xref: /linux/drivers/net/virtio_net.c (revision 6ebe6dbd6886af07b102aca42e44edbee94a22d9)
1 /* A network driver using virtio.
2  *
3  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License as published by
7  * the Free Software Foundation; either version 2 of the License, or
8  * (at your option) any later version.
9  *
10  * This program is distributed in the hope that it will be useful,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13  * GNU General Public License for more details.
14  *
15  * You should have received a copy of the GNU General Public License
16  * along with this program; if not, see <http://www.gnu.org/licenses/>.
17  */
18 //#define DEBUG
19 #include <linux/netdevice.h>
20 #include <linux/etherdevice.h>
21 #include <linux/ethtool.h>
22 #include <linux/module.h>
23 #include <linux/virtio.h>
24 #include <linux/virtio_net.h>
25 #include <linux/bpf.h>
26 #include <linux/bpf_trace.h>
27 #include <linux/scatterlist.h>
28 #include <linux/if_vlan.h>
29 #include <linux/slab.h>
30 #include <linux/cpu.h>
31 #include <linux/average.h>
32 #include <linux/filter.h>
33 #include <net/route.h>
34 
35 static int napi_weight = NAPI_POLL_WEIGHT;
36 module_param(napi_weight, int, 0444);
37 
38 static bool csum = true, gso = true, napi_tx;
39 module_param(csum, bool, 0444);
40 module_param(gso, bool, 0444);
41 module_param(napi_tx, bool, 0644);
42 
43 /* FIXME: MTU in config. */
44 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
45 #define GOOD_COPY_LEN	128
46 
47 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
48 
49 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
50 #define VIRTIO_XDP_HEADROOM 256
51 
52 /* RX packet size EWMA. The average packet size is used to determine the packet
53  * buffer size when refilling RX rings. As the entire RX ring may be refilled
54  * at once, the weight is chosen so that the EWMA will be insensitive to short-
55  * term, transient changes in packet size.
56  */
57 DECLARE_EWMA(pkt_len, 0, 64)
58 
59 #define VIRTNET_DRIVER_VERSION "1.0.0"
60 
61 static const unsigned long guest_offloads[] = {
62 	VIRTIO_NET_F_GUEST_TSO4,
63 	VIRTIO_NET_F_GUEST_TSO6,
64 	VIRTIO_NET_F_GUEST_ECN,
65 	VIRTIO_NET_F_GUEST_UFO
66 };
67 
68 struct virtnet_stats {
69 	struct u64_stats_sync tx_syncp;
70 	struct u64_stats_sync rx_syncp;
71 	u64 tx_bytes;
72 	u64 tx_packets;
73 
74 	u64 rx_bytes;
75 	u64 rx_packets;
76 };
77 
78 /* Internal representation of a send virtqueue */
79 struct send_queue {
80 	/* Virtqueue associated with this send _queue */
81 	struct virtqueue *vq;
82 
83 	/* TX: fragments + linear part + virtio header */
84 	struct scatterlist sg[MAX_SKB_FRAGS + 2];
85 
86 	/* Name of the send queue: output.$index */
87 	char name[40];
88 
89 	struct napi_struct napi;
90 };
91 
92 /* Internal representation of a receive virtqueue */
93 struct receive_queue {
94 	/* Virtqueue associated with this receive_queue */
95 	struct virtqueue *vq;
96 
97 	struct napi_struct napi;
98 
99 	struct bpf_prog __rcu *xdp_prog;
100 
101 	/* Chain pages by the private ptr. */
102 	struct page *pages;
103 
104 	/* Average packet length for mergeable receive buffers. */
105 	struct ewma_pkt_len mrg_avg_pkt_len;
106 
107 	/* Page frag for packet buffer allocation. */
108 	struct page_frag alloc_frag;
109 
110 	/* RX: fragments + linear part + virtio header */
111 	struct scatterlist sg[MAX_SKB_FRAGS + 2];
112 
113 	/* Min single buffer size for mergeable buffers case. */
114 	unsigned int min_buf_len;
115 
116 	/* Name of this receive queue: input.$index */
117 	char name[40];
118 };
119 
120 struct virtnet_info {
121 	struct virtio_device *vdev;
122 	struct virtqueue *cvq;
123 	struct net_device *dev;
124 	struct send_queue *sq;
125 	struct receive_queue *rq;
126 	unsigned int status;
127 
128 	/* Max # of queue pairs supported by the device */
129 	u16 max_queue_pairs;
130 
131 	/* # of queue pairs currently used by the driver */
132 	u16 curr_queue_pairs;
133 
134 	/* # of XDP queue pairs currently used by the driver */
135 	u16 xdp_queue_pairs;
136 
137 	/* I like... big packets and I cannot lie! */
138 	bool big_packets;
139 
140 	/* Host will merge rx buffers for big packets (shake it! shake it!) */
141 	bool mergeable_rx_bufs;
142 
143 	/* Has control virtqueue */
144 	bool has_cvq;
145 
146 	/* Host can handle any s/g split between our header and packet data */
147 	bool any_header_sg;
148 
149 	/* Packet virtio header size */
150 	u8 hdr_len;
151 
152 	/* Active statistics */
153 	struct virtnet_stats __percpu *stats;
154 
155 	/* Work struct for refilling if we run low on memory. */
156 	struct delayed_work refill;
157 
158 	/* Work struct for config space updates */
159 	struct work_struct config_work;
160 
161 	/* Does the affinity hint is set for virtqueues? */
162 	bool affinity_hint_set;
163 
164 	/* CPU hotplug instances for online & dead */
165 	struct hlist_node node;
166 	struct hlist_node node_dead;
167 
168 	/* Control VQ buffers: protected by the rtnl lock */
169 	struct virtio_net_ctrl_hdr ctrl_hdr;
170 	virtio_net_ctrl_ack ctrl_status;
171 	struct virtio_net_ctrl_mq ctrl_mq;
172 	u8 ctrl_promisc;
173 	u8 ctrl_allmulti;
174 	u16 ctrl_vid;
175 	u64 ctrl_offloads;
176 
177 	/* Ethtool settings */
178 	u8 duplex;
179 	u32 speed;
180 
181 	unsigned long guest_offloads;
182 };
183 
184 struct padded_vnet_hdr {
185 	struct virtio_net_hdr_mrg_rxbuf hdr;
186 	/*
187 	 * hdr is in a separate sg buffer, and data sg buffer shares same page
188 	 * with this header sg. This padding makes next sg 16 byte aligned
189 	 * after the header.
190 	 */
191 	char padding[4];
192 };
193 
194 /* Converting between virtqueue no. and kernel tx/rx queue no.
195  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
196  */
197 static int vq2txq(struct virtqueue *vq)
198 {
199 	return (vq->index - 1) / 2;
200 }
201 
202 static int txq2vq(int txq)
203 {
204 	return txq * 2 + 1;
205 }
206 
207 static int vq2rxq(struct virtqueue *vq)
208 {
209 	return vq->index / 2;
210 }
211 
212 static int rxq2vq(int rxq)
213 {
214 	return rxq * 2;
215 }
216 
217 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
218 {
219 	return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
220 }
221 
222 /*
223  * private is used to chain pages for big packets, put the whole
224  * most recent used list in the beginning for reuse
225  */
226 static void give_pages(struct receive_queue *rq, struct page *page)
227 {
228 	struct page *end;
229 
230 	/* Find end of list, sew whole thing into vi->rq.pages. */
231 	for (end = page; end->private; end = (struct page *)end->private);
232 	end->private = (unsigned long)rq->pages;
233 	rq->pages = page;
234 }
235 
236 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
237 {
238 	struct page *p = rq->pages;
239 
240 	if (p) {
241 		rq->pages = (struct page *)p->private;
242 		/* clear private here, it is used to chain pages */
243 		p->private = 0;
244 	} else
245 		p = alloc_page(gfp_mask);
246 	return p;
247 }
248 
249 static void virtqueue_napi_schedule(struct napi_struct *napi,
250 				    struct virtqueue *vq)
251 {
252 	if (napi_schedule_prep(napi)) {
253 		virtqueue_disable_cb(vq);
254 		__napi_schedule(napi);
255 	}
256 }
257 
258 static void virtqueue_napi_complete(struct napi_struct *napi,
259 				    struct virtqueue *vq, int processed)
260 {
261 	int opaque;
262 
263 	opaque = virtqueue_enable_cb_prepare(vq);
264 	if (napi_complete_done(napi, processed)) {
265 		if (unlikely(virtqueue_poll(vq, opaque)))
266 			virtqueue_napi_schedule(napi, vq);
267 	} else {
268 		virtqueue_disable_cb(vq);
269 	}
270 }
271 
272 static void skb_xmit_done(struct virtqueue *vq)
273 {
274 	struct virtnet_info *vi = vq->vdev->priv;
275 	struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
276 
277 	/* Suppress further interrupts. */
278 	virtqueue_disable_cb(vq);
279 
280 	if (napi->weight)
281 		virtqueue_napi_schedule(napi, vq);
282 	else
283 		/* We were probably waiting for more output buffers. */
284 		netif_wake_subqueue(vi->dev, vq2txq(vq));
285 }
286 
287 #define MRG_CTX_HEADER_SHIFT 22
288 static void *mergeable_len_to_ctx(unsigned int truesize,
289 				  unsigned int headroom)
290 {
291 	return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
292 }
293 
294 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
295 {
296 	return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
297 }
298 
299 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
300 {
301 	return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
302 }
303 
304 /* Called from bottom half context */
305 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
306 				   struct receive_queue *rq,
307 				   struct page *page, unsigned int offset,
308 				   unsigned int len, unsigned int truesize)
309 {
310 	struct sk_buff *skb;
311 	struct virtio_net_hdr_mrg_rxbuf *hdr;
312 	unsigned int copy, hdr_len, hdr_padded_len;
313 	char *p;
314 
315 	p = page_address(page) + offset;
316 
317 	/* copy small packet so we can reuse these pages for small data */
318 	skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
319 	if (unlikely(!skb))
320 		return NULL;
321 
322 	hdr = skb_vnet_hdr(skb);
323 
324 	hdr_len = vi->hdr_len;
325 	if (vi->mergeable_rx_bufs)
326 		hdr_padded_len = sizeof(*hdr);
327 	else
328 		hdr_padded_len = sizeof(struct padded_vnet_hdr);
329 
330 	memcpy(hdr, p, hdr_len);
331 
332 	len -= hdr_len;
333 	offset += hdr_padded_len;
334 	p += hdr_padded_len;
335 
336 	copy = len;
337 	if (copy > skb_tailroom(skb))
338 		copy = skb_tailroom(skb);
339 	skb_put_data(skb, p, copy);
340 
341 	len -= copy;
342 	offset += copy;
343 
344 	if (vi->mergeable_rx_bufs) {
345 		if (len)
346 			skb_add_rx_frag(skb, 0, page, offset, len, truesize);
347 		else
348 			put_page(page);
349 		return skb;
350 	}
351 
352 	/*
353 	 * Verify that we can indeed put this data into a skb.
354 	 * This is here to handle cases when the device erroneously
355 	 * tries to receive more than is possible. This is usually
356 	 * the case of a broken device.
357 	 */
358 	if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
359 		net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
360 		dev_kfree_skb(skb);
361 		return NULL;
362 	}
363 	BUG_ON(offset >= PAGE_SIZE);
364 	while (len) {
365 		unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
366 		skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
367 				frag_size, truesize);
368 		len -= frag_size;
369 		page = (struct page *)page->private;
370 		offset = 0;
371 	}
372 
373 	if (page)
374 		give_pages(rq, page);
375 
376 	return skb;
377 }
378 
379 static void virtnet_xdp_flush(struct net_device *dev)
380 {
381 	struct virtnet_info *vi = netdev_priv(dev);
382 	struct send_queue *sq;
383 	unsigned int qp;
384 
385 	qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
386 	sq = &vi->sq[qp];
387 
388 	virtqueue_kick(sq->vq);
389 }
390 
391 static bool __virtnet_xdp_xmit(struct virtnet_info *vi,
392 			       struct xdp_buff *xdp)
393 {
394 	struct virtio_net_hdr_mrg_rxbuf *hdr;
395 	unsigned int len;
396 	struct send_queue *sq;
397 	unsigned int qp;
398 	void *xdp_sent;
399 	int err;
400 
401 	qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
402 	sq = &vi->sq[qp];
403 
404 	/* Free up any pending old buffers before queueing new ones. */
405 	while ((xdp_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) {
406 		struct page *sent_page = virt_to_head_page(xdp_sent);
407 
408 		put_page(sent_page);
409 	}
410 
411 	xdp->data -= vi->hdr_len;
412 	/* Zero header and leave csum up to XDP layers */
413 	hdr = xdp->data;
414 	memset(hdr, 0, vi->hdr_len);
415 
416 	sg_init_one(sq->sg, xdp->data, xdp->data_end - xdp->data);
417 
418 	err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp->data, GFP_ATOMIC);
419 	if (unlikely(err)) {
420 		struct page *page = virt_to_head_page(xdp->data);
421 
422 		put_page(page);
423 		return false;
424 	}
425 
426 	return true;
427 }
428 
429 static int virtnet_xdp_xmit(struct net_device *dev, struct xdp_buff *xdp)
430 {
431 	struct virtnet_info *vi = netdev_priv(dev);
432 	bool sent = __virtnet_xdp_xmit(vi, xdp);
433 
434 	if (!sent)
435 		return -ENOSPC;
436 	return 0;
437 }
438 
439 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
440 {
441 	return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
442 }
443 
444 /* We copy the packet for XDP in the following cases:
445  *
446  * 1) Packet is scattered across multiple rx buffers.
447  * 2) Headroom space is insufficient.
448  *
449  * This is inefficient but it's a temporary condition that
450  * we hit right after XDP is enabled and until queue is refilled
451  * with large buffers with sufficient headroom - so it should affect
452  * at most queue size packets.
453  * Afterwards, the conditions to enable
454  * XDP should preclude the underlying device from sending packets
455  * across multiple buffers (num_buf > 1), and we make sure buffers
456  * have enough headroom.
457  */
458 static struct page *xdp_linearize_page(struct receive_queue *rq,
459 				       u16 *num_buf,
460 				       struct page *p,
461 				       int offset,
462 				       int page_off,
463 				       unsigned int *len)
464 {
465 	struct page *page = alloc_page(GFP_ATOMIC);
466 
467 	if (!page)
468 		return NULL;
469 
470 	memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
471 	page_off += *len;
472 
473 	while (--*num_buf) {
474 		unsigned int buflen;
475 		void *buf;
476 		int off;
477 
478 		buf = virtqueue_get_buf(rq->vq, &buflen);
479 		if (unlikely(!buf))
480 			goto err_buf;
481 
482 		p = virt_to_head_page(buf);
483 		off = buf - page_address(p);
484 
485 		/* guard against a misconfigured or uncooperative backend that
486 		 * is sending packet larger than the MTU.
487 		 */
488 		if ((page_off + buflen) > PAGE_SIZE) {
489 			put_page(p);
490 			goto err_buf;
491 		}
492 
493 		memcpy(page_address(page) + page_off,
494 		       page_address(p) + off, buflen);
495 		page_off += buflen;
496 		put_page(p);
497 	}
498 
499 	/* Headroom does not contribute to packet length */
500 	*len = page_off - VIRTIO_XDP_HEADROOM;
501 	return page;
502 err_buf:
503 	__free_pages(page, 0);
504 	return NULL;
505 }
506 
507 static struct sk_buff *receive_small(struct net_device *dev,
508 				     struct virtnet_info *vi,
509 				     struct receive_queue *rq,
510 				     void *buf, void *ctx,
511 				     unsigned int len,
512 				     bool *xdp_xmit)
513 {
514 	struct sk_buff *skb;
515 	struct bpf_prog *xdp_prog;
516 	unsigned int xdp_headroom = (unsigned long)ctx;
517 	unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
518 	unsigned int headroom = vi->hdr_len + header_offset;
519 	unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
520 			      SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
521 	struct page *page = virt_to_head_page(buf);
522 	unsigned int delta = 0, err;
523 	struct page *xdp_page;
524 	len -= vi->hdr_len;
525 
526 	rcu_read_lock();
527 	xdp_prog = rcu_dereference(rq->xdp_prog);
528 	if (xdp_prog) {
529 		struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
530 		struct xdp_buff xdp;
531 		void *orig_data;
532 		u32 act;
533 
534 		if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags))
535 			goto err_xdp;
536 
537 		if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
538 			int offset = buf - page_address(page) + header_offset;
539 			unsigned int tlen = len + vi->hdr_len;
540 			u16 num_buf = 1;
541 
542 			xdp_headroom = virtnet_get_headroom(vi);
543 			header_offset = VIRTNET_RX_PAD + xdp_headroom;
544 			headroom = vi->hdr_len + header_offset;
545 			buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
546 				 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
547 			xdp_page = xdp_linearize_page(rq, &num_buf, page,
548 						      offset, header_offset,
549 						      &tlen);
550 			if (!xdp_page)
551 				goto err_xdp;
552 
553 			buf = page_address(xdp_page);
554 			put_page(page);
555 			page = xdp_page;
556 		}
557 
558 		xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
559 		xdp.data = xdp.data_hard_start + xdp_headroom;
560 		xdp_set_data_meta_invalid(&xdp);
561 		xdp.data_end = xdp.data + len;
562 		orig_data = xdp.data;
563 		act = bpf_prog_run_xdp(xdp_prog, &xdp);
564 
565 		switch (act) {
566 		case XDP_PASS:
567 			/* Recalculate length in case bpf program changed it */
568 			delta = orig_data - xdp.data;
569 			break;
570 		case XDP_TX:
571 			if (unlikely(!__virtnet_xdp_xmit(vi, &xdp)))
572 				trace_xdp_exception(vi->dev, xdp_prog, act);
573 			else
574 				*xdp_xmit = true;
575 			rcu_read_unlock();
576 			goto xdp_xmit;
577 		case XDP_REDIRECT:
578 			err = xdp_do_redirect(dev, &xdp, xdp_prog);
579 			if (!err)
580 				*xdp_xmit = true;
581 			rcu_read_unlock();
582 			goto xdp_xmit;
583 		default:
584 			bpf_warn_invalid_xdp_action(act);
585 		case XDP_ABORTED:
586 			trace_xdp_exception(vi->dev, xdp_prog, act);
587 		case XDP_DROP:
588 			goto err_xdp;
589 		}
590 	}
591 	rcu_read_unlock();
592 
593 	skb = build_skb(buf, buflen);
594 	if (!skb) {
595 		put_page(page);
596 		goto err;
597 	}
598 	skb_reserve(skb, headroom - delta);
599 	skb_put(skb, len + delta);
600 	if (!delta) {
601 		buf += header_offset;
602 		memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
603 	} /* keep zeroed vnet hdr since packet was changed by bpf */
604 
605 err:
606 	return skb;
607 
608 err_xdp:
609 	rcu_read_unlock();
610 	dev->stats.rx_dropped++;
611 	put_page(page);
612 xdp_xmit:
613 	return NULL;
614 }
615 
616 static struct sk_buff *receive_big(struct net_device *dev,
617 				   struct virtnet_info *vi,
618 				   struct receive_queue *rq,
619 				   void *buf,
620 				   unsigned int len)
621 {
622 	struct page *page = buf;
623 	struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
624 
625 	if (unlikely(!skb))
626 		goto err;
627 
628 	return skb;
629 
630 err:
631 	dev->stats.rx_dropped++;
632 	give_pages(rq, page);
633 	return NULL;
634 }
635 
636 static struct sk_buff *receive_mergeable(struct net_device *dev,
637 					 struct virtnet_info *vi,
638 					 struct receive_queue *rq,
639 					 void *buf,
640 					 void *ctx,
641 					 unsigned int len,
642 					 bool *xdp_xmit)
643 {
644 	struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
645 	u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
646 	struct page *page = virt_to_head_page(buf);
647 	int offset = buf - page_address(page);
648 	struct sk_buff *head_skb, *curr_skb;
649 	struct bpf_prog *xdp_prog;
650 	unsigned int truesize;
651 	unsigned int headroom = mergeable_ctx_to_headroom(ctx);
652 	int err;
653 
654 	head_skb = NULL;
655 
656 	rcu_read_lock();
657 	xdp_prog = rcu_dereference(rq->xdp_prog);
658 	if (xdp_prog) {
659 		struct page *xdp_page;
660 		struct xdp_buff xdp;
661 		void *data;
662 		u32 act;
663 
664 		/* This happens when rx buffer size is underestimated */
665 		if (unlikely(num_buf > 1 ||
666 			     headroom < virtnet_get_headroom(vi))) {
667 			/* linearize data for XDP */
668 			xdp_page = xdp_linearize_page(rq, &num_buf,
669 						      page, offset,
670 						      VIRTIO_XDP_HEADROOM,
671 						      &len);
672 			if (!xdp_page)
673 				goto err_xdp;
674 			offset = VIRTIO_XDP_HEADROOM;
675 		} else {
676 			xdp_page = page;
677 		}
678 
679 		/* Transient failure which in theory could occur if
680 		 * in-flight packets from before XDP was enabled reach
681 		 * the receive path after XDP is loaded. In practice I
682 		 * was not able to create this condition.
683 		 */
684 		if (unlikely(hdr->hdr.gso_type))
685 			goto err_xdp;
686 
687 		/* Allow consuming headroom but reserve enough space to push
688 		 * the descriptor on if we get an XDP_TX return code.
689 		 */
690 		data = page_address(xdp_page) + offset;
691 		xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
692 		xdp.data = data + vi->hdr_len;
693 		xdp_set_data_meta_invalid(&xdp);
694 		xdp.data_end = xdp.data + (len - vi->hdr_len);
695 		act = bpf_prog_run_xdp(xdp_prog, &xdp);
696 
697 		if (act != XDP_PASS)
698 			ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
699 
700 		switch (act) {
701 		case XDP_PASS:
702 			/* recalculate offset to account for any header
703 			 * adjustments. Note other cases do not build an
704 			 * skb and avoid using offset
705 			 */
706 			offset = xdp.data -
707 					page_address(xdp_page) - vi->hdr_len;
708 
709 			/* We can only create skb based on xdp_page. */
710 			if (unlikely(xdp_page != page)) {
711 				rcu_read_unlock();
712 				put_page(page);
713 				head_skb = page_to_skb(vi, rq, xdp_page,
714 						       offset, len, PAGE_SIZE);
715 				return head_skb;
716 			}
717 			break;
718 		case XDP_TX:
719 			if (unlikely(!__virtnet_xdp_xmit(vi, &xdp)))
720 				trace_xdp_exception(vi->dev, xdp_prog, act);
721 			else
722 				*xdp_xmit = true;
723 			if (unlikely(xdp_page != page))
724 				goto err_xdp;
725 			rcu_read_unlock();
726 			goto xdp_xmit;
727 		case XDP_REDIRECT:
728 			err = xdp_do_redirect(dev, &xdp, xdp_prog);
729 			if (!err)
730 				*xdp_xmit = true;
731 			rcu_read_unlock();
732 			goto xdp_xmit;
733 		default:
734 			bpf_warn_invalid_xdp_action(act);
735 		case XDP_ABORTED:
736 			trace_xdp_exception(vi->dev, xdp_prog, act);
737 		case XDP_DROP:
738 			if (unlikely(xdp_page != page))
739 				__free_pages(xdp_page, 0);
740 			goto err_xdp;
741 		}
742 	}
743 	rcu_read_unlock();
744 
745 	truesize = mergeable_ctx_to_truesize(ctx);
746 	if (unlikely(len > truesize)) {
747 		pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
748 			 dev->name, len, (unsigned long)ctx);
749 		dev->stats.rx_length_errors++;
750 		goto err_skb;
751 	}
752 
753 	head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
754 	curr_skb = head_skb;
755 
756 	if (unlikely(!curr_skb))
757 		goto err_skb;
758 	while (--num_buf) {
759 		int num_skb_frags;
760 
761 		buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
762 		if (unlikely(!buf)) {
763 			pr_debug("%s: rx error: %d buffers out of %d missing\n",
764 				 dev->name, num_buf,
765 				 virtio16_to_cpu(vi->vdev,
766 						 hdr->num_buffers));
767 			dev->stats.rx_length_errors++;
768 			goto err_buf;
769 		}
770 
771 		page = virt_to_head_page(buf);
772 
773 		truesize = mergeable_ctx_to_truesize(ctx);
774 		if (unlikely(len > truesize)) {
775 			pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
776 				 dev->name, len, (unsigned long)ctx);
777 			dev->stats.rx_length_errors++;
778 			goto err_skb;
779 		}
780 
781 		num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
782 		if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
783 			struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
784 
785 			if (unlikely(!nskb))
786 				goto err_skb;
787 			if (curr_skb == head_skb)
788 				skb_shinfo(curr_skb)->frag_list = nskb;
789 			else
790 				curr_skb->next = nskb;
791 			curr_skb = nskb;
792 			head_skb->truesize += nskb->truesize;
793 			num_skb_frags = 0;
794 		}
795 		if (curr_skb != head_skb) {
796 			head_skb->data_len += len;
797 			head_skb->len += len;
798 			head_skb->truesize += truesize;
799 		}
800 		offset = buf - page_address(page);
801 		if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
802 			put_page(page);
803 			skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
804 					     len, truesize);
805 		} else {
806 			skb_add_rx_frag(curr_skb, num_skb_frags, page,
807 					offset, len, truesize);
808 		}
809 	}
810 
811 	ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
812 	return head_skb;
813 
814 err_xdp:
815 	rcu_read_unlock();
816 err_skb:
817 	put_page(page);
818 	while (--num_buf) {
819 		buf = virtqueue_get_buf(rq->vq, &len);
820 		if (unlikely(!buf)) {
821 			pr_debug("%s: rx error: %d buffers missing\n",
822 				 dev->name, num_buf);
823 			dev->stats.rx_length_errors++;
824 			break;
825 		}
826 		page = virt_to_head_page(buf);
827 		put_page(page);
828 	}
829 err_buf:
830 	dev->stats.rx_dropped++;
831 	dev_kfree_skb(head_skb);
832 xdp_xmit:
833 	return NULL;
834 }
835 
836 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
837 		       void *buf, unsigned int len, void **ctx, bool *xdp_xmit)
838 {
839 	struct net_device *dev = vi->dev;
840 	struct sk_buff *skb;
841 	struct virtio_net_hdr_mrg_rxbuf *hdr;
842 	int ret;
843 
844 	if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
845 		pr_debug("%s: short packet %i\n", dev->name, len);
846 		dev->stats.rx_length_errors++;
847 		if (vi->mergeable_rx_bufs) {
848 			put_page(virt_to_head_page(buf));
849 		} else if (vi->big_packets) {
850 			give_pages(rq, buf);
851 		} else {
852 			put_page(virt_to_head_page(buf));
853 		}
854 		return 0;
855 	}
856 
857 	if (vi->mergeable_rx_bufs)
858 		skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit);
859 	else if (vi->big_packets)
860 		skb = receive_big(dev, vi, rq, buf, len);
861 	else
862 		skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit);
863 
864 	if (unlikely(!skb))
865 		return 0;
866 
867 	hdr = skb_vnet_hdr(skb);
868 
869 	ret = skb->len;
870 
871 	if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
872 		skb->ip_summed = CHECKSUM_UNNECESSARY;
873 
874 	if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
875 				  virtio_is_little_endian(vi->vdev))) {
876 		net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
877 				     dev->name, hdr->hdr.gso_type,
878 				     hdr->hdr.gso_size);
879 		goto frame_err;
880 	}
881 
882 	skb->protocol = eth_type_trans(skb, dev);
883 	pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
884 		 ntohs(skb->protocol), skb->len, skb->pkt_type);
885 
886 	napi_gro_receive(&rq->napi, skb);
887 	return ret;
888 
889 frame_err:
890 	dev->stats.rx_frame_errors++;
891 	dev_kfree_skb(skb);
892 	return 0;
893 }
894 
895 /* Unlike mergeable buffers, all buffers are allocated to the
896  * same size, except for the headroom. For this reason we do
897  * not need to use  mergeable_len_to_ctx here - it is enough
898  * to store the headroom as the context ignoring the truesize.
899  */
900 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
901 			     gfp_t gfp)
902 {
903 	struct page_frag *alloc_frag = &rq->alloc_frag;
904 	char *buf;
905 	unsigned int xdp_headroom = virtnet_get_headroom(vi);
906 	void *ctx = (void *)(unsigned long)xdp_headroom;
907 	int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
908 	int err;
909 
910 	len = SKB_DATA_ALIGN(len) +
911 	      SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
912 	if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
913 		return -ENOMEM;
914 
915 	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
916 	get_page(alloc_frag->page);
917 	alloc_frag->offset += len;
918 	sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
919 		    vi->hdr_len + GOOD_PACKET_LEN);
920 	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
921 	if (err < 0)
922 		put_page(virt_to_head_page(buf));
923 	return err;
924 }
925 
926 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
927 			   gfp_t gfp)
928 {
929 	struct page *first, *list = NULL;
930 	char *p;
931 	int i, err, offset;
932 
933 	sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
934 
935 	/* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
936 	for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
937 		first = get_a_page(rq, gfp);
938 		if (!first) {
939 			if (list)
940 				give_pages(rq, list);
941 			return -ENOMEM;
942 		}
943 		sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
944 
945 		/* chain new page in list head to match sg */
946 		first->private = (unsigned long)list;
947 		list = first;
948 	}
949 
950 	first = get_a_page(rq, gfp);
951 	if (!first) {
952 		give_pages(rq, list);
953 		return -ENOMEM;
954 	}
955 	p = page_address(first);
956 
957 	/* rq->sg[0], rq->sg[1] share the same page */
958 	/* a separated rq->sg[0] for header - required in case !any_header_sg */
959 	sg_set_buf(&rq->sg[0], p, vi->hdr_len);
960 
961 	/* rq->sg[1] for data packet, from offset */
962 	offset = sizeof(struct padded_vnet_hdr);
963 	sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
964 
965 	/* chain first in list head */
966 	first->private = (unsigned long)list;
967 	err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
968 				  first, gfp);
969 	if (err < 0)
970 		give_pages(rq, first);
971 
972 	return err;
973 }
974 
975 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
976 					  struct ewma_pkt_len *avg_pkt_len)
977 {
978 	const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
979 	unsigned int len;
980 
981 	len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
982 				rq->min_buf_len, PAGE_SIZE - hdr_len);
983 	return ALIGN(len, L1_CACHE_BYTES);
984 }
985 
986 static int add_recvbuf_mergeable(struct virtnet_info *vi,
987 				 struct receive_queue *rq, gfp_t gfp)
988 {
989 	struct page_frag *alloc_frag = &rq->alloc_frag;
990 	unsigned int headroom = virtnet_get_headroom(vi);
991 	char *buf;
992 	void *ctx;
993 	int err;
994 	unsigned int len, hole;
995 
996 	len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len);
997 	if (unlikely(!skb_page_frag_refill(len + headroom, alloc_frag, gfp)))
998 		return -ENOMEM;
999 
1000 	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1001 	buf += headroom; /* advance address leaving hole at front of pkt */
1002 	get_page(alloc_frag->page);
1003 	alloc_frag->offset += len + headroom;
1004 	hole = alloc_frag->size - alloc_frag->offset;
1005 	if (hole < len + headroom) {
1006 		/* To avoid internal fragmentation, if there is very likely not
1007 		 * enough space for another buffer, add the remaining space to
1008 		 * the current buffer.
1009 		 */
1010 		len += hole;
1011 		alloc_frag->offset += hole;
1012 	}
1013 
1014 	sg_init_one(rq->sg, buf, len);
1015 	ctx = mergeable_len_to_ctx(len, headroom);
1016 	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1017 	if (err < 0)
1018 		put_page(virt_to_head_page(buf));
1019 
1020 	return err;
1021 }
1022 
1023 /*
1024  * Returns false if we couldn't fill entirely (OOM).
1025  *
1026  * Normally run in the receive path, but can also be run from ndo_open
1027  * before we're receiving packets, or from refill_work which is
1028  * careful to disable receiving (using napi_disable).
1029  */
1030 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1031 			  gfp_t gfp)
1032 {
1033 	int err;
1034 	bool oom;
1035 
1036 	do {
1037 		if (vi->mergeable_rx_bufs)
1038 			err = add_recvbuf_mergeable(vi, rq, gfp);
1039 		else if (vi->big_packets)
1040 			err = add_recvbuf_big(vi, rq, gfp);
1041 		else
1042 			err = add_recvbuf_small(vi, rq, gfp);
1043 
1044 		oom = err == -ENOMEM;
1045 		if (err)
1046 			break;
1047 	} while (rq->vq->num_free);
1048 	virtqueue_kick(rq->vq);
1049 	return !oom;
1050 }
1051 
1052 static void skb_recv_done(struct virtqueue *rvq)
1053 {
1054 	struct virtnet_info *vi = rvq->vdev->priv;
1055 	struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1056 
1057 	virtqueue_napi_schedule(&rq->napi, rvq);
1058 }
1059 
1060 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1061 {
1062 	napi_enable(napi);
1063 
1064 	/* If all buffers were filled by other side before we napi_enabled, we
1065 	 * won't get another interrupt, so process any outstanding packets now.
1066 	 * Call local_bh_enable after to trigger softIRQ processing.
1067 	 */
1068 	local_bh_disable();
1069 	virtqueue_napi_schedule(napi, vq);
1070 	local_bh_enable();
1071 }
1072 
1073 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1074 				   struct virtqueue *vq,
1075 				   struct napi_struct *napi)
1076 {
1077 	if (!napi->weight)
1078 		return;
1079 
1080 	/* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1081 	 * enable the feature if this is likely affine with the transmit path.
1082 	 */
1083 	if (!vi->affinity_hint_set) {
1084 		napi->weight = 0;
1085 		return;
1086 	}
1087 
1088 	return virtnet_napi_enable(vq, napi);
1089 }
1090 
1091 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1092 {
1093 	if (napi->weight)
1094 		napi_disable(napi);
1095 }
1096 
1097 static void refill_work(struct work_struct *work)
1098 {
1099 	struct virtnet_info *vi =
1100 		container_of(work, struct virtnet_info, refill.work);
1101 	bool still_empty;
1102 	int i;
1103 
1104 	for (i = 0; i < vi->curr_queue_pairs; i++) {
1105 		struct receive_queue *rq = &vi->rq[i];
1106 
1107 		napi_disable(&rq->napi);
1108 		still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1109 		virtnet_napi_enable(rq->vq, &rq->napi);
1110 
1111 		/* In theory, this can happen: if we don't get any buffers in
1112 		 * we will *never* try to fill again.
1113 		 */
1114 		if (still_empty)
1115 			schedule_delayed_work(&vi->refill, HZ/2);
1116 	}
1117 }
1118 
1119 static int virtnet_receive(struct receive_queue *rq, int budget, bool *xdp_xmit)
1120 {
1121 	struct virtnet_info *vi = rq->vq->vdev->priv;
1122 	unsigned int len, received = 0, bytes = 0;
1123 	void *buf;
1124 	struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1125 
1126 	if (!vi->big_packets || vi->mergeable_rx_bufs) {
1127 		void *ctx;
1128 
1129 		while (received < budget &&
1130 		       (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1131 			bytes += receive_buf(vi, rq, buf, len, ctx, xdp_xmit);
1132 			received++;
1133 		}
1134 	} else {
1135 		while (received < budget &&
1136 		       (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1137 			bytes += receive_buf(vi, rq, buf, len, NULL, xdp_xmit);
1138 			received++;
1139 		}
1140 	}
1141 
1142 	if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1143 		if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1144 			schedule_delayed_work(&vi->refill, 0);
1145 	}
1146 
1147 	u64_stats_update_begin(&stats->rx_syncp);
1148 	stats->rx_bytes += bytes;
1149 	stats->rx_packets += received;
1150 	u64_stats_update_end(&stats->rx_syncp);
1151 
1152 	return received;
1153 }
1154 
1155 static void free_old_xmit_skbs(struct send_queue *sq)
1156 {
1157 	struct sk_buff *skb;
1158 	unsigned int len;
1159 	struct virtnet_info *vi = sq->vq->vdev->priv;
1160 	struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1161 	unsigned int packets = 0;
1162 	unsigned int bytes = 0;
1163 
1164 	while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1165 		pr_debug("Sent skb %p\n", skb);
1166 
1167 		bytes += skb->len;
1168 		packets++;
1169 
1170 		dev_consume_skb_any(skb);
1171 	}
1172 
1173 	/* Avoid overhead when no packets have been processed
1174 	 * happens when called speculatively from start_xmit.
1175 	 */
1176 	if (!packets)
1177 		return;
1178 
1179 	u64_stats_update_begin(&stats->tx_syncp);
1180 	stats->tx_bytes += bytes;
1181 	stats->tx_packets += packets;
1182 	u64_stats_update_end(&stats->tx_syncp);
1183 }
1184 
1185 static void virtnet_poll_cleantx(struct receive_queue *rq)
1186 {
1187 	struct virtnet_info *vi = rq->vq->vdev->priv;
1188 	unsigned int index = vq2rxq(rq->vq);
1189 	struct send_queue *sq = &vi->sq[index];
1190 	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1191 
1192 	if (!sq->napi.weight)
1193 		return;
1194 
1195 	if (__netif_tx_trylock(txq)) {
1196 		free_old_xmit_skbs(sq);
1197 		__netif_tx_unlock(txq);
1198 	}
1199 
1200 	if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1201 		netif_tx_wake_queue(txq);
1202 }
1203 
1204 static int virtnet_poll(struct napi_struct *napi, int budget)
1205 {
1206 	struct receive_queue *rq =
1207 		container_of(napi, struct receive_queue, napi);
1208 	unsigned int received;
1209 	bool xdp_xmit = false;
1210 
1211 	virtnet_poll_cleantx(rq);
1212 
1213 	received = virtnet_receive(rq, budget, &xdp_xmit);
1214 
1215 	/* Out of packets? */
1216 	if (received < budget)
1217 		virtqueue_napi_complete(napi, rq->vq, received);
1218 
1219 	if (xdp_xmit)
1220 		xdp_do_flush_map();
1221 
1222 	return received;
1223 }
1224 
1225 static int virtnet_open(struct net_device *dev)
1226 {
1227 	struct virtnet_info *vi = netdev_priv(dev);
1228 	int i;
1229 
1230 	for (i = 0; i < vi->max_queue_pairs; i++) {
1231 		if (i < vi->curr_queue_pairs)
1232 			/* Make sure we have some buffers: if oom use wq. */
1233 			if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1234 				schedule_delayed_work(&vi->refill, 0);
1235 		virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1236 		virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1237 	}
1238 
1239 	return 0;
1240 }
1241 
1242 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1243 {
1244 	struct send_queue *sq = container_of(napi, struct send_queue, napi);
1245 	struct virtnet_info *vi = sq->vq->vdev->priv;
1246 	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
1247 
1248 	__netif_tx_lock(txq, raw_smp_processor_id());
1249 	free_old_xmit_skbs(sq);
1250 	__netif_tx_unlock(txq);
1251 
1252 	virtqueue_napi_complete(napi, sq->vq, 0);
1253 
1254 	if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1255 		netif_tx_wake_queue(txq);
1256 
1257 	return 0;
1258 }
1259 
1260 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1261 {
1262 	struct virtio_net_hdr_mrg_rxbuf *hdr;
1263 	const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1264 	struct virtnet_info *vi = sq->vq->vdev->priv;
1265 	int num_sg;
1266 	unsigned hdr_len = vi->hdr_len;
1267 	bool can_push;
1268 
1269 	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1270 
1271 	can_push = vi->any_header_sg &&
1272 		!((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1273 		!skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1274 	/* Even if we can, don't push here yet as this would skew
1275 	 * csum_start offset below. */
1276 	if (can_push)
1277 		hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1278 	else
1279 		hdr = skb_vnet_hdr(skb);
1280 
1281 	if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1282 				    virtio_is_little_endian(vi->vdev), false))
1283 		BUG();
1284 
1285 	if (vi->mergeable_rx_bufs)
1286 		hdr->num_buffers = 0;
1287 
1288 	sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1289 	if (can_push) {
1290 		__skb_push(skb, hdr_len);
1291 		num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1292 		if (unlikely(num_sg < 0))
1293 			return num_sg;
1294 		/* Pull header back to avoid skew in tx bytes calculations. */
1295 		__skb_pull(skb, hdr_len);
1296 	} else {
1297 		sg_set_buf(sq->sg, hdr, hdr_len);
1298 		num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1299 		if (unlikely(num_sg < 0))
1300 			return num_sg;
1301 		num_sg++;
1302 	}
1303 	return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1304 }
1305 
1306 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1307 {
1308 	struct virtnet_info *vi = netdev_priv(dev);
1309 	int qnum = skb_get_queue_mapping(skb);
1310 	struct send_queue *sq = &vi->sq[qnum];
1311 	int err;
1312 	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1313 	bool kick = !skb->xmit_more;
1314 	bool use_napi = sq->napi.weight;
1315 
1316 	/* Free up any pending old buffers before queueing new ones. */
1317 	free_old_xmit_skbs(sq);
1318 
1319 	if (use_napi && kick)
1320 		virtqueue_enable_cb_delayed(sq->vq);
1321 
1322 	/* timestamp packet in software */
1323 	skb_tx_timestamp(skb);
1324 
1325 	/* Try to transmit */
1326 	err = xmit_skb(sq, skb);
1327 
1328 	/* This should not happen! */
1329 	if (unlikely(err)) {
1330 		dev->stats.tx_fifo_errors++;
1331 		if (net_ratelimit())
1332 			dev_warn(&dev->dev,
1333 				 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1334 		dev->stats.tx_dropped++;
1335 		dev_kfree_skb_any(skb);
1336 		return NETDEV_TX_OK;
1337 	}
1338 
1339 	/* Don't wait up for transmitted skbs to be freed. */
1340 	if (!use_napi) {
1341 		skb_orphan(skb);
1342 		nf_reset(skb);
1343 	}
1344 
1345 	/* If running out of space, stop queue to avoid getting packets that we
1346 	 * are then unable to transmit.
1347 	 * An alternative would be to force queuing layer to requeue the skb by
1348 	 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1349 	 * returned in a normal path of operation: it means that driver is not
1350 	 * maintaining the TX queue stop/start state properly, and causes
1351 	 * the stack to do a non-trivial amount of useless work.
1352 	 * Since most packets only take 1 or 2 ring slots, stopping the queue
1353 	 * early means 16 slots are typically wasted.
1354 	 */
1355 	if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1356 		netif_stop_subqueue(dev, qnum);
1357 		if (!use_napi &&
1358 		    unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1359 			/* More just got used, free them then recheck. */
1360 			free_old_xmit_skbs(sq);
1361 			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1362 				netif_start_subqueue(dev, qnum);
1363 				virtqueue_disable_cb(sq->vq);
1364 			}
1365 		}
1366 	}
1367 
1368 	if (kick || netif_xmit_stopped(txq))
1369 		virtqueue_kick(sq->vq);
1370 
1371 	return NETDEV_TX_OK;
1372 }
1373 
1374 /*
1375  * Send command via the control virtqueue and check status.  Commands
1376  * supported by the hypervisor, as indicated by feature bits, should
1377  * never fail unless improperly formatted.
1378  */
1379 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1380 				 struct scatterlist *out)
1381 {
1382 	struct scatterlist *sgs[4], hdr, stat;
1383 	unsigned out_num = 0, tmp;
1384 
1385 	/* Caller should know better */
1386 	BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1387 
1388 	vi->ctrl_status = ~0;
1389 	vi->ctrl_hdr.class = class;
1390 	vi->ctrl_hdr.cmd = cmd;
1391 	/* Add header */
1392 	sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
1393 	sgs[out_num++] = &hdr;
1394 
1395 	if (out)
1396 		sgs[out_num++] = out;
1397 
1398 	/* Add return status. */
1399 	sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1400 	sgs[out_num] = &stat;
1401 
1402 	BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1403 	virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1404 
1405 	if (unlikely(!virtqueue_kick(vi->cvq)))
1406 		return vi->ctrl_status == VIRTIO_NET_OK;
1407 
1408 	/* Spin for a response, the kick causes an ioport write, trapping
1409 	 * into the hypervisor, so the request should be handled immediately.
1410 	 */
1411 	while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1412 	       !virtqueue_is_broken(vi->cvq))
1413 		cpu_relax();
1414 
1415 	return vi->ctrl_status == VIRTIO_NET_OK;
1416 }
1417 
1418 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1419 {
1420 	struct virtnet_info *vi = netdev_priv(dev);
1421 	struct virtio_device *vdev = vi->vdev;
1422 	int ret;
1423 	struct sockaddr *addr;
1424 	struct scatterlist sg;
1425 
1426 	addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1427 	if (!addr)
1428 		return -ENOMEM;
1429 
1430 	ret = eth_prepare_mac_addr_change(dev, addr);
1431 	if (ret)
1432 		goto out;
1433 
1434 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1435 		sg_init_one(&sg, addr->sa_data, dev->addr_len);
1436 		if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1437 					  VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1438 			dev_warn(&vdev->dev,
1439 				 "Failed to set mac address by vq command.\n");
1440 			ret = -EINVAL;
1441 			goto out;
1442 		}
1443 	} else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1444 		   !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1445 		unsigned int i;
1446 
1447 		/* Naturally, this has an atomicity problem. */
1448 		for (i = 0; i < dev->addr_len; i++)
1449 			virtio_cwrite8(vdev,
1450 				       offsetof(struct virtio_net_config, mac) +
1451 				       i, addr->sa_data[i]);
1452 	}
1453 
1454 	eth_commit_mac_addr_change(dev, p);
1455 	ret = 0;
1456 
1457 out:
1458 	kfree(addr);
1459 	return ret;
1460 }
1461 
1462 static void virtnet_stats(struct net_device *dev,
1463 			  struct rtnl_link_stats64 *tot)
1464 {
1465 	struct virtnet_info *vi = netdev_priv(dev);
1466 	int cpu;
1467 	unsigned int start;
1468 
1469 	for_each_possible_cpu(cpu) {
1470 		struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1471 		u64 tpackets, tbytes, rpackets, rbytes;
1472 
1473 		do {
1474 			start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1475 			tpackets = stats->tx_packets;
1476 			tbytes   = stats->tx_bytes;
1477 		} while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1478 
1479 		do {
1480 			start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1481 			rpackets = stats->rx_packets;
1482 			rbytes   = stats->rx_bytes;
1483 		} while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1484 
1485 		tot->rx_packets += rpackets;
1486 		tot->tx_packets += tpackets;
1487 		tot->rx_bytes   += rbytes;
1488 		tot->tx_bytes   += tbytes;
1489 	}
1490 
1491 	tot->tx_dropped = dev->stats.tx_dropped;
1492 	tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1493 	tot->rx_dropped = dev->stats.rx_dropped;
1494 	tot->rx_length_errors = dev->stats.rx_length_errors;
1495 	tot->rx_frame_errors = dev->stats.rx_frame_errors;
1496 }
1497 
1498 #ifdef CONFIG_NET_POLL_CONTROLLER
1499 static void virtnet_netpoll(struct net_device *dev)
1500 {
1501 	struct virtnet_info *vi = netdev_priv(dev);
1502 	int i;
1503 
1504 	for (i = 0; i < vi->curr_queue_pairs; i++)
1505 		napi_schedule(&vi->rq[i].napi);
1506 }
1507 #endif
1508 
1509 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1510 {
1511 	rtnl_lock();
1512 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1513 				  VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1514 		dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1515 	rtnl_unlock();
1516 }
1517 
1518 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1519 {
1520 	struct scatterlist sg;
1521 	struct net_device *dev = vi->dev;
1522 
1523 	if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1524 		return 0;
1525 
1526 	vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1527 	sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1528 
1529 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1530 				  VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1531 		dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1532 			 queue_pairs);
1533 		return -EINVAL;
1534 	} else {
1535 		vi->curr_queue_pairs = queue_pairs;
1536 		/* virtnet_open() will refill when device is going to up. */
1537 		if (dev->flags & IFF_UP)
1538 			schedule_delayed_work(&vi->refill, 0);
1539 	}
1540 
1541 	return 0;
1542 }
1543 
1544 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1545 {
1546 	int err;
1547 
1548 	rtnl_lock();
1549 	err = _virtnet_set_queues(vi, queue_pairs);
1550 	rtnl_unlock();
1551 	return err;
1552 }
1553 
1554 static int virtnet_close(struct net_device *dev)
1555 {
1556 	struct virtnet_info *vi = netdev_priv(dev);
1557 	int i;
1558 
1559 	/* Make sure refill_work doesn't re-enable napi! */
1560 	cancel_delayed_work_sync(&vi->refill);
1561 
1562 	for (i = 0; i < vi->max_queue_pairs; i++) {
1563 		napi_disable(&vi->rq[i].napi);
1564 		virtnet_napi_tx_disable(&vi->sq[i].napi);
1565 	}
1566 
1567 	return 0;
1568 }
1569 
1570 static void virtnet_set_rx_mode(struct net_device *dev)
1571 {
1572 	struct virtnet_info *vi = netdev_priv(dev);
1573 	struct scatterlist sg[2];
1574 	struct virtio_net_ctrl_mac *mac_data;
1575 	struct netdev_hw_addr *ha;
1576 	int uc_count;
1577 	int mc_count;
1578 	void *buf;
1579 	int i;
1580 
1581 	/* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1582 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1583 		return;
1584 
1585 	vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1586 	vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1587 
1588 	sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1589 
1590 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1591 				  VIRTIO_NET_CTRL_RX_PROMISC, sg))
1592 		dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1593 			 vi->ctrl_promisc ? "en" : "dis");
1594 
1595 	sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1596 
1597 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1598 				  VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1599 		dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1600 			 vi->ctrl_allmulti ? "en" : "dis");
1601 
1602 	uc_count = netdev_uc_count(dev);
1603 	mc_count = netdev_mc_count(dev);
1604 	/* MAC filter - use one buffer for both lists */
1605 	buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1606 		      (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1607 	mac_data = buf;
1608 	if (!buf)
1609 		return;
1610 
1611 	sg_init_table(sg, 2);
1612 
1613 	/* Store the unicast list and count in the front of the buffer */
1614 	mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1615 	i = 0;
1616 	netdev_for_each_uc_addr(ha, dev)
1617 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1618 
1619 	sg_set_buf(&sg[0], mac_data,
1620 		   sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1621 
1622 	/* multicast list and count fill the end */
1623 	mac_data = (void *)&mac_data->macs[uc_count][0];
1624 
1625 	mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1626 	i = 0;
1627 	netdev_for_each_mc_addr(ha, dev)
1628 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1629 
1630 	sg_set_buf(&sg[1], mac_data,
1631 		   sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1632 
1633 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1634 				  VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1635 		dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1636 
1637 	kfree(buf);
1638 }
1639 
1640 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1641 				   __be16 proto, u16 vid)
1642 {
1643 	struct virtnet_info *vi = netdev_priv(dev);
1644 	struct scatterlist sg;
1645 
1646 	vi->ctrl_vid = vid;
1647 	sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1648 
1649 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1650 				  VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1651 		dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1652 	return 0;
1653 }
1654 
1655 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1656 				    __be16 proto, u16 vid)
1657 {
1658 	struct virtnet_info *vi = netdev_priv(dev);
1659 	struct scatterlist sg;
1660 
1661 	vi->ctrl_vid = vid;
1662 	sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1663 
1664 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1665 				  VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1666 		dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1667 	return 0;
1668 }
1669 
1670 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1671 {
1672 	int i;
1673 
1674 	if (vi->affinity_hint_set) {
1675 		for (i = 0; i < vi->max_queue_pairs; i++) {
1676 			virtqueue_set_affinity(vi->rq[i].vq, -1);
1677 			virtqueue_set_affinity(vi->sq[i].vq, -1);
1678 		}
1679 
1680 		vi->affinity_hint_set = false;
1681 	}
1682 }
1683 
1684 static void virtnet_set_affinity(struct virtnet_info *vi)
1685 {
1686 	int i;
1687 	int cpu;
1688 
1689 	/* In multiqueue mode, when the number of cpu is equal to the number of
1690 	 * queue pairs, we let the queue pairs to be private to one cpu by
1691 	 * setting the affinity hint to eliminate the contention.
1692 	 */
1693 	if (vi->curr_queue_pairs == 1 ||
1694 	    vi->max_queue_pairs != num_online_cpus()) {
1695 		virtnet_clean_affinity(vi, -1);
1696 		return;
1697 	}
1698 
1699 	i = 0;
1700 	for_each_online_cpu(cpu) {
1701 		virtqueue_set_affinity(vi->rq[i].vq, cpu);
1702 		virtqueue_set_affinity(vi->sq[i].vq, cpu);
1703 		netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1704 		i++;
1705 	}
1706 
1707 	vi->affinity_hint_set = true;
1708 }
1709 
1710 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1711 {
1712 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1713 						   node);
1714 	virtnet_set_affinity(vi);
1715 	return 0;
1716 }
1717 
1718 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1719 {
1720 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1721 						   node_dead);
1722 	virtnet_set_affinity(vi);
1723 	return 0;
1724 }
1725 
1726 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1727 {
1728 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1729 						   node);
1730 
1731 	virtnet_clean_affinity(vi, cpu);
1732 	return 0;
1733 }
1734 
1735 static enum cpuhp_state virtionet_online;
1736 
1737 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1738 {
1739 	int ret;
1740 
1741 	ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1742 	if (ret)
1743 		return ret;
1744 	ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1745 					       &vi->node_dead);
1746 	if (!ret)
1747 		return ret;
1748 	cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1749 	return ret;
1750 }
1751 
1752 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1753 {
1754 	cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1755 	cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1756 					    &vi->node_dead);
1757 }
1758 
1759 static void virtnet_get_ringparam(struct net_device *dev,
1760 				struct ethtool_ringparam *ring)
1761 {
1762 	struct virtnet_info *vi = netdev_priv(dev);
1763 
1764 	ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1765 	ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1766 	ring->rx_pending = ring->rx_max_pending;
1767 	ring->tx_pending = ring->tx_max_pending;
1768 }
1769 
1770 
1771 static void virtnet_get_drvinfo(struct net_device *dev,
1772 				struct ethtool_drvinfo *info)
1773 {
1774 	struct virtnet_info *vi = netdev_priv(dev);
1775 	struct virtio_device *vdev = vi->vdev;
1776 
1777 	strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1778 	strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1779 	strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1780 
1781 }
1782 
1783 /* TODO: Eliminate OOO packets during switching */
1784 static int virtnet_set_channels(struct net_device *dev,
1785 				struct ethtool_channels *channels)
1786 {
1787 	struct virtnet_info *vi = netdev_priv(dev);
1788 	u16 queue_pairs = channels->combined_count;
1789 	int err;
1790 
1791 	/* We don't support separate rx/tx channels.
1792 	 * We don't allow setting 'other' channels.
1793 	 */
1794 	if (channels->rx_count || channels->tx_count || channels->other_count)
1795 		return -EINVAL;
1796 
1797 	if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1798 		return -EINVAL;
1799 
1800 	/* For now we don't support modifying channels while XDP is loaded
1801 	 * also when XDP is loaded all RX queues have XDP programs so we only
1802 	 * need to check a single RX queue.
1803 	 */
1804 	if (vi->rq[0].xdp_prog)
1805 		return -EINVAL;
1806 
1807 	get_online_cpus();
1808 	err = _virtnet_set_queues(vi, queue_pairs);
1809 	if (!err) {
1810 		netif_set_real_num_tx_queues(dev, queue_pairs);
1811 		netif_set_real_num_rx_queues(dev, queue_pairs);
1812 
1813 		virtnet_set_affinity(vi);
1814 	}
1815 	put_online_cpus();
1816 
1817 	return err;
1818 }
1819 
1820 static void virtnet_get_channels(struct net_device *dev,
1821 				 struct ethtool_channels *channels)
1822 {
1823 	struct virtnet_info *vi = netdev_priv(dev);
1824 
1825 	channels->combined_count = vi->curr_queue_pairs;
1826 	channels->max_combined = vi->max_queue_pairs;
1827 	channels->max_other = 0;
1828 	channels->rx_count = 0;
1829 	channels->tx_count = 0;
1830 	channels->other_count = 0;
1831 }
1832 
1833 /* Check if the user is trying to change anything besides speed/duplex */
1834 static bool
1835 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
1836 {
1837 	struct ethtool_link_ksettings diff1 = *cmd;
1838 	struct ethtool_link_ksettings diff2 = {};
1839 
1840 	/* cmd is always set so we need to clear it, validate the port type
1841 	 * and also without autonegotiation we can ignore advertising
1842 	 */
1843 	diff1.base.speed = 0;
1844 	diff2.base.port = PORT_OTHER;
1845 	ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
1846 	diff1.base.duplex = 0;
1847 	diff1.base.cmd = 0;
1848 	diff1.base.link_mode_masks_nwords = 0;
1849 
1850 	return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
1851 		bitmap_empty(diff1.link_modes.supported,
1852 			     __ETHTOOL_LINK_MODE_MASK_NBITS) &&
1853 		bitmap_empty(diff1.link_modes.advertising,
1854 			     __ETHTOOL_LINK_MODE_MASK_NBITS) &&
1855 		bitmap_empty(diff1.link_modes.lp_advertising,
1856 			     __ETHTOOL_LINK_MODE_MASK_NBITS);
1857 }
1858 
1859 static int virtnet_set_link_ksettings(struct net_device *dev,
1860 				      const struct ethtool_link_ksettings *cmd)
1861 {
1862 	struct virtnet_info *vi = netdev_priv(dev);
1863 	u32 speed;
1864 
1865 	speed = cmd->base.speed;
1866 	/* don't allow custom speed and duplex */
1867 	if (!ethtool_validate_speed(speed) ||
1868 	    !ethtool_validate_duplex(cmd->base.duplex) ||
1869 	    !virtnet_validate_ethtool_cmd(cmd))
1870 		return -EINVAL;
1871 	vi->speed = speed;
1872 	vi->duplex = cmd->base.duplex;
1873 
1874 	return 0;
1875 }
1876 
1877 static int virtnet_get_link_ksettings(struct net_device *dev,
1878 				      struct ethtool_link_ksettings *cmd)
1879 {
1880 	struct virtnet_info *vi = netdev_priv(dev);
1881 
1882 	cmd->base.speed = vi->speed;
1883 	cmd->base.duplex = vi->duplex;
1884 	cmd->base.port = PORT_OTHER;
1885 
1886 	return 0;
1887 }
1888 
1889 static void virtnet_init_settings(struct net_device *dev)
1890 {
1891 	struct virtnet_info *vi = netdev_priv(dev);
1892 
1893 	vi->speed = SPEED_UNKNOWN;
1894 	vi->duplex = DUPLEX_UNKNOWN;
1895 }
1896 
1897 static const struct ethtool_ops virtnet_ethtool_ops = {
1898 	.get_drvinfo = virtnet_get_drvinfo,
1899 	.get_link = ethtool_op_get_link,
1900 	.get_ringparam = virtnet_get_ringparam,
1901 	.set_channels = virtnet_set_channels,
1902 	.get_channels = virtnet_get_channels,
1903 	.get_ts_info = ethtool_op_get_ts_info,
1904 	.get_link_ksettings = virtnet_get_link_ksettings,
1905 	.set_link_ksettings = virtnet_set_link_ksettings,
1906 };
1907 
1908 static void virtnet_freeze_down(struct virtio_device *vdev)
1909 {
1910 	struct virtnet_info *vi = vdev->priv;
1911 	int i;
1912 
1913 	/* Make sure no work handler is accessing the device */
1914 	flush_work(&vi->config_work);
1915 
1916 	netif_device_detach(vi->dev);
1917 	netif_tx_disable(vi->dev);
1918 	cancel_delayed_work_sync(&vi->refill);
1919 
1920 	if (netif_running(vi->dev)) {
1921 		for (i = 0; i < vi->max_queue_pairs; i++) {
1922 			napi_disable(&vi->rq[i].napi);
1923 			virtnet_napi_tx_disable(&vi->sq[i].napi);
1924 		}
1925 	}
1926 }
1927 
1928 static int init_vqs(struct virtnet_info *vi);
1929 
1930 static int virtnet_restore_up(struct virtio_device *vdev)
1931 {
1932 	struct virtnet_info *vi = vdev->priv;
1933 	int err, i;
1934 
1935 	err = init_vqs(vi);
1936 	if (err)
1937 		return err;
1938 
1939 	virtio_device_ready(vdev);
1940 
1941 	if (netif_running(vi->dev)) {
1942 		for (i = 0; i < vi->curr_queue_pairs; i++)
1943 			if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1944 				schedule_delayed_work(&vi->refill, 0);
1945 
1946 		for (i = 0; i < vi->max_queue_pairs; i++) {
1947 			virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1948 			virtnet_napi_tx_enable(vi, vi->sq[i].vq,
1949 					       &vi->sq[i].napi);
1950 		}
1951 	}
1952 
1953 	netif_device_attach(vi->dev);
1954 	return err;
1955 }
1956 
1957 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
1958 {
1959 	struct scatterlist sg;
1960 	vi->ctrl_offloads = cpu_to_virtio64(vi->vdev, offloads);
1961 
1962 	sg_init_one(&sg, &vi->ctrl_offloads, sizeof(vi->ctrl_offloads));
1963 
1964 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
1965 				  VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
1966 		dev_warn(&vi->dev->dev, "Fail to set guest offload. \n");
1967 		return -EINVAL;
1968 	}
1969 
1970 	return 0;
1971 }
1972 
1973 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
1974 {
1975 	u64 offloads = 0;
1976 
1977 	if (!vi->guest_offloads)
1978 		return 0;
1979 
1980 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
1981 		offloads = 1ULL << VIRTIO_NET_F_GUEST_CSUM;
1982 
1983 	return virtnet_set_guest_offloads(vi, offloads);
1984 }
1985 
1986 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
1987 {
1988 	u64 offloads = vi->guest_offloads;
1989 
1990 	if (!vi->guest_offloads)
1991 		return 0;
1992 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
1993 		offloads |= 1ULL << VIRTIO_NET_F_GUEST_CSUM;
1994 
1995 	return virtnet_set_guest_offloads(vi, offloads);
1996 }
1997 
1998 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1999 			   struct netlink_ext_ack *extack)
2000 {
2001 	unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2002 	struct virtnet_info *vi = netdev_priv(dev);
2003 	struct bpf_prog *old_prog;
2004 	u16 xdp_qp = 0, curr_qp;
2005 	int i, err;
2006 
2007 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2008 	    && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2009 	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2010 	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2011 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO))) {
2012 		NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first");
2013 		return -EOPNOTSUPP;
2014 	}
2015 
2016 	if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2017 		NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2018 		return -EINVAL;
2019 	}
2020 
2021 	if (dev->mtu > max_sz) {
2022 		NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2023 		netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2024 		return -EINVAL;
2025 	}
2026 
2027 	curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2028 	if (prog)
2029 		xdp_qp = nr_cpu_ids;
2030 
2031 	/* XDP requires extra queues for XDP_TX */
2032 	if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2033 		NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2034 		netdev_warn(dev, "request %i queues but max is %i\n",
2035 			    curr_qp + xdp_qp, vi->max_queue_pairs);
2036 		return -ENOMEM;
2037 	}
2038 
2039 	if (prog) {
2040 		prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
2041 		if (IS_ERR(prog))
2042 			return PTR_ERR(prog);
2043 	}
2044 
2045 	/* Make sure NAPI is not using any XDP TX queues for RX. */
2046 	for (i = 0; i < vi->max_queue_pairs; i++)
2047 		napi_disable(&vi->rq[i].napi);
2048 
2049 	netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2050 	err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2051 	if (err)
2052 		goto err;
2053 	vi->xdp_queue_pairs = xdp_qp;
2054 
2055 	for (i = 0; i < vi->max_queue_pairs; i++) {
2056 		old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2057 		rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2058 		if (i == 0) {
2059 			if (!old_prog)
2060 				virtnet_clear_guest_offloads(vi);
2061 			if (!prog)
2062 				virtnet_restore_guest_offloads(vi);
2063 		}
2064 		if (old_prog)
2065 			bpf_prog_put(old_prog);
2066 		virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2067 	}
2068 
2069 	return 0;
2070 
2071 err:
2072 	for (i = 0; i < vi->max_queue_pairs; i++)
2073 		virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2074 	if (prog)
2075 		bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2076 	return err;
2077 }
2078 
2079 static u32 virtnet_xdp_query(struct net_device *dev)
2080 {
2081 	struct virtnet_info *vi = netdev_priv(dev);
2082 	const struct bpf_prog *xdp_prog;
2083 	int i;
2084 
2085 	for (i = 0; i < vi->max_queue_pairs; i++) {
2086 		xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2087 		if (xdp_prog)
2088 			return xdp_prog->aux->id;
2089 	}
2090 	return 0;
2091 }
2092 
2093 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2094 {
2095 	switch (xdp->command) {
2096 	case XDP_SETUP_PROG:
2097 		return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2098 	case XDP_QUERY_PROG:
2099 		xdp->prog_id = virtnet_xdp_query(dev);
2100 		xdp->prog_attached = !!xdp->prog_id;
2101 		return 0;
2102 	default:
2103 		return -EINVAL;
2104 	}
2105 }
2106 
2107 static const struct net_device_ops virtnet_netdev = {
2108 	.ndo_open            = virtnet_open,
2109 	.ndo_stop   	     = virtnet_close,
2110 	.ndo_start_xmit      = start_xmit,
2111 	.ndo_validate_addr   = eth_validate_addr,
2112 	.ndo_set_mac_address = virtnet_set_mac_address,
2113 	.ndo_set_rx_mode     = virtnet_set_rx_mode,
2114 	.ndo_get_stats64     = virtnet_stats,
2115 	.ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2116 	.ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2117 #ifdef CONFIG_NET_POLL_CONTROLLER
2118 	.ndo_poll_controller = virtnet_netpoll,
2119 #endif
2120 	.ndo_bpf		= virtnet_xdp,
2121 	.ndo_xdp_xmit		= virtnet_xdp_xmit,
2122 	.ndo_xdp_flush		= virtnet_xdp_flush,
2123 	.ndo_features_check	= passthru_features_check,
2124 };
2125 
2126 static void virtnet_config_changed_work(struct work_struct *work)
2127 {
2128 	struct virtnet_info *vi =
2129 		container_of(work, struct virtnet_info, config_work);
2130 	u16 v;
2131 
2132 	if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2133 				 struct virtio_net_config, status, &v) < 0)
2134 		return;
2135 
2136 	if (v & VIRTIO_NET_S_ANNOUNCE) {
2137 		netdev_notify_peers(vi->dev);
2138 		virtnet_ack_link_announce(vi);
2139 	}
2140 
2141 	/* Ignore unknown (future) status bits */
2142 	v &= VIRTIO_NET_S_LINK_UP;
2143 
2144 	if (vi->status == v)
2145 		return;
2146 
2147 	vi->status = v;
2148 
2149 	if (vi->status & VIRTIO_NET_S_LINK_UP) {
2150 		netif_carrier_on(vi->dev);
2151 		netif_tx_wake_all_queues(vi->dev);
2152 	} else {
2153 		netif_carrier_off(vi->dev);
2154 		netif_tx_stop_all_queues(vi->dev);
2155 	}
2156 }
2157 
2158 static void virtnet_config_changed(struct virtio_device *vdev)
2159 {
2160 	struct virtnet_info *vi = vdev->priv;
2161 
2162 	schedule_work(&vi->config_work);
2163 }
2164 
2165 static void virtnet_free_queues(struct virtnet_info *vi)
2166 {
2167 	int i;
2168 
2169 	for (i = 0; i < vi->max_queue_pairs; i++) {
2170 		napi_hash_del(&vi->rq[i].napi);
2171 		netif_napi_del(&vi->rq[i].napi);
2172 		netif_napi_del(&vi->sq[i].napi);
2173 	}
2174 
2175 	/* We called napi_hash_del() before netif_napi_del(),
2176 	 * we need to respect an RCU grace period before freeing vi->rq
2177 	 */
2178 	synchronize_net();
2179 
2180 	kfree(vi->rq);
2181 	kfree(vi->sq);
2182 }
2183 
2184 static void _free_receive_bufs(struct virtnet_info *vi)
2185 {
2186 	struct bpf_prog *old_prog;
2187 	int i;
2188 
2189 	for (i = 0; i < vi->max_queue_pairs; i++) {
2190 		while (vi->rq[i].pages)
2191 			__free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2192 
2193 		old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2194 		RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2195 		if (old_prog)
2196 			bpf_prog_put(old_prog);
2197 	}
2198 }
2199 
2200 static void free_receive_bufs(struct virtnet_info *vi)
2201 {
2202 	rtnl_lock();
2203 	_free_receive_bufs(vi);
2204 	rtnl_unlock();
2205 }
2206 
2207 static void free_receive_page_frags(struct virtnet_info *vi)
2208 {
2209 	int i;
2210 	for (i = 0; i < vi->max_queue_pairs; i++)
2211 		if (vi->rq[i].alloc_frag.page)
2212 			put_page(vi->rq[i].alloc_frag.page);
2213 }
2214 
2215 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
2216 {
2217 	if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
2218 		return false;
2219 	else if (q < vi->curr_queue_pairs)
2220 		return true;
2221 	else
2222 		return false;
2223 }
2224 
2225 static void free_unused_bufs(struct virtnet_info *vi)
2226 {
2227 	void *buf;
2228 	int i;
2229 
2230 	for (i = 0; i < vi->max_queue_pairs; i++) {
2231 		struct virtqueue *vq = vi->sq[i].vq;
2232 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2233 			if (!is_xdp_raw_buffer_queue(vi, i))
2234 				dev_kfree_skb(buf);
2235 			else
2236 				put_page(virt_to_head_page(buf));
2237 		}
2238 	}
2239 
2240 	for (i = 0; i < vi->max_queue_pairs; i++) {
2241 		struct virtqueue *vq = vi->rq[i].vq;
2242 
2243 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2244 			if (vi->mergeable_rx_bufs) {
2245 				put_page(virt_to_head_page(buf));
2246 			} else if (vi->big_packets) {
2247 				give_pages(&vi->rq[i], buf);
2248 			} else {
2249 				put_page(virt_to_head_page(buf));
2250 			}
2251 		}
2252 	}
2253 }
2254 
2255 static void virtnet_del_vqs(struct virtnet_info *vi)
2256 {
2257 	struct virtio_device *vdev = vi->vdev;
2258 
2259 	virtnet_clean_affinity(vi, -1);
2260 
2261 	vdev->config->del_vqs(vdev);
2262 
2263 	virtnet_free_queues(vi);
2264 }
2265 
2266 /* How large should a single buffer be so a queue full of these can fit at
2267  * least one full packet?
2268  * Logic below assumes the mergeable buffer header is used.
2269  */
2270 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2271 {
2272 	const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2273 	unsigned int rq_size = virtqueue_get_vring_size(vq);
2274 	unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2275 	unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2276 	unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2277 
2278 	return max(max(min_buf_len, hdr_len) - hdr_len,
2279 		   (unsigned int)GOOD_PACKET_LEN);
2280 }
2281 
2282 static int virtnet_find_vqs(struct virtnet_info *vi)
2283 {
2284 	vq_callback_t **callbacks;
2285 	struct virtqueue **vqs;
2286 	int ret = -ENOMEM;
2287 	int i, total_vqs;
2288 	const char **names;
2289 	bool *ctx;
2290 
2291 	/* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2292 	 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2293 	 * possible control vq.
2294 	 */
2295 	total_vqs = vi->max_queue_pairs * 2 +
2296 		    virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2297 
2298 	/* Allocate space for find_vqs parameters */
2299 	vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
2300 	if (!vqs)
2301 		goto err_vq;
2302 	callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
2303 	if (!callbacks)
2304 		goto err_callback;
2305 	names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
2306 	if (!names)
2307 		goto err_names;
2308 	if (!vi->big_packets || vi->mergeable_rx_bufs) {
2309 		ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL);
2310 		if (!ctx)
2311 			goto err_ctx;
2312 	} else {
2313 		ctx = NULL;
2314 	}
2315 
2316 	/* Parameters for control virtqueue, if any */
2317 	if (vi->has_cvq) {
2318 		callbacks[total_vqs - 1] = NULL;
2319 		names[total_vqs - 1] = "control";
2320 	}
2321 
2322 	/* Allocate/initialize parameters for send/receive virtqueues */
2323 	for (i = 0; i < vi->max_queue_pairs; i++) {
2324 		callbacks[rxq2vq(i)] = skb_recv_done;
2325 		callbacks[txq2vq(i)] = skb_xmit_done;
2326 		sprintf(vi->rq[i].name, "input.%d", i);
2327 		sprintf(vi->sq[i].name, "output.%d", i);
2328 		names[rxq2vq(i)] = vi->rq[i].name;
2329 		names[txq2vq(i)] = vi->sq[i].name;
2330 		if (ctx)
2331 			ctx[rxq2vq(i)] = true;
2332 	}
2333 
2334 	ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2335 					 names, ctx, NULL);
2336 	if (ret)
2337 		goto err_find;
2338 
2339 	if (vi->has_cvq) {
2340 		vi->cvq = vqs[total_vqs - 1];
2341 		if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2342 			vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2343 	}
2344 
2345 	for (i = 0; i < vi->max_queue_pairs; i++) {
2346 		vi->rq[i].vq = vqs[rxq2vq(i)];
2347 		vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2348 		vi->sq[i].vq = vqs[txq2vq(i)];
2349 	}
2350 
2351 	kfree(names);
2352 	kfree(callbacks);
2353 	kfree(vqs);
2354 	kfree(ctx);
2355 
2356 	return 0;
2357 
2358 err_find:
2359 	kfree(ctx);
2360 err_ctx:
2361 	kfree(names);
2362 err_names:
2363 	kfree(callbacks);
2364 err_callback:
2365 	kfree(vqs);
2366 err_vq:
2367 	return ret;
2368 }
2369 
2370 static int virtnet_alloc_queues(struct virtnet_info *vi)
2371 {
2372 	int i;
2373 
2374 	vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2375 	if (!vi->sq)
2376 		goto err_sq;
2377 	vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2378 	if (!vi->rq)
2379 		goto err_rq;
2380 
2381 	INIT_DELAYED_WORK(&vi->refill, refill_work);
2382 	for (i = 0; i < vi->max_queue_pairs; i++) {
2383 		vi->rq[i].pages = NULL;
2384 		netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2385 			       napi_weight);
2386 		netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2387 				  napi_tx ? napi_weight : 0);
2388 
2389 		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2390 		ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2391 		sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2392 	}
2393 
2394 	return 0;
2395 
2396 err_rq:
2397 	kfree(vi->sq);
2398 err_sq:
2399 	return -ENOMEM;
2400 }
2401 
2402 static int init_vqs(struct virtnet_info *vi)
2403 {
2404 	int ret;
2405 
2406 	/* Allocate send & receive queues */
2407 	ret = virtnet_alloc_queues(vi);
2408 	if (ret)
2409 		goto err;
2410 
2411 	ret = virtnet_find_vqs(vi);
2412 	if (ret)
2413 		goto err_free;
2414 
2415 	get_online_cpus();
2416 	virtnet_set_affinity(vi);
2417 	put_online_cpus();
2418 
2419 	return 0;
2420 
2421 err_free:
2422 	virtnet_free_queues(vi);
2423 err:
2424 	return ret;
2425 }
2426 
2427 #ifdef CONFIG_SYSFS
2428 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2429 		char *buf)
2430 {
2431 	struct virtnet_info *vi = netdev_priv(queue->dev);
2432 	unsigned int queue_index = get_netdev_rx_queue_index(queue);
2433 	struct ewma_pkt_len *avg;
2434 
2435 	BUG_ON(queue_index >= vi->max_queue_pairs);
2436 	avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2437 	return sprintf(buf, "%u\n",
2438 		       get_mergeable_buf_len(&vi->rq[queue_index], avg));
2439 }
2440 
2441 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2442 	__ATTR_RO(mergeable_rx_buffer_size);
2443 
2444 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2445 	&mergeable_rx_buffer_size_attribute.attr,
2446 	NULL
2447 };
2448 
2449 static const struct attribute_group virtio_net_mrg_rx_group = {
2450 	.name = "virtio_net",
2451 	.attrs = virtio_net_mrg_rx_attrs
2452 };
2453 #endif
2454 
2455 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2456 				    unsigned int fbit,
2457 				    const char *fname, const char *dname)
2458 {
2459 	if (!virtio_has_feature(vdev, fbit))
2460 		return false;
2461 
2462 	dev_err(&vdev->dev, "device advertises feature %s but not %s",
2463 		fname, dname);
2464 
2465 	return true;
2466 }
2467 
2468 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)			\
2469 	virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2470 
2471 static bool virtnet_validate_features(struct virtio_device *vdev)
2472 {
2473 	if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2474 	    (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2475 			     "VIRTIO_NET_F_CTRL_VQ") ||
2476 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2477 			     "VIRTIO_NET_F_CTRL_VQ") ||
2478 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2479 			     "VIRTIO_NET_F_CTRL_VQ") ||
2480 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2481 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2482 			     "VIRTIO_NET_F_CTRL_VQ"))) {
2483 		return false;
2484 	}
2485 
2486 	return true;
2487 }
2488 
2489 #define MIN_MTU ETH_MIN_MTU
2490 #define MAX_MTU ETH_MAX_MTU
2491 
2492 static int virtnet_validate(struct virtio_device *vdev)
2493 {
2494 	if (!vdev->config->get) {
2495 		dev_err(&vdev->dev, "%s failure: config access disabled\n",
2496 			__func__);
2497 		return -EINVAL;
2498 	}
2499 
2500 	if (!virtnet_validate_features(vdev))
2501 		return -EINVAL;
2502 
2503 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2504 		int mtu = virtio_cread16(vdev,
2505 					 offsetof(struct virtio_net_config,
2506 						  mtu));
2507 		if (mtu < MIN_MTU)
2508 			__virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2509 	}
2510 
2511 	return 0;
2512 }
2513 
2514 static int virtnet_probe(struct virtio_device *vdev)
2515 {
2516 	int i, err;
2517 	struct net_device *dev;
2518 	struct virtnet_info *vi;
2519 	u16 max_queue_pairs;
2520 	int mtu;
2521 
2522 	/* Find if host supports multiqueue virtio_net device */
2523 	err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2524 				   struct virtio_net_config,
2525 				   max_virtqueue_pairs, &max_queue_pairs);
2526 
2527 	/* We need at least 2 queue's */
2528 	if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2529 	    max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2530 	    !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2531 		max_queue_pairs = 1;
2532 
2533 	/* Allocate ourselves a network device with room for our info */
2534 	dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2535 	if (!dev)
2536 		return -ENOMEM;
2537 
2538 	/* Set up network device as normal. */
2539 	dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2540 	dev->netdev_ops = &virtnet_netdev;
2541 	dev->features = NETIF_F_HIGHDMA;
2542 
2543 	dev->ethtool_ops = &virtnet_ethtool_ops;
2544 	SET_NETDEV_DEV(dev, &vdev->dev);
2545 
2546 	/* Do we support "hardware" checksums? */
2547 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2548 		/* This opens up the world of extra features. */
2549 		dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2550 		if (csum)
2551 			dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2552 
2553 		if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2554 			dev->hw_features |= NETIF_F_TSO
2555 				| NETIF_F_TSO_ECN | NETIF_F_TSO6;
2556 		}
2557 		/* Individual feature bits: what can host handle? */
2558 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2559 			dev->hw_features |= NETIF_F_TSO;
2560 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2561 			dev->hw_features |= NETIF_F_TSO6;
2562 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2563 			dev->hw_features |= NETIF_F_TSO_ECN;
2564 
2565 		dev->features |= NETIF_F_GSO_ROBUST;
2566 
2567 		if (gso)
2568 			dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
2569 		/* (!csum && gso) case will be fixed by register_netdev() */
2570 	}
2571 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2572 		dev->features |= NETIF_F_RXCSUM;
2573 
2574 	dev->vlan_features = dev->features;
2575 
2576 	/* MTU range: 68 - 65535 */
2577 	dev->min_mtu = MIN_MTU;
2578 	dev->max_mtu = MAX_MTU;
2579 
2580 	/* Configuration may specify what MAC to use.  Otherwise random. */
2581 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2582 		virtio_cread_bytes(vdev,
2583 				   offsetof(struct virtio_net_config, mac),
2584 				   dev->dev_addr, dev->addr_len);
2585 	else
2586 		eth_hw_addr_random(dev);
2587 
2588 	/* Set up our device-specific information */
2589 	vi = netdev_priv(dev);
2590 	vi->dev = dev;
2591 	vi->vdev = vdev;
2592 	vdev->priv = vi;
2593 	vi->stats = alloc_percpu(struct virtnet_stats);
2594 	err = -ENOMEM;
2595 	if (vi->stats == NULL)
2596 		goto free;
2597 
2598 	for_each_possible_cpu(i) {
2599 		struct virtnet_stats *virtnet_stats;
2600 		virtnet_stats = per_cpu_ptr(vi->stats, i);
2601 		u64_stats_init(&virtnet_stats->tx_syncp);
2602 		u64_stats_init(&virtnet_stats->rx_syncp);
2603 	}
2604 
2605 	INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2606 
2607 	/* If we can receive ANY GSO packets, we must allocate large ones. */
2608 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2609 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2610 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2611 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2612 		vi->big_packets = true;
2613 
2614 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2615 		vi->mergeable_rx_bufs = true;
2616 
2617 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2618 	    virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2619 		vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2620 	else
2621 		vi->hdr_len = sizeof(struct virtio_net_hdr);
2622 
2623 	if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2624 	    virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2625 		vi->any_header_sg = true;
2626 
2627 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2628 		vi->has_cvq = true;
2629 
2630 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2631 		mtu = virtio_cread16(vdev,
2632 				     offsetof(struct virtio_net_config,
2633 					      mtu));
2634 		if (mtu < dev->min_mtu) {
2635 			/* Should never trigger: MTU was previously validated
2636 			 * in virtnet_validate.
2637 			 */
2638 			dev_err(&vdev->dev, "device MTU appears to have changed "
2639 				"it is now %d < %d", mtu, dev->min_mtu);
2640 			goto free_stats;
2641 		}
2642 
2643 		dev->mtu = mtu;
2644 		dev->max_mtu = mtu;
2645 
2646 		/* TODO: size buffers correctly in this case. */
2647 		if (dev->mtu > ETH_DATA_LEN)
2648 			vi->big_packets = true;
2649 	}
2650 
2651 	if (vi->any_header_sg)
2652 		dev->needed_headroom = vi->hdr_len;
2653 
2654 	/* Enable multiqueue by default */
2655 	if (num_online_cpus() >= max_queue_pairs)
2656 		vi->curr_queue_pairs = max_queue_pairs;
2657 	else
2658 		vi->curr_queue_pairs = num_online_cpus();
2659 	vi->max_queue_pairs = max_queue_pairs;
2660 
2661 	/* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2662 	err = init_vqs(vi);
2663 	if (err)
2664 		goto free_stats;
2665 
2666 #ifdef CONFIG_SYSFS
2667 	if (vi->mergeable_rx_bufs)
2668 		dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2669 #endif
2670 	netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2671 	netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2672 
2673 	virtnet_init_settings(dev);
2674 
2675 	err = register_netdev(dev);
2676 	if (err) {
2677 		pr_debug("virtio_net: registering device failed\n");
2678 		goto free_vqs;
2679 	}
2680 
2681 	virtio_device_ready(vdev);
2682 
2683 	err = virtnet_cpu_notif_add(vi);
2684 	if (err) {
2685 		pr_debug("virtio_net: registering cpu notifier failed\n");
2686 		goto free_unregister_netdev;
2687 	}
2688 
2689 	virtnet_set_queues(vi, vi->curr_queue_pairs);
2690 
2691 	/* Assume link up if device can't report link status,
2692 	   otherwise get link status from config. */
2693 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2694 		netif_carrier_off(dev);
2695 		schedule_work(&vi->config_work);
2696 	} else {
2697 		vi->status = VIRTIO_NET_S_LINK_UP;
2698 		netif_carrier_on(dev);
2699 	}
2700 
2701 	for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
2702 		if (virtio_has_feature(vi->vdev, guest_offloads[i]))
2703 			set_bit(guest_offloads[i], &vi->guest_offloads);
2704 
2705 	pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2706 		 dev->name, max_queue_pairs);
2707 
2708 	return 0;
2709 
2710 free_unregister_netdev:
2711 	vi->vdev->config->reset(vdev);
2712 
2713 	unregister_netdev(dev);
2714 free_vqs:
2715 	cancel_delayed_work_sync(&vi->refill);
2716 	free_receive_page_frags(vi);
2717 	virtnet_del_vqs(vi);
2718 free_stats:
2719 	free_percpu(vi->stats);
2720 free:
2721 	free_netdev(dev);
2722 	return err;
2723 }
2724 
2725 static void remove_vq_common(struct virtnet_info *vi)
2726 {
2727 	vi->vdev->config->reset(vi->vdev);
2728 
2729 	/* Free unused buffers in both send and recv, if any. */
2730 	free_unused_bufs(vi);
2731 
2732 	free_receive_bufs(vi);
2733 
2734 	free_receive_page_frags(vi);
2735 
2736 	virtnet_del_vqs(vi);
2737 }
2738 
2739 static void virtnet_remove(struct virtio_device *vdev)
2740 {
2741 	struct virtnet_info *vi = vdev->priv;
2742 
2743 	virtnet_cpu_notif_remove(vi);
2744 
2745 	/* Make sure no work handler is accessing the device. */
2746 	flush_work(&vi->config_work);
2747 
2748 	unregister_netdev(vi->dev);
2749 
2750 	remove_vq_common(vi);
2751 
2752 	free_percpu(vi->stats);
2753 	free_netdev(vi->dev);
2754 }
2755 
2756 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
2757 {
2758 	struct virtnet_info *vi = vdev->priv;
2759 
2760 	virtnet_cpu_notif_remove(vi);
2761 	virtnet_freeze_down(vdev);
2762 	remove_vq_common(vi);
2763 
2764 	return 0;
2765 }
2766 
2767 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
2768 {
2769 	struct virtnet_info *vi = vdev->priv;
2770 	int err;
2771 
2772 	err = virtnet_restore_up(vdev);
2773 	if (err)
2774 		return err;
2775 	virtnet_set_queues(vi, vi->curr_queue_pairs);
2776 
2777 	err = virtnet_cpu_notif_add(vi);
2778 	if (err)
2779 		return err;
2780 
2781 	return 0;
2782 }
2783 
2784 static struct virtio_device_id id_table[] = {
2785 	{ VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2786 	{ 0 },
2787 };
2788 
2789 #define VIRTNET_FEATURES \
2790 	VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2791 	VIRTIO_NET_F_MAC, \
2792 	VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2793 	VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2794 	VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2795 	VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2796 	VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2797 	VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2798 	VIRTIO_NET_F_CTRL_MAC_ADDR, \
2799 	VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
2800 
2801 static unsigned int features[] = {
2802 	VIRTNET_FEATURES,
2803 };
2804 
2805 static unsigned int features_legacy[] = {
2806 	VIRTNET_FEATURES,
2807 	VIRTIO_NET_F_GSO,
2808 	VIRTIO_F_ANY_LAYOUT,
2809 };
2810 
2811 static struct virtio_driver virtio_net_driver = {
2812 	.feature_table = features,
2813 	.feature_table_size = ARRAY_SIZE(features),
2814 	.feature_table_legacy = features_legacy,
2815 	.feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2816 	.driver.name =	KBUILD_MODNAME,
2817 	.driver.owner =	THIS_MODULE,
2818 	.id_table =	id_table,
2819 	.validate =	virtnet_validate,
2820 	.probe =	virtnet_probe,
2821 	.remove =	virtnet_remove,
2822 	.config_changed = virtnet_config_changed,
2823 #ifdef CONFIG_PM_SLEEP
2824 	.freeze =	virtnet_freeze,
2825 	.restore =	virtnet_restore,
2826 #endif
2827 };
2828 
2829 static __init int virtio_net_driver_init(void)
2830 {
2831 	int ret;
2832 
2833 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
2834 				      virtnet_cpu_online,
2835 				      virtnet_cpu_down_prep);
2836 	if (ret < 0)
2837 		goto out;
2838 	virtionet_online = ret;
2839 	ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
2840 				      NULL, virtnet_cpu_dead);
2841 	if (ret)
2842 		goto err_dead;
2843 
2844         ret = register_virtio_driver(&virtio_net_driver);
2845 	if (ret)
2846 		goto err_virtio;
2847 	return 0;
2848 err_virtio:
2849 	cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2850 err_dead:
2851 	cpuhp_remove_multi_state(virtionet_online);
2852 out:
2853 	return ret;
2854 }
2855 module_init(virtio_net_driver_init);
2856 
2857 static __exit void virtio_net_driver_exit(void)
2858 {
2859 	unregister_virtio_driver(&virtio_net_driver);
2860 	cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2861 	cpuhp_remove_multi_state(virtionet_online);
2862 }
2863 module_exit(virtio_net_driver_exit);
2864 
2865 MODULE_DEVICE_TABLE(virtio, id_table);
2866 MODULE_DESCRIPTION("Virtio network driver");
2867 MODULE_LICENSE("GPL");
2868