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