xref: /linux/drivers/net/virtio_net.c (revision 60684c2bd35064043360e6f716d1b7c20e967b7d)
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 };
67 
68 #define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
69 				(1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
70 				(1ULL << VIRTIO_NET_F_GUEST_ECN)  | \
71 				(1ULL << VIRTIO_NET_F_GUEST_UFO)  | \
72 				(1ULL << VIRTIO_NET_F_GUEST_USO4) | \
73 				(1ULL << VIRTIO_NET_F_GUEST_USO6))
74 
75 struct virtnet_stat_desc {
76 	char desc[ETH_GSTRING_LEN];
77 	size_t offset;
78 };
79 
80 struct virtnet_sq_stats {
81 	struct u64_stats_sync syncp;
82 	u64 packets;
83 	u64 bytes;
84 	u64 xdp_tx;
85 	u64 xdp_tx_drops;
86 	u64 kicks;
87 	u64 tx_timeouts;
88 };
89 
90 struct virtnet_rq_stats {
91 	struct u64_stats_sync syncp;
92 	u64 packets;
93 	u64 bytes;
94 	u64 drops;
95 	u64 xdp_packets;
96 	u64 xdp_tx;
97 	u64 xdp_redirects;
98 	u64 xdp_drops;
99 	u64 kicks;
100 };
101 
102 #define VIRTNET_SQ_STAT(m)	offsetof(struct virtnet_sq_stats, m)
103 #define VIRTNET_RQ_STAT(m)	offsetof(struct virtnet_rq_stats, m)
104 
105 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
106 	{ "packets",		VIRTNET_SQ_STAT(packets) },
107 	{ "bytes",		VIRTNET_SQ_STAT(bytes) },
108 	{ "xdp_tx",		VIRTNET_SQ_STAT(xdp_tx) },
109 	{ "xdp_tx_drops",	VIRTNET_SQ_STAT(xdp_tx_drops) },
110 	{ "kicks",		VIRTNET_SQ_STAT(kicks) },
111 	{ "tx_timeouts",	VIRTNET_SQ_STAT(tx_timeouts) },
112 };
113 
114 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
115 	{ "packets",		VIRTNET_RQ_STAT(packets) },
116 	{ "bytes",		VIRTNET_RQ_STAT(bytes) },
117 	{ "drops",		VIRTNET_RQ_STAT(drops) },
118 	{ "xdp_packets",	VIRTNET_RQ_STAT(xdp_packets) },
119 	{ "xdp_tx",		VIRTNET_RQ_STAT(xdp_tx) },
120 	{ "xdp_redirects",	VIRTNET_RQ_STAT(xdp_redirects) },
121 	{ "xdp_drops",		VIRTNET_RQ_STAT(xdp_drops) },
122 	{ "kicks",		VIRTNET_RQ_STAT(kicks) },
123 };
124 
125 #define VIRTNET_SQ_STATS_LEN	ARRAY_SIZE(virtnet_sq_stats_desc)
126 #define VIRTNET_RQ_STATS_LEN	ARRAY_SIZE(virtnet_rq_stats_desc)
127 
128 /* Internal representation of a send virtqueue */
129 struct send_queue {
130 	/* Virtqueue associated with this send _queue */
131 	struct virtqueue *vq;
132 
133 	/* TX: fragments + linear part + virtio header */
134 	struct scatterlist sg[MAX_SKB_FRAGS + 2];
135 
136 	/* Name of the send queue: output.$index */
137 	char name[16];
138 
139 	struct virtnet_sq_stats stats;
140 
141 	struct napi_struct napi;
142 
143 	/* Record whether sq is in reset state. */
144 	bool reset;
145 };
146 
147 /* Internal representation of a receive virtqueue */
148 struct receive_queue {
149 	/* Virtqueue associated with this receive_queue */
150 	struct virtqueue *vq;
151 
152 	struct napi_struct napi;
153 
154 	struct bpf_prog __rcu *xdp_prog;
155 
156 	struct virtnet_rq_stats stats;
157 
158 	/* Chain pages by the private ptr. */
159 	struct page *pages;
160 
161 	/* Average packet length for mergeable receive buffers. */
162 	struct ewma_pkt_len mrg_avg_pkt_len;
163 
164 	/* Page frag for packet buffer allocation. */
165 	struct page_frag alloc_frag;
166 
167 	/* RX: fragments + linear part + virtio header */
168 	struct scatterlist sg[MAX_SKB_FRAGS + 2];
169 
170 	/* Min single buffer size for mergeable buffers case. */
171 	unsigned int min_buf_len;
172 
173 	/* Name of this receive queue: input.$index */
174 	char name[16];
175 
176 	struct xdp_rxq_info xdp_rxq;
177 };
178 
179 /* This structure can contain rss message with maximum settings for indirection table and keysize
180  * Note, that default structure that describes RSS configuration virtio_net_rss_config
181  * contains same info but can't handle table values.
182  * In any case, structure would be passed to virtio hw through sg_buf split by parts
183  * because table sizes may be differ according to the device configuration.
184  */
185 #define VIRTIO_NET_RSS_MAX_KEY_SIZE     40
186 #define VIRTIO_NET_RSS_MAX_TABLE_LEN    128
187 struct virtio_net_ctrl_rss {
188 	u32 hash_types;
189 	u16 indirection_table_mask;
190 	u16 unclassified_queue;
191 	u16 indirection_table[VIRTIO_NET_RSS_MAX_TABLE_LEN];
192 	u16 max_tx_vq;
193 	u8 hash_key_length;
194 	u8 key[VIRTIO_NET_RSS_MAX_KEY_SIZE];
195 };
196 
197 /* Control VQ buffers: protected by the rtnl lock */
198 struct control_buf {
199 	struct virtio_net_ctrl_hdr hdr;
200 	virtio_net_ctrl_ack status;
201 	struct virtio_net_ctrl_mq mq;
202 	u8 promisc;
203 	u8 allmulti;
204 	__virtio16 vid;
205 	__virtio64 offloads;
206 	struct virtio_net_ctrl_rss rss;
207 };
208 
209 struct virtnet_info {
210 	struct virtio_device *vdev;
211 	struct virtqueue *cvq;
212 	struct net_device *dev;
213 	struct send_queue *sq;
214 	struct receive_queue *rq;
215 	unsigned int status;
216 
217 	/* Max # of queue pairs supported by the device */
218 	u16 max_queue_pairs;
219 
220 	/* # of queue pairs currently used by the driver */
221 	u16 curr_queue_pairs;
222 
223 	/* # of XDP queue pairs currently used by the driver */
224 	u16 xdp_queue_pairs;
225 
226 	/* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
227 	bool xdp_enabled;
228 
229 	/* I like... big packets and I cannot lie! */
230 	bool big_packets;
231 
232 	/* number of sg entries allocated for big packets */
233 	unsigned int big_packets_num_skbfrags;
234 
235 	/* Host will merge rx buffers for big packets (shake it! shake it!) */
236 	bool mergeable_rx_bufs;
237 
238 	/* Host supports rss and/or hash report */
239 	bool has_rss;
240 	bool has_rss_hash_report;
241 	u8 rss_key_size;
242 	u16 rss_indir_table_size;
243 	u32 rss_hash_types_supported;
244 	u32 rss_hash_types_saved;
245 
246 	/* Has control virtqueue */
247 	bool has_cvq;
248 
249 	/* Host can handle any s/g split between our header and packet data */
250 	bool any_header_sg;
251 
252 	/* Packet virtio header size */
253 	u8 hdr_len;
254 
255 	/* Work struct for delayed refilling if we run low on memory. */
256 	struct delayed_work refill;
257 
258 	/* Is delayed refill enabled? */
259 	bool refill_enabled;
260 
261 	/* The lock to synchronize the access to refill_enabled */
262 	spinlock_t refill_lock;
263 
264 	/* Work struct for config space updates */
265 	struct work_struct config_work;
266 
267 	/* Does the affinity hint is set for virtqueues? */
268 	bool affinity_hint_set;
269 
270 	/* CPU hotplug instances for online & dead */
271 	struct hlist_node node;
272 	struct hlist_node node_dead;
273 
274 	struct control_buf *ctrl;
275 
276 	/* Ethtool settings */
277 	u8 duplex;
278 	u32 speed;
279 
280 	/* Interrupt coalescing settings */
281 	u32 tx_usecs;
282 	u32 rx_usecs;
283 	u32 tx_max_packets;
284 	u32 rx_max_packets;
285 
286 	unsigned long guest_offloads;
287 	unsigned long guest_offloads_capable;
288 
289 	/* failover when STANDBY feature enabled */
290 	struct failover *failover;
291 };
292 
293 struct padded_vnet_hdr {
294 	struct virtio_net_hdr_v1_hash hdr;
295 	/*
296 	 * hdr is in a separate sg buffer, and data sg buffer shares same page
297 	 * with this header sg. This padding makes next sg 16 byte aligned
298 	 * after the header.
299 	 */
300 	char padding[12];
301 };
302 
303 static void virtnet_rq_free_unused_buf(struct virtqueue *vq, void *buf);
304 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf);
305 
306 static bool is_xdp_frame(void *ptr)
307 {
308 	return (unsigned long)ptr & VIRTIO_XDP_FLAG;
309 }
310 
311 static void *xdp_to_ptr(struct xdp_frame *ptr)
312 {
313 	return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
314 }
315 
316 static struct xdp_frame *ptr_to_xdp(void *ptr)
317 {
318 	return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
319 }
320 
321 /* Converting between virtqueue no. and kernel tx/rx queue no.
322  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
323  */
324 static int vq2txq(struct virtqueue *vq)
325 {
326 	return (vq->index - 1) / 2;
327 }
328 
329 static int txq2vq(int txq)
330 {
331 	return txq * 2 + 1;
332 }
333 
334 static int vq2rxq(struct virtqueue *vq)
335 {
336 	return vq->index / 2;
337 }
338 
339 static int rxq2vq(int rxq)
340 {
341 	return rxq * 2;
342 }
343 
344 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
345 {
346 	return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
347 }
348 
349 /*
350  * private is used to chain pages for big packets, put the whole
351  * most recent used list in the beginning for reuse
352  */
353 static void give_pages(struct receive_queue *rq, struct page *page)
354 {
355 	struct page *end;
356 
357 	/* Find end of list, sew whole thing into vi->rq.pages. */
358 	for (end = page; end->private; end = (struct page *)end->private);
359 	end->private = (unsigned long)rq->pages;
360 	rq->pages = page;
361 }
362 
363 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
364 {
365 	struct page *p = rq->pages;
366 
367 	if (p) {
368 		rq->pages = (struct page *)p->private;
369 		/* clear private here, it is used to chain pages */
370 		p->private = 0;
371 	} else
372 		p = alloc_page(gfp_mask);
373 	return p;
374 }
375 
376 static void enable_delayed_refill(struct virtnet_info *vi)
377 {
378 	spin_lock_bh(&vi->refill_lock);
379 	vi->refill_enabled = true;
380 	spin_unlock_bh(&vi->refill_lock);
381 }
382 
383 static void disable_delayed_refill(struct virtnet_info *vi)
384 {
385 	spin_lock_bh(&vi->refill_lock);
386 	vi->refill_enabled = false;
387 	spin_unlock_bh(&vi->refill_lock);
388 }
389 
390 static void virtqueue_napi_schedule(struct napi_struct *napi,
391 				    struct virtqueue *vq)
392 {
393 	if (napi_schedule_prep(napi)) {
394 		virtqueue_disable_cb(vq);
395 		__napi_schedule(napi);
396 	}
397 }
398 
399 static void virtqueue_napi_complete(struct napi_struct *napi,
400 				    struct virtqueue *vq, int processed)
401 {
402 	int opaque;
403 
404 	opaque = virtqueue_enable_cb_prepare(vq);
405 	if (napi_complete_done(napi, processed)) {
406 		if (unlikely(virtqueue_poll(vq, opaque)))
407 			virtqueue_napi_schedule(napi, vq);
408 	} else {
409 		virtqueue_disable_cb(vq);
410 	}
411 }
412 
413 static void skb_xmit_done(struct virtqueue *vq)
414 {
415 	struct virtnet_info *vi = vq->vdev->priv;
416 	struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
417 
418 	/* Suppress further interrupts. */
419 	virtqueue_disable_cb(vq);
420 
421 	if (napi->weight)
422 		virtqueue_napi_schedule(napi, vq);
423 	else
424 		/* We were probably waiting for more output buffers. */
425 		netif_wake_subqueue(vi->dev, vq2txq(vq));
426 }
427 
428 #define MRG_CTX_HEADER_SHIFT 22
429 static void *mergeable_len_to_ctx(unsigned int truesize,
430 				  unsigned int headroom)
431 {
432 	return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
433 }
434 
435 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
436 {
437 	return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
438 }
439 
440 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
441 {
442 	return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
443 }
444 
445 /* Called from bottom half context */
446 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
447 				   struct receive_queue *rq,
448 				   struct page *page, unsigned int offset,
449 				   unsigned int len, unsigned int truesize)
450 {
451 	struct sk_buff *skb;
452 	struct virtio_net_hdr_mrg_rxbuf *hdr;
453 	unsigned int copy, hdr_len, hdr_padded_len;
454 	struct page *page_to_free = NULL;
455 	int tailroom, shinfo_size;
456 	char *p, *hdr_p, *buf;
457 
458 	p = page_address(page) + offset;
459 	hdr_p = p;
460 
461 	hdr_len = vi->hdr_len;
462 	if (vi->mergeable_rx_bufs)
463 		hdr_padded_len = hdr_len;
464 	else
465 		hdr_padded_len = sizeof(struct padded_vnet_hdr);
466 
467 	buf = p;
468 	len -= hdr_len;
469 	offset += hdr_padded_len;
470 	p += hdr_padded_len;
471 	tailroom = truesize - hdr_padded_len - len;
472 
473 	shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
474 
475 	/* copy small packet so we can reuse these pages */
476 	if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
477 		skb = build_skb(buf, truesize);
478 		if (unlikely(!skb))
479 			return NULL;
480 
481 		skb_reserve(skb, p - buf);
482 		skb_put(skb, len);
483 
484 		page = (struct page *)page->private;
485 		if (page)
486 			give_pages(rq, page);
487 		goto ok;
488 	}
489 
490 	/* copy small packet so we can reuse these pages for small data */
491 	skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
492 	if (unlikely(!skb))
493 		return NULL;
494 
495 	/* Copy all frame if it fits skb->head, otherwise
496 	 * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
497 	 */
498 	if (len <= skb_tailroom(skb))
499 		copy = len;
500 	else
501 		copy = ETH_HLEN;
502 	skb_put_data(skb, p, copy);
503 
504 	len -= copy;
505 	offset += copy;
506 
507 	if (vi->mergeable_rx_bufs) {
508 		if (len)
509 			skb_add_rx_frag(skb, 0, page, offset, len, truesize);
510 		else
511 			page_to_free = page;
512 		goto ok;
513 	}
514 
515 	/*
516 	 * Verify that we can indeed put this data into a skb.
517 	 * This is here to handle cases when the device erroneously
518 	 * tries to receive more than is possible. This is usually
519 	 * the case of a broken device.
520 	 */
521 	if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
522 		net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
523 		dev_kfree_skb(skb);
524 		return NULL;
525 	}
526 	BUG_ON(offset >= PAGE_SIZE);
527 	while (len) {
528 		unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
529 		skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
530 				frag_size, truesize);
531 		len -= frag_size;
532 		page = (struct page *)page->private;
533 		offset = 0;
534 	}
535 
536 	if (page)
537 		give_pages(rq, page);
538 
539 ok:
540 	hdr = skb_vnet_hdr(skb);
541 	memcpy(hdr, hdr_p, hdr_len);
542 	if (page_to_free)
543 		put_page(page_to_free);
544 
545 	return skb;
546 }
547 
548 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
549 				   struct send_queue *sq,
550 				   struct xdp_frame *xdpf)
551 {
552 	struct virtio_net_hdr_mrg_rxbuf *hdr;
553 	struct skb_shared_info *shinfo;
554 	u8 nr_frags = 0;
555 	int err, i;
556 
557 	if (unlikely(xdpf->headroom < vi->hdr_len))
558 		return -EOVERFLOW;
559 
560 	if (unlikely(xdp_frame_has_frags(xdpf))) {
561 		shinfo = xdp_get_shared_info_from_frame(xdpf);
562 		nr_frags = shinfo->nr_frags;
563 	}
564 
565 	/* In wrapping function virtnet_xdp_xmit(), we need to free
566 	 * up the pending old buffers, where we need to calculate the
567 	 * position of skb_shared_info in xdp_get_frame_len() and
568 	 * xdp_return_frame(), which will involve to xdpf->data and
569 	 * xdpf->headroom. Therefore, we need to update the value of
570 	 * headroom synchronously here.
571 	 */
572 	xdpf->headroom -= vi->hdr_len;
573 	xdpf->data -= vi->hdr_len;
574 	/* Zero header and leave csum up to XDP layers */
575 	hdr = xdpf->data;
576 	memset(hdr, 0, vi->hdr_len);
577 	xdpf->len   += vi->hdr_len;
578 
579 	sg_init_table(sq->sg, nr_frags + 1);
580 	sg_set_buf(sq->sg, xdpf->data, xdpf->len);
581 	for (i = 0; i < nr_frags; i++) {
582 		skb_frag_t *frag = &shinfo->frags[i];
583 
584 		sg_set_page(&sq->sg[i + 1], skb_frag_page(frag),
585 			    skb_frag_size(frag), skb_frag_off(frag));
586 	}
587 
588 	err = virtqueue_add_outbuf(sq->vq, sq->sg, nr_frags + 1,
589 				   xdp_to_ptr(xdpf), GFP_ATOMIC);
590 	if (unlikely(err))
591 		return -ENOSPC; /* Caller handle free/refcnt */
592 
593 	return 0;
594 }
595 
596 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
597  * the current cpu, so it does not need to be locked.
598  *
599  * Here we use marco instead of inline functions because we have to deal with
600  * three issues at the same time: 1. the choice of sq. 2. judge and execute the
601  * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
602  * functions to perfectly solve these three problems at the same time.
603  */
604 #define virtnet_xdp_get_sq(vi) ({                                       \
605 	int cpu = smp_processor_id();                                   \
606 	struct netdev_queue *txq;                                       \
607 	typeof(vi) v = (vi);                                            \
608 	unsigned int qp;                                                \
609 									\
610 	if (v->curr_queue_pairs > nr_cpu_ids) {                         \
611 		qp = v->curr_queue_pairs - v->xdp_queue_pairs;          \
612 		qp += cpu;                                              \
613 		txq = netdev_get_tx_queue(v->dev, qp);                  \
614 		__netif_tx_acquire(txq);                                \
615 	} else {                                                        \
616 		qp = cpu % v->curr_queue_pairs;                         \
617 		txq = netdev_get_tx_queue(v->dev, qp);                  \
618 		__netif_tx_lock(txq, cpu);                              \
619 	}                                                               \
620 	v->sq + qp;                                                     \
621 })
622 
623 #define virtnet_xdp_put_sq(vi, q) {                                     \
624 	struct netdev_queue *txq;                                       \
625 	typeof(vi) v = (vi);                                            \
626 									\
627 	txq = netdev_get_tx_queue(v->dev, (q) - v->sq);                 \
628 	if (v->curr_queue_pairs > nr_cpu_ids)                           \
629 		__netif_tx_release(txq);                                \
630 	else                                                            \
631 		__netif_tx_unlock(txq);                                 \
632 }
633 
634 static int virtnet_xdp_xmit(struct net_device *dev,
635 			    int n, struct xdp_frame **frames, u32 flags)
636 {
637 	struct virtnet_info *vi = netdev_priv(dev);
638 	struct receive_queue *rq = vi->rq;
639 	struct bpf_prog *xdp_prog;
640 	struct send_queue *sq;
641 	unsigned int len;
642 	int packets = 0;
643 	int bytes = 0;
644 	int nxmit = 0;
645 	int kicks = 0;
646 	void *ptr;
647 	int ret;
648 	int i;
649 
650 	/* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
651 	 * indicate XDP resources have been successfully allocated.
652 	 */
653 	xdp_prog = rcu_access_pointer(rq->xdp_prog);
654 	if (!xdp_prog)
655 		return -ENXIO;
656 
657 	sq = virtnet_xdp_get_sq(vi);
658 
659 	if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
660 		ret = -EINVAL;
661 		goto out;
662 	}
663 
664 	/* Free up any pending old buffers before queueing new ones. */
665 	while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
666 		if (likely(is_xdp_frame(ptr))) {
667 			struct xdp_frame *frame = ptr_to_xdp(ptr);
668 
669 			bytes += xdp_get_frame_len(frame);
670 			xdp_return_frame(frame);
671 		} else {
672 			struct sk_buff *skb = ptr;
673 
674 			bytes += skb->len;
675 			napi_consume_skb(skb, false);
676 		}
677 		packets++;
678 	}
679 
680 	for (i = 0; i < n; i++) {
681 		struct xdp_frame *xdpf = frames[i];
682 
683 		if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
684 			break;
685 		nxmit++;
686 	}
687 	ret = nxmit;
688 
689 	if (flags & XDP_XMIT_FLUSH) {
690 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
691 			kicks = 1;
692 	}
693 out:
694 	u64_stats_update_begin(&sq->stats.syncp);
695 	sq->stats.bytes += bytes;
696 	sq->stats.packets += packets;
697 	sq->stats.xdp_tx += n;
698 	sq->stats.xdp_tx_drops += n - nxmit;
699 	sq->stats.kicks += kicks;
700 	u64_stats_update_end(&sq->stats.syncp);
701 
702 	virtnet_xdp_put_sq(vi, sq);
703 	return ret;
704 }
705 
706 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
707 {
708 	return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
709 }
710 
711 /* We copy the packet for XDP in the following cases:
712  *
713  * 1) Packet is scattered across multiple rx buffers.
714  * 2) Headroom space is insufficient.
715  *
716  * This is inefficient but it's a temporary condition that
717  * we hit right after XDP is enabled and until queue is refilled
718  * with large buffers with sufficient headroom - so it should affect
719  * at most queue size packets.
720  * Afterwards, the conditions to enable
721  * XDP should preclude the underlying device from sending packets
722  * across multiple buffers (num_buf > 1), and we make sure buffers
723  * have enough headroom.
724  */
725 static struct page *xdp_linearize_page(struct receive_queue *rq,
726 				       int *num_buf,
727 				       struct page *p,
728 				       int offset,
729 				       int page_off,
730 				       unsigned int *len)
731 {
732 	struct page *page = alloc_page(GFP_ATOMIC);
733 
734 	if (!page)
735 		return NULL;
736 
737 	memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
738 	page_off += *len;
739 
740 	while (--*num_buf) {
741 		int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
742 		unsigned int buflen;
743 		void *buf;
744 		int off;
745 
746 		buf = virtqueue_get_buf(rq->vq, &buflen);
747 		if (unlikely(!buf))
748 			goto err_buf;
749 
750 		p = virt_to_head_page(buf);
751 		off = buf - page_address(p);
752 
753 		/* guard against a misconfigured or uncooperative backend that
754 		 * is sending packet larger than the MTU.
755 		 */
756 		if ((page_off + buflen + tailroom) > PAGE_SIZE) {
757 			put_page(p);
758 			goto err_buf;
759 		}
760 
761 		memcpy(page_address(page) + page_off,
762 		       page_address(p) + off, buflen);
763 		page_off += buflen;
764 		put_page(p);
765 	}
766 
767 	/* Headroom does not contribute to packet length */
768 	*len = page_off - VIRTIO_XDP_HEADROOM;
769 	return page;
770 err_buf:
771 	__free_pages(page, 0);
772 	return NULL;
773 }
774 
775 static struct sk_buff *receive_small(struct net_device *dev,
776 				     struct virtnet_info *vi,
777 				     struct receive_queue *rq,
778 				     void *buf, void *ctx,
779 				     unsigned int len,
780 				     unsigned int *xdp_xmit,
781 				     struct virtnet_rq_stats *stats)
782 {
783 	struct sk_buff *skb;
784 	struct bpf_prog *xdp_prog;
785 	unsigned int xdp_headroom = (unsigned long)ctx;
786 	unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
787 	unsigned int headroom = vi->hdr_len + header_offset;
788 	unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
789 			      SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
790 	struct page *page = virt_to_head_page(buf);
791 	unsigned int delta = 0;
792 	struct page *xdp_page;
793 	int err;
794 	unsigned int metasize = 0;
795 
796 	len -= vi->hdr_len;
797 	stats->bytes += len;
798 
799 	if (unlikely(len > GOOD_PACKET_LEN)) {
800 		pr_debug("%s: rx error: len %u exceeds max size %d\n",
801 			 dev->name, len, GOOD_PACKET_LEN);
802 		dev->stats.rx_length_errors++;
803 		goto err;
804 	}
805 
806 	if (likely(!vi->xdp_enabled)) {
807 		xdp_prog = NULL;
808 		goto skip_xdp;
809 	}
810 
811 	rcu_read_lock();
812 	xdp_prog = rcu_dereference(rq->xdp_prog);
813 	if (xdp_prog) {
814 		struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
815 		struct xdp_frame *xdpf;
816 		struct xdp_buff xdp;
817 		void *orig_data;
818 		u32 act;
819 
820 		if (unlikely(hdr->hdr.gso_type))
821 			goto err_xdp;
822 
823 		if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
824 			int offset = buf - page_address(page) + header_offset;
825 			unsigned int tlen = len + vi->hdr_len;
826 			int num_buf = 1;
827 
828 			xdp_headroom = virtnet_get_headroom(vi);
829 			header_offset = VIRTNET_RX_PAD + xdp_headroom;
830 			headroom = vi->hdr_len + header_offset;
831 			buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
832 				 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
833 			xdp_page = xdp_linearize_page(rq, &num_buf, page,
834 						      offset, header_offset,
835 						      &tlen);
836 			if (!xdp_page)
837 				goto err_xdp;
838 
839 			buf = page_address(xdp_page);
840 			put_page(page);
841 			page = xdp_page;
842 		}
843 
844 		xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
845 		xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
846 				 xdp_headroom, len, true);
847 		orig_data = xdp.data;
848 		act = bpf_prog_run_xdp(xdp_prog, &xdp);
849 		stats->xdp_packets++;
850 
851 		switch (act) {
852 		case XDP_PASS:
853 			/* Recalculate length in case bpf program changed it */
854 			delta = orig_data - xdp.data;
855 			len = xdp.data_end - xdp.data;
856 			metasize = xdp.data - xdp.data_meta;
857 			break;
858 		case XDP_TX:
859 			stats->xdp_tx++;
860 			xdpf = xdp_convert_buff_to_frame(&xdp);
861 			if (unlikely(!xdpf))
862 				goto err_xdp;
863 			err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
864 			if (unlikely(!err)) {
865 				xdp_return_frame_rx_napi(xdpf);
866 			} else if (unlikely(err < 0)) {
867 				trace_xdp_exception(vi->dev, xdp_prog, act);
868 				goto err_xdp;
869 			}
870 			*xdp_xmit |= VIRTIO_XDP_TX;
871 			rcu_read_unlock();
872 			goto xdp_xmit;
873 		case XDP_REDIRECT:
874 			stats->xdp_redirects++;
875 			err = xdp_do_redirect(dev, &xdp, xdp_prog);
876 			if (err)
877 				goto err_xdp;
878 			*xdp_xmit |= VIRTIO_XDP_REDIR;
879 			rcu_read_unlock();
880 			goto xdp_xmit;
881 		default:
882 			bpf_warn_invalid_xdp_action(vi->dev, xdp_prog, act);
883 			fallthrough;
884 		case XDP_ABORTED:
885 			trace_xdp_exception(vi->dev, xdp_prog, act);
886 			goto err_xdp;
887 		case XDP_DROP:
888 			goto err_xdp;
889 		}
890 	}
891 	rcu_read_unlock();
892 
893 skip_xdp:
894 	skb = build_skb(buf, buflen);
895 	if (!skb)
896 		goto err;
897 	skb_reserve(skb, headroom - delta);
898 	skb_put(skb, len);
899 	if (!xdp_prog) {
900 		buf += header_offset;
901 		memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
902 	} /* keep zeroed vnet hdr since XDP is loaded */
903 
904 	if (metasize)
905 		skb_metadata_set(skb, metasize);
906 
907 	return skb;
908 
909 err_xdp:
910 	rcu_read_unlock();
911 	stats->xdp_drops++;
912 err:
913 	stats->drops++;
914 	put_page(page);
915 xdp_xmit:
916 	return NULL;
917 }
918 
919 static struct sk_buff *receive_big(struct net_device *dev,
920 				   struct virtnet_info *vi,
921 				   struct receive_queue *rq,
922 				   void *buf,
923 				   unsigned int len,
924 				   struct virtnet_rq_stats *stats)
925 {
926 	struct page *page = buf;
927 	struct sk_buff *skb =
928 		page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
929 
930 	stats->bytes += len - vi->hdr_len;
931 	if (unlikely(!skb))
932 		goto err;
933 
934 	return skb;
935 
936 err:
937 	stats->drops++;
938 	give_pages(rq, page);
939 	return NULL;
940 }
941 
942 /* Why not use xdp_build_skb_from_frame() ?
943  * XDP core assumes that xdp frags are PAGE_SIZE in length, while in
944  * virtio-net there are 2 points that do not match its requirements:
945  *  1. The size of the prefilled buffer is not fixed before xdp is set.
946  *  2. xdp_build_skb_from_frame() does more checks that we don't need,
947  *     like eth_type_trans() (which virtio-net does in receive_buf()).
948  */
949 static struct sk_buff *build_skb_from_xdp_buff(struct net_device *dev,
950 					       struct virtnet_info *vi,
951 					       struct xdp_buff *xdp,
952 					       unsigned int xdp_frags_truesz)
953 {
954 	struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
955 	unsigned int headroom, data_len;
956 	struct sk_buff *skb;
957 	int metasize;
958 	u8 nr_frags;
959 
960 	if (unlikely(xdp->data_end > xdp_data_hard_end(xdp))) {
961 		pr_debug("Error building skb as missing reserved tailroom for xdp");
962 		return NULL;
963 	}
964 
965 	if (unlikely(xdp_buff_has_frags(xdp)))
966 		nr_frags = sinfo->nr_frags;
967 
968 	skb = build_skb(xdp->data_hard_start, xdp->frame_sz);
969 	if (unlikely(!skb))
970 		return NULL;
971 
972 	headroom = xdp->data - xdp->data_hard_start;
973 	data_len = xdp->data_end - xdp->data;
974 	skb_reserve(skb, headroom);
975 	__skb_put(skb, data_len);
976 
977 	metasize = xdp->data - xdp->data_meta;
978 	metasize = metasize > 0 ? metasize : 0;
979 	if (metasize)
980 		skb_metadata_set(skb, metasize);
981 
982 	if (unlikely(xdp_buff_has_frags(xdp)))
983 		xdp_update_skb_shared_info(skb, nr_frags,
984 					   sinfo->xdp_frags_size,
985 					   xdp_frags_truesz,
986 					   xdp_buff_is_frag_pfmemalloc(xdp));
987 
988 	return skb;
989 }
990 
991 /* TODO: build xdp in big mode */
992 static int virtnet_build_xdp_buff_mrg(struct net_device *dev,
993 				      struct virtnet_info *vi,
994 				      struct receive_queue *rq,
995 				      struct xdp_buff *xdp,
996 				      void *buf,
997 				      unsigned int len,
998 				      unsigned int frame_sz,
999 				      int *num_buf,
1000 				      unsigned int *xdp_frags_truesize,
1001 				      struct virtnet_rq_stats *stats)
1002 {
1003 	struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1004 	unsigned int headroom, tailroom, room;
1005 	unsigned int truesize, cur_frag_size;
1006 	struct skb_shared_info *shinfo;
1007 	unsigned int xdp_frags_truesz = 0;
1008 	struct page *page;
1009 	skb_frag_t *frag;
1010 	int offset;
1011 	void *ctx;
1012 
1013 	xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq);
1014 	xdp_prepare_buff(xdp, buf - VIRTIO_XDP_HEADROOM,
1015 			 VIRTIO_XDP_HEADROOM + vi->hdr_len, len - vi->hdr_len, true);
1016 
1017 	if (!*num_buf)
1018 		return 0;
1019 
1020 	if (*num_buf > 1) {
1021 		/* If we want to build multi-buffer xdp, we need
1022 		 * to specify that the flags of xdp_buff have the
1023 		 * XDP_FLAGS_HAS_FRAG bit.
1024 		 */
1025 		if (!xdp_buff_has_frags(xdp))
1026 			xdp_buff_set_frags_flag(xdp);
1027 
1028 		shinfo = xdp_get_shared_info_from_buff(xdp);
1029 		shinfo->nr_frags = 0;
1030 		shinfo->xdp_frags_size = 0;
1031 	}
1032 
1033 	if (*num_buf > MAX_SKB_FRAGS + 1)
1034 		return -EINVAL;
1035 
1036 	while (--*num_buf > 0) {
1037 		buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1038 		if (unlikely(!buf)) {
1039 			pr_debug("%s: rx error: %d buffers out of %d missing\n",
1040 				 dev->name, *num_buf,
1041 				 virtio16_to_cpu(vi->vdev, hdr->num_buffers));
1042 			dev->stats.rx_length_errors++;
1043 			return -EINVAL;
1044 		}
1045 
1046 		stats->bytes += len;
1047 		page = virt_to_head_page(buf);
1048 		offset = buf - page_address(page);
1049 
1050 		truesize = mergeable_ctx_to_truesize(ctx);
1051 		headroom = mergeable_ctx_to_headroom(ctx);
1052 		tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1053 		room = SKB_DATA_ALIGN(headroom + tailroom);
1054 
1055 		cur_frag_size = truesize;
1056 		xdp_frags_truesz += cur_frag_size;
1057 		if (unlikely(len > truesize - room || cur_frag_size > PAGE_SIZE)) {
1058 			put_page(page);
1059 			pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1060 				 dev->name, len, (unsigned long)(truesize - room));
1061 			dev->stats.rx_length_errors++;
1062 			return -EINVAL;
1063 		}
1064 
1065 		frag = &shinfo->frags[shinfo->nr_frags++];
1066 		__skb_frag_set_page(frag, page);
1067 		skb_frag_off_set(frag, offset);
1068 		skb_frag_size_set(frag, len);
1069 		if (page_is_pfmemalloc(page))
1070 			xdp_buff_set_frag_pfmemalloc(xdp);
1071 
1072 		shinfo->xdp_frags_size += len;
1073 	}
1074 
1075 	*xdp_frags_truesize = xdp_frags_truesz;
1076 	return 0;
1077 }
1078 
1079 static struct sk_buff *receive_mergeable(struct net_device *dev,
1080 					 struct virtnet_info *vi,
1081 					 struct receive_queue *rq,
1082 					 void *buf,
1083 					 void *ctx,
1084 					 unsigned int len,
1085 					 unsigned int *xdp_xmit,
1086 					 struct virtnet_rq_stats *stats)
1087 {
1088 	struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1089 	int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
1090 	struct page *page = virt_to_head_page(buf);
1091 	int offset = buf - page_address(page);
1092 	struct sk_buff *head_skb, *curr_skb;
1093 	struct bpf_prog *xdp_prog;
1094 	unsigned int truesize = mergeable_ctx_to_truesize(ctx);
1095 	unsigned int headroom = mergeable_ctx_to_headroom(ctx);
1096 	unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1097 	unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1098 	unsigned int frame_sz, xdp_room;
1099 	int err;
1100 
1101 	head_skb = NULL;
1102 	stats->bytes += len - vi->hdr_len;
1103 
1104 	if (unlikely(len > truesize - room)) {
1105 		pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1106 			 dev->name, len, (unsigned long)(truesize - room));
1107 		dev->stats.rx_length_errors++;
1108 		goto err_skb;
1109 	}
1110 
1111 	if (likely(!vi->xdp_enabled)) {
1112 		xdp_prog = NULL;
1113 		goto skip_xdp;
1114 	}
1115 
1116 	rcu_read_lock();
1117 	xdp_prog = rcu_dereference(rq->xdp_prog);
1118 	if (xdp_prog) {
1119 		unsigned int xdp_frags_truesz = 0;
1120 		struct skb_shared_info *shinfo;
1121 		struct xdp_frame *xdpf;
1122 		struct page *xdp_page;
1123 		struct xdp_buff xdp;
1124 		void *data;
1125 		u32 act;
1126 		int i;
1127 
1128 		/* Transient failure which in theory could occur if
1129 		 * in-flight packets from before XDP was enabled reach
1130 		 * the receive path after XDP is loaded.
1131 		 */
1132 		if (unlikely(hdr->hdr.gso_type))
1133 			goto err_xdp;
1134 
1135 		/* Now XDP core assumes frag size is PAGE_SIZE, but buffers
1136 		 * with headroom may add hole in truesize, which
1137 		 * make their length exceed PAGE_SIZE. So we disabled the
1138 		 * hole mechanism for xdp. See add_recvbuf_mergeable().
1139 		 */
1140 		frame_sz = truesize;
1141 
1142 		/* This happens when headroom is not enough because
1143 		 * of the buffer was prefilled before XDP is set.
1144 		 * This should only happen for the first several packets.
1145 		 * In fact, vq reset can be used here to help us clean up
1146 		 * the prefilled buffers, but many existing devices do not
1147 		 * support it, and we don't want to bother users who are
1148 		 * using xdp normally.
1149 		 */
1150 		if (!xdp_prog->aux->xdp_has_frags &&
1151 		    (num_buf > 1 || headroom < virtnet_get_headroom(vi))) {
1152 			/* linearize data for XDP */
1153 			xdp_page = xdp_linearize_page(rq, &num_buf,
1154 						      page, offset,
1155 						      VIRTIO_XDP_HEADROOM,
1156 						      &len);
1157 			frame_sz = PAGE_SIZE;
1158 
1159 			if (!xdp_page)
1160 				goto err_xdp;
1161 			offset = VIRTIO_XDP_HEADROOM;
1162 		} else if (unlikely(headroom < virtnet_get_headroom(vi))) {
1163 			xdp_room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM +
1164 						  sizeof(struct skb_shared_info));
1165 			if (len + xdp_room > PAGE_SIZE)
1166 				goto err_xdp;
1167 
1168 			xdp_page = alloc_page(GFP_ATOMIC);
1169 			if (!xdp_page)
1170 				goto err_xdp;
1171 
1172 			memcpy(page_address(xdp_page) + VIRTIO_XDP_HEADROOM,
1173 			       page_address(page) + offset, len);
1174 			frame_sz = PAGE_SIZE;
1175 			offset = VIRTIO_XDP_HEADROOM;
1176 		} else {
1177 			xdp_page = page;
1178 		}
1179 
1180 		data = page_address(xdp_page) + offset;
1181 		err = virtnet_build_xdp_buff_mrg(dev, vi, rq, &xdp, data, len, frame_sz,
1182 						 &num_buf, &xdp_frags_truesz, stats);
1183 		if (unlikely(err))
1184 			goto err_xdp_frags;
1185 
1186 		act = bpf_prog_run_xdp(xdp_prog, &xdp);
1187 		stats->xdp_packets++;
1188 
1189 		switch (act) {
1190 		case XDP_PASS:
1191 			if (unlikely(xdp_page != page))
1192 				put_page(page);
1193 			head_skb = build_skb_from_xdp_buff(dev, vi, &xdp, xdp_frags_truesz);
1194 			rcu_read_unlock();
1195 			return head_skb;
1196 		case XDP_TX:
1197 			stats->xdp_tx++;
1198 			xdpf = xdp_convert_buff_to_frame(&xdp);
1199 			if (unlikely(!xdpf)) {
1200 				netdev_dbg(dev, "convert buff to frame failed for xdp\n");
1201 				goto err_xdp_frags;
1202 			}
1203 			err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
1204 			if (unlikely(!err)) {
1205 				xdp_return_frame_rx_napi(xdpf);
1206 			} else if (unlikely(err < 0)) {
1207 				trace_xdp_exception(vi->dev, xdp_prog, act);
1208 				goto err_xdp_frags;
1209 			}
1210 			*xdp_xmit |= VIRTIO_XDP_TX;
1211 			if (unlikely(xdp_page != page))
1212 				put_page(page);
1213 			rcu_read_unlock();
1214 			goto xdp_xmit;
1215 		case XDP_REDIRECT:
1216 			stats->xdp_redirects++;
1217 			err = xdp_do_redirect(dev, &xdp, xdp_prog);
1218 			if (err)
1219 				goto err_xdp_frags;
1220 			*xdp_xmit |= VIRTIO_XDP_REDIR;
1221 			if (unlikely(xdp_page != page))
1222 				put_page(page);
1223 			rcu_read_unlock();
1224 			goto xdp_xmit;
1225 		default:
1226 			bpf_warn_invalid_xdp_action(vi->dev, xdp_prog, act);
1227 			fallthrough;
1228 		case XDP_ABORTED:
1229 			trace_xdp_exception(vi->dev, xdp_prog, act);
1230 			fallthrough;
1231 		case XDP_DROP:
1232 			goto err_xdp_frags;
1233 		}
1234 err_xdp_frags:
1235 		if (unlikely(xdp_page != page))
1236 			__free_pages(xdp_page, 0);
1237 
1238 		if (xdp_buff_has_frags(&xdp)) {
1239 			shinfo = xdp_get_shared_info_from_buff(&xdp);
1240 			for (i = 0; i < shinfo->nr_frags; i++) {
1241 				xdp_page = skb_frag_page(&shinfo->frags[i]);
1242 				put_page(xdp_page);
1243 			}
1244 		}
1245 
1246 		goto err_xdp;
1247 	}
1248 	rcu_read_unlock();
1249 
1250 skip_xdp:
1251 	head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
1252 	curr_skb = head_skb;
1253 
1254 	if (unlikely(!curr_skb))
1255 		goto err_skb;
1256 	while (--num_buf) {
1257 		int num_skb_frags;
1258 
1259 		buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1260 		if (unlikely(!buf)) {
1261 			pr_debug("%s: rx error: %d buffers out of %d missing\n",
1262 				 dev->name, num_buf,
1263 				 virtio16_to_cpu(vi->vdev,
1264 						 hdr->num_buffers));
1265 			dev->stats.rx_length_errors++;
1266 			goto err_buf;
1267 		}
1268 
1269 		stats->bytes += len;
1270 		page = virt_to_head_page(buf);
1271 
1272 		truesize = mergeable_ctx_to_truesize(ctx);
1273 		headroom = mergeable_ctx_to_headroom(ctx);
1274 		tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1275 		room = SKB_DATA_ALIGN(headroom + tailroom);
1276 		if (unlikely(len > truesize - room)) {
1277 			pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1278 				 dev->name, len, (unsigned long)(truesize - room));
1279 			dev->stats.rx_length_errors++;
1280 			goto err_skb;
1281 		}
1282 
1283 		num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1284 		if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1285 			struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1286 
1287 			if (unlikely(!nskb))
1288 				goto err_skb;
1289 			if (curr_skb == head_skb)
1290 				skb_shinfo(curr_skb)->frag_list = nskb;
1291 			else
1292 				curr_skb->next = nskb;
1293 			curr_skb = nskb;
1294 			head_skb->truesize += nskb->truesize;
1295 			num_skb_frags = 0;
1296 		}
1297 		if (curr_skb != head_skb) {
1298 			head_skb->data_len += len;
1299 			head_skb->len += len;
1300 			head_skb->truesize += truesize;
1301 		}
1302 		offset = buf - page_address(page);
1303 		if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1304 			put_page(page);
1305 			skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1306 					     len, truesize);
1307 		} else {
1308 			skb_add_rx_frag(curr_skb, num_skb_frags, page,
1309 					offset, len, truesize);
1310 		}
1311 	}
1312 
1313 	ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1314 	return head_skb;
1315 
1316 err_xdp:
1317 	rcu_read_unlock();
1318 	stats->xdp_drops++;
1319 err_skb:
1320 	put_page(page);
1321 	while (num_buf-- > 1) {
1322 		buf = virtqueue_get_buf(rq->vq, &len);
1323 		if (unlikely(!buf)) {
1324 			pr_debug("%s: rx error: %d buffers missing\n",
1325 				 dev->name, num_buf);
1326 			dev->stats.rx_length_errors++;
1327 			break;
1328 		}
1329 		stats->bytes += len;
1330 		page = virt_to_head_page(buf);
1331 		put_page(page);
1332 	}
1333 err_buf:
1334 	stats->drops++;
1335 	dev_kfree_skb(head_skb);
1336 xdp_xmit:
1337 	return NULL;
1338 }
1339 
1340 static void virtio_skb_set_hash(const struct virtio_net_hdr_v1_hash *hdr_hash,
1341 				struct sk_buff *skb)
1342 {
1343 	enum pkt_hash_types rss_hash_type;
1344 
1345 	if (!hdr_hash || !skb)
1346 		return;
1347 
1348 	switch (__le16_to_cpu(hdr_hash->hash_report)) {
1349 	case VIRTIO_NET_HASH_REPORT_TCPv4:
1350 	case VIRTIO_NET_HASH_REPORT_UDPv4:
1351 	case VIRTIO_NET_HASH_REPORT_TCPv6:
1352 	case VIRTIO_NET_HASH_REPORT_UDPv6:
1353 	case VIRTIO_NET_HASH_REPORT_TCPv6_EX:
1354 	case VIRTIO_NET_HASH_REPORT_UDPv6_EX:
1355 		rss_hash_type = PKT_HASH_TYPE_L4;
1356 		break;
1357 	case VIRTIO_NET_HASH_REPORT_IPv4:
1358 	case VIRTIO_NET_HASH_REPORT_IPv6:
1359 	case VIRTIO_NET_HASH_REPORT_IPv6_EX:
1360 		rss_hash_type = PKT_HASH_TYPE_L3;
1361 		break;
1362 	case VIRTIO_NET_HASH_REPORT_NONE:
1363 	default:
1364 		rss_hash_type = PKT_HASH_TYPE_NONE;
1365 	}
1366 	skb_set_hash(skb, __le32_to_cpu(hdr_hash->hash_value), rss_hash_type);
1367 }
1368 
1369 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1370 			void *buf, unsigned int len, void **ctx,
1371 			unsigned int *xdp_xmit,
1372 			struct virtnet_rq_stats *stats)
1373 {
1374 	struct net_device *dev = vi->dev;
1375 	struct sk_buff *skb;
1376 	struct virtio_net_hdr_mrg_rxbuf *hdr;
1377 
1378 	if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1379 		pr_debug("%s: short packet %i\n", dev->name, len);
1380 		dev->stats.rx_length_errors++;
1381 		virtnet_rq_free_unused_buf(rq->vq, buf);
1382 		return;
1383 	}
1384 
1385 	if (vi->mergeable_rx_bufs)
1386 		skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1387 					stats);
1388 	else if (vi->big_packets)
1389 		skb = receive_big(dev, vi, rq, buf, len, stats);
1390 	else
1391 		skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1392 
1393 	if (unlikely(!skb))
1394 		return;
1395 
1396 	hdr = skb_vnet_hdr(skb);
1397 	if (dev->features & NETIF_F_RXHASH && vi->has_rss_hash_report)
1398 		virtio_skb_set_hash((const struct virtio_net_hdr_v1_hash *)hdr, skb);
1399 
1400 	if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1401 		skb->ip_summed = CHECKSUM_UNNECESSARY;
1402 
1403 	if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1404 				  virtio_is_little_endian(vi->vdev))) {
1405 		net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1406 				     dev->name, hdr->hdr.gso_type,
1407 				     hdr->hdr.gso_size);
1408 		goto frame_err;
1409 	}
1410 
1411 	skb_record_rx_queue(skb, vq2rxq(rq->vq));
1412 	skb->protocol = eth_type_trans(skb, dev);
1413 	pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1414 		 ntohs(skb->protocol), skb->len, skb->pkt_type);
1415 
1416 	napi_gro_receive(&rq->napi, skb);
1417 	return;
1418 
1419 frame_err:
1420 	dev->stats.rx_frame_errors++;
1421 	dev_kfree_skb(skb);
1422 }
1423 
1424 /* Unlike mergeable buffers, all buffers are allocated to the
1425  * same size, except for the headroom. For this reason we do
1426  * not need to use  mergeable_len_to_ctx here - it is enough
1427  * to store the headroom as the context ignoring the truesize.
1428  */
1429 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1430 			     gfp_t gfp)
1431 {
1432 	struct page_frag *alloc_frag = &rq->alloc_frag;
1433 	char *buf;
1434 	unsigned int xdp_headroom = virtnet_get_headroom(vi);
1435 	void *ctx = (void *)(unsigned long)xdp_headroom;
1436 	int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1437 	int err;
1438 
1439 	len = SKB_DATA_ALIGN(len) +
1440 	      SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1441 	if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1442 		return -ENOMEM;
1443 
1444 	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1445 	get_page(alloc_frag->page);
1446 	alloc_frag->offset += len;
1447 	sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1448 		    vi->hdr_len + GOOD_PACKET_LEN);
1449 	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1450 	if (err < 0)
1451 		put_page(virt_to_head_page(buf));
1452 	return err;
1453 }
1454 
1455 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1456 			   gfp_t gfp)
1457 {
1458 	struct page *first, *list = NULL;
1459 	char *p;
1460 	int i, err, offset;
1461 
1462 	sg_init_table(rq->sg, vi->big_packets_num_skbfrags + 2);
1463 
1464 	/* page in rq->sg[vi->big_packets_num_skbfrags + 1] is list tail */
1465 	for (i = vi->big_packets_num_skbfrags + 1; i > 1; --i) {
1466 		first = get_a_page(rq, gfp);
1467 		if (!first) {
1468 			if (list)
1469 				give_pages(rq, list);
1470 			return -ENOMEM;
1471 		}
1472 		sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1473 
1474 		/* chain new page in list head to match sg */
1475 		first->private = (unsigned long)list;
1476 		list = first;
1477 	}
1478 
1479 	first = get_a_page(rq, gfp);
1480 	if (!first) {
1481 		give_pages(rq, list);
1482 		return -ENOMEM;
1483 	}
1484 	p = page_address(first);
1485 
1486 	/* rq->sg[0], rq->sg[1] share the same page */
1487 	/* a separated rq->sg[0] for header - required in case !any_header_sg */
1488 	sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1489 
1490 	/* rq->sg[1] for data packet, from offset */
1491 	offset = sizeof(struct padded_vnet_hdr);
1492 	sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1493 
1494 	/* chain first in list head */
1495 	first->private = (unsigned long)list;
1496 	err = virtqueue_add_inbuf(rq->vq, rq->sg, vi->big_packets_num_skbfrags + 2,
1497 				  first, gfp);
1498 	if (err < 0)
1499 		give_pages(rq, first);
1500 
1501 	return err;
1502 }
1503 
1504 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1505 					  struct ewma_pkt_len *avg_pkt_len,
1506 					  unsigned int room)
1507 {
1508 	struct virtnet_info *vi = rq->vq->vdev->priv;
1509 	const size_t hdr_len = vi->hdr_len;
1510 	unsigned int len;
1511 
1512 	if (room)
1513 		return PAGE_SIZE - room;
1514 
1515 	len = hdr_len +	clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1516 				rq->min_buf_len, PAGE_SIZE - hdr_len);
1517 
1518 	return ALIGN(len, L1_CACHE_BYTES);
1519 }
1520 
1521 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1522 				 struct receive_queue *rq, gfp_t gfp)
1523 {
1524 	struct page_frag *alloc_frag = &rq->alloc_frag;
1525 	unsigned int headroom = virtnet_get_headroom(vi);
1526 	unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1527 	unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1528 	char *buf;
1529 	void *ctx;
1530 	int err;
1531 	unsigned int len, hole;
1532 
1533 	/* Extra tailroom is needed to satisfy XDP's assumption. This
1534 	 * means rx frags coalescing won't work, but consider we've
1535 	 * disabled GSO for XDP, it won't be a big issue.
1536 	 */
1537 	len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1538 	if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1539 		return -ENOMEM;
1540 
1541 	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1542 	buf += headroom; /* advance address leaving hole at front of pkt */
1543 	get_page(alloc_frag->page);
1544 	alloc_frag->offset += len + room;
1545 	hole = alloc_frag->size - alloc_frag->offset;
1546 	if (hole < len + room) {
1547 		/* To avoid internal fragmentation, if there is very likely not
1548 		 * enough space for another buffer, add the remaining space to
1549 		 * the current buffer.
1550 		 * XDP core assumes that frame_size of xdp_buff and the length
1551 		 * of the frag are PAGE_SIZE, so we disable the hole mechanism.
1552 		 */
1553 		if (!headroom)
1554 			len += hole;
1555 		alloc_frag->offset += hole;
1556 	}
1557 
1558 	sg_init_one(rq->sg, buf, len);
1559 	ctx = mergeable_len_to_ctx(len + room, headroom);
1560 	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1561 	if (err < 0)
1562 		put_page(virt_to_head_page(buf));
1563 
1564 	return err;
1565 }
1566 
1567 /*
1568  * Returns false if we couldn't fill entirely (OOM).
1569  *
1570  * Normally run in the receive path, but can also be run from ndo_open
1571  * before we're receiving packets, or from refill_work which is
1572  * careful to disable receiving (using napi_disable).
1573  */
1574 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1575 			  gfp_t gfp)
1576 {
1577 	int err;
1578 	bool oom;
1579 
1580 	do {
1581 		if (vi->mergeable_rx_bufs)
1582 			err = add_recvbuf_mergeable(vi, rq, gfp);
1583 		else if (vi->big_packets)
1584 			err = add_recvbuf_big(vi, rq, gfp);
1585 		else
1586 			err = add_recvbuf_small(vi, rq, gfp);
1587 
1588 		oom = err == -ENOMEM;
1589 		if (err)
1590 			break;
1591 	} while (rq->vq->num_free);
1592 	if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1593 		unsigned long flags;
1594 
1595 		flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1596 		rq->stats.kicks++;
1597 		u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1598 	}
1599 
1600 	return !oom;
1601 }
1602 
1603 static void skb_recv_done(struct virtqueue *rvq)
1604 {
1605 	struct virtnet_info *vi = rvq->vdev->priv;
1606 	struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1607 
1608 	virtqueue_napi_schedule(&rq->napi, rvq);
1609 }
1610 
1611 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1612 {
1613 	napi_enable(napi);
1614 
1615 	/* If all buffers were filled by other side before we napi_enabled, we
1616 	 * won't get another interrupt, so process any outstanding packets now.
1617 	 * Call local_bh_enable after to trigger softIRQ processing.
1618 	 */
1619 	local_bh_disable();
1620 	virtqueue_napi_schedule(napi, vq);
1621 	local_bh_enable();
1622 }
1623 
1624 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1625 				   struct virtqueue *vq,
1626 				   struct napi_struct *napi)
1627 {
1628 	if (!napi->weight)
1629 		return;
1630 
1631 	/* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1632 	 * enable the feature if this is likely affine with the transmit path.
1633 	 */
1634 	if (!vi->affinity_hint_set) {
1635 		napi->weight = 0;
1636 		return;
1637 	}
1638 
1639 	return virtnet_napi_enable(vq, napi);
1640 }
1641 
1642 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1643 {
1644 	if (napi->weight)
1645 		napi_disable(napi);
1646 }
1647 
1648 static void refill_work(struct work_struct *work)
1649 {
1650 	struct virtnet_info *vi =
1651 		container_of(work, struct virtnet_info, refill.work);
1652 	bool still_empty;
1653 	int i;
1654 
1655 	for (i = 0; i < vi->curr_queue_pairs; i++) {
1656 		struct receive_queue *rq = &vi->rq[i];
1657 
1658 		napi_disable(&rq->napi);
1659 		still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1660 		virtnet_napi_enable(rq->vq, &rq->napi);
1661 
1662 		/* In theory, this can happen: if we don't get any buffers in
1663 		 * we will *never* try to fill again.
1664 		 */
1665 		if (still_empty)
1666 			schedule_delayed_work(&vi->refill, HZ/2);
1667 	}
1668 }
1669 
1670 static int virtnet_receive(struct receive_queue *rq, int budget,
1671 			   unsigned int *xdp_xmit)
1672 {
1673 	struct virtnet_info *vi = rq->vq->vdev->priv;
1674 	struct virtnet_rq_stats stats = {};
1675 	unsigned int len;
1676 	void *buf;
1677 	int i;
1678 
1679 	if (!vi->big_packets || vi->mergeable_rx_bufs) {
1680 		void *ctx;
1681 
1682 		while (stats.packets < budget &&
1683 		       (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1684 			receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1685 			stats.packets++;
1686 		}
1687 	} else {
1688 		while (stats.packets < budget &&
1689 		       (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1690 			receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1691 			stats.packets++;
1692 		}
1693 	}
1694 
1695 	if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1696 		if (!try_fill_recv(vi, rq, GFP_ATOMIC)) {
1697 			spin_lock(&vi->refill_lock);
1698 			if (vi->refill_enabled)
1699 				schedule_delayed_work(&vi->refill, 0);
1700 			spin_unlock(&vi->refill_lock);
1701 		}
1702 	}
1703 
1704 	u64_stats_update_begin(&rq->stats.syncp);
1705 	for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1706 		size_t offset = virtnet_rq_stats_desc[i].offset;
1707 		u64 *item;
1708 
1709 		item = (u64 *)((u8 *)&rq->stats + offset);
1710 		*item += *(u64 *)((u8 *)&stats + offset);
1711 	}
1712 	u64_stats_update_end(&rq->stats.syncp);
1713 
1714 	return stats.packets;
1715 }
1716 
1717 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1718 {
1719 	unsigned int len;
1720 	unsigned int packets = 0;
1721 	unsigned int bytes = 0;
1722 	void *ptr;
1723 
1724 	while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1725 		if (likely(!is_xdp_frame(ptr))) {
1726 			struct sk_buff *skb = ptr;
1727 
1728 			pr_debug("Sent skb %p\n", skb);
1729 
1730 			bytes += skb->len;
1731 			napi_consume_skb(skb, in_napi);
1732 		} else {
1733 			struct xdp_frame *frame = ptr_to_xdp(ptr);
1734 
1735 			bytes += xdp_get_frame_len(frame);
1736 			xdp_return_frame(frame);
1737 		}
1738 		packets++;
1739 	}
1740 
1741 	/* Avoid overhead when no packets have been processed
1742 	 * happens when called speculatively from start_xmit.
1743 	 */
1744 	if (!packets)
1745 		return;
1746 
1747 	u64_stats_update_begin(&sq->stats.syncp);
1748 	sq->stats.bytes += bytes;
1749 	sq->stats.packets += packets;
1750 	u64_stats_update_end(&sq->stats.syncp);
1751 }
1752 
1753 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1754 {
1755 	if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1756 		return false;
1757 	else if (q < vi->curr_queue_pairs)
1758 		return true;
1759 	else
1760 		return false;
1761 }
1762 
1763 static void virtnet_poll_cleantx(struct receive_queue *rq)
1764 {
1765 	struct virtnet_info *vi = rq->vq->vdev->priv;
1766 	unsigned int index = vq2rxq(rq->vq);
1767 	struct send_queue *sq = &vi->sq[index];
1768 	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1769 
1770 	if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1771 		return;
1772 
1773 	if (__netif_tx_trylock(txq)) {
1774 		if (sq->reset) {
1775 			__netif_tx_unlock(txq);
1776 			return;
1777 		}
1778 
1779 		do {
1780 			virtqueue_disable_cb(sq->vq);
1781 			free_old_xmit_skbs(sq, true);
1782 		} while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1783 
1784 		if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1785 			netif_tx_wake_queue(txq);
1786 
1787 		__netif_tx_unlock(txq);
1788 	}
1789 }
1790 
1791 static int virtnet_poll(struct napi_struct *napi, int budget)
1792 {
1793 	struct receive_queue *rq =
1794 		container_of(napi, struct receive_queue, napi);
1795 	struct virtnet_info *vi = rq->vq->vdev->priv;
1796 	struct send_queue *sq;
1797 	unsigned int received;
1798 	unsigned int xdp_xmit = 0;
1799 
1800 	virtnet_poll_cleantx(rq);
1801 
1802 	received = virtnet_receive(rq, budget, &xdp_xmit);
1803 
1804 	if (xdp_xmit & VIRTIO_XDP_REDIR)
1805 		xdp_do_flush();
1806 
1807 	/* Out of packets? */
1808 	if (received < budget)
1809 		virtqueue_napi_complete(napi, rq->vq, received);
1810 
1811 	if (xdp_xmit & VIRTIO_XDP_TX) {
1812 		sq = virtnet_xdp_get_sq(vi);
1813 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1814 			u64_stats_update_begin(&sq->stats.syncp);
1815 			sq->stats.kicks++;
1816 			u64_stats_update_end(&sq->stats.syncp);
1817 		}
1818 		virtnet_xdp_put_sq(vi, sq);
1819 	}
1820 
1821 	return received;
1822 }
1823 
1824 static int virtnet_open(struct net_device *dev)
1825 {
1826 	struct virtnet_info *vi = netdev_priv(dev);
1827 	int i, err;
1828 
1829 	enable_delayed_refill(vi);
1830 
1831 	for (i = 0; i < vi->max_queue_pairs; i++) {
1832 		if (i < vi->curr_queue_pairs)
1833 			/* Make sure we have some buffers: if oom use wq. */
1834 			if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1835 				schedule_delayed_work(&vi->refill, 0);
1836 
1837 		err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id);
1838 		if (err < 0)
1839 			return err;
1840 
1841 		err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1842 						 MEM_TYPE_PAGE_SHARED, NULL);
1843 		if (err < 0) {
1844 			xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1845 			return err;
1846 		}
1847 
1848 		virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1849 		virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1850 	}
1851 
1852 	return 0;
1853 }
1854 
1855 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1856 {
1857 	struct send_queue *sq = container_of(napi, struct send_queue, napi);
1858 	struct virtnet_info *vi = sq->vq->vdev->priv;
1859 	unsigned int index = vq2txq(sq->vq);
1860 	struct netdev_queue *txq;
1861 	int opaque;
1862 	bool done;
1863 
1864 	if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1865 		/* We don't need to enable cb for XDP */
1866 		napi_complete_done(napi, 0);
1867 		return 0;
1868 	}
1869 
1870 	txq = netdev_get_tx_queue(vi->dev, index);
1871 	__netif_tx_lock(txq, raw_smp_processor_id());
1872 	virtqueue_disable_cb(sq->vq);
1873 	free_old_xmit_skbs(sq, true);
1874 
1875 	if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1876 		netif_tx_wake_queue(txq);
1877 
1878 	opaque = virtqueue_enable_cb_prepare(sq->vq);
1879 
1880 	done = napi_complete_done(napi, 0);
1881 
1882 	if (!done)
1883 		virtqueue_disable_cb(sq->vq);
1884 
1885 	__netif_tx_unlock(txq);
1886 
1887 	if (done) {
1888 		if (unlikely(virtqueue_poll(sq->vq, opaque))) {
1889 			if (napi_schedule_prep(napi)) {
1890 				__netif_tx_lock(txq, raw_smp_processor_id());
1891 				virtqueue_disable_cb(sq->vq);
1892 				__netif_tx_unlock(txq);
1893 				__napi_schedule(napi);
1894 			}
1895 		}
1896 	}
1897 
1898 	return 0;
1899 }
1900 
1901 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1902 {
1903 	struct virtio_net_hdr_mrg_rxbuf *hdr;
1904 	const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1905 	struct virtnet_info *vi = sq->vq->vdev->priv;
1906 	int num_sg;
1907 	unsigned hdr_len = vi->hdr_len;
1908 	bool can_push;
1909 
1910 	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1911 
1912 	can_push = vi->any_header_sg &&
1913 		!((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1914 		!skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1915 	/* Even if we can, don't push here yet as this would skew
1916 	 * csum_start offset below. */
1917 	if (can_push)
1918 		hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1919 	else
1920 		hdr = skb_vnet_hdr(skb);
1921 
1922 	if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1923 				    virtio_is_little_endian(vi->vdev), false,
1924 				    0))
1925 		return -EPROTO;
1926 
1927 	if (vi->mergeable_rx_bufs)
1928 		hdr->num_buffers = 0;
1929 
1930 	sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1931 	if (can_push) {
1932 		__skb_push(skb, hdr_len);
1933 		num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1934 		if (unlikely(num_sg < 0))
1935 			return num_sg;
1936 		/* Pull header back to avoid skew in tx bytes calculations. */
1937 		__skb_pull(skb, hdr_len);
1938 	} else {
1939 		sg_set_buf(sq->sg, hdr, hdr_len);
1940 		num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1941 		if (unlikely(num_sg < 0))
1942 			return num_sg;
1943 		num_sg++;
1944 	}
1945 	return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1946 }
1947 
1948 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1949 {
1950 	struct virtnet_info *vi = netdev_priv(dev);
1951 	int qnum = skb_get_queue_mapping(skb);
1952 	struct send_queue *sq = &vi->sq[qnum];
1953 	int err;
1954 	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1955 	bool kick = !netdev_xmit_more();
1956 	bool use_napi = sq->napi.weight;
1957 
1958 	/* Free up any pending old buffers before queueing new ones. */
1959 	do {
1960 		if (use_napi)
1961 			virtqueue_disable_cb(sq->vq);
1962 
1963 		free_old_xmit_skbs(sq, false);
1964 
1965 	} while (use_napi && kick &&
1966 	       unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1967 
1968 	/* timestamp packet in software */
1969 	skb_tx_timestamp(skb);
1970 
1971 	/* Try to transmit */
1972 	err = xmit_skb(sq, skb);
1973 
1974 	/* This should not happen! */
1975 	if (unlikely(err)) {
1976 		dev->stats.tx_fifo_errors++;
1977 		if (net_ratelimit())
1978 			dev_warn(&dev->dev,
1979 				 "Unexpected TXQ (%d) queue failure: %d\n",
1980 				 qnum, err);
1981 		dev->stats.tx_dropped++;
1982 		dev_kfree_skb_any(skb);
1983 		return NETDEV_TX_OK;
1984 	}
1985 
1986 	/* Don't wait up for transmitted skbs to be freed. */
1987 	if (!use_napi) {
1988 		skb_orphan(skb);
1989 		nf_reset_ct(skb);
1990 	}
1991 
1992 	/* If running out of space, stop queue to avoid getting packets that we
1993 	 * are then unable to transmit.
1994 	 * An alternative would be to force queuing layer to requeue the skb by
1995 	 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1996 	 * returned in a normal path of operation: it means that driver is not
1997 	 * maintaining the TX queue stop/start state properly, and causes
1998 	 * the stack to do a non-trivial amount of useless work.
1999 	 * Since most packets only take 1 or 2 ring slots, stopping the queue
2000 	 * early means 16 slots are typically wasted.
2001 	 */
2002 	if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
2003 		netif_stop_subqueue(dev, qnum);
2004 		if (use_napi) {
2005 			if (unlikely(!virtqueue_enable_cb_delayed(sq->vq)))
2006 				virtqueue_napi_schedule(&sq->napi, sq->vq);
2007 		} else if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
2008 			/* More just got used, free them then recheck. */
2009 			free_old_xmit_skbs(sq, false);
2010 			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
2011 				netif_start_subqueue(dev, qnum);
2012 				virtqueue_disable_cb(sq->vq);
2013 			}
2014 		}
2015 	}
2016 
2017 	if (kick || netif_xmit_stopped(txq)) {
2018 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
2019 			u64_stats_update_begin(&sq->stats.syncp);
2020 			sq->stats.kicks++;
2021 			u64_stats_update_end(&sq->stats.syncp);
2022 		}
2023 	}
2024 
2025 	return NETDEV_TX_OK;
2026 }
2027 
2028 static int virtnet_rx_resize(struct virtnet_info *vi,
2029 			     struct receive_queue *rq, u32 ring_num)
2030 {
2031 	bool running = netif_running(vi->dev);
2032 	int err, qindex;
2033 
2034 	qindex = rq - vi->rq;
2035 
2036 	if (running)
2037 		napi_disable(&rq->napi);
2038 
2039 	err = virtqueue_resize(rq->vq, ring_num, virtnet_rq_free_unused_buf);
2040 	if (err)
2041 		netdev_err(vi->dev, "resize rx fail: rx queue index: %d err: %d\n", qindex, err);
2042 
2043 	if (!try_fill_recv(vi, rq, GFP_KERNEL))
2044 		schedule_delayed_work(&vi->refill, 0);
2045 
2046 	if (running)
2047 		virtnet_napi_enable(rq->vq, &rq->napi);
2048 	return err;
2049 }
2050 
2051 static int virtnet_tx_resize(struct virtnet_info *vi,
2052 			     struct send_queue *sq, u32 ring_num)
2053 {
2054 	bool running = netif_running(vi->dev);
2055 	struct netdev_queue *txq;
2056 	int err, qindex;
2057 
2058 	qindex = sq - vi->sq;
2059 
2060 	if (running)
2061 		virtnet_napi_tx_disable(&sq->napi);
2062 
2063 	txq = netdev_get_tx_queue(vi->dev, qindex);
2064 
2065 	/* 1. wait all ximt complete
2066 	 * 2. fix the race of netif_stop_subqueue() vs netif_start_subqueue()
2067 	 */
2068 	__netif_tx_lock_bh(txq);
2069 
2070 	/* Prevent rx poll from accessing sq. */
2071 	sq->reset = true;
2072 
2073 	/* Prevent the upper layer from trying to send packets. */
2074 	netif_stop_subqueue(vi->dev, qindex);
2075 
2076 	__netif_tx_unlock_bh(txq);
2077 
2078 	err = virtqueue_resize(sq->vq, ring_num, virtnet_sq_free_unused_buf);
2079 	if (err)
2080 		netdev_err(vi->dev, "resize tx fail: tx queue index: %d err: %d\n", qindex, err);
2081 
2082 	__netif_tx_lock_bh(txq);
2083 	sq->reset = false;
2084 	netif_tx_wake_queue(txq);
2085 	__netif_tx_unlock_bh(txq);
2086 
2087 	if (running)
2088 		virtnet_napi_tx_enable(vi, sq->vq, &sq->napi);
2089 	return err;
2090 }
2091 
2092 /*
2093  * Send command via the control virtqueue and check status.  Commands
2094  * supported by the hypervisor, as indicated by feature bits, should
2095  * never fail unless improperly formatted.
2096  */
2097 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
2098 				 struct scatterlist *out)
2099 {
2100 	struct scatterlist *sgs[4], hdr, stat;
2101 	unsigned out_num = 0, tmp;
2102 	int ret;
2103 
2104 	/* Caller should know better */
2105 	BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
2106 
2107 	vi->ctrl->status = ~0;
2108 	vi->ctrl->hdr.class = class;
2109 	vi->ctrl->hdr.cmd = cmd;
2110 	/* Add header */
2111 	sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
2112 	sgs[out_num++] = &hdr;
2113 
2114 	if (out)
2115 		sgs[out_num++] = out;
2116 
2117 	/* Add return status. */
2118 	sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
2119 	sgs[out_num] = &stat;
2120 
2121 	BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
2122 	ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
2123 	if (ret < 0) {
2124 		dev_warn(&vi->vdev->dev,
2125 			 "Failed to add sgs for command vq: %d\n.", ret);
2126 		return false;
2127 	}
2128 
2129 	if (unlikely(!virtqueue_kick(vi->cvq)))
2130 		return vi->ctrl->status == VIRTIO_NET_OK;
2131 
2132 	/* Spin for a response, the kick causes an ioport write, trapping
2133 	 * into the hypervisor, so the request should be handled immediately.
2134 	 */
2135 	while (!virtqueue_get_buf(vi->cvq, &tmp) &&
2136 	       !virtqueue_is_broken(vi->cvq))
2137 		cpu_relax();
2138 
2139 	return vi->ctrl->status == VIRTIO_NET_OK;
2140 }
2141 
2142 static int virtnet_set_mac_address(struct net_device *dev, void *p)
2143 {
2144 	struct virtnet_info *vi = netdev_priv(dev);
2145 	struct virtio_device *vdev = vi->vdev;
2146 	int ret;
2147 	struct sockaddr *addr;
2148 	struct scatterlist sg;
2149 
2150 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2151 		return -EOPNOTSUPP;
2152 
2153 	addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
2154 	if (!addr)
2155 		return -ENOMEM;
2156 
2157 	ret = eth_prepare_mac_addr_change(dev, addr);
2158 	if (ret)
2159 		goto out;
2160 
2161 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
2162 		sg_init_one(&sg, addr->sa_data, dev->addr_len);
2163 		if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2164 					  VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
2165 			dev_warn(&vdev->dev,
2166 				 "Failed to set mac address by vq command.\n");
2167 			ret = -EINVAL;
2168 			goto out;
2169 		}
2170 	} else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
2171 		   !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
2172 		unsigned int i;
2173 
2174 		/* Naturally, this has an atomicity problem. */
2175 		for (i = 0; i < dev->addr_len; i++)
2176 			virtio_cwrite8(vdev,
2177 				       offsetof(struct virtio_net_config, mac) +
2178 				       i, addr->sa_data[i]);
2179 	}
2180 
2181 	eth_commit_mac_addr_change(dev, p);
2182 	ret = 0;
2183 
2184 out:
2185 	kfree(addr);
2186 	return ret;
2187 }
2188 
2189 static void virtnet_stats(struct net_device *dev,
2190 			  struct rtnl_link_stats64 *tot)
2191 {
2192 	struct virtnet_info *vi = netdev_priv(dev);
2193 	unsigned int start;
2194 	int i;
2195 
2196 	for (i = 0; i < vi->max_queue_pairs; i++) {
2197 		u64 tpackets, tbytes, terrors, rpackets, rbytes, rdrops;
2198 		struct receive_queue *rq = &vi->rq[i];
2199 		struct send_queue *sq = &vi->sq[i];
2200 
2201 		do {
2202 			start = u64_stats_fetch_begin(&sq->stats.syncp);
2203 			tpackets = sq->stats.packets;
2204 			tbytes   = sq->stats.bytes;
2205 			terrors  = sq->stats.tx_timeouts;
2206 		} while (u64_stats_fetch_retry(&sq->stats.syncp, start));
2207 
2208 		do {
2209 			start = u64_stats_fetch_begin(&rq->stats.syncp);
2210 			rpackets = rq->stats.packets;
2211 			rbytes   = rq->stats.bytes;
2212 			rdrops   = rq->stats.drops;
2213 		} while (u64_stats_fetch_retry(&rq->stats.syncp, start));
2214 
2215 		tot->rx_packets += rpackets;
2216 		tot->tx_packets += tpackets;
2217 		tot->rx_bytes   += rbytes;
2218 		tot->tx_bytes   += tbytes;
2219 		tot->rx_dropped += rdrops;
2220 		tot->tx_errors  += terrors;
2221 	}
2222 
2223 	tot->tx_dropped = dev->stats.tx_dropped;
2224 	tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
2225 	tot->rx_length_errors = dev->stats.rx_length_errors;
2226 	tot->rx_frame_errors = dev->stats.rx_frame_errors;
2227 }
2228 
2229 static void virtnet_ack_link_announce(struct virtnet_info *vi)
2230 {
2231 	rtnl_lock();
2232 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
2233 				  VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
2234 		dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
2235 	rtnl_unlock();
2236 }
2237 
2238 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
2239 {
2240 	struct scatterlist sg;
2241 	struct net_device *dev = vi->dev;
2242 
2243 	if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
2244 		return 0;
2245 
2246 	vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
2247 	sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
2248 
2249 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
2250 				  VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
2251 		dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
2252 			 queue_pairs);
2253 		return -EINVAL;
2254 	} else {
2255 		vi->curr_queue_pairs = queue_pairs;
2256 		/* virtnet_open() will refill when device is going to up. */
2257 		if (dev->flags & IFF_UP)
2258 			schedule_delayed_work(&vi->refill, 0);
2259 	}
2260 
2261 	return 0;
2262 }
2263 
2264 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
2265 {
2266 	int err;
2267 
2268 	rtnl_lock();
2269 	err = _virtnet_set_queues(vi, queue_pairs);
2270 	rtnl_unlock();
2271 	return err;
2272 }
2273 
2274 static int virtnet_close(struct net_device *dev)
2275 {
2276 	struct virtnet_info *vi = netdev_priv(dev);
2277 	int i;
2278 
2279 	/* Make sure NAPI doesn't schedule refill work */
2280 	disable_delayed_refill(vi);
2281 	/* Make sure refill_work doesn't re-enable napi! */
2282 	cancel_delayed_work_sync(&vi->refill);
2283 
2284 	for (i = 0; i < vi->max_queue_pairs; i++) {
2285 		virtnet_napi_tx_disable(&vi->sq[i].napi);
2286 		napi_disable(&vi->rq[i].napi);
2287 		xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
2288 	}
2289 
2290 	return 0;
2291 }
2292 
2293 static void virtnet_set_rx_mode(struct net_device *dev)
2294 {
2295 	struct virtnet_info *vi = netdev_priv(dev);
2296 	struct scatterlist sg[2];
2297 	struct virtio_net_ctrl_mac *mac_data;
2298 	struct netdev_hw_addr *ha;
2299 	int uc_count;
2300 	int mc_count;
2301 	void *buf;
2302 	int i;
2303 
2304 	/* We can't dynamically set ndo_set_rx_mode, so return gracefully */
2305 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
2306 		return;
2307 
2308 	vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
2309 	vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
2310 
2311 	sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
2312 
2313 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2314 				  VIRTIO_NET_CTRL_RX_PROMISC, sg))
2315 		dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
2316 			 vi->ctrl->promisc ? "en" : "dis");
2317 
2318 	sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
2319 
2320 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2321 				  VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
2322 		dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
2323 			 vi->ctrl->allmulti ? "en" : "dis");
2324 
2325 	uc_count = netdev_uc_count(dev);
2326 	mc_count = netdev_mc_count(dev);
2327 	/* MAC filter - use one buffer for both lists */
2328 	buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
2329 		      (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
2330 	mac_data = buf;
2331 	if (!buf)
2332 		return;
2333 
2334 	sg_init_table(sg, 2);
2335 
2336 	/* Store the unicast list and count in the front of the buffer */
2337 	mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
2338 	i = 0;
2339 	netdev_for_each_uc_addr(ha, dev)
2340 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2341 
2342 	sg_set_buf(&sg[0], mac_data,
2343 		   sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
2344 
2345 	/* multicast list and count fill the end */
2346 	mac_data = (void *)&mac_data->macs[uc_count][0];
2347 
2348 	mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
2349 	i = 0;
2350 	netdev_for_each_mc_addr(ha, dev)
2351 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2352 
2353 	sg_set_buf(&sg[1], mac_data,
2354 		   sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
2355 
2356 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2357 				  VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
2358 		dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
2359 
2360 	kfree(buf);
2361 }
2362 
2363 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
2364 				   __be16 proto, u16 vid)
2365 {
2366 	struct virtnet_info *vi = netdev_priv(dev);
2367 	struct scatterlist sg;
2368 
2369 	vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2370 	sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2371 
2372 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2373 				  VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2374 		dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2375 	return 0;
2376 }
2377 
2378 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2379 				    __be16 proto, u16 vid)
2380 {
2381 	struct virtnet_info *vi = netdev_priv(dev);
2382 	struct scatterlist sg;
2383 
2384 	vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2385 	sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2386 
2387 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2388 				  VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2389 		dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2390 	return 0;
2391 }
2392 
2393 static void virtnet_clean_affinity(struct virtnet_info *vi)
2394 {
2395 	int i;
2396 
2397 	if (vi->affinity_hint_set) {
2398 		for (i = 0; i < vi->max_queue_pairs; i++) {
2399 			virtqueue_set_affinity(vi->rq[i].vq, NULL);
2400 			virtqueue_set_affinity(vi->sq[i].vq, NULL);
2401 		}
2402 
2403 		vi->affinity_hint_set = false;
2404 	}
2405 }
2406 
2407 static void virtnet_set_affinity(struct virtnet_info *vi)
2408 {
2409 	cpumask_var_t mask;
2410 	int stragglers;
2411 	int group_size;
2412 	int i, j, cpu;
2413 	int num_cpu;
2414 	int stride;
2415 
2416 	if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2417 		virtnet_clean_affinity(vi);
2418 		return;
2419 	}
2420 
2421 	num_cpu = num_online_cpus();
2422 	stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2423 	stragglers = num_cpu >= vi->curr_queue_pairs ?
2424 			num_cpu % vi->curr_queue_pairs :
2425 			0;
2426 	cpu = cpumask_first(cpu_online_mask);
2427 
2428 	for (i = 0; i < vi->curr_queue_pairs; i++) {
2429 		group_size = stride + (i < stragglers ? 1 : 0);
2430 
2431 		for (j = 0; j < group_size; j++) {
2432 			cpumask_set_cpu(cpu, mask);
2433 			cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2434 						nr_cpu_ids, false);
2435 		}
2436 		virtqueue_set_affinity(vi->rq[i].vq, mask);
2437 		virtqueue_set_affinity(vi->sq[i].vq, mask);
2438 		__netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2439 		cpumask_clear(mask);
2440 	}
2441 
2442 	vi->affinity_hint_set = true;
2443 	free_cpumask_var(mask);
2444 }
2445 
2446 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2447 {
2448 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2449 						   node);
2450 	virtnet_set_affinity(vi);
2451 	return 0;
2452 }
2453 
2454 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2455 {
2456 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2457 						   node_dead);
2458 	virtnet_set_affinity(vi);
2459 	return 0;
2460 }
2461 
2462 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2463 {
2464 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2465 						   node);
2466 
2467 	virtnet_clean_affinity(vi);
2468 	return 0;
2469 }
2470 
2471 static enum cpuhp_state virtionet_online;
2472 
2473 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2474 {
2475 	int ret;
2476 
2477 	ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2478 	if (ret)
2479 		return ret;
2480 	ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2481 					       &vi->node_dead);
2482 	if (!ret)
2483 		return ret;
2484 	cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2485 	return ret;
2486 }
2487 
2488 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2489 {
2490 	cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2491 	cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2492 					    &vi->node_dead);
2493 }
2494 
2495 static void virtnet_get_ringparam(struct net_device *dev,
2496 				  struct ethtool_ringparam *ring,
2497 				  struct kernel_ethtool_ringparam *kernel_ring,
2498 				  struct netlink_ext_ack *extack)
2499 {
2500 	struct virtnet_info *vi = netdev_priv(dev);
2501 
2502 	ring->rx_max_pending = vi->rq[0].vq->num_max;
2503 	ring->tx_max_pending = vi->sq[0].vq->num_max;
2504 	ring->rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2505 	ring->tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2506 }
2507 
2508 static int virtnet_set_ringparam(struct net_device *dev,
2509 				 struct ethtool_ringparam *ring,
2510 				 struct kernel_ethtool_ringparam *kernel_ring,
2511 				 struct netlink_ext_ack *extack)
2512 {
2513 	struct virtnet_info *vi = netdev_priv(dev);
2514 	u32 rx_pending, tx_pending;
2515 	struct receive_queue *rq;
2516 	struct send_queue *sq;
2517 	int i, err;
2518 
2519 	if (ring->rx_mini_pending || ring->rx_jumbo_pending)
2520 		return -EINVAL;
2521 
2522 	rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2523 	tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2524 
2525 	if (ring->rx_pending == rx_pending &&
2526 	    ring->tx_pending == tx_pending)
2527 		return 0;
2528 
2529 	if (ring->rx_pending > vi->rq[0].vq->num_max)
2530 		return -EINVAL;
2531 
2532 	if (ring->tx_pending > vi->sq[0].vq->num_max)
2533 		return -EINVAL;
2534 
2535 	for (i = 0; i < vi->max_queue_pairs; i++) {
2536 		rq = vi->rq + i;
2537 		sq = vi->sq + i;
2538 
2539 		if (ring->tx_pending != tx_pending) {
2540 			err = virtnet_tx_resize(vi, sq, ring->tx_pending);
2541 			if (err)
2542 				return err;
2543 		}
2544 
2545 		if (ring->rx_pending != rx_pending) {
2546 			err = virtnet_rx_resize(vi, rq, ring->rx_pending);
2547 			if (err)
2548 				return err;
2549 		}
2550 	}
2551 
2552 	return 0;
2553 }
2554 
2555 static bool virtnet_commit_rss_command(struct virtnet_info *vi)
2556 {
2557 	struct net_device *dev = vi->dev;
2558 	struct scatterlist sgs[4];
2559 	unsigned int sg_buf_size;
2560 
2561 	/* prepare sgs */
2562 	sg_init_table(sgs, 4);
2563 
2564 	sg_buf_size = offsetof(struct virtio_net_ctrl_rss, indirection_table);
2565 	sg_set_buf(&sgs[0], &vi->ctrl->rss, sg_buf_size);
2566 
2567 	sg_buf_size = sizeof(uint16_t) * (vi->ctrl->rss.indirection_table_mask + 1);
2568 	sg_set_buf(&sgs[1], vi->ctrl->rss.indirection_table, sg_buf_size);
2569 
2570 	sg_buf_size = offsetof(struct virtio_net_ctrl_rss, key)
2571 			- offsetof(struct virtio_net_ctrl_rss, max_tx_vq);
2572 	sg_set_buf(&sgs[2], &vi->ctrl->rss.max_tx_vq, sg_buf_size);
2573 
2574 	sg_buf_size = vi->rss_key_size;
2575 	sg_set_buf(&sgs[3], vi->ctrl->rss.key, sg_buf_size);
2576 
2577 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
2578 				  vi->has_rss ? VIRTIO_NET_CTRL_MQ_RSS_CONFIG
2579 				  : VIRTIO_NET_CTRL_MQ_HASH_CONFIG, sgs)) {
2580 		dev_warn(&dev->dev, "VIRTIONET issue with committing RSS sgs\n");
2581 		return false;
2582 	}
2583 	return true;
2584 }
2585 
2586 static void virtnet_init_default_rss(struct virtnet_info *vi)
2587 {
2588 	u32 indir_val = 0;
2589 	int i = 0;
2590 
2591 	vi->ctrl->rss.hash_types = vi->rss_hash_types_supported;
2592 	vi->rss_hash_types_saved = vi->rss_hash_types_supported;
2593 	vi->ctrl->rss.indirection_table_mask = vi->rss_indir_table_size
2594 						? vi->rss_indir_table_size - 1 : 0;
2595 	vi->ctrl->rss.unclassified_queue = 0;
2596 
2597 	for (; i < vi->rss_indir_table_size; ++i) {
2598 		indir_val = ethtool_rxfh_indir_default(i, vi->curr_queue_pairs);
2599 		vi->ctrl->rss.indirection_table[i] = indir_val;
2600 	}
2601 
2602 	vi->ctrl->rss.max_tx_vq = vi->curr_queue_pairs;
2603 	vi->ctrl->rss.hash_key_length = vi->rss_key_size;
2604 
2605 	netdev_rss_key_fill(vi->ctrl->rss.key, vi->rss_key_size);
2606 }
2607 
2608 static void virtnet_get_hashflow(const struct virtnet_info *vi, struct ethtool_rxnfc *info)
2609 {
2610 	info->data = 0;
2611 	switch (info->flow_type) {
2612 	case TCP_V4_FLOW:
2613 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv4) {
2614 			info->data = RXH_IP_SRC | RXH_IP_DST |
2615 						 RXH_L4_B_0_1 | RXH_L4_B_2_3;
2616 		} else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
2617 			info->data = RXH_IP_SRC | RXH_IP_DST;
2618 		}
2619 		break;
2620 	case TCP_V6_FLOW:
2621 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv6) {
2622 			info->data = RXH_IP_SRC | RXH_IP_DST |
2623 						 RXH_L4_B_0_1 | RXH_L4_B_2_3;
2624 		} else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
2625 			info->data = RXH_IP_SRC | RXH_IP_DST;
2626 		}
2627 		break;
2628 	case UDP_V4_FLOW:
2629 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv4) {
2630 			info->data = RXH_IP_SRC | RXH_IP_DST |
2631 						 RXH_L4_B_0_1 | RXH_L4_B_2_3;
2632 		} else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
2633 			info->data = RXH_IP_SRC | RXH_IP_DST;
2634 		}
2635 		break;
2636 	case UDP_V6_FLOW:
2637 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv6) {
2638 			info->data = RXH_IP_SRC | RXH_IP_DST |
2639 						 RXH_L4_B_0_1 | RXH_L4_B_2_3;
2640 		} else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
2641 			info->data = RXH_IP_SRC | RXH_IP_DST;
2642 		}
2643 		break;
2644 	case IPV4_FLOW:
2645 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4)
2646 			info->data = RXH_IP_SRC | RXH_IP_DST;
2647 
2648 		break;
2649 	case IPV6_FLOW:
2650 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6)
2651 			info->data = RXH_IP_SRC | RXH_IP_DST;
2652 
2653 		break;
2654 	default:
2655 		info->data = 0;
2656 		break;
2657 	}
2658 }
2659 
2660 static bool virtnet_set_hashflow(struct virtnet_info *vi, struct ethtool_rxnfc *info)
2661 {
2662 	u32 new_hashtypes = vi->rss_hash_types_saved;
2663 	bool is_disable = info->data & RXH_DISCARD;
2664 	bool is_l4 = info->data == (RXH_IP_SRC | RXH_IP_DST | RXH_L4_B_0_1 | RXH_L4_B_2_3);
2665 
2666 	/* supports only 'sd', 'sdfn' and 'r' */
2667 	if (!((info->data == (RXH_IP_SRC | RXH_IP_DST)) | is_l4 | is_disable))
2668 		return false;
2669 
2670 	switch (info->flow_type) {
2671 	case TCP_V4_FLOW:
2672 		new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_TCPv4);
2673 		if (!is_disable)
2674 			new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
2675 				| (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv4 : 0);
2676 		break;
2677 	case UDP_V4_FLOW:
2678 		new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_UDPv4);
2679 		if (!is_disable)
2680 			new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
2681 				| (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv4 : 0);
2682 		break;
2683 	case IPV4_FLOW:
2684 		new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv4;
2685 		if (!is_disable)
2686 			new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv4;
2687 		break;
2688 	case TCP_V6_FLOW:
2689 		new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_TCPv6);
2690 		if (!is_disable)
2691 			new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
2692 				| (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv6 : 0);
2693 		break;
2694 	case UDP_V6_FLOW:
2695 		new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_UDPv6);
2696 		if (!is_disable)
2697 			new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
2698 				| (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv6 : 0);
2699 		break;
2700 	case IPV6_FLOW:
2701 		new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv6;
2702 		if (!is_disable)
2703 			new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv6;
2704 		break;
2705 	default:
2706 		/* unsupported flow */
2707 		return false;
2708 	}
2709 
2710 	/* if unsupported hashtype was set */
2711 	if (new_hashtypes != (new_hashtypes & vi->rss_hash_types_supported))
2712 		return false;
2713 
2714 	if (new_hashtypes != vi->rss_hash_types_saved) {
2715 		vi->rss_hash_types_saved = new_hashtypes;
2716 		vi->ctrl->rss.hash_types = vi->rss_hash_types_saved;
2717 		if (vi->dev->features & NETIF_F_RXHASH)
2718 			return virtnet_commit_rss_command(vi);
2719 	}
2720 
2721 	return true;
2722 }
2723 
2724 static void virtnet_get_drvinfo(struct net_device *dev,
2725 				struct ethtool_drvinfo *info)
2726 {
2727 	struct virtnet_info *vi = netdev_priv(dev);
2728 	struct virtio_device *vdev = vi->vdev;
2729 
2730 	strscpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2731 	strscpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2732 	strscpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2733 
2734 }
2735 
2736 /* TODO: Eliminate OOO packets during switching */
2737 static int virtnet_set_channels(struct net_device *dev,
2738 				struct ethtool_channels *channels)
2739 {
2740 	struct virtnet_info *vi = netdev_priv(dev);
2741 	u16 queue_pairs = channels->combined_count;
2742 	int err;
2743 
2744 	/* We don't support separate rx/tx channels.
2745 	 * We don't allow setting 'other' channels.
2746 	 */
2747 	if (channels->rx_count || channels->tx_count || channels->other_count)
2748 		return -EINVAL;
2749 
2750 	if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2751 		return -EINVAL;
2752 
2753 	/* For now we don't support modifying channels while XDP is loaded
2754 	 * also when XDP is loaded all RX queues have XDP programs so we only
2755 	 * need to check a single RX queue.
2756 	 */
2757 	if (vi->rq[0].xdp_prog)
2758 		return -EINVAL;
2759 
2760 	cpus_read_lock();
2761 	err = _virtnet_set_queues(vi, queue_pairs);
2762 	if (err) {
2763 		cpus_read_unlock();
2764 		goto err;
2765 	}
2766 	virtnet_set_affinity(vi);
2767 	cpus_read_unlock();
2768 
2769 	netif_set_real_num_tx_queues(dev, queue_pairs);
2770 	netif_set_real_num_rx_queues(dev, queue_pairs);
2771  err:
2772 	return err;
2773 }
2774 
2775 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2776 {
2777 	struct virtnet_info *vi = netdev_priv(dev);
2778 	unsigned int i, j;
2779 	u8 *p = data;
2780 
2781 	switch (stringset) {
2782 	case ETH_SS_STATS:
2783 		for (i = 0; i < vi->curr_queue_pairs; i++) {
2784 			for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
2785 				ethtool_sprintf(&p, "rx_queue_%u_%s", i,
2786 						virtnet_rq_stats_desc[j].desc);
2787 		}
2788 
2789 		for (i = 0; i < vi->curr_queue_pairs; i++) {
2790 			for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
2791 				ethtool_sprintf(&p, "tx_queue_%u_%s", i,
2792 						virtnet_sq_stats_desc[j].desc);
2793 		}
2794 		break;
2795 	}
2796 }
2797 
2798 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2799 {
2800 	struct virtnet_info *vi = netdev_priv(dev);
2801 
2802 	switch (sset) {
2803 	case ETH_SS_STATS:
2804 		return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2805 					       VIRTNET_SQ_STATS_LEN);
2806 	default:
2807 		return -EOPNOTSUPP;
2808 	}
2809 }
2810 
2811 static void virtnet_get_ethtool_stats(struct net_device *dev,
2812 				      struct ethtool_stats *stats, u64 *data)
2813 {
2814 	struct virtnet_info *vi = netdev_priv(dev);
2815 	unsigned int idx = 0, start, i, j;
2816 	const u8 *stats_base;
2817 	size_t offset;
2818 
2819 	for (i = 0; i < vi->curr_queue_pairs; i++) {
2820 		struct receive_queue *rq = &vi->rq[i];
2821 
2822 		stats_base = (u8 *)&rq->stats;
2823 		do {
2824 			start = u64_stats_fetch_begin(&rq->stats.syncp);
2825 			for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2826 				offset = virtnet_rq_stats_desc[j].offset;
2827 				data[idx + j] = *(u64 *)(stats_base + offset);
2828 			}
2829 		} while (u64_stats_fetch_retry(&rq->stats.syncp, start));
2830 		idx += VIRTNET_RQ_STATS_LEN;
2831 	}
2832 
2833 	for (i = 0; i < vi->curr_queue_pairs; i++) {
2834 		struct send_queue *sq = &vi->sq[i];
2835 
2836 		stats_base = (u8 *)&sq->stats;
2837 		do {
2838 			start = u64_stats_fetch_begin(&sq->stats.syncp);
2839 			for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2840 				offset = virtnet_sq_stats_desc[j].offset;
2841 				data[idx + j] = *(u64 *)(stats_base + offset);
2842 			}
2843 		} while (u64_stats_fetch_retry(&sq->stats.syncp, start));
2844 		idx += VIRTNET_SQ_STATS_LEN;
2845 	}
2846 }
2847 
2848 static void virtnet_get_channels(struct net_device *dev,
2849 				 struct ethtool_channels *channels)
2850 {
2851 	struct virtnet_info *vi = netdev_priv(dev);
2852 
2853 	channels->combined_count = vi->curr_queue_pairs;
2854 	channels->max_combined = vi->max_queue_pairs;
2855 	channels->max_other = 0;
2856 	channels->rx_count = 0;
2857 	channels->tx_count = 0;
2858 	channels->other_count = 0;
2859 }
2860 
2861 static int virtnet_set_link_ksettings(struct net_device *dev,
2862 				      const struct ethtool_link_ksettings *cmd)
2863 {
2864 	struct virtnet_info *vi = netdev_priv(dev);
2865 
2866 	return ethtool_virtdev_set_link_ksettings(dev, cmd,
2867 						  &vi->speed, &vi->duplex);
2868 }
2869 
2870 static int virtnet_get_link_ksettings(struct net_device *dev,
2871 				      struct ethtool_link_ksettings *cmd)
2872 {
2873 	struct virtnet_info *vi = netdev_priv(dev);
2874 
2875 	cmd->base.speed = vi->speed;
2876 	cmd->base.duplex = vi->duplex;
2877 	cmd->base.port = PORT_OTHER;
2878 
2879 	return 0;
2880 }
2881 
2882 static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
2883 				       struct ethtool_coalesce *ec)
2884 {
2885 	struct scatterlist sgs_tx, sgs_rx;
2886 	struct virtio_net_ctrl_coal_tx coal_tx;
2887 	struct virtio_net_ctrl_coal_rx coal_rx;
2888 
2889 	coal_tx.tx_usecs = cpu_to_le32(ec->tx_coalesce_usecs);
2890 	coal_tx.tx_max_packets = cpu_to_le32(ec->tx_max_coalesced_frames);
2891 	sg_init_one(&sgs_tx, &coal_tx, sizeof(coal_tx));
2892 
2893 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
2894 				  VIRTIO_NET_CTRL_NOTF_COAL_TX_SET,
2895 				  &sgs_tx))
2896 		return -EINVAL;
2897 
2898 	/* Save parameters */
2899 	vi->tx_usecs = ec->tx_coalesce_usecs;
2900 	vi->tx_max_packets = ec->tx_max_coalesced_frames;
2901 
2902 	coal_rx.rx_usecs = cpu_to_le32(ec->rx_coalesce_usecs);
2903 	coal_rx.rx_max_packets = cpu_to_le32(ec->rx_max_coalesced_frames);
2904 	sg_init_one(&sgs_rx, &coal_rx, sizeof(coal_rx));
2905 
2906 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
2907 				  VIRTIO_NET_CTRL_NOTF_COAL_RX_SET,
2908 				  &sgs_rx))
2909 		return -EINVAL;
2910 
2911 	/* Save parameters */
2912 	vi->rx_usecs = ec->rx_coalesce_usecs;
2913 	vi->rx_max_packets = ec->rx_max_coalesced_frames;
2914 
2915 	return 0;
2916 }
2917 
2918 static int virtnet_coal_params_supported(struct ethtool_coalesce *ec)
2919 {
2920 	/* usecs coalescing is supported only if VIRTIO_NET_F_NOTF_COAL
2921 	 * feature is negotiated.
2922 	 */
2923 	if (ec->rx_coalesce_usecs || ec->tx_coalesce_usecs)
2924 		return -EOPNOTSUPP;
2925 
2926 	if (ec->tx_max_coalesced_frames > 1 ||
2927 	    ec->rx_max_coalesced_frames != 1)
2928 		return -EINVAL;
2929 
2930 	return 0;
2931 }
2932 
2933 static int virtnet_set_coalesce(struct net_device *dev,
2934 				struct ethtool_coalesce *ec,
2935 				struct kernel_ethtool_coalesce *kernel_coal,
2936 				struct netlink_ext_ack *extack)
2937 {
2938 	struct virtnet_info *vi = netdev_priv(dev);
2939 	int ret, i, napi_weight;
2940 	bool update_napi = false;
2941 
2942 	/* Can't change NAPI weight if the link is up */
2943 	napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2944 	if (napi_weight ^ vi->sq[0].napi.weight) {
2945 		if (dev->flags & IFF_UP)
2946 			return -EBUSY;
2947 		else
2948 			update_napi = true;
2949 	}
2950 
2951 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL))
2952 		ret = virtnet_send_notf_coal_cmds(vi, ec);
2953 	else
2954 		ret = virtnet_coal_params_supported(ec);
2955 
2956 	if (ret)
2957 		return ret;
2958 
2959 	if (update_napi) {
2960 		for (i = 0; i < vi->max_queue_pairs; i++)
2961 			vi->sq[i].napi.weight = napi_weight;
2962 	}
2963 
2964 	return ret;
2965 }
2966 
2967 static int virtnet_get_coalesce(struct net_device *dev,
2968 				struct ethtool_coalesce *ec,
2969 				struct kernel_ethtool_coalesce *kernel_coal,
2970 				struct netlink_ext_ack *extack)
2971 {
2972 	struct virtnet_info *vi = netdev_priv(dev);
2973 
2974 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
2975 		ec->rx_coalesce_usecs = vi->rx_usecs;
2976 		ec->tx_coalesce_usecs = vi->tx_usecs;
2977 		ec->tx_max_coalesced_frames = vi->tx_max_packets;
2978 		ec->rx_max_coalesced_frames = vi->rx_max_packets;
2979 	} else {
2980 		ec->rx_max_coalesced_frames = 1;
2981 
2982 		if (vi->sq[0].napi.weight)
2983 			ec->tx_max_coalesced_frames = 1;
2984 	}
2985 
2986 	return 0;
2987 }
2988 
2989 static void virtnet_init_settings(struct net_device *dev)
2990 {
2991 	struct virtnet_info *vi = netdev_priv(dev);
2992 
2993 	vi->speed = SPEED_UNKNOWN;
2994 	vi->duplex = DUPLEX_UNKNOWN;
2995 }
2996 
2997 static void virtnet_update_settings(struct virtnet_info *vi)
2998 {
2999 	u32 speed;
3000 	u8 duplex;
3001 
3002 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
3003 		return;
3004 
3005 	virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
3006 
3007 	if (ethtool_validate_speed(speed))
3008 		vi->speed = speed;
3009 
3010 	virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
3011 
3012 	if (ethtool_validate_duplex(duplex))
3013 		vi->duplex = duplex;
3014 }
3015 
3016 static u32 virtnet_get_rxfh_key_size(struct net_device *dev)
3017 {
3018 	return ((struct virtnet_info *)netdev_priv(dev))->rss_key_size;
3019 }
3020 
3021 static u32 virtnet_get_rxfh_indir_size(struct net_device *dev)
3022 {
3023 	return ((struct virtnet_info *)netdev_priv(dev))->rss_indir_table_size;
3024 }
3025 
3026 static int virtnet_get_rxfh(struct net_device *dev, u32 *indir, u8 *key, u8 *hfunc)
3027 {
3028 	struct virtnet_info *vi = netdev_priv(dev);
3029 	int i;
3030 
3031 	if (indir) {
3032 		for (i = 0; i < vi->rss_indir_table_size; ++i)
3033 			indir[i] = vi->ctrl->rss.indirection_table[i];
3034 	}
3035 
3036 	if (key)
3037 		memcpy(key, vi->ctrl->rss.key, vi->rss_key_size);
3038 
3039 	if (hfunc)
3040 		*hfunc = ETH_RSS_HASH_TOP;
3041 
3042 	return 0;
3043 }
3044 
3045 static int virtnet_set_rxfh(struct net_device *dev, const u32 *indir, const u8 *key, const u8 hfunc)
3046 {
3047 	struct virtnet_info *vi = netdev_priv(dev);
3048 	int i;
3049 
3050 	if (hfunc != ETH_RSS_HASH_NO_CHANGE && hfunc != ETH_RSS_HASH_TOP)
3051 		return -EOPNOTSUPP;
3052 
3053 	if (indir) {
3054 		for (i = 0; i < vi->rss_indir_table_size; ++i)
3055 			vi->ctrl->rss.indirection_table[i] = indir[i];
3056 	}
3057 	if (key)
3058 		memcpy(vi->ctrl->rss.key, key, vi->rss_key_size);
3059 
3060 	virtnet_commit_rss_command(vi);
3061 
3062 	return 0;
3063 }
3064 
3065 static int virtnet_get_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info, u32 *rule_locs)
3066 {
3067 	struct virtnet_info *vi = netdev_priv(dev);
3068 	int rc = 0;
3069 
3070 	switch (info->cmd) {
3071 	case ETHTOOL_GRXRINGS:
3072 		info->data = vi->curr_queue_pairs;
3073 		break;
3074 	case ETHTOOL_GRXFH:
3075 		virtnet_get_hashflow(vi, info);
3076 		break;
3077 	default:
3078 		rc = -EOPNOTSUPP;
3079 	}
3080 
3081 	return rc;
3082 }
3083 
3084 static int virtnet_set_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info)
3085 {
3086 	struct virtnet_info *vi = netdev_priv(dev);
3087 	int rc = 0;
3088 
3089 	switch (info->cmd) {
3090 	case ETHTOOL_SRXFH:
3091 		if (!virtnet_set_hashflow(vi, info))
3092 			rc = -EINVAL;
3093 
3094 		break;
3095 	default:
3096 		rc = -EOPNOTSUPP;
3097 	}
3098 
3099 	return rc;
3100 }
3101 
3102 static const struct ethtool_ops virtnet_ethtool_ops = {
3103 	.supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES |
3104 		ETHTOOL_COALESCE_USECS,
3105 	.get_drvinfo = virtnet_get_drvinfo,
3106 	.get_link = ethtool_op_get_link,
3107 	.get_ringparam = virtnet_get_ringparam,
3108 	.set_ringparam = virtnet_set_ringparam,
3109 	.get_strings = virtnet_get_strings,
3110 	.get_sset_count = virtnet_get_sset_count,
3111 	.get_ethtool_stats = virtnet_get_ethtool_stats,
3112 	.set_channels = virtnet_set_channels,
3113 	.get_channels = virtnet_get_channels,
3114 	.get_ts_info = ethtool_op_get_ts_info,
3115 	.get_link_ksettings = virtnet_get_link_ksettings,
3116 	.set_link_ksettings = virtnet_set_link_ksettings,
3117 	.set_coalesce = virtnet_set_coalesce,
3118 	.get_coalesce = virtnet_get_coalesce,
3119 	.get_rxfh_key_size = virtnet_get_rxfh_key_size,
3120 	.get_rxfh_indir_size = virtnet_get_rxfh_indir_size,
3121 	.get_rxfh = virtnet_get_rxfh,
3122 	.set_rxfh = virtnet_set_rxfh,
3123 	.get_rxnfc = virtnet_get_rxnfc,
3124 	.set_rxnfc = virtnet_set_rxnfc,
3125 };
3126 
3127 static void virtnet_freeze_down(struct virtio_device *vdev)
3128 {
3129 	struct virtnet_info *vi = vdev->priv;
3130 
3131 	/* Make sure no work handler is accessing the device */
3132 	flush_work(&vi->config_work);
3133 
3134 	netif_tx_lock_bh(vi->dev);
3135 	netif_device_detach(vi->dev);
3136 	netif_tx_unlock_bh(vi->dev);
3137 	if (netif_running(vi->dev))
3138 		virtnet_close(vi->dev);
3139 }
3140 
3141 static int init_vqs(struct virtnet_info *vi);
3142 
3143 static int virtnet_restore_up(struct virtio_device *vdev)
3144 {
3145 	struct virtnet_info *vi = vdev->priv;
3146 	int err;
3147 
3148 	err = init_vqs(vi);
3149 	if (err)
3150 		return err;
3151 
3152 	virtio_device_ready(vdev);
3153 
3154 	enable_delayed_refill(vi);
3155 
3156 	if (netif_running(vi->dev)) {
3157 		err = virtnet_open(vi->dev);
3158 		if (err)
3159 			return err;
3160 	}
3161 
3162 	netif_tx_lock_bh(vi->dev);
3163 	netif_device_attach(vi->dev);
3164 	netif_tx_unlock_bh(vi->dev);
3165 	return err;
3166 }
3167 
3168 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
3169 {
3170 	struct scatterlist sg;
3171 	vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
3172 
3173 	sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
3174 
3175 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
3176 				  VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
3177 		dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
3178 		return -EINVAL;
3179 	}
3180 
3181 	return 0;
3182 }
3183 
3184 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
3185 {
3186 	u64 offloads = 0;
3187 
3188 	if (!vi->guest_offloads)
3189 		return 0;
3190 
3191 	return virtnet_set_guest_offloads(vi, offloads);
3192 }
3193 
3194 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
3195 {
3196 	u64 offloads = vi->guest_offloads;
3197 
3198 	if (!vi->guest_offloads)
3199 		return 0;
3200 
3201 	return virtnet_set_guest_offloads(vi, offloads);
3202 }
3203 
3204 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
3205 			   struct netlink_ext_ack *extack)
3206 {
3207 	unsigned int room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM +
3208 					   sizeof(struct skb_shared_info));
3209 	unsigned int max_sz = PAGE_SIZE - room - ETH_HLEN;
3210 	struct virtnet_info *vi = netdev_priv(dev);
3211 	struct bpf_prog *old_prog;
3212 	u16 xdp_qp = 0, curr_qp;
3213 	int i, err;
3214 
3215 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
3216 	    && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3217 	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3218 	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
3219 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
3220 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM) ||
3221 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) ||
3222 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6))) {
3223 		NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
3224 		return -EOPNOTSUPP;
3225 	}
3226 
3227 	if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
3228 		NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
3229 		return -EINVAL;
3230 	}
3231 
3232 	if (prog && !prog->aux->xdp_has_frags && dev->mtu > max_sz) {
3233 		NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP without frags");
3234 		netdev_warn(dev, "single-buffer XDP requires MTU less than %u\n", max_sz);
3235 		return -EINVAL;
3236 	}
3237 
3238 	curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
3239 	if (prog)
3240 		xdp_qp = nr_cpu_ids;
3241 
3242 	/* XDP requires extra queues for XDP_TX */
3243 	if (curr_qp + xdp_qp > vi->max_queue_pairs) {
3244 		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",
3245 				 curr_qp + xdp_qp, vi->max_queue_pairs);
3246 		xdp_qp = 0;
3247 	}
3248 
3249 	old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
3250 	if (!prog && !old_prog)
3251 		return 0;
3252 
3253 	if (prog)
3254 		bpf_prog_add(prog, vi->max_queue_pairs - 1);
3255 
3256 	/* Make sure NAPI is not using any XDP TX queues for RX. */
3257 	if (netif_running(dev)) {
3258 		for (i = 0; i < vi->max_queue_pairs; i++) {
3259 			napi_disable(&vi->rq[i].napi);
3260 			virtnet_napi_tx_disable(&vi->sq[i].napi);
3261 		}
3262 	}
3263 
3264 	if (!prog) {
3265 		for (i = 0; i < vi->max_queue_pairs; i++) {
3266 			rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
3267 			if (i == 0)
3268 				virtnet_restore_guest_offloads(vi);
3269 		}
3270 		synchronize_net();
3271 	}
3272 
3273 	err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
3274 	if (err)
3275 		goto err;
3276 	netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
3277 	vi->xdp_queue_pairs = xdp_qp;
3278 
3279 	if (prog) {
3280 		vi->xdp_enabled = true;
3281 		for (i = 0; i < vi->max_queue_pairs; i++) {
3282 			rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
3283 			if (i == 0 && !old_prog)
3284 				virtnet_clear_guest_offloads(vi);
3285 		}
3286 		if (!old_prog)
3287 			xdp_features_set_redirect_target(dev, true);
3288 	} else {
3289 		xdp_features_clear_redirect_target(dev);
3290 		vi->xdp_enabled = false;
3291 	}
3292 
3293 	for (i = 0; i < vi->max_queue_pairs; i++) {
3294 		if (old_prog)
3295 			bpf_prog_put(old_prog);
3296 		if (netif_running(dev)) {
3297 			virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
3298 			virtnet_napi_tx_enable(vi, vi->sq[i].vq,
3299 					       &vi->sq[i].napi);
3300 		}
3301 	}
3302 
3303 	return 0;
3304 
3305 err:
3306 	if (!prog) {
3307 		virtnet_clear_guest_offloads(vi);
3308 		for (i = 0; i < vi->max_queue_pairs; i++)
3309 			rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
3310 	}
3311 
3312 	if (netif_running(dev)) {
3313 		for (i = 0; i < vi->max_queue_pairs; i++) {
3314 			virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
3315 			virtnet_napi_tx_enable(vi, vi->sq[i].vq,
3316 					       &vi->sq[i].napi);
3317 		}
3318 	}
3319 	if (prog)
3320 		bpf_prog_sub(prog, vi->max_queue_pairs - 1);
3321 	return err;
3322 }
3323 
3324 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
3325 {
3326 	switch (xdp->command) {
3327 	case XDP_SETUP_PROG:
3328 		return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
3329 	default:
3330 		return -EINVAL;
3331 	}
3332 }
3333 
3334 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
3335 				      size_t len)
3336 {
3337 	struct virtnet_info *vi = netdev_priv(dev);
3338 	int ret;
3339 
3340 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
3341 		return -EOPNOTSUPP;
3342 
3343 	ret = snprintf(buf, len, "sby");
3344 	if (ret >= len)
3345 		return -EOPNOTSUPP;
3346 
3347 	return 0;
3348 }
3349 
3350 static int virtnet_set_features(struct net_device *dev,
3351 				netdev_features_t features)
3352 {
3353 	struct virtnet_info *vi = netdev_priv(dev);
3354 	u64 offloads;
3355 	int err;
3356 
3357 	if ((dev->features ^ features) & NETIF_F_GRO_HW) {
3358 		if (vi->xdp_enabled)
3359 			return -EBUSY;
3360 
3361 		if (features & NETIF_F_GRO_HW)
3362 			offloads = vi->guest_offloads_capable;
3363 		else
3364 			offloads = vi->guest_offloads_capable &
3365 				   ~GUEST_OFFLOAD_GRO_HW_MASK;
3366 
3367 		err = virtnet_set_guest_offloads(vi, offloads);
3368 		if (err)
3369 			return err;
3370 		vi->guest_offloads = offloads;
3371 	}
3372 
3373 	if ((dev->features ^ features) & NETIF_F_RXHASH) {
3374 		if (features & NETIF_F_RXHASH)
3375 			vi->ctrl->rss.hash_types = vi->rss_hash_types_saved;
3376 		else
3377 			vi->ctrl->rss.hash_types = VIRTIO_NET_HASH_REPORT_NONE;
3378 
3379 		if (!virtnet_commit_rss_command(vi))
3380 			return -EINVAL;
3381 	}
3382 
3383 	return 0;
3384 }
3385 
3386 static void virtnet_tx_timeout(struct net_device *dev, unsigned int txqueue)
3387 {
3388 	struct virtnet_info *priv = netdev_priv(dev);
3389 	struct send_queue *sq = &priv->sq[txqueue];
3390 	struct netdev_queue *txq = netdev_get_tx_queue(dev, txqueue);
3391 
3392 	u64_stats_update_begin(&sq->stats.syncp);
3393 	sq->stats.tx_timeouts++;
3394 	u64_stats_update_end(&sq->stats.syncp);
3395 
3396 	netdev_err(dev, "TX timeout on queue: %u, sq: %s, vq: 0x%x, name: %s, %u usecs ago\n",
3397 		   txqueue, sq->name, sq->vq->index, sq->vq->name,
3398 		   jiffies_to_usecs(jiffies - READ_ONCE(txq->trans_start)));
3399 }
3400 
3401 static const struct net_device_ops virtnet_netdev = {
3402 	.ndo_open            = virtnet_open,
3403 	.ndo_stop   	     = virtnet_close,
3404 	.ndo_start_xmit      = start_xmit,
3405 	.ndo_validate_addr   = eth_validate_addr,
3406 	.ndo_set_mac_address = virtnet_set_mac_address,
3407 	.ndo_set_rx_mode     = virtnet_set_rx_mode,
3408 	.ndo_get_stats64     = virtnet_stats,
3409 	.ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
3410 	.ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
3411 	.ndo_bpf		= virtnet_xdp,
3412 	.ndo_xdp_xmit		= virtnet_xdp_xmit,
3413 	.ndo_features_check	= passthru_features_check,
3414 	.ndo_get_phys_port_name	= virtnet_get_phys_port_name,
3415 	.ndo_set_features	= virtnet_set_features,
3416 	.ndo_tx_timeout		= virtnet_tx_timeout,
3417 };
3418 
3419 static void virtnet_config_changed_work(struct work_struct *work)
3420 {
3421 	struct virtnet_info *vi =
3422 		container_of(work, struct virtnet_info, config_work);
3423 	u16 v;
3424 
3425 	if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
3426 				 struct virtio_net_config, status, &v) < 0)
3427 		return;
3428 
3429 	if (v & VIRTIO_NET_S_ANNOUNCE) {
3430 		netdev_notify_peers(vi->dev);
3431 		virtnet_ack_link_announce(vi);
3432 	}
3433 
3434 	/* Ignore unknown (future) status bits */
3435 	v &= VIRTIO_NET_S_LINK_UP;
3436 
3437 	if (vi->status == v)
3438 		return;
3439 
3440 	vi->status = v;
3441 
3442 	if (vi->status & VIRTIO_NET_S_LINK_UP) {
3443 		virtnet_update_settings(vi);
3444 		netif_carrier_on(vi->dev);
3445 		netif_tx_wake_all_queues(vi->dev);
3446 	} else {
3447 		netif_carrier_off(vi->dev);
3448 		netif_tx_stop_all_queues(vi->dev);
3449 	}
3450 }
3451 
3452 static void virtnet_config_changed(struct virtio_device *vdev)
3453 {
3454 	struct virtnet_info *vi = vdev->priv;
3455 
3456 	schedule_work(&vi->config_work);
3457 }
3458 
3459 static void virtnet_free_queues(struct virtnet_info *vi)
3460 {
3461 	int i;
3462 
3463 	for (i = 0; i < vi->max_queue_pairs; i++) {
3464 		__netif_napi_del(&vi->rq[i].napi);
3465 		__netif_napi_del(&vi->sq[i].napi);
3466 	}
3467 
3468 	/* We called __netif_napi_del(),
3469 	 * we need to respect an RCU grace period before freeing vi->rq
3470 	 */
3471 	synchronize_net();
3472 
3473 	kfree(vi->rq);
3474 	kfree(vi->sq);
3475 	kfree(vi->ctrl);
3476 }
3477 
3478 static void _free_receive_bufs(struct virtnet_info *vi)
3479 {
3480 	struct bpf_prog *old_prog;
3481 	int i;
3482 
3483 	for (i = 0; i < vi->max_queue_pairs; i++) {
3484 		while (vi->rq[i].pages)
3485 			__free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
3486 
3487 		old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
3488 		RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
3489 		if (old_prog)
3490 			bpf_prog_put(old_prog);
3491 	}
3492 }
3493 
3494 static void free_receive_bufs(struct virtnet_info *vi)
3495 {
3496 	rtnl_lock();
3497 	_free_receive_bufs(vi);
3498 	rtnl_unlock();
3499 }
3500 
3501 static void free_receive_page_frags(struct virtnet_info *vi)
3502 {
3503 	int i;
3504 	for (i = 0; i < vi->max_queue_pairs; i++)
3505 		if (vi->rq[i].alloc_frag.page)
3506 			put_page(vi->rq[i].alloc_frag.page);
3507 }
3508 
3509 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf)
3510 {
3511 	if (!is_xdp_frame(buf))
3512 		dev_kfree_skb(buf);
3513 	else
3514 		xdp_return_frame(ptr_to_xdp(buf));
3515 }
3516 
3517 static void virtnet_rq_free_unused_buf(struct virtqueue *vq, void *buf)
3518 {
3519 	struct virtnet_info *vi = vq->vdev->priv;
3520 	int i = vq2rxq(vq);
3521 
3522 	if (vi->mergeable_rx_bufs)
3523 		put_page(virt_to_head_page(buf));
3524 	else if (vi->big_packets)
3525 		give_pages(&vi->rq[i], buf);
3526 	else
3527 		put_page(virt_to_head_page(buf));
3528 }
3529 
3530 static void free_unused_bufs(struct virtnet_info *vi)
3531 {
3532 	void *buf;
3533 	int i;
3534 
3535 	for (i = 0; i < vi->max_queue_pairs; i++) {
3536 		struct virtqueue *vq = vi->sq[i].vq;
3537 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
3538 			virtnet_sq_free_unused_buf(vq, buf);
3539 	}
3540 
3541 	for (i = 0; i < vi->max_queue_pairs; i++) {
3542 		struct virtqueue *vq = vi->rq[i].vq;
3543 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
3544 			virtnet_rq_free_unused_buf(vq, buf);
3545 	}
3546 }
3547 
3548 static void virtnet_del_vqs(struct virtnet_info *vi)
3549 {
3550 	struct virtio_device *vdev = vi->vdev;
3551 
3552 	virtnet_clean_affinity(vi);
3553 
3554 	vdev->config->del_vqs(vdev);
3555 
3556 	virtnet_free_queues(vi);
3557 }
3558 
3559 /* How large should a single buffer be so a queue full of these can fit at
3560  * least one full packet?
3561  * Logic below assumes the mergeable buffer header is used.
3562  */
3563 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
3564 {
3565 	const unsigned int hdr_len = vi->hdr_len;
3566 	unsigned int rq_size = virtqueue_get_vring_size(vq);
3567 	unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
3568 	unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
3569 	unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
3570 
3571 	return max(max(min_buf_len, hdr_len) - hdr_len,
3572 		   (unsigned int)GOOD_PACKET_LEN);
3573 }
3574 
3575 static int virtnet_find_vqs(struct virtnet_info *vi)
3576 {
3577 	vq_callback_t **callbacks;
3578 	struct virtqueue **vqs;
3579 	int ret = -ENOMEM;
3580 	int i, total_vqs;
3581 	const char **names;
3582 	bool *ctx;
3583 
3584 	/* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
3585 	 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
3586 	 * possible control vq.
3587 	 */
3588 	total_vqs = vi->max_queue_pairs * 2 +
3589 		    virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
3590 
3591 	/* Allocate space for find_vqs parameters */
3592 	vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
3593 	if (!vqs)
3594 		goto err_vq;
3595 	callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
3596 	if (!callbacks)
3597 		goto err_callback;
3598 	names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
3599 	if (!names)
3600 		goto err_names;
3601 	if (!vi->big_packets || vi->mergeable_rx_bufs) {
3602 		ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
3603 		if (!ctx)
3604 			goto err_ctx;
3605 	} else {
3606 		ctx = NULL;
3607 	}
3608 
3609 	/* Parameters for control virtqueue, if any */
3610 	if (vi->has_cvq) {
3611 		callbacks[total_vqs - 1] = NULL;
3612 		names[total_vqs - 1] = "control";
3613 	}
3614 
3615 	/* Allocate/initialize parameters for send/receive virtqueues */
3616 	for (i = 0; i < vi->max_queue_pairs; i++) {
3617 		callbacks[rxq2vq(i)] = skb_recv_done;
3618 		callbacks[txq2vq(i)] = skb_xmit_done;
3619 		sprintf(vi->rq[i].name, "input.%d", i);
3620 		sprintf(vi->sq[i].name, "output.%d", i);
3621 		names[rxq2vq(i)] = vi->rq[i].name;
3622 		names[txq2vq(i)] = vi->sq[i].name;
3623 		if (ctx)
3624 			ctx[rxq2vq(i)] = true;
3625 	}
3626 
3627 	ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
3628 				  names, ctx, NULL);
3629 	if (ret)
3630 		goto err_find;
3631 
3632 	if (vi->has_cvq) {
3633 		vi->cvq = vqs[total_vqs - 1];
3634 		if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
3635 			vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
3636 	}
3637 
3638 	for (i = 0; i < vi->max_queue_pairs; i++) {
3639 		vi->rq[i].vq = vqs[rxq2vq(i)];
3640 		vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
3641 		vi->sq[i].vq = vqs[txq2vq(i)];
3642 	}
3643 
3644 	/* run here: ret == 0. */
3645 
3646 
3647 err_find:
3648 	kfree(ctx);
3649 err_ctx:
3650 	kfree(names);
3651 err_names:
3652 	kfree(callbacks);
3653 err_callback:
3654 	kfree(vqs);
3655 err_vq:
3656 	return ret;
3657 }
3658 
3659 static int virtnet_alloc_queues(struct virtnet_info *vi)
3660 {
3661 	int i;
3662 
3663 	if (vi->has_cvq) {
3664 		vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
3665 		if (!vi->ctrl)
3666 			goto err_ctrl;
3667 	} else {
3668 		vi->ctrl = NULL;
3669 	}
3670 	vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
3671 	if (!vi->sq)
3672 		goto err_sq;
3673 	vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
3674 	if (!vi->rq)
3675 		goto err_rq;
3676 
3677 	INIT_DELAYED_WORK(&vi->refill, refill_work);
3678 	for (i = 0; i < vi->max_queue_pairs; i++) {
3679 		vi->rq[i].pages = NULL;
3680 		netif_napi_add_weight(vi->dev, &vi->rq[i].napi, virtnet_poll,
3681 				      napi_weight);
3682 		netif_napi_add_tx_weight(vi->dev, &vi->sq[i].napi,
3683 					 virtnet_poll_tx,
3684 					 napi_tx ? napi_weight : 0);
3685 
3686 		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
3687 		ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
3688 		sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
3689 
3690 		u64_stats_init(&vi->rq[i].stats.syncp);
3691 		u64_stats_init(&vi->sq[i].stats.syncp);
3692 	}
3693 
3694 	return 0;
3695 
3696 err_rq:
3697 	kfree(vi->sq);
3698 err_sq:
3699 	kfree(vi->ctrl);
3700 err_ctrl:
3701 	return -ENOMEM;
3702 }
3703 
3704 static int init_vqs(struct virtnet_info *vi)
3705 {
3706 	int ret;
3707 
3708 	/* Allocate send & receive queues */
3709 	ret = virtnet_alloc_queues(vi);
3710 	if (ret)
3711 		goto err;
3712 
3713 	ret = virtnet_find_vqs(vi);
3714 	if (ret)
3715 		goto err_free;
3716 
3717 	cpus_read_lock();
3718 	virtnet_set_affinity(vi);
3719 	cpus_read_unlock();
3720 
3721 	return 0;
3722 
3723 err_free:
3724 	virtnet_free_queues(vi);
3725 err:
3726 	return ret;
3727 }
3728 
3729 #ifdef CONFIG_SYSFS
3730 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
3731 		char *buf)
3732 {
3733 	struct virtnet_info *vi = netdev_priv(queue->dev);
3734 	unsigned int queue_index = get_netdev_rx_queue_index(queue);
3735 	unsigned int headroom = virtnet_get_headroom(vi);
3736 	unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
3737 	struct ewma_pkt_len *avg;
3738 
3739 	BUG_ON(queue_index >= vi->max_queue_pairs);
3740 	avg = &vi->rq[queue_index].mrg_avg_pkt_len;
3741 	return sprintf(buf, "%u\n",
3742 		       get_mergeable_buf_len(&vi->rq[queue_index], avg,
3743 				       SKB_DATA_ALIGN(headroom + tailroom)));
3744 }
3745 
3746 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
3747 	__ATTR_RO(mergeable_rx_buffer_size);
3748 
3749 static struct attribute *virtio_net_mrg_rx_attrs[] = {
3750 	&mergeable_rx_buffer_size_attribute.attr,
3751 	NULL
3752 };
3753 
3754 static const struct attribute_group virtio_net_mrg_rx_group = {
3755 	.name = "virtio_net",
3756 	.attrs = virtio_net_mrg_rx_attrs
3757 };
3758 #endif
3759 
3760 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
3761 				    unsigned int fbit,
3762 				    const char *fname, const char *dname)
3763 {
3764 	if (!virtio_has_feature(vdev, fbit))
3765 		return false;
3766 
3767 	dev_err(&vdev->dev, "device advertises feature %s but not %s",
3768 		fname, dname);
3769 
3770 	return true;
3771 }
3772 
3773 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)			\
3774 	virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
3775 
3776 static bool virtnet_validate_features(struct virtio_device *vdev)
3777 {
3778 	if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
3779 	    (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
3780 			     "VIRTIO_NET_F_CTRL_VQ") ||
3781 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
3782 			     "VIRTIO_NET_F_CTRL_VQ") ||
3783 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
3784 			     "VIRTIO_NET_F_CTRL_VQ") ||
3785 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
3786 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
3787 			     "VIRTIO_NET_F_CTRL_VQ") ||
3788 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_RSS,
3789 			     "VIRTIO_NET_F_CTRL_VQ") ||
3790 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_HASH_REPORT,
3791 			     "VIRTIO_NET_F_CTRL_VQ") ||
3792 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_NOTF_COAL,
3793 			     "VIRTIO_NET_F_CTRL_VQ"))) {
3794 		return false;
3795 	}
3796 
3797 	return true;
3798 }
3799 
3800 #define MIN_MTU ETH_MIN_MTU
3801 #define MAX_MTU ETH_MAX_MTU
3802 
3803 static int virtnet_validate(struct virtio_device *vdev)
3804 {
3805 	if (!vdev->config->get) {
3806 		dev_err(&vdev->dev, "%s failure: config access disabled\n",
3807 			__func__);
3808 		return -EINVAL;
3809 	}
3810 
3811 	if (!virtnet_validate_features(vdev))
3812 		return -EINVAL;
3813 
3814 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3815 		int mtu = virtio_cread16(vdev,
3816 					 offsetof(struct virtio_net_config,
3817 						  mtu));
3818 		if (mtu < MIN_MTU)
3819 			__virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
3820 	}
3821 
3822 	if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY) &&
3823 	    !virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
3824 		dev_warn(&vdev->dev, "device advertises feature VIRTIO_NET_F_STANDBY but not VIRTIO_NET_F_MAC, disabling standby");
3825 		__virtio_clear_bit(vdev, VIRTIO_NET_F_STANDBY);
3826 	}
3827 
3828 	return 0;
3829 }
3830 
3831 static bool virtnet_check_guest_gso(const struct virtnet_info *vi)
3832 {
3833 	return virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3834 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3835 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
3836 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
3837 		(virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) &&
3838 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6));
3839 }
3840 
3841 static void virtnet_set_big_packets(struct virtnet_info *vi, const int mtu)
3842 {
3843 	bool guest_gso = virtnet_check_guest_gso(vi);
3844 
3845 	/* If device can receive ANY guest GSO packets, regardless of mtu,
3846 	 * allocate packets of maximum size, otherwise limit it to only
3847 	 * mtu size worth only.
3848 	 */
3849 	if (mtu > ETH_DATA_LEN || guest_gso) {
3850 		vi->big_packets = true;
3851 		vi->big_packets_num_skbfrags = guest_gso ? MAX_SKB_FRAGS : DIV_ROUND_UP(mtu, PAGE_SIZE);
3852 	}
3853 }
3854 
3855 static int virtnet_probe(struct virtio_device *vdev)
3856 {
3857 	int i, err = -ENOMEM;
3858 	struct net_device *dev;
3859 	struct virtnet_info *vi;
3860 	u16 max_queue_pairs;
3861 	int mtu = 0;
3862 
3863 	/* Find if host supports multiqueue/rss virtio_net device */
3864 	max_queue_pairs = 1;
3865 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MQ) || virtio_has_feature(vdev, VIRTIO_NET_F_RSS))
3866 		max_queue_pairs =
3867 		     virtio_cread16(vdev, offsetof(struct virtio_net_config, max_virtqueue_pairs));
3868 
3869 	/* We need at least 2 queue's */
3870 	if (max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3871 	    max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3872 	    !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3873 		max_queue_pairs = 1;
3874 
3875 	/* Allocate ourselves a network device with room for our info */
3876 	dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3877 	if (!dev)
3878 		return -ENOMEM;
3879 
3880 	/* Set up network device as normal. */
3881 	dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
3882 			   IFF_TX_SKB_NO_LINEAR;
3883 	dev->netdev_ops = &virtnet_netdev;
3884 	dev->features = NETIF_F_HIGHDMA;
3885 
3886 	dev->ethtool_ops = &virtnet_ethtool_ops;
3887 	SET_NETDEV_DEV(dev, &vdev->dev);
3888 
3889 	/* Do we support "hardware" checksums? */
3890 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3891 		/* This opens up the world of extra features. */
3892 		dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3893 		if (csum)
3894 			dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3895 
3896 		if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3897 			dev->hw_features |= NETIF_F_TSO
3898 				| NETIF_F_TSO_ECN | NETIF_F_TSO6;
3899 		}
3900 		/* Individual feature bits: what can host handle? */
3901 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3902 			dev->hw_features |= NETIF_F_TSO;
3903 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3904 			dev->hw_features |= NETIF_F_TSO6;
3905 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3906 			dev->hw_features |= NETIF_F_TSO_ECN;
3907 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_USO))
3908 			dev->hw_features |= NETIF_F_GSO_UDP_L4;
3909 
3910 		dev->features |= NETIF_F_GSO_ROBUST;
3911 
3912 		if (gso)
3913 			dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3914 		/* (!csum && gso) case will be fixed by register_netdev() */
3915 	}
3916 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3917 		dev->features |= NETIF_F_RXCSUM;
3918 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3919 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3920 		dev->features |= NETIF_F_GRO_HW;
3921 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3922 		dev->hw_features |= NETIF_F_GRO_HW;
3923 
3924 	dev->vlan_features = dev->features;
3925 	dev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT;
3926 
3927 	/* MTU range: 68 - 65535 */
3928 	dev->min_mtu = MIN_MTU;
3929 	dev->max_mtu = MAX_MTU;
3930 
3931 	/* Configuration may specify what MAC to use.  Otherwise random. */
3932 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
3933 		u8 addr[ETH_ALEN];
3934 
3935 		virtio_cread_bytes(vdev,
3936 				   offsetof(struct virtio_net_config, mac),
3937 				   addr, ETH_ALEN);
3938 		eth_hw_addr_set(dev, addr);
3939 	} else {
3940 		eth_hw_addr_random(dev);
3941 		dev_info(&vdev->dev, "Assigned random MAC address %pM\n",
3942 			 dev->dev_addr);
3943 	}
3944 
3945 	/* Set up our device-specific information */
3946 	vi = netdev_priv(dev);
3947 	vi->dev = dev;
3948 	vi->vdev = vdev;
3949 	vdev->priv = vi;
3950 
3951 	INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3952 	spin_lock_init(&vi->refill_lock);
3953 
3954 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) {
3955 		vi->mergeable_rx_bufs = true;
3956 		dev->xdp_features |= NETDEV_XDP_ACT_RX_SG;
3957 	}
3958 
3959 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
3960 		vi->rx_usecs = 0;
3961 		vi->tx_usecs = 0;
3962 		vi->tx_max_packets = 0;
3963 		vi->rx_max_packets = 0;
3964 	}
3965 
3966 	if (virtio_has_feature(vdev, VIRTIO_NET_F_HASH_REPORT))
3967 		vi->has_rss_hash_report = true;
3968 
3969 	if (virtio_has_feature(vdev, VIRTIO_NET_F_RSS))
3970 		vi->has_rss = true;
3971 
3972 	if (vi->has_rss || vi->has_rss_hash_report) {
3973 		vi->rss_indir_table_size =
3974 			virtio_cread16(vdev, offsetof(struct virtio_net_config,
3975 				rss_max_indirection_table_length));
3976 		vi->rss_key_size =
3977 			virtio_cread8(vdev, offsetof(struct virtio_net_config, rss_max_key_size));
3978 
3979 		vi->rss_hash_types_supported =
3980 		    virtio_cread32(vdev, offsetof(struct virtio_net_config, supported_hash_types));
3981 		vi->rss_hash_types_supported &=
3982 				~(VIRTIO_NET_RSS_HASH_TYPE_IP_EX |
3983 				  VIRTIO_NET_RSS_HASH_TYPE_TCP_EX |
3984 				  VIRTIO_NET_RSS_HASH_TYPE_UDP_EX);
3985 
3986 		dev->hw_features |= NETIF_F_RXHASH;
3987 	}
3988 
3989 	if (vi->has_rss_hash_report)
3990 		vi->hdr_len = sizeof(struct virtio_net_hdr_v1_hash);
3991 	else if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3992 		 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3993 		vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3994 	else
3995 		vi->hdr_len = sizeof(struct virtio_net_hdr);
3996 
3997 	if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3998 	    virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3999 		vi->any_header_sg = true;
4000 
4001 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
4002 		vi->has_cvq = true;
4003 
4004 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
4005 		mtu = virtio_cread16(vdev,
4006 				     offsetof(struct virtio_net_config,
4007 					      mtu));
4008 		if (mtu < dev->min_mtu) {
4009 			/* Should never trigger: MTU was previously validated
4010 			 * in virtnet_validate.
4011 			 */
4012 			dev_err(&vdev->dev,
4013 				"device MTU appears to have changed it is now %d < %d",
4014 				mtu, dev->min_mtu);
4015 			err = -EINVAL;
4016 			goto free;
4017 		}
4018 
4019 		dev->mtu = mtu;
4020 		dev->max_mtu = mtu;
4021 	}
4022 
4023 	virtnet_set_big_packets(vi, mtu);
4024 
4025 	if (vi->any_header_sg)
4026 		dev->needed_headroom = vi->hdr_len;
4027 
4028 	/* Enable multiqueue by default */
4029 	if (num_online_cpus() >= max_queue_pairs)
4030 		vi->curr_queue_pairs = max_queue_pairs;
4031 	else
4032 		vi->curr_queue_pairs = num_online_cpus();
4033 	vi->max_queue_pairs = max_queue_pairs;
4034 
4035 	/* Allocate/initialize the rx/tx queues, and invoke find_vqs */
4036 	err = init_vqs(vi);
4037 	if (err)
4038 		goto free;
4039 
4040 #ifdef CONFIG_SYSFS
4041 	if (vi->mergeable_rx_bufs)
4042 		dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
4043 #endif
4044 	netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
4045 	netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
4046 
4047 	virtnet_init_settings(dev);
4048 
4049 	if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
4050 		vi->failover = net_failover_create(vi->dev);
4051 		if (IS_ERR(vi->failover)) {
4052 			err = PTR_ERR(vi->failover);
4053 			goto free_vqs;
4054 		}
4055 	}
4056 
4057 	if (vi->has_rss || vi->has_rss_hash_report)
4058 		virtnet_init_default_rss(vi);
4059 
4060 	/* serialize netdev register + virtio_device_ready() with ndo_open() */
4061 	rtnl_lock();
4062 
4063 	err = register_netdevice(dev);
4064 	if (err) {
4065 		pr_debug("virtio_net: registering device failed\n");
4066 		rtnl_unlock();
4067 		goto free_failover;
4068 	}
4069 
4070 	virtio_device_ready(vdev);
4071 
4072 	/* a random MAC address has been assigned, notify the device.
4073 	 * We don't fail probe if VIRTIO_NET_F_CTRL_MAC_ADDR is not there
4074 	 * because many devices work fine without getting MAC explicitly
4075 	 */
4076 	if (!virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
4077 	    virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
4078 		struct scatterlist sg;
4079 
4080 		sg_init_one(&sg, dev->dev_addr, dev->addr_len);
4081 		if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
4082 					  VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
4083 			pr_debug("virtio_net: setting MAC address failed\n");
4084 			rtnl_unlock();
4085 			err = -EINVAL;
4086 			goto free_unregister_netdev;
4087 		}
4088 	}
4089 
4090 	rtnl_unlock();
4091 
4092 	err = virtnet_cpu_notif_add(vi);
4093 	if (err) {
4094 		pr_debug("virtio_net: registering cpu notifier failed\n");
4095 		goto free_unregister_netdev;
4096 	}
4097 
4098 	virtnet_set_queues(vi, vi->curr_queue_pairs);
4099 
4100 	/* Assume link up if device can't report link status,
4101 	   otherwise get link status from config. */
4102 	netif_carrier_off(dev);
4103 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
4104 		schedule_work(&vi->config_work);
4105 	} else {
4106 		vi->status = VIRTIO_NET_S_LINK_UP;
4107 		virtnet_update_settings(vi);
4108 		netif_carrier_on(dev);
4109 	}
4110 
4111 	for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
4112 		if (virtio_has_feature(vi->vdev, guest_offloads[i]))
4113 			set_bit(guest_offloads[i], &vi->guest_offloads);
4114 	vi->guest_offloads_capable = vi->guest_offloads;
4115 
4116 	pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
4117 		 dev->name, max_queue_pairs);
4118 
4119 	return 0;
4120 
4121 free_unregister_netdev:
4122 	unregister_netdev(dev);
4123 free_failover:
4124 	net_failover_destroy(vi->failover);
4125 free_vqs:
4126 	virtio_reset_device(vdev);
4127 	cancel_delayed_work_sync(&vi->refill);
4128 	free_receive_page_frags(vi);
4129 	virtnet_del_vqs(vi);
4130 free:
4131 	free_netdev(dev);
4132 	return err;
4133 }
4134 
4135 static void remove_vq_common(struct virtnet_info *vi)
4136 {
4137 	virtio_reset_device(vi->vdev);
4138 
4139 	/* Free unused buffers in both send and recv, if any. */
4140 	free_unused_bufs(vi);
4141 
4142 	free_receive_bufs(vi);
4143 
4144 	free_receive_page_frags(vi);
4145 
4146 	virtnet_del_vqs(vi);
4147 }
4148 
4149 static void virtnet_remove(struct virtio_device *vdev)
4150 {
4151 	struct virtnet_info *vi = vdev->priv;
4152 
4153 	virtnet_cpu_notif_remove(vi);
4154 
4155 	/* Make sure no work handler is accessing the device. */
4156 	flush_work(&vi->config_work);
4157 
4158 	unregister_netdev(vi->dev);
4159 
4160 	net_failover_destroy(vi->failover);
4161 
4162 	remove_vq_common(vi);
4163 
4164 	free_netdev(vi->dev);
4165 }
4166 
4167 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
4168 {
4169 	struct virtnet_info *vi = vdev->priv;
4170 
4171 	virtnet_cpu_notif_remove(vi);
4172 	virtnet_freeze_down(vdev);
4173 	remove_vq_common(vi);
4174 
4175 	return 0;
4176 }
4177 
4178 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
4179 {
4180 	struct virtnet_info *vi = vdev->priv;
4181 	int err;
4182 
4183 	err = virtnet_restore_up(vdev);
4184 	if (err)
4185 		return err;
4186 	virtnet_set_queues(vi, vi->curr_queue_pairs);
4187 
4188 	err = virtnet_cpu_notif_add(vi);
4189 	if (err) {
4190 		virtnet_freeze_down(vdev);
4191 		remove_vq_common(vi);
4192 		return err;
4193 	}
4194 
4195 	return 0;
4196 }
4197 
4198 static struct virtio_device_id id_table[] = {
4199 	{ VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
4200 	{ 0 },
4201 };
4202 
4203 #define VIRTNET_FEATURES \
4204 	VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
4205 	VIRTIO_NET_F_MAC, \
4206 	VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
4207 	VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
4208 	VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
4209 	VIRTIO_NET_F_HOST_USO, VIRTIO_NET_F_GUEST_USO4, VIRTIO_NET_F_GUEST_USO6, \
4210 	VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
4211 	VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
4212 	VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
4213 	VIRTIO_NET_F_CTRL_MAC_ADDR, \
4214 	VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
4215 	VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY, \
4216 	VIRTIO_NET_F_RSS, VIRTIO_NET_F_HASH_REPORT, VIRTIO_NET_F_NOTF_COAL
4217 
4218 static unsigned int features[] = {
4219 	VIRTNET_FEATURES,
4220 };
4221 
4222 static unsigned int features_legacy[] = {
4223 	VIRTNET_FEATURES,
4224 	VIRTIO_NET_F_GSO,
4225 	VIRTIO_F_ANY_LAYOUT,
4226 };
4227 
4228 static struct virtio_driver virtio_net_driver = {
4229 	.feature_table = features,
4230 	.feature_table_size = ARRAY_SIZE(features),
4231 	.feature_table_legacy = features_legacy,
4232 	.feature_table_size_legacy = ARRAY_SIZE(features_legacy),
4233 	.driver.name =	KBUILD_MODNAME,
4234 	.driver.owner =	THIS_MODULE,
4235 	.id_table =	id_table,
4236 	.validate =	virtnet_validate,
4237 	.probe =	virtnet_probe,
4238 	.remove =	virtnet_remove,
4239 	.config_changed = virtnet_config_changed,
4240 #ifdef CONFIG_PM_SLEEP
4241 	.freeze =	virtnet_freeze,
4242 	.restore =	virtnet_restore,
4243 #endif
4244 };
4245 
4246 static __init int virtio_net_driver_init(void)
4247 {
4248 	int ret;
4249 
4250 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
4251 				      virtnet_cpu_online,
4252 				      virtnet_cpu_down_prep);
4253 	if (ret < 0)
4254 		goto out;
4255 	virtionet_online = ret;
4256 	ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
4257 				      NULL, virtnet_cpu_dead);
4258 	if (ret)
4259 		goto err_dead;
4260 	ret = register_virtio_driver(&virtio_net_driver);
4261 	if (ret)
4262 		goto err_virtio;
4263 	return 0;
4264 err_virtio:
4265 	cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
4266 err_dead:
4267 	cpuhp_remove_multi_state(virtionet_online);
4268 out:
4269 	return ret;
4270 }
4271 module_init(virtio_net_driver_init);
4272 
4273 static __exit void virtio_net_driver_exit(void)
4274 {
4275 	unregister_virtio_driver(&virtio_net_driver);
4276 	cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
4277 	cpuhp_remove_multi_state(virtionet_online);
4278 }
4279 module_exit(virtio_net_driver_exit);
4280 
4281 MODULE_DEVICE_TABLE(virtio, id_table);
4282 MODULE_DESCRIPTION("Virtio network driver");
4283 MODULE_LICENSE("GPL");
4284