xref: /openbmc/linux/drivers/net/virtio_net.c (revision 24b1944f)
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->max_queue_pairs; i++) {
640 		if (i < vi->curr_queue_pairs)
641 			/* Make sure we have some buffers: if oom use wq. */
642 			if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
643 				schedule_delayed_work(&vi->refill, 0);
644 		virtnet_napi_enable(&vi->rq[i]);
645 	}
646 
647 	return 0;
648 }
649 
650 static void free_old_xmit_skbs(struct send_queue *sq)
651 {
652 	struct sk_buff *skb;
653 	unsigned int len;
654 	struct virtnet_info *vi = sq->vq->vdev->priv;
655 	struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
656 
657 	while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
658 		pr_debug("Sent skb %p\n", skb);
659 
660 		u64_stats_update_begin(&stats->tx_syncp);
661 		stats->tx_bytes += skb->len;
662 		stats->tx_packets++;
663 		u64_stats_update_end(&stats->tx_syncp);
664 
665 		dev_kfree_skb_any(skb);
666 	}
667 }
668 
669 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
670 {
671 	struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
672 	const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
673 	struct virtnet_info *vi = sq->vq->vdev->priv;
674 	unsigned num_sg;
675 
676 	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
677 
678 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
679 		hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
680 		hdr->hdr.csum_start = skb_checksum_start_offset(skb);
681 		hdr->hdr.csum_offset = skb->csum_offset;
682 	} else {
683 		hdr->hdr.flags = 0;
684 		hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
685 	}
686 
687 	if (skb_is_gso(skb)) {
688 		hdr->hdr.hdr_len = skb_headlen(skb);
689 		hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
690 		if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
691 			hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
692 		else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
693 			hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
694 		else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
695 			hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
696 		else
697 			BUG();
698 		if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
699 			hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
700 	} else {
701 		hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
702 		hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
703 	}
704 
705 	hdr->mhdr.num_buffers = 0;
706 
707 	/* Encode metadata header at front. */
708 	if (vi->mergeable_rx_bufs)
709 		sg_set_buf(sq->sg, &hdr->mhdr, sizeof hdr->mhdr);
710 	else
711 		sg_set_buf(sq->sg, &hdr->hdr, sizeof hdr->hdr);
712 
713 	num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
714 	return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
715 }
716 
717 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
718 {
719 	struct virtnet_info *vi = netdev_priv(dev);
720 	int qnum = skb_get_queue_mapping(skb);
721 	struct send_queue *sq = &vi->sq[qnum];
722 	int err;
723 
724 	/* Free up any pending old buffers before queueing new ones. */
725 	free_old_xmit_skbs(sq);
726 
727 	/* Try to transmit */
728 	err = xmit_skb(sq, skb);
729 
730 	/* This should not happen! */
731 	if (unlikely(err)) {
732 		dev->stats.tx_fifo_errors++;
733 		if (net_ratelimit())
734 			dev_warn(&dev->dev,
735 				 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
736 		dev->stats.tx_dropped++;
737 		kfree_skb(skb);
738 		return NETDEV_TX_OK;
739 	}
740 	virtqueue_kick(sq->vq);
741 
742 	/* Don't wait up for transmitted skbs to be freed. */
743 	skb_orphan(skb);
744 	nf_reset(skb);
745 
746 	/* Apparently nice girls don't return TX_BUSY; stop the queue
747 	 * before it gets out of hand.  Naturally, this wastes entries. */
748 	if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
749 		netif_stop_subqueue(dev, qnum);
750 		if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
751 			/* More just got used, free them then recheck. */
752 			free_old_xmit_skbs(sq);
753 			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
754 				netif_start_subqueue(dev, qnum);
755 				virtqueue_disable_cb(sq->vq);
756 			}
757 		}
758 	}
759 
760 	return NETDEV_TX_OK;
761 }
762 
763 /*
764  * Send command via the control virtqueue and check status.  Commands
765  * supported by the hypervisor, as indicated by feature bits, should
766  * never fail unless improperly formated.
767  */
768 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
769 				 struct scatterlist *out,
770 				 struct scatterlist *in)
771 {
772 	struct scatterlist *sgs[4], hdr, stat;
773 	struct virtio_net_ctrl_hdr ctrl;
774 	virtio_net_ctrl_ack status = ~0;
775 	unsigned out_num = 0, in_num = 0, tmp;
776 
777 	/* Caller should know better */
778 	BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
779 
780 	ctrl.class = class;
781 	ctrl.cmd = cmd;
782 	/* Add header */
783 	sg_init_one(&hdr, &ctrl, sizeof(ctrl));
784 	sgs[out_num++] = &hdr;
785 
786 	if (out)
787 		sgs[out_num++] = out;
788 	if (in)
789 		sgs[out_num + in_num++] = in;
790 
791 	/* Add return status. */
792 	sg_init_one(&stat, &status, sizeof(status));
793 	sgs[out_num + in_num++] = &stat;
794 
795 	BUG_ON(out_num + in_num > ARRAY_SIZE(sgs));
796 	BUG_ON(virtqueue_add_sgs(vi->cvq, sgs, out_num, in_num, vi, GFP_ATOMIC)
797 	       < 0);
798 
799 	virtqueue_kick(vi->cvq);
800 
801 	/* Spin for a response, the kick causes an ioport write, trapping
802 	 * into the hypervisor, so the request should be handled immediately.
803 	 */
804 	while (!virtqueue_get_buf(vi->cvq, &tmp))
805 		cpu_relax();
806 
807 	return status == VIRTIO_NET_OK;
808 }
809 
810 static int virtnet_set_mac_address(struct net_device *dev, void *p)
811 {
812 	struct virtnet_info *vi = netdev_priv(dev);
813 	struct virtio_device *vdev = vi->vdev;
814 	int ret;
815 	struct sockaddr *addr = p;
816 	struct scatterlist sg;
817 
818 	ret = eth_prepare_mac_addr_change(dev, p);
819 	if (ret)
820 		return ret;
821 
822 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
823 		sg_init_one(&sg, addr->sa_data, dev->addr_len);
824 		if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
825 					  VIRTIO_NET_CTRL_MAC_ADDR_SET,
826 					  &sg, NULL)) {
827 			dev_warn(&vdev->dev,
828 				 "Failed to set mac address by vq command.\n");
829 			return -EINVAL;
830 		}
831 	} else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
832 		vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
833 				  addr->sa_data, dev->addr_len);
834 	}
835 
836 	eth_commit_mac_addr_change(dev, p);
837 
838 	return 0;
839 }
840 
841 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
842 					       struct rtnl_link_stats64 *tot)
843 {
844 	struct virtnet_info *vi = netdev_priv(dev);
845 	int cpu;
846 	unsigned int start;
847 
848 	for_each_possible_cpu(cpu) {
849 		struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
850 		u64 tpackets, tbytes, rpackets, rbytes;
851 
852 		do {
853 			start = u64_stats_fetch_begin_bh(&stats->tx_syncp);
854 			tpackets = stats->tx_packets;
855 			tbytes   = stats->tx_bytes;
856 		} while (u64_stats_fetch_retry_bh(&stats->tx_syncp, start));
857 
858 		do {
859 			start = u64_stats_fetch_begin_bh(&stats->rx_syncp);
860 			rpackets = stats->rx_packets;
861 			rbytes   = stats->rx_bytes;
862 		} while (u64_stats_fetch_retry_bh(&stats->rx_syncp, start));
863 
864 		tot->rx_packets += rpackets;
865 		tot->tx_packets += tpackets;
866 		tot->rx_bytes   += rbytes;
867 		tot->tx_bytes   += tbytes;
868 	}
869 
870 	tot->tx_dropped = dev->stats.tx_dropped;
871 	tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
872 	tot->rx_dropped = dev->stats.rx_dropped;
873 	tot->rx_length_errors = dev->stats.rx_length_errors;
874 	tot->rx_frame_errors = dev->stats.rx_frame_errors;
875 
876 	return tot;
877 }
878 
879 #ifdef CONFIG_NET_POLL_CONTROLLER
880 static void virtnet_netpoll(struct net_device *dev)
881 {
882 	struct virtnet_info *vi = netdev_priv(dev);
883 	int i;
884 
885 	for (i = 0; i < vi->curr_queue_pairs; i++)
886 		napi_schedule(&vi->rq[i].napi);
887 }
888 #endif
889 
890 static void virtnet_ack_link_announce(struct virtnet_info *vi)
891 {
892 	rtnl_lock();
893 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
894 				  VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL, NULL))
895 		dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
896 	rtnl_unlock();
897 }
898 
899 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
900 {
901 	struct scatterlist sg;
902 	struct virtio_net_ctrl_mq s;
903 	struct net_device *dev = vi->dev;
904 	int i;
905 
906 	if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
907 		return 0;
908 
909 	s.virtqueue_pairs = queue_pairs;
910 	sg_init_one(&sg, &s, sizeof(s));
911 
912 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
913 				  VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg, NULL)) {
914 		dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
915 			 queue_pairs);
916 		return -EINVAL;
917 	} else {
918 		for (i = vi->curr_queue_pairs; i < queue_pairs; i++)
919 			if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
920 				schedule_delayed_work(&vi->refill, 0);
921 		vi->curr_queue_pairs = queue_pairs;
922 	}
923 
924 	return 0;
925 }
926 
927 static int virtnet_close(struct net_device *dev)
928 {
929 	struct virtnet_info *vi = netdev_priv(dev);
930 	int i;
931 
932 	/* Make sure refill_work doesn't re-enable napi! */
933 	cancel_delayed_work_sync(&vi->refill);
934 
935 	for (i = 0; i < vi->max_queue_pairs; i++)
936 		napi_disable(&vi->rq[i].napi);
937 
938 	return 0;
939 }
940 
941 static void virtnet_set_rx_mode(struct net_device *dev)
942 {
943 	struct virtnet_info *vi = netdev_priv(dev);
944 	struct scatterlist sg[2];
945 	u8 promisc, allmulti;
946 	struct virtio_net_ctrl_mac *mac_data;
947 	struct netdev_hw_addr *ha;
948 	int uc_count;
949 	int mc_count;
950 	void *buf;
951 	int i;
952 
953 	/* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
954 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
955 		return;
956 
957 	promisc = ((dev->flags & IFF_PROMISC) != 0);
958 	allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
959 
960 	sg_init_one(sg, &promisc, sizeof(promisc));
961 
962 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
963 				  VIRTIO_NET_CTRL_RX_PROMISC,
964 				  sg, NULL))
965 		dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
966 			 promisc ? "en" : "dis");
967 
968 	sg_init_one(sg, &allmulti, sizeof(allmulti));
969 
970 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
971 				  VIRTIO_NET_CTRL_RX_ALLMULTI,
972 				  sg, NULL))
973 		dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
974 			 allmulti ? "en" : "dis");
975 
976 	uc_count = netdev_uc_count(dev);
977 	mc_count = netdev_mc_count(dev);
978 	/* MAC filter - use one buffer for both lists */
979 	buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
980 		      (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
981 	mac_data = buf;
982 	if (!buf)
983 		return;
984 
985 	sg_init_table(sg, 2);
986 
987 	/* Store the unicast list and count in the front of the buffer */
988 	mac_data->entries = uc_count;
989 	i = 0;
990 	netdev_for_each_uc_addr(ha, dev)
991 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
992 
993 	sg_set_buf(&sg[0], mac_data,
994 		   sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
995 
996 	/* multicast list and count fill the end */
997 	mac_data = (void *)&mac_data->macs[uc_count][0];
998 
999 	mac_data->entries = mc_count;
1000 	i = 0;
1001 	netdev_for_each_mc_addr(ha, dev)
1002 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1003 
1004 	sg_set_buf(&sg[1], mac_data,
1005 		   sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1006 
1007 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1008 				  VIRTIO_NET_CTRL_MAC_TABLE_SET,
1009 				  sg, NULL))
1010 		dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
1011 
1012 	kfree(buf);
1013 }
1014 
1015 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1016 				   __be16 proto, u16 vid)
1017 {
1018 	struct virtnet_info *vi = netdev_priv(dev);
1019 	struct scatterlist sg;
1020 
1021 	sg_init_one(&sg, &vid, sizeof(vid));
1022 
1023 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1024 				  VIRTIO_NET_CTRL_VLAN_ADD, &sg, NULL))
1025 		dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1026 	return 0;
1027 }
1028 
1029 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1030 				    __be16 proto, u16 vid)
1031 {
1032 	struct virtnet_info *vi = netdev_priv(dev);
1033 	struct scatterlist sg;
1034 
1035 	sg_init_one(&sg, &vid, sizeof(vid));
1036 
1037 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1038 				  VIRTIO_NET_CTRL_VLAN_DEL, &sg, NULL))
1039 		dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1040 	return 0;
1041 }
1042 
1043 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1044 {
1045 	int i;
1046 	int cpu;
1047 
1048 	if (vi->affinity_hint_set) {
1049 		for (i = 0; i < vi->max_queue_pairs; i++) {
1050 			virtqueue_set_affinity(vi->rq[i].vq, -1);
1051 			virtqueue_set_affinity(vi->sq[i].vq, -1);
1052 		}
1053 
1054 		vi->affinity_hint_set = false;
1055 	}
1056 
1057 	i = 0;
1058 	for_each_online_cpu(cpu) {
1059 		if (cpu == hcpu) {
1060 			*per_cpu_ptr(vi->vq_index, cpu) = -1;
1061 		} else {
1062 			*per_cpu_ptr(vi->vq_index, cpu) =
1063 				++i % vi->curr_queue_pairs;
1064 		}
1065 	}
1066 }
1067 
1068 static void virtnet_set_affinity(struct virtnet_info *vi)
1069 {
1070 	int i;
1071 	int cpu;
1072 
1073 	/* In multiqueue mode, when the number of cpu is equal to the number of
1074 	 * queue pairs, we let the queue pairs to be private to one cpu by
1075 	 * setting the affinity hint to eliminate the contention.
1076 	 */
1077 	if (vi->curr_queue_pairs == 1 ||
1078 	    vi->max_queue_pairs != num_online_cpus()) {
1079 		virtnet_clean_affinity(vi, -1);
1080 		return;
1081 	}
1082 
1083 	i = 0;
1084 	for_each_online_cpu(cpu) {
1085 		virtqueue_set_affinity(vi->rq[i].vq, cpu);
1086 		virtqueue_set_affinity(vi->sq[i].vq, cpu);
1087 		*per_cpu_ptr(vi->vq_index, cpu) = i;
1088 		i++;
1089 	}
1090 
1091 	vi->affinity_hint_set = true;
1092 }
1093 
1094 static int virtnet_cpu_callback(struct notifier_block *nfb,
1095 			        unsigned long action, void *hcpu)
1096 {
1097 	struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1098 
1099 	switch(action & ~CPU_TASKS_FROZEN) {
1100 	case CPU_ONLINE:
1101 	case CPU_DOWN_FAILED:
1102 	case CPU_DEAD:
1103 		virtnet_set_affinity(vi);
1104 		break;
1105 	case CPU_DOWN_PREPARE:
1106 		virtnet_clean_affinity(vi, (long)hcpu);
1107 		break;
1108 	default:
1109 		break;
1110 	}
1111 	return NOTIFY_OK;
1112 }
1113 
1114 static void virtnet_get_ringparam(struct net_device *dev,
1115 				struct ethtool_ringparam *ring)
1116 {
1117 	struct virtnet_info *vi = netdev_priv(dev);
1118 
1119 	ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1120 	ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1121 	ring->rx_pending = ring->rx_max_pending;
1122 	ring->tx_pending = ring->tx_max_pending;
1123 }
1124 
1125 
1126 static void virtnet_get_drvinfo(struct net_device *dev,
1127 				struct ethtool_drvinfo *info)
1128 {
1129 	struct virtnet_info *vi = netdev_priv(dev);
1130 	struct virtio_device *vdev = vi->vdev;
1131 
1132 	strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1133 	strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1134 	strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1135 
1136 }
1137 
1138 /* TODO: Eliminate OOO packets during switching */
1139 static int virtnet_set_channels(struct net_device *dev,
1140 				struct ethtool_channels *channels)
1141 {
1142 	struct virtnet_info *vi = netdev_priv(dev);
1143 	u16 queue_pairs = channels->combined_count;
1144 	int err;
1145 
1146 	/* We don't support separate rx/tx channels.
1147 	 * We don't allow setting 'other' channels.
1148 	 */
1149 	if (channels->rx_count || channels->tx_count || channels->other_count)
1150 		return -EINVAL;
1151 
1152 	if (queue_pairs > vi->max_queue_pairs)
1153 		return -EINVAL;
1154 
1155 	get_online_cpus();
1156 	err = virtnet_set_queues(vi, queue_pairs);
1157 	if (!err) {
1158 		netif_set_real_num_tx_queues(dev, queue_pairs);
1159 		netif_set_real_num_rx_queues(dev, queue_pairs);
1160 
1161 		virtnet_set_affinity(vi);
1162 	}
1163 	put_online_cpus();
1164 
1165 	return err;
1166 }
1167 
1168 static void virtnet_get_channels(struct net_device *dev,
1169 				 struct ethtool_channels *channels)
1170 {
1171 	struct virtnet_info *vi = netdev_priv(dev);
1172 
1173 	channels->combined_count = vi->curr_queue_pairs;
1174 	channels->max_combined = vi->max_queue_pairs;
1175 	channels->max_other = 0;
1176 	channels->rx_count = 0;
1177 	channels->tx_count = 0;
1178 	channels->other_count = 0;
1179 }
1180 
1181 static const struct ethtool_ops virtnet_ethtool_ops = {
1182 	.get_drvinfo = virtnet_get_drvinfo,
1183 	.get_link = ethtool_op_get_link,
1184 	.get_ringparam = virtnet_get_ringparam,
1185 	.set_channels = virtnet_set_channels,
1186 	.get_channels = virtnet_get_channels,
1187 };
1188 
1189 #define MIN_MTU 68
1190 #define MAX_MTU 65535
1191 
1192 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1193 {
1194 	if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1195 		return -EINVAL;
1196 	dev->mtu = new_mtu;
1197 	return 0;
1198 }
1199 
1200 /* To avoid contending a lock hold by a vcpu who would exit to host, select the
1201  * txq based on the processor id.
1202  */
1203 static u16 virtnet_select_queue(struct net_device *dev, struct sk_buff *skb)
1204 {
1205 	int txq;
1206 	struct virtnet_info *vi = netdev_priv(dev);
1207 
1208 	if (skb_rx_queue_recorded(skb)) {
1209 		txq = skb_get_rx_queue(skb);
1210 	} else {
1211 		txq = *__this_cpu_ptr(vi->vq_index);
1212 		if (txq == -1)
1213 			txq = 0;
1214 	}
1215 
1216 	while (unlikely(txq >= dev->real_num_tx_queues))
1217 		txq -= dev->real_num_tx_queues;
1218 
1219 	return txq;
1220 }
1221 
1222 static const struct net_device_ops virtnet_netdev = {
1223 	.ndo_open            = virtnet_open,
1224 	.ndo_stop   	     = virtnet_close,
1225 	.ndo_start_xmit      = start_xmit,
1226 	.ndo_validate_addr   = eth_validate_addr,
1227 	.ndo_set_mac_address = virtnet_set_mac_address,
1228 	.ndo_set_rx_mode     = virtnet_set_rx_mode,
1229 	.ndo_change_mtu	     = virtnet_change_mtu,
1230 	.ndo_get_stats64     = virtnet_stats,
1231 	.ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1232 	.ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1233 	.ndo_select_queue     = virtnet_select_queue,
1234 #ifdef CONFIG_NET_POLL_CONTROLLER
1235 	.ndo_poll_controller = virtnet_netpoll,
1236 #endif
1237 };
1238 
1239 static void virtnet_config_changed_work(struct work_struct *work)
1240 {
1241 	struct virtnet_info *vi =
1242 		container_of(work, struct virtnet_info, config_work);
1243 	u16 v;
1244 
1245 	mutex_lock(&vi->config_lock);
1246 	if (!vi->config_enable)
1247 		goto done;
1248 
1249 	if (virtio_config_val(vi->vdev, VIRTIO_NET_F_STATUS,
1250 			      offsetof(struct virtio_net_config, status),
1251 			      &v) < 0)
1252 		goto done;
1253 
1254 	if (v & VIRTIO_NET_S_ANNOUNCE) {
1255 		netdev_notify_peers(vi->dev);
1256 		virtnet_ack_link_announce(vi);
1257 	}
1258 
1259 	/* Ignore unknown (future) status bits */
1260 	v &= VIRTIO_NET_S_LINK_UP;
1261 
1262 	if (vi->status == v)
1263 		goto done;
1264 
1265 	vi->status = v;
1266 
1267 	if (vi->status & VIRTIO_NET_S_LINK_UP) {
1268 		netif_carrier_on(vi->dev);
1269 		netif_tx_wake_all_queues(vi->dev);
1270 	} else {
1271 		netif_carrier_off(vi->dev);
1272 		netif_tx_stop_all_queues(vi->dev);
1273 	}
1274 done:
1275 	mutex_unlock(&vi->config_lock);
1276 }
1277 
1278 static void virtnet_config_changed(struct virtio_device *vdev)
1279 {
1280 	struct virtnet_info *vi = vdev->priv;
1281 
1282 	schedule_work(&vi->config_work);
1283 }
1284 
1285 static void virtnet_free_queues(struct virtnet_info *vi)
1286 {
1287 	kfree(vi->rq);
1288 	kfree(vi->sq);
1289 }
1290 
1291 static void free_receive_bufs(struct virtnet_info *vi)
1292 {
1293 	int i;
1294 
1295 	for (i = 0; i < vi->max_queue_pairs; i++) {
1296 		while (vi->rq[i].pages)
1297 			__free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1298 	}
1299 }
1300 
1301 static void free_unused_bufs(struct virtnet_info *vi)
1302 {
1303 	void *buf;
1304 	int i;
1305 
1306 	for (i = 0; i < vi->max_queue_pairs; i++) {
1307 		struct virtqueue *vq = vi->sq[i].vq;
1308 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1309 			dev_kfree_skb(buf);
1310 	}
1311 
1312 	for (i = 0; i < vi->max_queue_pairs; i++) {
1313 		struct virtqueue *vq = vi->rq[i].vq;
1314 
1315 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1316 			if (vi->mergeable_rx_bufs || vi->big_packets)
1317 				give_pages(&vi->rq[i], buf);
1318 			else
1319 				dev_kfree_skb(buf);
1320 			--vi->rq[i].num;
1321 		}
1322 		BUG_ON(vi->rq[i].num != 0);
1323 	}
1324 }
1325 
1326 static void virtnet_del_vqs(struct virtnet_info *vi)
1327 {
1328 	struct virtio_device *vdev = vi->vdev;
1329 
1330 	virtnet_clean_affinity(vi, -1);
1331 
1332 	vdev->config->del_vqs(vdev);
1333 
1334 	virtnet_free_queues(vi);
1335 }
1336 
1337 static int virtnet_find_vqs(struct virtnet_info *vi)
1338 {
1339 	vq_callback_t **callbacks;
1340 	struct virtqueue **vqs;
1341 	int ret = -ENOMEM;
1342 	int i, total_vqs;
1343 	const char **names;
1344 
1345 	/* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1346 	 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1347 	 * possible control vq.
1348 	 */
1349 	total_vqs = vi->max_queue_pairs * 2 +
1350 		    virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1351 
1352 	/* Allocate space for find_vqs parameters */
1353 	vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1354 	if (!vqs)
1355 		goto err_vq;
1356 	callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1357 	if (!callbacks)
1358 		goto err_callback;
1359 	names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1360 	if (!names)
1361 		goto err_names;
1362 
1363 	/* Parameters for control virtqueue, if any */
1364 	if (vi->has_cvq) {
1365 		callbacks[total_vqs - 1] = NULL;
1366 		names[total_vqs - 1] = "control";
1367 	}
1368 
1369 	/* Allocate/initialize parameters for send/receive virtqueues */
1370 	for (i = 0; i < vi->max_queue_pairs; i++) {
1371 		callbacks[rxq2vq(i)] = skb_recv_done;
1372 		callbacks[txq2vq(i)] = skb_xmit_done;
1373 		sprintf(vi->rq[i].name, "input.%d", i);
1374 		sprintf(vi->sq[i].name, "output.%d", i);
1375 		names[rxq2vq(i)] = vi->rq[i].name;
1376 		names[txq2vq(i)] = vi->sq[i].name;
1377 	}
1378 
1379 	ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1380 					 names);
1381 	if (ret)
1382 		goto err_find;
1383 
1384 	if (vi->has_cvq) {
1385 		vi->cvq = vqs[total_vqs - 1];
1386 		if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1387 			vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1388 	}
1389 
1390 	for (i = 0; i < vi->max_queue_pairs; i++) {
1391 		vi->rq[i].vq = vqs[rxq2vq(i)];
1392 		vi->sq[i].vq = vqs[txq2vq(i)];
1393 	}
1394 
1395 	kfree(names);
1396 	kfree(callbacks);
1397 	kfree(vqs);
1398 
1399 	return 0;
1400 
1401 err_find:
1402 	kfree(names);
1403 err_names:
1404 	kfree(callbacks);
1405 err_callback:
1406 	kfree(vqs);
1407 err_vq:
1408 	return ret;
1409 }
1410 
1411 static int virtnet_alloc_queues(struct virtnet_info *vi)
1412 {
1413 	int i;
1414 
1415 	vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1416 	if (!vi->sq)
1417 		goto err_sq;
1418 	vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1419 	if (!vi->rq)
1420 		goto err_rq;
1421 
1422 	INIT_DELAYED_WORK(&vi->refill, refill_work);
1423 	for (i = 0; i < vi->max_queue_pairs; i++) {
1424 		vi->rq[i].pages = NULL;
1425 		netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1426 			       napi_weight);
1427 
1428 		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1429 		sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1430 	}
1431 
1432 	return 0;
1433 
1434 err_rq:
1435 	kfree(vi->sq);
1436 err_sq:
1437 	return -ENOMEM;
1438 }
1439 
1440 static int init_vqs(struct virtnet_info *vi)
1441 {
1442 	int ret;
1443 
1444 	/* Allocate send & receive queues */
1445 	ret = virtnet_alloc_queues(vi);
1446 	if (ret)
1447 		goto err;
1448 
1449 	ret = virtnet_find_vqs(vi);
1450 	if (ret)
1451 		goto err_free;
1452 
1453 	get_online_cpus();
1454 	virtnet_set_affinity(vi);
1455 	put_online_cpus();
1456 
1457 	return 0;
1458 
1459 err_free:
1460 	virtnet_free_queues(vi);
1461 err:
1462 	return ret;
1463 }
1464 
1465 static int virtnet_probe(struct virtio_device *vdev)
1466 {
1467 	int i, err;
1468 	struct net_device *dev;
1469 	struct virtnet_info *vi;
1470 	u16 max_queue_pairs;
1471 
1472 	/* Find if host supports multiqueue virtio_net device */
1473 	err = virtio_config_val(vdev, VIRTIO_NET_F_MQ,
1474 				offsetof(struct virtio_net_config,
1475 				max_virtqueue_pairs), &max_queue_pairs);
1476 
1477 	/* We need at least 2 queue's */
1478 	if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1479 	    max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1480 	    !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1481 		max_queue_pairs = 1;
1482 
1483 	/* Allocate ourselves a network device with room for our info */
1484 	dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1485 	if (!dev)
1486 		return -ENOMEM;
1487 
1488 	/* Set up network device as normal. */
1489 	dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1490 	dev->netdev_ops = &virtnet_netdev;
1491 	dev->features = NETIF_F_HIGHDMA;
1492 
1493 	SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1494 	SET_NETDEV_DEV(dev, &vdev->dev);
1495 
1496 	/* Do we support "hardware" checksums? */
1497 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1498 		/* This opens up the world of extra features. */
1499 		dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1500 		if (csum)
1501 			dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1502 
1503 		if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1504 			dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1505 				| NETIF_F_TSO_ECN | NETIF_F_TSO6;
1506 		}
1507 		/* Individual feature bits: what can host handle? */
1508 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1509 			dev->hw_features |= NETIF_F_TSO;
1510 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1511 			dev->hw_features |= NETIF_F_TSO6;
1512 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1513 			dev->hw_features |= NETIF_F_TSO_ECN;
1514 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1515 			dev->hw_features |= NETIF_F_UFO;
1516 
1517 		if (gso)
1518 			dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1519 		/* (!csum && gso) case will be fixed by register_netdev() */
1520 	}
1521 
1522 	dev->vlan_features = dev->features;
1523 
1524 	/* Configuration may specify what MAC to use.  Otherwise random. */
1525 	if (virtio_config_val_len(vdev, VIRTIO_NET_F_MAC,
1526 				  offsetof(struct virtio_net_config, mac),
1527 				  dev->dev_addr, dev->addr_len) < 0)
1528 		eth_hw_addr_random(dev);
1529 
1530 	/* Set up our device-specific information */
1531 	vi = netdev_priv(dev);
1532 	vi->dev = dev;
1533 	vi->vdev = vdev;
1534 	vdev->priv = vi;
1535 	vi->stats = alloc_percpu(struct virtnet_stats);
1536 	err = -ENOMEM;
1537 	if (vi->stats == NULL)
1538 		goto free;
1539 
1540 	vi->vq_index = alloc_percpu(int);
1541 	if (vi->vq_index == NULL)
1542 		goto free_stats;
1543 
1544 	mutex_init(&vi->config_lock);
1545 	vi->config_enable = true;
1546 	INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1547 
1548 	/* If we can receive ANY GSO packets, we must allocate large ones. */
1549 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1550 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1551 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1552 		vi->big_packets = true;
1553 
1554 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1555 		vi->mergeable_rx_bufs = true;
1556 
1557 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1558 		vi->has_cvq = true;
1559 
1560 	/* Use single tx/rx queue pair as default */
1561 	vi->curr_queue_pairs = 1;
1562 	vi->max_queue_pairs = max_queue_pairs;
1563 
1564 	/* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1565 	err = init_vqs(vi);
1566 	if (err)
1567 		goto free_index;
1568 
1569 	netif_set_real_num_tx_queues(dev, 1);
1570 	netif_set_real_num_rx_queues(dev, 1);
1571 
1572 	err = register_netdev(dev);
1573 	if (err) {
1574 		pr_debug("virtio_net: registering device failed\n");
1575 		goto free_vqs;
1576 	}
1577 
1578 	/* Last of all, set up some receive buffers. */
1579 	for (i = 0; i < vi->curr_queue_pairs; i++) {
1580 		try_fill_recv(&vi->rq[i], GFP_KERNEL);
1581 
1582 		/* If we didn't even get one input buffer, we're useless. */
1583 		if (vi->rq[i].num == 0) {
1584 			free_unused_bufs(vi);
1585 			err = -ENOMEM;
1586 			goto free_recv_bufs;
1587 		}
1588 	}
1589 
1590 	vi->nb.notifier_call = &virtnet_cpu_callback;
1591 	err = register_hotcpu_notifier(&vi->nb);
1592 	if (err) {
1593 		pr_debug("virtio_net: registering cpu notifier failed\n");
1594 		goto free_recv_bufs;
1595 	}
1596 
1597 	/* Assume link up if device can't report link status,
1598 	   otherwise get link status from config. */
1599 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1600 		netif_carrier_off(dev);
1601 		schedule_work(&vi->config_work);
1602 	} else {
1603 		vi->status = VIRTIO_NET_S_LINK_UP;
1604 		netif_carrier_on(dev);
1605 	}
1606 
1607 	pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1608 		 dev->name, max_queue_pairs);
1609 
1610 	return 0;
1611 
1612 free_recv_bufs:
1613 	free_receive_bufs(vi);
1614 	unregister_netdev(dev);
1615 free_vqs:
1616 	cancel_delayed_work_sync(&vi->refill);
1617 	virtnet_del_vqs(vi);
1618 free_index:
1619 	free_percpu(vi->vq_index);
1620 free_stats:
1621 	free_percpu(vi->stats);
1622 free:
1623 	free_netdev(dev);
1624 	return err;
1625 }
1626 
1627 static void remove_vq_common(struct virtnet_info *vi)
1628 {
1629 	vi->vdev->config->reset(vi->vdev);
1630 
1631 	/* Free unused buffers in both send and recv, if any. */
1632 	free_unused_bufs(vi);
1633 
1634 	free_receive_bufs(vi);
1635 
1636 	virtnet_del_vqs(vi);
1637 }
1638 
1639 static void virtnet_remove(struct virtio_device *vdev)
1640 {
1641 	struct virtnet_info *vi = vdev->priv;
1642 
1643 	unregister_hotcpu_notifier(&vi->nb);
1644 
1645 	/* Prevent config work handler from accessing the device. */
1646 	mutex_lock(&vi->config_lock);
1647 	vi->config_enable = false;
1648 	mutex_unlock(&vi->config_lock);
1649 
1650 	unregister_netdev(vi->dev);
1651 
1652 	remove_vq_common(vi);
1653 
1654 	flush_work(&vi->config_work);
1655 
1656 	free_percpu(vi->vq_index);
1657 	free_percpu(vi->stats);
1658 	free_netdev(vi->dev);
1659 }
1660 
1661 #ifdef CONFIG_PM
1662 static int virtnet_freeze(struct virtio_device *vdev)
1663 {
1664 	struct virtnet_info *vi = vdev->priv;
1665 	int i;
1666 
1667 	/* Prevent config work handler from accessing the device */
1668 	mutex_lock(&vi->config_lock);
1669 	vi->config_enable = false;
1670 	mutex_unlock(&vi->config_lock);
1671 
1672 	netif_device_detach(vi->dev);
1673 	cancel_delayed_work_sync(&vi->refill);
1674 
1675 	if (netif_running(vi->dev))
1676 		for (i = 0; i < vi->max_queue_pairs; i++) {
1677 			napi_disable(&vi->rq[i].napi);
1678 			netif_napi_del(&vi->rq[i].napi);
1679 		}
1680 
1681 	remove_vq_common(vi);
1682 
1683 	flush_work(&vi->config_work);
1684 
1685 	return 0;
1686 }
1687 
1688 static int virtnet_restore(struct virtio_device *vdev)
1689 {
1690 	struct virtnet_info *vi = vdev->priv;
1691 	int err, i;
1692 
1693 	err = init_vqs(vi);
1694 	if (err)
1695 		return err;
1696 
1697 	if (netif_running(vi->dev))
1698 		for (i = 0; i < vi->max_queue_pairs; i++)
1699 			virtnet_napi_enable(&vi->rq[i]);
1700 
1701 	netif_device_attach(vi->dev);
1702 
1703 	for (i = 0; i < vi->curr_queue_pairs; i++)
1704 		if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1705 			schedule_delayed_work(&vi->refill, 0);
1706 
1707 	mutex_lock(&vi->config_lock);
1708 	vi->config_enable = true;
1709 	mutex_unlock(&vi->config_lock);
1710 
1711 	virtnet_set_queues(vi, vi->curr_queue_pairs);
1712 
1713 	return 0;
1714 }
1715 #endif
1716 
1717 static struct virtio_device_id id_table[] = {
1718 	{ VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1719 	{ 0 },
1720 };
1721 
1722 static unsigned int features[] = {
1723 	VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1724 	VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1725 	VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1726 	VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1727 	VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1728 	VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1729 	VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1730 	VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1731 	VIRTIO_NET_F_CTRL_MAC_ADDR,
1732 };
1733 
1734 static struct virtio_driver virtio_net_driver = {
1735 	.feature_table = features,
1736 	.feature_table_size = ARRAY_SIZE(features),
1737 	.driver.name =	KBUILD_MODNAME,
1738 	.driver.owner =	THIS_MODULE,
1739 	.id_table =	id_table,
1740 	.probe =	virtnet_probe,
1741 	.remove =	virtnet_remove,
1742 	.config_changed = virtnet_config_changed,
1743 #ifdef CONFIG_PM
1744 	.freeze =	virtnet_freeze,
1745 	.restore =	virtnet_restore,
1746 #endif
1747 };
1748 
1749 module_virtio_driver(virtio_net_driver);
1750 
1751 MODULE_DEVICE_TABLE(virtio, id_table);
1752 MODULE_DESCRIPTION("Virtio network driver");
1753 MODULE_LICENSE("GPL");
1754