xref: /openbmc/linux/drivers/net/virtio_net.c (revision a36954f5)
1 /* A network driver using virtio.
2  *
3  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License as published by
7  * the Free Software Foundation; either version 2 of the License, or
8  * (at your option) any later version.
9  *
10  * This program is distributed in the hope that it will be useful,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13  * GNU General Public License for more details.
14  *
15  * You should have received a copy of the GNU General Public License
16  * along with this program; if not, see <http://www.gnu.org/licenses/>.
17  */
18 //#define DEBUG
19 #include <linux/netdevice.h>
20 #include <linux/etherdevice.h>
21 #include <linux/ethtool.h>
22 #include <linux/module.h>
23 #include <linux/virtio.h>
24 #include <linux/virtio_net.h>
25 #include <linux/bpf.h>
26 #include <linux/bpf_trace.h>
27 #include <linux/scatterlist.h>
28 #include <linux/if_vlan.h>
29 #include <linux/slab.h>
30 #include <linux/cpu.h>
31 #include <linux/average.h>
32 #include <net/route.h>
33 
34 static int napi_weight = NAPI_POLL_WEIGHT;
35 module_param(napi_weight, int, 0444);
36 
37 static bool csum = true, gso = true, napi_tx;
38 module_param(csum, bool, 0444);
39 module_param(gso, bool, 0444);
40 module_param(napi_tx, bool, 0644);
41 
42 /* FIXME: MTU in config. */
43 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
44 #define GOOD_COPY_LEN	128
45 
46 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
47 
48 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
49 #define VIRTIO_XDP_HEADROOM 256
50 
51 /* RX packet size EWMA. The average packet size is used to determine the packet
52  * buffer size when refilling RX rings. As the entire RX ring may be refilled
53  * at once, the weight is chosen so that the EWMA will be insensitive to short-
54  * term, transient changes in packet size.
55  */
56 DECLARE_EWMA(pkt_len, 0, 64)
57 
58 #define VIRTNET_DRIVER_VERSION "1.0.0"
59 
60 struct virtnet_stats {
61 	struct u64_stats_sync tx_syncp;
62 	struct u64_stats_sync rx_syncp;
63 	u64 tx_bytes;
64 	u64 tx_packets;
65 
66 	u64 rx_bytes;
67 	u64 rx_packets;
68 };
69 
70 /* Internal representation of a send virtqueue */
71 struct send_queue {
72 	/* Virtqueue associated with this send _queue */
73 	struct virtqueue *vq;
74 
75 	/* TX: fragments + linear part + virtio header */
76 	struct scatterlist sg[MAX_SKB_FRAGS + 2];
77 
78 	/* Name of the send queue: output.$index */
79 	char name[40];
80 
81 	struct napi_struct napi;
82 };
83 
84 /* Internal representation of a receive virtqueue */
85 struct receive_queue {
86 	/* Virtqueue associated with this receive_queue */
87 	struct virtqueue *vq;
88 
89 	struct napi_struct napi;
90 
91 	struct bpf_prog __rcu *xdp_prog;
92 
93 	/* Chain pages by the private ptr. */
94 	struct page *pages;
95 
96 	/* Average packet length for mergeable receive buffers. */
97 	struct ewma_pkt_len mrg_avg_pkt_len;
98 
99 	/* Page frag for packet buffer allocation. */
100 	struct page_frag alloc_frag;
101 
102 	/* RX: fragments + linear part + virtio header */
103 	struct scatterlist sg[MAX_SKB_FRAGS + 2];
104 
105 	/* Min single buffer size for mergeable buffers case. */
106 	unsigned int min_buf_len;
107 
108 	/* Name of this receive queue: input.$index */
109 	char name[40];
110 };
111 
112 struct virtnet_info {
113 	struct virtio_device *vdev;
114 	struct virtqueue *cvq;
115 	struct net_device *dev;
116 	struct send_queue *sq;
117 	struct receive_queue *rq;
118 	unsigned int status;
119 
120 	/* Max # of queue pairs supported by the device */
121 	u16 max_queue_pairs;
122 
123 	/* # of queue pairs currently used by the driver */
124 	u16 curr_queue_pairs;
125 
126 	/* # of XDP queue pairs currently used by the driver */
127 	u16 xdp_queue_pairs;
128 
129 	/* I like... big packets and I cannot lie! */
130 	bool big_packets;
131 
132 	/* Host will merge rx buffers for big packets (shake it! shake it!) */
133 	bool mergeable_rx_bufs;
134 
135 	/* Has control virtqueue */
136 	bool has_cvq;
137 
138 	/* Host can handle any s/g split between our header and packet data */
139 	bool any_header_sg;
140 
141 	/* Packet virtio header size */
142 	u8 hdr_len;
143 
144 	/* Active statistics */
145 	struct virtnet_stats __percpu *stats;
146 
147 	/* Work struct for refilling if we run low on memory. */
148 	struct delayed_work refill;
149 
150 	/* Work struct for config space updates */
151 	struct work_struct config_work;
152 
153 	/* Does the affinity hint is set for virtqueues? */
154 	bool affinity_hint_set;
155 
156 	/* CPU hotplug instances for online & dead */
157 	struct hlist_node node;
158 	struct hlist_node node_dead;
159 
160 	/* Control VQ buffers: protected by the rtnl lock */
161 	struct virtio_net_ctrl_hdr ctrl_hdr;
162 	virtio_net_ctrl_ack ctrl_status;
163 	struct virtio_net_ctrl_mq ctrl_mq;
164 	u8 ctrl_promisc;
165 	u8 ctrl_allmulti;
166 	u16 ctrl_vid;
167 
168 	/* Ethtool settings */
169 	u8 duplex;
170 	u32 speed;
171 };
172 
173 struct padded_vnet_hdr {
174 	struct virtio_net_hdr_mrg_rxbuf hdr;
175 	/*
176 	 * hdr is in a separate sg buffer, and data sg buffer shares same page
177 	 * with this header sg. This padding makes next sg 16 byte aligned
178 	 * after the header.
179 	 */
180 	char padding[4];
181 };
182 
183 /* Converting between virtqueue no. and kernel tx/rx queue no.
184  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
185  */
186 static int vq2txq(struct virtqueue *vq)
187 {
188 	return (vq->index - 1) / 2;
189 }
190 
191 static int txq2vq(int txq)
192 {
193 	return txq * 2 + 1;
194 }
195 
196 static int vq2rxq(struct virtqueue *vq)
197 {
198 	return vq->index / 2;
199 }
200 
201 static int rxq2vq(int rxq)
202 {
203 	return rxq * 2;
204 }
205 
206 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
207 {
208 	return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
209 }
210 
211 /*
212  * private is used to chain pages for big packets, put the whole
213  * most recent used list in the beginning for reuse
214  */
215 static void give_pages(struct receive_queue *rq, struct page *page)
216 {
217 	struct page *end;
218 
219 	/* Find end of list, sew whole thing into vi->rq.pages. */
220 	for (end = page; end->private; end = (struct page *)end->private);
221 	end->private = (unsigned long)rq->pages;
222 	rq->pages = page;
223 }
224 
225 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
226 {
227 	struct page *p = rq->pages;
228 
229 	if (p) {
230 		rq->pages = (struct page *)p->private;
231 		/* clear private here, it is used to chain pages */
232 		p->private = 0;
233 	} else
234 		p = alloc_page(gfp_mask);
235 	return p;
236 }
237 
238 static void virtqueue_napi_schedule(struct napi_struct *napi,
239 				    struct virtqueue *vq)
240 {
241 	if (napi_schedule_prep(napi)) {
242 		virtqueue_disable_cb(vq);
243 		__napi_schedule(napi);
244 	}
245 }
246 
247 static void virtqueue_napi_complete(struct napi_struct *napi,
248 				    struct virtqueue *vq, int processed)
249 {
250 	int opaque;
251 
252 	opaque = virtqueue_enable_cb_prepare(vq);
253 	if (napi_complete_done(napi, processed) &&
254 	    unlikely(virtqueue_poll(vq, opaque)))
255 		virtqueue_napi_schedule(napi, vq);
256 }
257 
258 static void skb_xmit_done(struct virtqueue *vq)
259 {
260 	struct virtnet_info *vi = vq->vdev->priv;
261 	struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
262 
263 	/* Suppress further interrupts. */
264 	virtqueue_disable_cb(vq);
265 
266 	if (napi->weight)
267 		virtqueue_napi_schedule(napi, vq);
268 	else
269 		/* We were probably waiting for more output buffers. */
270 		netif_wake_subqueue(vi->dev, vq2txq(vq));
271 }
272 
273 /* Called from bottom half context */
274 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
275 				   struct receive_queue *rq,
276 				   struct page *page, unsigned int offset,
277 				   unsigned int len, unsigned int truesize)
278 {
279 	struct sk_buff *skb;
280 	struct virtio_net_hdr_mrg_rxbuf *hdr;
281 	unsigned int copy, hdr_len, hdr_padded_len;
282 	char *p;
283 
284 	p = page_address(page) + offset;
285 
286 	/* copy small packet so we can reuse these pages for small data */
287 	skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
288 	if (unlikely(!skb))
289 		return NULL;
290 
291 	hdr = skb_vnet_hdr(skb);
292 
293 	hdr_len = vi->hdr_len;
294 	if (vi->mergeable_rx_bufs)
295 		hdr_padded_len = sizeof *hdr;
296 	else
297 		hdr_padded_len = sizeof(struct padded_vnet_hdr);
298 
299 	memcpy(hdr, p, hdr_len);
300 
301 	len -= hdr_len;
302 	offset += hdr_padded_len;
303 	p += hdr_padded_len;
304 
305 	copy = len;
306 	if (copy > skb_tailroom(skb))
307 		copy = skb_tailroom(skb);
308 	memcpy(skb_put(skb, copy), p, copy);
309 
310 	len -= copy;
311 	offset += copy;
312 
313 	if (vi->mergeable_rx_bufs) {
314 		if (len)
315 			skb_add_rx_frag(skb, 0, page, offset, len, truesize);
316 		else
317 			put_page(page);
318 		return skb;
319 	}
320 
321 	/*
322 	 * Verify that we can indeed put this data into a skb.
323 	 * This is here to handle cases when the device erroneously
324 	 * tries to receive more than is possible. This is usually
325 	 * the case of a broken device.
326 	 */
327 	if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
328 		net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
329 		dev_kfree_skb(skb);
330 		return NULL;
331 	}
332 	BUG_ON(offset >= PAGE_SIZE);
333 	while (len) {
334 		unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
335 		skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
336 				frag_size, truesize);
337 		len -= frag_size;
338 		page = (struct page *)page->private;
339 		offset = 0;
340 	}
341 
342 	if (page)
343 		give_pages(rq, page);
344 
345 	return skb;
346 }
347 
348 static bool virtnet_xdp_xmit(struct virtnet_info *vi,
349 			     struct receive_queue *rq,
350 			     struct xdp_buff *xdp)
351 {
352 	struct virtio_net_hdr_mrg_rxbuf *hdr;
353 	unsigned int len;
354 	struct send_queue *sq;
355 	unsigned int qp;
356 	void *xdp_sent;
357 	int err;
358 
359 	qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
360 	sq = &vi->sq[qp];
361 
362 	/* Free up any pending old buffers before queueing new ones. */
363 	while ((xdp_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) {
364 		struct page *sent_page = virt_to_head_page(xdp_sent);
365 
366 		put_page(sent_page);
367 	}
368 
369 	xdp->data -= vi->hdr_len;
370 	/* Zero header and leave csum up to XDP layers */
371 	hdr = xdp->data;
372 	memset(hdr, 0, vi->hdr_len);
373 
374 	sg_init_one(sq->sg, xdp->data, xdp->data_end - xdp->data);
375 
376 	err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp->data, GFP_ATOMIC);
377 	if (unlikely(err)) {
378 		struct page *page = virt_to_head_page(xdp->data);
379 
380 		put_page(page);
381 		return false;
382 	}
383 
384 	virtqueue_kick(sq->vq);
385 	return true;
386 }
387 
388 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
389 {
390 	return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
391 }
392 
393 static struct sk_buff *receive_small(struct net_device *dev,
394 				     struct virtnet_info *vi,
395 				     struct receive_queue *rq,
396 				     void *buf, unsigned int len)
397 {
398 	struct sk_buff *skb;
399 	struct bpf_prog *xdp_prog;
400 	unsigned int xdp_headroom = virtnet_get_headroom(vi);
401 	unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
402 	unsigned int headroom = vi->hdr_len + header_offset;
403 	unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
404 			      SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
405 	unsigned int delta = 0;
406 	len -= vi->hdr_len;
407 
408 	rcu_read_lock();
409 	xdp_prog = rcu_dereference(rq->xdp_prog);
410 	if (xdp_prog) {
411 		struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
412 		struct xdp_buff xdp;
413 		void *orig_data;
414 		u32 act;
415 
416 		if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags))
417 			goto err_xdp;
418 
419 		xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
420 		xdp.data = xdp.data_hard_start + xdp_headroom;
421 		xdp.data_end = xdp.data + len;
422 		orig_data = xdp.data;
423 		act = bpf_prog_run_xdp(xdp_prog, &xdp);
424 
425 		switch (act) {
426 		case XDP_PASS:
427 			/* Recalculate length in case bpf program changed it */
428 			delta = orig_data - xdp.data;
429 			break;
430 		case XDP_TX:
431 			if (unlikely(!virtnet_xdp_xmit(vi, rq, &xdp)))
432 				trace_xdp_exception(vi->dev, xdp_prog, act);
433 			rcu_read_unlock();
434 			goto xdp_xmit;
435 		default:
436 			bpf_warn_invalid_xdp_action(act);
437 		case XDP_ABORTED:
438 			trace_xdp_exception(vi->dev, xdp_prog, act);
439 		case XDP_DROP:
440 			goto err_xdp;
441 		}
442 	}
443 	rcu_read_unlock();
444 
445 	skb = build_skb(buf, buflen);
446 	if (!skb) {
447 		put_page(virt_to_head_page(buf));
448 		goto err;
449 	}
450 	skb_reserve(skb, headroom - delta);
451 	skb_put(skb, len + delta);
452 	if (!delta) {
453 		buf += header_offset;
454 		memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
455 	} /* keep zeroed vnet hdr since packet was changed by bpf */
456 
457 err:
458 	return skb;
459 
460 err_xdp:
461 	rcu_read_unlock();
462 	dev->stats.rx_dropped++;
463 	put_page(virt_to_head_page(buf));
464 xdp_xmit:
465 	return NULL;
466 }
467 
468 static struct sk_buff *receive_big(struct net_device *dev,
469 				   struct virtnet_info *vi,
470 				   struct receive_queue *rq,
471 				   void *buf,
472 				   unsigned int len)
473 {
474 	struct page *page = buf;
475 	struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
476 
477 	if (unlikely(!skb))
478 		goto err;
479 
480 	return skb;
481 
482 err:
483 	dev->stats.rx_dropped++;
484 	give_pages(rq, page);
485 	return NULL;
486 }
487 
488 /* The conditions to enable XDP should preclude the underlying device from
489  * sending packets across multiple buffers (num_buf > 1). However per spec
490  * it does not appear to be illegal to do so but rather just against convention.
491  * So in order to avoid making a system unresponsive the packets are pushed
492  * into a page and the XDP program is run. This will be extremely slow and we
493  * push a warning to the user to fix this as soon as possible. Fixing this may
494  * require resolving the underlying hardware to determine why multiple buffers
495  * are being received or simply loading the XDP program in the ingress stack
496  * after the skb is built because there is no advantage to running it here
497  * anymore.
498  */
499 static struct page *xdp_linearize_page(struct receive_queue *rq,
500 				       u16 *num_buf,
501 				       struct page *p,
502 				       int offset,
503 				       unsigned int *len)
504 {
505 	struct page *page = alloc_page(GFP_ATOMIC);
506 	unsigned int page_off = VIRTIO_XDP_HEADROOM;
507 
508 	if (!page)
509 		return NULL;
510 
511 	memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
512 	page_off += *len;
513 
514 	while (--*num_buf) {
515 		unsigned int buflen;
516 		void *buf;
517 		int off;
518 
519 		buf = virtqueue_get_buf(rq->vq, &buflen);
520 		if (unlikely(!buf))
521 			goto err_buf;
522 
523 		p = virt_to_head_page(buf);
524 		off = buf - page_address(p);
525 
526 		/* guard against a misconfigured or uncooperative backend that
527 		 * is sending packet larger than the MTU.
528 		 */
529 		if ((page_off + buflen) > PAGE_SIZE) {
530 			put_page(p);
531 			goto err_buf;
532 		}
533 
534 		memcpy(page_address(page) + page_off,
535 		       page_address(p) + off, buflen);
536 		page_off += buflen;
537 		put_page(p);
538 	}
539 
540 	/* Headroom does not contribute to packet length */
541 	*len = page_off - VIRTIO_XDP_HEADROOM;
542 	return page;
543 err_buf:
544 	__free_pages(page, 0);
545 	return NULL;
546 }
547 
548 static struct sk_buff *receive_mergeable(struct net_device *dev,
549 					 struct virtnet_info *vi,
550 					 struct receive_queue *rq,
551 					 void *buf,
552 					 void *ctx,
553 					 unsigned int len)
554 {
555 	struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
556 	u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
557 	struct page *page = virt_to_head_page(buf);
558 	int offset = buf - page_address(page);
559 	struct sk_buff *head_skb, *curr_skb;
560 	struct bpf_prog *xdp_prog;
561 	unsigned int truesize;
562 
563 	head_skb = NULL;
564 
565 	rcu_read_lock();
566 	xdp_prog = rcu_dereference(rq->xdp_prog);
567 	if (xdp_prog) {
568 		struct page *xdp_page;
569 		struct xdp_buff xdp;
570 		void *data;
571 		u32 act;
572 
573 		/* This happens when rx buffer size is underestimated */
574 		if (unlikely(num_buf > 1)) {
575 			/* linearize data for XDP */
576 			xdp_page = xdp_linearize_page(rq, &num_buf,
577 						      page, offset, &len);
578 			if (!xdp_page)
579 				goto err_xdp;
580 			offset = VIRTIO_XDP_HEADROOM;
581 		} else {
582 			xdp_page = page;
583 		}
584 
585 		/* Transient failure which in theory could occur if
586 		 * in-flight packets from before XDP was enabled reach
587 		 * the receive path after XDP is loaded. In practice I
588 		 * was not able to create this condition.
589 		 */
590 		if (unlikely(hdr->hdr.gso_type))
591 			goto err_xdp;
592 
593 		/* Allow consuming headroom but reserve enough space to push
594 		 * the descriptor on if we get an XDP_TX return code.
595 		 */
596 		data = page_address(xdp_page) + offset;
597 		xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
598 		xdp.data = data + vi->hdr_len;
599 		xdp.data_end = xdp.data + (len - vi->hdr_len);
600 		act = bpf_prog_run_xdp(xdp_prog, &xdp);
601 
602 		switch (act) {
603 		case XDP_PASS:
604 			/* recalculate offset to account for any header
605 			 * adjustments. Note other cases do not build an
606 			 * skb and avoid using offset
607 			 */
608 			offset = xdp.data -
609 					page_address(xdp_page) - vi->hdr_len;
610 
611 			/* We can only create skb based on xdp_page. */
612 			if (unlikely(xdp_page != page)) {
613 				rcu_read_unlock();
614 				put_page(page);
615 				head_skb = page_to_skb(vi, rq, xdp_page,
616 						       offset, len, PAGE_SIZE);
617 				ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
618 				return head_skb;
619 			}
620 			break;
621 		case XDP_TX:
622 			if (unlikely(!virtnet_xdp_xmit(vi, rq, &xdp)))
623 				trace_xdp_exception(vi->dev, xdp_prog, act);
624 			ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
625 			if (unlikely(xdp_page != page))
626 				goto err_xdp;
627 			rcu_read_unlock();
628 			goto xdp_xmit;
629 		default:
630 			bpf_warn_invalid_xdp_action(act);
631 		case XDP_ABORTED:
632 			trace_xdp_exception(vi->dev, xdp_prog, act);
633 		case XDP_DROP:
634 			if (unlikely(xdp_page != page))
635 				__free_pages(xdp_page, 0);
636 			ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
637 			goto err_xdp;
638 		}
639 	}
640 	rcu_read_unlock();
641 
642 	if (unlikely(len > (unsigned long)ctx)) {
643 		pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
644 			 dev->name, len, (unsigned long)ctx);
645 		dev->stats.rx_length_errors++;
646 		goto err_skb;
647 	}
648 	truesize = (unsigned long)ctx;
649 	head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
650 	curr_skb = head_skb;
651 
652 	if (unlikely(!curr_skb))
653 		goto err_skb;
654 	while (--num_buf) {
655 		int num_skb_frags;
656 
657 		buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
658 		if (unlikely(!ctx)) {
659 			pr_debug("%s: rx error: %d buffers out of %d missing\n",
660 				 dev->name, num_buf,
661 				 virtio16_to_cpu(vi->vdev,
662 						 hdr->num_buffers));
663 			dev->stats.rx_length_errors++;
664 			goto err_buf;
665 		}
666 
667 		page = virt_to_head_page(buf);
668 		if (unlikely(len > (unsigned long)ctx)) {
669 			pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
670 				 dev->name, len, (unsigned long)ctx);
671 			dev->stats.rx_length_errors++;
672 			goto err_skb;
673 		}
674 		truesize = (unsigned long)ctx;
675 
676 		num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
677 		if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
678 			struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
679 
680 			if (unlikely(!nskb))
681 				goto err_skb;
682 			if (curr_skb == head_skb)
683 				skb_shinfo(curr_skb)->frag_list = nskb;
684 			else
685 				curr_skb->next = nskb;
686 			curr_skb = nskb;
687 			head_skb->truesize += nskb->truesize;
688 			num_skb_frags = 0;
689 		}
690 		if (curr_skb != head_skb) {
691 			head_skb->data_len += len;
692 			head_skb->len += len;
693 			head_skb->truesize += truesize;
694 		}
695 		offset = buf - page_address(page);
696 		if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
697 			put_page(page);
698 			skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
699 					     len, truesize);
700 		} else {
701 			skb_add_rx_frag(curr_skb, num_skb_frags, page,
702 					offset, len, truesize);
703 		}
704 	}
705 
706 	ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
707 	return head_skb;
708 
709 err_xdp:
710 	rcu_read_unlock();
711 err_skb:
712 	put_page(page);
713 	while (--num_buf) {
714 		buf = virtqueue_get_buf(rq->vq, &len);
715 		if (unlikely(!buf)) {
716 			pr_debug("%s: rx error: %d buffers missing\n",
717 				 dev->name, num_buf);
718 			dev->stats.rx_length_errors++;
719 			break;
720 		}
721 		page = virt_to_head_page(buf);
722 		put_page(page);
723 	}
724 err_buf:
725 	dev->stats.rx_dropped++;
726 	dev_kfree_skb(head_skb);
727 xdp_xmit:
728 	return NULL;
729 }
730 
731 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
732 		       void *buf, unsigned int len, void **ctx)
733 {
734 	struct net_device *dev = vi->dev;
735 	struct sk_buff *skb;
736 	struct virtio_net_hdr_mrg_rxbuf *hdr;
737 	int ret;
738 
739 	if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
740 		pr_debug("%s: short packet %i\n", dev->name, len);
741 		dev->stats.rx_length_errors++;
742 		if (vi->mergeable_rx_bufs) {
743 			put_page(virt_to_head_page(buf));
744 		} else if (vi->big_packets) {
745 			give_pages(rq, buf);
746 		} else {
747 			put_page(virt_to_head_page(buf));
748 		}
749 		return 0;
750 	}
751 
752 	if (vi->mergeable_rx_bufs)
753 		skb = receive_mergeable(dev, vi, rq, buf, ctx, len);
754 	else if (vi->big_packets)
755 		skb = receive_big(dev, vi, rq, buf, len);
756 	else
757 		skb = receive_small(dev, vi, rq, buf, len);
758 
759 	if (unlikely(!skb))
760 		return 0;
761 
762 	hdr = skb_vnet_hdr(skb);
763 
764 	ret = skb->len;
765 
766 	if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
767 		skb->ip_summed = CHECKSUM_UNNECESSARY;
768 
769 	if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
770 				  virtio_is_little_endian(vi->vdev))) {
771 		net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
772 				     dev->name, hdr->hdr.gso_type,
773 				     hdr->hdr.gso_size);
774 		goto frame_err;
775 	}
776 
777 	skb->protocol = eth_type_trans(skb, dev);
778 	pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
779 		 ntohs(skb->protocol), skb->len, skb->pkt_type);
780 
781 	napi_gro_receive(&rq->napi, skb);
782 	return ret;
783 
784 frame_err:
785 	dev->stats.rx_frame_errors++;
786 	dev_kfree_skb(skb);
787 	return 0;
788 }
789 
790 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
791 			     gfp_t gfp)
792 {
793 	struct page_frag *alloc_frag = &rq->alloc_frag;
794 	char *buf;
795 	unsigned int xdp_headroom = virtnet_get_headroom(vi);
796 	int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
797 	int err;
798 
799 	len = SKB_DATA_ALIGN(len) +
800 	      SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
801 	if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
802 		return -ENOMEM;
803 
804 	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
805 	get_page(alloc_frag->page);
806 	alloc_frag->offset += len;
807 	sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
808 		    vi->hdr_len + GOOD_PACKET_LEN);
809 	err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, buf, gfp);
810 	if (err < 0)
811 		put_page(virt_to_head_page(buf));
812 
813 	return err;
814 }
815 
816 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
817 			   gfp_t gfp)
818 {
819 	struct page *first, *list = NULL;
820 	char *p;
821 	int i, err, offset;
822 
823 	sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
824 
825 	/* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
826 	for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
827 		first = get_a_page(rq, gfp);
828 		if (!first) {
829 			if (list)
830 				give_pages(rq, list);
831 			return -ENOMEM;
832 		}
833 		sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
834 
835 		/* chain new page in list head to match sg */
836 		first->private = (unsigned long)list;
837 		list = first;
838 	}
839 
840 	first = get_a_page(rq, gfp);
841 	if (!first) {
842 		give_pages(rq, list);
843 		return -ENOMEM;
844 	}
845 	p = page_address(first);
846 
847 	/* rq->sg[0], rq->sg[1] share the same page */
848 	/* a separated rq->sg[0] for header - required in case !any_header_sg */
849 	sg_set_buf(&rq->sg[0], p, vi->hdr_len);
850 
851 	/* rq->sg[1] for data packet, from offset */
852 	offset = sizeof(struct padded_vnet_hdr);
853 	sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
854 
855 	/* chain first in list head */
856 	first->private = (unsigned long)list;
857 	err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
858 				  first, gfp);
859 	if (err < 0)
860 		give_pages(rq, first);
861 
862 	return err;
863 }
864 
865 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
866 					  struct ewma_pkt_len *avg_pkt_len)
867 {
868 	const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
869 	unsigned int len;
870 
871 	len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
872 				rq->min_buf_len - hdr_len, PAGE_SIZE - hdr_len);
873 	return ALIGN(len, L1_CACHE_BYTES);
874 }
875 
876 static int add_recvbuf_mergeable(struct virtnet_info *vi,
877 				 struct receive_queue *rq, gfp_t gfp)
878 {
879 	struct page_frag *alloc_frag = &rq->alloc_frag;
880 	unsigned int headroom = virtnet_get_headroom(vi);
881 	char *buf;
882 	void *ctx;
883 	int err;
884 	unsigned int len, hole;
885 
886 	len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len);
887 	if (unlikely(!skb_page_frag_refill(len + headroom, alloc_frag, gfp)))
888 		return -ENOMEM;
889 
890 	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
891 	buf += headroom; /* advance address leaving hole at front of pkt */
892 	ctx = (void *)(unsigned long)len;
893 	get_page(alloc_frag->page);
894 	alloc_frag->offset += len + headroom;
895 	hole = alloc_frag->size - alloc_frag->offset;
896 	if (hole < len + headroom) {
897 		/* To avoid internal fragmentation, if there is very likely not
898 		 * enough space for another buffer, add the remaining space to
899 		 * the current buffer. This extra space is not included in
900 		 * the truesize stored in ctx.
901 		 */
902 		len += hole;
903 		alloc_frag->offset += hole;
904 	}
905 
906 	sg_init_one(rq->sg, buf, len);
907 	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
908 	if (err < 0)
909 		put_page(virt_to_head_page(buf));
910 
911 	return err;
912 }
913 
914 /*
915  * Returns false if we couldn't fill entirely (OOM).
916  *
917  * Normally run in the receive path, but can also be run from ndo_open
918  * before we're receiving packets, or from refill_work which is
919  * careful to disable receiving (using napi_disable).
920  */
921 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
922 			  gfp_t gfp)
923 {
924 	int err;
925 	bool oom;
926 
927 	gfp |= __GFP_COLD;
928 	do {
929 		if (vi->mergeable_rx_bufs)
930 			err = add_recvbuf_mergeable(vi, rq, gfp);
931 		else if (vi->big_packets)
932 			err = add_recvbuf_big(vi, rq, gfp);
933 		else
934 			err = add_recvbuf_small(vi, rq, gfp);
935 
936 		oom = err == -ENOMEM;
937 		if (err)
938 			break;
939 	} while (rq->vq->num_free);
940 	virtqueue_kick(rq->vq);
941 	return !oom;
942 }
943 
944 static void skb_recv_done(struct virtqueue *rvq)
945 {
946 	struct virtnet_info *vi = rvq->vdev->priv;
947 	struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
948 
949 	virtqueue_napi_schedule(&rq->napi, rvq);
950 }
951 
952 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
953 {
954 	napi_enable(napi);
955 
956 	/* If all buffers were filled by other side before we napi_enabled, we
957 	 * won't get another interrupt, so process any outstanding packets now.
958 	 * Call local_bh_enable after to trigger softIRQ processing.
959 	 */
960 	local_bh_disable();
961 	virtqueue_napi_schedule(napi, vq);
962 	local_bh_enable();
963 }
964 
965 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
966 				   struct virtqueue *vq,
967 				   struct napi_struct *napi)
968 {
969 	if (!napi->weight)
970 		return;
971 
972 	/* Tx napi touches cachelines on the cpu handling tx interrupts. Only
973 	 * enable the feature if this is likely affine with the transmit path.
974 	 */
975 	if (!vi->affinity_hint_set) {
976 		napi->weight = 0;
977 		return;
978 	}
979 
980 	return virtnet_napi_enable(vq, napi);
981 }
982 
983 static void virtnet_napi_tx_disable(struct napi_struct *napi)
984 {
985 	if (napi->weight)
986 		napi_disable(napi);
987 }
988 
989 static void refill_work(struct work_struct *work)
990 {
991 	struct virtnet_info *vi =
992 		container_of(work, struct virtnet_info, refill.work);
993 	bool still_empty;
994 	int i;
995 
996 	for (i = 0; i < vi->curr_queue_pairs; i++) {
997 		struct receive_queue *rq = &vi->rq[i];
998 
999 		napi_disable(&rq->napi);
1000 		still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1001 		virtnet_napi_enable(rq->vq, &rq->napi);
1002 
1003 		/* In theory, this can happen: if we don't get any buffers in
1004 		 * we will *never* try to fill again.
1005 		 */
1006 		if (still_empty)
1007 			schedule_delayed_work(&vi->refill, HZ/2);
1008 	}
1009 }
1010 
1011 static int virtnet_receive(struct receive_queue *rq, int budget)
1012 {
1013 	struct virtnet_info *vi = rq->vq->vdev->priv;
1014 	unsigned int len, received = 0, bytes = 0;
1015 	void *buf;
1016 	struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1017 
1018 	if (vi->mergeable_rx_bufs) {
1019 		void *ctx;
1020 
1021 		while (received < budget &&
1022 		       (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1023 			bytes += receive_buf(vi, rq, buf, len, ctx);
1024 			received++;
1025 		}
1026 	} else {
1027 		while (received < budget &&
1028 		       (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1029 			bytes += receive_buf(vi, rq, buf, len, NULL);
1030 			received++;
1031 		}
1032 	}
1033 
1034 	if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1035 		if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1036 			schedule_delayed_work(&vi->refill, 0);
1037 	}
1038 
1039 	u64_stats_update_begin(&stats->rx_syncp);
1040 	stats->rx_bytes += bytes;
1041 	stats->rx_packets += received;
1042 	u64_stats_update_end(&stats->rx_syncp);
1043 
1044 	return received;
1045 }
1046 
1047 static void free_old_xmit_skbs(struct send_queue *sq)
1048 {
1049 	struct sk_buff *skb;
1050 	unsigned int len;
1051 	struct virtnet_info *vi = sq->vq->vdev->priv;
1052 	struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1053 	unsigned int packets = 0;
1054 	unsigned int bytes = 0;
1055 
1056 	while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1057 		pr_debug("Sent skb %p\n", skb);
1058 
1059 		bytes += skb->len;
1060 		packets++;
1061 
1062 		dev_kfree_skb_any(skb);
1063 	}
1064 
1065 	/* Avoid overhead when no packets have been processed
1066 	 * happens when called speculatively from start_xmit.
1067 	 */
1068 	if (!packets)
1069 		return;
1070 
1071 	u64_stats_update_begin(&stats->tx_syncp);
1072 	stats->tx_bytes += bytes;
1073 	stats->tx_packets += packets;
1074 	u64_stats_update_end(&stats->tx_syncp);
1075 }
1076 
1077 static void virtnet_poll_cleantx(struct receive_queue *rq)
1078 {
1079 	struct virtnet_info *vi = rq->vq->vdev->priv;
1080 	unsigned int index = vq2rxq(rq->vq);
1081 	struct send_queue *sq = &vi->sq[index];
1082 	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1083 
1084 	if (!sq->napi.weight)
1085 		return;
1086 
1087 	if (__netif_tx_trylock(txq)) {
1088 		free_old_xmit_skbs(sq);
1089 		__netif_tx_unlock(txq);
1090 	}
1091 
1092 	if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1093 		netif_tx_wake_queue(txq);
1094 }
1095 
1096 static int virtnet_poll(struct napi_struct *napi, int budget)
1097 {
1098 	struct receive_queue *rq =
1099 		container_of(napi, struct receive_queue, napi);
1100 	unsigned int received;
1101 
1102 	virtnet_poll_cleantx(rq);
1103 
1104 	received = virtnet_receive(rq, budget);
1105 
1106 	/* Out of packets? */
1107 	if (received < budget)
1108 		virtqueue_napi_complete(napi, rq->vq, received);
1109 
1110 	return received;
1111 }
1112 
1113 static int virtnet_open(struct net_device *dev)
1114 {
1115 	struct virtnet_info *vi = netdev_priv(dev);
1116 	int i;
1117 
1118 	for (i = 0; i < vi->max_queue_pairs; i++) {
1119 		if (i < vi->curr_queue_pairs)
1120 			/* Make sure we have some buffers: if oom use wq. */
1121 			if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1122 				schedule_delayed_work(&vi->refill, 0);
1123 		virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1124 		virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1125 	}
1126 
1127 	return 0;
1128 }
1129 
1130 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1131 {
1132 	struct send_queue *sq = container_of(napi, struct send_queue, napi);
1133 	struct virtnet_info *vi = sq->vq->vdev->priv;
1134 	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
1135 
1136 	__netif_tx_lock(txq, raw_smp_processor_id());
1137 	free_old_xmit_skbs(sq);
1138 	__netif_tx_unlock(txq);
1139 
1140 	virtqueue_napi_complete(napi, sq->vq, 0);
1141 
1142 	if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1143 		netif_tx_wake_queue(txq);
1144 
1145 	return 0;
1146 }
1147 
1148 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1149 {
1150 	struct virtio_net_hdr_mrg_rxbuf *hdr;
1151 	const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1152 	struct virtnet_info *vi = sq->vq->vdev->priv;
1153 	unsigned num_sg;
1154 	unsigned hdr_len = vi->hdr_len;
1155 	bool can_push;
1156 
1157 	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1158 
1159 	can_push = vi->any_header_sg &&
1160 		!((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1161 		!skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1162 	/* Even if we can, don't push here yet as this would skew
1163 	 * csum_start offset below. */
1164 	if (can_push)
1165 		hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1166 	else
1167 		hdr = skb_vnet_hdr(skb);
1168 
1169 	if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1170 				    virtio_is_little_endian(vi->vdev), false))
1171 		BUG();
1172 
1173 	if (vi->mergeable_rx_bufs)
1174 		hdr->num_buffers = 0;
1175 
1176 	sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1177 	if (can_push) {
1178 		__skb_push(skb, hdr_len);
1179 		num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1180 		/* Pull header back to avoid skew in tx bytes calculations. */
1181 		__skb_pull(skb, hdr_len);
1182 	} else {
1183 		sg_set_buf(sq->sg, hdr, hdr_len);
1184 		num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
1185 	}
1186 	return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1187 }
1188 
1189 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1190 {
1191 	struct virtnet_info *vi = netdev_priv(dev);
1192 	int qnum = skb_get_queue_mapping(skb);
1193 	struct send_queue *sq = &vi->sq[qnum];
1194 	int err;
1195 	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1196 	bool kick = !skb->xmit_more;
1197 	bool use_napi = sq->napi.weight;
1198 
1199 	/* Free up any pending old buffers before queueing new ones. */
1200 	free_old_xmit_skbs(sq);
1201 
1202 	if (use_napi && kick)
1203 		virtqueue_enable_cb_delayed(sq->vq);
1204 
1205 	/* timestamp packet in software */
1206 	skb_tx_timestamp(skb);
1207 
1208 	/* Try to transmit */
1209 	err = xmit_skb(sq, skb);
1210 
1211 	/* This should not happen! */
1212 	if (unlikely(err)) {
1213 		dev->stats.tx_fifo_errors++;
1214 		if (net_ratelimit())
1215 			dev_warn(&dev->dev,
1216 				 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1217 		dev->stats.tx_dropped++;
1218 		dev_kfree_skb_any(skb);
1219 		return NETDEV_TX_OK;
1220 	}
1221 
1222 	/* Don't wait up for transmitted skbs to be freed. */
1223 	if (!use_napi) {
1224 		skb_orphan(skb);
1225 		nf_reset(skb);
1226 	}
1227 
1228 	/* If running out of space, stop queue to avoid getting packets that we
1229 	 * are then unable to transmit.
1230 	 * An alternative would be to force queuing layer to requeue the skb by
1231 	 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1232 	 * returned in a normal path of operation: it means that driver is not
1233 	 * maintaining the TX queue stop/start state properly, and causes
1234 	 * the stack to do a non-trivial amount of useless work.
1235 	 * Since most packets only take 1 or 2 ring slots, stopping the queue
1236 	 * early means 16 slots are typically wasted.
1237 	 */
1238 	if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1239 		netif_stop_subqueue(dev, qnum);
1240 		if (!use_napi &&
1241 		    unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1242 			/* More just got used, free them then recheck. */
1243 			free_old_xmit_skbs(sq);
1244 			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1245 				netif_start_subqueue(dev, qnum);
1246 				virtqueue_disable_cb(sq->vq);
1247 			}
1248 		}
1249 	}
1250 
1251 	if (kick || netif_xmit_stopped(txq))
1252 		virtqueue_kick(sq->vq);
1253 
1254 	return NETDEV_TX_OK;
1255 }
1256 
1257 /*
1258  * Send command via the control virtqueue and check status.  Commands
1259  * supported by the hypervisor, as indicated by feature bits, should
1260  * never fail unless improperly formatted.
1261  */
1262 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1263 				 struct scatterlist *out)
1264 {
1265 	struct scatterlist *sgs[4], hdr, stat;
1266 	unsigned out_num = 0, tmp;
1267 
1268 	/* Caller should know better */
1269 	BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1270 
1271 	vi->ctrl_status = ~0;
1272 	vi->ctrl_hdr.class = class;
1273 	vi->ctrl_hdr.cmd = cmd;
1274 	/* Add header */
1275 	sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
1276 	sgs[out_num++] = &hdr;
1277 
1278 	if (out)
1279 		sgs[out_num++] = out;
1280 
1281 	/* Add return status. */
1282 	sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1283 	sgs[out_num] = &stat;
1284 
1285 	BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1286 	virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1287 
1288 	if (unlikely(!virtqueue_kick(vi->cvq)))
1289 		return vi->ctrl_status == VIRTIO_NET_OK;
1290 
1291 	/* Spin for a response, the kick causes an ioport write, trapping
1292 	 * into the hypervisor, so the request should be handled immediately.
1293 	 */
1294 	while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1295 	       !virtqueue_is_broken(vi->cvq))
1296 		cpu_relax();
1297 
1298 	return vi->ctrl_status == VIRTIO_NET_OK;
1299 }
1300 
1301 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1302 {
1303 	struct virtnet_info *vi = netdev_priv(dev);
1304 	struct virtio_device *vdev = vi->vdev;
1305 	int ret;
1306 	struct sockaddr *addr;
1307 	struct scatterlist sg;
1308 
1309 	addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1310 	if (!addr)
1311 		return -ENOMEM;
1312 
1313 	ret = eth_prepare_mac_addr_change(dev, addr);
1314 	if (ret)
1315 		goto out;
1316 
1317 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1318 		sg_init_one(&sg, addr->sa_data, dev->addr_len);
1319 		if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1320 					  VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1321 			dev_warn(&vdev->dev,
1322 				 "Failed to set mac address by vq command.\n");
1323 			ret = -EINVAL;
1324 			goto out;
1325 		}
1326 	} else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1327 		   !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1328 		unsigned int i;
1329 
1330 		/* Naturally, this has an atomicity problem. */
1331 		for (i = 0; i < dev->addr_len; i++)
1332 			virtio_cwrite8(vdev,
1333 				       offsetof(struct virtio_net_config, mac) +
1334 				       i, addr->sa_data[i]);
1335 	}
1336 
1337 	eth_commit_mac_addr_change(dev, p);
1338 	ret = 0;
1339 
1340 out:
1341 	kfree(addr);
1342 	return ret;
1343 }
1344 
1345 static void virtnet_stats(struct net_device *dev,
1346 			  struct rtnl_link_stats64 *tot)
1347 {
1348 	struct virtnet_info *vi = netdev_priv(dev);
1349 	int cpu;
1350 	unsigned int start;
1351 
1352 	for_each_possible_cpu(cpu) {
1353 		struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1354 		u64 tpackets, tbytes, rpackets, rbytes;
1355 
1356 		do {
1357 			start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1358 			tpackets = stats->tx_packets;
1359 			tbytes   = stats->tx_bytes;
1360 		} while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1361 
1362 		do {
1363 			start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1364 			rpackets = stats->rx_packets;
1365 			rbytes   = stats->rx_bytes;
1366 		} while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1367 
1368 		tot->rx_packets += rpackets;
1369 		tot->tx_packets += tpackets;
1370 		tot->rx_bytes   += rbytes;
1371 		tot->tx_bytes   += tbytes;
1372 	}
1373 
1374 	tot->tx_dropped = dev->stats.tx_dropped;
1375 	tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1376 	tot->rx_dropped = dev->stats.rx_dropped;
1377 	tot->rx_length_errors = dev->stats.rx_length_errors;
1378 	tot->rx_frame_errors = dev->stats.rx_frame_errors;
1379 }
1380 
1381 #ifdef CONFIG_NET_POLL_CONTROLLER
1382 static void virtnet_netpoll(struct net_device *dev)
1383 {
1384 	struct virtnet_info *vi = netdev_priv(dev);
1385 	int i;
1386 
1387 	for (i = 0; i < vi->curr_queue_pairs; i++)
1388 		napi_schedule(&vi->rq[i].napi);
1389 }
1390 #endif
1391 
1392 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1393 {
1394 	rtnl_lock();
1395 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1396 				  VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1397 		dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1398 	rtnl_unlock();
1399 }
1400 
1401 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1402 {
1403 	struct scatterlist sg;
1404 	struct net_device *dev = vi->dev;
1405 
1406 	if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1407 		return 0;
1408 
1409 	vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1410 	sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1411 
1412 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1413 				  VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1414 		dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1415 			 queue_pairs);
1416 		return -EINVAL;
1417 	} else {
1418 		vi->curr_queue_pairs = queue_pairs;
1419 		/* virtnet_open() will refill when device is going to up. */
1420 		if (dev->flags & IFF_UP)
1421 			schedule_delayed_work(&vi->refill, 0);
1422 	}
1423 
1424 	return 0;
1425 }
1426 
1427 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1428 {
1429 	int err;
1430 
1431 	rtnl_lock();
1432 	err = _virtnet_set_queues(vi, queue_pairs);
1433 	rtnl_unlock();
1434 	return err;
1435 }
1436 
1437 static int virtnet_close(struct net_device *dev)
1438 {
1439 	struct virtnet_info *vi = netdev_priv(dev);
1440 	int i;
1441 
1442 	/* Make sure refill_work doesn't re-enable napi! */
1443 	cancel_delayed_work_sync(&vi->refill);
1444 
1445 	for (i = 0; i < vi->max_queue_pairs; i++) {
1446 		napi_disable(&vi->rq[i].napi);
1447 		virtnet_napi_tx_disable(&vi->sq[i].napi);
1448 	}
1449 
1450 	return 0;
1451 }
1452 
1453 static void virtnet_set_rx_mode(struct net_device *dev)
1454 {
1455 	struct virtnet_info *vi = netdev_priv(dev);
1456 	struct scatterlist sg[2];
1457 	struct virtio_net_ctrl_mac *mac_data;
1458 	struct netdev_hw_addr *ha;
1459 	int uc_count;
1460 	int mc_count;
1461 	void *buf;
1462 	int i;
1463 
1464 	/* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1465 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1466 		return;
1467 
1468 	vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1469 	vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1470 
1471 	sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1472 
1473 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1474 				  VIRTIO_NET_CTRL_RX_PROMISC, sg))
1475 		dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1476 			 vi->ctrl_promisc ? "en" : "dis");
1477 
1478 	sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1479 
1480 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1481 				  VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1482 		dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1483 			 vi->ctrl_allmulti ? "en" : "dis");
1484 
1485 	uc_count = netdev_uc_count(dev);
1486 	mc_count = netdev_mc_count(dev);
1487 	/* MAC filter - use one buffer for both lists */
1488 	buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1489 		      (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1490 	mac_data = buf;
1491 	if (!buf)
1492 		return;
1493 
1494 	sg_init_table(sg, 2);
1495 
1496 	/* Store the unicast list and count in the front of the buffer */
1497 	mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1498 	i = 0;
1499 	netdev_for_each_uc_addr(ha, dev)
1500 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1501 
1502 	sg_set_buf(&sg[0], mac_data,
1503 		   sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1504 
1505 	/* multicast list and count fill the end */
1506 	mac_data = (void *)&mac_data->macs[uc_count][0];
1507 
1508 	mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1509 	i = 0;
1510 	netdev_for_each_mc_addr(ha, dev)
1511 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1512 
1513 	sg_set_buf(&sg[1], mac_data,
1514 		   sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1515 
1516 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1517 				  VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1518 		dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1519 
1520 	kfree(buf);
1521 }
1522 
1523 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1524 				   __be16 proto, u16 vid)
1525 {
1526 	struct virtnet_info *vi = netdev_priv(dev);
1527 	struct scatterlist sg;
1528 
1529 	vi->ctrl_vid = vid;
1530 	sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1531 
1532 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1533 				  VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1534 		dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1535 	return 0;
1536 }
1537 
1538 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1539 				    __be16 proto, u16 vid)
1540 {
1541 	struct virtnet_info *vi = netdev_priv(dev);
1542 	struct scatterlist sg;
1543 
1544 	vi->ctrl_vid = vid;
1545 	sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1546 
1547 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1548 				  VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1549 		dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1550 	return 0;
1551 }
1552 
1553 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1554 {
1555 	int i;
1556 
1557 	if (vi->affinity_hint_set) {
1558 		for (i = 0; i < vi->max_queue_pairs; i++) {
1559 			virtqueue_set_affinity(vi->rq[i].vq, -1);
1560 			virtqueue_set_affinity(vi->sq[i].vq, -1);
1561 		}
1562 
1563 		vi->affinity_hint_set = false;
1564 	}
1565 }
1566 
1567 static void virtnet_set_affinity(struct virtnet_info *vi)
1568 {
1569 	int i;
1570 	int cpu;
1571 
1572 	/* In multiqueue mode, when the number of cpu is equal to the number of
1573 	 * queue pairs, we let the queue pairs to be private to one cpu by
1574 	 * setting the affinity hint to eliminate the contention.
1575 	 */
1576 	if (vi->curr_queue_pairs == 1 ||
1577 	    vi->max_queue_pairs != num_online_cpus()) {
1578 		virtnet_clean_affinity(vi, -1);
1579 		return;
1580 	}
1581 
1582 	i = 0;
1583 	for_each_online_cpu(cpu) {
1584 		virtqueue_set_affinity(vi->rq[i].vq, cpu);
1585 		virtqueue_set_affinity(vi->sq[i].vq, cpu);
1586 		netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1587 		i++;
1588 	}
1589 
1590 	vi->affinity_hint_set = true;
1591 }
1592 
1593 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1594 {
1595 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1596 						   node);
1597 	virtnet_set_affinity(vi);
1598 	return 0;
1599 }
1600 
1601 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1602 {
1603 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1604 						   node_dead);
1605 	virtnet_set_affinity(vi);
1606 	return 0;
1607 }
1608 
1609 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1610 {
1611 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1612 						   node);
1613 
1614 	virtnet_clean_affinity(vi, cpu);
1615 	return 0;
1616 }
1617 
1618 static enum cpuhp_state virtionet_online;
1619 
1620 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1621 {
1622 	int ret;
1623 
1624 	ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1625 	if (ret)
1626 		return ret;
1627 	ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1628 					       &vi->node_dead);
1629 	if (!ret)
1630 		return ret;
1631 	cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1632 	return ret;
1633 }
1634 
1635 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1636 {
1637 	cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1638 	cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1639 					    &vi->node_dead);
1640 }
1641 
1642 static void virtnet_get_ringparam(struct net_device *dev,
1643 				struct ethtool_ringparam *ring)
1644 {
1645 	struct virtnet_info *vi = netdev_priv(dev);
1646 
1647 	ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1648 	ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1649 	ring->rx_pending = ring->rx_max_pending;
1650 	ring->tx_pending = ring->tx_max_pending;
1651 }
1652 
1653 
1654 static void virtnet_get_drvinfo(struct net_device *dev,
1655 				struct ethtool_drvinfo *info)
1656 {
1657 	struct virtnet_info *vi = netdev_priv(dev);
1658 	struct virtio_device *vdev = vi->vdev;
1659 
1660 	strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1661 	strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1662 	strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1663 
1664 }
1665 
1666 /* TODO: Eliminate OOO packets during switching */
1667 static int virtnet_set_channels(struct net_device *dev,
1668 				struct ethtool_channels *channels)
1669 {
1670 	struct virtnet_info *vi = netdev_priv(dev);
1671 	u16 queue_pairs = channels->combined_count;
1672 	int err;
1673 
1674 	/* We don't support separate rx/tx channels.
1675 	 * We don't allow setting 'other' channels.
1676 	 */
1677 	if (channels->rx_count || channels->tx_count || channels->other_count)
1678 		return -EINVAL;
1679 
1680 	if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1681 		return -EINVAL;
1682 
1683 	/* For now we don't support modifying channels while XDP is loaded
1684 	 * also when XDP is loaded all RX queues have XDP programs so we only
1685 	 * need to check a single RX queue.
1686 	 */
1687 	if (vi->rq[0].xdp_prog)
1688 		return -EINVAL;
1689 
1690 	get_online_cpus();
1691 	err = _virtnet_set_queues(vi, queue_pairs);
1692 	if (!err) {
1693 		netif_set_real_num_tx_queues(dev, queue_pairs);
1694 		netif_set_real_num_rx_queues(dev, queue_pairs);
1695 
1696 		virtnet_set_affinity(vi);
1697 	}
1698 	put_online_cpus();
1699 
1700 	return err;
1701 }
1702 
1703 static void virtnet_get_channels(struct net_device *dev,
1704 				 struct ethtool_channels *channels)
1705 {
1706 	struct virtnet_info *vi = netdev_priv(dev);
1707 
1708 	channels->combined_count = vi->curr_queue_pairs;
1709 	channels->max_combined = vi->max_queue_pairs;
1710 	channels->max_other = 0;
1711 	channels->rx_count = 0;
1712 	channels->tx_count = 0;
1713 	channels->other_count = 0;
1714 }
1715 
1716 /* Check if the user is trying to change anything besides speed/duplex */
1717 static bool
1718 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
1719 {
1720 	struct ethtool_link_ksettings diff1 = *cmd;
1721 	struct ethtool_link_ksettings diff2 = {};
1722 
1723 	/* cmd is always set so we need to clear it, validate the port type
1724 	 * and also without autonegotiation we can ignore advertising
1725 	 */
1726 	diff1.base.speed = 0;
1727 	diff2.base.port = PORT_OTHER;
1728 	ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
1729 	diff1.base.duplex = 0;
1730 	diff1.base.cmd = 0;
1731 	diff1.base.link_mode_masks_nwords = 0;
1732 
1733 	return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
1734 		bitmap_empty(diff1.link_modes.supported,
1735 			     __ETHTOOL_LINK_MODE_MASK_NBITS) &&
1736 		bitmap_empty(diff1.link_modes.advertising,
1737 			     __ETHTOOL_LINK_MODE_MASK_NBITS) &&
1738 		bitmap_empty(diff1.link_modes.lp_advertising,
1739 			     __ETHTOOL_LINK_MODE_MASK_NBITS);
1740 }
1741 
1742 static int virtnet_set_link_ksettings(struct net_device *dev,
1743 				      const struct ethtool_link_ksettings *cmd)
1744 {
1745 	struct virtnet_info *vi = netdev_priv(dev);
1746 	u32 speed;
1747 
1748 	speed = cmd->base.speed;
1749 	/* don't allow custom speed and duplex */
1750 	if (!ethtool_validate_speed(speed) ||
1751 	    !ethtool_validate_duplex(cmd->base.duplex) ||
1752 	    !virtnet_validate_ethtool_cmd(cmd))
1753 		return -EINVAL;
1754 	vi->speed = speed;
1755 	vi->duplex = cmd->base.duplex;
1756 
1757 	return 0;
1758 }
1759 
1760 static int virtnet_get_link_ksettings(struct net_device *dev,
1761 				      struct ethtool_link_ksettings *cmd)
1762 {
1763 	struct virtnet_info *vi = netdev_priv(dev);
1764 
1765 	cmd->base.speed = vi->speed;
1766 	cmd->base.duplex = vi->duplex;
1767 	cmd->base.port = PORT_OTHER;
1768 
1769 	return 0;
1770 }
1771 
1772 static void virtnet_init_settings(struct net_device *dev)
1773 {
1774 	struct virtnet_info *vi = netdev_priv(dev);
1775 
1776 	vi->speed = SPEED_UNKNOWN;
1777 	vi->duplex = DUPLEX_UNKNOWN;
1778 }
1779 
1780 static const struct ethtool_ops virtnet_ethtool_ops = {
1781 	.get_drvinfo = virtnet_get_drvinfo,
1782 	.get_link = ethtool_op_get_link,
1783 	.get_ringparam = virtnet_get_ringparam,
1784 	.set_channels = virtnet_set_channels,
1785 	.get_channels = virtnet_get_channels,
1786 	.get_ts_info = ethtool_op_get_ts_info,
1787 	.get_link_ksettings = virtnet_get_link_ksettings,
1788 	.set_link_ksettings = virtnet_set_link_ksettings,
1789 };
1790 
1791 static void virtnet_freeze_down(struct virtio_device *vdev)
1792 {
1793 	struct virtnet_info *vi = vdev->priv;
1794 	int i;
1795 
1796 	/* Make sure no work handler is accessing the device */
1797 	flush_work(&vi->config_work);
1798 
1799 	netif_device_detach(vi->dev);
1800 	cancel_delayed_work_sync(&vi->refill);
1801 
1802 	if (netif_running(vi->dev)) {
1803 		for (i = 0; i < vi->max_queue_pairs; i++) {
1804 			napi_disable(&vi->rq[i].napi);
1805 			virtnet_napi_tx_disable(&vi->sq[i].napi);
1806 		}
1807 	}
1808 }
1809 
1810 static int init_vqs(struct virtnet_info *vi);
1811 static void _remove_vq_common(struct virtnet_info *vi);
1812 
1813 static int virtnet_restore_up(struct virtio_device *vdev)
1814 {
1815 	struct virtnet_info *vi = vdev->priv;
1816 	int err, i;
1817 
1818 	err = init_vqs(vi);
1819 	if (err)
1820 		return err;
1821 
1822 	virtio_device_ready(vdev);
1823 
1824 	if (netif_running(vi->dev)) {
1825 		for (i = 0; i < vi->curr_queue_pairs; i++)
1826 			if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1827 				schedule_delayed_work(&vi->refill, 0);
1828 
1829 		for (i = 0; i < vi->max_queue_pairs; i++) {
1830 			virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1831 			virtnet_napi_tx_enable(vi, vi->sq[i].vq,
1832 					       &vi->sq[i].napi);
1833 		}
1834 	}
1835 
1836 	netif_device_attach(vi->dev);
1837 	return err;
1838 }
1839 
1840 static int virtnet_reset(struct virtnet_info *vi, int curr_qp, int xdp_qp)
1841 {
1842 	struct virtio_device *dev = vi->vdev;
1843 	int ret;
1844 
1845 	virtio_config_disable(dev);
1846 	dev->failed = dev->config->get_status(dev) & VIRTIO_CONFIG_S_FAILED;
1847 	virtnet_freeze_down(dev);
1848 	_remove_vq_common(vi);
1849 
1850 	virtio_add_status(dev, VIRTIO_CONFIG_S_ACKNOWLEDGE);
1851 	virtio_add_status(dev, VIRTIO_CONFIG_S_DRIVER);
1852 
1853 	ret = virtio_finalize_features(dev);
1854 	if (ret)
1855 		goto err;
1856 
1857 	vi->xdp_queue_pairs = xdp_qp;
1858 	ret = virtnet_restore_up(dev);
1859 	if (ret)
1860 		goto err;
1861 	ret = _virtnet_set_queues(vi, curr_qp);
1862 	if (ret)
1863 		goto err;
1864 
1865 	virtio_add_status(dev, VIRTIO_CONFIG_S_DRIVER_OK);
1866 	virtio_config_enable(dev);
1867 	return 0;
1868 err:
1869 	virtio_add_status(dev, VIRTIO_CONFIG_S_FAILED);
1870 	return ret;
1871 }
1872 
1873 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1874 			   struct netlink_ext_ack *extack)
1875 {
1876 	unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
1877 	struct virtnet_info *vi = netdev_priv(dev);
1878 	struct bpf_prog *old_prog;
1879 	u16 xdp_qp = 0, curr_qp;
1880 	int i, err;
1881 
1882 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1883 	    virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1884 	    virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
1885 	    virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO)) {
1886 		NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first");
1887 		return -EOPNOTSUPP;
1888 	}
1889 
1890 	if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
1891 		NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
1892 		return -EINVAL;
1893 	}
1894 
1895 	if (dev->mtu > max_sz) {
1896 		NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
1897 		netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
1898 		return -EINVAL;
1899 	}
1900 
1901 	curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
1902 	if (prog)
1903 		xdp_qp = nr_cpu_ids;
1904 
1905 	/* XDP requires extra queues for XDP_TX */
1906 	if (curr_qp + xdp_qp > vi->max_queue_pairs) {
1907 		NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
1908 		netdev_warn(dev, "request %i queues but max is %i\n",
1909 			    curr_qp + xdp_qp, vi->max_queue_pairs);
1910 		return -ENOMEM;
1911 	}
1912 
1913 	if (prog) {
1914 		prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
1915 		if (IS_ERR(prog))
1916 			return PTR_ERR(prog);
1917 	}
1918 
1919 	/* Changing the headroom in buffers is a disruptive operation because
1920 	 * existing buffers must be flushed and reallocated. This will happen
1921 	 * when a xdp program is initially added or xdp is disabled by removing
1922 	 * the xdp program resulting in number of XDP queues changing.
1923 	 */
1924 	if (vi->xdp_queue_pairs != xdp_qp) {
1925 		err = virtnet_reset(vi, curr_qp + xdp_qp, xdp_qp);
1926 		if (err) {
1927 			dev_warn(&dev->dev, "XDP reset failure.\n");
1928 			goto virtio_reset_err;
1929 		}
1930 	}
1931 
1932 	netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
1933 
1934 	for (i = 0; i < vi->max_queue_pairs; i++) {
1935 		old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1936 		rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
1937 		if (old_prog)
1938 			bpf_prog_put(old_prog);
1939 	}
1940 
1941 	return 0;
1942 
1943 virtio_reset_err:
1944 	/* On reset error do our best to unwind XDP changes inflight and return
1945 	 * error up to user space for resolution. The underlying reset hung on
1946 	 * us so not much we can do here.
1947 	 */
1948 	if (prog)
1949 		bpf_prog_sub(prog, vi->max_queue_pairs - 1);
1950 	return err;
1951 }
1952 
1953 static bool virtnet_xdp_query(struct net_device *dev)
1954 {
1955 	struct virtnet_info *vi = netdev_priv(dev);
1956 	int i;
1957 
1958 	for (i = 0; i < vi->max_queue_pairs; i++) {
1959 		if (vi->rq[i].xdp_prog)
1960 			return true;
1961 	}
1962 	return false;
1963 }
1964 
1965 static int virtnet_xdp(struct net_device *dev, struct netdev_xdp *xdp)
1966 {
1967 	switch (xdp->command) {
1968 	case XDP_SETUP_PROG:
1969 		return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
1970 	case XDP_QUERY_PROG:
1971 		xdp->prog_attached = virtnet_xdp_query(dev);
1972 		return 0;
1973 	default:
1974 		return -EINVAL;
1975 	}
1976 }
1977 
1978 static const struct net_device_ops virtnet_netdev = {
1979 	.ndo_open            = virtnet_open,
1980 	.ndo_stop   	     = virtnet_close,
1981 	.ndo_start_xmit      = start_xmit,
1982 	.ndo_validate_addr   = eth_validate_addr,
1983 	.ndo_set_mac_address = virtnet_set_mac_address,
1984 	.ndo_set_rx_mode     = virtnet_set_rx_mode,
1985 	.ndo_get_stats64     = virtnet_stats,
1986 	.ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1987 	.ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1988 #ifdef CONFIG_NET_POLL_CONTROLLER
1989 	.ndo_poll_controller = virtnet_netpoll,
1990 #endif
1991 	.ndo_xdp		= virtnet_xdp,
1992 };
1993 
1994 static void virtnet_config_changed_work(struct work_struct *work)
1995 {
1996 	struct virtnet_info *vi =
1997 		container_of(work, struct virtnet_info, config_work);
1998 	u16 v;
1999 
2000 	if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2001 				 struct virtio_net_config, status, &v) < 0)
2002 		return;
2003 
2004 	if (v & VIRTIO_NET_S_ANNOUNCE) {
2005 		netdev_notify_peers(vi->dev);
2006 		virtnet_ack_link_announce(vi);
2007 	}
2008 
2009 	/* Ignore unknown (future) status bits */
2010 	v &= VIRTIO_NET_S_LINK_UP;
2011 
2012 	if (vi->status == v)
2013 		return;
2014 
2015 	vi->status = v;
2016 
2017 	if (vi->status & VIRTIO_NET_S_LINK_UP) {
2018 		netif_carrier_on(vi->dev);
2019 		netif_tx_wake_all_queues(vi->dev);
2020 	} else {
2021 		netif_carrier_off(vi->dev);
2022 		netif_tx_stop_all_queues(vi->dev);
2023 	}
2024 }
2025 
2026 static void virtnet_config_changed(struct virtio_device *vdev)
2027 {
2028 	struct virtnet_info *vi = vdev->priv;
2029 
2030 	schedule_work(&vi->config_work);
2031 }
2032 
2033 static void virtnet_free_queues(struct virtnet_info *vi)
2034 {
2035 	int i;
2036 
2037 	for (i = 0; i < vi->max_queue_pairs; i++) {
2038 		napi_hash_del(&vi->rq[i].napi);
2039 		netif_napi_del(&vi->rq[i].napi);
2040 		netif_napi_del(&vi->sq[i].napi);
2041 	}
2042 
2043 	/* We called napi_hash_del() before netif_napi_del(),
2044 	 * we need to respect an RCU grace period before freeing vi->rq
2045 	 */
2046 	synchronize_net();
2047 
2048 	kfree(vi->rq);
2049 	kfree(vi->sq);
2050 }
2051 
2052 static void _free_receive_bufs(struct virtnet_info *vi)
2053 {
2054 	struct bpf_prog *old_prog;
2055 	int i;
2056 
2057 	for (i = 0; i < vi->max_queue_pairs; i++) {
2058 		while (vi->rq[i].pages)
2059 			__free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2060 
2061 		old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2062 		RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2063 		if (old_prog)
2064 			bpf_prog_put(old_prog);
2065 	}
2066 }
2067 
2068 static void free_receive_bufs(struct virtnet_info *vi)
2069 {
2070 	rtnl_lock();
2071 	_free_receive_bufs(vi);
2072 	rtnl_unlock();
2073 }
2074 
2075 static void free_receive_page_frags(struct virtnet_info *vi)
2076 {
2077 	int i;
2078 	for (i = 0; i < vi->max_queue_pairs; i++)
2079 		if (vi->rq[i].alloc_frag.page)
2080 			put_page(vi->rq[i].alloc_frag.page);
2081 }
2082 
2083 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
2084 {
2085 	if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
2086 		return false;
2087 	else if (q < vi->curr_queue_pairs)
2088 		return true;
2089 	else
2090 		return false;
2091 }
2092 
2093 static void free_unused_bufs(struct virtnet_info *vi)
2094 {
2095 	void *buf;
2096 	int i;
2097 
2098 	for (i = 0; i < vi->max_queue_pairs; i++) {
2099 		struct virtqueue *vq = vi->sq[i].vq;
2100 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2101 			if (!is_xdp_raw_buffer_queue(vi, i))
2102 				dev_kfree_skb(buf);
2103 			else
2104 				put_page(virt_to_head_page(buf));
2105 		}
2106 	}
2107 
2108 	for (i = 0; i < vi->max_queue_pairs; i++) {
2109 		struct virtqueue *vq = vi->rq[i].vq;
2110 
2111 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2112 			if (vi->mergeable_rx_bufs) {
2113 				put_page(virt_to_head_page(buf));
2114 			} else if (vi->big_packets) {
2115 				give_pages(&vi->rq[i], buf);
2116 			} else {
2117 				put_page(virt_to_head_page(buf));
2118 			}
2119 		}
2120 	}
2121 }
2122 
2123 static void virtnet_del_vqs(struct virtnet_info *vi)
2124 {
2125 	struct virtio_device *vdev = vi->vdev;
2126 
2127 	virtnet_clean_affinity(vi, -1);
2128 
2129 	vdev->config->del_vqs(vdev);
2130 
2131 	virtnet_free_queues(vi);
2132 }
2133 
2134 /* How large should a single buffer be so a queue full of these can fit at
2135  * least one full packet?
2136  * Logic below assumes the mergeable buffer header is used.
2137  */
2138 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2139 {
2140 	const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2141 	unsigned int rq_size = virtqueue_get_vring_size(vq);
2142 	unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2143 	unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2144 	unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2145 
2146 	return max(min_buf_len, hdr_len);
2147 }
2148 
2149 static int virtnet_find_vqs(struct virtnet_info *vi)
2150 {
2151 	vq_callback_t **callbacks;
2152 	struct virtqueue **vqs;
2153 	int ret = -ENOMEM;
2154 	int i, total_vqs;
2155 	const char **names;
2156 	bool *ctx;
2157 
2158 	/* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2159 	 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2160 	 * possible control vq.
2161 	 */
2162 	total_vqs = vi->max_queue_pairs * 2 +
2163 		    virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2164 
2165 	/* Allocate space for find_vqs parameters */
2166 	vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
2167 	if (!vqs)
2168 		goto err_vq;
2169 	callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
2170 	if (!callbacks)
2171 		goto err_callback;
2172 	names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
2173 	if (!names)
2174 		goto err_names;
2175 	if (vi->mergeable_rx_bufs) {
2176 		ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL);
2177 		if (!ctx)
2178 			goto err_ctx;
2179 	} else {
2180 		ctx = NULL;
2181 	}
2182 
2183 	/* Parameters for control virtqueue, if any */
2184 	if (vi->has_cvq) {
2185 		callbacks[total_vqs - 1] = NULL;
2186 		names[total_vqs - 1] = "control";
2187 	}
2188 
2189 	/* Allocate/initialize parameters for send/receive virtqueues */
2190 	for (i = 0; i < vi->max_queue_pairs; i++) {
2191 		callbacks[rxq2vq(i)] = skb_recv_done;
2192 		callbacks[txq2vq(i)] = skb_xmit_done;
2193 		sprintf(vi->rq[i].name, "input.%d", i);
2194 		sprintf(vi->sq[i].name, "output.%d", i);
2195 		names[rxq2vq(i)] = vi->rq[i].name;
2196 		names[txq2vq(i)] = vi->sq[i].name;
2197 		if (ctx)
2198 			ctx[rxq2vq(i)] = true;
2199 	}
2200 
2201 	ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2202 					 names, ctx, NULL);
2203 	if (ret)
2204 		goto err_find;
2205 
2206 	if (vi->has_cvq) {
2207 		vi->cvq = vqs[total_vqs - 1];
2208 		if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2209 			vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2210 	}
2211 
2212 	for (i = 0; i < vi->max_queue_pairs; i++) {
2213 		vi->rq[i].vq = vqs[rxq2vq(i)];
2214 		vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2215 		vi->sq[i].vq = vqs[txq2vq(i)];
2216 	}
2217 
2218 	kfree(names);
2219 	kfree(callbacks);
2220 	kfree(vqs);
2221 
2222 	return 0;
2223 
2224 err_find:
2225 	kfree(ctx);
2226 err_ctx:
2227 	kfree(names);
2228 err_names:
2229 	kfree(callbacks);
2230 err_callback:
2231 	kfree(vqs);
2232 err_vq:
2233 	return ret;
2234 }
2235 
2236 static int virtnet_alloc_queues(struct virtnet_info *vi)
2237 {
2238 	int i;
2239 
2240 	vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2241 	if (!vi->sq)
2242 		goto err_sq;
2243 	vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2244 	if (!vi->rq)
2245 		goto err_rq;
2246 
2247 	INIT_DELAYED_WORK(&vi->refill, refill_work);
2248 	for (i = 0; i < vi->max_queue_pairs; i++) {
2249 		vi->rq[i].pages = NULL;
2250 		netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2251 			       napi_weight);
2252 		netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2253 				  napi_tx ? napi_weight : 0);
2254 
2255 		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2256 		ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2257 		sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2258 	}
2259 
2260 	return 0;
2261 
2262 err_rq:
2263 	kfree(vi->sq);
2264 err_sq:
2265 	return -ENOMEM;
2266 }
2267 
2268 static int init_vqs(struct virtnet_info *vi)
2269 {
2270 	int ret;
2271 
2272 	/* Allocate send & receive queues */
2273 	ret = virtnet_alloc_queues(vi);
2274 	if (ret)
2275 		goto err;
2276 
2277 	ret = virtnet_find_vqs(vi);
2278 	if (ret)
2279 		goto err_free;
2280 
2281 	get_online_cpus();
2282 	virtnet_set_affinity(vi);
2283 	put_online_cpus();
2284 
2285 	return 0;
2286 
2287 err_free:
2288 	virtnet_free_queues(vi);
2289 err:
2290 	return ret;
2291 }
2292 
2293 #ifdef CONFIG_SYSFS
2294 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2295 		struct rx_queue_attribute *attribute, char *buf)
2296 {
2297 	struct virtnet_info *vi = netdev_priv(queue->dev);
2298 	unsigned int queue_index = get_netdev_rx_queue_index(queue);
2299 	struct ewma_pkt_len *avg;
2300 
2301 	BUG_ON(queue_index >= vi->max_queue_pairs);
2302 	avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2303 	return sprintf(buf, "%u\n",
2304 		       get_mergeable_buf_len(&vi->rq[queue_index], avg));
2305 }
2306 
2307 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2308 	__ATTR_RO(mergeable_rx_buffer_size);
2309 
2310 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2311 	&mergeable_rx_buffer_size_attribute.attr,
2312 	NULL
2313 };
2314 
2315 static const struct attribute_group virtio_net_mrg_rx_group = {
2316 	.name = "virtio_net",
2317 	.attrs = virtio_net_mrg_rx_attrs
2318 };
2319 #endif
2320 
2321 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2322 				    unsigned int fbit,
2323 				    const char *fname, const char *dname)
2324 {
2325 	if (!virtio_has_feature(vdev, fbit))
2326 		return false;
2327 
2328 	dev_err(&vdev->dev, "device advertises feature %s but not %s",
2329 		fname, dname);
2330 
2331 	return true;
2332 }
2333 
2334 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)			\
2335 	virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2336 
2337 static bool virtnet_validate_features(struct virtio_device *vdev)
2338 {
2339 	if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2340 	    (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2341 			     "VIRTIO_NET_F_CTRL_VQ") ||
2342 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2343 			     "VIRTIO_NET_F_CTRL_VQ") ||
2344 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2345 			     "VIRTIO_NET_F_CTRL_VQ") ||
2346 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2347 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2348 			     "VIRTIO_NET_F_CTRL_VQ"))) {
2349 		return false;
2350 	}
2351 
2352 	return true;
2353 }
2354 
2355 #define MIN_MTU ETH_MIN_MTU
2356 #define MAX_MTU ETH_MAX_MTU
2357 
2358 static int virtnet_validate(struct virtio_device *vdev)
2359 {
2360 	if (!vdev->config->get) {
2361 		dev_err(&vdev->dev, "%s failure: config access disabled\n",
2362 			__func__);
2363 		return -EINVAL;
2364 	}
2365 
2366 	if (!virtnet_validate_features(vdev))
2367 		return -EINVAL;
2368 
2369 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2370 		int mtu = virtio_cread16(vdev,
2371 					 offsetof(struct virtio_net_config,
2372 						  mtu));
2373 		if (mtu < MIN_MTU)
2374 			__virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2375 	}
2376 
2377 	return 0;
2378 }
2379 
2380 static int virtnet_probe(struct virtio_device *vdev)
2381 {
2382 	int i, err;
2383 	struct net_device *dev;
2384 	struct virtnet_info *vi;
2385 	u16 max_queue_pairs;
2386 	int mtu;
2387 
2388 	/* Find if host supports multiqueue virtio_net device */
2389 	err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2390 				   struct virtio_net_config,
2391 				   max_virtqueue_pairs, &max_queue_pairs);
2392 
2393 	/* We need at least 2 queue's */
2394 	if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2395 	    max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2396 	    !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2397 		max_queue_pairs = 1;
2398 
2399 	/* Allocate ourselves a network device with room for our info */
2400 	dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2401 	if (!dev)
2402 		return -ENOMEM;
2403 
2404 	/* Set up network device as normal. */
2405 	dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2406 	dev->netdev_ops = &virtnet_netdev;
2407 	dev->features = NETIF_F_HIGHDMA;
2408 
2409 	dev->ethtool_ops = &virtnet_ethtool_ops;
2410 	SET_NETDEV_DEV(dev, &vdev->dev);
2411 
2412 	/* Do we support "hardware" checksums? */
2413 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2414 		/* This opens up the world of extra features. */
2415 		dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2416 		if (csum)
2417 			dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2418 
2419 		if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2420 			dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
2421 				| NETIF_F_TSO_ECN | NETIF_F_TSO6;
2422 		}
2423 		/* Individual feature bits: what can host handle? */
2424 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2425 			dev->hw_features |= NETIF_F_TSO;
2426 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2427 			dev->hw_features |= NETIF_F_TSO6;
2428 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2429 			dev->hw_features |= NETIF_F_TSO_ECN;
2430 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
2431 			dev->hw_features |= NETIF_F_UFO;
2432 
2433 		dev->features |= NETIF_F_GSO_ROBUST;
2434 
2435 		if (gso)
2436 			dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
2437 		/* (!csum && gso) case will be fixed by register_netdev() */
2438 	}
2439 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2440 		dev->features |= NETIF_F_RXCSUM;
2441 
2442 	dev->vlan_features = dev->features;
2443 
2444 	/* MTU range: 68 - 65535 */
2445 	dev->min_mtu = MIN_MTU;
2446 	dev->max_mtu = MAX_MTU;
2447 
2448 	/* Configuration may specify what MAC to use.  Otherwise random. */
2449 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2450 		virtio_cread_bytes(vdev,
2451 				   offsetof(struct virtio_net_config, mac),
2452 				   dev->dev_addr, dev->addr_len);
2453 	else
2454 		eth_hw_addr_random(dev);
2455 
2456 	/* Set up our device-specific information */
2457 	vi = netdev_priv(dev);
2458 	vi->dev = dev;
2459 	vi->vdev = vdev;
2460 	vdev->priv = vi;
2461 	vi->stats = alloc_percpu(struct virtnet_stats);
2462 	err = -ENOMEM;
2463 	if (vi->stats == NULL)
2464 		goto free;
2465 
2466 	for_each_possible_cpu(i) {
2467 		struct virtnet_stats *virtnet_stats;
2468 		virtnet_stats = per_cpu_ptr(vi->stats, i);
2469 		u64_stats_init(&virtnet_stats->tx_syncp);
2470 		u64_stats_init(&virtnet_stats->rx_syncp);
2471 	}
2472 
2473 	INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2474 
2475 	/* If we can receive ANY GSO packets, we must allocate large ones. */
2476 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2477 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2478 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2479 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2480 		vi->big_packets = true;
2481 
2482 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2483 		vi->mergeable_rx_bufs = true;
2484 
2485 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2486 	    virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2487 		vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2488 	else
2489 		vi->hdr_len = sizeof(struct virtio_net_hdr);
2490 
2491 	if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2492 	    virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2493 		vi->any_header_sg = true;
2494 
2495 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2496 		vi->has_cvq = true;
2497 
2498 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2499 		mtu = virtio_cread16(vdev,
2500 				     offsetof(struct virtio_net_config,
2501 					      mtu));
2502 		if (mtu < dev->min_mtu) {
2503 			/* Should never trigger: MTU was previously validated
2504 			 * in virtnet_validate.
2505 			 */
2506 			dev_err(&vdev->dev, "device MTU appears to have changed "
2507 				"it is now %d < %d", mtu, dev->min_mtu);
2508 			goto free_stats;
2509 		}
2510 
2511 		dev->mtu = mtu;
2512 		dev->max_mtu = mtu;
2513 
2514 		/* TODO: size buffers correctly in this case. */
2515 		if (dev->mtu > ETH_DATA_LEN)
2516 			vi->big_packets = true;
2517 	}
2518 
2519 	if (vi->any_header_sg)
2520 		dev->needed_headroom = vi->hdr_len;
2521 
2522 	/* Enable multiqueue by default */
2523 	if (num_online_cpus() >= max_queue_pairs)
2524 		vi->curr_queue_pairs = max_queue_pairs;
2525 	else
2526 		vi->curr_queue_pairs = num_online_cpus();
2527 	vi->max_queue_pairs = max_queue_pairs;
2528 
2529 	/* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2530 	err = init_vqs(vi);
2531 	if (err)
2532 		goto free_stats;
2533 
2534 #ifdef CONFIG_SYSFS
2535 	if (vi->mergeable_rx_bufs)
2536 		dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2537 #endif
2538 	netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2539 	netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2540 
2541 	virtnet_init_settings(dev);
2542 
2543 	err = register_netdev(dev);
2544 	if (err) {
2545 		pr_debug("virtio_net: registering device failed\n");
2546 		goto free_vqs;
2547 	}
2548 
2549 	virtio_device_ready(vdev);
2550 
2551 	err = virtnet_cpu_notif_add(vi);
2552 	if (err) {
2553 		pr_debug("virtio_net: registering cpu notifier failed\n");
2554 		goto free_unregister_netdev;
2555 	}
2556 
2557 	virtnet_set_queues(vi, vi->curr_queue_pairs);
2558 
2559 	/* Assume link up if device can't report link status,
2560 	   otherwise get link status from config. */
2561 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2562 		netif_carrier_off(dev);
2563 		schedule_work(&vi->config_work);
2564 	} else {
2565 		vi->status = VIRTIO_NET_S_LINK_UP;
2566 		netif_carrier_on(dev);
2567 	}
2568 
2569 	pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2570 		 dev->name, max_queue_pairs);
2571 
2572 	return 0;
2573 
2574 free_unregister_netdev:
2575 	vi->vdev->config->reset(vdev);
2576 
2577 	unregister_netdev(dev);
2578 free_vqs:
2579 	cancel_delayed_work_sync(&vi->refill);
2580 	free_receive_page_frags(vi);
2581 	virtnet_del_vqs(vi);
2582 free_stats:
2583 	free_percpu(vi->stats);
2584 free:
2585 	free_netdev(dev);
2586 	return err;
2587 }
2588 
2589 static void _remove_vq_common(struct virtnet_info *vi)
2590 {
2591 	vi->vdev->config->reset(vi->vdev);
2592 	free_unused_bufs(vi);
2593 	_free_receive_bufs(vi);
2594 	free_receive_page_frags(vi);
2595 	virtnet_del_vqs(vi);
2596 }
2597 
2598 static void remove_vq_common(struct virtnet_info *vi)
2599 {
2600 	vi->vdev->config->reset(vi->vdev);
2601 
2602 	/* Free unused buffers in both send and recv, if any. */
2603 	free_unused_bufs(vi);
2604 
2605 	free_receive_bufs(vi);
2606 
2607 	free_receive_page_frags(vi);
2608 
2609 	virtnet_del_vqs(vi);
2610 }
2611 
2612 static void virtnet_remove(struct virtio_device *vdev)
2613 {
2614 	struct virtnet_info *vi = vdev->priv;
2615 
2616 	virtnet_cpu_notif_remove(vi);
2617 
2618 	/* Make sure no work handler is accessing the device. */
2619 	flush_work(&vi->config_work);
2620 
2621 	unregister_netdev(vi->dev);
2622 
2623 	remove_vq_common(vi);
2624 
2625 	free_percpu(vi->stats);
2626 	free_netdev(vi->dev);
2627 }
2628 
2629 #ifdef CONFIG_PM_SLEEP
2630 static int virtnet_freeze(struct virtio_device *vdev)
2631 {
2632 	struct virtnet_info *vi = vdev->priv;
2633 
2634 	virtnet_cpu_notif_remove(vi);
2635 	virtnet_freeze_down(vdev);
2636 	remove_vq_common(vi);
2637 
2638 	return 0;
2639 }
2640 
2641 static int virtnet_restore(struct virtio_device *vdev)
2642 {
2643 	struct virtnet_info *vi = vdev->priv;
2644 	int err;
2645 
2646 	err = virtnet_restore_up(vdev);
2647 	if (err)
2648 		return err;
2649 	virtnet_set_queues(vi, vi->curr_queue_pairs);
2650 
2651 	err = virtnet_cpu_notif_add(vi);
2652 	if (err)
2653 		return err;
2654 
2655 	return 0;
2656 }
2657 #endif
2658 
2659 static struct virtio_device_id id_table[] = {
2660 	{ VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2661 	{ 0 },
2662 };
2663 
2664 #define VIRTNET_FEATURES \
2665 	VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2666 	VIRTIO_NET_F_MAC, \
2667 	VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2668 	VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2669 	VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2670 	VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2671 	VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2672 	VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2673 	VIRTIO_NET_F_CTRL_MAC_ADDR, \
2674 	VIRTIO_NET_F_MTU
2675 
2676 static unsigned int features[] = {
2677 	VIRTNET_FEATURES,
2678 };
2679 
2680 static unsigned int features_legacy[] = {
2681 	VIRTNET_FEATURES,
2682 	VIRTIO_NET_F_GSO,
2683 	VIRTIO_F_ANY_LAYOUT,
2684 };
2685 
2686 static struct virtio_driver virtio_net_driver = {
2687 	.feature_table = features,
2688 	.feature_table_size = ARRAY_SIZE(features),
2689 	.feature_table_legacy = features_legacy,
2690 	.feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2691 	.driver.name =	KBUILD_MODNAME,
2692 	.driver.owner =	THIS_MODULE,
2693 	.id_table =	id_table,
2694 	.validate =	virtnet_validate,
2695 	.probe =	virtnet_probe,
2696 	.remove =	virtnet_remove,
2697 	.config_changed = virtnet_config_changed,
2698 #ifdef CONFIG_PM_SLEEP
2699 	.freeze =	virtnet_freeze,
2700 	.restore =	virtnet_restore,
2701 #endif
2702 };
2703 
2704 static __init int virtio_net_driver_init(void)
2705 {
2706 	int ret;
2707 
2708 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
2709 				      virtnet_cpu_online,
2710 				      virtnet_cpu_down_prep);
2711 	if (ret < 0)
2712 		goto out;
2713 	virtionet_online = ret;
2714 	ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
2715 				      NULL, virtnet_cpu_dead);
2716 	if (ret)
2717 		goto err_dead;
2718 
2719         ret = register_virtio_driver(&virtio_net_driver);
2720 	if (ret)
2721 		goto err_virtio;
2722 	return 0;
2723 err_virtio:
2724 	cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2725 err_dead:
2726 	cpuhp_remove_multi_state(virtionet_online);
2727 out:
2728 	return ret;
2729 }
2730 module_init(virtio_net_driver_init);
2731 
2732 static __exit void virtio_net_driver_exit(void)
2733 {
2734 	cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2735 	cpuhp_remove_multi_state(virtionet_online);
2736 	unregister_virtio_driver(&virtio_net_driver);
2737 }
2738 module_exit(virtio_net_driver_exit);
2739 
2740 MODULE_DEVICE_TABLE(virtio, id_table);
2741 MODULE_DESCRIPTION("Virtio network driver");
2742 MODULE_LICENSE("GPL");
2743