Loading...
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 <linux/dim.h>
23#include <net/route.h>
24#include <net/xdp.h>
25#include <net/net_failover.h>
26#include <net/netdev_rx_queue.h>
27
28static int napi_weight = NAPI_POLL_WEIGHT;
29module_param(napi_weight, int, 0444);
30
31static bool csum = true, gso = true, napi_tx = true;
32module_param(csum, bool, 0444);
33module_param(gso, bool, 0444);
34module_param(napi_tx, bool, 0644);
35
36/* FIXME: MTU in config. */
37#define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
38#define GOOD_COPY_LEN 128
39
40#define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
41
42/* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
43#define VIRTIO_XDP_HEADROOM 256
44
45/* Separating two types of XDP xmit */
46#define VIRTIO_XDP_TX BIT(0)
47#define VIRTIO_XDP_REDIR BIT(1)
48
49#define VIRTIO_XDP_FLAG BIT(0)
50
51/* RX packet size EWMA. The average packet size is used to determine the packet
52 * buffer size when refilling RX rings. As the entire RX ring may be refilled
53 * at once, the weight is chosen so that the EWMA will be insensitive to short-
54 * term, transient changes in packet size.
55 */
56DECLARE_EWMA(pkt_len, 0, 64)
57
58#define VIRTNET_DRIVER_VERSION "1.0.0"
59
60static const unsigned long guest_offloads[] = {
61 VIRTIO_NET_F_GUEST_TSO4,
62 VIRTIO_NET_F_GUEST_TSO6,
63 VIRTIO_NET_F_GUEST_ECN,
64 VIRTIO_NET_F_GUEST_UFO,
65 VIRTIO_NET_F_GUEST_CSUM,
66 VIRTIO_NET_F_GUEST_USO4,
67 VIRTIO_NET_F_GUEST_USO6,
68 VIRTIO_NET_F_GUEST_HDRLEN
69};
70
71#define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
72 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
73 (1ULL << VIRTIO_NET_F_GUEST_ECN) | \
74 (1ULL << VIRTIO_NET_F_GUEST_UFO) | \
75 (1ULL << VIRTIO_NET_F_GUEST_USO4) | \
76 (1ULL << VIRTIO_NET_F_GUEST_USO6))
77
78struct virtnet_stat_desc {
79 char desc[ETH_GSTRING_LEN];
80 size_t offset;
81};
82
83struct virtnet_sq_stats {
84 struct u64_stats_sync syncp;
85 u64_stats_t packets;
86 u64_stats_t bytes;
87 u64_stats_t xdp_tx;
88 u64_stats_t xdp_tx_drops;
89 u64_stats_t kicks;
90 u64_stats_t tx_timeouts;
91};
92
93struct virtnet_rq_stats {
94 struct u64_stats_sync syncp;
95 u64_stats_t packets;
96 u64_stats_t bytes;
97 u64_stats_t drops;
98 u64_stats_t xdp_packets;
99 u64_stats_t xdp_tx;
100 u64_stats_t xdp_redirects;
101 u64_stats_t xdp_drops;
102 u64_stats_t kicks;
103};
104
105#define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m)
106#define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m)
107
108static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
109 { "packets", VIRTNET_SQ_STAT(packets) },
110 { "bytes", VIRTNET_SQ_STAT(bytes) },
111 { "xdp_tx", VIRTNET_SQ_STAT(xdp_tx) },
112 { "xdp_tx_drops", VIRTNET_SQ_STAT(xdp_tx_drops) },
113 { "kicks", VIRTNET_SQ_STAT(kicks) },
114 { "tx_timeouts", VIRTNET_SQ_STAT(tx_timeouts) },
115};
116
117static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
118 { "packets", VIRTNET_RQ_STAT(packets) },
119 { "bytes", VIRTNET_RQ_STAT(bytes) },
120 { "drops", VIRTNET_RQ_STAT(drops) },
121 { "xdp_packets", VIRTNET_RQ_STAT(xdp_packets) },
122 { "xdp_tx", VIRTNET_RQ_STAT(xdp_tx) },
123 { "xdp_redirects", VIRTNET_RQ_STAT(xdp_redirects) },
124 { "xdp_drops", VIRTNET_RQ_STAT(xdp_drops) },
125 { "kicks", VIRTNET_RQ_STAT(kicks) },
126};
127
128#define VIRTNET_SQ_STATS_LEN ARRAY_SIZE(virtnet_sq_stats_desc)
129#define VIRTNET_RQ_STATS_LEN ARRAY_SIZE(virtnet_rq_stats_desc)
130
131struct virtnet_interrupt_coalesce {
132 u32 max_packets;
133 u32 max_usecs;
134};
135
136/* The dma information of pages allocated at a time. */
137struct virtnet_rq_dma {
138 dma_addr_t addr;
139 u32 ref;
140 u16 len;
141 u16 need_sync;
142};
143
144/* Internal representation of a send virtqueue */
145struct send_queue {
146 /* Virtqueue associated with this send _queue */
147 struct virtqueue *vq;
148
149 /* TX: fragments + linear part + virtio header */
150 struct scatterlist sg[MAX_SKB_FRAGS + 2];
151
152 /* Name of the send queue: output.$index */
153 char name[16];
154
155 struct virtnet_sq_stats stats;
156
157 struct virtnet_interrupt_coalesce intr_coal;
158
159 struct napi_struct napi;
160
161 /* Record whether sq is in reset state. */
162 bool reset;
163};
164
165/* Internal representation of a receive virtqueue */
166struct receive_queue {
167 /* Virtqueue associated with this receive_queue */
168 struct virtqueue *vq;
169
170 struct napi_struct napi;
171
172 struct bpf_prog __rcu *xdp_prog;
173
174 struct virtnet_rq_stats stats;
175
176 /* The number of rx notifications */
177 u16 calls;
178
179 /* Is dynamic interrupt moderation enabled? */
180 bool dim_enabled;
181
182 /* Dynamic Interrupt Moderation */
183 struct dim dim;
184
185 u32 packets_in_napi;
186
187 struct virtnet_interrupt_coalesce intr_coal;
188
189 /* Chain pages by the private ptr. */
190 struct page *pages;
191
192 /* Average packet length for mergeable receive buffers. */
193 struct ewma_pkt_len mrg_avg_pkt_len;
194
195 /* Page frag for packet buffer allocation. */
196 struct page_frag alloc_frag;
197
198 /* RX: fragments + linear part + virtio header */
199 struct scatterlist sg[MAX_SKB_FRAGS + 2];
200
201 /* Min single buffer size for mergeable buffers case. */
202 unsigned int min_buf_len;
203
204 /* Name of this receive queue: input.$index */
205 char name[16];
206
207 struct xdp_rxq_info xdp_rxq;
208
209 /* Record the last dma info to free after new pages is allocated. */
210 struct virtnet_rq_dma *last_dma;
211
212 /* Do dma by self */
213 bool do_dma;
214};
215
216/* This structure can contain rss message with maximum settings for indirection table and keysize
217 * Note, that default structure that describes RSS configuration virtio_net_rss_config
218 * contains same info but can't handle table values.
219 * In any case, structure would be passed to virtio hw through sg_buf split by parts
220 * because table sizes may be differ according to the device configuration.
221 */
222#define VIRTIO_NET_RSS_MAX_KEY_SIZE 40
223#define VIRTIO_NET_RSS_MAX_TABLE_LEN 128
224struct virtio_net_ctrl_rss {
225 u32 hash_types;
226 u16 indirection_table_mask;
227 u16 unclassified_queue;
228 u16 indirection_table[VIRTIO_NET_RSS_MAX_TABLE_LEN];
229 u16 max_tx_vq;
230 u8 hash_key_length;
231 u8 key[VIRTIO_NET_RSS_MAX_KEY_SIZE];
232};
233
234/* Control VQ buffers: protected by the rtnl lock */
235struct control_buf {
236 struct virtio_net_ctrl_hdr hdr;
237 virtio_net_ctrl_ack status;
238 struct virtio_net_ctrl_mq mq;
239 u8 promisc;
240 u8 allmulti;
241 __virtio16 vid;
242 __virtio64 offloads;
243 struct virtio_net_ctrl_rss rss;
244 struct virtio_net_ctrl_coal_tx coal_tx;
245 struct virtio_net_ctrl_coal_rx coal_rx;
246 struct virtio_net_ctrl_coal_vq coal_vq;
247};
248
249struct virtnet_info {
250 struct virtio_device *vdev;
251 struct virtqueue *cvq;
252 struct net_device *dev;
253 struct send_queue *sq;
254 struct receive_queue *rq;
255 unsigned int status;
256
257 /* Max # of queue pairs supported by the device */
258 u16 max_queue_pairs;
259
260 /* # of queue pairs currently used by the driver */
261 u16 curr_queue_pairs;
262
263 /* # of XDP queue pairs currently used by the driver */
264 u16 xdp_queue_pairs;
265
266 /* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
267 bool xdp_enabled;
268
269 /* I like... big packets and I cannot lie! */
270 bool big_packets;
271
272 /* number of sg entries allocated for big packets */
273 unsigned int big_packets_num_skbfrags;
274
275 /* Host will merge rx buffers for big packets (shake it! shake it!) */
276 bool mergeable_rx_bufs;
277
278 /* Host supports rss and/or hash report */
279 bool has_rss;
280 bool has_rss_hash_report;
281 u8 rss_key_size;
282 u16 rss_indir_table_size;
283 u32 rss_hash_types_supported;
284 u32 rss_hash_types_saved;
285
286 /* Has control virtqueue */
287 bool has_cvq;
288
289 /* Host can handle any s/g split between our header and packet data */
290 bool any_header_sg;
291
292 /* Packet virtio header size */
293 u8 hdr_len;
294
295 /* Work struct for delayed refilling if we run low on memory. */
296 struct delayed_work refill;
297
298 /* Is delayed refill enabled? */
299 bool refill_enabled;
300
301 /* The lock to synchronize the access to refill_enabled */
302 spinlock_t refill_lock;
303
304 /* Work struct for config space updates */
305 struct work_struct config_work;
306
307 /* Does the affinity hint is set for virtqueues? */
308 bool affinity_hint_set;
309
310 /* CPU hotplug instances for online & dead */
311 struct hlist_node node;
312 struct hlist_node node_dead;
313
314 struct control_buf *ctrl;
315
316 /* Ethtool settings */
317 u8 duplex;
318 u32 speed;
319
320 /* Is rx dynamic interrupt moderation enabled? */
321 bool rx_dim_enabled;
322
323 /* Interrupt coalescing settings */
324 struct virtnet_interrupt_coalesce intr_coal_tx;
325 struct virtnet_interrupt_coalesce intr_coal_rx;
326
327 unsigned long guest_offloads;
328 unsigned long guest_offloads_capable;
329
330 /* failover when STANDBY feature enabled */
331 struct failover *failover;
332};
333
334struct padded_vnet_hdr {
335 struct virtio_net_hdr_v1_hash hdr;
336 /*
337 * hdr is in a separate sg buffer, and data sg buffer shares same page
338 * with this header sg. This padding makes next sg 16 byte aligned
339 * after the header.
340 */
341 char padding[12];
342};
343
344struct virtio_net_common_hdr {
345 union {
346 struct virtio_net_hdr hdr;
347 struct virtio_net_hdr_mrg_rxbuf mrg_hdr;
348 struct virtio_net_hdr_v1_hash hash_v1_hdr;
349 };
350};
351
352static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf);
353
354static bool is_xdp_frame(void *ptr)
355{
356 return (unsigned long)ptr & VIRTIO_XDP_FLAG;
357}
358
359static void *xdp_to_ptr(struct xdp_frame *ptr)
360{
361 return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
362}
363
364static struct xdp_frame *ptr_to_xdp(void *ptr)
365{
366 return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
367}
368
369/* Converting between virtqueue no. and kernel tx/rx queue no.
370 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
371 */
372static int vq2txq(struct virtqueue *vq)
373{
374 return (vq->index - 1) / 2;
375}
376
377static int txq2vq(int txq)
378{
379 return txq * 2 + 1;
380}
381
382static int vq2rxq(struct virtqueue *vq)
383{
384 return vq->index / 2;
385}
386
387static int rxq2vq(int rxq)
388{
389 return rxq * 2;
390}
391
392static inline struct virtio_net_common_hdr *
393skb_vnet_common_hdr(struct sk_buff *skb)
394{
395 return (struct virtio_net_common_hdr *)skb->cb;
396}
397
398/*
399 * private is used to chain pages for big packets, put the whole
400 * most recent used list in the beginning for reuse
401 */
402static void give_pages(struct receive_queue *rq, struct page *page)
403{
404 struct page *end;
405
406 /* Find end of list, sew whole thing into vi->rq.pages. */
407 for (end = page; end->private; end = (struct page *)end->private);
408 end->private = (unsigned long)rq->pages;
409 rq->pages = page;
410}
411
412static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
413{
414 struct page *p = rq->pages;
415
416 if (p) {
417 rq->pages = (struct page *)p->private;
418 /* clear private here, it is used to chain pages */
419 p->private = 0;
420 } else
421 p = alloc_page(gfp_mask);
422 return p;
423}
424
425static void virtnet_rq_free_buf(struct virtnet_info *vi,
426 struct receive_queue *rq, void *buf)
427{
428 if (vi->mergeable_rx_bufs)
429 put_page(virt_to_head_page(buf));
430 else if (vi->big_packets)
431 give_pages(rq, buf);
432 else
433 put_page(virt_to_head_page(buf));
434}
435
436static void enable_delayed_refill(struct virtnet_info *vi)
437{
438 spin_lock_bh(&vi->refill_lock);
439 vi->refill_enabled = true;
440 spin_unlock_bh(&vi->refill_lock);
441}
442
443static void disable_delayed_refill(struct virtnet_info *vi)
444{
445 spin_lock_bh(&vi->refill_lock);
446 vi->refill_enabled = false;
447 spin_unlock_bh(&vi->refill_lock);
448}
449
450static void virtqueue_napi_schedule(struct napi_struct *napi,
451 struct virtqueue *vq)
452{
453 if (napi_schedule_prep(napi)) {
454 virtqueue_disable_cb(vq);
455 __napi_schedule(napi);
456 }
457}
458
459static bool virtqueue_napi_complete(struct napi_struct *napi,
460 struct virtqueue *vq, int processed)
461{
462 int opaque;
463
464 opaque = virtqueue_enable_cb_prepare(vq);
465 if (napi_complete_done(napi, processed)) {
466 if (unlikely(virtqueue_poll(vq, opaque)))
467 virtqueue_napi_schedule(napi, vq);
468 else
469 return true;
470 } else {
471 virtqueue_disable_cb(vq);
472 }
473
474 return false;
475}
476
477static void skb_xmit_done(struct virtqueue *vq)
478{
479 struct virtnet_info *vi = vq->vdev->priv;
480 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
481
482 /* Suppress further interrupts. */
483 virtqueue_disable_cb(vq);
484
485 if (napi->weight)
486 virtqueue_napi_schedule(napi, vq);
487 else
488 /* We were probably waiting for more output buffers. */
489 netif_wake_subqueue(vi->dev, vq2txq(vq));
490}
491
492#define MRG_CTX_HEADER_SHIFT 22
493static void *mergeable_len_to_ctx(unsigned int truesize,
494 unsigned int headroom)
495{
496 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
497}
498
499static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
500{
501 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
502}
503
504static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
505{
506 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
507}
508
509static struct sk_buff *virtnet_build_skb(void *buf, unsigned int buflen,
510 unsigned int headroom,
511 unsigned int len)
512{
513 struct sk_buff *skb;
514
515 skb = build_skb(buf, buflen);
516 if (unlikely(!skb))
517 return NULL;
518
519 skb_reserve(skb, headroom);
520 skb_put(skb, len);
521
522 return skb;
523}
524
525/* Called from bottom half context */
526static struct sk_buff *page_to_skb(struct virtnet_info *vi,
527 struct receive_queue *rq,
528 struct page *page, unsigned int offset,
529 unsigned int len, unsigned int truesize,
530 unsigned int headroom)
531{
532 struct sk_buff *skb;
533 struct virtio_net_common_hdr *hdr;
534 unsigned int copy, hdr_len, hdr_padded_len;
535 struct page *page_to_free = NULL;
536 int tailroom, shinfo_size;
537 char *p, *hdr_p, *buf;
538
539 p = page_address(page) + offset;
540 hdr_p = p;
541
542 hdr_len = vi->hdr_len;
543 if (vi->mergeable_rx_bufs)
544 hdr_padded_len = hdr_len;
545 else
546 hdr_padded_len = sizeof(struct padded_vnet_hdr);
547
548 buf = p - headroom;
549 len -= hdr_len;
550 offset += hdr_padded_len;
551 p += hdr_padded_len;
552 tailroom = truesize - headroom - hdr_padded_len - len;
553
554 shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
555
556 /* copy small packet so we can reuse these pages */
557 if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
558 skb = virtnet_build_skb(buf, truesize, p - buf, len);
559 if (unlikely(!skb))
560 return NULL;
561
562 page = (struct page *)page->private;
563 if (page)
564 give_pages(rq, page);
565 goto ok;
566 }
567
568 /* copy small packet so we can reuse these pages for small data */
569 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
570 if (unlikely(!skb))
571 return NULL;
572
573 /* Copy all frame if it fits skb->head, otherwise
574 * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
575 */
576 if (len <= skb_tailroom(skb))
577 copy = len;
578 else
579 copy = ETH_HLEN;
580 skb_put_data(skb, p, copy);
581
582 len -= copy;
583 offset += copy;
584
585 if (vi->mergeable_rx_bufs) {
586 if (len)
587 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
588 else
589 page_to_free = page;
590 goto ok;
591 }
592
593 /*
594 * Verify that we can indeed put this data into a skb.
595 * This is here to handle cases when the device erroneously
596 * tries to receive more than is possible. This is usually
597 * the case of a broken device.
598 */
599 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
600 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
601 dev_kfree_skb(skb);
602 return NULL;
603 }
604 BUG_ON(offset >= PAGE_SIZE);
605 while (len) {
606 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
607 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
608 frag_size, truesize);
609 len -= frag_size;
610 page = (struct page *)page->private;
611 offset = 0;
612 }
613
614 if (page)
615 give_pages(rq, page);
616
617ok:
618 hdr = skb_vnet_common_hdr(skb);
619 memcpy(hdr, hdr_p, hdr_len);
620 if (page_to_free)
621 put_page(page_to_free);
622
623 return skb;
624}
625
626static void virtnet_rq_unmap(struct receive_queue *rq, void *buf, u32 len)
627{
628 struct page *page = virt_to_head_page(buf);
629 struct virtnet_rq_dma *dma;
630 void *head;
631 int offset;
632
633 head = page_address(page);
634
635 dma = head;
636
637 --dma->ref;
638
639 if (dma->need_sync && len) {
640 offset = buf - (head + sizeof(*dma));
641
642 virtqueue_dma_sync_single_range_for_cpu(rq->vq, dma->addr,
643 offset, len,
644 DMA_FROM_DEVICE);
645 }
646
647 if (dma->ref)
648 return;
649
650 virtqueue_dma_unmap_single_attrs(rq->vq, dma->addr, dma->len,
651 DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
652 put_page(page);
653}
654
655static void *virtnet_rq_get_buf(struct receive_queue *rq, u32 *len, void **ctx)
656{
657 void *buf;
658
659 buf = virtqueue_get_buf_ctx(rq->vq, len, ctx);
660 if (buf && rq->do_dma)
661 virtnet_rq_unmap(rq, buf, *len);
662
663 return buf;
664}
665
666static void virtnet_rq_init_one_sg(struct receive_queue *rq, void *buf, u32 len)
667{
668 struct virtnet_rq_dma *dma;
669 dma_addr_t addr;
670 u32 offset;
671 void *head;
672
673 if (!rq->do_dma) {
674 sg_init_one(rq->sg, buf, len);
675 return;
676 }
677
678 head = page_address(rq->alloc_frag.page);
679
680 offset = buf - head;
681
682 dma = head;
683
684 addr = dma->addr - sizeof(*dma) + offset;
685
686 sg_init_table(rq->sg, 1);
687 rq->sg[0].dma_address = addr;
688 rq->sg[0].length = len;
689}
690
691static void *virtnet_rq_alloc(struct receive_queue *rq, u32 size, gfp_t gfp)
692{
693 struct page_frag *alloc_frag = &rq->alloc_frag;
694 struct virtnet_rq_dma *dma;
695 void *buf, *head;
696 dma_addr_t addr;
697
698 if (unlikely(!skb_page_frag_refill(size, alloc_frag, gfp)))
699 return NULL;
700
701 head = page_address(alloc_frag->page);
702
703 if (rq->do_dma) {
704 dma = head;
705
706 /* new pages */
707 if (!alloc_frag->offset) {
708 if (rq->last_dma) {
709 /* Now, the new page is allocated, the last dma
710 * will not be used. So the dma can be unmapped
711 * if the ref is 0.
712 */
713 virtnet_rq_unmap(rq, rq->last_dma, 0);
714 rq->last_dma = NULL;
715 }
716
717 dma->len = alloc_frag->size - sizeof(*dma);
718
719 addr = virtqueue_dma_map_single_attrs(rq->vq, dma + 1,
720 dma->len, DMA_FROM_DEVICE, 0);
721 if (virtqueue_dma_mapping_error(rq->vq, addr))
722 return NULL;
723
724 dma->addr = addr;
725 dma->need_sync = virtqueue_dma_need_sync(rq->vq, addr);
726
727 /* Add a reference to dma to prevent the entire dma from
728 * being released during error handling. This reference
729 * will be freed after the pages are no longer used.
730 */
731 get_page(alloc_frag->page);
732 dma->ref = 1;
733 alloc_frag->offset = sizeof(*dma);
734
735 rq->last_dma = dma;
736 }
737
738 ++dma->ref;
739 }
740
741 buf = head + alloc_frag->offset;
742
743 get_page(alloc_frag->page);
744 alloc_frag->offset += size;
745
746 return buf;
747}
748
749static void virtnet_rq_set_premapped(struct virtnet_info *vi)
750{
751 int i;
752
753 /* disable for big mode */
754 if (!vi->mergeable_rx_bufs && vi->big_packets)
755 return;
756
757 for (i = 0; i < vi->max_queue_pairs; i++) {
758 if (virtqueue_set_dma_premapped(vi->rq[i].vq))
759 continue;
760
761 vi->rq[i].do_dma = true;
762 }
763}
764
765static void virtnet_rq_unmap_free_buf(struct virtqueue *vq, void *buf)
766{
767 struct virtnet_info *vi = vq->vdev->priv;
768 struct receive_queue *rq;
769 int i = vq2rxq(vq);
770
771 rq = &vi->rq[i];
772
773 if (rq->do_dma)
774 virtnet_rq_unmap(rq, buf, 0);
775
776 virtnet_rq_free_buf(vi, rq, buf);
777}
778
779static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
780{
781 unsigned int len;
782 unsigned int packets = 0;
783 unsigned int bytes = 0;
784 void *ptr;
785
786 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
787 if (likely(!is_xdp_frame(ptr))) {
788 struct sk_buff *skb = ptr;
789
790 pr_debug("Sent skb %p\n", skb);
791
792 bytes += skb->len;
793 napi_consume_skb(skb, in_napi);
794 } else {
795 struct xdp_frame *frame = ptr_to_xdp(ptr);
796
797 bytes += xdp_get_frame_len(frame);
798 xdp_return_frame(frame);
799 }
800 packets++;
801 }
802
803 /* Avoid overhead when no packets have been processed
804 * happens when called speculatively from start_xmit.
805 */
806 if (!packets)
807 return;
808
809 u64_stats_update_begin(&sq->stats.syncp);
810 u64_stats_add(&sq->stats.bytes, bytes);
811 u64_stats_add(&sq->stats.packets, packets);
812 u64_stats_update_end(&sq->stats.syncp);
813}
814
815static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
816{
817 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
818 return false;
819 else if (q < vi->curr_queue_pairs)
820 return true;
821 else
822 return false;
823}
824
825static void check_sq_full_and_disable(struct virtnet_info *vi,
826 struct net_device *dev,
827 struct send_queue *sq)
828{
829 bool use_napi = sq->napi.weight;
830 int qnum;
831
832 qnum = sq - vi->sq;
833
834 /* If running out of space, stop queue to avoid getting packets that we
835 * are then unable to transmit.
836 * An alternative would be to force queuing layer to requeue the skb by
837 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
838 * returned in a normal path of operation: it means that driver is not
839 * maintaining the TX queue stop/start state properly, and causes
840 * the stack to do a non-trivial amount of useless work.
841 * Since most packets only take 1 or 2 ring slots, stopping the queue
842 * early means 16 slots are typically wasted.
843 */
844 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
845 netif_stop_subqueue(dev, qnum);
846 if (use_napi) {
847 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq)))
848 virtqueue_napi_schedule(&sq->napi, sq->vq);
849 } else if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
850 /* More just got used, free them then recheck. */
851 free_old_xmit_skbs(sq, false);
852 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
853 netif_start_subqueue(dev, qnum);
854 virtqueue_disable_cb(sq->vq);
855 }
856 }
857 }
858}
859
860static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
861 struct send_queue *sq,
862 struct xdp_frame *xdpf)
863{
864 struct virtio_net_hdr_mrg_rxbuf *hdr;
865 struct skb_shared_info *shinfo;
866 u8 nr_frags = 0;
867 int err, i;
868
869 if (unlikely(xdpf->headroom < vi->hdr_len))
870 return -EOVERFLOW;
871
872 if (unlikely(xdp_frame_has_frags(xdpf))) {
873 shinfo = xdp_get_shared_info_from_frame(xdpf);
874 nr_frags = shinfo->nr_frags;
875 }
876
877 /* In wrapping function virtnet_xdp_xmit(), we need to free
878 * up the pending old buffers, where we need to calculate the
879 * position of skb_shared_info in xdp_get_frame_len() and
880 * xdp_return_frame(), which will involve to xdpf->data and
881 * xdpf->headroom. Therefore, we need to update the value of
882 * headroom synchronously here.
883 */
884 xdpf->headroom -= vi->hdr_len;
885 xdpf->data -= vi->hdr_len;
886 /* Zero header and leave csum up to XDP layers */
887 hdr = xdpf->data;
888 memset(hdr, 0, vi->hdr_len);
889 xdpf->len += vi->hdr_len;
890
891 sg_init_table(sq->sg, nr_frags + 1);
892 sg_set_buf(sq->sg, xdpf->data, xdpf->len);
893 for (i = 0; i < nr_frags; i++) {
894 skb_frag_t *frag = &shinfo->frags[i];
895
896 sg_set_page(&sq->sg[i + 1], skb_frag_page(frag),
897 skb_frag_size(frag), skb_frag_off(frag));
898 }
899
900 err = virtqueue_add_outbuf(sq->vq, sq->sg, nr_frags + 1,
901 xdp_to_ptr(xdpf), GFP_ATOMIC);
902 if (unlikely(err))
903 return -ENOSPC; /* Caller handle free/refcnt */
904
905 return 0;
906}
907
908/* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
909 * the current cpu, so it does not need to be locked.
910 *
911 * Here we use marco instead of inline functions because we have to deal with
912 * three issues at the same time: 1. the choice of sq. 2. judge and execute the
913 * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
914 * functions to perfectly solve these three problems at the same time.
915 */
916#define virtnet_xdp_get_sq(vi) ({ \
917 int cpu = smp_processor_id(); \
918 struct netdev_queue *txq; \
919 typeof(vi) v = (vi); \
920 unsigned int qp; \
921 \
922 if (v->curr_queue_pairs > nr_cpu_ids) { \
923 qp = v->curr_queue_pairs - v->xdp_queue_pairs; \
924 qp += cpu; \
925 txq = netdev_get_tx_queue(v->dev, qp); \
926 __netif_tx_acquire(txq); \
927 } else { \
928 qp = cpu % v->curr_queue_pairs; \
929 txq = netdev_get_tx_queue(v->dev, qp); \
930 __netif_tx_lock(txq, cpu); \
931 } \
932 v->sq + qp; \
933})
934
935#define virtnet_xdp_put_sq(vi, q) { \
936 struct netdev_queue *txq; \
937 typeof(vi) v = (vi); \
938 \
939 txq = netdev_get_tx_queue(v->dev, (q) - v->sq); \
940 if (v->curr_queue_pairs > nr_cpu_ids) \
941 __netif_tx_release(txq); \
942 else \
943 __netif_tx_unlock(txq); \
944}
945
946static int virtnet_xdp_xmit(struct net_device *dev,
947 int n, struct xdp_frame **frames, u32 flags)
948{
949 struct virtnet_info *vi = netdev_priv(dev);
950 struct receive_queue *rq = vi->rq;
951 struct bpf_prog *xdp_prog;
952 struct send_queue *sq;
953 unsigned int len;
954 int packets = 0;
955 int bytes = 0;
956 int nxmit = 0;
957 int kicks = 0;
958 void *ptr;
959 int ret;
960 int i;
961
962 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
963 * indicate XDP resources have been successfully allocated.
964 */
965 xdp_prog = rcu_access_pointer(rq->xdp_prog);
966 if (!xdp_prog)
967 return -ENXIO;
968
969 sq = virtnet_xdp_get_sq(vi);
970
971 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
972 ret = -EINVAL;
973 goto out;
974 }
975
976 /* Free up any pending old buffers before queueing new ones. */
977 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
978 if (likely(is_xdp_frame(ptr))) {
979 struct xdp_frame *frame = ptr_to_xdp(ptr);
980
981 bytes += xdp_get_frame_len(frame);
982 xdp_return_frame(frame);
983 } else {
984 struct sk_buff *skb = ptr;
985
986 bytes += skb->len;
987 napi_consume_skb(skb, false);
988 }
989 packets++;
990 }
991
992 for (i = 0; i < n; i++) {
993 struct xdp_frame *xdpf = frames[i];
994
995 if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
996 break;
997 nxmit++;
998 }
999 ret = nxmit;
1000
1001 if (!is_xdp_raw_buffer_queue(vi, sq - vi->sq))
1002 check_sq_full_and_disable(vi, dev, sq);
1003
1004 if (flags & XDP_XMIT_FLUSH) {
1005 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
1006 kicks = 1;
1007 }
1008out:
1009 u64_stats_update_begin(&sq->stats.syncp);
1010 u64_stats_add(&sq->stats.bytes, bytes);
1011 u64_stats_add(&sq->stats.packets, packets);
1012 u64_stats_add(&sq->stats.xdp_tx, n);
1013 u64_stats_add(&sq->stats.xdp_tx_drops, n - nxmit);
1014 u64_stats_add(&sq->stats.kicks, kicks);
1015 u64_stats_update_end(&sq->stats.syncp);
1016
1017 virtnet_xdp_put_sq(vi, sq);
1018 return ret;
1019}
1020
1021static void put_xdp_frags(struct xdp_buff *xdp)
1022{
1023 struct skb_shared_info *shinfo;
1024 struct page *xdp_page;
1025 int i;
1026
1027 if (xdp_buff_has_frags(xdp)) {
1028 shinfo = xdp_get_shared_info_from_buff(xdp);
1029 for (i = 0; i < shinfo->nr_frags; i++) {
1030 xdp_page = skb_frag_page(&shinfo->frags[i]);
1031 put_page(xdp_page);
1032 }
1033 }
1034}
1035
1036static int virtnet_xdp_handler(struct bpf_prog *xdp_prog, struct xdp_buff *xdp,
1037 struct net_device *dev,
1038 unsigned int *xdp_xmit,
1039 struct virtnet_rq_stats *stats)
1040{
1041 struct xdp_frame *xdpf;
1042 int err;
1043 u32 act;
1044
1045 act = bpf_prog_run_xdp(xdp_prog, xdp);
1046 u64_stats_inc(&stats->xdp_packets);
1047
1048 switch (act) {
1049 case XDP_PASS:
1050 return act;
1051
1052 case XDP_TX:
1053 u64_stats_inc(&stats->xdp_tx);
1054 xdpf = xdp_convert_buff_to_frame(xdp);
1055 if (unlikely(!xdpf)) {
1056 netdev_dbg(dev, "convert buff to frame failed for xdp\n");
1057 return XDP_DROP;
1058 }
1059
1060 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
1061 if (unlikely(!err)) {
1062 xdp_return_frame_rx_napi(xdpf);
1063 } else if (unlikely(err < 0)) {
1064 trace_xdp_exception(dev, xdp_prog, act);
1065 return XDP_DROP;
1066 }
1067 *xdp_xmit |= VIRTIO_XDP_TX;
1068 return act;
1069
1070 case XDP_REDIRECT:
1071 u64_stats_inc(&stats->xdp_redirects);
1072 err = xdp_do_redirect(dev, xdp, xdp_prog);
1073 if (err)
1074 return XDP_DROP;
1075
1076 *xdp_xmit |= VIRTIO_XDP_REDIR;
1077 return act;
1078
1079 default:
1080 bpf_warn_invalid_xdp_action(dev, xdp_prog, act);
1081 fallthrough;
1082 case XDP_ABORTED:
1083 trace_xdp_exception(dev, xdp_prog, act);
1084 fallthrough;
1085 case XDP_DROP:
1086 return XDP_DROP;
1087 }
1088}
1089
1090static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
1091{
1092 return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
1093}
1094
1095/* We copy the packet for XDP in the following cases:
1096 *
1097 * 1) Packet is scattered across multiple rx buffers.
1098 * 2) Headroom space is insufficient.
1099 *
1100 * This is inefficient but it's a temporary condition that
1101 * we hit right after XDP is enabled and until queue is refilled
1102 * with large buffers with sufficient headroom - so it should affect
1103 * at most queue size packets.
1104 * Afterwards, the conditions to enable
1105 * XDP should preclude the underlying device from sending packets
1106 * across multiple buffers (num_buf > 1), and we make sure buffers
1107 * have enough headroom.
1108 */
1109static struct page *xdp_linearize_page(struct receive_queue *rq,
1110 int *num_buf,
1111 struct page *p,
1112 int offset,
1113 int page_off,
1114 unsigned int *len)
1115{
1116 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1117 struct page *page;
1118
1119 if (page_off + *len + tailroom > PAGE_SIZE)
1120 return NULL;
1121
1122 page = alloc_page(GFP_ATOMIC);
1123 if (!page)
1124 return NULL;
1125
1126 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
1127 page_off += *len;
1128
1129 while (--*num_buf) {
1130 unsigned int buflen;
1131 void *buf;
1132 int off;
1133
1134 buf = virtnet_rq_get_buf(rq, &buflen, NULL);
1135 if (unlikely(!buf))
1136 goto err_buf;
1137
1138 p = virt_to_head_page(buf);
1139 off = buf - page_address(p);
1140
1141 /* guard against a misconfigured or uncooperative backend that
1142 * is sending packet larger than the MTU.
1143 */
1144 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
1145 put_page(p);
1146 goto err_buf;
1147 }
1148
1149 memcpy(page_address(page) + page_off,
1150 page_address(p) + off, buflen);
1151 page_off += buflen;
1152 put_page(p);
1153 }
1154
1155 /* Headroom does not contribute to packet length */
1156 *len = page_off - VIRTIO_XDP_HEADROOM;
1157 return page;
1158err_buf:
1159 __free_pages(page, 0);
1160 return NULL;
1161}
1162
1163static struct sk_buff *receive_small_build_skb(struct virtnet_info *vi,
1164 unsigned int xdp_headroom,
1165 void *buf,
1166 unsigned int len)
1167{
1168 unsigned int header_offset;
1169 unsigned int headroom;
1170 unsigned int buflen;
1171 struct sk_buff *skb;
1172
1173 header_offset = VIRTNET_RX_PAD + xdp_headroom;
1174 headroom = vi->hdr_len + header_offset;
1175 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
1176 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1177
1178 skb = virtnet_build_skb(buf, buflen, headroom, len);
1179 if (unlikely(!skb))
1180 return NULL;
1181
1182 buf += header_offset;
1183 memcpy(skb_vnet_common_hdr(skb), buf, vi->hdr_len);
1184
1185 return skb;
1186}
1187
1188static struct sk_buff *receive_small_xdp(struct net_device *dev,
1189 struct virtnet_info *vi,
1190 struct receive_queue *rq,
1191 struct bpf_prog *xdp_prog,
1192 void *buf,
1193 unsigned int xdp_headroom,
1194 unsigned int len,
1195 unsigned int *xdp_xmit,
1196 struct virtnet_rq_stats *stats)
1197{
1198 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
1199 unsigned int headroom = vi->hdr_len + header_offset;
1200 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
1201 struct page *page = virt_to_head_page(buf);
1202 struct page *xdp_page;
1203 unsigned int buflen;
1204 struct xdp_buff xdp;
1205 struct sk_buff *skb;
1206 unsigned int metasize = 0;
1207 u32 act;
1208
1209 if (unlikely(hdr->hdr.gso_type))
1210 goto err_xdp;
1211
1212 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
1213 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1214
1215 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
1216 int offset = buf - page_address(page) + header_offset;
1217 unsigned int tlen = len + vi->hdr_len;
1218 int num_buf = 1;
1219
1220 xdp_headroom = virtnet_get_headroom(vi);
1221 header_offset = VIRTNET_RX_PAD + xdp_headroom;
1222 headroom = vi->hdr_len + header_offset;
1223 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
1224 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1225 xdp_page = xdp_linearize_page(rq, &num_buf, page,
1226 offset, header_offset,
1227 &tlen);
1228 if (!xdp_page)
1229 goto err_xdp;
1230
1231 buf = page_address(xdp_page);
1232 put_page(page);
1233 page = xdp_page;
1234 }
1235
1236 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
1237 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
1238 xdp_headroom, len, true);
1239
1240 act = virtnet_xdp_handler(xdp_prog, &xdp, dev, xdp_xmit, stats);
1241
1242 switch (act) {
1243 case XDP_PASS:
1244 /* Recalculate length in case bpf program changed it */
1245 len = xdp.data_end - xdp.data;
1246 metasize = xdp.data - xdp.data_meta;
1247 break;
1248
1249 case XDP_TX:
1250 case XDP_REDIRECT:
1251 goto xdp_xmit;
1252
1253 default:
1254 goto err_xdp;
1255 }
1256
1257 skb = virtnet_build_skb(buf, buflen, xdp.data - buf, len);
1258 if (unlikely(!skb))
1259 goto err;
1260
1261 if (metasize)
1262 skb_metadata_set(skb, metasize);
1263
1264 return skb;
1265
1266err_xdp:
1267 u64_stats_inc(&stats->xdp_drops);
1268err:
1269 u64_stats_inc(&stats->drops);
1270 put_page(page);
1271xdp_xmit:
1272 return NULL;
1273}
1274
1275static struct sk_buff *receive_small(struct net_device *dev,
1276 struct virtnet_info *vi,
1277 struct receive_queue *rq,
1278 void *buf, void *ctx,
1279 unsigned int len,
1280 unsigned int *xdp_xmit,
1281 struct virtnet_rq_stats *stats)
1282{
1283 unsigned int xdp_headroom = (unsigned long)ctx;
1284 struct page *page = virt_to_head_page(buf);
1285 struct sk_buff *skb;
1286
1287 len -= vi->hdr_len;
1288 u64_stats_add(&stats->bytes, len);
1289
1290 if (unlikely(len > GOOD_PACKET_LEN)) {
1291 pr_debug("%s: rx error: len %u exceeds max size %d\n",
1292 dev->name, len, GOOD_PACKET_LEN);
1293 DEV_STATS_INC(dev, rx_length_errors);
1294 goto err;
1295 }
1296
1297 if (unlikely(vi->xdp_enabled)) {
1298 struct bpf_prog *xdp_prog;
1299
1300 rcu_read_lock();
1301 xdp_prog = rcu_dereference(rq->xdp_prog);
1302 if (xdp_prog) {
1303 skb = receive_small_xdp(dev, vi, rq, xdp_prog, buf,
1304 xdp_headroom, len, xdp_xmit,
1305 stats);
1306 rcu_read_unlock();
1307 return skb;
1308 }
1309 rcu_read_unlock();
1310 }
1311
1312 skb = receive_small_build_skb(vi, xdp_headroom, buf, len);
1313 if (likely(skb))
1314 return skb;
1315
1316err:
1317 u64_stats_inc(&stats->drops);
1318 put_page(page);
1319 return NULL;
1320}
1321
1322static struct sk_buff *receive_big(struct net_device *dev,
1323 struct virtnet_info *vi,
1324 struct receive_queue *rq,
1325 void *buf,
1326 unsigned int len,
1327 struct virtnet_rq_stats *stats)
1328{
1329 struct page *page = buf;
1330 struct sk_buff *skb =
1331 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, 0);
1332
1333 u64_stats_add(&stats->bytes, len - vi->hdr_len);
1334 if (unlikely(!skb))
1335 goto err;
1336
1337 return skb;
1338
1339err:
1340 u64_stats_inc(&stats->drops);
1341 give_pages(rq, page);
1342 return NULL;
1343}
1344
1345static void mergeable_buf_free(struct receive_queue *rq, int num_buf,
1346 struct net_device *dev,
1347 struct virtnet_rq_stats *stats)
1348{
1349 struct page *page;
1350 void *buf;
1351 int len;
1352
1353 while (num_buf-- > 1) {
1354 buf = virtnet_rq_get_buf(rq, &len, NULL);
1355 if (unlikely(!buf)) {
1356 pr_debug("%s: rx error: %d buffers missing\n",
1357 dev->name, num_buf);
1358 DEV_STATS_INC(dev, rx_length_errors);
1359 break;
1360 }
1361 u64_stats_add(&stats->bytes, len);
1362 page = virt_to_head_page(buf);
1363 put_page(page);
1364 }
1365}
1366
1367/* Why not use xdp_build_skb_from_frame() ?
1368 * XDP core assumes that xdp frags are PAGE_SIZE in length, while in
1369 * virtio-net there are 2 points that do not match its requirements:
1370 * 1. The size of the prefilled buffer is not fixed before xdp is set.
1371 * 2. xdp_build_skb_from_frame() does more checks that we don't need,
1372 * like eth_type_trans() (which virtio-net does in receive_buf()).
1373 */
1374static struct sk_buff *build_skb_from_xdp_buff(struct net_device *dev,
1375 struct virtnet_info *vi,
1376 struct xdp_buff *xdp,
1377 unsigned int xdp_frags_truesz)
1378{
1379 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
1380 unsigned int headroom, data_len;
1381 struct sk_buff *skb;
1382 int metasize;
1383 u8 nr_frags;
1384
1385 if (unlikely(xdp->data_end > xdp_data_hard_end(xdp))) {
1386 pr_debug("Error building skb as missing reserved tailroom for xdp");
1387 return NULL;
1388 }
1389
1390 if (unlikely(xdp_buff_has_frags(xdp)))
1391 nr_frags = sinfo->nr_frags;
1392
1393 skb = build_skb(xdp->data_hard_start, xdp->frame_sz);
1394 if (unlikely(!skb))
1395 return NULL;
1396
1397 headroom = xdp->data - xdp->data_hard_start;
1398 data_len = xdp->data_end - xdp->data;
1399 skb_reserve(skb, headroom);
1400 __skb_put(skb, data_len);
1401
1402 metasize = xdp->data - xdp->data_meta;
1403 metasize = metasize > 0 ? metasize : 0;
1404 if (metasize)
1405 skb_metadata_set(skb, metasize);
1406
1407 if (unlikely(xdp_buff_has_frags(xdp)))
1408 xdp_update_skb_shared_info(skb, nr_frags,
1409 sinfo->xdp_frags_size,
1410 xdp_frags_truesz,
1411 xdp_buff_is_frag_pfmemalloc(xdp));
1412
1413 return skb;
1414}
1415
1416/* TODO: build xdp in big mode */
1417static int virtnet_build_xdp_buff_mrg(struct net_device *dev,
1418 struct virtnet_info *vi,
1419 struct receive_queue *rq,
1420 struct xdp_buff *xdp,
1421 void *buf,
1422 unsigned int len,
1423 unsigned int frame_sz,
1424 int *num_buf,
1425 unsigned int *xdp_frags_truesize,
1426 struct virtnet_rq_stats *stats)
1427{
1428 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1429 unsigned int headroom, tailroom, room;
1430 unsigned int truesize, cur_frag_size;
1431 struct skb_shared_info *shinfo;
1432 unsigned int xdp_frags_truesz = 0;
1433 struct page *page;
1434 skb_frag_t *frag;
1435 int offset;
1436 void *ctx;
1437
1438 xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq);
1439 xdp_prepare_buff(xdp, buf - VIRTIO_XDP_HEADROOM,
1440 VIRTIO_XDP_HEADROOM + vi->hdr_len, len - vi->hdr_len, true);
1441
1442 if (!*num_buf)
1443 return 0;
1444
1445 if (*num_buf > 1) {
1446 /* If we want to build multi-buffer xdp, we need
1447 * to specify that the flags of xdp_buff have the
1448 * XDP_FLAGS_HAS_FRAG bit.
1449 */
1450 if (!xdp_buff_has_frags(xdp))
1451 xdp_buff_set_frags_flag(xdp);
1452
1453 shinfo = xdp_get_shared_info_from_buff(xdp);
1454 shinfo->nr_frags = 0;
1455 shinfo->xdp_frags_size = 0;
1456 }
1457
1458 if (*num_buf > MAX_SKB_FRAGS + 1)
1459 return -EINVAL;
1460
1461 while (--*num_buf > 0) {
1462 buf = virtnet_rq_get_buf(rq, &len, &ctx);
1463 if (unlikely(!buf)) {
1464 pr_debug("%s: rx error: %d buffers out of %d missing\n",
1465 dev->name, *num_buf,
1466 virtio16_to_cpu(vi->vdev, hdr->num_buffers));
1467 DEV_STATS_INC(dev, rx_length_errors);
1468 goto err;
1469 }
1470
1471 u64_stats_add(&stats->bytes, len);
1472 page = virt_to_head_page(buf);
1473 offset = buf - page_address(page);
1474
1475 truesize = mergeable_ctx_to_truesize(ctx);
1476 headroom = mergeable_ctx_to_headroom(ctx);
1477 tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1478 room = SKB_DATA_ALIGN(headroom + tailroom);
1479
1480 cur_frag_size = truesize;
1481 xdp_frags_truesz += cur_frag_size;
1482 if (unlikely(len > truesize - room || cur_frag_size > PAGE_SIZE)) {
1483 put_page(page);
1484 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1485 dev->name, len, (unsigned long)(truesize - room));
1486 DEV_STATS_INC(dev, rx_length_errors);
1487 goto err;
1488 }
1489
1490 frag = &shinfo->frags[shinfo->nr_frags++];
1491 skb_frag_fill_page_desc(frag, page, offset, len);
1492 if (page_is_pfmemalloc(page))
1493 xdp_buff_set_frag_pfmemalloc(xdp);
1494
1495 shinfo->xdp_frags_size += len;
1496 }
1497
1498 *xdp_frags_truesize = xdp_frags_truesz;
1499 return 0;
1500
1501err:
1502 put_xdp_frags(xdp);
1503 return -EINVAL;
1504}
1505
1506static void *mergeable_xdp_get_buf(struct virtnet_info *vi,
1507 struct receive_queue *rq,
1508 struct bpf_prog *xdp_prog,
1509 void *ctx,
1510 unsigned int *frame_sz,
1511 int *num_buf,
1512 struct page **page,
1513 int offset,
1514 unsigned int *len,
1515 struct virtio_net_hdr_mrg_rxbuf *hdr)
1516{
1517 unsigned int truesize = mergeable_ctx_to_truesize(ctx);
1518 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
1519 struct page *xdp_page;
1520 unsigned int xdp_room;
1521
1522 /* Transient failure which in theory could occur if
1523 * in-flight packets from before XDP was enabled reach
1524 * the receive path after XDP is loaded.
1525 */
1526 if (unlikely(hdr->hdr.gso_type))
1527 return NULL;
1528
1529 /* Now XDP core assumes frag size is PAGE_SIZE, but buffers
1530 * with headroom may add hole in truesize, which
1531 * make their length exceed PAGE_SIZE. So we disabled the
1532 * hole mechanism for xdp. See add_recvbuf_mergeable().
1533 */
1534 *frame_sz = truesize;
1535
1536 if (likely(headroom >= virtnet_get_headroom(vi) &&
1537 (*num_buf == 1 || xdp_prog->aux->xdp_has_frags))) {
1538 return page_address(*page) + offset;
1539 }
1540
1541 /* This happens when headroom is not enough because
1542 * of the buffer was prefilled before XDP is set.
1543 * This should only happen for the first several packets.
1544 * In fact, vq reset can be used here to help us clean up
1545 * the prefilled buffers, but many existing devices do not
1546 * support it, and we don't want to bother users who are
1547 * using xdp normally.
1548 */
1549 if (!xdp_prog->aux->xdp_has_frags) {
1550 /* linearize data for XDP */
1551 xdp_page = xdp_linearize_page(rq, num_buf,
1552 *page, offset,
1553 VIRTIO_XDP_HEADROOM,
1554 len);
1555 if (!xdp_page)
1556 return NULL;
1557 } else {
1558 xdp_room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM +
1559 sizeof(struct skb_shared_info));
1560 if (*len + xdp_room > PAGE_SIZE)
1561 return NULL;
1562
1563 xdp_page = alloc_page(GFP_ATOMIC);
1564 if (!xdp_page)
1565 return NULL;
1566
1567 memcpy(page_address(xdp_page) + VIRTIO_XDP_HEADROOM,
1568 page_address(*page) + offset, *len);
1569 }
1570
1571 *frame_sz = PAGE_SIZE;
1572
1573 put_page(*page);
1574
1575 *page = xdp_page;
1576
1577 return page_address(*page) + VIRTIO_XDP_HEADROOM;
1578}
1579
1580static struct sk_buff *receive_mergeable_xdp(struct net_device *dev,
1581 struct virtnet_info *vi,
1582 struct receive_queue *rq,
1583 struct bpf_prog *xdp_prog,
1584 void *buf,
1585 void *ctx,
1586 unsigned int len,
1587 unsigned int *xdp_xmit,
1588 struct virtnet_rq_stats *stats)
1589{
1590 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1591 int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
1592 struct page *page = virt_to_head_page(buf);
1593 int offset = buf - page_address(page);
1594 unsigned int xdp_frags_truesz = 0;
1595 struct sk_buff *head_skb;
1596 unsigned int frame_sz;
1597 struct xdp_buff xdp;
1598 void *data;
1599 u32 act;
1600 int err;
1601
1602 data = mergeable_xdp_get_buf(vi, rq, xdp_prog, ctx, &frame_sz, &num_buf, &page,
1603 offset, &len, hdr);
1604 if (unlikely(!data))
1605 goto err_xdp;
1606
1607 err = virtnet_build_xdp_buff_mrg(dev, vi, rq, &xdp, data, len, frame_sz,
1608 &num_buf, &xdp_frags_truesz, stats);
1609 if (unlikely(err))
1610 goto err_xdp;
1611
1612 act = virtnet_xdp_handler(xdp_prog, &xdp, dev, xdp_xmit, stats);
1613
1614 switch (act) {
1615 case XDP_PASS:
1616 head_skb = build_skb_from_xdp_buff(dev, vi, &xdp, xdp_frags_truesz);
1617 if (unlikely(!head_skb))
1618 break;
1619 return head_skb;
1620
1621 case XDP_TX:
1622 case XDP_REDIRECT:
1623 return NULL;
1624
1625 default:
1626 break;
1627 }
1628
1629 put_xdp_frags(&xdp);
1630
1631err_xdp:
1632 put_page(page);
1633 mergeable_buf_free(rq, num_buf, dev, stats);
1634
1635 u64_stats_inc(&stats->xdp_drops);
1636 u64_stats_inc(&stats->drops);
1637 return NULL;
1638}
1639
1640static struct sk_buff *receive_mergeable(struct net_device *dev,
1641 struct virtnet_info *vi,
1642 struct receive_queue *rq,
1643 void *buf,
1644 void *ctx,
1645 unsigned int len,
1646 unsigned int *xdp_xmit,
1647 struct virtnet_rq_stats *stats)
1648{
1649 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1650 int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
1651 struct page *page = virt_to_head_page(buf);
1652 int offset = buf - page_address(page);
1653 struct sk_buff *head_skb, *curr_skb;
1654 unsigned int truesize = mergeable_ctx_to_truesize(ctx);
1655 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
1656 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1657 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1658
1659 head_skb = NULL;
1660 u64_stats_add(&stats->bytes, len - vi->hdr_len);
1661
1662 if (unlikely(len > truesize - room)) {
1663 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1664 dev->name, len, (unsigned long)(truesize - room));
1665 DEV_STATS_INC(dev, rx_length_errors);
1666 goto err_skb;
1667 }
1668
1669 if (unlikely(vi->xdp_enabled)) {
1670 struct bpf_prog *xdp_prog;
1671
1672 rcu_read_lock();
1673 xdp_prog = rcu_dereference(rq->xdp_prog);
1674 if (xdp_prog) {
1675 head_skb = receive_mergeable_xdp(dev, vi, rq, xdp_prog, buf, ctx,
1676 len, xdp_xmit, stats);
1677 rcu_read_unlock();
1678 return head_skb;
1679 }
1680 rcu_read_unlock();
1681 }
1682
1683 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, headroom);
1684 curr_skb = head_skb;
1685
1686 if (unlikely(!curr_skb))
1687 goto err_skb;
1688 while (--num_buf) {
1689 int num_skb_frags;
1690
1691 buf = virtnet_rq_get_buf(rq, &len, &ctx);
1692 if (unlikely(!buf)) {
1693 pr_debug("%s: rx error: %d buffers out of %d missing\n",
1694 dev->name, num_buf,
1695 virtio16_to_cpu(vi->vdev,
1696 hdr->num_buffers));
1697 DEV_STATS_INC(dev, rx_length_errors);
1698 goto err_buf;
1699 }
1700
1701 u64_stats_add(&stats->bytes, len);
1702 page = virt_to_head_page(buf);
1703
1704 truesize = mergeable_ctx_to_truesize(ctx);
1705 headroom = mergeable_ctx_to_headroom(ctx);
1706 tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1707 room = SKB_DATA_ALIGN(headroom + tailroom);
1708 if (unlikely(len > truesize - room)) {
1709 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1710 dev->name, len, (unsigned long)(truesize - room));
1711 DEV_STATS_INC(dev, rx_length_errors);
1712 goto err_skb;
1713 }
1714
1715 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1716 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1717 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1718
1719 if (unlikely(!nskb))
1720 goto err_skb;
1721 if (curr_skb == head_skb)
1722 skb_shinfo(curr_skb)->frag_list = nskb;
1723 else
1724 curr_skb->next = nskb;
1725 curr_skb = nskb;
1726 head_skb->truesize += nskb->truesize;
1727 num_skb_frags = 0;
1728 }
1729 if (curr_skb != head_skb) {
1730 head_skb->data_len += len;
1731 head_skb->len += len;
1732 head_skb->truesize += truesize;
1733 }
1734 offset = buf - page_address(page);
1735 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1736 put_page(page);
1737 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1738 len, truesize);
1739 } else {
1740 skb_add_rx_frag(curr_skb, num_skb_frags, page,
1741 offset, len, truesize);
1742 }
1743 }
1744
1745 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1746 return head_skb;
1747
1748err_skb:
1749 put_page(page);
1750 mergeable_buf_free(rq, num_buf, dev, stats);
1751
1752err_buf:
1753 u64_stats_inc(&stats->drops);
1754 dev_kfree_skb(head_skb);
1755 return NULL;
1756}
1757
1758static void virtio_skb_set_hash(const struct virtio_net_hdr_v1_hash *hdr_hash,
1759 struct sk_buff *skb)
1760{
1761 enum pkt_hash_types rss_hash_type;
1762
1763 if (!hdr_hash || !skb)
1764 return;
1765
1766 switch (__le16_to_cpu(hdr_hash->hash_report)) {
1767 case VIRTIO_NET_HASH_REPORT_TCPv4:
1768 case VIRTIO_NET_HASH_REPORT_UDPv4:
1769 case VIRTIO_NET_HASH_REPORT_TCPv6:
1770 case VIRTIO_NET_HASH_REPORT_UDPv6:
1771 case VIRTIO_NET_HASH_REPORT_TCPv6_EX:
1772 case VIRTIO_NET_HASH_REPORT_UDPv6_EX:
1773 rss_hash_type = PKT_HASH_TYPE_L4;
1774 break;
1775 case VIRTIO_NET_HASH_REPORT_IPv4:
1776 case VIRTIO_NET_HASH_REPORT_IPv6:
1777 case VIRTIO_NET_HASH_REPORT_IPv6_EX:
1778 rss_hash_type = PKT_HASH_TYPE_L3;
1779 break;
1780 case VIRTIO_NET_HASH_REPORT_NONE:
1781 default:
1782 rss_hash_type = PKT_HASH_TYPE_NONE;
1783 }
1784 skb_set_hash(skb, __le32_to_cpu(hdr_hash->hash_value), rss_hash_type);
1785}
1786
1787static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1788 void *buf, unsigned int len, void **ctx,
1789 unsigned int *xdp_xmit,
1790 struct virtnet_rq_stats *stats)
1791{
1792 struct net_device *dev = vi->dev;
1793 struct sk_buff *skb;
1794 struct virtio_net_common_hdr *hdr;
1795
1796 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1797 pr_debug("%s: short packet %i\n", dev->name, len);
1798 DEV_STATS_INC(dev, rx_length_errors);
1799 virtnet_rq_free_buf(vi, rq, buf);
1800 return;
1801 }
1802
1803 if (vi->mergeable_rx_bufs)
1804 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1805 stats);
1806 else if (vi->big_packets)
1807 skb = receive_big(dev, vi, rq, buf, len, stats);
1808 else
1809 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1810
1811 if (unlikely(!skb))
1812 return;
1813
1814 hdr = skb_vnet_common_hdr(skb);
1815 if (dev->features & NETIF_F_RXHASH && vi->has_rss_hash_report)
1816 virtio_skb_set_hash(&hdr->hash_v1_hdr, skb);
1817
1818 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1819 skb->ip_summed = CHECKSUM_UNNECESSARY;
1820
1821 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1822 virtio_is_little_endian(vi->vdev))) {
1823 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1824 dev->name, hdr->hdr.gso_type,
1825 hdr->hdr.gso_size);
1826 goto frame_err;
1827 }
1828
1829 skb_record_rx_queue(skb, vq2rxq(rq->vq));
1830 skb->protocol = eth_type_trans(skb, dev);
1831 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1832 ntohs(skb->protocol), skb->len, skb->pkt_type);
1833
1834 napi_gro_receive(&rq->napi, skb);
1835 return;
1836
1837frame_err:
1838 DEV_STATS_INC(dev, rx_frame_errors);
1839 dev_kfree_skb(skb);
1840}
1841
1842/* Unlike mergeable buffers, all buffers are allocated to the
1843 * same size, except for the headroom. For this reason we do
1844 * not need to use mergeable_len_to_ctx here - it is enough
1845 * to store the headroom as the context ignoring the truesize.
1846 */
1847static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1848 gfp_t gfp)
1849{
1850 char *buf;
1851 unsigned int xdp_headroom = virtnet_get_headroom(vi);
1852 void *ctx = (void *)(unsigned long)xdp_headroom;
1853 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1854 int err;
1855
1856 len = SKB_DATA_ALIGN(len) +
1857 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1858
1859 buf = virtnet_rq_alloc(rq, len, gfp);
1860 if (unlikely(!buf))
1861 return -ENOMEM;
1862
1863 virtnet_rq_init_one_sg(rq, buf + VIRTNET_RX_PAD + xdp_headroom,
1864 vi->hdr_len + GOOD_PACKET_LEN);
1865
1866 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1867 if (err < 0) {
1868 if (rq->do_dma)
1869 virtnet_rq_unmap(rq, buf, 0);
1870 put_page(virt_to_head_page(buf));
1871 }
1872
1873 return err;
1874}
1875
1876static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1877 gfp_t gfp)
1878{
1879 struct page *first, *list = NULL;
1880 char *p;
1881 int i, err, offset;
1882
1883 sg_init_table(rq->sg, vi->big_packets_num_skbfrags + 2);
1884
1885 /* page in rq->sg[vi->big_packets_num_skbfrags + 1] is list tail */
1886 for (i = vi->big_packets_num_skbfrags + 1; i > 1; --i) {
1887 first = get_a_page(rq, gfp);
1888 if (!first) {
1889 if (list)
1890 give_pages(rq, list);
1891 return -ENOMEM;
1892 }
1893 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1894
1895 /* chain new page in list head to match sg */
1896 first->private = (unsigned long)list;
1897 list = first;
1898 }
1899
1900 first = get_a_page(rq, gfp);
1901 if (!first) {
1902 give_pages(rq, list);
1903 return -ENOMEM;
1904 }
1905 p = page_address(first);
1906
1907 /* rq->sg[0], rq->sg[1] share the same page */
1908 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1909 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1910
1911 /* rq->sg[1] for data packet, from offset */
1912 offset = sizeof(struct padded_vnet_hdr);
1913 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1914
1915 /* chain first in list head */
1916 first->private = (unsigned long)list;
1917 err = virtqueue_add_inbuf(rq->vq, rq->sg, vi->big_packets_num_skbfrags + 2,
1918 first, gfp);
1919 if (err < 0)
1920 give_pages(rq, first);
1921
1922 return err;
1923}
1924
1925static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1926 struct ewma_pkt_len *avg_pkt_len,
1927 unsigned int room)
1928{
1929 struct virtnet_info *vi = rq->vq->vdev->priv;
1930 const size_t hdr_len = vi->hdr_len;
1931 unsigned int len;
1932
1933 if (room)
1934 return PAGE_SIZE - room;
1935
1936 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1937 rq->min_buf_len, PAGE_SIZE - hdr_len);
1938
1939 return ALIGN(len, L1_CACHE_BYTES);
1940}
1941
1942static int add_recvbuf_mergeable(struct virtnet_info *vi,
1943 struct receive_queue *rq, gfp_t gfp)
1944{
1945 struct page_frag *alloc_frag = &rq->alloc_frag;
1946 unsigned int headroom = virtnet_get_headroom(vi);
1947 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1948 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1949 unsigned int len, hole;
1950 void *ctx;
1951 char *buf;
1952 int err;
1953
1954 /* Extra tailroom is needed to satisfy XDP's assumption. This
1955 * means rx frags coalescing won't work, but consider we've
1956 * disabled GSO for XDP, it won't be a big issue.
1957 */
1958 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1959
1960 buf = virtnet_rq_alloc(rq, len + room, gfp);
1961 if (unlikely(!buf))
1962 return -ENOMEM;
1963
1964 buf += headroom; /* advance address leaving hole at front of pkt */
1965 hole = alloc_frag->size - alloc_frag->offset;
1966 if (hole < len + room) {
1967 /* To avoid internal fragmentation, if there is very likely not
1968 * enough space for another buffer, add the remaining space to
1969 * the current buffer.
1970 * XDP core assumes that frame_size of xdp_buff and the length
1971 * of the frag are PAGE_SIZE, so we disable the hole mechanism.
1972 */
1973 if (!headroom)
1974 len += hole;
1975 alloc_frag->offset += hole;
1976 }
1977
1978 virtnet_rq_init_one_sg(rq, buf, len);
1979
1980 ctx = mergeable_len_to_ctx(len + room, headroom);
1981 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1982 if (err < 0) {
1983 if (rq->do_dma)
1984 virtnet_rq_unmap(rq, buf, 0);
1985 put_page(virt_to_head_page(buf));
1986 }
1987
1988 return err;
1989}
1990
1991/*
1992 * Returns false if we couldn't fill entirely (OOM).
1993 *
1994 * Normally run in the receive path, but can also be run from ndo_open
1995 * before we're receiving packets, or from refill_work which is
1996 * careful to disable receiving (using napi_disable).
1997 */
1998static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1999 gfp_t gfp)
2000{
2001 int err;
2002 bool oom;
2003
2004 do {
2005 if (vi->mergeable_rx_bufs)
2006 err = add_recvbuf_mergeable(vi, rq, gfp);
2007 else if (vi->big_packets)
2008 err = add_recvbuf_big(vi, rq, gfp);
2009 else
2010 err = add_recvbuf_small(vi, rq, gfp);
2011
2012 oom = err == -ENOMEM;
2013 if (err)
2014 break;
2015 } while (rq->vq->num_free);
2016 if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
2017 unsigned long flags;
2018
2019 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
2020 u64_stats_inc(&rq->stats.kicks);
2021 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
2022 }
2023
2024 return !oom;
2025}
2026
2027static void skb_recv_done(struct virtqueue *rvq)
2028{
2029 struct virtnet_info *vi = rvq->vdev->priv;
2030 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
2031
2032 rq->calls++;
2033 virtqueue_napi_schedule(&rq->napi, rvq);
2034}
2035
2036static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
2037{
2038 napi_enable(napi);
2039
2040 /* If all buffers were filled by other side before we napi_enabled, we
2041 * won't get another interrupt, so process any outstanding packets now.
2042 * Call local_bh_enable after to trigger softIRQ processing.
2043 */
2044 local_bh_disable();
2045 virtqueue_napi_schedule(napi, vq);
2046 local_bh_enable();
2047}
2048
2049static void virtnet_napi_tx_enable(struct virtnet_info *vi,
2050 struct virtqueue *vq,
2051 struct napi_struct *napi)
2052{
2053 if (!napi->weight)
2054 return;
2055
2056 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
2057 * enable the feature if this is likely affine with the transmit path.
2058 */
2059 if (!vi->affinity_hint_set) {
2060 napi->weight = 0;
2061 return;
2062 }
2063
2064 return virtnet_napi_enable(vq, napi);
2065}
2066
2067static void virtnet_napi_tx_disable(struct napi_struct *napi)
2068{
2069 if (napi->weight)
2070 napi_disable(napi);
2071}
2072
2073static void refill_work(struct work_struct *work)
2074{
2075 struct virtnet_info *vi =
2076 container_of(work, struct virtnet_info, refill.work);
2077 bool still_empty;
2078 int i;
2079
2080 for (i = 0; i < vi->curr_queue_pairs; i++) {
2081 struct receive_queue *rq = &vi->rq[i];
2082
2083 napi_disable(&rq->napi);
2084 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
2085 virtnet_napi_enable(rq->vq, &rq->napi);
2086
2087 /* In theory, this can happen: if we don't get any buffers in
2088 * we will *never* try to fill again.
2089 */
2090 if (still_empty)
2091 schedule_delayed_work(&vi->refill, HZ/2);
2092 }
2093}
2094
2095static int virtnet_receive(struct receive_queue *rq, int budget,
2096 unsigned int *xdp_xmit)
2097{
2098 struct virtnet_info *vi = rq->vq->vdev->priv;
2099 struct virtnet_rq_stats stats = {};
2100 unsigned int len;
2101 int packets = 0;
2102 void *buf;
2103 int i;
2104
2105 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2106 void *ctx;
2107
2108 while (packets < budget &&
2109 (buf = virtnet_rq_get_buf(rq, &len, &ctx))) {
2110 receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
2111 packets++;
2112 }
2113 } else {
2114 while (packets < budget &&
2115 (buf = virtnet_rq_get_buf(rq, &len, NULL)) != NULL) {
2116 receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
2117 packets++;
2118 }
2119 }
2120
2121 if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
2122 if (!try_fill_recv(vi, rq, GFP_ATOMIC)) {
2123 spin_lock(&vi->refill_lock);
2124 if (vi->refill_enabled)
2125 schedule_delayed_work(&vi->refill, 0);
2126 spin_unlock(&vi->refill_lock);
2127 }
2128 }
2129
2130 u64_stats_set(&stats.packets, packets);
2131 u64_stats_update_begin(&rq->stats.syncp);
2132 for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
2133 size_t offset = virtnet_rq_stats_desc[i].offset;
2134 u64_stats_t *item, *src;
2135
2136 item = (u64_stats_t *)((u8 *)&rq->stats + offset);
2137 src = (u64_stats_t *)((u8 *)&stats + offset);
2138 u64_stats_add(item, u64_stats_read(src));
2139 }
2140 u64_stats_update_end(&rq->stats.syncp);
2141
2142 return packets;
2143}
2144
2145static void virtnet_poll_cleantx(struct receive_queue *rq)
2146{
2147 struct virtnet_info *vi = rq->vq->vdev->priv;
2148 unsigned int index = vq2rxq(rq->vq);
2149 struct send_queue *sq = &vi->sq[index];
2150 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
2151
2152 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
2153 return;
2154
2155 if (__netif_tx_trylock(txq)) {
2156 if (sq->reset) {
2157 __netif_tx_unlock(txq);
2158 return;
2159 }
2160
2161 do {
2162 virtqueue_disable_cb(sq->vq);
2163 free_old_xmit_skbs(sq, true);
2164 } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
2165
2166 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
2167 netif_tx_wake_queue(txq);
2168
2169 __netif_tx_unlock(txq);
2170 }
2171}
2172
2173static void virtnet_rx_dim_update(struct virtnet_info *vi, struct receive_queue *rq)
2174{
2175 struct dim_sample cur_sample = {};
2176
2177 if (!rq->packets_in_napi)
2178 return;
2179
2180 u64_stats_update_begin(&rq->stats.syncp);
2181 dim_update_sample(rq->calls,
2182 u64_stats_read(&rq->stats.packets),
2183 u64_stats_read(&rq->stats.bytes),
2184 &cur_sample);
2185 u64_stats_update_end(&rq->stats.syncp);
2186
2187 net_dim(&rq->dim, cur_sample);
2188 rq->packets_in_napi = 0;
2189}
2190
2191static int virtnet_poll(struct napi_struct *napi, int budget)
2192{
2193 struct receive_queue *rq =
2194 container_of(napi, struct receive_queue, napi);
2195 struct virtnet_info *vi = rq->vq->vdev->priv;
2196 struct send_queue *sq;
2197 unsigned int received;
2198 unsigned int xdp_xmit = 0;
2199 bool napi_complete;
2200
2201 virtnet_poll_cleantx(rq);
2202
2203 received = virtnet_receive(rq, budget, &xdp_xmit);
2204 rq->packets_in_napi += received;
2205
2206 if (xdp_xmit & VIRTIO_XDP_REDIR)
2207 xdp_do_flush();
2208
2209 /* Out of packets? */
2210 if (received < budget) {
2211 napi_complete = virtqueue_napi_complete(napi, rq->vq, received);
2212 if (napi_complete && rq->dim_enabled)
2213 virtnet_rx_dim_update(vi, rq);
2214 }
2215
2216 if (xdp_xmit & VIRTIO_XDP_TX) {
2217 sq = virtnet_xdp_get_sq(vi);
2218 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
2219 u64_stats_update_begin(&sq->stats.syncp);
2220 u64_stats_inc(&sq->stats.kicks);
2221 u64_stats_update_end(&sq->stats.syncp);
2222 }
2223 virtnet_xdp_put_sq(vi, sq);
2224 }
2225
2226 return received;
2227}
2228
2229static void virtnet_disable_queue_pair(struct virtnet_info *vi, int qp_index)
2230{
2231 virtnet_napi_tx_disable(&vi->sq[qp_index].napi);
2232 napi_disable(&vi->rq[qp_index].napi);
2233 xdp_rxq_info_unreg(&vi->rq[qp_index].xdp_rxq);
2234}
2235
2236static int virtnet_enable_queue_pair(struct virtnet_info *vi, int qp_index)
2237{
2238 struct net_device *dev = vi->dev;
2239 int err;
2240
2241 err = xdp_rxq_info_reg(&vi->rq[qp_index].xdp_rxq, dev, qp_index,
2242 vi->rq[qp_index].napi.napi_id);
2243 if (err < 0)
2244 return err;
2245
2246 err = xdp_rxq_info_reg_mem_model(&vi->rq[qp_index].xdp_rxq,
2247 MEM_TYPE_PAGE_SHARED, NULL);
2248 if (err < 0)
2249 goto err_xdp_reg_mem_model;
2250
2251 virtnet_napi_enable(vi->rq[qp_index].vq, &vi->rq[qp_index].napi);
2252 virtnet_napi_tx_enable(vi, vi->sq[qp_index].vq, &vi->sq[qp_index].napi);
2253
2254 return 0;
2255
2256err_xdp_reg_mem_model:
2257 xdp_rxq_info_unreg(&vi->rq[qp_index].xdp_rxq);
2258 return err;
2259}
2260
2261static int virtnet_open(struct net_device *dev)
2262{
2263 struct virtnet_info *vi = netdev_priv(dev);
2264 int i, err;
2265
2266 enable_delayed_refill(vi);
2267
2268 for (i = 0; i < vi->max_queue_pairs; i++) {
2269 if (i < vi->curr_queue_pairs)
2270 /* Make sure we have some buffers: if oom use wq. */
2271 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2272 schedule_delayed_work(&vi->refill, 0);
2273
2274 err = virtnet_enable_queue_pair(vi, i);
2275 if (err < 0)
2276 goto err_enable_qp;
2277 }
2278
2279 return 0;
2280
2281err_enable_qp:
2282 disable_delayed_refill(vi);
2283 cancel_delayed_work_sync(&vi->refill);
2284
2285 for (i--; i >= 0; i--) {
2286 virtnet_disable_queue_pair(vi, i);
2287 cancel_work_sync(&vi->rq[i].dim.work);
2288 }
2289
2290 return err;
2291}
2292
2293static int virtnet_poll_tx(struct napi_struct *napi, int budget)
2294{
2295 struct send_queue *sq = container_of(napi, struct send_queue, napi);
2296 struct virtnet_info *vi = sq->vq->vdev->priv;
2297 unsigned int index = vq2txq(sq->vq);
2298 struct netdev_queue *txq;
2299 int opaque;
2300 bool done;
2301
2302 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
2303 /* We don't need to enable cb for XDP */
2304 napi_complete_done(napi, 0);
2305 return 0;
2306 }
2307
2308 txq = netdev_get_tx_queue(vi->dev, index);
2309 __netif_tx_lock(txq, raw_smp_processor_id());
2310 virtqueue_disable_cb(sq->vq);
2311 free_old_xmit_skbs(sq, true);
2312
2313 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
2314 netif_tx_wake_queue(txq);
2315
2316 opaque = virtqueue_enable_cb_prepare(sq->vq);
2317
2318 done = napi_complete_done(napi, 0);
2319
2320 if (!done)
2321 virtqueue_disable_cb(sq->vq);
2322
2323 __netif_tx_unlock(txq);
2324
2325 if (done) {
2326 if (unlikely(virtqueue_poll(sq->vq, opaque))) {
2327 if (napi_schedule_prep(napi)) {
2328 __netif_tx_lock(txq, raw_smp_processor_id());
2329 virtqueue_disable_cb(sq->vq);
2330 __netif_tx_unlock(txq);
2331 __napi_schedule(napi);
2332 }
2333 }
2334 }
2335
2336 return 0;
2337}
2338
2339static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
2340{
2341 struct virtio_net_hdr_mrg_rxbuf *hdr;
2342 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
2343 struct virtnet_info *vi = sq->vq->vdev->priv;
2344 int num_sg;
2345 unsigned hdr_len = vi->hdr_len;
2346 bool can_push;
2347
2348 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
2349
2350 can_push = vi->any_header_sg &&
2351 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
2352 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
2353 /* Even if we can, don't push here yet as this would skew
2354 * csum_start offset below. */
2355 if (can_push)
2356 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
2357 else
2358 hdr = &skb_vnet_common_hdr(skb)->mrg_hdr;
2359
2360 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
2361 virtio_is_little_endian(vi->vdev), false,
2362 0))
2363 return -EPROTO;
2364
2365 if (vi->mergeable_rx_bufs)
2366 hdr->num_buffers = 0;
2367
2368 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
2369 if (can_push) {
2370 __skb_push(skb, hdr_len);
2371 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
2372 if (unlikely(num_sg < 0))
2373 return num_sg;
2374 /* Pull header back to avoid skew in tx bytes calculations. */
2375 __skb_pull(skb, hdr_len);
2376 } else {
2377 sg_set_buf(sq->sg, hdr, hdr_len);
2378 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
2379 if (unlikely(num_sg < 0))
2380 return num_sg;
2381 num_sg++;
2382 }
2383 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
2384}
2385
2386static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
2387{
2388 struct virtnet_info *vi = netdev_priv(dev);
2389 int qnum = skb_get_queue_mapping(skb);
2390 struct send_queue *sq = &vi->sq[qnum];
2391 int err;
2392 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
2393 bool kick = !netdev_xmit_more();
2394 bool use_napi = sq->napi.weight;
2395
2396 /* Free up any pending old buffers before queueing new ones. */
2397 do {
2398 if (use_napi)
2399 virtqueue_disable_cb(sq->vq);
2400
2401 free_old_xmit_skbs(sq, false);
2402
2403 } while (use_napi && kick &&
2404 unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
2405
2406 /* timestamp packet in software */
2407 skb_tx_timestamp(skb);
2408
2409 /* Try to transmit */
2410 err = xmit_skb(sq, skb);
2411
2412 /* This should not happen! */
2413 if (unlikely(err)) {
2414 DEV_STATS_INC(dev, tx_fifo_errors);
2415 if (net_ratelimit())
2416 dev_warn(&dev->dev,
2417 "Unexpected TXQ (%d) queue failure: %d\n",
2418 qnum, err);
2419 DEV_STATS_INC(dev, tx_dropped);
2420 dev_kfree_skb_any(skb);
2421 return NETDEV_TX_OK;
2422 }
2423
2424 /* Don't wait up for transmitted skbs to be freed. */
2425 if (!use_napi) {
2426 skb_orphan(skb);
2427 nf_reset_ct(skb);
2428 }
2429
2430 check_sq_full_and_disable(vi, dev, sq);
2431
2432 if (kick || netif_xmit_stopped(txq)) {
2433 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
2434 u64_stats_update_begin(&sq->stats.syncp);
2435 u64_stats_inc(&sq->stats.kicks);
2436 u64_stats_update_end(&sq->stats.syncp);
2437 }
2438 }
2439
2440 return NETDEV_TX_OK;
2441}
2442
2443static int virtnet_rx_resize(struct virtnet_info *vi,
2444 struct receive_queue *rq, u32 ring_num)
2445{
2446 bool running = netif_running(vi->dev);
2447 int err, qindex;
2448
2449 qindex = rq - vi->rq;
2450
2451 if (running) {
2452 napi_disable(&rq->napi);
2453 cancel_work_sync(&rq->dim.work);
2454 }
2455
2456 err = virtqueue_resize(rq->vq, ring_num, virtnet_rq_unmap_free_buf);
2457 if (err)
2458 netdev_err(vi->dev, "resize rx fail: rx queue index: %d err: %d\n", qindex, err);
2459
2460 if (!try_fill_recv(vi, rq, GFP_KERNEL))
2461 schedule_delayed_work(&vi->refill, 0);
2462
2463 if (running)
2464 virtnet_napi_enable(rq->vq, &rq->napi);
2465 return err;
2466}
2467
2468static int virtnet_tx_resize(struct virtnet_info *vi,
2469 struct send_queue *sq, u32 ring_num)
2470{
2471 bool running = netif_running(vi->dev);
2472 struct netdev_queue *txq;
2473 int err, qindex;
2474
2475 qindex = sq - vi->sq;
2476
2477 if (running)
2478 virtnet_napi_tx_disable(&sq->napi);
2479
2480 txq = netdev_get_tx_queue(vi->dev, qindex);
2481
2482 /* 1. wait all ximt complete
2483 * 2. fix the race of netif_stop_subqueue() vs netif_start_subqueue()
2484 */
2485 __netif_tx_lock_bh(txq);
2486
2487 /* Prevent rx poll from accessing sq. */
2488 sq->reset = true;
2489
2490 /* Prevent the upper layer from trying to send packets. */
2491 netif_stop_subqueue(vi->dev, qindex);
2492
2493 __netif_tx_unlock_bh(txq);
2494
2495 err = virtqueue_resize(sq->vq, ring_num, virtnet_sq_free_unused_buf);
2496 if (err)
2497 netdev_err(vi->dev, "resize tx fail: tx queue index: %d err: %d\n", qindex, err);
2498
2499 __netif_tx_lock_bh(txq);
2500 sq->reset = false;
2501 netif_tx_wake_queue(txq);
2502 __netif_tx_unlock_bh(txq);
2503
2504 if (running)
2505 virtnet_napi_tx_enable(vi, sq->vq, &sq->napi);
2506 return err;
2507}
2508
2509/*
2510 * Send command via the control virtqueue and check status. Commands
2511 * supported by the hypervisor, as indicated by feature bits, should
2512 * never fail unless improperly formatted.
2513 */
2514static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
2515 struct scatterlist *out)
2516{
2517 struct scatterlist *sgs[4], hdr, stat;
2518 unsigned out_num = 0, tmp;
2519 int ret;
2520
2521 /* Caller should know better */
2522 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
2523
2524 vi->ctrl->status = ~0;
2525 vi->ctrl->hdr.class = class;
2526 vi->ctrl->hdr.cmd = cmd;
2527 /* Add header */
2528 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
2529 sgs[out_num++] = &hdr;
2530
2531 if (out)
2532 sgs[out_num++] = out;
2533
2534 /* Add return status. */
2535 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
2536 sgs[out_num] = &stat;
2537
2538 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
2539 ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
2540 if (ret < 0) {
2541 dev_warn(&vi->vdev->dev,
2542 "Failed to add sgs for command vq: %d\n.", ret);
2543 return false;
2544 }
2545
2546 if (unlikely(!virtqueue_kick(vi->cvq)))
2547 return vi->ctrl->status == VIRTIO_NET_OK;
2548
2549 /* Spin for a response, the kick causes an ioport write, trapping
2550 * into the hypervisor, so the request should be handled immediately.
2551 */
2552 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
2553 !virtqueue_is_broken(vi->cvq))
2554 cpu_relax();
2555
2556 return vi->ctrl->status == VIRTIO_NET_OK;
2557}
2558
2559static int virtnet_set_mac_address(struct net_device *dev, void *p)
2560{
2561 struct virtnet_info *vi = netdev_priv(dev);
2562 struct virtio_device *vdev = vi->vdev;
2563 int ret;
2564 struct sockaddr *addr;
2565 struct scatterlist sg;
2566
2567 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2568 return -EOPNOTSUPP;
2569
2570 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
2571 if (!addr)
2572 return -ENOMEM;
2573
2574 ret = eth_prepare_mac_addr_change(dev, addr);
2575 if (ret)
2576 goto out;
2577
2578 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
2579 sg_init_one(&sg, addr->sa_data, dev->addr_len);
2580 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2581 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
2582 dev_warn(&vdev->dev,
2583 "Failed to set mac address by vq command.\n");
2584 ret = -EINVAL;
2585 goto out;
2586 }
2587 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
2588 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
2589 unsigned int i;
2590
2591 /* Naturally, this has an atomicity problem. */
2592 for (i = 0; i < dev->addr_len; i++)
2593 virtio_cwrite8(vdev,
2594 offsetof(struct virtio_net_config, mac) +
2595 i, addr->sa_data[i]);
2596 }
2597
2598 eth_commit_mac_addr_change(dev, p);
2599 ret = 0;
2600
2601out:
2602 kfree(addr);
2603 return ret;
2604}
2605
2606static void virtnet_stats(struct net_device *dev,
2607 struct rtnl_link_stats64 *tot)
2608{
2609 struct virtnet_info *vi = netdev_priv(dev);
2610 unsigned int start;
2611 int i;
2612
2613 for (i = 0; i < vi->max_queue_pairs; i++) {
2614 u64 tpackets, tbytes, terrors, rpackets, rbytes, rdrops;
2615 struct receive_queue *rq = &vi->rq[i];
2616 struct send_queue *sq = &vi->sq[i];
2617
2618 do {
2619 start = u64_stats_fetch_begin(&sq->stats.syncp);
2620 tpackets = u64_stats_read(&sq->stats.packets);
2621 tbytes = u64_stats_read(&sq->stats.bytes);
2622 terrors = u64_stats_read(&sq->stats.tx_timeouts);
2623 } while (u64_stats_fetch_retry(&sq->stats.syncp, start));
2624
2625 do {
2626 start = u64_stats_fetch_begin(&rq->stats.syncp);
2627 rpackets = u64_stats_read(&rq->stats.packets);
2628 rbytes = u64_stats_read(&rq->stats.bytes);
2629 rdrops = u64_stats_read(&rq->stats.drops);
2630 } while (u64_stats_fetch_retry(&rq->stats.syncp, start));
2631
2632 tot->rx_packets += rpackets;
2633 tot->tx_packets += tpackets;
2634 tot->rx_bytes += rbytes;
2635 tot->tx_bytes += tbytes;
2636 tot->rx_dropped += rdrops;
2637 tot->tx_errors += terrors;
2638 }
2639
2640 tot->tx_dropped = DEV_STATS_READ(dev, tx_dropped);
2641 tot->tx_fifo_errors = DEV_STATS_READ(dev, tx_fifo_errors);
2642 tot->rx_length_errors = DEV_STATS_READ(dev, rx_length_errors);
2643 tot->rx_frame_errors = DEV_STATS_READ(dev, rx_frame_errors);
2644}
2645
2646static void virtnet_ack_link_announce(struct virtnet_info *vi)
2647{
2648 rtnl_lock();
2649 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
2650 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
2651 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
2652 rtnl_unlock();
2653}
2654
2655static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
2656{
2657 struct scatterlist sg;
2658 struct net_device *dev = vi->dev;
2659
2660 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
2661 return 0;
2662
2663 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
2664 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
2665
2666 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
2667 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
2668 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
2669 queue_pairs);
2670 return -EINVAL;
2671 } else {
2672 vi->curr_queue_pairs = queue_pairs;
2673 /* virtnet_open() will refill when device is going to up. */
2674 if (dev->flags & IFF_UP)
2675 schedule_delayed_work(&vi->refill, 0);
2676 }
2677
2678 return 0;
2679}
2680
2681static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
2682{
2683 int err;
2684
2685 rtnl_lock();
2686 err = _virtnet_set_queues(vi, queue_pairs);
2687 rtnl_unlock();
2688 return err;
2689}
2690
2691static int virtnet_close(struct net_device *dev)
2692{
2693 struct virtnet_info *vi = netdev_priv(dev);
2694 int i;
2695
2696 /* Make sure NAPI doesn't schedule refill work */
2697 disable_delayed_refill(vi);
2698 /* Make sure refill_work doesn't re-enable napi! */
2699 cancel_delayed_work_sync(&vi->refill);
2700
2701 for (i = 0; i < vi->max_queue_pairs; i++) {
2702 virtnet_disable_queue_pair(vi, i);
2703 cancel_work_sync(&vi->rq[i].dim.work);
2704 }
2705
2706 return 0;
2707}
2708
2709static void virtnet_set_rx_mode(struct net_device *dev)
2710{
2711 struct virtnet_info *vi = netdev_priv(dev);
2712 struct scatterlist sg[2];
2713 struct virtio_net_ctrl_mac *mac_data;
2714 struct netdev_hw_addr *ha;
2715 int uc_count;
2716 int mc_count;
2717 void *buf;
2718 int i;
2719
2720 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
2721 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
2722 return;
2723
2724 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
2725 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
2726
2727 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
2728
2729 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2730 VIRTIO_NET_CTRL_RX_PROMISC, sg))
2731 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
2732 vi->ctrl->promisc ? "en" : "dis");
2733
2734 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
2735
2736 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2737 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
2738 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
2739 vi->ctrl->allmulti ? "en" : "dis");
2740
2741 uc_count = netdev_uc_count(dev);
2742 mc_count = netdev_mc_count(dev);
2743 /* MAC filter - use one buffer for both lists */
2744 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
2745 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
2746 mac_data = buf;
2747 if (!buf)
2748 return;
2749
2750 sg_init_table(sg, 2);
2751
2752 /* Store the unicast list and count in the front of the buffer */
2753 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
2754 i = 0;
2755 netdev_for_each_uc_addr(ha, dev)
2756 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2757
2758 sg_set_buf(&sg[0], mac_data,
2759 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
2760
2761 /* multicast list and count fill the end */
2762 mac_data = (void *)&mac_data->macs[uc_count][0];
2763
2764 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
2765 i = 0;
2766 netdev_for_each_mc_addr(ha, dev)
2767 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2768
2769 sg_set_buf(&sg[1], mac_data,
2770 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
2771
2772 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2773 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
2774 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
2775
2776 kfree(buf);
2777}
2778
2779static int virtnet_vlan_rx_add_vid(struct net_device *dev,
2780 __be16 proto, u16 vid)
2781{
2782 struct virtnet_info *vi = netdev_priv(dev);
2783 struct scatterlist sg;
2784
2785 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2786 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2787
2788 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2789 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2790 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2791 return 0;
2792}
2793
2794static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2795 __be16 proto, u16 vid)
2796{
2797 struct virtnet_info *vi = netdev_priv(dev);
2798 struct scatterlist sg;
2799
2800 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2801 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2802
2803 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2804 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2805 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2806 return 0;
2807}
2808
2809static void virtnet_clean_affinity(struct virtnet_info *vi)
2810{
2811 int i;
2812
2813 if (vi->affinity_hint_set) {
2814 for (i = 0; i < vi->max_queue_pairs; i++) {
2815 virtqueue_set_affinity(vi->rq[i].vq, NULL);
2816 virtqueue_set_affinity(vi->sq[i].vq, NULL);
2817 }
2818
2819 vi->affinity_hint_set = false;
2820 }
2821}
2822
2823static void virtnet_set_affinity(struct virtnet_info *vi)
2824{
2825 cpumask_var_t mask;
2826 int stragglers;
2827 int group_size;
2828 int i, j, cpu;
2829 int num_cpu;
2830 int stride;
2831
2832 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2833 virtnet_clean_affinity(vi);
2834 return;
2835 }
2836
2837 num_cpu = num_online_cpus();
2838 stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2839 stragglers = num_cpu >= vi->curr_queue_pairs ?
2840 num_cpu % vi->curr_queue_pairs :
2841 0;
2842 cpu = cpumask_first(cpu_online_mask);
2843
2844 for (i = 0; i < vi->curr_queue_pairs; i++) {
2845 group_size = stride + (i < stragglers ? 1 : 0);
2846
2847 for (j = 0; j < group_size; j++) {
2848 cpumask_set_cpu(cpu, mask);
2849 cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2850 nr_cpu_ids, false);
2851 }
2852 virtqueue_set_affinity(vi->rq[i].vq, mask);
2853 virtqueue_set_affinity(vi->sq[i].vq, mask);
2854 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2855 cpumask_clear(mask);
2856 }
2857
2858 vi->affinity_hint_set = true;
2859 free_cpumask_var(mask);
2860}
2861
2862static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2863{
2864 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2865 node);
2866 virtnet_set_affinity(vi);
2867 return 0;
2868}
2869
2870static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2871{
2872 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2873 node_dead);
2874 virtnet_set_affinity(vi);
2875 return 0;
2876}
2877
2878static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2879{
2880 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2881 node);
2882
2883 virtnet_clean_affinity(vi);
2884 return 0;
2885}
2886
2887static enum cpuhp_state virtionet_online;
2888
2889static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2890{
2891 int ret;
2892
2893 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2894 if (ret)
2895 return ret;
2896 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2897 &vi->node_dead);
2898 if (!ret)
2899 return ret;
2900 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2901 return ret;
2902}
2903
2904static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2905{
2906 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2907 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2908 &vi->node_dead);
2909}
2910
2911static int virtnet_send_ctrl_coal_vq_cmd(struct virtnet_info *vi,
2912 u16 vqn, u32 max_usecs, u32 max_packets)
2913{
2914 struct scatterlist sgs;
2915
2916 vi->ctrl->coal_vq.vqn = cpu_to_le16(vqn);
2917 vi->ctrl->coal_vq.coal.max_usecs = cpu_to_le32(max_usecs);
2918 vi->ctrl->coal_vq.coal.max_packets = cpu_to_le32(max_packets);
2919 sg_init_one(&sgs, &vi->ctrl->coal_vq, sizeof(vi->ctrl->coal_vq));
2920
2921 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
2922 VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET,
2923 &sgs))
2924 return -EINVAL;
2925
2926 return 0;
2927}
2928
2929static int virtnet_send_rx_ctrl_coal_vq_cmd(struct virtnet_info *vi,
2930 u16 queue, u32 max_usecs,
2931 u32 max_packets)
2932{
2933 int err;
2934
2935 err = virtnet_send_ctrl_coal_vq_cmd(vi, rxq2vq(queue),
2936 max_usecs, max_packets);
2937 if (err)
2938 return err;
2939
2940 vi->rq[queue].intr_coal.max_usecs = max_usecs;
2941 vi->rq[queue].intr_coal.max_packets = max_packets;
2942
2943 return 0;
2944}
2945
2946static int virtnet_send_tx_ctrl_coal_vq_cmd(struct virtnet_info *vi,
2947 u16 queue, u32 max_usecs,
2948 u32 max_packets)
2949{
2950 int err;
2951
2952 err = virtnet_send_ctrl_coal_vq_cmd(vi, txq2vq(queue),
2953 max_usecs, max_packets);
2954 if (err)
2955 return err;
2956
2957 vi->sq[queue].intr_coal.max_usecs = max_usecs;
2958 vi->sq[queue].intr_coal.max_packets = max_packets;
2959
2960 return 0;
2961}
2962
2963static void virtnet_get_ringparam(struct net_device *dev,
2964 struct ethtool_ringparam *ring,
2965 struct kernel_ethtool_ringparam *kernel_ring,
2966 struct netlink_ext_ack *extack)
2967{
2968 struct virtnet_info *vi = netdev_priv(dev);
2969
2970 ring->rx_max_pending = vi->rq[0].vq->num_max;
2971 ring->tx_max_pending = vi->sq[0].vq->num_max;
2972 ring->rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2973 ring->tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2974}
2975
2976static int virtnet_set_ringparam(struct net_device *dev,
2977 struct ethtool_ringparam *ring,
2978 struct kernel_ethtool_ringparam *kernel_ring,
2979 struct netlink_ext_ack *extack)
2980{
2981 struct virtnet_info *vi = netdev_priv(dev);
2982 u32 rx_pending, tx_pending;
2983 struct receive_queue *rq;
2984 struct send_queue *sq;
2985 int i, err;
2986
2987 if (ring->rx_mini_pending || ring->rx_jumbo_pending)
2988 return -EINVAL;
2989
2990 rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2991 tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2992
2993 if (ring->rx_pending == rx_pending &&
2994 ring->tx_pending == tx_pending)
2995 return 0;
2996
2997 if (ring->rx_pending > vi->rq[0].vq->num_max)
2998 return -EINVAL;
2999
3000 if (ring->tx_pending > vi->sq[0].vq->num_max)
3001 return -EINVAL;
3002
3003 for (i = 0; i < vi->max_queue_pairs; i++) {
3004 rq = vi->rq + i;
3005 sq = vi->sq + i;
3006
3007 if (ring->tx_pending != tx_pending) {
3008 err = virtnet_tx_resize(vi, sq, ring->tx_pending);
3009 if (err)
3010 return err;
3011
3012 /* Upon disabling and re-enabling a transmit virtqueue, the device must
3013 * set the coalescing parameters of the virtqueue to those configured
3014 * through the VIRTIO_NET_CTRL_NOTF_COAL_TX_SET command, or, if the driver
3015 * did not set any TX coalescing parameters, to 0.
3016 */
3017 err = virtnet_send_tx_ctrl_coal_vq_cmd(vi, i,
3018 vi->intr_coal_tx.max_usecs,
3019 vi->intr_coal_tx.max_packets);
3020 if (err)
3021 return err;
3022 }
3023
3024 if (ring->rx_pending != rx_pending) {
3025 err = virtnet_rx_resize(vi, rq, ring->rx_pending);
3026 if (err)
3027 return err;
3028
3029 /* The reason is same as the transmit virtqueue reset */
3030 err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, i,
3031 vi->intr_coal_rx.max_usecs,
3032 vi->intr_coal_rx.max_packets);
3033 if (err)
3034 return err;
3035 }
3036 }
3037
3038 return 0;
3039}
3040
3041static bool virtnet_commit_rss_command(struct virtnet_info *vi)
3042{
3043 struct net_device *dev = vi->dev;
3044 struct scatterlist sgs[4];
3045 unsigned int sg_buf_size;
3046
3047 /* prepare sgs */
3048 sg_init_table(sgs, 4);
3049
3050 sg_buf_size = offsetof(struct virtio_net_ctrl_rss, indirection_table);
3051 sg_set_buf(&sgs[0], &vi->ctrl->rss, sg_buf_size);
3052
3053 sg_buf_size = sizeof(uint16_t) * (vi->ctrl->rss.indirection_table_mask + 1);
3054 sg_set_buf(&sgs[1], vi->ctrl->rss.indirection_table, sg_buf_size);
3055
3056 sg_buf_size = offsetof(struct virtio_net_ctrl_rss, key)
3057 - offsetof(struct virtio_net_ctrl_rss, max_tx_vq);
3058 sg_set_buf(&sgs[2], &vi->ctrl->rss.max_tx_vq, sg_buf_size);
3059
3060 sg_buf_size = vi->rss_key_size;
3061 sg_set_buf(&sgs[3], vi->ctrl->rss.key, sg_buf_size);
3062
3063 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
3064 vi->has_rss ? VIRTIO_NET_CTRL_MQ_RSS_CONFIG
3065 : VIRTIO_NET_CTRL_MQ_HASH_CONFIG, sgs)) {
3066 dev_warn(&dev->dev, "VIRTIONET issue with committing RSS sgs\n");
3067 return false;
3068 }
3069 return true;
3070}
3071
3072static void virtnet_init_default_rss(struct virtnet_info *vi)
3073{
3074 u32 indir_val = 0;
3075 int i = 0;
3076
3077 vi->ctrl->rss.hash_types = vi->rss_hash_types_supported;
3078 vi->rss_hash_types_saved = vi->rss_hash_types_supported;
3079 vi->ctrl->rss.indirection_table_mask = vi->rss_indir_table_size
3080 ? vi->rss_indir_table_size - 1 : 0;
3081 vi->ctrl->rss.unclassified_queue = 0;
3082
3083 for (; i < vi->rss_indir_table_size; ++i) {
3084 indir_val = ethtool_rxfh_indir_default(i, vi->curr_queue_pairs);
3085 vi->ctrl->rss.indirection_table[i] = indir_val;
3086 }
3087
3088 vi->ctrl->rss.max_tx_vq = vi->has_rss ? vi->curr_queue_pairs : 0;
3089 vi->ctrl->rss.hash_key_length = vi->rss_key_size;
3090
3091 netdev_rss_key_fill(vi->ctrl->rss.key, vi->rss_key_size);
3092}
3093
3094static void virtnet_get_hashflow(const struct virtnet_info *vi, struct ethtool_rxnfc *info)
3095{
3096 info->data = 0;
3097 switch (info->flow_type) {
3098 case TCP_V4_FLOW:
3099 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv4) {
3100 info->data = RXH_IP_SRC | RXH_IP_DST |
3101 RXH_L4_B_0_1 | RXH_L4_B_2_3;
3102 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
3103 info->data = RXH_IP_SRC | RXH_IP_DST;
3104 }
3105 break;
3106 case TCP_V6_FLOW:
3107 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv6) {
3108 info->data = RXH_IP_SRC | RXH_IP_DST |
3109 RXH_L4_B_0_1 | RXH_L4_B_2_3;
3110 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
3111 info->data = RXH_IP_SRC | RXH_IP_DST;
3112 }
3113 break;
3114 case UDP_V4_FLOW:
3115 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv4) {
3116 info->data = RXH_IP_SRC | RXH_IP_DST |
3117 RXH_L4_B_0_1 | RXH_L4_B_2_3;
3118 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
3119 info->data = RXH_IP_SRC | RXH_IP_DST;
3120 }
3121 break;
3122 case UDP_V6_FLOW:
3123 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv6) {
3124 info->data = RXH_IP_SRC | RXH_IP_DST |
3125 RXH_L4_B_0_1 | RXH_L4_B_2_3;
3126 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
3127 info->data = RXH_IP_SRC | RXH_IP_DST;
3128 }
3129 break;
3130 case IPV4_FLOW:
3131 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4)
3132 info->data = RXH_IP_SRC | RXH_IP_DST;
3133
3134 break;
3135 case IPV6_FLOW:
3136 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6)
3137 info->data = RXH_IP_SRC | RXH_IP_DST;
3138
3139 break;
3140 default:
3141 info->data = 0;
3142 break;
3143 }
3144}
3145
3146static bool virtnet_set_hashflow(struct virtnet_info *vi, struct ethtool_rxnfc *info)
3147{
3148 u32 new_hashtypes = vi->rss_hash_types_saved;
3149 bool is_disable = info->data & RXH_DISCARD;
3150 bool is_l4 = info->data == (RXH_IP_SRC | RXH_IP_DST | RXH_L4_B_0_1 | RXH_L4_B_2_3);
3151
3152 /* supports only 'sd', 'sdfn' and 'r' */
3153 if (!((info->data == (RXH_IP_SRC | RXH_IP_DST)) | is_l4 | is_disable))
3154 return false;
3155
3156 switch (info->flow_type) {
3157 case TCP_V4_FLOW:
3158 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_TCPv4);
3159 if (!is_disable)
3160 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
3161 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv4 : 0);
3162 break;
3163 case UDP_V4_FLOW:
3164 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_UDPv4);
3165 if (!is_disable)
3166 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
3167 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv4 : 0);
3168 break;
3169 case IPV4_FLOW:
3170 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv4;
3171 if (!is_disable)
3172 new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv4;
3173 break;
3174 case TCP_V6_FLOW:
3175 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_TCPv6);
3176 if (!is_disable)
3177 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
3178 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv6 : 0);
3179 break;
3180 case UDP_V6_FLOW:
3181 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_UDPv6);
3182 if (!is_disable)
3183 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
3184 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv6 : 0);
3185 break;
3186 case IPV6_FLOW:
3187 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv6;
3188 if (!is_disable)
3189 new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv6;
3190 break;
3191 default:
3192 /* unsupported flow */
3193 return false;
3194 }
3195
3196 /* if unsupported hashtype was set */
3197 if (new_hashtypes != (new_hashtypes & vi->rss_hash_types_supported))
3198 return false;
3199
3200 if (new_hashtypes != vi->rss_hash_types_saved) {
3201 vi->rss_hash_types_saved = new_hashtypes;
3202 vi->ctrl->rss.hash_types = vi->rss_hash_types_saved;
3203 if (vi->dev->features & NETIF_F_RXHASH)
3204 return virtnet_commit_rss_command(vi);
3205 }
3206
3207 return true;
3208}
3209
3210static void virtnet_get_drvinfo(struct net_device *dev,
3211 struct ethtool_drvinfo *info)
3212{
3213 struct virtnet_info *vi = netdev_priv(dev);
3214 struct virtio_device *vdev = vi->vdev;
3215
3216 strscpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
3217 strscpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
3218 strscpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
3219
3220}
3221
3222/* TODO: Eliminate OOO packets during switching */
3223static int virtnet_set_channels(struct net_device *dev,
3224 struct ethtool_channels *channels)
3225{
3226 struct virtnet_info *vi = netdev_priv(dev);
3227 u16 queue_pairs = channels->combined_count;
3228 int err;
3229
3230 /* We don't support separate rx/tx channels.
3231 * We don't allow setting 'other' channels.
3232 */
3233 if (channels->rx_count || channels->tx_count || channels->other_count)
3234 return -EINVAL;
3235
3236 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
3237 return -EINVAL;
3238
3239 /* For now we don't support modifying channels while XDP is loaded
3240 * also when XDP is loaded all RX queues have XDP programs so we only
3241 * need to check a single RX queue.
3242 */
3243 if (vi->rq[0].xdp_prog)
3244 return -EINVAL;
3245
3246 cpus_read_lock();
3247 err = _virtnet_set_queues(vi, queue_pairs);
3248 if (err) {
3249 cpus_read_unlock();
3250 goto err;
3251 }
3252 virtnet_set_affinity(vi);
3253 cpus_read_unlock();
3254
3255 netif_set_real_num_tx_queues(dev, queue_pairs);
3256 netif_set_real_num_rx_queues(dev, queue_pairs);
3257 err:
3258 return err;
3259}
3260
3261static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
3262{
3263 struct virtnet_info *vi = netdev_priv(dev);
3264 unsigned int i, j;
3265 u8 *p = data;
3266
3267 switch (stringset) {
3268 case ETH_SS_STATS:
3269 for (i = 0; i < vi->curr_queue_pairs; i++) {
3270 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
3271 ethtool_sprintf(&p, "rx_queue_%u_%s", i,
3272 virtnet_rq_stats_desc[j].desc);
3273 }
3274
3275 for (i = 0; i < vi->curr_queue_pairs; i++) {
3276 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
3277 ethtool_sprintf(&p, "tx_queue_%u_%s", i,
3278 virtnet_sq_stats_desc[j].desc);
3279 }
3280 break;
3281 }
3282}
3283
3284static int virtnet_get_sset_count(struct net_device *dev, int sset)
3285{
3286 struct virtnet_info *vi = netdev_priv(dev);
3287
3288 switch (sset) {
3289 case ETH_SS_STATS:
3290 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
3291 VIRTNET_SQ_STATS_LEN);
3292 default:
3293 return -EOPNOTSUPP;
3294 }
3295}
3296
3297static void virtnet_get_ethtool_stats(struct net_device *dev,
3298 struct ethtool_stats *stats, u64 *data)
3299{
3300 struct virtnet_info *vi = netdev_priv(dev);
3301 unsigned int idx = 0, start, i, j;
3302 const u8 *stats_base;
3303 const u64_stats_t *p;
3304 size_t offset;
3305
3306 for (i = 0; i < vi->curr_queue_pairs; i++) {
3307 struct receive_queue *rq = &vi->rq[i];
3308
3309 stats_base = (const u8 *)&rq->stats;
3310 do {
3311 start = u64_stats_fetch_begin(&rq->stats.syncp);
3312 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
3313 offset = virtnet_rq_stats_desc[j].offset;
3314 p = (const u64_stats_t *)(stats_base + offset);
3315 data[idx + j] = u64_stats_read(p);
3316 }
3317 } while (u64_stats_fetch_retry(&rq->stats.syncp, start));
3318 idx += VIRTNET_RQ_STATS_LEN;
3319 }
3320
3321 for (i = 0; i < vi->curr_queue_pairs; i++) {
3322 struct send_queue *sq = &vi->sq[i];
3323
3324 stats_base = (const u8 *)&sq->stats;
3325 do {
3326 start = u64_stats_fetch_begin(&sq->stats.syncp);
3327 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
3328 offset = virtnet_sq_stats_desc[j].offset;
3329 p = (const u64_stats_t *)(stats_base + offset);
3330 data[idx + j] = u64_stats_read(p);
3331 }
3332 } while (u64_stats_fetch_retry(&sq->stats.syncp, start));
3333 idx += VIRTNET_SQ_STATS_LEN;
3334 }
3335}
3336
3337static void virtnet_get_channels(struct net_device *dev,
3338 struct ethtool_channels *channels)
3339{
3340 struct virtnet_info *vi = netdev_priv(dev);
3341
3342 channels->combined_count = vi->curr_queue_pairs;
3343 channels->max_combined = vi->max_queue_pairs;
3344 channels->max_other = 0;
3345 channels->rx_count = 0;
3346 channels->tx_count = 0;
3347 channels->other_count = 0;
3348}
3349
3350static int virtnet_set_link_ksettings(struct net_device *dev,
3351 const struct ethtool_link_ksettings *cmd)
3352{
3353 struct virtnet_info *vi = netdev_priv(dev);
3354
3355 return ethtool_virtdev_set_link_ksettings(dev, cmd,
3356 &vi->speed, &vi->duplex);
3357}
3358
3359static int virtnet_get_link_ksettings(struct net_device *dev,
3360 struct ethtool_link_ksettings *cmd)
3361{
3362 struct virtnet_info *vi = netdev_priv(dev);
3363
3364 cmd->base.speed = vi->speed;
3365 cmd->base.duplex = vi->duplex;
3366 cmd->base.port = PORT_OTHER;
3367
3368 return 0;
3369}
3370
3371static int virtnet_send_tx_notf_coal_cmds(struct virtnet_info *vi,
3372 struct ethtool_coalesce *ec)
3373{
3374 struct scatterlist sgs_tx;
3375 int i;
3376
3377 vi->ctrl->coal_tx.tx_usecs = cpu_to_le32(ec->tx_coalesce_usecs);
3378 vi->ctrl->coal_tx.tx_max_packets = cpu_to_le32(ec->tx_max_coalesced_frames);
3379 sg_init_one(&sgs_tx, &vi->ctrl->coal_tx, sizeof(vi->ctrl->coal_tx));
3380
3381 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
3382 VIRTIO_NET_CTRL_NOTF_COAL_TX_SET,
3383 &sgs_tx))
3384 return -EINVAL;
3385
3386 vi->intr_coal_tx.max_usecs = ec->tx_coalesce_usecs;
3387 vi->intr_coal_tx.max_packets = ec->tx_max_coalesced_frames;
3388 for (i = 0; i < vi->max_queue_pairs; i++) {
3389 vi->sq[i].intr_coal.max_usecs = ec->tx_coalesce_usecs;
3390 vi->sq[i].intr_coal.max_packets = ec->tx_max_coalesced_frames;
3391 }
3392
3393 return 0;
3394}
3395
3396static int virtnet_send_rx_notf_coal_cmds(struct virtnet_info *vi,
3397 struct ethtool_coalesce *ec)
3398{
3399 bool rx_ctrl_dim_on = !!ec->use_adaptive_rx_coalesce;
3400 struct scatterlist sgs_rx;
3401 int i;
3402
3403 if (rx_ctrl_dim_on && !virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL))
3404 return -EOPNOTSUPP;
3405
3406 if (rx_ctrl_dim_on && (ec->rx_coalesce_usecs != vi->intr_coal_rx.max_usecs ||
3407 ec->rx_max_coalesced_frames != vi->intr_coal_rx.max_packets))
3408 return -EINVAL;
3409
3410 if (rx_ctrl_dim_on && !vi->rx_dim_enabled) {
3411 vi->rx_dim_enabled = true;
3412 for (i = 0; i < vi->max_queue_pairs; i++)
3413 vi->rq[i].dim_enabled = true;
3414 return 0;
3415 }
3416
3417 if (!rx_ctrl_dim_on && vi->rx_dim_enabled) {
3418 vi->rx_dim_enabled = false;
3419 for (i = 0; i < vi->max_queue_pairs; i++)
3420 vi->rq[i].dim_enabled = false;
3421 }
3422
3423 /* Since the per-queue coalescing params can be set,
3424 * we need apply the global new params even if they
3425 * are not updated.
3426 */
3427 vi->ctrl->coal_rx.rx_usecs = cpu_to_le32(ec->rx_coalesce_usecs);
3428 vi->ctrl->coal_rx.rx_max_packets = cpu_to_le32(ec->rx_max_coalesced_frames);
3429 sg_init_one(&sgs_rx, &vi->ctrl->coal_rx, sizeof(vi->ctrl->coal_rx));
3430
3431 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
3432 VIRTIO_NET_CTRL_NOTF_COAL_RX_SET,
3433 &sgs_rx))
3434 return -EINVAL;
3435
3436 vi->intr_coal_rx.max_usecs = ec->rx_coalesce_usecs;
3437 vi->intr_coal_rx.max_packets = ec->rx_max_coalesced_frames;
3438 for (i = 0; i < vi->max_queue_pairs; i++) {
3439 vi->rq[i].intr_coal.max_usecs = ec->rx_coalesce_usecs;
3440 vi->rq[i].intr_coal.max_packets = ec->rx_max_coalesced_frames;
3441 }
3442
3443 return 0;
3444}
3445
3446static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
3447 struct ethtool_coalesce *ec)
3448{
3449 int err;
3450
3451 err = virtnet_send_tx_notf_coal_cmds(vi, ec);
3452 if (err)
3453 return err;
3454
3455 err = virtnet_send_rx_notf_coal_cmds(vi, ec);
3456 if (err)
3457 return err;
3458
3459 return 0;
3460}
3461
3462static int virtnet_send_rx_notf_coal_vq_cmds(struct virtnet_info *vi,
3463 struct ethtool_coalesce *ec,
3464 u16 queue)
3465{
3466 bool rx_ctrl_dim_on = !!ec->use_adaptive_rx_coalesce;
3467 bool cur_rx_dim = vi->rq[queue].dim_enabled;
3468 u32 max_usecs, max_packets;
3469 int err;
3470
3471 max_usecs = vi->rq[queue].intr_coal.max_usecs;
3472 max_packets = vi->rq[queue].intr_coal.max_packets;
3473
3474 if (rx_ctrl_dim_on && (ec->rx_coalesce_usecs != max_usecs ||
3475 ec->rx_max_coalesced_frames != max_packets))
3476 return -EINVAL;
3477
3478 if (rx_ctrl_dim_on && !cur_rx_dim) {
3479 vi->rq[queue].dim_enabled = true;
3480 return 0;
3481 }
3482
3483 if (!rx_ctrl_dim_on && cur_rx_dim)
3484 vi->rq[queue].dim_enabled = false;
3485
3486 /* If no params are updated, userspace ethtool will
3487 * reject the modification.
3488 */
3489 err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, queue,
3490 ec->rx_coalesce_usecs,
3491 ec->rx_max_coalesced_frames);
3492 if (err)
3493 return err;
3494
3495 return 0;
3496}
3497
3498static int virtnet_send_notf_coal_vq_cmds(struct virtnet_info *vi,
3499 struct ethtool_coalesce *ec,
3500 u16 queue)
3501{
3502 int err;
3503
3504 err = virtnet_send_rx_notf_coal_vq_cmds(vi, ec, queue);
3505 if (err)
3506 return err;
3507
3508 err = virtnet_send_tx_ctrl_coal_vq_cmd(vi, queue,
3509 ec->tx_coalesce_usecs,
3510 ec->tx_max_coalesced_frames);
3511 if (err)
3512 return err;
3513
3514 return 0;
3515}
3516
3517static void virtnet_rx_dim_work(struct work_struct *work)
3518{
3519 struct dim *dim = container_of(work, struct dim, work);
3520 struct receive_queue *rq = container_of(dim,
3521 struct receive_queue, dim);
3522 struct virtnet_info *vi = rq->vq->vdev->priv;
3523 struct net_device *dev = vi->dev;
3524 struct dim_cq_moder update_moder;
3525 int i, qnum, err;
3526
3527 if (!rtnl_trylock())
3528 return;
3529
3530 /* Each rxq's work is queued by "net_dim()->schedule_work()"
3531 * in response to NAPI traffic changes. Note that dim->profile_ix
3532 * for each rxq is updated prior to the queuing action.
3533 * So we only need to traverse and update profiles for all rxqs
3534 * in the work which is holding rtnl_lock.
3535 */
3536 for (i = 0; i < vi->curr_queue_pairs; i++) {
3537 rq = &vi->rq[i];
3538 dim = &rq->dim;
3539 qnum = rq - vi->rq;
3540
3541 if (!rq->dim_enabled)
3542 continue;
3543
3544 update_moder = net_dim_get_rx_moderation(dim->mode, dim->profile_ix);
3545 if (update_moder.usec != rq->intr_coal.max_usecs ||
3546 update_moder.pkts != rq->intr_coal.max_packets) {
3547 err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, qnum,
3548 update_moder.usec,
3549 update_moder.pkts);
3550 if (err)
3551 pr_debug("%s: Failed to send dim parameters on rxq%d\n",
3552 dev->name, qnum);
3553 dim->state = DIM_START_MEASURE;
3554 }
3555 }
3556
3557 rtnl_unlock();
3558}
3559
3560static int virtnet_coal_params_supported(struct ethtool_coalesce *ec)
3561{
3562 /* usecs coalescing is supported only if VIRTIO_NET_F_NOTF_COAL
3563 * or VIRTIO_NET_F_VQ_NOTF_COAL feature is negotiated.
3564 */
3565 if (ec->rx_coalesce_usecs || ec->tx_coalesce_usecs)
3566 return -EOPNOTSUPP;
3567
3568 if (ec->tx_max_coalesced_frames > 1 ||
3569 ec->rx_max_coalesced_frames != 1)
3570 return -EINVAL;
3571
3572 return 0;
3573}
3574
3575static int virtnet_should_update_vq_weight(int dev_flags, int weight,
3576 int vq_weight, bool *should_update)
3577{
3578 if (weight ^ vq_weight) {
3579 if (dev_flags & IFF_UP)
3580 return -EBUSY;
3581 *should_update = true;
3582 }
3583
3584 return 0;
3585}
3586
3587static int virtnet_set_coalesce(struct net_device *dev,
3588 struct ethtool_coalesce *ec,
3589 struct kernel_ethtool_coalesce *kernel_coal,
3590 struct netlink_ext_ack *extack)
3591{
3592 struct virtnet_info *vi = netdev_priv(dev);
3593 int ret, queue_number, napi_weight;
3594 bool update_napi = false;
3595
3596 /* Can't change NAPI weight if the link is up */
3597 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
3598 for (queue_number = 0; queue_number < vi->max_queue_pairs; queue_number++) {
3599 ret = virtnet_should_update_vq_weight(dev->flags, napi_weight,
3600 vi->sq[queue_number].napi.weight,
3601 &update_napi);
3602 if (ret)
3603 return ret;
3604
3605 if (update_napi) {
3606 /* All queues that belong to [queue_number, vi->max_queue_pairs] will be
3607 * updated for the sake of simplicity, which might not be necessary
3608 */
3609 break;
3610 }
3611 }
3612
3613 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL))
3614 ret = virtnet_send_notf_coal_cmds(vi, ec);
3615 else
3616 ret = virtnet_coal_params_supported(ec);
3617
3618 if (ret)
3619 return ret;
3620
3621 if (update_napi) {
3622 for (; queue_number < vi->max_queue_pairs; queue_number++)
3623 vi->sq[queue_number].napi.weight = napi_weight;
3624 }
3625
3626 return ret;
3627}
3628
3629static int virtnet_get_coalesce(struct net_device *dev,
3630 struct ethtool_coalesce *ec,
3631 struct kernel_ethtool_coalesce *kernel_coal,
3632 struct netlink_ext_ack *extack)
3633{
3634 struct virtnet_info *vi = netdev_priv(dev);
3635
3636 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
3637 ec->rx_coalesce_usecs = vi->intr_coal_rx.max_usecs;
3638 ec->tx_coalesce_usecs = vi->intr_coal_tx.max_usecs;
3639 ec->tx_max_coalesced_frames = vi->intr_coal_tx.max_packets;
3640 ec->rx_max_coalesced_frames = vi->intr_coal_rx.max_packets;
3641 ec->use_adaptive_rx_coalesce = vi->rx_dim_enabled;
3642 } else {
3643 ec->rx_max_coalesced_frames = 1;
3644
3645 if (vi->sq[0].napi.weight)
3646 ec->tx_max_coalesced_frames = 1;
3647 }
3648
3649 return 0;
3650}
3651
3652static int virtnet_set_per_queue_coalesce(struct net_device *dev,
3653 u32 queue,
3654 struct ethtool_coalesce *ec)
3655{
3656 struct virtnet_info *vi = netdev_priv(dev);
3657 int ret, napi_weight;
3658 bool update_napi = false;
3659
3660 if (queue >= vi->max_queue_pairs)
3661 return -EINVAL;
3662
3663 /* Can't change NAPI weight if the link is up */
3664 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
3665 ret = virtnet_should_update_vq_weight(dev->flags, napi_weight,
3666 vi->sq[queue].napi.weight,
3667 &update_napi);
3668 if (ret)
3669 return ret;
3670
3671 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL))
3672 ret = virtnet_send_notf_coal_vq_cmds(vi, ec, queue);
3673 else
3674 ret = virtnet_coal_params_supported(ec);
3675
3676 if (ret)
3677 return ret;
3678
3679 if (update_napi)
3680 vi->sq[queue].napi.weight = napi_weight;
3681
3682 return 0;
3683}
3684
3685static int virtnet_get_per_queue_coalesce(struct net_device *dev,
3686 u32 queue,
3687 struct ethtool_coalesce *ec)
3688{
3689 struct virtnet_info *vi = netdev_priv(dev);
3690
3691 if (queue >= vi->max_queue_pairs)
3692 return -EINVAL;
3693
3694 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL)) {
3695 ec->rx_coalesce_usecs = vi->rq[queue].intr_coal.max_usecs;
3696 ec->tx_coalesce_usecs = vi->sq[queue].intr_coal.max_usecs;
3697 ec->tx_max_coalesced_frames = vi->sq[queue].intr_coal.max_packets;
3698 ec->rx_max_coalesced_frames = vi->rq[queue].intr_coal.max_packets;
3699 ec->use_adaptive_rx_coalesce = vi->rq[queue].dim_enabled;
3700 } else {
3701 ec->rx_max_coalesced_frames = 1;
3702
3703 if (vi->sq[queue].napi.weight)
3704 ec->tx_max_coalesced_frames = 1;
3705 }
3706
3707 return 0;
3708}
3709
3710static void virtnet_init_settings(struct net_device *dev)
3711{
3712 struct virtnet_info *vi = netdev_priv(dev);
3713
3714 vi->speed = SPEED_UNKNOWN;
3715 vi->duplex = DUPLEX_UNKNOWN;
3716}
3717
3718static void virtnet_update_settings(struct virtnet_info *vi)
3719{
3720 u32 speed;
3721 u8 duplex;
3722
3723 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
3724 return;
3725
3726 virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
3727
3728 if (ethtool_validate_speed(speed))
3729 vi->speed = speed;
3730
3731 virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
3732
3733 if (ethtool_validate_duplex(duplex))
3734 vi->duplex = duplex;
3735}
3736
3737static u32 virtnet_get_rxfh_key_size(struct net_device *dev)
3738{
3739 return ((struct virtnet_info *)netdev_priv(dev))->rss_key_size;
3740}
3741
3742static u32 virtnet_get_rxfh_indir_size(struct net_device *dev)
3743{
3744 return ((struct virtnet_info *)netdev_priv(dev))->rss_indir_table_size;
3745}
3746
3747static int virtnet_get_rxfh(struct net_device *dev,
3748 struct ethtool_rxfh_param *rxfh)
3749{
3750 struct virtnet_info *vi = netdev_priv(dev);
3751 int i;
3752
3753 if (rxfh->indir) {
3754 for (i = 0; i < vi->rss_indir_table_size; ++i)
3755 rxfh->indir[i] = vi->ctrl->rss.indirection_table[i];
3756 }
3757
3758 if (rxfh->key)
3759 memcpy(rxfh->key, vi->ctrl->rss.key, vi->rss_key_size);
3760
3761 rxfh->hfunc = ETH_RSS_HASH_TOP;
3762
3763 return 0;
3764}
3765
3766static int virtnet_set_rxfh(struct net_device *dev,
3767 struct ethtool_rxfh_param *rxfh,
3768 struct netlink_ext_ack *extack)
3769{
3770 struct virtnet_info *vi = netdev_priv(dev);
3771 int i;
3772
3773 if (rxfh->hfunc != ETH_RSS_HASH_NO_CHANGE &&
3774 rxfh->hfunc != ETH_RSS_HASH_TOP)
3775 return -EOPNOTSUPP;
3776
3777 if (rxfh->indir) {
3778 for (i = 0; i < vi->rss_indir_table_size; ++i)
3779 vi->ctrl->rss.indirection_table[i] = rxfh->indir[i];
3780 }
3781 if (rxfh->key)
3782 memcpy(vi->ctrl->rss.key, rxfh->key, vi->rss_key_size);
3783
3784 virtnet_commit_rss_command(vi);
3785
3786 return 0;
3787}
3788
3789static int virtnet_get_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info, u32 *rule_locs)
3790{
3791 struct virtnet_info *vi = netdev_priv(dev);
3792 int rc = 0;
3793
3794 switch (info->cmd) {
3795 case ETHTOOL_GRXRINGS:
3796 info->data = vi->curr_queue_pairs;
3797 break;
3798 case ETHTOOL_GRXFH:
3799 virtnet_get_hashflow(vi, info);
3800 break;
3801 default:
3802 rc = -EOPNOTSUPP;
3803 }
3804
3805 return rc;
3806}
3807
3808static int virtnet_set_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info)
3809{
3810 struct virtnet_info *vi = netdev_priv(dev);
3811 int rc = 0;
3812
3813 switch (info->cmd) {
3814 case ETHTOOL_SRXFH:
3815 if (!virtnet_set_hashflow(vi, info))
3816 rc = -EINVAL;
3817
3818 break;
3819 default:
3820 rc = -EOPNOTSUPP;
3821 }
3822
3823 return rc;
3824}
3825
3826static const struct ethtool_ops virtnet_ethtool_ops = {
3827 .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES |
3828 ETHTOOL_COALESCE_USECS | ETHTOOL_COALESCE_USE_ADAPTIVE_RX,
3829 .get_drvinfo = virtnet_get_drvinfo,
3830 .get_link = ethtool_op_get_link,
3831 .get_ringparam = virtnet_get_ringparam,
3832 .set_ringparam = virtnet_set_ringparam,
3833 .get_strings = virtnet_get_strings,
3834 .get_sset_count = virtnet_get_sset_count,
3835 .get_ethtool_stats = virtnet_get_ethtool_stats,
3836 .set_channels = virtnet_set_channels,
3837 .get_channels = virtnet_get_channels,
3838 .get_ts_info = ethtool_op_get_ts_info,
3839 .get_link_ksettings = virtnet_get_link_ksettings,
3840 .set_link_ksettings = virtnet_set_link_ksettings,
3841 .set_coalesce = virtnet_set_coalesce,
3842 .get_coalesce = virtnet_get_coalesce,
3843 .set_per_queue_coalesce = virtnet_set_per_queue_coalesce,
3844 .get_per_queue_coalesce = virtnet_get_per_queue_coalesce,
3845 .get_rxfh_key_size = virtnet_get_rxfh_key_size,
3846 .get_rxfh_indir_size = virtnet_get_rxfh_indir_size,
3847 .get_rxfh = virtnet_get_rxfh,
3848 .set_rxfh = virtnet_set_rxfh,
3849 .get_rxnfc = virtnet_get_rxnfc,
3850 .set_rxnfc = virtnet_set_rxnfc,
3851};
3852
3853static void virtnet_freeze_down(struct virtio_device *vdev)
3854{
3855 struct virtnet_info *vi = vdev->priv;
3856
3857 /* Make sure no work handler is accessing the device */
3858 flush_work(&vi->config_work);
3859
3860 netif_tx_lock_bh(vi->dev);
3861 netif_device_detach(vi->dev);
3862 netif_tx_unlock_bh(vi->dev);
3863 if (netif_running(vi->dev))
3864 virtnet_close(vi->dev);
3865}
3866
3867static int init_vqs(struct virtnet_info *vi);
3868
3869static int virtnet_restore_up(struct virtio_device *vdev)
3870{
3871 struct virtnet_info *vi = vdev->priv;
3872 int err;
3873
3874 err = init_vqs(vi);
3875 if (err)
3876 return err;
3877
3878 virtio_device_ready(vdev);
3879
3880 enable_delayed_refill(vi);
3881
3882 if (netif_running(vi->dev)) {
3883 err = virtnet_open(vi->dev);
3884 if (err)
3885 return err;
3886 }
3887
3888 netif_tx_lock_bh(vi->dev);
3889 netif_device_attach(vi->dev);
3890 netif_tx_unlock_bh(vi->dev);
3891 return err;
3892}
3893
3894static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
3895{
3896 struct scatterlist sg;
3897 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
3898
3899 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
3900
3901 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
3902 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
3903 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
3904 return -EINVAL;
3905 }
3906
3907 return 0;
3908}
3909
3910static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
3911{
3912 u64 offloads = 0;
3913
3914 if (!vi->guest_offloads)
3915 return 0;
3916
3917 return virtnet_set_guest_offloads(vi, offloads);
3918}
3919
3920static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
3921{
3922 u64 offloads = vi->guest_offloads;
3923
3924 if (!vi->guest_offloads)
3925 return 0;
3926
3927 return virtnet_set_guest_offloads(vi, offloads);
3928}
3929
3930static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
3931 struct netlink_ext_ack *extack)
3932{
3933 unsigned int room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM +
3934 sizeof(struct skb_shared_info));
3935 unsigned int max_sz = PAGE_SIZE - room - ETH_HLEN;
3936 struct virtnet_info *vi = netdev_priv(dev);
3937 struct bpf_prog *old_prog;
3938 u16 xdp_qp = 0, curr_qp;
3939 int i, err;
3940
3941 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
3942 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3943 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3944 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
3945 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
3946 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM) ||
3947 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) ||
3948 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6))) {
3949 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
3950 return -EOPNOTSUPP;
3951 }
3952
3953 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
3954 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
3955 return -EINVAL;
3956 }
3957
3958 if (prog && !prog->aux->xdp_has_frags && dev->mtu > max_sz) {
3959 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP without frags");
3960 netdev_warn(dev, "single-buffer XDP requires MTU less than %u\n", max_sz);
3961 return -EINVAL;
3962 }
3963
3964 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
3965 if (prog)
3966 xdp_qp = nr_cpu_ids;
3967
3968 /* XDP requires extra queues for XDP_TX */
3969 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
3970 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",
3971 curr_qp + xdp_qp, vi->max_queue_pairs);
3972 xdp_qp = 0;
3973 }
3974
3975 old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
3976 if (!prog && !old_prog)
3977 return 0;
3978
3979 if (prog)
3980 bpf_prog_add(prog, vi->max_queue_pairs - 1);
3981
3982 /* Make sure NAPI is not using any XDP TX queues for RX. */
3983 if (netif_running(dev)) {
3984 for (i = 0; i < vi->max_queue_pairs; i++) {
3985 napi_disable(&vi->rq[i].napi);
3986 virtnet_napi_tx_disable(&vi->sq[i].napi);
3987 }
3988 }
3989
3990 if (!prog) {
3991 for (i = 0; i < vi->max_queue_pairs; i++) {
3992 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
3993 if (i == 0)
3994 virtnet_restore_guest_offloads(vi);
3995 }
3996 synchronize_net();
3997 }
3998
3999 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
4000 if (err)
4001 goto err;
4002 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
4003 vi->xdp_queue_pairs = xdp_qp;
4004
4005 if (prog) {
4006 vi->xdp_enabled = true;
4007 for (i = 0; i < vi->max_queue_pairs; i++) {
4008 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
4009 if (i == 0 && !old_prog)
4010 virtnet_clear_guest_offloads(vi);
4011 }
4012 if (!old_prog)
4013 xdp_features_set_redirect_target(dev, true);
4014 } else {
4015 xdp_features_clear_redirect_target(dev);
4016 vi->xdp_enabled = false;
4017 }
4018
4019 for (i = 0; i < vi->max_queue_pairs; i++) {
4020 if (old_prog)
4021 bpf_prog_put(old_prog);
4022 if (netif_running(dev)) {
4023 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
4024 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
4025 &vi->sq[i].napi);
4026 }
4027 }
4028
4029 return 0;
4030
4031err:
4032 if (!prog) {
4033 virtnet_clear_guest_offloads(vi);
4034 for (i = 0; i < vi->max_queue_pairs; i++)
4035 rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
4036 }
4037
4038 if (netif_running(dev)) {
4039 for (i = 0; i < vi->max_queue_pairs; i++) {
4040 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
4041 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
4042 &vi->sq[i].napi);
4043 }
4044 }
4045 if (prog)
4046 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
4047 return err;
4048}
4049
4050static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
4051{
4052 switch (xdp->command) {
4053 case XDP_SETUP_PROG:
4054 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
4055 default:
4056 return -EINVAL;
4057 }
4058}
4059
4060static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
4061 size_t len)
4062{
4063 struct virtnet_info *vi = netdev_priv(dev);
4064 int ret;
4065
4066 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
4067 return -EOPNOTSUPP;
4068
4069 ret = snprintf(buf, len, "sby");
4070 if (ret >= len)
4071 return -EOPNOTSUPP;
4072
4073 return 0;
4074}
4075
4076static int virtnet_set_features(struct net_device *dev,
4077 netdev_features_t features)
4078{
4079 struct virtnet_info *vi = netdev_priv(dev);
4080 u64 offloads;
4081 int err;
4082
4083 if ((dev->features ^ features) & NETIF_F_GRO_HW) {
4084 if (vi->xdp_enabled)
4085 return -EBUSY;
4086
4087 if (features & NETIF_F_GRO_HW)
4088 offloads = vi->guest_offloads_capable;
4089 else
4090 offloads = vi->guest_offloads_capable &
4091 ~GUEST_OFFLOAD_GRO_HW_MASK;
4092
4093 err = virtnet_set_guest_offloads(vi, offloads);
4094 if (err)
4095 return err;
4096 vi->guest_offloads = offloads;
4097 }
4098
4099 if ((dev->features ^ features) & NETIF_F_RXHASH) {
4100 if (features & NETIF_F_RXHASH)
4101 vi->ctrl->rss.hash_types = vi->rss_hash_types_saved;
4102 else
4103 vi->ctrl->rss.hash_types = VIRTIO_NET_HASH_REPORT_NONE;
4104
4105 if (!virtnet_commit_rss_command(vi))
4106 return -EINVAL;
4107 }
4108
4109 return 0;
4110}
4111
4112static void virtnet_tx_timeout(struct net_device *dev, unsigned int txqueue)
4113{
4114 struct virtnet_info *priv = netdev_priv(dev);
4115 struct send_queue *sq = &priv->sq[txqueue];
4116 struct netdev_queue *txq = netdev_get_tx_queue(dev, txqueue);
4117
4118 u64_stats_update_begin(&sq->stats.syncp);
4119 u64_stats_inc(&sq->stats.tx_timeouts);
4120 u64_stats_update_end(&sq->stats.syncp);
4121
4122 netdev_err(dev, "TX timeout on queue: %u, sq: %s, vq: 0x%x, name: %s, %u usecs ago\n",
4123 txqueue, sq->name, sq->vq->index, sq->vq->name,
4124 jiffies_to_usecs(jiffies - READ_ONCE(txq->trans_start)));
4125}
4126
4127static const struct net_device_ops virtnet_netdev = {
4128 .ndo_open = virtnet_open,
4129 .ndo_stop = virtnet_close,
4130 .ndo_start_xmit = start_xmit,
4131 .ndo_validate_addr = eth_validate_addr,
4132 .ndo_set_mac_address = virtnet_set_mac_address,
4133 .ndo_set_rx_mode = virtnet_set_rx_mode,
4134 .ndo_get_stats64 = virtnet_stats,
4135 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
4136 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
4137 .ndo_bpf = virtnet_xdp,
4138 .ndo_xdp_xmit = virtnet_xdp_xmit,
4139 .ndo_features_check = passthru_features_check,
4140 .ndo_get_phys_port_name = virtnet_get_phys_port_name,
4141 .ndo_set_features = virtnet_set_features,
4142 .ndo_tx_timeout = virtnet_tx_timeout,
4143};
4144
4145static void virtnet_config_changed_work(struct work_struct *work)
4146{
4147 struct virtnet_info *vi =
4148 container_of(work, struct virtnet_info, config_work);
4149 u16 v;
4150
4151 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
4152 struct virtio_net_config, status, &v) < 0)
4153 return;
4154
4155 if (v & VIRTIO_NET_S_ANNOUNCE) {
4156 netdev_notify_peers(vi->dev);
4157 virtnet_ack_link_announce(vi);
4158 }
4159
4160 /* Ignore unknown (future) status bits */
4161 v &= VIRTIO_NET_S_LINK_UP;
4162
4163 if (vi->status == v)
4164 return;
4165
4166 vi->status = v;
4167
4168 if (vi->status & VIRTIO_NET_S_LINK_UP) {
4169 virtnet_update_settings(vi);
4170 netif_carrier_on(vi->dev);
4171 netif_tx_wake_all_queues(vi->dev);
4172 } else {
4173 netif_carrier_off(vi->dev);
4174 netif_tx_stop_all_queues(vi->dev);
4175 }
4176}
4177
4178static void virtnet_config_changed(struct virtio_device *vdev)
4179{
4180 struct virtnet_info *vi = vdev->priv;
4181
4182 schedule_work(&vi->config_work);
4183}
4184
4185static void virtnet_free_queues(struct virtnet_info *vi)
4186{
4187 int i;
4188
4189 for (i = 0; i < vi->max_queue_pairs; i++) {
4190 __netif_napi_del(&vi->rq[i].napi);
4191 __netif_napi_del(&vi->sq[i].napi);
4192 }
4193
4194 /* We called __netif_napi_del(),
4195 * we need to respect an RCU grace period before freeing vi->rq
4196 */
4197 synchronize_net();
4198
4199 kfree(vi->rq);
4200 kfree(vi->sq);
4201 kfree(vi->ctrl);
4202}
4203
4204static void _free_receive_bufs(struct virtnet_info *vi)
4205{
4206 struct bpf_prog *old_prog;
4207 int i;
4208
4209 for (i = 0; i < vi->max_queue_pairs; i++) {
4210 while (vi->rq[i].pages)
4211 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
4212
4213 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
4214 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
4215 if (old_prog)
4216 bpf_prog_put(old_prog);
4217 }
4218}
4219
4220static void free_receive_bufs(struct virtnet_info *vi)
4221{
4222 rtnl_lock();
4223 _free_receive_bufs(vi);
4224 rtnl_unlock();
4225}
4226
4227static void free_receive_page_frags(struct virtnet_info *vi)
4228{
4229 int i;
4230 for (i = 0; i < vi->max_queue_pairs; i++)
4231 if (vi->rq[i].alloc_frag.page) {
4232 if (vi->rq[i].do_dma && vi->rq[i].last_dma)
4233 virtnet_rq_unmap(&vi->rq[i], vi->rq[i].last_dma, 0);
4234 put_page(vi->rq[i].alloc_frag.page);
4235 }
4236}
4237
4238static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf)
4239{
4240 if (!is_xdp_frame(buf))
4241 dev_kfree_skb(buf);
4242 else
4243 xdp_return_frame(ptr_to_xdp(buf));
4244}
4245
4246static void free_unused_bufs(struct virtnet_info *vi)
4247{
4248 void *buf;
4249 int i;
4250
4251 for (i = 0; i < vi->max_queue_pairs; i++) {
4252 struct virtqueue *vq = vi->sq[i].vq;
4253 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
4254 virtnet_sq_free_unused_buf(vq, buf);
4255 cond_resched();
4256 }
4257
4258 for (i = 0; i < vi->max_queue_pairs; i++) {
4259 struct virtqueue *vq = vi->rq[i].vq;
4260
4261 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
4262 virtnet_rq_unmap_free_buf(vq, buf);
4263 cond_resched();
4264 }
4265}
4266
4267static void virtnet_del_vqs(struct virtnet_info *vi)
4268{
4269 struct virtio_device *vdev = vi->vdev;
4270
4271 virtnet_clean_affinity(vi);
4272
4273 vdev->config->del_vqs(vdev);
4274
4275 virtnet_free_queues(vi);
4276}
4277
4278/* How large should a single buffer be so a queue full of these can fit at
4279 * least one full packet?
4280 * Logic below assumes the mergeable buffer header is used.
4281 */
4282static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
4283{
4284 const unsigned int hdr_len = vi->hdr_len;
4285 unsigned int rq_size = virtqueue_get_vring_size(vq);
4286 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
4287 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
4288 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
4289
4290 return max(max(min_buf_len, hdr_len) - hdr_len,
4291 (unsigned int)GOOD_PACKET_LEN);
4292}
4293
4294static int virtnet_find_vqs(struct virtnet_info *vi)
4295{
4296 vq_callback_t **callbacks;
4297 struct virtqueue **vqs;
4298 const char **names;
4299 int ret = -ENOMEM;
4300 int total_vqs;
4301 bool *ctx;
4302 u16 i;
4303
4304 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
4305 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
4306 * possible control vq.
4307 */
4308 total_vqs = vi->max_queue_pairs * 2 +
4309 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
4310
4311 /* Allocate space for find_vqs parameters */
4312 vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
4313 if (!vqs)
4314 goto err_vq;
4315 callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
4316 if (!callbacks)
4317 goto err_callback;
4318 names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
4319 if (!names)
4320 goto err_names;
4321 if (!vi->big_packets || vi->mergeable_rx_bufs) {
4322 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
4323 if (!ctx)
4324 goto err_ctx;
4325 } else {
4326 ctx = NULL;
4327 }
4328
4329 /* Parameters for control virtqueue, if any */
4330 if (vi->has_cvq) {
4331 callbacks[total_vqs - 1] = NULL;
4332 names[total_vqs - 1] = "control";
4333 }
4334
4335 /* Allocate/initialize parameters for send/receive virtqueues */
4336 for (i = 0; i < vi->max_queue_pairs; i++) {
4337 callbacks[rxq2vq(i)] = skb_recv_done;
4338 callbacks[txq2vq(i)] = skb_xmit_done;
4339 sprintf(vi->rq[i].name, "input.%u", i);
4340 sprintf(vi->sq[i].name, "output.%u", i);
4341 names[rxq2vq(i)] = vi->rq[i].name;
4342 names[txq2vq(i)] = vi->sq[i].name;
4343 if (ctx)
4344 ctx[rxq2vq(i)] = true;
4345 }
4346
4347 ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
4348 names, ctx, NULL);
4349 if (ret)
4350 goto err_find;
4351
4352 if (vi->has_cvq) {
4353 vi->cvq = vqs[total_vqs - 1];
4354 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
4355 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
4356 }
4357
4358 for (i = 0; i < vi->max_queue_pairs; i++) {
4359 vi->rq[i].vq = vqs[rxq2vq(i)];
4360 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
4361 vi->sq[i].vq = vqs[txq2vq(i)];
4362 }
4363
4364 /* run here: ret == 0. */
4365
4366
4367err_find:
4368 kfree(ctx);
4369err_ctx:
4370 kfree(names);
4371err_names:
4372 kfree(callbacks);
4373err_callback:
4374 kfree(vqs);
4375err_vq:
4376 return ret;
4377}
4378
4379static int virtnet_alloc_queues(struct virtnet_info *vi)
4380{
4381 int i;
4382
4383 if (vi->has_cvq) {
4384 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
4385 if (!vi->ctrl)
4386 goto err_ctrl;
4387 } else {
4388 vi->ctrl = NULL;
4389 }
4390 vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
4391 if (!vi->sq)
4392 goto err_sq;
4393 vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
4394 if (!vi->rq)
4395 goto err_rq;
4396
4397 INIT_DELAYED_WORK(&vi->refill, refill_work);
4398 for (i = 0; i < vi->max_queue_pairs; i++) {
4399 vi->rq[i].pages = NULL;
4400 netif_napi_add_weight(vi->dev, &vi->rq[i].napi, virtnet_poll,
4401 napi_weight);
4402 netif_napi_add_tx_weight(vi->dev, &vi->sq[i].napi,
4403 virtnet_poll_tx,
4404 napi_tx ? napi_weight : 0);
4405
4406 INIT_WORK(&vi->rq[i].dim.work, virtnet_rx_dim_work);
4407 vi->rq[i].dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
4408
4409 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
4410 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
4411 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
4412
4413 u64_stats_init(&vi->rq[i].stats.syncp);
4414 u64_stats_init(&vi->sq[i].stats.syncp);
4415 }
4416
4417 return 0;
4418
4419err_rq:
4420 kfree(vi->sq);
4421err_sq:
4422 kfree(vi->ctrl);
4423err_ctrl:
4424 return -ENOMEM;
4425}
4426
4427static int init_vqs(struct virtnet_info *vi)
4428{
4429 int ret;
4430
4431 /* Allocate send & receive queues */
4432 ret = virtnet_alloc_queues(vi);
4433 if (ret)
4434 goto err;
4435
4436 ret = virtnet_find_vqs(vi);
4437 if (ret)
4438 goto err_free;
4439
4440 virtnet_rq_set_premapped(vi);
4441
4442 cpus_read_lock();
4443 virtnet_set_affinity(vi);
4444 cpus_read_unlock();
4445
4446 return 0;
4447
4448err_free:
4449 virtnet_free_queues(vi);
4450err:
4451 return ret;
4452}
4453
4454#ifdef CONFIG_SYSFS
4455static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
4456 char *buf)
4457{
4458 struct virtnet_info *vi = netdev_priv(queue->dev);
4459 unsigned int queue_index = get_netdev_rx_queue_index(queue);
4460 unsigned int headroom = virtnet_get_headroom(vi);
4461 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
4462 struct ewma_pkt_len *avg;
4463
4464 BUG_ON(queue_index >= vi->max_queue_pairs);
4465 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
4466 return sprintf(buf, "%u\n",
4467 get_mergeable_buf_len(&vi->rq[queue_index], avg,
4468 SKB_DATA_ALIGN(headroom + tailroom)));
4469}
4470
4471static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
4472 __ATTR_RO(mergeable_rx_buffer_size);
4473
4474static struct attribute *virtio_net_mrg_rx_attrs[] = {
4475 &mergeable_rx_buffer_size_attribute.attr,
4476 NULL
4477};
4478
4479static const struct attribute_group virtio_net_mrg_rx_group = {
4480 .name = "virtio_net",
4481 .attrs = virtio_net_mrg_rx_attrs
4482};
4483#endif
4484
4485static bool virtnet_fail_on_feature(struct virtio_device *vdev,
4486 unsigned int fbit,
4487 const char *fname, const char *dname)
4488{
4489 if (!virtio_has_feature(vdev, fbit))
4490 return false;
4491
4492 dev_err(&vdev->dev, "device advertises feature %s but not %s",
4493 fname, dname);
4494
4495 return true;
4496}
4497
4498#define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
4499 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
4500
4501static bool virtnet_validate_features(struct virtio_device *vdev)
4502{
4503 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
4504 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
4505 "VIRTIO_NET_F_CTRL_VQ") ||
4506 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
4507 "VIRTIO_NET_F_CTRL_VQ") ||
4508 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
4509 "VIRTIO_NET_F_CTRL_VQ") ||
4510 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
4511 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
4512 "VIRTIO_NET_F_CTRL_VQ") ||
4513 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_RSS,
4514 "VIRTIO_NET_F_CTRL_VQ") ||
4515 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_HASH_REPORT,
4516 "VIRTIO_NET_F_CTRL_VQ") ||
4517 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_NOTF_COAL,
4518 "VIRTIO_NET_F_CTRL_VQ") ||
4519 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_VQ_NOTF_COAL,
4520 "VIRTIO_NET_F_CTRL_VQ"))) {
4521 return false;
4522 }
4523
4524 return true;
4525}
4526
4527#define MIN_MTU ETH_MIN_MTU
4528#define MAX_MTU ETH_MAX_MTU
4529
4530static int virtnet_validate(struct virtio_device *vdev)
4531{
4532 if (!vdev->config->get) {
4533 dev_err(&vdev->dev, "%s failure: config access disabled\n",
4534 __func__);
4535 return -EINVAL;
4536 }
4537
4538 if (!virtnet_validate_features(vdev))
4539 return -EINVAL;
4540
4541 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
4542 int mtu = virtio_cread16(vdev,
4543 offsetof(struct virtio_net_config,
4544 mtu));
4545 if (mtu < MIN_MTU)
4546 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
4547 }
4548
4549 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY) &&
4550 !virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
4551 dev_warn(&vdev->dev, "device advertises feature VIRTIO_NET_F_STANDBY but not VIRTIO_NET_F_MAC, disabling standby");
4552 __virtio_clear_bit(vdev, VIRTIO_NET_F_STANDBY);
4553 }
4554
4555 return 0;
4556}
4557
4558static bool virtnet_check_guest_gso(const struct virtnet_info *vi)
4559{
4560 return virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
4561 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
4562 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
4563 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
4564 (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) &&
4565 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6));
4566}
4567
4568static void virtnet_set_big_packets(struct virtnet_info *vi, const int mtu)
4569{
4570 bool guest_gso = virtnet_check_guest_gso(vi);
4571
4572 /* If device can receive ANY guest GSO packets, regardless of mtu,
4573 * allocate packets of maximum size, otherwise limit it to only
4574 * mtu size worth only.
4575 */
4576 if (mtu > ETH_DATA_LEN || guest_gso) {
4577 vi->big_packets = true;
4578 vi->big_packets_num_skbfrags = guest_gso ? MAX_SKB_FRAGS : DIV_ROUND_UP(mtu, PAGE_SIZE);
4579 }
4580}
4581
4582static int virtnet_probe(struct virtio_device *vdev)
4583{
4584 int i, err = -ENOMEM;
4585 struct net_device *dev;
4586 struct virtnet_info *vi;
4587 u16 max_queue_pairs;
4588 int mtu = 0;
4589
4590 /* Find if host supports multiqueue/rss virtio_net device */
4591 max_queue_pairs = 1;
4592 if (virtio_has_feature(vdev, VIRTIO_NET_F_MQ) || virtio_has_feature(vdev, VIRTIO_NET_F_RSS))
4593 max_queue_pairs =
4594 virtio_cread16(vdev, offsetof(struct virtio_net_config, max_virtqueue_pairs));
4595
4596 /* We need at least 2 queue's */
4597 if (max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
4598 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
4599 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
4600 max_queue_pairs = 1;
4601
4602 /* Allocate ourselves a network device with room for our info */
4603 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
4604 if (!dev)
4605 return -ENOMEM;
4606
4607 /* Set up network device as normal. */
4608 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
4609 IFF_TX_SKB_NO_LINEAR;
4610 dev->netdev_ops = &virtnet_netdev;
4611 dev->features = NETIF_F_HIGHDMA;
4612
4613 dev->ethtool_ops = &virtnet_ethtool_ops;
4614 SET_NETDEV_DEV(dev, &vdev->dev);
4615
4616 /* Do we support "hardware" checksums? */
4617 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
4618 /* This opens up the world of extra features. */
4619 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
4620 if (csum)
4621 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
4622
4623 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
4624 dev->hw_features |= NETIF_F_TSO
4625 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
4626 }
4627 /* Individual feature bits: what can host handle? */
4628 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
4629 dev->hw_features |= NETIF_F_TSO;
4630 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
4631 dev->hw_features |= NETIF_F_TSO6;
4632 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
4633 dev->hw_features |= NETIF_F_TSO_ECN;
4634 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_USO))
4635 dev->hw_features |= NETIF_F_GSO_UDP_L4;
4636
4637 dev->features |= NETIF_F_GSO_ROBUST;
4638
4639 if (gso)
4640 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
4641 /* (!csum && gso) case will be fixed by register_netdev() */
4642 }
4643 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
4644 dev->features |= NETIF_F_RXCSUM;
4645 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
4646 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
4647 dev->features |= NETIF_F_GRO_HW;
4648 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
4649 dev->hw_features |= NETIF_F_GRO_HW;
4650
4651 dev->vlan_features = dev->features;
4652 dev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT;
4653
4654 /* MTU range: 68 - 65535 */
4655 dev->min_mtu = MIN_MTU;
4656 dev->max_mtu = MAX_MTU;
4657
4658 /* Configuration may specify what MAC to use. Otherwise random. */
4659 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
4660 u8 addr[ETH_ALEN];
4661
4662 virtio_cread_bytes(vdev,
4663 offsetof(struct virtio_net_config, mac),
4664 addr, ETH_ALEN);
4665 eth_hw_addr_set(dev, addr);
4666 } else {
4667 eth_hw_addr_random(dev);
4668 dev_info(&vdev->dev, "Assigned random MAC address %pM\n",
4669 dev->dev_addr);
4670 }
4671
4672 /* Set up our device-specific information */
4673 vi = netdev_priv(dev);
4674 vi->dev = dev;
4675 vi->vdev = vdev;
4676 vdev->priv = vi;
4677
4678 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
4679 spin_lock_init(&vi->refill_lock);
4680
4681 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) {
4682 vi->mergeable_rx_bufs = true;
4683 dev->xdp_features |= NETDEV_XDP_ACT_RX_SG;
4684 }
4685
4686 if (virtio_has_feature(vdev, VIRTIO_NET_F_HASH_REPORT))
4687 vi->has_rss_hash_report = true;
4688
4689 if (virtio_has_feature(vdev, VIRTIO_NET_F_RSS))
4690 vi->has_rss = true;
4691
4692 if (vi->has_rss || vi->has_rss_hash_report) {
4693 vi->rss_indir_table_size =
4694 virtio_cread16(vdev, offsetof(struct virtio_net_config,
4695 rss_max_indirection_table_length));
4696 vi->rss_key_size =
4697 virtio_cread8(vdev, offsetof(struct virtio_net_config, rss_max_key_size));
4698
4699 vi->rss_hash_types_supported =
4700 virtio_cread32(vdev, offsetof(struct virtio_net_config, supported_hash_types));
4701 vi->rss_hash_types_supported &=
4702 ~(VIRTIO_NET_RSS_HASH_TYPE_IP_EX |
4703 VIRTIO_NET_RSS_HASH_TYPE_TCP_EX |
4704 VIRTIO_NET_RSS_HASH_TYPE_UDP_EX);
4705
4706 dev->hw_features |= NETIF_F_RXHASH;
4707 }
4708
4709 if (vi->has_rss_hash_report)
4710 vi->hdr_len = sizeof(struct virtio_net_hdr_v1_hash);
4711 else if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
4712 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
4713 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
4714 else
4715 vi->hdr_len = sizeof(struct virtio_net_hdr);
4716
4717 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
4718 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
4719 vi->any_header_sg = true;
4720
4721 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
4722 vi->has_cvq = true;
4723
4724 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
4725 mtu = virtio_cread16(vdev,
4726 offsetof(struct virtio_net_config,
4727 mtu));
4728 if (mtu < dev->min_mtu) {
4729 /* Should never trigger: MTU was previously validated
4730 * in virtnet_validate.
4731 */
4732 dev_err(&vdev->dev,
4733 "device MTU appears to have changed it is now %d < %d",
4734 mtu, dev->min_mtu);
4735 err = -EINVAL;
4736 goto free;
4737 }
4738
4739 dev->mtu = mtu;
4740 dev->max_mtu = mtu;
4741 }
4742
4743 virtnet_set_big_packets(vi, mtu);
4744
4745 if (vi->any_header_sg)
4746 dev->needed_headroom = vi->hdr_len;
4747
4748 /* Enable multiqueue by default */
4749 if (num_online_cpus() >= max_queue_pairs)
4750 vi->curr_queue_pairs = max_queue_pairs;
4751 else
4752 vi->curr_queue_pairs = num_online_cpus();
4753 vi->max_queue_pairs = max_queue_pairs;
4754
4755 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
4756 err = init_vqs(vi);
4757 if (err)
4758 goto free;
4759
4760 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
4761 vi->intr_coal_rx.max_usecs = 0;
4762 vi->intr_coal_tx.max_usecs = 0;
4763 vi->intr_coal_rx.max_packets = 0;
4764
4765 /* Keep the default values of the coalescing parameters
4766 * aligned with the default napi_tx state.
4767 */
4768 if (vi->sq[0].napi.weight)
4769 vi->intr_coal_tx.max_packets = 1;
4770 else
4771 vi->intr_coal_tx.max_packets = 0;
4772 }
4773
4774 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL)) {
4775 /* The reason is the same as VIRTIO_NET_F_NOTF_COAL. */
4776 for (i = 0; i < vi->max_queue_pairs; i++)
4777 if (vi->sq[i].napi.weight)
4778 vi->sq[i].intr_coal.max_packets = 1;
4779 }
4780
4781#ifdef CONFIG_SYSFS
4782 if (vi->mergeable_rx_bufs)
4783 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
4784#endif
4785 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
4786 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
4787
4788 virtnet_init_settings(dev);
4789
4790 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
4791 vi->failover = net_failover_create(vi->dev);
4792 if (IS_ERR(vi->failover)) {
4793 err = PTR_ERR(vi->failover);
4794 goto free_vqs;
4795 }
4796 }
4797
4798 if (vi->has_rss || vi->has_rss_hash_report)
4799 virtnet_init_default_rss(vi);
4800
4801 /* serialize netdev register + virtio_device_ready() with ndo_open() */
4802 rtnl_lock();
4803
4804 err = register_netdevice(dev);
4805 if (err) {
4806 pr_debug("virtio_net: registering device failed\n");
4807 rtnl_unlock();
4808 goto free_failover;
4809 }
4810
4811 virtio_device_ready(vdev);
4812
4813 _virtnet_set_queues(vi, vi->curr_queue_pairs);
4814
4815 /* a random MAC address has been assigned, notify the device.
4816 * We don't fail probe if VIRTIO_NET_F_CTRL_MAC_ADDR is not there
4817 * because many devices work fine without getting MAC explicitly
4818 */
4819 if (!virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
4820 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
4821 struct scatterlist sg;
4822
4823 sg_init_one(&sg, dev->dev_addr, dev->addr_len);
4824 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
4825 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
4826 pr_debug("virtio_net: setting MAC address failed\n");
4827 rtnl_unlock();
4828 err = -EINVAL;
4829 goto free_unregister_netdev;
4830 }
4831 }
4832
4833 rtnl_unlock();
4834
4835 err = virtnet_cpu_notif_add(vi);
4836 if (err) {
4837 pr_debug("virtio_net: registering cpu notifier failed\n");
4838 goto free_unregister_netdev;
4839 }
4840
4841 /* Assume link up if device can't report link status,
4842 otherwise get link status from config. */
4843 netif_carrier_off(dev);
4844 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
4845 schedule_work(&vi->config_work);
4846 } else {
4847 vi->status = VIRTIO_NET_S_LINK_UP;
4848 virtnet_update_settings(vi);
4849 netif_carrier_on(dev);
4850 }
4851
4852 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
4853 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
4854 set_bit(guest_offloads[i], &vi->guest_offloads);
4855 vi->guest_offloads_capable = vi->guest_offloads;
4856
4857 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
4858 dev->name, max_queue_pairs);
4859
4860 return 0;
4861
4862free_unregister_netdev:
4863 unregister_netdev(dev);
4864free_failover:
4865 net_failover_destroy(vi->failover);
4866free_vqs:
4867 virtio_reset_device(vdev);
4868 cancel_delayed_work_sync(&vi->refill);
4869 free_receive_page_frags(vi);
4870 virtnet_del_vqs(vi);
4871free:
4872 free_netdev(dev);
4873 return err;
4874}
4875
4876static void remove_vq_common(struct virtnet_info *vi)
4877{
4878 virtio_reset_device(vi->vdev);
4879
4880 /* Free unused buffers in both send and recv, if any. */
4881 free_unused_bufs(vi);
4882
4883 free_receive_bufs(vi);
4884
4885 free_receive_page_frags(vi);
4886
4887 virtnet_del_vqs(vi);
4888}
4889
4890static void virtnet_remove(struct virtio_device *vdev)
4891{
4892 struct virtnet_info *vi = vdev->priv;
4893
4894 virtnet_cpu_notif_remove(vi);
4895
4896 /* Make sure no work handler is accessing the device. */
4897 flush_work(&vi->config_work);
4898
4899 unregister_netdev(vi->dev);
4900
4901 net_failover_destroy(vi->failover);
4902
4903 remove_vq_common(vi);
4904
4905 free_netdev(vi->dev);
4906}
4907
4908static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
4909{
4910 struct virtnet_info *vi = vdev->priv;
4911
4912 virtnet_cpu_notif_remove(vi);
4913 virtnet_freeze_down(vdev);
4914 remove_vq_common(vi);
4915
4916 return 0;
4917}
4918
4919static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
4920{
4921 struct virtnet_info *vi = vdev->priv;
4922 int err;
4923
4924 err = virtnet_restore_up(vdev);
4925 if (err)
4926 return err;
4927 virtnet_set_queues(vi, vi->curr_queue_pairs);
4928
4929 err = virtnet_cpu_notif_add(vi);
4930 if (err) {
4931 virtnet_freeze_down(vdev);
4932 remove_vq_common(vi);
4933 return err;
4934 }
4935
4936 return 0;
4937}
4938
4939static struct virtio_device_id id_table[] = {
4940 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
4941 { 0 },
4942};
4943
4944#define VIRTNET_FEATURES \
4945 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
4946 VIRTIO_NET_F_MAC, \
4947 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
4948 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
4949 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
4950 VIRTIO_NET_F_HOST_USO, VIRTIO_NET_F_GUEST_USO4, VIRTIO_NET_F_GUEST_USO6, \
4951 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
4952 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
4953 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
4954 VIRTIO_NET_F_CTRL_MAC_ADDR, \
4955 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
4956 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY, \
4957 VIRTIO_NET_F_RSS, VIRTIO_NET_F_HASH_REPORT, VIRTIO_NET_F_NOTF_COAL, \
4958 VIRTIO_NET_F_VQ_NOTF_COAL, \
4959 VIRTIO_NET_F_GUEST_HDRLEN
4960
4961static unsigned int features[] = {
4962 VIRTNET_FEATURES,
4963};
4964
4965static unsigned int features_legacy[] = {
4966 VIRTNET_FEATURES,
4967 VIRTIO_NET_F_GSO,
4968 VIRTIO_F_ANY_LAYOUT,
4969};
4970
4971static struct virtio_driver virtio_net_driver = {
4972 .feature_table = features,
4973 .feature_table_size = ARRAY_SIZE(features),
4974 .feature_table_legacy = features_legacy,
4975 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
4976 .driver.name = KBUILD_MODNAME,
4977 .driver.owner = THIS_MODULE,
4978 .id_table = id_table,
4979 .validate = virtnet_validate,
4980 .probe = virtnet_probe,
4981 .remove = virtnet_remove,
4982 .config_changed = virtnet_config_changed,
4983#ifdef CONFIG_PM_SLEEP
4984 .freeze = virtnet_freeze,
4985 .restore = virtnet_restore,
4986#endif
4987};
4988
4989static __init int virtio_net_driver_init(void)
4990{
4991 int ret;
4992
4993 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
4994 virtnet_cpu_online,
4995 virtnet_cpu_down_prep);
4996 if (ret < 0)
4997 goto out;
4998 virtionet_online = ret;
4999 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
5000 NULL, virtnet_cpu_dead);
5001 if (ret)
5002 goto err_dead;
5003 ret = register_virtio_driver(&virtio_net_driver);
5004 if (ret)
5005 goto err_virtio;
5006 return 0;
5007err_virtio:
5008 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
5009err_dead:
5010 cpuhp_remove_multi_state(virtionet_online);
5011out:
5012 return ret;
5013}
5014module_init(virtio_net_driver_init);
5015
5016static __exit void virtio_net_driver_exit(void)
5017{
5018 unregister_virtio_driver(&virtio_net_driver);
5019 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
5020 cpuhp_remove_multi_state(virtionet_online);
5021}
5022module_exit(virtio_net_driver_exit);
5023
5024MODULE_DEVICE_TABLE(virtio, id_table);
5025MODULE_DESCRIPTION("Virtio network driver");
5026MODULE_LICENSE("GPL");
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 <linux/dim.h>
23#include <net/route.h>
24#include <net/xdp.h>
25#include <net/net_failover.h>
26#include <net/netdev_rx_queue.h>
27#include <net/netdev_queues.h>
28#include <net/xdp_sock_drv.h>
29
30static int napi_weight = NAPI_POLL_WEIGHT;
31module_param(napi_weight, int, 0444);
32
33static bool csum = true, gso = true, napi_tx = true;
34module_param(csum, bool, 0444);
35module_param(gso, bool, 0444);
36module_param(napi_tx, bool, 0644);
37
38/* FIXME: MTU in config. */
39#define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
40#define GOOD_COPY_LEN 128
41
42#define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
43
44/* Separating two types of XDP xmit */
45#define VIRTIO_XDP_TX BIT(0)
46#define VIRTIO_XDP_REDIR BIT(1)
47
48/* RX packet size EWMA. The average packet size is used to determine the packet
49 * buffer size when refilling RX rings. As the entire RX ring may be refilled
50 * at once, the weight is chosen so that the EWMA will be insensitive to short-
51 * term, transient changes in packet size.
52 */
53DECLARE_EWMA(pkt_len, 0, 64)
54
55#define VIRTNET_DRIVER_VERSION "1.0.0"
56
57static const unsigned long guest_offloads[] = {
58 VIRTIO_NET_F_GUEST_TSO4,
59 VIRTIO_NET_F_GUEST_TSO6,
60 VIRTIO_NET_F_GUEST_ECN,
61 VIRTIO_NET_F_GUEST_UFO,
62 VIRTIO_NET_F_GUEST_CSUM,
63 VIRTIO_NET_F_GUEST_USO4,
64 VIRTIO_NET_F_GUEST_USO6,
65 VIRTIO_NET_F_GUEST_HDRLEN
66};
67
68#define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
69 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
70 (1ULL << VIRTIO_NET_F_GUEST_ECN) | \
71 (1ULL << VIRTIO_NET_F_GUEST_UFO) | \
72 (1ULL << VIRTIO_NET_F_GUEST_USO4) | \
73 (1ULL << VIRTIO_NET_F_GUEST_USO6))
74
75struct virtnet_stat_desc {
76 char desc[ETH_GSTRING_LEN];
77 size_t offset;
78 size_t qstat_offset;
79};
80
81struct virtnet_sq_free_stats {
82 u64 packets;
83 u64 bytes;
84 u64 napi_packets;
85 u64 napi_bytes;
86 u64 xsk;
87};
88
89struct virtnet_sq_stats {
90 struct u64_stats_sync syncp;
91 u64_stats_t packets;
92 u64_stats_t bytes;
93 u64_stats_t xdp_tx;
94 u64_stats_t xdp_tx_drops;
95 u64_stats_t kicks;
96 u64_stats_t tx_timeouts;
97 u64_stats_t stop;
98 u64_stats_t wake;
99};
100
101struct virtnet_rq_stats {
102 struct u64_stats_sync syncp;
103 u64_stats_t packets;
104 u64_stats_t bytes;
105 u64_stats_t drops;
106 u64_stats_t xdp_packets;
107 u64_stats_t xdp_tx;
108 u64_stats_t xdp_redirects;
109 u64_stats_t xdp_drops;
110 u64_stats_t kicks;
111};
112
113#define VIRTNET_SQ_STAT(name, m) {name, offsetof(struct virtnet_sq_stats, m), -1}
114#define VIRTNET_RQ_STAT(name, m) {name, offsetof(struct virtnet_rq_stats, m), -1}
115
116#define VIRTNET_SQ_STAT_QSTAT(name, m) \
117 { \
118 name, \
119 offsetof(struct virtnet_sq_stats, m), \
120 offsetof(struct netdev_queue_stats_tx, m), \
121 }
122
123#define VIRTNET_RQ_STAT_QSTAT(name, m) \
124 { \
125 name, \
126 offsetof(struct virtnet_rq_stats, m), \
127 offsetof(struct netdev_queue_stats_rx, m), \
128 }
129
130static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
131 VIRTNET_SQ_STAT("xdp_tx", xdp_tx),
132 VIRTNET_SQ_STAT("xdp_tx_drops", xdp_tx_drops),
133 VIRTNET_SQ_STAT("kicks", kicks),
134 VIRTNET_SQ_STAT("tx_timeouts", tx_timeouts),
135};
136
137static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
138 VIRTNET_RQ_STAT("drops", drops),
139 VIRTNET_RQ_STAT("xdp_packets", xdp_packets),
140 VIRTNET_RQ_STAT("xdp_tx", xdp_tx),
141 VIRTNET_RQ_STAT("xdp_redirects", xdp_redirects),
142 VIRTNET_RQ_STAT("xdp_drops", xdp_drops),
143 VIRTNET_RQ_STAT("kicks", kicks),
144};
145
146static const struct virtnet_stat_desc virtnet_sq_stats_desc_qstat[] = {
147 VIRTNET_SQ_STAT_QSTAT("packets", packets),
148 VIRTNET_SQ_STAT_QSTAT("bytes", bytes),
149 VIRTNET_SQ_STAT_QSTAT("stop", stop),
150 VIRTNET_SQ_STAT_QSTAT("wake", wake),
151};
152
153static const struct virtnet_stat_desc virtnet_rq_stats_desc_qstat[] = {
154 VIRTNET_RQ_STAT_QSTAT("packets", packets),
155 VIRTNET_RQ_STAT_QSTAT("bytes", bytes),
156};
157
158#define VIRTNET_STATS_DESC_CQ(name) \
159 {#name, offsetof(struct virtio_net_stats_cvq, name), -1}
160
161#define VIRTNET_STATS_DESC_RX(class, name) \
162 {#name, offsetof(struct virtio_net_stats_rx_ ## class, rx_ ## name), -1}
163
164#define VIRTNET_STATS_DESC_TX(class, name) \
165 {#name, offsetof(struct virtio_net_stats_tx_ ## class, tx_ ## name), -1}
166
167
168static const struct virtnet_stat_desc virtnet_stats_cvq_desc[] = {
169 VIRTNET_STATS_DESC_CQ(command_num),
170 VIRTNET_STATS_DESC_CQ(ok_num),
171};
172
173static const struct virtnet_stat_desc virtnet_stats_rx_basic_desc[] = {
174 VIRTNET_STATS_DESC_RX(basic, packets),
175 VIRTNET_STATS_DESC_RX(basic, bytes),
176
177 VIRTNET_STATS_DESC_RX(basic, notifications),
178 VIRTNET_STATS_DESC_RX(basic, interrupts),
179};
180
181static const struct virtnet_stat_desc virtnet_stats_tx_basic_desc[] = {
182 VIRTNET_STATS_DESC_TX(basic, packets),
183 VIRTNET_STATS_DESC_TX(basic, bytes),
184
185 VIRTNET_STATS_DESC_TX(basic, notifications),
186 VIRTNET_STATS_DESC_TX(basic, interrupts),
187};
188
189static const struct virtnet_stat_desc virtnet_stats_rx_csum_desc[] = {
190 VIRTNET_STATS_DESC_RX(csum, needs_csum),
191};
192
193static const struct virtnet_stat_desc virtnet_stats_tx_gso_desc[] = {
194 VIRTNET_STATS_DESC_TX(gso, gso_packets_noseg),
195 VIRTNET_STATS_DESC_TX(gso, gso_bytes_noseg),
196};
197
198static const struct virtnet_stat_desc virtnet_stats_rx_speed_desc[] = {
199 VIRTNET_STATS_DESC_RX(speed, ratelimit_bytes),
200};
201
202static const struct virtnet_stat_desc virtnet_stats_tx_speed_desc[] = {
203 VIRTNET_STATS_DESC_TX(speed, ratelimit_bytes),
204};
205
206#define VIRTNET_STATS_DESC_RX_QSTAT(class, name, qstat_field) \
207 { \
208 #name, \
209 offsetof(struct virtio_net_stats_rx_ ## class, rx_ ## name), \
210 offsetof(struct netdev_queue_stats_rx, qstat_field), \
211 }
212
213#define VIRTNET_STATS_DESC_TX_QSTAT(class, name, qstat_field) \
214 { \
215 #name, \
216 offsetof(struct virtio_net_stats_tx_ ## class, tx_ ## name), \
217 offsetof(struct netdev_queue_stats_tx, qstat_field), \
218 }
219
220static const struct virtnet_stat_desc virtnet_stats_rx_basic_desc_qstat[] = {
221 VIRTNET_STATS_DESC_RX_QSTAT(basic, drops, hw_drops),
222 VIRTNET_STATS_DESC_RX_QSTAT(basic, drop_overruns, hw_drop_overruns),
223};
224
225static const struct virtnet_stat_desc virtnet_stats_tx_basic_desc_qstat[] = {
226 VIRTNET_STATS_DESC_TX_QSTAT(basic, drops, hw_drops),
227 VIRTNET_STATS_DESC_TX_QSTAT(basic, drop_malformed, hw_drop_errors),
228};
229
230static const struct virtnet_stat_desc virtnet_stats_rx_csum_desc_qstat[] = {
231 VIRTNET_STATS_DESC_RX_QSTAT(csum, csum_valid, csum_unnecessary),
232 VIRTNET_STATS_DESC_RX_QSTAT(csum, csum_none, csum_none),
233 VIRTNET_STATS_DESC_RX_QSTAT(csum, csum_bad, csum_bad),
234};
235
236static const struct virtnet_stat_desc virtnet_stats_tx_csum_desc_qstat[] = {
237 VIRTNET_STATS_DESC_TX_QSTAT(csum, csum_none, csum_none),
238 VIRTNET_STATS_DESC_TX_QSTAT(csum, needs_csum, needs_csum),
239};
240
241static const struct virtnet_stat_desc virtnet_stats_rx_gso_desc_qstat[] = {
242 VIRTNET_STATS_DESC_RX_QSTAT(gso, gso_packets, hw_gro_packets),
243 VIRTNET_STATS_DESC_RX_QSTAT(gso, gso_bytes, hw_gro_bytes),
244 VIRTNET_STATS_DESC_RX_QSTAT(gso, gso_packets_coalesced, hw_gro_wire_packets),
245 VIRTNET_STATS_DESC_RX_QSTAT(gso, gso_bytes_coalesced, hw_gro_wire_bytes),
246};
247
248static const struct virtnet_stat_desc virtnet_stats_tx_gso_desc_qstat[] = {
249 VIRTNET_STATS_DESC_TX_QSTAT(gso, gso_packets, hw_gso_packets),
250 VIRTNET_STATS_DESC_TX_QSTAT(gso, gso_bytes, hw_gso_bytes),
251 VIRTNET_STATS_DESC_TX_QSTAT(gso, gso_segments, hw_gso_wire_packets),
252 VIRTNET_STATS_DESC_TX_QSTAT(gso, gso_segments_bytes, hw_gso_wire_bytes),
253};
254
255static const struct virtnet_stat_desc virtnet_stats_rx_speed_desc_qstat[] = {
256 VIRTNET_STATS_DESC_RX_QSTAT(speed, ratelimit_packets, hw_drop_ratelimits),
257};
258
259static const struct virtnet_stat_desc virtnet_stats_tx_speed_desc_qstat[] = {
260 VIRTNET_STATS_DESC_TX_QSTAT(speed, ratelimit_packets, hw_drop_ratelimits),
261};
262
263#define VIRTNET_Q_TYPE_RX 0
264#define VIRTNET_Q_TYPE_TX 1
265#define VIRTNET_Q_TYPE_CQ 2
266
267struct virtnet_interrupt_coalesce {
268 u32 max_packets;
269 u32 max_usecs;
270};
271
272/* The dma information of pages allocated at a time. */
273struct virtnet_rq_dma {
274 dma_addr_t addr;
275 u32 ref;
276 u16 len;
277 u16 need_sync;
278};
279
280/* Internal representation of a send virtqueue */
281struct send_queue {
282 /* Virtqueue associated with this send _queue */
283 struct virtqueue *vq;
284
285 /* TX: fragments + linear part + virtio header */
286 struct scatterlist sg[MAX_SKB_FRAGS + 2];
287
288 /* Name of the send queue: output.$index */
289 char name[16];
290
291 struct virtnet_sq_stats stats;
292
293 struct virtnet_interrupt_coalesce intr_coal;
294
295 struct napi_struct napi;
296
297 /* Record whether sq is in reset state. */
298 bool reset;
299
300 struct xsk_buff_pool *xsk_pool;
301
302 dma_addr_t xsk_hdr_dma_addr;
303};
304
305/* Internal representation of a receive virtqueue */
306struct receive_queue {
307 /* Virtqueue associated with this receive_queue */
308 struct virtqueue *vq;
309
310 struct napi_struct napi;
311
312 struct bpf_prog __rcu *xdp_prog;
313
314 struct virtnet_rq_stats stats;
315
316 /* The number of rx notifications */
317 u16 calls;
318
319 /* Is dynamic interrupt moderation enabled? */
320 bool dim_enabled;
321
322 /* Used to protect dim_enabled and inter_coal */
323 struct mutex dim_lock;
324
325 /* Dynamic Interrupt Moderation */
326 struct dim dim;
327
328 u32 packets_in_napi;
329
330 struct virtnet_interrupt_coalesce intr_coal;
331
332 /* Chain pages by the private ptr. */
333 struct page *pages;
334
335 /* Average packet length for mergeable receive buffers. */
336 struct ewma_pkt_len mrg_avg_pkt_len;
337
338 /* Page frag for packet buffer allocation. */
339 struct page_frag alloc_frag;
340
341 /* RX: fragments + linear part + virtio header */
342 struct scatterlist sg[MAX_SKB_FRAGS + 2];
343
344 /* Min single buffer size for mergeable buffers case. */
345 unsigned int min_buf_len;
346
347 /* Name of this receive queue: input.$index */
348 char name[16];
349
350 struct xdp_rxq_info xdp_rxq;
351
352 /* Record the last dma info to free after new pages is allocated. */
353 struct virtnet_rq_dma *last_dma;
354
355 struct xsk_buff_pool *xsk_pool;
356
357 /* xdp rxq used by xsk */
358 struct xdp_rxq_info xsk_rxq_info;
359
360 struct xdp_buff **xsk_buffs;
361};
362
363/* This structure can contain rss message with maximum settings for indirection table and keysize
364 * Note, that default structure that describes RSS configuration virtio_net_rss_config
365 * contains same info but can't handle table values.
366 * In any case, structure would be passed to virtio hw through sg_buf split by parts
367 * because table sizes may be differ according to the device configuration.
368 */
369#define VIRTIO_NET_RSS_MAX_KEY_SIZE 40
370struct virtio_net_ctrl_rss {
371 u32 hash_types;
372 u16 indirection_table_mask;
373 u16 unclassified_queue;
374 u16 hash_cfg_reserved; /* for HASH_CONFIG (see virtio_net_hash_config for details) */
375 u16 max_tx_vq;
376 u8 hash_key_length;
377 u8 key[VIRTIO_NET_RSS_MAX_KEY_SIZE];
378
379 u16 *indirection_table;
380};
381
382/* Control VQ buffers: protected by the rtnl lock */
383struct control_buf {
384 struct virtio_net_ctrl_hdr hdr;
385 virtio_net_ctrl_ack status;
386};
387
388struct virtnet_info {
389 struct virtio_device *vdev;
390 struct virtqueue *cvq;
391 struct net_device *dev;
392 struct send_queue *sq;
393 struct receive_queue *rq;
394 unsigned int status;
395
396 /* Max # of queue pairs supported by the device */
397 u16 max_queue_pairs;
398
399 /* # of queue pairs currently used by the driver */
400 u16 curr_queue_pairs;
401
402 /* # of XDP queue pairs currently used by the driver */
403 u16 xdp_queue_pairs;
404
405 /* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
406 bool xdp_enabled;
407
408 /* I like... big packets and I cannot lie! */
409 bool big_packets;
410
411 /* number of sg entries allocated for big packets */
412 unsigned int big_packets_num_skbfrags;
413
414 /* Host will merge rx buffers for big packets (shake it! shake it!) */
415 bool mergeable_rx_bufs;
416
417 /* Host supports rss and/or hash report */
418 bool has_rss;
419 bool has_rss_hash_report;
420 u8 rss_key_size;
421 u16 rss_indir_table_size;
422 u32 rss_hash_types_supported;
423 u32 rss_hash_types_saved;
424 struct virtio_net_ctrl_rss rss;
425
426 /* Has control virtqueue */
427 bool has_cvq;
428
429 /* Lock to protect the control VQ */
430 struct mutex cvq_lock;
431
432 /* Host can handle any s/g split between our header and packet data */
433 bool any_header_sg;
434
435 /* Packet virtio header size */
436 u8 hdr_len;
437
438 /* Work struct for delayed refilling if we run low on memory. */
439 struct delayed_work refill;
440
441 /* Is delayed refill enabled? */
442 bool refill_enabled;
443
444 /* The lock to synchronize the access to refill_enabled */
445 spinlock_t refill_lock;
446
447 /* Work struct for config space updates */
448 struct work_struct config_work;
449
450 /* Work struct for setting rx mode */
451 struct work_struct rx_mode_work;
452
453 /* OK to queue work setting RX mode? */
454 bool rx_mode_work_enabled;
455
456 /* Does the affinity hint is set for virtqueues? */
457 bool affinity_hint_set;
458
459 /* CPU hotplug instances for online & dead */
460 struct hlist_node node;
461 struct hlist_node node_dead;
462
463 struct control_buf *ctrl;
464
465 /* Ethtool settings */
466 u8 duplex;
467 u32 speed;
468
469 /* Is rx dynamic interrupt moderation enabled? */
470 bool rx_dim_enabled;
471
472 /* Interrupt coalescing settings */
473 struct virtnet_interrupt_coalesce intr_coal_tx;
474 struct virtnet_interrupt_coalesce intr_coal_rx;
475
476 unsigned long guest_offloads;
477 unsigned long guest_offloads_capable;
478
479 /* failover when STANDBY feature enabled */
480 struct failover *failover;
481
482 u64 device_stats_cap;
483};
484
485struct padded_vnet_hdr {
486 struct virtio_net_hdr_v1_hash hdr;
487 /*
488 * hdr is in a separate sg buffer, and data sg buffer shares same page
489 * with this header sg. This padding makes next sg 16 byte aligned
490 * after the header.
491 */
492 char padding[12];
493};
494
495struct virtio_net_common_hdr {
496 union {
497 struct virtio_net_hdr hdr;
498 struct virtio_net_hdr_mrg_rxbuf mrg_hdr;
499 struct virtio_net_hdr_v1_hash hash_v1_hdr;
500 };
501};
502
503static struct virtio_net_common_hdr xsk_hdr;
504
505static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf);
506static void virtnet_sq_free_unused_buf_done(struct virtqueue *vq);
507static int virtnet_xdp_handler(struct bpf_prog *xdp_prog, struct xdp_buff *xdp,
508 struct net_device *dev,
509 unsigned int *xdp_xmit,
510 struct virtnet_rq_stats *stats);
511static void virtnet_receive_done(struct virtnet_info *vi, struct receive_queue *rq,
512 struct sk_buff *skb, u8 flags);
513static struct sk_buff *virtnet_skb_append_frag(struct sk_buff *head_skb,
514 struct sk_buff *curr_skb,
515 struct page *page, void *buf,
516 int len, int truesize);
517static void virtnet_xsk_completed(struct send_queue *sq, int num);
518
519enum virtnet_xmit_type {
520 VIRTNET_XMIT_TYPE_SKB,
521 VIRTNET_XMIT_TYPE_SKB_ORPHAN,
522 VIRTNET_XMIT_TYPE_XDP,
523 VIRTNET_XMIT_TYPE_XSK,
524};
525
526static int rss_indirection_table_alloc(struct virtio_net_ctrl_rss *rss, u16 indir_table_size)
527{
528 if (!indir_table_size) {
529 rss->indirection_table = NULL;
530 return 0;
531 }
532
533 rss->indirection_table = kmalloc_array(indir_table_size, sizeof(u16), GFP_KERNEL);
534 if (!rss->indirection_table)
535 return -ENOMEM;
536
537 return 0;
538}
539
540static void rss_indirection_table_free(struct virtio_net_ctrl_rss *rss)
541{
542 kfree(rss->indirection_table);
543}
544
545/* We use the last two bits of the pointer to distinguish the xmit type. */
546#define VIRTNET_XMIT_TYPE_MASK (BIT(0) | BIT(1))
547
548#define VIRTIO_XSK_FLAG_OFFSET 2
549
550static enum virtnet_xmit_type virtnet_xmit_ptr_unpack(void **ptr)
551{
552 unsigned long p = (unsigned long)*ptr;
553
554 *ptr = (void *)(p & ~VIRTNET_XMIT_TYPE_MASK);
555
556 return p & VIRTNET_XMIT_TYPE_MASK;
557}
558
559static void *virtnet_xmit_ptr_pack(void *ptr, enum virtnet_xmit_type type)
560{
561 return (void *)((unsigned long)ptr | type);
562}
563
564static int virtnet_add_outbuf(struct send_queue *sq, int num, void *data,
565 enum virtnet_xmit_type type)
566{
567 return virtqueue_add_outbuf(sq->vq, sq->sg, num,
568 virtnet_xmit_ptr_pack(data, type),
569 GFP_ATOMIC);
570}
571
572static u32 virtnet_ptr_to_xsk_buff_len(void *ptr)
573{
574 return ((unsigned long)ptr) >> VIRTIO_XSK_FLAG_OFFSET;
575}
576
577static void sg_fill_dma(struct scatterlist *sg, dma_addr_t addr, u32 len)
578{
579 sg_dma_address(sg) = addr;
580 sg_dma_len(sg) = len;
581}
582
583static void __free_old_xmit(struct send_queue *sq, struct netdev_queue *txq,
584 bool in_napi, struct virtnet_sq_free_stats *stats)
585{
586 struct xdp_frame *frame;
587 struct sk_buff *skb;
588 unsigned int len;
589 void *ptr;
590
591 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
592 switch (virtnet_xmit_ptr_unpack(&ptr)) {
593 case VIRTNET_XMIT_TYPE_SKB:
594 skb = ptr;
595
596 pr_debug("Sent skb %p\n", skb);
597 stats->napi_packets++;
598 stats->napi_bytes += skb->len;
599 napi_consume_skb(skb, in_napi);
600 break;
601
602 case VIRTNET_XMIT_TYPE_SKB_ORPHAN:
603 skb = ptr;
604
605 stats->packets++;
606 stats->bytes += skb->len;
607 napi_consume_skb(skb, in_napi);
608 break;
609
610 case VIRTNET_XMIT_TYPE_XDP:
611 frame = ptr;
612
613 stats->packets++;
614 stats->bytes += xdp_get_frame_len(frame);
615 xdp_return_frame(frame);
616 break;
617
618 case VIRTNET_XMIT_TYPE_XSK:
619 stats->bytes += virtnet_ptr_to_xsk_buff_len(ptr);
620 stats->xsk++;
621 break;
622 }
623 }
624 netdev_tx_completed_queue(txq, stats->napi_packets, stats->napi_bytes);
625}
626
627static void virtnet_free_old_xmit(struct send_queue *sq,
628 struct netdev_queue *txq,
629 bool in_napi,
630 struct virtnet_sq_free_stats *stats)
631{
632 __free_old_xmit(sq, txq, in_napi, stats);
633
634 if (stats->xsk)
635 virtnet_xsk_completed(sq, stats->xsk);
636}
637
638/* Converting between virtqueue no. and kernel tx/rx queue no.
639 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
640 */
641static int vq2txq(struct virtqueue *vq)
642{
643 return (vq->index - 1) / 2;
644}
645
646static int txq2vq(int txq)
647{
648 return txq * 2 + 1;
649}
650
651static int vq2rxq(struct virtqueue *vq)
652{
653 return vq->index / 2;
654}
655
656static int rxq2vq(int rxq)
657{
658 return rxq * 2;
659}
660
661static int vq_type(struct virtnet_info *vi, int qid)
662{
663 if (qid == vi->max_queue_pairs * 2)
664 return VIRTNET_Q_TYPE_CQ;
665
666 if (qid % 2)
667 return VIRTNET_Q_TYPE_TX;
668
669 return VIRTNET_Q_TYPE_RX;
670}
671
672static inline struct virtio_net_common_hdr *
673skb_vnet_common_hdr(struct sk_buff *skb)
674{
675 return (struct virtio_net_common_hdr *)skb->cb;
676}
677
678/*
679 * private is used to chain pages for big packets, put the whole
680 * most recent used list in the beginning for reuse
681 */
682static void give_pages(struct receive_queue *rq, struct page *page)
683{
684 struct page *end;
685
686 /* Find end of list, sew whole thing into vi->rq.pages. */
687 for (end = page; end->private; end = (struct page *)end->private);
688 end->private = (unsigned long)rq->pages;
689 rq->pages = page;
690}
691
692static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
693{
694 struct page *p = rq->pages;
695
696 if (p) {
697 rq->pages = (struct page *)p->private;
698 /* clear private here, it is used to chain pages */
699 p->private = 0;
700 } else
701 p = alloc_page(gfp_mask);
702 return p;
703}
704
705static void virtnet_rq_free_buf(struct virtnet_info *vi,
706 struct receive_queue *rq, void *buf)
707{
708 if (vi->mergeable_rx_bufs)
709 put_page(virt_to_head_page(buf));
710 else if (vi->big_packets)
711 give_pages(rq, buf);
712 else
713 put_page(virt_to_head_page(buf));
714}
715
716static void enable_delayed_refill(struct virtnet_info *vi)
717{
718 spin_lock_bh(&vi->refill_lock);
719 vi->refill_enabled = true;
720 spin_unlock_bh(&vi->refill_lock);
721}
722
723static void disable_delayed_refill(struct virtnet_info *vi)
724{
725 spin_lock_bh(&vi->refill_lock);
726 vi->refill_enabled = false;
727 spin_unlock_bh(&vi->refill_lock);
728}
729
730static void enable_rx_mode_work(struct virtnet_info *vi)
731{
732 rtnl_lock();
733 vi->rx_mode_work_enabled = true;
734 rtnl_unlock();
735}
736
737static void disable_rx_mode_work(struct virtnet_info *vi)
738{
739 rtnl_lock();
740 vi->rx_mode_work_enabled = false;
741 rtnl_unlock();
742}
743
744static void virtqueue_napi_schedule(struct napi_struct *napi,
745 struct virtqueue *vq)
746{
747 if (napi_schedule_prep(napi)) {
748 virtqueue_disable_cb(vq);
749 __napi_schedule(napi);
750 }
751}
752
753static bool virtqueue_napi_complete(struct napi_struct *napi,
754 struct virtqueue *vq, int processed)
755{
756 int opaque;
757
758 opaque = virtqueue_enable_cb_prepare(vq);
759 if (napi_complete_done(napi, processed)) {
760 if (unlikely(virtqueue_poll(vq, opaque)))
761 virtqueue_napi_schedule(napi, vq);
762 else
763 return true;
764 } else {
765 virtqueue_disable_cb(vq);
766 }
767
768 return false;
769}
770
771static void skb_xmit_done(struct virtqueue *vq)
772{
773 struct virtnet_info *vi = vq->vdev->priv;
774 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
775
776 /* Suppress further interrupts. */
777 virtqueue_disable_cb(vq);
778
779 if (napi->weight)
780 virtqueue_napi_schedule(napi, vq);
781 else
782 /* We were probably waiting for more output buffers. */
783 netif_wake_subqueue(vi->dev, vq2txq(vq));
784}
785
786#define MRG_CTX_HEADER_SHIFT 22
787static void *mergeable_len_to_ctx(unsigned int truesize,
788 unsigned int headroom)
789{
790 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
791}
792
793static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
794{
795 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
796}
797
798static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
799{
800 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
801}
802
803static struct sk_buff *virtnet_build_skb(void *buf, unsigned int buflen,
804 unsigned int headroom,
805 unsigned int len)
806{
807 struct sk_buff *skb;
808
809 skb = build_skb(buf, buflen);
810 if (unlikely(!skb))
811 return NULL;
812
813 skb_reserve(skb, headroom);
814 skb_put(skb, len);
815
816 return skb;
817}
818
819/* Called from bottom half context */
820static struct sk_buff *page_to_skb(struct virtnet_info *vi,
821 struct receive_queue *rq,
822 struct page *page, unsigned int offset,
823 unsigned int len, unsigned int truesize,
824 unsigned int headroom)
825{
826 struct sk_buff *skb;
827 struct virtio_net_common_hdr *hdr;
828 unsigned int copy, hdr_len, hdr_padded_len;
829 struct page *page_to_free = NULL;
830 int tailroom, shinfo_size;
831 char *p, *hdr_p, *buf;
832
833 p = page_address(page) + offset;
834 hdr_p = p;
835
836 hdr_len = vi->hdr_len;
837 if (vi->mergeable_rx_bufs)
838 hdr_padded_len = hdr_len;
839 else
840 hdr_padded_len = sizeof(struct padded_vnet_hdr);
841
842 buf = p - headroom;
843 len -= hdr_len;
844 offset += hdr_padded_len;
845 p += hdr_padded_len;
846 tailroom = truesize - headroom - hdr_padded_len - len;
847
848 shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
849
850 if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
851 skb = virtnet_build_skb(buf, truesize, p - buf, len);
852 if (unlikely(!skb))
853 return NULL;
854
855 page = (struct page *)page->private;
856 if (page)
857 give_pages(rq, page);
858 goto ok;
859 }
860
861 /* copy small packet so we can reuse these pages for small data */
862 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
863 if (unlikely(!skb))
864 return NULL;
865
866 /* Copy all frame if it fits skb->head, otherwise
867 * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
868 */
869 if (len <= skb_tailroom(skb))
870 copy = len;
871 else
872 copy = ETH_HLEN;
873 skb_put_data(skb, p, copy);
874
875 len -= copy;
876 offset += copy;
877
878 if (vi->mergeable_rx_bufs) {
879 if (len)
880 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
881 else
882 page_to_free = page;
883 goto ok;
884 }
885
886 /*
887 * Verify that we can indeed put this data into a skb.
888 * This is here to handle cases when the device erroneously
889 * tries to receive more than is possible. This is usually
890 * the case of a broken device.
891 */
892 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
893 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
894 dev_kfree_skb(skb);
895 return NULL;
896 }
897 BUG_ON(offset >= PAGE_SIZE);
898 while (len) {
899 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
900 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
901 frag_size, truesize);
902 len -= frag_size;
903 page = (struct page *)page->private;
904 offset = 0;
905 }
906
907 if (page)
908 give_pages(rq, page);
909
910ok:
911 hdr = skb_vnet_common_hdr(skb);
912 memcpy(hdr, hdr_p, hdr_len);
913 if (page_to_free)
914 put_page(page_to_free);
915
916 return skb;
917}
918
919static void virtnet_rq_unmap(struct receive_queue *rq, void *buf, u32 len)
920{
921 struct virtnet_info *vi = rq->vq->vdev->priv;
922 struct page *page = virt_to_head_page(buf);
923 struct virtnet_rq_dma *dma;
924 void *head;
925 int offset;
926
927 BUG_ON(vi->big_packets && !vi->mergeable_rx_bufs);
928
929 head = page_address(page);
930
931 dma = head;
932
933 --dma->ref;
934
935 if (dma->need_sync && len) {
936 offset = buf - (head + sizeof(*dma));
937
938 virtqueue_dma_sync_single_range_for_cpu(rq->vq, dma->addr,
939 offset, len,
940 DMA_FROM_DEVICE);
941 }
942
943 if (dma->ref)
944 return;
945
946 virtqueue_dma_unmap_single_attrs(rq->vq, dma->addr, dma->len,
947 DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
948 put_page(page);
949}
950
951static void *virtnet_rq_get_buf(struct receive_queue *rq, u32 *len, void **ctx)
952{
953 struct virtnet_info *vi = rq->vq->vdev->priv;
954 void *buf;
955
956 BUG_ON(vi->big_packets && !vi->mergeable_rx_bufs);
957
958 buf = virtqueue_get_buf_ctx(rq->vq, len, ctx);
959 if (buf)
960 virtnet_rq_unmap(rq, buf, *len);
961
962 return buf;
963}
964
965static void virtnet_rq_init_one_sg(struct receive_queue *rq, void *buf, u32 len)
966{
967 struct virtnet_info *vi = rq->vq->vdev->priv;
968 struct virtnet_rq_dma *dma;
969 dma_addr_t addr;
970 u32 offset;
971 void *head;
972
973 BUG_ON(vi->big_packets && !vi->mergeable_rx_bufs);
974
975 head = page_address(rq->alloc_frag.page);
976
977 offset = buf - head;
978
979 dma = head;
980
981 addr = dma->addr - sizeof(*dma) + offset;
982
983 sg_init_table(rq->sg, 1);
984 sg_fill_dma(rq->sg, addr, len);
985}
986
987static void *virtnet_rq_alloc(struct receive_queue *rq, u32 size, gfp_t gfp)
988{
989 struct page_frag *alloc_frag = &rq->alloc_frag;
990 struct virtnet_info *vi = rq->vq->vdev->priv;
991 struct virtnet_rq_dma *dma;
992 void *buf, *head;
993 dma_addr_t addr;
994
995 BUG_ON(vi->big_packets && !vi->mergeable_rx_bufs);
996
997 head = page_address(alloc_frag->page);
998
999 dma = head;
1000
1001 /* new pages */
1002 if (!alloc_frag->offset) {
1003 if (rq->last_dma) {
1004 /* Now, the new page is allocated, the last dma
1005 * will not be used. So the dma can be unmapped
1006 * if the ref is 0.
1007 */
1008 virtnet_rq_unmap(rq, rq->last_dma, 0);
1009 rq->last_dma = NULL;
1010 }
1011
1012 dma->len = alloc_frag->size - sizeof(*dma);
1013
1014 addr = virtqueue_dma_map_single_attrs(rq->vq, dma + 1,
1015 dma->len, DMA_FROM_DEVICE, 0);
1016 if (virtqueue_dma_mapping_error(rq->vq, addr))
1017 return NULL;
1018
1019 dma->addr = addr;
1020 dma->need_sync = virtqueue_dma_need_sync(rq->vq, addr);
1021
1022 /* Add a reference to dma to prevent the entire dma from
1023 * being released during error handling. This reference
1024 * will be freed after the pages are no longer used.
1025 */
1026 get_page(alloc_frag->page);
1027 dma->ref = 1;
1028 alloc_frag->offset = sizeof(*dma);
1029
1030 rq->last_dma = dma;
1031 }
1032
1033 ++dma->ref;
1034
1035 buf = head + alloc_frag->offset;
1036
1037 get_page(alloc_frag->page);
1038 alloc_frag->offset += size;
1039
1040 return buf;
1041}
1042
1043static void virtnet_rq_unmap_free_buf(struct virtqueue *vq, void *buf)
1044{
1045 struct virtnet_info *vi = vq->vdev->priv;
1046 struct receive_queue *rq;
1047 int i = vq2rxq(vq);
1048
1049 rq = &vi->rq[i];
1050
1051 if (rq->xsk_pool) {
1052 xsk_buff_free((struct xdp_buff *)buf);
1053 return;
1054 }
1055
1056 if (!vi->big_packets || vi->mergeable_rx_bufs)
1057 virtnet_rq_unmap(rq, buf, 0);
1058
1059 virtnet_rq_free_buf(vi, rq, buf);
1060}
1061
1062static void free_old_xmit(struct send_queue *sq, struct netdev_queue *txq,
1063 bool in_napi)
1064{
1065 struct virtnet_sq_free_stats stats = {0};
1066
1067 virtnet_free_old_xmit(sq, txq, in_napi, &stats);
1068
1069 /* Avoid overhead when no packets have been processed
1070 * happens when called speculatively from start_xmit.
1071 */
1072 if (!stats.packets && !stats.napi_packets)
1073 return;
1074
1075 u64_stats_update_begin(&sq->stats.syncp);
1076 u64_stats_add(&sq->stats.bytes, stats.bytes + stats.napi_bytes);
1077 u64_stats_add(&sq->stats.packets, stats.packets + stats.napi_packets);
1078 u64_stats_update_end(&sq->stats.syncp);
1079}
1080
1081static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1082{
1083 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1084 return false;
1085 else if (q < vi->curr_queue_pairs)
1086 return true;
1087 else
1088 return false;
1089}
1090
1091static void check_sq_full_and_disable(struct virtnet_info *vi,
1092 struct net_device *dev,
1093 struct send_queue *sq)
1094{
1095 bool use_napi = sq->napi.weight;
1096 int qnum;
1097
1098 qnum = sq - vi->sq;
1099
1100 /* If running out of space, stop queue to avoid getting packets that we
1101 * are then unable to transmit.
1102 * An alternative would be to force queuing layer to requeue the skb by
1103 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1104 * returned in a normal path of operation: it means that driver is not
1105 * maintaining the TX queue stop/start state properly, and causes
1106 * the stack to do a non-trivial amount of useless work.
1107 * Since most packets only take 1 or 2 ring slots, stopping the queue
1108 * early means 16 slots are typically wasted.
1109 */
1110 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1111 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1112
1113 netif_tx_stop_queue(txq);
1114 u64_stats_update_begin(&sq->stats.syncp);
1115 u64_stats_inc(&sq->stats.stop);
1116 u64_stats_update_end(&sq->stats.syncp);
1117 if (use_napi) {
1118 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq)))
1119 virtqueue_napi_schedule(&sq->napi, sq->vq);
1120 } else if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1121 /* More just got used, free them then recheck. */
1122 free_old_xmit(sq, txq, false);
1123 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1124 netif_start_subqueue(dev, qnum);
1125 u64_stats_update_begin(&sq->stats.syncp);
1126 u64_stats_inc(&sq->stats.wake);
1127 u64_stats_update_end(&sq->stats.syncp);
1128 virtqueue_disable_cb(sq->vq);
1129 }
1130 }
1131 }
1132}
1133
1134static struct xdp_buff *buf_to_xdp(struct virtnet_info *vi,
1135 struct receive_queue *rq, void *buf, u32 len)
1136{
1137 struct xdp_buff *xdp;
1138 u32 bufsize;
1139
1140 xdp = (struct xdp_buff *)buf;
1141
1142 bufsize = xsk_pool_get_rx_frame_size(rq->xsk_pool) + vi->hdr_len;
1143
1144 if (unlikely(len > bufsize)) {
1145 pr_debug("%s: rx error: len %u exceeds truesize %u\n",
1146 vi->dev->name, len, bufsize);
1147 DEV_STATS_INC(vi->dev, rx_length_errors);
1148 xsk_buff_free(xdp);
1149 return NULL;
1150 }
1151
1152 xsk_buff_set_size(xdp, len);
1153 xsk_buff_dma_sync_for_cpu(xdp);
1154
1155 return xdp;
1156}
1157
1158static struct sk_buff *xsk_construct_skb(struct receive_queue *rq,
1159 struct xdp_buff *xdp)
1160{
1161 unsigned int metasize = xdp->data - xdp->data_meta;
1162 struct sk_buff *skb;
1163 unsigned int size;
1164
1165 size = xdp->data_end - xdp->data_hard_start;
1166 skb = napi_alloc_skb(&rq->napi, size);
1167 if (unlikely(!skb)) {
1168 xsk_buff_free(xdp);
1169 return NULL;
1170 }
1171
1172 skb_reserve(skb, xdp->data_meta - xdp->data_hard_start);
1173
1174 size = xdp->data_end - xdp->data_meta;
1175 memcpy(__skb_put(skb, size), xdp->data_meta, size);
1176
1177 if (metasize) {
1178 __skb_pull(skb, metasize);
1179 skb_metadata_set(skb, metasize);
1180 }
1181
1182 xsk_buff_free(xdp);
1183
1184 return skb;
1185}
1186
1187static struct sk_buff *virtnet_receive_xsk_small(struct net_device *dev, struct virtnet_info *vi,
1188 struct receive_queue *rq, struct xdp_buff *xdp,
1189 unsigned int *xdp_xmit,
1190 struct virtnet_rq_stats *stats)
1191{
1192 struct bpf_prog *prog;
1193 u32 ret;
1194
1195 ret = XDP_PASS;
1196 rcu_read_lock();
1197 prog = rcu_dereference(rq->xdp_prog);
1198 if (prog)
1199 ret = virtnet_xdp_handler(prog, xdp, dev, xdp_xmit, stats);
1200 rcu_read_unlock();
1201
1202 switch (ret) {
1203 case XDP_PASS:
1204 return xsk_construct_skb(rq, xdp);
1205
1206 case XDP_TX:
1207 case XDP_REDIRECT:
1208 return NULL;
1209
1210 default:
1211 /* drop packet */
1212 xsk_buff_free(xdp);
1213 u64_stats_inc(&stats->drops);
1214 return NULL;
1215 }
1216}
1217
1218static void xsk_drop_follow_bufs(struct net_device *dev,
1219 struct receive_queue *rq,
1220 u32 num_buf,
1221 struct virtnet_rq_stats *stats)
1222{
1223 struct xdp_buff *xdp;
1224 u32 len;
1225
1226 while (num_buf-- > 1) {
1227 xdp = virtqueue_get_buf(rq->vq, &len);
1228 if (unlikely(!xdp)) {
1229 pr_debug("%s: rx error: %d buffers missing\n",
1230 dev->name, num_buf);
1231 DEV_STATS_INC(dev, rx_length_errors);
1232 break;
1233 }
1234 u64_stats_add(&stats->bytes, len);
1235 xsk_buff_free(xdp);
1236 }
1237}
1238
1239static int xsk_append_merge_buffer(struct virtnet_info *vi,
1240 struct receive_queue *rq,
1241 struct sk_buff *head_skb,
1242 u32 num_buf,
1243 struct virtio_net_hdr_mrg_rxbuf *hdr,
1244 struct virtnet_rq_stats *stats)
1245{
1246 struct sk_buff *curr_skb;
1247 struct xdp_buff *xdp;
1248 u32 len, truesize;
1249 struct page *page;
1250 void *buf;
1251
1252 curr_skb = head_skb;
1253
1254 while (--num_buf) {
1255 buf = virtqueue_get_buf(rq->vq, &len);
1256 if (unlikely(!buf)) {
1257 pr_debug("%s: rx error: %d buffers out of %d missing\n",
1258 vi->dev->name, num_buf,
1259 virtio16_to_cpu(vi->vdev,
1260 hdr->num_buffers));
1261 DEV_STATS_INC(vi->dev, rx_length_errors);
1262 return -EINVAL;
1263 }
1264
1265 u64_stats_add(&stats->bytes, len);
1266
1267 xdp = buf_to_xdp(vi, rq, buf, len);
1268 if (!xdp)
1269 goto err;
1270
1271 buf = napi_alloc_frag(len);
1272 if (!buf) {
1273 xsk_buff_free(xdp);
1274 goto err;
1275 }
1276
1277 memcpy(buf, xdp->data - vi->hdr_len, len);
1278
1279 xsk_buff_free(xdp);
1280
1281 page = virt_to_page(buf);
1282
1283 truesize = len;
1284
1285 curr_skb = virtnet_skb_append_frag(head_skb, curr_skb, page,
1286 buf, len, truesize);
1287 if (!curr_skb) {
1288 put_page(page);
1289 goto err;
1290 }
1291 }
1292
1293 return 0;
1294
1295err:
1296 xsk_drop_follow_bufs(vi->dev, rq, num_buf, stats);
1297 return -EINVAL;
1298}
1299
1300static struct sk_buff *virtnet_receive_xsk_merge(struct net_device *dev, struct virtnet_info *vi,
1301 struct receive_queue *rq, struct xdp_buff *xdp,
1302 unsigned int *xdp_xmit,
1303 struct virtnet_rq_stats *stats)
1304{
1305 struct virtio_net_hdr_mrg_rxbuf *hdr;
1306 struct bpf_prog *prog;
1307 struct sk_buff *skb;
1308 u32 ret, num_buf;
1309
1310 hdr = xdp->data - vi->hdr_len;
1311 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
1312
1313 ret = XDP_PASS;
1314 rcu_read_lock();
1315 prog = rcu_dereference(rq->xdp_prog);
1316 /* TODO: support multi buffer. */
1317 if (prog && num_buf == 1)
1318 ret = virtnet_xdp_handler(prog, xdp, dev, xdp_xmit, stats);
1319 rcu_read_unlock();
1320
1321 switch (ret) {
1322 case XDP_PASS:
1323 skb = xsk_construct_skb(rq, xdp);
1324 if (!skb)
1325 goto drop_bufs;
1326
1327 if (xsk_append_merge_buffer(vi, rq, skb, num_buf, hdr, stats)) {
1328 dev_kfree_skb(skb);
1329 goto drop;
1330 }
1331
1332 return skb;
1333
1334 case XDP_TX:
1335 case XDP_REDIRECT:
1336 return NULL;
1337
1338 default:
1339 /* drop packet */
1340 xsk_buff_free(xdp);
1341 }
1342
1343drop_bufs:
1344 xsk_drop_follow_bufs(dev, rq, num_buf, stats);
1345
1346drop:
1347 u64_stats_inc(&stats->drops);
1348 return NULL;
1349}
1350
1351static void virtnet_receive_xsk_buf(struct virtnet_info *vi, struct receive_queue *rq,
1352 void *buf, u32 len,
1353 unsigned int *xdp_xmit,
1354 struct virtnet_rq_stats *stats)
1355{
1356 struct net_device *dev = vi->dev;
1357 struct sk_buff *skb = NULL;
1358 struct xdp_buff *xdp;
1359 u8 flags;
1360
1361 len -= vi->hdr_len;
1362
1363 u64_stats_add(&stats->bytes, len);
1364
1365 xdp = buf_to_xdp(vi, rq, buf, len);
1366 if (!xdp)
1367 return;
1368
1369 if (unlikely(len < ETH_HLEN)) {
1370 pr_debug("%s: short packet %i\n", dev->name, len);
1371 DEV_STATS_INC(dev, rx_length_errors);
1372 xsk_buff_free(xdp);
1373 return;
1374 }
1375
1376 flags = ((struct virtio_net_common_hdr *)(xdp->data - vi->hdr_len))->hdr.flags;
1377
1378 if (!vi->mergeable_rx_bufs)
1379 skb = virtnet_receive_xsk_small(dev, vi, rq, xdp, xdp_xmit, stats);
1380 else
1381 skb = virtnet_receive_xsk_merge(dev, vi, rq, xdp, xdp_xmit, stats);
1382
1383 if (skb)
1384 virtnet_receive_done(vi, rq, skb, flags);
1385}
1386
1387static int virtnet_add_recvbuf_xsk(struct virtnet_info *vi, struct receive_queue *rq,
1388 struct xsk_buff_pool *pool, gfp_t gfp)
1389{
1390 struct xdp_buff **xsk_buffs;
1391 dma_addr_t addr;
1392 int err = 0;
1393 u32 len, i;
1394 int num;
1395
1396 xsk_buffs = rq->xsk_buffs;
1397
1398 num = xsk_buff_alloc_batch(pool, xsk_buffs, rq->vq->num_free);
1399 if (!num)
1400 return -ENOMEM;
1401
1402 len = xsk_pool_get_rx_frame_size(pool) + vi->hdr_len;
1403
1404 for (i = 0; i < num; ++i) {
1405 /* Use the part of XDP_PACKET_HEADROOM as the virtnet hdr space.
1406 * We assume XDP_PACKET_HEADROOM is larger than hdr->len.
1407 * (see function virtnet_xsk_pool_enable)
1408 */
1409 addr = xsk_buff_xdp_get_dma(xsk_buffs[i]) - vi->hdr_len;
1410
1411 sg_init_table(rq->sg, 1);
1412 sg_fill_dma(rq->sg, addr, len);
1413
1414 err = virtqueue_add_inbuf_premapped(rq->vq, rq->sg, 1,
1415 xsk_buffs[i], NULL, gfp);
1416 if (err)
1417 goto err;
1418 }
1419
1420 return num;
1421
1422err:
1423 for (; i < num; ++i)
1424 xsk_buff_free(xsk_buffs[i]);
1425
1426 return err;
1427}
1428
1429static void *virtnet_xsk_to_ptr(u32 len)
1430{
1431 unsigned long p;
1432
1433 p = len << VIRTIO_XSK_FLAG_OFFSET;
1434
1435 return virtnet_xmit_ptr_pack((void *)p, VIRTNET_XMIT_TYPE_XSK);
1436}
1437
1438static int virtnet_xsk_xmit_one(struct send_queue *sq,
1439 struct xsk_buff_pool *pool,
1440 struct xdp_desc *desc)
1441{
1442 struct virtnet_info *vi;
1443 dma_addr_t addr;
1444
1445 vi = sq->vq->vdev->priv;
1446
1447 addr = xsk_buff_raw_get_dma(pool, desc->addr);
1448 xsk_buff_raw_dma_sync_for_device(pool, addr, desc->len);
1449
1450 sg_init_table(sq->sg, 2);
1451 sg_fill_dma(sq->sg, sq->xsk_hdr_dma_addr, vi->hdr_len);
1452 sg_fill_dma(sq->sg + 1, addr, desc->len);
1453
1454 return virtqueue_add_outbuf_premapped(sq->vq, sq->sg, 2,
1455 virtnet_xsk_to_ptr(desc->len),
1456 GFP_ATOMIC);
1457}
1458
1459static int virtnet_xsk_xmit_batch(struct send_queue *sq,
1460 struct xsk_buff_pool *pool,
1461 unsigned int budget,
1462 u64 *kicks)
1463{
1464 struct xdp_desc *descs = pool->tx_descs;
1465 bool kick = false;
1466 u32 nb_pkts, i;
1467 int err;
1468
1469 budget = min_t(u32, budget, sq->vq->num_free);
1470
1471 nb_pkts = xsk_tx_peek_release_desc_batch(pool, budget);
1472 if (!nb_pkts)
1473 return 0;
1474
1475 for (i = 0; i < nb_pkts; i++) {
1476 err = virtnet_xsk_xmit_one(sq, pool, &descs[i]);
1477 if (unlikely(err)) {
1478 xsk_tx_completed(sq->xsk_pool, nb_pkts - i);
1479 break;
1480 }
1481
1482 kick = true;
1483 }
1484
1485 if (kick && virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
1486 (*kicks)++;
1487
1488 return i;
1489}
1490
1491static bool virtnet_xsk_xmit(struct send_queue *sq, struct xsk_buff_pool *pool,
1492 int budget)
1493{
1494 struct virtnet_info *vi = sq->vq->vdev->priv;
1495 struct virtnet_sq_free_stats stats = {};
1496 struct net_device *dev = vi->dev;
1497 u64 kicks = 0;
1498 int sent;
1499
1500 /* Avoid to wakeup napi meanless, so call __free_old_xmit instead of
1501 * free_old_xmit().
1502 */
1503 __free_old_xmit(sq, netdev_get_tx_queue(dev, sq - vi->sq), true, &stats);
1504
1505 if (stats.xsk)
1506 xsk_tx_completed(sq->xsk_pool, stats.xsk);
1507
1508 sent = virtnet_xsk_xmit_batch(sq, pool, budget, &kicks);
1509
1510 if (!is_xdp_raw_buffer_queue(vi, sq - vi->sq))
1511 check_sq_full_and_disable(vi, vi->dev, sq);
1512
1513 if (sent) {
1514 struct netdev_queue *txq;
1515
1516 txq = netdev_get_tx_queue(vi->dev, sq - vi->sq);
1517 txq_trans_cond_update(txq);
1518 }
1519
1520 u64_stats_update_begin(&sq->stats.syncp);
1521 u64_stats_add(&sq->stats.packets, stats.packets);
1522 u64_stats_add(&sq->stats.bytes, stats.bytes);
1523 u64_stats_add(&sq->stats.kicks, kicks);
1524 u64_stats_add(&sq->stats.xdp_tx, sent);
1525 u64_stats_update_end(&sq->stats.syncp);
1526
1527 if (xsk_uses_need_wakeup(pool))
1528 xsk_set_tx_need_wakeup(pool);
1529
1530 return sent;
1531}
1532
1533static void xsk_wakeup(struct send_queue *sq)
1534{
1535 if (napi_if_scheduled_mark_missed(&sq->napi))
1536 return;
1537
1538 local_bh_disable();
1539 virtqueue_napi_schedule(&sq->napi, sq->vq);
1540 local_bh_enable();
1541}
1542
1543static int virtnet_xsk_wakeup(struct net_device *dev, u32 qid, u32 flag)
1544{
1545 struct virtnet_info *vi = netdev_priv(dev);
1546 struct send_queue *sq;
1547
1548 if (!netif_running(dev))
1549 return -ENETDOWN;
1550
1551 if (qid >= vi->curr_queue_pairs)
1552 return -EINVAL;
1553
1554 sq = &vi->sq[qid];
1555
1556 xsk_wakeup(sq);
1557 return 0;
1558}
1559
1560static void virtnet_xsk_completed(struct send_queue *sq, int num)
1561{
1562 xsk_tx_completed(sq->xsk_pool, num);
1563
1564 /* If this is called by rx poll, start_xmit and xdp xmit we should
1565 * wakeup the tx napi to consume the xsk tx queue, because the tx
1566 * interrupt may not be triggered.
1567 */
1568 xsk_wakeup(sq);
1569}
1570
1571static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
1572 struct send_queue *sq,
1573 struct xdp_frame *xdpf)
1574{
1575 struct virtio_net_hdr_mrg_rxbuf *hdr;
1576 struct skb_shared_info *shinfo;
1577 u8 nr_frags = 0;
1578 int err, i;
1579
1580 if (unlikely(xdpf->headroom < vi->hdr_len))
1581 return -EOVERFLOW;
1582
1583 if (unlikely(xdp_frame_has_frags(xdpf))) {
1584 shinfo = xdp_get_shared_info_from_frame(xdpf);
1585 nr_frags = shinfo->nr_frags;
1586 }
1587
1588 /* In wrapping function virtnet_xdp_xmit(), we need to free
1589 * up the pending old buffers, where we need to calculate the
1590 * position of skb_shared_info in xdp_get_frame_len() and
1591 * xdp_return_frame(), which will involve to xdpf->data and
1592 * xdpf->headroom. Therefore, we need to update the value of
1593 * headroom synchronously here.
1594 */
1595 xdpf->headroom -= vi->hdr_len;
1596 xdpf->data -= vi->hdr_len;
1597 /* Zero header and leave csum up to XDP layers */
1598 hdr = xdpf->data;
1599 memset(hdr, 0, vi->hdr_len);
1600 xdpf->len += vi->hdr_len;
1601
1602 sg_init_table(sq->sg, nr_frags + 1);
1603 sg_set_buf(sq->sg, xdpf->data, xdpf->len);
1604 for (i = 0; i < nr_frags; i++) {
1605 skb_frag_t *frag = &shinfo->frags[i];
1606
1607 sg_set_page(&sq->sg[i + 1], skb_frag_page(frag),
1608 skb_frag_size(frag), skb_frag_off(frag));
1609 }
1610
1611 err = virtnet_add_outbuf(sq, nr_frags + 1, xdpf, VIRTNET_XMIT_TYPE_XDP);
1612 if (unlikely(err))
1613 return -ENOSPC; /* Caller handle free/refcnt */
1614
1615 return 0;
1616}
1617
1618/* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
1619 * the current cpu, so it does not need to be locked.
1620 *
1621 * Here we use marco instead of inline functions because we have to deal with
1622 * three issues at the same time: 1. the choice of sq. 2. judge and execute the
1623 * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
1624 * functions to perfectly solve these three problems at the same time.
1625 */
1626#define virtnet_xdp_get_sq(vi) ({ \
1627 int cpu = smp_processor_id(); \
1628 struct netdev_queue *txq; \
1629 typeof(vi) v = (vi); \
1630 unsigned int qp; \
1631 \
1632 if (v->curr_queue_pairs > nr_cpu_ids) { \
1633 qp = v->curr_queue_pairs - v->xdp_queue_pairs; \
1634 qp += cpu; \
1635 txq = netdev_get_tx_queue(v->dev, qp); \
1636 __netif_tx_acquire(txq); \
1637 } else { \
1638 qp = cpu % v->curr_queue_pairs; \
1639 txq = netdev_get_tx_queue(v->dev, qp); \
1640 __netif_tx_lock(txq, cpu); \
1641 } \
1642 v->sq + qp; \
1643})
1644
1645#define virtnet_xdp_put_sq(vi, q) { \
1646 struct netdev_queue *txq; \
1647 typeof(vi) v = (vi); \
1648 \
1649 txq = netdev_get_tx_queue(v->dev, (q) - v->sq); \
1650 if (v->curr_queue_pairs > nr_cpu_ids) \
1651 __netif_tx_release(txq); \
1652 else \
1653 __netif_tx_unlock(txq); \
1654}
1655
1656static int virtnet_xdp_xmit(struct net_device *dev,
1657 int n, struct xdp_frame **frames, u32 flags)
1658{
1659 struct virtnet_info *vi = netdev_priv(dev);
1660 struct virtnet_sq_free_stats stats = {0};
1661 struct receive_queue *rq = vi->rq;
1662 struct bpf_prog *xdp_prog;
1663 struct send_queue *sq;
1664 int nxmit = 0;
1665 int kicks = 0;
1666 int ret;
1667 int i;
1668
1669 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
1670 * indicate XDP resources have been successfully allocated.
1671 */
1672 xdp_prog = rcu_access_pointer(rq->xdp_prog);
1673 if (!xdp_prog)
1674 return -ENXIO;
1675
1676 sq = virtnet_xdp_get_sq(vi);
1677
1678 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
1679 ret = -EINVAL;
1680 goto out;
1681 }
1682
1683 /* Free up any pending old buffers before queueing new ones. */
1684 virtnet_free_old_xmit(sq, netdev_get_tx_queue(dev, sq - vi->sq),
1685 false, &stats);
1686
1687 for (i = 0; i < n; i++) {
1688 struct xdp_frame *xdpf = frames[i];
1689
1690 if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
1691 break;
1692 nxmit++;
1693 }
1694 ret = nxmit;
1695
1696 if (!is_xdp_raw_buffer_queue(vi, sq - vi->sq))
1697 check_sq_full_and_disable(vi, dev, sq);
1698
1699 if (flags & XDP_XMIT_FLUSH) {
1700 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
1701 kicks = 1;
1702 }
1703out:
1704 u64_stats_update_begin(&sq->stats.syncp);
1705 u64_stats_add(&sq->stats.bytes, stats.bytes);
1706 u64_stats_add(&sq->stats.packets, stats.packets);
1707 u64_stats_add(&sq->stats.xdp_tx, n);
1708 u64_stats_add(&sq->stats.xdp_tx_drops, n - nxmit);
1709 u64_stats_add(&sq->stats.kicks, kicks);
1710 u64_stats_update_end(&sq->stats.syncp);
1711
1712 virtnet_xdp_put_sq(vi, sq);
1713 return ret;
1714}
1715
1716static void put_xdp_frags(struct xdp_buff *xdp)
1717{
1718 struct skb_shared_info *shinfo;
1719 struct page *xdp_page;
1720 int i;
1721
1722 if (xdp_buff_has_frags(xdp)) {
1723 shinfo = xdp_get_shared_info_from_buff(xdp);
1724 for (i = 0; i < shinfo->nr_frags; i++) {
1725 xdp_page = skb_frag_page(&shinfo->frags[i]);
1726 put_page(xdp_page);
1727 }
1728 }
1729}
1730
1731static int virtnet_xdp_handler(struct bpf_prog *xdp_prog, struct xdp_buff *xdp,
1732 struct net_device *dev,
1733 unsigned int *xdp_xmit,
1734 struct virtnet_rq_stats *stats)
1735{
1736 struct xdp_frame *xdpf;
1737 int err;
1738 u32 act;
1739
1740 act = bpf_prog_run_xdp(xdp_prog, xdp);
1741 u64_stats_inc(&stats->xdp_packets);
1742
1743 switch (act) {
1744 case XDP_PASS:
1745 return act;
1746
1747 case XDP_TX:
1748 u64_stats_inc(&stats->xdp_tx);
1749 xdpf = xdp_convert_buff_to_frame(xdp);
1750 if (unlikely(!xdpf)) {
1751 netdev_dbg(dev, "convert buff to frame failed for xdp\n");
1752 return XDP_DROP;
1753 }
1754
1755 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
1756 if (unlikely(!err)) {
1757 xdp_return_frame_rx_napi(xdpf);
1758 } else if (unlikely(err < 0)) {
1759 trace_xdp_exception(dev, xdp_prog, act);
1760 return XDP_DROP;
1761 }
1762 *xdp_xmit |= VIRTIO_XDP_TX;
1763 return act;
1764
1765 case XDP_REDIRECT:
1766 u64_stats_inc(&stats->xdp_redirects);
1767 err = xdp_do_redirect(dev, xdp, xdp_prog);
1768 if (err)
1769 return XDP_DROP;
1770
1771 *xdp_xmit |= VIRTIO_XDP_REDIR;
1772 return act;
1773
1774 default:
1775 bpf_warn_invalid_xdp_action(dev, xdp_prog, act);
1776 fallthrough;
1777 case XDP_ABORTED:
1778 trace_xdp_exception(dev, xdp_prog, act);
1779 fallthrough;
1780 case XDP_DROP:
1781 return XDP_DROP;
1782 }
1783}
1784
1785static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
1786{
1787 return vi->xdp_enabled ? XDP_PACKET_HEADROOM : 0;
1788}
1789
1790/* We copy the packet for XDP in the following cases:
1791 *
1792 * 1) Packet is scattered across multiple rx buffers.
1793 * 2) Headroom space is insufficient.
1794 *
1795 * This is inefficient but it's a temporary condition that
1796 * we hit right after XDP is enabled and until queue is refilled
1797 * with large buffers with sufficient headroom - so it should affect
1798 * at most queue size packets.
1799 * Afterwards, the conditions to enable
1800 * XDP should preclude the underlying device from sending packets
1801 * across multiple buffers (num_buf > 1), and we make sure buffers
1802 * have enough headroom.
1803 */
1804static struct page *xdp_linearize_page(struct receive_queue *rq,
1805 int *num_buf,
1806 struct page *p,
1807 int offset,
1808 int page_off,
1809 unsigned int *len)
1810{
1811 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1812 struct page *page;
1813
1814 if (page_off + *len + tailroom > PAGE_SIZE)
1815 return NULL;
1816
1817 page = alloc_page(GFP_ATOMIC);
1818 if (!page)
1819 return NULL;
1820
1821 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
1822 page_off += *len;
1823
1824 while (--*num_buf) {
1825 unsigned int buflen;
1826 void *buf;
1827 int off;
1828
1829 buf = virtnet_rq_get_buf(rq, &buflen, NULL);
1830 if (unlikely(!buf))
1831 goto err_buf;
1832
1833 p = virt_to_head_page(buf);
1834 off = buf - page_address(p);
1835
1836 /* guard against a misconfigured or uncooperative backend that
1837 * is sending packet larger than the MTU.
1838 */
1839 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
1840 put_page(p);
1841 goto err_buf;
1842 }
1843
1844 memcpy(page_address(page) + page_off,
1845 page_address(p) + off, buflen);
1846 page_off += buflen;
1847 put_page(p);
1848 }
1849
1850 /* Headroom does not contribute to packet length */
1851 *len = page_off - XDP_PACKET_HEADROOM;
1852 return page;
1853err_buf:
1854 __free_pages(page, 0);
1855 return NULL;
1856}
1857
1858static struct sk_buff *receive_small_build_skb(struct virtnet_info *vi,
1859 unsigned int xdp_headroom,
1860 void *buf,
1861 unsigned int len)
1862{
1863 unsigned int header_offset;
1864 unsigned int headroom;
1865 unsigned int buflen;
1866 struct sk_buff *skb;
1867
1868 header_offset = VIRTNET_RX_PAD + xdp_headroom;
1869 headroom = vi->hdr_len + header_offset;
1870 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
1871 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1872
1873 skb = virtnet_build_skb(buf, buflen, headroom, len);
1874 if (unlikely(!skb))
1875 return NULL;
1876
1877 buf += header_offset;
1878 memcpy(skb_vnet_common_hdr(skb), buf, vi->hdr_len);
1879
1880 return skb;
1881}
1882
1883static struct sk_buff *receive_small_xdp(struct net_device *dev,
1884 struct virtnet_info *vi,
1885 struct receive_queue *rq,
1886 struct bpf_prog *xdp_prog,
1887 void *buf,
1888 unsigned int xdp_headroom,
1889 unsigned int len,
1890 unsigned int *xdp_xmit,
1891 struct virtnet_rq_stats *stats)
1892{
1893 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
1894 unsigned int headroom = vi->hdr_len + header_offset;
1895 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
1896 struct page *page = virt_to_head_page(buf);
1897 struct page *xdp_page;
1898 unsigned int buflen;
1899 struct xdp_buff xdp;
1900 struct sk_buff *skb;
1901 unsigned int metasize = 0;
1902 u32 act;
1903
1904 if (unlikely(hdr->hdr.gso_type))
1905 goto err_xdp;
1906
1907 /* Partially checksummed packets must be dropped. */
1908 if (unlikely(hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM))
1909 goto err_xdp;
1910
1911 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
1912 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1913
1914 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
1915 int offset = buf - page_address(page) + header_offset;
1916 unsigned int tlen = len + vi->hdr_len;
1917 int num_buf = 1;
1918
1919 xdp_headroom = virtnet_get_headroom(vi);
1920 header_offset = VIRTNET_RX_PAD + xdp_headroom;
1921 headroom = vi->hdr_len + header_offset;
1922 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
1923 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1924 xdp_page = xdp_linearize_page(rq, &num_buf, page,
1925 offset, header_offset,
1926 &tlen);
1927 if (!xdp_page)
1928 goto err_xdp;
1929
1930 buf = page_address(xdp_page);
1931 put_page(page);
1932 page = xdp_page;
1933 }
1934
1935 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
1936 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
1937 xdp_headroom, len, true);
1938
1939 act = virtnet_xdp_handler(xdp_prog, &xdp, dev, xdp_xmit, stats);
1940
1941 switch (act) {
1942 case XDP_PASS:
1943 /* Recalculate length in case bpf program changed it */
1944 len = xdp.data_end - xdp.data;
1945 metasize = xdp.data - xdp.data_meta;
1946 break;
1947
1948 case XDP_TX:
1949 case XDP_REDIRECT:
1950 goto xdp_xmit;
1951
1952 default:
1953 goto err_xdp;
1954 }
1955
1956 skb = virtnet_build_skb(buf, buflen, xdp.data - buf, len);
1957 if (unlikely(!skb))
1958 goto err;
1959
1960 if (metasize)
1961 skb_metadata_set(skb, metasize);
1962
1963 return skb;
1964
1965err_xdp:
1966 u64_stats_inc(&stats->xdp_drops);
1967err:
1968 u64_stats_inc(&stats->drops);
1969 put_page(page);
1970xdp_xmit:
1971 return NULL;
1972}
1973
1974static struct sk_buff *receive_small(struct net_device *dev,
1975 struct virtnet_info *vi,
1976 struct receive_queue *rq,
1977 void *buf, void *ctx,
1978 unsigned int len,
1979 unsigned int *xdp_xmit,
1980 struct virtnet_rq_stats *stats)
1981{
1982 unsigned int xdp_headroom = (unsigned long)ctx;
1983 struct page *page = virt_to_head_page(buf);
1984 struct sk_buff *skb;
1985
1986 /* We passed the address of virtnet header to virtio-core,
1987 * so truncate the padding.
1988 */
1989 buf -= VIRTNET_RX_PAD + xdp_headroom;
1990
1991 len -= vi->hdr_len;
1992 u64_stats_add(&stats->bytes, len);
1993
1994 if (unlikely(len > GOOD_PACKET_LEN)) {
1995 pr_debug("%s: rx error: len %u exceeds max size %d\n",
1996 dev->name, len, GOOD_PACKET_LEN);
1997 DEV_STATS_INC(dev, rx_length_errors);
1998 goto err;
1999 }
2000
2001 if (unlikely(vi->xdp_enabled)) {
2002 struct bpf_prog *xdp_prog;
2003
2004 rcu_read_lock();
2005 xdp_prog = rcu_dereference(rq->xdp_prog);
2006 if (xdp_prog) {
2007 skb = receive_small_xdp(dev, vi, rq, xdp_prog, buf,
2008 xdp_headroom, len, xdp_xmit,
2009 stats);
2010 rcu_read_unlock();
2011 return skb;
2012 }
2013 rcu_read_unlock();
2014 }
2015
2016 skb = receive_small_build_skb(vi, xdp_headroom, buf, len);
2017 if (likely(skb))
2018 return skb;
2019
2020err:
2021 u64_stats_inc(&stats->drops);
2022 put_page(page);
2023 return NULL;
2024}
2025
2026static struct sk_buff *receive_big(struct net_device *dev,
2027 struct virtnet_info *vi,
2028 struct receive_queue *rq,
2029 void *buf,
2030 unsigned int len,
2031 struct virtnet_rq_stats *stats)
2032{
2033 struct page *page = buf;
2034 struct sk_buff *skb =
2035 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, 0);
2036
2037 u64_stats_add(&stats->bytes, len - vi->hdr_len);
2038 if (unlikely(!skb))
2039 goto err;
2040
2041 return skb;
2042
2043err:
2044 u64_stats_inc(&stats->drops);
2045 give_pages(rq, page);
2046 return NULL;
2047}
2048
2049static void mergeable_buf_free(struct receive_queue *rq, int num_buf,
2050 struct net_device *dev,
2051 struct virtnet_rq_stats *stats)
2052{
2053 struct page *page;
2054 void *buf;
2055 int len;
2056
2057 while (num_buf-- > 1) {
2058 buf = virtnet_rq_get_buf(rq, &len, NULL);
2059 if (unlikely(!buf)) {
2060 pr_debug("%s: rx error: %d buffers missing\n",
2061 dev->name, num_buf);
2062 DEV_STATS_INC(dev, rx_length_errors);
2063 break;
2064 }
2065 u64_stats_add(&stats->bytes, len);
2066 page = virt_to_head_page(buf);
2067 put_page(page);
2068 }
2069}
2070
2071/* Why not use xdp_build_skb_from_frame() ?
2072 * XDP core assumes that xdp frags are PAGE_SIZE in length, while in
2073 * virtio-net there are 2 points that do not match its requirements:
2074 * 1. The size of the prefilled buffer is not fixed before xdp is set.
2075 * 2. xdp_build_skb_from_frame() does more checks that we don't need,
2076 * like eth_type_trans() (which virtio-net does in receive_buf()).
2077 */
2078static struct sk_buff *build_skb_from_xdp_buff(struct net_device *dev,
2079 struct virtnet_info *vi,
2080 struct xdp_buff *xdp,
2081 unsigned int xdp_frags_truesz)
2082{
2083 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
2084 unsigned int headroom, data_len;
2085 struct sk_buff *skb;
2086 int metasize;
2087 u8 nr_frags;
2088
2089 if (unlikely(xdp->data_end > xdp_data_hard_end(xdp))) {
2090 pr_debug("Error building skb as missing reserved tailroom for xdp");
2091 return NULL;
2092 }
2093
2094 if (unlikely(xdp_buff_has_frags(xdp)))
2095 nr_frags = sinfo->nr_frags;
2096
2097 skb = build_skb(xdp->data_hard_start, xdp->frame_sz);
2098 if (unlikely(!skb))
2099 return NULL;
2100
2101 headroom = xdp->data - xdp->data_hard_start;
2102 data_len = xdp->data_end - xdp->data;
2103 skb_reserve(skb, headroom);
2104 __skb_put(skb, data_len);
2105
2106 metasize = xdp->data - xdp->data_meta;
2107 metasize = metasize > 0 ? metasize : 0;
2108 if (metasize)
2109 skb_metadata_set(skb, metasize);
2110
2111 if (unlikely(xdp_buff_has_frags(xdp)))
2112 xdp_update_skb_shared_info(skb, nr_frags,
2113 sinfo->xdp_frags_size,
2114 xdp_frags_truesz,
2115 xdp_buff_is_frag_pfmemalloc(xdp));
2116
2117 return skb;
2118}
2119
2120/* TODO: build xdp in big mode */
2121static int virtnet_build_xdp_buff_mrg(struct net_device *dev,
2122 struct virtnet_info *vi,
2123 struct receive_queue *rq,
2124 struct xdp_buff *xdp,
2125 void *buf,
2126 unsigned int len,
2127 unsigned int frame_sz,
2128 int *num_buf,
2129 unsigned int *xdp_frags_truesize,
2130 struct virtnet_rq_stats *stats)
2131{
2132 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
2133 unsigned int headroom, tailroom, room;
2134 unsigned int truesize, cur_frag_size;
2135 struct skb_shared_info *shinfo;
2136 unsigned int xdp_frags_truesz = 0;
2137 struct page *page;
2138 skb_frag_t *frag;
2139 int offset;
2140 void *ctx;
2141
2142 xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq);
2143 xdp_prepare_buff(xdp, buf - XDP_PACKET_HEADROOM,
2144 XDP_PACKET_HEADROOM + vi->hdr_len, len - vi->hdr_len, true);
2145
2146 if (!*num_buf)
2147 return 0;
2148
2149 if (*num_buf > 1) {
2150 /* If we want to build multi-buffer xdp, we need
2151 * to specify that the flags of xdp_buff have the
2152 * XDP_FLAGS_HAS_FRAG bit.
2153 */
2154 if (!xdp_buff_has_frags(xdp))
2155 xdp_buff_set_frags_flag(xdp);
2156
2157 shinfo = xdp_get_shared_info_from_buff(xdp);
2158 shinfo->nr_frags = 0;
2159 shinfo->xdp_frags_size = 0;
2160 }
2161
2162 if (*num_buf > MAX_SKB_FRAGS + 1)
2163 return -EINVAL;
2164
2165 while (--*num_buf > 0) {
2166 buf = virtnet_rq_get_buf(rq, &len, &ctx);
2167 if (unlikely(!buf)) {
2168 pr_debug("%s: rx error: %d buffers out of %d missing\n",
2169 dev->name, *num_buf,
2170 virtio16_to_cpu(vi->vdev, hdr->num_buffers));
2171 DEV_STATS_INC(dev, rx_length_errors);
2172 goto err;
2173 }
2174
2175 u64_stats_add(&stats->bytes, len);
2176 page = virt_to_head_page(buf);
2177 offset = buf - page_address(page);
2178
2179 truesize = mergeable_ctx_to_truesize(ctx);
2180 headroom = mergeable_ctx_to_headroom(ctx);
2181 tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2182 room = SKB_DATA_ALIGN(headroom + tailroom);
2183
2184 cur_frag_size = truesize;
2185 xdp_frags_truesz += cur_frag_size;
2186 if (unlikely(len > truesize - room || cur_frag_size > PAGE_SIZE)) {
2187 put_page(page);
2188 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
2189 dev->name, len, (unsigned long)(truesize - room));
2190 DEV_STATS_INC(dev, rx_length_errors);
2191 goto err;
2192 }
2193
2194 frag = &shinfo->frags[shinfo->nr_frags++];
2195 skb_frag_fill_page_desc(frag, page, offset, len);
2196 if (page_is_pfmemalloc(page))
2197 xdp_buff_set_frag_pfmemalloc(xdp);
2198
2199 shinfo->xdp_frags_size += len;
2200 }
2201
2202 *xdp_frags_truesize = xdp_frags_truesz;
2203 return 0;
2204
2205err:
2206 put_xdp_frags(xdp);
2207 return -EINVAL;
2208}
2209
2210static void *mergeable_xdp_get_buf(struct virtnet_info *vi,
2211 struct receive_queue *rq,
2212 struct bpf_prog *xdp_prog,
2213 void *ctx,
2214 unsigned int *frame_sz,
2215 int *num_buf,
2216 struct page **page,
2217 int offset,
2218 unsigned int *len,
2219 struct virtio_net_hdr_mrg_rxbuf *hdr)
2220{
2221 unsigned int truesize = mergeable_ctx_to_truesize(ctx);
2222 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
2223 struct page *xdp_page;
2224 unsigned int xdp_room;
2225
2226 /* Transient failure which in theory could occur if
2227 * in-flight packets from before XDP was enabled reach
2228 * the receive path after XDP is loaded.
2229 */
2230 if (unlikely(hdr->hdr.gso_type))
2231 return NULL;
2232
2233 /* Partially checksummed packets must be dropped. */
2234 if (unlikely(hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM))
2235 return NULL;
2236
2237 /* Now XDP core assumes frag size is PAGE_SIZE, but buffers
2238 * with headroom may add hole in truesize, which
2239 * make their length exceed PAGE_SIZE. So we disabled the
2240 * hole mechanism for xdp. See add_recvbuf_mergeable().
2241 */
2242 *frame_sz = truesize;
2243
2244 if (likely(headroom >= virtnet_get_headroom(vi) &&
2245 (*num_buf == 1 || xdp_prog->aux->xdp_has_frags))) {
2246 return page_address(*page) + offset;
2247 }
2248
2249 /* This happens when headroom is not enough because
2250 * of the buffer was prefilled before XDP is set.
2251 * This should only happen for the first several packets.
2252 * In fact, vq reset can be used here to help us clean up
2253 * the prefilled buffers, but many existing devices do not
2254 * support it, and we don't want to bother users who are
2255 * using xdp normally.
2256 */
2257 if (!xdp_prog->aux->xdp_has_frags) {
2258 /* linearize data for XDP */
2259 xdp_page = xdp_linearize_page(rq, num_buf,
2260 *page, offset,
2261 XDP_PACKET_HEADROOM,
2262 len);
2263 if (!xdp_page)
2264 return NULL;
2265 } else {
2266 xdp_room = SKB_DATA_ALIGN(XDP_PACKET_HEADROOM +
2267 sizeof(struct skb_shared_info));
2268 if (*len + xdp_room > PAGE_SIZE)
2269 return NULL;
2270
2271 xdp_page = alloc_page(GFP_ATOMIC);
2272 if (!xdp_page)
2273 return NULL;
2274
2275 memcpy(page_address(xdp_page) + XDP_PACKET_HEADROOM,
2276 page_address(*page) + offset, *len);
2277 }
2278
2279 *frame_sz = PAGE_SIZE;
2280
2281 put_page(*page);
2282
2283 *page = xdp_page;
2284
2285 return page_address(*page) + XDP_PACKET_HEADROOM;
2286}
2287
2288static struct sk_buff *receive_mergeable_xdp(struct net_device *dev,
2289 struct virtnet_info *vi,
2290 struct receive_queue *rq,
2291 struct bpf_prog *xdp_prog,
2292 void *buf,
2293 void *ctx,
2294 unsigned int len,
2295 unsigned int *xdp_xmit,
2296 struct virtnet_rq_stats *stats)
2297{
2298 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
2299 int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
2300 struct page *page = virt_to_head_page(buf);
2301 int offset = buf - page_address(page);
2302 unsigned int xdp_frags_truesz = 0;
2303 struct sk_buff *head_skb;
2304 unsigned int frame_sz;
2305 struct xdp_buff xdp;
2306 void *data;
2307 u32 act;
2308 int err;
2309
2310 data = mergeable_xdp_get_buf(vi, rq, xdp_prog, ctx, &frame_sz, &num_buf, &page,
2311 offset, &len, hdr);
2312 if (unlikely(!data))
2313 goto err_xdp;
2314
2315 err = virtnet_build_xdp_buff_mrg(dev, vi, rq, &xdp, data, len, frame_sz,
2316 &num_buf, &xdp_frags_truesz, stats);
2317 if (unlikely(err))
2318 goto err_xdp;
2319
2320 act = virtnet_xdp_handler(xdp_prog, &xdp, dev, xdp_xmit, stats);
2321
2322 switch (act) {
2323 case XDP_PASS:
2324 head_skb = build_skb_from_xdp_buff(dev, vi, &xdp, xdp_frags_truesz);
2325 if (unlikely(!head_skb))
2326 break;
2327 return head_skb;
2328
2329 case XDP_TX:
2330 case XDP_REDIRECT:
2331 return NULL;
2332
2333 default:
2334 break;
2335 }
2336
2337 put_xdp_frags(&xdp);
2338
2339err_xdp:
2340 put_page(page);
2341 mergeable_buf_free(rq, num_buf, dev, stats);
2342
2343 u64_stats_inc(&stats->xdp_drops);
2344 u64_stats_inc(&stats->drops);
2345 return NULL;
2346}
2347
2348static struct sk_buff *virtnet_skb_append_frag(struct sk_buff *head_skb,
2349 struct sk_buff *curr_skb,
2350 struct page *page, void *buf,
2351 int len, int truesize)
2352{
2353 int num_skb_frags;
2354 int offset;
2355
2356 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
2357 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
2358 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
2359
2360 if (unlikely(!nskb))
2361 return NULL;
2362
2363 if (curr_skb == head_skb)
2364 skb_shinfo(curr_skb)->frag_list = nskb;
2365 else
2366 curr_skb->next = nskb;
2367 curr_skb = nskb;
2368 head_skb->truesize += nskb->truesize;
2369 num_skb_frags = 0;
2370 }
2371
2372 if (curr_skb != head_skb) {
2373 head_skb->data_len += len;
2374 head_skb->len += len;
2375 head_skb->truesize += truesize;
2376 }
2377
2378 offset = buf - page_address(page);
2379 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
2380 put_page(page);
2381 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
2382 len, truesize);
2383 } else {
2384 skb_add_rx_frag(curr_skb, num_skb_frags, page,
2385 offset, len, truesize);
2386 }
2387
2388 return curr_skb;
2389}
2390
2391static struct sk_buff *receive_mergeable(struct net_device *dev,
2392 struct virtnet_info *vi,
2393 struct receive_queue *rq,
2394 void *buf,
2395 void *ctx,
2396 unsigned int len,
2397 unsigned int *xdp_xmit,
2398 struct virtnet_rq_stats *stats)
2399{
2400 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
2401 int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
2402 struct page *page = virt_to_head_page(buf);
2403 int offset = buf - page_address(page);
2404 struct sk_buff *head_skb, *curr_skb;
2405 unsigned int truesize = mergeable_ctx_to_truesize(ctx);
2406 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
2407 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2408 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
2409
2410 head_skb = NULL;
2411 u64_stats_add(&stats->bytes, len - vi->hdr_len);
2412
2413 if (unlikely(len > truesize - room)) {
2414 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
2415 dev->name, len, (unsigned long)(truesize - room));
2416 DEV_STATS_INC(dev, rx_length_errors);
2417 goto err_skb;
2418 }
2419
2420 if (unlikely(vi->xdp_enabled)) {
2421 struct bpf_prog *xdp_prog;
2422
2423 rcu_read_lock();
2424 xdp_prog = rcu_dereference(rq->xdp_prog);
2425 if (xdp_prog) {
2426 head_skb = receive_mergeable_xdp(dev, vi, rq, xdp_prog, buf, ctx,
2427 len, xdp_xmit, stats);
2428 rcu_read_unlock();
2429 return head_skb;
2430 }
2431 rcu_read_unlock();
2432 }
2433
2434 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, headroom);
2435 curr_skb = head_skb;
2436
2437 if (unlikely(!curr_skb))
2438 goto err_skb;
2439 while (--num_buf) {
2440 buf = virtnet_rq_get_buf(rq, &len, &ctx);
2441 if (unlikely(!buf)) {
2442 pr_debug("%s: rx error: %d buffers out of %d missing\n",
2443 dev->name, num_buf,
2444 virtio16_to_cpu(vi->vdev,
2445 hdr->num_buffers));
2446 DEV_STATS_INC(dev, rx_length_errors);
2447 goto err_buf;
2448 }
2449
2450 u64_stats_add(&stats->bytes, len);
2451 page = virt_to_head_page(buf);
2452
2453 truesize = mergeable_ctx_to_truesize(ctx);
2454 headroom = mergeable_ctx_to_headroom(ctx);
2455 tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2456 room = SKB_DATA_ALIGN(headroom + tailroom);
2457 if (unlikely(len > truesize - room)) {
2458 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
2459 dev->name, len, (unsigned long)(truesize - room));
2460 DEV_STATS_INC(dev, rx_length_errors);
2461 goto err_skb;
2462 }
2463
2464 curr_skb = virtnet_skb_append_frag(head_skb, curr_skb, page,
2465 buf, len, truesize);
2466 if (!curr_skb)
2467 goto err_skb;
2468 }
2469
2470 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
2471 return head_skb;
2472
2473err_skb:
2474 put_page(page);
2475 mergeable_buf_free(rq, num_buf, dev, stats);
2476
2477err_buf:
2478 u64_stats_inc(&stats->drops);
2479 dev_kfree_skb(head_skb);
2480 return NULL;
2481}
2482
2483static void virtio_skb_set_hash(const struct virtio_net_hdr_v1_hash *hdr_hash,
2484 struct sk_buff *skb)
2485{
2486 enum pkt_hash_types rss_hash_type;
2487
2488 if (!hdr_hash || !skb)
2489 return;
2490
2491 switch (__le16_to_cpu(hdr_hash->hash_report)) {
2492 case VIRTIO_NET_HASH_REPORT_TCPv4:
2493 case VIRTIO_NET_HASH_REPORT_UDPv4:
2494 case VIRTIO_NET_HASH_REPORT_TCPv6:
2495 case VIRTIO_NET_HASH_REPORT_UDPv6:
2496 case VIRTIO_NET_HASH_REPORT_TCPv6_EX:
2497 case VIRTIO_NET_HASH_REPORT_UDPv6_EX:
2498 rss_hash_type = PKT_HASH_TYPE_L4;
2499 break;
2500 case VIRTIO_NET_HASH_REPORT_IPv4:
2501 case VIRTIO_NET_HASH_REPORT_IPv6:
2502 case VIRTIO_NET_HASH_REPORT_IPv6_EX:
2503 rss_hash_type = PKT_HASH_TYPE_L3;
2504 break;
2505 case VIRTIO_NET_HASH_REPORT_NONE:
2506 default:
2507 rss_hash_type = PKT_HASH_TYPE_NONE;
2508 }
2509 skb_set_hash(skb, __le32_to_cpu(hdr_hash->hash_value), rss_hash_type);
2510}
2511
2512static void virtnet_receive_done(struct virtnet_info *vi, struct receive_queue *rq,
2513 struct sk_buff *skb, u8 flags)
2514{
2515 struct virtio_net_common_hdr *hdr;
2516 struct net_device *dev = vi->dev;
2517
2518 hdr = skb_vnet_common_hdr(skb);
2519 if (dev->features & NETIF_F_RXHASH && vi->has_rss_hash_report)
2520 virtio_skb_set_hash(&hdr->hash_v1_hdr, skb);
2521
2522 if (flags & VIRTIO_NET_HDR_F_DATA_VALID)
2523 skb->ip_summed = CHECKSUM_UNNECESSARY;
2524
2525 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
2526 virtio_is_little_endian(vi->vdev))) {
2527 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
2528 dev->name, hdr->hdr.gso_type,
2529 hdr->hdr.gso_size);
2530 goto frame_err;
2531 }
2532
2533 skb_record_rx_queue(skb, vq2rxq(rq->vq));
2534 skb->protocol = eth_type_trans(skb, dev);
2535 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
2536 ntohs(skb->protocol), skb->len, skb->pkt_type);
2537
2538 napi_gro_receive(&rq->napi, skb);
2539 return;
2540
2541frame_err:
2542 DEV_STATS_INC(dev, rx_frame_errors);
2543 dev_kfree_skb(skb);
2544}
2545
2546static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
2547 void *buf, unsigned int len, void **ctx,
2548 unsigned int *xdp_xmit,
2549 struct virtnet_rq_stats *stats)
2550{
2551 struct net_device *dev = vi->dev;
2552 struct sk_buff *skb;
2553 u8 flags;
2554
2555 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
2556 pr_debug("%s: short packet %i\n", dev->name, len);
2557 DEV_STATS_INC(dev, rx_length_errors);
2558 virtnet_rq_free_buf(vi, rq, buf);
2559 return;
2560 }
2561
2562 /* 1. Save the flags early, as the XDP program might overwrite them.
2563 * These flags ensure packets marked as VIRTIO_NET_HDR_F_DATA_VALID
2564 * stay valid after XDP processing.
2565 * 2. XDP doesn't work with partially checksummed packets (refer to
2566 * virtnet_xdp_set()), so packets marked as
2567 * VIRTIO_NET_HDR_F_NEEDS_CSUM get dropped during XDP processing.
2568 */
2569 flags = ((struct virtio_net_common_hdr *)buf)->hdr.flags;
2570
2571 if (vi->mergeable_rx_bufs)
2572 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
2573 stats);
2574 else if (vi->big_packets)
2575 skb = receive_big(dev, vi, rq, buf, len, stats);
2576 else
2577 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
2578
2579 if (unlikely(!skb))
2580 return;
2581
2582 virtnet_receive_done(vi, rq, skb, flags);
2583}
2584
2585/* Unlike mergeable buffers, all buffers are allocated to the
2586 * same size, except for the headroom. For this reason we do
2587 * not need to use mergeable_len_to_ctx here - it is enough
2588 * to store the headroom as the context ignoring the truesize.
2589 */
2590static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
2591 gfp_t gfp)
2592{
2593 char *buf;
2594 unsigned int xdp_headroom = virtnet_get_headroom(vi);
2595 void *ctx = (void *)(unsigned long)xdp_headroom;
2596 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
2597 int err;
2598
2599 len = SKB_DATA_ALIGN(len) +
2600 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
2601
2602 if (unlikely(!skb_page_frag_refill(len, &rq->alloc_frag, gfp)))
2603 return -ENOMEM;
2604
2605 buf = virtnet_rq_alloc(rq, len, gfp);
2606 if (unlikely(!buf))
2607 return -ENOMEM;
2608
2609 buf += VIRTNET_RX_PAD + xdp_headroom;
2610
2611 virtnet_rq_init_one_sg(rq, buf, vi->hdr_len + GOOD_PACKET_LEN);
2612
2613 err = virtqueue_add_inbuf_premapped(rq->vq, rq->sg, 1, buf, ctx, gfp);
2614 if (err < 0) {
2615 virtnet_rq_unmap(rq, buf, 0);
2616 put_page(virt_to_head_page(buf));
2617 }
2618
2619 return err;
2620}
2621
2622static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
2623 gfp_t gfp)
2624{
2625 struct page *first, *list = NULL;
2626 char *p;
2627 int i, err, offset;
2628
2629 sg_init_table(rq->sg, vi->big_packets_num_skbfrags + 2);
2630
2631 /* page in rq->sg[vi->big_packets_num_skbfrags + 1] is list tail */
2632 for (i = vi->big_packets_num_skbfrags + 1; i > 1; --i) {
2633 first = get_a_page(rq, gfp);
2634 if (!first) {
2635 if (list)
2636 give_pages(rq, list);
2637 return -ENOMEM;
2638 }
2639 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
2640
2641 /* chain new page in list head to match sg */
2642 first->private = (unsigned long)list;
2643 list = first;
2644 }
2645
2646 first = get_a_page(rq, gfp);
2647 if (!first) {
2648 give_pages(rq, list);
2649 return -ENOMEM;
2650 }
2651 p = page_address(first);
2652
2653 /* rq->sg[0], rq->sg[1] share the same page */
2654 /* a separated rq->sg[0] for header - required in case !any_header_sg */
2655 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
2656
2657 /* rq->sg[1] for data packet, from offset */
2658 offset = sizeof(struct padded_vnet_hdr);
2659 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
2660
2661 /* chain first in list head */
2662 first->private = (unsigned long)list;
2663 err = virtqueue_add_inbuf(rq->vq, rq->sg, vi->big_packets_num_skbfrags + 2,
2664 first, gfp);
2665 if (err < 0)
2666 give_pages(rq, first);
2667
2668 return err;
2669}
2670
2671static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
2672 struct ewma_pkt_len *avg_pkt_len,
2673 unsigned int room)
2674{
2675 struct virtnet_info *vi = rq->vq->vdev->priv;
2676 const size_t hdr_len = vi->hdr_len;
2677 unsigned int len;
2678
2679 if (room)
2680 return PAGE_SIZE - room;
2681
2682 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
2683 rq->min_buf_len, PAGE_SIZE - hdr_len);
2684
2685 return ALIGN(len, L1_CACHE_BYTES);
2686}
2687
2688static int add_recvbuf_mergeable(struct virtnet_info *vi,
2689 struct receive_queue *rq, gfp_t gfp)
2690{
2691 struct page_frag *alloc_frag = &rq->alloc_frag;
2692 unsigned int headroom = virtnet_get_headroom(vi);
2693 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2694 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
2695 unsigned int len, hole;
2696 void *ctx;
2697 char *buf;
2698 int err;
2699
2700 /* Extra tailroom is needed to satisfy XDP's assumption. This
2701 * means rx frags coalescing won't work, but consider we've
2702 * disabled GSO for XDP, it won't be a big issue.
2703 */
2704 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
2705
2706 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
2707 return -ENOMEM;
2708
2709 if (!alloc_frag->offset && len + room + sizeof(struct virtnet_rq_dma) > alloc_frag->size)
2710 len -= sizeof(struct virtnet_rq_dma);
2711
2712 buf = virtnet_rq_alloc(rq, len + room, gfp);
2713 if (unlikely(!buf))
2714 return -ENOMEM;
2715
2716 buf += headroom; /* advance address leaving hole at front of pkt */
2717 hole = alloc_frag->size - alloc_frag->offset;
2718 if (hole < len + room) {
2719 /* To avoid internal fragmentation, if there is very likely not
2720 * enough space for another buffer, add the remaining space to
2721 * the current buffer.
2722 * XDP core assumes that frame_size of xdp_buff and the length
2723 * of the frag are PAGE_SIZE, so we disable the hole mechanism.
2724 */
2725 if (!headroom)
2726 len += hole;
2727 alloc_frag->offset += hole;
2728 }
2729
2730 virtnet_rq_init_one_sg(rq, buf, len);
2731
2732 ctx = mergeable_len_to_ctx(len + room, headroom);
2733 err = virtqueue_add_inbuf_premapped(rq->vq, rq->sg, 1, buf, ctx, gfp);
2734 if (err < 0) {
2735 virtnet_rq_unmap(rq, buf, 0);
2736 put_page(virt_to_head_page(buf));
2737 }
2738
2739 return err;
2740}
2741
2742/*
2743 * Returns false if we couldn't fill entirely (OOM).
2744 *
2745 * Normally run in the receive path, but can also be run from ndo_open
2746 * before we're receiving packets, or from refill_work which is
2747 * careful to disable receiving (using napi_disable).
2748 */
2749static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
2750 gfp_t gfp)
2751{
2752 int err;
2753
2754 if (rq->xsk_pool) {
2755 err = virtnet_add_recvbuf_xsk(vi, rq, rq->xsk_pool, gfp);
2756 goto kick;
2757 }
2758
2759 do {
2760 if (vi->mergeable_rx_bufs)
2761 err = add_recvbuf_mergeable(vi, rq, gfp);
2762 else if (vi->big_packets)
2763 err = add_recvbuf_big(vi, rq, gfp);
2764 else
2765 err = add_recvbuf_small(vi, rq, gfp);
2766
2767 if (err)
2768 break;
2769 } while (rq->vq->num_free);
2770
2771kick:
2772 if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
2773 unsigned long flags;
2774
2775 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
2776 u64_stats_inc(&rq->stats.kicks);
2777 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
2778 }
2779
2780 return err != -ENOMEM;
2781}
2782
2783static void skb_recv_done(struct virtqueue *rvq)
2784{
2785 struct virtnet_info *vi = rvq->vdev->priv;
2786 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
2787
2788 rq->calls++;
2789 virtqueue_napi_schedule(&rq->napi, rvq);
2790}
2791
2792static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
2793{
2794 napi_enable(napi);
2795
2796 /* If all buffers were filled by other side before we napi_enabled, we
2797 * won't get another interrupt, so process any outstanding packets now.
2798 * Call local_bh_enable after to trigger softIRQ processing.
2799 */
2800 local_bh_disable();
2801 virtqueue_napi_schedule(napi, vq);
2802 local_bh_enable();
2803}
2804
2805static void virtnet_napi_tx_enable(struct virtnet_info *vi,
2806 struct virtqueue *vq,
2807 struct napi_struct *napi)
2808{
2809 if (!napi->weight)
2810 return;
2811
2812 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
2813 * enable the feature if this is likely affine with the transmit path.
2814 */
2815 if (!vi->affinity_hint_set) {
2816 napi->weight = 0;
2817 return;
2818 }
2819
2820 return virtnet_napi_enable(vq, napi);
2821}
2822
2823static void virtnet_napi_tx_disable(struct napi_struct *napi)
2824{
2825 if (napi->weight)
2826 napi_disable(napi);
2827}
2828
2829static void refill_work(struct work_struct *work)
2830{
2831 struct virtnet_info *vi =
2832 container_of(work, struct virtnet_info, refill.work);
2833 bool still_empty;
2834 int i;
2835
2836 for (i = 0; i < vi->curr_queue_pairs; i++) {
2837 struct receive_queue *rq = &vi->rq[i];
2838
2839 napi_disable(&rq->napi);
2840 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
2841 virtnet_napi_enable(rq->vq, &rq->napi);
2842
2843 /* In theory, this can happen: if we don't get any buffers in
2844 * we will *never* try to fill again.
2845 */
2846 if (still_empty)
2847 schedule_delayed_work(&vi->refill, HZ/2);
2848 }
2849}
2850
2851static int virtnet_receive_xsk_bufs(struct virtnet_info *vi,
2852 struct receive_queue *rq,
2853 int budget,
2854 unsigned int *xdp_xmit,
2855 struct virtnet_rq_stats *stats)
2856{
2857 unsigned int len;
2858 int packets = 0;
2859 void *buf;
2860
2861 while (packets < budget) {
2862 buf = virtqueue_get_buf(rq->vq, &len);
2863 if (!buf)
2864 break;
2865
2866 virtnet_receive_xsk_buf(vi, rq, buf, len, xdp_xmit, stats);
2867 packets++;
2868 }
2869
2870 return packets;
2871}
2872
2873static int virtnet_receive_packets(struct virtnet_info *vi,
2874 struct receive_queue *rq,
2875 int budget,
2876 unsigned int *xdp_xmit,
2877 struct virtnet_rq_stats *stats)
2878{
2879 unsigned int len;
2880 int packets = 0;
2881 void *buf;
2882
2883 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2884 void *ctx;
2885 while (packets < budget &&
2886 (buf = virtnet_rq_get_buf(rq, &len, &ctx))) {
2887 receive_buf(vi, rq, buf, len, ctx, xdp_xmit, stats);
2888 packets++;
2889 }
2890 } else {
2891 while (packets < budget &&
2892 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
2893 receive_buf(vi, rq, buf, len, NULL, xdp_xmit, stats);
2894 packets++;
2895 }
2896 }
2897
2898 return packets;
2899}
2900
2901static int virtnet_receive(struct receive_queue *rq, int budget,
2902 unsigned int *xdp_xmit)
2903{
2904 struct virtnet_info *vi = rq->vq->vdev->priv;
2905 struct virtnet_rq_stats stats = {};
2906 int i, packets;
2907
2908 if (rq->xsk_pool)
2909 packets = virtnet_receive_xsk_bufs(vi, rq, budget, xdp_xmit, &stats);
2910 else
2911 packets = virtnet_receive_packets(vi, rq, budget, xdp_xmit, &stats);
2912
2913 if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
2914 if (!try_fill_recv(vi, rq, GFP_ATOMIC)) {
2915 spin_lock(&vi->refill_lock);
2916 if (vi->refill_enabled)
2917 schedule_delayed_work(&vi->refill, 0);
2918 spin_unlock(&vi->refill_lock);
2919 }
2920 }
2921
2922 u64_stats_set(&stats.packets, packets);
2923 u64_stats_update_begin(&rq->stats.syncp);
2924 for (i = 0; i < ARRAY_SIZE(virtnet_rq_stats_desc); i++) {
2925 size_t offset = virtnet_rq_stats_desc[i].offset;
2926 u64_stats_t *item, *src;
2927
2928 item = (u64_stats_t *)((u8 *)&rq->stats + offset);
2929 src = (u64_stats_t *)((u8 *)&stats + offset);
2930 u64_stats_add(item, u64_stats_read(src));
2931 }
2932
2933 u64_stats_add(&rq->stats.packets, u64_stats_read(&stats.packets));
2934 u64_stats_add(&rq->stats.bytes, u64_stats_read(&stats.bytes));
2935
2936 u64_stats_update_end(&rq->stats.syncp);
2937
2938 return packets;
2939}
2940
2941static void virtnet_poll_cleantx(struct receive_queue *rq, int budget)
2942{
2943 struct virtnet_info *vi = rq->vq->vdev->priv;
2944 unsigned int index = vq2rxq(rq->vq);
2945 struct send_queue *sq = &vi->sq[index];
2946 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
2947
2948 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
2949 return;
2950
2951 if (__netif_tx_trylock(txq)) {
2952 if (sq->reset) {
2953 __netif_tx_unlock(txq);
2954 return;
2955 }
2956
2957 do {
2958 virtqueue_disable_cb(sq->vq);
2959 free_old_xmit(sq, txq, !!budget);
2960 } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
2961
2962 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) {
2963 if (netif_tx_queue_stopped(txq)) {
2964 u64_stats_update_begin(&sq->stats.syncp);
2965 u64_stats_inc(&sq->stats.wake);
2966 u64_stats_update_end(&sq->stats.syncp);
2967 }
2968 netif_tx_wake_queue(txq);
2969 }
2970
2971 __netif_tx_unlock(txq);
2972 }
2973}
2974
2975static void virtnet_rx_dim_update(struct virtnet_info *vi, struct receive_queue *rq)
2976{
2977 struct dim_sample cur_sample = {};
2978
2979 if (!rq->packets_in_napi)
2980 return;
2981
2982 /* Don't need protection when fetching stats, since fetcher and
2983 * updater of the stats are in same context
2984 */
2985 dim_update_sample(rq->calls,
2986 u64_stats_read(&rq->stats.packets),
2987 u64_stats_read(&rq->stats.bytes),
2988 &cur_sample);
2989
2990 net_dim(&rq->dim, &cur_sample);
2991 rq->packets_in_napi = 0;
2992}
2993
2994static int virtnet_poll(struct napi_struct *napi, int budget)
2995{
2996 struct receive_queue *rq =
2997 container_of(napi, struct receive_queue, napi);
2998 struct virtnet_info *vi = rq->vq->vdev->priv;
2999 struct send_queue *sq;
3000 unsigned int received;
3001 unsigned int xdp_xmit = 0;
3002 bool napi_complete;
3003
3004 virtnet_poll_cleantx(rq, budget);
3005
3006 received = virtnet_receive(rq, budget, &xdp_xmit);
3007 rq->packets_in_napi += received;
3008
3009 if (xdp_xmit & VIRTIO_XDP_REDIR)
3010 xdp_do_flush();
3011
3012 /* Out of packets? */
3013 if (received < budget) {
3014 napi_complete = virtqueue_napi_complete(napi, rq->vq, received);
3015 /* Intentionally not taking dim_lock here. This may result in a
3016 * spurious net_dim call. But if that happens virtnet_rx_dim_work
3017 * will not act on the scheduled work.
3018 */
3019 if (napi_complete && rq->dim_enabled)
3020 virtnet_rx_dim_update(vi, rq);
3021 }
3022
3023 if (xdp_xmit & VIRTIO_XDP_TX) {
3024 sq = virtnet_xdp_get_sq(vi);
3025 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
3026 u64_stats_update_begin(&sq->stats.syncp);
3027 u64_stats_inc(&sq->stats.kicks);
3028 u64_stats_update_end(&sq->stats.syncp);
3029 }
3030 virtnet_xdp_put_sq(vi, sq);
3031 }
3032
3033 return received;
3034}
3035
3036static void virtnet_disable_queue_pair(struct virtnet_info *vi, int qp_index)
3037{
3038 virtnet_napi_tx_disable(&vi->sq[qp_index].napi);
3039 napi_disable(&vi->rq[qp_index].napi);
3040 xdp_rxq_info_unreg(&vi->rq[qp_index].xdp_rxq);
3041}
3042
3043static int virtnet_enable_queue_pair(struct virtnet_info *vi, int qp_index)
3044{
3045 struct net_device *dev = vi->dev;
3046 int err;
3047
3048 err = xdp_rxq_info_reg(&vi->rq[qp_index].xdp_rxq, dev, qp_index,
3049 vi->rq[qp_index].napi.napi_id);
3050 if (err < 0)
3051 return err;
3052
3053 err = xdp_rxq_info_reg_mem_model(&vi->rq[qp_index].xdp_rxq,
3054 MEM_TYPE_PAGE_SHARED, NULL);
3055 if (err < 0)
3056 goto err_xdp_reg_mem_model;
3057
3058 virtnet_napi_enable(vi->rq[qp_index].vq, &vi->rq[qp_index].napi);
3059 virtnet_napi_tx_enable(vi, vi->sq[qp_index].vq, &vi->sq[qp_index].napi);
3060
3061 return 0;
3062
3063err_xdp_reg_mem_model:
3064 xdp_rxq_info_unreg(&vi->rq[qp_index].xdp_rxq);
3065 return err;
3066}
3067
3068static void virtnet_cancel_dim(struct virtnet_info *vi, struct dim *dim)
3069{
3070 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL))
3071 return;
3072 net_dim_work_cancel(dim);
3073}
3074
3075static void virtnet_update_settings(struct virtnet_info *vi)
3076{
3077 u32 speed;
3078 u8 duplex;
3079
3080 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
3081 return;
3082
3083 virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
3084
3085 if (ethtool_validate_speed(speed))
3086 vi->speed = speed;
3087
3088 virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
3089
3090 if (ethtool_validate_duplex(duplex))
3091 vi->duplex = duplex;
3092}
3093
3094static int virtnet_open(struct net_device *dev)
3095{
3096 struct virtnet_info *vi = netdev_priv(dev);
3097 int i, err;
3098
3099 enable_delayed_refill(vi);
3100
3101 for (i = 0; i < vi->max_queue_pairs; i++) {
3102 if (i < vi->curr_queue_pairs)
3103 /* Make sure we have some buffers: if oom use wq. */
3104 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
3105 schedule_delayed_work(&vi->refill, 0);
3106
3107 err = virtnet_enable_queue_pair(vi, i);
3108 if (err < 0)
3109 goto err_enable_qp;
3110 }
3111
3112 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3113 if (vi->status & VIRTIO_NET_S_LINK_UP)
3114 netif_carrier_on(vi->dev);
3115 virtio_config_driver_enable(vi->vdev);
3116 } else {
3117 vi->status = VIRTIO_NET_S_LINK_UP;
3118 netif_carrier_on(dev);
3119 }
3120
3121 return 0;
3122
3123err_enable_qp:
3124 disable_delayed_refill(vi);
3125 cancel_delayed_work_sync(&vi->refill);
3126
3127 for (i--; i >= 0; i--) {
3128 virtnet_disable_queue_pair(vi, i);
3129 virtnet_cancel_dim(vi, &vi->rq[i].dim);
3130 }
3131
3132 return err;
3133}
3134
3135static int virtnet_poll_tx(struct napi_struct *napi, int budget)
3136{
3137 struct send_queue *sq = container_of(napi, struct send_queue, napi);
3138 struct virtnet_info *vi = sq->vq->vdev->priv;
3139 unsigned int index = vq2txq(sq->vq);
3140 struct netdev_queue *txq;
3141 int opaque, xsk_done = 0;
3142 bool done;
3143
3144 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
3145 /* We don't need to enable cb for XDP */
3146 napi_complete_done(napi, 0);
3147 return 0;
3148 }
3149
3150 txq = netdev_get_tx_queue(vi->dev, index);
3151 __netif_tx_lock(txq, raw_smp_processor_id());
3152 virtqueue_disable_cb(sq->vq);
3153
3154 if (sq->xsk_pool)
3155 xsk_done = virtnet_xsk_xmit(sq, sq->xsk_pool, budget);
3156 else
3157 free_old_xmit(sq, txq, !!budget);
3158
3159 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) {
3160 if (netif_tx_queue_stopped(txq)) {
3161 u64_stats_update_begin(&sq->stats.syncp);
3162 u64_stats_inc(&sq->stats.wake);
3163 u64_stats_update_end(&sq->stats.syncp);
3164 }
3165 netif_tx_wake_queue(txq);
3166 }
3167
3168 if (xsk_done >= budget) {
3169 __netif_tx_unlock(txq);
3170 return budget;
3171 }
3172
3173 opaque = virtqueue_enable_cb_prepare(sq->vq);
3174
3175 done = napi_complete_done(napi, 0);
3176
3177 if (!done)
3178 virtqueue_disable_cb(sq->vq);
3179
3180 __netif_tx_unlock(txq);
3181
3182 if (done) {
3183 if (unlikely(virtqueue_poll(sq->vq, opaque))) {
3184 if (napi_schedule_prep(napi)) {
3185 __netif_tx_lock(txq, raw_smp_processor_id());
3186 virtqueue_disable_cb(sq->vq);
3187 __netif_tx_unlock(txq);
3188 __napi_schedule(napi);
3189 }
3190 }
3191 }
3192
3193 return 0;
3194}
3195
3196static int xmit_skb(struct send_queue *sq, struct sk_buff *skb, bool orphan)
3197{
3198 struct virtio_net_hdr_mrg_rxbuf *hdr;
3199 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
3200 struct virtnet_info *vi = sq->vq->vdev->priv;
3201 int num_sg;
3202 unsigned hdr_len = vi->hdr_len;
3203 bool can_push;
3204
3205 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
3206
3207 can_push = vi->any_header_sg &&
3208 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
3209 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
3210 /* Even if we can, don't push here yet as this would skew
3211 * csum_start offset below. */
3212 if (can_push)
3213 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
3214 else
3215 hdr = &skb_vnet_common_hdr(skb)->mrg_hdr;
3216
3217 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
3218 virtio_is_little_endian(vi->vdev), false,
3219 0))
3220 return -EPROTO;
3221
3222 if (vi->mergeable_rx_bufs)
3223 hdr->num_buffers = 0;
3224
3225 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
3226 if (can_push) {
3227 __skb_push(skb, hdr_len);
3228 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
3229 if (unlikely(num_sg < 0))
3230 return num_sg;
3231 /* Pull header back to avoid skew in tx bytes calculations. */
3232 __skb_pull(skb, hdr_len);
3233 } else {
3234 sg_set_buf(sq->sg, hdr, hdr_len);
3235 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
3236 if (unlikely(num_sg < 0))
3237 return num_sg;
3238 num_sg++;
3239 }
3240
3241 return virtnet_add_outbuf(sq, num_sg, skb,
3242 orphan ? VIRTNET_XMIT_TYPE_SKB_ORPHAN : VIRTNET_XMIT_TYPE_SKB);
3243}
3244
3245static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
3246{
3247 struct virtnet_info *vi = netdev_priv(dev);
3248 int qnum = skb_get_queue_mapping(skb);
3249 struct send_queue *sq = &vi->sq[qnum];
3250 int err;
3251 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
3252 bool xmit_more = netdev_xmit_more();
3253 bool use_napi = sq->napi.weight;
3254 bool kick;
3255
3256 /* Free up any pending old buffers before queueing new ones. */
3257 do {
3258 if (use_napi)
3259 virtqueue_disable_cb(sq->vq);
3260
3261 free_old_xmit(sq, txq, false);
3262
3263 } while (use_napi && !xmit_more &&
3264 unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
3265
3266 /* timestamp packet in software */
3267 skb_tx_timestamp(skb);
3268
3269 /* Try to transmit */
3270 err = xmit_skb(sq, skb, !use_napi);
3271
3272 /* This should not happen! */
3273 if (unlikely(err)) {
3274 DEV_STATS_INC(dev, tx_fifo_errors);
3275 if (net_ratelimit())
3276 dev_warn(&dev->dev,
3277 "Unexpected TXQ (%d) queue failure: %d\n",
3278 qnum, err);
3279 DEV_STATS_INC(dev, tx_dropped);
3280 dev_kfree_skb_any(skb);
3281 return NETDEV_TX_OK;
3282 }
3283
3284 /* Don't wait up for transmitted skbs to be freed. */
3285 if (!use_napi) {
3286 skb_orphan(skb);
3287 nf_reset_ct(skb);
3288 }
3289
3290 check_sq_full_and_disable(vi, dev, sq);
3291
3292 kick = use_napi ? __netdev_tx_sent_queue(txq, skb->len, xmit_more) :
3293 !xmit_more || netif_xmit_stopped(txq);
3294 if (kick) {
3295 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
3296 u64_stats_update_begin(&sq->stats.syncp);
3297 u64_stats_inc(&sq->stats.kicks);
3298 u64_stats_update_end(&sq->stats.syncp);
3299 }
3300 }
3301
3302 return NETDEV_TX_OK;
3303}
3304
3305static void virtnet_rx_pause(struct virtnet_info *vi, struct receive_queue *rq)
3306{
3307 bool running = netif_running(vi->dev);
3308
3309 if (running) {
3310 napi_disable(&rq->napi);
3311 virtnet_cancel_dim(vi, &rq->dim);
3312 }
3313}
3314
3315static void virtnet_rx_resume(struct virtnet_info *vi, struct receive_queue *rq)
3316{
3317 bool running = netif_running(vi->dev);
3318
3319 if (!try_fill_recv(vi, rq, GFP_KERNEL))
3320 schedule_delayed_work(&vi->refill, 0);
3321
3322 if (running)
3323 virtnet_napi_enable(rq->vq, &rq->napi);
3324}
3325
3326static int virtnet_rx_resize(struct virtnet_info *vi,
3327 struct receive_queue *rq, u32 ring_num)
3328{
3329 int err, qindex;
3330
3331 qindex = rq - vi->rq;
3332
3333 virtnet_rx_pause(vi, rq);
3334
3335 err = virtqueue_resize(rq->vq, ring_num, virtnet_rq_unmap_free_buf, NULL);
3336 if (err)
3337 netdev_err(vi->dev, "resize rx fail: rx queue index: %d err: %d\n", qindex, err);
3338
3339 virtnet_rx_resume(vi, rq);
3340 return err;
3341}
3342
3343static void virtnet_tx_pause(struct virtnet_info *vi, struct send_queue *sq)
3344{
3345 bool running = netif_running(vi->dev);
3346 struct netdev_queue *txq;
3347 int qindex;
3348
3349 qindex = sq - vi->sq;
3350
3351 if (running)
3352 virtnet_napi_tx_disable(&sq->napi);
3353
3354 txq = netdev_get_tx_queue(vi->dev, qindex);
3355
3356 /* 1. wait all ximt complete
3357 * 2. fix the race of netif_stop_subqueue() vs netif_start_subqueue()
3358 */
3359 __netif_tx_lock_bh(txq);
3360
3361 /* Prevent rx poll from accessing sq. */
3362 sq->reset = true;
3363
3364 /* Prevent the upper layer from trying to send packets. */
3365 netif_stop_subqueue(vi->dev, qindex);
3366
3367 __netif_tx_unlock_bh(txq);
3368}
3369
3370static void virtnet_tx_resume(struct virtnet_info *vi, struct send_queue *sq)
3371{
3372 bool running = netif_running(vi->dev);
3373 struct netdev_queue *txq;
3374 int qindex;
3375
3376 qindex = sq - vi->sq;
3377
3378 txq = netdev_get_tx_queue(vi->dev, qindex);
3379
3380 __netif_tx_lock_bh(txq);
3381 sq->reset = false;
3382 netif_tx_wake_queue(txq);
3383 __netif_tx_unlock_bh(txq);
3384
3385 if (running)
3386 virtnet_napi_tx_enable(vi, sq->vq, &sq->napi);
3387}
3388
3389static int virtnet_tx_resize(struct virtnet_info *vi, struct send_queue *sq,
3390 u32 ring_num)
3391{
3392 int qindex, err;
3393
3394 qindex = sq - vi->sq;
3395
3396 virtnet_tx_pause(vi, sq);
3397
3398 err = virtqueue_resize(sq->vq, ring_num, virtnet_sq_free_unused_buf,
3399 virtnet_sq_free_unused_buf_done);
3400 if (err)
3401 netdev_err(vi->dev, "resize tx fail: tx queue index: %d err: %d\n", qindex, err);
3402
3403 virtnet_tx_resume(vi, sq);
3404
3405 return err;
3406}
3407
3408/*
3409 * Send command via the control virtqueue and check status. Commands
3410 * supported by the hypervisor, as indicated by feature bits, should
3411 * never fail unless improperly formatted.
3412 */
3413static bool virtnet_send_command_reply(struct virtnet_info *vi, u8 class, u8 cmd,
3414 struct scatterlist *out,
3415 struct scatterlist *in)
3416{
3417 struct scatterlist *sgs[5], hdr, stat;
3418 u32 out_num = 0, tmp, in_num = 0;
3419 bool ok;
3420 int ret;
3421
3422 /* Caller should know better */
3423 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
3424
3425 mutex_lock(&vi->cvq_lock);
3426 vi->ctrl->status = ~0;
3427 vi->ctrl->hdr.class = class;
3428 vi->ctrl->hdr.cmd = cmd;
3429 /* Add header */
3430 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
3431 sgs[out_num++] = &hdr;
3432
3433 if (out)
3434 sgs[out_num++] = out;
3435
3436 /* Add return status. */
3437 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
3438 sgs[out_num + in_num++] = &stat;
3439
3440 if (in)
3441 sgs[out_num + in_num++] = in;
3442
3443 BUG_ON(out_num + in_num > ARRAY_SIZE(sgs));
3444 ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, in_num, vi, GFP_ATOMIC);
3445 if (ret < 0) {
3446 dev_warn(&vi->vdev->dev,
3447 "Failed to add sgs for command vq: %d\n.", ret);
3448 mutex_unlock(&vi->cvq_lock);
3449 return false;
3450 }
3451
3452 if (unlikely(!virtqueue_kick(vi->cvq)))
3453 goto unlock;
3454
3455 /* Spin for a response, the kick causes an ioport write, trapping
3456 * into the hypervisor, so the request should be handled immediately.
3457 */
3458 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
3459 !virtqueue_is_broken(vi->cvq)) {
3460 cond_resched();
3461 cpu_relax();
3462 }
3463
3464unlock:
3465 ok = vi->ctrl->status == VIRTIO_NET_OK;
3466 mutex_unlock(&vi->cvq_lock);
3467 return ok;
3468}
3469
3470static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
3471 struct scatterlist *out)
3472{
3473 return virtnet_send_command_reply(vi, class, cmd, out, NULL);
3474}
3475
3476static int virtnet_set_mac_address(struct net_device *dev, void *p)
3477{
3478 struct virtnet_info *vi = netdev_priv(dev);
3479 struct virtio_device *vdev = vi->vdev;
3480 int ret;
3481 struct sockaddr *addr;
3482 struct scatterlist sg;
3483
3484 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
3485 return -EOPNOTSUPP;
3486
3487 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
3488 if (!addr)
3489 return -ENOMEM;
3490
3491 ret = eth_prepare_mac_addr_change(dev, addr);
3492 if (ret)
3493 goto out;
3494
3495 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
3496 sg_init_one(&sg, addr->sa_data, dev->addr_len);
3497 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
3498 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
3499 dev_warn(&vdev->dev,
3500 "Failed to set mac address by vq command.\n");
3501 ret = -EINVAL;
3502 goto out;
3503 }
3504 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
3505 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
3506 unsigned int i;
3507
3508 /* Naturally, this has an atomicity problem. */
3509 for (i = 0; i < dev->addr_len; i++)
3510 virtio_cwrite8(vdev,
3511 offsetof(struct virtio_net_config, mac) +
3512 i, addr->sa_data[i]);
3513 }
3514
3515 eth_commit_mac_addr_change(dev, p);
3516 ret = 0;
3517
3518out:
3519 kfree(addr);
3520 return ret;
3521}
3522
3523static void virtnet_stats(struct net_device *dev,
3524 struct rtnl_link_stats64 *tot)
3525{
3526 struct virtnet_info *vi = netdev_priv(dev);
3527 unsigned int start;
3528 int i;
3529
3530 for (i = 0; i < vi->max_queue_pairs; i++) {
3531 u64 tpackets, tbytes, terrors, rpackets, rbytes, rdrops;
3532 struct receive_queue *rq = &vi->rq[i];
3533 struct send_queue *sq = &vi->sq[i];
3534
3535 do {
3536 start = u64_stats_fetch_begin(&sq->stats.syncp);
3537 tpackets = u64_stats_read(&sq->stats.packets);
3538 tbytes = u64_stats_read(&sq->stats.bytes);
3539 terrors = u64_stats_read(&sq->stats.tx_timeouts);
3540 } while (u64_stats_fetch_retry(&sq->stats.syncp, start));
3541
3542 do {
3543 start = u64_stats_fetch_begin(&rq->stats.syncp);
3544 rpackets = u64_stats_read(&rq->stats.packets);
3545 rbytes = u64_stats_read(&rq->stats.bytes);
3546 rdrops = u64_stats_read(&rq->stats.drops);
3547 } while (u64_stats_fetch_retry(&rq->stats.syncp, start));
3548
3549 tot->rx_packets += rpackets;
3550 tot->tx_packets += tpackets;
3551 tot->rx_bytes += rbytes;
3552 tot->tx_bytes += tbytes;
3553 tot->rx_dropped += rdrops;
3554 tot->tx_errors += terrors;
3555 }
3556
3557 tot->tx_dropped = DEV_STATS_READ(dev, tx_dropped);
3558 tot->tx_fifo_errors = DEV_STATS_READ(dev, tx_fifo_errors);
3559 tot->rx_length_errors = DEV_STATS_READ(dev, rx_length_errors);
3560 tot->rx_frame_errors = DEV_STATS_READ(dev, rx_frame_errors);
3561}
3562
3563static void virtnet_ack_link_announce(struct virtnet_info *vi)
3564{
3565 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
3566 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
3567 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
3568}
3569
3570static bool virtnet_commit_rss_command(struct virtnet_info *vi);
3571
3572static void virtnet_rss_update_by_qpairs(struct virtnet_info *vi, u16 queue_pairs)
3573{
3574 u32 indir_val = 0;
3575 int i = 0;
3576
3577 for (; i < vi->rss_indir_table_size; ++i) {
3578 indir_val = ethtool_rxfh_indir_default(i, queue_pairs);
3579 vi->rss.indirection_table[i] = indir_val;
3580 }
3581 vi->rss.max_tx_vq = queue_pairs;
3582}
3583
3584static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
3585{
3586 struct virtio_net_ctrl_mq *mq __free(kfree) = NULL;
3587 struct virtio_net_ctrl_rss old_rss;
3588 struct net_device *dev = vi->dev;
3589 struct scatterlist sg;
3590
3591 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
3592 return 0;
3593
3594 /* Firstly check if we need update rss. Do updating if both (1) rss enabled and
3595 * (2) no user configuration.
3596 *
3597 * During rss command processing, device updates queue_pairs using rss.max_tx_vq. That is,
3598 * the device updates queue_pairs together with rss, so we can skip the sperate queue_pairs
3599 * update (VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET below) and return directly.
3600 */
3601 if (vi->has_rss && !netif_is_rxfh_configured(dev)) {
3602 memcpy(&old_rss, &vi->rss, sizeof(old_rss));
3603 if (rss_indirection_table_alloc(&vi->rss, vi->rss_indir_table_size)) {
3604 vi->rss.indirection_table = old_rss.indirection_table;
3605 return -ENOMEM;
3606 }
3607
3608 virtnet_rss_update_by_qpairs(vi, queue_pairs);
3609
3610 if (!virtnet_commit_rss_command(vi)) {
3611 /* restore ctrl_rss if commit_rss_command failed */
3612 rss_indirection_table_free(&vi->rss);
3613 memcpy(&vi->rss, &old_rss, sizeof(old_rss));
3614
3615 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d, because committing RSS failed\n",
3616 queue_pairs);
3617 return -EINVAL;
3618 }
3619 rss_indirection_table_free(&old_rss);
3620 goto succ;
3621 }
3622
3623 mq = kzalloc(sizeof(*mq), GFP_KERNEL);
3624 if (!mq)
3625 return -ENOMEM;
3626
3627 mq->virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
3628 sg_init_one(&sg, mq, sizeof(*mq));
3629
3630 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
3631 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
3632 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
3633 queue_pairs);
3634 return -EINVAL;
3635 }
3636succ:
3637 vi->curr_queue_pairs = queue_pairs;
3638 /* virtnet_open() will refill when device is going to up. */
3639 if (dev->flags & IFF_UP)
3640 schedule_delayed_work(&vi->refill, 0);
3641
3642 return 0;
3643}
3644
3645static int virtnet_close(struct net_device *dev)
3646{
3647 struct virtnet_info *vi = netdev_priv(dev);
3648 int i;
3649
3650 /* Make sure NAPI doesn't schedule refill work */
3651 disable_delayed_refill(vi);
3652 /* Make sure refill_work doesn't re-enable napi! */
3653 cancel_delayed_work_sync(&vi->refill);
3654 /* Prevent the config change callback from changing carrier
3655 * after close
3656 */
3657 virtio_config_driver_disable(vi->vdev);
3658 /* Stop getting status/speed updates: we don't care until next
3659 * open
3660 */
3661 cancel_work_sync(&vi->config_work);
3662
3663 for (i = 0; i < vi->max_queue_pairs; i++) {
3664 virtnet_disable_queue_pair(vi, i);
3665 virtnet_cancel_dim(vi, &vi->rq[i].dim);
3666 }
3667
3668 netif_carrier_off(dev);
3669
3670 return 0;
3671}
3672
3673static void virtnet_rx_mode_work(struct work_struct *work)
3674{
3675 struct virtnet_info *vi =
3676 container_of(work, struct virtnet_info, rx_mode_work);
3677 u8 *promisc_allmulti __free(kfree) = NULL;
3678 struct net_device *dev = vi->dev;
3679 struct scatterlist sg[2];
3680 struct virtio_net_ctrl_mac *mac_data;
3681 struct netdev_hw_addr *ha;
3682 int uc_count;
3683 int mc_count;
3684 void *buf;
3685 int i;
3686
3687 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
3688 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
3689 return;
3690
3691 promisc_allmulti = kzalloc(sizeof(*promisc_allmulti), GFP_KERNEL);
3692 if (!promisc_allmulti) {
3693 dev_warn(&dev->dev, "Failed to set RX mode, no memory.\n");
3694 return;
3695 }
3696
3697 rtnl_lock();
3698
3699 *promisc_allmulti = !!(dev->flags & IFF_PROMISC);
3700 sg_init_one(sg, promisc_allmulti, sizeof(*promisc_allmulti));
3701
3702 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
3703 VIRTIO_NET_CTRL_RX_PROMISC, sg))
3704 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
3705 *promisc_allmulti ? "en" : "dis");
3706
3707 *promisc_allmulti = !!(dev->flags & IFF_ALLMULTI);
3708 sg_init_one(sg, promisc_allmulti, sizeof(*promisc_allmulti));
3709
3710 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
3711 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
3712 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
3713 *promisc_allmulti ? "en" : "dis");
3714
3715 netif_addr_lock_bh(dev);
3716
3717 uc_count = netdev_uc_count(dev);
3718 mc_count = netdev_mc_count(dev);
3719 /* MAC filter - use one buffer for both lists */
3720 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
3721 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
3722 mac_data = buf;
3723 if (!buf) {
3724 netif_addr_unlock_bh(dev);
3725 rtnl_unlock();
3726 return;
3727 }
3728
3729 sg_init_table(sg, 2);
3730
3731 /* Store the unicast list and count in the front of the buffer */
3732 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
3733 i = 0;
3734 netdev_for_each_uc_addr(ha, dev)
3735 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
3736
3737 sg_set_buf(&sg[0], mac_data,
3738 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
3739
3740 /* multicast list and count fill the end */
3741 mac_data = (void *)&mac_data->macs[uc_count][0];
3742
3743 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
3744 i = 0;
3745 netdev_for_each_mc_addr(ha, dev)
3746 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
3747
3748 netif_addr_unlock_bh(dev);
3749
3750 sg_set_buf(&sg[1], mac_data,
3751 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
3752
3753 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
3754 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
3755 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
3756
3757 rtnl_unlock();
3758
3759 kfree(buf);
3760}
3761
3762static void virtnet_set_rx_mode(struct net_device *dev)
3763{
3764 struct virtnet_info *vi = netdev_priv(dev);
3765
3766 if (vi->rx_mode_work_enabled)
3767 schedule_work(&vi->rx_mode_work);
3768}
3769
3770static int virtnet_vlan_rx_add_vid(struct net_device *dev,
3771 __be16 proto, u16 vid)
3772{
3773 struct virtnet_info *vi = netdev_priv(dev);
3774 __virtio16 *_vid __free(kfree) = NULL;
3775 struct scatterlist sg;
3776
3777 _vid = kzalloc(sizeof(*_vid), GFP_KERNEL);
3778 if (!_vid)
3779 return -ENOMEM;
3780
3781 *_vid = cpu_to_virtio16(vi->vdev, vid);
3782 sg_init_one(&sg, _vid, sizeof(*_vid));
3783
3784 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
3785 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
3786 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
3787 return 0;
3788}
3789
3790static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
3791 __be16 proto, u16 vid)
3792{
3793 struct virtnet_info *vi = netdev_priv(dev);
3794 __virtio16 *_vid __free(kfree) = NULL;
3795 struct scatterlist sg;
3796
3797 _vid = kzalloc(sizeof(*_vid), GFP_KERNEL);
3798 if (!_vid)
3799 return -ENOMEM;
3800
3801 *_vid = cpu_to_virtio16(vi->vdev, vid);
3802 sg_init_one(&sg, _vid, sizeof(*_vid));
3803
3804 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
3805 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
3806 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
3807 return 0;
3808}
3809
3810static void virtnet_clean_affinity(struct virtnet_info *vi)
3811{
3812 int i;
3813
3814 if (vi->affinity_hint_set) {
3815 for (i = 0; i < vi->max_queue_pairs; i++) {
3816 virtqueue_set_affinity(vi->rq[i].vq, NULL);
3817 virtqueue_set_affinity(vi->sq[i].vq, NULL);
3818 }
3819
3820 vi->affinity_hint_set = false;
3821 }
3822}
3823
3824static void virtnet_set_affinity(struct virtnet_info *vi)
3825{
3826 cpumask_var_t mask;
3827 int stragglers;
3828 int group_size;
3829 int i, j, cpu;
3830 int num_cpu;
3831 int stride;
3832
3833 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
3834 virtnet_clean_affinity(vi);
3835 return;
3836 }
3837
3838 num_cpu = num_online_cpus();
3839 stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
3840 stragglers = num_cpu >= vi->curr_queue_pairs ?
3841 num_cpu % vi->curr_queue_pairs :
3842 0;
3843 cpu = cpumask_first(cpu_online_mask);
3844
3845 for (i = 0; i < vi->curr_queue_pairs; i++) {
3846 group_size = stride + (i < stragglers ? 1 : 0);
3847
3848 for (j = 0; j < group_size; j++) {
3849 cpumask_set_cpu(cpu, mask);
3850 cpu = cpumask_next_wrap(cpu, cpu_online_mask,
3851 nr_cpu_ids, false);
3852 }
3853 virtqueue_set_affinity(vi->rq[i].vq, mask);
3854 virtqueue_set_affinity(vi->sq[i].vq, mask);
3855 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
3856 cpumask_clear(mask);
3857 }
3858
3859 vi->affinity_hint_set = true;
3860 free_cpumask_var(mask);
3861}
3862
3863static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
3864{
3865 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
3866 node);
3867 virtnet_set_affinity(vi);
3868 return 0;
3869}
3870
3871static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
3872{
3873 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
3874 node_dead);
3875 virtnet_set_affinity(vi);
3876 return 0;
3877}
3878
3879static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
3880{
3881 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
3882 node);
3883
3884 virtnet_clean_affinity(vi);
3885 return 0;
3886}
3887
3888static enum cpuhp_state virtionet_online;
3889
3890static int virtnet_cpu_notif_add(struct virtnet_info *vi)
3891{
3892 int ret;
3893
3894 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
3895 if (ret)
3896 return ret;
3897 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
3898 &vi->node_dead);
3899 if (!ret)
3900 return ret;
3901 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
3902 return ret;
3903}
3904
3905static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
3906{
3907 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
3908 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
3909 &vi->node_dead);
3910}
3911
3912static int virtnet_send_ctrl_coal_vq_cmd(struct virtnet_info *vi,
3913 u16 vqn, u32 max_usecs, u32 max_packets)
3914{
3915 struct virtio_net_ctrl_coal_vq *coal_vq __free(kfree) = NULL;
3916 struct scatterlist sgs;
3917
3918 coal_vq = kzalloc(sizeof(*coal_vq), GFP_KERNEL);
3919 if (!coal_vq)
3920 return -ENOMEM;
3921
3922 coal_vq->vqn = cpu_to_le16(vqn);
3923 coal_vq->coal.max_usecs = cpu_to_le32(max_usecs);
3924 coal_vq->coal.max_packets = cpu_to_le32(max_packets);
3925 sg_init_one(&sgs, coal_vq, sizeof(*coal_vq));
3926
3927 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
3928 VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET,
3929 &sgs))
3930 return -EINVAL;
3931
3932 return 0;
3933}
3934
3935static int virtnet_send_rx_ctrl_coal_vq_cmd(struct virtnet_info *vi,
3936 u16 queue, u32 max_usecs,
3937 u32 max_packets)
3938{
3939 int err;
3940
3941 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL))
3942 return -EOPNOTSUPP;
3943
3944 err = virtnet_send_ctrl_coal_vq_cmd(vi, rxq2vq(queue),
3945 max_usecs, max_packets);
3946 if (err)
3947 return err;
3948
3949 vi->rq[queue].intr_coal.max_usecs = max_usecs;
3950 vi->rq[queue].intr_coal.max_packets = max_packets;
3951
3952 return 0;
3953}
3954
3955static int virtnet_send_tx_ctrl_coal_vq_cmd(struct virtnet_info *vi,
3956 u16 queue, u32 max_usecs,
3957 u32 max_packets)
3958{
3959 int err;
3960
3961 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL))
3962 return -EOPNOTSUPP;
3963
3964 err = virtnet_send_ctrl_coal_vq_cmd(vi, txq2vq(queue),
3965 max_usecs, max_packets);
3966 if (err)
3967 return err;
3968
3969 vi->sq[queue].intr_coal.max_usecs = max_usecs;
3970 vi->sq[queue].intr_coal.max_packets = max_packets;
3971
3972 return 0;
3973}
3974
3975static void virtnet_get_ringparam(struct net_device *dev,
3976 struct ethtool_ringparam *ring,
3977 struct kernel_ethtool_ringparam *kernel_ring,
3978 struct netlink_ext_ack *extack)
3979{
3980 struct virtnet_info *vi = netdev_priv(dev);
3981
3982 ring->rx_max_pending = vi->rq[0].vq->num_max;
3983 ring->tx_max_pending = vi->sq[0].vq->num_max;
3984 ring->rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
3985 ring->tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
3986}
3987
3988static int virtnet_set_ringparam(struct net_device *dev,
3989 struct ethtool_ringparam *ring,
3990 struct kernel_ethtool_ringparam *kernel_ring,
3991 struct netlink_ext_ack *extack)
3992{
3993 struct virtnet_info *vi = netdev_priv(dev);
3994 u32 rx_pending, tx_pending;
3995 struct receive_queue *rq;
3996 struct send_queue *sq;
3997 int i, err;
3998
3999 if (ring->rx_mini_pending || ring->rx_jumbo_pending)
4000 return -EINVAL;
4001
4002 rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
4003 tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
4004
4005 if (ring->rx_pending == rx_pending &&
4006 ring->tx_pending == tx_pending)
4007 return 0;
4008
4009 if (ring->rx_pending > vi->rq[0].vq->num_max)
4010 return -EINVAL;
4011
4012 if (ring->tx_pending > vi->sq[0].vq->num_max)
4013 return -EINVAL;
4014
4015 for (i = 0; i < vi->max_queue_pairs; i++) {
4016 rq = vi->rq + i;
4017 sq = vi->sq + i;
4018
4019 if (ring->tx_pending != tx_pending) {
4020 err = virtnet_tx_resize(vi, sq, ring->tx_pending);
4021 if (err)
4022 return err;
4023
4024 /* Upon disabling and re-enabling a transmit virtqueue, the device must
4025 * set the coalescing parameters of the virtqueue to those configured
4026 * through the VIRTIO_NET_CTRL_NOTF_COAL_TX_SET command, or, if the driver
4027 * did not set any TX coalescing parameters, to 0.
4028 */
4029 err = virtnet_send_tx_ctrl_coal_vq_cmd(vi, i,
4030 vi->intr_coal_tx.max_usecs,
4031 vi->intr_coal_tx.max_packets);
4032
4033 /* Don't break the tx resize action if the vq coalescing is not
4034 * supported. The same is true for rx resize below.
4035 */
4036 if (err && err != -EOPNOTSUPP)
4037 return err;
4038 }
4039
4040 if (ring->rx_pending != rx_pending) {
4041 err = virtnet_rx_resize(vi, rq, ring->rx_pending);
4042 if (err)
4043 return err;
4044
4045 /* The reason is same as the transmit virtqueue reset */
4046 mutex_lock(&vi->rq[i].dim_lock);
4047 err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, i,
4048 vi->intr_coal_rx.max_usecs,
4049 vi->intr_coal_rx.max_packets);
4050 mutex_unlock(&vi->rq[i].dim_lock);
4051 if (err && err != -EOPNOTSUPP)
4052 return err;
4053 }
4054 }
4055
4056 return 0;
4057}
4058
4059static bool virtnet_commit_rss_command(struct virtnet_info *vi)
4060{
4061 struct net_device *dev = vi->dev;
4062 struct scatterlist sgs[4];
4063 unsigned int sg_buf_size;
4064
4065 /* prepare sgs */
4066 sg_init_table(sgs, 4);
4067
4068 sg_buf_size = offsetof(struct virtio_net_ctrl_rss, hash_cfg_reserved);
4069 sg_set_buf(&sgs[0], &vi->rss, sg_buf_size);
4070
4071 if (vi->has_rss) {
4072 sg_buf_size = sizeof(uint16_t) * vi->rss_indir_table_size;
4073 sg_set_buf(&sgs[1], vi->rss.indirection_table, sg_buf_size);
4074 } else {
4075 sg_set_buf(&sgs[1], &vi->rss.hash_cfg_reserved, sizeof(uint16_t));
4076 }
4077
4078 sg_buf_size = offsetof(struct virtio_net_ctrl_rss, key)
4079 - offsetof(struct virtio_net_ctrl_rss, max_tx_vq);
4080 sg_set_buf(&sgs[2], &vi->rss.max_tx_vq, sg_buf_size);
4081
4082 sg_buf_size = vi->rss_key_size;
4083 sg_set_buf(&sgs[3], vi->rss.key, sg_buf_size);
4084
4085 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
4086 vi->has_rss ? VIRTIO_NET_CTRL_MQ_RSS_CONFIG
4087 : VIRTIO_NET_CTRL_MQ_HASH_CONFIG, sgs))
4088 goto err;
4089
4090 return true;
4091
4092err:
4093 dev_warn(&dev->dev, "VIRTIONET issue with committing RSS sgs\n");
4094 return false;
4095
4096}
4097
4098static void virtnet_init_default_rss(struct virtnet_info *vi)
4099{
4100 vi->rss.hash_types = vi->rss_hash_types_supported;
4101 vi->rss_hash_types_saved = vi->rss_hash_types_supported;
4102 vi->rss.indirection_table_mask = vi->rss_indir_table_size
4103 ? vi->rss_indir_table_size - 1 : 0;
4104 vi->rss.unclassified_queue = 0;
4105
4106 virtnet_rss_update_by_qpairs(vi, vi->curr_queue_pairs);
4107
4108 vi->rss.hash_key_length = vi->rss_key_size;
4109
4110 netdev_rss_key_fill(vi->rss.key, vi->rss_key_size);
4111}
4112
4113static void virtnet_get_hashflow(const struct virtnet_info *vi, struct ethtool_rxnfc *info)
4114{
4115 info->data = 0;
4116 switch (info->flow_type) {
4117 case TCP_V4_FLOW:
4118 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv4) {
4119 info->data = RXH_IP_SRC | RXH_IP_DST |
4120 RXH_L4_B_0_1 | RXH_L4_B_2_3;
4121 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
4122 info->data = RXH_IP_SRC | RXH_IP_DST;
4123 }
4124 break;
4125 case TCP_V6_FLOW:
4126 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv6) {
4127 info->data = RXH_IP_SRC | RXH_IP_DST |
4128 RXH_L4_B_0_1 | RXH_L4_B_2_3;
4129 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
4130 info->data = RXH_IP_SRC | RXH_IP_DST;
4131 }
4132 break;
4133 case UDP_V4_FLOW:
4134 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv4) {
4135 info->data = RXH_IP_SRC | RXH_IP_DST |
4136 RXH_L4_B_0_1 | RXH_L4_B_2_3;
4137 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
4138 info->data = RXH_IP_SRC | RXH_IP_DST;
4139 }
4140 break;
4141 case UDP_V6_FLOW:
4142 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv6) {
4143 info->data = RXH_IP_SRC | RXH_IP_DST |
4144 RXH_L4_B_0_1 | RXH_L4_B_2_3;
4145 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
4146 info->data = RXH_IP_SRC | RXH_IP_DST;
4147 }
4148 break;
4149 case IPV4_FLOW:
4150 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4)
4151 info->data = RXH_IP_SRC | RXH_IP_DST;
4152
4153 break;
4154 case IPV6_FLOW:
4155 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6)
4156 info->data = RXH_IP_SRC | RXH_IP_DST;
4157
4158 break;
4159 default:
4160 info->data = 0;
4161 break;
4162 }
4163}
4164
4165static bool virtnet_set_hashflow(struct virtnet_info *vi, struct ethtool_rxnfc *info)
4166{
4167 u32 new_hashtypes = vi->rss_hash_types_saved;
4168 bool is_disable = info->data & RXH_DISCARD;
4169 bool is_l4 = info->data == (RXH_IP_SRC | RXH_IP_DST | RXH_L4_B_0_1 | RXH_L4_B_2_3);
4170
4171 /* supports only 'sd', 'sdfn' and 'r' */
4172 if (!((info->data == (RXH_IP_SRC | RXH_IP_DST)) | is_l4 | is_disable))
4173 return false;
4174
4175 switch (info->flow_type) {
4176 case TCP_V4_FLOW:
4177 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_TCPv4);
4178 if (!is_disable)
4179 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
4180 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv4 : 0);
4181 break;
4182 case UDP_V4_FLOW:
4183 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_UDPv4);
4184 if (!is_disable)
4185 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
4186 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv4 : 0);
4187 break;
4188 case IPV4_FLOW:
4189 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv4;
4190 if (!is_disable)
4191 new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv4;
4192 break;
4193 case TCP_V6_FLOW:
4194 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_TCPv6);
4195 if (!is_disable)
4196 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
4197 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv6 : 0);
4198 break;
4199 case UDP_V6_FLOW:
4200 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_UDPv6);
4201 if (!is_disable)
4202 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
4203 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv6 : 0);
4204 break;
4205 case IPV6_FLOW:
4206 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv6;
4207 if (!is_disable)
4208 new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv6;
4209 break;
4210 default:
4211 /* unsupported flow */
4212 return false;
4213 }
4214
4215 /* if unsupported hashtype was set */
4216 if (new_hashtypes != (new_hashtypes & vi->rss_hash_types_supported))
4217 return false;
4218
4219 if (new_hashtypes != vi->rss_hash_types_saved) {
4220 vi->rss_hash_types_saved = new_hashtypes;
4221 vi->rss.hash_types = vi->rss_hash_types_saved;
4222 if (vi->dev->features & NETIF_F_RXHASH)
4223 return virtnet_commit_rss_command(vi);
4224 }
4225
4226 return true;
4227}
4228
4229static void virtnet_get_drvinfo(struct net_device *dev,
4230 struct ethtool_drvinfo *info)
4231{
4232 struct virtnet_info *vi = netdev_priv(dev);
4233 struct virtio_device *vdev = vi->vdev;
4234
4235 strscpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
4236 strscpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
4237 strscpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
4238
4239}
4240
4241/* TODO: Eliminate OOO packets during switching */
4242static int virtnet_set_channels(struct net_device *dev,
4243 struct ethtool_channels *channels)
4244{
4245 struct virtnet_info *vi = netdev_priv(dev);
4246 u16 queue_pairs = channels->combined_count;
4247 int err;
4248
4249 /* We don't support separate rx/tx channels.
4250 * We don't allow setting 'other' channels.
4251 */
4252 if (channels->rx_count || channels->tx_count || channels->other_count)
4253 return -EINVAL;
4254
4255 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
4256 return -EINVAL;
4257
4258 /* For now we don't support modifying channels while XDP is loaded
4259 * also when XDP is loaded all RX queues have XDP programs so we only
4260 * need to check a single RX queue.
4261 */
4262 if (vi->rq[0].xdp_prog)
4263 return -EINVAL;
4264
4265 cpus_read_lock();
4266 err = virtnet_set_queues(vi, queue_pairs);
4267 if (err) {
4268 cpus_read_unlock();
4269 goto err;
4270 }
4271 virtnet_set_affinity(vi);
4272 cpus_read_unlock();
4273
4274 netif_set_real_num_tx_queues(dev, queue_pairs);
4275 netif_set_real_num_rx_queues(dev, queue_pairs);
4276 err:
4277 return err;
4278}
4279
4280static void virtnet_stats_sprintf(u8 **p, const char *fmt, const char *noq_fmt,
4281 int num, int qid, const struct virtnet_stat_desc *desc)
4282{
4283 int i;
4284
4285 if (qid < 0) {
4286 for (i = 0; i < num; ++i)
4287 ethtool_sprintf(p, noq_fmt, desc[i].desc);
4288 } else {
4289 for (i = 0; i < num; ++i)
4290 ethtool_sprintf(p, fmt, qid, desc[i].desc);
4291 }
4292}
4293
4294/* qid == -1: for rx/tx queue total field */
4295static void virtnet_get_stats_string(struct virtnet_info *vi, int type, int qid, u8 **data)
4296{
4297 const struct virtnet_stat_desc *desc;
4298 const char *fmt, *noq_fmt;
4299 u8 *p = *data;
4300 u32 num;
4301
4302 if (type == VIRTNET_Q_TYPE_CQ && qid >= 0) {
4303 noq_fmt = "cq_hw_%s";
4304
4305 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_CVQ) {
4306 desc = &virtnet_stats_cvq_desc[0];
4307 num = ARRAY_SIZE(virtnet_stats_cvq_desc);
4308
4309 virtnet_stats_sprintf(&p, NULL, noq_fmt, num, -1, desc);
4310 }
4311 }
4312
4313 if (type == VIRTNET_Q_TYPE_RX) {
4314 fmt = "rx%u_%s";
4315 noq_fmt = "rx_%s";
4316
4317 desc = &virtnet_rq_stats_desc[0];
4318 num = ARRAY_SIZE(virtnet_rq_stats_desc);
4319
4320 virtnet_stats_sprintf(&p, fmt, noq_fmt, num, qid, desc);
4321
4322 fmt = "rx%u_hw_%s";
4323 noq_fmt = "rx_hw_%s";
4324
4325 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_RX_BASIC) {
4326 desc = &virtnet_stats_rx_basic_desc[0];
4327 num = ARRAY_SIZE(virtnet_stats_rx_basic_desc);
4328
4329 virtnet_stats_sprintf(&p, fmt, noq_fmt, num, qid, desc);
4330 }
4331
4332 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_RX_CSUM) {
4333 desc = &virtnet_stats_rx_csum_desc[0];
4334 num = ARRAY_SIZE(virtnet_stats_rx_csum_desc);
4335
4336 virtnet_stats_sprintf(&p, fmt, noq_fmt, num, qid, desc);
4337 }
4338
4339 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_RX_SPEED) {
4340 desc = &virtnet_stats_rx_speed_desc[0];
4341 num = ARRAY_SIZE(virtnet_stats_rx_speed_desc);
4342
4343 virtnet_stats_sprintf(&p, fmt, noq_fmt, num, qid, desc);
4344 }
4345 }
4346
4347 if (type == VIRTNET_Q_TYPE_TX) {
4348 fmt = "tx%u_%s";
4349 noq_fmt = "tx_%s";
4350
4351 desc = &virtnet_sq_stats_desc[0];
4352 num = ARRAY_SIZE(virtnet_sq_stats_desc);
4353
4354 virtnet_stats_sprintf(&p, fmt, noq_fmt, num, qid, desc);
4355
4356 fmt = "tx%u_hw_%s";
4357 noq_fmt = "tx_hw_%s";
4358
4359 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_TX_BASIC) {
4360 desc = &virtnet_stats_tx_basic_desc[0];
4361 num = ARRAY_SIZE(virtnet_stats_tx_basic_desc);
4362
4363 virtnet_stats_sprintf(&p, fmt, noq_fmt, num, qid, desc);
4364 }
4365
4366 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_TX_GSO) {
4367 desc = &virtnet_stats_tx_gso_desc[0];
4368 num = ARRAY_SIZE(virtnet_stats_tx_gso_desc);
4369
4370 virtnet_stats_sprintf(&p, fmt, noq_fmt, num, qid, desc);
4371 }
4372
4373 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_TX_SPEED) {
4374 desc = &virtnet_stats_tx_speed_desc[0];
4375 num = ARRAY_SIZE(virtnet_stats_tx_speed_desc);
4376
4377 virtnet_stats_sprintf(&p, fmt, noq_fmt, num, qid, desc);
4378 }
4379 }
4380
4381 *data = p;
4382}
4383
4384struct virtnet_stats_ctx {
4385 /* The stats are write to qstats or ethtool -S */
4386 bool to_qstat;
4387
4388 /* Used to calculate the offset inside the output buffer. */
4389 u32 desc_num[3];
4390
4391 /* The actual supported stat types. */
4392 u64 bitmap[3];
4393
4394 /* Used to calculate the reply buffer size. */
4395 u32 size[3];
4396
4397 /* Record the output buffer. */
4398 u64 *data;
4399};
4400
4401static void virtnet_stats_ctx_init(struct virtnet_info *vi,
4402 struct virtnet_stats_ctx *ctx,
4403 u64 *data, bool to_qstat)
4404{
4405 u32 queue_type;
4406
4407 ctx->data = data;
4408 ctx->to_qstat = to_qstat;
4409
4410 if (to_qstat) {
4411 ctx->desc_num[VIRTNET_Q_TYPE_RX] = ARRAY_SIZE(virtnet_rq_stats_desc_qstat);
4412 ctx->desc_num[VIRTNET_Q_TYPE_TX] = ARRAY_SIZE(virtnet_sq_stats_desc_qstat);
4413
4414 queue_type = VIRTNET_Q_TYPE_RX;
4415
4416 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_RX_BASIC) {
4417 ctx->bitmap[queue_type] |= VIRTIO_NET_STATS_TYPE_RX_BASIC;
4418 ctx->desc_num[queue_type] += ARRAY_SIZE(virtnet_stats_rx_basic_desc_qstat);
4419 ctx->size[queue_type] += sizeof(struct virtio_net_stats_rx_basic);
4420 }
4421
4422 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_RX_CSUM) {
4423 ctx->bitmap[queue_type] |= VIRTIO_NET_STATS_TYPE_RX_CSUM;
4424 ctx->desc_num[queue_type] += ARRAY_SIZE(virtnet_stats_rx_csum_desc_qstat);
4425 ctx->size[queue_type] += sizeof(struct virtio_net_stats_rx_csum);
4426 }
4427
4428 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_RX_GSO) {
4429 ctx->bitmap[queue_type] |= VIRTIO_NET_STATS_TYPE_RX_GSO;
4430 ctx->desc_num[queue_type] += ARRAY_SIZE(virtnet_stats_rx_gso_desc_qstat);
4431 ctx->size[queue_type] += sizeof(struct virtio_net_stats_rx_gso);
4432 }
4433
4434 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_RX_SPEED) {
4435 ctx->bitmap[queue_type] |= VIRTIO_NET_STATS_TYPE_RX_SPEED;
4436 ctx->desc_num[queue_type] += ARRAY_SIZE(virtnet_stats_rx_speed_desc_qstat);
4437 ctx->size[queue_type] += sizeof(struct virtio_net_stats_rx_speed);
4438 }
4439
4440 queue_type = VIRTNET_Q_TYPE_TX;
4441
4442 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_TX_BASIC) {
4443 ctx->bitmap[queue_type] |= VIRTIO_NET_STATS_TYPE_TX_BASIC;
4444 ctx->desc_num[queue_type] += ARRAY_SIZE(virtnet_stats_tx_basic_desc_qstat);
4445 ctx->size[queue_type] += sizeof(struct virtio_net_stats_tx_basic);
4446 }
4447
4448 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_TX_CSUM) {
4449 ctx->bitmap[queue_type] |= VIRTIO_NET_STATS_TYPE_TX_CSUM;
4450 ctx->desc_num[queue_type] += ARRAY_SIZE(virtnet_stats_tx_csum_desc_qstat);
4451 ctx->size[queue_type] += sizeof(struct virtio_net_stats_tx_csum);
4452 }
4453
4454 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_TX_GSO) {
4455 ctx->bitmap[queue_type] |= VIRTIO_NET_STATS_TYPE_TX_GSO;
4456 ctx->desc_num[queue_type] += ARRAY_SIZE(virtnet_stats_tx_gso_desc_qstat);
4457 ctx->size[queue_type] += sizeof(struct virtio_net_stats_tx_gso);
4458 }
4459
4460 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_TX_SPEED) {
4461 ctx->bitmap[queue_type] |= VIRTIO_NET_STATS_TYPE_TX_SPEED;
4462 ctx->desc_num[queue_type] += ARRAY_SIZE(virtnet_stats_tx_speed_desc_qstat);
4463 ctx->size[queue_type] += sizeof(struct virtio_net_stats_tx_speed);
4464 }
4465
4466 return;
4467 }
4468
4469 ctx->desc_num[VIRTNET_Q_TYPE_RX] = ARRAY_SIZE(virtnet_rq_stats_desc);
4470 ctx->desc_num[VIRTNET_Q_TYPE_TX] = ARRAY_SIZE(virtnet_sq_stats_desc);
4471
4472 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_CVQ) {
4473 queue_type = VIRTNET_Q_TYPE_CQ;
4474
4475 ctx->bitmap[queue_type] |= VIRTIO_NET_STATS_TYPE_CVQ;
4476 ctx->desc_num[queue_type] += ARRAY_SIZE(virtnet_stats_cvq_desc);
4477 ctx->size[queue_type] += sizeof(struct virtio_net_stats_cvq);
4478 }
4479
4480 queue_type = VIRTNET_Q_TYPE_RX;
4481
4482 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_RX_BASIC) {
4483 ctx->bitmap[queue_type] |= VIRTIO_NET_STATS_TYPE_RX_BASIC;
4484 ctx->desc_num[queue_type] += ARRAY_SIZE(virtnet_stats_rx_basic_desc);
4485 ctx->size[queue_type] += sizeof(struct virtio_net_stats_rx_basic);
4486 }
4487
4488 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_RX_CSUM) {
4489 ctx->bitmap[queue_type] |= VIRTIO_NET_STATS_TYPE_RX_CSUM;
4490 ctx->desc_num[queue_type] += ARRAY_SIZE(virtnet_stats_rx_csum_desc);
4491 ctx->size[queue_type] += sizeof(struct virtio_net_stats_rx_csum);
4492 }
4493
4494 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_RX_SPEED) {
4495 ctx->bitmap[queue_type] |= VIRTIO_NET_STATS_TYPE_RX_SPEED;
4496 ctx->desc_num[queue_type] += ARRAY_SIZE(virtnet_stats_rx_speed_desc);
4497 ctx->size[queue_type] += sizeof(struct virtio_net_stats_rx_speed);
4498 }
4499
4500 queue_type = VIRTNET_Q_TYPE_TX;
4501
4502 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_TX_BASIC) {
4503 ctx->bitmap[queue_type] |= VIRTIO_NET_STATS_TYPE_TX_BASIC;
4504 ctx->desc_num[queue_type] += ARRAY_SIZE(virtnet_stats_tx_basic_desc);
4505 ctx->size[queue_type] += sizeof(struct virtio_net_stats_tx_basic);
4506 }
4507
4508 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_TX_GSO) {
4509 ctx->bitmap[queue_type] |= VIRTIO_NET_STATS_TYPE_TX_GSO;
4510 ctx->desc_num[queue_type] += ARRAY_SIZE(virtnet_stats_tx_gso_desc);
4511 ctx->size[queue_type] += sizeof(struct virtio_net_stats_tx_gso);
4512 }
4513
4514 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_TX_SPEED) {
4515 ctx->bitmap[queue_type] |= VIRTIO_NET_STATS_TYPE_TX_SPEED;
4516 ctx->desc_num[queue_type] += ARRAY_SIZE(virtnet_stats_tx_speed_desc);
4517 ctx->size[queue_type] += sizeof(struct virtio_net_stats_tx_speed);
4518 }
4519}
4520
4521/* stats_sum_queue - Calculate the sum of the same fields in sq or rq.
4522 * @sum: the position to store the sum values
4523 * @num: field num
4524 * @q_value: the first queue fields
4525 * @q_num: number of the queues
4526 */
4527static void stats_sum_queue(u64 *sum, u32 num, u64 *q_value, u32 q_num)
4528{
4529 u32 step = num;
4530 int i, j;
4531 u64 *p;
4532
4533 for (i = 0; i < num; ++i) {
4534 p = sum + i;
4535 *p = 0;
4536
4537 for (j = 0; j < q_num; ++j)
4538 *p += *(q_value + i + j * step);
4539 }
4540}
4541
4542static void virtnet_fill_total_fields(struct virtnet_info *vi,
4543 struct virtnet_stats_ctx *ctx)
4544{
4545 u64 *data, *first_rx_q, *first_tx_q;
4546 u32 num_cq, num_rx, num_tx;
4547
4548 num_cq = ctx->desc_num[VIRTNET_Q_TYPE_CQ];
4549 num_rx = ctx->desc_num[VIRTNET_Q_TYPE_RX];
4550 num_tx = ctx->desc_num[VIRTNET_Q_TYPE_TX];
4551
4552 first_rx_q = ctx->data + num_rx + num_tx + num_cq;
4553 first_tx_q = first_rx_q + vi->curr_queue_pairs * num_rx;
4554
4555 data = ctx->data;
4556
4557 stats_sum_queue(data, num_rx, first_rx_q, vi->curr_queue_pairs);
4558
4559 data = ctx->data + num_rx;
4560
4561 stats_sum_queue(data, num_tx, first_tx_q, vi->curr_queue_pairs);
4562}
4563
4564static void virtnet_fill_stats_qstat(struct virtnet_info *vi, u32 qid,
4565 struct virtnet_stats_ctx *ctx,
4566 const u8 *base, bool drv_stats, u8 reply_type)
4567{
4568 const struct virtnet_stat_desc *desc;
4569 const u64_stats_t *v_stat;
4570 u64 offset, bitmap;
4571 const __le64 *v;
4572 u32 queue_type;
4573 int i, num;
4574
4575 queue_type = vq_type(vi, qid);
4576 bitmap = ctx->bitmap[queue_type];
4577
4578 if (drv_stats) {
4579 if (queue_type == VIRTNET_Q_TYPE_RX) {
4580 desc = &virtnet_rq_stats_desc_qstat[0];
4581 num = ARRAY_SIZE(virtnet_rq_stats_desc_qstat);
4582 } else {
4583 desc = &virtnet_sq_stats_desc_qstat[0];
4584 num = ARRAY_SIZE(virtnet_sq_stats_desc_qstat);
4585 }
4586
4587 for (i = 0; i < num; ++i) {
4588 offset = desc[i].qstat_offset / sizeof(*ctx->data);
4589 v_stat = (const u64_stats_t *)(base + desc[i].offset);
4590 ctx->data[offset] = u64_stats_read(v_stat);
4591 }
4592 return;
4593 }
4594
4595 if (bitmap & VIRTIO_NET_STATS_TYPE_RX_BASIC) {
4596 desc = &virtnet_stats_rx_basic_desc_qstat[0];
4597 num = ARRAY_SIZE(virtnet_stats_rx_basic_desc_qstat);
4598 if (reply_type == VIRTIO_NET_STATS_TYPE_REPLY_RX_BASIC)
4599 goto found;
4600 }
4601
4602 if (bitmap & VIRTIO_NET_STATS_TYPE_RX_CSUM) {
4603 desc = &virtnet_stats_rx_csum_desc_qstat[0];
4604 num = ARRAY_SIZE(virtnet_stats_rx_csum_desc_qstat);
4605 if (reply_type == VIRTIO_NET_STATS_TYPE_REPLY_RX_CSUM)
4606 goto found;
4607 }
4608
4609 if (bitmap & VIRTIO_NET_STATS_TYPE_RX_GSO) {
4610 desc = &virtnet_stats_rx_gso_desc_qstat[0];
4611 num = ARRAY_SIZE(virtnet_stats_rx_gso_desc_qstat);
4612 if (reply_type == VIRTIO_NET_STATS_TYPE_REPLY_RX_GSO)
4613 goto found;
4614 }
4615
4616 if (bitmap & VIRTIO_NET_STATS_TYPE_RX_SPEED) {
4617 desc = &virtnet_stats_rx_speed_desc_qstat[0];
4618 num = ARRAY_SIZE(virtnet_stats_rx_speed_desc_qstat);
4619 if (reply_type == VIRTIO_NET_STATS_TYPE_REPLY_RX_SPEED)
4620 goto found;
4621 }
4622
4623 if (bitmap & VIRTIO_NET_STATS_TYPE_TX_BASIC) {
4624 desc = &virtnet_stats_tx_basic_desc_qstat[0];
4625 num = ARRAY_SIZE(virtnet_stats_tx_basic_desc_qstat);
4626 if (reply_type == VIRTIO_NET_STATS_TYPE_REPLY_TX_BASIC)
4627 goto found;
4628 }
4629
4630 if (bitmap & VIRTIO_NET_STATS_TYPE_TX_CSUM) {
4631 desc = &virtnet_stats_tx_csum_desc_qstat[0];
4632 num = ARRAY_SIZE(virtnet_stats_tx_csum_desc_qstat);
4633 if (reply_type == VIRTIO_NET_STATS_TYPE_REPLY_TX_CSUM)
4634 goto found;
4635 }
4636
4637 if (bitmap & VIRTIO_NET_STATS_TYPE_TX_GSO) {
4638 desc = &virtnet_stats_tx_gso_desc_qstat[0];
4639 num = ARRAY_SIZE(virtnet_stats_tx_gso_desc_qstat);
4640 if (reply_type == VIRTIO_NET_STATS_TYPE_REPLY_TX_GSO)
4641 goto found;
4642 }
4643
4644 if (bitmap & VIRTIO_NET_STATS_TYPE_TX_SPEED) {
4645 desc = &virtnet_stats_tx_speed_desc_qstat[0];
4646 num = ARRAY_SIZE(virtnet_stats_tx_speed_desc_qstat);
4647 if (reply_type == VIRTIO_NET_STATS_TYPE_REPLY_TX_SPEED)
4648 goto found;
4649 }
4650
4651 return;
4652
4653found:
4654 for (i = 0; i < num; ++i) {
4655 offset = desc[i].qstat_offset / sizeof(*ctx->data);
4656 v = (const __le64 *)(base + desc[i].offset);
4657 ctx->data[offset] = le64_to_cpu(*v);
4658 }
4659}
4660
4661/* virtnet_fill_stats - copy the stats to qstats or ethtool -S
4662 * The stats source is the device or the driver.
4663 *
4664 * @vi: virtio net info
4665 * @qid: the vq id
4666 * @ctx: stats ctx (initiated by virtnet_stats_ctx_init())
4667 * @base: pointer to the device reply or the driver stats structure.
4668 * @drv_stats: designate the base type (device reply, driver stats)
4669 * @type: the type of the device reply (if drv_stats is true, this must be zero)
4670 */
4671static void virtnet_fill_stats(struct virtnet_info *vi, u32 qid,
4672 struct virtnet_stats_ctx *ctx,
4673 const u8 *base, bool drv_stats, u8 reply_type)
4674{
4675 u32 queue_type, num_rx, num_tx, num_cq;
4676 const struct virtnet_stat_desc *desc;
4677 const u64_stats_t *v_stat;
4678 u64 offset, bitmap;
4679 const __le64 *v;
4680 int i, num;
4681
4682 if (ctx->to_qstat)
4683 return virtnet_fill_stats_qstat(vi, qid, ctx, base, drv_stats, reply_type);
4684
4685 num_cq = ctx->desc_num[VIRTNET_Q_TYPE_CQ];
4686 num_rx = ctx->desc_num[VIRTNET_Q_TYPE_RX];
4687 num_tx = ctx->desc_num[VIRTNET_Q_TYPE_TX];
4688
4689 queue_type = vq_type(vi, qid);
4690 bitmap = ctx->bitmap[queue_type];
4691
4692 /* skip the total fields of pairs */
4693 offset = num_rx + num_tx;
4694
4695 if (queue_type == VIRTNET_Q_TYPE_TX) {
4696 offset += num_cq + num_rx * vi->curr_queue_pairs + num_tx * (qid / 2);
4697
4698 num = ARRAY_SIZE(virtnet_sq_stats_desc);
4699 if (drv_stats) {
4700 desc = &virtnet_sq_stats_desc[0];
4701 goto drv_stats;
4702 }
4703
4704 offset += num;
4705
4706 } else if (queue_type == VIRTNET_Q_TYPE_RX) {
4707 offset += num_cq + num_rx * (qid / 2);
4708
4709 num = ARRAY_SIZE(virtnet_rq_stats_desc);
4710 if (drv_stats) {
4711 desc = &virtnet_rq_stats_desc[0];
4712 goto drv_stats;
4713 }
4714
4715 offset += num;
4716 }
4717
4718 if (bitmap & VIRTIO_NET_STATS_TYPE_CVQ) {
4719 desc = &virtnet_stats_cvq_desc[0];
4720 num = ARRAY_SIZE(virtnet_stats_cvq_desc);
4721 if (reply_type == VIRTIO_NET_STATS_TYPE_REPLY_CVQ)
4722 goto found;
4723
4724 offset += num;
4725 }
4726
4727 if (bitmap & VIRTIO_NET_STATS_TYPE_RX_BASIC) {
4728 desc = &virtnet_stats_rx_basic_desc[0];
4729 num = ARRAY_SIZE(virtnet_stats_rx_basic_desc);
4730 if (reply_type == VIRTIO_NET_STATS_TYPE_REPLY_RX_BASIC)
4731 goto found;
4732
4733 offset += num;
4734 }
4735
4736 if (bitmap & VIRTIO_NET_STATS_TYPE_RX_CSUM) {
4737 desc = &virtnet_stats_rx_csum_desc[0];
4738 num = ARRAY_SIZE(virtnet_stats_rx_csum_desc);
4739 if (reply_type == VIRTIO_NET_STATS_TYPE_REPLY_RX_CSUM)
4740 goto found;
4741
4742 offset += num;
4743 }
4744
4745 if (bitmap & VIRTIO_NET_STATS_TYPE_RX_SPEED) {
4746 desc = &virtnet_stats_rx_speed_desc[0];
4747 num = ARRAY_SIZE(virtnet_stats_rx_speed_desc);
4748 if (reply_type == VIRTIO_NET_STATS_TYPE_REPLY_RX_SPEED)
4749 goto found;
4750
4751 offset += num;
4752 }
4753
4754 if (bitmap & VIRTIO_NET_STATS_TYPE_TX_BASIC) {
4755 desc = &virtnet_stats_tx_basic_desc[0];
4756 num = ARRAY_SIZE(virtnet_stats_tx_basic_desc);
4757 if (reply_type == VIRTIO_NET_STATS_TYPE_REPLY_TX_BASIC)
4758 goto found;
4759
4760 offset += num;
4761 }
4762
4763 if (bitmap & VIRTIO_NET_STATS_TYPE_TX_GSO) {
4764 desc = &virtnet_stats_tx_gso_desc[0];
4765 num = ARRAY_SIZE(virtnet_stats_tx_gso_desc);
4766 if (reply_type == VIRTIO_NET_STATS_TYPE_REPLY_TX_GSO)
4767 goto found;
4768
4769 offset += num;
4770 }
4771
4772 if (bitmap & VIRTIO_NET_STATS_TYPE_TX_SPEED) {
4773 desc = &virtnet_stats_tx_speed_desc[0];
4774 num = ARRAY_SIZE(virtnet_stats_tx_speed_desc);
4775 if (reply_type == VIRTIO_NET_STATS_TYPE_REPLY_TX_SPEED)
4776 goto found;
4777
4778 offset += num;
4779 }
4780
4781 return;
4782
4783found:
4784 for (i = 0; i < num; ++i) {
4785 v = (const __le64 *)(base + desc[i].offset);
4786 ctx->data[offset + i] = le64_to_cpu(*v);
4787 }
4788
4789 return;
4790
4791drv_stats:
4792 for (i = 0; i < num; ++i) {
4793 v_stat = (const u64_stats_t *)(base + desc[i].offset);
4794 ctx->data[offset + i] = u64_stats_read(v_stat);
4795 }
4796}
4797
4798static int __virtnet_get_hw_stats(struct virtnet_info *vi,
4799 struct virtnet_stats_ctx *ctx,
4800 struct virtio_net_ctrl_queue_stats *req,
4801 int req_size, void *reply, int res_size)
4802{
4803 struct virtio_net_stats_reply_hdr *hdr;
4804 struct scatterlist sgs_in, sgs_out;
4805 void *p;
4806 u32 qid;
4807 int ok;
4808
4809 sg_init_one(&sgs_out, req, req_size);
4810 sg_init_one(&sgs_in, reply, res_size);
4811
4812 ok = virtnet_send_command_reply(vi, VIRTIO_NET_CTRL_STATS,
4813 VIRTIO_NET_CTRL_STATS_GET,
4814 &sgs_out, &sgs_in);
4815
4816 if (!ok)
4817 return ok;
4818
4819 for (p = reply; p - reply < res_size; p += le16_to_cpu(hdr->size)) {
4820 hdr = p;
4821 qid = le16_to_cpu(hdr->vq_index);
4822 virtnet_fill_stats(vi, qid, ctx, p, false, hdr->type);
4823 }
4824
4825 return 0;
4826}
4827
4828static void virtnet_make_stat_req(struct virtnet_info *vi,
4829 struct virtnet_stats_ctx *ctx,
4830 struct virtio_net_ctrl_queue_stats *req,
4831 int qid, int *idx)
4832{
4833 int qtype = vq_type(vi, qid);
4834 u64 bitmap = ctx->bitmap[qtype];
4835
4836 if (!bitmap)
4837 return;
4838
4839 req->stats[*idx].vq_index = cpu_to_le16(qid);
4840 req->stats[*idx].types_bitmap[0] = cpu_to_le64(bitmap);
4841 *idx += 1;
4842}
4843
4844/* qid: -1: get stats of all vq.
4845 * > 0: get the stats for the special vq. This must not be cvq.
4846 */
4847static int virtnet_get_hw_stats(struct virtnet_info *vi,
4848 struct virtnet_stats_ctx *ctx, int qid)
4849{
4850 int qnum, i, j, res_size, qtype, last_vq, first_vq;
4851 struct virtio_net_ctrl_queue_stats *req;
4852 bool enable_cvq;
4853 void *reply;
4854 int ok;
4855
4856 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_DEVICE_STATS))
4857 return 0;
4858
4859 if (qid == -1) {
4860 last_vq = vi->curr_queue_pairs * 2 - 1;
4861 first_vq = 0;
4862 enable_cvq = true;
4863 } else {
4864 last_vq = qid;
4865 first_vq = qid;
4866 enable_cvq = false;
4867 }
4868
4869 qnum = 0;
4870 res_size = 0;
4871 for (i = first_vq; i <= last_vq ; ++i) {
4872 qtype = vq_type(vi, i);
4873 if (ctx->bitmap[qtype]) {
4874 ++qnum;
4875 res_size += ctx->size[qtype];
4876 }
4877 }
4878
4879 if (enable_cvq && ctx->bitmap[VIRTNET_Q_TYPE_CQ]) {
4880 res_size += ctx->size[VIRTNET_Q_TYPE_CQ];
4881 qnum += 1;
4882 }
4883
4884 req = kcalloc(qnum, sizeof(*req), GFP_KERNEL);
4885 if (!req)
4886 return -ENOMEM;
4887
4888 reply = kmalloc(res_size, GFP_KERNEL);
4889 if (!reply) {
4890 kfree(req);
4891 return -ENOMEM;
4892 }
4893
4894 j = 0;
4895 for (i = first_vq; i <= last_vq ; ++i)
4896 virtnet_make_stat_req(vi, ctx, req, i, &j);
4897
4898 if (enable_cvq)
4899 virtnet_make_stat_req(vi, ctx, req, vi->max_queue_pairs * 2, &j);
4900
4901 ok = __virtnet_get_hw_stats(vi, ctx, req, sizeof(*req) * j, reply, res_size);
4902
4903 kfree(req);
4904 kfree(reply);
4905
4906 return ok;
4907}
4908
4909static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
4910{
4911 struct virtnet_info *vi = netdev_priv(dev);
4912 unsigned int i;
4913 u8 *p = data;
4914
4915 switch (stringset) {
4916 case ETH_SS_STATS:
4917 /* Generate the total field names. */
4918 virtnet_get_stats_string(vi, VIRTNET_Q_TYPE_RX, -1, &p);
4919 virtnet_get_stats_string(vi, VIRTNET_Q_TYPE_TX, -1, &p);
4920
4921 virtnet_get_stats_string(vi, VIRTNET_Q_TYPE_CQ, 0, &p);
4922
4923 for (i = 0; i < vi->curr_queue_pairs; ++i)
4924 virtnet_get_stats_string(vi, VIRTNET_Q_TYPE_RX, i, &p);
4925
4926 for (i = 0; i < vi->curr_queue_pairs; ++i)
4927 virtnet_get_stats_string(vi, VIRTNET_Q_TYPE_TX, i, &p);
4928 break;
4929 }
4930}
4931
4932static int virtnet_get_sset_count(struct net_device *dev, int sset)
4933{
4934 struct virtnet_info *vi = netdev_priv(dev);
4935 struct virtnet_stats_ctx ctx = {0};
4936 u32 pair_count;
4937
4938 switch (sset) {
4939 case ETH_SS_STATS:
4940 virtnet_stats_ctx_init(vi, &ctx, NULL, false);
4941
4942 pair_count = ctx.desc_num[VIRTNET_Q_TYPE_RX] + ctx.desc_num[VIRTNET_Q_TYPE_TX];
4943
4944 return pair_count + ctx.desc_num[VIRTNET_Q_TYPE_CQ] +
4945 vi->curr_queue_pairs * pair_count;
4946 default:
4947 return -EOPNOTSUPP;
4948 }
4949}
4950
4951static void virtnet_get_ethtool_stats(struct net_device *dev,
4952 struct ethtool_stats *stats, u64 *data)
4953{
4954 struct virtnet_info *vi = netdev_priv(dev);
4955 struct virtnet_stats_ctx ctx = {0};
4956 unsigned int start, i;
4957 const u8 *stats_base;
4958
4959 virtnet_stats_ctx_init(vi, &ctx, data, false);
4960 if (virtnet_get_hw_stats(vi, &ctx, -1))
4961 dev_warn(&vi->dev->dev, "Failed to get hw stats.\n");
4962
4963 for (i = 0; i < vi->curr_queue_pairs; i++) {
4964 struct receive_queue *rq = &vi->rq[i];
4965 struct send_queue *sq = &vi->sq[i];
4966
4967 stats_base = (const u8 *)&rq->stats;
4968 do {
4969 start = u64_stats_fetch_begin(&rq->stats.syncp);
4970 virtnet_fill_stats(vi, i * 2, &ctx, stats_base, true, 0);
4971 } while (u64_stats_fetch_retry(&rq->stats.syncp, start));
4972
4973 stats_base = (const u8 *)&sq->stats;
4974 do {
4975 start = u64_stats_fetch_begin(&sq->stats.syncp);
4976 virtnet_fill_stats(vi, i * 2 + 1, &ctx, stats_base, true, 0);
4977 } while (u64_stats_fetch_retry(&sq->stats.syncp, start));
4978 }
4979
4980 virtnet_fill_total_fields(vi, &ctx);
4981}
4982
4983static void virtnet_get_channels(struct net_device *dev,
4984 struct ethtool_channels *channels)
4985{
4986 struct virtnet_info *vi = netdev_priv(dev);
4987
4988 channels->combined_count = vi->curr_queue_pairs;
4989 channels->max_combined = vi->max_queue_pairs;
4990 channels->max_other = 0;
4991 channels->rx_count = 0;
4992 channels->tx_count = 0;
4993 channels->other_count = 0;
4994}
4995
4996static int virtnet_set_link_ksettings(struct net_device *dev,
4997 const struct ethtool_link_ksettings *cmd)
4998{
4999 struct virtnet_info *vi = netdev_priv(dev);
5000
5001 return ethtool_virtdev_set_link_ksettings(dev, cmd,
5002 &vi->speed, &vi->duplex);
5003}
5004
5005static int virtnet_get_link_ksettings(struct net_device *dev,
5006 struct ethtool_link_ksettings *cmd)
5007{
5008 struct virtnet_info *vi = netdev_priv(dev);
5009
5010 cmd->base.speed = vi->speed;
5011 cmd->base.duplex = vi->duplex;
5012 cmd->base.port = PORT_OTHER;
5013
5014 return 0;
5015}
5016
5017static int virtnet_send_tx_notf_coal_cmds(struct virtnet_info *vi,
5018 struct ethtool_coalesce *ec)
5019{
5020 struct virtio_net_ctrl_coal_tx *coal_tx __free(kfree) = NULL;
5021 struct scatterlist sgs_tx;
5022 int i;
5023
5024 coal_tx = kzalloc(sizeof(*coal_tx), GFP_KERNEL);
5025 if (!coal_tx)
5026 return -ENOMEM;
5027
5028 coal_tx->tx_usecs = cpu_to_le32(ec->tx_coalesce_usecs);
5029 coal_tx->tx_max_packets = cpu_to_le32(ec->tx_max_coalesced_frames);
5030 sg_init_one(&sgs_tx, coal_tx, sizeof(*coal_tx));
5031
5032 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
5033 VIRTIO_NET_CTRL_NOTF_COAL_TX_SET,
5034 &sgs_tx))
5035 return -EINVAL;
5036
5037 vi->intr_coal_tx.max_usecs = ec->tx_coalesce_usecs;
5038 vi->intr_coal_tx.max_packets = ec->tx_max_coalesced_frames;
5039 for (i = 0; i < vi->max_queue_pairs; i++) {
5040 vi->sq[i].intr_coal.max_usecs = ec->tx_coalesce_usecs;
5041 vi->sq[i].intr_coal.max_packets = ec->tx_max_coalesced_frames;
5042 }
5043
5044 return 0;
5045}
5046
5047static int virtnet_send_rx_notf_coal_cmds(struct virtnet_info *vi,
5048 struct ethtool_coalesce *ec)
5049{
5050 struct virtio_net_ctrl_coal_rx *coal_rx __free(kfree) = NULL;
5051 bool rx_ctrl_dim_on = !!ec->use_adaptive_rx_coalesce;
5052 struct scatterlist sgs_rx;
5053 int i;
5054
5055 if (rx_ctrl_dim_on && !virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL))
5056 return -EOPNOTSUPP;
5057
5058 if (rx_ctrl_dim_on && (ec->rx_coalesce_usecs != vi->intr_coal_rx.max_usecs ||
5059 ec->rx_max_coalesced_frames != vi->intr_coal_rx.max_packets))
5060 return -EINVAL;
5061
5062 if (rx_ctrl_dim_on && !vi->rx_dim_enabled) {
5063 vi->rx_dim_enabled = true;
5064 for (i = 0; i < vi->max_queue_pairs; i++) {
5065 mutex_lock(&vi->rq[i].dim_lock);
5066 vi->rq[i].dim_enabled = true;
5067 mutex_unlock(&vi->rq[i].dim_lock);
5068 }
5069 return 0;
5070 }
5071
5072 coal_rx = kzalloc(sizeof(*coal_rx), GFP_KERNEL);
5073 if (!coal_rx)
5074 return -ENOMEM;
5075
5076 if (!rx_ctrl_dim_on && vi->rx_dim_enabled) {
5077 vi->rx_dim_enabled = false;
5078 for (i = 0; i < vi->max_queue_pairs; i++) {
5079 mutex_lock(&vi->rq[i].dim_lock);
5080 vi->rq[i].dim_enabled = false;
5081 mutex_unlock(&vi->rq[i].dim_lock);
5082 }
5083 }
5084
5085 /* Since the per-queue coalescing params can be set,
5086 * we need apply the global new params even if they
5087 * are not updated.
5088 */
5089 coal_rx->rx_usecs = cpu_to_le32(ec->rx_coalesce_usecs);
5090 coal_rx->rx_max_packets = cpu_to_le32(ec->rx_max_coalesced_frames);
5091 sg_init_one(&sgs_rx, coal_rx, sizeof(*coal_rx));
5092
5093 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
5094 VIRTIO_NET_CTRL_NOTF_COAL_RX_SET,
5095 &sgs_rx))
5096 return -EINVAL;
5097
5098 vi->intr_coal_rx.max_usecs = ec->rx_coalesce_usecs;
5099 vi->intr_coal_rx.max_packets = ec->rx_max_coalesced_frames;
5100 for (i = 0; i < vi->max_queue_pairs; i++) {
5101 mutex_lock(&vi->rq[i].dim_lock);
5102 vi->rq[i].intr_coal.max_usecs = ec->rx_coalesce_usecs;
5103 vi->rq[i].intr_coal.max_packets = ec->rx_max_coalesced_frames;
5104 mutex_unlock(&vi->rq[i].dim_lock);
5105 }
5106
5107 return 0;
5108}
5109
5110static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
5111 struct ethtool_coalesce *ec)
5112{
5113 int err;
5114
5115 err = virtnet_send_tx_notf_coal_cmds(vi, ec);
5116 if (err)
5117 return err;
5118
5119 err = virtnet_send_rx_notf_coal_cmds(vi, ec);
5120 if (err)
5121 return err;
5122
5123 return 0;
5124}
5125
5126static int virtnet_send_rx_notf_coal_vq_cmds(struct virtnet_info *vi,
5127 struct ethtool_coalesce *ec,
5128 u16 queue)
5129{
5130 bool rx_ctrl_dim_on = !!ec->use_adaptive_rx_coalesce;
5131 u32 max_usecs, max_packets;
5132 bool cur_rx_dim;
5133 int err;
5134
5135 mutex_lock(&vi->rq[queue].dim_lock);
5136 cur_rx_dim = vi->rq[queue].dim_enabled;
5137 max_usecs = vi->rq[queue].intr_coal.max_usecs;
5138 max_packets = vi->rq[queue].intr_coal.max_packets;
5139
5140 if (rx_ctrl_dim_on && (ec->rx_coalesce_usecs != max_usecs ||
5141 ec->rx_max_coalesced_frames != max_packets)) {
5142 mutex_unlock(&vi->rq[queue].dim_lock);
5143 return -EINVAL;
5144 }
5145
5146 if (rx_ctrl_dim_on && !cur_rx_dim) {
5147 vi->rq[queue].dim_enabled = true;
5148 mutex_unlock(&vi->rq[queue].dim_lock);
5149 return 0;
5150 }
5151
5152 if (!rx_ctrl_dim_on && cur_rx_dim)
5153 vi->rq[queue].dim_enabled = false;
5154
5155 /* If no params are updated, userspace ethtool will
5156 * reject the modification.
5157 */
5158 err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, queue,
5159 ec->rx_coalesce_usecs,
5160 ec->rx_max_coalesced_frames);
5161 mutex_unlock(&vi->rq[queue].dim_lock);
5162 return err;
5163}
5164
5165static int virtnet_send_notf_coal_vq_cmds(struct virtnet_info *vi,
5166 struct ethtool_coalesce *ec,
5167 u16 queue)
5168{
5169 int err;
5170
5171 err = virtnet_send_rx_notf_coal_vq_cmds(vi, ec, queue);
5172 if (err)
5173 return err;
5174
5175 err = virtnet_send_tx_ctrl_coal_vq_cmd(vi, queue,
5176 ec->tx_coalesce_usecs,
5177 ec->tx_max_coalesced_frames);
5178 if (err)
5179 return err;
5180
5181 return 0;
5182}
5183
5184static void virtnet_rx_dim_work(struct work_struct *work)
5185{
5186 struct dim *dim = container_of(work, struct dim, work);
5187 struct receive_queue *rq = container_of(dim,
5188 struct receive_queue, dim);
5189 struct virtnet_info *vi = rq->vq->vdev->priv;
5190 struct net_device *dev = vi->dev;
5191 struct dim_cq_moder update_moder;
5192 int qnum, err;
5193
5194 qnum = rq - vi->rq;
5195
5196 mutex_lock(&rq->dim_lock);
5197 if (!rq->dim_enabled)
5198 goto out;
5199
5200 update_moder = net_dim_get_rx_irq_moder(dev, dim);
5201 if (update_moder.usec != rq->intr_coal.max_usecs ||
5202 update_moder.pkts != rq->intr_coal.max_packets) {
5203 err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, qnum,
5204 update_moder.usec,
5205 update_moder.pkts);
5206 if (err)
5207 pr_debug("%s: Failed to send dim parameters on rxq%d\n",
5208 dev->name, qnum);
5209 }
5210out:
5211 dim->state = DIM_START_MEASURE;
5212 mutex_unlock(&rq->dim_lock);
5213}
5214
5215static int virtnet_coal_params_supported(struct ethtool_coalesce *ec)
5216{
5217 /* usecs coalescing is supported only if VIRTIO_NET_F_NOTF_COAL
5218 * or VIRTIO_NET_F_VQ_NOTF_COAL feature is negotiated.
5219 */
5220 if (ec->rx_coalesce_usecs || ec->tx_coalesce_usecs)
5221 return -EOPNOTSUPP;
5222
5223 if (ec->tx_max_coalesced_frames > 1 ||
5224 ec->rx_max_coalesced_frames != 1)
5225 return -EINVAL;
5226
5227 return 0;
5228}
5229
5230static int virtnet_should_update_vq_weight(int dev_flags, int weight,
5231 int vq_weight, bool *should_update)
5232{
5233 if (weight ^ vq_weight) {
5234 if (dev_flags & IFF_UP)
5235 return -EBUSY;
5236 *should_update = true;
5237 }
5238
5239 return 0;
5240}
5241
5242static int virtnet_set_coalesce(struct net_device *dev,
5243 struct ethtool_coalesce *ec,
5244 struct kernel_ethtool_coalesce *kernel_coal,
5245 struct netlink_ext_ack *extack)
5246{
5247 struct virtnet_info *vi = netdev_priv(dev);
5248 int ret, queue_number, napi_weight, i;
5249 bool update_napi = false;
5250
5251 /* Can't change NAPI weight if the link is up */
5252 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
5253 for (queue_number = 0; queue_number < vi->max_queue_pairs; queue_number++) {
5254 ret = virtnet_should_update_vq_weight(dev->flags, napi_weight,
5255 vi->sq[queue_number].napi.weight,
5256 &update_napi);
5257 if (ret)
5258 return ret;
5259
5260 if (update_napi) {
5261 /* All queues that belong to [queue_number, vi->max_queue_pairs] will be
5262 * updated for the sake of simplicity, which might not be necessary
5263 */
5264 break;
5265 }
5266 }
5267
5268 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL))
5269 ret = virtnet_send_notf_coal_cmds(vi, ec);
5270 else
5271 ret = virtnet_coal_params_supported(ec);
5272
5273 if (ret)
5274 return ret;
5275
5276 if (update_napi) {
5277 /* xsk xmit depends on the tx napi. So if xsk is active,
5278 * prevent modifications to tx napi.
5279 */
5280 for (i = queue_number; i < vi->max_queue_pairs; i++) {
5281 if (vi->sq[i].xsk_pool)
5282 return -EBUSY;
5283 }
5284
5285 for (; queue_number < vi->max_queue_pairs; queue_number++)
5286 vi->sq[queue_number].napi.weight = napi_weight;
5287 }
5288
5289 return ret;
5290}
5291
5292static int virtnet_get_coalesce(struct net_device *dev,
5293 struct ethtool_coalesce *ec,
5294 struct kernel_ethtool_coalesce *kernel_coal,
5295 struct netlink_ext_ack *extack)
5296{
5297 struct virtnet_info *vi = netdev_priv(dev);
5298
5299 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
5300 ec->rx_coalesce_usecs = vi->intr_coal_rx.max_usecs;
5301 ec->tx_coalesce_usecs = vi->intr_coal_tx.max_usecs;
5302 ec->tx_max_coalesced_frames = vi->intr_coal_tx.max_packets;
5303 ec->rx_max_coalesced_frames = vi->intr_coal_rx.max_packets;
5304 ec->use_adaptive_rx_coalesce = vi->rx_dim_enabled;
5305 } else {
5306 ec->rx_max_coalesced_frames = 1;
5307
5308 if (vi->sq[0].napi.weight)
5309 ec->tx_max_coalesced_frames = 1;
5310 }
5311
5312 return 0;
5313}
5314
5315static int virtnet_set_per_queue_coalesce(struct net_device *dev,
5316 u32 queue,
5317 struct ethtool_coalesce *ec)
5318{
5319 struct virtnet_info *vi = netdev_priv(dev);
5320 int ret, napi_weight;
5321 bool update_napi = false;
5322
5323 if (queue >= vi->max_queue_pairs)
5324 return -EINVAL;
5325
5326 /* Can't change NAPI weight if the link is up */
5327 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
5328 ret = virtnet_should_update_vq_weight(dev->flags, napi_weight,
5329 vi->sq[queue].napi.weight,
5330 &update_napi);
5331 if (ret)
5332 return ret;
5333
5334 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL))
5335 ret = virtnet_send_notf_coal_vq_cmds(vi, ec, queue);
5336 else
5337 ret = virtnet_coal_params_supported(ec);
5338
5339 if (ret)
5340 return ret;
5341
5342 if (update_napi)
5343 vi->sq[queue].napi.weight = napi_weight;
5344
5345 return 0;
5346}
5347
5348static int virtnet_get_per_queue_coalesce(struct net_device *dev,
5349 u32 queue,
5350 struct ethtool_coalesce *ec)
5351{
5352 struct virtnet_info *vi = netdev_priv(dev);
5353
5354 if (queue >= vi->max_queue_pairs)
5355 return -EINVAL;
5356
5357 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL)) {
5358 mutex_lock(&vi->rq[queue].dim_lock);
5359 ec->rx_coalesce_usecs = vi->rq[queue].intr_coal.max_usecs;
5360 ec->tx_coalesce_usecs = vi->sq[queue].intr_coal.max_usecs;
5361 ec->tx_max_coalesced_frames = vi->sq[queue].intr_coal.max_packets;
5362 ec->rx_max_coalesced_frames = vi->rq[queue].intr_coal.max_packets;
5363 ec->use_adaptive_rx_coalesce = vi->rq[queue].dim_enabled;
5364 mutex_unlock(&vi->rq[queue].dim_lock);
5365 } else {
5366 ec->rx_max_coalesced_frames = 1;
5367
5368 if (vi->sq[queue].napi.weight)
5369 ec->tx_max_coalesced_frames = 1;
5370 }
5371
5372 return 0;
5373}
5374
5375static void virtnet_init_settings(struct net_device *dev)
5376{
5377 struct virtnet_info *vi = netdev_priv(dev);
5378
5379 vi->speed = SPEED_UNKNOWN;
5380 vi->duplex = DUPLEX_UNKNOWN;
5381}
5382
5383static u32 virtnet_get_rxfh_key_size(struct net_device *dev)
5384{
5385 return ((struct virtnet_info *)netdev_priv(dev))->rss_key_size;
5386}
5387
5388static u32 virtnet_get_rxfh_indir_size(struct net_device *dev)
5389{
5390 return ((struct virtnet_info *)netdev_priv(dev))->rss_indir_table_size;
5391}
5392
5393static int virtnet_get_rxfh(struct net_device *dev,
5394 struct ethtool_rxfh_param *rxfh)
5395{
5396 struct virtnet_info *vi = netdev_priv(dev);
5397 int i;
5398
5399 if (rxfh->indir) {
5400 for (i = 0; i < vi->rss_indir_table_size; ++i)
5401 rxfh->indir[i] = vi->rss.indirection_table[i];
5402 }
5403
5404 if (rxfh->key)
5405 memcpy(rxfh->key, vi->rss.key, vi->rss_key_size);
5406
5407 rxfh->hfunc = ETH_RSS_HASH_TOP;
5408
5409 return 0;
5410}
5411
5412static int virtnet_set_rxfh(struct net_device *dev,
5413 struct ethtool_rxfh_param *rxfh,
5414 struct netlink_ext_ack *extack)
5415{
5416 struct virtnet_info *vi = netdev_priv(dev);
5417 bool update = false;
5418 int i;
5419
5420 if (rxfh->hfunc != ETH_RSS_HASH_NO_CHANGE &&
5421 rxfh->hfunc != ETH_RSS_HASH_TOP)
5422 return -EOPNOTSUPP;
5423
5424 if (rxfh->indir) {
5425 if (!vi->has_rss)
5426 return -EOPNOTSUPP;
5427
5428 for (i = 0; i < vi->rss_indir_table_size; ++i)
5429 vi->rss.indirection_table[i] = rxfh->indir[i];
5430 update = true;
5431 }
5432
5433 if (rxfh->key) {
5434 /* If either _F_HASH_REPORT or _F_RSS are negotiated, the
5435 * device provides hash calculation capabilities, that is,
5436 * hash_key is configured.
5437 */
5438 if (!vi->has_rss && !vi->has_rss_hash_report)
5439 return -EOPNOTSUPP;
5440
5441 memcpy(vi->rss.key, rxfh->key, vi->rss_key_size);
5442 update = true;
5443 }
5444
5445 if (update)
5446 virtnet_commit_rss_command(vi);
5447
5448 return 0;
5449}
5450
5451static int virtnet_get_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info, u32 *rule_locs)
5452{
5453 struct virtnet_info *vi = netdev_priv(dev);
5454 int rc = 0;
5455
5456 switch (info->cmd) {
5457 case ETHTOOL_GRXRINGS:
5458 info->data = vi->curr_queue_pairs;
5459 break;
5460 case ETHTOOL_GRXFH:
5461 virtnet_get_hashflow(vi, info);
5462 break;
5463 default:
5464 rc = -EOPNOTSUPP;
5465 }
5466
5467 return rc;
5468}
5469
5470static int virtnet_set_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info)
5471{
5472 struct virtnet_info *vi = netdev_priv(dev);
5473 int rc = 0;
5474
5475 switch (info->cmd) {
5476 case ETHTOOL_SRXFH:
5477 if (!virtnet_set_hashflow(vi, info))
5478 rc = -EINVAL;
5479
5480 break;
5481 default:
5482 rc = -EOPNOTSUPP;
5483 }
5484
5485 return rc;
5486}
5487
5488static const struct ethtool_ops virtnet_ethtool_ops = {
5489 .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES |
5490 ETHTOOL_COALESCE_USECS | ETHTOOL_COALESCE_USE_ADAPTIVE_RX,
5491 .get_drvinfo = virtnet_get_drvinfo,
5492 .get_link = ethtool_op_get_link,
5493 .get_ringparam = virtnet_get_ringparam,
5494 .set_ringparam = virtnet_set_ringparam,
5495 .get_strings = virtnet_get_strings,
5496 .get_sset_count = virtnet_get_sset_count,
5497 .get_ethtool_stats = virtnet_get_ethtool_stats,
5498 .set_channels = virtnet_set_channels,
5499 .get_channels = virtnet_get_channels,
5500 .get_ts_info = ethtool_op_get_ts_info,
5501 .get_link_ksettings = virtnet_get_link_ksettings,
5502 .set_link_ksettings = virtnet_set_link_ksettings,
5503 .set_coalesce = virtnet_set_coalesce,
5504 .get_coalesce = virtnet_get_coalesce,
5505 .set_per_queue_coalesce = virtnet_set_per_queue_coalesce,
5506 .get_per_queue_coalesce = virtnet_get_per_queue_coalesce,
5507 .get_rxfh_key_size = virtnet_get_rxfh_key_size,
5508 .get_rxfh_indir_size = virtnet_get_rxfh_indir_size,
5509 .get_rxfh = virtnet_get_rxfh,
5510 .set_rxfh = virtnet_set_rxfh,
5511 .get_rxnfc = virtnet_get_rxnfc,
5512 .set_rxnfc = virtnet_set_rxnfc,
5513};
5514
5515static void virtnet_get_queue_stats_rx(struct net_device *dev, int i,
5516 struct netdev_queue_stats_rx *stats)
5517{
5518 struct virtnet_info *vi = netdev_priv(dev);
5519 struct receive_queue *rq = &vi->rq[i];
5520 struct virtnet_stats_ctx ctx = {0};
5521
5522 virtnet_stats_ctx_init(vi, &ctx, (void *)stats, true);
5523
5524 virtnet_get_hw_stats(vi, &ctx, i * 2);
5525 virtnet_fill_stats(vi, i * 2, &ctx, (void *)&rq->stats, true, 0);
5526}
5527
5528static void virtnet_get_queue_stats_tx(struct net_device *dev, int i,
5529 struct netdev_queue_stats_tx *stats)
5530{
5531 struct virtnet_info *vi = netdev_priv(dev);
5532 struct send_queue *sq = &vi->sq[i];
5533 struct virtnet_stats_ctx ctx = {0};
5534
5535 virtnet_stats_ctx_init(vi, &ctx, (void *)stats, true);
5536
5537 virtnet_get_hw_stats(vi, &ctx, i * 2 + 1);
5538 virtnet_fill_stats(vi, i * 2 + 1, &ctx, (void *)&sq->stats, true, 0);
5539}
5540
5541static void virtnet_get_base_stats(struct net_device *dev,
5542 struct netdev_queue_stats_rx *rx,
5543 struct netdev_queue_stats_tx *tx)
5544{
5545 struct virtnet_info *vi = netdev_priv(dev);
5546
5547 /* The queue stats of the virtio-net will not be reset. So here we
5548 * return 0.
5549 */
5550 rx->bytes = 0;
5551 rx->packets = 0;
5552
5553 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_RX_BASIC) {
5554 rx->hw_drops = 0;
5555 rx->hw_drop_overruns = 0;
5556 }
5557
5558 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_RX_CSUM) {
5559 rx->csum_unnecessary = 0;
5560 rx->csum_none = 0;
5561 rx->csum_bad = 0;
5562 }
5563
5564 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_RX_GSO) {
5565 rx->hw_gro_packets = 0;
5566 rx->hw_gro_bytes = 0;
5567 rx->hw_gro_wire_packets = 0;
5568 rx->hw_gro_wire_bytes = 0;
5569 }
5570
5571 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_RX_SPEED)
5572 rx->hw_drop_ratelimits = 0;
5573
5574 tx->bytes = 0;
5575 tx->packets = 0;
5576 tx->stop = 0;
5577 tx->wake = 0;
5578
5579 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_TX_BASIC) {
5580 tx->hw_drops = 0;
5581 tx->hw_drop_errors = 0;
5582 }
5583
5584 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_TX_CSUM) {
5585 tx->csum_none = 0;
5586 tx->needs_csum = 0;
5587 }
5588
5589 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_TX_GSO) {
5590 tx->hw_gso_packets = 0;
5591 tx->hw_gso_bytes = 0;
5592 tx->hw_gso_wire_packets = 0;
5593 tx->hw_gso_wire_bytes = 0;
5594 }
5595
5596 if (vi->device_stats_cap & VIRTIO_NET_STATS_TYPE_TX_SPEED)
5597 tx->hw_drop_ratelimits = 0;
5598}
5599
5600static const struct netdev_stat_ops virtnet_stat_ops = {
5601 .get_queue_stats_rx = virtnet_get_queue_stats_rx,
5602 .get_queue_stats_tx = virtnet_get_queue_stats_tx,
5603 .get_base_stats = virtnet_get_base_stats,
5604};
5605
5606static void virtnet_freeze_down(struct virtio_device *vdev)
5607{
5608 struct virtnet_info *vi = vdev->priv;
5609
5610 /* Make sure no work handler is accessing the device */
5611 flush_work(&vi->config_work);
5612 disable_rx_mode_work(vi);
5613 flush_work(&vi->rx_mode_work);
5614
5615 netif_tx_lock_bh(vi->dev);
5616 netif_device_detach(vi->dev);
5617 netif_tx_unlock_bh(vi->dev);
5618 if (netif_running(vi->dev))
5619 virtnet_close(vi->dev);
5620}
5621
5622static int init_vqs(struct virtnet_info *vi);
5623
5624static int virtnet_restore_up(struct virtio_device *vdev)
5625{
5626 struct virtnet_info *vi = vdev->priv;
5627 int err;
5628
5629 err = init_vqs(vi);
5630 if (err)
5631 return err;
5632
5633 virtio_device_ready(vdev);
5634
5635 enable_delayed_refill(vi);
5636 enable_rx_mode_work(vi);
5637
5638 if (netif_running(vi->dev)) {
5639 err = virtnet_open(vi->dev);
5640 if (err)
5641 return err;
5642 }
5643
5644 netif_tx_lock_bh(vi->dev);
5645 netif_device_attach(vi->dev);
5646 netif_tx_unlock_bh(vi->dev);
5647 return err;
5648}
5649
5650static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
5651{
5652 __virtio64 *_offloads __free(kfree) = NULL;
5653 struct scatterlist sg;
5654
5655 _offloads = kzalloc(sizeof(*_offloads), GFP_KERNEL);
5656 if (!_offloads)
5657 return -ENOMEM;
5658
5659 *_offloads = cpu_to_virtio64(vi->vdev, offloads);
5660
5661 sg_init_one(&sg, _offloads, sizeof(*_offloads));
5662
5663 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
5664 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
5665 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
5666 return -EINVAL;
5667 }
5668
5669 return 0;
5670}
5671
5672static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
5673{
5674 u64 offloads = 0;
5675
5676 if (!vi->guest_offloads)
5677 return 0;
5678
5679 return virtnet_set_guest_offloads(vi, offloads);
5680}
5681
5682static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
5683{
5684 u64 offloads = vi->guest_offloads;
5685
5686 if (!vi->guest_offloads)
5687 return 0;
5688
5689 return virtnet_set_guest_offloads(vi, offloads);
5690}
5691
5692static int virtnet_rq_bind_xsk_pool(struct virtnet_info *vi, struct receive_queue *rq,
5693 struct xsk_buff_pool *pool)
5694{
5695 int err, qindex;
5696
5697 qindex = rq - vi->rq;
5698
5699 if (pool) {
5700 err = xdp_rxq_info_reg(&rq->xsk_rxq_info, vi->dev, qindex, rq->napi.napi_id);
5701 if (err < 0)
5702 return err;
5703
5704 err = xdp_rxq_info_reg_mem_model(&rq->xsk_rxq_info,
5705 MEM_TYPE_XSK_BUFF_POOL, NULL);
5706 if (err < 0)
5707 goto unreg;
5708
5709 xsk_pool_set_rxq_info(pool, &rq->xsk_rxq_info);
5710 }
5711
5712 virtnet_rx_pause(vi, rq);
5713
5714 err = virtqueue_reset(rq->vq, virtnet_rq_unmap_free_buf, NULL);
5715 if (err) {
5716 netdev_err(vi->dev, "reset rx fail: rx queue index: %d err: %d\n", qindex, err);
5717
5718 pool = NULL;
5719 }
5720
5721 rq->xsk_pool = pool;
5722
5723 virtnet_rx_resume(vi, rq);
5724
5725 if (pool)
5726 return 0;
5727
5728unreg:
5729 xdp_rxq_info_unreg(&rq->xsk_rxq_info);
5730 return err;
5731}
5732
5733static int virtnet_sq_bind_xsk_pool(struct virtnet_info *vi,
5734 struct send_queue *sq,
5735 struct xsk_buff_pool *pool)
5736{
5737 int err, qindex;
5738
5739 qindex = sq - vi->sq;
5740
5741 virtnet_tx_pause(vi, sq);
5742
5743 err = virtqueue_reset(sq->vq, virtnet_sq_free_unused_buf,
5744 virtnet_sq_free_unused_buf_done);
5745 if (err) {
5746 netdev_err(vi->dev, "reset tx fail: tx queue index: %d err: %d\n", qindex, err);
5747 pool = NULL;
5748 }
5749
5750 sq->xsk_pool = pool;
5751
5752 virtnet_tx_resume(vi, sq);
5753
5754 return err;
5755}
5756
5757static int virtnet_xsk_pool_enable(struct net_device *dev,
5758 struct xsk_buff_pool *pool,
5759 u16 qid)
5760{
5761 struct virtnet_info *vi = netdev_priv(dev);
5762 struct receive_queue *rq;
5763 struct device *dma_dev;
5764 struct send_queue *sq;
5765 dma_addr_t hdr_dma;
5766 int err, size;
5767
5768 if (vi->hdr_len > xsk_pool_get_headroom(pool))
5769 return -EINVAL;
5770
5771 /* In big_packets mode, xdp cannot work, so there is no need to
5772 * initialize xsk of rq.
5773 */
5774 if (vi->big_packets && !vi->mergeable_rx_bufs)
5775 return -ENOENT;
5776
5777 if (qid >= vi->curr_queue_pairs)
5778 return -EINVAL;
5779
5780 sq = &vi->sq[qid];
5781 rq = &vi->rq[qid];
5782
5783 /* xsk assumes that tx and rx must have the same dma device. The af-xdp
5784 * may use one buffer to receive from the rx and reuse this buffer to
5785 * send by the tx. So the dma dev of sq and rq must be the same one.
5786 *
5787 * But vq->dma_dev allows every vq has the respective dma dev. So I
5788 * check the dma dev of vq and sq is the same dev.
5789 */
5790 if (virtqueue_dma_dev(rq->vq) != virtqueue_dma_dev(sq->vq))
5791 return -EINVAL;
5792
5793 dma_dev = virtqueue_dma_dev(rq->vq);
5794 if (!dma_dev)
5795 return -EINVAL;
5796
5797 size = virtqueue_get_vring_size(rq->vq);
5798
5799 rq->xsk_buffs = kvcalloc(size, sizeof(*rq->xsk_buffs), GFP_KERNEL);
5800 if (!rq->xsk_buffs)
5801 return -ENOMEM;
5802
5803 hdr_dma = virtqueue_dma_map_single_attrs(sq->vq, &xsk_hdr, vi->hdr_len,
5804 DMA_TO_DEVICE, 0);
5805 if (virtqueue_dma_mapping_error(sq->vq, hdr_dma))
5806 return -ENOMEM;
5807
5808 err = xsk_pool_dma_map(pool, dma_dev, 0);
5809 if (err)
5810 goto err_xsk_map;
5811
5812 err = virtnet_rq_bind_xsk_pool(vi, rq, pool);
5813 if (err)
5814 goto err_rq;
5815
5816 err = virtnet_sq_bind_xsk_pool(vi, sq, pool);
5817 if (err)
5818 goto err_sq;
5819
5820 /* Now, we do not support tx offload(such as tx csum), so all the tx
5821 * virtnet hdr is zero. So all the tx packets can share a single hdr.
5822 */
5823 sq->xsk_hdr_dma_addr = hdr_dma;
5824
5825 return 0;
5826
5827err_sq:
5828 virtnet_rq_bind_xsk_pool(vi, rq, NULL);
5829err_rq:
5830 xsk_pool_dma_unmap(pool, 0);
5831err_xsk_map:
5832 virtqueue_dma_unmap_single_attrs(rq->vq, hdr_dma, vi->hdr_len,
5833 DMA_TO_DEVICE, 0);
5834 return err;
5835}
5836
5837static int virtnet_xsk_pool_disable(struct net_device *dev, u16 qid)
5838{
5839 struct virtnet_info *vi = netdev_priv(dev);
5840 struct xsk_buff_pool *pool;
5841 struct receive_queue *rq;
5842 struct send_queue *sq;
5843 int err;
5844
5845 if (qid >= vi->curr_queue_pairs)
5846 return -EINVAL;
5847
5848 sq = &vi->sq[qid];
5849 rq = &vi->rq[qid];
5850
5851 pool = rq->xsk_pool;
5852
5853 err = virtnet_rq_bind_xsk_pool(vi, rq, NULL);
5854 err |= virtnet_sq_bind_xsk_pool(vi, sq, NULL);
5855
5856 xsk_pool_dma_unmap(pool, 0);
5857
5858 virtqueue_dma_unmap_single_attrs(sq->vq, sq->xsk_hdr_dma_addr,
5859 vi->hdr_len, DMA_TO_DEVICE, 0);
5860 kvfree(rq->xsk_buffs);
5861
5862 return err;
5863}
5864
5865static int virtnet_xsk_pool_setup(struct net_device *dev, struct netdev_bpf *xdp)
5866{
5867 if (xdp->xsk.pool)
5868 return virtnet_xsk_pool_enable(dev, xdp->xsk.pool,
5869 xdp->xsk.queue_id);
5870 else
5871 return virtnet_xsk_pool_disable(dev, xdp->xsk.queue_id);
5872}
5873
5874static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
5875 struct netlink_ext_ack *extack)
5876{
5877 unsigned int room = SKB_DATA_ALIGN(XDP_PACKET_HEADROOM +
5878 sizeof(struct skb_shared_info));
5879 unsigned int max_sz = PAGE_SIZE - room - ETH_HLEN;
5880 struct virtnet_info *vi = netdev_priv(dev);
5881 struct bpf_prog *old_prog;
5882 u16 xdp_qp = 0, curr_qp;
5883 int i, err;
5884
5885 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
5886 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
5887 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
5888 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
5889 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
5890 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM) ||
5891 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) ||
5892 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6))) {
5893 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
5894 return -EOPNOTSUPP;
5895 }
5896
5897 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
5898 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
5899 return -EINVAL;
5900 }
5901
5902 if (prog && !prog->aux->xdp_has_frags && dev->mtu > max_sz) {
5903 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP without frags");
5904 netdev_warn(dev, "single-buffer XDP requires MTU less than %u\n", max_sz);
5905 return -EINVAL;
5906 }
5907
5908 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
5909 if (prog)
5910 xdp_qp = nr_cpu_ids;
5911
5912 /* XDP requires extra queues for XDP_TX */
5913 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
5914 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",
5915 curr_qp + xdp_qp, vi->max_queue_pairs);
5916 xdp_qp = 0;
5917 }
5918
5919 old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
5920 if (!prog && !old_prog)
5921 return 0;
5922
5923 if (prog)
5924 bpf_prog_add(prog, vi->max_queue_pairs - 1);
5925
5926 /* Make sure NAPI is not using any XDP TX queues for RX. */
5927 if (netif_running(dev)) {
5928 for (i = 0; i < vi->max_queue_pairs; i++) {
5929 napi_disable(&vi->rq[i].napi);
5930 virtnet_napi_tx_disable(&vi->sq[i].napi);
5931 }
5932 }
5933
5934 if (!prog) {
5935 for (i = 0; i < vi->max_queue_pairs; i++) {
5936 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
5937 if (i == 0)
5938 virtnet_restore_guest_offloads(vi);
5939 }
5940 synchronize_net();
5941 }
5942
5943 err = virtnet_set_queues(vi, curr_qp + xdp_qp);
5944 if (err)
5945 goto err;
5946 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
5947 vi->xdp_queue_pairs = xdp_qp;
5948
5949 if (prog) {
5950 vi->xdp_enabled = true;
5951 for (i = 0; i < vi->max_queue_pairs; i++) {
5952 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
5953 if (i == 0 && !old_prog)
5954 virtnet_clear_guest_offloads(vi);
5955 }
5956 if (!old_prog)
5957 xdp_features_set_redirect_target(dev, true);
5958 } else {
5959 xdp_features_clear_redirect_target(dev);
5960 vi->xdp_enabled = false;
5961 }
5962
5963 for (i = 0; i < vi->max_queue_pairs; i++) {
5964 if (old_prog)
5965 bpf_prog_put(old_prog);
5966 if (netif_running(dev)) {
5967 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
5968 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
5969 &vi->sq[i].napi);
5970 }
5971 }
5972
5973 return 0;
5974
5975err:
5976 if (!prog) {
5977 virtnet_clear_guest_offloads(vi);
5978 for (i = 0; i < vi->max_queue_pairs; i++)
5979 rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
5980 }
5981
5982 if (netif_running(dev)) {
5983 for (i = 0; i < vi->max_queue_pairs; i++) {
5984 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
5985 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
5986 &vi->sq[i].napi);
5987 }
5988 }
5989 if (prog)
5990 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
5991 return err;
5992}
5993
5994static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
5995{
5996 switch (xdp->command) {
5997 case XDP_SETUP_PROG:
5998 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
5999 case XDP_SETUP_XSK_POOL:
6000 return virtnet_xsk_pool_setup(dev, xdp);
6001 default:
6002 return -EINVAL;
6003 }
6004}
6005
6006static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
6007 size_t len)
6008{
6009 struct virtnet_info *vi = netdev_priv(dev);
6010 int ret;
6011
6012 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
6013 return -EOPNOTSUPP;
6014
6015 ret = snprintf(buf, len, "sby");
6016 if (ret >= len)
6017 return -EOPNOTSUPP;
6018
6019 return 0;
6020}
6021
6022static int virtnet_set_features(struct net_device *dev,
6023 netdev_features_t features)
6024{
6025 struct virtnet_info *vi = netdev_priv(dev);
6026 u64 offloads;
6027 int err;
6028
6029 if ((dev->features ^ features) & NETIF_F_GRO_HW) {
6030 if (vi->xdp_enabled)
6031 return -EBUSY;
6032
6033 if (features & NETIF_F_GRO_HW)
6034 offloads = vi->guest_offloads_capable;
6035 else
6036 offloads = vi->guest_offloads_capable &
6037 ~GUEST_OFFLOAD_GRO_HW_MASK;
6038
6039 err = virtnet_set_guest_offloads(vi, offloads);
6040 if (err)
6041 return err;
6042 vi->guest_offloads = offloads;
6043 }
6044
6045 if ((dev->features ^ features) & NETIF_F_RXHASH) {
6046 if (features & NETIF_F_RXHASH)
6047 vi->rss.hash_types = vi->rss_hash_types_saved;
6048 else
6049 vi->rss.hash_types = VIRTIO_NET_HASH_REPORT_NONE;
6050
6051 if (!virtnet_commit_rss_command(vi))
6052 return -EINVAL;
6053 }
6054
6055 return 0;
6056}
6057
6058static void virtnet_tx_timeout(struct net_device *dev, unsigned int txqueue)
6059{
6060 struct virtnet_info *priv = netdev_priv(dev);
6061 struct send_queue *sq = &priv->sq[txqueue];
6062 struct netdev_queue *txq = netdev_get_tx_queue(dev, txqueue);
6063
6064 u64_stats_update_begin(&sq->stats.syncp);
6065 u64_stats_inc(&sq->stats.tx_timeouts);
6066 u64_stats_update_end(&sq->stats.syncp);
6067
6068 netdev_err(dev, "TX timeout on queue: %u, sq: %s, vq: 0x%x, name: %s, %u usecs ago\n",
6069 txqueue, sq->name, sq->vq->index, sq->vq->name,
6070 jiffies_to_usecs(jiffies - READ_ONCE(txq->trans_start)));
6071}
6072
6073static int virtnet_init_irq_moder(struct virtnet_info *vi)
6074{
6075 u8 profile_flags = 0, coal_flags = 0;
6076 int ret, i;
6077
6078 profile_flags |= DIM_PROFILE_RX;
6079 coal_flags |= DIM_COALESCE_USEC | DIM_COALESCE_PKTS;
6080 ret = net_dim_init_irq_moder(vi->dev, profile_flags, coal_flags,
6081 DIM_CQ_PERIOD_MODE_START_FROM_EQE,
6082 0, virtnet_rx_dim_work, NULL);
6083
6084 if (ret)
6085 return ret;
6086
6087 for (i = 0; i < vi->max_queue_pairs; i++)
6088 net_dim_setting(vi->dev, &vi->rq[i].dim, false);
6089
6090 return 0;
6091}
6092
6093static void virtnet_free_irq_moder(struct virtnet_info *vi)
6094{
6095 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL))
6096 return;
6097
6098 rtnl_lock();
6099 net_dim_free_irq_moder(vi->dev);
6100 rtnl_unlock();
6101}
6102
6103static const struct net_device_ops virtnet_netdev = {
6104 .ndo_open = virtnet_open,
6105 .ndo_stop = virtnet_close,
6106 .ndo_start_xmit = start_xmit,
6107 .ndo_validate_addr = eth_validate_addr,
6108 .ndo_set_mac_address = virtnet_set_mac_address,
6109 .ndo_set_rx_mode = virtnet_set_rx_mode,
6110 .ndo_get_stats64 = virtnet_stats,
6111 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
6112 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
6113 .ndo_bpf = virtnet_xdp,
6114 .ndo_xdp_xmit = virtnet_xdp_xmit,
6115 .ndo_xsk_wakeup = virtnet_xsk_wakeup,
6116 .ndo_features_check = passthru_features_check,
6117 .ndo_get_phys_port_name = virtnet_get_phys_port_name,
6118 .ndo_set_features = virtnet_set_features,
6119 .ndo_tx_timeout = virtnet_tx_timeout,
6120};
6121
6122static void virtnet_config_changed_work(struct work_struct *work)
6123{
6124 struct virtnet_info *vi =
6125 container_of(work, struct virtnet_info, config_work);
6126 u16 v;
6127
6128 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
6129 struct virtio_net_config, status, &v) < 0)
6130 return;
6131
6132 if (v & VIRTIO_NET_S_ANNOUNCE) {
6133 netdev_notify_peers(vi->dev);
6134 virtnet_ack_link_announce(vi);
6135 }
6136
6137 /* Ignore unknown (future) status bits */
6138 v &= VIRTIO_NET_S_LINK_UP;
6139
6140 if (vi->status == v)
6141 return;
6142
6143 vi->status = v;
6144
6145 if (vi->status & VIRTIO_NET_S_LINK_UP) {
6146 virtnet_update_settings(vi);
6147 netif_carrier_on(vi->dev);
6148 netif_tx_wake_all_queues(vi->dev);
6149 } else {
6150 netif_carrier_off(vi->dev);
6151 netif_tx_stop_all_queues(vi->dev);
6152 }
6153}
6154
6155static void virtnet_config_changed(struct virtio_device *vdev)
6156{
6157 struct virtnet_info *vi = vdev->priv;
6158
6159 schedule_work(&vi->config_work);
6160}
6161
6162static void virtnet_free_queues(struct virtnet_info *vi)
6163{
6164 int i;
6165
6166 for (i = 0; i < vi->max_queue_pairs; i++) {
6167 __netif_napi_del(&vi->rq[i].napi);
6168 __netif_napi_del(&vi->sq[i].napi);
6169 }
6170
6171 /* We called __netif_napi_del(),
6172 * we need to respect an RCU grace period before freeing vi->rq
6173 */
6174 synchronize_net();
6175
6176 kfree(vi->rq);
6177 kfree(vi->sq);
6178 kfree(vi->ctrl);
6179}
6180
6181static void _free_receive_bufs(struct virtnet_info *vi)
6182{
6183 struct bpf_prog *old_prog;
6184 int i;
6185
6186 for (i = 0; i < vi->max_queue_pairs; i++) {
6187 while (vi->rq[i].pages)
6188 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
6189
6190 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
6191 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
6192 if (old_prog)
6193 bpf_prog_put(old_prog);
6194 }
6195}
6196
6197static void free_receive_bufs(struct virtnet_info *vi)
6198{
6199 rtnl_lock();
6200 _free_receive_bufs(vi);
6201 rtnl_unlock();
6202}
6203
6204static void free_receive_page_frags(struct virtnet_info *vi)
6205{
6206 int i;
6207 for (i = 0; i < vi->max_queue_pairs; i++)
6208 if (vi->rq[i].alloc_frag.page) {
6209 if (vi->rq[i].last_dma)
6210 virtnet_rq_unmap(&vi->rq[i], vi->rq[i].last_dma, 0);
6211 put_page(vi->rq[i].alloc_frag.page);
6212 }
6213}
6214
6215static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf)
6216{
6217 struct virtnet_info *vi = vq->vdev->priv;
6218 struct send_queue *sq;
6219 int i = vq2txq(vq);
6220
6221 sq = &vi->sq[i];
6222
6223 switch (virtnet_xmit_ptr_unpack(&buf)) {
6224 case VIRTNET_XMIT_TYPE_SKB:
6225 case VIRTNET_XMIT_TYPE_SKB_ORPHAN:
6226 dev_kfree_skb(buf);
6227 break;
6228
6229 case VIRTNET_XMIT_TYPE_XDP:
6230 xdp_return_frame(buf);
6231 break;
6232
6233 case VIRTNET_XMIT_TYPE_XSK:
6234 xsk_tx_completed(sq->xsk_pool, 1);
6235 break;
6236 }
6237}
6238
6239static void virtnet_sq_free_unused_buf_done(struct virtqueue *vq)
6240{
6241 struct virtnet_info *vi = vq->vdev->priv;
6242 int i = vq2txq(vq);
6243
6244 netdev_tx_reset_queue(netdev_get_tx_queue(vi->dev, i));
6245}
6246
6247static void free_unused_bufs(struct virtnet_info *vi)
6248{
6249 void *buf;
6250 int i;
6251
6252 for (i = 0; i < vi->max_queue_pairs; i++) {
6253 struct virtqueue *vq = vi->sq[i].vq;
6254 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
6255 virtnet_sq_free_unused_buf(vq, buf);
6256 cond_resched();
6257 }
6258
6259 for (i = 0; i < vi->max_queue_pairs; i++) {
6260 struct virtqueue *vq = vi->rq[i].vq;
6261
6262 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
6263 virtnet_rq_unmap_free_buf(vq, buf);
6264 cond_resched();
6265 }
6266}
6267
6268static void virtnet_del_vqs(struct virtnet_info *vi)
6269{
6270 struct virtio_device *vdev = vi->vdev;
6271
6272 virtnet_clean_affinity(vi);
6273
6274 vdev->config->del_vqs(vdev);
6275
6276 virtnet_free_queues(vi);
6277}
6278
6279/* How large should a single buffer be so a queue full of these can fit at
6280 * least one full packet?
6281 * Logic below assumes the mergeable buffer header is used.
6282 */
6283static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
6284{
6285 const unsigned int hdr_len = vi->hdr_len;
6286 unsigned int rq_size = virtqueue_get_vring_size(vq);
6287 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
6288 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
6289 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
6290
6291 return max(max(min_buf_len, hdr_len) - hdr_len,
6292 (unsigned int)GOOD_PACKET_LEN);
6293}
6294
6295static int virtnet_find_vqs(struct virtnet_info *vi)
6296{
6297 struct virtqueue_info *vqs_info;
6298 struct virtqueue **vqs;
6299 int ret = -ENOMEM;
6300 int total_vqs;
6301 bool *ctx;
6302 u16 i;
6303
6304 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
6305 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
6306 * possible control vq.
6307 */
6308 total_vqs = vi->max_queue_pairs * 2 +
6309 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
6310
6311 /* Allocate space for find_vqs parameters */
6312 vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
6313 if (!vqs)
6314 goto err_vq;
6315 vqs_info = kcalloc(total_vqs, sizeof(*vqs_info), GFP_KERNEL);
6316 if (!vqs_info)
6317 goto err_vqs_info;
6318 if (!vi->big_packets || vi->mergeable_rx_bufs) {
6319 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
6320 if (!ctx)
6321 goto err_ctx;
6322 } else {
6323 ctx = NULL;
6324 }
6325
6326 /* Parameters for control virtqueue, if any */
6327 if (vi->has_cvq) {
6328 vqs_info[total_vqs - 1].name = "control";
6329 }
6330
6331 /* Allocate/initialize parameters for send/receive virtqueues */
6332 for (i = 0; i < vi->max_queue_pairs; i++) {
6333 vqs_info[rxq2vq(i)].callback = skb_recv_done;
6334 vqs_info[txq2vq(i)].callback = skb_xmit_done;
6335 sprintf(vi->rq[i].name, "input.%u", i);
6336 sprintf(vi->sq[i].name, "output.%u", i);
6337 vqs_info[rxq2vq(i)].name = vi->rq[i].name;
6338 vqs_info[txq2vq(i)].name = vi->sq[i].name;
6339 if (ctx)
6340 vqs_info[rxq2vq(i)].ctx = true;
6341 }
6342
6343 ret = virtio_find_vqs(vi->vdev, total_vqs, vqs, vqs_info, NULL);
6344 if (ret)
6345 goto err_find;
6346
6347 if (vi->has_cvq) {
6348 vi->cvq = vqs[total_vqs - 1];
6349 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
6350 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
6351 }
6352
6353 for (i = 0; i < vi->max_queue_pairs; i++) {
6354 vi->rq[i].vq = vqs[rxq2vq(i)];
6355 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
6356 vi->sq[i].vq = vqs[txq2vq(i)];
6357 }
6358
6359 /* run here: ret == 0. */
6360
6361
6362err_find:
6363 kfree(ctx);
6364err_ctx:
6365 kfree(vqs_info);
6366err_vqs_info:
6367 kfree(vqs);
6368err_vq:
6369 return ret;
6370}
6371
6372static int virtnet_alloc_queues(struct virtnet_info *vi)
6373{
6374 int i;
6375
6376 if (vi->has_cvq) {
6377 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
6378 if (!vi->ctrl)
6379 goto err_ctrl;
6380 } else {
6381 vi->ctrl = NULL;
6382 }
6383 vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
6384 if (!vi->sq)
6385 goto err_sq;
6386 vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
6387 if (!vi->rq)
6388 goto err_rq;
6389
6390 INIT_DELAYED_WORK(&vi->refill, refill_work);
6391 for (i = 0; i < vi->max_queue_pairs; i++) {
6392 vi->rq[i].pages = NULL;
6393 netif_napi_add_weight(vi->dev, &vi->rq[i].napi, virtnet_poll,
6394 napi_weight);
6395 netif_napi_add_tx_weight(vi->dev, &vi->sq[i].napi,
6396 virtnet_poll_tx,
6397 napi_tx ? napi_weight : 0);
6398
6399 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
6400 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
6401 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
6402
6403 u64_stats_init(&vi->rq[i].stats.syncp);
6404 u64_stats_init(&vi->sq[i].stats.syncp);
6405 mutex_init(&vi->rq[i].dim_lock);
6406 }
6407
6408 return 0;
6409
6410err_rq:
6411 kfree(vi->sq);
6412err_sq:
6413 kfree(vi->ctrl);
6414err_ctrl:
6415 return -ENOMEM;
6416}
6417
6418static int init_vqs(struct virtnet_info *vi)
6419{
6420 int ret;
6421
6422 /* Allocate send & receive queues */
6423 ret = virtnet_alloc_queues(vi);
6424 if (ret)
6425 goto err;
6426
6427 ret = virtnet_find_vqs(vi);
6428 if (ret)
6429 goto err_free;
6430
6431 cpus_read_lock();
6432 virtnet_set_affinity(vi);
6433 cpus_read_unlock();
6434
6435 return 0;
6436
6437err_free:
6438 virtnet_free_queues(vi);
6439err:
6440 return ret;
6441}
6442
6443#ifdef CONFIG_SYSFS
6444static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
6445 char *buf)
6446{
6447 struct virtnet_info *vi = netdev_priv(queue->dev);
6448 unsigned int queue_index = get_netdev_rx_queue_index(queue);
6449 unsigned int headroom = virtnet_get_headroom(vi);
6450 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
6451 struct ewma_pkt_len *avg;
6452
6453 BUG_ON(queue_index >= vi->max_queue_pairs);
6454 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
6455 return sprintf(buf, "%u\n",
6456 get_mergeable_buf_len(&vi->rq[queue_index], avg,
6457 SKB_DATA_ALIGN(headroom + tailroom)));
6458}
6459
6460static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
6461 __ATTR_RO(mergeable_rx_buffer_size);
6462
6463static struct attribute *virtio_net_mrg_rx_attrs[] = {
6464 &mergeable_rx_buffer_size_attribute.attr,
6465 NULL
6466};
6467
6468static const struct attribute_group virtio_net_mrg_rx_group = {
6469 .name = "virtio_net",
6470 .attrs = virtio_net_mrg_rx_attrs
6471};
6472#endif
6473
6474static bool virtnet_fail_on_feature(struct virtio_device *vdev,
6475 unsigned int fbit,
6476 const char *fname, const char *dname)
6477{
6478 if (!virtio_has_feature(vdev, fbit))
6479 return false;
6480
6481 dev_err(&vdev->dev, "device advertises feature %s but not %s",
6482 fname, dname);
6483
6484 return true;
6485}
6486
6487#define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
6488 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
6489
6490static bool virtnet_validate_features(struct virtio_device *vdev)
6491{
6492 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
6493 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
6494 "VIRTIO_NET_F_CTRL_VQ") ||
6495 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
6496 "VIRTIO_NET_F_CTRL_VQ") ||
6497 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
6498 "VIRTIO_NET_F_CTRL_VQ") ||
6499 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
6500 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
6501 "VIRTIO_NET_F_CTRL_VQ") ||
6502 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_RSS,
6503 "VIRTIO_NET_F_CTRL_VQ") ||
6504 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_HASH_REPORT,
6505 "VIRTIO_NET_F_CTRL_VQ") ||
6506 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_NOTF_COAL,
6507 "VIRTIO_NET_F_CTRL_VQ") ||
6508 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_VQ_NOTF_COAL,
6509 "VIRTIO_NET_F_CTRL_VQ"))) {
6510 return false;
6511 }
6512
6513 return true;
6514}
6515
6516#define MIN_MTU ETH_MIN_MTU
6517#define MAX_MTU ETH_MAX_MTU
6518
6519static int virtnet_validate(struct virtio_device *vdev)
6520{
6521 if (!vdev->config->get) {
6522 dev_err(&vdev->dev, "%s failure: config access disabled\n",
6523 __func__);
6524 return -EINVAL;
6525 }
6526
6527 if (!virtnet_validate_features(vdev))
6528 return -EINVAL;
6529
6530 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
6531 int mtu = virtio_cread16(vdev,
6532 offsetof(struct virtio_net_config,
6533 mtu));
6534 if (mtu < MIN_MTU)
6535 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
6536 }
6537
6538 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY) &&
6539 !virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
6540 dev_warn(&vdev->dev, "device advertises feature VIRTIO_NET_F_STANDBY but not VIRTIO_NET_F_MAC, disabling standby");
6541 __virtio_clear_bit(vdev, VIRTIO_NET_F_STANDBY);
6542 }
6543
6544 return 0;
6545}
6546
6547static bool virtnet_check_guest_gso(const struct virtnet_info *vi)
6548{
6549 return virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
6550 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
6551 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
6552 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
6553 (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) &&
6554 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6));
6555}
6556
6557static void virtnet_set_big_packets(struct virtnet_info *vi, const int mtu)
6558{
6559 bool guest_gso = virtnet_check_guest_gso(vi);
6560
6561 /* If device can receive ANY guest GSO packets, regardless of mtu,
6562 * allocate packets of maximum size, otherwise limit it to only
6563 * mtu size worth only.
6564 */
6565 if (mtu > ETH_DATA_LEN || guest_gso) {
6566 vi->big_packets = true;
6567 vi->big_packets_num_skbfrags = guest_gso ? MAX_SKB_FRAGS : DIV_ROUND_UP(mtu, PAGE_SIZE);
6568 }
6569}
6570
6571#define VIRTIO_NET_HASH_REPORT_MAX_TABLE 10
6572static enum xdp_rss_hash_type
6573virtnet_xdp_rss_type[VIRTIO_NET_HASH_REPORT_MAX_TABLE] = {
6574 [VIRTIO_NET_HASH_REPORT_NONE] = XDP_RSS_TYPE_NONE,
6575 [VIRTIO_NET_HASH_REPORT_IPv4] = XDP_RSS_TYPE_L3_IPV4,
6576 [VIRTIO_NET_HASH_REPORT_TCPv4] = XDP_RSS_TYPE_L4_IPV4_TCP,
6577 [VIRTIO_NET_HASH_REPORT_UDPv4] = XDP_RSS_TYPE_L4_IPV4_UDP,
6578 [VIRTIO_NET_HASH_REPORT_IPv6] = XDP_RSS_TYPE_L3_IPV6,
6579 [VIRTIO_NET_HASH_REPORT_TCPv6] = XDP_RSS_TYPE_L4_IPV6_TCP,
6580 [VIRTIO_NET_HASH_REPORT_UDPv6] = XDP_RSS_TYPE_L4_IPV6_UDP,
6581 [VIRTIO_NET_HASH_REPORT_IPv6_EX] = XDP_RSS_TYPE_L3_IPV6_EX,
6582 [VIRTIO_NET_HASH_REPORT_TCPv6_EX] = XDP_RSS_TYPE_L4_IPV6_TCP_EX,
6583 [VIRTIO_NET_HASH_REPORT_UDPv6_EX] = XDP_RSS_TYPE_L4_IPV6_UDP_EX
6584};
6585
6586static int virtnet_xdp_rx_hash(const struct xdp_md *_ctx, u32 *hash,
6587 enum xdp_rss_hash_type *rss_type)
6588{
6589 const struct xdp_buff *xdp = (void *)_ctx;
6590 struct virtio_net_hdr_v1_hash *hdr_hash;
6591 struct virtnet_info *vi;
6592 u16 hash_report;
6593
6594 if (!(xdp->rxq->dev->features & NETIF_F_RXHASH))
6595 return -ENODATA;
6596
6597 vi = netdev_priv(xdp->rxq->dev);
6598 hdr_hash = (struct virtio_net_hdr_v1_hash *)(xdp->data - vi->hdr_len);
6599 hash_report = __le16_to_cpu(hdr_hash->hash_report);
6600
6601 if (hash_report >= VIRTIO_NET_HASH_REPORT_MAX_TABLE)
6602 hash_report = VIRTIO_NET_HASH_REPORT_NONE;
6603
6604 *rss_type = virtnet_xdp_rss_type[hash_report];
6605 *hash = __le32_to_cpu(hdr_hash->hash_value);
6606 return 0;
6607}
6608
6609static const struct xdp_metadata_ops virtnet_xdp_metadata_ops = {
6610 .xmo_rx_hash = virtnet_xdp_rx_hash,
6611};
6612
6613static int virtnet_probe(struct virtio_device *vdev)
6614{
6615 int i, err = -ENOMEM;
6616 struct net_device *dev;
6617 struct virtnet_info *vi;
6618 u16 max_queue_pairs;
6619 int mtu = 0;
6620
6621 /* Find if host supports multiqueue/rss virtio_net device */
6622 max_queue_pairs = 1;
6623 if (virtio_has_feature(vdev, VIRTIO_NET_F_MQ) || virtio_has_feature(vdev, VIRTIO_NET_F_RSS))
6624 max_queue_pairs =
6625 virtio_cread16(vdev, offsetof(struct virtio_net_config, max_virtqueue_pairs));
6626
6627 /* We need at least 2 queue's */
6628 if (max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
6629 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
6630 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
6631 max_queue_pairs = 1;
6632
6633 /* Allocate ourselves a network device with room for our info */
6634 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
6635 if (!dev)
6636 return -ENOMEM;
6637
6638 /* Set up network device as normal. */
6639 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
6640 IFF_TX_SKB_NO_LINEAR;
6641 dev->netdev_ops = &virtnet_netdev;
6642 dev->stat_ops = &virtnet_stat_ops;
6643 dev->features = NETIF_F_HIGHDMA;
6644
6645 dev->ethtool_ops = &virtnet_ethtool_ops;
6646 SET_NETDEV_DEV(dev, &vdev->dev);
6647
6648 /* Do we support "hardware" checksums? */
6649 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
6650 /* This opens up the world of extra features. */
6651 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
6652 if (csum)
6653 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
6654
6655 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
6656 dev->hw_features |= NETIF_F_TSO
6657 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
6658 }
6659 /* Individual feature bits: what can host handle? */
6660 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
6661 dev->hw_features |= NETIF_F_TSO;
6662 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
6663 dev->hw_features |= NETIF_F_TSO6;
6664 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
6665 dev->hw_features |= NETIF_F_TSO_ECN;
6666 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_USO))
6667 dev->hw_features |= NETIF_F_GSO_UDP_L4;
6668
6669 dev->features |= NETIF_F_GSO_ROBUST;
6670
6671 if (gso)
6672 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
6673 /* (!csum && gso) case will be fixed by register_netdev() */
6674 }
6675
6676 /* 1. With VIRTIO_NET_F_GUEST_CSUM negotiation, the driver doesn't
6677 * need to calculate checksums for partially checksummed packets,
6678 * as they're considered valid by the upper layer.
6679 * 2. Without VIRTIO_NET_F_GUEST_CSUM negotiation, the driver only
6680 * receives fully checksummed packets. The device may assist in
6681 * validating these packets' checksums, so the driver won't have to.
6682 */
6683 dev->features |= NETIF_F_RXCSUM;
6684
6685 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
6686 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
6687 dev->features |= NETIF_F_GRO_HW;
6688 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
6689 dev->hw_features |= NETIF_F_GRO_HW;
6690
6691 dev->vlan_features = dev->features;
6692 dev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT |
6693 NETDEV_XDP_ACT_XSK_ZEROCOPY;
6694
6695 /* MTU range: 68 - 65535 */
6696 dev->min_mtu = MIN_MTU;
6697 dev->max_mtu = MAX_MTU;
6698
6699 /* Configuration may specify what MAC to use. Otherwise random. */
6700 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
6701 u8 addr[ETH_ALEN];
6702
6703 virtio_cread_bytes(vdev,
6704 offsetof(struct virtio_net_config, mac),
6705 addr, ETH_ALEN);
6706 eth_hw_addr_set(dev, addr);
6707 } else {
6708 eth_hw_addr_random(dev);
6709 dev_info(&vdev->dev, "Assigned random MAC address %pM\n",
6710 dev->dev_addr);
6711 }
6712
6713 /* Set up our device-specific information */
6714 vi = netdev_priv(dev);
6715 vi->dev = dev;
6716 vi->vdev = vdev;
6717 vdev->priv = vi;
6718
6719 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
6720 INIT_WORK(&vi->rx_mode_work, virtnet_rx_mode_work);
6721 spin_lock_init(&vi->refill_lock);
6722
6723 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) {
6724 vi->mergeable_rx_bufs = true;
6725 dev->xdp_features |= NETDEV_XDP_ACT_RX_SG;
6726 }
6727
6728 if (virtio_has_feature(vdev, VIRTIO_NET_F_HASH_REPORT))
6729 vi->has_rss_hash_report = true;
6730
6731 if (virtio_has_feature(vdev, VIRTIO_NET_F_RSS)) {
6732 vi->has_rss = true;
6733
6734 vi->rss_indir_table_size =
6735 virtio_cread16(vdev, offsetof(struct virtio_net_config,
6736 rss_max_indirection_table_length));
6737 }
6738 err = rss_indirection_table_alloc(&vi->rss, vi->rss_indir_table_size);
6739 if (err)
6740 goto free;
6741
6742 if (vi->has_rss || vi->has_rss_hash_report) {
6743 vi->rss_key_size =
6744 virtio_cread8(vdev, offsetof(struct virtio_net_config, rss_max_key_size));
6745 if (vi->rss_key_size > VIRTIO_NET_RSS_MAX_KEY_SIZE) {
6746 dev_err(&vdev->dev, "rss_max_key_size=%u exceeds the limit %u.\n",
6747 vi->rss_key_size, VIRTIO_NET_RSS_MAX_KEY_SIZE);
6748 err = -EINVAL;
6749 goto free;
6750 }
6751
6752 vi->rss_hash_types_supported =
6753 virtio_cread32(vdev, offsetof(struct virtio_net_config, supported_hash_types));
6754 vi->rss_hash_types_supported &=
6755 ~(VIRTIO_NET_RSS_HASH_TYPE_IP_EX |
6756 VIRTIO_NET_RSS_HASH_TYPE_TCP_EX |
6757 VIRTIO_NET_RSS_HASH_TYPE_UDP_EX);
6758
6759 dev->hw_features |= NETIF_F_RXHASH;
6760 dev->xdp_metadata_ops = &virtnet_xdp_metadata_ops;
6761 }
6762
6763 if (vi->has_rss_hash_report)
6764 vi->hdr_len = sizeof(struct virtio_net_hdr_v1_hash);
6765 else if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
6766 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
6767 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
6768 else
6769 vi->hdr_len = sizeof(struct virtio_net_hdr);
6770
6771 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
6772 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
6773 vi->any_header_sg = true;
6774
6775 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
6776 vi->has_cvq = true;
6777
6778 mutex_init(&vi->cvq_lock);
6779
6780 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
6781 mtu = virtio_cread16(vdev,
6782 offsetof(struct virtio_net_config,
6783 mtu));
6784 if (mtu < dev->min_mtu) {
6785 /* Should never trigger: MTU was previously validated
6786 * in virtnet_validate.
6787 */
6788 dev_err(&vdev->dev,
6789 "device MTU appears to have changed it is now %d < %d",
6790 mtu, dev->min_mtu);
6791 err = -EINVAL;
6792 goto free;
6793 }
6794
6795 dev->mtu = mtu;
6796 dev->max_mtu = mtu;
6797 }
6798
6799 virtnet_set_big_packets(vi, mtu);
6800
6801 if (vi->any_header_sg)
6802 dev->needed_headroom = vi->hdr_len;
6803
6804 /* Enable multiqueue by default */
6805 if (num_online_cpus() >= max_queue_pairs)
6806 vi->curr_queue_pairs = max_queue_pairs;
6807 else
6808 vi->curr_queue_pairs = num_online_cpus();
6809 vi->max_queue_pairs = max_queue_pairs;
6810
6811 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
6812 err = init_vqs(vi);
6813 if (err)
6814 goto free;
6815
6816 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
6817 vi->intr_coal_rx.max_usecs = 0;
6818 vi->intr_coal_tx.max_usecs = 0;
6819 vi->intr_coal_rx.max_packets = 0;
6820
6821 /* Keep the default values of the coalescing parameters
6822 * aligned with the default napi_tx state.
6823 */
6824 if (vi->sq[0].napi.weight)
6825 vi->intr_coal_tx.max_packets = 1;
6826 else
6827 vi->intr_coal_tx.max_packets = 0;
6828 }
6829
6830 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL)) {
6831 /* The reason is the same as VIRTIO_NET_F_NOTF_COAL. */
6832 for (i = 0; i < vi->max_queue_pairs; i++)
6833 if (vi->sq[i].napi.weight)
6834 vi->sq[i].intr_coal.max_packets = 1;
6835
6836 err = virtnet_init_irq_moder(vi);
6837 if (err)
6838 goto free;
6839 }
6840
6841#ifdef CONFIG_SYSFS
6842 if (vi->mergeable_rx_bufs)
6843 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
6844#endif
6845 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
6846 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
6847
6848 virtnet_init_settings(dev);
6849
6850 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
6851 vi->failover = net_failover_create(vi->dev);
6852 if (IS_ERR(vi->failover)) {
6853 err = PTR_ERR(vi->failover);
6854 goto free_vqs;
6855 }
6856 }
6857
6858 if (vi->has_rss || vi->has_rss_hash_report)
6859 virtnet_init_default_rss(vi);
6860
6861 enable_rx_mode_work(vi);
6862
6863 /* serialize netdev register + virtio_device_ready() with ndo_open() */
6864 rtnl_lock();
6865
6866 err = register_netdevice(dev);
6867 if (err) {
6868 pr_debug("virtio_net: registering device failed\n");
6869 rtnl_unlock();
6870 goto free_failover;
6871 }
6872
6873 /* Disable config change notification until ndo_open. */
6874 virtio_config_driver_disable(vi->vdev);
6875
6876 virtio_device_ready(vdev);
6877
6878 if (vi->has_rss || vi->has_rss_hash_report) {
6879 if (!virtnet_commit_rss_command(vi)) {
6880 dev_warn(&vdev->dev, "RSS disabled because committing failed.\n");
6881 dev->hw_features &= ~NETIF_F_RXHASH;
6882 vi->has_rss_hash_report = false;
6883 vi->has_rss = false;
6884 }
6885 }
6886
6887 virtnet_set_queues(vi, vi->curr_queue_pairs);
6888
6889 /* a random MAC address has been assigned, notify the device.
6890 * We don't fail probe if VIRTIO_NET_F_CTRL_MAC_ADDR is not there
6891 * because many devices work fine without getting MAC explicitly
6892 */
6893 if (!virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
6894 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
6895 struct scatterlist sg;
6896
6897 sg_init_one(&sg, dev->dev_addr, dev->addr_len);
6898 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
6899 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
6900 pr_debug("virtio_net: setting MAC address failed\n");
6901 rtnl_unlock();
6902 err = -EINVAL;
6903 goto free_unregister_netdev;
6904 }
6905 }
6906
6907 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_DEVICE_STATS)) {
6908 struct virtio_net_stats_capabilities *stats_cap __free(kfree) = NULL;
6909 struct scatterlist sg;
6910 __le64 v;
6911
6912 stats_cap = kzalloc(sizeof(*stats_cap), GFP_KERNEL);
6913 if (!stats_cap) {
6914 rtnl_unlock();
6915 err = -ENOMEM;
6916 goto free_unregister_netdev;
6917 }
6918
6919 sg_init_one(&sg, stats_cap, sizeof(*stats_cap));
6920
6921 if (!virtnet_send_command_reply(vi, VIRTIO_NET_CTRL_STATS,
6922 VIRTIO_NET_CTRL_STATS_QUERY,
6923 NULL, &sg)) {
6924 pr_debug("virtio_net: fail to get stats capability\n");
6925 rtnl_unlock();
6926 err = -EINVAL;
6927 goto free_unregister_netdev;
6928 }
6929
6930 v = stats_cap->supported_stats_types[0];
6931 vi->device_stats_cap = le64_to_cpu(v);
6932 }
6933
6934 /* Assume link up if device can't report link status,
6935 otherwise get link status from config. */
6936 netif_carrier_off(dev);
6937 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
6938 virtnet_config_changed_work(&vi->config_work);
6939 } else {
6940 vi->status = VIRTIO_NET_S_LINK_UP;
6941 virtnet_update_settings(vi);
6942 netif_carrier_on(dev);
6943 }
6944
6945 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
6946 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
6947 set_bit(guest_offloads[i], &vi->guest_offloads);
6948 vi->guest_offloads_capable = vi->guest_offloads;
6949
6950 rtnl_unlock();
6951
6952 err = virtnet_cpu_notif_add(vi);
6953 if (err) {
6954 pr_debug("virtio_net: registering cpu notifier failed\n");
6955 goto free_unregister_netdev;
6956 }
6957
6958 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
6959 dev->name, max_queue_pairs);
6960
6961 return 0;
6962
6963free_unregister_netdev:
6964 unregister_netdev(dev);
6965free_failover:
6966 net_failover_destroy(vi->failover);
6967free_vqs:
6968 virtio_reset_device(vdev);
6969 cancel_delayed_work_sync(&vi->refill);
6970 free_receive_page_frags(vi);
6971 virtnet_del_vqs(vi);
6972free:
6973 free_netdev(dev);
6974 return err;
6975}
6976
6977static void remove_vq_common(struct virtnet_info *vi)
6978{
6979 int i;
6980
6981 virtio_reset_device(vi->vdev);
6982
6983 /* Free unused buffers in both send and recv, if any. */
6984 free_unused_bufs(vi);
6985
6986 /*
6987 * Rule of thumb is netdev_tx_reset_queue() should follow any
6988 * skb freeing not followed by netdev_tx_completed_queue()
6989 */
6990 for (i = 0; i < vi->max_queue_pairs; i++)
6991 netdev_tx_reset_queue(netdev_get_tx_queue(vi->dev, i));
6992
6993 free_receive_bufs(vi);
6994
6995 free_receive_page_frags(vi);
6996
6997 virtnet_del_vqs(vi);
6998}
6999
7000static void virtnet_remove(struct virtio_device *vdev)
7001{
7002 struct virtnet_info *vi = vdev->priv;
7003
7004 virtnet_cpu_notif_remove(vi);
7005
7006 /* Make sure no work handler is accessing the device. */
7007 flush_work(&vi->config_work);
7008 disable_rx_mode_work(vi);
7009 flush_work(&vi->rx_mode_work);
7010
7011 virtnet_free_irq_moder(vi);
7012
7013 unregister_netdev(vi->dev);
7014
7015 net_failover_destroy(vi->failover);
7016
7017 remove_vq_common(vi);
7018
7019 rss_indirection_table_free(&vi->rss);
7020
7021 free_netdev(vi->dev);
7022}
7023
7024static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
7025{
7026 struct virtnet_info *vi = vdev->priv;
7027
7028 virtnet_cpu_notif_remove(vi);
7029 virtnet_freeze_down(vdev);
7030 remove_vq_common(vi);
7031
7032 return 0;
7033}
7034
7035static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
7036{
7037 struct virtnet_info *vi = vdev->priv;
7038 int err;
7039
7040 err = virtnet_restore_up(vdev);
7041 if (err)
7042 return err;
7043 virtnet_set_queues(vi, vi->curr_queue_pairs);
7044
7045 err = virtnet_cpu_notif_add(vi);
7046 if (err) {
7047 virtnet_freeze_down(vdev);
7048 remove_vq_common(vi);
7049 return err;
7050 }
7051
7052 return 0;
7053}
7054
7055static struct virtio_device_id id_table[] = {
7056 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
7057 { 0 },
7058};
7059
7060#define VIRTNET_FEATURES \
7061 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
7062 VIRTIO_NET_F_MAC, \
7063 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
7064 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
7065 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
7066 VIRTIO_NET_F_HOST_USO, VIRTIO_NET_F_GUEST_USO4, VIRTIO_NET_F_GUEST_USO6, \
7067 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
7068 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
7069 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
7070 VIRTIO_NET_F_CTRL_MAC_ADDR, \
7071 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
7072 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY, \
7073 VIRTIO_NET_F_RSS, VIRTIO_NET_F_HASH_REPORT, VIRTIO_NET_F_NOTF_COAL, \
7074 VIRTIO_NET_F_VQ_NOTF_COAL, \
7075 VIRTIO_NET_F_GUEST_HDRLEN, VIRTIO_NET_F_DEVICE_STATS
7076
7077static unsigned int features[] = {
7078 VIRTNET_FEATURES,
7079};
7080
7081static unsigned int features_legacy[] = {
7082 VIRTNET_FEATURES,
7083 VIRTIO_NET_F_GSO,
7084 VIRTIO_F_ANY_LAYOUT,
7085};
7086
7087static struct virtio_driver virtio_net_driver = {
7088 .feature_table = features,
7089 .feature_table_size = ARRAY_SIZE(features),
7090 .feature_table_legacy = features_legacy,
7091 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
7092 .driver.name = KBUILD_MODNAME,
7093 .id_table = id_table,
7094 .validate = virtnet_validate,
7095 .probe = virtnet_probe,
7096 .remove = virtnet_remove,
7097 .config_changed = virtnet_config_changed,
7098#ifdef CONFIG_PM_SLEEP
7099 .freeze = virtnet_freeze,
7100 .restore = virtnet_restore,
7101#endif
7102};
7103
7104static __init int virtio_net_driver_init(void)
7105{
7106 int ret;
7107
7108 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
7109 virtnet_cpu_online,
7110 virtnet_cpu_down_prep);
7111 if (ret < 0)
7112 goto out;
7113 virtionet_online = ret;
7114 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
7115 NULL, virtnet_cpu_dead);
7116 if (ret)
7117 goto err_dead;
7118 ret = register_virtio_driver(&virtio_net_driver);
7119 if (ret)
7120 goto err_virtio;
7121 return 0;
7122err_virtio:
7123 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
7124err_dead:
7125 cpuhp_remove_multi_state(virtionet_online);
7126out:
7127 return ret;
7128}
7129module_init(virtio_net_driver_init);
7130
7131static __exit void virtio_net_driver_exit(void)
7132{
7133 unregister_virtio_driver(&virtio_net_driver);
7134 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
7135 cpuhp_remove_multi_state(virtionet_online);
7136}
7137module_exit(virtio_net_driver_exit);
7138
7139MODULE_DEVICE_TABLE(virtio, id_table);
7140MODULE_DESCRIPTION("Virtio network driver");
7141MODULE_LICENSE("GPL");