198 	struct page *p = rq->pages;
199 
200 	if (p) {
201 		rq->pages = (struct page *)p->private;
202 		/* clear private here, it is used to chain pages */
203 		p->private = 0;
204 	} else
205 		p = alloc_page(gfp_mask);
206 	return p;
207 }
208 
209 static void skb_xmit_done(struct virtqueue *vq)
210 {
211 	struct virtnet_info *vi = vq->vdev->priv;
212 
213 	/* Suppress further interrupts. */
214 	virtqueue_disable_cb(vq);
215 
216 	/* We were probably waiting for more output buffers. */
217 	netif_wake_subqueue(vi->dev, vq2txq(vq));
218 }
219 
220 static void set_skb_frag(struct sk_buff *skb, struct page *page,
221 			 unsigned int offset, unsigned int *len)
222 {
223 	int size = min((unsigned)PAGE_SIZE - offset, *len);
224 	int i = skb_shinfo(skb)->nr_frags;
225 
226 	__skb_fill_page_desc(skb, i, page, offset, size);
227 
228 	skb->data_len += size;
229 	skb->len += size;
230 	skb->truesize += PAGE_SIZE;
231 	skb_shinfo(skb)->nr_frags++;
232 	skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
233 	*len -= size;
234 }
235 
236 /* Called from bottom half context */
237 static struct sk_buff *page_to_skb(struct receive_queue *rq,
238 				   struct page *page, unsigned int len)
239 {
240 	struct virtnet_info *vi = rq->vq->vdev->priv;
241 	struct sk_buff *skb;
242 	struct skb_vnet_hdr *hdr;
243 	unsigned int copy, hdr_len, offset;
244 	char *p;
245 
246 	p = page_address(page);
247 
248 	/* copy small packet so we can reuse these pages for small data */
249 	skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
250 	if (unlikely(!skb))
251 		return NULL;
252 
253 	hdr = skb_vnet_hdr(skb);
254 
255 	if (vi->mergeable_rx_bufs) {
256 		hdr_len = sizeof hdr->mhdr;
257 		offset = hdr_len;
258 	} else {
259 		hdr_len = sizeof hdr->hdr;
260 		offset = sizeof(struct padded_vnet_hdr);
261 	}
262 
263 	memcpy(hdr, p, hdr_len);
264 
265 	len -= hdr_len;
266 	p += offset;
267 
268 	copy = len;
269 	if (copy > skb_tailroom(skb))
270 		copy = skb_tailroom(skb);
271 	memcpy(skb_put(skb, copy), p, copy);
272 
273 	len -= copy;
274 	offset += copy;
275 
276 	/*
277 	 * Verify that we can indeed put this data into a skb.
278 	 * This is here to handle cases when the device erroneously
279 	 * tries to receive more than is possible. This is usually
280 	 * the case of a broken device.
281 	 */
282 	if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
283 		net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
284 		dev_kfree_skb(skb);
285 		return NULL;
286 	}
287 
288 	while (len) {
289 		set_skb_frag(skb, page, offset, &len);
290 		page = (struct page *)page->private;
291 		offset = 0;
292 	}
293 
294 	if (page)
295 		give_pages(rq, page);
296 
297 	return skb;
298 }
299 
300 static int receive_mergeable(struct receive_queue *rq, struct sk_buff *skb)
301 {
302 	struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
303 	struct page *page;
304 	int num_buf, i, len;
305 
306 	num_buf = hdr->mhdr.num_buffers;
307 	while (--num_buf) {
308 		i = skb_shinfo(skb)->nr_frags;
309 		if (i >= MAX_SKB_FRAGS) {
310 			pr_debug("%s: packet too long\n", skb->dev->name);
311 			skb->dev->stats.rx_length_errors++;
312 			return -EINVAL;
313 		}
314 		page = virtqueue_get_buf(rq->vq, &len);
315 		if (!page) {
316 			pr_debug("%s: rx error: %d buffers missing\n",
317 				 skb->dev->name, hdr->mhdr.num_buffers);
318 			skb->dev->stats.rx_length_errors++;
319 			return -EINVAL;
320 		}
321 
322 		if (len > PAGE_SIZE)
323 			len = PAGE_SIZE;
324 
325 		set_skb_frag(skb, page, 0, &len);
326 
327 		--rq->num;
328 	}
329 	return 0;
330 }
331 
332 static void receive_buf(struct receive_queue *rq, void *buf, unsigned int len)
333 {
334 	struct virtnet_info *vi = rq->vq->vdev->priv;
335 	struct net_device *dev = vi->dev;
336 	struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
337 	struct sk_buff *skb;
338 	struct page *page;
339 	struct skb_vnet_hdr *hdr;
340 
341 	if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
342 		pr_debug("%s: short packet %i\n", dev->name, len);
343 		dev->stats.rx_length_errors++;
344 		if (vi->mergeable_rx_bufs || vi->big_packets)
345 			give_pages(rq, buf);
346 		else
347 			dev_kfree_skb(buf);
348 		return;
349 	}
350 
351 	if (!vi->mergeable_rx_bufs && !vi->big_packets) {
352 		skb = buf;
353 		len -= sizeof(struct virtio_net_hdr);
354 		skb_trim(skb, len);
355 	} else {
356 		page = buf;
357 		skb = page_to_skb(rq, page, len);
358 		if (unlikely(!skb)) {
359 			dev->stats.rx_dropped++;
360 			give_pages(rq, page);
361 			return;
362 		}
363 		if (vi->mergeable_rx_bufs)
364 			if (receive_mergeable(rq, skb)) {
365 				dev_kfree_skb(skb);
366 				return;
367 			}
368 	}
369 
370 	hdr = skb_vnet_hdr(skb);
371 
372 	u64_stats_update_begin(&stats->rx_syncp);
373 	stats->rx_bytes += skb->len;
374 	stats->rx_packets++;
375 	u64_stats_update_end(&stats->rx_syncp);
376 
377 	if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
378 		pr_debug("Needs csum!\n");
379 		if (!skb_partial_csum_set(skb,
380 					  hdr->hdr.csum_start,
381 					  hdr->hdr.csum_offset))
382 			goto frame_err;
383 	} else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
384 		skb->ip_summed = CHECKSUM_UNNECESSARY;
385 	}
386 
387 	skb->protocol = eth_type_trans(skb, dev);
388 	pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
389 		 ntohs(skb->protocol), skb->len, skb->pkt_type);
390 
391 	if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
392 		pr_debug("GSO!\n");
393 		switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
394 		case VIRTIO_NET_HDR_GSO_TCPV4:
395 			skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
396 			break;
397 		case VIRTIO_NET_HDR_GSO_UDP:
398 			skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
399 			break;
400 		case VIRTIO_NET_HDR_GSO_TCPV6:
401 			skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
402 			break;
403 		default:
404 			net_warn_ratelimited("%s: bad gso type %u.\n",
405 					     dev->name, hdr->hdr.gso_type);
406 			goto frame_err;
407 		}
408 
409 		if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
410 			skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
411 
412 		skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
413 		if (skb_shinfo(skb)->gso_size == 0) {
414 			net_warn_ratelimited("%s: zero gso size.\n", dev->name);
415 			goto frame_err;
416 		}
417 
418 		/* Header must be checked, and gso_segs computed. */
419 		skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
420 		skb_shinfo(skb)->gso_segs = 0;
421 	}
422 
423 	netif_receive_skb(skb);
424 	return;
425 
426 frame_err:
427 	dev->stats.rx_frame_errors++;
428 	dev_kfree_skb(skb);
429 }
430 
431 static int add_recvbuf_small(struct receive_queue *rq, gfp_t gfp)
432 {
433 	struct virtnet_info *vi = rq->vq->vdev->priv;
434 	struct sk_buff *skb;
435 	struct skb_vnet_hdr *hdr;
436 	int err;
437 
438 	skb = __netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN, gfp);
439 	if (unlikely(!skb))
440 		return -ENOMEM;
441 
442 	skb_put(skb, MAX_PACKET_LEN);
443 
444 	hdr = skb_vnet_hdr(skb);
445 	sg_set_buf(rq->sg, &hdr->hdr, sizeof hdr->hdr);
446 
447 	skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
448 
449 	err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
450 	if (err < 0)
451 		dev_kfree_skb(skb);
452 
453 	return err;
454 }
455 
456 static int add_recvbuf_big(struct receive_queue *rq, gfp_t gfp)
457 {
458 	struct page *first, *list = NULL;
459 	char *p;
460 	int i, err, offset;
461 
462 	/* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
463 	for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
464 		first = get_a_page(rq, gfp);
465 		if (!first) {
466 			if (list)
467 				give_pages(rq, list);
468 			return -ENOMEM;
469 		}
470 		sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
471 
472 		/* chain new page in list head to match sg */
473 		first->private = (unsigned long)list;
474 		list = first;
475 	}
476 
477 	first = get_a_page(rq, gfp);
478 	if (!first) {
479 		give_pages(rq, list);
480 		return -ENOMEM;
481 	}
482 	p = page_address(first);
483 
484 	/* rq->sg[0], rq->sg[1] share the same page */
485 	/* a separated rq->sg[0] for virtio_net_hdr only due to QEMU bug */
486 	sg_set_buf(&rq->sg[0], p, sizeof(struct virtio_net_hdr));
487 
488 	/* rq->sg[1] for data packet, from offset */
489 	offset = sizeof(struct padded_vnet_hdr);
490 	sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
491 
492 	/* chain first in list head */
493 	first->private = (unsigned long)list;
494 	err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
495 				  first, gfp);
496 	if (err < 0)
497 		give_pages(rq, first);
498 
499 	return err;
500 }
501 
502 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
503 {
504 	struct page *page;
505 	int err;
506 
507 	page = get_a_page(rq, gfp);
508 	if (!page)
509 		return -ENOMEM;
510 
511 	sg_init_one(rq->sg, page_address(page), PAGE_SIZE);
512 
513 	err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, page, gfp);
514 	if (err < 0)
515 		give_pages(rq, page);
516 
517 	return err;
518 }
519 
520 /*
521  * Returns false if we couldn't fill entirely (OOM).
522  *
523  * Normally run in the receive path, but can also be run from ndo_open
524  * before we're receiving packets, or from refill_work which is
525  * careful to disable receiving (using napi_disable).
526  */
527 static bool try_fill_recv(struct receive_queue *rq, gfp_t gfp)
528 {
529 	struct virtnet_info *vi = rq->vq->vdev->priv;
530 	int err;
531 	bool oom;
532 
533 	do {
534 		if (vi->mergeable_rx_bufs)
535 			err = add_recvbuf_mergeable(rq, gfp);
536 		else if (vi->big_packets)
537 			err = add_recvbuf_big(rq, gfp);
538 		else
539 			err = add_recvbuf_small(rq, gfp);
540 
541 		oom = err == -ENOMEM;
542 		if (err)
543 			break;
544 		++rq->num;
545 	} while (rq->vq->num_free);
546 	if (unlikely(rq->num > rq->max))
547 		rq->max = rq->num;
548 	virtqueue_kick(rq->vq);
549 	return !oom;
550 }
551 
552 static void skb_recv_done(struct virtqueue *rvq)
553 {
554 	struct virtnet_info *vi = rvq->vdev->priv;
555 	struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
556 
557 	/* Schedule NAPI, Suppress further interrupts if successful. */
558 	if (napi_schedule_prep(&rq->napi)) {
559 		virtqueue_disable_cb(rvq);
560 		__napi_schedule(&rq->napi);
561 	}
562 }
563 
564 static void virtnet_napi_enable(struct receive_queue *rq)
565 {
566 	napi_enable(&rq->napi);
567 
568 	/* If all buffers were filled by other side before we napi_enabled, we
569 	 * won't get another interrupt, so process any outstanding packets
570 	 * now.  virtnet_poll wants re-enable the queue, so we disable here.
571 	 * We synchronize against interrupts via NAPI_STATE_SCHED */
572 	if (napi_schedule_prep(&rq->napi)) {
573 		virtqueue_disable_cb(rq->vq);
574 		local_bh_disable();
575 		__napi_schedule(&rq->napi);
576 		local_bh_enable();
577 	}
578 }
579 
580 static void refill_work(struct work_struct *work)
581 {
582 	struct virtnet_info *vi =
583 		container_of(work, struct virtnet_info, refill.work);
584 	bool still_empty;
585 	int i;
586 
587 	for (i = 0; i < vi->curr_queue_pairs; i++) {
588 		struct receive_queue *rq = &vi->rq[i];
589 
590 		napi_disable(&rq->napi);
591 		still_empty = !try_fill_recv(rq, GFP_KERNEL);
592 		virtnet_napi_enable(rq);
593 
594 		/* In theory, this can happen: if we don't get any buffers in
595 		 * we will *never* try to fill again.
596 		 */
597 		if (still_empty)
598 			schedule_delayed_work(&vi->refill, HZ/2);
599 	}
600 }
601 
602 static int virtnet_poll(struct napi_struct *napi, int budget)
603 {
604 	struct receive_queue *rq =
605 		container_of(napi, struct receive_queue, napi);
606 	struct virtnet_info *vi = rq->vq->vdev->priv;
607 	void *buf;
608 	unsigned int r, len, received = 0;
609 
610 again:
611 	while (received < budget &&
612 	       (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
613 		receive_buf(rq, buf, len);
614 		--rq->num;
615 		received++;
616 	}
617 
618 	if (rq->num < rq->max / 2) {
619 		if (!try_fill_recv(rq, GFP_ATOMIC))
620 			schedule_delayed_work(&vi->refill, 0);
621 	}
622 
623 	/* Out of packets? */
624 	if (received < budget) {
625 		r = virtqueue_enable_cb_prepare(rq->vq);
626 		napi_complete(napi);
627 		if (unlikely(virtqueue_poll(rq->vq, r)) &&
628 		    napi_schedule_prep(napi)) {
629 			virtqueue_disable_cb(rq->vq);
630 			__napi_schedule(napi);
631 			goto again;
632 		}
633 	}
634 
635 	return received;
636 }
637 
638 static int virtnet_open(struct net_device *dev)
639 {
640 	struct virtnet_info *vi = netdev_priv(dev);
641 	int i;
642 
643 	for (i = 0; i < vi->max_queue_pairs; i++) {
644 		if (i < vi->curr_queue_pairs)
645 			/* Make sure we have some buffers: if oom use wq. */
646 			if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
647 				schedule_delayed_work(&vi->refill, 0);
648 		virtnet_napi_enable(&vi->rq[i]);
649 	}
650 
651 	return 0;
652 }
653 
654 static void free_old_xmit_skbs(struct send_queue *sq)
655 {
656 	struct sk_buff *skb;
657 	unsigned int len;
658 	struct virtnet_info *vi = sq->vq->vdev->priv;
659 	struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
660 
661 	while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
662 		pr_debug("Sent skb %p\n", skb);
663 
664 		u64_stats_update_begin(&stats->tx_syncp);
665 		stats->tx_bytes += skb->len;
666 		stats->tx_packets++;
667 		u64_stats_update_end(&stats->tx_syncp);
668 
669 		dev_kfree_skb_any(skb);
670 	}
671 }
672 
673 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
674 {
675 	struct skb_vnet_hdr *hdr;
676 	const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
677 	struct virtnet_info *vi = sq->vq->vdev->priv;
678 	unsigned num_sg;
679 	unsigned hdr_len;
680 	bool can_push;
681 
682 	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
683 	if (vi->mergeable_rx_bufs)
684 		hdr_len = sizeof hdr->mhdr;
685 	else
686 		hdr_len = sizeof hdr->hdr;
687 
688 	can_push = vi->any_header_sg &&
689 		!((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
690 		!skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
691 	/* Even if we can, don't push here yet as this would skew
692 	 * csum_start offset below. */
693 	if (can_push)
694 		hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
695 	else
696 		hdr = skb_vnet_hdr(skb);
697 
698 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
699 		hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
700 		hdr->hdr.csum_start = skb_checksum_start_offset(skb);
701 		hdr->hdr.csum_offset = skb->csum_offset;
702 	} else {
703 		hdr->hdr.flags = 0;
704 		hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
705 	}
706 
707 	if (skb_is_gso(skb)) {
708 		hdr->hdr.hdr_len = skb_headlen(skb);
709 		hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
710 		if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
711 			hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
712 		else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
713 			hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
714 		else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
715 			hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
716 		else
717 			BUG();
718 		if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
719 			hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
720 	} else {
721 		hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
722 		hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
723 	}
724 
725 	if (vi->mergeable_rx_bufs)
726 		hdr->mhdr.num_buffers = 0;
727 
728 	if (can_push) {
729 		__skb_push(skb, hdr_len);
730 		num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
731 		/* Pull header back to avoid skew in tx bytes calculations. */
732 		__skb_pull(skb, hdr_len);
733 	} else {
734 		sg_set_buf(sq->sg, hdr, hdr_len);
735 		num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
736 	}
737 	return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
738 }
739 
740 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
741 {
742 	struct virtnet_info *vi = netdev_priv(dev);
743 	int qnum = skb_get_queue_mapping(skb);
744 	struct send_queue *sq = &vi->sq[qnum];
745 	int err;
746 
747 	/* Free up any pending old buffers before queueing new ones. */
748 	free_old_xmit_skbs(sq);
749 
750 	/* Try to transmit */
751 	err = xmit_skb(sq, skb);
752 
753 	/* This should not happen! */
754 	if (unlikely(err)) {
755 		dev->stats.tx_fifo_errors++;
756 		if (net_ratelimit())
757 			dev_warn(&dev->dev,
758 				 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
759 		dev->stats.tx_dropped++;
760 		kfree_skb(skb);
761 		return NETDEV_TX_OK;
762 	}
763 	virtqueue_kick(sq->vq);
764 
765 	/* Don't wait up for transmitted skbs to be freed. */
766 	skb_orphan(skb);
767 	nf_reset(skb);
768 
769 	/* Apparently nice girls don't return TX_BUSY; stop the queue
770 	 * before it gets out of hand.  Naturally, this wastes entries. */
771 	if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
772 		netif_stop_subqueue(dev, qnum);
773 		if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
774 			/* More just got used, free them then recheck. */
775 			free_old_xmit_skbs(sq);
776 			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
777 				netif_start_subqueue(dev, qnum);
778 				virtqueue_disable_cb(sq->vq);
779 			}
780 		}
781 	}
782 
783 	return NETDEV_TX_OK;
784 }
785 
786 /*
787  * Send command via the control virtqueue and check status.  Commands
788  * supported by the hypervisor, as indicated by feature bits, should
789  * never fail unless improperly formated.
790  */
791 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
792 				 struct scatterlist *out,
793 				 struct scatterlist *in)
794 {
795 	struct scatterlist *sgs[4], hdr, stat;
796 	struct virtio_net_ctrl_hdr ctrl;
797 	virtio_net_ctrl_ack status = ~0;
798 	unsigned out_num = 0, in_num = 0, tmp;
799 
800 	/* Caller should know better */
801 	BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
802 
803 	ctrl.class = class;
804 	ctrl.cmd = cmd;
805 	/* Add header */
806 	sg_init_one(&hdr, &ctrl, sizeof(ctrl));
807 	sgs[out_num++] = &hdr;
808 
809 	if (out)
810 		sgs[out_num++] = out;
811 	if (in)
812 		sgs[out_num + in_num++] = in;
813 
814 	/* Add return status. */
815 	sg_init_one(&stat, &status, sizeof(status));
816 	sgs[out_num + in_num++] = &stat;
817 
818 	BUG_ON(out_num + in_num > ARRAY_SIZE(sgs));
819 	BUG_ON(virtqueue_add_sgs(vi->cvq, sgs, out_num, in_num, vi, GFP_ATOMIC)
820 	       < 0);
821 
822 	virtqueue_kick(vi->cvq);
823 
824 	/* Spin for a response, the kick causes an ioport write, trapping
825 	 * into the hypervisor, so the request should be handled immediately.
826 	 */
827 	while (!virtqueue_get_buf(vi->cvq, &tmp))
828 		cpu_relax();
829 
830 	return status == VIRTIO_NET_OK;
831 }
832 
833 static int virtnet_set_mac_address(struct net_device *dev, void *p)
834 {
835 	struct virtnet_info *vi = netdev_priv(dev);
836 	struct virtio_device *vdev = vi->vdev;
837 	int ret;
838 	struct sockaddr *addr = p;
839 	struct scatterlist sg;
840 
841 	ret = eth_prepare_mac_addr_change(dev, p);
842 	if (ret)
843 		return ret;
844 
845 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
846 		sg_init_one(&sg, addr->sa_data, dev->addr_len);
847 		if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
848 					  VIRTIO_NET_CTRL_MAC_ADDR_SET,
849 					  &sg, NULL)) {
850 			dev_warn(&vdev->dev,
851 				 "Failed to set mac address by vq command.\n");
852 			return -EINVAL;
853 		}
854 	} else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
855 		vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
856 				  addr->sa_data, dev->addr_len);
857 	}
858 
859 	eth_commit_mac_addr_change(dev, p);
860 
861 	return 0;
862 }
863 
864 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
865 					       struct rtnl_link_stats64 *tot)
866 {
867 	struct virtnet_info *vi = netdev_priv(dev);
868 	int cpu;
869 	unsigned int start;
870 
871 	for_each_possible_cpu(cpu) {
872 		struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
873 		u64 tpackets, tbytes, rpackets, rbytes;
874 
875 		do {
876 			start = u64_stats_fetch_begin_bh(&stats->tx_syncp);
877 			tpackets = stats->tx_packets;
878 			tbytes   = stats->tx_bytes;
879 		} while (u64_stats_fetch_retry_bh(&stats->tx_syncp, start));
880 
881 		do {
882 			start = u64_stats_fetch_begin_bh(&stats->rx_syncp);
883 			rpackets = stats->rx_packets;
884 			rbytes   = stats->rx_bytes;
885 		} while (u64_stats_fetch_retry_bh(&stats->rx_syncp, start));
886 
887 		tot->rx_packets += rpackets;
888 		tot->tx_packets += tpackets;
889 		tot->rx_bytes   += rbytes;
890 		tot->tx_bytes   += tbytes;
891 	}
892 
893 	tot->tx_dropped = dev->stats.tx_dropped;
894 	tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
895 	tot->rx_dropped = dev->stats.rx_dropped;
896 	tot->rx_length_errors = dev->stats.rx_length_errors;
897 	tot->rx_frame_errors = dev->stats.rx_frame_errors;
898 
899 	return tot;
900 }
901 
902 #ifdef CONFIG_NET_POLL_CONTROLLER
903 static void virtnet_netpoll(struct net_device *dev)
904 {
905 	struct virtnet_info *vi = netdev_priv(dev);
906 	int i;
907 
908 	for (i = 0; i < vi->curr_queue_pairs; i++)
909 		napi_schedule(&vi->rq[i].napi);
910 }
911 #endif
912 
913 static void virtnet_ack_link_announce(struct virtnet_info *vi)
914 {
915 	rtnl_lock();
916 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
917 				  VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL, NULL))
918 		dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
919 	rtnl_unlock();
920 }
921 
922 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
923 {
924 	struct scatterlist sg;
925 	struct virtio_net_ctrl_mq s;
926 	struct net_device *dev = vi->dev;
927 
928 	if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
929 		return 0;
930 
931 	s.virtqueue_pairs = queue_pairs;
932 	sg_init_one(&sg, &s, sizeof(s));
933 
934 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
935 				  VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg, NULL)) {
936 		dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
937 			 queue_pairs);
938 		return -EINVAL;
939 	} else {
940 		vi->curr_queue_pairs = queue_pairs;
941 		/* virtnet_open() will refill when device is going to up. */
942 		if (dev->flags & IFF_UP)
943 			schedule_delayed_work(&vi->refill, 0);
944 	}
945 
946 	return 0;
947 }
948 
949 static int virtnet_close(struct net_device *dev)
950 {
951 	struct virtnet_info *vi = netdev_priv(dev);
952 	int i;
953 
954 	/* Make sure refill_work doesn't re-enable napi! */
955 	cancel_delayed_work_sync(&vi->refill);
956 
957 	for (i = 0; i < vi->max_queue_pairs; i++)
958 		napi_disable(&vi->rq[i].napi);
959 
960 	return 0;
961 }
962 
963 static void virtnet_set_rx_mode(struct net_device *dev)
964 {
965 	struct virtnet_info *vi = netdev_priv(dev);
966 	struct scatterlist sg[2];
967 	u8 promisc, allmulti;
968 	struct virtio_net_ctrl_mac *mac_data;
969 	struct netdev_hw_addr *ha;
970 	int uc_count;
971 	int mc_count;
972 	void *buf;
973 	int i;
974 
975 	/* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
976 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
977 		return;
978 
979 	promisc = ((dev->flags & IFF_PROMISC) != 0);
980 	allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
981 
982 	sg_init_one(sg, &promisc, sizeof(promisc));
983 
984 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
985 				  VIRTIO_NET_CTRL_RX_PROMISC,
986 				  sg, NULL))
987 		dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
988 			 promisc ? "en" : "dis");
989 
990 	sg_init_one(sg, &allmulti, sizeof(allmulti));
991 
992 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
993 				  VIRTIO_NET_CTRL_RX_ALLMULTI,
994 				  sg, NULL))
995 		dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
996 			 allmulti ? "en" : "dis");
997 
998 	uc_count = netdev_uc_count(dev);
999 	mc_count = netdev_mc_count(dev);
1000 	/* MAC filter - use one buffer for both lists */
1001 	buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1002 		      (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1003 	mac_data = buf;
1004 	if (!buf)
1005 		return;
1006 
1007 	sg_init_table(sg, 2);
1008 
1009 	/* Store the unicast list and count in the front of the buffer */
1010 	mac_data->entries = uc_count;
1011 	i = 0;
1012 	netdev_for_each_uc_addr(ha, dev)
1013 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1014 
1015 	sg_set_buf(&sg[0], mac_data,
1016 		   sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1017 
1018 	/* multicast list and count fill the end */
1019 	mac_data = (void *)&mac_data->macs[uc_count][0];
1020 
1021 	mac_data->entries = mc_count;
1022 	i = 0;
1023 	netdev_for_each_mc_addr(ha, dev)
1024 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1025 
1026 	sg_set_buf(&sg[1], mac_data,
1027 		   sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1028 
1029 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1030 				  VIRTIO_NET_CTRL_MAC_TABLE_SET,
1031 				  sg, NULL))
1032 		dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
1033 
1034 	kfree(buf);
1035 }
1036 
1037 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1038 				   __be16 proto, u16 vid)
1039 {
1040 	struct virtnet_info *vi = netdev_priv(dev);
1041 	struct scatterlist sg;
1042 
1043 	sg_init_one(&sg, &vid, sizeof(vid));
1044 
1045 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1046 				  VIRTIO_NET_CTRL_VLAN_ADD, &sg, NULL))
1047 		dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1048 	return 0;
1049 }
1050 
1051 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1052 				    __be16 proto, u16 vid)
1053 {
1054 	struct virtnet_info *vi = netdev_priv(dev);
1055 	struct scatterlist sg;
1056 
1057 	sg_init_one(&sg, &vid, sizeof(vid));
1058 
1059 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1060 				  VIRTIO_NET_CTRL_VLAN_DEL, &sg, NULL))
1061 		dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1062 	return 0;
1063 }
1064 
1065 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1066 {
1067 	int i;
1068 	int cpu;
1069 
1070 	if (vi->affinity_hint_set) {
1071 		for (i = 0; i < vi->max_queue_pairs; i++) {
1072 			virtqueue_set_affinity(vi->rq[i].vq, -1);
1073 			virtqueue_set_affinity(vi->sq[i].vq, -1);
1074 		}
1075 
1076 		vi->affinity_hint_set = false;
1077 	}
1078 
1079 	i = 0;
1080 	for_each_online_cpu(cpu) {
1081 		if (cpu == hcpu) {
1082 			*per_cpu_ptr(vi->vq_index, cpu) = -1;
1083 		} else {
1084 			*per_cpu_ptr(vi->vq_index, cpu) =
1085 				++i % vi->curr_queue_pairs;
1086 		}
1087 	}
1088 }
1089 
1090 static void virtnet_set_affinity(struct virtnet_info *vi)
1091 {
1092 	int i;
1093 	int cpu;
1094 
1095 	/* In multiqueue mode, when the number of cpu is equal to the number of
1096 	 * queue pairs, we let the queue pairs to be private to one cpu by
1097 	 * setting the affinity hint to eliminate the contention.
1098 	 */
1099 	if (vi->curr_queue_pairs == 1 ||
1100 	    vi->max_queue_pairs != num_online_cpus()) {
1101 		virtnet_clean_affinity(vi, -1);
1102 		return;
1103 	}
1104 
1105 	i = 0;
1106 	for_each_online_cpu(cpu) {
1107 		virtqueue_set_affinity(vi->rq[i].vq, cpu);
1108 		virtqueue_set_affinity(vi->sq[i].vq, cpu);
1109 		*per_cpu_ptr(vi->vq_index, cpu) = i;
1110 		i++;
1111 	}
1112 
1113 	vi->affinity_hint_set = true;
1114 }
1115 
1116 static int virtnet_cpu_callback(struct notifier_block *nfb,
1117 			        unsigned long action, void *hcpu)
1118 {
1119 	struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1120 
1121 	switch(action & ~CPU_TASKS_FROZEN) {
1122 	case CPU_ONLINE:
1123 	case CPU_DOWN_FAILED:
1124 	case CPU_DEAD:
1125 		virtnet_set_affinity(vi);
1126 		break;
1127 	case CPU_DOWN_PREPARE:
1128 		virtnet_clean_affinity(vi, (long)hcpu);
1129 		break;
1130 	default:
1131 		break;
1132 	}
1133 
1134 	return NOTIFY_OK;
1135 }
1136 
1137 static void virtnet_get_ringparam(struct net_device *dev,
1138 				struct ethtool_ringparam *ring)
1139 {
1140 	struct virtnet_info *vi = netdev_priv(dev);
1141 
1142 	ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1143 	ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1144 	ring->rx_pending = ring->rx_max_pending;
1145 	ring->tx_pending = ring->tx_max_pending;
1146 }
1147 
1148 
1149 static void virtnet_get_drvinfo(struct net_device *dev,
1150 				struct ethtool_drvinfo *info)
1151 {
1152 	struct virtnet_info *vi = netdev_priv(dev);
1153 	struct virtio_device *vdev = vi->vdev;
1154 
1155 	strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1156 	strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1157 	strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1158 
1159 }
1160 
1161 /* TODO: Eliminate OOO packets during switching */
1162 static int virtnet_set_channels(struct net_device *dev,
1163 				struct ethtool_channels *channels)
1164 {
1165 	struct virtnet_info *vi = netdev_priv(dev);
1166 	u16 queue_pairs = channels->combined_count;
1167 	int err;
1168 
1169 	/* We don't support separate rx/tx channels.
1170 	 * We don't allow setting 'other' channels.
1171 	 */
1172 	if (channels->rx_count || channels->tx_count || channels->other_count)
1173 		return -EINVAL;
1174 
1175 	if (queue_pairs > vi->max_queue_pairs)
1176 		return -EINVAL;
1177 
1178 	get_online_cpus();
1179 	err = virtnet_set_queues(vi, queue_pairs);
1180 	if (!err) {
1181 		netif_set_real_num_tx_queues(dev, queue_pairs);
1182 		netif_set_real_num_rx_queues(dev, queue_pairs);
1183 
1184 		virtnet_set_affinity(vi);
1185 	}
1186 	put_online_cpus();
1187 
1188 	return err;
1189 }
1190 
1191 static void virtnet_get_channels(struct net_device *dev,
1192 				 struct ethtool_channels *channels)
1193 {
1194 	struct virtnet_info *vi = netdev_priv(dev);
1195 
1196 	channels->combined_count = vi->curr_queue_pairs;
1197 	channels->max_combined = vi->max_queue_pairs;
1198 	channels->max_other = 0;
1199 	channels->rx_count = 0;
1200 	channels->tx_count = 0;
1201 	channels->other_count = 0;
1202 }
1203 
1204 static const struct ethtool_ops virtnet_ethtool_ops = {
1205 	.get_drvinfo = virtnet_get_drvinfo,
1206 	.get_link = ethtool_op_get_link,
1207 	.get_ringparam = virtnet_get_ringparam,
1208 	.set_channels = virtnet_set_channels,
1209 	.get_channels = virtnet_get_channels,
1210 };
1211 
1212 #define MIN_MTU 68
1213 #define MAX_MTU 65535
1214 
1215 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1216 {
1217 	if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1218 		return -EINVAL;
1219 	dev->mtu = new_mtu;
1220 	return 0;
1221 }
1222 
1223 /* To avoid contending a lock hold by a vcpu who would exit to host, select the
1224  * txq based on the processor id.
1225  */
1226 static u16 virtnet_select_queue(struct net_device *dev, struct sk_buff *skb)
1227 {
1228 	int txq;
1229 	struct virtnet_info *vi = netdev_priv(dev);
1230 
1231 	if (skb_rx_queue_recorded(skb)) {
1232 		txq = skb_get_rx_queue(skb);
1233 	} else {
1234 		txq = *__this_cpu_ptr(vi->vq_index);
1235 		if (txq == -1)
1236 			txq = 0;
1237 	}
1238 
1239 	while (unlikely(txq >= dev->real_num_tx_queues))
1240 		txq -= dev->real_num_tx_queues;
1241 
1242 	return txq;
1243 }
1244 
1245 static const struct net_device_ops virtnet_netdev = {
1246 	.ndo_open            = virtnet_open,
1247 	.ndo_stop   	     = virtnet_close,
1248 	.ndo_start_xmit      = start_xmit,
1249 	.ndo_validate_addr   = eth_validate_addr,
1250 	.ndo_set_mac_address = virtnet_set_mac_address,
1251 	.ndo_set_rx_mode     = virtnet_set_rx_mode,
1252 	.ndo_change_mtu	     = virtnet_change_mtu,
1253 	.ndo_get_stats64     = virtnet_stats,
1254 	.ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1255 	.ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1256 	.ndo_select_queue     = virtnet_select_queue,
1257 #ifdef CONFIG_NET_POLL_CONTROLLER
1258 	.ndo_poll_controller = virtnet_netpoll,
1259 #endif
1260 };
1261 
1262 static void virtnet_config_changed_work(struct work_struct *work)
1263 {
1264 	struct virtnet_info *vi =
1265 		container_of(work, struct virtnet_info, config_work);
1266 	u16 v;
1267 
1268 	mutex_lock(&vi->config_lock);
1269 	if (!vi->config_enable)
1270 		goto done;
1271 
1272 	if (virtio_config_val(vi->vdev, VIRTIO_NET_F_STATUS,
1273 			      offsetof(struct virtio_net_config, status),
1274 			      &v) < 0)
1275 		goto done;
1276 
1277 	if (v & VIRTIO_NET_S_ANNOUNCE) {
1278 		netdev_notify_peers(vi->dev);
1279 		virtnet_ack_link_announce(vi);
1280 	}
1281 
1282 	/* Ignore unknown (future) status bits */
1283 	v &= VIRTIO_NET_S_LINK_UP;
1284 
1285 	if (vi->status == v)
1286 		goto done;
1287 
1288 	vi->status = v;
1289 
1290 	if (vi->status & VIRTIO_NET_S_LINK_UP) {
1291 		netif_carrier_on(vi->dev);
1292 		netif_tx_wake_all_queues(vi->dev);
1293 	} else {
1294 		netif_carrier_off(vi->dev);
1295 		netif_tx_stop_all_queues(vi->dev);
1296 	}
1297 done:
1298 	mutex_unlock(&vi->config_lock);
1299 }
1300 
1301 static void virtnet_config_changed(struct virtio_device *vdev)
1302 {
1303 	struct virtnet_info *vi = vdev->priv;
1304 
1305 	schedule_work(&vi->config_work);
1306 }
1307 
1308 static void virtnet_free_queues(struct virtnet_info *vi)
1309 {
1310 	kfree(vi->rq);
1311 	kfree(vi->sq);
1312 }
1313 
1314 static void free_receive_bufs(struct virtnet_info *vi)
1315 {
1316 	int i;
1317 
1318 	for (i = 0; i < vi->max_queue_pairs; i++) {
1319 		while (vi->rq[i].pages)
1320 			__free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1321 	}
1322 }
1323 
1324 static void free_unused_bufs(struct virtnet_info *vi)
1325 {
1326 	void *buf;
1327 	int i;
1328 
1329 	for (i = 0; i < vi->max_queue_pairs; i++) {
1330 		struct virtqueue *vq = vi->sq[i].vq;
1331 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1332 			dev_kfree_skb(buf);
1333 	}
1334 
1335 	for (i = 0; i < vi->max_queue_pairs; i++) {
1336 		struct virtqueue *vq = vi->rq[i].vq;
1337 
1338 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1339 			if (vi->mergeable_rx_bufs || vi->big_packets)
1340 				give_pages(&vi->rq[i], buf);
1341 			else
1342 				dev_kfree_skb(buf);
1343 			--vi->rq[i].num;
1344 		}
1345 		BUG_ON(vi->rq[i].num != 0);
1346 	}
1347 }
1348 
1349 static void virtnet_del_vqs(struct virtnet_info *vi)
1350 {
1351 	struct virtio_device *vdev = vi->vdev;
1352 
1353 	virtnet_clean_affinity(vi, -1);
1354 
1355 	vdev->config->del_vqs(vdev);
1356 
1357 	virtnet_free_queues(vi);
1358 }
1359 
1360 static int virtnet_find_vqs(struct virtnet_info *vi)
1361 {
1362 	vq_callback_t **callbacks;
1363 	struct virtqueue **vqs;
1364 	int ret = -ENOMEM;
1365 	int i, total_vqs;
1366 	const char **names;
1367 
1368 	/* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1369 	 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1370 	 * possible control vq.
1371 	 */
1372 	total_vqs = vi->max_queue_pairs * 2 +
1373 		    virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1374 
1375 	/* Allocate space for find_vqs parameters */
1376 	vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1377 	if (!vqs)
1378 		goto err_vq;
1379 	callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1380 	if (!callbacks)
1381 		goto err_callback;
1382 	names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1383 	if (!names)
1384 		goto err_names;
1385 
1386 	/* Parameters for control virtqueue, if any */
1387 	if (vi->has_cvq) {
1388 		callbacks[total_vqs - 1] = NULL;
1389 		names[total_vqs - 1] = "control";
1390 	}
1391 
1392 	/* Allocate/initialize parameters for send/receive virtqueues */
1393 	for (i = 0; i < vi->max_queue_pairs; i++) {
1394 		callbacks[rxq2vq(i)] = skb_recv_done;
1395 		callbacks[txq2vq(i)] = skb_xmit_done;
1396 		sprintf(vi->rq[i].name, "input.%d", i);
1397 		sprintf(vi->sq[i].name, "output.%d", i);
1398 		names[rxq2vq(i)] = vi->rq[i].name;
1399 		names[txq2vq(i)] = vi->sq[i].name;
1400 	}
1401 
1402 	ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1403 					 names);
1404 	if (ret)
1405 		goto err_find;
1406 
1407 	if (vi->has_cvq) {
1408 		vi->cvq = vqs[total_vqs - 1];
1409 		if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1410 			vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1411 	}
1412 
1413 	for (i = 0; i < vi->max_queue_pairs; i++) {
1414 		vi->rq[i].vq = vqs[rxq2vq(i)];
1415 		vi->sq[i].vq = vqs[txq2vq(i)];
1416 	}
1417 
1418 	kfree(names);
1419 	kfree(callbacks);
1420 	kfree(vqs);
1421 
1422 	return 0;
1423 
1424 err_find:
1425 	kfree(names);
1426 err_names:
1427 	kfree(callbacks);
1428 err_callback:
1429 	kfree(vqs);
1430 err_vq:
1431 	return ret;
1432 }
1433 
1434 static int virtnet_alloc_queues(struct virtnet_info *vi)
1435 {
1436 	int i;
1437 
1438 	vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1439 	if (!vi->sq)
1440 		goto err_sq;
1441 	vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1442 	if (!vi->rq)
1443 		goto err_rq;
1444 
1445 	INIT_DELAYED_WORK(&vi->refill, refill_work);
1446 	for (i = 0; i < vi->max_queue_pairs; i++) {
1447 		vi->rq[i].pages = NULL;
1448 		netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1449 			       napi_weight);
1450 
1451 		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1452 		sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1453 	}
1454 
1455 	return 0;
1456 
1457 err_rq:
1458 	kfree(vi->sq);
1459 err_sq:
1460 	return -ENOMEM;
1461 }
1462 
1463 static int init_vqs(struct virtnet_info *vi)
1464 {
1465 	int ret;
1466 
1467 	/* Allocate send & receive queues */
1468 	ret = virtnet_alloc_queues(vi);
1469 	if (ret)
1470 		goto err;
1471 
1472 	ret = virtnet_find_vqs(vi);
1473 	if (ret)
1474 		goto err_free;
1475 
1476 	get_online_cpus();
1477 	virtnet_set_affinity(vi);
1478 	put_online_cpus();
1479 
1480 	return 0;
1481 
1482 err_free:
1483 	virtnet_free_queues(vi);
1484 err:
1485 	return ret;
1486 }
1487 
1488 static int virtnet_probe(struct virtio_device *vdev)
1489 {
1490 	int i, err;
1491 	struct net_device *dev;
1492 	struct virtnet_info *vi;
1493 	u16 max_queue_pairs;
1494 
1495 	/* Find if host supports multiqueue virtio_net device */
1496 	err = virtio_config_val(vdev, VIRTIO_NET_F_MQ,
1497 				offsetof(struct virtio_net_config,
1498 				max_virtqueue_pairs), &max_queue_pairs);
1499 
1500 	/* We need at least 2 queue's */
1501 	if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1502 	    max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1503 	    !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1504 		max_queue_pairs = 1;
1505 
1506 	/* Allocate ourselves a network device with room for our info */
1507 	dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1508 	if (!dev)
1509 		return -ENOMEM;
1510 
1511 	/* Set up network device as normal. */
1512 	dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1513 	dev->netdev_ops = &virtnet_netdev;
1514 	dev->features = NETIF_F_HIGHDMA;
1515 
1516 	SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1517 	SET_NETDEV_DEV(dev, &vdev->dev);
1518 
1519 	/* Do we support "hardware" checksums? */
1520 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1521 		/* This opens up the world of extra features. */
1522 		dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1523 		if (csum)
1524 			dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1525 
1526 		if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1527 			dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1528 				| NETIF_F_TSO_ECN | NETIF_F_TSO6;
1529 		}
1530 		/* Individual feature bits: what can host handle? */
1531 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1532 			dev->hw_features |= NETIF_F_TSO;
1533 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1534 			dev->hw_features |= NETIF_F_TSO6;
1535 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1536 			dev->hw_features |= NETIF_F_TSO_ECN;
1537 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1538 			dev->hw_features |= NETIF_F_UFO;
1539 
1540 		if (gso)
1541 			dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1542 		/* (!csum && gso) case will be fixed by register_netdev() */
1543 	}
1544 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1545 		dev->features |= NETIF_F_RXCSUM;
1546 
1547 	dev->vlan_features = dev->features;
1548 
1549 	/* Configuration may specify what MAC to use.  Otherwise random. */
1550 	if (virtio_config_val_len(vdev, VIRTIO_NET_F_MAC,
1551 				  offsetof(struct virtio_net_config, mac),
1552 				  dev->dev_addr, dev->addr_len) < 0)
1553 		eth_hw_addr_random(dev);
1554 
1555 	/* Set up our device-specific information */
1556 	vi = netdev_priv(dev);
1557 	vi->dev = dev;
1558 	vi->vdev = vdev;
1559 	vdev->priv = vi;
1560 	vi->stats = alloc_percpu(struct virtnet_stats);
1561 	err = -ENOMEM;
1562 	if (vi->stats == NULL)
1563 		goto free;
1564 
1565 	vi->vq_index = alloc_percpu(int);
1566 	if (vi->vq_index == NULL)
1567 		goto free_stats;
1568 
1569 	mutex_init(&vi->config_lock);
1570 	vi->config_enable = true;
1571 	INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1572 
1573 	/* If we can receive ANY GSO packets, we must allocate large ones. */
1574 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1575 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1576 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1577 		vi->big_packets = true;
1578 
1579 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1580 		vi->mergeable_rx_bufs = true;
1581 
1582 	if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1583 		vi->any_header_sg = true;
1584 
1585 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1586 		vi->has_cvq = true;
1587 
1588 	/* Use single tx/rx queue pair as default */
1589 	vi->curr_queue_pairs = 1;
1590 	vi->max_queue_pairs = max_queue_pairs;
1591 
1592 	/* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1593 	err = init_vqs(vi);
1594 	if (err)
1595 		goto free_index;
1596 
1597 	netif_set_real_num_tx_queues(dev, 1);
1598 	netif_set_real_num_rx_queues(dev, 1);
1599 
1600 	err = register_netdev(dev);
1601 	if (err) {
1602 		pr_debug("virtio_net: registering device failed\n");
1603 		goto free_vqs;
1604 	}
1605 
1606 	/* Last of all, set up some receive buffers. */
1607 	for (i = 0; i < vi->curr_queue_pairs; i++) {
1608 		try_fill_recv(&vi->rq[i], GFP_KERNEL);
1609 
1610 		/* If we didn't even get one input buffer, we're useless. */
1611 		if (vi->rq[i].num == 0) {
1612 			free_unused_bufs(vi);
1613 			err = -ENOMEM;
1614 			goto free_recv_bufs;
1615 		}
1616 	}
1617 
1618 	vi->nb.notifier_call = &virtnet_cpu_callback;
1619 	err = register_hotcpu_notifier(&vi->nb);
1620 	if (err) {
1621 		pr_debug("virtio_net: registering cpu notifier failed\n");
1622 		goto free_recv_bufs;
1623 	}
1624 
1625 	/* Assume link up if device can't report link status,
1626 	   otherwise get link status from config. */
1627 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1628 		netif_carrier_off(dev);
1629 		schedule_work(&vi->config_work);
1630 	} else {
1631 		vi->status = VIRTIO_NET_S_LINK_UP;
1632 		netif_carrier_on(dev);
1633 	}
1634 
1635 	pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1636 		 dev->name, max_queue_pairs);
1637 
1638 	return 0;
1639 
1640 free_recv_bufs:
1641 	free_receive_bufs(vi);
1642 	unregister_netdev(dev);
1643 free_vqs:
1644 	cancel_delayed_work_sync(&vi->refill);
1645 	virtnet_del_vqs(vi);
1646 free_index:
1647 	free_percpu(vi->vq_index);
1648 free_stats:
1649 	free_percpu(vi->stats);
1650 free:
1651 	free_netdev(dev);
1652 	return err;
1653 }
1654 
1655 static void remove_vq_common(struct virtnet_info *vi)
1656 {
1657 	vi->vdev->config->reset(vi->vdev);
1658 
1659 	/* Free unused buffers in both send and recv, if any. */
1660 	free_unused_bufs(vi);
1661 
1662 	free_receive_bufs(vi);
1663 
1664 	virtnet_del_vqs(vi);
1665 }
1666 
1667 static void virtnet_remove(struct virtio_device *vdev)
1668 {
1669 	struct virtnet_info *vi = vdev->priv;
1670 
1671 	unregister_hotcpu_notifier(&vi->nb);
1672 
1673 	/* Prevent config work handler from accessing the device. */
1674 	mutex_lock(&vi->config_lock);
1675 	vi->config_enable = false;
1676 	mutex_unlock(&vi->config_lock);
1677 
1678 	unregister_netdev(vi->dev);
1679 
1680 	remove_vq_common(vi);
1681 
1682 	flush_work(&vi->config_work);
1683 
1684 	free_percpu(vi->vq_index);
1685 	free_percpu(vi->stats);
1686 	free_netdev(vi->dev);
1687 }
1688 
1689 #ifdef CONFIG_PM
1690 static int virtnet_freeze(struct virtio_device *vdev)
1691 {
1692 	struct virtnet_info *vi = vdev->priv;
1693 	int i;
1694 
1695 	unregister_hotcpu_notifier(&vi->nb);
1696 
1697 	/* Prevent config work handler from accessing the device */
1698 	mutex_lock(&vi->config_lock);
1699 	vi->config_enable = false;
1700 	mutex_unlock(&vi->config_lock);
1701 
1702 	netif_device_detach(vi->dev);
1703 	cancel_delayed_work_sync(&vi->refill);
1704 
1705 	if (netif_running(vi->dev))
1706 		for (i = 0; i < vi->max_queue_pairs; i++) {
1707 			napi_disable(&vi->rq[i].napi);
1708 			netif_napi_del(&vi->rq[i].napi);
1709 		}
1710 
1711 	remove_vq_common(vi);
1712 
1713 	flush_work(&vi->config_work);
1714 
1715 	return 0;
1716 }
1717 
1718 static int virtnet_restore(struct virtio_device *vdev)
1719 {
1720 	struct virtnet_info *vi = vdev->priv;
1721 	int err, i;
1722 
1723 	err = init_vqs(vi);
1724 	if (err)
1725 		return err;
1726 
1727 	if (netif_running(vi->dev))
1728 		for (i = 0; i < vi->max_queue_pairs; i++)
1729 			virtnet_napi_enable(&vi->rq[i]);
1730 
1731 	netif_device_attach(vi->dev);
1732 
1733 	for (i = 0; i < vi->curr_queue_pairs; i++)
1734 		if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1735 			schedule_delayed_work(&vi->refill, 0);
1736 
1737 	mutex_lock(&vi->config_lock);
1738 	vi->config_enable = true;
1739 	mutex_unlock(&vi->config_lock);
1740 
1741 	rtnl_lock();
1742 	virtnet_set_queues(vi, vi->curr_queue_pairs);
1743 	rtnl_unlock();
1744 
1745 	err = register_hotcpu_notifier(&vi->nb);
1746 	if (err)
1747 		return err;
1748 
1749 	return 0;
1750 }
1751 #endif
1752 
1753 static struct virtio_device_id id_table[] = {
1754 	{ VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1755 	{ 0 },
1756 };
1757 
1758 static unsigned int features[] = {
1759 	VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1760 	VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1761 	VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1762 	VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1763 	VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1764 	VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1765 	VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1766 	VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1767 	VIRTIO_NET_F_CTRL_MAC_ADDR,
1768 	VIRTIO_F_ANY_LAYOUT,
1769 };
1770 
1771 static struct virtio_driver virtio_net_driver = {
1772 	.feature_table = features,
1773 	.feature_table_size = ARRAY_SIZE(features),
1774 	.driver.name =	KBUILD_MODNAME,
1775 	.driver.owner =	THIS_MODULE,
1776 	.id_table =	id_table,
1777 	.probe =	virtnet_probe,
1778 	.remove =	virtnet_remove,
1779 	.config_changed = virtnet_config_changed,
1780 #ifdef CONFIG_PM
1781 	.freeze =	virtnet_freeze,
1782 	.restore =	virtnet_restore,
1783 #endif
1784 };
1785 
1786 module_virtio_driver(virtio_net_driver);
1787 
1788 MODULE_DEVICE_TABLE(virtio, id_table);
1789 MODULE_DESCRIPTION("Virtio network driver");
1790 MODULE_LICENSE("GPL");
1791