xref: /openbmc/linux/drivers/net/virtio_net.c (revision 00765f8ed74240419091a1708c195405b55fe243)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* A network driver using virtio.
3  *
4  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
5  */
6 //#define DEBUG
7 #include <linux/netdevice.h>
8 #include <linux/etherdevice.h>
9 #include <linux/ethtool.h>
10 #include <linux/module.h>
11 #include <linux/virtio.h>
12 #include <linux/virtio_net.h>
13 #include <linux/bpf.h>
14 #include <linux/bpf_trace.h>
15 #include <linux/scatterlist.h>
16 #include <linux/if_vlan.h>
17 #include <linux/slab.h>
18 #include <linux/cpu.h>
19 #include <linux/average.h>
20 #include <linux/filter.h>
21 #include <linux/kernel.h>
22 #include <net/route.h>
23 #include <net/xdp.h>
24 #include <net/net_failover.h>
25 
26 static int napi_weight = NAPI_POLL_WEIGHT;
27 module_param(napi_weight, int, 0444);
28 
29 static bool csum = true, gso = true, napi_tx = true;
30 module_param(csum, bool, 0444);
31 module_param(gso, bool, 0444);
32 module_param(napi_tx, bool, 0644);
33 
34 /* FIXME: MTU in config. */
35 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
36 #define GOOD_COPY_LEN	128
37 
38 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
39 
40 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
41 #define VIRTIO_XDP_HEADROOM 256
42 
43 /* Separating two types of XDP xmit */
44 #define VIRTIO_XDP_TX		BIT(0)
45 #define VIRTIO_XDP_REDIR	BIT(1)
46 
47 #define VIRTIO_XDP_FLAG	BIT(0)
48 
49 /* RX packet size EWMA. The average packet size is used to determine the packet
50  * buffer size when refilling RX rings. As the entire RX ring may be refilled
51  * at once, the weight is chosen so that the EWMA will be insensitive to short-
52  * term, transient changes in packet size.
53  */
54 DECLARE_EWMA(pkt_len, 0, 64)
55 
56 #define VIRTNET_DRIVER_VERSION "1.0.0"
57 
58 static const unsigned long guest_offloads[] = {
59 	VIRTIO_NET_F_GUEST_TSO4,
60 	VIRTIO_NET_F_GUEST_TSO6,
61 	VIRTIO_NET_F_GUEST_ECN,
62 	VIRTIO_NET_F_GUEST_UFO,
63 	VIRTIO_NET_F_GUEST_CSUM,
64 	VIRTIO_NET_F_GUEST_USO4,
65 	VIRTIO_NET_F_GUEST_USO6,
66 	VIRTIO_NET_F_GUEST_HDRLEN
67 };
68 
69 #define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
70 				(1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
71 				(1ULL << VIRTIO_NET_F_GUEST_ECN)  | \
72 				(1ULL << VIRTIO_NET_F_GUEST_UFO)  | \
73 				(1ULL << VIRTIO_NET_F_GUEST_USO4) | \
74 				(1ULL << VIRTIO_NET_F_GUEST_USO6))
75 
76 struct virtnet_stat_desc {
77 	char desc[ETH_GSTRING_LEN];
78 	size_t offset;
79 };
80 
81 struct virtnet_sq_stats {
82 	struct u64_stats_sync syncp;
83 	u64 packets;
84 	u64 bytes;
85 	u64 xdp_tx;
86 	u64 xdp_tx_drops;
87 	u64 kicks;
88 	u64 tx_timeouts;
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 	{ "tx_timeouts",	VIRTNET_SQ_STAT(tx_timeouts) },
113 };
114 
115 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
116 	{ "packets",		VIRTNET_RQ_STAT(packets) },
117 	{ "bytes",		VIRTNET_RQ_STAT(bytes) },
118 	{ "drops",		VIRTNET_RQ_STAT(drops) },
119 	{ "xdp_packets",	VIRTNET_RQ_STAT(xdp_packets) },
120 	{ "xdp_tx",		VIRTNET_RQ_STAT(xdp_tx) },
121 	{ "xdp_redirects",	VIRTNET_RQ_STAT(xdp_redirects) },
122 	{ "xdp_drops",		VIRTNET_RQ_STAT(xdp_drops) },
123 	{ "kicks",		VIRTNET_RQ_STAT(kicks) },
124 };
125 
126 #define VIRTNET_SQ_STATS_LEN	ARRAY_SIZE(virtnet_sq_stats_desc)
127 #define VIRTNET_RQ_STATS_LEN	ARRAY_SIZE(virtnet_rq_stats_desc)
128 
129 /* Internal representation of a send virtqueue */
130 struct send_queue {
131 	/* Virtqueue associated with this send _queue */
132 	struct virtqueue *vq;
133 
134 	/* TX: fragments + linear part + virtio header */
135 	struct scatterlist sg[MAX_SKB_FRAGS + 2];
136 
137 	/* Name of the send queue: output.$index */
138 	char name[16];
139 
140 	struct virtnet_sq_stats stats;
141 
142 	struct napi_struct napi;
143 
144 	/* Record whether sq is in reset state. */
145 	bool reset;
146 };
147 
148 /* Internal representation of a receive virtqueue */
149 struct receive_queue {
150 	/* Virtqueue associated with this receive_queue */
151 	struct virtqueue *vq;
152 
153 	struct napi_struct napi;
154 
155 	struct bpf_prog __rcu *xdp_prog;
156 
157 	struct virtnet_rq_stats stats;
158 
159 	/* Chain pages by the private ptr. */
160 	struct page *pages;
161 
162 	/* Average packet length for mergeable receive buffers. */
163 	struct ewma_pkt_len mrg_avg_pkt_len;
164 
165 	/* Page frag for packet buffer allocation. */
166 	struct page_frag alloc_frag;
167 
168 	/* RX: fragments + linear part + virtio header */
169 	struct scatterlist sg[MAX_SKB_FRAGS + 2];
170 
171 	/* Min single buffer size for mergeable buffers case. */
172 	unsigned int min_buf_len;
173 
174 	/* Name of this receive queue: input.$index */
175 	char name[16];
176 
177 	struct xdp_rxq_info xdp_rxq;
178 };
179 
180 /* This structure can contain rss message with maximum settings for indirection table and keysize
181  * Note, that default structure that describes RSS configuration virtio_net_rss_config
182  * contains same info but can't handle table values.
183  * In any case, structure would be passed to virtio hw through sg_buf split by parts
184  * because table sizes may be differ according to the device configuration.
185  */
186 #define VIRTIO_NET_RSS_MAX_KEY_SIZE     40
187 #define VIRTIO_NET_RSS_MAX_TABLE_LEN    128
188 struct virtio_net_ctrl_rss {
189 	u32 hash_types;
190 	u16 indirection_table_mask;
191 	u16 unclassified_queue;
192 	u16 indirection_table[VIRTIO_NET_RSS_MAX_TABLE_LEN];
193 	u16 max_tx_vq;
194 	u8 hash_key_length;
195 	u8 key[VIRTIO_NET_RSS_MAX_KEY_SIZE];
196 };
197 
198 /* Control VQ buffers: protected by the rtnl lock */
199 struct control_buf {
200 	struct virtio_net_ctrl_hdr hdr;
201 	virtio_net_ctrl_ack status;
202 	struct virtio_net_ctrl_mq mq;
203 	u8 promisc;
204 	u8 allmulti;
205 	__virtio16 vid;
206 	__virtio64 offloads;
207 	struct virtio_net_ctrl_rss rss;
208 };
209 
210 struct virtnet_info {
211 	struct virtio_device *vdev;
212 	struct virtqueue *cvq;
213 	struct net_device *dev;
214 	struct send_queue *sq;
215 	struct receive_queue *rq;
216 	unsigned int status;
217 
218 	/* Max # of queue pairs supported by the device */
219 	u16 max_queue_pairs;
220 
221 	/* # of queue pairs currently used by the driver */
222 	u16 curr_queue_pairs;
223 
224 	/* # of XDP queue pairs currently used by the driver */
225 	u16 xdp_queue_pairs;
226 
227 	/* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
228 	bool xdp_enabled;
229 
230 	/* I like... big packets and I cannot lie! */
231 	bool big_packets;
232 
233 	/* number of sg entries allocated for big packets */
234 	unsigned int big_packets_num_skbfrags;
235 
236 	/* Host will merge rx buffers for big packets (shake it! shake it!) */
237 	bool mergeable_rx_bufs;
238 
239 	/* Host supports rss and/or hash report */
240 	bool has_rss;
241 	bool has_rss_hash_report;
242 	u8 rss_key_size;
243 	u16 rss_indir_table_size;
244 	u32 rss_hash_types_supported;
245 	u32 rss_hash_types_saved;
246 
247 	/* Has control virtqueue */
248 	bool has_cvq;
249 
250 	/* Host can handle any s/g split between our header and packet data */
251 	bool any_header_sg;
252 
253 	/* Packet virtio header size */
254 	u8 hdr_len;
255 
256 	/* Work struct for delayed refilling if we run low on memory. */
257 	struct delayed_work refill;
258 
259 	/* Is delayed refill enabled? */
260 	bool refill_enabled;
261 
262 	/* The lock to synchronize the access to refill_enabled */
263 	spinlock_t refill_lock;
264 
265 	/* Work struct for config space updates */
266 	struct work_struct config_work;
267 
268 	/* Does the affinity hint is set for virtqueues? */
269 	bool affinity_hint_set;
270 
271 	/* CPU hotplug instances for online & dead */
272 	struct hlist_node node;
273 	struct hlist_node node_dead;
274 
275 	struct control_buf *ctrl;
276 
277 	/* Ethtool settings */
278 	u8 duplex;
279 	u32 speed;
280 
281 	/* Interrupt coalescing settings */
282 	u32 tx_usecs;
283 	u32 rx_usecs;
284 	u32 tx_max_packets;
285 	u32 rx_max_packets;
286 
287 	unsigned long guest_offloads;
288 	unsigned long guest_offloads_capable;
289 
290 	/* failover when STANDBY feature enabled */
291 	struct failover *failover;
292 };
293 
294 struct padded_vnet_hdr {
295 	struct virtio_net_hdr_v1_hash hdr;
296 	/*
297 	 * hdr is in a separate sg buffer, and data sg buffer shares same page
298 	 * with this header sg. This padding makes next sg 16 byte aligned
299 	 * after the header.
300 	 */
301 	char padding[12];
302 };
303 
304 static void virtnet_rq_free_unused_buf(struct virtqueue *vq, void *buf);
305 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf);
306 
307 static bool is_xdp_frame(void *ptr)
308 {
309 	return (unsigned long)ptr & VIRTIO_XDP_FLAG;
310 }
311 
312 static void *xdp_to_ptr(struct xdp_frame *ptr)
313 {
314 	return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
315 }
316 
317 static struct xdp_frame *ptr_to_xdp(void *ptr)
318 {
319 	return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
320 }
321 
322 /* Converting between virtqueue no. and kernel tx/rx queue no.
323  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
324  */
325 static int vq2txq(struct virtqueue *vq)
326 {
327 	return (vq->index - 1) / 2;
328 }
329 
330 static int txq2vq(int txq)
331 {
332 	return txq * 2 + 1;
333 }
334 
335 static int vq2rxq(struct virtqueue *vq)
336 {
337 	return vq->index / 2;
338 }
339 
340 static int rxq2vq(int rxq)
341 {
342 	return rxq * 2;
343 }
344 
345 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
346 {
347 	return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
348 }
349 
350 /*
351  * private is used to chain pages for big packets, put the whole
352  * most recent used list in the beginning for reuse
353  */
354 static void give_pages(struct receive_queue *rq, struct page *page)
355 {
356 	struct page *end;
357 
358 	/* Find end of list, sew whole thing into vi->rq.pages. */
359 	for (end = page; end->private; end = (struct page *)end->private);
360 	end->private = (unsigned long)rq->pages;
361 	rq->pages = page;
362 }
363 
364 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
365 {
366 	struct page *p = rq->pages;
367 
368 	if (p) {
369 		rq->pages = (struct page *)p->private;
370 		/* clear private here, it is used to chain pages */
371 		p->private = 0;
372 	} else
373 		p = alloc_page(gfp_mask);
374 	return p;
375 }
376 
377 static void enable_delayed_refill(struct virtnet_info *vi)
378 {
379 	spin_lock_bh(&vi->refill_lock);
380 	vi->refill_enabled = true;
381 	spin_unlock_bh(&vi->refill_lock);
382 }
383 
384 static void disable_delayed_refill(struct virtnet_info *vi)
385 {
386 	spin_lock_bh(&vi->refill_lock);
387 	vi->refill_enabled = false;
388 	spin_unlock_bh(&vi->refill_lock);
389 }
390 
391 static void virtqueue_napi_schedule(struct napi_struct *napi,
392 				    struct virtqueue *vq)
393 {
394 	if (napi_schedule_prep(napi)) {
395 		virtqueue_disable_cb(vq);
396 		__napi_schedule(napi);
397 	}
398 }
399 
400 static void virtqueue_napi_complete(struct napi_struct *napi,
401 				    struct virtqueue *vq, int processed)
402 {
403 	int opaque;
404 
405 	opaque = virtqueue_enable_cb_prepare(vq);
406 	if (napi_complete_done(napi, processed)) {
407 		if (unlikely(virtqueue_poll(vq, opaque)))
408 			virtqueue_napi_schedule(napi, vq);
409 	} else {
410 		virtqueue_disable_cb(vq);
411 	}
412 }
413 
414 static void skb_xmit_done(struct virtqueue *vq)
415 {
416 	struct virtnet_info *vi = vq->vdev->priv;
417 	struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
418 
419 	/* Suppress further interrupts. */
420 	virtqueue_disable_cb(vq);
421 
422 	if (napi->weight)
423 		virtqueue_napi_schedule(napi, vq);
424 	else
425 		/* We were probably waiting for more output buffers. */
426 		netif_wake_subqueue(vi->dev, vq2txq(vq));
427 }
428 
429 #define MRG_CTX_HEADER_SHIFT 22
430 static void *mergeable_len_to_ctx(unsigned int truesize,
431 				  unsigned int headroom)
432 {
433 	return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
434 }
435 
436 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
437 {
438 	return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
439 }
440 
441 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
442 {
443 	return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
444 }
445 
446 /* Called from bottom half context */
447 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
448 				   struct receive_queue *rq,
449 				   struct page *page, unsigned int offset,
450 				   unsigned int len, unsigned int truesize,
451 				   unsigned int headroom)
452 {
453 	struct sk_buff *skb;
454 	struct virtio_net_hdr_mrg_rxbuf *hdr;
455 	unsigned int copy, hdr_len, hdr_padded_len;
456 	struct page *page_to_free = NULL;
457 	int tailroom, shinfo_size;
458 	char *p, *hdr_p, *buf;
459 
460 	p = page_address(page) + offset;
461 	hdr_p = p;
462 
463 	hdr_len = vi->hdr_len;
464 	if (vi->mergeable_rx_bufs)
465 		hdr_padded_len = hdr_len;
466 	else
467 		hdr_padded_len = sizeof(struct padded_vnet_hdr);
468 
469 	buf = p - headroom;
470 	len -= hdr_len;
471 	offset += hdr_padded_len;
472 	p += hdr_padded_len;
473 	tailroom = truesize - headroom  - hdr_padded_len - len;
474 
475 	shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
476 
477 	/* copy small packet so we can reuse these pages */
478 	if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
479 		skb = build_skb(buf, truesize);
480 		if (unlikely(!skb))
481 			return NULL;
482 
483 		skb_reserve(skb, p - buf);
484 		skb_put(skb, len);
485 
486 		page = (struct page *)page->private;
487 		if (page)
488 			give_pages(rq, page);
489 		goto ok;
490 	}
491 
492 	/* copy small packet so we can reuse these pages for small data */
493 	skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
494 	if (unlikely(!skb))
495 		return NULL;
496 
497 	/* Copy all frame if it fits skb->head, otherwise
498 	 * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
499 	 */
500 	if (len <= skb_tailroom(skb))
501 		copy = len;
502 	else
503 		copy = ETH_HLEN;
504 	skb_put_data(skb, p, copy);
505 
506 	len -= copy;
507 	offset += copy;
508 
509 	if (vi->mergeable_rx_bufs) {
510 		if (len)
511 			skb_add_rx_frag(skb, 0, page, offset, len, truesize);
512 		else
513 			page_to_free = page;
514 		goto ok;
515 	}
516 
517 	/*
518 	 * Verify that we can indeed put this data into a skb.
519 	 * This is here to handle cases when the device erroneously
520 	 * tries to receive more than is possible. This is usually
521 	 * the case of a broken device.
522 	 */
523 	if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
524 		net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
525 		dev_kfree_skb(skb);
526 		return NULL;
527 	}
528 	BUG_ON(offset >= PAGE_SIZE);
529 	while (len) {
530 		unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
531 		skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
532 				frag_size, truesize);
533 		len -= frag_size;
534 		page = (struct page *)page->private;
535 		offset = 0;
536 	}
537 
538 	if (page)
539 		give_pages(rq, page);
540 
541 ok:
542 	hdr = skb_vnet_hdr(skb);
543 	memcpy(hdr, hdr_p, hdr_len);
544 	if (page_to_free)
545 		put_page(page_to_free);
546 
547 	return skb;
548 }
549 
550 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
551 {
552 	unsigned int len;
553 	unsigned int packets = 0;
554 	unsigned int bytes = 0;
555 	void *ptr;
556 
557 	while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
558 		if (likely(!is_xdp_frame(ptr))) {
559 			struct sk_buff *skb = ptr;
560 
561 			pr_debug("Sent skb %p\n", skb);
562 
563 			bytes += skb->len;
564 			napi_consume_skb(skb, in_napi);
565 		} else {
566 			struct xdp_frame *frame = ptr_to_xdp(ptr);
567 
568 			bytes += xdp_get_frame_len(frame);
569 			xdp_return_frame(frame);
570 		}
571 		packets++;
572 	}
573 
574 	/* Avoid overhead when no packets have been processed
575 	 * happens when called speculatively from start_xmit.
576 	 */
577 	if (!packets)
578 		return;
579 
580 	u64_stats_update_begin(&sq->stats.syncp);
581 	sq->stats.bytes += bytes;
582 	sq->stats.packets += packets;
583 	u64_stats_update_end(&sq->stats.syncp);
584 }
585 
586 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
587 {
588 	if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
589 		return false;
590 	else if (q < vi->curr_queue_pairs)
591 		return true;
592 	else
593 		return false;
594 }
595 
596 static void check_sq_full_and_disable(struct virtnet_info *vi,
597 				      struct net_device *dev,
598 				      struct send_queue *sq)
599 {
600 	bool use_napi = sq->napi.weight;
601 	int qnum;
602 
603 	qnum = sq - vi->sq;
604 
605 	/* If running out of space, stop queue to avoid getting packets that we
606 	 * are then unable to transmit.
607 	 * An alternative would be to force queuing layer to requeue the skb by
608 	 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
609 	 * returned in a normal path of operation: it means that driver is not
610 	 * maintaining the TX queue stop/start state properly, and causes
611 	 * the stack to do a non-trivial amount of useless work.
612 	 * Since most packets only take 1 or 2 ring slots, stopping the queue
613 	 * early means 16 slots are typically wasted.
614 	 */
615 	if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
616 		netif_stop_subqueue(dev, qnum);
617 		if (use_napi) {
618 			if (unlikely(!virtqueue_enable_cb_delayed(sq->vq)))
619 				virtqueue_napi_schedule(&sq->napi, sq->vq);
620 		} else if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
621 			/* More just got used, free them then recheck. */
622 			free_old_xmit_skbs(sq, false);
623 			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
624 				netif_start_subqueue(dev, qnum);
625 				virtqueue_disable_cb(sq->vq);
626 			}
627 		}
628 	}
629 }
630 
631 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
632 				   struct send_queue *sq,
633 				   struct xdp_frame *xdpf)
634 {
635 	struct virtio_net_hdr_mrg_rxbuf *hdr;
636 	struct skb_shared_info *shinfo;
637 	u8 nr_frags = 0;
638 	int err, i;
639 
640 	if (unlikely(xdpf->headroom < vi->hdr_len))
641 		return -EOVERFLOW;
642 
643 	if (unlikely(xdp_frame_has_frags(xdpf))) {
644 		shinfo = xdp_get_shared_info_from_frame(xdpf);
645 		nr_frags = shinfo->nr_frags;
646 	}
647 
648 	/* In wrapping function virtnet_xdp_xmit(), we need to free
649 	 * up the pending old buffers, where we need to calculate the
650 	 * position of skb_shared_info in xdp_get_frame_len() and
651 	 * xdp_return_frame(), which will involve to xdpf->data and
652 	 * xdpf->headroom. Therefore, we need to update the value of
653 	 * headroom synchronously here.
654 	 */
655 	xdpf->headroom -= vi->hdr_len;
656 	xdpf->data -= vi->hdr_len;
657 	/* Zero header and leave csum up to XDP layers */
658 	hdr = xdpf->data;
659 	memset(hdr, 0, vi->hdr_len);
660 	xdpf->len   += vi->hdr_len;
661 
662 	sg_init_table(sq->sg, nr_frags + 1);
663 	sg_set_buf(sq->sg, xdpf->data, xdpf->len);
664 	for (i = 0; i < nr_frags; i++) {
665 		skb_frag_t *frag = &shinfo->frags[i];
666 
667 		sg_set_page(&sq->sg[i + 1], skb_frag_page(frag),
668 			    skb_frag_size(frag), skb_frag_off(frag));
669 	}
670 
671 	err = virtqueue_add_outbuf(sq->vq, sq->sg, nr_frags + 1,
672 				   xdp_to_ptr(xdpf), GFP_ATOMIC);
673 	if (unlikely(err))
674 		return -ENOSPC; /* Caller handle free/refcnt */
675 
676 	return 0;
677 }
678 
679 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
680  * the current cpu, so it does not need to be locked.
681  *
682  * Here we use marco instead of inline functions because we have to deal with
683  * three issues at the same time: 1. the choice of sq. 2. judge and execute the
684  * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
685  * functions to perfectly solve these three problems at the same time.
686  */
687 #define virtnet_xdp_get_sq(vi) ({                                       \
688 	int cpu = smp_processor_id();                                   \
689 	struct netdev_queue *txq;                                       \
690 	typeof(vi) v = (vi);                                            \
691 	unsigned int qp;                                                \
692 									\
693 	if (v->curr_queue_pairs > nr_cpu_ids) {                         \
694 		qp = v->curr_queue_pairs - v->xdp_queue_pairs;          \
695 		qp += cpu;                                              \
696 		txq = netdev_get_tx_queue(v->dev, qp);                  \
697 		__netif_tx_acquire(txq);                                \
698 	} else {                                                        \
699 		qp = cpu % v->curr_queue_pairs;                         \
700 		txq = netdev_get_tx_queue(v->dev, qp);                  \
701 		__netif_tx_lock(txq, cpu);                              \
702 	}                                                               \
703 	v->sq + qp;                                                     \
704 })
705 
706 #define virtnet_xdp_put_sq(vi, q) {                                     \
707 	struct netdev_queue *txq;                                       \
708 	typeof(vi) v = (vi);                                            \
709 									\
710 	txq = netdev_get_tx_queue(v->dev, (q) - v->sq);                 \
711 	if (v->curr_queue_pairs > nr_cpu_ids)                           \
712 		__netif_tx_release(txq);                                \
713 	else                                                            \
714 		__netif_tx_unlock(txq);                                 \
715 }
716 
717 static int virtnet_xdp_xmit(struct net_device *dev,
718 			    int n, struct xdp_frame **frames, u32 flags)
719 {
720 	struct virtnet_info *vi = netdev_priv(dev);
721 	struct receive_queue *rq = vi->rq;
722 	struct bpf_prog *xdp_prog;
723 	struct send_queue *sq;
724 	unsigned int len;
725 	int packets = 0;
726 	int bytes = 0;
727 	int nxmit = 0;
728 	int kicks = 0;
729 	void *ptr;
730 	int ret;
731 	int i;
732 
733 	/* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
734 	 * indicate XDP resources have been successfully allocated.
735 	 */
736 	xdp_prog = rcu_access_pointer(rq->xdp_prog);
737 	if (!xdp_prog)
738 		return -ENXIO;
739 
740 	sq = virtnet_xdp_get_sq(vi);
741 
742 	if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
743 		ret = -EINVAL;
744 		goto out;
745 	}
746 
747 	/* Free up any pending old buffers before queueing new ones. */
748 	while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
749 		if (likely(is_xdp_frame(ptr))) {
750 			struct xdp_frame *frame = ptr_to_xdp(ptr);
751 
752 			bytes += xdp_get_frame_len(frame);
753 			xdp_return_frame(frame);
754 		} else {
755 			struct sk_buff *skb = ptr;
756 
757 			bytes += skb->len;
758 			napi_consume_skb(skb, false);
759 		}
760 		packets++;
761 	}
762 
763 	for (i = 0; i < n; i++) {
764 		struct xdp_frame *xdpf = frames[i];
765 
766 		if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
767 			break;
768 		nxmit++;
769 	}
770 	ret = nxmit;
771 
772 	if (!is_xdp_raw_buffer_queue(vi, sq - vi->sq))
773 		check_sq_full_and_disable(vi, dev, sq);
774 
775 	if (flags & XDP_XMIT_FLUSH) {
776 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
777 			kicks = 1;
778 	}
779 out:
780 	u64_stats_update_begin(&sq->stats.syncp);
781 	sq->stats.bytes += bytes;
782 	sq->stats.packets += packets;
783 	sq->stats.xdp_tx += n;
784 	sq->stats.xdp_tx_drops += n - nxmit;
785 	sq->stats.kicks += kicks;
786 	u64_stats_update_end(&sq->stats.syncp);
787 
788 	virtnet_xdp_put_sq(vi, sq);
789 	return ret;
790 }
791 
792 static int virtnet_xdp_handler(struct bpf_prog *xdp_prog, struct xdp_buff *xdp,
793 			       struct net_device *dev,
794 			       unsigned int *xdp_xmit,
795 			       struct virtnet_rq_stats *stats)
796 {
797 	struct xdp_frame *xdpf;
798 	int err;
799 	u32 act;
800 
801 	act = bpf_prog_run_xdp(xdp_prog, xdp);
802 	stats->xdp_packets++;
803 
804 	switch (act) {
805 	case XDP_PASS:
806 		return act;
807 
808 	case XDP_TX:
809 		stats->xdp_tx++;
810 		xdpf = xdp_convert_buff_to_frame(xdp);
811 		if (unlikely(!xdpf)) {
812 			netdev_dbg(dev, "convert buff to frame failed for xdp\n");
813 			return XDP_DROP;
814 		}
815 
816 		err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
817 		if (unlikely(!err)) {
818 			xdp_return_frame_rx_napi(xdpf);
819 		} else if (unlikely(err < 0)) {
820 			trace_xdp_exception(dev, xdp_prog, act);
821 			return XDP_DROP;
822 		}
823 		*xdp_xmit |= VIRTIO_XDP_TX;
824 		return act;
825 
826 	case XDP_REDIRECT:
827 		stats->xdp_redirects++;
828 		err = xdp_do_redirect(dev, xdp, xdp_prog);
829 		if (err)
830 			return XDP_DROP;
831 
832 		*xdp_xmit |= VIRTIO_XDP_REDIR;
833 		return act;
834 
835 	default:
836 		bpf_warn_invalid_xdp_action(dev, xdp_prog, act);
837 		fallthrough;
838 	case XDP_ABORTED:
839 		trace_xdp_exception(dev, xdp_prog, act);
840 		fallthrough;
841 	case XDP_DROP:
842 		return XDP_DROP;
843 	}
844 }
845 
846 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
847 {
848 	return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
849 }
850 
851 /* We copy the packet for XDP in the following cases:
852  *
853  * 1) Packet is scattered across multiple rx buffers.
854  * 2) Headroom space is insufficient.
855  *
856  * This is inefficient but it's a temporary condition that
857  * we hit right after XDP is enabled and until queue is refilled
858  * with large buffers with sufficient headroom - so it should affect
859  * at most queue size packets.
860  * Afterwards, the conditions to enable
861  * XDP should preclude the underlying device from sending packets
862  * across multiple buffers (num_buf > 1), and we make sure buffers
863  * have enough headroom.
864  */
865 static struct page *xdp_linearize_page(struct receive_queue *rq,
866 				       int *num_buf,
867 				       struct page *p,
868 				       int offset,
869 				       int page_off,
870 				       unsigned int *len)
871 {
872 	int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
873 	struct page *page;
874 
875 	if (page_off + *len + tailroom > PAGE_SIZE)
876 		return NULL;
877 
878 	page = alloc_page(GFP_ATOMIC);
879 	if (!page)
880 		return NULL;
881 
882 	memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
883 	page_off += *len;
884 
885 	while (--*num_buf) {
886 		unsigned int buflen;
887 		void *buf;
888 		int off;
889 
890 		buf = virtqueue_get_buf(rq->vq, &buflen);
891 		if (unlikely(!buf))
892 			goto err_buf;
893 
894 		p = virt_to_head_page(buf);
895 		off = buf - page_address(p);
896 
897 		/* guard against a misconfigured or uncooperative backend that
898 		 * is sending packet larger than the MTU.
899 		 */
900 		if ((page_off + buflen + tailroom) > PAGE_SIZE) {
901 			put_page(p);
902 			goto err_buf;
903 		}
904 
905 		memcpy(page_address(page) + page_off,
906 		       page_address(p) + off, buflen);
907 		page_off += buflen;
908 		put_page(p);
909 	}
910 
911 	/* Headroom does not contribute to packet length */
912 	*len = page_off - VIRTIO_XDP_HEADROOM;
913 	return page;
914 err_buf:
915 	__free_pages(page, 0);
916 	return NULL;
917 }
918 
919 static struct sk_buff *receive_small(struct net_device *dev,
920 				     struct virtnet_info *vi,
921 				     struct receive_queue *rq,
922 				     void *buf, void *ctx,
923 				     unsigned int len,
924 				     unsigned int *xdp_xmit,
925 				     struct virtnet_rq_stats *stats)
926 {
927 	struct sk_buff *skb;
928 	struct bpf_prog *xdp_prog;
929 	unsigned int xdp_headroom = (unsigned long)ctx;
930 	unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
931 	unsigned int headroom = vi->hdr_len + header_offset;
932 	unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
933 			      SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
934 	struct page *page = virt_to_head_page(buf);
935 	unsigned int delta = 0;
936 	struct page *xdp_page;
937 	unsigned int metasize = 0;
938 
939 	len -= vi->hdr_len;
940 	stats->bytes += len;
941 
942 	if (unlikely(len > GOOD_PACKET_LEN)) {
943 		pr_debug("%s: rx error: len %u exceeds max size %d\n",
944 			 dev->name, len, GOOD_PACKET_LEN);
945 		dev->stats.rx_length_errors++;
946 		goto err;
947 	}
948 
949 	if (likely(!vi->xdp_enabled)) {
950 		xdp_prog = NULL;
951 		goto skip_xdp;
952 	}
953 
954 	rcu_read_lock();
955 	xdp_prog = rcu_dereference(rq->xdp_prog);
956 	if (xdp_prog) {
957 		struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
958 		struct xdp_buff xdp;
959 		void *orig_data;
960 		u32 act;
961 
962 		if (unlikely(hdr->hdr.gso_type))
963 			goto err_xdp;
964 
965 		if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
966 			int offset = buf - page_address(page) + header_offset;
967 			unsigned int tlen = len + vi->hdr_len;
968 			int num_buf = 1;
969 
970 			xdp_headroom = virtnet_get_headroom(vi);
971 			header_offset = VIRTNET_RX_PAD + xdp_headroom;
972 			headroom = vi->hdr_len + header_offset;
973 			buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
974 				 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
975 			xdp_page = xdp_linearize_page(rq, &num_buf, page,
976 						      offset, header_offset,
977 						      &tlen);
978 			if (!xdp_page)
979 				goto err_xdp;
980 
981 			buf = page_address(xdp_page);
982 			put_page(page);
983 			page = xdp_page;
984 		}
985 
986 		xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
987 		xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
988 				 xdp_headroom, len, true);
989 		orig_data = xdp.data;
990 
991 		act = virtnet_xdp_handler(xdp_prog, &xdp, dev, xdp_xmit, stats);
992 
993 		switch (act) {
994 		case XDP_PASS:
995 			/* Recalculate length in case bpf program changed it */
996 			delta = orig_data - xdp.data;
997 			len = xdp.data_end - xdp.data;
998 			metasize = xdp.data - xdp.data_meta;
999 			break;
1000 		case XDP_TX:
1001 		case XDP_REDIRECT:
1002 			rcu_read_unlock();
1003 			goto xdp_xmit;
1004 		default:
1005 			goto err_xdp;
1006 		}
1007 	}
1008 	rcu_read_unlock();
1009 
1010 skip_xdp:
1011 	skb = build_skb(buf, buflen);
1012 	if (!skb)
1013 		goto err;
1014 	skb_reserve(skb, headroom - delta);
1015 	skb_put(skb, len);
1016 	if (!xdp_prog) {
1017 		buf += header_offset;
1018 		memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
1019 	} /* keep zeroed vnet hdr since XDP is loaded */
1020 
1021 	if (metasize)
1022 		skb_metadata_set(skb, metasize);
1023 
1024 	return skb;
1025 
1026 err_xdp:
1027 	rcu_read_unlock();
1028 	stats->xdp_drops++;
1029 err:
1030 	stats->drops++;
1031 	put_page(page);
1032 xdp_xmit:
1033 	return NULL;
1034 }
1035 
1036 static struct sk_buff *receive_big(struct net_device *dev,
1037 				   struct virtnet_info *vi,
1038 				   struct receive_queue *rq,
1039 				   void *buf,
1040 				   unsigned int len,
1041 				   struct virtnet_rq_stats *stats)
1042 {
1043 	struct page *page = buf;
1044 	struct sk_buff *skb =
1045 		page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, 0);
1046 
1047 	stats->bytes += len - vi->hdr_len;
1048 	if (unlikely(!skb))
1049 		goto err;
1050 
1051 	return skb;
1052 
1053 err:
1054 	stats->drops++;
1055 	give_pages(rq, page);
1056 	return NULL;
1057 }
1058 
1059 /* Why not use xdp_build_skb_from_frame() ?
1060  * XDP core assumes that xdp frags are PAGE_SIZE in length, while in
1061  * virtio-net there are 2 points that do not match its requirements:
1062  *  1. The size of the prefilled buffer is not fixed before xdp is set.
1063  *  2. xdp_build_skb_from_frame() does more checks that we don't need,
1064  *     like eth_type_trans() (which virtio-net does in receive_buf()).
1065  */
1066 static struct sk_buff *build_skb_from_xdp_buff(struct net_device *dev,
1067 					       struct virtnet_info *vi,
1068 					       struct xdp_buff *xdp,
1069 					       unsigned int xdp_frags_truesz)
1070 {
1071 	struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
1072 	unsigned int headroom, data_len;
1073 	struct sk_buff *skb;
1074 	int metasize;
1075 	u8 nr_frags;
1076 
1077 	if (unlikely(xdp->data_end > xdp_data_hard_end(xdp))) {
1078 		pr_debug("Error building skb as missing reserved tailroom for xdp");
1079 		return NULL;
1080 	}
1081 
1082 	if (unlikely(xdp_buff_has_frags(xdp)))
1083 		nr_frags = sinfo->nr_frags;
1084 
1085 	skb = build_skb(xdp->data_hard_start, xdp->frame_sz);
1086 	if (unlikely(!skb))
1087 		return NULL;
1088 
1089 	headroom = xdp->data - xdp->data_hard_start;
1090 	data_len = xdp->data_end - xdp->data;
1091 	skb_reserve(skb, headroom);
1092 	__skb_put(skb, data_len);
1093 
1094 	metasize = xdp->data - xdp->data_meta;
1095 	metasize = metasize > 0 ? metasize : 0;
1096 	if (metasize)
1097 		skb_metadata_set(skb, metasize);
1098 
1099 	if (unlikely(xdp_buff_has_frags(xdp)))
1100 		xdp_update_skb_shared_info(skb, nr_frags,
1101 					   sinfo->xdp_frags_size,
1102 					   xdp_frags_truesz,
1103 					   xdp_buff_is_frag_pfmemalloc(xdp));
1104 
1105 	return skb;
1106 }
1107 
1108 /* TODO: build xdp in big mode */
1109 static int virtnet_build_xdp_buff_mrg(struct net_device *dev,
1110 				      struct virtnet_info *vi,
1111 				      struct receive_queue *rq,
1112 				      struct xdp_buff *xdp,
1113 				      void *buf,
1114 				      unsigned int len,
1115 				      unsigned int frame_sz,
1116 				      int *num_buf,
1117 				      unsigned int *xdp_frags_truesize,
1118 				      struct virtnet_rq_stats *stats)
1119 {
1120 	struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1121 	unsigned int headroom, tailroom, room;
1122 	unsigned int truesize, cur_frag_size;
1123 	struct skb_shared_info *shinfo;
1124 	unsigned int xdp_frags_truesz = 0;
1125 	struct page *page;
1126 	skb_frag_t *frag;
1127 	int offset;
1128 	void *ctx;
1129 
1130 	xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq);
1131 	xdp_prepare_buff(xdp, buf - VIRTIO_XDP_HEADROOM,
1132 			 VIRTIO_XDP_HEADROOM + vi->hdr_len, len - vi->hdr_len, true);
1133 
1134 	if (!*num_buf)
1135 		return 0;
1136 
1137 	if (*num_buf > 1) {
1138 		/* If we want to build multi-buffer xdp, we need
1139 		 * to specify that the flags of xdp_buff have the
1140 		 * XDP_FLAGS_HAS_FRAG bit.
1141 		 */
1142 		if (!xdp_buff_has_frags(xdp))
1143 			xdp_buff_set_frags_flag(xdp);
1144 
1145 		shinfo = xdp_get_shared_info_from_buff(xdp);
1146 		shinfo->nr_frags = 0;
1147 		shinfo->xdp_frags_size = 0;
1148 	}
1149 
1150 	if (*num_buf > MAX_SKB_FRAGS + 1)
1151 		return -EINVAL;
1152 
1153 	while (--*num_buf > 0) {
1154 		buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1155 		if (unlikely(!buf)) {
1156 			pr_debug("%s: rx error: %d buffers out of %d missing\n",
1157 				 dev->name, *num_buf,
1158 				 virtio16_to_cpu(vi->vdev, hdr->num_buffers));
1159 			dev->stats.rx_length_errors++;
1160 			return -EINVAL;
1161 		}
1162 
1163 		stats->bytes += len;
1164 		page = virt_to_head_page(buf);
1165 		offset = buf - page_address(page);
1166 
1167 		truesize = mergeable_ctx_to_truesize(ctx);
1168 		headroom = mergeable_ctx_to_headroom(ctx);
1169 		tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1170 		room = SKB_DATA_ALIGN(headroom + tailroom);
1171 
1172 		cur_frag_size = truesize;
1173 		xdp_frags_truesz += cur_frag_size;
1174 		if (unlikely(len > truesize - room || cur_frag_size > PAGE_SIZE)) {
1175 			put_page(page);
1176 			pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1177 				 dev->name, len, (unsigned long)(truesize - room));
1178 			dev->stats.rx_length_errors++;
1179 			return -EINVAL;
1180 		}
1181 
1182 		frag = &shinfo->frags[shinfo->nr_frags++];
1183 		__skb_frag_set_page(frag, page);
1184 		skb_frag_off_set(frag, offset);
1185 		skb_frag_size_set(frag, len);
1186 		if (page_is_pfmemalloc(page))
1187 			xdp_buff_set_frag_pfmemalloc(xdp);
1188 
1189 		shinfo->xdp_frags_size += len;
1190 	}
1191 
1192 	*xdp_frags_truesize = xdp_frags_truesz;
1193 	return 0;
1194 }
1195 
1196 static void *mergeable_xdp_get_buf(struct virtnet_info *vi,
1197 				   struct receive_queue *rq,
1198 				   struct bpf_prog *xdp_prog,
1199 				   void *ctx,
1200 				   unsigned int *frame_sz,
1201 				   int *num_buf,
1202 				   struct page **page,
1203 				   int offset,
1204 				   unsigned int *len,
1205 				   struct virtio_net_hdr_mrg_rxbuf *hdr)
1206 {
1207 	unsigned int truesize = mergeable_ctx_to_truesize(ctx);
1208 	unsigned int headroom = mergeable_ctx_to_headroom(ctx);
1209 	struct page *xdp_page;
1210 	unsigned int xdp_room;
1211 
1212 	/* Transient failure which in theory could occur if
1213 	 * in-flight packets from before XDP was enabled reach
1214 	 * the receive path after XDP is loaded.
1215 	 */
1216 	if (unlikely(hdr->hdr.gso_type))
1217 		return NULL;
1218 
1219 	/* Now XDP core assumes frag size is PAGE_SIZE, but buffers
1220 	 * with headroom may add hole in truesize, which
1221 	 * make their length exceed PAGE_SIZE. So we disabled the
1222 	 * hole mechanism for xdp. See add_recvbuf_mergeable().
1223 	 */
1224 	*frame_sz = truesize;
1225 
1226 	if (likely(headroom >= virtnet_get_headroom(vi) &&
1227 		   (*num_buf == 1 || xdp_prog->aux->xdp_has_frags))) {
1228 		return page_address(*page) + offset;
1229 	}
1230 
1231 	/* This happens when headroom is not enough because
1232 	 * of the buffer was prefilled before XDP is set.
1233 	 * This should only happen for the first several packets.
1234 	 * In fact, vq reset can be used here to help us clean up
1235 	 * the prefilled buffers, but many existing devices do not
1236 	 * support it, and we don't want to bother users who are
1237 	 * using xdp normally.
1238 	 */
1239 	if (!xdp_prog->aux->xdp_has_frags) {
1240 		/* linearize data for XDP */
1241 		xdp_page = xdp_linearize_page(rq, num_buf,
1242 					      *page, offset,
1243 					      VIRTIO_XDP_HEADROOM,
1244 					      len);
1245 		if (!xdp_page)
1246 			return NULL;
1247 	} else {
1248 		xdp_room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM +
1249 					  sizeof(struct skb_shared_info));
1250 		if (*len + xdp_room > PAGE_SIZE)
1251 			return NULL;
1252 
1253 		xdp_page = alloc_page(GFP_ATOMIC);
1254 		if (!xdp_page)
1255 			return NULL;
1256 
1257 		memcpy(page_address(xdp_page) + VIRTIO_XDP_HEADROOM,
1258 		       page_address(*page) + offset, *len);
1259 	}
1260 
1261 	*frame_sz = PAGE_SIZE;
1262 
1263 	put_page(*page);
1264 
1265 	*page = xdp_page;
1266 
1267 	return page_address(*page) + VIRTIO_XDP_HEADROOM;
1268 }
1269 
1270 static struct sk_buff *receive_mergeable(struct net_device *dev,
1271 					 struct virtnet_info *vi,
1272 					 struct receive_queue *rq,
1273 					 void *buf,
1274 					 void *ctx,
1275 					 unsigned int len,
1276 					 unsigned int *xdp_xmit,
1277 					 struct virtnet_rq_stats *stats)
1278 {
1279 	struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1280 	int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
1281 	struct page *page = virt_to_head_page(buf);
1282 	int offset = buf - page_address(page);
1283 	struct sk_buff *head_skb, *curr_skb;
1284 	struct bpf_prog *xdp_prog;
1285 	unsigned int truesize = mergeable_ctx_to_truesize(ctx);
1286 	unsigned int headroom = mergeable_ctx_to_headroom(ctx);
1287 	unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1288 	unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1289 	unsigned int frame_sz;
1290 	int err;
1291 
1292 	head_skb = NULL;
1293 	stats->bytes += len - vi->hdr_len;
1294 
1295 	if (unlikely(len > truesize - room)) {
1296 		pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1297 			 dev->name, len, (unsigned long)(truesize - room));
1298 		dev->stats.rx_length_errors++;
1299 		goto err_skb;
1300 	}
1301 
1302 	if (likely(!vi->xdp_enabled)) {
1303 		xdp_prog = NULL;
1304 		goto skip_xdp;
1305 	}
1306 
1307 	rcu_read_lock();
1308 	xdp_prog = rcu_dereference(rq->xdp_prog);
1309 	if (xdp_prog) {
1310 		unsigned int xdp_frags_truesz = 0;
1311 		struct skb_shared_info *shinfo;
1312 		struct page *xdp_page;
1313 		struct xdp_buff xdp;
1314 		void *data;
1315 		u32 act;
1316 		int i;
1317 
1318 		data = mergeable_xdp_get_buf(vi, rq, xdp_prog, ctx, &frame_sz,
1319 					     &num_buf, &page, offset, &len, hdr);
1320 		if (unlikely(!data))
1321 			goto err_xdp;
1322 
1323 		err = virtnet_build_xdp_buff_mrg(dev, vi, rq, &xdp, data, len, frame_sz,
1324 						 &num_buf, &xdp_frags_truesz, stats);
1325 		if (unlikely(err))
1326 			goto err_xdp_frags;
1327 
1328 		act = virtnet_xdp_handler(xdp_prog, &xdp, dev, xdp_xmit, stats);
1329 
1330 		switch (act) {
1331 		case XDP_PASS:
1332 			head_skb = build_skb_from_xdp_buff(dev, vi, &xdp, xdp_frags_truesz);
1333 			if (unlikely(!head_skb))
1334 				goto err_xdp_frags;
1335 
1336 			rcu_read_unlock();
1337 			return head_skb;
1338 		case XDP_TX:
1339 		case XDP_REDIRECT:
1340 			rcu_read_unlock();
1341 			goto xdp_xmit;
1342 		default:
1343 			break;
1344 		}
1345 err_xdp_frags:
1346 		if (xdp_buff_has_frags(&xdp)) {
1347 			shinfo = xdp_get_shared_info_from_buff(&xdp);
1348 			for (i = 0; i < shinfo->nr_frags; i++) {
1349 				xdp_page = skb_frag_page(&shinfo->frags[i]);
1350 				put_page(xdp_page);
1351 			}
1352 		}
1353 
1354 		goto err_xdp;
1355 	}
1356 	rcu_read_unlock();
1357 
1358 skip_xdp:
1359 	head_skb = page_to_skb(vi, rq, page, offset, len, truesize, headroom);
1360 	curr_skb = head_skb;
1361 
1362 	if (unlikely(!curr_skb))
1363 		goto err_skb;
1364 	while (--num_buf) {
1365 		int num_skb_frags;
1366 
1367 		buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1368 		if (unlikely(!buf)) {
1369 			pr_debug("%s: rx error: %d buffers out of %d missing\n",
1370 				 dev->name, num_buf,
1371 				 virtio16_to_cpu(vi->vdev,
1372 						 hdr->num_buffers));
1373 			dev->stats.rx_length_errors++;
1374 			goto err_buf;
1375 		}
1376 
1377 		stats->bytes += len;
1378 		page = virt_to_head_page(buf);
1379 
1380 		truesize = mergeable_ctx_to_truesize(ctx);
1381 		headroom = mergeable_ctx_to_headroom(ctx);
1382 		tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1383 		room = SKB_DATA_ALIGN(headroom + tailroom);
1384 		if (unlikely(len > truesize - room)) {
1385 			pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1386 				 dev->name, len, (unsigned long)(truesize - room));
1387 			dev->stats.rx_length_errors++;
1388 			goto err_skb;
1389 		}
1390 
1391 		num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1392 		if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1393 			struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1394 
1395 			if (unlikely(!nskb))
1396 				goto err_skb;
1397 			if (curr_skb == head_skb)
1398 				skb_shinfo(curr_skb)->frag_list = nskb;
1399 			else
1400 				curr_skb->next = nskb;
1401 			curr_skb = nskb;
1402 			head_skb->truesize += nskb->truesize;
1403 			num_skb_frags = 0;
1404 		}
1405 		if (curr_skb != head_skb) {
1406 			head_skb->data_len += len;
1407 			head_skb->len += len;
1408 			head_skb->truesize += truesize;
1409 		}
1410 		offset = buf - page_address(page);
1411 		if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1412 			put_page(page);
1413 			skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1414 					     len, truesize);
1415 		} else {
1416 			skb_add_rx_frag(curr_skb, num_skb_frags, page,
1417 					offset, len, truesize);
1418 		}
1419 	}
1420 
1421 	ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1422 	return head_skb;
1423 
1424 err_xdp:
1425 	rcu_read_unlock();
1426 	stats->xdp_drops++;
1427 err_skb:
1428 	put_page(page);
1429 	while (num_buf-- > 1) {
1430 		buf = virtqueue_get_buf(rq->vq, &len);
1431 		if (unlikely(!buf)) {
1432 			pr_debug("%s: rx error: %d buffers missing\n",
1433 				 dev->name, num_buf);
1434 			dev->stats.rx_length_errors++;
1435 			break;
1436 		}
1437 		stats->bytes += len;
1438 		page = virt_to_head_page(buf);
1439 		put_page(page);
1440 	}
1441 err_buf:
1442 	stats->drops++;
1443 	dev_kfree_skb(head_skb);
1444 xdp_xmit:
1445 	return NULL;
1446 }
1447 
1448 static void virtio_skb_set_hash(const struct virtio_net_hdr_v1_hash *hdr_hash,
1449 				struct sk_buff *skb)
1450 {
1451 	enum pkt_hash_types rss_hash_type;
1452 
1453 	if (!hdr_hash || !skb)
1454 		return;
1455 
1456 	switch (__le16_to_cpu(hdr_hash->hash_report)) {
1457 	case VIRTIO_NET_HASH_REPORT_TCPv4:
1458 	case VIRTIO_NET_HASH_REPORT_UDPv4:
1459 	case VIRTIO_NET_HASH_REPORT_TCPv6:
1460 	case VIRTIO_NET_HASH_REPORT_UDPv6:
1461 	case VIRTIO_NET_HASH_REPORT_TCPv6_EX:
1462 	case VIRTIO_NET_HASH_REPORT_UDPv6_EX:
1463 		rss_hash_type = PKT_HASH_TYPE_L4;
1464 		break;
1465 	case VIRTIO_NET_HASH_REPORT_IPv4:
1466 	case VIRTIO_NET_HASH_REPORT_IPv6:
1467 	case VIRTIO_NET_HASH_REPORT_IPv6_EX:
1468 		rss_hash_type = PKT_HASH_TYPE_L3;
1469 		break;
1470 	case VIRTIO_NET_HASH_REPORT_NONE:
1471 	default:
1472 		rss_hash_type = PKT_HASH_TYPE_NONE;
1473 	}
1474 	skb_set_hash(skb, __le32_to_cpu(hdr_hash->hash_value), rss_hash_type);
1475 }
1476 
1477 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1478 			void *buf, unsigned int len, void **ctx,
1479 			unsigned int *xdp_xmit,
1480 			struct virtnet_rq_stats *stats)
1481 {
1482 	struct net_device *dev = vi->dev;
1483 	struct sk_buff *skb;
1484 	struct virtio_net_hdr_mrg_rxbuf *hdr;
1485 
1486 	if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1487 		pr_debug("%s: short packet %i\n", dev->name, len);
1488 		dev->stats.rx_length_errors++;
1489 		virtnet_rq_free_unused_buf(rq->vq, buf);
1490 		return;
1491 	}
1492 
1493 	if (vi->mergeable_rx_bufs)
1494 		skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1495 					stats);
1496 	else if (vi->big_packets)
1497 		skb = receive_big(dev, vi, rq, buf, len, stats);
1498 	else
1499 		skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1500 
1501 	if (unlikely(!skb))
1502 		return;
1503 
1504 	hdr = skb_vnet_hdr(skb);
1505 	if (dev->features & NETIF_F_RXHASH && vi->has_rss_hash_report)
1506 		virtio_skb_set_hash((const struct virtio_net_hdr_v1_hash *)hdr, skb);
1507 
1508 	if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1509 		skb->ip_summed = CHECKSUM_UNNECESSARY;
1510 
1511 	if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1512 				  virtio_is_little_endian(vi->vdev))) {
1513 		net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1514 				     dev->name, hdr->hdr.gso_type,
1515 				     hdr->hdr.gso_size);
1516 		goto frame_err;
1517 	}
1518 
1519 	skb_record_rx_queue(skb, vq2rxq(rq->vq));
1520 	skb->protocol = eth_type_trans(skb, dev);
1521 	pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1522 		 ntohs(skb->protocol), skb->len, skb->pkt_type);
1523 
1524 	napi_gro_receive(&rq->napi, skb);
1525 	return;
1526 
1527 frame_err:
1528 	dev->stats.rx_frame_errors++;
1529 	dev_kfree_skb(skb);
1530 }
1531 
1532 /* Unlike mergeable buffers, all buffers are allocated to the
1533  * same size, except for the headroom. For this reason we do
1534  * not need to use  mergeable_len_to_ctx here - it is enough
1535  * to store the headroom as the context ignoring the truesize.
1536  */
1537 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1538 			     gfp_t gfp)
1539 {
1540 	struct page_frag *alloc_frag = &rq->alloc_frag;
1541 	char *buf;
1542 	unsigned int xdp_headroom = virtnet_get_headroom(vi);
1543 	void *ctx = (void *)(unsigned long)xdp_headroom;
1544 	int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1545 	int err;
1546 
1547 	len = SKB_DATA_ALIGN(len) +
1548 	      SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1549 	if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1550 		return -ENOMEM;
1551 
1552 	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1553 	get_page(alloc_frag->page);
1554 	alloc_frag->offset += len;
1555 	sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1556 		    vi->hdr_len + GOOD_PACKET_LEN);
1557 	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1558 	if (err < 0)
1559 		put_page(virt_to_head_page(buf));
1560 	return err;
1561 }
1562 
1563 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1564 			   gfp_t gfp)
1565 {
1566 	struct page *first, *list = NULL;
1567 	char *p;
1568 	int i, err, offset;
1569 
1570 	sg_init_table(rq->sg, vi->big_packets_num_skbfrags + 2);
1571 
1572 	/* page in rq->sg[vi->big_packets_num_skbfrags + 1] is list tail */
1573 	for (i = vi->big_packets_num_skbfrags + 1; i > 1; --i) {
1574 		first = get_a_page(rq, gfp);
1575 		if (!first) {
1576 			if (list)
1577 				give_pages(rq, list);
1578 			return -ENOMEM;
1579 		}
1580 		sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1581 
1582 		/* chain new page in list head to match sg */
1583 		first->private = (unsigned long)list;
1584 		list = first;
1585 	}
1586 
1587 	first = get_a_page(rq, gfp);
1588 	if (!first) {
1589 		give_pages(rq, list);
1590 		return -ENOMEM;
1591 	}
1592 	p = page_address(first);
1593 
1594 	/* rq->sg[0], rq->sg[1] share the same page */
1595 	/* a separated rq->sg[0] for header - required in case !any_header_sg */
1596 	sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1597 
1598 	/* rq->sg[1] for data packet, from offset */
1599 	offset = sizeof(struct padded_vnet_hdr);
1600 	sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1601 
1602 	/* chain first in list head */
1603 	first->private = (unsigned long)list;
1604 	err = virtqueue_add_inbuf(rq->vq, rq->sg, vi->big_packets_num_skbfrags + 2,
1605 				  first, gfp);
1606 	if (err < 0)
1607 		give_pages(rq, first);
1608 
1609 	return err;
1610 }
1611 
1612 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1613 					  struct ewma_pkt_len *avg_pkt_len,
1614 					  unsigned int room)
1615 {
1616 	struct virtnet_info *vi = rq->vq->vdev->priv;
1617 	const size_t hdr_len = vi->hdr_len;
1618 	unsigned int len;
1619 
1620 	if (room)
1621 		return PAGE_SIZE - room;
1622 
1623 	len = hdr_len +	clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1624 				rq->min_buf_len, PAGE_SIZE - hdr_len);
1625 
1626 	return ALIGN(len, L1_CACHE_BYTES);
1627 }
1628 
1629 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1630 				 struct receive_queue *rq, gfp_t gfp)
1631 {
1632 	struct page_frag *alloc_frag = &rq->alloc_frag;
1633 	unsigned int headroom = virtnet_get_headroom(vi);
1634 	unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1635 	unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1636 	char *buf;
1637 	void *ctx;
1638 	int err;
1639 	unsigned int len, hole;
1640 
1641 	/* Extra tailroom is needed to satisfy XDP's assumption. This
1642 	 * means rx frags coalescing won't work, but consider we've
1643 	 * disabled GSO for XDP, it won't be a big issue.
1644 	 */
1645 	len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1646 	if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1647 		return -ENOMEM;
1648 
1649 	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1650 	buf += headroom; /* advance address leaving hole at front of pkt */
1651 	get_page(alloc_frag->page);
1652 	alloc_frag->offset += len + room;
1653 	hole = alloc_frag->size - alloc_frag->offset;
1654 	if (hole < len + room) {
1655 		/* To avoid internal fragmentation, if there is very likely not
1656 		 * enough space for another buffer, add the remaining space to
1657 		 * the current buffer.
1658 		 * XDP core assumes that frame_size of xdp_buff and the length
1659 		 * of the frag are PAGE_SIZE, so we disable the hole mechanism.
1660 		 */
1661 		if (!headroom)
1662 			len += hole;
1663 		alloc_frag->offset += hole;
1664 	}
1665 
1666 	sg_init_one(rq->sg, buf, len);
1667 	ctx = mergeable_len_to_ctx(len + room, headroom);
1668 	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1669 	if (err < 0)
1670 		put_page(virt_to_head_page(buf));
1671 
1672 	return err;
1673 }
1674 
1675 /*
1676  * Returns false if we couldn't fill entirely (OOM).
1677  *
1678  * Normally run in the receive path, but can also be run from ndo_open
1679  * before we're receiving packets, or from refill_work which is
1680  * careful to disable receiving (using napi_disable).
1681  */
1682 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1683 			  gfp_t gfp)
1684 {
1685 	int err;
1686 	bool oom;
1687 
1688 	do {
1689 		if (vi->mergeable_rx_bufs)
1690 			err = add_recvbuf_mergeable(vi, rq, gfp);
1691 		else if (vi->big_packets)
1692 			err = add_recvbuf_big(vi, rq, gfp);
1693 		else
1694 			err = add_recvbuf_small(vi, rq, gfp);
1695 
1696 		oom = err == -ENOMEM;
1697 		if (err)
1698 			break;
1699 	} while (rq->vq->num_free);
1700 	if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1701 		unsigned long flags;
1702 
1703 		flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1704 		rq->stats.kicks++;
1705 		u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1706 	}
1707 
1708 	return !oom;
1709 }
1710 
1711 static void skb_recv_done(struct virtqueue *rvq)
1712 {
1713 	struct virtnet_info *vi = rvq->vdev->priv;
1714 	struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1715 
1716 	virtqueue_napi_schedule(&rq->napi, rvq);
1717 }
1718 
1719 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1720 {
1721 	napi_enable(napi);
1722 
1723 	/* If all buffers were filled by other side before we napi_enabled, we
1724 	 * won't get another interrupt, so process any outstanding packets now.
1725 	 * Call local_bh_enable after to trigger softIRQ processing.
1726 	 */
1727 	local_bh_disable();
1728 	virtqueue_napi_schedule(napi, vq);
1729 	local_bh_enable();
1730 }
1731 
1732 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1733 				   struct virtqueue *vq,
1734 				   struct napi_struct *napi)
1735 {
1736 	if (!napi->weight)
1737 		return;
1738 
1739 	/* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1740 	 * enable the feature if this is likely affine with the transmit path.
1741 	 */
1742 	if (!vi->affinity_hint_set) {
1743 		napi->weight = 0;
1744 		return;
1745 	}
1746 
1747 	return virtnet_napi_enable(vq, napi);
1748 }
1749 
1750 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1751 {
1752 	if (napi->weight)
1753 		napi_disable(napi);
1754 }
1755 
1756 static void refill_work(struct work_struct *work)
1757 {
1758 	struct virtnet_info *vi =
1759 		container_of(work, struct virtnet_info, refill.work);
1760 	bool still_empty;
1761 	int i;
1762 
1763 	for (i = 0; i < vi->curr_queue_pairs; i++) {
1764 		struct receive_queue *rq = &vi->rq[i];
1765 
1766 		napi_disable(&rq->napi);
1767 		still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1768 		virtnet_napi_enable(rq->vq, &rq->napi);
1769 
1770 		/* In theory, this can happen: if we don't get any buffers in
1771 		 * we will *never* try to fill again.
1772 		 */
1773 		if (still_empty)
1774 			schedule_delayed_work(&vi->refill, HZ/2);
1775 	}
1776 }
1777 
1778 static int virtnet_receive(struct receive_queue *rq, int budget,
1779 			   unsigned int *xdp_xmit)
1780 {
1781 	struct virtnet_info *vi = rq->vq->vdev->priv;
1782 	struct virtnet_rq_stats stats = {};
1783 	unsigned int len;
1784 	void *buf;
1785 	int i;
1786 
1787 	if (!vi->big_packets || vi->mergeable_rx_bufs) {
1788 		void *ctx;
1789 
1790 		while (stats.packets < budget &&
1791 		       (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1792 			receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1793 			stats.packets++;
1794 		}
1795 	} else {
1796 		while (stats.packets < budget &&
1797 		       (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1798 			receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1799 			stats.packets++;
1800 		}
1801 	}
1802 
1803 	if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1804 		if (!try_fill_recv(vi, rq, GFP_ATOMIC)) {
1805 			spin_lock(&vi->refill_lock);
1806 			if (vi->refill_enabled)
1807 				schedule_delayed_work(&vi->refill, 0);
1808 			spin_unlock(&vi->refill_lock);
1809 		}
1810 	}
1811 
1812 	u64_stats_update_begin(&rq->stats.syncp);
1813 	for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1814 		size_t offset = virtnet_rq_stats_desc[i].offset;
1815 		u64 *item;
1816 
1817 		item = (u64 *)((u8 *)&rq->stats + offset);
1818 		*item += *(u64 *)((u8 *)&stats + offset);
1819 	}
1820 	u64_stats_update_end(&rq->stats.syncp);
1821 
1822 	return stats.packets;
1823 }
1824 
1825 static void virtnet_poll_cleantx(struct receive_queue *rq)
1826 {
1827 	struct virtnet_info *vi = rq->vq->vdev->priv;
1828 	unsigned int index = vq2rxq(rq->vq);
1829 	struct send_queue *sq = &vi->sq[index];
1830 	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1831 
1832 	if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1833 		return;
1834 
1835 	if (__netif_tx_trylock(txq)) {
1836 		if (sq->reset) {
1837 			__netif_tx_unlock(txq);
1838 			return;
1839 		}
1840 
1841 		do {
1842 			virtqueue_disable_cb(sq->vq);
1843 			free_old_xmit_skbs(sq, true);
1844 		} while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1845 
1846 		if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1847 			netif_tx_wake_queue(txq);
1848 
1849 		__netif_tx_unlock(txq);
1850 	}
1851 }
1852 
1853 static int virtnet_poll(struct napi_struct *napi, int budget)
1854 {
1855 	struct receive_queue *rq =
1856 		container_of(napi, struct receive_queue, napi);
1857 	struct virtnet_info *vi = rq->vq->vdev->priv;
1858 	struct send_queue *sq;
1859 	unsigned int received;
1860 	unsigned int xdp_xmit = 0;
1861 
1862 	virtnet_poll_cleantx(rq);
1863 
1864 	received = virtnet_receive(rq, budget, &xdp_xmit);
1865 
1866 	if (xdp_xmit & VIRTIO_XDP_REDIR)
1867 		xdp_do_flush();
1868 
1869 	/* Out of packets? */
1870 	if (received < budget)
1871 		virtqueue_napi_complete(napi, rq->vq, received);
1872 
1873 	if (xdp_xmit & VIRTIO_XDP_TX) {
1874 		sq = virtnet_xdp_get_sq(vi);
1875 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1876 			u64_stats_update_begin(&sq->stats.syncp);
1877 			sq->stats.kicks++;
1878 			u64_stats_update_end(&sq->stats.syncp);
1879 		}
1880 		virtnet_xdp_put_sq(vi, sq);
1881 	}
1882 
1883 	return received;
1884 }
1885 
1886 static int virtnet_open(struct net_device *dev)
1887 {
1888 	struct virtnet_info *vi = netdev_priv(dev);
1889 	int i, err;
1890 
1891 	enable_delayed_refill(vi);
1892 
1893 	for (i = 0; i < vi->max_queue_pairs; i++) {
1894 		if (i < vi->curr_queue_pairs)
1895 			/* Make sure we have some buffers: if oom use wq. */
1896 			if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1897 				schedule_delayed_work(&vi->refill, 0);
1898 
1899 		err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id);
1900 		if (err < 0)
1901 			return err;
1902 
1903 		err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1904 						 MEM_TYPE_PAGE_SHARED, NULL);
1905 		if (err < 0) {
1906 			xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1907 			return err;
1908 		}
1909 
1910 		virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1911 		virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1912 	}
1913 
1914 	return 0;
1915 }
1916 
1917 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1918 {
1919 	struct send_queue *sq = container_of(napi, struct send_queue, napi);
1920 	struct virtnet_info *vi = sq->vq->vdev->priv;
1921 	unsigned int index = vq2txq(sq->vq);
1922 	struct netdev_queue *txq;
1923 	int opaque;
1924 	bool done;
1925 
1926 	if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1927 		/* We don't need to enable cb for XDP */
1928 		napi_complete_done(napi, 0);
1929 		return 0;
1930 	}
1931 
1932 	txq = netdev_get_tx_queue(vi->dev, index);
1933 	__netif_tx_lock(txq, raw_smp_processor_id());
1934 	virtqueue_disable_cb(sq->vq);
1935 	free_old_xmit_skbs(sq, true);
1936 
1937 	if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1938 		netif_tx_wake_queue(txq);
1939 
1940 	opaque = virtqueue_enable_cb_prepare(sq->vq);
1941 
1942 	done = napi_complete_done(napi, 0);
1943 
1944 	if (!done)
1945 		virtqueue_disable_cb(sq->vq);
1946 
1947 	__netif_tx_unlock(txq);
1948 
1949 	if (done) {
1950 		if (unlikely(virtqueue_poll(sq->vq, opaque))) {
1951 			if (napi_schedule_prep(napi)) {
1952 				__netif_tx_lock(txq, raw_smp_processor_id());
1953 				virtqueue_disable_cb(sq->vq);
1954 				__netif_tx_unlock(txq);
1955 				__napi_schedule(napi);
1956 			}
1957 		}
1958 	}
1959 
1960 	return 0;
1961 }
1962 
1963 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1964 {
1965 	struct virtio_net_hdr_mrg_rxbuf *hdr;
1966 	const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1967 	struct virtnet_info *vi = sq->vq->vdev->priv;
1968 	int num_sg;
1969 	unsigned hdr_len = vi->hdr_len;
1970 	bool can_push;
1971 
1972 	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1973 
1974 	can_push = vi->any_header_sg &&
1975 		!((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1976 		!skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1977 	/* Even if we can, don't push here yet as this would skew
1978 	 * csum_start offset below. */
1979 	if (can_push)
1980 		hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1981 	else
1982 		hdr = skb_vnet_hdr(skb);
1983 
1984 	if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1985 				    virtio_is_little_endian(vi->vdev), false,
1986 				    0))
1987 		return -EPROTO;
1988 
1989 	if (vi->mergeable_rx_bufs)
1990 		hdr->num_buffers = 0;
1991 
1992 	sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1993 	if (can_push) {
1994 		__skb_push(skb, hdr_len);
1995 		num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1996 		if (unlikely(num_sg < 0))
1997 			return num_sg;
1998 		/* Pull header back to avoid skew in tx bytes calculations. */
1999 		__skb_pull(skb, hdr_len);
2000 	} else {
2001 		sg_set_buf(sq->sg, hdr, hdr_len);
2002 		num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
2003 		if (unlikely(num_sg < 0))
2004 			return num_sg;
2005 		num_sg++;
2006 	}
2007 	return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
2008 }
2009 
2010 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
2011 {
2012 	struct virtnet_info *vi = netdev_priv(dev);
2013 	int qnum = skb_get_queue_mapping(skb);
2014 	struct send_queue *sq = &vi->sq[qnum];
2015 	int err;
2016 	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
2017 	bool kick = !netdev_xmit_more();
2018 	bool use_napi = sq->napi.weight;
2019 
2020 	/* Free up any pending old buffers before queueing new ones. */
2021 	do {
2022 		if (use_napi)
2023 			virtqueue_disable_cb(sq->vq);
2024 
2025 		free_old_xmit_skbs(sq, false);
2026 
2027 	} while (use_napi && kick &&
2028 	       unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
2029 
2030 	/* timestamp packet in software */
2031 	skb_tx_timestamp(skb);
2032 
2033 	/* Try to transmit */
2034 	err = xmit_skb(sq, skb);
2035 
2036 	/* This should not happen! */
2037 	if (unlikely(err)) {
2038 		dev->stats.tx_fifo_errors++;
2039 		if (net_ratelimit())
2040 			dev_warn(&dev->dev,
2041 				 "Unexpected TXQ (%d) queue failure: %d\n",
2042 				 qnum, err);
2043 		dev->stats.tx_dropped++;
2044 		dev_kfree_skb_any(skb);
2045 		return NETDEV_TX_OK;
2046 	}
2047 
2048 	/* Don't wait up for transmitted skbs to be freed. */
2049 	if (!use_napi) {
2050 		skb_orphan(skb);
2051 		nf_reset_ct(skb);
2052 	}
2053 
2054 	check_sq_full_and_disable(vi, dev, sq);
2055 
2056 	if (kick || netif_xmit_stopped(txq)) {
2057 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
2058 			u64_stats_update_begin(&sq->stats.syncp);
2059 			sq->stats.kicks++;
2060 			u64_stats_update_end(&sq->stats.syncp);
2061 		}
2062 	}
2063 
2064 	return NETDEV_TX_OK;
2065 }
2066 
2067 static int virtnet_rx_resize(struct virtnet_info *vi,
2068 			     struct receive_queue *rq, u32 ring_num)
2069 {
2070 	bool running = netif_running(vi->dev);
2071 	int err, qindex;
2072 
2073 	qindex = rq - vi->rq;
2074 
2075 	if (running)
2076 		napi_disable(&rq->napi);
2077 
2078 	err = virtqueue_resize(rq->vq, ring_num, virtnet_rq_free_unused_buf);
2079 	if (err)
2080 		netdev_err(vi->dev, "resize rx fail: rx queue index: %d err: %d\n", qindex, err);
2081 
2082 	if (!try_fill_recv(vi, rq, GFP_KERNEL))
2083 		schedule_delayed_work(&vi->refill, 0);
2084 
2085 	if (running)
2086 		virtnet_napi_enable(rq->vq, &rq->napi);
2087 	return err;
2088 }
2089 
2090 static int virtnet_tx_resize(struct virtnet_info *vi,
2091 			     struct send_queue *sq, u32 ring_num)
2092 {
2093 	bool running = netif_running(vi->dev);
2094 	struct netdev_queue *txq;
2095 	int err, qindex;
2096 
2097 	qindex = sq - vi->sq;
2098 
2099 	if (running)
2100 		virtnet_napi_tx_disable(&sq->napi);
2101 
2102 	txq = netdev_get_tx_queue(vi->dev, qindex);
2103 
2104 	/* 1. wait all ximt complete
2105 	 * 2. fix the race of netif_stop_subqueue() vs netif_start_subqueue()
2106 	 */
2107 	__netif_tx_lock_bh(txq);
2108 
2109 	/* Prevent rx poll from accessing sq. */
2110 	sq->reset = true;
2111 
2112 	/* Prevent the upper layer from trying to send packets. */
2113 	netif_stop_subqueue(vi->dev, qindex);
2114 
2115 	__netif_tx_unlock_bh(txq);
2116 
2117 	err = virtqueue_resize(sq->vq, ring_num, virtnet_sq_free_unused_buf);
2118 	if (err)
2119 		netdev_err(vi->dev, "resize tx fail: tx queue index: %d err: %d\n", qindex, err);
2120 
2121 	__netif_tx_lock_bh(txq);
2122 	sq->reset = false;
2123 	netif_tx_wake_queue(txq);
2124 	__netif_tx_unlock_bh(txq);
2125 
2126 	if (running)
2127 		virtnet_napi_tx_enable(vi, sq->vq, &sq->napi);
2128 	return err;
2129 }
2130 
2131 /*
2132  * Send command via the control virtqueue and check status.  Commands
2133  * supported by the hypervisor, as indicated by feature bits, should
2134  * never fail unless improperly formatted.
2135  */
2136 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
2137 				 struct scatterlist *out)
2138 {
2139 	struct scatterlist *sgs[4], hdr, stat;
2140 	unsigned out_num = 0, tmp;
2141 	int ret;
2142 
2143 	/* Caller should know better */
2144 	BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
2145 
2146 	vi->ctrl->status = ~0;
2147 	vi->ctrl->hdr.class = class;
2148 	vi->ctrl->hdr.cmd = cmd;
2149 	/* Add header */
2150 	sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
2151 	sgs[out_num++] = &hdr;
2152 
2153 	if (out)
2154 		sgs[out_num++] = out;
2155 
2156 	/* Add return status. */
2157 	sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
2158 	sgs[out_num] = &stat;
2159 
2160 	BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
2161 	ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
2162 	if (ret < 0) {
2163 		dev_warn(&vi->vdev->dev,
2164 			 "Failed to add sgs for command vq: %d\n.", ret);
2165 		return false;
2166 	}
2167 
2168 	if (unlikely(!virtqueue_kick(vi->cvq)))
2169 		return vi->ctrl->status == VIRTIO_NET_OK;
2170 
2171 	/* Spin for a response, the kick causes an ioport write, trapping
2172 	 * into the hypervisor, so the request should be handled immediately.
2173 	 */
2174 	while (!virtqueue_get_buf(vi->cvq, &tmp) &&
2175 	       !virtqueue_is_broken(vi->cvq))
2176 		cpu_relax();
2177 
2178 	return vi->ctrl->status == VIRTIO_NET_OK;
2179 }
2180 
2181 static int virtnet_set_mac_address(struct net_device *dev, void *p)
2182 {
2183 	struct virtnet_info *vi = netdev_priv(dev);
2184 	struct virtio_device *vdev = vi->vdev;
2185 	int ret;
2186 	struct sockaddr *addr;
2187 	struct scatterlist sg;
2188 
2189 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2190 		return -EOPNOTSUPP;
2191 
2192 	addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
2193 	if (!addr)
2194 		return -ENOMEM;
2195 
2196 	ret = eth_prepare_mac_addr_change(dev, addr);
2197 	if (ret)
2198 		goto out;
2199 
2200 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
2201 		sg_init_one(&sg, addr->sa_data, dev->addr_len);
2202 		if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2203 					  VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
2204 			dev_warn(&vdev->dev,
2205 				 "Failed to set mac address by vq command.\n");
2206 			ret = -EINVAL;
2207 			goto out;
2208 		}
2209 	} else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
2210 		   !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
2211 		unsigned int i;
2212 
2213 		/* Naturally, this has an atomicity problem. */
2214 		for (i = 0; i < dev->addr_len; i++)
2215 			virtio_cwrite8(vdev,
2216 				       offsetof(struct virtio_net_config, mac) +
2217 				       i, addr->sa_data[i]);
2218 	}
2219 
2220 	eth_commit_mac_addr_change(dev, p);
2221 	ret = 0;
2222 
2223 out:
2224 	kfree(addr);
2225 	return ret;
2226 }
2227 
2228 static void virtnet_stats(struct net_device *dev,
2229 			  struct rtnl_link_stats64 *tot)
2230 {
2231 	struct virtnet_info *vi = netdev_priv(dev);
2232 	unsigned int start;
2233 	int i;
2234 
2235 	for (i = 0; i < vi->max_queue_pairs; i++) {
2236 		u64 tpackets, tbytes, terrors, rpackets, rbytes, rdrops;
2237 		struct receive_queue *rq = &vi->rq[i];
2238 		struct send_queue *sq = &vi->sq[i];
2239 
2240 		do {
2241 			start = u64_stats_fetch_begin(&sq->stats.syncp);
2242 			tpackets = sq->stats.packets;
2243 			tbytes   = sq->stats.bytes;
2244 			terrors  = sq->stats.tx_timeouts;
2245 		} while (u64_stats_fetch_retry(&sq->stats.syncp, start));
2246 
2247 		do {
2248 			start = u64_stats_fetch_begin(&rq->stats.syncp);
2249 			rpackets = rq->stats.packets;
2250 			rbytes   = rq->stats.bytes;
2251 			rdrops   = rq->stats.drops;
2252 		} while (u64_stats_fetch_retry(&rq->stats.syncp, start));
2253 
2254 		tot->rx_packets += rpackets;
2255 		tot->tx_packets += tpackets;
2256 		tot->rx_bytes   += rbytes;
2257 		tot->tx_bytes   += tbytes;
2258 		tot->rx_dropped += rdrops;
2259 		tot->tx_errors  += terrors;
2260 	}
2261 
2262 	tot->tx_dropped = dev->stats.tx_dropped;
2263 	tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
2264 	tot->rx_length_errors = dev->stats.rx_length_errors;
2265 	tot->rx_frame_errors = dev->stats.rx_frame_errors;
2266 }
2267 
2268 static void virtnet_ack_link_announce(struct virtnet_info *vi)
2269 {
2270 	rtnl_lock();
2271 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
2272 				  VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
2273 		dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
2274 	rtnl_unlock();
2275 }
2276 
2277 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
2278 {
2279 	struct scatterlist sg;
2280 	struct net_device *dev = vi->dev;
2281 
2282 	if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
2283 		return 0;
2284 
2285 	vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
2286 	sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
2287 
2288 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
2289 				  VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
2290 		dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
2291 			 queue_pairs);
2292 		return -EINVAL;
2293 	} else {
2294 		vi->curr_queue_pairs = queue_pairs;
2295 		/* virtnet_open() will refill when device is going to up. */
2296 		if (dev->flags & IFF_UP)
2297 			schedule_delayed_work(&vi->refill, 0);
2298 	}
2299 
2300 	return 0;
2301 }
2302 
2303 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
2304 {
2305 	int err;
2306 
2307 	rtnl_lock();
2308 	err = _virtnet_set_queues(vi, queue_pairs);
2309 	rtnl_unlock();
2310 	return err;
2311 }
2312 
2313 static int virtnet_close(struct net_device *dev)
2314 {
2315 	struct virtnet_info *vi = netdev_priv(dev);
2316 	int i;
2317 
2318 	/* Make sure NAPI doesn't schedule refill work */
2319 	disable_delayed_refill(vi);
2320 	/* Make sure refill_work doesn't re-enable napi! */
2321 	cancel_delayed_work_sync(&vi->refill);
2322 
2323 	for (i = 0; i < vi->max_queue_pairs; i++) {
2324 		virtnet_napi_tx_disable(&vi->sq[i].napi);
2325 		napi_disable(&vi->rq[i].napi);
2326 		xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
2327 	}
2328 
2329 	return 0;
2330 }
2331 
2332 static void virtnet_set_rx_mode(struct net_device *dev)
2333 {
2334 	struct virtnet_info *vi = netdev_priv(dev);
2335 	struct scatterlist sg[2];
2336 	struct virtio_net_ctrl_mac *mac_data;
2337 	struct netdev_hw_addr *ha;
2338 	int uc_count;
2339 	int mc_count;
2340 	void *buf;
2341 	int i;
2342 
2343 	/* We can't dynamically set ndo_set_rx_mode, so return gracefully */
2344 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
2345 		return;
2346 
2347 	vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
2348 	vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
2349 
2350 	sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
2351 
2352 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2353 				  VIRTIO_NET_CTRL_RX_PROMISC, sg))
2354 		dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
2355 			 vi->ctrl->promisc ? "en" : "dis");
2356 
2357 	sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
2358 
2359 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2360 				  VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
2361 		dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
2362 			 vi->ctrl->allmulti ? "en" : "dis");
2363 
2364 	uc_count = netdev_uc_count(dev);
2365 	mc_count = netdev_mc_count(dev);
2366 	/* MAC filter - use one buffer for both lists */
2367 	buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
2368 		      (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
2369 	mac_data = buf;
2370 	if (!buf)
2371 		return;
2372 
2373 	sg_init_table(sg, 2);
2374 
2375 	/* Store the unicast list and count in the front of the buffer */
2376 	mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
2377 	i = 0;
2378 	netdev_for_each_uc_addr(ha, dev)
2379 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2380 
2381 	sg_set_buf(&sg[0], mac_data,
2382 		   sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
2383 
2384 	/* multicast list and count fill the end */
2385 	mac_data = (void *)&mac_data->macs[uc_count][0];
2386 
2387 	mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
2388 	i = 0;
2389 	netdev_for_each_mc_addr(ha, dev)
2390 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2391 
2392 	sg_set_buf(&sg[1], mac_data,
2393 		   sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
2394 
2395 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2396 				  VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
2397 		dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
2398 
2399 	kfree(buf);
2400 }
2401 
2402 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
2403 				   __be16 proto, u16 vid)
2404 {
2405 	struct virtnet_info *vi = netdev_priv(dev);
2406 	struct scatterlist sg;
2407 
2408 	vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2409 	sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2410 
2411 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2412 				  VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2413 		dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2414 	return 0;
2415 }
2416 
2417 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2418 				    __be16 proto, u16 vid)
2419 {
2420 	struct virtnet_info *vi = netdev_priv(dev);
2421 	struct scatterlist sg;
2422 
2423 	vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2424 	sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2425 
2426 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2427 				  VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2428 		dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2429 	return 0;
2430 }
2431 
2432 static void virtnet_clean_affinity(struct virtnet_info *vi)
2433 {
2434 	int i;
2435 
2436 	if (vi->affinity_hint_set) {
2437 		for (i = 0; i < vi->max_queue_pairs; i++) {
2438 			virtqueue_set_affinity(vi->rq[i].vq, NULL);
2439 			virtqueue_set_affinity(vi->sq[i].vq, NULL);
2440 		}
2441 
2442 		vi->affinity_hint_set = false;
2443 	}
2444 }
2445 
2446 static void virtnet_set_affinity(struct virtnet_info *vi)
2447 {
2448 	cpumask_var_t mask;
2449 	int stragglers;
2450 	int group_size;
2451 	int i, j, cpu;
2452 	int num_cpu;
2453 	int stride;
2454 
2455 	if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2456 		virtnet_clean_affinity(vi);
2457 		return;
2458 	}
2459 
2460 	num_cpu = num_online_cpus();
2461 	stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2462 	stragglers = num_cpu >= vi->curr_queue_pairs ?
2463 			num_cpu % vi->curr_queue_pairs :
2464 			0;
2465 	cpu = cpumask_first(cpu_online_mask);
2466 
2467 	for (i = 0; i < vi->curr_queue_pairs; i++) {
2468 		group_size = stride + (i < stragglers ? 1 : 0);
2469 
2470 		for (j = 0; j < group_size; j++) {
2471 			cpumask_set_cpu(cpu, mask);
2472 			cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2473 						nr_cpu_ids, false);
2474 		}
2475 		virtqueue_set_affinity(vi->rq[i].vq, mask);
2476 		virtqueue_set_affinity(vi->sq[i].vq, mask);
2477 		__netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2478 		cpumask_clear(mask);
2479 	}
2480 
2481 	vi->affinity_hint_set = true;
2482 	free_cpumask_var(mask);
2483 }
2484 
2485 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2486 {
2487 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2488 						   node);
2489 	virtnet_set_affinity(vi);
2490 	return 0;
2491 }
2492 
2493 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2494 {
2495 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2496 						   node_dead);
2497 	virtnet_set_affinity(vi);
2498 	return 0;
2499 }
2500 
2501 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2502 {
2503 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2504 						   node);
2505 
2506 	virtnet_clean_affinity(vi);
2507 	return 0;
2508 }
2509 
2510 static enum cpuhp_state virtionet_online;
2511 
2512 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2513 {
2514 	int ret;
2515 
2516 	ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2517 	if (ret)
2518 		return ret;
2519 	ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2520 					       &vi->node_dead);
2521 	if (!ret)
2522 		return ret;
2523 	cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2524 	return ret;
2525 }
2526 
2527 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2528 {
2529 	cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2530 	cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2531 					    &vi->node_dead);
2532 }
2533 
2534 static void virtnet_get_ringparam(struct net_device *dev,
2535 				  struct ethtool_ringparam *ring,
2536 				  struct kernel_ethtool_ringparam *kernel_ring,
2537 				  struct netlink_ext_ack *extack)
2538 {
2539 	struct virtnet_info *vi = netdev_priv(dev);
2540 
2541 	ring->rx_max_pending = vi->rq[0].vq->num_max;
2542 	ring->tx_max_pending = vi->sq[0].vq->num_max;
2543 	ring->rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2544 	ring->tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2545 }
2546 
2547 static int virtnet_set_ringparam(struct net_device *dev,
2548 				 struct ethtool_ringparam *ring,
2549 				 struct kernel_ethtool_ringparam *kernel_ring,
2550 				 struct netlink_ext_ack *extack)
2551 {
2552 	struct virtnet_info *vi = netdev_priv(dev);
2553 	u32 rx_pending, tx_pending;
2554 	struct receive_queue *rq;
2555 	struct send_queue *sq;
2556 	int i, err;
2557 
2558 	if (ring->rx_mini_pending || ring->rx_jumbo_pending)
2559 		return -EINVAL;
2560 
2561 	rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2562 	tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2563 
2564 	if (ring->rx_pending == rx_pending &&
2565 	    ring->tx_pending == tx_pending)
2566 		return 0;
2567 
2568 	if (ring->rx_pending > vi->rq[0].vq->num_max)
2569 		return -EINVAL;
2570 
2571 	if (ring->tx_pending > vi->sq[0].vq->num_max)
2572 		return -EINVAL;
2573 
2574 	for (i = 0; i < vi->max_queue_pairs; i++) {
2575 		rq = vi->rq + i;
2576 		sq = vi->sq + i;
2577 
2578 		if (ring->tx_pending != tx_pending) {
2579 			err = virtnet_tx_resize(vi, sq, ring->tx_pending);
2580 			if (err)
2581 				return err;
2582 		}
2583 
2584 		if (ring->rx_pending != rx_pending) {
2585 			err = virtnet_rx_resize(vi, rq, ring->rx_pending);
2586 			if (err)
2587 				return err;
2588 		}
2589 	}
2590 
2591 	return 0;
2592 }
2593 
2594 static bool virtnet_commit_rss_command(struct virtnet_info *vi)
2595 {
2596 	struct net_device *dev = vi->dev;
2597 	struct scatterlist sgs[4];
2598 	unsigned int sg_buf_size;
2599 
2600 	/* prepare sgs */
2601 	sg_init_table(sgs, 4);
2602 
2603 	sg_buf_size = offsetof(struct virtio_net_ctrl_rss, indirection_table);
2604 	sg_set_buf(&sgs[0], &vi->ctrl->rss, sg_buf_size);
2605 
2606 	sg_buf_size = sizeof(uint16_t) * (vi->ctrl->rss.indirection_table_mask + 1);
2607 	sg_set_buf(&sgs[1], vi->ctrl->rss.indirection_table, sg_buf_size);
2608 
2609 	sg_buf_size = offsetof(struct virtio_net_ctrl_rss, key)
2610 			- offsetof(struct virtio_net_ctrl_rss, max_tx_vq);
2611 	sg_set_buf(&sgs[2], &vi->ctrl->rss.max_tx_vq, sg_buf_size);
2612 
2613 	sg_buf_size = vi->rss_key_size;
2614 	sg_set_buf(&sgs[3], vi->ctrl->rss.key, sg_buf_size);
2615 
2616 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
2617 				  vi->has_rss ? VIRTIO_NET_CTRL_MQ_RSS_CONFIG
2618 				  : VIRTIO_NET_CTRL_MQ_HASH_CONFIG, sgs)) {
2619 		dev_warn(&dev->dev, "VIRTIONET issue with committing RSS sgs\n");
2620 		return false;
2621 	}
2622 	return true;
2623 }
2624 
2625 static void virtnet_init_default_rss(struct virtnet_info *vi)
2626 {
2627 	u32 indir_val = 0;
2628 	int i = 0;
2629 
2630 	vi->ctrl->rss.hash_types = vi->rss_hash_types_supported;
2631 	vi->rss_hash_types_saved = vi->rss_hash_types_supported;
2632 	vi->ctrl->rss.indirection_table_mask = vi->rss_indir_table_size
2633 						? vi->rss_indir_table_size - 1 : 0;
2634 	vi->ctrl->rss.unclassified_queue = 0;
2635 
2636 	for (; i < vi->rss_indir_table_size; ++i) {
2637 		indir_val = ethtool_rxfh_indir_default(i, vi->curr_queue_pairs);
2638 		vi->ctrl->rss.indirection_table[i] = indir_val;
2639 	}
2640 
2641 	vi->ctrl->rss.max_tx_vq = vi->curr_queue_pairs;
2642 	vi->ctrl->rss.hash_key_length = vi->rss_key_size;
2643 
2644 	netdev_rss_key_fill(vi->ctrl->rss.key, vi->rss_key_size);
2645 }
2646 
2647 static void virtnet_get_hashflow(const struct virtnet_info *vi, struct ethtool_rxnfc *info)
2648 {
2649 	info->data = 0;
2650 	switch (info->flow_type) {
2651 	case TCP_V4_FLOW:
2652 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv4) {
2653 			info->data = RXH_IP_SRC | RXH_IP_DST |
2654 						 RXH_L4_B_0_1 | RXH_L4_B_2_3;
2655 		} else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
2656 			info->data = RXH_IP_SRC | RXH_IP_DST;
2657 		}
2658 		break;
2659 	case TCP_V6_FLOW:
2660 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv6) {
2661 			info->data = RXH_IP_SRC | RXH_IP_DST |
2662 						 RXH_L4_B_0_1 | RXH_L4_B_2_3;
2663 		} else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
2664 			info->data = RXH_IP_SRC | RXH_IP_DST;
2665 		}
2666 		break;
2667 	case UDP_V4_FLOW:
2668 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv4) {
2669 			info->data = RXH_IP_SRC | RXH_IP_DST |
2670 						 RXH_L4_B_0_1 | RXH_L4_B_2_3;
2671 		} else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
2672 			info->data = RXH_IP_SRC | RXH_IP_DST;
2673 		}
2674 		break;
2675 	case UDP_V6_FLOW:
2676 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv6) {
2677 			info->data = RXH_IP_SRC | RXH_IP_DST |
2678 						 RXH_L4_B_0_1 | RXH_L4_B_2_3;
2679 		} else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
2680 			info->data = RXH_IP_SRC | RXH_IP_DST;
2681 		}
2682 		break;
2683 	case IPV4_FLOW:
2684 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4)
2685 			info->data = RXH_IP_SRC | RXH_IP_DST;
2686 
2687 		break;
2688 	case IPV6_FLOW:
2689 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6)
2690 			info->data = RXH_IP_SRC | RXH_IP_DST;
2691 
2692 		break;
2693 	default:
2694 		info->data = 0;
2695 		break;
2696 	}
2697 }
2698 
2699 static bool virtnet_set_hashflow(struct virtnet_info *vi, struct ethtool_rxnfc *info)
2700 {
2701 	u32 new_hashtypes = vi->rss_hash_types_saved;
2702 	bool is_disable = info->data & RXH_DISCARD;
2703 	bool is_l4 = info->data == (RXH_IP_SRC | RXH_IP_DST | RXH_L4_B_0_1 | RXH_L4_B_2_3);
2704 
2705 	/* supports only 'sd', 'sdfn' and 'r' */
2706 	if (!((info->data == (RXH_IP_SRC | RXH_IP_DST)) | is_l4 | is_disable))
2707 		return false;
2708 
2709 	switch (info->flow_type) {
2710 	case TCP_V4_FLOW:
2711 		new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_TCPv4);
2712 		if (!is_disable)
2713 			new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
2714 				| (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv4 : 0);
2715 		break;
2716 	case UDP_V4_FLOW:
2717 		new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_UDPv4);
2718 		if (!is_disable)
2719 			new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
2720 				| (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv4 : 0);
2721 		break;
2722 	case IPV4_FLOW:
2723 		new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv4;
2724 		if (!is_disable)
2725 			new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv4;
2726 		break;
2727 	case TCP_V6_FLOW:
2728 		new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_TCPv6);
2729 		if (!is_disable)
2730 			new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
2731 				| (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv6 : 0);
2732 		break;
2733 	case UDP_V6_FLOW:
2734 		new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_UDPv6);
2735 		if (!is_disable)
2736 			new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
2737 				| (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv6 : 0);
2738 		break;
2739 	case IPV6_FLOW:
2740 		new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv6;
2741 		if (!is_disable)
2742 			new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv6;
2743 		break;
2744 	default:
2745 		/* unsupported flow */
2746 		return false;
2747 	}
2748 
2749 	/* if unsupported hashtype was set */
2750 	if (new_hashtypes != (new_hashtypes & vi->rss_hash_types_supported))
2751 		return false;
2752 
2753 	if (new_hashtypes != vi->rss_hash_types_saved) {
2754 		vi->rss_hash_types_saved = new_hashtypes;
2755 		vi->ctrl->rss.hash_types = vi->rss_hash_types_saved;
2756 		if (vi->dev->features & NETIF_F_RXHASH)
2757 			return virtnet_commit_rss_command(vi);
2758 	}
2759 
2760 	return true;
2761 }
2762 
2763 static void virtnet_get_drvinfo(struct net_device *dev,
2764 				struct ethtool_drvinfo *info)
2765 {
2766 	struct virtnet_info *vi = netdev_priv(dev);
2767 	struct virtio_device *vdev = vi->vdev;
2768 
2769 	strscpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2770 	strscpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2771 	strscpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2772 
2773 }
2774 
2775 /* TODO: Eliminate OOO packets during switching */
2776 static int virtnet_set_channels(struct net_device *dev,
2777 				struct ethtool_channels *channels)
2778 {
2779 	struct virtnet_info *vi = netdev_priv(dev);
2780 	u16 queue_pairs = channels->combined_count;
2781 	int err;
2782 
2783 	/* We don't support separate rx/tx channels.
2784 	 * We don't allow setting 'other' channels.
2785 	 */
2786 	if (channels->rx_count || channels->tx_count || channels->other_count)
2787 		return -EINVAL;
2788 
2789 	if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2790 		return -EINVAL;
2791 
2792 	/* For now we don't support modifying channels while XDP is loaded
2793 	 * also when XDP is loaded all RX queues have XDP programs so we only
2794 	 * need to check a single RX queue.
2795 	 */
2796 	if (vi->rq[0].xdp_prog)
2797 		return -EINVAL;
2798 
2799 	cpus_read_lock();
2800 	err = _virtnet_set_queues(vi, queue_pairs);
2801 	if (err) {
2802 		cpus_read_unlock();
2803 		goto err;
2804 	}
2805 	virtnet_set_affinity(vi);
2806 	cpus_read_unlock();
2807 
2808 	netif_set_real_num_tx_queues(dev, queue_pairs);
2809 	netif_set_real_num_rx_queues(dev, queue_pairs);
2810  err:
2811 	return err;
2812 }
2813 
2814 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2815 {
2816 	struct virtnet_info *vi = netdev_priv(dev);
2817 	unsigned int i, j;
2818 	u8 *p = data;
2819 
2820 	switch (stringset) {
2821 	case ETH_SS_STATS:
2822 		for (i = 0; i < vi->curr_queue_pairs; i++) {
2823 			for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
2824 				ethtool_sprintf(&p, "rx_queue_%u_%s", i,
2825 						virtnet_rq_stats_desc[j].desc);
2826 		}
2827 
2828 		for (i = 0; i < vi->curr_queue_pairs; i++) {
2829 			for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
2830 				ethtool_sprintf(&p, "tx_queue_%u_%s", i,
2831 						virtnet_sq_stats_desc[j].desc);
2832 		}
2833 		break;
2834 	}
2835 }
2836 
2837 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2838 {
2839 	struct virtnet_info *vi = netdev_priv(dev);
2840 
2841 	switch (sset) {
2842 	case ETH_SS_STATS:
2843 		return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2844 					       VIRTNET_SQ_STATS_LEN);
2845 	default:
2846 		return -EOPNOTSUPP;
2847 	}
2848 }
2849 
2850 static void virtnet_get_ethtool_stats(struct net_device *dev,
2851 				      struct ethtool_stats *stats, u64 *data)
2852 {
2853 	struct virtnet_info *vi = netdev_priv(dev);
2854 	unsigned int idx = 0, start, i, j;
2855 	const u8 *stats_base;
2856 	size_t offset;
2857 
2858 	for (i = 0; i < vi->curr_queue_pairs; i++) {
2859 		struct receive_queue *rq = &vi->rq[i];
2860 
2861 		stats_base = (u8 *)&rq->stats;
2862 		do {
2863 			start = u64_stats_fetch_begin(&rq->stats.syncp);
2864 			for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2865 				offset = virtnet_rq_stats_desc[j].offset;
2866 				data[idx + j] = *(u64 *)(stats_base + offset);
2867 			}
2868 		} while (u64_stats_fetch_retry(&rq->stats.syncp, start));
2869 		idx += VIRTNET_RQ_STATS_LEN;
2870 	}
2871 
2872 	for (i = 0; i < vi->curr_queue_pairs; i++) {
2873 		struct send_queue *sq = &vi->sq[i];
2874 
2875 		stats_base = (u8 *)&sq->stats;
2876 		do {
2877 			start = u64_stats_fetch_begin(&sq->stats.syncp);
2878 			for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2879 				offset = virtnet_sq_stats_desc[j].offset;
2880 				data[idx + j] = *(u64 *)(stats_base + offset);
2881 			}
2882 		} while (u64_stats_fetch_retry(&sq->stats.syncp, start));
2883 		idx += VIRTNET_SQ_STATS_LEN;
2884 	}
2885 }
2886 
2887 static void virtnet_get_channels(struct net_device *dev,
2888 				 struct ethtool_channels *channels)
2889 {
2890 	struct virtnet_info *vi = netdev_priv(dev);
2891 
2892 	channels->combined_count = vi->curr_queue_pairs;
2893 	channels->max_combined = vi->max_queue_pairs;
2894 	channels->max_other = 0;
2895 	channels->rx_count = 0;
2896 	channels->tx_count = 0;
2897 	channels->other_count = 0;
2898 }
2899 
2900 static int virtnet_set_link_ksettings(struct net_device *dev,
2901 				      const struct ethtool_link_ksettings *cmd)
2902 {
2903 	struct virtnet_info *vi = netdev_priv(dev);
2904 
2905 	return ethtool_virtdev_set_link_ksettings(dev, cmd,
2906 						  &vi->speed, &vi->duplex);
2907 }
2908 
2909 static int virtnet_get_link_ksettings(struct net_device *dev,
2910 				      struct ethtool_link_ksettings *cmd)
2911 {
2912 	struct virtnet_info *vi = netdev_priv(dev);
2913 
2914 	cmd->base.speed = vi->speed;
2915 	cmd->base.duplex = vi->duplex;
2916 	cmd->base.port = PORT_OTHER;
2917 
2918 	return 0;
2919 }
2920 
2921 static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
2922 				       struct ethtool_coalesce *ec)
2923 {
2924 	struct scatterlist sgs_tx, sgs_rx;
2925 	struct virtio_net_ctrl_coal_tx coal_tx;
2926 	struct virtio_net_ctrl_coal_rx coal_rx;
2927 
2928 	coal_tx.tx_usecs = cpu_to_le32(ec->tx_coalesce_usecs);
2929 	coal_tx.tx_max_packets = cpu_to_le32(ec->tx_max_coalesced_frames);
2930 	sg_init_one(&sgs_tx, &coal_tx, sizeof(coal_tx));
2931 
2932 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
2933 				  VIRTIO_NET_CTRL_NOTF_COAL_TX_SET,
2934 				  &sgs_tx))
2935 		return -EINVAL;
2936 
2937 	/* Save parameters */
2938 	vi->tx_usecs = ec->tx_coalesce_usecs;
2939 	vi->tx_max_packets = ec->tx_max_coalesced_frames;
2940 
2941 	coal_rx.rx_usecs = cpu_to_le32(ec->rx_coalesce_usecs);
2942 	coal_rx.rx_max_packets = cpu_to_le32(ec->rx_max_coalesced_frames);
2943 	sg_init_one(&sgs_rx, &coal_rx, sizeof(coal_rx));
2944 
2945 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
2946 				  VIRTIO_NET_CTRL_NOTF_COAL_RX_SET,
2947 				  &sgs_rx))
2948 		return -EINVAL;
2949 
2950 	/* Save parameters */
2951 	vi->rx_usecs = ec->rx_coalesce_usecs;
2952 	vi->rx_max_packets = ec->rx_max_coalesced_frames;
2953 
2954 	return 0;
2955 }
2956 
2957 static int virtnet_coal_params_supported(struct ethtool_coalesce *ec)
2958 {
2959 	/* usecs coalescing is supported only if VIRTIO_NET_F_NOTF_COAL
2960 	 * feature is negotiated.
2961 	 */
2962 	if (ec->rx_coalesce_usecs || ec->tx_coalesce_usecs)
2963 		return -EOPNOTSUPP;
2964 
2965 	if (ec->tx_max_coalesced_frames > 1 ||
2966 	    ec->rx_max_coalesced_frames != 1)
2967 		return -EINVAL;
2968 
2969 	return 0;
2970 }
2971 
2972 static int virtnet_set_coalesce(struct net_device *dev,
2973 				struct ethtool_coalesce *ec,
2974 				struct kernel_ethtool_coalesce *kernel_coal,
2975 				struct netlink_ext_ack *extack)
2976 {
2977 	struct virtnet_info *vi = netdev_priv(dev);
2978 	int ret, i, napi_weight;
2979 	bool update_napi = false;
2980 
2981 	/* Can't change NAPI weight if the link is up */
2982 	napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2983 	if (napi_weight ^ vi->sq[0].napi.weight) {
2984 		if (dev->flags & IFF_UP)
2985 			return -EBUSY;
2986 		else
2987 			update_napi = true;
2988 	}
2989 
2990 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL))
2991 		ret = virtnet_send_notf_coal_cmds(vi, ec);
2992 	else
2993 		ret = virtnet_coal_params_supported(ec);
2994 
2995 	if (ret)
2996 		return ret;
2997 
2998 	if (update_napi) {
2999 		for (i = 0; i < vi->max_queue_pairs; i++)
3000 			vi->sq[i].napi.weight = napi_weight;
3001 	}
3002 
3003 	return ret;
3004 }
3005 
3006 static int virtnet_get_coalesce(struct net_device *dev,
3007 				struct ethtool_coalesce *ec,
3008 				struct kernel_ethtool_coalesce *kernel_coal,
3009 				struct netlink_ext_ack *extack)
3010 {
3011 	struct virtnet_info *vi = netdev_priv(dev);
3012 
3013 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
3014 		ec->rx_coalesce_usecs = vi->rx_usecs;
3015 		ec->tx_coalesce_usecs = vi->tx_usecs;
3016 		ec->tx_max_coalesced_frames = vi->tx_max_packets;
3017 		ec->rx_max_coalesced_frames = vi->rx_max_packets;
3018 	} else {
3019 		ec->rx_max_coalesced_frames = 1;
3020 
3021 		if (vi->sq[0].napi.weight)
3022 			ec->tx_max_coalesced_frames = 1;
3023 	}
3024 
3025 	return 0;
3026 }
3027 
3028 static void virtnet_init_settings(struct net_device *dev)
3029 {
3030 	struct virtnet_info *vi = netdev_priv(dev);
3031 
3032 	vi->speed = SPEED_UNKNOWN;
3033 	vi->duplex = DUPLEX_UNKNOWN;
3034 }
3035 
3036 static void virtnet_update_settings(struct virtnet_info *vi)
3037 {
3038 	u32 speed;
3039 	u8 duplex;
3040 
3041 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
3042 		return;
3043 
3044 	virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
3045 
3046 	if (ethtool_validate_speed(speed))
3047 		vi->speed = speed;
3048 
3049 	virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
3050 
3051 	if (ethtool_validate_duplex(duplex))
3052 		vi->duplex = duplex;
3053 }
3054 
3055 static u32 virtnet_get_rxfh_key_size(struct net_device *dev)
3056 {
3057 	return ((struct virtnet_info *)netdev_priv(dev))->rss_key_size;
3058 }
3059 
3060 static u32 virtnet_get_rxfh_indir_size(struct net_device *dev)
3061 {
3062 	return ((struct virtnet_info *)netdev_priv(dev))->rss_indir_table_size;
3063 }
3064 
3065 static int virtnet_get_rxfh(struct net_device *dev, u32 *indir, u8 *key, u8 *hfunc)
3066 {
3067 	struct virtnet_info *vi = netdev_priv(dev);
3068 	int i;
3069 
3070 	if (indir) {
3071 		for (i = 0; i < vi->rss_indir_table_size; ++i)
3072 			indir[i] = vi->ctrl->rss.indirection_table[i];
3073 	}
3074 
3075 	if (key)
3076 		memcpy(key, vi->ctrl->rss.key, vi->rss_key_size);
3077 
3078 	if (hfunc)
3079 		*hfunc = ETH_RSS_HASH_TOP;
3080 
3081 	return 0;
3082 }
3083 
3084 static int virtnet_set_rxfh(struct net_device *dev, const u32 *indir, const u8 *key, const u8 hfunc)
3085 {
3086 	struct virtnet_info *vi = netdev_priv(dev);
3087 	int i;
3088 
3089 	if (hfunc != ETH_RSS_HASH_NO_CHANGE && hfunc != ETH_RSS_HASH_TOP)
3090 		return -EOPNOTSUPP;
3091 
3092 	if (indir) {
3093 		for (i = 0; i < vi->rss_indir_table_size; ++i)
3094 			vi->ctrl->rss.indirection_table[i] = indir[i];
3095 	}
3096 	if (key)
3097 		memcpy(vi->ctrl->rss.key, key, vi->rss_key_size);
3098 
3099 	virtnet_commit_rss_command(vi);
3100 
3101 	return 0;
3102 }
3103 
3104 static int virtnet_get_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info, u32 *rule_locs)
3105 {
3106 	struct virtnet_info *vi = netdev_priv(dev);
3107 	int rc = 0;
3108 
3109 	switch (info->cmd) {
3110 	case ETHTOOL_GRXRINGS:
3111 		info->data = vi->curr_queue_pairs;
3112 		break;
3113 	case ETHTOOL_GRXFH:
3114 		virtnet_get_hashflow(vi, info);
3115 		break;
3116 	default:
3117 		rc = -EOPNOTSUPP;
3118 	}
3119 
3120 	return rc;
3121 }
3122 
3123 static int virtnet_set_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info)
3124 {
3125 	struct virtnet_info *vi = netdev_priv(dev);
3126 	int rc = 0;
3127 
3128 	switch (info->cmd) {
3129 	case ETHTOOL_SRXFH:
3130 		if (!virtnet_set_hashflow(vi, info))
3131 			rc = -EINVAL;
3132 
3133 		break;
3134 	default:
3135 		rc = -EOPNOTSUPP;
3136 	}
3137 
3138 	return rc;
3139 }
3140 
3141 static const struct ethtool_ops virtnet_ethtool_ops = {
3142 	.supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES |
3143 		ETHTOOL_COALESCE_USECS,
3144 	.get_drvinfo = virtnet_get_drvinfo,
3145 	.get_link = ethtool_op_get_link,
3146 	.get_ringparam = virtnet_get_ringparam,
3147 	.set_ringparam = virtnet_set_ringparam,
3148 	.get_strings = virtnet_get_strings,
3149 	.get_sset_count = virtnet_get_sset_count,
3150 	.get_ethtool_stats = virtnet_get_ethtool_stats,
3151 	.set_channels = virtnet_set_channels,
3152 	.get_channels = virtnet_get_channels,
3153 	.get_ts_info = ethtool_op_get_ts_info,
3154 	.get_link_ksettings = virtnet_get_link_ksettings,
3155 	.set_link_ksettings = virtnet_set_link_ksettings,
3156 	.set_coalesce = virtnet_set_coalesce,
3157 	.get_coalesce = virtnet_get_coalesce,
3158 	.get_rxfh_key_size = virtnet_get_rxfh_key_size,
3159 	.get_rxfh_indir_size = virtnet_get_rxfh_indir_size,
3160 	.get_rxfh = virtnet_get_rxfh,
3161 	.set_rxfh = virtnet_set_rxfh,
3162 	.get_rxnfc = virtnet_get_rxnfc,
3163 	.set_rxnfc = virtnet_set_rxnfc,
3164 };
3165 
3166 static void virtnet_freeze_down(struct virtio_device *vdev)
3167 {
3168 	struct virtnet_info *vi = vdev->priv;
3169 
3170 	/* Make sure no work handler is accessing the device */
3171 	flush_work(&vi->config_work);
3172 
3173 	netif_tx_lock_bh(vi->dev);
3174 	netif_device_detach(vi->dev);
3175 	netif_tx_unlock_bh(vi->dev);
3176 	if (netif_running(vi->dev))
3177 		virtnet_close(vi->dev);
3178 }
3179 
3180 static int init_vqs(struct virtnet_info *vi);
3181 
3182 static int virtnet_restore_up(struct virtio_device *vdev)
3183 {
3184 	struct virtnet_info *vi = vdev->priv;
3185 	int err;
3186 
3187 	err = init_vqs(vi);
3188 	if (err)
3189 		return err;
3190 
3191 	virtio_device_ready(vdev);
3192 
3193 	enable_delayed_refill(vi);
3194 
3195 	if (netif_running(vi->dev)) {
3196 		err = virtnet_open(vi->dev);
3197 		if (err)
3198 			return err;
3199 	}
3200 
3201 	netif_tx_lock_bh(vi->dev);
3202 	netif_device_attach(vi->dev);
3203 	netif_tx_unlock_bh(vi->dev);
3204 	return err;
3205 }
3206 
3207 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
3208 {
3209 	struct scatterlist sg;
3210 	vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
3211 
3212 	sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
3213 
3214 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
3215 				  VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
3216 		dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
3217 		return -EINVAL;
3218 	}
3219 
3220 	return 0;
3221 }
3222 
3223 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
3224 {
3225 	u64 offloads = 0;
3226 
3227 	if (!vi->guest_offloads)
3228 		return 0;
3229 
3230 	return virtnet_set_guest_offloads(vi, offloads);
3231 }
3232 
3233 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
3234 {
3235 	u64 offloads = vi->guest_offloads;
3236 
3237 	if (!vi->guest_offloads)
3238 		return 0;
3239 
3240 	return virtnet_set_guest_offloads(vi, offloads);
3241 }
3242 
3243 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
3244 			   struct netlink_ext_ack *extack)
3245 {
3246 	unsigned int room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM +
3247 					   sizeof(struct skb_shared_info));
3248 	unsigned int max_sz = PAGE_SIZE - room - ETH_HLEN;
3249 	struct virtnet_info *vi = netdev_priv(dev);
3250 	struct bpf_prog *old_prog;
3251 	u16 xdp_qp = 0, curr_qp;
3252 	int i, err;
3253 
3254 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
3255 	    && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3256 	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3257 	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
3258 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
3259 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM) ||
3260 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) ||
3261 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6))) {
3262 		NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
3263 		return -EOPNOTSUPP;
3264 	}
3265 
3266 	if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
3267 		NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
3268 		return -EINVAL;
3269 	}
3270 
3271 	if (prog && !prog->aux->xdp_has_frags && dev->mtu > max_sz) {
3272 		NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP without frags");
3273 		netdev_warn(dev, "single-buffer XDP requires MTU less than %u\n", max_sz);
3274 		return -EINVAL;
3275 	}
3276 
3277 	curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
3278 	if (prog)
3279 		xdp_qp = nr_cpu_ids;
3280 
3281 	/* XDP requires extra queues for XDP_TX */
3282 	if (curr_qp + xdp_qp > vi->max_queue_pairs) {
3283 		netdev_warn_once(dev, "XDP request %i queues but max is %i. XDP_TX and XDP_REDIRECT will operate in a slower locked tx mode.\n",
3284 				 curr_qp + xdp_qp, vi->max_queue_pairs);
3285 		xdp_qp = 0;
3286 	}
3287 
3288 	old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
3289 	if (!prog && !old_prog)
3290 		return 0;
3291 
3292 	if (prog)
3293 		bpf_prog_add(prog, vi->max_queue_pairs - 1);
3294 
3295 	/* Make sure NAPI is not using any XDP TX queues for RX. */
3296 	if (netif_running(dev)) {
3297 		for (i = 0; i < vi->max_queue_pairs; i++) {
3298 			napi_disable(&vi->rq[i].napi);
3299 			virtnet_napi_tx_disable(&vi->sq[i].napi);
3300 		}
3301 	}
3302 
3303 	if (!prog) {
3304 		for (i = 0; i < vi->max_queue_pairs; i++) {
3305 			rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
3306 			if (i == 0)
3307 				virtnet_restore_guest_offloads(vi);
3308 		}
3309 		synchronize_net();
3310 	}
3311 
3312 	err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
3313 	if (err)
3314 		goto err;
3315 	netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
3316 	vi->xdp_queue_pairs = xdp_qp;
3317 
3318 	if (prog) {
3319 		vi->xdp_enabled = true;
3320 		for (i = 0; i < vi->max_queue_pairs; i++) {
3321 			rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
3322 			if (i == 0 && !old_prog)
3323 				virtnet_clear_guest_offloads(vi);
3324 		}
3325 		if (!old_prog)
3326 			xdp_features_set_redirect_target(dev, true);
3327 	} else {
3328 		xdp_features_clear_redirect_target(dev);
3329 		vi->xdp_enabled = false;
3330 	}
3331 
3332 	for (i = 0; i < vi->max_queue_pairs; i++) {
3333 		if (old_prog)
3334 			bpf_prog_put(old_prog);
3335 		if (netif_running(dev)) {
3336 			virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
3337 			virtnet_napi_tx_enable(vi, vi->sq[i].vq,
3338 					       &vi->sq[i].napi);
3339 		}
3340 	}
3341 
3342 	return 0;
3343 
3344 err:
3345 	if (!prog) {
3346 		virtnet_clear_guest_offloads(vi);
3347 		for (i = 0; i < vi->max_queue_pairs; i++)
3348 			rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
3349 	}
3350 
3351 	if (netif_running(dev)) {
3352 		for (i = 0; i < vi->max_queue_pairs; i++) {
3353 			virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
3354 			virtnet_napi_tx_enable(vi, vi->sq[i].vq,
3355 					       &vi->sq[i].napi);
3356 		}
3357 	}
3358 	if (prog)
3359 		bpf_prog_sub(prog, vi->max_queue_pairs - 1);
3360 	return err;
3361 }
3362 
3363 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
3364 {
3365 	switch (xdp->command) {
3366 	case XDP_SETUP_PROG:
3367 		return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
3368 	default:
3369 		return -EINVAL;
3370 	}
3371 }
3372 
3373 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
3374 				      size_t len)
3375 {
3376 	struct virtnet_info *vi = netdev_priv(dev);
3377 	int ret;
3378 
3379 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
3380 		return -EOPNOTSUPP;
3381 
3382 	ret = snprintf(buf, len, "sby");
3383 	if (ret >= len)
3384 		return -EOPNOTSUPP;
3385 
3386 	return 0;
3387 }
3388 
3389 static int virtnet_set_features(struct net_device *dev,
3390 				netdev_features_t features)
3391 {
3392 	struct virtnet_info *vi = netdev_priv(dev);
3393 	u64 offloads;
3394 	int err;
3395 
3396 	if ((dev->features ^ features) & NETIF_F_GRO_HW) {
3397 		if (vi->xdp_enabled)
3398 			return -EBUSY;
3399 
3400 		if (features & NETIF_F_GRO_HW)
3401 			offloads = vi->guest_offloads_capable;
3402 		else
3403 			offloads = vi->guest_offloads_capable &
3404 				   ~GUEST_OFFLOAD_GRO_HW_MASK;
3405 
3406 		err = virtnet_set_guest_offloads(vi, offloads);
3407 		if (err)
3408 			return err;
3409 		vi->guest_offloads = offloads;
3410 	}
3411 
3412 	if ((dev->features ^ features) & NETIF_F_RXHASH) {
3413 		if (features & NETIF_F_RXHASH)
3414 			vi->ctrl->rss.hash_types = vi->rss_hash_types_saved;
3415 		else
3416 			vi->ctrl->rss.hash_types = VIRTIO_NET_HASH_REPORT_NONE;
3417 
3418 		if (!virtnet_commit_rss_command(vi))
3419 			return -EINVAL;
3420 	}
3421 
3422 	return 0;
3423 }
3424 
3425 static void virtnet_tx_timeout(struct net_device *dev, unsigned int txqueue)
3426 {
3427 	struct virtnet_info *priv = netdev_priv(dev);
3428 	struct send_queue *sq = &priv->sq[txqueue];
3429 	struct netdev_queue *txq = netdev_get_tx_queue(dev, txqueue);
3430 
3431 	u64_stats_update_begin(&sq->stats.syncp);
3432 	sq->stats.tx_timeouts++;
3433 	u64_stats_update_end(&sq->stats.syncp);
3434 
3435 	netdev_err(dev, "TX timeout on queue: %u, sq: %s, vq: 0x%x, name: %s, %u usecs ago\n",
3436 		   txqueue, sq->name, sq->vq->index, sq->vq->name,
3437 		   jiffies_to_usecs(jiffies - READ_ONCE(txq->trans_start)));
3438 }
3439 
3440 static const struct net_device_ops virtnet_netdev = {
3441 	.ndo_open            = virtnet_open,
3442 	.ndo_stop   	     = virtnet_close,
3443 	.ndo_start_xmit      = start_xmit,
3444 	.ndo_validate_addr   = eth_validate_addr,
3445 	.ndo_set_mac_address = virtnet_set_mac_address,
3446 	.ndo_set_rx_mode     = virtnet_set_rx_mode,
3447 	.ndo_get_stats64     = virtnet_stats,
3448 	.ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
3449 	.ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
3450 	.ndo_bpf		= virtnet_xdp,
3451 	.ndo_xdp_xmit		= virtnet_xdp_xmit,
3452 	.ndo_features_check	= passthru_features_check,
3453 	.ndo_get_phys_port_name	= virtnet_get_phys_port_name,
3454 	.ndo_set_features	= virtnet_set_features,
3455 	.ndo_tx_timeout		= virtnet_tx_timeout,
3456 };
3457 
3458 static void virtnet_config_changed_work(struct work_struct *work)
3459 {
3460 	struct virtnet_info *vi =
3461 		container_of(work, struct virtnet_info, config_work);
3462 	u16 v;
3463 
3464 	if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
3465 				 struct virtio_net_config, status, &v) < 0)
3466 		return;
3467 
3468 	if (v & VIRTIO_NET_S_ANNOUNCE) {
3469 		netdev_notify_peers(vi->dev);
3470 		virtnet_ack_link_announce(vi);
3471 	}
3472 
3473 	/* Ignore unknown (future) status bits */
3474 	v &= VIRTIO_NET_S_LINK_UP;
3475 
3476 	if (vi->status == v)
3477 		return;
3478 
3479 	vi->status = v;
3480 
3481 	if (vi->status & VIRTIO_NET_S_LINK_UP) {
3482 		virtnet_update_settings(vi);
3483 		netif_carrier_on(vi->dev);
3484 		netif_tx_wake_all_queues(vi->dev);
3485 	} else {
3486 		netif_carrier_off(vi->dev);
3487 		netif_tx_stop_all_queues(vi->dev);
3488 	}
3489 }
3490 
3491 static void virtnet_config_changed(struct virtio_device *vdev)
3492 {
3493 	struct virtnet_info *vi = vdev->priv;
3494 
3495 	schedule_work(&vi->config_work);
3496 }
3497 
3498 static void virtnet_free_queues(struct virtnet_info *vi)
3499 {
3500 	int i;
3501 
3502 	for (i = 0; i < vi->max_queue_pairs; i++) {
3503 		__netif_napi_del(&vi->rq[i].napi);
3504 		__netif_napi_del(&vi->sq[i].napi);
3505 	}
3506 
3507 	/* We called __netif_napi_del(),
3508 	 * we need to respect an RCU grace period before freeing vi->rq
3509 	 */
3510 	synchronize_net();
3511 
3512 	kfree(vi->rq);
3513 	kfree(vi->sq);
3514 	kfree(vi->ctrl);
3515 }
3516 
3517 static void _free_receive_bufs(struct virtnet_info *vi)
3518 {
3519 	struct bpf_prog *old_prog;
3520 	int i;
3521 
3522 	for (i = 0; i < vi->max_queue_pairs; i++) {
3523 		while (vi->rq[i].pages)
3524 			__free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
3525 
3526 		old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
3527 		RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
3528 		if (old_prog)
3529 			bpf_prog_put(old_prog);
3530 	}
3531 }
3532 
3533 static void free_receive_bufs(struct virtnet_info *vi)
3534 {
3535 	rtnl_lock();
3536 	_free_receive_bufs(vi);
3537 	rtnl_unlock();
3538 }
3539 
3540 static void free_receive_page_frags(struct virtnet_info *vi)
3541 {
3542 	int i;
3543 	for (i = 0; i < vi->max_queue_pairs; i++)
3544 		if (vi->rq[i].alloc_frag.page)
3545 			put_page(vi->rq[i].alloc_frag.page);
3546 }
3547 
3548 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf)
3549 {
3550 	if (!is_xdp_frame(buf))
3551 		dev_kfree_skb(buf);
3552 	else
3553 		xdp_return_frame(ptr_to_xdp(buf));
3554 }
3555 
3556 static void virtnet_rq_free_unused_buf(struct virtqueue *vq, void *buf)
3557 {
3558 	struct virtnet_info *vi = vq->vdev->priv;
3559 	int i = vq2rxq(vq);
3560 
3561 	if (vi->mergeable_rx_bufs)
3562 		put_page(virt_to_head_page(buf));
3563 	else if (vi->big_packets)
3564 		give_pages(&vi->rq[i], buf);
3565 	else
3566 		put_page(virt_to_head_page(buf));
3567 }
3568 
3569 static void free_unused_bufs(struct virtnet_info *vi)
3570 {
3571 	void *buf;
3572 	int i;
3573 
3574 	for (i = 0; i < vi->max_queue_pairs; i++) {
3575 		struct virtqueue *vq = vi->sq[i].vq;
3576 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
3577 			virtnet_sq_free_unused_buf(vq, buf);
3578 		cond_resched();
3579 	}
3580 
3581 	for (i = 0; i < vi->max_queue_pairs; i++) {
3582 		struct virtqueue *vq = vi->rq[i].vq;
3583 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
3584 			virtnet_rq_free_unused_buf(vq, buf);
3585 		cond_resched();
3586 	}
3587 }
3588 
3589 static void virtnet_del_vqs(struct virtnet_info *vi)
3590 {
3591 	struct virtio_device *vdev = vi->vdev;
3592 
3593 	virtnet_clean_affinity(vi);
3594 
3595 	vdev->config->del_vqs(vdev);
3596 
3597 	virtnet_free_queues(vi);
3598 }
3599 
3600 /* How large should a single buffer be so a queue full of these can fit at
3601  * least one full packet?
3602  * Logic below assumes the mergeable buffer header is used.
3603  */
3604 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
3605 {
3606 	const unsigned int hdr_len = vi->hdr_len;
3607 	unsigned int rq_size = virtqueue_get_vring_size(vq);
3608 	unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
3609 	unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
3610 	unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
3611 
3612 	return max(max(min_buf_len, hdr_len) - hdr_len,
3613 		   (unsigned int)GOOD_PACKET_LEN);
3614 }
3615 
3616 static int virtnet_find_vqs(struct virtnet_info *vi)
3617 {
3618 	vq_callback_t **callbacks;
3619 	struct virtqueue **vqs;
3620 	int ret = -ENOMEM;
3621 	int i, total_vqs;
3622 	const char **names;
3623 	bool *ctx;
3624 
3625 	/* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
3626 	 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
3627 	 * possible control vq.
3628 	 */
3629 	total_vqs = vi->max_queue_pairs * 2 +
3630 		    virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
3631 
3632 	/* Allocate space for find_vqs parameters */
3633 	vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
3634 	if (!vqs)
3635 		goto err_vq;
3636 	callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
3637 	if (!callbacks)
3638 		goto err_callback;
3639 	names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
3640 	if (!names)
3641 		goto err_names;
3642 	if (!vi->big_packets || vi->mergeable_rx_bufs) {
3643 		ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
3644 		if (!ctx)
3645 			goto err_ctx;
3646 	} else {
3647 		ctx = NULL;
3648 	}
3649 
3650 	/* Parameters for control virtqueue, if any */
3651 	if (vi->has_cvq) {
3652 		callbacks[total_vqs - 1] = NULL;
3653 		names[total_vqs - 1] = "control";
3654 	}
3655 
3656 	/* Allocate/initialize parameters for send/receive virtqueues */
3657 	for (i = 0; i < vi->max_queue_pairs; i++) {
3658 		callbacks[rxq2vq(i)] = skb_recv_done;
3659 		callbacks[txq2vq(i)] = skb_xmit_done;
3660 		sprintf(vi->rq[i].name, "input.%d", i);
3661 		sprintf(vi->sq[i].name, "output.%d", i);
3662 		names[rxq2vq(i)] = vi->rq[i].name;
3663 		names[txq2vq(i)] = vi->sq[i].name;
3664 		if (ctx)
3665 			ctx[rxq2vq(i)] = true;
3666 	}
3667 
3668 	ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
3669 				  names, ctx, NULL);
3670 	if (ret)
3671 		goto err_find;
3672 
3673 	if (vi->has_cvq) {
3674 		vi->cvq = vqs[total_vqs - 1];
3675 		if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
3676 			vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
3677 	}
3678 
3679 	for (i = 0; i < vi->max_queue_pairs; i++) {
3680 		vi->rq[i].vq = vqs[rxq2vq(i)];
3681 		vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
3682 		vi->sq[i].vq = vqs[txq2vq(i)];
3683 	}
3684 
3685 	/* run here: ret == 0. */
3686 
3687 
3688 err_find:
3689 	kfree(ctx);
3690 err_ctx:
3691 	kfree(names);
3692 err_names:
3693 	kfree(callbacks);
3694 err_callback:
3695 	kfree(vqs);
3696 err_vq:
3697 	return ret;
3698 }
3699 
3700 static int virtnet_alloc_queues(struct virtnet_info *vi)
3701 {
3702 	int i;
3703 
3704 	if (vi->has_cvq) {
3705 		vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
3706 		if (!vi->ctrl)
3707 			goto err_ctrl;
3708 	} else {
3709 		vi->ctrl = NULL;
3710 	}
3711 	vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
3712 	if (!vi->sq)
3713 		goto err_sq;
3714 	vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
3715 	if (!vi->rq)
3716 		goto err_rq;
3717 
3718 	INIT_DELAYED_WORK(&vi->refill, refill_work);
3719 	for (i = 0; i < vi->max_queue_pairs; i++) {
3720 		vi->rq[i].pages = NULL;
3721 		netif_napi_add_weight(vi->dev, &vi->rq[i].napi, virtnet_poll,
3722 				      napi_weight);
3723 		netif_napi_add_tx_weight(vi->dev, &vi->sq[i].napi,
3724 					 virtnet_poll_tx,
3725 					 napi_tx ? napi_weight : 0);
3726 
3727 		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
3728 		ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
3729 		sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
3730 
3731 		u64_stats_init(&vi->rq[i].stats.syncp);
3732 		u64_stats_init(&vi->sq[i].stats.syncp);
3733 	}
3734 
3735 	return 0;
3736 
3737 err_rq:
3738 	kfree(vi->sq);
3739 err_sq:
3740 	kfree(vi->ctrl);
3741 err_ctrl:
3742 	return -ENOMEM;
3743 }
3744 
3745 static int init_vqs(struct virtnet_info *vi)
3746 {
3747 	int ret;
3748 
3749 	/* Allocate send & receive queues */
3750 	ret = virtnet_alloc_queues(vi);
3751 	if (ret)
3752 		goto err;
3753 
3754 	ret = virtnet_find_vqs(vi);
3755 	if (ret)
3756 		goto err_free;
3757 
3758 	cpus_read_lock();
3759 	virtnet_set_affinity(vi);
3760 	cpus_read_unlock();
3761 
3762 	return 0;
3763 
3764 err_free:
3765 	virtnet_free_queues(vi);
3766 err:
3767 	return ret;
3768 }
3769 
3770 #ifdef CONFIG_SYSFS
3771 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
3772 		char *buf)
3773 {
3774 	struct virtnet_info *vi = netdev_priv(queue->dev);
3775 	unsigned int queue_index = get_netdev_rx_queue_index(queue);
3776 	unsigned int headroom = virtnet_get_headroom(vi);
3777 	unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
3778 	struct ewma_pkt_len *avg;
3779 
3780 	BUG_ON(queue_index >= vi->max_queue_pairs);
3781 	avg = &vi->rq[queue_index].mrg_avg_pkt_len;
3782 	return sprintf(buf, "%u\n",
3783 		       get_mergeable_buf_len(&vi->rq[queue_index], avg,
3784 				       SKB_DATA_ALIGN(headroom + tailroom)));
3785 }
3786 
3787 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
3788 	__ATTR_RO(mergeable_rx_buffer_size);
3789 
3790 static struct attribute *virtio_net_mrg_rx_attrs[] = {
3791 	&mergeable_rx_buffer_size_attribute.attr,
3792 	NULL
3793 };
3794 
3795 static const struct attribute_group virtio_net_mrg_rx_group = {
3796 	.name = "virtio_net",
3797 	.attrs = virtio_net_mrg_rx_attrs
3798 };
3799 #endif
3800 
3801 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
3802 				    unsigned int fbit,
3803 				    const char *fname, const char *dname)
3804 {
3805 	if (!virtio_has_feature(vdev, fbit))
3806 		return false;
3807 
3808 	dev_err(&vdev->dev, "device advertises feature %s but not %s",
3809 		fname, dname);
3810 
3811 	return true;
3812 }
3813 
3814 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)			\
3815 	virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
3816 
3817 static bool virtnet_validate_features(struct virtio_device *vdev)
3818 {
3819 	if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
3820 	    (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
3821 			     "VIRTIO_NET_F_CTRL_VQ") ||
3822 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
3823 			     "VIRTIO_NET_F_CTRL_VQ") ||
3824 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
3825 			     "VIRTIO_NET_F_CTRL_VQ") ||
3826 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
3827 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
3828 			     "VIRTIO_NET_F_CTRL_VQ") ||
3829 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_RSS,
3830 			     "VIRTIO_NET_F_CTRL_VQ") ||
3831 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_HASH_REPORT,
3832 			     "VIRTIO_NET_F_CTRL_VQ") ||
3833 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_NOTF_COAL,
3834 			     "VIRTIO_NET_F_CTRL_VQ"))) {
3835 		return false;
3836 	}
3837 
3838 	return true;
3839 }
3840 
3841 #define MIN_MTU ETH_MIN_MTU
3842 #define MAX_MTU ETH_MAX_MTU
3843 
3844 static int virtnet_validate(struct virtio_device *vdev)
3845 {
3846 	if (!vdev->config->get) {
3847 		dev_err(&vdev->dev, "%s failure: config access disabled\n",
3848 			__func__);
3849 		return -EINVAL;
3850 	}
3851 
3852 	if (!virtnet_validate_features(vdev))
3853 		return -EINVAL;
3854 
3855 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3856 		int mtu = virtio_cread16(vdev,
3857 					 offsetof(struct virtio_net_config,
3858 						  mtu));
3859 		if (mtu < MIN_MTU)
3860 			__virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
3861 	}
3862 
3863 	if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY) &&
3864 	    !virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
3865 		dev_warn(&vdev->dev, "device advertises feature VIRTIO_NET_F_STANDBY but not VIRTIO_NET_F_MAC, disabling standby");
3866 		__virtio_clear_bit(vdev, VIRTIO_NET_F_STANDBY);
3867 	}
3868 
3869 	return 0;
3870 }
3871 
3872 static bool virtnet_check_guest_gso(const struct virtnet_info *vi)
3873 {
3874 	return virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3875 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3876 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
3877 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
3878 		(virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) &&
3879 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6));
3880 }
3881 
3882 static void virtnet_set_big_packets(struct virtnet_info *vi, const int mtu)
3883 {
3884 	bool guest_gso = virtnet_check_guest_gso(vi);
3885 
3886 	/* If device can receive ANY guest GSO packets, regardless of mtu,
3887 	 * allocate packets of maximum size, otherwise limit it to only
3888 	 * mtu size worth only.
3889 	 */
3890 	if (mtu > ETH_DATA_LEN || guest_gso) {
3891 		vi->big_packets = true;
3892 		vi->big_packets_num_skbfrags = guest_gso ? MAX_SKB_FRAGS : DIV_ROUND_UP(mtu, PAGE_SIZE);
3893 	}
3894 }
3895 
3896 static int virtnet_probe(struct virtio_device *vdev)
3897 {
3898 	int i, err = -ENOMEM;
3899 	struct net_device *dev;
3900 	struct virtnet_info *vi;
3901 	u16 max_queue_pairs;
3902 	int mtu = 0;
3903 
3904 	/* Find if host supports multiqueue/rss virtio_net device */
3905 	max_queue_pairs = 1;
3906 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MQ) || virtio_has_feature(vdev, VIRTIO_NET_F_RSS))
3907 		max_queue_pairs =
3908 		     virtio_cread16(vdev, offsetof(struct virtio_net_config, max_virtqueue_pairs));
3909 
3910 	/* We need at least 2 queue's */
3911 	if (max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3912 	    max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3913 	    !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3914 		max_queue_pairs = 1;
3915 
3916 	/* Allocate ourselves a network device with room for our info */
3917 	dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3918 	if (!dev)
3919 		return -ENOMEM;
3920 
3921 	/* Set up network device as normal. */
3922 	dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
3923 			   IFF_TX_SKB_NO_LINEAR;
3924 	dev->netdev_ops = &virtnet_netdev;
3925 	dev->features = NETIF_F_HIGHDMA;
3926 
3927 	dev->ethtool_ops = &virtnet_ethtool_ops;
3928 	SET_NETDEV_DEV(dev, &vdev->dev);
3929 
3930 	/* Do we support "hardware" checksums? */
3931 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3932 		/* This opens up the world of extra features. */
3933 		dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3934 		if (csum)
3935 			dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3936 
3937 		if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3938 			dev->hw_features |= NETIF_F_TSO
3939 				| NETIF_F_TSO_ECN | NETIF_F_TSO6;
3940 		}
3941 		/* Individual feature bits: what can host handle? */
3942 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3943 			dev->hw_features |= NETIF_F_TSO;
3944 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3945 			dev->hw_features |= NETIF_F_TSO6;
3946 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3947 			dev->hw_features |= NETIF_F_TSO_ECN;
3948 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_USO))
3949 			dev->hw_features |= NETIF_F_GSO_UDP_L4;
3950 
3951 		dev->features |= NETIF_F_GSO_ROBUST;
3952 
3953 		if (gso)
3954 			dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3955 		/* (!csum && gso) case will be fixed by register_netdev() */
3956 	}
3957 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3958 		dev->features |= NETIF_F_RXCSUM;
3959 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3960 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3961 		dev->features |= NETIF_F_GRO_HW;
3962 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3963 		dev->hw_features |= NETIF_F_GRO_HW;
3964 
3965 	dev->vlan_features = dev->features;
3966 	dev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT;
3967 
3968 	/* MTU range: 68 - 65535 */
3969 	dev->min_mtu = MIN_MTU;
3970 	dev->max_mtu = MAX_MTU;
3971 
3972 	/* Configuration may specify what MAC to use.  Otherwise random. */
3973 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
3974 		u8 addr[ETH_ALEN];
3975 
3976 		virtio_cread_bytes(vdev,
3977 				   offsetof(struct virtio_net_config, mac),
3978 				   addr, ETH_ALEN);
3979 		eth_hw_addr_set(dev, addr);
3980 	} else {
3981 		eth_hw_addr_random(dev);
3982 		dev_info(&vdev->dev, "Assigned random MAC address %pM\n",
3983 			 dev->dev_addr);
3984 	}
3985 
3986 	/* Set up our device-specific information */
3987 	vi = netdev_priv(dev);
3988 	vi->dev = dev;
3989 	vi->vdev = vdev;
3990 	vdev->priv = vi;
3991 
3992 	INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3993 	spin_lock_init(&vi->refill_lock);
3994 
3995 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) {
3996 		vi->mergeable_rx_bufs = true;
3997 		dev->xdp_features |= NETDEV_XDP_ACT_RX_SG;
3998 	}
3999 
4000 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
4001 		vi->rx_usecs = 0;
4002 		vi->tx_usecs = 0;
4003 		vi->tx_max_packets = 0;
4004 		vi->rx_max_packets = 0;
4005 	}
4006 
4007 	if (virtio_has_feature(vdev, VIRTIO_NET_F_HASH_REPORT))
4008 		vi->has_rss_hash_report = true;
4009 
4010 	if (virtio_has_feature(vdev, VIRTIO_NET_F_RSS))
4011 		vi->has_rss = true;
4012 
4013 	if (vi->has_rss || vi->has_rss_hash_report) {
4014 		vi->rss_indir_table_size =
4015 			virtio_cread16(vdev, offsetof(struct virtio_net_config,
4016 				rss_max_indirection_table_length));
4017 		vi->rss_key_size =
4018 			virtio_cread8(vdev, offsetof(struct virtio_net_config, rss_max_key_size));
4019 
4020 		vi->rss_hash_types_supported =
4021 		    virtio_cread32(vdev, offsetof(struct virtio_net_config, supported_hash_types));
4022 		vi->rss_hash_types_supported &=
4023 				~(VIRTIO_NET_RSS_HASH_TYPE_IP_EX |
4024 				  VIRTIO_NET_RSS_HASH_TYPE_TCP_EX |
4025 				  VIRTIO_NET_RSS_HASH_TYPE_UDP_EX);
4026 
4027 		dev->hw_features |= NETIF_F_RXHASH;
4028 	}
4029 
4030 	if (vi->has_rss_hash_report)
4031 		vi->hdr_len = sizeof(struct virtio_net_hdr_v1_hash);
4032 	else if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
4033 		 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
4034 		vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
4035 	else
4036 		vi->hdr_len = sizeof(struct virtio_net_hdr);
4037 
4038 	if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
4039 	    virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
4040 		vi->any_header_sg = true;
4041 
4042 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
4043 		vi->has_cvq = true;
4044 
4045 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
4046 		mtu = virtio_cread16(vdev,
4047 				     offsetof(struct virtio_net_config,
4048 					      mtu));
4049 		if (mtu < dev->min_mtu) {
4050 			/* Should never trigger: MTU was previously validated
4051 			 * in virtnet_validate.
4052 			 */
4053 			dev_err(&vdev->dev,
4054 				"device MTU appears to have changed it is now %d < %d",
4055 				mtu, dev->min_mtu);
4056 			err = -EINVAL;
4057 			goto free;
4058 		}
4059 
4060 		dev->mtu = mtu;
4061 		dev->max_mtu = mtu;
4062 	}
4063 
4064 	virtnet_set_big_packets(vi, mtu);
4065 
4066 	if (vi->any_header_sg)
4067 		dev->needed_headroom = vi->hdr_len;
4068 
4069 	/* Enable multiqueue by default */
4070 	if (num_online_cpus() >= max_queue_pairs)
4071 		vi->curr_queue_pairs = max_queue_pairs;
4072 	else
4073 		vi->curr_queue_pairs = num_online_cpus();
4074 	vi->max_queue_pairs = max_queue_pairs;
4075 
4076 	/* Allocate/initialize the rx/tx queues, and invoke find_vqs */
4077 	err = init_vqs(vi);
4078 	if (err)
4079 		goto free;
4080 
4081 #ifdef CONFIG_SYSFS
4082 	if (vi->mergeable_rx_bufs)
4083 		dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
4084 #endif
4085 	netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
4086 	netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
4087 
4088 	virtnet_init_settings(dev);
4089 
4090 	if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
4091 		vi->failover = net_failover_create(vi->dev);
4092 		if (IS_ERR(vi->failover)) {
4093 			err = PTR_ERR(vi->failover);
4094 			goto free_vqs;
4095 		}
4096 	}
4097 
4098 	if (vi->has_rss || vi->has_rss_hash_report)
4099 		virtnet_init_default_rss(vi);
4100 
4101 	/* serialize netdev register + virtio_device_ready() with ndo_open() */
4102 	rtnl_lock();
4103 
4104 	err = register_netdevice(dev);
4105 	if (err) {
4106 		pr_debug("virtio_net: registering device failed\n");
4107 		rtnl_unlock();
4108 		goto free_failover;
4109 	}
4110 
4111 	virtio_device_ready(vdev);
4112 
4113 	/* a random MAC address has been assigned, notify the device.
4114 	 * We don't fail probe if VIRTIO_NET_F_CTRL_MAC_ADDR is not there
4115 	 * because many devices work fine without getting MAC explicitly
4116 	 */
4117 	if (!virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
4118 	    virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
4119 		struct scatterlist sg;
4120 
4121 		sg_init_one(&sg, dev->dev_addr, dev->addr_len);
4122 		if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
4123 					  VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
4124 			pr_debug("virtio_net: setting MAC address failed\n");
4125 			rtnl_unlock();
4126 			err = -EINVAL;
4127 			goto free_unregister_netdev;
4128 		}
4129 	}
4130 
4131 	rtnl_unlock();
4132 
4133 	err = virtnet_cpu_notif_add(vi);
4134 	if (err) {
4135 		pr_debug("virtio_net: registering cpu notifier failed\n");
4136 		goto free_unregister_netdev;
4137 	}
4138 
4139 	virtnet_set_queues(vi, vi->curr_queue_pairs);
4140 
4141 	/* Assume link up if device can't report link status,
4142 	   otherwise get link status from config. */
4143 	netif_carrier_off(dev);
4144 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
4145 		schedule_work(&vi->config_work);
4146 	} else {
4147 		vi->status = VIRTIO_NET_S_LINK_UP;
4148 		virtnet_update_settings(vi);
4149 		netif_carrier_on(dev);
4150 	}
4151 
4152 	for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
4153 		if (virtio_has_feature(vi->vdev, guest_offloads[i]))
4154 			set_bit(guest_offloads[i], &vi->guest_offloads);
4155 	vi->guest_offloads_capable = vi->guest_offloads;
4156 
4157 	pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
4158 		 dev->name, max_queue_pairs);
4159 
4160 	return 0;
4161 
4162 free_unregister_netdev:
4163 	unregister_netdev(dev);
4164 free_failover:
4165 	net_failover_destroy(vi->failover);
4166 free_vqs:
4167 	virtio_reset_device(vdev);
4168 	cancel_delayed_work_sync(&vi->refill);
4169 	free_receive_page_frags(vi);
4170 	virtnet_del_vqs(vi);
4171 free:
4172 	free_netdev(dev);
4173 	return err;
4174 }
4175 
4176 static void remove_vq_common(struct virtnet_info *vi)
4177 {
4178 	virtio_reset_device(vi->vdev);
4179 
4180 	/* Free unused buffers in both send and recv, if any. */
4181 	free_unused_bufs(vi);
4182 
4183 	free_receive_bufs(vi);
4184 
4185 	free_receive_page_frags(vi);
4186 
4187 	virtnet_del_vqs(vi);
4188 }
4189 
4190 static void virtnet_remove(struct virtio_device *vdev)
4191 {
4192 	struct virtnet_info *vi = vdev->priv;
4193 
4194 	virtnet_cpu_notif_remove(vi);
4195 
4196 	/* Make sure no work handler is accessing the device. */
4197 	flush_work(&vi->config_work);
4198 
4199 	unregister_netdev(vi->dev);
4200 
4201 	net_failover_destroy(vi->failover);
4202 
4203 	remove_vq_common(vi);
4204 
4205 	free_netdev(vi->dev);
4206 }
4207 
4208 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
4209 {
4210 	struct virtnet_info *vi = vdev->priv;
4211 
4212 	virtnet_cpu_notif_remove(vi);
4213 	virtnet_freeze_down(vdev);
4214 	remove_vq_common(vi);
4215 
4216 	return 0;
4217 }
4218 
4219 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
4220 {
4221 	struct virtnet_info *vi = vdev->priv;
4222 	int err;
4223 
4224 	err = virtnet_restore_up(vdev);
4225 	if (err)
4226 		return err;
4227 	virtnet_set_queues(vi, vi->curr_queue_pairs);
4228 
4229 	err = virtnet_cpu_notif_add(vi);
4230 	if (err) {
4231 		virtnet_freeze_down(vdev);
4232 		remove_vq_common(vi);
4233 		return err;
4234 	}
4235 
4236 	return 0;
4237 }
4238 
4239 static struct virtio_device_id id_table[] = {
4240 	{ VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
4241 	{ 0 },
4242 };
4243 
4244 #define VIRTNET_FEATURES \
4245 	VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
4246 	VIRTIO_NET_F_MAC, \
4247 	VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
4248 	VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
4249 	VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
4250 	VIRTIO_NET_F_HOST_USO, VIRTIO_NET_F_GUEST_USO4, VIRTIO_NET_F_GUEST_USO6, \
4251 	VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
4252 	VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
4253 	VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
4254 	VIRTIO_NET_F_CTRL_MAC_ADDR, \
4255 	VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
4256 	VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY, \
4257 	VIRTIO_NET_F_RSS, VIRTIO_NET_F_HASH_REPORT, VIRTIO_NET_F_NOTF_COAL, \
4258 	VIRTIO_NET_F_GUEST_HDRLEN
4259 
4260 static unsigned int features[] = {
4261 	VIRTNET_FEATURES,
4262 };
4263 
4264 static unsigned int features_legacy[] = {
4265 	VIRTNET_FEATURES,
4266 	VIRTIO_NET_F_GSO,
4267 	VIRTIO_F_ANY_LAYOUT,
4268 };
4269 
4270 static struct virtio_driver virtio_net_driver = {
4271 	.feature_table = features,
4272 	.feature_table_size = ARRAY_SIZE(features),
4273 	.feature_table_legacy = features_legacy,
4274 	.feature_table_size_legacy = ARRAY_SIZE(features_legacy),
4275 	.driver.name =	KBUILD_MODNAME,
4276 	.driver.owner =	THIS_MODULE,
4277 	.id_table =	id_table,
4278 	.validate =	virtnet_validate,
4279 	.probe =	virtnet_probe,
4280 	.remove =	virtnet_remove,
4281 	.config_changed = virtnet_config_changed,
4282 #ifdef CONFIG_PM_SLEEP
4283 	.freeze =	virtnet_freeze,
4284 	.restore =	virtnet_restore,
4285 #endif
4286 };
4287 
4288 static __init int virtio_net_driver_init(void)
4289 {
4290 	int ret;
4291 
4292 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
4293 				      virtnet_cpu_online,
4294 				      virtnet_cpu_down_prep);
4295 	if (ret < 0)
4296 		goto out;
4297 	virtionet_online = ret;
4298 	ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
4299 				      NULL, virtnet_cpu_dead);
4300 	if (ret)
4301 		goto err_dead;
4302 	ret = register_virtio_driver(&virtio_net_driver);
4303 	if (ret)
4304 		goto err_virtio;
4305 	return 0;
4306 err_virtio:
4307 	cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
4308 err_dead:
4309 	cpuhp_remove_multi_state(virtionet_online);
4310 out:
4311 	return ret;
4312 }
4313 module_init(virtio_net_driver_init);
4314 
4315 static __exit void virtio_net_driver_exit(void)
4316 {
4317 	unregister_virtio_driver(&virtio_net_driver);
4318 	cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
4319 	cpuhp_remove_multi_state(virtionet_online);
4320 }
4321 module_exit(virtio_net_driver_exit);
4322 
4323 MODULE_DEVICE_TABLE(virtio, id_table);
4324 MODULE_DESCRIPTION("Virtio network driver");
4325 MODULE_LICENSE("GPL");
4326