xref: /openbmc/linux/drivers/net/virtio_net.c (revision a8f7d696)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* A network driver using virtio.
3  *
4  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
5  */
6 //#define DEBUG
7 #include <linux/netdevice.h>
8 #include <linux/etherdevice.h>
9 #include <linux/ethtool.h>
10 #include <linux/module.h>
11 #include <linux/virtio.h>
12 #include <linux/virtio_net.h>
13 #include <linux/bpf.h>
14 #include <linux/bpf_trace.h>
15 #include <linux/scatterlist.h>
16 #include <linux/if_vlan.h>
17 #include <linux/slab.h>
18 #include <linux/cpu.h>
19 #include <linux/average.h>
20 #include <linux/filter.h>
21 #include <linux/kernel.h>
22 #include <net/route.h>
23 #include <net/xdp.h>
24 #include <net/net_failover.h>
25 #include <net/netdev_rx_queue.h>
26 
27 static int napi_weight = NAPI_POLL_WEIGHT;
28 module_param(napi_weight, int, 0444);
29 
30 static bool csum = true, gso = true, napi_tx = true;
31 module_param(csum, bool, 0444);
32 module_param(gso, bool, 0444);
33 module_param(napi_tx, bool, 0644);
34 
35 /* FIXME: MTU in config. */
36 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
37 #define GOOD_COPY_LEN	128
38 
39 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
40 
41 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
42 #define VIRTIO_XDP_HEADROOM 256
43 
44 /* Separating two types of XDP xmit */
45 #define VIRTIO_XDP_TX		BIT(0)
46 #define VIRTIO_XDP_REDIR	BIT(1)
47 
48 #define VIRTIO_XDP_FLAG	BIT(0)
49 
50 /* RX packet size EWMA. The average packet size is used to determine the packet
51  * buffer size when refilling RX rings. As the entire RX ring may be refilled
52  * at once, the weight is chosen so that the EWMA will be insensitive to short-
53  * term, transient changes in packet size.
54  */
55 DECLARE_EWMA(pkt_len, 0, 64)
56 
57 #define VIRTNET_DRIVER_VERSION "1.0.0"
58 
59 static const unsigned long guest_offloads[] = {
60 	VIRTIO_NET_F_GUEST_TSO4,
61 	VIRTIO_NET_F_GUEST_TSO6,
62 	VIRTIO_NET_F_GUEST_ECN,
63 	VIRTIO_NET_F_GUEST_UFO,
64 	VIRTIO_NET_F_GUEST_CSUM,
65 	VIRTIO_NET_F_GUEST_USO4,
66 	VIRTIO_NET_F_GUEST_USO6,
67 	VIRTIO_NET_F_GUEST_HDRLEN
68 };
69 
70 #define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
71 				(1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
72 				(1ULL << VIRTIO_NET_F_GUEST_ECN)  | \
73 				(1ULL << VIRTIO_NET_F_GUEST_UFO)  | \
74 				(1ULL << VIRTIO_NET_F_GUEST_USO4) | \
75 				(1ULL << VIRTIO_NET_F_GUEST_USO6))
76 
77 struct virtnet_stat_desc {
78 	char desc[ETH_GSTRING_LEN];
79 	size_t offset;
80 };
81 
82 struct virtnet_sq_stats {
83 	struct u64_stats_sync syncp;
84 	u64_stats_t packets;
85 	u64_stats_t bytes;
86 	u64_stats_t xdp_tx;
87 	u64_stats_t xdp_tx_drops;
88 	u64_stats_t kicks;
89 	u64_stats_t tx_timeouts;
90 };
91 
92 struct virtnet_rq_stats {
93 	struct u64_stats_sync syncp;
94 	u64_stats_t packets;
95 	u64_stats_t bytes;
96 	u64_stats_t drops;
97 	u64_stats_t xdp_packets;
98 	u64_stats_t xdp_tx;
99 	u64_stats_t xdp_redirects;
100 	u64_stats_t xdp_drops;
101 	u64_stats_t kicks;
102 };
103 
104 #define VIRTNET_SQ_STAT(m)	offsetof(struct virtnet_sq_stats, m)
105 #define VIRTNET_RQ_STAT(m)	offsetof(struct virtnet_rq_stats, m)
106 
107 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
108 	{ "packets",		VIRTNET_SQ_STAT(packets) },
109 	{ "bytes",		VIRTNET_SQ_STAT(bytes) },
110 	{ "xdp_tx",		VIRTNET_SQ_STAT(xdp_tx) },
111 	{ "xdp_tx_drops",	VIRTNET_SQ_STAT(xdp_tx_drops) },
112 	{ "kicks",		VIRTNET_SQ_STAT(kicks) },
113 	{ "tx_timeouts",	VIRTNET_SQ_STAT(tx_timeouts) },
114 };
115 
116 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
117 	{ "packets",		VIRTNET_RQ_STAT(packets) },
118 	{ "bytes",		VIRTNET_RQ_STAT(bytes) },
119 	{ "drops",		VIRTNET_RQ_STAT(drops) },
120 	{ "xdp_packets",	VIRTNET_RQ_STAT(xdp_packets) },
121 	{ "xdp_tx",		VIRTNET_RQ_STAT(xdp_tx) },
122 	{ "xdp_redirects",	VIRTNET_RQ_STAT(xdp_redirects) },
123 	{ "xdp_drops",		VIRTNET_RQ_STAT(xdp_drops) },
124 	{ "kicks",		VIRTNET_RQ_STAT(kicks) },
125 };
126 
127 #define VIRTNET_SQ_STATS_LEN	ARRAY_SIZE(virtnet_sq_stats_desc)
128 #define VIRTNET_RQ_STATS_LEN	ARRAY_SIZE(virtnet_rq_stats_desc)
129 
130 struct virtnet_interrupt_coalesce {
131 	u32 max_packets;
132 	u32 max_usecs;
133 };
134 
135 /* The dma information of pages allocated at a time. */
136 struct virtnet_rq_dma {
137 	dma_addr_t addr;
138 	u32 ref;
139 	u16 len;
140 	u16 need_sync;
141 };
142 
143 /* Internal representation of a send virtqueue */
144 struct send_queue {
145 	/* Virtqueue associated with this send _queue */
146 	struct virtqueue *vq;
147 
148 	/* TX: fragments + linear part + virtio header */
149 	struct scatterlist sg[MAX_SKB_FRAGS + 2];
150 
151 	/* Name of the send queue: output.$index */
152 	char name[16];
153 
154 	struct virtnet_sq_stats stats;
155 
156 	struct virtnet_interrupt_coalesce intr_coal;
157 
158 	struct napi_struct napi;
159 
160 	/* Record whether sq is in reset state. */
161 	bool reset;
162 };
163 
164 /* Internal representation of a receive virtqueue */
165 struct receive_queue {
166 	/* Virtqueue associated with this receive_queue */
167 	struct virtqueue *vq;
168 
169 	struct napi_struct napi;
170 
171 	struct bpf_prog __rcu *xdp_prog;
172 
173 	struct virtnet_rq_stats stats;
174 
175 	struct virtnet_interrupt_coalesce intr_coal;
176 
177 	/* Chain pages by the private ptr. */
178 	struct page *pages;
179 
180 	/* Average packet length for mergeable receive buffers. */
181 	struct ewma_pkt_len mrg_avg_pkt_len;
182 
183 	/* Page frag for packet buffer allocation. */
184 	struct page_frag alloc_frag;
185 
186 	/* RX: fragments + linear part + virtio header */
187 	struct scatterlist sg[MAX_SKB_FRAGS + 2];
188 
189 	/* Min single buffer size for mergeable buffers case. */
190 	unsigned int min_buf_len;
191 
192 	/* Name of this receive queue: input.$index */
193 	char name[16];
194 
195 	struct xdp_rxq_info xdp_rxq;
196 
197 	/* Record the last dma info to free after new pages is allocated. */
198 	struct virtnet_rq_dma *last_dma;
199 
200 	/* Do dma by self */
201 	bool do_dma;
202 };
203 
204 /* This structure can contain rss message with maximum settings for indirection table and keysize
205  * Note, that default structure that describes RSS configuration virtio_net_rss_config
206  * contains same info but can't handle table values.
207  * In any case, structure would be passed to virtio hw through sg_buf split by parts
208  * because table sizes may be differ according to the device configuration.
209  */
210 #define VIRTIO_NET_RSS_MAX_KEY_SIZE     40
211 #define VIRTIO_NET_RSS_MAX_TABLE_LEN    128
212 struct virtio_net_ctrl_rss {
213 	u32 hash_types;
214 	u16 indirection_table_mask;
215 	u16 unclassified_queue;
216 	u16 indirection_table[VIRTIO_NET_RSS_MAX_TABLE_LEN];
217 	u16 max_tx_vq;
218 	u8 hash_key_length;
219 	u8 key[VIRTIO_NET_RSS_MAX_KEY_SIZE];
220 };
221 
222 /* Control VQ buffers: protected by the rtnl lock */
223 struct control_buf {
224 	struct virtio_net_ctrl_hdr hdr;
225 	virtio_net_ctrl_ack status;
226 	struct virtio_net_ctrl_mq mq;
227 	u8 promisc;
228 	u8 allmulti;
229 	__virtio16 vid;
230 	__virtio64 offloads;
231 	struct virtio_net_ctrl_rss rss;
232 	struct virtio_net_ctrl_coal_tx coal_tx;
233 	struct virtio_net_ctrl_coal_rx coal_rx;
234 	struct virtio_net_ctrl_coal_vq coal_vq;
235 };
236 
237 struct virtnet_info {
238 	struct virtio_device *vdev;
239 	struct virtqueue *cvq;
240 	struct net_device *dev;
241 	struct send_queue *sq;
242 	struct receive_queue *rq;
243 	unsigned int status;
244 
245 	/* Max # of queue pairs supported by the device */
246 	u16 max_queue_pairs;
247 
248 	/* # of queue pairs currently used by the driver */
249 	u16 curr_queue_pairs;
250 
251 	/* # of XDP queue pairs currently used by the driver */
252 	u16 xdp_queue_pairs;
253 
254 	/* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
255 	bool xdp_enabled;
256 
257 	/* I like... big packets and I cannot lie! */
258 	bool big_packets;
259 
260 	/* number of sg entries allocated for big packets */
261 	unsigned int big_packets_num_skbfrags;
262 
263 	/* Host will merge rx buffers for big packets (shake it! shake it!) */
264 	bool mergeable_rx_bufs;
265 
266 	/* Host supports rss and/or hash report */
267 	bool has_rss;
268 	bool has_rss_hash_report;
269 	u8 rss_key_size;
270 	u16 rss_indir_table_size;
271 	u32 rss_hash_types_supported;
272 	u32 rss_hash_types_saved;
273 
274 	/* Has control virtqueue */
275 	bool has_cvq;
276 
277 	/* Host can handle any s/g split between our header and packet data */
278 	bool any_header_sg;
279 
280 	/* Packet virtio header size */
281 	u8 hdr_len;
282 
283 	/* Work struct for delayed refilling if we run low on memory. */
284 	struct delayed_work refill;
285 
286 	/* Is delayed refill enabled? */
287 	bool refill_enabled;
288 
289 	/* The lock to synchronize the access to refill_enabled */
290 	spinlock_t refill_lock;
291 
292 	/* Work struct for config space updates */
293 	struct work_struct config_work;
294 
295 	/* Does the affinity hint is set for virtqueues? */
296 	bool affinity_hint_set;
297 
298 	/* CPU hotplug instances for online & dead */
299 	struct hlist_node node;
300 	struct hlist_node node_dead;
301 
302 	struct control_buf *ctrl;
303 
304 	/* Ethtool settings */
305 	u8 duplex;
306 	u32 speed;
307 
308 	/* Interrupt coalescing settings */
309 	struct virtnet_interrupt_coalesce intr_coal_tx;
310 	struct virtnet_interrupt_coalesce intr_coal_rx;
311 
312 	unsigned long guest_offloads;
313 	unsigned long guest_offloads_capable;
314 
315 	/* failover when STANDBY feature enabled */
316 	struct failover *failover;
317 };
318 
319 struct padded_vnet_hdr {
320 	struct virtio_net_hdr_v1_hash hdr;
321 	/*
322 	 * hdr is in a separate sg buffer, and data sg buffer shares same page
323 	 * with this header sg. This padding makes next sg 16 byte aligned
324 	 * after the header.
325 	 */
326 	char padding[12];
327 };
328 
329 struct virtio_net_common_hdr {
330 	union {
331 		struct virtio_net_hdr hdr;
332 		struct virtio_net_hdr_mrg_rxbuf	mrg_hdr;
333 		struct virtio_net_hdr_v1_hash hash_v1_hdr;
334 	};
335 };
336 
337 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf);
338 
is_xdp_frame(void * ptr)339 static bool is_xdp_frame(void *ptr)
340 {
341 	return (unsigned long)ptr & VIRTIO_XDP_FLAG;
342 }
343 
xdp_to_ptr(struct xdp_frame * ptr)344 static void *xdp_to_ptr(struct xdp_frame *ptr)
345 {
346 	return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
347 }
348 
ptr_to_xdp(void * ptr)349 static struct xdp_frame *ptr_to_xdp(void *ptr)
350 {
351 	return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
352 }
353 
354 /* Converting between virtqueue no. and kernel tx/rx queue no.
355  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
356  */
vq2txq(struct virtqueue * vq)357 static int vq2txq(struct virtqueue *vq)
358 {
359 	return (vq->index - 1) / 2;
360 }
361 
txq2vq(int txq)362 static int txq2vq(int txq)
363 {
364 	return txq * 2 + 1;
365 }
366 
vq2rxq(struct virtqueue * vq)367 static int vq2rxq(struct virtqueue *vq)
368 {
369 	return vq->index / 2;
370 }
371 
rxq2vq(int rxq)372 static int rxq2vq(int rxq)
373 {
374 	return rxq * 2;
375 }
376 
377 static inline struct virtio_net_common_hdr *
skb_vnet_common_hdr(struct sk_buff * skb)378 skb_vnet_common_hdr(struct sk_buff *skb)
379 {
380 	return (struct virtio_net_common_hdr *)skb->cb;
381 }
382 
383 /*
384  * private is used to chain pages for big packets, put the whole
385  * most recent used list in the beginning for reuse
386  */
give_pages(struct receive_queue * rq,struct page * page)387 static void give_pages(struct receive_queue *rq, struct page *page)
388 {
389 	struct page *end;
390 
391 	/* Find end of list, sew whole thing into vi->rq.pages. */
392 	for (end = page; end->private; end = (struct page *)end->private);
393 	end->private = (unsigned long)rq->pages;
394 	rq->pages = page;
395 }
396 
get_a_page(struct receive_queue * rq,gfp_t gfp_mask)397 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
398 {
399 	struct page *p = rq->pages;
400 
401 	if (p) {
402 		rq->pages = (struct page *)p->private;
403 		/* clear private here, it is used to chain pages */
404 		p->private = 0;
405 	} else
406 		p = alloc_page(gfp_mask);
407 	return p;
408 }
409 
virtnet_rq_free_buf(struct virtnet_info * vi,struct receive_queue * rq,void * buf)410 static void virtnet_rq_free_buf(struct virtnet_info *vi,
411 				struct receive_queue *rq, void *buf)
412 {
413 	if (vi->mergeable_rx_bufs)
414 		put_page(virt_to_head_page(buf));
415 	else if (vi->big_packets)
416 		give_pages(rq, buf);
417 	else
418 		put_page(virt_to_head_page(buf));
419 }
420 
enable_delayed_refill(struct virtnet_info * vi)421 static void enable_delayed_refill(struct virtnet_info *vi)
422 {
423 	spin_lock_bh(&vi->refill_lock);
424 	vi->refill_enabled = true;
425 	spin_unlock_bh(&vi->refill_lock);
426 }
427 
disable_delayed_refill(struct virtnet_info * vi)428 static void disable_delayed_refill(struct virtnet_info *vi)
429 {
430 	spin_lock_bh(&vi->refill_lock);
431 	vi->refill_enabled = false;
432 	spin_unlock_bh(&vi->refill_lock);
433 }
434 
virtqueue_napi_schedule(struct napi_struct * napi,struct virtqueue * vq)435 static void virtqueue_napi_schedule(struct napi_struct *napi,
436 				    struct virtqueue *vq)
437 {
438 	if (napi_schedule_prep(napi)) {
439 		virtqueue_disable_cb(vq);
440 		__napi_schedule(napi);
441 	}
442 }
443 
virtqueue_napi_complete(struct napi_struct * napi,struct virtqueue * vq,int processed)444 static void virtqueue_napi_complete(struct napi_struct *napi,
445 				    struct virtqueue *vq, int processed)
446 {
447 	int opaque;
448 
449 	opaque = virtqueue_enable_cb_prepare(vq);
450 	if (napi_complete_done(napi, processed)) {
451 		if (unlikely(virtqueue_poll(vq, opaque)))
452 			virtqueue_napi_schedule(napi, vq);
453 	} else {
454 		virtqueue_disable_cb(vq);
455 	}
456 }
457 
skb_xmit_done(struct virtqueue * vq)458 static void skb_xmit_done(struct virtqueue *vq)
459 {
460 	struct virtnet_info *vi = vq->vdev->priv;
461 	struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
462 
463 	/* Suppress further interrupts. */
464 	virtqueue_disable_cb(vq);
465 
466 	if (napi->weight)
467 		virtqueue_napi_schedule(napi, vq);
468 	else
469 		/* We were probably waiting for more output buffers. */
470 		netif_wake_subqueue(vi->dev, vq2txq(vq));
471 }
472 
473 #define MRG_CTX_HEADER_SHIFT 22
mergeable_len_to_ctx(unsigned int truesize,unsigned int headroom)474 static void *mergeable_len_to_ctx(unsigned int truesize,
475 				  unsigned int headroom)
476 {
477 	return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
478 }
479 
mergeable_ctx_to_headroom(void * mrg_ctx)480 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
481 {
482 	return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
483 }
484 
mergeable_ctx_to_truesize(void * mrg_ctx)485 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
486 {
487 	return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
488 }
489 
virtnet_build_skb(void * buf,unsigned int buflen,unsigned int headroom,unsigned int len)490 static struct sk_buff *virtnet_build_skb(void *buf, unsigned int buflen,
491 					 unsigned int headroom,
492 					 unsigned int len)
493 {
494 	struct sk_buff *skb;
495 
496 	skb = build_skb(buf, buflen);
497 	if (unlikely(!skb))
498 		return NULL;
499 
500 	skb_reserve(skb, headroom);
501 	skb_put(skb, len);
502 
503 	return skb;
504 }
505 
506 /* Called from bottom half context */
page_to_skb(struct virtnet_info * vi,struct receive_queue * rq,struct page * page,unsigned int offset,unsigned int len,unsigned int truesize,unsigned int headroom)507 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
508 				   struct receive_queue *rq,
509 				   struct page *page, unsigned int offset,
510 				   unsigned int len, unsigned int truesize,
511 				   unsigned int headroom)
512 {
513 	struct sk_buff *skb;
514 	struct virtio_net_common_hdr *hdr;
515 	unsigned int copy, hdr_len, hdr_padded_len;
516 	struct page *page_to_free = NULL;
517 	int tailroom, shinfo_size;
518 	char *p, *hdr_p, *buf;
519 
520 	p = page_address(page) + offset;
521 	hdr_p = p;
522 
523 	hdr_len = vi->hdr_len;
524 	if (vi->mergeable_rx_bufs)
525 		hdr_padded_len = hdr_len;
526 	else
527 		hdr_padded_len = sizeof(struct padded_vnet_hdr);
528 
529 	buf = p - headroom;
530 	len -= hdr_len;
531 	offset += hdr_padded_len;
532 	p += hdr_padded_len;
533 	tailroom = truesize - headroom  - hdr_padded_len - len;
534 
535 	shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
536 
537 	/* copy small packet so we can reuse these pages */
538 	if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
539 		skb = virtnet_build_skb(buf, truesize, p - buf, len);
540 		if (unlikely(!skb))
541 			return NULL;
542 
543 		page = (struct page *)page->private;
544 		if (page)
545 			give_pages(rq, page);
546 		goto ok;
547 	}
548 
549 	/* copy small packet so we can reuse these pages for small data */
550 	skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
551 	if (unlikely(!skb))
552 		return NULL;
553 
554 	/* Copy all frame if it fits skb->head, otherwise
555 	 * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
556 	 */
557 	if (len <= skb_tailroom(skb))
558 		copy = len;
559 	else
560 		copy = ETH_HLEN;
561 	skb_put_data(skb, p, copy);
562 
563 	len -= copy;
564 	offset += copy;
565 
566 	if (vi->mergeable_rx_bufs) {
567 		if (len)
568 			skb_add_rx_frag(skb, 0, page, offset, len, truesize);
569 		else
570 			page_to_free = page;
571 		goto ok;
572 	}
573 
574 	/*
575 	 * Verify that we can indeed put this data into a skb.
576 	 * This is here to handle cases when the device erroneously
577 	 * tries to receive more than is possible. This is usually
578 	 * the case of a broken device.
579 	 */
580 	if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
581 		net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
582 		dev_kfree_skb(skb);
583 		return NULL;
584 	}
585 	BUG_ON(offset >= PAGE_SIZE);
586 	while (len) {
587 		unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
588 		skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
589 				frag_size, truesize);
590 		len -= frag_size;
591 		page = (struct page *)page->private;
592 		offset = 0;
593 	}
594 
595 	if (page)
596 		give_pages(rq, page);
597 
598 ok:
599 	hdr = skb_vnet_common_hdr(skb);
600 	memcpy(hdr, hdr_p, hdr_len);
601 	if (page_to_free)
602 		put_page(page_to_free);
603 
604 	return skb;
605 }
606 
virtnet_rq_unmap(struct receive_queue * rq,void * buf,u32 len)607 static void virtnet_rq_unmap(struct receive_queue *rq, void *buf, u32 len)
608 {
609 	struct page *page = virt_to_head_page(buf);
610 	struct virtnet_rq_dma *dma;
611 	void *head;
612 	int offset;
613 
614 	head = page_address(page);
615 
616 	dma = head;
617 
618 	--dma->ref;
619 
620 	if (dma->need_sync && len) {
621 		offset = buf - (head + sizeof(*dma));
622 
623 		virtqueue_dma_sync_single_range_for_cpu(rq->vq, dma->addr,
624 							offset, len,
625 							DMA_FROM_DEVICE);
626 	}
627 
628 	if (dma->ref)
629 		return;
630 
631 	virtqueue_dma_unmap_single_attrs(rq->vq, dma->addr, dma->len,
632 					 DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
633 	put_page(page);
634 }
635 
virtnet_rq_get_buf(struct receive_queue * rq,u32 * len,void ** ctx)636 static void *virtnet_rq_get_buf(struct receive_queue *rq, u32 *len, void **ctx)
637 {
638 	void *buf;
639 
640 	buf = virtqueue_get_buf_ctx(rq->vq, len, ctx);
641 	if (buf && rq->do_dma)
642 		virtnet_rq_unmap(rq, buf, *len);
643 
644 	return buf;
645 }
646 
virtnet_rq_init_one_sg(struct receive_queue * rq,void * buf,u32 len)647 static void virtnet_rq_init_one_sg(struct receive_queue *rq, void *buf, u32 len)
648 {
649 	struct virtnet_rq_dma *dma;
650 	dma_addr_t addr;
651 	u32 offset;
652 	void *head;
653 
654 	if (!rq->do_dma) {
655 		sg_init_one(rq->sg, buf, len);
656 		return;
657 	}
658 
659 	head = page_address(rq->alloc_frag.page);
660 
661 	offset = buf - head;
662 
663 	dma = head;
664 
665 	addr = dma->addr - sizeof(*dma) + offset;
666 
667 	sg_init_table(rq->sg, 1);
668 	rq->sg[0].dma_address = addr;
669 	rq->sg[0].length = len;
670 }
671 
virtnet_rq_alloc(struct receive_queue * rq,u32 size,gfp_t gfp)672 static void *virtnet_rq_alloc(struct receive_queue *rq, u32 size, gfp_t gfp)
673 {
674 	struct page_frag *alloc_frag = &rq->alloc_frag;
675 	struct virtnet_rq_dma *dma;
676 	void *buf, *head;
677 	dma_addr_t addr;
678 
679 	head = page_address(alloc_frag->page);
680 
681 	if (rq->do_dma) {
682 		dma = head;
683 
684 		/* new pages */
685 		if (!alloc_frag->offset) {
686 			if (rq->last_dma) {
687 				/* Now, the new page is allocated, the last dma
688 				 * will not be used. So the dma can be unmapped
689 				 * if the ref is 0.
690 				 */
691 				virtnet_rq_unmap(rq, rq->last_dma, 0);
692 				rq->last_dma = NULL;
693 			}
694 
695 			dma->len = alloc_frag->size - sizeof(*dma);
696 
697 			addr = virtqueue_dma_map_single_attrs(rq->vq, dma + 1,
698 							      dma->len, DMA_FROM_DEVICE, 0);
699 			if (virtqueue_dma_mapping_error(rq->vq, addr))
700 				return NULL;
701 
702 			dma->addr = addr;
703 			dma->need_sync = virtqueue_dma_need_sync(rq->vq, addr);
704 
705 			/* Add a reference to dma to prevent the entire dma from
706 			 * being released during error handling. This reference
707 			 * will be freed after the pages are no longer used.
708 			 */
709 			get_page(alloc_frag->page);
710 			dma->ref = 1;
711 			alloc_frag->offset = sizeof(*dma);
712 
713 			rq->last_dma = dma;
714 		}
715 
716 		++dma->ref;
717 	}
718 
719 	buf = head + alloc_frag->offset;
720 
721 	get_page(alloc_frag->page);
722 	alloc_frag->offset += size;
723 
724 	return buf;
725 }
726 
virtnet_rq_set_premapped(struct virtnet_info * vi)727 static void virtnet_rq_set_premapped(struct virtnet_info *vi)
728 {
729 	int i;
730 
731 	/* disable for big mode */
732 	if (!vi->mergeable_rx_bufs && vi->big_packets)
733 		return;
734 
735 	for (i = 0; i < vi->max_queue_pairs; i++) {
736 		if (virtqueue_set_dma_premapped(vi->rq[i].vq))
737 			continue;
738 
739 		vi->rq[i].do_dma = true;
740 	}
741 }
742 
virtnet_rq_unmap_free_buf(struct virtqueue * vq,void * buf)743 static void virtnet_rq_unmap_free_buf(struct virtqueue *vq, void *buf)
744 {
745 	struct virtnet_info *vi = vq->vdev->priv;
746 	struct receive_queue *rq;
747 	int i = vq2rxq(vq);
748 
749 	rq = &vi->rq[i];
750 
751 	if (rq->do_dma)
752 		virtnet_rq_unmap(rq, buf, 0);
753 
754 	virtnet_rq_free_buf(vi, rq, buf);
755 }
756 
free_old_xmit_skbs(struct send_queue * sq,bool in_napi)757 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
758 {
759 	unsigned int len;
760 	unsigned int packets = 0;
761 	unsigned int bytes = 0;
762 	void *ptr;
763 
764 	while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
765 		if (likely(!is_xdp_frame(ptr))) {
766 			struct sk_buff *skb = ptr;
767 
768 			pr_debug("Sent skb %p\n", skb);
769 
770 			bytes += skb->len;
771 			napi_consume_skb(skb, in_napi);
772 		} else {
773 			struct xdp_frame *frame = ptr_to_xdp(ptr);
774 
775 			bytes += xdp_get_frame_len(frame);
776 			xdp_return_frame(frame);
777 		}
778 		packets++;
779 	}
780 
781 	/* Avoid overhead when no packets have been processed
782 	 * happens when called speculatively from start_xmit.
783 	 */
784 	if (!packets)
785 		return;
786 
787 	u64_stats_update_begin(&sq->stats.syncp);
788 	u64_stats_add(&sq->stats.bytes, bytes);
789 	u64_stats_add(&sq->stats.packets, packets);
790 	u64_stats_update_end(&sq->stats.syncp);
791 }
792 
is_xdp_raw_buffer_queue(struct virtnet_info * vi,int q)793 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
794 {
795 	if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
796 		return false;
797 	else if (q < vi->curr_queue_pairs)
798 		return true;
799 	else
800 		return false;
801 }
802 
check_sq_full_and_disable(struct virtnet_info * vi,struct net_device * dev,struct send_queue * sq)803 static void check_sq_full_and_disable(struct virtnet_info *vi,
804 				      struct net_device *dev,
805 				      struct send_queue *sq)
806 {
807 	bool use_napi = sq->napi.weight;
808 	int qnum;
809 
810 	qnum = sq - vi->sq;
811 
812 	/* If running out of space, stop queue to avoid getting packets that we
813 	 * are then unable to transmit.
814 	 * An alternative would be to force queuing layer to requeue the skb by
815 	 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
816 	 * returned in a normal path of operation: it means that driver is not
817 	 * maintaining the TX queue stop/start state properly, and causes
818 	 * the stack to do a non-trivial amount of useless work.
819 	 * Since most packets only take 1 or 2 ring slots, stopping the queue
820 	 * early means 16 slots are typically wasted.
821 	 */
822 	if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
823 		netif_stop_subqueue(dev, qnum);
824 		if (use_napi) {
825 			if (unlikely(!virtqueue_enable_cb_delayed(sq->vq)))
826 				virtqueue_napi_schedule(&sq->napi, sq->vq);
827 		} else if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
828 			/* More just got used, free them then recheck. */
829 			free_old_xmit_skbs(sq, false);
830 			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
831 				netif_start_subqueue(dev, qnum);
832 				virtqueue_disable_cb(sq->vq);
833 			}
834 		}
835 	}
836 }
837 
__virtnet_xdp_xmit_one(struct virtnet_info * vi,struct send_queue * sq,struct xdp_frame * xdpf)838 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
839 				   struct send_queue *sq,
840 				   struct xdp_frame *xdpf)
841 {
842 	struct virtio_net_hdr_mrg_rxbuf *hdr;
843 	struct skb_shared_info *shinfo;
844 	u8 nr_frags = 0;
845 	int err, i;
846 
847 	if (unlikely(xdpf->headroom < vi->hdr_len))
848 		return -EOVERFLOW;
849 
850 	if (unlikely(xdp_frame_has_frags(xdpf))) {
851 		shinfo = xdp_get_shared_info_from_frame(xdpf);
852 		nr_frags = shinfo->nr_frags;
853 	}
854 
855 	/* In wrapping function virtnet_xdp_xmit(), we need to free
856 	 * up the pending old buffers, where we need to calculate the
857 	 * position of skb_shared_info in xdp_get_frame_len() and
858 	 * xdp_return_frame(), which will involve to xdpf->data and
859 	 * xdpf->headroom. Therefore, we need to update the value of
860 	 * headroom synchronously here.
861 	 */
862 	xdpf->headroom -= vi->hdr_len;
863 	xdpf->data -= vi->hdr_len;
864 	/* Zero header and leave csum up to XDP layers */
865 	hdr = xdpf->data;
866 	memset(hdr, 0, vi->hdr_len);
867 	xdpf->len   += vi->hdr_len;
868 
869 	sg_init_table(sq->sg, nr_frags + 1);
870 	sg_set_buf(sq->sg, xdpf->data, xdpf->len);
871 	for (i = 0; i < nr_frags; i++) {
872 		skb_frag_t *frag = &shinfo->frags[i];
873 
874 		sg_set_page(&sq->sg[i + 1], skb_frag_page(frag),
875 			    skb_frag_size(frag), skb_frag_off(frag));
876 	}
877 
878 	err = virtqueue_add_outbuf(sq->vq, sq->sg, nr_frags + 1,
879 				   xdp_to_ptr(xdpf), GFP_ATOMIC);
880 	if (unlikely(err))
881 		return -ENOSPC; /* Caller handle free/refcnt */
882 
883 	return 0;
884 }
885 
886 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
887  * the current cpu, so it does not need to be locked.
888  *
889  * Here we use marco instead of inline functions because we have to deal with
890  * three issues at the same time: 1. the choice of sq. 2. judge and execute the
891  * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
892  * functions to perfectly solve these three problems at the same time.
893  */
894 #define virtnet_xdp_get_sq(vi) ({                                       \
895 	int cpu = smp_processor_id();                                   \
896 	struct netdev_queue *txq;                                       \
897 	typeof(vi) v = (vi);                                            \
898 	unsigned int qp;                                                \
899 									\
900 	if (v->curr_queue_pairs > nr_cpu_ids) {                         \
901 		qp = v->curr_queue_pairs - v->xdp_queue_pairs;          \
902 		qp += cpu;                                              \
903 		txq = netdev_get_tx_queue(v->dev, qp);                  \
904 		__netif_tx_acquire(txq);                                \
905 	} else {                                                        \
906 		qp = cpu % v->curr_queue_pairs;                         \
907 		txq = netdev_get_tx_queue(v->dev, qp);                  \
908 		__netif_tx_lock(txq, cpu);                              \
909 	}                                                               \
910 	v->sq + qp;                                                     \
911 })
912 
913 #define virtnet_xdp_put_sq(vi, q) {                                     \
914 	struct netdev_queue *txq;                                       \
915 	typeof(vi) v = (vi);                                            \
916 									\
917 	txq = netdev_get_tx_queue(v->dev, (q) - v->sq);                 \
918 	if (v->curr_queue_pairs > nr_cpu_ids)                           \
919 		__netif_tx_release(txq);                                \
920 	else                                                            \
921 		__netif_tx_unlock(txq);                                 \
922 }
923 
virtnet_xdp_xmit(struct net_device * dev,int n,struct xdp_frame ** frames,u32 flags)924 static int virtnet_xdp_xmit(struct net_device *dev,
925 			    int n, struct xdp_frame **frames, u32 flags)
926 {
927 	struct virtnet_info *vi = netdev_priv(dev);
928 	struct receive_queue *rq = vi->rq;
929 	struct bpf_prog *xdp_prog;
930 	struct send_queue *sq;
931 	unsigned int len;
932 	int packets = 0;
933 	int bytes = 0;
934 	int nxmit = 0;
935 	int kicks = 0;
936 	void *ptr;
937 	int ret;
938 	int i;
939 
940 	/* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
941 	 * indicate XDP resources have been successfully allocated.
942 	 */
943 	xdp_prog = rcu_access_pointer(rq->xdp_prog);
944 	if (!xdp_prog)
945 		return -ENXIO;
946 
947 	sq = virtnet_xdp_get_sq(vi);
948 
949 	if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
950 		ret = -EINVAL;
951 		goto out;
952 	}
953 
954 	/* Free up any pending old buffers before queueing new ones. */
955 	while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
956 		if (likely(is_xdp_frame(ptr))) {
957 			struct xdp_frame *frame = ptr_to_xdp(ptr);
958 
959 			bytes += xdp_get_frame_len(frame);
960 			xdp_return_frame(frame);
961 		} else {
962 			struct sk_buff *skb = ptr;
963 
964 			bytes += skb->len;
965 			napi_consume_skb(skb, false);
966 		}
967 		packets++;
968 	}
969 
970 	for (i = 0; i < n; i++) {
971 		struct xdp_frame *xdpf = frames[i];
972 
973 		if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
974 			break;
975 		nxmit++;
976 	}
977 	ret = nxmit;
978 
979 	if (!is_xdp_raw_buffer_queue(vi, sq - vi->sq))
980 		check_sq_full_and_disable(vi, dev, sq);
981 
982 	if (flags & XDP_XMIT_FLUSH) {
983 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
984 			kicks = 1;
985 	}
986 out:
987 	u64_stats_update_begin(&sq->stats.syncp);
988 	u64_stats_add(&sq->stats.bytes, bytes);
989 	u64_stats_add(&sq->stats.packets, packets);
990 	u64_stats_add(&sq->stats.xdp_tx, n);
991 	u64_stats_add(&sq->stats.xdp_tx_drops, n - nxmit);
992 	u64_stats_add(&sq->stats.kicks, kicks);
993 	u64_stats_update_end(&sq->stats.syncp);
994 
995 	virtnet_xdp_put_sq(vi, sq);
996 	return ret;
997 }
998 
put_xdp_frags(struct xdp_buff * xdp)999 static void put_xdp_frags(struct xdp_buff *xdp)
1000 {
1001 	struct skb_shared_info *shinfo;
1002 	struct page *xdp_page;
1003 	int i;
1004 
1005 	if (xdp_buff_has_frags(xdp)) {
1006 		shinfo = xdp_get_shared_info_from_buff(xdp);
1007 		for (i = 0; i < shinfo->nr_frags; i++) {
1008 			xdp_page = skb_frag_page(&shinfo->frags[i]);
1009 			put_page(xdp_page);
1010 		}
1011 	}
1012 }
1013 
virtnet_xdp_handler(struct bpf_prog * xdp_prog,struct xdp_buff * xdp,struct net_device * dev,unsigned int * xdp_xmit,struct virtnet_rq_stats * stats)1014 static int virtnet_xdp_handler(struct bpf_prog *xdp_prog, struct xdp_buff *xdp,
1015 			       struct net_device *dev,
1016 			       unsigned int *xdp_xmit,
1017 			       struct virtnet_rq_stats *stats)
1018 {
1019 	struct xdp_frame *xdpf;
1020 	int err;
1021 	u32 act;
1022 
1023 	act = bpf_prog_run_xdp(xdp_prog, xdp);
1024 	u64_stats_inc(&stats->xdp_packets);
1025 
1026 	switch (act) {
1027 	case XDP_PASS:
1028 		return act;
1029 
1030 	case XDP_TX:
1031 		u64_stats_inc(&stats->xdp_tx);
1032 		xdpf = xdp_convert_buff_to_frame(xdp);
1033 		if (unlikely(!xdpf)) {
1034 			netdev_dbg(dev, "convert buff to frame failed for xdp\n");
1035 			return XDP_DROP;
1036 		}
1037 
1038 		err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
1039 		if (unlikely(!err)) {
1040 			xdp_return_frame_rx_napi(xdpf);
1041 		} else if (unlikely(err < 0)) {
1042 			trace_xdp_exception(dev, xdp_prog, act);
1043 			return XDP_DROP;
1044 		}
1045 		*xdp_xmit |= VIRTIO_XDP_TX;
1046 		return act;
1047 
1048 	case XDP_REDIRECT:
1049 		u64_stats_inc(&stats->xdp_redirects);
1050 		err = xdp_do_redirect(dev, xdp, xdp_prog);
1051 		if (err)
1052 			return XDP_DROP;
1053 
1054 		*xdp_xmit |= VIRTIO_XDP_REDIR;
1055 		return act;
1056 
1057 	default:
1058 		bpf_warn_invalid_xdp_action(dev, xdp_prog, act);
1059 		fallthrough;
1060 	case XDP_ABORTED:
1061 		trace_xdp_exception(dev, xdp_prog, act);
1062 		fallthrough;
1063 	case XDP_DROP:
1064 		return XDP_DROP;
1065 	}
1066 }
1067 
virtnet_get_headroom(struct virtnet_info * vi)1068 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
1069 {
1070 	return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
1071 }
1072 
1073 /* We copy the packet for XDP in the following cases:
1074  *
1075  * 1) Packet is scattered across multiple rx buffers.
1076  * 2) Headroom space is insufficient.
1077  *
1078  * This is inefficient but it's a temporary condition that
1079  * we hit right after XDP is enabled and until queue is refilled
1080  * with large buffers with sufficient headroom - so it should affect
1081  * at most queue size packets.
1082  * Afterwards, the conditions to enable
1083  * XDP should preclude the underlying device from sending packets
1084  * across multiple buffers (num_buf > 1), and we make sure buffers
1085  * have enough headroom.
1086  */
xdp_linearize_page(struct receive_queue * rq,int * num_buf,struct page * p,int offset,int page_off,unsigned int * len)1087 static struct page *xdp_linearize_page(struct receive_queue *rq,
1088 				       int *num_buf,
1089 				       struct page *p,
1090 				       int offset,
1091 				       int page_off,
1092 				       unsigned int *len)
1093 {
1094 	int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1095 	struct page *page;
1096 
1097 	if (page_off + *len + tailroom > PAGE_SIZE)
1098 		return NULL;
1099 
1100 	page = alloc_page(GFP_ATOMIC);
1101 	if (!page)
1102 		return NULL;
1103 
1104 	memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
1105 	page_off += *len;
1106 
1107 	while (--*num_buf) {
1108 		unsigned int buflen;
1109 		void *buf;
1110 		int off;
1111 
1112 		buf = virtnet_rq_get_buf(rq, &buflen, NULL);
1113 		if (unlikely(!buf))
1114 			goto err_buf;
1115 
1116 		p = virt_to_head_page(buf);
1117 		off = buf - page_address(p);
1118 
1119 		/* guard against a misconfigured or uncooperative backend that
1120 		 * is sending packet larger than the MTU.
1121 		 */
1122 		if ((page_off + buflen + tailroom) > PAGE_SIZE) {
1123 			put_page(p);
1124 			goto err_buf;
1125 		}
1126 
1127 		memcpy(page_address(page) + page_off,
1128 		       page_address(p) + off, buflen);
1129 		page_off += buflen;
1130 		put_page(p);
1131 	}
1132 
1133 	/* Headroom does not contribute to packet length */
1134 	*len = page_off - VIRTIO_XDP_HEADROOM;
1135 	return page;
1136 err_buf:
1137 	__free_pages(page, 0);
1138 	return NULL;
1139 }
1140 
receive_small_build_skb(struct virtnet_info * vi,unsigned int xdp_headroom,void * buf,unsigned int len)1141 static struct sk_buff *receive_small_build_skb(struct virtnet_info *vi,
1142 					       unsigned int xdp_headroom,
1143 					       void *buf,
1144 					       unsigned int len)
1145 {
1146 	unsigned int header_offset;
1147 	unsigned int headroom;
1148 	unsigned int buflen;
1149 	struct sk_buff *skb;
1150 
1151 	header_offset = VIRTNET_RX_PAD + xdp_headroom;
1152 	headroom = vi->hdr_len + header_offset;
1153 	buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
1154 		SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1155 
1156 	skb = virtnet_build_skb(buf, buflen, headroom, len);
1157 	if (unlikely(!skb))
1158 		return NULL;
1159 
1160 	buf += header_offset;
1161 	memcpy(skb_vnet_common_hdr(skb), buf, vi->hdr_len);
1162 
1163 	return skb;
1164 }
1165 
receive_small_xdp(struct net_device * dev,struct virtnet_info * vi,struct receive_queue * rq,struct bpf_prog * xdp_prog,void * buf,unsigned int xdp_headroom,unsigned int len,unsigned int * xdp_xmit,struct virtnet_rq_stats * stats)1166 static struct sk_buff *receive_small_xdp(struct net_device *dev,
1167 					 struct virtnet_info *vi,
1168 					 struct receive_queue *rq,
1169 					 struct bpf_prog *xdp_prog,
1170 					 void *buf,
1171 					 unsigned int xdp_headroom,
1172 					 unsigned int len,
1173 					 unsigned int *xdp_xmit,
1174 					 struct virtnet_rq_stats *stats)
1175 {
1176 	unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
1177 	unsigned int headroom = vi->hdr_len + header_offset;
1178 	struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
1179 	struct page *page = virt_to_head_page(buf);
1180 	struct page *xdp_page;
1181 	unsigned int buflen;
1182 	struct xdp_buff xdp;
1183 	struct sk_buff *skb;
1184 	unsigned int metasize = 0;
1185 	u32 act;
1186 
1187 	if (unlikely(hdr->hdr.gso_type))
1188 		goto err_xdp;
1189 
1190 	/* Partially checksummed packets must be dropped. */
1191 	if (unlikely(hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM))
1192 		goto err_xdp;
1193 
1194 	buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
1195 		SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1196 
1197 	if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
1198 		int offset = buf - page_address(page) + header_offset;
1199 		unsigned int tlen = len + vi->hdr_len;
1200 		int num_buf = 1;
1201 
1202 		xdp_headroom = virtnet_get_headroom(vi);
1203 		header_offset = VIRTNET_RX_PAD + xdp_headroom;
1204 		headroom = vi->hdr_len + header_offset;
1205 		buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
1206 			SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1207 		xdp_page = xdp_linearize_page(rq, &num_buf, page,
1208 					      offset, header_offset,
1209 					      &tlen);
1210 		if (!xdp_page)
1211 			goto err_xdp;
1212 
1213 		buf = page_address(xdp_page);
1214 		put_page(page);
1215 		page = xdp_page;
1216 	}
1217 
1218 	xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
1219 	xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
1220 			 xdp_headroom, len, true);
1221 
1222 	act = virtnet_xdp_handler(xdp_prog, &xdp, dev, xdp_xmit, stats);
1223 
1224 	switch (act) {
1225 	case XDP_PASS:
1226 		/* Recalculate length in case bpf program changed it */
1227 		len = xdp.data_end - xdp.data;
1228 		metasize = xdp.data - xdp.data_meta;
1229 		break;
1230 
1231 	case XDP_TX:
1232 	case XDP_REDIRECT:
1233 		goto xdp_xmit;
1234 
1235 	default:
1236 		goto err_xdp;
1237 	}
1238 
1239 	skb = virtnet_build_skb(buf, buflen, xdp.data - buf, len);
1240 	if (unlikely(!skb))
1241 		goto err;
1242 
1243 	if (metasize)
1244 		skb_metadata_set(skb, metasize);
1245 
1246 	return skb;
1247 
1248 err_xdp:
1249 	u64_stats_inc(&stats->xdp_drops);
1250 err:
1251 	u64_stats_inc(&stats->drops);
1252 	put_page(page);
1253 xdp_xmit:
1254 	return NULL;
1255 }
1256 
receive_small(struct net_device * dev,struct virtnet_info * vi,struct receive_queue * rq,void * buf,void * ctx,unsigned int len,unsigned int * xdp_xmit,struct virtnet_rq_stats * stats)1257 static struct sk_buff *receive_small(struct net_device *dev,
1258 				     struct virtnet_info *vi,
1259 				     struct receive_queue *rq,
1260 				     void *buf, void *ctx,
1261 				     unsigned int len,
1262 				     unsigned int *xdp_xmit,
1263 				     struct virtnet_rq_stats *stats)
1264 {
1265 	unsigned int xdp_headroom = (unsigned long)ctx;
1266 	struct page *page = virt_to_head_page(buf);
1267 	struct sk_buff *skb;
1268 
1269 	/* We passed the address of virtnet header to virtio-core,
1270 	 * so truncate the padding.
1271 	 */
1272 	buf -= VIRTNET_RX_PAD + xdp_headroom;
1273 
1274 	len -= vi->hdr_len;
1275 	u64_stats_add(&stats->bytes, len);
1276 
1277 	if (unlikely(len > GOOD_PACKET_LEN)) {
1278 		pr_debug("%s: rx error: len %u exceeds max size %d\n",
1279 			 dev->name, len, GOOD_PACKET_LEN);
1280 		DEV_STATS_INC(dev, rx_length_errors);
1281 		goto err;
1282 	}
1283 
1284 	if (unlikely(vi->xdp_enabled)) {
1285 		struct bpf_prog *xdp_prog;
1286 
1287 		rcu_read_lock();
1288 		xdp_prog = rcu_dereference(rq->xdp_prog);
1289 		if (xdp_prog) {
1290 			skb = receive_small_xdp(dev, vi, rq, xdp_prog, buf,
1291 						xdp_headroom, len, xdp_xmit,
1292 						stats);
1293 			rcu_read_unlock();
1294 			return skb;
1295 		}
1296 		rcu_read_unlock();
1297 	}
1298 
1299 	skb = receive_small_build_skb(vi, xdp_headroom, buf, len);
1300 	if (likely(skb))
1301 		return skb;
1302 
1303 err:
1304 	u64_stats_inc(&stats->drops);
1305 	put_page(page);
1306 	return NULL;
1307 }
1308 
receive_big(struct net_device * dev,struct virtnet_info * vi,struct receive_queue * rq,void * buf,unsigned int len,struct virtnet_rq_stats * stats)1309 static struct sk_buff *receive_big(struct net_device *dev,
1310 				   struct virtnet_info *vi,
1311 				   struct receive_queue *rq,
1312 				   void *buf,
1313 				   unsigned int len,
1314 				   struct virtnet_rq_stats *stats)
1315 {
1316 	struct page *page = buf;
1317 	struct sk_buff *skb =
1318 		page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, 0);
1319 
1320 	u64_stats_add(&stats->bytes, len - vi->hdr_len);
1321 	if (unlikely(!skb))
1322 		goto err;
1323 
1324 	return skb;
1325 
1326 err:
1327 	u64_stats_inc(&stats->drops);
1328 	give_pages(rq, page);
1329 	return NULL;
1330 }
1331 
mergeable_buf_free(struct receive_queue * rq,int num_buf,struct net_device * dev,struct virtnet_rq_stats * stats)1332 static void mergeable_buf_free(struct receive_queue *rq, int num_buf,
1333 			       struct net_device *dev,
1334 			       struct virtnet_rq_stats *stats)
1335 {
1336 	struct page *page;
1337 	void *buf;
1338 	int len;
1339 
1340 	while (num_buf-- > 1) {
1341 		buf = virtnet_rq_get_buf(rq, &len, NULL);
1342 		if (unlikely(!buf)) {
1343 			pr_debug("%s: rx error: %d buffers missing\n",
1344 				 dev->name, num_buf);
1345 			DEV_STATS_INC(dev, rx_length_errors);
1346 			break;
1347 		}
1348 		u64_stats_add(&stats->bytes, len);
1349 		page = virt_to_head_page(buf);
1350 		put_page(page);
1351 	}
1352 }
1353 
1354 /* Why not use xdp_build_skb_from_frame() ?
1355  * XDP core assumes that xdp frags are PAGE_SIZE in length, while in
1356  * virtio-net there are 2 points that do not match its requirements:
1357  *  1. The size of the prefilled buffer is not fixed before xdp is set.
1358  *  2. xdp_build_skb_from_frame() does more checks that we don't need,
1359  *     like eth_type_trans() (which virtio-net does in receive_buf()).
1360  */
build_skb_from_xdp_buff(struct net_device * dev,struct virtnet_info * vi,struct xdp_buff * xdp,unsigned int xdp_frags_truesz)1361 static struct sk_buff *build_skb_from_xdp_buff(struct net_device *dev,
1362 					       struct virtnet_info *vi,
1363 					       struct xdp_buff *xdp,
1364 					       unsigned int xdp_frags_truesz)
1365 {
1366 	struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
1367 	unsigned int headroom, data_len;
1368 	struct sk_buff *skb;
1369 	int metasize;
1370 	u8 nr_frags;
1371 
1372 	if (unlikely(xdp->data_end > xdp_data_hard_end(xdp))) {
1373 		pr_debug("Error building skb as missing reserved tailroom for xdp");
1374 		return NULL;
1375 	}
1376 
1377 	if (unlikely(xdp_buff_has_frags(xdp)))
1378 		nr_frags = sinfo->nr_frags;
1379 
1380 	skb = build_skb(xdp->data_hard_start, xdp->frame_sz);
1381 	if (unlikely(!skb))
1382 		return NULL;
1383 
1384 	headroom = xdp->data - xdp->data_hard_start;
1385 	data_len = xdp->data_end - xdp->data;
1386 	skb_reserve(skb, headroom);
1387 	__skb_put(skb, data_len);
1388 
1389 	metasize = xdp->data - xdp->data_meta;
1390 	metasize = metasize > 0 ? metasize : 0;
1391 	if (metasize)
1392 		skb_metadata_set(skb, metasize);
1393 
1394 	if (unlikely(xdp_buff_has_frags(xdp)))
1395 		xdp_update_skb_shared_info(skb, nr_frags,
1396 					   sinfo->xdp_frags_size,
1397 					   xdp_frags_truesz,
1398 					   xdp_buff_is_frag_pfmemalloc(xdp));
1399 
1400 	return skb;
1401 }
1402 
1403 /* TODO: build xdp in big mode */
virtnet_build_xdp_buff_mrg(struct net_device * dev,struct virtnet_info * vi,struct receive_queue * rq,struct xdp_buff * xdp,void * buf,unsigned int len,unsigned int frame_sz,int * num_buf,unsigned int * xdp_frags_truesize,struct virtnet_rq_stats * stats)1404 static int virtnet_build_xdp_buff_mrg(struct net_device *dev,
1405 				      struct virtnet_info *vi,
1406 				      struct receive_queue *rq,
1407 				      struct xdp_buff *xdp,
1408 				      void *buf,
1409 				      unsigned int len,
1410 				      unsigned int frame_sz,
1411 				      int *num_buf,
1412 				      unsigned int *xdp_frags_truesize,
1413 				      struct virtnet_rq_stats *stats)
1414 {
1415 	struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1416 	unsigned int headroom, tailroom, room;
1417 	unsigned int truesize, cur_frag_size;
1418 	struct skb_shared_info *shinfo;
1419 	unsigned int xdp_frags_truesz = 0;
1420 	struct page *page;
1421 	skb_frag_t *frag;
1422 	int offset;
1423 	void *ctx;
1424 
1425 	xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq);
1426 	xdp_prepare_buff(xdp, buf - VIRTIO_XDP_HEADROOM,
1427 			 VIRTIO_XDP_HEADROOM + vi->hdr_len, len - vi->hdr_len, true);
1428 
1429 	if (!*num_buf)
1430 		return 0;
1431 
1432 	if (*num_buf > 1) {
1433 		/* If we want to build multi-buffer xdp, we need
1434 		 * to specify that the flags of xdp_buff have the
1435 		 * XDP_FLAGS_HAS_FRAG bit.
1436 		 */
1437 		if (!xdp_buff_has_frags(xdp))
1438 			xdp_buff_set_frags_flag(xdp);
1439 
1440 		shinfo = xdp_get_shared_info_from_buff(xdp);
1441 		shinfo->nr_frags = 0;
1442 		shinfo->xdp_frags_size = 0;
1443 	}
1444 
1445 	if (*num_buf > MAX_SKB_FRAGS + 1)
1446 		return -EINVAL;
1447 
1448 	while (--*num_buf > 0) {
1449 		buf = virtnet_rq_get_buf(rq, &len, &ctx);
1450 		if (unlikely(!buf)) {
1451 			pr_debug("%s: rx error: %d buffers out of %d missing\n",
1452 				 dev->name, *num_buf,
1453 				 virtio16_to_cpu(vi->vdev, hdr->num_buffers));
1454 			DEV_STATS_INC(dev, rx_length_errors);
1455 			goto err;
1456 		}
1457 
1458 		u64_stats_add(&stats->bytes, len);
1459 		page = virt_to_head_page(buf);
1460 		offset = buf - page_address(page);
1461 
1462 		truesize = mergeable_ctx_to_truesize(ctx);
1463 		headroom = mergeable_ctx_to_headroom(ctx);
1464 		tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1465 		room = SKB_DATA_ALIGN(headroom + tailroom);
1466 
1467 		cur_frag_size = truesize;
1468 		xdp_frags_truesz += cur_frag_size;
1469 		if (unlikely(len > truesize - room || cur_frag_size > PAGE_SIZE)) {
1470 			put_page(page);
1471 			pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1472 				 dev->name, len, (unsigned long)(truesize - room));
1473 			DEV_STATS_INC(dev, rx_length_errors);
1474 			goto err;
1475 		}
1476 
1477 		frag = &shinfo->frags[shinfo->nr_frags++];
1478 		skb_frag_fill_page_desc(frag, page, offset, len);
1479 		if (page_is_pfmemalloc(page))
1480 			xdp_buff_set_frag_pfmemalloc(xdp);
1481 
1482 		shinfo->xdp_frags_size += len;
1483 	}
1484 
1485 	*xdp_frags_truesize = xdp_frags_truesz;
1486 	return 0;
1487 
1488 err:
1489 	put_xdp_frags(xdp);
1490 	return -EINVAL;
1491 }
1492 
mergeable_xdp_get_buf(struct virtnet_info * vi,struct receive_queue * rq,struct bpf_prog * xdp_prog,void * ctx,unsigned int * frame_sz,int * num_buf,struct page ** page,int offset,unsigned int * len,struct virtio_net_hdr_mrg_rxbuf * hdr)1493 static void *mergeable_xdp_get_buf(struct virtnet_info *vi,
1494 				   struct receive_queue *rq,
1495 				   struct bpf_prog *xdp_prog,
1496 				   void *ctx,
1497 				   unsigned int *frame_sz,
1498 				   int *num_buf,
1499 				   struct page **page,
1500 				   int offset,
1501 				   unsigned int *len,
1502 				   struct virtio_net_hdr_mrg_rxbuf *hdr)
1503 {
1504 	unsigned int truesize = mergeable_ctx_to_truesize(ctx);
1505 	unsigned int headroom = mergeable_ctx_to_headroom(ctx);
1506 	struct page *xdp_page;
1507 	unsigned int xdp_room;
1508 
1509 	/* Transient failure which in theory could occur if
1510 	 * in-flight packets from before XDP was enabled reach
1511 	 * the receive path after XDP is loaded.
1512 	 */
1513 	if (unlikely(hdr->hdr.gso_type))
1514 		return NULL;
1515 
1516 	/* Partially checksummed packets must be dropped. */
1517 	if (unlikely(hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM))
1518 		return NULL;
1519 
1520 	/* Now XDP core assumes frag size is PAGE_SIZE, but buffers
1521 	 * with headroom may add hole in truesize, which
1522 	 * make their length exceed PAGE_SIZE. So we disabled the
1523 	 * hole mechanism for xdp. See add_recvbuf_mergeable().
1524 	 */
1525 	*frame_sz = truesize;
1526 
1527 	if (likely(headroom >= virtnet_get_headroom(vi) &&
1528 		   (*num_buf == 1 || xdp_prog->aux->xdp_has_frags))) {
1529 		return page_address(*page) + offset;
1530 	}
1531 
1532 	/* This happens when headroom is not enough because
1533 	 * of the buffer was prefilled before XDP is set.
1534 	 * This should only happen for the first several packets.
1535 	 * In fact, vq reset can be used here to help us clean up
1536 	 * the prefilled buffers, but many existing devices do not
1537 	 * support it, and we don't want to bother users who are
1538 	 * using xdp normally.
1539 	 */
1540 	if (!xdp_prog->aux->xdp_has_frags) {
1541 		/* linearize data for XDP */
1542 		xdp_page = xdp_linearize_page(rq, num_buf,
1543 					      *page, offset,
1544 					      VIRTIO_XDP_HEADROOM,
1545 					      len);
1546 		if (!xdp_page)
1547 			return NULL;
1548 	} else {
1549 		xdp_room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM +
1550 					  sizeof(struct skb_shared_info));
1551 		if (*len + xdp_room > PAGE_SIZE)
1552 			return NULL;
1553 
1554 		xdp_page = alloc_page(GFP_ATOMIC);
1555 		if (!xdp_page)
1556 			return NULL;
1557 
1558 		memcpy(page_address(xdp_page) + VIRTIO_XDP_HEADROOM,
1559 		       page_address(*page) + offset, *len);
1560 	}
1561 
1562 	*frame_sz = PAGE_SIZE;
1563 
1564 	put_page(*page);
1565 
1566 	*page = xdp_page;
1567 
1568 	return page_address(*page) + VIRTIO_XDP_HEADROOM;
1569 }
1570 
receive_mergeable_xdp(struct net_device * dev,struct virtnet_info * vi,struct receive_queue * rq,struct bpf_prog * xdp_prog,void * buf,void * ctx,unsigned int len,unsigned int * xdp_xmit,struct virtnet_rq_stats * stats)1571 static struct sk_buff *receive_mergeable_xdp(struct net_device *dev,
1572 					     struct virtnet_info *vi,
1573 					     struct receive_queue *rq,
1574 					     struct bpf_prog *xdp_prog,
1575 					     void *buf,
1576 					     void *ctx,
1577 					     unsigned int len,
1578 					     unsigned int *xdp_xmit,
1579 					     struct virtnet_rq_stats *stats)
1580 {
1581 	struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1582 	int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
1583 	struct page *page = virt_to_head_page(buf);
1584 	int offset = buf - page_address(page);
1585 	unsigned int xdp_frags_truesz = 0;
1586 	struct sk_buff *head_skb;
1587 	unsigned int frame_sz;
1588 	struct xdp_buff xdp;
1589 	void *data;
1590 	u32 act;
1591 	int err;
1592 
1593 	data = mergeable_xdp_get_buf(vi, rq, xdp_prog, ctx, &frame_sz, &num_buf, &page,
1594 				     offset, &len, hdr);
1595 	if (unlikely(!data))
1596 		goto err_xdp;
1597 
1598 	err = virtnet_build_xdp_buff_mrg(dev, vi, rq, &xdp, data, len, frame_sz,
1599 					 &num_buf, &xdp_frags_truesz, stats);
1600 	if (unlikely(err))
1601 		goto err_xdp;
1602 
1603 	act = virtnet_xdp_handler(xdp_prog, &xdp, dev, xdp_xmit, stats);
1604 
1605 	switch (act) {
1606 	case XDP_PASS:
1607 		head_skb = build_skb_from_xdp_buff(dev, vi, &xdp, xdp_frags_truesz);
1608 		if (unlikely(!head_skb))
1609 			break;
1610 		return head_skb;
1611 
1612 	case XDP_TX:
1613 	case XDP_REDIRECT:
1614 		return NULL;
1615 
1616 	default:
1617 		break;
1618 	}
1619 
1620 	put_xdp_frags(&xdp);
1621 
1622 err_xdp:
1623 	put_page(page);
1624 	mergeable_buf_free(rq, num_buf, dev, stats);
1625 
1626 	u64_stats_inc(&stats->xdp_drops);
1627 	u64_stats_inc(&stats->drops);
1628 	return NULL;
1629 }
1630 
receive_mergeable(struct net_device * dev,struct virtnet_info * vi,struct receive_queue * rq,void * buf,void * ctx,unsigned int len,unsigned int * xdp_xmit,struct virtnet_rq_stats * stats)1631 static struct sk_buff *receive_mergeable(struct net_device *dev,
1632 					 struct virtnet_info *vi,
1633 					 struct receive_queue *rq,
1634 					 void *buf,
1635 					 void *ctx,
1636 					 unsigned int len,
1637 					 unsigned int *xdp_xmit,
1638 					 struct virtnet_rq_stats *stats)
1639 {
1640 	struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1641 	int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
1642 	struct page *page = virt_to_head_page(buf);
1643 	int offset = buf - page_address(page);
1644 	struct sk_buff *head_skb, *curr_skb;
1645 	unsigned int truesize = mergeable_ctx_to_truesize(ctx);
1646 	unsigned int headroom = mergeable_ctx_to_headroom(ctx);
1647 	unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1648 	unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1649 
1650 	head_skb = NULL;
1651 	u64_stats_add(&stats->bytes, len - vi->hdr_len);
1652 
1653 	if (unlikely(len > truesize - room)) {
1654 		pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1655 			 dev->name, len, (unsigned long)(truesize - room));
1656 		DEV_STATS_INC(dev, rx_length_errors);
1657 		goto err_skb;
1658 	}
1659 
1660 	if (unlikely(vi->xdp_enabled)) {
1661 		struct bpf_prog *xdp_prog;
1662 
1663 		rcu_read_lock();
1664 		xdp_prog = rcu_dereference(rq->xdp_prog);
1665 		if (xdp_prog) {
1666 			head_skb = receive_mergeable_xdp(dev, vi, rq, xdp_prog, buf, ctx,
1667 							 len, xdp_xmit, stats);
1668 			rcu_read_unlock();
1669 			return head_skb;
1670 		}
1671 		rcu_read_unlock();
1672 	}
1673 
1674 	head_skb = page_to_skb(vi, rq, page, offset, len, truesize, headroom);
1675 	curr_skb = head_skb;
1676 
1677 	if (unlikely(!curr_skb))
1678 		goto err_skb;
1679 	while (--num_buf) {
1680 		int num_skb_frags;
1681 
1682 		buf = virtnet_rq_get_buf(rq, &len, &ctx);
1683 		if (unlikely(!buf)) {
1684 			pr_debug("%s: rx error: %d buffers out of %d missing\n",
1685 				 dev->name, num_buf,
1686 				 virtio16_to_cpu(vi->vdev,
1687 						 hdr->num_buffers));
1688 			DEV_STATS_INC(dev, rx_length_errors);
1689 			goto err_buf;
1690 		}
1691 
1692 		u64_stats_add(&stats->bytes, len);
1693 		page = virt_to_head_page(buf);
1694 
1695 		truesize = mergeable_ctx_to_truesize(ctx);
1696 		headroom = mergeable_ctx_to_headroom(ctx);
1697 		tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1698 		room = SKB_DATA_ALIGN(headroom + tailroom);
1699 		if (unlikely(len > truesize - room)) {
1700 			pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1701 				 dev->name, len, (unsigned long)(truesize - room));
1702 			DEV_STATS_INC(dev, rx_length_errors);
1703 			goto err_skb;
1704 		}
1705 
1706 		num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1707 		if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1708 			struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1709 
1710 			if (unlikely(!nskb))
1711 				goto err_skb;
1712 			if (curr_skb == head_skb)
1713 				skb_shinfo(curr_skb)->frag_list = nskb;
1714 			else
1715 				curr_skb->next = nskb;
1716 			curr_skb = nskb;
1717 			head_skb->truesize += nskb->truesize;
1718 			num_skb_frags = 0;
1719 		}
1720 		if (curr_skb != head_skb) {
1721 			head_skb->data_len += len;
1722 			head_skb->len += len;
1723 			head_skb->truesize += truesize;
1724 		}
1725 		offset = buf - page_address(page);
1726 		if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1727 			put_page(page);
1728 			skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1729 					     len, truesize);
1730 		} else {
1731 			skb_add_rx_frag(curr_skb, num_skb_frags, page,
1732 					offset, len, truesize);
1733 		}
1734 	}
1735 
1736 	ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1737 	return head_skb;
1738 
1739 err_skb:
1740 	put_page(page);
1741 	mergeable_buf_free(rq, num_buf, dev, stats);
1742 
1743 err_buf:
1744 	u64_stats_inc(&stats->drops);
1745 	dev_kfree_skb(head_skb);
1746 	return NULL;
1747 }
1748 
virtio_skb_set_hash(const struct virtio_net_hdr_v1_hash * hdr_hash,struct sk_buff * skb)1749 static void virtio_skb_set_hash(const struct virtio_net_hdr_v1_hash *hdr_hash,
1750 				struct sk_buff *skb)
1751 {
1752 	enum pkt_hash_types rss_hash_type;
1753 
1754 	if (!hdr_hash || !skb)
1755 		return;
1756 
1757 	switch (__le16_to_cpu(hdr_hash->hash_report)) {
1758 	case VIRTIO_NET_HASH_REPORT_TCPv4:
1759 	case VIRTIO_NET_HASH_REPORT_UDPv4:
1760 	case VIRTIO_NET_HASH_REPORT_TCPv6:
1761 	case VIRTIO_NET_HASH_REPORT_UDPv6:
1762 	case VIRTIO_NET_HASH_REPORT_TCPv6_EX:
1763 	case VIRTIO_NET_HASH_REPORT_UDPv6_EX:
1764 		rss_hash_type = PKT_HASH_TYPE_L4;
1765 		break;
1766 	case VIRTIO_NET_HASH_REPORT_IPv4:
1767 	case VIRTIO_NET_HASH_REPORT_IPv6:
1768 	case VIRTIO_NET_HASH_REPORT_IPv6_EX:
1769 		rss_hash_type = PKT_HASH_TYPE_L3;
1770 		break;
1771 	case VIRTIO_NET_HASH_REPORT_NONE:
1772 	default:
1773 		rss_hash_type = PKT_HASH_TYPE_NONE;
1774 	}
1775 	skb_set_hash(skb, __le32_to_cpu(hdr_hash->hash_value), rss_hash_type);
1776 }
1777 
receive_buf(struct virtnet_info * vi,struct receive_queue * rq,void * buf,unsigned int len,void ** ctx,unsigned int * xdp_xmit,struct virtnet_rq_stats * stats)1778 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1779 			void *buf, unsigned int len, void **ctx,
1780 			unsigned int *xdp_xmit,
1781 			struct virtnet_rq_stats *stats)
1782 {
1783 	struct net_device *dev = vi->dev;
1784 	struct sk_buff *skb;
1785 	struct virtio_net_common_hdr *hdr;
1786 	u8 flags;
1787 
1788 	if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1789 		pr_debug("%s: short packet %i\n", dev->name, len);
1790 		DEV_STATS_INC(dev, rx_length_errors);
1791 		virtnet_rq_free_buf(vi, rq, buf);
1792 		return;
1793 	}
1794 
1795 	/* 1. Save the flags early, as the XDP program might overwrite them.
1796 	 * These flags ensure packets marked as VIRTIO_NET_HDR_F_DATA_VALID
1797 	 * stay valid after XDP processing.
1798 	 * 2. XDP doesn't work with partially checksummed packets (refer to
1799 	 * virtnet_xdp_set()), so packets marked as
1800 	 * VIRTIO_NET_HDR_F_NEEDS_CSUM get dropped during XDP processing.
1801 	 */
1802 	flags = ((struct virtio_net_common_hdr *)buf)->hdr.flags;
1803 
1804 	if (vi->mergeable_rx_bufs)
1805 		skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1806 					stats);
1807 	else if (vi->big_packets)
1808 		skb = receive_big(dev, vi, rq, buf, len, stats);
1809 	else
1810 		skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1811 
1812 	if (unlikely(!skb))
1813 		return;
1814 
1815 	hdr = skb_vnet_common_hdr(skb);
1816 	if (dev->features & NETIF_F_RXHASH && vi->has_rss_hash_report)
1817 		virtio_skb_set_hash(&hdr->hash_v1_hdr, skb);
1818 
1819 	if (flags & VIRTIO_NET_HDR_F_DATA_VALID)
1820 		skb->ip_summed = CHECKSUM_UNNECESSARY;
1821 
1822 	if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1823 				  virtio_is_little_endian(vi->vdev))) {
1824 		net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1825 				     dev->name, hdr->hdr.gso_type,
1826 				     hdr->hdr.gso_size);
1827 		goto frame_err;
1828 	}
1829 
1830 	skb_record_rx_queue(skb, vq2rxq(rq->vq));
1831 	skb->protocol = eth_type_trans(skb, dev);
1832 	pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1833 		 ntohs(skb->protocol), skb->len, skb->pkt_type);
1834 
1835 	napi_gro_receive(&rq->napi, skb);
1836 	return;
1837 
1838 frame_err:
1839 	DEV_STATS_INC(dev, rx_frame_errors);
1840 	dev_kfree_skb(skb);
1841 }
1842 
1843 /* Unlike mergeable buffers, all buffers are allocated to the
1844  * same size, except for the headroom. For this reason we do
1845  * not need to use  mergeable_len_to_ctx here - it is enough
1846  * to store the headroom as the context ignoring the truesize.
1847  */
add_recvbuf_small(struct virtnet_info * vi,struct receive_queue * rq,gfp_t gfp)1848 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1849 			     gfp_t gfp)
1850 {
1851 	char *buf;
1852 	unsigned int xdp_headroom = virtnet_get_headroom(vi);
1853 	void *ctx = (void *)(unsigned long)xdp_headroom;
1854 	int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1855 	int err;
1856 
1857 	len = SKB_DATA_ALIGN(len) +
1858 	      SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1859 
1860 	if (unlikely(!skb_page_frag_refill(len, &rq->alloc_frag, gfp)))
1861 		return -ENOMEM;
1862 
1863 	buf = virtnet_rq_alloc(rq, len, gfp);
1864 	if (unlikely(!buf))
1865 		return -ENOMEM;
1866 
1867 	buf += VIRTNET_RX_PAD + xdp_headroom;
1868 
1869 	virtnet_rq_init_one_sg(rq, buf, vi->hdr_len + GOOD_PACKET_LEN);
1870 
1871 	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1872 	if (err < 0) {
1873 		if (rq->do_dma)
1874 			virtnet_rq_unmap(rq, buf, 0);
1875 		put_page(virt_to_head_page(buf));
1876 	}
1877 
1878 	return err;
1879 }
1880 
add_recvbuf_big(struct virtnet_info * vi,struct receive_queue * rq,gfp_t gfp)1881 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1882 			   gfp_t gfp)
1883 {
1884 	struct page *first, *list = NULL;
1885 	char *p;
1886 	int i, err, offset;
1887 
1888 	sg_init_table(rq->sg, vi->big_packets_num_skbfrags + 2);
1889 
1890 	/* page in rq->sg[vi->big_packets_num_skbfrags + 1] is list tail */
1891 	for (i = vi->big_packets_num_skbfrags + 1; i > 1; --i) {
1892 		first = get_a_page(rq, gfp);
1893 		if (!first) {
1894 			if (list)
1895 				give_pages(rq, list);
1896 			return -ENOMEM;
1897 		}
1898 		sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1899 
1900 		/* chain new page in list head to match sg */
1901 		first->private = (unsigned long)list;
1902 		list = first;
1903 	}
1904 
1905 	first = get_a_page(rq, gfp);
1906 	if (!first) {
1907 		give_pages(rq, list);
1908 		return -ENOMEM;
1909 	}
1910 	p = page_address(first);
1911 
1912 	/* rq->sg[0], rq->sg[1] share the same page */
1913 	/* a separated rq->sg[0] for header - required in case !any_header_sg */
1914 	sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1915 
1916 	/* rq->sg[1] for data packet, from offset */
1917 	offset = sizeof(struct padded_vnet_hdr);
1918 	sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1919 
1920 	/* chain first in list head */
1921 	first->private = (unsigned long)list;
1922 	err = virtqueue_add_inbuf(rq->vq, rq->sg, vi->big_packets_num_skbfrags + 2,
1923 				  first, gfp);
1924 	if (err < 0)
1925 		give_pages(rq, first);
1926 
1927 	return err;
1928 }
1929 
get_mergeable_buf_len(struct receive_queue * rq,struct ewma_pkt_len * avg_pkt_len,unsigned int room)1930 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1931 					  struct ewma_pkt_len *avg_pkt_len,
1932 					  unsigned int room)
1933 {
1934 	struct virtnet_info *vi = rq->vq->vdev->priv;
1935 	const size_t hdr_len = vi->hdr_len;
1936 	unsigned int len;
1937 
1938 	if (room)
1939 		return PAGE_SIZE - room;
1940 
1941 	len = hdr_len +	clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1942 				rq->min_buf_len, PAGE_SIZE - hdr_len);
1943 
1944 	return ALIGN(len, L1_CACHE_BYTES);
1945 }
1946 
add_recvbuf_mergeable(struct virtnet_info * vi,struct receive_queue * rq,gfp_t gfp)1947 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1948 				 struct receive_queue *rq, gfp_t gfp)
1949 {
1950 	struct page_frag *alloc_frag = &rq->alloc_frag;
1951 	unsigned int headroom = virtnet_get_headroom(vi);
1952 	unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1953 	unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1954 	unsigned int len, hole;
1955 	void *ctx;
1956 	char *buf;
1957 	int err;
1958 
1959 	/* Extra tailroom is needed to satisfy XDP's assumption. This
1960 	 * means rx frags coalescing won't work, but consider we've
1961 	 * disabled GSO for XDP, it won't be a big issue.
1962 	 */
1963 	len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1964 
1965 	if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1966 		return -ENOMEM;
1967 
1968 	if (!alloc_frag->offset && len + room + sizeof(struct virtnet_rq_dma) > alloc_frag->size)
1969 		len -= sizeof(struct virtnet_rq_dma);
1970 
1971 	buf = virtnet_rq_alloc(rq, len + room, gfp);
1972 	if (unlikely(!buf))
1973 		return -ENOMEM;
1974 
1975 	buf += headroom; /* advance address leaving hole at front of pkt */
1976 	hole = alloc_frag->size - alloc_frag->offset;
1977 	if (hole < len + room) {
1978 		/* To avoid internal fragmentation, if there is very likely not
1979 		 * enough space for another buffer, add the remaining space to
1980 		 * the current buffer.
1981 		 * XDP core assumes that frame_size of xdp_buff and the length
1982 		 * of the frag are PAGE_SIZE, so we disable the hole mechanism.
1983 		 */
1984 		if (!headroom)
1985 			len += hole;
1986 		alloc_frag->offset += hole;
1987 	}
1988 
1989 	virtnet_rq_init_one_sg(rq, buf, len);
1990 
1991 	ctx = mergeable_len_to_ctx(len + room, headroom);
1992 	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1993 	if (err < 0) {
1994 		if (rq->do_dma)
1995 			virtnet_rq_unmap(rq, buf, 0);
1996 		put_page(virt_to_head_page(buf));
1997 	}
1998 
1999 	return err;
2000 }
2001 
2002 /*
2003  * Returns false if we couldn't fill entirely (OOM).
2004  *
2005  * Normally run in the receive path, but can also be run from ndo_open
2006  * before we're receiving packets, or from refill_work which is
2007  * careful to disable receiving (using napi_disable).
2008  */
try_fill_recv(struct virtnet_info * vi,struct receive_queue * rq,gfp_t gfp)2009 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
2010 			  gfp_t gfp)
2011 {
2012 	int err;
2013 	bool oom;
2014 
2015 	do {
2016 		if (vi->mergeable_rx_bufs)
2017 			err = add_recvbuf_mergeable(vi, rq, gfp);
2018 		else if (vi->big_packets)
2019 			err = add_recvbuf_big(vi, rq, gfp);
2020 		else
2021 			err = add_recvbuf_small(vi, rq, gfp);
2022 
2023 		oom = err == -ENOMEM;
2024 		if (err)
2025 			break;
2026 	} while (rq->vq->num_free);
2027 	if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
2028 		unsigned long flags;
2029 
2030 		flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
2031 		u64_stats_inc(&rq->stats.kicks);
2032 		u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
2033 	}
2034 
2035 	return !oom;
2036 }
2037 
skb_recv_done(struct virtqueue * rvq)2038 static void skb_recv_done(struct virtqueue *rvq)
2039 {
2040 	struct virtnet_info *vi = rvq->vdev->priv;
2041 	struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
2042 
2043 	virtqueue_napi_schedule(&rq->napi, rvq);
2044 }
2045 
virtnet_napi_enable(struct virtqueue * vq,struct napi_struct * napi)2046 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
2047 {
2048 	napi_enable(napi);
2049 
2050 	/* If all buffers were filled by other side before we napi_enabled, we
2051 	 * won't get another interrupt, so process any outstanding packets now.
2052 	 * Call local_bh_enable after to trigger softIRQ processing.
2053 	 */
2054 	local_bh_disable();
2055 	virtqueue_napi_schedule(napi, vq);
2056 	local_bh_enable();
2057 }
2058 
virtnet_napi_tx_enable(struct virtnet_info * vi,struct virtqueue * vq,struct napi_struct * napi)2059 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
2060 				   struct virtqueue *vq,
2061 				   struct napi_struct *napi)
2062 {
2063 	if (!napi->weight)
2064 		return;
2065 
2066 	/* Tx napi touches cachelines on the cpu handling tx interrupts. Only
2067 	 * enable the feature if this is likely affine with the transmit path.
2068 	 */
2069 	if (!vi->affinity_hint_set) {
2070 		napi->weight = 0;
2071 		return;
2072 	}
2073 
2074 	return virtnet_napi_enable(vq, napi);
2075 }
2076 
virtnet_napi_tx_disable(struct napi_struct * napi)2077 static void virtnet_napi_tx_disable(struct napi_struct *napi)
2078 {
2079 	if (napi->weight)
2080 		napi_disable(napi);
2081 }
2082 
refill_work(struct work_struct * work)2083 static void refill_work(struct work_struct *work)
2084 {
2085 	struct virtnet_info *vi =
2086 		container_of(work, struct virtnet_info, refill.work);
2087 	bool still_empty;
2088 	int i;
2089 
2090 	for (i = 0; i < vi->curr_queue_pairs; i++) {
2091 		struct receive_queue *rq = &vi->rq[i];
2092 
2093 		napi_disable(&rq->napi);
2094 		still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
2095 		virtnet_napi_enable(rq->vq, &rq->napi);
2096 
2097 		/* In theory, this can happen: if we don't get any buffers in
2098 		 * we will *never* try to fill again.
2099 		 */
2100 		if (still_empty)
2101 			schedule_delayed_work(&vi->refill, HZ/2);
2102 	}
2103 }
2104 
virtnet_receive(struct receive_queue * rq,int budget,unsigned int * xdp_xmit)2105 static int virtnet_receive(struct receive_queue *rq, int budget,
2106 			   unsigned int *xdp_xmit)
2107 {
2108 	struct virtnet_info *vi = rq->vq->vdev->priv;
2109 	struct virtnet_rq_stats stats = {};
2110 	unsigned int len;
2111 	int packets = 0;
2112 	void *buf;
2113 	int i;
2114 
2115 	if (!vi->big_packets || vi->mergeable_rx_bufs) {
2116 		void *ctx;
2117 
2118 		while (packets < budget &&
2119 		       (buf = virtnet_rq_get_buf(rq, &len, &ctx))) {
2120 			receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
2121 			packets++;
2122 		}
2123 	} else {
2124 		while (packets < budget &&
2125 		       (buf = virtnet_rq_get_buf(rq, &len, NULL)) != NULL) {
2126 			receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
2127 			packets++;
2128 		}
2129 	}
2130 
2131 	if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
2132 		if (!try_fill_recv(vi, rq, GFP_ATOMIC)) {
2133 			spin_lock(&vi->refill_lock);
2134 			if (vi->refill_enabled)
2135 				schedule_delayed_work(&vi->refill, 0);
2136 			spin_unlock(&vi->refill_lock);
2137 		}
2138 	}
2139 
2140 	u64_stats_set(&stats.packets, packets);
2141 	u64_stats_update_begin(&rq->stats.syncp);
2142 	for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
2143 		size_t offset = virtnet_rq_stats_desc[i].offset;
2144 		u64_stats_t *item, *src;
2145 
2146 		item = (u64_stats_t *)((u8 *)&rq->stats + offset);
2147 		src = (u64_stats_t *)((u8 *)&stats + offset);
2148 		u64_stats_add(item, u64_stats_read(src));
2149 	}
2150 	u64_stats_update_end(&rq->stats.syncp);
2151 
2152 	return packets;
2153 }
2154 
virtnet_poll_cleantx(struct receive_queue * rq,int budget)2155 static void virtnet_poll_cleantx(struct receive_queue *rq, int budget)
2156 {
2157 	struct virtnet_info *vi = rq->vq->vdev->priv;
2158 	unsigned int index = vq2rxq(rq->vq);
2159 	struct send_queue *sq = &vi->sq[index];
2160 	struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
2161 
2162 	if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
2163 		return;
2164 
2165 	if (__netif_tx_trylock(txq)) {
2166 		if (sq->reset) {
2167 			__netif_tx_unlock(txq);
2168 			return;
2169 		}
2170 
2171 		do {
2172 			virtqueue_disable_cb(sq->vq);
2173 			free_old_xmit_skbs(sq, !!budget);
2174 		} while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
2175 
2176 		if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
2177 			netif_tx_wake_queue(txq);
2178 
2179 		__netif_tx_unlock(txq);
2180 	}
2181 }
2182 
virtnet_poll(struct napi_struct * napi,int budget)2183 static int virtnet_poll(struct napi_struct *napi, int budget)
2184 {
2185 	struct receive_queue *rq =
2186 		container_of(napi, struct receive_queue, napi);
2187 	struct virtnet_info *vi = rq->vq->vdev->priv;
2188 	struct send_queue *sq;
2189 	unsigned int received;
2190 	unsigned int xdp_xmit = 0;
2191 
2192 	virtnet_poll_cleantx(rq, budget);
2193 
2194 	received = virtnet_receive(rq, budget, &xdp_xmit);
2195 
2196 	if (xdp_xmit & VIRTIO_XDP_REDIR)
2197 		xdp_do_flush();
2198 
2199 	/* Out of packets? */
2200 	if (received < budget)
2201 		virtqueue_napi_complete(napi, rq->vq, received);
2202 
2203 	if (xdp_xmit & VIRTIO_XDP_TX) {
2204 		sq = virtnet_xdp_get_sq(vi);
2205 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
2206 			u64_stats_update_begin(&sq->stats.syncp);
2207 			u64_stats_inc(&sq->stats.kicks);
2208 			u64_stats_update_end(&sq->stats.syncp);
2209 		}
2210 		virtnet_xdp_put_sq(vi, sq);
2211 	}
2212 
2213 	return received;
2214 }
2215 
virtnet_disable_queue_pair(struct virtnet_info * vi,int qp_index)2216 static void virtnet_disable_queue_pair(struct virtnet_info *vi, int qp_index)
2217 {
2218 	virtnet_napi_tx_disable(&vi->sq[qp_index].napi);
2219 	napi_disable(&vi->rq[qp_index].napi);
2220 	xdp_rxq_info_unreg(&vi->rq[qp_index].xdp_rxq);
2221 }
2222 
virtnet_enable_queue_pair(struct virtnet_info * vi,int qp_index)2223 static int virtnet_enable_queue_pair(struct virtnet_info *vi, int qp_index)
2224 {
2225 	struct net_device *dev = vi->dev;
2226 	int err;
2227 
2228 	err = xdp_rxq_info_reg(&vi->rq[qp_index].xdp_rxq, dev, qp_index,
2229 			       vi->rq[qp_index].napi.napi_id);
2230 	if (err < 0)
2231 		return err;
2232 
2233 	err = xdp_rxq_info_reg_mem_model(&vi->rq[qp_index].xdp_rxq,
2234 					 MEM_TYPE_PAGE_SHARED, NULL);
2235 	if (err < 0)
2236 		goto err_xdp_reg_mem_model;
2237 
2238 	virtnet_napi_enable(vi->rq[qp_index].vq, &vi->rq[qp_index].napi);
2239 	virtnet_napi_tx_enable(vi, vi->sq[qp_index].vq, &vi->sq[qp_index].napi);
2240 
2241 	return 0;
2242 
2243 err_xdp_reg_mem_model:
2244 	xdp_rxq_info_unreg(&vi->rq[qp_index].xdp_rxq);
2245 	return err;
2246 }
2247 
virtnet_open(struct net_device * dev)2248 static int virtnet_open(struct net_device *dev)
2249 {
2250 	struct virtnet_info *vi = netdev_priv(dev);
2251 	int i, err;
2252 
2253 	enable_delayed_refill(vi);
2254 
2255 	for (i = 0; i < vi->max_queue_pairs; i++) {
2256 		if (i < vi->curr_queue_pairs)
2257 			/* Make sure we have some buffers: if oom use wq. */
2258 			if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2259 				schedule_delayed_work(&vi->refill, 0);
2260 
2261 		err = virtnet_enable_queue_pair(vi, i);
2262 		if (err < 0)
2263 			goto err_enable_qp;
2264 	}
2265 
2266 	return 0;
2267 
2268 err_enable_qp:
2269 	disable_delayed_refill(vi);
2270 	cancel_delayed_work_sync(&vi->refill);
2271 
2272 	for (i--; i >= 0; i--)
2273 		virtnet_disable_queue_pair(vi, i);
2274 	return err;
2275 }
2276 
virtnet_poll_tx(struct napi_struct * napi,int budget)2277 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
2278 {
2279 	struct send_queue *sq = container_of(napi, struct send_queue, napi);
2280 	struct virtnet_info *vi = sq->vq->vdev->priv;
2281 	unsigned int index = vq2txq(sq->vq);
2282 	struct netdev_queue *txq;
2283 	int opaque;
2284 	bool done;
2285 
2286 	if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
2287 		/* We don't need to enable cb for XDP */
2288 		napi_complete_done(napi, 0);
2289 		return 0;
2290 	}
2291 
2292 	txq = netdev_get_tx_queue(vi->dev, index);
2293 	__netif_tx_lock(txq, raw_smp_processor_id());
2294 	virtqueue_disable_cb(sq->vq);
2295 	free_old_xmit_skbs(sq, !!budget);
2296 
2297 	if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
2298 		netif_tx_wake_queue(txq);
2299 
2300 	opaque = virtqueue_enable_cb_prepare(sq->vq);
2301 
2302 	done = napi_complete_done(napi, 0);
2303 
2304 	if (!done)
2305 		virtqueue_disable_cb(sq->vq);
2306 
2307 	__netif_tx_unlock(txq);
2308 
2309 	if (done) {
2310 		if (unlikely(virtqueue_poll(sq->vq, opaque))) {
2311 			if (napi_schedule_prep(napi)) {
2312 				__netif_tx_lock(txq, raw_smp_processor_id());
2313 				virtqueue_disable_cb(sq->vq);
2314 				__netif_tx_unlock(txq);
2315 				__napi_schedule(napi);
2316 			}
2317 		}
2318 	}
2319 
2320 	return 0;
2321 }
2322 
xmit_skb(struct send_queue * sq,struct sk_buff * skb)2323 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
2324 {
2325 	struct virtio_net_hdr_mrg_rxbuf *hdr;
2326 	const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
2327 	struct virtnet_info *vi = sq->vq->vdev->priv;
2328 	int num_sg;
2329 	unsigned hdr_len = vi->hdr_len;
2330 	bool can_push;
2331 
2332 	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
2333 
2334 	can_push = vi->any_header_sg &&
2335 		!((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
2336 		!skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
2337 	/* Even if we can, don't push here yet as this would skew
2338 	 * csum_start offset below. */
2339 	if (can_push)
2340 		hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
2341 	else
2342 		hdr = &skb_vnet_common_hdr(skb)->mrg_hdr;
2343 
2344 	if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
2345 				    virtio_is_little_endian(vi->vdev), false,
2346 				    0))
2347 		return -EPROTO;
2348 
2349 	if (vi->mergeable_rx_bufs)
2350 		hdr->num_buffers = 0;
2351 
2352 	sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
2353 	if (can_push) {
2354 		__skb_push(skb, hdr_len);
2355 		num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
2356 		if (unlikely(num_sg < 0))
2357 			return num_sg;
2358 		/* Pull header back to avoid skew in tx bytes calculations. */
2359 		__skb_pull(skb, hdr_len);
2360 	} else {
2361 		sg_set_buf(sq->sg, hdr, hdr_len);
2362 		num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
2363 		if (unlikely(num_sg < 0))
2364 			return num_sg;
2365 		num_sg++;
2366 	}
2367 	return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
2368 }
2369 
start_xmit(struct sk_buff * skb,struct net_device * dev)2370 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
2371 {
2372 	struct virtnet_info *vi = netdev_priv(dev);
2373 	int qnum = skb_get_queue_mapping(skb);
2374 	struct send_queue *sq = &vi->sq[qnum];
2375 	int err;
2376 	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
2377 	bool kick = !netdev_xmit_more();
2378 	bool use_napi = sq->napi.weight;
2379 
2380 	/* Free up any pending old buffers before queueing new ones. */
2381 	do {
2382 		if (use_napi)
2383 			virtqueue_disable_cb(sq->vq);
2384 
2385 		free_old_xmit_skbs(sq, false);
2386 
2387 	} while (use_napi && kick &&
2388 	       unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
2389 
2390 	/* timestamp packet in software */
2391 	skb_tx_timestamp(skb);
2392 
2393 	/* Try to transmit */
2394 	err = xmit_skb(sq, skb);
2395 
2396 	/* This should not happen! */
2397 	if (unlikely(err)) {
2398 		DEV_STATS_INC(dev, tx_fifo_errors);
2399 		if (net_ratelimit())
2400 			dev_warn(&dev->dev,
2401 				 "Unexpected TXQ (%d) queue failure: %d\n",
2402 				 qnum, err);
2403 		DEV_STATS_INC(dev, tx_dropped);
2404 		dev_kfree_skb_any(skb);
2405 		return NETDEV_TX_OK;
2406 	}
2407 
2408 	/* Don't wait up for transmitted skbs to be freed. */
2409 	if (!use_napi) {
2410 		skb_orphan(skb);
2411 		nf_reset_ct(skb);
2412 	}
2413 
2414 	check_sq_full_and_disable(vi, dev, sq);
2415 
2416 	if (kick || netif_xmit_stopped(txq)) {
2417 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
2418 			u64_stats_update_begin(&sq->stats.syncp);
2419 			u64_stats_inc(&sq->stats.kicks);
2420 			u64_stats_update_end(&sq->stats.syncp);
2421 		}
2422 	}
2423 
2424 	return NETDEV_TX_OK;
2425 }
2426 
virtnet_rx_resize(struct virtnet_info * vi,struct receive_queue * rq,u32 ring_num)2427 static int virtnet_rx_resize(struct virtnet_info *vi,
2428 			     struct receive_queue *rq, u32 ring_num)
2429 {
2430 	bool running = netif_running(vi->dev);
2431 	int err, qindex;
2432 
2433 	qindex = rq - vi->rq;
2434 
2435 	if (running)
2436 		napi_disable(&rq->napi);
2437 
2438 	err = virtqueue_resize(rq->vq, ring_num, virtnet_rq_unmap_free_buf);
2439 	if (err)
2440 		netdev_err(vi->dev, "resize rx fail: rx queue index: %d err: %d\n", qindex, err);
2441 
2442 	if (!try_fill_recv(vi, rq, GFP_KERNEL))
2443 		schedule_delayed_work(&vi->refill, 0);
2444 
2445 	if (running)
2446 		virtnet_napi_enable(rq->vq, &rq->napi);
2447 	return err;
2448 }
2449 
virtnet_tx_resize(struct virtnet_info * vi,struct send_queue * sq,u32 ring_num)2450 static int virtnet_tx_resize(struct virtnet_info *vi,
2451 			     struct send_queue *sq, u32 ring_num)
2452 {
2453 	bool running = netif_running(vi->dev);
2454 	struct netdev_queue *txq;
2455 	int err, qindex;
2456 
2457 	qindex = sq - vi->sq;
2458 
2459 	if (running)
2460 		virtnet_napi_tx_disable(&sq->napi);
2461 
2462 	txq = netdev_get_tx_queue(vi->dev, qindex);
2463 
2464 	/* 1. wait all ximt complete
2465 	 * 2. fix the race of netif_stop_subqueue() vs netif_start_subqueue()
2466 	 */
2467 	__netif_tx_lock_bh(txq);
2468 
2469 	/* Prevent rx poll from accessing sq. */
2470 	sq->reset = true;
2471 
2472 	/* Prevent the upper layer from trying to send packets. */
2473 	netif_stop_subqueue(vi->dev, qindex);
2474 
2475 	__netif_tx_unlock_bh(txq);
2476 
2477 	err = virtqueue_resize(sq->vq, ring_num, virtnet_sq_free_unused_buf);
2478 	if (err)
2479 		netdev_err(vi->dev, "resize tx fail: tx queue index: %d err: %d\n", qindex, err);
2480 
2481 	__netif_tx_lock_bh(txq);
2482 	sq->reset = false;
2483 	netif_tx_wake_queue(txq);
2484 	__netif_tx_unlock_bh(txq);
2485 
2486 	if (running)
2487 		virtnet_napi_tx_enable(vi, sq->vq, &sq->napi);
2488 	return err;
2489 }
2490 
2491 /*
2492  * Send command via the control virtqueue and check status.  Commands
2493  * supported by the hypervisor, as indicated by feature bits, should
2494  * never fail unless improperly formatted.
2495  */
virtnet_send_command(struct virtnet_info * vi,u8 class,u8 cmd,struct scatterlist * out)2496 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
2497 				 struct scatterlist *out)
2498 {
2499 	struct scatterlist *sgs[4], hdr, stat;
2500 	unsigned out_num = 0, tmp;
2501 	int ret;
2502 
2503 	/* Caller should know better */
2504 	BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
2505 
2506 	vi->ctrl->status = ~0;
2507 	vi->ctrl->hdr.class = class;
2508 	vi->ctrl->hdr.cmd = cmd;
2509 	/* Add header */
2510 	sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
2511 	sgs[out_num++] = &hdr;
2512 
2513 	if (out)
2514 		sgs[out_num++] = out;
2515 
2516 	/* Add return status. */
2517 	sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
2518 	sgs[out_num] = &stat;
2519 
2520 	BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
2521 	ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
2522 	if (ret < 0) {
2523 		dev_warn(&vi->vdev->dev,
2524 			 "Failed to add sgs for command vq: %d\n.", ret);
2525 		return false;
2526 	}
2527 
2528 	if (unlikely(!virtqueue_kick(vi->cvq)))
2529 		return vi->ctrl->status == VIRTIO_NET_OK;
2530 
2531 	/* Spin for a response, the kick causes an ioport write, trapping
2532 	 * into the hypervisor, so the request should be handled immediately.
2533 	 */
2534 	while (!virtqueue_get_buf(vi->cvq, &tmp) &&
2535 	       !virtqueue_is_broken(vi->cvq))
2536 		cpu_relax();
2537 
2538 	return vi->ctrl->status == VIRTIO_NET_OK;
2539 }
2540 
virtnet_set_mac_address(struct net_device * dev,void * p)2541 static int virtnet_set_mac_address(struct net_device *dev, void *p)
2542 {
2543 	struct virtnet_info *vi = netdev_priv(dev);
2544 	struct virtio_device *vdev = vi->vdev;
2545 	int ret;
2546 	struct sockaddr *addr;
2547 	struct scatterlist sg;
2548 
2549 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2550 		return -EOPNOTSUPP;
2551 
2552 	addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
2553 	if (!addr)
2554 		return -ENOMEM;
2555 
2556 	ret = eth_prepare_mac_addr_change(dev, addr);
2557 	if (ret)
2558 		goto out;
2559 
2560 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
2561 		sg_init_one(&sg, addr->sa_data, dev->addr_len);
2562 		if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2563 					  VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
2564 			dev_warn(&vdev->dev,
2565 				 "Failed to set mac address by vq command.\n");
2566 			ret = -EINVAL;
2567 			goto out;
2568 		}
2569 	} else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
2570 		   !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
2571 		unsigned int i;
2572 
2573 		/* Naturally, this has an atomicity problem. */
2574 		for (i = 0; i < dev->addr_len; i++)
2575 			virtio_cwrite8(vdev,
2576 				       offsetof(struct virtio_net_config, mac) +
2577 				       i, addr->sa_data[i]);
2578 	}
2579 
2580 	eth_commit_mac_addr_change(dev, p);
2581 	ret = 0;
2582 
2583 out:
2584 	kfree(addr);
2585 	return ret;
2586 }
2587 
virtnet_stats(struct net_device * dev,struct rtnl_link_stats64 * tot)2588 static void virtnet_stats(struct net_device *dev,
2589 			  struct rtnl_link_stats64 *tot)
2590 {
2591 	struct virtnet_info *vi = netdev_priv(dev);
2592 	unsigned int start;
2593 	int i;
2594 
2595 	for (i = 0; i < vi->max_queue_pairs; i++) {
2596 		u64 tpackets, tbytes, terrors, rpackets, rbytes, rdrops;
2597 		struct receive_queue *rq = &vi->rq[i];
2598 		struct send_queue *sq = &vi->sq[i];
2599 
2600 		do {
2601 			start = u64_stats_fetch_begin(&sq->stats.syncp);
2602 			tpackets = u64_stats_read(&sq->stats.packets);
2603 			tbytes   = u64_stats_read(&sq->stats.bytes);
2604 			terrors  = u64_stats_read(&sq->stats.tx_timeouts);
2605 		} while (u64_stats_fetch_retry(&sq->stats.syncp, start));
2606 
2607 		do {
2608 			start = u64_stats_fetch_begin(&rq->stats.syncp);
2609 			rpackets = u64_stats_read(&rq->stats.packets);
2610 			rbytes   = u64_stats_read(&rq->stats.bytes);
2611 			rdrops   = u64_stats_read(&rq->stats.drops);
2612 		} while (u64_stats_fetch_retry(&rq->stats.syncp, start));
2613 
2614 		tot->rx_packets += rpackets;
2615 		tot->tx_packets += tpackets;
2616 		tot->rx_bytes   += rbytes;
2617 		tot->tx_bytes   += tbytes;
2618 		tot->rx_dropped += rdrops;
2619 		tot->tx_errors  += terrors;
2620 	}
2621 
2622 	tot->tx_dropped = DEV_STATS_READ(dev, tx_dropped);
2623 	tot->tx_fifo_errors = DEV_STATS_READ(dev, tx_fifo_errors);
2624 	tot->rx_length_errors = DEV_STATS_READ(dev, rx_length_errors);
2625 	tot->rx_frame_errors = DEV_STATS_READ(dev, rx_frame_errors);
2626 }
2627 
virtnet_ack_link_announce(struct virtnet_info * vi)2628 static void virtnet_ack_link_announce(struct virtnet_info *vi)
2629 {
2630 	rtnl_lock();
2631 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
2632 				  VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
2633 		dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
2634 	rtnl_unlock();
2635 }
2636 
_virtnet_set_queues(struct virtnet_info * vi,u16 queue_pairs)2637 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
2638 {
2639 	struct scatterlist sg;
2640 	struct net_device *dev = vi->dev;
2641 
2642 	if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
2643 		return 0;
2644 
2645 	vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
2646 	sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
2647 
2648 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
2649 				  VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
2650 		dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
2651 			 queue_pairs);
2652 		return -EINVAL;
2653 	} else {
2654 		vi->curr_queue_pairs = queue_pairs;
2655 		/* virtnet_open() will refill when device is going to up. */
2656 		if (dev->flags & IFF_UP)
2657 			schedule_delayed_work(&vi->refill, 0);
2658 	}
2659 
2660 	return 0;
2661 }
2662 
virtnet_set_queues(struct virtnet_info * vi,u16 queue_pairs)2663 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
2664 {
2665 	int err;
2666 
2667 	rtnl_lock();
2668 	err = _virtnet_set_queues(vi, queue_pairs);
2669 	rtnl_unlock();
2670 	return err;
2671 }
2672 
virtnet_close(struct net_device * dev)2673 static int virtnet_close(struct net_device *dev)
2674 {
2675 	struct virtnet_info *vi = netdev_priv(dev);
2676 	int i;
2677 
2678 	/* Make sure NAPI doesn't schedule refill work */
2679 	disable_delayed_refill(vi);
2680 	/* Make sure refill_work doesn't re-enable napi! */
2681 	cancel_delayed_work_sync(&vi->refill);
2682 
2683 	for (i = 0; i < vi->max_queue_pairs; i++)
2684 		virtnet_disable_queue_pair(vi, i);
2685 
2686 	return 0;
2687 }
2688 
virtnet_set_rx_mode(struct net_device * dev)2689 static void virtnet_set_rx_mode(struct net_device *dev)
2690 {
2691 	struct virtnet_info *vi = netdev_priv(dev);
2692 	struct scatterlist sg[2];
2693 	struct virtio_net_ctrl_mac *mac_data;
2694 	struct netdev_hw_addr *ha;
2695 	int uc_count;
2696 	int mc_count;
2697 	void *buf;
2698 	int i;
2699 
2700 	/* We can't dynamically set ndo_set_rx_mode, so return gracefully */
2701 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
2702 		return;
2703 
2704 	vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
2705 	vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
2706 
2707 	sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
2708 
2709 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2710 				  VIRTIO_NET_CTRL_RX_PROMISC, sg))
2711 		dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
2712 			 vi->ctrl->promisc ? "en" : "dis");
2713 
2714 	sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
2715 
2716 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2717 				  VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
2718 		dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
2719 			 vi->ctrl->allmulti ? "en" : "dis");
2720 
2721 	uc_count = netdev_uc_count(dev);
2722 	mc_count = netdev_mc_count(dev);
2723 	/* MAC filter - use one buffer for both lists */
2724 	buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
2725 		      (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
2726 	mac_data = buf;
2727 	if (!buf)
2728 		return;
2729 
2730 	sg_init_table(sg, 2);
2731 
2732 	/* Store the unicast list and count in the front of the buffer */
2733 	mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
2734 	i = 0;
2735 	netdev_for_each_uc_addr(ha, dev)
2736 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2737 
2738 	sg_set_buf(&sg[0], mac_data,
2739 		   sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
2740 
2741 	/* multicast list and count fill the end */
2742 	mac_data = (void *)&mac_data->macs[uc_count][0];
2743 
2744 	mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
2745 	i = 0;
2746 	netdev_for_each_mc_addr(ha, dev)
2747 		memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2748 
2749 	sg_set_buf(&sg[1], mac_data,
2750 		   sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
2751 
2752 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2753 				  VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
2754 		dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
2755 
2756 	kfree(buf);
2757 }
2758 
virtnet_vlan_rx_add_vid(struct net_device * dev,__be16 proto,u16 vid)2759 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
2760 				   __be16 proto, u16 vid)
2761 {
2762 	struct virtnet_info *vi = netdev_priv(dev);
2763 	struct scatterlist sg;
2764 
2765 	vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2766 	sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2767 
2768 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2769 				  VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2770 		dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2771 	return 0;
2772 }
2773 
virtnet_vlan_rx_kill_vid(struct net_device * dev,__be16 proto,u16 vid)2774 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2775 				    __be16 proto, u16 vid)
2776 {
2777 	struct virtnet_info *vi = netdev_priv(dev);
2778 	struct scatterlist sg;
2779 
2780 	vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2781 	sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2782 
2783 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2784 				  VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2785 		dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2786 	return 0;
2787 }
2788 
virtnet_clean_affinity(struct virtnet_info * vi)2789 static void virtnet_clean_affinity(struct virtnet_info *vi)
2790 {
2791 	int i;
2792 
2793 	if (vi->affinity_hint_set) {
2794 		for (i = 0; i < vi->max_queue_pairs; i++) {
2795 			virtqueue_set_affinity(vi->rq[i].vq, NULL);
2796 			virtqueue_set_affinity(vi->sq[i].vq, NULL);
2797 		}
2798 
2799 		vi->affinity_hint_set = false;
2800 	}
2801 }
2802 
virtnet_set_affinity(struct virtnet_info * vi)2803 static void virtnet_set_affinity(struct virtnet_info *vi)
2804 {
2805 	cpumask_var_t mask;
2806 	int stragglers;
2807 	int group_size;
2808 	int i, j, cpu;
2809 	int num_cpu;
2810 	int stride;
2811 
2812 	if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2813 		virtnet_clean_affinity(vi);
2814 		return;
2815 	}
2816 
2817 	num_cpu = num_online_cpus();
2818 	stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2819 	stragglers = num_cpu >= vi->curr_queue_pairs ?
2820 			num_cpu % vi->curr_queue_pairs :
2821 			0;
2822 	cpu = cpumask_first(cpu_online_mask);
2823 
2824 	for (i = 0; i < vi->curr_queue_pairs; i++) {
2825 		group_size = stride + (i < stragglers ? 1 : 0);
2826 
2827 		for (j = 0; j < group_size; j++) {
2828 			cpumask_set_cpu(cpu, mask);
2829 			cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2830 						nr_cpu_ids, false);
2831 		}
2832 		virtqueue_set_affinity(vi->rq[i].vq, mask);
2833 		virtqueue_set_affinity(vi->sq[i].vq, mask);
2834 		__netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2835 		cpumask_clear(mask);
2836 	}
2837 
2838 	vi->affinity_hint_set = true;
2839 	free_cpumask_var(mask);
2840 }
2841 
virtnet_cpu_online(unsigned int cpu,struct hlist_node * node)2842 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2843 {
2844 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2845 						   node);
2846 	virtnet_set_affinity(vi);
2847 	return 0;
2848 }
2849 
virtnet_cpu_dead(unsigned int cpu,struct hlist_node * node)2850 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2851 {
2852 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2853 						   node_dead);
2854 	virtnet_set_affinity(vi);
2855 	return 0;
2856 }
2857 
virtnet_cpu_down_prep(unsigned int cpu,struct hlist_node * node)2858 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2859 {
2860 	struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2861 						   node);
2862 
2863 	virtnet_clean_affinity(vi);
2864 	return 0;
2865 }
2866 
2867 static enum cpuhp_state virtionet_online;
2868 
virtnet_cpu_notif_add(struct virtnet_info * vi)2869 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2870 {
2871 	int ret;
2872 
2873 	ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2874 	if (ret)
2875 		return ret;
2876 	ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2877 					       &vi->node_dead);
2878 	if (!ret)
2879 		return ret;
2880 	cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2881 	return ret;
2882 }
2883 
virtnet_cpu_notif_remove(struct virtnet_info * vi)2884 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2885 {
2886 	cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2887 	cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2888 					    &vi->node_dead);
2889 }
2890 
virtnet_get_ringparam(struct net_device * dev,struct ethtool_ringparam * ring,struct kernel_ethtool_ringparam * kernel_ring,struct netlink_ext_ack * extack)2891 static void virtnet_get_ringparam(struct net_device *dev,
2892 				  struct ethtool_ringparam *ring,
2893 				  struct kernel_ethtool_ringparam *kernel_ring,
2894 				  struct netlink_ext_ack *extack)
2895 {
2896 	struct virtnet_info *vi = netdev_priv(dev);
2897 
2898 	ring->rx_max_pending = vi->rq[0].vq->num_max;
2899 	ring->tx_max_pending = vi->sq[0].vq->num_max;
2900 	ring->rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2901 	ring->tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2902 }
2903 
2904 static int virtnet_send_ctrl_coal_vq_cmd(struct virtnet_info *vi,
2905 					 u16 vqn, u32 max_usecs, u32 max_packets);
2906 
virtnet_set_ringparam(struct net_device * dev,struct ethtool_ringparam * ring,struct kernel_ethtool_ringparam * kernel_ring,struct netlink_ext_ack * extack)2907 static int virtnet_set_ringparam(struct net_device *dev,
2908 				 struct ethtool_ringparam *ring,
2909 				 struct kernel_ethtool_ringparam *kernel_ring,
2910 				 struct netlink_ext_ack *extack)
2911 {
2912 	struct virtnet_info *vi = netdev_priv(dev);
2913 	u32 rx_pending, tx_pending;
2914 	struct receive_queue *rq;
2915 	struct send_queue *sq;
2916 	int i, err;
2917 
2918 	if (ring->rx_mini_pending || ring->rx_jumbo_pending)
2919 		return -EINVAL;
2920 
2921 	rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2922 	tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2923 
2924 	if (ring->rx_pending == rx_pending &&
2925 	    ring->tx_pending == tx_pending)
2926 		return 0;
2927 
2928 	if (ring->rx_pending > vi->rq[0].vq->num_max)
2929 		return -EINVAL;
2930 
2931 	if (ring->tx_pending > vi->sq[0].vq->num_max)
2932 		return -EINVAL;
2933 
2934 	for (i = 0; i < vi->max_queue_pairs; i++) {
2935 		rq = vi->rq + i;
2936 		sq = vi->sq + i;
2937 
2938 		if (ring->tx_pending != tx_pending) {
2939 			err = virtnet_tx_resize(vi, sq, ring->tx_pending);
2940 			if (err)
2941 				return err;
2942 
2943 			/* Upon disabling and re-enabling a transmit virtqueue, the device must
2944 			 * set the coalescing parameters of the virtqueue to those configured
2945 			 * through the VIRTIO_NET_CTRL_NOTF_COAL_TX_SET command, or, if the driver
2946 			 * did not set any TX coalescing parameters, to 0.
2947 			 */
2948 			err = virtnet_send_ctrl_coal_vq_cmd(vi, txq2vq(i),
2949 							    vi->intr_coal_tx.max_usecs,
2950 							    vi->intr_coal_tx.max_packets);
2951 			if (err)
2952 				return err;
2953 
2954 			vi->sq[i].intr_coal.max_usecs = vi->intr_coal_tx.max_usecs;
2955 			vi->sq[i].intr_coal.max_packets = vi->intr_coal_tx.max_packets;
2956 		}
2957 
2958 		if (ring->rx_pending != rx_pending) {
2959 			err = virtnet_rx_resize(vi, rq, ring->rx_pending);
2960 			if (err)
2961 				return err;
2962 
2963 			/* The reason is same as the transmit virtqueue reset */
2964 			err = virtnet_send_ctrl_coal_vq_cmd(vi, rxq2vq(i),
2965 							    vi->intr_coal_rx.max_usecs,
2966 							    vi->intr_coal_rx.max_packets);
2967 			if (err)
2968 				return err;
2969 
2970 			vi->rq[i].intr_coal.max_usecs = vi->intr_coal_rx.max_usecs;
2971 			vi->rq[i].intr_coal.max_packets = vi->intr_coal_rx.max_packets;
2972 		}
2973 	}
2974 
2975 	return 0;
2976 }
2977 
virtnet_commit_rss_command(struct virtnet_info * vi)2978 static bool virtnet_commit_rss_command(struct virtnet_info *vi)
2979 {
2980 	struct net_device *dev = vi->dev;
2981 	struct scatterlist sgs[4];
2982 	unsigned int sg_buf_size;
2983 
2984 	/* prepare sgs */
2985 	sg_init_table(sgs, 4);
2986 
2987 	sg_buf_size = offsetof(struct virtio_net_ctrl_rss, indirection_table);
2988 	sg_set_buf(&sgs[0], &vi->ctrl->rss, sg_buf_size);
2989 
2990 	sg_buf_size = sizeof(uint16_t) * (vi->ctrl->rss.indirection_table_mask + 1);
2991 	sg_set_buf(&sgs[1], vi->ctrl->rss.indirection_table, sg_buf_size);
2992 
2993 	sg_buf_size = offsetof(struct virtio_net_ctrl_rss, key)
2994 			- offsetof(struct virtio_net_ctrl_rss, max_tx_vq);
2995 	sg_set_buf(&sgs[2], &vi->ctrl->rss.max_tx_vq, sg_buf_size);
2996 
2997 	sg_buf_size = vi->rss_key_size;
2998 	sg_set_buf(&sgs[3], vi->ctrl->rss.key, sg_buf_size);
2999 
3000 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
3001 				  vi->has_rss ? VIRTIO_NET_CTRL_MQ_RSS_CONFIG
3002 				  : VIRTIO_NET_CTRL_MQ_HASH_CONFIG, sgs)) {
3003 		dev_warn(&dev->dev, "VIRTIONET issue with committing RSS sgs\n");
3004 		return false;
3005 	}
3006 	return true;
3007 }
3008 
virtnet_init_default_rss(struct virtnet_info * vi)3009 static void virtnet_init_default_rss(struct virtnet_info *vi)
3010 {
3011 	u32 indir_val = 0;
3012 	int i = 0;
3013 
3014 	vi->ctrl->rss.hash_types = vi->rss_hash_types_supported;
3015 	vi->rss_hash_types_saved = vi->rss_hash_types_supported;
3016 	vi->ctrl->rss.indirection_table_mask = vi->rss_indir_table_size
3017 						? vi->rss_indir_table_size - 1 : 0;
3018 	vi->ctrl->rss.unclassified_queue = 0;
3019 
3020 	for (; i < vi->rss_indir_table_size; ++i) {
3021 		indir_val = ethtool_rxfh_indir_default(i, vi->curr_queue_pairs);
3022 		vi->ctrl->rss.indirection_table[i] = indir_val;
3023 	}
3024 
3025 	vi->ctrl->rss.max_tx_vq = vi->has_rss ? vi->curr_queue_pairs : 0;
3026 	vi->ctrl->rss.hash_key_length = vi->rss_key_size;
3027 
3028 	netdev_rss_key_fill(vi->ctrl->rss.key, vi->rss_key_size);
3029 }
3030 
virtnet_get_hashflow(const struct virtnet_info * vi,struct ethtool_rxnfc * info)3031 static void virtnet_get_hashflow(const struct virtnet_info *vi, struct ethtool_rxnfc *info)
3032 {
3033 	info->data = 0;
3034 	switch (info->flow_type) {
3035 	case TCP_V4_FLOW:
3036 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv4) {
3037 			info->data = RXH_IP_SRC | RXH_IP_DST |
3038 						 RXH_L4_B_0_1 | RXH_L4_B_2_3;
3039 		} else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
3040 			info->data = RXH_IP_SRC | RXH_IP_DST;
3041 		}
3042 		break;
3043 	case TCP_V6_FLOW:
3044 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv6) {
3045 			info->data = RXH_IP_SRC | RXH_IP_DST |
3046 						 RXH_L4_B_0_1 | RXH_L4_B_2_3;
3047 		} else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
3048 			info->data = RXH_IP_SRC | RXH_IP_DST;
3049 		}
3050 		break;
3051 	case UDP_V4_FLOW:
3052 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv4) {
3053 			info->data = RXH_IP_SRC | RXH_IP_DST |
3054 						 RXH_L4_B_0_1 | RXH_L4_B_2_3;
3055 		} else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
3056 			info->data = RXH_IP_SRC | RXH_IP_DST;
3057 		}
3058 		break;
3059 	case UDP_V6_FLOW:
3060 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv6) {
3061 			info->data = RXH_IP_SRC | RXH_IP_DST |
3062 						 RXH_L4_B_0_1 | RXH_L4_B_2_3;
3063 		} else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
3064 			info->data = RXH_IP_SRC | RXH_IP_DST;
3065 		}
3066 		break;
3067 	case IPV4_FLOW:
3068 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4)
3069 			info->data = RXH_IP_SRC | RXH_IP_DST;
3070 
3071 		break;
3072 	case IPV6_FLOW:
3073 		if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6)
3074 			info->data = RXH_IP_SRC | RXH_IP_DST;
3075 
3076 		break;
3077 	default:
3078 		info->data = 0;
3079 		break;
3080 	}
3081 }
3082 
virtnet_set_hashflow(struct virtnet_info * vi,struct ethtool_rxnfc * info)3083 static bool virtnet_set_hashflow(struct virtnet_info *vi, struct ethtool_rxnfc *info)
3084 {
3085 	u32 new_hashtypes = vi->rss_hash_types_saved;
3086 	bool is_disable = info->data & RXH_DISCARD;
3087 	bool is_l4 = info->data == (RXH_IP_SRC | RXH_IP_DST | RXH_L4_B_0_1 | RXH_L4_B_2_3);
3088 
3089 	/* supports only 'sd', 'sdfn' and 'r' */
3090 	if (!((info->data == (RXH_IP_SRC | RXH_IP_DST)) | is_l4 | is_disable))
3091 		return false;
3092 
3093 	switch (info->flow_type) {
3094 	case TCP_V4_FLOW:
3095 		new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_TCPv4);
3096 		if (!is_disable)
3097 			new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
3098 				| (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv4 : 0);
3099 		break;
3100 	case UDP_V4_FLOW:
3101 		new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_UDPv4);
3102 		if (!is_disable)
3103 			new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
3104 				| (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv4 : 0);
3105 		break;
3106 	case IPV4_FLOW:
3107 		new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv4;
3108 		if (!is_disable)
3109 			new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv4;
3110 		break;
3111 	case TCP_V6_FLOW:
3112 		new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_TCPv6);
3113 		if (!is_disable)
3114 			new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
3115 				| (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv6 : 0);
3116 		break;
3117 	case UDP_V6_FLOW:
3118 		new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_UDPv6);
3119 		if (!is_disable)
3120 			new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
3121 				| (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv6 : 0);
3122 		break;
3123 	case IPV6_FLOW:
3124 		new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv6;
3125 		if (!is_disable)
3126 			new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv6;
3127 		break;
3128 	default:
3129 		/* unsupported flow */
3130 		return false;
3131 	}
3132 
3133 	/* if unsupported hashtype was set */
3134 	if (new_hashtypes != (new_hashtypes & vi->rss_hash_types_supported))
3135 		return false;
3136 
3137 	if (new_hashtypes != vi->rss_hash_types_saved) {
3138 		vi->rss_hash_types_saved = new_hashtypes;
3139 		vi->ctrl->rss.hash_types = vi->rss_hash_types_saved;
3140 		if (vi->dev->features & NETIF_F_RXHASH)
3141 			return virtnet_commit_rss_command(vi);
3142 	}
3143 
3144 	return true;
3145 }
3146 
virtnet_get_drvinfo(struct net_device * dev,struct ethtool_drvinfo * info)3147 static void virtnet_get_drvinfo(struct net_device *dev,
3148 				struct ethtool_drvinfo *info)
3149 {
3150 	struct virtnet_info *vi = netdev_priv(dev);
3151 	struct virtio_device *vdev = vi->vdev;
3152 
3153 	strscpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
3154 	strscpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
3155 	strscpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
3156 
3157 }
3158 
3159 /* TODO: Eliminate OOO packets during switching */
virtnet_set_channels(struct net_device * dev,struct ethtool_channels * channels)3160 static int virtnet_set_channels(struct net_device *dev,
3161 				struct ethtool_channels *channels)
3162 {
3163 	struct virtnet_info *vi = netdev_priv(dev);
3164 	u16 queue_pairs = channels->combined_count;
3165 	int err;
3166 
3167 	/* We don't support separate rx/tx channels.
3168 	 * We don't allow setting 'other' channels.
3169 	 */
3170 	if (channels->rx_count || channels->tx_count || channels->other_count)
3171 		return -EINVAL;
3172 
3173 	if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
3174 		return -EINVAL;
3175 
3176 	/* For now we don't support modifying channels while XDP is loaded
3177 	 * also when XDP is loaded all RX queues have XDP programs so we only
3178 	 * need to check a single RX queue.
3179 	 */
3180 	if (vi->rq[0].xdp_prog)
3181 		return -EINVAL;
3182 
3183 	cpus_read_lock();
3184 	err = _virtnet_set_queues(vi, queue_pairs);
3185 	if (err) {
3186 		cpus_read_unlock();
3187 		goto err;
3188 	}
3189 	virtnet_set_affinity(vi);
3190 	cpus_read_unlock();
3191 
3192 	netif_set_real_num_tx_queues(dev, queue_pairs);
3193 	netif_set_real_num_rx_queues(dev, queue_pairs);
3194  err:
3195 	return err;
3196 }
3197 
virtnet_get_strings(struct net_device * dev,u32 stringset,u8 * data)3198 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
3199 {
3200 	struct virtnet_info *vi = netdev_priv(dev);
3201 	unsigned int i, j;
3202 	u8 *p = data;
3203 
3204 	switch (stringset) {
3205 	case ETH_SS_STATS:
3206 		for (i = 0; i < vi->curr_queue_pairs; i++) {
3207 			for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
3208 				ethtool_sprintf(&p, "rx_queue_%u_%s", i,
3209 						virtnet_rq_stats_desc[j].desc);
3210 		}
3211 
3212 		for (i = 0; i < vi->curr_queue_pairs; i++) {
3213 			for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
3214 				ethtool_sprintf(&p, "tx_queue_%u_%s", i,
3215 						virtnet_sq_stats_desc[j].desc);
3216 		}
3217 		break;
3218 	}
3219 }
3220 
virtnet_get_sset_count(struct net_device * dev,int sset)3221 static int virtnet_get_sset_count(struct net_device *dev, int sset)
3222 {
3223 	struct virtnet_info *vi = netdev_priv(dev);
3224 
3225 	switch (sset) {
3226 	case ETH_SS_STATS:
3227 		return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
3228 					       VIRTNET_SQ_STATS_LEN);
3229 	default:
3230 		return -EOPNOTSUPP;
3231 	}
3232 }
3233 
virtnet_get_ethtool_stats(struct net_device * dev,struct ethtool_stats * stats,u64 * data)3234 static void virtnet_get_ethtool_stats(struct net_device *dev,
3235 				      struct ethtool_stats *stats, u64 *data)
3236 {
3237 	struct virtnet_info *vi = netdev_priv(dev);
3238 	unsigned int idx = 0, start, i, j;
3239 	const u8 *stats_base;
3240 	const u64_stats_t *p;
3241 	size_t offset;
3242 
3243 	for (i = 0; i < vi->curr_queue_pairs; i++) {
3244 		struct receive_queue *rq = &vi->rq[i];
3245 
3246 		stats_base = (const u8 *)&rq->stats;
3247 		do {
3248 			start = u64_stats_fetch_begin(&rq->stats.syncp);
3249 			for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
3250 				offset = virtnet_rq_stats_desc[j].offset;
3251 				p = (const u64_stats_t *)(stats_base + offset);
3252 				data[idx + j] = u64_stats_read(p);
3253 			}
3254 		} while (u64_stats_fetch_retry(&rq->stats.syncp, start));
3255 		idx += VIRTNET_RQ_STATS_LEN;
3256 	}
3257 
3258 	for (i = 0; i < vi->curr_queue_pairs; i++) {
3259 		struct send_queue *sq = &vi->sq[i];
3260 
3261 		stats_base = (const u8 *)&sq->stats;
3262 		do {
3263 			start = u64_stats_fetch_begin(&sq->stats.syncp);
3264 			for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
3265 				offset = virtnet_sq_stats_desc[j].offset;
3266 				p = (const u64_stats_t *)(stats_base + offset);
3267 				data[idx + j] = u64_stats_read(p);
3268 			}
3269 		} while (u64_stats_fetch_retry(&sq->stats.syncp, start));
3270 		idx += VIRTNET_SQ_STATS_LEN;
3271 	}
3272 }
3273 
virtnet_get_channels(struct net_device * dev,struct ethtool_channels * channels)3274 static void virtnet_get_channels(struct net_device *dev,
3275 				 struct ethtool_channels *channels)
3276 {
3277 	struct virtnet_info *vi = netdev_priv(dev);
3278 
3279 	channels->combined_count = vi->curr_queue_pairs;
3280 	channels->max_combined = vi->max_queue_pairs;
3281 	channels->max_other = 0;
3282 	channels->rx_count = 0;
3283 	channels->tx_count = 0;
3284 	channels->other_count = 0;
3285 }
3286 
virtnet_set_link_ksettings(struct net_device * dev,const struct ethtool_link_ksettings * cmd)3287 static int virtnet_set_link_ksettings(struct net_device *dev,
3288 				      const struct ethtool_link_ksettings *cmd)
3289 {
3290 	struct virtnet_info *vi = netdev_priv(dev);
3291 
3292 	return ethtool_virtdev_set_link_ksettings(dev, cmd,
3293 						  &vi->speed, &vi->duplex);
3294 }
3295 
virtnet_get_link_ksettings(struct net_device * dev,struct ethtool_link_ksettings * cmd)3296 static int virtnet_get_link_ksettings(struct net_device *dev,
3297 				      struct ethtool_link_ksettings *cmd)
3298 {
3299 	struct virtnet_info *vi = netdev_priv(dev);
3300 
3301 	cmd->base.speed = vi->speed;
3302 	cmd->base.duplex = vi->duplex;
3303 	cmd->base.port = PORT_OTHER;
3304 
3305 	return 0;
3306 }
3307 
virtnet_send_notf_coal_cmds(struct virtnet_info * vi,struct ethtool_coalesce * ec)3308 static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
3309 				       struct ethtool_coalesce *ec)
3310 {
3311 	struct scatterlist sgs_tx, sgs_rx;
3312 	int i;
3313 
3314 	vi->ctrl->coal_tx.tx_usecs = cpu_to_le32(ec->tx_coalesce_usecs);
3315 	vi->ctrl->coal_tx.tx_max_packets = cpu_to_le32(ec->tx_max_coalesced_frames);
3316 	sg_init_one(&sgs_tx, &vi->ctrl->coal_tx, sizeof(vi->ctrl->coal_tx));
3317 
3318 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
3319 				  VIRTIO_NET_CTRL_NOTF_COAL_TX_SET,
3320 				  &sgs_tx))
3321 		return -EINVAL;
3322 
3323 	/* Save parameters */
3324 	vi->intr_coal_tx.max_usecs = ec->tx_coalesce_usecs;
3325 	vi->intr_coal_tx.max_packets = ec->tx_max_coalesced_frames;
3326 	for (i = 0; i < vi->max_queue_pairs; i++) {
3327 		vi->sq[i].intr_coal.max_usecs = ec->tx_coalesce_usecs;
3328 		vi->sq[i].intr_coal.max_packets = ec->tx_max_coalesced_frames;
3329 	}
3330 
3331 	vi->ctrl->coal_rx.rx_usecs = cpu_to_le32(ec->rx_coalesce_usecs);
3332 	vi->ctrl->coal_rx.rx_max_packets = cpu_to_le32(ec->rx_max_coalesced_frames);
3333 	sg_init_one(&sgs_rx, &vi->ctrl->coal_rx, sizeof(vi->ctrl->coal_rx));
3334 
3335 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
3336 				  VIRTIO_NET_CTRL_NOTF_COAL_RX_SET,
3337 				  &sgs_rx))
3338 		return -EINVAL;
3339 
3340 	/* Save parameters */
3341 	vi->intr_coal_rx.max_usecs = ec->rx_coalesce_usecs;
3342 	vi->intr_coal_rx.max_packets = ec->rx_max_coalesced_frames;
3343 	for (i = 0; i < vi->max_queue_pairs; i++) {
3344 		vi->rq[i].intr_coal.max_usecs = ec->rx_coalesce_usecs;
3345 		vi->rq[i].intr_coal.max_packets = ec->rx_max_coalesced_frames;
3346 	}
3347 
3348 	return 0;
3349 }
3350 
virtnet_send_ctrl_coal_vq_cmd(struct virtnet_info * vi,u16 vqn,u32 max_usecs,u32 max_packets)3351 static int virtnet_send_ctrl_coal_vq_cmd(struct virtnet_info *vi,
3352 					 u16 vqn, u32 max_usecs, u32 max_packets)
3353 {
3354 	struct scatterlist sgs;
3355 
3356 	vi->ctrl->coal_vq.vqn = cpu_to_le16(vqn);
3357 	vi->ctrl->coal_vq.coal.max_usecs = cpu_to_le32(max_usecs);
3358 	vi->ctrl->coal_vq.coal.max_packets = cpu_to_le32(max_packets);
3359 	sg_init_one(&sgs, &vi->ctrl->coal_vq, sizeof(vi->ctrl->coal_vq));
3360 
3361 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
3362 				  VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET,
3363 				  &sgs))
3364 		return -EINVAL;
3365 
3366 	return 0;
3367 }
3368 
virtnet_send_notf_coal_vq_cmds(struct virtnet_info * vi,struct ethtool_coalesce * ec,u16 queue)3369 static int virtnet_send_notf_coal_vq_cmds(struct virtnet_info *vi,
3370 					  struct ethtool_coalesce *ec,
3371 					  u16 queue)
3372 {
3373 	int err;
3374 
3375 	err = virtnet_send_ctrl_coal_vq_cmd(vi, rxq2vq(queue),
3376 					    ec->rx_coalesce_usecs,
3377 					    ec->rx_max_coalesced_frames);
3378 	if (err)
3379 		return err;
3380 
3381 	vi->rq[queue].intr_coal.max_usecs = ec->rx_coalesce_usecs;
3382 	vi->rq[queue].intr_coal.max_packets = ec->rx_max_coalesced_frames;
3383 
3384 	err = virtnet_send_ctrl_coal_vq_cmd(vi, txq2vq(queue),
3385 					    ec->tx_coalesce_usecs,
3386 					    ec->tx_max_coalesced_frames);
3387 	if (err)
3388 		return err;
3389 
3390 	vi->sq[queue].intr_coal.max_usecs = ec->tx_coalesce_usecs;
3391 	vi->sq[queue].intr_coal.max_packets = ec->tx_max_coalesced_frames;
3392 
3393 	return 0;
3394 }
3395 
virtnet_coal_params_supported(struct ethtool_coalesce * ec)3396 static int virtnet_coal_params_supported(struct ethtool_coalesce *ec)
3397 {
3398 	/* usecs coalescing is supported only if VIRTIO_NET_F_NOTF_COAL
3399 	 * feature is negotiated.
3400 	 */
3401 	if (ec->rx_coalesce_usecs || ec->tx_coalesce_usecs)
3402 		return -EOPNOTSUPP;
3403 
3404 	if (ec->tx_max_coalesced_frames > 1 ||
3405 	    ec->rx_max_coalesced_frames != 1)
3406 		return -EINVAL;
3407 
3408 	return 0;
3409 }
3410 
virtnet_should_update_vq_weight(int dev_flags,int weight,int vq_weight,bool * should_update)3411 static int virtnet_should_update_vq_weight(int dev_flags, int weight,
3412 					   int vq_weight, bool *should_update)
3413 {
3414 	if (weight ^ vq_weight) {
3415 		if (dev_flags & IFF_UP)
3416 			return -EBUSY;
3417 		*should_update = true;
3418 	}
3419 
3420 	return 0;
3421 }
3422 
virtnet_set_coalesce(struct net_device * dev,struct ethtool_coalesce * ec,struct kernel_ethtool_coalesce * kernel_coal,struct netlink_ext_ack * extack)3423 static int virtnet_set_coalesce(struct net_device *dev,
3424 				struct ethtool_coalesce *ec,
3425 				struct kernel_ethtool_coalesce *kernel_coal,
3426 				struct netlink_ext_ack *extack)
3427 {
3428 	struct virtnet_info *vi = netdev_priv(dev);
3429 	int ret, queue_number, napi_weight;
3430 	bool update_napi = false;
3431 
3432 	/* Can't change NAPI weight if the link is up */
3433 	napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
3434 	for (queue_number = 0; queue_number < vi->max_queue_pairs; queue_number++) {
3435 		ret = virtnet_should_update_vq_weight(dev->flags, napi_weight,
3436 						      vi->sq[queue_number].napi.weight,
3437 						      &update_napi);
3438 		if (ret)
3439 			return ret;
3440 
3441 		if (update_napi) {
3442 			/* All queues that belong to [queue_number, vi->max_queue_pairs] will be
3443 			 * updated for the sake of simplicity, which might not be necessary
3444 			 */
3445 			break;
3446 		}
3447 	}
3448 
3449 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL))
3450 		ret = virtnet_send_notf_coal_cmds(vi, ec);
3451 	else
3452 		ret = virtnet_coal_params_supported(ec);
3453 
3454 	if (ret)
3455 		return ret;
3456 
3457 	if (update_napi) {
3458 		for (; queue_number < vi->max_queue_pairs; queue_number++)
3459 			vi->sq[queue_number].napi.weight = napi_weight;
3460 	}
3461 
3462 	return ret;
3463 }
3464 
virtnet_get_coalesce(struct net_device * dev,struct ethtool_coalesce * ec,struct kernel_ethtool_coalesce * kernel_coal,struct netlink_ext_ack * extack)3465 static int virtnet_get_coalesce(struct net_device *dev,
3466 				struct ethtool_coalesce *ec,
3467 				struct kernel_ethtool_coalesce *kernel_coal,
3468 				struct netlink_ext_ack *extack)
3469 {
3470 	struct virtnet_info *vi = netdev_priv(dev);
3471 
3472 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
3473 		ec->rx_coalesce_usecs = vi->intr_coal_rx.max_usecs;
3474 		ec->tx_coalesce_usecs = vi->intr_coal_tx.max_usecs;
3475 		ec->tx_max_coalesced_frames = vi->intr_coal_tx.max_packets;
3476 		ec->rx_max_coalesced_frames = vi->intr_coal_rx.max_packets;
3477 	} else {
3478 		ec->rx_max_coalesced_frames = 1;
3479 
3480 		if (vi->sq[0].napi.weight)
3481 			ec->tx_max_coalesced_frames = 1;
3482 	}
3483 
3484 	return 0;
3485 }
3486 
virtnet_set_per_queue_coalesce(struct net_device * dev,u32 queue,struct ethtool_coalesce * ec)3487 static int virtnet_set_per_queue_coalesce(struct net_device *dev,
3488 					  u32 queue,
3489 					  struct ethtool_coalesce *ec)
3490 {
3491 	struct virtnet_info *vi = netdev_priv(dev);
3492 	int ret, napi_weight;
3493 	bool update_napi = false;
3494 
3495 	if (queue >= vi->max_queue_pairs)
3496 		return -EINVAL;
3497 
3498 	/* Can't change NAPI weight if the link is up */
3499 	napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
3500 	ret = virtnet_should_update_vq_weight(dev->flags, napi_weight,
3501 					      vi->sq[queue].napi.weight,
3502 					      &update_napi);
3503 	if (ret)
3504 		return ret;
3505 
3506 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL))
3507 		ret = virtnet_send_notf_coal_vq_cmds(vi, ec, queue);
3508 	else
3509 		ret = virtnet_coal_params_supported(ec);
3510 
3511 	if (ret)
3512 		return ret;
3513 
3514 	if (update_napi)
3515 		vi->sq[queue].napi.weight = napi_weight;
3516 
3517 	return 0;
3518 }
3519 
virtnet_get_per_queue_coalesce(struct net_device * dev,u32 queue,struct ethtool_coalesce * ec)3520 static int virtnet_get_per_queue_coalesce(struct net_device *dev,
3521 					  u32 queue,
3522 					  struct ethtool_coalesce *ec)
3523 {
3524 	struct virtnet_info *vi = netdev_priv(dev);
3525 
3526 	if (queue >= vi->max_queue_pairs)
3527 		return -EINVAL;
3528 
3529 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL)) {
3530 		ec->rx_coalesce_usecs = vi->rq[queue].intr_coal.max_usecs;
3531 		ec->tx_coalesce_usecs = vi->sq[queue].intr_coal.max_usecs;
3532 		ec->tx_max_coalesced_frames = vi->sq[queue].intr_coal.max_packets;
3533 		ec->rx_max_coalesced_frames = vi->rq[queue].intr_coal.max_packets;
3534 	} else {
3535 		ec->rx_max_coalesced_frames = 1;
3536 
3537 		if (vi->sq[queue].napi.weight)
3538 			ec->tx_max_coalesced_frames = 1;
3539 	}
3540 
3541 	return 0;
3542 }
3543 
virtnet_init_settings(struct net_device * dev)3544 static void virtnet_init_settings(struct net_device *dev)
3545 {
3546 	struct virtnet_info *vi = netdev_priv(dev);
3547 
3548 	vi->speed = SPEED_UNKNOWN;
3549 	vi->duplex = DUPLEX_UNKNOWN;
3550 }
3551 
virtnet_update_settings(struct virtnet_info * vi)3552 static void virtnet_update_settings(struct virtnet_info *vi)
3553 {
3554 	u32 speed;
3555 	u8 duplex;
3556 
3557 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
3558 		return;
3559 
3560 	virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
3561 
3562 	if (ethtool_validate_speed(speed))
3563 		vi->speed = speed;
3564 
3565 	virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
3566 
3567 	if (ethtool_validate_duplex(duplex))
3568 		vi->duplex = duplex;
3569 }
3570 
virtnet_get_rxfh_key_size(struct net_device * dev)3571 static u32 virtnet_get_rxfh_key_size(struct net_device *dev)
3572 {
3573 	return ((struct virtnet_info *)netdev_priv(dev))->rss_key_size;
3574 }
3575 
virtnet_get_rxfh_indir_size(struct net_device * dev)3576 static u32 virtnet_get_rxfh_indir_size(struct net_device *dev)
3577 {
3578 	return ((struct virtnet_info *)netdev_priv(dev))->rss_indir_table_size;
3579 }
3580 
virtnet_get_rxfh(struct net_device * dev,u32 * indir,u8 * key,u8 * hfunc)3581 static int virtnet_get_rxfh(struct net_device *dev, u32 *indir, u8 *key, u8 *hfunc)
3582 {
3583 	struct virtnet_info *vi = netdev_priv(dev);
3584 	int i;
3585 
3586 	if (indir) {
3587 		for (i = 0; i < vi->rss_indir_table_size; ++i)
3588 			indir[i] = vi->ctrl->rss.indirection_table[i];
3589 	}
3590 
3591 	if (key)
3592 		memcpy(key, vi->ctrl->rss.key, vi->rss_key_size);
3593 
3594 	if (hfunc)
3595 		*hfunc = ETH_RSS_HASH_TOP;
3596 
3597 	return 0;
3598 }
3599 
virtnet_set_rxfh(struct net_device * dev,const u32 * indir,const u8 * key,const u8 hfunc)3600 static int virtnet_set_rxfh(struct net_device *dev, const u32 *indir, const u8 *key, const u8 hfunc)
3601 {
3602 	struct virtnet_info *vi = netdev_priv(dev);
3603 	bool update = false;
3604 	int i;
3605 
3606 	if (hfunc != ETH_RSS_HASH_NO_CHANGE && hfunc != ETH_RSS_HASH_TOP)
3607 		return -EOPNOTSUPP;
3608 
3609 	if (indir) {
3610 		if (!vi->has_rss)
3611 			return -EOPNOTSUPP;
3612 
3613 		for (i = 0; i < vi->rss_indir_table_size; ++i)
3614 			vi->ctrl->rss.indirection_table[i] = indir[i];
3615 		update = true;
3616 	}
3617 	if (key) {
3618 		/* If either _F_HASH_REPORT or _F_RSS are negotiated, the
3619 		 * device provides hash calculation capabilities, that is,
3620 		 * hash_key is configured.
3621 		 */
3622 		if (!vi->has_rss && !vi->has_rss_hash_report)
3623 			return -EOPNOTSUPP;
3624 
3625 		memcpy(vi->ctrl->rss.key, key, vi->rss_key_size);
3626 		update = true;
3627 	}
3628 
3629 	if (update)
3630 		virtnet_commit_rss_command(vi);
3631 
3632 	return 0;
3633 }
3634 
virtnet_get_rxnfc(struct net_device * dev,struct ethtool_rxnfc * info,u32 * rule_locs)3635 static int virtnet_get_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info, u32 *rule_locs)
3636 {
3637 	struct virtnet_info *vi = netdev_priv(dev);
3638 	int rc = 0;
3639 
3640 	switch (info->cmd) {
3641 	case ETHTOOL_GRXRINGS:
3642 		info->data = vi->curr_queue_pairs;
3643 		break;
3644 	case ETHTOOL_GRXFH:
3645 		virtnet_get_hashflow(vi, info);
3646 		break;
3647 	default:
3648 		rc = -EOPNOTSUPP;
3649 	}
3650 
3651 	return rc;
3652 }
3653 
virtnet_set_rxnfc(struct net_device * dev,struct ethtool_rxnfc * info)3654 static int virtnet_set_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info)
3655 {
3656 	struct virtnet_info *vi = netdev_priv(dev);
3657 	int rc = 0;
3658 
3659 	switch (info->cmd) {
3660 	case ETHTOOL_SRXFH:
3661 		if (!virtnet_set_hashflow(vi, info))
3662 			rc = -EINVAL;
3663 
3664 		break;
3665 	default:
3666 		rc = -EOPNOTSUPP;
3667 	}
3668 
3669 	return rc;
3670 }
3671 
3672 static const struct ethtool_ops virtnet_ethtool_ops = {
3673 	.supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES |
3674 		ETHTOOL_COALESCE_USECS,
3675 	.get_drvinfo = virtnet_get_drvinfo,
3676 	.get_link = ethtool_op_get_link,
3677 	.get_ringparam = virtnet_get_ringparam,
3678 	.set_ringparam = virtnet_set_ringparam,
3679 	.get_strings = virtnet_get_strings,
3680 	.get_sset_count = virtnet_get_sset_count,
3681 	.get_ethtool_stats = virtnet_get_ethtool_stats,
3682 	.set_channels = virtnet_set_channels,
3683 	.get_channels = virtnet_get_channels,
3684 	.get_ts_info = ethtool_op_get_ts_info,
3685 	.get_link_ksettings = virtnet_get_link_ksettings,
3686 	.set_link_ksettings = virtnet_set_link_ksettings,
3687 	.set_coalesce = virtnet_set_coalesce,
3688 	.get_coalesce = virtnet_get_coalesce,
3689 	.set_per_queue_coalesce = virtnet_set_per_queue_coalesce,
3690 	.get_per_queue_coalesce = virtnet_get_per_queue_coalesce,
3691 	.get_rxfh_key_size = virtnet_get_rxfh_key_size,
3692 	.get_rxfh_indir_size = virtnet_get_rxfh_indir_size,
3693 	.get_rxfh = virtnet_get_rxfh,
3694 	.set_rxfh = virtnet_set_rxfh,
3695 	.get_rxnfc = virtnet_get_rxnfc,
3696 	.set_rxnfc = virtnet_set_rxnfc,
3697 };
3698 
virtnet_freeze_down(struct virtio_device * vdev)3699 static void virtnet_freeze_down(struct virtio_device *vdev)
3700 {
3701 	struct virtnet_info *vi = vdev->priv;
3702 
3703 	/* Make sure no work handler is accessing the device */
3704 	flush_work(&vi->config_work);
3705 
3706 	netif_tx_lock_bh(vi->dev);
3707 	netif_device_detach(vi->dev);
3708 	netif_tx_unlock_bh(vi->dev);
3709 	if (netif_running(vi->dev))
3710 		virtnet_close(vi->dev);
3711 }
3712 
3713 static int init_vqs(struct virtnet_info *vi);
3714 
virtnet_restore_up(struct virtio_device * vdev)3715 static int virtnet_restore_up(struct virtio_device *vdev)
3716 {
3717 	struct virtnet_info *vi = vdev->priv;
3718 	int err;
3719 
3720 	err = init_vqs(vi);
3721 	if (err)
3722 		return err;
3723 
3724 	virtio_device_ready(vdev);
3725 
3726 	enable_delayed_refill(vi);
3727 
3728 	if (netif_running(vi->dev)) {
3729 		err = virtnet_open(vi->dev);
3730 		if (err)
3731 			return err;
3732 	}
3733 
3734 	netif_tx_lock_bh(vi->dev);
3735 	netif_device_attach(vi->dev);
3736 	netif_tx_unlock_bh(vi->dev);
3737 	return err;
3738 }
3739 
virtnet_set_guest_offloads(struct virtnet_info * vi,u64 offloads)3740 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
3741 {
3742 	struct scatterlist sg;
3743 	vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
3744 
3745 	sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
3746 
3747 	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
3748 				  VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
3749 		dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
3750 		return -EINVAL;
3751 	}
3752 
3753 	return 0;
3754 }
3755 
virtnet_clear_guest_offloads(struct virtnet_info * vi)3756 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
3757 {
3758 	u64 offloads = 0;
3759 
3760 	if (!vi->guest_offloads)
3761 		return 0;
3762 
3763 	return virtnet_set_guest_offloads(vi, offloads);
3764 }
3765 
virtnet_restore_guest_offloads(struct virtnet_info * vi)3766 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
3767 {
3768 	u64 offloads = vi->guest_offloads;
3769 
3770 	if (!vi->guest_offloads)
3771 		return 0;
3772 
3773 	return virtnet_set_guest_offloads(vi, offloads);
3774 }
3775 
virtnet_xdp_set(struct net_device * dev,struct bpf_prog * prog,struct netlink_ext_ack * extack)3776 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
3777 			   struct netlink_ext_ack *extack)
3778 {
3779 	unsigned int room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM +
3780 					   sizeof(struct skb_shared_info));
3781 	unsigned int max_sz = PAGE_SIZE - room - ETH_HLEN;
3782 	struct virtnet_info *vi = netdev_priv(dev);
3783 	struct bpf_prog *old_prog;
3784 	u16 xdp_qp = 0, curr_qp;
3785 	int i, err;
3786 
3787 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
3788 	    && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3789 	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3790 	        virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
3791 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
3792 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM) ||
3793 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) ||
3794 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6))) {
3795 		NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
3796 		return -EOPNOTSUPP;
3797 	}
3798 
3799 	if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
3800 		NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
3801 		return -EINVAL;
3802 	}
3803 
3804 	if (prog && !prog->aux->xdp_has_frags && dev->mtu > max_sz) {
3805 		NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP without frags");
3806 		netdev_warn(dev, "single-buffer XDP requires MTU less than %u\n", max_sz);
3807 		return -EINVAL;
3808 	}
3809 
3810 	curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
3811 	if (prog)
3812 		xdp_qp = nr_cpu_ids;
3813 
3814 	/* XDP requires extra queues for XDP_TX */
3815 	if (curr_qp + xdp_qp > vi->max_queue_pairs) {
3816 		netdev_warn_once(dev, "XDP request %i queues but max is %i. XDP_TX and XDP_REDIRECT will operate in a slower locked tx mode.\n",
3817 				 curr_qp + xdp_qp, vi->max_queue_pairs);
3818 		xdp_qp = 0;
3819 	}
3820 
3821 	old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
3822 	if (!prog && !old_prog)
3823 		return 0;
3824 
3825 	if (prog)
3826 		bpf_prog_add(prog, vi->max_queue_pairs - 1);
3827 
3828 	/* Make sure NAPI is not using any XDP TX queues for RX. */
3829 	if (netif_running(dev)) {
3830 		for (i = 0; i < vi->max_queue_pairs; i++) {
3831 			napi_disable(&vi->rq[i].napi);
3832 			virtnet_napi_tx_disable(&vi->sq[i].napi);
3833 		}
3834 	}
3835 
3836 	if (!prog) {
3837 		for (i = 0; i < vi->max_queue_pairs; i++) {
3838 			rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
3839 			if (i == 0)
3840 				virtnet_restore_guest_offloads(vi);
3841 		}
3842 		synchronize_net();
3843 	}
3844 
3845 	err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
3846 	if (err)
3847 		goto err;
3848 	netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
3849 	vi->xdp_queue_pairs = xdp_qp;
3850 
3851 	if (prog) {
3852 		vi->xdp_enabled = true;
3853 		for (i = 0; i < vi->max_queue_pairs; i++) {
3854 			rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
3855 			if (i == 0 && !old_prog)
3856 				virtnet_clear_guest_offloads(vi);
3857 		}
3858 		if (!old_prog)
3859 			xdp_features_set_redirect_target(dev, true);
3860 	} else {
3861 		xdp_features_clear_redirect_target(dev);
3862 		vi->xdp_enabled = false;
3863 	}
3864 
3865 	for (i = 0; i < vi->max_queue_pairs; i++) {
3866 		if (old_prog)
3867 			bpf_prog_put(old_prog);
3868 		if (netif_running(dev)) {
3869 			virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
3870 			virtnet_napi_tx_enable(vi, vi->sq[i].vq,
3871 					       &vi->sq[i].napi);
3872 		}
3873 	}
3874 
3875 	return 0;
3876 
3877 err:
3878 	if (!prog) {
3879 		virtnet_clear_guest_offloads(vi);
3880 		for (i = 0; i < vi->max_queue_pairs; i++)
3881 			rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
3882 	}
3883 
3884 	if (netif_running(dev)) {
3885 		for (i = 0; i < vi->max_queue_pairs; i++) {
3886 			virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
3887 			virtnet_napi_tx_enable(vi, vi->sq[i].vq,
3888 					       &vi->sq[i].napi);
3889 		}
3890 	}
3891 	if (prog)
3892 		bpf_prog_sub(prog, vi->max_queue_pairs - 1);
3893 	return err;
3894 }
3895 
virtnet_xdp(struct net_device * dev,struct netdev_bpf * xdp)3896 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
3897 {
3898 	switch (xdp->command) {
3899 	case XDP_SETUP_PROG:
3900 		return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
3901 	default:
3902 		return -EINVAL;
3903 	}
3904 }
3905 
virtnet_get_phys_port_name(struct net_device * dev,char * buf,size_t len)3906 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
3907 				      size_t len)
3908 {
3909 	struct virtnet_info *vi = netdev_priv(dev);
3910 	int ret;
3911 
3912 	if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
3913 		return -EOPNOTSUPP;
3914 
3915 	ret = snprintf(buf, len, "sby");
3916 	if (ret >= len)
3917 		return -EOPNOTSUPP;
3918 
3919 	return 0;
3920 }
3921 
virtnet_set_features(struct net_device * dev,netdev_features_t features)3922 static int virtnet_set_features(struct net_device *dev,
3923 				netdev_features_t features)
3924 {
3925 	struct virtnet_info *vi = netdev_priv(dev);
3926 	u64 offloads;
3927 	int err;
3928 
3929 	if ((dev->features ^ features) & NETIF_F_GRO_HW) {
3930 		if (vi->xdp_enabled)
3931 			return -EBUSY;
3932 
3933 		if (features & NETIF_F_GRO_HW)
3934 			offloads = vi->guest_offloads_capable;
3935 		else
3936 			offloads = vi->guest_offloads_capable &
3937 				   ~GUEST_OFFLOAD_GRO_HW_MASK;
3938 
3939 		err = virtnet_set_guest_offloads(vi, offloads);
3940 		if (err)
3941 			return err;
3942 		vi->guest_offloads = offloads;
3943 	}
3944 
3945 	if ((dev->features ^ features) & NETIF_F_RXHASH) {
3946 		if (features & NETIF_F_RXHASH)
3947 			vi->ctrl->rss.hash_types = vi->rss_hash_types_saved;
3948 		else
3949 			vi->ctrl->rss.hash_types = VIRTIO_NET_HASH_REPORT_NONE;
3950 
3951 		if (!virtnet_commit_rss_command(vi))
3952 			return -EINVAL;
3953 	}
3954 
3955 	return 0;
3956 }
3957 
virtnet_tx_timeout(struct net_device * dev,unsigned int txqueue)3958 static void virtnet_tx_timeout(struct net_device *dev, unsigned int txqueue)
3959 {
3960 	struct virtnet_info *priv = netdev_priv(dev);
3961 	struct send_queue *sq = &priv->sq[txqueue];
3962 	struct netdev_queue *txq = netdev_get_tx_queue(dev, txqueue);
3963 
3964 	u64_stats_update_begin(&sq->stats.syncp);
3965 	u64_stats_inc(&sq->stats.tx_timeouts);
3966 	u64_stats_update_end(&sq->stats.syncp);
3967 
3968 	netdev_err(dev, "TX timeout on queue: %u, sq: %s, vq: 0x%x, name: %s, %u usecs ago\n",
3969 		   txqueue, sq->name, sq->vq->index, sq->vq->name,
3970 		   jiffies_to_usecs(jiffies - READ_ONCE(txq->trans_start)));
3971 }
3972 
3973 static const struct net_device_ops virtnet_netdev = {
3974 	.ndo_open            = virtnet_open,
3975 	.ndo_stop   	     = virtnet_close,
3976 	.ndo_start_xmit      = start_xmit,
3977 	.ndo_validate_addr   = eth_validate_addr,
3978 	.ndo_set_mac_address = virtnet_set_mac_address,
3979 	.ndo_set_rx_mode     = virtnet_set_rx_mode,
3980 	.ndo_get_stats64     = virtnet_stats,
3981 	.ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
3982 	.ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
3983 	.ndo_bpf		= virtnet_xdp,
3984 	.ndo_xdp_xmit		= virtnet_xdp_xmit,
3985 	.ndo_features_check	= passthru_features_check,
3986 	.ndo_get_phys_port_name	= virtnet_get_phys_port_name,
3987 	.ndo_set_features	= virtnet_set_features,
3988 	.ndo_tx_timeout		= virtnet_tx_timeout,
3989 };
3990 
virtnet_config_changed_work(struct work_struct * work)3991 static void virtnet_config_changed_work(struct work_struct *work)
3992 {
3993 	struct virtnet_info *vi =
3994 		container_of(work, struct virtnet_info, config_work);
3995 	u16 v;
3996 
3997 	if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
3998 				 struct virtio_net_config, status, &v) < 0)
3999 		return;
4000 
4001 	if (v & VIRTIO_NET_S_ANNOUNCE) {
4002 		netdev_notify_peers(vi->dev);
4003 		virtnet_ack_link_announce(vi);
4004 	}
4005 
4006 	/* Ignore unknown (future) status bits */
4007 	v &= VIRTIO_NET_S_LINK_UP;
4008 
4009 	if (vi->status == v)
4010 		return;
4011 
4012 	vi->status = v;
4013 
4014 	if (vi->status & VIRTIO_NET_S_LINK_UP) {
4015 		virtnet_update_settings(vi);
4016 		netif_carrier_on(vi->dev);
4017 		netif_tx_wake_all_queues(vi->dev);
4018 	} else {
4019 		netif_carrier_off(vi->dev);
4020 		netif_tx_stop_all_queues(vi->dev);
4021 	}
4022 }
4023 
virtnet_config_changed(struct virtio_device * vdev)4024 static void virtnet_config_changed(struct virtio_device *vdev)
4025 {
4026 	struct virtnet_info *vi = vdev->priv;
4027 
4028 	schedule_work(&vi->config_work);
4029 }
4030 
virtnet_free_queues(struct virtnet_info * vi)4031 static void virtnet_free_queues(struct virtnet_info *vi)
4032 {
4033 	int i;
4034 
4035 	for (i = 0; i < vi->max_queue_pairs; i++) {
4036 		__netif_napi_del(&vi->rq[i].napi);
4037 		__netif_napi_del(&vi->sq[i].napi);
4038 	}
4039 
4040 	/* We called __netif_napi_del(),
4041 	 * we need to respect an RCU grace period before freeing vi->rq
4042 	 */
4043 	synchronize_net();
4044 
4045 	kfree(vi->rq);
4046 	kfree(vi->sq);
4047 	kfree(vi->ctrl);
4048 }
4049 
_free_receive_bufs(struct virtnet_info * vi)4050 static void _free_receive_bufs(struct virtnet_info *vi)
4051 {
4052 	struct bpf_prog *old_prog;
4053 	int i;
4054 
4055 	for (i = 0; i < vi->max_queue_pairs; i++) {
4056 		while (vi->rq[i].pages)
4057 			__free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
4058 
4059 		old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
4060 		RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
4061 		if (old_prog)
4062 			bpf_prog_put(old_prog);
4063 	}
4064 }
4065 
free_receive_bufs(struct virtnet_info * vi)4066 static void free_receive_bufs(struct virtnet_info *vi)
4067 {
4068 	rtnl_lock();
4069 	_free_receive_bufs(vi);
4070 	rtnl_unlock();
4071 }
4072 
free_receive_page_frags(struct virtnet_info * vi)4073 static void free_receive_page_frags(struct virtnet_info *vi)
4074 {
4075 	int i;
4076 	for (i = 0; i < vi->max_queue_pairs; i++)
4077 		if (vi->rq[i].alloc_frag.page) {
4078 			if (vi->rq[i].do_dma && vi->rq[i].last_dma)
4079 				virtnet_rq_unmap(&vi->rq[i], vi->rq[i].last_dma, 0);
4080 			put_page(vi->rq[i].alloc_frag.page);
4081 		}
4082 }
4083 
virtnet_sq_free_unused_buf(struct virtqueue * vq,void * buf)4084 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf)
4085 {
4086 	if (!is_xdp_frame(buf))
4087 		dev_kfree_skb(buf);
4088 	else
4089 		xdp_return_frame(ptr_to_xdp(buf));
4090 }
4091 
free_unused_bufs(struct virtnet_info * vi)4092 static void free_unused_bufs(struct virtnet_info *vi)
4093 {
4094 	void *buf;
4095 	int i;
4096 
4097 	for (i = 0; i < vi->max_queue_pairs; i++) {
4098 		struct virtqueue *vq = vi->sq[i].vq;
4099 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
4100 			virtnet_sq_free_unused_buf(vq, buf);
4101 		cond_resched();
4102 	}
4103 
4104 	for (i = 0; i < vi->max_queue_pairs; i++) {
4105 		struct virtqueue *vq = vi->rq[i].vq;
4106 
4107 		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
4108 			virtnet_rq_unmap_free_buf(vq, buf);
4109 		cond_resched();
4110 	}
4111 }
4112 
virtnet_del_vqs(struct virtnet_info * vi)4113 static void virtnet_del_vqs(struct virtnet_info *vi)
4114 {
4115 	struct virtio_device *vdev = vi->vdev;
4116 
4117 	virtnet_clean_affinity(vi);
4118 
4119 	vdev->config->del_vqs(vdev);
4120 
4121 	virtnet_free_queues(vi);
4122 }
4123 
4124 /* How large should a single buffer be so a queue full of these can fit at
4125  * least one full packet?
4126  * Logic below assumes the mergeable buffer header is used.
4127  */
mergeable_min_buf_len(struct virtnet_info * vi,struct virtqueue * vq)4128 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
4129 {
4130 	const unsigned int hdr_len = vi->hdr_len;
4131 	unsigned int rq_size = virtqueue_get_vring_size(vq);
4132 	unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
4133 	unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
4134 	unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
4135 
4136 	return max(max(min_buf_len, hdr_len) - hdr_len,
4137 		   (unsigned int)GOOD_PACKET_LEN);
4138 }
4139 
virtnet_find_vqs(struct virtnet_info * vi)4140 static int virtnet_find_vqs(struct virtnet_info *vi)
4141 {
4142 	vq_callback_t **callbacks;
4143 	struct virtqueue **vqs;
4144 	const char **names;
4145 	int ret = -ENOMEM;
4146 	int total_vqs;
4147 	bool *ctx;
4148 	u16 i;
4149 
4150 	/* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
4151 	 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
4152 	 * possible control vq.
4153 	 */
4154 	total_vqs = vi->max_queue_pairs * 2 +
4155 		    virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
4156 
4157 	/* Allocate space for find_vqs parameters */
4158 	vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
4159 	if (!vqs)
4160 		goto err_vq;
4161 	callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
4162 	if (!callbacks)
4163 		goto err_callback;
4164 	names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
4165 	if (!names)
4166 		goto err_names;
4167 	if (!vi->big_packets || vi->mergeable_rx_bufs) {
4168 		ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
4169 		if (!ctx)
4170 			goto err_ctx;
4171 	} else {
4172 		ctx = NULL;
4173 	}
4174 
4175 	/* Parameters for control virtqueue, if any */
4176 	if (vi->has_cvq) {
4177 		callbacks[total_vqs - 1] = NULL;
4178 		names[total_vqs - 1] = "control";
4179 	}
4180 
4181 	/* Allocate/initialize parameters for send/receive virtqueues */
4182 	for (i = 0; i < vi->max_queue_pairs; i++) {
4183 		callbacks[rxq2vq(i)] = skb_recv_done;
4184 		callbacks[txq2vq(i)] = skb_xmit_done;
4185 		sprintf(vi->rq[i].name, "input.%u", i);
4186 		sprintf(vi->sq[i].name, "output.%u", i);
4187 		names[rxq2vq(i)] = vi->rq[i].name;
4188 		names[txq2vq(i)] = vi->sq[i].name;
4189 		if (ctx)
4190 			ctx[rxq2vq(i)] = true;
4191 	}
4192 
4193 	ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
4194 				  names, ctx, NULL);
4195 	if (ret)
4196 		goto err_find;
4197 
4198 	if (vi->has_cvq) {
4199 		vi->cvq = vqs[total_vqs - 1];
4200 		if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
4201 			vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
4202 	}
4203 
4204 	for (i = 0; i < vi->max_queue_pairs; i++) {
4205 		vi->rq[i].vq = vqs[rxq2vq(i)];
4206 		vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
4207 		vi->sq[i].vq = vqs[txq2vq(i)];
4208 	}
4209 
4210 	/* run here: ret == 0. */
4211 
4212 
4213 err_find:
4214 	kfree(ctx);
4215 err_ctx:
4216 	kfree(names);
4217 err_names:
4218 	kfree(callbacks);
4219 err_callback:
4220 	kfree(vqs);
4221 err_vq:
4222 	return ret;
4223 }
4224 
virtnet_alloc_queues(struct virtnet_info * vi)4225 static int virtnet_alloc_queues(struct virtnet_info *vi)
4226 {
4227 	int i;
4228 
4229 	if (vi->has_cvq) {
4230 		vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
4231 		if (!vi->ctrl)
4232 			goto err_ctrl;
4233 	} else {
4234 		vi->ctrl = NULL;
4235 	}
4236 	vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
4237 	if (!vi->sq)
4238 		goto err_sq;
4239 	vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
4240 	if (!vi->rq)
4241 		goto err_rq;
4242 
4243 	INIT_DELAYED_WORK(&vi->refill, refill_work);
4244 	for (i = 0; i < vi->max_queue_pairs; i++) {
4245 		vi->rq[i].pages = NULL;
4246 		netif_napi_add_weight(vi->dev, &vi->rq[i].napi, virtnet_poll,
4247 				      napi_weight);
4248 		netif_napi_add_tx_weight(vi->dev, &vi->sq[i].napi,
4249 					 virtnet_poll_tx,
4250 					 napi_tx ? napi_weight : 0);
4251 
4252 		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
4253 		ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
4254 		sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
4255 
4256 		u64_stats_init(&vi->rq[i].stats.syncp);
4257 		u64_stats_init(&vi->sq[i].stats.syncp);
4258 	}
4259 
4260 	return 0;
4261 
4262 err_rq:
4263 	kfree(vi->sq);
4264 err_sq:
4265 	kfree(vi->ctrl);
4266 err_ctrl:
4267 	return -ENOMEM;
4268 }
4269 
init_vqs(struct virtnet_info * vi)4270 static int init_vqs(struct virtnet_info *vi)
4271 {
4272 	int ret;
4273 
4274 	/* Allocate send & receive queues */
4275 	ret = virtnet_alloc_queues(vi);
4276 	if (ret)
4277 		goto err;
4278 
4279 	ret = virtnet_find_vqs(vi);
4280 	if (ret)
4281 		goto err_free;
4282 
4283 	virtnet_rq_set_premapped(vi);
4284 
4285 	cpus_read_lock();
4286 	virtnet_set_affinity(vi);
4287 	cpus_read_unlock();
4288 
4289 	return 0;
4290 
4291 err_free:
4292 	virtnet_free_queues(vi);
4293 err:
4294 	return ret;
4295 }
4296 
4297 #ifdef CONFIG_SYSFS
mergeable_rx_buffer_size_show(struct netdev_rx_queue * queue,char * buf)4298 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
4299 		char *buf)
4300 {
4301 	struct virtnet_info *vi = netdev_priv(queue->dev);
4302 	unsigned int queue_index = get_netdev_rx_queue_index(queue);
4303 	unsigned int headroom = virtnet_get_headroom(vi);
4304 	unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
4305 	struct ewma_pkt_len *avg;
4306 
4307 	BUG_ON(queue_index >= vi->max_queue_pairs);
4308 	avg = &vi->rq[queue_index].mrg_avg_pkt_len;
4309 	return sprintf(buf, "%u\n",
4310 		       get_mergeable_buf_len(&vi->rq[queue_index], avg,
4311 				       SKB_DATA_ALIGN(headroom + tailroom)));
4312 }
4313 
4314 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
4315 	__ATTR_RO(mergeable_rx_buffer_size);
4316 
4317 static struct attribute *virtio_net_mrg_rx_attrs[] = {
4318 	&mergeable_rx_buffer_size_attribute.attr,
4319 	NULL
4320 };
4321 
4322 static const struct attribute_group virtio_net_mrg_rx_group = {
4323 	.name = "virtio_net",
4324 	.attrs = virtio_net_mrg_rx_attrs
4325 };
4326 #endif
4327 
virtnet_fail_on_feature(struct virtio_device * vdev,unsigned int fbit,const char * fname,const char * dname)4328 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
4329 				    unsigned int fbit,
4330 				    const char *fname, const char *dname)
4331 {
4332 	if (!virtio_has_feature(vdev, fbit))
4333 		return false;
4334 
4335 	dev_err(&vdev->dev, "device advertises feature %s but not %s",
4336 		fname, dname);
4337 
4338 	return true;
4339 }
4340 
4341 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)			\
4342 	virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
4343 
virtnet_validate_features(struct virtio_device * vdev)4344 static bool virtnet_validate_features(struct virtio_device *vdev)
4345 {
4346 	if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
4347 	    (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
4348 			     "VIRTIO_NET_F_CTRL_VQ") ||
4349 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
4350 			     "VIRTIO_NET_F_CTRL_VQ") ||
4351 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
4352 			     "VIRTIO_NET_F_CTRL_VQ") ||
4353 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
4354 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
4355 			     "VIRTIO_NET_F_CTRL_VQ") ||
4356 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_RSS,
4357 			     "VIRTIO_NET_F_CTRL_VQ") ||
4358 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_HASH_REPORT,
4359 			     "VIRTIO_NET_F_CTRL_VQ") ||
4360 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_NOTF_COAL,
4361 			     "VIRTIO_NET_F_CTRL_VQ") ||
4362 	     VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_VQ_NOTF_COAL,
4363 			     "VIRTIO_NET_F_CTRL_VQ"))) {
4364 		return false;
4365 	}
4366 
4367 	return true;
4368 }
4369 
4370 #define MIN_MTU ETH_MIN_MTU
4371 #define MAX_MTU ETH_MAX_MTU
4372 
virtnet_validate(struct virtio_device * vdev)4373 static int virtnet_validate(struct virtio_device *vdev)
4374 {
4375 	if (!vdev->config->get) {
4376 		dev_err(&vdev->dev, "%s failure: config access disabled\n",
4377 			__func__);
4378 		return -EINVAL;
4379 	}
4380 
4381 	if (!virtnet_validate_features(vdev))
4382 		return -EINVAL;
4383 
4384 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
4385 		int mtu = virtio_cread16(vdev,
4386 					 offsetof(struct virtio_net_config,
4387 						  mtu));
4388 		if (mtu < MIN_MTU)
4389 			__virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
4390 	}
4391 
4392 	if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY) &&
4393 	    !virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
4394 		dev_warn(&vdev->dev, "device advertises feature VIRTIO_NET_F_STANDBY but not VIRTIO_NET_F_MAC, disabling standby");
4395 		__virtio_clear_bit(vdev, VIRTIO_NET_F_STANDBY);
4396 	}
4397 
4398 	return 0;
4399 }
4400 
virtnet_check_guest_gso(const struct virtnet_info * vi)4401 static bool virtnet_check_guest_gso(const struct virtnet_info *vi)
4402 {
4403 	return virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
4404 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
4405 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
4406 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
4407 		(virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) &&
4408 		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6));
4409 }
4410 
virtnet_set_big_packets(struct virtnet_info * vi,const int mtu)4411 static void virtnet_set_big_packets(struct virtnet_info *vi, const int mtu)
4412 {
4413 	bool guest_gso = virtnet_check_guest_gso(vi);
4414 
4415 	/* If device can receive ANY guest GSO packets, regardless of mtu,
4416 	 * allocate packets of maximum size, otherwise limit it to only
4417 	 * mtu size worth only.
4418 	 */
4419 	if (mtu > ETH_DATA_LEN || guest_gso) {
4420 		vi->big_packets = true;
4421 		vi->big_packets_num_skbfrags = guest_gso ? MAX_SKB_FRAGS : DIV_ROUND_UP(mtu, PAGE_SIZE);
4422 	}
4423 }
4424 
virtnet_probe(struct virtio_device * vdev)4425 static int virtnet_probe(struct virtio_device *vdev)
4426 {
4427 	int i, err = -ENOMEM;
4428 	struct net_device *dev;
4429 	struct virtnet_info *vi;
4430 	u16 max_queue_pairs;
4431 	int mtu = 0;
4432 
4433 	/* Find if host supports multiqueue/rss virtio_net device */
4434 	max_queue_pairs = 1;
4435 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MQ) || virtio_has_feature(vdev, VIRTIO_NET_F_RSS))
4436 		max_queue_pairs =
4437 		     virtio_cread16(vdev, offsetof(struct virtio_net_config, max_virtqueue_pairs));
4438 
4439 	/* We need at least 2 queue's */
4440 	if (max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
4441 	    max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
4442 	    !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
4443 		max_queue_pairs = 1;
4444 
4445 	/* Allocate ourselves a network device with room for our info */
4446 	dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
4447 	if (!dev)
4448 		return -ENOMEM;
4449 
4450 	/* Set up network device as normal. */
4451 	dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
4452 			   IFF_TX_SKB_NO_LINEAR;
4453 	dev->netdev_ops = &virtnet_netdev;
4454 	dev->features = NETIF_F_HIGHDMA;
4455 
4456 	dev->ethtool_ops = &virtnet_ethtool_ops;
4457 	SET_NETDEV_DEV(dev, &vdev->dev);
4458 
4459 	/* Do we support "hardware" checksums? */
4460 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
4461 		/* This opens up the world of extra features. */
4462 		dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
4463 		if (csum)
4464 			dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
4465 
4466 		if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
4467 			dev->hw_features |= NETIF_F_TSO
4468 				| NETIF_F_TSO_ECN | NETIF_F_TSO6;
4469 		}
4470 		/* Individual feature bits: what can host handle? */
4471 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
4472 			dev->hw_features |= NETIF_F_TSO;
4473 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
4474 			dev->hw_features |= NETIF_F_TSO6;
4475 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
4476 			dev->hw_features |= NETIF_F_TSO_ECN;
4477 		if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_USO))
4478 			dev->hw_features |= NETIF_F_GSO_UDP_L4;
4479 
4480 		dev->features |= NETIF_F_GSO_ROBUST;
4481 
4482 		if (gso)
4483 			dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
4484 		/* (!csum && gso) case will be fixed by register_netdev() */
4485 	}
4486 
4487 	/* 1. With VIRTIO_NET_F_GUEST_CSUM negotiation, the driver doesn't
4488 	 * need to calculate checksums for partially checksummed packets,
4489 	 * as they're considered valid by the upper layer.
4490 	 * 2. Without VIRTIO_NET_F_GUEST_CSUM negotiation, the driver only
4491 	 * receives fully checksummed packets. The device may assist in
4492 	 * validating these packets' checksums, so the driver won't have to.
4493 	 */
4494 	dev->features |= NETIF_F_RXCSUM;
4495 
4496 	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
4497 	    virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
4498 		dev->features |= NETIF_F_GRO_HW;
4499 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
4500 		dev->hw_features |= NETIF_F_GRO_HW;
4501 
4502 	dev->vlan_features = dev->features;
4503 	dev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT;
4504 
4505 	/* MTU range: 68 - 65535 */
4506 	dev->min_mtu = MIN_MTU;
4507 	dev->max_mtu = MAX_MTU;
4508 
4509 	/* Configuration may specify what MAC to use.  Otherwise random. */
4510 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
4511 		u8 addr[ETH_ALEN];
4512 
4513 		virtio_cread_bytes(vdev,
4514 				   offsetof(struct virtio_net_config, mac),
4515 				   addr, ETH_ALEN);
4516 		eth_hw_addr_set(dev, addr);
4517 	} else {
4518 		eth_hw_addr_random(dev);
4519 		dev_info(&vdev->dev, "Assigned random MAC address %pM\n",
4520 			 dev->dev_addr);
4521 	}
4522 
4523 	/* Set up our device-specific information */
4524 	vi = netdev_priv(dev);
4525 	vi->dev = dev;
4526 	vi->vdev = vdev;
4527 	vdev->priv = vi;
4528 
4529 	INIT_WORK(&vi->config_work, virtnet_config_changed_work);
4530 	spin_lock_init(&vi->refill_lock);
4531 
4532 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) {
4533 		vi->mergeable_rx_bufs = true;
4534 		dev->xdp_features |= NETDEV_XDP_ACT_RX_SG;
4535 	}
4536 
4537 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
4538 		vi->intr_coal_rx.max_usecs = 0;
4539 		vi->intr_coal_tx.max_usecs = 0;
4540 		vi->intr_coal_tx.max_packets = 0;
4541 		vi->intr_coal_rx.max_packets = 0;
4542 	}
4543 
4544 	if (virtio_has_feature(vdev, VIRTIO_NET_F_HASH_REPORT))
4545 		vi->has_rss_hash_report = true;
4546 
4547 	if (virtio_has_feature(vdev, VIRTIO_NET_F_RSS)) {
4548 		vi->has_rss = true;
4549 
4550 		vi->rss_indir_table_size =
4551 			virtio_cread16(vdev, offsetof(struct virtio_net_config,
4552 				rss_max_indirection_table_length));
4553 	}
4554 
4555 	if (vi->has_rss || vi->has_rss_hash_report) {
4556 		vi->rss_key_size =
4557 			virtio_cread8(vdev, offsetof(struct virtio_net_config, rss_max_key_size));
4558 		if (vi->rss_key_size > VIRTIO_NET_RSS_MAX_KEY_SIZE) {
4559 			dev_err(&vdev->dev, "rss_max_key_size=%u exceeds the limit %u.\n",
4560 				vi->rss_key_size, VIRTIO_NET_RSS_MAX_KEY_SIZE);
4561 			err = -EINVAL;
4562 			goto free;
4563 		}
4564 
4565 		vi->rss_hash_types_supported =
4566 		    virtio_cread32(vdev, offsetof(struct virtio_net_config, supported_hash_types));
4567 		vi->rss_hash_types_supported &=
4568 				~(VIRTIO_NET_RSS_HASH_TYPE_IP_EX |
4569 				  VIRTIO_NET_RSS_HASH_TYPE_TCP_EX |
4570 				  VIRTIO_NET_RSS_HASH_TYPE_UDP_EX);
4571 
4572 		dev->hw_features |= NETIF_F_RXHASH;
4573 	}
4574 
4575 	if (vi->has_rss_hash_report)
4576 		vi->hdr_len = sizeof(struct virtio_net_hdr_v1_hash);
4577 	else if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
4578 		 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
4579 		vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
4580 	else
4581 		vi->hdr_len = sizeof(struct virtio_net_hdr);
4582 
4583 	if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
4584 	    virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
4585 		vi->any_header_sg = true;
4586 
4587 	if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
4588 		vi->has_cvq = true;
4589 
4590 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
4591 		mtu = virtio_cread16(vdev,
4592 				     offsetof(struct virtio_net_config,
4593 					      mtu));
4594 		if (mtu < dev->min_mtu) {
4595 			/* Should never trigger: MTU was previously validated
4596 			 * in virtnet_validate.
4597 			 */
4598 			dev_err(&vdev->dev,
4599 				"device MTU appears to have changed it is now %d < %d",
4600 				mtu, dev->min_mtu);
4601 			err = -EINVAL;
4602 			goto free;
4603 		}
4604 
4605 		dev->mtu = mtu;
4606 		dev->max_mtu = mtu;
4607 	}
4608 
4609 	virtnet_set_big_packets(vi, mtu);
4610 
4611 	if (vi->any_header_sg)
4612 		dev->needed_headroom = vi->hdr_len;
4613 
4614 	/* Enable multiqueue by default */
4615 	if (num_online_cpus() >= max_queue_pairs)
4616 		vi->curr_queue_pairs = max_queue_pairs;
4617 	else
4618 		vi->curr_queue_pairs = num_online_cpus();
4619 	vi->max_queue_pairs = max_queue_pairs;
4620 
4621 	/* Allocate/initialize the rx/tx queues, and invoke find_vqs */
4622 	err = init_vqs(vi);
4623 	if (err)
4624 		goto free;
4625 
4626 #ifdef CONFIG_SYSFS
4627 	if (vi->mergeable_rx_bufs)
4628 		dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
4629 #endif
4630 	netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
4631 	netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
4632 
4633 	virtnet_init_settings(dev);
4634 
4635 	if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
4636 		vi->failover = net_failover_create(vi->dev);
4637 		if (IS_ERR(vi->failover)) {
4638 			err = PTR_ERR(vi->failover);
4639 			goto free_vqs;
4640 		}
4641 	}
4642 
4643 	if (vi->has_rss || vi->has_rss_hash_report)
4644 		virtnet_init_default_rss(vi);
4645 
4646 	/* serialize netdev register + virtio_device_ready() with ndo_open() */
4647 	rtnl_lock();
4648 
4649 	err = register_netdevice(dev);
4650 	if (err) {
4651 		pr_debug("virtio_net: registering device failed\n");
4652 		rtnl_unlock();
4653 		goto free_failover;
4654 	}
4655 
4656 	virtio_device_ready(vdev);
4657 
4658 	_virtnet_set_queues(vi, vi->curr_queue_pairs);
4659 
4660 	/* a random MAC address has been assigned, notify the device.
4661 	 * We don't fail probe if VIRTIO_NET_F_CTRL_MAC_ADDR is not there
4662 	 * because many devices work fine without getting MAC explicitly
4663 	 */
4664 	if (!virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
4665 	    virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
4666 		struct scatterlist sg;
4667 
4668 		sg_init_one(&sg, dev->dev_addr, dev->addr_len);
4669 		if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
4670 					  VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
4671 			pr_debug("virtio_net: setting MAC address failed\n");
4672 			rtnl_unlock();
4673 			err = -EINVAL;
4674 			goto free_unregister_netdev;
4675 		}
4676 	}
4677 
4678 	rtnl_unlock();
4679 
4680 	err = virtnet_cpu_notif_add(vi);
4681 	if (err) {
4682 		pr_debug("virtio_net: registering cpu notifier failed\n");
4683 		goto free_unregister_netdev;
4684 	}
4685 
4686 	/* Assume link up if device can't report link status,
4687 	   otherwise get link status from config. */
4688 	netif_carrier_off(dev);
4689 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
4690 		schedule_work(&vi->config_work);
4691 	} else {
4692 		vi->status = VIRTIO_NET_S_LINK_UP;
4693 		virtnet_update_settings(vi);
4694 		netif_carrier_on(dev);
4695 	}
4696 
4697 	for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
4698 		if (virtio_has_feature(vi->vdev, guest_offloads[i]))
4699 			set_bit(guest_offloads[i], &vi->guest_offloads);
4700 	vi->guest_offloads_capable = vi->guest_offloads;
4701 
4702 	pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
4703 		 dev->name, max_queue_pairs);
4704 
4705 	return 0;
4706 
4707 free_unregister_netdev:
4708 	unregister_netdev(dev);
4709 free_failover:
4710 	net_failover_destroy(vi->failover);
4711 free_vqs:
4712 	virtio_reset_device(vdev);
4713 	cancel_delayed_work_sync(&vi->refill);
4714 	free_receive_page_frags(vi);
4715 	virtnet_del_vqs(vi);
4716 free:
4717 	free_netdev(dev);
4718 	return err;
4719 }
4720 
remove_vq_common(struct virtnet_info * vi)4721 static void remove_vq_common(struct virtnet_info *vi)
4722 {
4723 	virtio_reset_device(vi->vdev);
4724 
4725 	/* Free unused buffers in both send and recv, if any. */
4726 	free_unused_bufs(vi);
4727 
4728 	free_receive_bufs(vi);
4729 
4730 	free_receive_page_frags(vi);
4731 
4732 	virtnet_del_vqs(vi);
4733 }
4734 
virtnet_remove(struct virtio_device * vdev)4735 static void virtnet_remove(struct virtio_device *vdev)
4736 {
4737 	struct virtnet_info *vi = vdev->priv;
4738 
4739 	virtnet_cpu_notif_remove(vi);
4740 
4741 	/* Make sure no work handler is accessing the device. */
4742 	flush_work(&vi->config_work);
4743 
4744 	unregister_netdev(vi->dev);
4745 
4746 	net_failover_destroy(vi->failover);
4747 
4748 	remove_vq_common(vi);
4749 
4750 	free_netdev(vi->dev);
4751 }
4752 
virtnet_freeze(struct virtio_device * vdev)4753 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
4754 {
4755 	struct virtnet_info *vi = vdev->priv;
4756 
4757 	virtnet_cpu_notif_remove(vi);
4758 	virtnet_freeze_down(vdev);
4759 	remove_vq_common(vi);
4760 
4761 	return 0;
4762 }
4763 
virtnet_restore(struct virtio_device * vdev)4764 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
4765 {
4766 	struct virtnet_info *vi = vdev->priv;
4767 	int err;
4768 
4769 	err = virtnet_restore_up(vdev);
4770 	if (err)
4771 		return err;
4772 	virtnet_set_queues(vi, vi->curr_queue_pairs);
4773 
4774 	err = virtnet_cpu_notif_add(vi);
4775 	if (err) {
4776 		virtnet_freeze_down(vdev);
4777 		remove_vq_common(vi);
4778 		return err;
4779 	}
4780 
4781 	return 0;
4782 }
4783 
4784 static struct virtio_device_id id_table[] = {
4785 	{ VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
4786 	{ 0 },
4787 };
4788 
4789 #define VIRTNET_FEATURES \
4790 	VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
4791 	VIRTIO_NET_F_MAC, \
4792 	VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
4793 	VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
4794 	VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
4795 	VIRTIO_NET_F_HOST_USO, VIRTIO_NET_F_GUEST_USO4, VIRTIO_NET_F_GUEST_USO6, \
4796 	VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
4797 	VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
4798 	VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
4799 	VIRTIO_NET_F_CTRL_MAC_ADDR, \
4800 	VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
4801 	VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY, \
4802 	VIRTIO_NET_F_RSS, VIRTIO_NET_F_HASH_REPORT, VIRTIO_NET_F_NOTF_COAL, \
4803 	VIRTIO_NET_F_VQ_NOTF_COAL, \
4804 	VIRTIO_NET_F_GUEST_HDRLEN
4805 
4806 static unsigned int features[] = {
4807 	VIRTNET_FEATURES,
4808 };
4809 
4810 static unsigned int features_legacy[] = {
4811 	VIRTNET_FEATURES,
4812 	VIRTIO_NET_F_GSO,
4813 	VIRTIO_F_ANY_LAYOUT,
4814 };
4815 
4816 static struct virtio_driver virtio_net_driver = {
4817 	.feature_table = features,
4818 	.feature_table_size = ARRAY_SIZE(features),
4819 	.feature_table_legacy = features_legacy,
4820 	.feature_table_size_legacy = ARRAY_SIZE(features_legacy),
4821 	.driver.name =	KBUILD_MODNAME,
4822 	.driver.owner =	THIS_MODULE,
4823 	.id_table =	id_table,
4824 	.validate =	virtnet_validate,
4825 	.probe =	virtnet_probe,
4826 	.remove =	virtnet_remove,
4827 	.config_changed = virtnet_config_changed,
4828 #ifdef CONFIG_PM_SLEEP
4829 	.freeze =	virtnet_freeze,
4830 	.restore =	virtnet_restore,
4831 #endif
4832 };
4833 
virtio_net_driver_init(void)4834 static __init int virtio_net_driver_init(void)
4835 {
4836 	int ret;
4837 
4838 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
4839 				      virtnet_cpu_online,
4840 				      virtnet_cpu_down_prep);
4841 	if (ret < 0)
4842 		goto out;
4843 	virtionet_online = ret;
4844 	ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
4845 				      NULL, virtnet_cpu_dead);
4846 	if (ret)
4847 		goto err_dead;
4848 	ret = register_virtio_driver(&virtio_net_driver);
4849 	if (ret)
4850 		goto err_virtio;
4851 	return 0;
4852 err_virtio:
4853 	cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
4854 err_dead:
4855 	cpuhp_remove_multi_state(virtionet_online);
4856 out:
4857 	return ret;
4858 }
4859 module_init(virtio_net_driver_init);
4860 
virtio_net_driver_exit(void)4861 static __exit void virtio_net_driver_exit(void)
4862 {
4863 	unregister_virtio_driver(&virtio_net_driver);
4864 	cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
4865 	cpuhp_remove_multi_state(virtionet_online);
4866 }
4867 module_exit(virtio_net_driver_exit);
4868 
4869 MODULE_DEVICE_TABLE(virtio, id_table);
4870 MODULE_DESCRIPTION("Virtio network driver");
4871 MODULE_LICENSE("GPL");
4872