xref: /openbmc/linux/drivers/net/virtio_net.c (revision 458a445deb9c9fb13cec46fe9b179a84d2ff514f)
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 	VIRTIO_NET_F_GUEST_CSUM
75 };
76 
77 struct virtnet_stat_desc {
78 	char desc[ETH_GSTRING_LEN];
79 	size_t offset;
80 };
81 
82 struct virtnet_sq_stats {
83 	struct u64_stats_sync syncp;
84 	u64 packets;
85 	u64 bytes;
86 	u64 xdp_tx;
87 	u64 xdp_tx_drops;
88 	u64 kicks;
89 };
90 
91 struct virtnet_rq_stats {
92 	struct u64_stats_sync syncp;
93 	u64 packets;
94 	u64 bytes;
95 	u64 drops;
96 	u64 xdp_packets;
97 	u64 xdp_tx;
98 	u64 xdp_redirects;
99 	u64 xdp_drops;
100 	u64 kicks;
101 };
102 
103 #define VIRTNET_SQ_STAT(m)	offsetof(struct virtnet_sq_stats, m)
104 #define VIRTNET_RQ_STAT(m)	offsetof(struct virtnet_rq_stats, m)
105 
106 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
107 	{ "packets",		VIRTNET_SQ_STAT(packets) },
108 	{ "bytes",		VIRTNET_SQ_STAT(bytes) },
109 	{ "xdp_tx",		VIRTNET_SQ_STAT(xdp_tx) },
110 	{ "xdp_tx_drops",	VIRTNET_SQ_STAT(xdp_tx_drops) },
111 	{ "kicks",		VIRTNET_SQ_STAT(kicks) },
112 };
113 
114 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
115 	{ "packets",		VIRTNET_RQ_STAT(packets) },
116 	{ "bytes",		VIRTNET_RQ_STAT(bytes) },
117 	{ "drops",		VIRTNET_RQ_STAT(drops) },
118 	{ "xdp_packets",	VIRTNET_RQ_STAT(xdp_packets) },
119 	{ "xdp_tx",		VIRTNET_RQ_STAT(xdp_tx) },
120 	{ "xdp_redirects",	VIRTNET_RQ_STAT(xdp_redirects) },
121 	{ "xdp_drops",		VIRTNET_RQ_STAT(xdp_drops) },
122 	{ "kicks",		VIRTNET_RQ_STAT(kicks) },
123 };
124 
125 #define VIRTNET_SQ_STATS_LEN	ARRAY_SIZE(virtnet_sq_stats_desc)
126 #define VIRTNET_RQ_STATS_LEN	ARRAY_SIZE(virtnet_rq_stats_desc)
127 
128 /* Internal representation of a send virtqueue */
129 struct send_queue {
130 	/* Virtqueue associated with this send _queue */
131 	struct virtqueue *vq;
132 
133 	/* TX: fragments + linear part + virtio header */
134 	struct scatterlist sg[MAX_SKB_FRAGS + 2];
135 
136 	/* Name of the send queue: output.$index */
137 	char name[40];
138 
139 	struct virtnet_sq_stats stats;
140 
141 	struct napi_struct napi;
142 };
143 
144 /* Internal representation of a receive virtqueue */
145 struct receive_queue {
146 	/* Virtqueue associated with this receive_queue */
147 	struct virtqueue *vq;
148 
149 	struct napi_struct napi;
150 
151 	struct bpf_prog __rcu *xdp_prog;
152 
153 	struct virtnet_rq_stats stats;
154 
155 	/* Chain pages by the private ptr. */
156 	struct page *pages;
157 
158 	/* Average packet length for mergeable receive buffers. */
159 	struct ewma_pkt_len mrg_avg_pkt_len;
160 
161 	/* Page frag for packet buffer allocation. */
162 	struct page_frag alloc_frag;
163 
164 	/* RX: fragments + linear part + virtio header */
165 	struct scatterlist sg[MAX_SKB_FRAGS + 2];
166 
167 	/* Min single buffer size for mergeable buffers case. */
168 	unsigned int min_buf_len;
169 
170 	/* Name of this receive queue: input.$index */
171 	char name[40];
172 
173 	struct xdp_rxq_info xdp_rxq;
174 };
175 
176 /* Control VQ buffers: protected by the rtnl lock */
177 struct control_buf {
178 	struct virtio_net_ctrl_hdr hdr;
179 	virtio_net_ctrl_ack status;
180 	struct virtio_net_ctrl_mq mq;
181 	u8 promisc;
182 	u8 allmulti;
183 	__virtio16 vid;
184 	__virtio64 offloads;
185 };
186 
187 struct virtnet_info {
188 	struct virtio_device *vdev;
189 	struct virtqueue *cvq;
190 	struct net_device *dev;
191 	struct send_queue *sq;
192 	struct receive_queue *rq;
193 	unsigned int status;
194 
195 	/* Max # of queue pairs supported by the device */
196 	u16 max_queue_pairs;
197 
198 	/* # of queue pairs currently used by the driver */
199 	u16 curr_queue_pairs;
200 
201 	/* # of XDP queue pairs currently used by the driver */
202 	u16 xdp_queue_pairs;
203 
204 	/* I like... big packets and I cannot lie! */
205 	bool big_packets;
206 
207 	/* Host will merge rx buffers for big packets (shake it! shake it!) */
208 	bool mergeable_rx_bufs;
209 
210 	/* Has control virtqueue */
211 	bool has_cvq;
212 
213 	/* Host can handle any s/g split between our header and packet data */
214 	bool any_header_sg;
215 
216 	/* Packet virtio header size */
217 	u8 hdr_len;
218 
219 	/* Work struct for refilling if we run low on memory. */
220 	struct delayed_work refill;
221 
222 	/* Work struct for config space updates */
223 	struct work_struct config_work;
224 
225 	/* Does the affinity hint is set for virtqueues? */
226 	bool affinity_hint_set;
227 
228 	/* CPU hotplug instances for online & dead */
229 	struct hlist_node node;
230 	struct hlist_node node_dead;
231 
232 	struct control_buf *ctrl;
233 
234 	/* Ethtool settings */
235 	u8 duplex;
236 	u32 speed;
237 
238 	unsigned long guest_offloads;
239 
240 	/* failover when STANDBY feature enabled */
241 	struct failover *failover;
242 };
243 
244 struct padded_vnet_hdr {
245 	struct virtio_net_hdr_mrg_rxbuf hdr;
246 	/*
247 	 * hdr is in a separate sg buffer, and data sg buffer shares same page
248 	 * with this header sg. This padding makes next sg 16 byte aligned
249 	 * after the header.
250 	 */
251 	char padding[4];
252 };
253 
254 /* Converting between virtqueue no. and kernel tx/rx queue no.
255  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
256  */
257 static int vq2txq(struct virtqueue *vq)
258 {
259 	return (vq->index - 1) / 2;
260 }
261 
262 static int txq2vq(int txq)
263 {
264 	return txq * 2 + 1;
265 }
266 
267 static int vq2rxq(struct virtqueue *vq)
268 {
269 	return vq->index / 2;
270 }
271 
272 static int rxq2vq(int rxq)
273 {
274 	return rxq * 2;
275 }
276 
277 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
278 {
279 	return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
280 }
281 
282 /*
283  * private is used to chain pages for big packets, put the whole
284  * most recent used list in the beginning for reuse
285  */
286 static void give_pages(struct receive_queue *rq, struct page *page)
287 {
288 	struct page *end;
289 
290 	/* Find end of list, sew whole thing into vi->rq.pages. */
291 	for (end = page; end->private; end = (struct page *)end->private);
292 	end->private = (unsigned long)rq->pages;
293 	rq->pages = page;
294 }
295 
296 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
297 {
298 	struct page *p = rq->pages;
299 
300 	if (p) {
301 		rq->pages = (struct page *)p->private;
302 		/* clear private here, it is used to chain pages */
303 		p->private = 0;
304 	} else
305 		p = alloc_page(gfp_mask);
306 	return p;
307 }
308 
309 static void virtqueue_napi_schedule(struct napi_struct *napi,
310 				    struct virtqueue *vq)
311 {
312 	if (napi_schedule_prep(napi)) {
313 		virtqueue_disable_cb(vq);
314 		__napi_schedule(napi);
315 	}
316 }
317 
318 static void virtqueue_napi_complete(struct napi_struct *napi,
319 				    struct virtqueue *vq, int processed)
320 {
321 	int opaque;
322 
323 	opaque = virtqueue_enable_cb_prepare(vq);
324 	if (napi_complete_done(napi, processed)) {
325 		if (unlikely(virtqueue_poll(vq, opaque)))
326 			virtqueue_napi_schedule(napi, vq);
327 	} else {
328 		virtqueue_disable_cb(vq);
329 	}
330 }
331 
332 static void skb_xmit_done(struct virtqueue *vq)
333 {
334 	struct virtnet_info *vi = vq->vdev->priv;
335 	struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
336 
337 	/* Suppress further interrupts. */
338 	virtqueue_disable_cb(vq);
339 
340 	if (napi->weight)
341 		virtqueue_napi_schedule(napi, vq);
342 	else
343 		/* We were probably waiting for more output buffers. */
344 		netif_wake_subqueue(vi->dev, vq2txq(vq));
345 }
346 
347 #define MRG_CTX_HEADER_SHIFT 22
348 static void *mergeable_len_to_ctx(unsigned int truesize,
349 				  unsigned int headroom)
350 {
351 	return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
352 }
353 
354 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
355 {
356 	return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
357 }
358 
359 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
360 {
361 	return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
362 }
363 
364 /* Called from bottom half context */
365 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
366 				   struct receive_queue *rq,
367 				   struct page *page, unsigned int offset,
368 				   unsigned int len, unsigned int truesize)
369 {
370 	struct sk_buff *skb;
371 	struct virtio_net_hdr_mrg_rxbuf *hdr;
372 	unsigned int copy, hdr_len, hdr_padded_len;
373 	char *p;
374 
375 	p = page_address(page) + offset;
376 
377 	/* copy small packet so we can reuse these pages for small data */
378 	skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
379 	if (unlikely(!skb))
380 		return NULL;
381 
382 	hdr = skb_vnet_hdr(skb);
383 
384 	hdr_len = vi->hdr_len;
385 	if (vi->mergeable_rx_bufs)
386 		hdr_padded_len = sizeof(*hdr);
387 	else
388 		hdr_padded_len = sizeof(struct padded_vnet_hdr);
389 
390 	memcpy(hdr, p, hdr_len);
391 
392 	len -= hdr_len;
393 	offset += hdr_padded_len;
394 	p += hdr_padded_len;
395 
396 	copy = len;
397 	if (copy > skb_tailroom(skb))
398 		copy = skb_tailroom(skb);
399 	skb_put_data(skb, p, copy);
400 
401 	len -= copy;
402 	offset += copy;
403 
404 	if (vi->mergeable_rx_bufs) {
405 		if (len)
406 			skb_add_rx_frag(skb, 0, page, offset, len, truesize);
407 		else
408 			put_page(page);
409 		return skb;
410 	}
411 
412 	/*
413 	 * Verify that we can indeed put this data into a skb.
414 	 * This is here to handle cases when the device erroneously
415 	 * tries to receive more than is possible. This is usually
416 	 * the case of a broken device.
417 	 */
418 	if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
419 		net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
420 		dev_kfree_skb(skb);
421 		return NULL;
422 	}
423 	BUG_ON(offset >= PAGE_SIZE);
424 	while (len) {
425 		unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
426 		skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
427 				frag_size, truesize);
428 		len -= frag_size;
429 		page = (struct page *)page->private;
430 		offset = 0;
431 	}
432 
433 	if (page)
434 		give_pages(rq, page);
435 
436 	return skb;
437 }
438 
439 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
440 				   struct send_queue *sq,
441 				   struct xdp_frame *xdpf)
442 {
443 	struct virtio_net_hdr_mrg_rxbuf *hdr;
444 	int err;
445 
446 	/* virtqueue want to use data area in-front of packet */
447 	if (unlikely(xdpf->metasize > 0))
448 		return -EOPNOTSUPP;
449 
450 	if (unlikely(xdpf->headroom < vi->hdr_len))
451 		return -EOVERFLOW;
452 
453 	/* Make room for virtqueue hdr (also change xdpf->headroom?) */
454 	xdpf->data -= vi->hdr_len;
455 	/* Zero header and leave csum up to XDP layers */
456 	hdr = xdpf->data;
457 	memset(hdr, 0, vi->hdr_len);
458 	xdpf->len   += vi->hdr_len;
459 
460 	sg_init_one(sq->sg, xdpf->data, xdpf->len);
461 
462 	err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdpf, GFP_ATOMIC);
463 	if (unlikely(err))
464 		return -ENOSPC; /* Caller handle free/refcnt */
465 
466 	return 0;
467 }
468 
469 static struct send_queue *virtnet_xdp_sq(struct virtnet_info *vi)
470 {
471 	unsigned int qp;
472 
473 	qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
474 	return &vi->sq[qp];
475 }
476 
477 static int virtnet_xdp_xmit(struct net_device *dev,
478 			    int n, struct xdp_frame **frames, u32 flags)
479 {
480 	struct virtnet_info *vi = netdev_priv(dev);
481 	struct receive_queue *rq = vi->rq;
482 	struct xdp_frame *xdpf_sent;
483 	struct bpf_prog *xdp_prog;
484 	struct send_queue *sq;
485 	unsigned int len;
486 	int drops = 0;
487 	int kicks = 0;
488 	int ret, err;
489 	int i;
490 
491 	sq = virtnet_xdp_sq(vi);
492 
493 	if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
494 		ret = -EINVAL;
495 		drops = n;
496 		goto out;
497 	}
498 
499 	/* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
500 	 * indicate XDP resources have been successfully allocated.
501 	 */
502 	xdp_prog = rcu_dereference(rq->xdp_prog);
503 	if (!xdp_prog) {
504 		ret = -ENXIO;
505 		drops = n;
506 		goto out;
507 	}
508 
509 	/* Free up any pending old buffers before queueing new ones. */
510 	while ((xdpf_sent = virtqueue_get_buf(sq->vq, &len)) != NULL)
511 		xdp_return_frame(xdpf_sent);
512 
513 	for (i = 0; i < n; i++) {
514 		struct xdp_frame *xdpf = frames[i];
515 
516 		err = __virtnet_xdp_xmit_one(vi, sq, xdpf);
517 		if (err) {
518 			xdp_return_frame_rx_napi(xdpf);
519 			drops++;
520 		}
521 	}
522 	ret = n - drops;
523 
524 	if (flags & XDP_XMIT_FLUSH) {
525 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
526 			kicks = 1;
527 	}
528 out:
529 	u64_stats_update_begin(&sq->stats.syncp);
530 	sq->stats.xdp_tx += n;
531 	sq->stats.xdp_tx_drops += drops;
532 	sq->stats.kicks += kicks;
533 	u64_stats_update_end(&sq->stats.syncp);
534 
535 	return ret;
536 }
537 
538 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
539 {
540 	return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
541 }
542 
543 /* We copy the packet for XDP in the following cases:
544  *
545  * 1) Packet is scattered across multiple rx buffers.
546  * 2) Headroom space is insufficient.
547  *
548  * This is inefficient but it's a temporary condition that
549  * we hit right after XDP is enabled and until queue is refilled
550  * with large buffers with sufficient headroom - so it should affect
551  * at most queue size packets.
552  * Afterwards, the conditions to enable
553  * XDP should preclude the underlying device from sending packets
554  * across multiple buffers (num_buf > 1), and we make sure buffers
555  * have enough headroom.
556  */
557 static struct page *xdp_linearize_page(struct receive_queue *rq,
558 				       u16 *num_buf,
559 				       struct page *p,
560 				       int offset,
561 				       int page_off,
562 				       unsigned int *len)
563 {
564 	struct page *page = alloc_page(GFP_ATOMIC);
565 
566 	if (!page)
567 		return NULL;
568 
569 	memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
570 	page_off += *len;
571 
572 	while (--*num_buf) {
573 		int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
574 		unsigned int buflen;
575 		void *buf;
576 		int off;
577 
578 		buf = virtqueue_get_buf(rq->vq, &buflen);
579 		if (unlikely(!buf))
580 			goto err_buf;
581 
582 		p = virt_to_head_page(buf);
583 		off = buf - page_address(p);
584 
585 		/* guard against a misconfigured or uncooperative backend that
586 		 * is sending packet larger than the MTU.
587 		 */
588 		if ((page_off + buflen + tailroom) > PAGE_SIZE) {
589 			put_page(p);
590 			goto err_buf;
591 		}
592 
593 		memcpy(page_address(page) + page_off,
594 		       page_address(p) + off, buflen);
595 		page_off += buflen;
596 		put_page(p);
597 	}
598 
599 	/* Headroom does not contribute to packet length */
600 	*len = page_off - VIRTIO_XDP_HEADROOM;
601 	return page;
602 err_buf:
603 	__free_pages(page, 0);
604 	return NULL;
605 }
606 
607 static struct sk_buff *receive_small(struct net_device *dev,
608 				     struct virtnet_info *vi,
609 				     struct receive_queue *rq,
610 				     void *buf, void *ctx,
611 				     unsigned int len,
612 				     unsigned int *xdp_xmit,
613 				     struct virtnet_rq_stats *stats)
614 {
615 	struct sk_buff *skb;
616 	struct bpf_prog *xdp_prog;
617 	unsigned int xdp_headroom = (unsigned long)ctx;
618 	unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
619 	unsigned int headroom = vi->hdr_len + header_offset;
620 	unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
621 			      SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
622 	struct page *page = virt_to_head_page(buf);
623 	unsigned int delta = 0;
624 	struct page *xdp_page;
625 	int err;
626 
627 	len -= vi->hdr_len;
628 	stats->bytes += len;
629 
630 	rcu_read_lock();
631 	xdp_prog = rcu_dereference(rq->xdp_prog);
632 	if (xdp_prog) {
633 		struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
634 		struct xdp_frame *xdpf;
635 		struct xdp_buff xdp;
636 		void *orig_data;
637 		u32 act;
638 
639 		if (unlikely(hdr->hdr.gso_type))
640 			goto err_xdp;
641 
642 		if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
643 			int offset = buf - page_address(page) + header_offset;
644 			unsigned int tlen = len + vi->hdr_len;
645 			u16 num_buf = 1;
646 
647 			xdp_headroom = virtnet_get_headroom(vi);
648 			header_offset = VIRTNET_RX_PAD + xdp_headroom;
649 			headroom = vi->hdr_len + header_offset;
650 			buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
651 				 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
652 			xdp_page = xdp_linearize_page(rq, &num_buf, page,
653 						      offset, header_offset,
654 						      &tlen);
655 			if (!xdp_page)
656 				goto err_xdp;
657 
658 			buf = page_address(xdp_page);
659 			put_page(page);
660 			page = xdp_page;
661 		}
662 
663 		xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
664 		xdp.data = xdp.data_hard_start + xdp_headroom;
665 		xdp_set_data_meta_invalid(&xdp);
666 		xdp.data_end = xdp.data + len;
667 		xdp.rxq = &rq->xdp_rxq;
668 		orig_data = xdp.data;
669 		act = bpf_prog_run_xdp(xdp_prog, &xdp);
670 		stats->xdp_packets++;
671 
672 		switch (act) {
673 		case XDP_PASS:
674 			/* Recalculate length in case bpf program changed it */
675 			delta = orig_data - xdp.data;
676 			len = xdp.data_end - xdp.data;
677 			break;
678 		case XDP_TX:
679 			stats->xdp_tx++;
680 			xdpf = convert_to_xdp_frame(&xdp);
681 			if (unlikely(!xdpf))
682 				goto err_xdp;
683 			err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
684 			if (unlikely(err < 0)) {
685 				trace_xdp_exception(vi->dev, xdp_prog, act);
686 				goto err_xdp;
687 			}
688 			*xdp_xmit |= VIRTIO_XDP_TX;
689 			rcu_read_unlock();
690 			goto xdp_xmit;
691 		case XDP_REDIRECT:
692 			stats->xdp_redirects++;
693 			err = xdp_do_redirect(dev, &xdp, xdp_prog);
694 			if (err)
695 				goto err_xdp;
696 			*xdp_xmit |= VIRTIO_XDP_REDIR;
697 			rcu_read_unlock();
698 			goto xdp_xmit;
699 		default:
700 			bpf_warn_invalid_xdp_action(act);
701 			/* fall through */
702 		case XDP_ABORTED:
703 			trace_xdp_exception(vi->dev, xdp_prog, act);
704 		case XDP_DROP:
705 			goto err_xdp;
706 		}
707 	}
708 	rcu_read_unlock();
709 
710 	skb = build_skb(buf, buflen);
711 	if (!skb) {
712 		put_page(page);
713 		goto err;
714 	}
715 	skb_reserve(skb, headroom - delta);
716 	skb_put(skb, len);
717 	if (!delta) {
718 		buf += header_offset;
719 		memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
720 	} /* keep zeroed vnet hdr since packet was changed by bpf */
721 
722 err:
723 	return skb;
724 
725 err_xdp:
726 	rcu_read_unlock();
727 	stats->xdp_drops++;
728 	stats->drops++;
729 	put_page(page);
730 xdp_xmit:
731 	return NULL;
732 }
733 
734 static struct sk_buff *receive_big(struct net_device *dev,
735 				   struct virtnet_info *vi,
736 				   struct receive_queue *rq,
737 				   void *buf,
738 				   unsigned int len,
739 				   struct virtnet_rq_stats *stats)
740 {
741 	struct page *page = buf;
742 	struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
743 
744 	stats->bytes += len - vi->hdr_len;
745 	if (unlikely(!skb))
746 		goto err;
747 
748 	return skb;
749 
750 err:
751 	stats->drops++;
752 	give_pages(rq, page);
753 	return NULL;
754 }
755 
756 static struct sk_buff *receive_mergeable(struct net_device *dev,
757 					 struct virtnet_info *vi,
758 					 struct receive_queue *rq,
759 					 void *buf,
760 					 void *ctx,
761 					 unsigned int len,
762 					 unsigned int *xdp_xmit,
763 					 struct virtnet_rq_stats *stats)
764 {
765 	struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
766 	u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
767 	struct page *page = virt_to_head_page(buf);
768 	int offset = buf - page_address(page);
769 	struct sk_buff *head_skb, *curr_skb;
770 	struct bpf_prog *xdp_prog;
771 	unsigned int truesize;
772 	unsigned int headroom = mergeable_ctx_to_headroom(ctx);
773 	int err;
774 
775 	head_skb = NULL;
776 	stats->bytes += len - vi->hdr_len;
777 
778 	rcu_read_lock();
779 	xdp_prog = rcu_dereference(rq->xdp_prog);
780 	if (xdp_prog) {
781 		struct xdp_frame *xdpf;
782 		struct page *xdp_page;
783 		struct xdp_buff xdp;
784 		void *data;
785 		u32 act;
786 
787 		/* Transient failure which in theory could occur if
788 		 * in-flight packets from before XDP was enabled reach
789 		 * the receive path after XDP is loaded.
790 		 */
791 		if (unlikely(hdr->hdr.gso_type))
792 			goto err_xdp;
793 
794 		/* This happens when rx buffer size is underestimated
795 		 * or headroom is not enough because of the buffer
796 		 * was refilled before XDP is set. This should only
797 		 * happen for the first several packets, so we don't
798 		 * care much about its performance.
799 		 */
800 		if (unlikely(num_buf > 1 ||
801 			     headroom < virtnet_get_headroom(vi))) {
802 			/* linearize data for XDP */
803 			xdp_page = xdp_linearize_page(rq, &num_buf,
804 						      page, offset,
805 						      VIRTIO_XDP_HEADROOM,
806 						      &len);
807 			if (!xdp_page)
808 				goto err_xdp;
809 			offset = VIRTIO_XDP_HEADROOM;
810 		} else {
811 			xdp_page = page;
812 		}
813 
814 		/* Allow consuming headroom but reserve enough space to push
815 		 * the descriptor on if we get an XDP_TX return code.
816 		 */
817 		data = page_address(xdp_page) + offset;
818 		xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
819 		xdp.data = data + vi->hdr_len;
820 		xdp_set_data_meta_invalid(&xdp);
821 		xdp.data_end = xdp.data + (len - vi->hdr_len);
822 		xdp.rxq = &rq->xdp_rxq;
823 
824 		act = bpf_prog_run_xdp(xdp_prog, &xdp);
825 		stats->xdp_packets++;
826 
827 		switch (act) {
828 		case XDP_PASS:
829 			/* recalculate offset to account for any header
830 			 * adjustments. Note other cases do not build an
831 			 * skb and avoid using offset
832 			 */
833 			offset = xdp.data -
834 					page_address(xdp_page) - vi->hdr_len;
835 
836 			/* recalculate len if xdp.data or xdp.data_end were
837 			 * adjusted
838 			 */
839 			len = xdp.data_end - xdp.data + vi->hdr_len;
840 			/* We can only create skb based on xdp_page. */
841 			if (unlikely(xdp_page != page)) {
842 				rcu_read_unlock();
843 				put_page(page);
844 				head_skb = page_to_skb(vi, rq, xdp_page,
845 						       offset, len, PAGE_SIZE);
846 				return head_skb;
847 			}
848 			break;
849 		case XDP_TX:
850 			stats->xdp_tx++;
851 			xdpf = convert_to_xdp_frame(&xdp);
852 			if (unlikely(!xdpf))
853 				goto err_xdp;
854 			err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
855 			if (unlikely(err < 0)) {
856 				trace_xdp_exception(vi->dev, xdp_prog, act);
857 				if (unlikely(xdp_page != page))
858 					put_page(xdp_page);
859 				goto err_xdp;
860 			}
861 			*xdp_xmit |= VIRTIO_XDP_TX;
862 			if (unlikely(xdp_page != page))
863 				put_page(page);
864 			rcu_read_unlock();
865 			goto xdp_xmit;
866 		case XDP_REDIRECT:
867 			stats->xdp_redirects++;
868 			err = xdp_do_redirect(dev, &xdp, xdp_prog);
869 			if (err) {
870 				if (unlikely(xdp_page != page))
871 					put_page(xdp_page);
872 				goto err_xdp;
873 			}
874 			*xdp_xmit |= VIRTIO_XDP_REDIR;
875 			if (unlikely(xdp_page != page))
876 				put_page(page);
877 			rcu_read_unlock();
878 			goto xdp_xmit;
879 		default:
880 			bpf_warn_invalid_xdp_action(act);
881 			/* fall through */
882 		case XDP_ABORTED:
883 			trace_xdp_exception(vi->dev, xdp_prog, act);
884 			/* fall through */
885 		case XDP_DROP:
886 			if (unlikely(xdp_page != page))
887 				__free_pages(xdp_page, 0);
888 			goto err_xdp;
889 		}
890 	}
891 	rcu_read_unlock();
892 
893 	truesize = mergeable_ctx_to_truesize(ctx);
894 	if (unlikely(len > truesize)) {
895 		pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
896 			 dev->name, len, (unsigned long)ctx);
897 		dev->stats.rx_length_errors++;
898 		goto err_skb;
899 	}
900 
901 	head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
902 	curr_skb = head_skb;
903 
904 	if (unlikely(!curr_skb))
905 		goto err_skb;
906 	while (--num_buf) {
907 		int num_skb_frags;
908 
909 		buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
910 		if (unlikely(!buf)) {
911 			pr_debug("%s: rx error: %d buffers out of %d missing\n",
912 				 dev->name, num_buf,
913 				 virtio16_to_cpu(vi->vdev,
914 						 hdr->num_buffers));
915 			dev->stats.rx_length_errors++;
916 			goto err_buf;
917 		}
918 
919 		stats->bytes += len;
920 		page = virt_to_head_page(buf);
921 
922 		truesize = mergeable_ctx_to_truesize(ctx);
923 		if (unlikely(len > truesize)) {
924 			pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
925 				 dev->name, len, (unsigned long)ctx);
926 			dev->stats.rx_length_errors++;
927 			goto err_skb;
928 		}
929 
930 		num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
931 		if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
932 			struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
933 
934 			if (unlikely(!nskb))
935 				goto err_skb;
936 			if (curr_skb == head_skb)
937 				skb_shinfo(curr_skb)->frag_list = nskb;
938 			else
939 				curr_skb->next = nskb;
940 			curr_skb = nskb;
941 			head_skb->truesize += nskb->truesize;
942 			num_skb_frags = 0;
943 		}
944 		if (curr_skb != head_skb) {
945 			head_skb->data_len += len;
946 			head_skb->len += len;
947 			head_skb->truesize += truesize;
948 		}
949 		offset = buf - page_address(page);
950 		if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
951 			put_page(page);
952 			skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
953 					     len, truesize);
954 		} else {
955 			skb_add_rx_frag(curr_skb, num_skb_frags, page,
956 					offset, len, truesize);
957 		}
958 	}
959 
960 	ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
961 	return head_skb;
962 
963 err_xdp:
964 	rcu_read_unlock();
965 	stats->xdp_drops++;
966 err_skb:
967 	put_page(page);
968 	while (num_buf-- > 1) {
969 		buf = virtqueue_get_buf(rq->vq, &len);
970 		if (unlikely(!buf)) {
971 			pr_debug("%s: rx error: %d buffers missing\n",
972 				 dev->name, num_buf);
973 			dev->stats.rx_length_errors++;
974 			break;
975 		}
976 		stats->bytes += len;
977 		page = virt_to_head_page(buf);
978 		put_page(page);
979 	}
980 err_buf:
981 	stats->drops++;
982 	dev_kfree_skb(head_skb);
983 xdp_xmit:
984 	return NULL;
985 }
986 
987 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
988 			void *buf, unsigned int len, void **ctx,
989 			unsigned int *xdp_xmit,
990 			struct virtnet_rq_stats *stats)
991 {
992 	struct net_device *dev = vi->dev;
993 	struct sk_buff *skb;
994 	struct virtio_net_hdr_mrg_rxbuf *hdr;
995 
996 	if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
997 		pr_debug("%s: short packet %i\n", dev->name, len);
998 		dev->stats.rx_length_errors++;
999 		if (vi->mergeable_rx_bufs) {
1000 			put_page(virt_to_head_page(buf));
1001 		} else if (vi->big_packets) {
1002 			give_pages(rq, buf);
1003 		} else {
1004 			put_page(virt_to_head_page(buf));
1005 		}
1006 		return;
1007 	}
1008 
1009 	if (vi->mergeable_rx_bufs)
1010 		skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1011 					stats);
1012 	else if (vi->big_packets)
1013 		skb = receive_big(dev, vi, rq, buf, len, stats);
1014 	else
1015 		skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1016 
1017 	if (unlikely(!skb))
1018 		return;
1019 
1020 	hdr = skb_vnet_hdr(skb);
1021 
1022 	if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1023 		skb->ip_summed = CHECKSUM_UNNECESSARY;
1024 
1025 	if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1026 				  virtio_is_little_endian(vi->vdev))) {
1027 		net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1028 				     dev->name, hdr->hdr.gso_type,
1029 				     hdr->hdr.gso_size);
1030 		goto frame_err;
1031 	}
1032 
1033 	skb->protocol = eth_type_trans(skb, dev);
1034 	pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1035 		 ntohs(skb->protocol), skb->len, skb->pkt_type);
1036 
1037 	napi_gro_receive(&rq->napi, skb);
1038 	return;
1039 
1040 frame_err:
1041 	dev->stats.rx_frame_errors++;
1042 	dev_kfree_skb(skb);
1043 }
1044 
1045 /* Unlike mergeable buffers, all buffers are allocated to the
1046  * same size, except for the headroom. For this reason we do
1047  * not need to use  mergeable_len_to_ctx here - it is enough
1048  * to store the headroom as the context ignoring the truesize.
1049  */
1050 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1051 			     gfp_t gfp)
1052 {
1053 	struct page_frag *alloc_frag = &rq->alloc_frag;
1054 	char *buf;
1055 	unsigned int xdp_headroom = virtnet_get_headroom(vi);
1056 	void *ctx = (void *)(unsigned long)xdp_headroom;
1057 	int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1058 	int err;
1059 
1060 	len = SKB_DATA_ALIGN(len) +
1061 	      SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1062 	if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1063 		return -ENOMEM;
1064 
1065 	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1066 	get_page(alloc_frag->page);
1067 	alloc_frag->offset += len;
1068 	sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1069 		    vi->hdr_len + GOOD_PACKET_LEN);
1070 	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1071 	if (err < 0)
1072 		put_page(virt_to_head_page(buf));
1073 	return err;
1074 }
1075 
1076 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1077 			   gfp_t gfp)
1078 {
1079 	struct page *first, *list = NULL;
1080 	char *p;
1081 	int i, err, offset;
1082 
1083 	sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1084 
1085 	/* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1086 	for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1087 		first = get_a_page(rq, gfp);
1088 		if (!first) {
1089 			if (list)
1090 				give_pages(rq, list);
1091 			return -ENOMEM;
1092 		}
1093 		sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1094 
1095 		/* chain new page in list head to match sg */
1096 		first->private = (unsigned long)list;
1097 		list = first;
1098 	}
1099 
1100 	first = get_a_page(rq, gfp);
1101 	if (!first) {
1102 		give_pages(rq, list);
1103 		return -ENOMEM;
1104 	}
1105 	p = page_address(first);
1106 
1107 	/* rq->sg[0], rq->sg[1] share the same page */
1108 	/* a separated rq->sg[0] for header - required in case !any_header_sg */
1109 	sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1110 
1111 	/* rq->sg[1] for data packet, from offset */
1112 	offset = sizeof(struct padded_vnet_hdr);
1113 	sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1114 
1115 	/* chain first in list head */
1116 	first->private = (unsigned long)list;
1117 	err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1118 				  first, gfp);
1119 	if (err < 0)
1120 		give_pages(rq, first);
1121 
1122 	return err;
1123 }
1124 
1125 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1126 					  struct ewma_pkt_len *avg_pkt_len,
1127 					  unsigned int room)
1128 {
1129 	const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1130 	unsigned int len;
1131 
1132 	if (room)
1133 		return PAGE_SIZE - room;
1134 
1135 	len = hdr_len +	clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1136 				rq->min_buf_len, PAGE_SIZE - hdr_len);
1137 
1138 	return ALIGN(len, L1_CACHE_BYTES);
1139 }
1140 
1141 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1142 				 struct receive_queue *rq, gfp_t gfp)
1143 {
1144 	struct page_frag *alloc_frag = &rq->alloc_frag;
1145 	unsigned int headroom = virtnet_get_headroom(vi);
1146 	unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1147 	unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1148 	char *buf;
1149 	void *ctx;
1150 	int err;
1151 	unsigned int len, hole;
1152 
1153 	/* Extra tailroom is needed to satisfy XDP's assumption. This
1154 	 * means rx frags coalescing won't work, but consider we've
1155 	 * disabled GSO for XDP, it won't be a big issue.
1156 	 */
1157 	len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1158 	if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1159 		return -ENOMEM;
1160 
1161 	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1162 	buf += headroom; /* advance address leaving hole at front of pkt */
1163 	get_page(alloc_frag->page);
1164 	alloc_frag->offset += len + room;
1165 	hole = alloc_frag->size - alloc_frag->offset;
1166 	if (hole < len + room) {
1167 		/* To avoid internal fragmentation, if there is very likely not
1168 		 * enough space for another buffer, add the remaining space to
1169 		 * the current buffer.
1170 		 */
1171 		len += hole;
1172 		alloc_frag->offset += hole;
1173 	}
1174 
1175 	sg_init_one(rq->sg, buf, len);
1176 	ctx = mergeable_len_to_ctx(len, headroom);
1177 	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1178 	if (err < 0)
1179 		put_page(virt_to_head_page(buf));
1180 
1181 	return err;
1182 }
1183 
1184 /*
1185  * Returns false if we couldn't fill entirely (OOM).
1186  *
1187  * Normally run in the receive path, but can also be run from ndo_open
1188  * before we're receiving packets, or from refill_work which is
1189  * careful to disable receiving (using napi_disable).
1190  */
1191 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1192 			  gfp_t gfp)
1193 {
1194 	int err;
1195 	bool oom;
1196 
1197 	do {
1198 		if (vi->mergeable_rx_bufs)
1199 			err = add_recvbuf_mergeable(vi, rq, gfp);
1200 		else if (vi->big_packets)
1201 			err = add_recvbuf_big(vi, rq, gfp);
1202 		else
1203 			err = add_recvbuf_small(vi, rq, gfp);
1204 
1205 		oom = err == -ENOMEM;
1206 		if (err)
1207 			break;
1208 	} while (rq->vq->num_free);
1209 	if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1210 		u64_stats_update_begin(&rq->stats.syncp);
1211 		rq->stats.kicks++;
1212 		u64_stats_update_end(&rq->stats.syncp);
1213 	}
1214 
1215 	return !oom;
1216 }
1217 
1218 static void skb_recv_done(struct virtqueue *rvq)
1219 {
1220 	struct virtnet_info *vi = rvq->vdev->priv;
1221 	struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1222 
1223 	virtqueue_napi_schedule(&rq->napi, rvq);
1224 }
1225 
1226 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1227 {
1228 	napi_enable(napi);
1229 
1230 	/* If all buffers were filled by other side before we napi_enabled, we
1231 	 * won't get another interrupt, so process any outstanding packets now.
1232 	 * Call local_bh_enable after to trigger softIRQ processing.
1233 	 */
1234 	local_bh_disable();
1235 	virtqueue_napi_schedule(napi, vq);
1236 	local_bh_enable();
1237 }
1238 
1239 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1240 				   struct virtqueue *vq,
1241 				   struct napi_struct *napi)
1242 {
1243 	if (!napi->weight)
1244 		return;
1245 
1246 	/* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1247 	 * enable the feature if this is likely affine with the transmit path.
1248 	 */
1249 	if (!vi->affinity_hint_set) {
1250 		napi->weight = 0;
1251 		return;
1252 	}
1253 
1254 	return virtnet_napi_enable(vq, napi);
1255 }
1256 
1257 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1258 {
1259 	if (napi->weight)
1260 		napi_disable(napi);
1261 }
1262 
1263 static void refill_work(struct work_struct *work)
1264 {
1265 	struct virtnet_info *vi =
1266 		container_of(work, struct virtnet_info, refill.work);
1267 	bool still_empty;
1268 	int i;
1269 
1270 	for (i = 0; i < vi->curr_queue_pairs; i++) {
1271 		struct receive_queue *rq = &vi->rq[i];
1272 
1273 		napi_disable(&rq->napi);
1274 		still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1275 		virtnet_napi_enable(rq->vq, &rq->napi);
1276 
1277 		/* In theory, this can happen: if we don't get any buffers in
1278 		 * we will *never* try to fill again.
1279 		 */
1280 		if (still_empty)
1281 			schedule_delayed_work(&vi->refill, HZ/2);
1282 	}
1283 }
1284 
1285 static int virtnet_receive(struct receive_queue *rq, int budget,
1286 			   unsigned int *xdp_xmit)
1287 {
1288 	struct virtnet_info *vi = rq->vq->vdev->priv;
1289 	struct virtnet_rq_stats stats = {};
1290 	unsigned int len;
1291 	void *buf;
1292 	int i;
1293 
1294 	if (!vi->big_packets || vi->mergeable_rx_bufs) {
1295 		void *ctx;
1296 
1297 		while (stats.packets < budget &&
1298 		       (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1299 			receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1300 			stats.packets++;
1301 		}
1302 	} else {
1303 		while (stats.packets < budget &&
1304 		       (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1305 			receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1306 			stats.packets++;
1307 		}
1308 	}
1309 
1310 	if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1311 		if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1312 			schedule_delayed_work(&vi->refill, 0);
1313 	}
1314 
1315 	u64_stats_update_begin(&rq->stats.syncp);
1316 	for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1317 		size_t offset = virtnet_rq_stats_desc[i].offset;
1318 		u64 *item;
1319 
1320 		item = (u64 *)((u8 *)&rq->stats + offset);
1321 		*item += *(u64 *)((u8 *)&stats + offset);
1322 	}
1323 	u64_stats_update_end(&rq->stats.syncp);
1324 
1325 	return stats.packets;
1326 }
1327 
1328 static void free_old_xmit_skbs(struct send_queue *sq)
1329 {
1330 	struct sk_buff *skb;
1331 	unsigned int len;
1332 	unsigned int packets = 0;
1333 	unsigned int bytes = 0;
1334 
1335 	while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1336 		pr_debug("Sent skb %p\n", skb);
1337 
1338 		bytes += skb->len;
1339 		packets++;
1340 
1341 		dev_consume_skb_any(skb);
1342 	}
1343 
1344 	/* Avoid overhead when no packets have been processed
1345 	 * happens when called speculatively from start_xmit.
1346 	 */
1347 	if (!packets)
1348 		return;
1349 
1350 	u64_stats_update_begin(&sq->stats.syncp);
1351 	sq->stats.bytes += bytes;
1352 	sq->stats.packets += packets;
1353 	u64_stats_update_end(&sq->stats.syncp);
1354 }
1355 
1356 static void virtnet_poll_cleantx(struct receive_queue *rq)
1357 {
1358 	struct virtnet_info *vi = rq->vq->vdev->priv;
1359 	unsigned int index = vq2rxq(rq->vq);
1360 	struct send_queue *sq = &vi->sq[index];
1361 	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1362 
1363 	if (!sq->napi.weight)
1364 		return;
1365 
1366 	if (__netif_tx_trylock(txq)) {
1367 		free_old_xmit_skbs(sq);
1368 		__netif_tx_unlock(txq);
1369 	}
1370 
1371 	if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1372 		netif_tx_wake_queue(txq);
1373 }
1374 
1375 static int virtnet_poll(struct napi_struct *napi, int budget)
1376 {
1377 	struct receive_queue *rq =
1378 		container_of(napi, struct receive_queue, napi);
1379 	struct virtnet_info *vi = rq->vq->vdev->priv;
1380 	struct send_queue *sq;
1381 	unsigned int received;
1382 	unsigned int xdp_xmit = 0;
1383 
1384 	virtnet_poll_cleantx(rq);
1385 
1386 	received = virtnet_receive(rq, budget, &xdp_xmit);
1387 
1388 	/* Out of packets? */
1389 	if (received < budget)
1390 		virtqueue_napi_complete(napi, rq->vq, received);
1391 
1392 	if (xdp_xmit & VIRTIO_XDP_REDIR)
1393 		xdp_do_flush_map();
1394 
1395 	if (xdp_xmit & VIRTIO_XDP_TX) {
1396 		sq = virtnet_xdp_sq(vi);
1397 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1398 			u64_stats_update_begin(&sq->stats.syncp);
1399 			sq->stats.kicks++;
1400 			u64_stats_update_end(&sq->stats.syncp);
1401 		}
1402 	}
1403 
1404 	return received;
1405 }
1406 
1407 static int virtnet_open(struct net_device *dev)
1408 {
1409 	struct virtnet_info *vi = netdev_priv(dev);
1410 	int i, err;
1411 
1412 	for (i = 0; i < vi->max_queue_pairs; i++) {
1413 		if (i < vi->curr_queue_pairs)
1414 			/* Make sure we have some buffers: if oom use wq. */
1415 			if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1416 				schedule_delayed_work(&vi->refill, 0);
1417 
1418 		err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i);
1419 		if (err < 0)
1420 			return err;
1421 
1422 		err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1423 						 MEM_TYPE_PAGE_SHARED, NULL);
1424 		if (err < 0) {
1425 			xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1426 			return err;
1427 		}
1428 
1429 		virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1430 		virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1431 	}
1432 
1433 	return 0;
1434 }
1435 
1436 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1437 {
1438 	struct send_queue *sq = container_of(napi, struct send_queue, napi);
1439 	struct virtnet_info *vi = sq->vq->vdev->priv;
1440 	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
1441 
1442 	__netif_tx_lock(txq, raw_smp_processor_id());
1443 	free_old_xmit_skbs(sq);
1444 	__netif_tx_unlock(txq);
1445 
1446 	virtqueue_napi_complete(napi, sq->vq, 0);
1447 
1448 	if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1449 		netif_tx_wake_queue(txq);
1450 
1451 	return 0;
1452 }
1453 
1454 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1455 {
1456 	struct virtio_net_hdr_mrg_rxbuf *hdr;
1457 	const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1458 	struct virtnet_info *vi = sq->vq->vdev->priv;
1459 	int num_sg;
1460 	unsigned hdr_len = vi->hdr_len;
1461 	bool can_push;
1462 
1463 	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1464 
1465 	can_push = vi->any_header_sg &&
1466 		!((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1467 		!skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1468 	/* Even if we can, don't push here yet as this would skew
1469 	 * csum_start offset below. */
1470 	if (can_push)
1471 		hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1472 	else
1473 		hdr = skb_vnet_hdr(skb);
1474 
1475 	if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1476 				    virtio_is_little_endian(vi->vdev), false,
1477 				    0))
1478 		BUG();
1479 
1480 	if (vi->mergeable_rx_bufs)
1481 		hdr->num_buffers = 0;
1482 
1483 	sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1484 	if (can_push) {
1485 		__skb_push(skb, hdr_len);
1486 		num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1487 		if (unlikely(num_sg < 0))
1488 			return num_sg;
1489 		/* Pull header back to avoid skew in tx bytes calculations. */
1490 		__skb_pull(skb, hdr_len);
1491 	} else {
1492 		sg_set_buf(sq->sg, hdr, hdr_len);
1493 		num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1494 		if (unlikely(num_sg < 0))
1495 			return num_sg;
1496 		num_sg++;
1497 	}
1498 	return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1499 }
1500 
1501 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1502 {
1503 	struct virtnet_info *vi = netdev_priv(dev);
1504 	int qnum = skb_get_queue_mapping(skb);
1505 	struct send_queue *sq = &vi->sq[qnum];
1506 	int err;
1507 	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1508 	bool kick = !skb->xmit_more;
1509 	bool use_napi = sq->napi.weight;
1510 
1511 	/* Free up any pending old buffers before queueing new ones. */
1512 	free_old_xmit_skbs(sq);
1513 
1514 	if (use_napi && kick)
1515 		virtqueue_enable_cb_delayed(sq->vq);
1516 
1517 	/* timestamp packet in software */
1518 	skb_tx_timestamp(skb);
1519 
1520 	/* Try to transmit */
1521 	err = xmit_skb(sq, skb);
1522 
1523 	/* This should not happen! */
1524 	if (unlikely(err)) {
1525 		dev->stats.tx_fifo_errors++;
1526 		if (net_ratelimit())
1527 			dev_warn(&dev->dev,
1528 				 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1529 		dev->stats.tx_dropped++;
1530 		dev_kfree_skb_any(skb);
1531 		return NETDEV_TX_OK;
1532 	}
1533 
1534 	/* Don't wait up for transmitted skbs to be freed. */
1535 	if (!use_napi) {
1536 		skb_orphan(skb);
1537 		nf_reset(skb);
1538 	}
1539 
1540 	/* If running out of space, stop queue to avoid getting packets that we
1541 	 * are then unable to transmit.
1542 	 * An alternative would be to force queuing layer to requeue the skb by
1543 	 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1544 	 * returned in a normal path of operation: it means that driver is not
1545 	 * maintaining the TX queue stop/start state properly, and causes
1546 	 * the stack to do a non-trivial amount of useless work.
1547 	 * Since most packets only take 1 or 2 ring slots, stopping the queue
1548 	 * early means 16 slots are typically wasted.
1549 	 */
1550 	if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1551 		netif_stop_subqueue(dev, qnum);
1552 		if (!use_napi &&
1553 		    unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1554 			/* More just got used, free them then recheck. */
1555 			free_old_xmit_skbs(sq);
1556 			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1557 				netif_start_subqueue(dev, qnum);
1558 				virtqueue_disable_cb(sq->vq);
1559 			}
1560 		}
1561 	}
1562 
1563 	if (kick || netif_xmit_stopped(txq)) {
1564 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1565 			u64_stats_update_begin(&sq->stats.syncp);
1566 			sq->stats.kicks++;
1567 			u64_stats_update_end(&sq->stats.syncp);
1568 		}
1569 	}
1570 
1571 	return NETDEV_TX_OK;
1572 }
1573 
1574 /*
1575  * Send command via the control virtqueue and check status.  Commands
1576  * supported by the hypervisor, as indicated by feature bits, should
1577  * never fail unless improperly formatted.
1578  */
1579 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1580 				 struct scatterlist *out)
1581 {
1582 	struct scatterlist *sgs[4], hdr, stat;
1583 	unsigned out_num = 0, tmp;
1584 
1585 	/* Caller should know better */
1586 	BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1587 
1588 	vi->ctrl->status = ~0;
1589 	vi->ctrl->hdr.class = class;
1590 	vi->ctrl->hdr.cmd = cmd;
1591 	/* Add header */
1592 	sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1593 	sgs[out_num++] = &hdr;
1594 
1595 	if (out)
1596 		sgs[out_num++] = out;
1597 
1598 	/* Add return status. */
1599 	sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1600 	sgs[out_num] = &stat;
1601 
1602 	BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1603 	virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1604 
1605 	if (unlikely(!virtqueue_kick(vi->cvq)))
1606 		return vi->ctrl->status == VIRTIO_NET_OK;
1607 
1608 	/* Spin for a response, the kick causes an ioport write, trapping
1609 	 * into the hypervisor, so the request should be handled immediately.
1610 	 */
1611 	while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1612 	       !virtqueue_is_broken(vi->cvq))
1613 		cpu_relax();
1614 
1615 	return vi->ctrl->status == VIRTIO_NET_OK;
1616 }
1617 
1618 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1619 {
1620 	struct virtnet_info *vi = netdev_priv(dev);
1621 	struct virtio_device *vdev = vi->vdev;
1622 	int ret;
1623 	struct sockaddr *addr;
1624 	struct scatterlist sg;
1625 
1626 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1627 		return -EOPNOTSUPP;
1628 
1629 	addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1630 	if (!addr)
1631 		return -ENOMEM;
1632 
1633 	ret = eth_prepare_mac_addr_change(dev, addr);
1634 	if (ret)
1635 		goto out;
1636 
1637 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1638 		sg_init_one(&sg, addr->sa_data, dev->addr_len);
1639 		if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1640 					  VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1641 			dev_warn(&vdev->dev,
1642 				 "Failed to set mac address by vq command.\n");
1643 			ret = -EINVAL;
1644 			goto out;
1645 		}
1646 	} else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1647 		   !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1648 		unsigned int i;
1649 
1650 		/* Naturally, this has an atomicity problem. */
1651 		for (i = 0; i < dev->addr_len; i++)
1652 			virtio_cwrite8(vdev,
1653 				       offsetof(struct virtio_net_config, mac) +
1654 				       i, addr->sa_data[i]);
1655 	}
1656 
1657 	eth_commit_mac_addr_change(dev, p);
1658 	ret = 0;
1659 
1660 out:
1661 	kfree(addr);
1662 	return ret;
1663 }
1664 
1665 static void virtnet_stats(struct net_device *dev,
1666 			  struct rtnl_link_stats64 *tot)
1667 {
1668 	struct virtnet_info *vi = netdev_priv(dev);
1669 	unsigned int start;
1670 	int i;
1671 
1672 	for (i = 0; i < vi->max_queue_pairs; i++) {
1673 		u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1674 		struct receive_queue *rq = &vi->rq[i];
1675 		struct send_queue *sq = &vi->sq[i];
1676 
1677 		do {
1678 			start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1679 			tpackets = sq->stats.packets;
1680 			tbytes   = sq->stats.bytes;
1681 		} while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1682 
1683 		do {
1684 			start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1685 			rpackets = rq->stats.packets;
1686 			rbytes   = rq->stats.bytes;
1687 			rdrops   = rq->stats.drops;
1688 		} while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1689 
1690 		tot->rx_packets += rpackets;
1691 		tot->tx_packets += tpackets;
1692 		tot->rx_bytes   += rbytes;
1693 		tot->tx_bytes   += tbytes;
1694 		tot->rx_dropped += rdrops;
1695 	}
1696 
1697 	tot->tx_dropped = dev->stats.tx_dropped;
1698 	tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1699 	tot->rx_length_errors = dev->stats.rx_length_errors;
1700 	tot->rx_frame_errors = dev->stats.rx_frame_errors;
1701 }
1702 
1703 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1704 {
1705 	rtnl_lock();
1706 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1707 				  VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1708 		dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1709 	rtnl_unlock();
1710 }
1711 
1712 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1713 {
1714 	struct scatterlist sg;
1715 	struct net_device *dev = vi->dev;
1716 
1717 	if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1718 		return 0;
1719 
1720 	vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1721 	sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1722 
1723 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1724 				  VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1725 		dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1726 			 queue_pairs);
1727 		return -EINVAL;
1728 	} else {
1729 		vi->curr_queue_pairs = queue_pairs;
1730 		/* virtnet_open() will refill when device is going to up. */
1731 		if (dev->flags & IFF_UP)
1732 			schedule_delayed_work(&vi->refill, 0);
1733 	}
1734 
1735 	return 0;
1736 }
1737 
1738 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1739 {
1740 	int err;
1741 
1742 	rtnl_lock();
1743 	err = _virtnet_set_queues(vi, queue_pairs);
1744 	rtnl_unlock();
1745 	return err;
1746 }
1747 
1748 static int virtnet_close(struct net_device *dev)
1749 {
1750 	struct virtnet_info *vi = netdev_priv(dev);
1751 	int i;
1752 
1753 	/* Make sure refill_work doesn't re-enable napi! */
1754 	cancel_delayed_work_sync(&vi->refill);
1755 
1756 	for (i = 0; i < vi->max_queue_pairs; i++) {
1757 		xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1758 		napi_disable(&vi->rq[i].napi);
1759 		virtnet_napi_tx_disable(&vi->sq[i].napi);
1760 	}
1761 
1762 	return 0;
1763 }
1764 
1765 static void virtnet_set_rx_mode(struct net_device *dev)
1766 {
1767 	struct virtnet_info *vi = netdev_priv(dev);
1768 	struct scatterlist sg[2];
1769 	struct virtio_net_ctrl_mac *mac_data;
1770 	struct netdev_hw_addr *ha;
1771 	int uc_count;
1772 	int mc_count;
1773 	void *buf;
1774 	int i;
1775 
1776 	/* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1777 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1778 		return;
1779 
1780 	vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1781 	vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1782 
1783 	sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1784 
1785 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1786 				  VIRTIO_NET_CTRL_RX_PROMISC, sg))
1787 		dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1788 			 vi->ctrl->promisc ? "en" : "dis");
1789 
1790 	sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1791 
1792 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1793 				  VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1794 		dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1795 			 vi->ctrl->allmulti ? "en" : "dis");
1796 
1797 	uc_count = netdev_uc_count(dev);
1798 	mc_count = netdev_mc_count(dev);
1799 	/* MAC filter - use one buffer for both lists */
1800 	buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1801 		      (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1802 	mac_data = buf;
1803 	if (!buf)
1804 		return;
1805 
1806 	sg_init_table(sg, 2);
1807 
1808 	/* Store the unicast list and count in the front of the buffer */
1809 	mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1810 	i = 0;
1811 	netdev_for_each_uc_addr(ha, dev)
1812 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1813 
1814 	sg_set_buf(&sg[0], mac_data,
1815 		   sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1816 
1817 	/* multicast list and count fill the end */
1818 	mac_data = (void *)&mac_data->macs[uc_count][0];
1819 
1820 	mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1821 	i = 0;
1822 	netdev_for_each_mc_addr(ha, dev)
1823 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1824 
1825 	sg_set_buf(&sg[1], mac_data,
1826 		   sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1827 
1828 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1829 				  VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1830 		dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1831 
1832 	kfree(buf);
1833 }
1834 
1835 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1836 				   __be16 proto, u16 vid)
1837 {
1838 	struct virtnet_info *vi = netdev_priv(dev);
1839 	struct scatterlist sg;
1840 
1841 	vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1842 	sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1843 
1844 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1845 				  VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1846 		dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1847 	return 0;
1848 }
1849 
1850 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1851 				    __be16 proto, u16 vid)
1852 {
1853 	struct virtnet_info *vi = netdev_priv(dev);
1854 	struct scatterlist sg;
1855 
1856 	vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1857 	sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1858 
1859 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1860 				  VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1861 		dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1862 	return 0;
1863 }
1864 
1865 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1866 {
1867 	int i;
1868 
1869 	if (vi->affinity_hint_set) {
1870 		for (i = 0; i < vi->max_queue_pairs; i++) {
1871 			virtqueue_set_affinity(vi->rq[i].vq, NULL);
1872 			virtqueue_set_affinity(vi->sq[i].vq, NULL);
1873 		}
1874 
1875 		vi->affinity_hint_set = false;
1876 	}
1877 }
1878 
1879 static void virtnet_set_affinity(struct virtnet_info *vi)
1880 {
1881 	cpumask_var_t mask;
1882 	int stragglers;
1883 	int group_size;
1884 	int i, j, cpu;
1885 	int num_cpu;
1886 	int stride;
1887 
1888 	if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
1889 		virtnet_clean_affinity(vi, -1);
1890 		return;
1891 	}
1892 
1893 	num_cpu = num_online_cpus();
1894 	stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
1895 	stragglers = num_cpu >= vi->curr_queue_pairs ?
1896 			num_cpu % vi->curr_queue_pairs :
1897 			0;
1898 	cpu = cpumask_next(-1, cpu_online_mask);
1899 
1900 	for (i = 0; i < vi->curr_queue_pairs; i++) {
1901 		group_size = stride + (i < stragglers ? 1 : 0);
1902 
1903 		for (j = 0; j < group_size; j++) {
1904 			cpumask_set_cpu(cpu, mask);
1905 			cpu = cpumask_next_wrap(cpu, cpu_online_mask,
1906 						nr_cpu_ids, false);
1907 		}
1908 		virtqueue_set_affinity(vi->rq[i].vq, mask);
1909 		virtqueue_set_affinity(vi->sq[i].vq, mask);
1910 		__netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, false);
1911 		cpumask_clear(mask);
1912 	}
1913 
1914 	vi->affinity_hint_set = true;
1915 	free_cpumask_var(mask);
1916 }
1917 
1918 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1919 {
1920 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1921 						   node);
1922 	virtnet_set_affinity(vi);
1923 	return 0;
1924 }
1925 
1926 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1927 {
1928 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1929 						   node_dead);
1930 	virtnet_set_affinity(vi);
1931 	return 0;
1932 }
1933 
1934 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1935 {
1936 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1937 						   node);
1938 
1939 	virtnet_clean_affinity(vi, cpu);
1940 	return 0;
1941 }
1942 
1943 static enum cpuhp_state virtionet_online;
1944 
1945 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1946 {
1947 	int ret;
1948 
1949 	ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1950 	if (ret)
1951 		return ret;
1952 	ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1953 					       &vi->node_dead);
1954 	if (!ret)
1955 		return ret;
1956 	cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1957 	return ret;
1958 }
1959 
1960 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1961 {
1962 	cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1963 	cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1964 					    &vi->node_dead);
1965 }
1966 
1967 static void virtnet_get_ringparam(struct net_device *dev,
1968 				struct ethtool_ringparam *ring)
1969 {
1970 	struct virtnet_info *vi = netdev_priv(dev);
1971 
1972 	ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1973 	ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1974 	ring->rx_pending = ring->rx_max_pending;
1975 	ring->tx_pending = ring->tx_max_pending;
1976 }
1977 
1978 
1979 static void virtnet_get_drvinfo(struct net_device *dev,
1980 				struct ethtool_drvinfo *info)
1981 {
1982 	struct virtnet_info *vi = netdev_priv(dev);
1983 	struct virtio_device *vdev = vi->vdev;
1984 
1985 	strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1986 	strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1987 	strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1988 
1989 }
1990 
1991 /* TODO: Eliminate OOO packets during switching */
1992 static int virtnet_set_channels(struct net_device *dev,
1993 				struct ethtool_channels *channels)
1994 {
1995 	struct virtnet_info *vi = netdev_priv(dev);
1996 	u16 queue_pairs = channels->combined_count;
1997 	int err;
1998 
1999 	/* We don't support separate rx/tx channels.
2000 	 * We don't allow setting 'other' channels.
2001 	 */
2002 	if (channels->rx_count || channels->tx_count || channels->other_count)
2003 		return -EINVAL;
2004 
2005 	if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2006 		return -EINVAL;
2007 
2008 	/* For now we don't support modifying channels while XDP is loaded
2009 	 * also when XDP is loaded all RX queues have XDP programs so we only
2010 	 * need to check a single RX queue.
2011 	 */
2012 	if (vi->rq[0].xdp_prog)
2013 		return -EINVAL;
2014 
2015 	get_online_cpus();
2016 	err = _virtnet_set_queues(vi, queue_pairs);
2017 	if (!err) {
2018 		netif_set_real_num_tx_queues(dev, queue_pairs);
2019 		netif_set_real_num_rx_queues(dev, queue_pairs);
2020 
2021 		virtnet_set_affinity(vi);
2022 	}
2023 	put_online_cpus();
2024 
2025 	return err;
2026 }
2027 
2028 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2029 {
2030 	struct virtnet_info *vi = netdev_priv(dev);
2031 	char *p = (char *)data;
2032 	unsigned int i, j;
2033 
2034 	switch (stringset) {
2035 	case ETH_SS_STATS:
2036 		for (i = 0; i < vi->curr_queue_pairs; i++) {
2037 			for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2038 				snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s",
2039 					 i, virtnet_rq_stats_desc[j].desc);
2040 				p += ETH_GSTRING_LEN;
2041 			}
2042 		}
2043 
2044 		for (i = 0; i < vi->curr_queue_pairs; i++) {
2045 			for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2046 				snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s",
2047 					 i, virtnet_sq_stats_desc[j].desc);
2048 				p += ETH_GSTRING_LEN;
2049 			}
2050 		}
2051 		break;
2052 	}
2053 }
2054 
2055 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2056 {
2057 	struct virtnet_info *vi = netdev_priv(dev);
2058 
2059 	switch (sset) {
2060 	case ETH_SS_STATS:
2061 		return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2062 					       VIRTNET_SQ_STATS_LEN);
2063 	default:
2064 		return -EOPNOTSUPP;
2065 	}
2066 }
2067 
2068 static void virtnet_get_ethtool_stats(struct net_device *dev,
2069 				      struct ethtool_stats *stats, u64 *data)
2070 {
2071 	struct virtnet_info *vi = netdev_priv(dev);
2072 	unsigned int idx = 0, start, i, j;
2073 	const u8 *stats_base;
2074 	size_t offset;
2075 
2076 	for (i = 0; i < vi->curr_queue_pairs; i++) {
2077 		struct receive_queue *rq = &vi->rq[i];
2078 
2079 		stats_base = (u8 *)&rq->stats;
2080 		do {
2081 			start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2082 			for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2083 				offset = virtnet_rq_stats_desc[j].offset;
2084 				data[idx + j] = *(u64 *)(stats_base + offset);
2085 			}
2086 		} while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2087 		idx += VIRTNET_RQ_STATS_LEN;
2088 	}
2089 
2090 	for (i = 0; i < vi->curr_queue_pairs; i++) {
2091 		struct send_queue *sq = &vi->sq[i];
2092 
2093 		stats_base = (u8 *)&sq->stats;
2094 		do {
2095 			start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2096 			for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2097 				offset = virtnet_sq_stats_desc[j].offset;
2098 				data[idx + j] = *(u64 *)(stats_base + offset);
2099 			}
2100 		} while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2101 		idx += VIRTNET_SQ_STATS_LEN;
2102 	}
2103 }
2104 
2105 static void virtnet_get_channels(struct net_device *dev,
2106 				 struct ethtool_channels *channels)
2107 {
2108 	struct virtnet_info *vi = netdev_priv(dev);
2109 
2110 	channels->combined_count = vi->curr_queue_pairs;
2111 	channels->max_combined = vi->max_queue_pairs;
2112 	channels->max_other = 0;
2113 	channels->rx_count = 0;
2114 	channels->tx_count = 0;
2115 	channels->other_count = 0;
2116 }
2117 
2118 /* Check if the user is trying to change anything besides speed/duplex */
2119 static bool
2120 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
2121 {
2122 	struct ethtool_link_ksettings diff1 = *cmd;
2123 	struct ethtool_link_ksettings diff2 = {};
2124 
2125 	/* cmd is always set so we need to clear it, validate the port type
2126 	 * and also without autonegotiation we can ignore advertising
2127 	 */
2128 	diff1.base.speed = 0;
2129 	diff2.base.port = PORT_OTHER;
2130 	ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
2131 	diff1.base.duplex = 0;
2132 	diff1.base.cmd = 0;
2133 	diff1.base.link_mode_masks_nwords = 0;
2134 
2135 	return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
2136 		bitmap_empty(diff1.link_modes.supported,
2137 			     __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2138 		bitmap_empty(diff1.link_modes.advertising,
2139 			     __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2140 		bitmap_empty(diff1.link_modes.lp_advertising,
2141 			     __ETHTOOL_LINK_MODE_MASK_NBITS);
2142 }
2143 
2144 static int virtnet_set_link_ksettings(struct net_device *dev,
2145 				      const struct ethtool_link_ksettings *cmd)
2146 {
2147 	struct virtnet_info *vi = netdev_priv(dev);
2148 	u32 speed;
2149 
2150 	speed = cmd->base.speed;
2151 	/* don't allow custom speed and duplex */
2152 	if (!ethtool_validate_speed(speed) ||
2153 	    !ethtool_validate_duplex(cmd->base.duplex) ||
2154 	    !virtnet_validate_ethtool_cmd(cmd))
2155 		return -EINVAL;
2156 	vi->speed = speed;
2157 	vi->duplex = cmd->base.duplex;
2158 
2159 	return 0;
2160 }
2161 
2162 static int virtnet_get_link_ksettings(struct net_device *dev,
2163 				      struct ethtool_link_ksettings *cmd)
2164 {
2165 	struct virtnet_info *vi = netdev_priv(dev);
2166 
2167 	cmd->base.speed = vi->speed;
2168 	cmd->base.duplex = vi->duplex;
2169 	cmd->base.port = PORT_OTHER;
2170 
2171 	return 0;
2172 }
2173 
2174 static int virtnet_set_coalesce(struct net_device *dev,
2175 				struct ethtool_coalesce *ec)
2176 {
2177 	struct ethtool_coalesce ec_default = {
2178 		.cmd = ETHTOOL_SCOALESCE,
2179 		.rx_max_coalesced_frames = 1,
2180 	};
2181 	struct virtnet_info *vi = netdev_priv(dev);
2182 	int i, napi_weight;
2183 
2184 	if (ec->tx_max_coalesced_frames > 1)
2185 		return -EINVAL;
2186 
2187 	ec_default.tx_max_coalesced_frames = ec->tx_max_coalesced_frames;
2188 	napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2189 
2190 	/* disallow changes to fields not explicitly tested above */
2191 	if (memcmp(ec, &ec_default, sizeof(ec_default)))
2192 		return -EINVAL;
2193 
2194 	if (napi_weight ^ vi->sq[0].napi.weight) {
2195 		if (dev->flags & IFF_UP)
2196 			return -EBUSY;
2197 		for (i = 0; i < vi->max_queue_pairs; i++)
2198 			vi->sq[i].napi.weight = napi_weight;
2199 	}
2200 
2201 	return 0;
2202 }
2203 
2204 static int virtnet_get_coalesce(struct net_device *dev,
2205 				struct ethtool_coalesce *ec)
2206 {
2207 	struct ethtool_coalesce ec_default = {
2208 		.cmd = ETHTOOL_GCOALESCE,
2209 		.rx_max_coalesced_frames = 1,
2210 	};
2211 	struct virtnet_info *vi = netdev_priv(dev);
2212 
2213 	memcpy(ec, &ec_default, sizeof(ec_default));
2214 
2215 	if (vi->sq[0].napi.weight)
2216 		ec->tx_max_coalesced_frames = 1;
2217 
2218 	return 0;
2219 }
2220 
2221 static void virtnet_init_settings(struct net_device *dev)
2222 {
2223 	struct virtnet_info *vi = netdev_priv(dev);
2224 
2225 	vi->speed = SPEED_UNKNOWN;
2226 	vi->duplex = DUPLEX_UNKNOWN;
2227 }
2228 
2229 static void virtnet_update_settings(struct virtnet_info *vi)
2230 {
2231 	u32 speed;
2232 	u8 duplex;
2233 
2234 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2235 		return;
2236 
2237 	speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config,
2238 						  speed));
2239 	if (ethtool_validate_speed(speed))
2240 		vi->speed = speed;
2241 	duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config,
2242 						  duplex));
2243 	if (ethtool_validate_duplex(duplex))
2244 		vi->duplex = duplex;
2245 }
2246 
2247 static const struct ethtool_ops virtnet_ethtool_ops = {
2248 	.get_drvinfo = virtnet_get_drvinfo,
2249 	.get_link = ethtool_op_get_link,
2250 	.get_ringparam = virtnet_get_ringparam,
2251 	.get_strings = virtnet_get_strings,
2252 	.get_sset_count = virtnet_get_sset_count,
2253 	.get_ethtool_stats = virtnet_get_ethtool_stats,
2254 	.set_channels = virtnet_set_channels,
2255 	.get_channels = virtnet_get_channels,
2256 	.get_ts_info = ethtool_op_get_ts_info,
2257 	.get_link_ksettings = virtnet_get_link_ksettings,
2258 	.set_link_ksettings = virtnet_set_link_ksettings,
2259 	.set_coalesce = virtnet_set_coalesce,
2260 	.get_coalesce = virtnet_get_coalesce,
2261 };
2262 
2263 static void virtnet_freeze_down(struct virtio_device *vdev)
2264 {
2265 	struct virtnet_info *vi = vdev->priv;
2266 	int i;
2267 
2268 	/* Make sure no work handler is accessing the device */
2269 	flush_work(&vi->config_work);
2270 
2271 	netif_tx_lock_bh(vi->dev);
2272 	netif_device_detach(vi->dev);
2273 	netif_tx_unlock_bh(vi->dev);
2274 	cancel_delayed_work_sync(&vi->refill);
2275 
2276 	if (netif_running(vi->dev)) {
2277 		for (i = 0; i < vi->max_queue_pairs; i++) {
2278 			napi_disable(&vi->rq[i].napi);
2279 			virtnet_napi_tx_disable(&vi->sq[i].napi);
2280 		}
2281 	}
2282 }
2283 
2284 static int init_vqs(struct virtnet_info *vi);
2285 
2286 static int virtnet_restore_up(struct virtio_device *vdev)
2287 {
2288 	struct virtnet_info *vi = vdev->priv;
2289 	int err, i;
2290 
2291 	err = init_vqs(vi);
2292 	if (err)
2293 		return err;
2294 
2295 	virtio_device_ready(vdev);
2296 
2297 	if (netif_running(vi->dev)) {
2298 		for (i = 0; i < vi->curr_queue_pairs; i++)
2299 			if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2300 				schedule_delayed_work(&vi->refill, 0);
2301 
2302 		for (i = 0; i < vi->max_queue_pairs; i++) {
2303 			virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2304 			virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2305 					       &vi->sq[i].napi);
2306 		}
2307 	}
2308 
2309 	netif_tx_lock_bh(vi->dev);
2310 	netif_device_attach(vi->dev);
2311 	netif_tx_unlock_bh(vi->dev);
2312 	return err;
2313 }
2314 
2315 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2316 {
2317 	struct scatterlist sg;
2318 	vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2319 
2320 	sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2321 
2322 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2323 				  VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2324 		dev_warn(&vi->dev->dev, "Fail to set guest offload. \n");
2325 		return -EINVAL;
2326 	}
2327 
2328 	return 0;
2329 }
2330 
2331 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2332 {
2333 	u64 offloads = 0;
2334 
2335 	if (!vi->guest_offloads)
2336 		return 0;
2337 
2338 	return virtnet_set_guest_offloads(vi, offloads);
2339 }
2340 
2341 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2342 {
2343 	u64 offloads = vi->guest_offloads;
2344 
2345 	if (!vi->guest_offloads)
2346 		return 0;
2347 
2348 	return virtnet_set_guest_offloads(vi, offloads);
2349 }
2350 
2351 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2352 			   struct netlink_ext_ack *extack)
2353 {
2354 	unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2355 	struct virtnet_info *vi = netdev_priv(dev);
2356 	struct bpf_prog *old_prog;
2357 	u16 xdp_qp = 0, curr_qp;
2358 	int i, err;
2359 
2360 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2361 	    && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2362 	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2363 	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2364 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2365 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2366 		NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
2367 		return -EOPNOTSUPP;
2368 	}
2369 
2370 	if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2371 		NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2372 		return -EINVAL;
2373 	}
2374 
2375 	if (dev->mtu > max_sz) {
2376 		NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2377 		netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2378 		return -EINVAL;
2379 	}
2380 
2381 	curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2382 	if (prog)
2383 		xdp_qp = nr_cpu_ids;
2384 
2385 	/* XDP requires extra queues for XDP_TX */
2386 	if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2387 		NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2388 		netdev_warn(dev, "request %i queues but max is %i\n",
2389 			    curr_qp + xdp_qp, vi->max_queue_pairs);
2390 		return -ENOMEM;
2391 	}
2392 
2393 	if (prog) {
2394 		prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
2395 		if (IS_ERR(prog))
2396 			return PTR_ERR(prog);
2397 	}
2398 
2399 	/* Make sure NAPI is not using any XDP TX queues for RX. */
2400 	if (netif_running(dev))
2401 		for (i = 0; i < vi->max_queue_pairs; i++)
2402 			napi_disable(&vi->rq[i].napi);
2403 
2404 	netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2405 	err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2406 	if (err)
2407 		goto err;
2408 	vi->xdp_queue_pairs = xdp_qp;
2409 
2410 	for (i = 0; i < vi->max_queue_pairs; i++) {
2411 		old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2412 		rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2413 		if (i == 0) {
2414 			if (!old_prog)
2415 				virtnet_clear_guest_offloads(vi);
2416 			if (!prog)
2417 				virtnet_restore_guest_offloads(vi);
2418 		}
2419 		if (old_prog)
2420 			bpf_prog_put(old_prog);
2421 		if (netif_running(dev))
2422 			virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2423 	}
2424 
2425 	return 0;
2426 
2427 err:
2428 	for (i = 0; i < vi->max_queue_pairs; i++)
2429 		virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2430 	if (prog)
2431 		bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2432 	return err;
2433 }
2434 
2435 static u32 virtnet_xdp_query(struct net_device *dev)
2436 {
2437 	struct virtnet_info *vi = netdev_priv(dev);
2438 	const struct bpf_prog *xdp_prog;
2439 	int i;
2440 
2441 	for (i = 0; i < vi->max_queue_pairs; i++) {
2442 		xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2443 		if (xdp_prog)
2444 			return xdp_prog->aux->id;
2445 	}
2446 	return 0;
2447 }
2448 
2449 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2450 {
2451 	switch (xdp->command) {
2452 	case XDP_SETUP_PROG:
2453 		return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2454 	case XDP_QUERY_PROG:
2455 		xdp->prog_id = virtnet_xdp_query(dev);
2456 		return 0;
2457 	default:
2458 		return -EINVAL;
2459 	}
2460 }
2461 
2462 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2463 				      size_t len)
2464 {
2465 	struct virtnet_info *vi = netdev_priv(dev);
2466 	int ret;
2467 
2468 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2469 		return -EOPNOTSUPP;
2470 
2471 	ret = snprintf(buf, len, "sby");
2472 	if (ret >= len)
2473 		return -EOPNOTSUPP;
2474 
2475 	return 0;
2476 }
2477 
2478 static const struct net_device_ops virtnet_netdev = {
2479 	.ndo_open            = virtnet_open,
2480 	.ndo_stop   	     = virtnet_close,
2481 	.ndo_start_xmit      = start_xmit,
2482 	.ndo_validate_addr   = eth_validate_addr,
2483 	.ndo_set_mac_address = virtnet_set_mac_address,
2484 	.ndo_set_rx_mode     = virtnet_set_rx_mode,
2485 	.ndo_get_stats64     = virtnet_stats,
2486 	.ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2487 	.ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2488 	.ndo_bpf		= virtnet_xdp,
2489 	.ndo_xdp_xmit		= virtnet_xdp_xmit,
2490 	.ndo_features_check	= passthru_features_check,
2491 	.ndo_get_phys_port_name	= virtnet_get_phys_port_name,
2492 };
2493 
2494 static void virtnet_config_changed_work(struct work_struct *work)
2495 {
2496 	struct virtnet_info *vi =
2497 		container_of(work, struct virtnet_info, config_work);
2498 	u16 v;
2499 
2500 	if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2501 				 struct virtio_net_config, status, &v) < 0)
2502 		return;
2503 
2504 	if (v & VIRTIO_NET_S_ANNOUNCE) {
2505 		netdev_notify_peers(vi->dev);
2506 		virtnet_ack_link_announce(vi);
2507 	}
2508 
2509 	/* Ignore unknown (future) status bits */
2510 	v &= VIRTIO_NET_S_LINK_UP;
2511 
2512 	if (vi->status == v)
2513 		return;
2514 
2515 	vi->status = v;
2516 
2517 	if (vi->status & VIRTIO_NET_S_LINK_UP) {
2518 		virtnet_update_settings(vi);
2519 		netif_carrier_on(vi->dev);
2520 		netif_tx_wake_all_queues(vi->dev);
2521 	} else {
2522 		netif_carrier_off(vi->dev);
2523 		netif_tx_stop_all_queues(vi->dev);
2524 	}
2525 }
2526 
2527 static void virtnet_config_changed(struct virtio_device *vdev)
2528 {
2529 	struct virtnet_info *vi = vdev->priv;
2530 
2531 	schedule_work(&vi->config_work);
2532 }
2533 
2534 static void virtnet_free_queues(struct virtnet_info *vi)
2535 {
2536 	int i;
2537 
2538 	for (i = 0; i < vi->max_queue_pairs; i++) {
2539 		napi_hash_del(&vi->rq[i].napi);
2540 		netif_napi_del(&vi->rq[i].napi);
2541 		netif_napi_del(&vi->sq[i].napi);
2542 	}
2543 
2544 	/* We called napi_hash_del() before netif_napi_del(),
2545 	 * we need to respect an RCU grace period before freeing vi->rq
2546 	 */
2547 	synchronize_net();
2548 
2549 	kfree(vi->rq);
2550 	kfree(vi->sq);
2551 	kfree(vi->ctrl);
2552 }
2553 
2554 static void _free_receive_bufs(struct virtnet_info *vi)
2555 {
2556 	struct bpf_prog *old_prog;
2557 	int i;
2558 
2559 	for (i = 0; i < vi->max_queue_pairs; i++) {
2560 		while (vi->rq[i].pages)
2561 			__free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2562 
2563 		old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2564 		RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2565 		if (old_prog)
2566 			bpf_prog_put(old_prog);
2567 	}
2568 }
2569 
2570 static void free_receive_bufs(struct virtnet_info *vi)
2571 {
2572 	rtnl_lock();
2573 	_free_receive_bufs(vi);
2574 	rtnl_unlock();
2575 }
2576 
2577 static void free_receive_page_frags(struct virtnet_info *vi)
2578 {
2579 	int i;
2580 	for (i = 0; i < vi->max_queue_pairs; i++)
2581 		if (vi->rq[i].alloc_frag.page)
2582 			put_page(vi->rq[i].alloc_frag.page);
2583 }
2584 
2585 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
2586 {
2587 	if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
2588 		return false;
2589 	else if (q < vi->curr_queue_pairs)
2590 		return true;
2591 	else
2592 		return false;
2593 }
2594 
2595 static void free_unused_bufs(struct virtnet_info *vi)
2596 {
2597 	void *buf;
2598 	int i;
2599 
2600 	for (i = 0; i < vi->max_queue_pairs; i++) {
2601 		struct virtqueue *vq = vi->sq[i].vq;
2602 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2603 			if (!is_xdp_raw_buffer_queue(vi, i))
2604 				dev_kfree_skb(buf);
2605 			else
2606 				put_page(virt_to_head_page(buf));
2607 		}
2608 	}
2609 
2610 	for (i = 0; i < vi->max_queue_pairs; i++) {
2611 		struct virtqueue *vq = vi->rq[i].vq;
2612 
2613 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2614 			if (vi->mergeable_rx_bufs) {
2615 				put_page(virt_to_head_page(buf));
2616 			} else if (vi->big_packets) {
2617 				give_pages(&vi->rq[i], buf);
2618 			} else {
2619 				put_page(virt_to_head_page(buf));
2620 			}
2621 		}
2622 	}
2623 }
2624 
2625 static void virtnet_del_vqs(struct virtnet_info *vi)
2626 {
2627 	struct virtio_device *vdev = vi->vdev;
2628 
2629 	virtnet_clean_affinity(vi, -1);
2630 
2631 	vdev->config->del_vqs(vdev);
2632 
2633 	virtnet_free_queues(vi);
2634 }
2635 
2636 /* How large should a single buffer be so a queue full of these can fit at
2637  * least one full packet?
2638  * Logic below assumes the mergeable buffer header is used.
2639  */
2640 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2641 {
2642 	const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2643 	unsigned int rq_size = virtqueue_get_vring_size(vq);
2644 	unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2645 	unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2646 	unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2647 
2648 	return max(max(min_buf_len, hdr_len) - hdr_len,
2649 		   (unsigned int)GOOD_PACKET_LEN);
2650 }
2651 
2652 static int virtnet_find_vqs(struct virtnet_info *vi)
2653 {
2654 	vq_callback_t **callbacks;
2655 	struct virtqueue **vqs;
2656 	int ret = -ENOMEM;
2657 	int i, total_vqs;
2658 	const char **names;
2659 	bool *ctx;
2660 
2661 	/* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2662 	 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2663 	 * possible control vq.
2664 	 */
2665 	total_vqs = vi->max_queue_pairs * 2 +
2666 		    virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2667 
2668 	/* Allocate space for find_vqs parameters */
2669 	vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2670 	if (!vqs)
2671 		goto err_vq;
2672 	callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2673 	if (!callbacks)
2674 		goto err_callback;
2675 	names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2676 	if (!names)
2677 		goto err_names;
2678 	if (!vi->big_packets || vi->mergeable_rx_bufs) {
2679 		ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2680 		if (!ctx)
2681 			goto err_ctx;
2682 	} else {
2683 		ctx = NULL;
2684 	}
2685 
2686 	/* Parameters for control virtqueue, if any */
2687 	if (vi->has_cvq) {
2688 		callbacks[total_vqs - 1] = NULL;
2689 		names[total_vqs - 1] = "control";
2690 	}
2691 
2692 	/* Allocate/initialize parameters for send/receive virtqueues */
2693 	for (i = 0; i < vi->max_queue_pairs; i++) {
2694 		callbacks[rxq2vq(i)] = skb_recv_done;
2695 		callbacks[txq2vq(i)] = skb_xmit_done;
2696 		sprintf(vi->rq[i].name, "input.%d", i);
2697 		sprintf(vi->sq[i].name, "output.%d", i);
2698 		names[rxq2vq(i)] = vi->rq[i].name;
2699 		names[txq2vq(i)] = vi->sq[i].name;
2700 		if (ctx)
2701 			ctx[rxq2vq(i)] = true;
2702 	}
2703 
2704 	ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2705 					 names, ctx, NULL);
2706 	if (ret)
2707 		goto err_find;
2708 
2709 	if (vi->has_cvq) {
2710 		vi->cvq = vqs[total_vqs - 1];
2711 		if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2712 			vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2713 	}
2714 
2715 	for (i = 0; i < vi->max_queue_pairs; i++) {
2716 		vi->rq[i].vq = vqs[rxq2vq(i)];
2717 		vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2718 		vi->sq[i].vq = vqs[txq2vq(i)];
2719 	}
2720 
2721 	/* run here: ret == 0. */
2722 
2723 
2724 err_find:
2725 	kfree(ctx);
2726 err_ctx:
2727 	kfree(names);
2728 err_names:
2729 	kfree(callbacks);
2730 err_callback:
2731 	kfree(vqs);
2732 err_vq:
2733 	return ret;
2734 }
2735 
2736 static int virtnet_alloc_queues(struct virtnet_info *vi)
2737 {
2738 	int i;
2739 
2740 	vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2741 	if (!vi->ctrl)
2742 		goto err_ctrl;
2743 	vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2744 	if (!vi->sq)
2745 		goto err_sq;
2746 	vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2747 	if (!vi->rq)
2748 		goto err_rq;
2749 
2750 	INIT_DELAYED_WORK(&vi->refill, refill_work);
2751 	for (i = 0; i < vi->max_queue_pairs; i++) {
2752 		vi->rq[i].pages = NULL;
2753 		netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2754 			       napi_weight);
2755 		netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2756 				  napi_tx ? napi_weight : 0);
2757 
2758 		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2759 		ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2760 		sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2761 
2762 		u64_stats_init(&vi->rq[i].stats.syncp);
2763 		u64_stats_init(&vi->sq[i].stats.syncp);
2764 	}
2765 
2766 	return 0;
2767 
2768 err_rq:
2769 	kfree(vi->sq);
2770 err_sq:
2771 	kfree(vi->ctrl);
2772 err_ctrl:
2773 	return -ENOMEM;
2774 }
2775 
2776 static int init_vqs(struct virtnet_info *vi)
2777 {
2778 	int ret;
2779 
2780 	/* Allocate send & receive queues */
2781 	ret = virtnet_alloc_queues(vi);
2782 	if (ret)
2783 		goto err;
2784 
2785 	ret = virtnet_find_vqs(vi);
2786 	if (ret)
2787 		goto err_free;
2788 
2789 	get_online_cpus();
2790 	virtnet_set_affinity(vi);
2791 	put_online_cpus();
2792 
2793 	return 0;
2794 
2795 err_free:
2796 	virtnet_free_queues(vi);
2797 err:
2798 	return ret;
2799 }
2800 
2801 #ifdef CONFIG_SYSFS
2802 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2803 		char *buf)
2804 {
2805 	struct virtnet_info *vi = netdev_priv(queue->dev);
2806 	unsigned int queue_index = get_netdev_rx_queue_index(queue);
2807 	unsigned int headroom = virtnet_get_headroom(vi);
2808 	unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2809 	struct ewma_pkt_len *avg;
2810 
2811 	BUG_ON(queue_index >= vi->max_queue_pairs);
2812 	avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2813 	return sprintf(buf, "%u\n",
2814 		       get_mergeable_buf_len(&vi->rq[queue_index], avg,
2815 				       SKB_DATA_ALIGN(headroom + tailroom)));
2816 }
2817 
2818 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2819 	__ATTR_RO(mergeable_rx_buffer_size);
2820 
2821 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2822 	&mergeable_rx_buffer_size_attribute.attr,
2823 	NULL
2824 };
2825 
2826 static const struct attribute_group virtio_net_mrg_rx_group = {
2827 	.name = "virtio_net",
2828 	.attrs = virtio_net_mrg_rx_attrs
2829 };
2830 #endif
2831 
2832 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2833 				    unsigned int fbit,
2834 				    const char *fname, const char *dname)
2835 {
2836 	if (!virtio_has_feature(vdev, fbit))
2837 		return false;
2838 
2839 	dev_err(&vdev->dev, "device advertises feature %s but not %s",
2840 		fname, dname);
2841 
2842 	return true;
2843 }
2844 
2845 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)			\
2846 	virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2847 
2848 static bool virtnet_validate_features(struct virtio_device *vdev)
2849 {
2850 	if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2851 	    (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2852 			     "VIRTIO_NET_F_CTRL_VQ") ||
2853 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2854 			     "VIRTIO_NET_F_CTRL_VQ") ||
2855 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2856 			     "VIRTIO_NET_F_CTRL_VQ") ||
2857 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2858 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2859 			     "VIRTIO_NET_F_CTRL_VQ"))) {
2860 		return false;
2861 	}
2862 
2863 	return true;
2864 }
2865 
2866 #define MIN_MTU ETH_MIN_MTU
2867 #define MAX_MTU ETH_MAX_MTU
2868 
2869 static int virtnet_validate(struct virtio_device *vdev)
2870 {
2871 	if (!vdev->config->get) {
2872 		dev_err(&vdev->dev, "%s failure: config access disabled\n",
2873 			__func__);
2874 		return -EINVAL;
2875 	}
2876 
2877 	if (!virtnet_validate_features(vdev))
2878 		return -EINVAL;
2879 
2880 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2881 		int mtu = virtio_cread16(vdev,
2882 					 offsetof(struct virtio_net_config,
2883 						  mtu));
2884 		if (mtu < MIN_MTU)
2885 			__virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2886 	}
2887 
2888 	return 0;
2889 }
2890 
2891 static int virtnet_probe(struct virtio_device *vdev)
2892 {
2893 	int i, err = -ENOMEM;
2894 	struct net_device *dev;
2895 	struct virtnet_info *vi;
2896 	u16 max_queue_pairs;
2897 	int mtu;
2898 
2899 	/* Find if host supports multiqueue virtio_net device */
2900 	err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2901 				   struct virtio_net_config,
2902 				   max_virtqueue_pairs, &max_queue_pairs);
2903 
2904 	/* We need at least 2 queue's */
2905 	if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2906 	    max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2907 	    !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2908 		max_queue_pairs = 1;
2909 
2910 	/* Allocate ourselves a network device with room for our info */
2911 	dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2912 	if (!dev)
2913 		return -ENOMEM;
2914 
2915 	/* Set up network device as normal. */
2916 	dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2917 	dev->netdev_ops = &virtnet_netdev;
2918 	dev->features = NETIF_F_HIGHDMA;
2919 
2920 	dev->ethtool_ops = &virtnet_ethtool_ops;
2921 	SET_NETDEV_DEV(dev, &vdev->dev);
2922 
2923 	/* Do we support "hardware" checksums? */
2924 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2925 		/* This opens up the world of extra features. */
2926 		dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2927 		if (csum)
2928 			dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2929 
2930 		if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2931 			dev->hw_features |= NETIF_F_TSO
2932 				| NETIF_F_TSO_ECN | NETIF_F_TSO6;
2933 		}
2934 		/* Individual feature bits: what can host handle? */
2935 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2936 			dev->hw_features |= NETIF_F_TSO;
2937 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2938 			dev->hw_features |= NETIF_F_TSO6;
2939 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2940 			dev->hw_features |= NETIF_F_TSO_ECN;
2941 
2942 		dev->features |= NETIF_F_GSO_ROBUST;
2943 
2944 		if (gso)
2945 			dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
2946 		/* (!csum && gso) case will be fixed by register_netdev() */
2947 	}
2948 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2949 		dev->features |= NETIF_F_RXCSUM;
2950 
2951 	dev->vlan_features = dev->features;
2952 
2953 	/* MTU range: 68 - 65535 */
2954 	dev->min_mtu = MIN_MTU;
2955 	dev->max_mtu = MAX_MTU;
2956 
2957 	/* Configuration may specify what MAC to use.  Otherwise random. */
2958 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2959 		virtio_cread_bytes(vdev,
2960 				   offsetof(struct virtio_net_config, mac),
2961 				   dev->dev_addr, dev->addr_len);
2962 	else
2963 		eth_hw_addr_random(dev);
2964 
2965 	/* Set up our device-specific information */
2966 	vi = netdev_priv(dev);
2967 	vi->dev = dev;
2968 	vi->vdev = vdev;
2969 	vdev->priv = vi;
2970 
2971 	INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2972 
2973 	/* If we can receive ANY GSO packets, we must allocate large ones. */
2974 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2975 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2976 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2977 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2978 		vi->big_packets = true;
2979 
2980 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2981 		vi->mergeable_rx_bufs = true;
2982 
2983 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2984 	    virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2985 		vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2986 	else
2987 		vi->hdr_len = sizeof(struct virtio_net_hdr);
2988 
2989 	if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2990 	    virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2991 		vi->any_header_sg = true;
2992 
2993 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2994 		vi->has_cvq = true;
2995 
2996 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2997 		mtu = virtio_cread16(vdev,
2998 				     offsetof(struct virtio_net_config,
2999 					      mtu));
3000 		if (mtu < dev->min_mtu) {
3001 			/* Should never trigger: MTU was previously validated
3002 			 * in virtnet_validate.
3003 			 */
3004 			dev_err(&vdev->dev, "device MTU appears to have changed "
3005 				"it is now %d < %d", mtu, dev->min_mtu);
3006 			goto free;
3007 		}
3008 
3009 		dev->mtu = mtu;
3010 		dev->max_mtu = mtu;
3011 
3012 		/* TODO: size buffers correctly in this case. */
3013 		if (dev->mtu > ETH_DATA_LEN)
3014 			vi->big_packets = true;
3015 	}
3016 
3017 	if (vi->any_header_sg)
3018 		dev->needed_headroom = vi->hdr_len;
3019 
3020 	/* Enable multiqueue by default */
3021 	if (num_online_cpus() >= max_queue_pairs)
3022 		vi->curr_queue_pairs = max_queue_pairs;
3023 	else
3024 		vi->curr_queue_pairs = num_online_cpus();
3025 	vi->max_queue_pairs = max_queue_pairs;
3026 
3027 	/* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3028 	err = init_vqs(vi);
3029 	if (err)
3030 		goto free;
3031 
3032 #ifdef CONFIG_SYSFS
3033 	if (vi->mergeable_rx_bufs)
3034 		dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3035 #endif
3036 	netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3037 	netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3038 
3039 	virtnet_init_settings(dev);
3040 
3041 	if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3042 		vi->failover = net_failover_create(vi->dev);
3043 		if (IS_ERR(vi->failover)) {
3044 			err = PTR_ERR(vi->failover);
3045 			goto free_vqs;
3046 		}
3047 	}
3048 
3049 	err = register_netdev(dev);
3050 	if (err) {
3051 		pr_debug("virtio_net: registering device failed\n");
3052 		goto free_failover;
3053 	}
3054 
3055 	virtio_device_ready(vdev);
3056 
3057 	err = virtnet_cpu_notif_add(vi);
3058 	if (err) {
3059 		pr_debug("virtio_net: registering cpu notifier failed\n");
3060 		goto free_unregister_netdev;
3061 	}
3062 
3063 	virtnet_set_queues(vi, vi->curr_queue_pairs);
3064 
3065 	/* Assume link up if device can't report link status,
3066 	   otherwise get link status from config. */
3067 	netif_carrier_off(dev);
3068 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3069 		schedule_work(&vi->config_work);
3070 	} else {
3071 		vi->status = VIRTIO_NET_S_LINK_UP;
3072 		virtnet_update_settings(vi);
3073 		netif_carrier_on(dev);
3074 	}
3075 
3076 	for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3077 		if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3078 			set_bit(guest_offloads[i], &vi->guest_offloads);
3079 
3080 	pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3081 		 dev->name, max_queue_pairs);
3082 
3083 	return 0;
3084 
3085 free_unregister_netdev:
3086 	vi->vdev->config->reset(vdev);
3087 
3088 	unregister_netdev(dev);
3089 free_failover:
3090 	net_failover_destroy(vi->failover);
3091 free_vqs:
3092 	cancel_delayed_work_sync(&vi->refill);
3093 	free_receive_page_frags(vi);
3094 	virtnet_del_vqs(vi);
3095 free:
3096 	free_netdev(dev);
3097 	return err;
3098 }
3099 
3100 static void remove_vq_common(struct virtnet_info *vi)
3101 {
3102 	vi->vdev->config->reset(vi->vdev);
3103 
3104 	/* Free unused buffers in both send and recv, if any. */
3105 	free_unused_bufs(vi);
3106 
3107 	free_receive_bufs(vi);
3108 
3109 	free_receive_page_frags(vi);
3110 
3111 	virtnet_del_vqs(vi);
3112 }
3113 
3114 static void virtnet_remove(struct virtio_device *vdev)
3115 {
3116 	struct virtnet_info *vi = vdev->priv;
3117 
3118 	virtnet_cpu_notif_remove(vi);
3119 
3120 	/* Make sure no work handler is accessing the device. */
3121 	flush_work(&vi->config_work);
3122 
3123 	unregister_netdev(vi->dev);
3124 
3125 	net_failover_destroy(vi->failover);
3126 
3127 	remove_vq_common(vi);
3128 
3129 	free_netdev(vi->dev);
3130 }
3131 
3132 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3133 {
3134 	struct virtnet_info *vi = vdev->priv;
3135 
3136 	virtnet_cpu_notif_remove(vi);
3137 	virtnet_freeze_down(vdev);
3138 	remove_vq_common(vi);
3139 
3140 	return 0;
3141 }
3142 
3143 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3144 {
3145 	struct virtnet_info *vi = vdev->priv;
3146 	int err;
3147 
3148 	err = virtnet_restore_up(vdev);
3149 	if (err)
3150 		return err;
3151 	virtnet_set_queues(vi, vi->curr_queue_pairs);
3152 
3153 	err = virtnet_cpu_notif_add(vi);
3154 	if (err)
3155 		return err;
3156 
3157 	return 0;
3158 }
3159 
3160 static struct virtio_device_id id_table[] = {
3161 	{ VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3162 	{ 0 },
3163 };
3164 
3165 #define VIRTNET_FEATURES \
3166 	VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3167 	VIRTIO_NET_F_MAC, \
3168 	VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3169 	VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3170 	VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3171 	VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3172 	VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3173 	VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3174 	VIRTIO_NET_F_CTRL_MAC_ADDR, \
3175 	VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3176 	VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3177 
3178 static unsigned int features[] = {
3179 	VIRTNET_FEATURES,
3180 };
3181 
3182 static unsigned int features_legacy[] = {
3183 	VIRTNET_FEATURES,
3184 	VIRTIO_NET_F_GSO,
3185 	VIRTIO_F_ANY_LAYOUT,
3186 };
3187 
3188 static struct virtio_driver virtio_net_driver = {
3189 	.feature_table = features,
3190 	.feature_table_size = ARRAY_SIZE(features),
3191 	.feature_table_legacy = features_legacy,
3192 	.feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3193 	.driver.name =	KBUILD_MODNAME,
3194 	.driver.owner =	THIS_MODULE,
3195 	.id_table =	id_table,
3196 	.validate =	virtnet_validate,
3197 	.probe =	virtnet_probe,
3198 	.remove =	virtnet_remove,
3199 	.config_changed = virtnet_config_changed,
3200 #ifdef CONFIG_PM_SLEEP
3201 	.freeze =	virtnet_freeze,
3202 	.restore =	virtnet_restore,
3203 #endif
3204 };
3205 
3206 static __init int virtio_net_driver_init(void)
3207 {
3208 	int ret;
3209 
3210 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3211 				      virtnet_cpu_online,
3212 				      virtnet_cpu_down_prep);
3213 	if (ret < 0)
3214 		goto out;
3215 	virtionet_online = ret;
3216 	ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3217 				      NULL, virtnet_cpu_dead);
3218 	if (ret)
3219 		goto err_dead;
3220 
3221         ret = register_virtio_driver(&virtio_net_driver);
3222 	if (ret)
3223 		goto err_virtio;
3224 	return 0;
3225 err_virtio:
3226 	cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3227 err_dead:
3228 	cpuhp_remove_multi_state(virtionet_online);
3229 out:
3230 	return ret;
3231 }
3232 module_init(virtio_net_driver_init);
3233 
3234 static __exit void virtio_net_driver_exit(void)
3235 {
3236 	unregister_virtio_driver(&virtio_net_driver);
3237 	cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3238 	cpuhp_remove_multi_state(virtionet_online);
3239 }
3240 module_exit(virtio_net_driver_exit);
3241 
3242 MODULE_DEVICE_TABLE(virtio, id_table);
3243 MODULE_DESCRIPTION("Virtio network driver");
3244 MODULE_LICENSE("GPL");
3245