xref: /openbmc/linux/drivers/net/virtio_net.c (revision 6d99a79c)
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, see <http://www.gnu.org/licenses/>.
17  */
18 //#define DEBUG
19 #include <linux/netdevice.h>
20 #include <linux/etherdevice.h>
21 #include <linux/ethtool.h>
22 #include <linux/module.h>
23 #include <linux/virtio.h>
24 #include <linux/virtio_net.h>
25 #include <linux/bpf.h>
26 #include <linux/bpf_trace.h>
27 #include <linux/scatterlist.h>
28 #include <linux/if_vlan.h>
29 #include <linux/slab.h>
30 #include <linux/cpu.h>
31 #include <linux/average.h>
32 #include <linux/filter.h>
33 #include <linux/kernel.h>
34 #include <linux/pci.h>
35 #include <net/route.h>
36 #include <net/xdp.h>
37 #include <net/net_failover.h>
38 
39 static int napi_weight = NAPI_POLL_WEIGHT;
40 module_param(napi_weight, int, 0444);
41 
42 static bool csum = true, gso = true, napi_tx;
43 module_param(csum, bool, 0444);
44 module_param(gso, bool, 0444);
45 module_param(napi_tx, bool, 0644);
46 
47 /* FIXME: MTU in config. */
48 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
49 #define GOOD_COPY_LEN	128
50 
51 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
52 
53 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
54 #define VIRTIO_XDP_HEADROOM 256
55 
56 /* Separating two types of XDP xmit */
57 #define VIRTIO_XDP_TX		BIT(0)
58 #define VIRTIO_XDP_REDIR	BIT(1)
59 
60 /* RX packet size EWMA. The average packet size is used to determine the packet
61  * buffer size when refilling RX rings. As the entire RX ring may be refilled
62  * at once, the weight is chosen so that the EWMA will be insensitive to short-
63  * term, transient changes in packet size.
64  */
65 DECLARE_EWMA(pkt_len, 0, 64)
66 
67 #define VIRTNET_DRIVER_VERSION "1.0.0"
68 
69 static const unsigned long guest_offloads[] = {
70 	VIRTIO_NET_F_GUEST_TSO4,
71 	VIRTIO_NET_F_GUEST_TSO6,
72 	VIRTIO_NET_F_GUEST_ECN,
73 	VIRTIO_NET_F_GUEST_UFO
74 };
75 
76 struct virtnet_stat_desc {
77 	char desc[ETH_GSTRING_LEN];
78 	size_t offset;
79 };
80 
81 struct virtnet_sq_stats {
82 	struct u64_stats_sync syncp;
83 	u64 packets;
84 	u64 bytes;
85 	u64 xdp_tx;
86 	u64 xdp_tx_drops;
87 	u64 kicks;
88 };
89 
90 struct virtnet_rq_stats {
91 	struct u64_stats_sync syncp;
92 	u64 packets;
93 	u64 bytes;
94 	u64 drops;
95 	u64 xdp_packets;
96 	u64 xdp_tx;
97 	u64 xdp_redirects;
98 	u64 xdp_drops;
99 	u64 kicks;
100 };
101 
102 #define VIRTNET_SQ_STAT(m)	offsetof(struct virtnet_sq_stats, m)
103 #define VIRTNET_RQ_STAT(m)	offsetof(struct virtnet_rq_stats, m)
104 
105 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
106 	{ "packets",		VIRTNET_SQ_STAT(packets) },
107 	{ "bytes",		VIRTNET_SQ_STAT(bytes) },
108 	{ "xdp_tx",		VIRTNET_SQ_STAT(xdp_tx) },
109 	{ "xdp_tx_drops",	VIRTNET_SQ_STAT(xdp_tx_drops) },
110 	{ "kicks",		VIRTNET_SQ_STAT(kicks) },
111 };
112 
113 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
114 	{ "packets",		VIRTNET_RQ_STAT(packets) },
115 	{ "bytes",		VIRTNET_RQ_STAT(bytes) },
116 	{ "drops",		VIRTNET_RQ_STAT(drops) },
117 	{ "xdp_packets",	VIRTNET_RQ_STAT(xdp_packets) },
118 	{ "xdp_tx",		VIRTNET_RQ_STAT(xdp_tx) },
119 	{ "xdp_redirects",	VIRTNET_RQ_STAT(xdp_redirects) },
120 	{ "xdp_drops",		VIRTNET_RQ_STAT(xdp_drops) },
121 	{ "kicks",		VIRTNET_RQ_STAT(kicks) },
122 };
123 
124 #define VIRTNET_SQ_STATS_LEN	ARRAY_SIZE(virtnet_sq_stats_desc)
125 #define VIRTNET_RQ_STATS_LEN	ARRAY_SIZE(virtnet_rq_stats_desc)
126 
127 /* Internal representation of a send virtqueue */
128 struct send_queue {
129 	/* Virtqueue associated with this send _queue */
130 	struct virtqueue *vq;
131 
132 	/* TX: fragments + linear part + virtio header */
133 	struct scatterlist sg[MAX_SKB_FRAGS + 2];
134 
135 	/* Name of the send queue: output.$index */
136 	char name[40];
137 
138 	struct virtnet_sq_stats stats;
139 
140 	struct napi_struct napi;
141 };
142 
143 /* Internal representation of a receive virtqueue */
144 struct receive_queue {
145 	/* Virtqueue associated with this receive_queue */
146 	struct virtqueue *vq;
147 
148 	struct napi_struct napi;
149 
150 	struct bpf_prog __rcu *xdp_prog;
151 
152 	struct virtnet_rq_stats stats;
153 
154 	/* Chain pages by the private ptr. */
155 	struct page *pages;
156 
157 	/* Average packet length for mergeable receive buffers. */
158 	struct ewma_pkt_len mrg_avg_pkt_len;
159 
160 	/* Page frag for packet buffer allocation. */
161 	struct page_frag alloc_frag;
162 
163 	/* RX: fragments + linear part + virtio header */
164 	struct scatterlist sg[MAX_SKB_FRAGS + 2];
165 
166 	/* Min single buffer size for mergeable buffers case. */
167 	unsigned int min_buf_len;
168 
169 	/* Name of this receive queue: input.$index */
170 	char name[40];
171 
172 	struct xdp_rxq_info xdp_rxq;
173 };
174 
175 /* Control VQ buffers: protected by the rtnl lock */
176 struct control_buf {
177 	struct virtio_net_ctrl_hdr hdr;
178 	virtio_net_ctrl_ack status;
179 	struct virtio_net_ctrl_mq mq;
180 	u8 promisc;
181 	u8 allmulti;
182 	__virtio16 vid;
183 	__virtio64 offloads;
184 };
185 
186 struct virtnet_info {
187 	struct virtio_device *vdev;
188 	struct virtqueue *cvq;
189 	struct net_device *dev;
190 	struct send_queue *sq;
191 	struct receive_queue *rq;
192 	unsigned int status;
193 
194 	/* Max # of queue pairs supported by the device */
195 	u16 max_queue_pairs;
196 
197 	/* # of queue pairs currently used by the driver */
198 	u16 curr_queue_pairs;
199 
200 	/* # of XDP queue pairs currently used by the driver */
201 	u16 xdp_queue_pairs;
202 
203 	/* I like... big packets and I cannot lie! */
204 	bool big_packets;
205 
206 	/* Host will merge rx buffers for big packets (shake it! shake it!) */
207 	bool mergeable_rx_bufs;
208 
209 	/* Has control virtqueue */
210 	bool has_cvq;
211 
212 	/* Host can handle any s/g split between our header and packet data */
213 	bool any_header_sg;
214 
215 	/* Packet virtio header size */
216 	u8 hdr_len;
217 
218 	/* Work struct for refilling if we run low on memory. */
219 	struct delayed_work refill;
220 
221 	/* Work struct for config space updates */
222 	struct work_struct config_work;
223 
224 	/* Does the affinity hint is set for virtqueues? */
225 	bool affinity_hint_set;
226 
227 	/* CPU hotplug instances for online & dead */
228 	struct hlist_node node;
229 	struct hlist_node node_dead;
230 
231 	struct control_buf *ctrl;
232 
233 	/* Ethtool settings */
234 	u8 duplex;
235 	u32 speed;
236 
237 	unsigned long guest_offloads;
238 
239 	/* failover when STANDBY feature enabled */
240 	struct failover *failover;
241 };
242 
243 struct padded_vnet_hdr {
244 	struct virtio_net_hdr_mrg_rxbuf hdr;
245 	/*
246 	 * hdr is in a separate sg buffer, and data sg buffer shares same page
247 	 * with this header sg. This padding makes next sg 16 byte aligned
248 	 * after the header.
249 	 */
250 	char padding[4];
251 };
252 
253 /* Converting between virtqueue no. and kernel tx/rx queue no.
254  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
255  */
256 static int vq2txq(struct virtqueue *vq)
257 {
258 	return (vq->index - 1) / 2;
259 }
260 
261 static int txq2vq(int txq)
262 {
263 	return txq * 2 + 1;
264 }
265 
266 static int vq2rxq(struct virtqueue *vq)
267 {
268 	return vq->index / 2;
269 }
270 
271 static int rxq2vq(int rxq)
272 {
273 	return rxq * 2;
274 }
275 
276 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
277 {
278 	return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
279 }
280 
281 /*
282  * private is used to chain pages for big packets, put the whole
283  * most recent used list in the beginning for reuse
284  */
285 static void give_pages(struct receive_queue *rq, struct page *page)
286 {
287 	struct page *end;
288 
289 	/* Find end of list, sew whole thing into vi->rq.pages. */
290 	for (end = page; end->private; end = (struct page *)end->private);
291 	end->private = (unsigned long)rq->pages;
292 	rq->pages = page;
293 }
294 
295 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
296 {
297 	struct page *p = rq->pages;
298 
299 	if (p) {
300 		rq->pages = (struct page *)p->private;
301 		/* clear private here, it is used to chain pages */
302 		p->private = 0;
303 	} else
304 		p = alloc_page(gfp_mask);
305 	return p;
306 }
307 
308 static void virtqueue_napi_schedule(struct napi_struct *napi,
309 				    struct virtqueue *vq)
310 {
311 	if (napi_schedule_prep(napi)) {
312 		virtqueue_disable_cb(vq);
313 		__napi_schedule(napi);
314 	}
315 }
316 
317 static void virtqueue_napi_complete(struct napi_struct *napi,
318 				    struct virtqueue *vq, int processed)
319 {
320 	int opaque;
321 
322 	opaque = virtqueue_enable_cb_prepare(vq);
323 	if (napi_complete_done(napi, processed)) {
324 		if (unlikely(virtqueue_poll(vq, opaque)))
325 			virtqueue_napi_schedule(napi, vq);
326 	} else {
327 		virtqueue_disable_cb(vq);
328 	}
329 }
330 
331 static void skb_xmit_done(struct virtqueue *vq)
332 {
333 	struct virtnet_info *vi = vq->vdev->priv;
334 	struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
335 
336 	/* Suppress further interrupts. */
337 	virtqueue_disable_cb(vq);
338 
339 	if (napi->weight)
340 		virtqueue_napi_schedule(napi, vq);
341 	else
342 		/* We were probably waiting for more output buffers. */
343 		netif_wake_subqueue(vi->dev, vq2txq(vq));
344 }
345 
346 #define MRG_CTX_HEADER_SHIFT 22
347 static void *mergeable_len_to_ctx(unsigned int truesize,
348 				  unsigned int headroom)
349 {
350 	return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
351 }
352 
353 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
354 {
355 	return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
356 }
357 
358 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
359 {
360 	return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
361 }
362 
363 /* Called from bottom half context */
364 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
365 				   struct receive_queue *rq,
366 				   struct page *page, unsigned int offset,
367 				   unsigned int len, unsigned int truesize)
368 {
369 	struct sk_buff *skb;
370 	struct virtio_net_hdr_mrg_rxbuf *hdr;
371 	unsigned int copy, hdr_len, hdr_padded_len;
372 	char *p;
373 
374 	p = page_address(page) + offset;
375 
376 	/* copy small packet so we can reuse these pages for small data */
377 	skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
378 	if (unlikely(!skb))
379 		return NULL;
380 
381 	hdr = skb_vnet_hdr(skb);
382 
383 	hdr_len = vi->hdr_len;
384 	if (vi->mergeable_rx_bufs)
385 		hdr_padded_len = sizeof(*hdr);
386 	else
387 		hdr_padded_len = sizeof(struct padded_vnet_hdr);
388 
389 	memcpy(hdr, p, hdr_len);
390 
391 	len -= hdr_len;
392 	offset += hdr_padded_len;
393 	p += hdr_padded_len;
394 
395 	copy = len;
396 	if (copy > skb_tailroom(skb))
397 		copy = skb_tailroom(skb);
398 	skb_put_data(skb, p, copy);
399 
400 	len -= copy;
401 	offset += copy;
402 
403 	if (vi->mergeable_rx_bufs) {
404 		if (len)
405 			skb_add_rx_frag(skb, 0, page, offset, len, truesize);
406 		else
407 			put_page(page);
408 		return skb;
409 	}
410 
411 	/*
412 	 * Verify that we can indeed put this data into a skb.
413 	 * This is here to handle cases when the device erroneously
414 	 * tries to receive more than is possible. This is usually
415 	 * the case of a broken device.
416 	 */
417 	if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
418 		net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
419 		dev_kfree_skb(skb);
420 		return NULL;
421 	}
422 	BUG_ON(offset >= PAGE_SIZE);
423 	while (len) {
424 		unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
425 		skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
426 				frag_size, truesize);
427 		len -= frag_size;
428 		page = (struct page *)page->private;
429 		offset = 0;
430 	}
431 
432 	if (page)
433 		give_pages(rq, page);
434 
435 	return skb;
436 }
437 
438 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
439 				   struct send_queue *sq,
440 				   struct xdp_frame *xdpf)
441 {
442 	struct virtio_net_hdr_mrg_rxbuf *hdr;
443 	int err;
444 
445 	/* virtqueue want to use data area in-front of packet */
446 	if (unlikely(xdpf->metasize > 0))
447 		return -EOPNOTSUPP;
448 
449 	if (unlikely(xdpf->headroom < vi->hdr_len))
450 		return -EOVERFLOW;
451 
452 	/* Make room for virtqueue hdr (also change xdpf->headroom?) */
453 	xdpf->data -= vi->hdr_len;
454 	/* Zero header and leave csum up to XDP layers */
455 	hdr = xdpf->data;
456 	memset(hdr, 0, vi->hdr_len);
457 	xdpf->len   += vi->hdr_len;
458 
459 	sg_init_one(sq->sg, xdpf->data, xdpf->len);
460 
461 	err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdpf, GFP_ATOMIC);
462 	if (unlikely(err))
463 		return -ENOSPC; /* Caller handle free/refcnt */
464 
465 	return 0;
466 }
467 
468 static struct send_queue *virtnet_xdp_sq(struct virtnet_info *vi)
469 {
470 	unsigned int qp;
471 
472 	qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
473 	return &vi->sq[qp];
474 }
475 
476 static int virtnet_xdp_xmit(struct net_device *dev,
477 			    int n, struct xdp_frame **frames, u32 flags)
478 {
479 	struct virtnet_info *vi = netdev_priv(dev);
480 	struct receive_queue *rq = vi->rq;
481 	struct xdp_frame *xdpf_sent;
482 	struct bpf_prog *xdp_prog;
483 	struct send_queue *sq;
484 	unsigned int len;
485 	int drops = 0;
486 	int kicks = 0;
487 	int ret, err;
488 	int i;
489 
490 	sq = virtnet_xdp_sq(vi);
491 
492 	if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
493 		ret = -EINVAL;
494 		drops = n;
495 		goto out;
496 	}
497 
498 	/* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
499 	 * indicate XDP resources have been successfully allocated.
500 	 */
501 	xdp_prog = rcu_dereference(rq->xdp_prog);
502 	if (!xdp_prog) {
503 		ret = -ENXIO;
504 		drops = n;
505 		goto out;
506 	}
507 
508 	/* Free up any pending old buffers before queueing new ones. */
509 	while ((xdpf_sent = virtqueue_get_buf(sq->vq, &len)) != NULL)
510 		xdp_return_frame(xdpf_sent);
511 
512 	for (i = 0; i < n; i++) {
513 		struct xdp_frame *xdpf = frames[i];
514 
515 		err = __virtnet_xdp_xmit_one(vi, sq, xdpf);
516 		if (err) {
517 			xdp_return_frame_rx_napi(xdpf);
518 			drops++;
519 		}
520 	}
521 	ret = n - drops;
522 
523 	if (flags & XDP_XMIT_FLUSH) {
524 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
525 			kicks = 1;
526 	}
527 out:
528 	u64_stats_update_begin(&sq->stats.syncp);
529 	sq->stats.xdp_tx += n;
530 	sq->stats.xdp_tx_drops += drops;
531 	sq->stats.kicks += kicks;
532 	u64_stats_update_end(&sq->stats.syncp);
533 
534 	return ret;
535 }
536 
537 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
538 {
539 	return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
540 }
541 
542 /* We copy the packet for XDP in the following cases:
543  *
544  * 1) Packet is scattered across multiple rx buffers.
545  * 2) Headroom space is insufficient.
546  *
547  * This is inefficient but it's a temporary condition that
548  * we hit right after XDP is enabled and until queue is refilled
549  * with large buffers with sufficient headroom - so it should affect
550  * at most queue size packets.
551  * Afterwards, the conditions to enable
552  * XDP should preclude the underlying device from sending packets
553  * across multiple buffers (num_buf > 1), and we make sure buffers
554  * have enough headroom.
555  */
556 static struct page *xdp_linearize_page(struct receive_queue *rq,
557 				       u16 *num_buf,
558 				       struct page *p,
559 				       int offset,
560 				       int page_off,
561 				       unsigned int *len)
562 {
563 	struct page *page = alloc_page(GFP_ATOMIC);
564 
565 	if (!page)
566 		return NULL;
567 
568 	memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
569 	page_off += *len;
570 
571 	while (--*num_buf) {
572 		int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
573 		unsigned int buflen;
574 		void *buf;
575 		int off;
576 
577 		buf = virtqueue_get_buf(rq->vq, &buflen);
578 		if (unlikely(!buf))
579 			goto err_buf;
580 
581 		p = virt_to_head_page(buf);
582 		off = buf - page_address(p);
583 
584 		/* guard against a misconfigured or uncooperative backend that
585 		 * is sending packet larger than the MTU.
586 		 */
587 		if ((page_off + buflen + tailroom) > PAGE_SIZE) {
588 			put_page(p);
589 			goto err_buf;
590 		}
591 
592 		memcpy(page_address(page) + page_off,
593 		       page_address(p) + off, buflen);
594 		page_off += buflen;
595 		put_page(p);
596 	}
597 
598 	/* Headroom does not contribute to packet length */
599 	*len = page_off - VIRTIO_XDP_HEADROOM;
600 	return page;
601 err_buf:
602 	__free_pages(page, 0);
603 	return NULL;
604 }
605 
606 static struct sk_buff *receive_small(struct net_device *dev,
607 				     struct virtnet_info *vi,
608 				     struct receive_queue *rq,
609 				     void *buf, void *ctx,
610 				     unsigned int len,
611 				     unsigned int *xdp_xmit,
612 				     struct virtnet_rq_stats *stats)
613 {
614 	struct sk_buff *skb;
615 	struct bpf_prog *xdp_prog;
616 	unsigned int xdp_headroom = (unsigned long)ctx;
617 	unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
618 	unsigned int headroom = vi->hdr_len + header_offset;
619 	unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
620 			      SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
621 	struct page *page = virt_to_head_page(buf);
622 	unsigned int delta = 0;
623 	struct page *xdp_page;
624 	int err;
625 
626 	len -= vi->hdr_len;
627 	stats->bytes += len;
628 
629 	rcu_read_lock();
630 	xdp_prog = rcu_dereference(rq->xdp_prog);
631 	if (xdp_prog) {
632 		struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
633 		struct xdp_frame *xdpf;
634 		struct xdp_buff xdp;
635 		void *orig_data;
636 		u32 act;
637 
638 		if (unlikely(hdr->hdr.gso_type))
639 			goto err_xdp;
640 
641 		if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
642 			int offset = buf - page_address(page) + header_offset;
643 			unsigned int tlen = len + vi->hdr_len;
644 			u16 num_buf = 1;
645 
646 			xdp_headroom = virtnet_get_headroom(vi);
647 			header_offset = VIRTNET_RX_PAD + xdp_headroom;
648 			headroom = vi->hdr_len + header_offset;
649 			buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
650 				 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
651 			xdp_page = xdp_linearize_page(rq, &num_buf, page,
652 						      offset, header_offset,
653 						      &tlen);
654 			if (!xdp_page)
655 				goto err_xdp;
656 
657 			buf = page_address(xdp_page);
658 			put_page(page);
659 			page = xdp_page;
660 		}
661 
662 		xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
663 		xdp.data = xdp.data_hard_start + xdp_headroom;
664 		xdp_set_data_meta_invalid(&xdp);
665 		xdp.data_end = xdp.data + len;
666 		xdp.rxq = &rq->xdp_rxq;
667 		orig_data = xdp.data;
668 		act = bpf_prog_run_xdp(xdp_prog, &xdp);
669 		stats->xdp_packets++;
670 
671 		switch (act) {
672 		case XDP_PASS:
673 			/* Recalculate length in case bpf program changed it */
674 			delta = orig_data - xdp.data;
675 			len = xdp.data_end - xdp.data;
676 			break;
677 		case XDP_TX:
678 			stats->xdp_tx++;
679 			xdpf = convert_to_xdp_frame(&xdp);
680 			if (unlikely(!xdpf))
681 				goto err_xdp;
682 			err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
683 			if (unlikely(err < 0)) {
684 				trace_xdp_exception(vi->dev, xdp_prog, act);
685 				goto err_xdp;
686 			}
687 			*xdp_xmit |= VIRTIO_XDP_TX;
688 			rcu_read_unlock();
689 			goto xdp_xmit;
690 		case XDP_REDIRECT:
691 			stats->xdp_redirects++;
692 			err = xdp_do_redirect(dev, &xdp, xdp_prog);
693 			if (err)
694 				goto err_xdp;
695 			*xdp_xmit |= VIRTIO_XDP_REDIR;
696 			rcu_read_unlock();
697 			goto xdp_xmit;
698 		default:
699 			bpf_warn_invalid_xdp_action(act);
700 			/* fall through */
701 		case XDP_ABORTED:
702 			trace_xdp_exception(vi->dev, xdp_prog, act);
703 		case XDP_DROP:
704 			goto err_xdp;
705 		}
706 	}
707 	rcu_read_unlock();
708 
709 	skb = build_skb(buf, buflen);
710 	if (!skb) {
711 		put_page(page);
712 		goto err;
713 	}
714 	skb_reserve(skb, headroom - delta);
715 	skb_put(skb, len);
716 	if (!delta) {
717 		buf += header_offset;
718 		memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
719 	} /* keep zeroed vnet hdr since packet was changed by bpf */
720 
721 err:
722 	return skb;
723 
724 err_xdp:
725 	rcu_read_unlock();
726 	stats->xdp_drops++;
727 	stats->drops++;
728 	put_page(page);
729 xdp_xmit:
730 	return NULL;
731 }
732 
733 static struct sk_buff *receive_big(struct net_device *dev,
734 				   struct virtnet_info *vi,
735 				   struct receive_queue *rq,
736 				   void *buf,
737 				   unsigned int len,
738 				   struct virtnet_rq_stats *stats)
739 {
740 	struct page *page = buf;
741 	struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
742 
743 	stats->bytes += len - vi->hdr_len;
744 	if (unlikely(!skb))
745 		goto err;
746 
747 	return skb;
748 
749 err:
750 	stats->drops++;
751 	give_pages(rq, page);
752 	return NULL;
753 }
754 
755 static struct sk_buff *receive_mergeable(struct net_device *dev,
756 					 struct virtnet_info *vi,
757 					 struct receive_queue *rq,
758 					 void *buf,
759 					 void *ctx,
760 					 unsigned int len,
761 					 unsigned int *xdp_xmit,
762 					 struct virtnet_rq_stats *stats)
763 {
764 	struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
765 	u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
766 	struct page *page = virt_to_head_page(buf);
767 	int offset = buf - page_address(page);
768 	struct sk_buff *head_skb, *curr_skb;
769 	struct bpf_prog *xdp_prog;
770 	unsigned int truesize;
771 	unsigned int headroom = mergeable_ctx_to_headroom(ctx);
772 	int err;
773 
774 	head_skb = NULL;
775 	stats->bytes += len - vi->hdr_len;
776 
777 	rcu_read_lock();
778 	xdp_prog = rcu_dereference(rq->xdp_prog);
779 	if (xdp_prog) {
780 		struct xdp_frame *xdpf;
781 		struct page *xdp_page;
782 		struct xdp_buff xdp;
783 		void *data;
784 		u32 act;
785 
786 		/* Transient failure which in theory could occur if
787 		 * in-flight packets from before XDP was enabled reach
788 		 * the receive path after XDP is loaded.
789 		 */
790 		if (unlikely(hdr->hdr.gso_type))
791 			goto err_xdp;
792 
793 		/* This happens when rx buffer size is underestimated
794 		 * or headroom is not enough because of the buffer
795 		 * was refilled before XDP is set. This should only
796 		 * happen for the first several packets, so we don't
797 		 * care much about its performance.
798 		 */
799 		if (unlikely(num_buf > 1 ||
800 			     headroom < virtnet_get_headroom(vi))) {
801 			/* linearize data for XDP */
802 			xdp_page = xdp_linearize_page(rq, &num_buf,
803 						      page, offset,
804 						      VIRTIO_XDP_HEADROOM,
805 						      &len);
806 			if (!xdp_page)
807 				goto err_xdp;
808 			offset = VIRTIO_XDP_HEADROOM;
809 		} else {
810 			xdp_page = page;
811 		}
812 
813 		/* Allow consuming headroom but reserve enough space to push
814 		 * the descriptor on if we get an XDP_TX return code.
815 		 */
816 		data = page_address(xdp_page) + offset;
817 		xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
818 		xdp.data = data + vi->hdr_len;
819 		xdp_set_data_meta_invalid(&xdp);
820 		xdp.data_end = xdp.data + (len - vi->hdr_len);
821 		xdp.rxq = &rq->xdp_rxq;
822 
823 		act = bpf_prog_run_xdp(xdp_prog, &xdp);
824 		stats->xdp_packets++;
825 
826 		switch (act) {
827 		case XDP_PASS:
828 			/* recalculate offset to account for any header
829 			 * adjustments. Note other cases do not build an
830 			 * skb and avoid using offset
831 			 */
832 			offset = xdp.data -
833 					page_address(xdp_page) - vi->hdr_len;
834 
835 			/* recalculate len if xdp.data or xdp.data_end were
836 			 * adjusted
837 			 */
838 			len = xdp.data_end - xdp.data + vi->hdr_len;
839 			/* We can only create skb based on xdp_page. */
840 			if (unlikely(xdp_page != page)) {
841 				rcu_read_unlock();
842 				put_page(page);
843 				head_skb = page_to_skb(vi, rq, xdp_page,
844 						       offset, len, PAGE_SIZE);
845 				return head_skb;
846 			}
847 			break;
848 		case XDP_TX:
849 			stats->xdp_tx++;
850 			xdpf = convert_to_xdp_frame(&xdp);
851 			if (unlikely(!xdpf))
852 				goto err_xdp;
853 			err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
854 			if (unlikely(err < 0)) {
855 				trace_xdp_exception(vi->dev, xdp_prog, act);
856 				if (unlikely(xdp_page != page))
857 					put_page(xdp_page);
858 				goto err_xdp;
859 			}
860 			*xdp_xmit |= VIRTIO_XDP_TX;
861 			if (unlikely(xdp_page != page))
862 				put_page(page);
863 			rcu_read_unlock();
864 			goto xdp_xmit;
865 		case XDP_REDIRECT:
866 			stats->xdp_redirects++;
867 			err = xdp_do_redirect(dev, &xdp, xdp_prog);
868 			if (err) {
869 				if (unlikely(xdp_page != page))
870 					put_page(xdp_page);
871 				goto err_xdp;
872 			}
873 			*xdp_xmit |= VIRTIO_XDP_REDIR;
874 			if (unlikely(xdp_page != page))
875 				put_page(page);
876 			rcu_read_unlock();
877 			goto xdp_xmit;
878 		default:
879 			bpf_warn_invalid_xdp_action(act);
880 			/* fall through */
881 		case XDP_ABORTED:
882 			trace_xdp_exception(vi->dev, xdp_prog, act);
883 			/* fall through */
884 		case XDP_DROP:
885 			if (unlikely(xdp_page != page))
886 				__free_pages(xdp_page, 0);
887 			goto err_xdp;
888 		}
889 	}
890 	rcu_read_unlock();
891 
892 	truesize = mergeable_ctx_to_truesize(ctx);
893 	if (unlikely(len > truesize)) {
894 		pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
895 			 dev->name, len, (unsigned long)ctx);
896 		dev->stats.rx_length_errors++;
897 		goto err_skb;
898 	}
899 
900 	head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
901 	curr_skb = head_skb;
902 
903 	if (unlikely(!curr_skb))
904 		goto err_skb;
905 	while (--num_buf) {
906 		int num_skb_frags;
907 
908 		buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
909 		if (unlikely(!buf)) {
910 			pr_debug("%s: rx error: %d buffers out of %d missing\n",
911 				 dev->name, num_buf,
912 				 virtio16_to_cpu(vi->vdev,
913 						 hdr->num_buffers));
914 			dev->stats.rx_length_errors++;
915 			goto err_buf;
916 		}
917 
918 		stats->bytes += len;
919 		page = virt_to_head_page(buf);
920 
921 		truesize = mergeable_ctx_to_truesize(ctx);
922 		if (unlikely(len > truesize)) {
923 			pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
924 				 dev->name, len, (unsigned long)ctx);
925 			dev->stats.rx_length_errors++;
926 			goto err_skb;
927 		}
928 
929 		num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
930 		if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
931 			struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
932 
933 			if (unlikely(!nskb))
934 				goto err_skb;
935 			if (curr_skb == head_skb)
936 				skb_shinfo(curr_skb)->frag_list = nskb;
937 			else
938 				curr_skb->next = nskb;
939 			curr_skb = nskb;
940 			head_skb->truesize += nskb->truesize;
941 			num_skb_frags = 0;
942 		}
943 		if (curr_skb != head_skb) {
944 			head_skb->data_len += len;
945 			head_skb->len += len;
946 			head_skb->truesize += truesize;
947 		}
948 		offset = buf - page_address(page);
949 		if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
950 			put_page(page);
951 			skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
952 					     len, truesize);
953 		} else {
954 			skb_add_rx_frag(curr_skb, num_skb_frags, page,
955 					offset, len, truesize);
956 		}
957 	}
958 
959 	ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
960 	return head_skb;
961 
962 err_xdp:
963 	rcu_read_unlock();
964 	stats->xdp_drops++;
965 err_skb:
966 	put_page(page);
967 	while (num_buf-- > 1) {
968 		buf = virtqueue_get_buf(rq->vq, &len);
969 		if (unlikely(!buf)) {
970 			pr_debug("%s: rx error: %d buffers missing\n",
971 				 dev->name, num_buf);
972 			dev->stats.rx_length_errors++;
973 			break;
974 		}
975 		stats->bytes += len;
976 		page = virt_to_head_page(buf);
977 		put_page(page);
978 	}
979 err_buf:
980 	stats->drops++;
981 	dev_kfree_skb(head_skb);
982 xdp_xmit:
983 	return NULL;
984 }
985 
986 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
987 			void *buf, unsigned int len, void **ctx,
988 			unsigned int *xdp_xmit,
989 			struct virtnet_rq_stats *stats)
990 {
991 	struct net_device *dev = vi->dev;
992 	struct sk_buff *skb;
993 	struct virtio_net_hdr_mrg_rxbuf *hdr;
994 
995 	if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
996 		pr_debug("%s: short packet %i\n", dev->name, len);
997 		dev->stats.rx_length_errors++;
998 		if (vi->mergeable_rx_bufs) {
999 			put_page(virt_to_head_page(buf));
1000 		} else if (vi->big_packets) {
1001 			give_pages(rq, buf);
1002 		} else {
1003 			put_page(virt_to_head_page(buf));
1004 		}
1005 		return;
1006 	}
1007 
1008 	if (vi->mergeable_rx_bufs)
1009 		skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1010 					stats);
1011 	else if (vi->big_packets)
1012 		skb = receive_big(dev, vi, rq, buf, len, stats);
1013 	else
1014 		skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1015 
1016 	if (unlikely(!skb))
1017 		return;
1018 
1019 	hdr = skb_vnet_hdr(skb);
1020 
1021 	if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1022 		skb->ip_summed = CHECKSUM_UNNECESSARY;
1023 
1024 	if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1025 				  virtio_is_little_endian(vi->vdev))) {
1026 		net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1027 				     dev->name, hdr->hdr.gso_type,
1028 				     hdr->hdr.gso_size);
1029 		goto frame_err;
1030 	}
1031 
1032 	skb->protocol = eth_type_trans(skb, dev);
1033 	pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1034 		 ntohs(skb->protocol), skb->len, skb->pkt_type);
1035 
1036 	napi_gro_receive(&rq->napi, skb);
1037 	return;
1038 
1039 frame_err:
1040 	dev->stats.rx_frame_errors++;
1041 	dev_kfree_skb(skb);
1042 }
1043 
1044 /* Unlike mergeable buffers, all buffers are allocated to the
1045  * same size, except for the headroom. For this reason we do
1046  * not need to use  mergeable_len_to_ctx here - it is enough
1047  * to store the headroom as the context ignoring the truesize.
1048  */
1049 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1050 			     gfp_t gfp)
1051 {
1052 	struct page_frag *alloc_frag = &rq->alloc_frag;
1053 	char *buf;
1054 	unsigned int xdp_headroom = virtnet_get_headroom(vi);
1055 	void *ctx = (void *)(unsigned long)xdp_headroom;
1056 	int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1057 	int err;
1058 
1059 	len = SKB_DATA_ALIGN(len) +
1060 	      SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1061 	if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1062 		return -ENOMEM;
1063 
1064 	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1065 	get_page(alloc_frag->page);
1066 	alloc_frag->offset += len;
1067 	sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1068 		    vi->hdr_len + GOOD_PACKET_LEN);
1069 	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1070 	if (err < 0)
1071 		put_page(virt_to_head_page(buf));
1072 	return err;
1073 }
1074 
1075 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1076 			   gfp_t gfp)
1077 {
1078 	struct page *first, *list = NULL;
1079 	char *p;
1080 	int i, err, offset;
1081 
1082 	sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1083 
1084 	/* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1085 	for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1086 		first = get_a_page(rq, gfp);
1087 		if (!first) {
1088 			if (list)
1089 				give_pages(rq, list);
1090 			return -ENOMEM;
1091 		}
1092 		sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1093 
1094 		/* chain new page in list head to match sg */
1095 		first->private = (unsigned long)list;
1096 		list = first;
1097 	}
1098 
1099 	first = get_a_page(rq, gfp);
1100 	if (!first) {
1101 		give_pages(rq, list);
1102 		return -ENOMEM;
1103 	}
1104 	p = page_address(first);
1105 
1106 	/* rq->sg[0], rq->sg[1] share the same page */
1107 	/* a separated rq->sg[0] for header - required in case !any_header_sg */
1108 	sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1109 
1110 	/* rq->sg[1] for data packet, from offset */
1111 	offset = sizeof(struct padded_vnet_hdr);
1112 	sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1113 
1114 	/* chain first in list head */
1115 	first->private = (unsigned long)list;
1116 	err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1117 				  first, gfp);
1118 	if (err < 0)
1119 		give_pages(rq, first);
1120 
1121 	return err;
1122 }
1123 
1124 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1125 					  struct ewma_pkt_len *avg_pkt_len,
1126 					  unsigned int room)
1127 {
1128 	const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1129 	unsigned int len;
1130 
1131 	if (room)
1132 		return PAGE_SIZE - room;
1133 
1134 	len = hdr_len +	clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1135 				rq->min_buf_len, PAGE_SIZE - hdr_len);
1136 
1137 	return ALIGN(len, L1_CACHE_BYTES);
1138 }
1139 
1140 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1141 				 struct receive_queue *rq, gfp_t gfp)
1142 {
1143 	struct page_frag *alloc_frag = &rq->alloc_frag;
1144 	unsigned int headroom = virtnet_get_headroom(vi);
1145 	unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1146 	unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1147 	char *buf;
1148 	void *ctx;
1149 	int err;
1150 	unsigned int len, hole;
1151 
1152 	/* Extra tailroom is needed to satisfy XDP's assumption. This
1153 	 * means rx frags coalescing won't work, but consider we've
1154 	 * disabled GSO for XDP, it won't be a big issue.
1155 	 */
1156 	len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1157 	if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1158 		return -ENOMEM;
1159 
1160 	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1161 	buf += headroom; /* advance address leaving hole at front of pkt */
1162 	get_page(alloc_frag->page);
1163 	alloc_frag->offset += len + room;
1164 	hole = alloc_frag->size - alloc_frag->offset;
1165 	if (hole < len + room) {
1166 		/* To avoid internal fragmentation, if there is very likely not
1167 		 * enough space for another buffer, add the remaining space to
1168 		 * the current buffer.
1169 		 */
1170 		len += hole;
1171 		alloc_frag->offset += hole;
1172 	}
1173 
1174 	sg_init_one(rq->sg, buf, len);
1175 	ctx = mergeable_len_to_ctx(len, headroom);
1176 	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1177 	if (err < 0)
1178 		put_page(virt_to_head_page(buf));
1179 
1180 	return err;
1181 }
1182 
1183 /*
1184  * Returns false if we couldn't fill entirely (OOM).
1185  *
1186  * Normally run in the receive path, but can also be run from ndo_open
1187  * before we're receiving packets, or from refill_work which is
1188  * careful to disable receiving (using napi_disable).
1189  */
1190 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1191 			  gfp_t gfp)
1192 {
1193 	int err;
1194 	bool oom;
1195 
1196 	do {
1197 		if (vi->mergeable_rx_bufs)
1198 			err = add_recvbuf_mergeable(vi, rq, gfp);
1199 		else if (vi->big_packets)
1200 			err = add_recvbuf_big(vi, rq, gfp);
1201 		else
1202 			err = add_recvbuf_small(vi, rq, gfp);
1203 
1204 		oom = err == -ENOMEM;
1205 		if (err)
1206 			break;
1207 	} while (rq->vq->num_free);
1208 	if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1209 		u64_stats_update_begin(&rq->stats.syncp);
1210 		rq->stats.kicks++;
1211 		u64_stats_update_end(&rq->stats.syncp);
1212 	}
1213 
1214 	return !oom;
1215 }
1216 
1217 static void skb_recv_done(struct virtqueue *rvq)
1218 {
1219 	struct virtnet_info *vi = rvq->vdev->priv;
1220 	struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1221 
1222 	virtqueue_napi_schedule(&rq->napi, rvq);
1223 }
1224 
1225 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1226 {
1227 	napi_enable(napi);
1228 
1229 	/* If all buffers were filled by other side before we napi_enabled, we
1230 	 * won't get another interrupt, so process any outstanding packets now.
1231 	 * Call local_bh_enable after to trigger softIRQ processing.
1232 	 */
1233 	local_bh_disable();
1234 	virtqueue_napi_schedule(napi, vq);
1235 	local_bh_enable();
1236 }
1237 
1238 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1239 				   struct virtqueue *vq,
1240 				   struct napi_struct *napi)
1241 {
1242 	if (!napi->weight)
1243 		return;
1244 
1245 	/* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1246 	 * enable the feature if this is likely affine with the transmit path.
1247 	 */
1248 	if (!vi->affinity_hint_set) {
1249 		napi->weight = 0;
1250 		return;
1251 	}
1252 
1253 	return virtnet_napi_enable(vq, napi);
1254 }
1255 
1256 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1257 {
1258 	if (napi->weight)
1259 		napi_disable(napi);
1260 }
1261 
1262 static void refill_work(struct work_struct *work)
1263 {
1264 	struct virtnet_info *vi =
1265 		container_of(work, struct virtnet_info, refill.work);
1266 	bool still_empty;
1267 	int i;
1268 
1269 	for (i = 0; i < vi->curr_queue_pairs; i++) {
1270 		struct receive_queue *rq = &vi->rq[i];
1271 
1272 		napi_disable(&rq->napi);
1273 		still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1274 		virtnet_napi_enable(rq->vq, &rq->napi);
1275 
1276 		/* In theory, this can happen: if we don't get any buffers in
1277 		 * we will *never* try to fill again.
1278 		 */
1279 		if (still_empty)
1280 			schedule_delayed_work(&vi->refill, HZ/2);
1281 	}
1282 }
1283 
1284 static int virtnet_receive(struct receive_queue *rq, int budget,
1285 			   unsigned int *xdp_xmit)
1286 {
1287 	struct virtnet_info *vi = rq->vq->vdev->priv;
1288 	struct virtnet_rq_stats stats = {};
1289 	unsigned int len;
1290 	void *buf;
1291 	int i;
1292 
1293 	if (!vi->big_packets || vi->mergeable_rx_bufs) {
1294 		void *ctx;
1295 
1296 		while (stats.packets < budget &&
1297 		       (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1298 			receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1299 			stats.packets++;
1300 		}
1301 	} else {
1302 		while (stats.packets < budget &&
1303 		       (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1304 			receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1305 			stats.packets++;
1306 		}
1307 	}
1308 
1309 	if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1310 		if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1311 			schedule_delayed_work(&vi->refill, 0);
1312 	}
1313 
1314 	u64_stats_update_begin(&rq->stats.syncp);
1315 	for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1316 		size_t offset = virtnet_rq_stats_desc[i].offset;
1317 		u64 *item;
1318 
1319 		item = (u64 *)((u8 *)&rq->stats + offset);
1320 		*item += *(u64 *)((u8 *)&stats + offset);
1321 	}
1322 	u64_stats_update_end(&rq->stats.syncp);
1323 
1324 	return stats.packets;
1325 }
1326 
1327 static void free_old_xmit_skbs(struct send_queue *sq)
1328 {
1329 	struct sk_buff *skb;
1330 	unsigned int len;
1331 	unsigned int packets = 0;
1332 	unsigned int bytes = 0;
1333 
1334 	while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1335 		pr_debug("Sent skb %p\n", skb);
1336 
1337 		bytes += skb->len;
1338 		packets++;
1339 
1340 		dev_consume_skb_any(skb);
1341 	}
1342 
1343 	/* Avoid overhead when no packets have been processed
1344 	 * happens when called speculatively from start_xmit.
1345 	 */
1346 	if (!packets)
1347 		return;
1348 
1349 	u64_stats_update_begin(&sq->stats.syncp);
1350 	sq->stats.bytes += bytes;
1351 	sq->stats.packets += packets;
1352 	u64_stats_update_end(&sq->stats.syncp);
1353 }
1354 
1355 static void virtnet_poll_cleantx(struct receive_queue *rq)
1356 {
1357 	struct virtnet_info *vi = rq->vq->vdev->priv;
1358 	unsigned int index = vq2rxq(rq->vq);
1359 	struct send_queue *sq = &vi->sq[index];
1360 	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1361 
1362 	if (!sq->napi.weight)
1363 		return;
1364 
1365 	if (__netif_tx_trylock(txq)) {
1366 		free_old_xmit_skbs(sq);
1367 		__netif_tx_unlock(txq);
1368 	}
1369 
1370 	if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1371 		netif_tx_wake_queue(txq);
1372 }
1373 
1374 static int virtnet_poll(struct napi_struct *napi, int budget)
1375 {
1376 	struct receive_queue *rq =
1377 		container_of(napi, struct receive_queue, napi);
1378 	struct virtnet_info *vi = rq->vq->vdev->priv;
1379 	struct send_queue *sq;
1380 	unsigned int received;
1381 	unsigned int xdp_xmit = 0;
1382 
1383 	virtnet_poll_cleantx(rq);
1384 
1385 	received = virtnet_receive(rq, budget, &xdp_xmit);
1386 
1387 	/* Out of packets? */
1388 	if (received < budget)
1389 		virtqueue_napi_complete(napi, rq->vq, received);
1390 
1391 	if (xdp_xmit & VIRTIO_XDP_REDIR)
1392 		xdp_do_flush_map();
1393 
1394 	if (xdp_xmit & VIRTIO_XDP_TX) {
1395 		sq = virtnet_xdp_sq(vi);
1396 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1397 			u64_stats_update_begin(&sq->stats.syncp);
1398 			sq->stats.kicks++;
1399 			u64_stats_update_end(&sq->stats.syncp);
1400 		}
1401 	}
1402 
1403 	return received;
1404 }
1405 
1406 static int virtnet_open(struct net_device *dev)
1407 {
1408 	struct virtnet_info *vi = netdev_priv(dev);
1409 	int i, err;
1410 
1411 	for (i = 0; i < vi->max_queue_pairs; i++) {
1412 		if (i < vi->curr_queue_pairs)
1413 			/* Make sure we have some buffers: if oom use wq. */
1414 			if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1415 				schedule_delayed_work(&vi->refill, 0);
1416 
1417 		err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i);
1418 		if (err < 0)
1419 			return err;
1420 
1421 		err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1422 						 MEM_TYPE_PAGE_SHARED, NULL);
1423 		if (err < 0) {
1424 			xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1425 			return err;
1426 		}
1427 
1428 		virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1429 		virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1430 	}
1431 
1432 	return 0;
1433 }
1434 
1435 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1436 {
1437 	struct send_queue *sq = container_of(napi, struct send_queue, napi);
1438 	struct virtnet_info *vi = sq->vq->vdev->priv;
1439 	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
1440 
1441 	__netif_tx_lock(txq, raw_smp_processor_id());
1442 	free_old_xmit_skbs(sq);
1443 	__netif_tx_unlock(txq);
1444 
1445 	virtqueue_napi_complete(napi, sq->vq, 0);
1446 
1447 	if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1448 		netif_tx_wake_queue(txq);
1449 
1450 	return 0;
1451 }
1452 
1453 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1454 {
1455 	struct virtio_net_hdr_mrg_rxbuf *hdr;
1456 	const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1457 	struct virtnet_info *vi = sq->vq->vdev->priv;
1458 	int num_sg;
1459 	unsigned hdr_len = vi->hdr_len;
1460 	bool can_push;
1461 
1462 	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1463 
1464 	can_push = vi->any_header_sg &&
1465 		!((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1466 		!skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1467 	/* Even if we can, don't push here yet as this would skew
1468 	 * csum_start offset below. */
1469 	if (can_push)
1470 		hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1471 	else
1472 		hdr = skb_vnet_hdr(skb);
1473 
1474 	if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1475 				    virtio_is_little_endian(vi->vdev), false,
1476 				    0))
1477 		BUG();
1478 
1479 	if (vi->mergeable_rx_bufs)
1480 		hdr->num_buffers = 0;
1481 
1482 	sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1483 	if (can_push) {
1484 		__skb_push(skb, hdr_len);
1485 		num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1486 		if (unlikely(num_sg < 0))
1487 			return num_sg;
1488 		/* Pull header back to avoid skew in tx bytes calculations. */
1489 		__skb_pull(skb, hdr_len);
1490 	} else {
1491 		sg_set_buf(sq->sg, hdr, hdr_len);
1492 		num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1493 		if (unlikely(num_sg < 0))
1494 			return num_sg;
1495 		num_sg++;
1496 	}
1497 	return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1498 }
1499 
1500 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1501 {
1502 	struct virtnet_info *vi = netdev_priv(dev);
1503 	int qnum = skb_get_queue_mapping(skb);
1504 	struct send_queue *sq = &vi->sq[qnum];
1505 	int err;
1506 	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1507 	bool kick = !skb->xmit_more;
1508 	bool use_napi = sq->napi.weight;
1509 
1510 	/* Free up any pending old buffers before queueing new ones. */
1511 	free_old_xmit_skbs(sq);
1512 
1513 	if (use_napi && kick)
1514 		virtqueue_enable_cb_delayed(sq->vq);
1515 
1516 	/* timestamp packet in software */
1517 	skb_tx_timestamp(skb);
1518 
1519 	/* Try to transmit */
1520 	err = xmit_skb(sq, skb);
1521 
1522 	/* This should not happen! */
1523 	if (unlikely(err)) {
1524 		dev->stats.tx_fifo_errors++;
1525 		if (net_ratelimit())
1526 			dev_warn(&dev->dev,
1527 				 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1528 		dev->stats.tx_dropped++;
1529 		dev_kfree_skb_any(skb);
1530 		return NETDEV_TX_OK;
1531 	}
1532 
1533 	/* Don't wait up for transmitted skbs to be freed. */
1534 	if (!use_napi) {
1535 		skb_orphan(skb);
1536 		nf_reset(skb);
1537 	}
1538 
1539 	/* If running out of space, stop queue to avoid getting packets that we
1540 	 * are then unable to transmit.
1541 	 * An alternative would be to force queuing layer to requeue the skb by
1542 	 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1543 	 * returned in a normal path of operation: it means that driver is not
1544 	 * maintaining the TX queue stop/start state properly, and causes
1545 	 * the stack to do a non-trivial amount of useless work.
1546 	 * Since most packets only take 1 or 2 ring slots, stopping the queue
1547 	 * early means 16 slots are typically wasted.
1548 	 */
1549 	if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1550 		netif_stop_subqueue(dev, qnum);
1551 		if (!use_napi &&
1552 		    unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1553 			/* More just got used, free them then recheck. */
1554 			free_old_xmit_skbs(sq);
1555 			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1556 				netif_start_subqueue(dev, qnum);
1557 				virtqueue_disable_cb(sq->vq);
1558 			}
1559 		}
1560 	}
1561 
1562 	if (kick || netif_xmit_stopped(txq)) {
1563 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1564 			u64_stats_update_begin(&sq->stats.syncp);
1565 			sq->stats.kicks++;
1566 			u64_stats_update_end(&sq->stats.syncp);
1567 		}
1568 	}
1569 
1570 	return NETDEV_TX_OK;
1571 }
1572 
1573 /*
1574  * Send command via the control virtqueue and check status.  Commands
1575  * supported by the hypervisor, as indicated by feature bits, should
1576  * never fail unless improperly formatted.
1577  */
1578 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1579 				 struct scatterlist *out)
1580 {
1581 	struct scatterlist *sgs[4], hdr, stat;
1582 	unsigned out_num = 0, tmp;
1583 
1584 	/* Caller should know better */
1585 	BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1586 
1587 	vi->ctrl->status = ~0;
1588 	vi->ctrl->hdr.class = class;
1589 	vi->ctrl->hdr.cmd = cmd;
1590 	/* Add header */
1591 	sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1592 	sgs[out_num++] = &hdr;
1593 
1594 	if (out)
1595 		sgs[out_num++] = out;
1596 
1597 	/* Add return status. */
1598 	sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1599 	sgs[out_num] = &stat;
1600 
1601 	BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1602 	virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1603 
1604 	if (unlikely(!virtqueue_kick(vi->cvq)))
1605 		return vi->ctrl->status == VIRTIO_NET_OK;
1606 
1607 	/* Spin for a response, the kick causes an ioport write, trapping
1608 	 * into the hypervisor, so the request should be handled immediately.
1609 	 */
1610 	while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1611 	       !virtqueue_is_broken(vi->cvq))
1612 		cpu_relax();
1613 
1614 	return vi->ctrl->status == VIRTIO_NET_OK;
1615 }
1616 
1617 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1618 {
1619 	struct virtnet_info *vi = netdev_priv(dev);
1620 	struct virtio_device *vdev = vi->vdev;
1621 	int ret;
1622 	struct sockaddr *addr;
1623 	struct scatterlist sg;
1624 
1625 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1626 		return -EOPNOTSUPP;
1627 
1628 	addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1629 	if (!addr)
1630 		return -ENOMEM;
1631 
1632 	ret = eth_prepare_mac_addr_change(dev, addr);
1633 	if (ret)
1634 		goto out;
1635 
1636 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1637 		sg_init_one(&sg, addr->sa_data, dev->addr_len);
1638 		if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1639 					  VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1640 			dev_warn(&vdev->dev,
1641 				 "Failed to set mac address by vq command.\n");
1642 			ret = -EINVAL;
1643 			goto out;
1644 		}
1645 	} else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1646 		   !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1647 		unsigned int i;
1648 
1649 		/* Naturally, this has an atomicity problem. */
1650 		for (i = 0; i < dev->addr_len; i++)
1651 			virtio_cwrite8(vdev,
1652 				       offsetof(struct virtio_net_config, mac) +
1653 				       i, addr->sa_data[i]);
1654 	}
1655 
1656 	eth_commit_mac_addr_change(dev, p);
1657 	ret = 0;
1658 
1659 out:
1660 	kfree(addr);
1661 	return ret;
1662 }
1663 
1664 static void virtnet_stats(struct net_device *dev,
1665 			  struct rtnl_link_stats64 *tot)
1666 {
1667 	struct virtnet_info *vi = netdev_priv(dev);
1668 	unsigned int start;
1669 	int i;
1670 
1671 	for (i = 0; i < vi->max_queue_pairs; i++) {
1672 		u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1673 		struct receive_queue *rq = &vi->rq[i];
1674 		struct send_queue *sq = &vi->sq[i];
1675 
1676 		do {
1677 			start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1678 			tpackets = sq->stats.packets;
1679 			tbytes   = sq->stats.bytes;
1680 		} while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1681 
1682 		do {
1683 			start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1684 			rpackets = rq->stats.packets;
1685 			rbytes   = rq->stats.bytes;
1686 			rdrops   = rq->stats.drops;
1687 		} while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1688 
1689 		tot->rx_packets += rpackets;
1690 		tot->tx_packets += tpackets;
1691 		tot->rx_bytes   += rbytes;
1692 		tot->tx_bytes   += tbytes;
1693 		tot->rx_dropped += rdrops;
1694 	}
1695 
1696 	tot->tx_dropped = dev->stats.tx_dropped;
1697 	tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1698 	tot->rx_length_errors = dev->stats.rx_length_errors;
1699 	tot->rx_frame_errors = dev->stats.rx_frame_errors;
1700 }
1701 
1702 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1703 {
1704 	rtnl_lock();
1705 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1706 				  VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1707 		dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1708 	rtnl_unlock();
1709 }
1710 
1711 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1712 {
1713 	struct scatterlist sg;
1714 	struct net_device *dev = vi->dev;
1715 
1716 	if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1717 		return 0;
1718 
1719 	vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1720 	sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1721 
1722 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1723 				  VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1724 		dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1725 			 queue_pairs);
1726 		return -EINVAL;
1727 	} else {
1728 		vi->curr_queue_pairs = queue_pairs;
1729 		/* virtnet_open() will refill when device is going to up. */
1730 		if (dev->flags & IFF_UP)
1731 			schedule_delayed_work(&vi->refill, 0);
1732 	}
1733 
1734 	return 0;
1735 }
1736 
1737 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1738 {
1739 	int err;
1740 
1741 	rtnl_lock();
1742 	err = _virtnet_set_queues(vi, queue_pairs);
1743 	rtnl_unlock();
1744 	return err;
1745 }
1746 
1747 static int virtnet_close(struct net_device *dev)
1748 {
1749 	struct virtnet_info *vi = netdev_priv(dev);
1750 	int i;
1751 
1752 	/* Make sure refill_work doesn't re-enable napi! */
1753 	cancel_delayed_work_sync(&vi->refill);
1754 
1755 	for (i = 0; i < vi->max_queue_pairs; i++) {
1756 		xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1757 		napi_disable(&vi->rq[i].napi);
1758 		virtnet_napi_tx_disable(&vi->sq[i].napi);
1759 	}
1760 
1761 	return 0;
1762 }
1763 
1764 static void virtnet_set_rx_mode(struct net_device *dev)
1765 {
1766 	struct virtnet_info *vi = netdev_priv(dev);
1767 	struct scatterlist sg[2];
1768 	struct virtio_net_ctrl_mac *mac_data;
1769 	struct netdev_hw_addr *ha;
1770 	int uc_count;
1771 	int mc_count;
1772 	void *buf;
1773 	int i;
1774 
1775 	/* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1776 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1777 		return;
1778 
1779 	vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1780 	vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1781 
1782 	sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1783 
1784 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1785 				  VIRTIO_NET_CTRL_RX_PROMISC, sg))
1786 		dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1787 			 vi->ctrl->promisc ? "en" : "dis");
1788 
1789 	sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1790 
1791 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1792 				  VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1793 		dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1794 			 vi->ctrl->allmulti ? "en" : "dis");
1795 
1796 	uc_count = netdev_uc_count(dev);
1797 	mc_count = netdev_mc_count(dev);
1798 	/* MAC filter - use one buffer for both lists */
1799 	buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1800 		      (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1801 	mac_data = buf;
1802 	if (!buf)
1803 		return;
1804 
1805 	sg_init_table(sg, 2);
1806 
1807 	/* Store the unicast list and count in the front of the buffer */
1808 	mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1809 	i = 0;
1810 	netdev_for_each_uc_addr(ha, dev)
1811 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1812 
1813 	sg_set_buf(&sg[0], mac_data,
1814 		   sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1815 
1816 	/* multicast list and count fill the end */
1817 	mac_data = (void *)&mac_data->macs[uc_count][0];
1818 
1819 	mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1820 	i = 0;
1821 	netdev_for_each_mc_addr(ha, dev)
1822 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1823 
1824 	sg_set_buf(&sg[1], mac_data,
1825 		   sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1826 
1827 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1828 				  VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1829 		dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1830 
1831 	kfree(buf);
1832 }
1833 
1834 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1835 				   __be16 proto, u16 vid)
1836 {
1837 	struct virtnet_info *vi = netdev_priv(dev);
1838 	struct scatterlist sg;
1839 
1840 	vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1841 	sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1842 
1843 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1844 				  VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1845 		dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1846 	return 0;
1847 }
1848 
1849 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1850 				    __be16 proto, u16 vid)
1851 {
1852 	struct virtnet_info *vi = netdev_priv(dev);
1853 	struct scatterlist sg;
1854 
1855 	vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1856 	sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1857 
1858 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1859 				  VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1860 		dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1861 	return 0;
1862 }
1863 
1864 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1865 {
1866 	int i;
1867 
1868 	if (vi->affinity_hint_set) {
1869 		for (i = 0; i < vi->max_queue_pairs; i++) {
1870 			virtqueue_set_affinity(vi->rq[i].vq, NULL);
1871 			virtqueue_set_affinity(vi->sq[i].vq, NULL);
1872 		}
1873 
1874 		vi->affinity_hint_set = false;
1875 	}
1876 }
1877 
1878 static void virtnet_set_affinity(struct virtnet_info *vi)
1879 {
1880 	cpumask_var_t mask;
1881 	int stragglers;
1882 	int group_size;
1883 	int i, j, cpu;
1884 	int num_cpu;
1885 	int stride;
1886 
1887 	if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
1888 		virtnet_clean_affinity(vi, -1);
1889 		return;
1890 	}
1891 
1892 	num_cpu = num_online_cpus();
1893 	stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
1894 	stragglers = num_cpu >= vi->curr_queue_pairs ?
1895 			num_cpu % vi->curr_queue_pairs :
1896 			0;
1897 	cpu = cpumask_next(-1, cpu_online_mask);
1898 
1899 	for (i = 0; i < vi->curr_queue_pairs; i++) {
1900 		group_size = stride + (i < stragglers ? 1 : 0);
1901 
1902 		for (j = 0; j < group_size; j++) {
1903 			cpumask_set_cpu(cpu, mask);
1904 			cpu = cpumask_next_wrap(cpu, cpu_online_mask,
1905 						nr_cpu_ids, false);
1906 		}
1907 		virtqueue_set_affinity(vi->rq[i].vq, mask);
1908 		virtqueue_set_affinity(vi->sq[i].vq, mask);
1909 		__netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, false);
1910 		cpumask_clear(mask);
1911 	}
1912 
1913 	vi->affinity_hint_set = true;
1914 	free_cpumask_var(mask);
1915 }
1916 
1917 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1918 {
1919 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1920 						   node);
1921 	virtnet_set_affinity(vi);
1922 	return 0;
1923 }
1924 
1925 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1926 {
1927 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1928 						   node_dead);
1929 	virtnet_set_affinity(vi);
1930 	return 0;
1931 }
1932 
1933 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1934 {
1935 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1936 						   node);
1937 
1938 	virtnet_clean_affinity(vi, cpu);
1939 	return 0;
1940 }
1941 
1942 static enum cpuhp_state virtionet_online;
1943 
1944 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1945 {
1946 	int ret;
1947 
1948 	ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1949 	if (ret)
1950 		return ret;
1951 	ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1952 					       &vi->node_dead);
1953 	if (!ret)
1954 		return ret;
1955 	cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1956 	return ret;
1957 }
1958 
1959 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1960 {
1961 	cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1962 	cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1963 					    &vi->node_dead);
1964 }
1965 
1966 static void virtnet_get_ringparam(struct net_device *dev,
1967 				struct ethtool_ringparam *ring)
1968 {
1969 	struct virtnet_info *vi = netdev_priv(dev);
1970 
1971 	ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1972 	ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1973 	ring->rx_pending = ring->rx_max_pending;
1974 	ring->tx_pending = ring->tx_max_pending;
1975 }
1976 
1977 
1978 static void virtnet_get_drvinfo(struct net_device *dev,
1979 				struct ethtool_drvinfo *info)
1980 {
1981 	struct virtnet_info *vi = netdev_priv(dev);
1982 	struct virtio_device *vdev = vi->vdev;
1983 
1984 	strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1985 	strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1986 	strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1987 
1988 }
1989 
1990 /* TODO: Eliminate OOO packets during switching */
1991 static int virtnet_set_channels(struct net_device *dev,
1992 				struct ethtool_channels *channels)
1993 {
1994 	struct virtnet_info *vi = netdev_priv(dev);
1995 	u16 queue_pairs = channels->combined_count;
1996 	int err;
1997 
1998 	/* We don't support separate rx/tx channels.
1999 	 * We don't allow setting 'other' channels.
2000 	 */
2001 	if (channels->rx_count || channels->tx_count || channels->other_count)
2002 		return -EINVAL;
2003 
2004 	if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2005 		return -EINVAL;
2006 
2007 	/* For now we don't support modifying channels while XDP is loaded
2008 	 * also when XDP is loaded all RX queues have XDP programs so we only
2009 	 * need to check a single RX queue.
2010 	 */
2011 	if (vi->rq[0].xdp_prog)
2012 		return -EINVAL;
2013 
2014 	get_online_cpus();
2015 	err = _virtnet_set_queues(vi, queue_pairs);
2016 	if (!err) {
2017 		netif_set_real_num_tx_queues(dev, queue_pairs);
2018 		netif_set_real_num_rx_queues(dev, queue_pairs);
2019 
2020 		virtnet_set_affinity(vi);
2021 	}
2022 	put_online_cpus();
2023 
2024 	return err;
2025 }
2026 
2027 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2028 {
2029 	struct virtnet_info *vi = netdev_priv(dev);
2030 	char *p = (char *)data;
2031 	unsigned int i, j;
2032 
2033 	switch (stringset) {
2034 	case ETH_SS_STATS:
2035 		for (i = 0; i < vi->curr_queue_pairs; i++) {
2036 			for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2037 				snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s",
2038 					 i, virtnet_rq_stats_desc[j].desc);
2039 				p += ETH_GSTRING_LEN;
2040 			}
2041 		}
2042 
2043 		for (i = 0; i < vi->curr_queue_pairs; i++) {
2044 			for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2045 				snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s",
2046 					 i, virtnet_sq_stats_desc[j].desc);
2047 				p += ETH_GSTRING_LEN;
2048 			}
2049 		}
2050 		break;
2051 	}
2052 }
2053 
2054 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2055 {
2056 	struct virtnet_info *vi = netdev_priv(dev);
2057 
2058 	switch (sset) {
2059 	case ETH_SS_STATS:
2060 		return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2061 					       VIRTNET_SQ_STATS_LEN);
2062 	default:
2063 		return -EOPNOTSUPP;
2064 	}
2065 }
2066 
2067 static void virtnet_get_ethtool_stats(struct net_device *dev,
2068 				      struct ethtool_stats *stats, u64 *data)
2069 {
2070 	struct virtnet_info *vi = netdev_priv(dev);
2071 	unsigned int idx = 0, start, i, j;
2072 	const u8 *stats_base;
2073 	size_t offset;
2074 
2075 	for (i = 0; i < vi->curr_queue_pairs; i++) {
2076 		struct receive_queue *rq = &vi->rq[i];
2077 
2078 		stats_base = (u8 *)&rq->stats;
2079 		do {
2080 			start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2081 			for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2082 				offset = virtnet_rq_stats_desc[j].offset;
2083 				data[idx + j] = *(u64 *)(stats_base + offset);
2084 			}
2085 		} while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2086 		idx += VIRTNET_RQ_STATS_LEN;
2087 	}
2088 
2089 	for (i = 0; i < vi->curr_queue_pairs; i++) {
2090 		struct send_queue *sq = &vi->sq[i];
2091 
2092 		stats_base = (u8 *)&sq->stats;
2093 		do {
2094 			start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2095 			for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2096 				offset = virtnet_sq_stats_desc[j].offset;
2097 				data[idx + j] = *(u64 *)(stats_base + offset);
2098 			}
2099 		} while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2100 		idx += VIRTNET_SQ_STATS_LEN;
2101 	}
2102 }
2103 
2104 static void virtnet_get_channels(struct net_device *dev,
2105 				 struct ethtool_channels *channels)
2106 {
2107 	struct virtnet_info *vi = netdev_priv(dev);
2108 
2109 	channels->combined_count = vi->curr_queue_pairs;
2110 	channels->max_combined = vi->max_queue_pairs;
2111 	channels->max_other = 0;
2112 	channels->rx_count = 0;
2113 	channels->tx_count = 0;
2114 	channels->other_count = 0;
2115 }
2116 
2117 /* Check if the user is trying to change anything besides speed/duplex */
2118 static bool
2119 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
2120 {
2121 	struct ethtool_link_ksettings diff1 = *cmd;
2122 	struct ethtool_link_ksettings diff2 = {};
2123 
2124 	/* cmd is always set so we need to clear it, validate the port type
2125 	 * and also without autonegotiation we can ignore advertising
2126 	 */
2127 	diff1.base.speed = 0;
2128 	diff2.base.port = PORT_OTHER;
2129 	ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
2130 	diff1.base.duplex = 0;
2131 	diff1.base.cmd = 0;
2132 	diff1.base.link_mode_masks_nwords = 0;
2133 
2134 	return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
2135 		bitmap_empty(diff1.link_modes.supported,
2136 			     __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2137 		bitmap_empty(diff1.link_modes.advertising,
2138 			     __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2139 		bitmap_empty(diff1.link_modes.lp_advertising,
2140 			     __ETHTOOL_LINK_MODE_MASK_NBITS);
2141 }
2142 
2143 static int virtnet_set_link_ksettings(struct net_device *dev,
2144 				      const struct ethtool_link_ksettings *cmd)
2145 {
2146 	struct virtnet_info *vi = netdev_priv(dev);
2147 	u32 speed;
2148 
2149 	speed = cmd->base.speed;
2150 	/* don't allow custom speed and duplex */
2151 	if (!ethtool_validate_speed(speed) ||
2152 	    !ethtool_validate_duplex(cmd->base.duplex) ||
2153 	    !virtnet_validate_ethtool_cmd(cmd))
2154 		return -EINVAL;
2155 	vi->speed = speed;
2156 	vi->duplex = cmd->base.duplex;
2157 
2158 	return 0;
2159 }
2160 
2161 static int virtnet_get_link_ksettings(struct net_device *dev,
2162 				      struct ethtool_link_ksettings *cmd)
2163 {
2164 	struct virtnet_info *vi = netdev_priv(dev);
2165 
2166 	cmd->base.speed = vi->speed;
2167 	cmd->base.duplex = vi->duplex;
2168 	cmd->base.port = PORT_OTHER;
2169 
2170 	return 0;
2171 }
2172 
2173 static int virtnet_set_coalesce(struct net_device *dev,
2174 				struct ethtool_coalesce *ec)
2175 {
2176 	struct ethtool_coalesce ec_default = {
2177 		.cmd = ETHTOOL_SCOALESCE,
2178 		.rx_max_coalesced_frames = 1,
2179 	};
2180 	struct virtnet_info *vi = netdev_priv(dev);
2181 	int i, napi_weight;
2182 
2183 	if (ec->tx_max_coalesced_frames > 1)
2184 		return -EINVAL;
2185 
2186 	ec_default.tx_max_coalesced_frames = ec->tx_max_coalesced_frames;
2187 	napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2188 
2189 	/* disallow changes to fields not explicitly tested above */
2190 	if (memcmp(ec, &ec_default, sizeof(ec_default)))
2191 		return -EINVAL;
2192 
2193 	if (napi_weight ^ vi->sq[0].napi.weight) {
2194 		if (dev->flags & IFF_UP)
2195 			return -EBUSY;
2196 		for (i = 0; i < vi->max_queue_pairs; i++)
2197 			vi->sq[i].napi.weight = napi_weight;
2198 	}
2199 
2200 	return 0;
2201 }
2202 
2203 static int virtnet_get_coalesce(struct net_device *dev,
2204 				struct ethtool_coalesce *ec)
2205 {
2206 	struct ethtool_coalesce ec_default = {
2207 		.cmd = ETHTOOL_GCOALESCE,
2208 		.rx_max_coalesced_frames = 1,
2209 	};
2210 	struct virtnet_info *vi = netdev_priv(dev);
2211 
2212 	memcpy(ec, &ec_default, sizeof(ec_default));
2213 
2214 	if (vi->sq[0].napi.weight)
2215 		ec->tx_max_coalesced_frames = 1;
2216 
2217 	return 0;
2218 }
2219 
2220 static void virtnet_init_settings(struct net_device *dev)
2221 {
2222 	struct virtnet_info *vi = netdev_priv(dev);
2223 
2224 	vi->speed = SPEED_UNKNOWN;
2225 	vi->duplex = DUPLEX_UNKNOWN;
2226 }
2227 
2228 static void virtnet_update_settings(struct virtnet_info *vi)
2229 {
2230 	u32 speed;
2231 	u8 duplex;
2232 
2233 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2234 		return;
2235 
2236 	speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config,
2237 						  speed));
2238 	if (ethtool_validate_speed(speed))
2239 		vi->speed = speed;
2240 	duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config,
2241 						  duplex));
2242 	if (ethtool_validate_duplex(duplex))
2243 		vi->duplex = duplex;
2244 }
2245 
2246 static const struct ethtool_ops virtnet_ethtool_ops = {
2247 	.get_drvinfo = virtnet_get_drvinfo,
2248 	.get_link = ethtool_op_get_link,
2249 	.get_ringparam = virtnet_get_ringparam,
2250 	.get_strings = virtnet_get_strings,
2251 	.get_sset_count = virtnet_get_sset_count,
2252 	.get_ethtool_stats = virtnet_get_ethtool_stats,
2253 	.set_channels = virtnet_set_channels,
2254 	.get_channels = virtnet_get_channels,
2255 	.get_ts_info = ethtool_op_get_ts_info,
2256 	.get_link_ksettings = virtnet_get_link_ksettings,
2257 	.set_link_ksettings = virtnet_set_link_ksettings,
2258 	.set_coalesce = virtnet_set_coalesce,
2259 	.get_coalesce = virtnet_get_coalesce,
2260 };
2261 
2262 static void virtnet_freeze_down(struct virtio_device *vdev)
2263 {
2264 	struct virtnet_info *vi = vdev->priv;
2265 	int i;
2266 
2267 	/* Make sure no work handler is accessing the device */
2268 	flush_work(&vi->config_work);
2269 
2270 	netif_tx_lock_bh(vi->dev);
2271 	netif_device_detach(vi->dev);
2272 	netif_tx_unlock_bh(vi->dev);
2273 	cancel_delayed_work_sync(&vi->refill);
2274 
2275 	if (netif_running(vi->dev)) {
2276 		for (i = 0; i < vi->max_queue_pairs; i++) {
2277 			napi_disable(&vi->rq[i].napi);
2278 			virtnet_napi_tx_disable(&vi->sq[i].napi);
2279 		}
2280 	}
2281 }
2282 
2283 static int init_vqs(struct virtnet_info *vi);
2284 
2285 static int virtnet_restore_up(struct virtio_device *vdev)
2286 {
2287 	struct virtnet_info *vi = vdev->priv;
2288 	int err, i;
2289 
2290 	err = init_vqs(vi);
2291 	if (err)
2292 		return err;
2293 
2294 	virtio_device_ready(vdev);
2295 
2296 	if (netif_running(vi->dev)) {
2297 		for (i = 0; i < vi->curr_queue_pairs; i++)
2298 			if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2299 				schedule_delayed_work(&vi->refill, 0);
2300 
2301 		for (i = 0; i < vi->max_queue_pairs; i++) {
2302 			virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2303 			virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2304 					       &vi->sq[i].napi);
2305 		}
2306 	}
2307 
2308 	netif_tx_lock_bh(vi->dev);
2309 	netif_device_attach(vi->dev);
2310 	netif_tx_unlock_bh(vi->dev);
2311 	return err;
2312 }
2313 
2314 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2315 {
2316 	struct scatterlist sg;
2317 	vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2318 
2319 	sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2320 
2321 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2322 				  VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2323 		dev_warn(&vi->dev->dev, "Fail to set guest offload. \n");
2324 		return -EINVAL;
2325 	}
2326 
2327 	return 0;
2328 }
2329 
2330 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2331 {
2332 	u64 offloads = 0;
2333 
2334 	if (!vi->guest_offloads)
2335 		return 0;
2336 
2337 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
2338 		offloads = 1ULL << VIRTIO_NET_F_GUEST_CSUM;
2339 
2340 	return virtnet_set_guest_offloads(vi, offloads);
2341 }
2342 
2343 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2344 {
2345 	u64 offloads = vi->guest_offloads;
2346 
2347 	if (!vi->guest_offloads)
2348 		return 0;
2349 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
2350 		offloads |= 1ULL << VIRTIO_NET_F_GUEST_CSUM;
2351 
2352 	return virtnet_set_guest_offloads(vi, offloads);
2353 }
2354 
2355 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2356 			   struct netlink_ext_ack *extack)
2357 {
2358 	unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2359 	struct virtnet_info *vi = netdev_priv(dev);
2360 	struct bpf_prog *old_prog;
2361 	u16 xdp_qp = 0, curr_qp;
2362 	int i, err;
2363 
2364 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2365 	    && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2366 	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2367 	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2368 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO))) {
2369 		NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first");
2370 		return -EOPNOTSUPP;
2371 	}
2372 
2373 	if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2374 		NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2375 		return -EINVAL;
2376 	}
2377 
2378 	if (dev->mtu > max_sz) {
2379 		NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2380 		netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2381 		return -EINVAL;
2382 	}
2383 
2384 	curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2385 	if (prog)
2386 		xdp_qp = nr_cpu_ids;
2387 
2388 	/* XDP requires extra queues for XDP_TX */
2389 	if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2390 		NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2391 		netdev_warn(dev, "request %i queues but max is %i\n",
2392 			    curr_qp + xdp_qp, vi->max_queue_pairs);
2393 		return -ENOMEM;
2394 	}
2395 
2396 	if (prog) {
2397 		prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
2398 		if (IS_ERR(prog))
2399 			return PTR_ERR(prog);
2400 	}
2401 
2402 	/* Make sure NAPI is not using any XDP TX queues for RX. */
2403 	if (netif_running(dev))
2404 		for (i = 0; i < vi->max_queue_pairs; i++)
2405 			napi_disable(&vi->rq[i].napi);
2406 
2407 	netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2408 	err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2409 	if (err)
2410 		goto err;
2411 	vi->xdp_queue_pairs = xdp_qp;
2412 
2413 	for (i = 0; i < vi->max_queue_pairs; i++) {
2414 		old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2415 		rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2416 		if (i == 0) {
2417 			if (!old_prog)
2418 				virtnet_clear_guest_offloads(vi);
2419 			if (!prog)
2420 				virtnet_restore_guest_offloads(vi);
2421 		}
2422 		if (old_prog)
2423 			bpf_prog_put(old_prog);
2424 		if (netif_running(dev))
2425 			virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2426 	}
2427 
2428 	return 0;
2429 
2430 err:
2431 	for (i = 0; i < vi->max_queue_pairs; i++)
2432 		virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2433 	if (prog)
2434 		bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2435 	return err;
2436 }
2437 
2438 static u32 virtnet_xdp_query(struct net_device *dev)
2439 {
2440 	struct virtnet_info *vi = netdev_priv(dev);
2441 	const struct bpf_prog *xdp_prog;
2442 	int i;
2443 
2444 	for (i = 0; i < vi->max_queue_pairs; i++) {
2445 		xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2446 		if (xdp_prog)
2447 			return xdp_prog->aux->id;
2448 	}
2449 	return 0;
2450 }
2451 
2452 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2453 {
2454 	switch (xdp->command) {
2455 	case XDP_SETUP_PROG:
2456 		return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2457 	case XDP_QUERY_PROG:
2458 		xdp->prog_id = virtnet_xdp_query(dev);
2459 		return 0;
2460 	default:
2461 		return -EINVAL;
2462 	}
2463 }
2464 
2465 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2466 				      size_t len)
2467 {
2468 	struct virtnet_info *vi = netdev_priv(dev);
2469 	int ret;
2470 
2471 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2472 		return -EOPNOTSUPP;
2473 
2474 	ret = snprintf(buf, len, "sby");
2475 	if (ret >= len)
2476 		return -EOPNOTSUPP;
2477 
2478 	return 0;
2479 }
2480 
2481 static const struct net_device_ops virtnet_netdev = {
2482 	.ndo_open            = virtnet_open,
2483 	.ndo_stop   	     = virtnet_close,
2484 	.ndo_start_xmit      = start_xmit,
2485 	.ndo_validate_addr   = eth_validate_addr,
2486 	.ndo_set_mac_address = virtnet_set_mac_address,
2487 	.ndo_set_rx_mode     = virtnet_set_rx_mode,
2488 	.ndo_get_stats64     = virtnet_stats,
2489 	.ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2490 	.ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2491 	.ndo_bpf		= virtnet_xdp,
2492 	.ndo_xdp_xmit		= virtnet_xdp_xmit,
2493 	.ndo_features_check	= passthru_features_check,
2494 	.ndo_get_phys_port_name	= virtnet_get_phys_port_name,
2495 };
2496 
2497 static void virtnet_config_changed_work(struct work_struct *work)
2498 {
2499 	struct virtnet_info *vi =
2500 		container_of(work, struct virtnet_info, config_work);
2501 	u16 v;
2502 
2503 	if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2504 				 struct virtio_net_config, status, &v) < 0)
2505 		return;
2506 
2507 	if (v & VIRTIO_NET_S_ANNOUNCE) {
2508 		netdev_notify_peers(vi->dev);
2509 		virtnet_ack_link_announce(vi);
2510 	}
2511 
2512 	/* Ignore unknown (future) status bits */
2513 	v &= VIRTIO_NET_S_LINK_UP;
2514 
2515 	if (vi->status == v)
2516 		return;
2517 
2518 	vi->status = v;
2519 
2520 	if (vi->status & VIRTIO_NET_S_LINK_UP) {
2521 		virtnet_update_settings(vi);
2522 		netif_carrier_on(vi->dev);
2523 		netif_tx_wake_all_queues(vi->dev);
2524 	} else {
2525 		netif_carrier_off(vi->dev);
2526 		netif_tx_stop_all_queues(vi->dev);
2527 	}
2528 }
2529 
2530 static void virtnet_config_changed(struct virtio_device *vdev)
2531 {
2532 	struct virtnet_info *vi = vdev->priv;
2533 
2534 	schedule_work(&vi->config_work);
2535 }
2536 
2537 static void virtnet_free_queues(struct virtnet_info *vi)
2538 {
2539 	int i;
2540 
2541 	for (i = 0; i < vi->max_queue_pairs; i++) {
2542 		napi_hash_del(&vi->rq[i].napi);
2543 		netif_napi_del(&vi->rq[i].napi);
2544 		netif_napi_del(&vi->sq[i].napi);
2545 	}
2546 
2547 	/* We called napi_hash_del() before netif_napi_del(),
2548 	 * we need to respect an RCU grace period before freeing vi->rq
2549 	 */
2550 	synchronize_net();
2551 
2552 	kfree(vi->rq);
2553 	kfree(vi->sq);
2554 	kfree(vi->ctrl);
2555 }
2556 
2557 static void _free_receive_bufs(struct virtnet_info *vi)
2558 {
2559 	struct bpf_prog *old_prog;
2560 	int i;
2561 
2562 	for (i = 0; i < vi->max_queue_pairs; i++) {
2563 		while (vi->rq[i].pages)
2564 			__free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2565 
2566 		old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2567 		RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2568 		if (old_prog)
2569 			bpf_prog_put(old_prog);
2570 	}
2571 }
2572 
2573 static void free_receive_bufs(struct virtnet_info *vi)
2574 {
2575 	rtnl_lock();
2576 	_free_receive_bufs(vi);
2577 	rtnl_unlock();
2578 }
2579 
2580 static void free_receive_page_frags(struct virtnet_info *vi)
2581 {
2582 	int i;
2583 	for (i = 0; i < vi->max_queue_pairs; i++)
2584 		if (vi->rq[i].alloc_frag.page)
2585 			put_page(vi->rq[i].alloc_frag.page);
2586 }
2587 
2588 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
2589 {
2590 	if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
2591 		return false;
2592 	else if (q < vi->curr_queue_pairs)
2593 		return true;
2594 	else
2595 		return false;
2596 }
2597 
2598 static void free_unused_bufs(struct virtnet_info *vi)
2599 {
2600 	void *buf;
2601 	int i;
2602 
2603 	for (i = 0; i < vi->max_queue_pairs; i++) {
2604 		struct virtqueue *vq = vi->sq[i].vq;
2605 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2606 			if (!is_xdp_raw_buffer_queue(vi, i))
2607 				dev_kfree_skb(buf);
2608 			else
2609 				put_page(virt_to_head_page(buf));
2610 		}
2611 	}
2612 
2613 	for (i = 0; i < vi->max_queue_pairs; i++) {
2614 		struct virtqueue *vq = vi->rq[i].vq;
2615 
2616 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2617 			if (vi->mergeable_rx_bufs) {
2618 				put_page(virt_to_head_page(buf));
2619 			} else if (vi->big_packets) {
2620 				give_pages(&vi->rq[i], buf);
2621 			} else {
2622 				put_page(virt_to_head_page(buf));
2623 			}
2624 		}
2625 	}
2626 }
2627 
2628 static void virtnet_del_vqs(struct virtnet_info *vi)
2629 {
2630 	struct virtio_device *vdev = vi->vdev;
2631 
2632 	virtnet_clean_affinity(vi, -1);
2633 
2634 	vdev->config->del_vqs(vdev);
2635 
2636 	virtnet_free_queues(vi);
2637 }
2638 
2639 /* How large should a single buffer be so a queue full of these can fit at
2640  * least one full packet?
2641  * Logic below assumes the mergeable buffer header is used.
2642  */
2643 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2644 {
2645 	const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2646 	unsigned int rq_size = virtqueue_get_vring_size(vq);
2647 	unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2648 	unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2649 	unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2650 
2651 	return max(max(min_buf_len, hdr_len) - hdr_len,
2652 		   (unsigned int)GOOD_PACKET_LEN);
2653 }
2654 
2655 static int virtnet_find_vqs(struct virtnet_info *vi)
2656 {
2657 	vq_callback_t **callbacks;
2658 	struct virtqueue **vqs;
2659 	int ret = -ENOMEM;
2660 	int i, total_vqs;
2661 	const char **names;
2662 	bool *ctx;
2663 
2664 	/* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2665 	 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2666 	 * possible control vq.
2667 	 */
2668 	total_vqs = vi->max_queue_pairs * 2 +
2669 		    virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2670 
2671 	/* Allocate space for find_vqs parameters */
2672 	vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2673 	if (!vqs)
2674 		goto err_vq;
2675 	callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2676 	if (!callbacks)
2677 		goto err_callback;
2678 	names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2679 	if (!names)
2680 		goto err_names;
2681 	if (!vi->big_packets || vi->mergeable_rx_bufs) {
2682 		ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2683 		if (!ctx)
2684 			goto err_ctx;
2685 	} else {
2686 		ctx = NULL;
2687 	}
2688 
2689 	/* Parameters for control virtqueue, if any */
2690 	if (vi->has_cvq) {
2691 		callbacks[total_vqs - 1] = NULL;
2692 		names[total_vqs - 1] = "control";
2693 	}
2694 
2695 	/* Allocate/initialize parameters for send/receive virtqueues */
2696 	for (i = 0; i < vi->max_queue_pairs; i++) {
2697 		callbacks[rxq2vq(i)] = skb_recv_done;
2698 		callbacks[txq2vq(i)] = skb_xmit_done;
2699 		sprintf(vi->rq[i].name, "input.%d", i);
2700 		sprintf(vi->sq[i].name, "output.%d", i);
2701 		names[rxq2vq(i)] = vi->rq[i].name;
2702 		names[txq2vq(i)] = vi->sq[i].name;
2703 		if (ctx)
2704 			ctx[rxq2vq(i)] = true;
2705 	}
2706 
2707 	ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2708 					 names, ctx, NULL);
2709 	if (ret)
2710 		goto err_find;
2711 
2712 	if (vi->has_cvq) {
2713 		vi->cvq = vqs[total_vqs - 1];
2714 		if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2715 			vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2716 	}
2717 
2718 	for (i = 0; i < vi->max_queue_pairs; i++) {
2719 		vi->rq[i].vq = vqs[rxq2vq(i)];
2720 		vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2721 		vi->sq[i].vq = vqs[txq2vq(i)];
2722 	}
2723 
2724 	/* run here: ret == 0. */
2725 
2726 
2727 err_find:
2728 	kfree(ctx);
2729 err_ctx:
2730 	kfree(names);
2731 err_names:
2732 	kfree(callbacks);
2733 err_callback:
2734 	kfree(vqs);
2735 err_vq:
2736 	return ret;
2737 }
2738 
2739 static int virtnet_alloc_queues(struct virtnet_info *vi)
2740 {
2741 	int i;
2742 
2743 	vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2744 	if (!vi->ctrl)
2745 		goto err_ctrl;
2746 	vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2747 	if (!vi->sq)
2748 		goto err_sq;
2749 	vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2750 	if (!vi->rq)
2751 		goto err_rq;
2752 
2753 	INIT_DELAYED_WORK(&vi->refill, refill_work);
2754 	for (i = 0; i < vi->max_queue_pairs; i++) {
2755 		vi->rq[i].pages = NULL;
2756 		netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2757 			       napi_weight);
2758 		netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2759 				  napi_tx ? napi_weight : 0);
2760 
2761 		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2762 		ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2763 		sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2764 
2765 		u64_stats_init(&vi->rq[i].stats.syncp);
2766 		u64_stats_init(&vi->sq[i].stats.syncp);
2767 	}
2768 
2769 	return 0;
2770 
2771 err_rq:
2772 	kfree(vi->sq);
2773 err_sq:
2774 	kfree(vi->ctrl);
2775 err_ctrl:
2776 	return -ENOMEM;
2777 }
2778 
2779 static int init_vqs(struct virtnet_info *vi)
2780 {
2781 	int ret;
2782 
2783 	/* Allocate send & receive queues */
2784 	ret = virtnet_alloc_queues(vi);
2785 	if (ret)
2786 		goto err;
2787 
2788 	ret = virtnet_find_vqs(vi);
2789 	if (ret)
2790 		goto err_free;
2791 
2792 	get_online_cpus();
2793 	virtnet_set_affinity(vi);
2794 	put_online_cpus();
2795 
2796 	return 0;
2797 
2798 err_free:
2799 	virtnet_free_queues(vi);
2800 err:
2801 	return ret;
2802 }
2803 
2804 #ifdef CONFIG_SYSFS
2805 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2806 		char *buf)
2807 {
2808 	struct virtnet_info *vi = netdev_priv(queue->dev);
2809 	unsigned int queue_index = get_netdev_rx_queue_index(queue);
2810 	unsigned int headroom = virtnet_get_headroom(vi);
2811 	unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2812 	struct ewma_pkt_len *avg;
2813 
2814 	BUG_ON(queue_index >= vi->max_queue_pairs);
2815 	avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2816 	return sprintf(buf, "%u\n",
2817 		       get_mergeable_buf_len(&vi->rq[queue_index], avg,
2818 				       SKB_DATA_ALIGN(headroom + tailroom)));
2819 }
2820 
2821 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2822 	__ATTR_RO(mergeable_rx_buffer_size);
2823 
2824 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2825 	&mergeable_rx_buffer_size_attribute.attr,
2826 	NULL
2827 };
2828 
2829 static const struct attribute_group virtio_net_mrg_rx_group = {
2830 	.name = "virtio_net",
2831 	.attrs = virtio_net_mrg_rx_attrs
2832 };
2833 #endif
2834 
2835 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2836 				    unsigned int fbit,
2837 				    const char *fname, const char *dname)
2838 {
2839 	if (!virtio_has_feature(vdev, fbit))
2840 		return false;
2841 
2842 	dev_err(&vdev->dev, "device advertises feature %s but not %s",
2843 		fname, dname);
2844 
2845 	return true;
2846 }
2847 
2848 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)			\
2849 	virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2850 
2851 static bool virtnet_validate_features(struct virtio_device *vdev)
2852 {
2853 	if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2854 	    (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2855 			     "VIRTIO_NET_F_CTRL_VQ") ||
2856 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2857 			     "VIRTIO_NET_F_CTRL_VQ") ||
2858 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2859 			     "VIRTIO_NET_F_CTRL_VQ") ||
2860 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2861 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2862 			     "VIRTIO_NET_F_CTRL_VQ"))) {
2863 		return false;
2864 	}
2865 
2866 	return true;
2867 }
2868 
2869 #define MIN_MTU ETH_MIN_MTU
2870 #define MAX_MTU ETH_MAX_MTU
2871 
2872 static int virtnet_validate(struct virtio_device *vdev)
2873 {
2874 	if (!vdev->config->get) {
2875 		dev_err(&vdev->dev, "%s failure: config access disabled\n",
2876 			__func__);
2877 		return -EINVAL;
2878 	}
2879 
2880 	if (!virtnet_validate_features(vdev))
2881 		return -EINVAL;
2882 
2883 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2884 		int mtu = virtio_cread16(vdev,
2885 					 offsetof(struct virtio_net_config,
2886 						  mtu));
2887 		if (mtu < MIN_MTU)
2888 			__virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2889 	}
2890 
2891 	return 0;
2892 }
2893 
2894 static int virtnet_probe(struct virtio_device *vdev)
2895 {
2896 	int i, err = -ENOMEM;
2897 	struct net_device *dev;
2898 	struct virtnet_info *vi;
2899 	u16 max_queue_pairs;
2900 	int mtu;
2901 
2902 	/* Find if host supports multiqueue virtio_net device */
2903 	err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2904 				   struct virtio_net_config,
2905 				   max_virtqueue_pairs, &max_queue_pairs);
2906 
2907 	/* We need at least 2 queue's */
2908 	if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2909 	    max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2910 	    !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2911 		max_queue_pairs = 1;
2912 
2913 	/* Allocate ourselves a network device with room for our info */
2914 	dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2915 	if (!dev)
2916 		return -ENOMEM;
2917 
2918 	/* Set up network device as normal. */
2919 	dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2920 	dev->netdev_ops = &virtnet_netdev;
2921 	dev->features = NETIF_F_HIGHDMA;
2922 
2923 	dev->ethtool_ops = &virtnet_ethtool_ops;
2924 	SET_NETDEV_DEV(dev, &vdev->dev);
2925 
2926 	/* Do we support "hardware" checksums? */
2927 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2928 		/* This opens up the world of extra features. */
2929 		dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2930 		if (csum)
2931 			dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2932 
2933 		if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2934 			dev->hw_features |= NETIF_F_TSO
2935 				| NETIF_F_TSO_ECN | NETIF_F_TSO6;
2936 		}
2937 		/* Individual feature bits: what can host handle? */
2938 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2939 			dev->hw_features |= NETIF_F_TSO;
2940 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2941 			dev->hw_features |= NETIF_F_TSO6;
2942 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2943 			dev->hw_features |= NETIF_F_TSO_ECN;
2944 
2945 		dev->features |= NETIF_F_GSO_ROBUST;
2946 
2947 		if (gso)
2948 			dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
2949 		/* (!csum && gso) case will be fixed by register_netdev() */
2950 	}
2951 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2952 		dev->features |= NETIF_F_RXCSUM;
2953 
2954 	dev->vlan_features = dev->features;
2955 
2956 	/* MTU range: 68 - 65535 */
2957 	dev->min_mtu = MIN_MTU;
2958 	dev->max_mtu = MAX_MTU;
2959 
2960 	/* Configuration may specify what MAC to use.  Otherwise random. */
2961 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2962 		virtio_cread_bytes(vdev,
2963 				   offsetof(struct virtio_net_config, mac),
2964 				   dev->dev_addr, dev->addr_len);
2965 	else
2966 		eth_hw_addr_random(dev);
2967 
2968 	/* Set up our device-specific information */
2969 	vi = netdev_priv(dev);
2970 	vi->dev = dev;
2971 	vi->vdev = vdev;
2972 	vdev->priv = vi;
2973 
2974 	INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2975 
2976 	/* If we can receive ANY GSO packets, we must allocate large ones. */
2977 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2978 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2979 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2980 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2981 		vi->big_packets = true;
2982 
2983 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2984 		vi->mergeable_rx_bufs = true;
2985 
2986 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2987 	    virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2988 		vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2989 	else
2990 		vi->hdr_len = sizeof(struct virtio_net_hdr);
2991 
2992 	if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2993 	    virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2994 		vi->any_header_sg = true;
2995 
2996 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2997 		vi->has_cvq = true;
2998 
2999 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3000 		mtu = virtio_cread16(vdev,
3001 				     offsetof(struct virtio_net_config,
3002 					      mtu));
3003 		if (mtu < dev->min_mtu) {
3004 			/* Should never trigger: MTU was previously validated
3005 			 * in virtnet_validate.
3006 			 */
3007 			dev_err(&vdev->dev, "device MTU appears to have changed "
3008 				"it is now %d < %d", mtu, dev->min_mtu);
3009 			goto free;
3010 		}
3011 
3012 		dev->mtu = mtu;
3013 		dev->max_mtu = mtu;
3014 
3015 		/* TODO: size buffers correctly in this case. */
3016 		if (dev->mtu > ETH_DATA_LEN)
3017 			vi->big_packets = true;
3018 	}
3019 
3020 	if (vi->any_header_sg)
3021 		dev->needed_headroom = vi->hdr_len;
3022 
3023 	/* Enable multiqueue by default */
3024 	if (num_online_cpus() >= max_queue_pairs)
3025 		vi->curr_queue_pairs = max_queue_pairs;
3026 	else
3027 		vi->curr_queue_pairs = num_online_cpus();
3028 	vi->max_queue_pairs = max_queue_pairs;
3029 
3030 	/* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3031 	err = init_vqs(vi);
3032 	if (err)
3033 		goto free;
3034 
3035 #ifdef CONFIG_SYSFS
3036 	if (vi->mergeable_rx_bufs)
3037 		dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3038 #endif
3039 	netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3040 	netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3041 
3042 	virtnet_init_settings(dev);
3043 
3044 	if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3045 		vi->failover = net_failover_create(vi->dev);
3046 		if (IS_ERR(vi->failover)) {
3047 			err = PTR_ERR(vi->failover);
3048 			goto free_vqs;
3049 		}
3050 	}
3051 
3052 	err = register_netdev(dev);
3053 	if (err) {
3054 		pr_debug("virtio_net: registering device failed\n");
3055 		goto free_failover;
3056 	}
3057 
3058 	virtio_device_ready(vdev);
3059 
3060 	err = virtnet_cpu_notif_add(vi);
3061 	if (err) {
3062 		pr_debug("virtio_net: registering cpu notifier failed\n");
3063 		goto free_unregister_netdev;
3064 	}
3065 
3066 	virtnet_set_queues(vi, vi->curr_queue_pairs);
3067 
3068 	/* Assume link up if device can't report link status,
3069 	   otherwise get link status from config. */
3070 	netif_carrier_off(dev);
3071 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3072 		schedule_work(&vi->config_work);
3073 	} else {
3074 		vi->status = VIRTIO_NET_S_LINK_UP;
3075 		virtnet_update_settings(vi);
3076 		netif_carrier_on(dev);
3077 	}
3078 
3079 	for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3080 		if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3081 			set_bit(guest_offloads[i], &vi->guest_offloads);
3082 
3083 	pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3084 		 dev->name, max_queue_pairs);
3085 
3086 	return 0;
3087 
3088 free_unregister_netdev:
3089 	vi->vdev->config->reset(vdev);
3090 
3091 	unregister_netdev(dev);
3092 free_failover:
3093 	net_failover_destroy(vi->failover);
3094 free_vqs:
3095 	cancel_delayed_work_sync(&vi->refill);
3096 	free_receive_page_frags(vi);
3097 	virtnet_del_vqs(vi);
3098 free:
3099 	free_netdev(dev);
3100 	return err;
3101 }
3102 
3103 static void remove_vq_common(struct virtnet_info *vi)
3104 {
3105 	vi->vdev->config->reset(vi->vdev);
3106 
3107 	/* Free unused buffers in both send and recv, if any. */
3108 	free_unused_bufs(vi);
3109 
3110 	free_receive_bufs(vi);
3111 
3112 	free_receive_page_frags(vi);
3113 
3114 	virtnet_del_vqs(vi);
3115 }
3116 
3117 static void virtnet_remove(struct virtio_device *vdev)
3118 {
3119 	struct virtnet_info *vi = vdev->priv;
3120 
3121 	virtnet_cpu_notif_remove(vi);
3122 
3123 	/* Make sure no work handler is accessing the device. */
3124 	flush_work(&vi->config_work);
3125 
3126 	unregister_netdev(vi->dev);
3127 
3128 	net_failover_destroy(vi->failover);
3129 
3130 	remove_vq_common(vi);
3131 
3132 	free_netdev(vi->dev);
3133 }
3134 
3135 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3136 {
3137 	struct virtnet_info *vi = vdev->priv;
3138 
3139 	virtnet_cpu_notif_remove(vi);
3140 	virtnet_freeze_down(vdev);
3141 	remove_vq_common(vi);
3142 
3143 	return 0;
3144 }
3145 
3146 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3147 {
3148 	struct virtnet_info *vi = vdev->priv;
3149 	int err;
3150 
3151 	err = virtnet_restore_up(vdev);
3152 	if (err)
3153 		return err;
3154 	virtnet_set_queues(vi, vi->curr_queue_pairs);
3155 
3156 	err = virtnet_cpu_notif_add(vi);
3157 	if (err)
3158 		return err;
3159 
3160 	return 0;
3161 }
3162 
3163 static struct virtio_device_id id_table[] = {
3164 	{ VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3165 	{ 0 },
3166 };
3167 
3168 #define VIRTNET_FEATURES \
3169 	VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3170 	VIRTIO_NET_F_MAC, \
3171 	VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3172 	VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3173 	VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3174 	VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3175 	VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3176 	VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3177 	VIRTIO_NET_F_CTRL_MAC_ADDR, \
3178 	VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3179 	VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3180 
3181 static unsigned int features[] = {
3182 	VIRTNET_FEATURES,
3183 };
3184 
3185 static unsigned int features_legacy[] = {
3186 	VIRTNET_FEATURES,
3187 	VIRTIO_NET_F_GSO,
3188 	VIRTIO_F_ANY_LAYOUT,
3189 };
3190 
3191 static struct virtio_driver virtio_net_driver = {
3192 	.feature_table = features,
3193 	.feature_table_size = ARRAY_SIZE(features),
3194 	.feature_table_legacy = features_legacy,
3195 	.feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3196 	.driver.name =	KBUILD_MODNAME,
3197 	.driver.owner =	THIS_MODULE,
3198 	.id_table =	id_table,
3199 	.validate =	virtnet_validate,
3200 	.probe =	virtnet_probe,
3201 	.remove =	virtnet_remove,
3202 	.config_changed = virtnet_config_changed,
3203 #ifdef CONFIG_PM_SLEEP
3204 	.freeze =	virtnet_freeze,
3205 	.restore =	virtnet_restore,
3206 #endif
3207 };
3208 
3209 static __init int virtio_net_driver_init(void)
3210 {
3211 	int ret;
3212 
3213 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3214 				      virtnet_cpu_online,
3215 				      virtnet_cpu_down_prep);
3216 	if (ret < 0)
3217 		goto out;
3218 	virtionet_online = ret;
3219 	ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3220 				      NULL, virtnet_cpu_dead);
3221 	if (ret)
3222 		goto err_dead;
3223 
3224         ret = register_virtio_driver(&virtio_net_driver);
3225 	if (ret)
3226 		goto err_virtio;
3227 	return 0;
3228 err_virtio:
3229 	cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3230 err_dead:
3231 	cpuhp_remove_multi_state(virtionet_online);
3232 out:
3233 	return ret;
3234 }
3235 module_init(virtio_net_driver_init);
3236 
3237 static __exit void virtio_net_driver_exit(void)
3238 {
3239 	unregister_virtio_driver(&virtio_net_driver);
3240 	cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3241 	cpuhp_remove_multi_state(virtionet_online);
3242 }
3243 module_exit(virtio_net_driver_exit);
3244 
3245 MODULE_DEVICE_TABLE(virtio, id_table);
3246 MODULE_DESCRIPTION("Virtio network driver");
3247 MODULE_LICENSE("GPL");
3248