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 <net/route.h>
23#include <net/xdp.h>
24#include <net/net_failover.h>
25
26static int napi_weight = NAPI_POLL_WEIGHT;
27module_param(napi_weight, int, 0444);
28
29static bool csum = true, gso = true, napi_tx = true;
30module_param(csum, bool, 0444);
31module_param(gso, bool, 0444);
32module_param(napi_tx, bool, 0644);
33
34/* FIXME: MTU in config. */
35#define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
36#define GOOD_COPY_LEN 128
37
38#define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
39
40/* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
41#define VIRTIO_XDP_HEADROOM 256
42
43/* Separating two types of XDP xmit */
44#define VIRTIO_XDP_TX BIT(0)
45#define VIRTIO_XDP_REDIR BIT(1)
46
47#define VIRTIO_XDP_FLAG BIT(0)
48
49/* RX packet size EWMA. The average packet size is used to determine the packet
50 * buffer size when refilling RX rings. As the entire RX ring may be refilled
51 * at once, the weight is chosen so that the EWMA will be insensitive to short-
52 * term, transient changes in packet size.
53 */
54DECLARE_EWMA(pkt_len, 0, 64)
55
56#define VIRTNET_DRIVER_VERSION "1.0.0"
57
58static const unsigned long guest_offloads[] = {
59 VIRTIO_NET_F_GUEST_TSO4,
60 VIRTIO_NET_F_GUEST_TSO6,
61 VIRTIO_NET_F_GUEST_ECN,
62 VIRTIO_NET_F_GUEST_UFO,
63 VIRTIO_NET_F_GUEST_CSUM
64};
65
66#define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
67 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
68 (1ULL << VIRTIO_NET_F_GUEST_ECN) | \
69 (1ULL << VIRTIO_NET_F_GUEST_UFO))
70
71struct virtnet_stat_desc {
72 char desc[ETH_GSTRING_LEN];
73 size_t offset;
74};
75
76struct virtnet_sq_stats {
77 struct u64_stats_sync syncp;
78 u64 packets;
79 u64 bytes;
80 u64 xdp_tx;
81 u64 xdp_tx_drops;
82 u64 kicks;
83};
84
85struct virtnet_rq_stats {
86 struct u64_stats_sync syncp;
87 u64 packets;
88 u64 bytes;
89 u64 drops;
90 u64 xdp_packets;
91 u64 xdp_tx;
92 u64 xdp_redirects;
93 u64 xdp_drops;
94 u64 kicks;
95};
96
97#define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m)
98#define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m)
99
100static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
101 { "packets", VIRTNET_SQ_STAT(packets) },
102 { "bytes", VIRTNET_SQ_STAT(bytes) },
103 { "xdp_tx", VIRTNET_SQ_STAT(xdp_tx) },
104 { "xdp_tx_drops", VIRTNET_SQ_STAT(xdp_tx_drops) },
105 { "kicks", VIRTNET_SQ_STAT(kicks) },
106};
107
108static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
109 { "packets", VIRTNET_RQ_STAT(packets) },
110 { "bytes", VIRTNET_RQ_STAT(bytes) },
111 { "drops", VIRTNET_RQ_STAT(drops) },
112 { "xdp_packets", VIRTNET_RQ_STAT(xdp_packets) },
113 { "xdp_tx", VIRTNET_RQ_STAT(xdp_tx) },
114 { "xdp_redirects", VIRTNET_RQ_STAT(xdp_redirects) },
115 { "xdp_drops", VIRTNET_RQ_STAT(xdp_drops) },
116 { "kicks", VIRTNET_RQ_STAT(kicks) },
117};
118
119#define VIRTNET_SQ_STATS_LEN ARRAY_SIZE(virtnet_sq_stats_desc)
120#define VIRTNET_RQ_STATS_LEN ARRAY_SIZE(virtnet_rq_stats_desc)
121
122/* Internal representation of a send virtqueue */
123struct send_queue {
124 /* Virtqueue associated with this send _queue */
125 struct virtqueue *vq;
126
127 /* TX: fragments + linear part + virtio header */
128 struct scatterlist sg[MAX_SKB_FRAGS + 2];
129
130 /* Name of the send queue: output.$index */
131 char name[40];
132
133 struct virtnet_sq_stats stats;
134
135 struct napi_struct napi;
136};
137
138/* Internal representation of a receive virtqueue */
139struct receive_queue {
140 /* Virtqueue associated with this receive_queue */
141 struct virtqueue *vq;
142
143 struct napi_struct napi;
144
145 struct bpf_prog __rcu *xdp_prog;
146
147 struct virtnet_rq_stats stats;
148
149 /* Chain pages by the private ptr. */
150 struct page *pages;
151
152 /* Average packet length for mergeable receive buffers. */
153 struct ewma_pkt_len mrg_avg_pkt_len;
154
155 /* Page frag for packet buffer allocation. */
156 struct page_frag alloc_frag;
157
158 /* RX: fragments + linear part + virtio header */
159 struct scatterlist sg[MAX_SKB_FRAGS + 2];
160
161 /* Min single buffer size for mergeable buffers case. */
162 unsigned int min_buf_len;
163
164 /* Name of this receive queue: input.$index */
165 char name[40];
166
167 struct xdp_rxq_info xdp_rxq;
168};
169
170/* Control VQ buffers: protected by the rtnl lock */
171struct control_buf {
172 struct virtio_net_ctrl_hdr hdr;
173 virtio_net_ctrl_ack status;
174 struct virtio_net_ctrl_mq mq;
175 u8 promisc;
176 u8 allmulti;
177 __virtio16 vid;
178 __virtio64 offloads;
179};
180
181struct virtnet_info {
182 struct virtio_device *vdev;
183 struct virtqueue *cvq;
184 struct net_device *dev;
185 struct send_queue *sq;
186 struct receive_queue *rq;
187 unsigned int status;
188
189 /* Max # of queue pairs supported by the device */
190 u16 max_queue_pairs;
191
192 /* # of queue pairs currently used by the driver */
193 u16 curr_queue_pairs;
194
195 /* # of XDP queue pairs currently used by the driver */
196 u16 xdp_queue_pairs;
197
198 /* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
199 bool xdp_enabled;
200
201 /* I like... big packets and I cannot lie! */
202 bool big_packets;
203
204 /* Host will merge rx buffers for big packets (shake it! shake it!) */
205 bool mergeable_rx_bufs;
206
207 /* Has control virtqueue */
208 bool has_cvq;
209
210 /* Host can handle any s/g split between our header and packet data */
211 bool any_header_sg;
212
213 /* Packet virtio header size */
214 u8 hdr_len;
215
216 /* Work struct for refilling if we run low on memory. */
217 struct delayed_work refill;
218
219 /* Work struct for config space updates */
220 struct work_struct config_work;
221
222 /* Does the affinity hint is set for virtqueues? */
223 bool affinity_hint_set;
224
225 /* CPU hotplug instances for online & dead */
226 struct hlist_node node;
227 struct hlist_node node_dead;
228
229 struct control_buf *ctrl;
230
231 /* Ethtool settings */
232 u8 duplex;
233 u32 speed;
234
235 unsigned long guest_offloads;
236 unsigned long guest_offloads_capable;
237
238 /* failover when STANDBY feature enabled */
239 struct failover *failover;
240};
241
242struct padded_vnet_hdr {
243 struct virtio_net_hdr_mrg_rxbuf hdr;
244 /*
245 * hdr is in a separate sg buffer, and data sg buffer shares same page
246 * with this header sg. This padding makes next sg 16 byte aligned
247 * after the header.
248 */
249 char padding[4];
250};
251
252static bool is_xdp_frame(void *ptr)
253{
254 return (unsigned long)ptr & VIRTIO_XDP_FLAG;
255}
256
257static void *xdp_to_ptr(struct xdp_frame *ptr)
258{
259 return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
260}
261
262static struct xdp_frame *ptr_to_xdp(void *ptr)
263{
264 return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
265}
266
267/* Converting between virtqueue no. and kernel tx/rx queue no.
268 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
269 */
270static int vq2txq(struct virtqueue *vq)
271{
272 return (vq->index - 1) / 2;
273}
274
275static int txq2vq(int txq)
276{
277 return txq * 2 + 1;
278}
279
280static int vq2rxq(struct virtqueue *vq)
281{
282 return vq->index / 2;
283}
284
285static int rxq2vq(int rxq)
286{
287 return rxq * 2;
288}
289
290static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
291{
292 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
293}
294
295/*
296 * private is used to chain pages for big packets, put the whole
297 * most recent used list in the beginning for reuse
298 */
299static void give_pages(struct receive_queue *rq, struct page *page)
300{
301 struct page *end;
302
303 /* Find end of list, sew whole thing into vi->rq.pages. */
304 for (end = page; end->private; end = (struct page *)end->private);
305 end->private = (unsigned long)rq->pages;
306 rq->pages = page;
307}
308
309static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
310{
311 struct page *p = rq->pages;
312
313 if (p) {
314 rq->pages = (struct page *)p->private;
315 /* clear private here, it is used to chain pages */
316 p->private = 0;
317 } else
318 p = alloc_page(gfp_mask);
319 return p;
320}
321
322static void virtqueue_napi_schedule(struct napi_struct *napi,
323 struct virtqueue *vq)
324{
325 if (napi_schedule_prep(napi)) {
326 virtqueue_disable_cb(vq);
327 __napi_schedule(napi);
328 }
329}
330
331static void virtqueue_napi_complete(struct napi_struct *napi,
332 struct virtqueue *vq, int processed)
333{
334 int opaque;
335
336 opaque = virtqueue_enable_cb_prepare(vq);
337 if (napi_complete_done(napi, processed)) {
338 if (unlikely(virtqueue_poll(vq, opaque)))
339 virtqueue_napi_schedule(napi, vq);
340 } else {
341 virtqueue_disable_cb(vq);
342 }
343}
344
345static void skb_xmit_done(struct virtqueue *vq)
346{
347 struct virtnet_info *vi = vq->vdev->priv;
348 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
349
350 /* Suppress further interrupts. */
351 virtqueue_disable_cb(vq);
352
353 if (napi->weight)
354 virtqueue_napi_schedule(napi, vq);
355 else
356 /* We were probably waiting for more output buffers. */
357 netif_wake_subqueue(vi->dev, vq2txq(vq));
358}
359
360#define MRG_CTX_HEADER_SHIFT 22
361static void *mergeable_len_to_ctx(unsigned int truesize,
362 unsigned int headroom)
363{
364 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
365}
366
367static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
368{
369 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
370}
371
372static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
373{
374 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
375}
376
377/* Called from bottom half context */
378static struct sk_buff *page_to_skb(struct virtnet_info *vi,
379 struct receive_queue *rq,
380 struct page *page, unsigned int offset,
381 unsigned int len, unsigned int truesize,
382 bool hdr_valid, unsigned int metasize,
383 bool whole_page)
384{
385 struct sk_buff *skb;
386 struct virtio_net_hdr_mrg_rxbuf *hdr;
387 unsigned int copy, hdr_len, hdr_padded_len;
388 struct page *page_to_free = NULL;
389 int tailroom, shinfo_size;
390 char *p, *hdr_p, *buf;
391
392 p = page_address(page) + offset;
393 hdr_p = p;
394
395 hdr_len = vi->hdr_len;
396 if (vi->mergeable_rx_bufs)
397 hdr_padded_len = sizeof(*hdr);
398 else
399 hdr_padded_len = sizeof(struct padded_vnet_hdr);
400
401 /* If whole_page, there is an offset between the beginning of the
402 * data and the allocated space, otherwise the data and the allocated
403 * space are aligned.
404 *
405 * Buffers with headroom use PAGE_SIZE as alloc size, see
406 * add_recvbuf_mergeable() + get_mergeable_buf_len()
407 */
408 if (whole_page) {
409 /* Buffers with whole_page use PAGE_SIZE as alloc size,
410 * see add_recvbuf_mergeable() + get_mergeable_buf_len()
411 */
412 truesize = PAGE_SIZE;
413
414 /* page maybe head page, so we should get the buf by p, not the
415 * page
416 */
417 tailroom = truesize - len - offset_in_page(p);
418 buf = (char *)((unsigned long)p & PAGE_MASK);
419 } else {
420 tailroom = truesize - len;
421 buf = p;
422 }
423
424 len -= hdr_len;
425 offset += hdr_padded_len;
426 p += hdr_padded_len;
427
428 shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
429
430 /* copy small packet so we can reuse these pages */
431 if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
432 skb = build_skb(buf, truesize);
433 if (unlikely(!skb))
434 return NULL;
435
436 skb_reserve(skb, p - buf);
437 skb_put(skb, len);
438
439 page = (struct page *)page->private;
440 if (page)
441 give_pages(rq, page);
442 goto ok;
443 }
444
445 /* copy small packet so we can reuse these pages for small data */
446 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
447 if (unlikely(!skb))
448 return NULL;
449
450 /* Copy all frame if it fits skb->head, otherwise
451 * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
452 */
453 if (len <= skb_tailroom(skb))
454 copy = len;
455 else
456 copy = ETH_HLEN + metasize;
457 skb_put_data(skb, p, copy);
458
459 len -= copy;
460 offset += copy;
461
462 if (vi->mergeable_rx_bufs) {
463 if (len)
464 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
465 else
466 page_to_free = page;
467 goto ok;
468 }
469
470 /*
471 * Verify that we can indeed put this data into a skb.
472 * This is here to handle cases when the device erroneously
473 * tries to receive more than is possible. This is usually
474 * the case of a broken device.
475 */
476 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
477 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
478 dev_kfree_skb(skb);
479 return NULL;
480 }
481 BUG_ON(offset >= PAGE_SIZE);
482 while (len) {
483 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
484 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
485 frag_size, truesize);
486 len -= frag_size;
487 page = (struct page *)page->private;
488 offset = 0;
489 }
490
491 if (page)
492 give_pages(rq, page);
493
494ok:
495 /* hdr_valid means no XDP, so we can copy the vnet header */
496 if (hdr_valid) {
497 hdr = skb_vnet_hdr(skb);
498 memcpy(hdr, hdr_p, hdr_len);
499 }
500 if (page_to_free)
501 put_page(page_to_free);
502
503 if (metasize) {
504 __skb_pull(skb, metasize);
505 skb_metadata_set(skb, metasize);
506 }
507
508 return skb;
509}
510
511static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
512 struct send_queue *sq,
513 struct xdp_frame *xdpf)
514{
515 struct virtio_net_hdr_mrg_rxbuf *hdr;
516 int err;
517
518 if (unlikely(xdpf->headroom < vi->hdr_len))
519 return -EOVERFLOW;
520
521 /* Make room for virtqueue hdr (also change xdpf->headroom?) */
522 xdpf->data -= vi->hdr_len;
523 /* Zero header and leave csum up to XDP layers */
524 hdr = xdpf->data;
525 memset(hdr, 0, vi->hdr_len);
526 xdpf->len += vi->hdr_len;
527
528 sg_init_one(sq->sg, xdpf->data, xdpf->len);
529
530 err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
531 GFP_ATOMIC);
532 if (unlikely(err))
533 return -ENOSPC; /* Caller handle free/refcnt */
534
535 return 0;
536}
537
538/* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
539 * the current cpu, so it does not need to be locked.
540 *
541 * Here we use marco instead of inline functions because we have to deal with
542 * three issues at the same time: 1. the choice of sq. 2. judge and execute the
543 * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
544 * functions to perfectly solve these three problems at the same time.
545 */
546#define virtnet_xdp_get_sq(vi) ({ \
547 struct netdev_queue *txq; \
548 typeof(vi) v = (vi); \
549 unsigned int qp; \
550 \
551 if (v->curr_queue_pairs > nr_cpu_ids) { \
552 qp = v->curr_queue_pairs - v->xdp_queue_pairs; \
553 qp += smp_processor_id(); \
554 txq = netdev_get_tx_queue(v->dev, qp); \
555 __netif_tx_acquire(txq); \
556 } else { \
557 qp = smp_processor_id() % v->curr_queue_pairs; \
558 txq = netdev_get_tx_queue(v->dev, qp); \
559 __netif_tx_lock(txq, raw_smp_processor_id()); \
560 } \
561 v->sq + qp; \
562})
563
564#define virtnet_xdp_put_sq(vi, q) { \
565 struct netdev_queue *txq; \
566 typeof(vi) v = (vi); \
567 \
568 txq = netdev_get_tx_queue(v->dev, (q) - v->sq); \
569 if (v->curr_queue_pairs > nr_cpu_ids) \
570 __netif_tx_release(txq); \
571 else \
572 __netif_tx_unlock(txq); \
573}
574
575static int virtnet_xdp_xmit(struct net_device *dev,
576 int n, struct xdp_frame **frames, u32 flags)
577{
578 struct virtnet_info *vi = netdev_priv(dev);
579 struct receive_queue *rq = vi->rq;
580 struct bpf_prog *xdp_prog;
581 struct send_queue *sq;
582 unsigned int len;
583 int packets = 0;
584 int bytes = 0;
585 int nxmit = 0;
586 int kicks = 0;
587 void *ptr;
588 int ret;
589 int i;
590
591 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
592 * indicate XDP resources have been successfully allocated.
593 */
594 xdp_prog = rcu_access_pointer(rq->xdp_prog);
595 if (!xdp_prog)
596 return -ENXIO;
597
598 sq = virtnet_xdp_get_sq(vi);
599
600 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
601 ret = -EINVAL;
602 goto out;
603 }
604
605 /* Free up any pending old buffers before queueing new ones. */
606 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
607 if (likely(is_xdp_frame(ptr))) {
608 struct xdp_frame *frame = ptr_to_xdp(ptr);
609
610 bytes += frame->len;
611 xdp_return_frame(frame);
612 } else {
613 struct sk_buff *skb = ptr;
614
615 bytes += skb->len;
616 napi_consume_skb(skb, false);
617 }
618 packets++;
619 }
620
621 for (i = 0; i < n; i++) {
622 struct xdp_frame *xdpf = frames[i];
623
624 if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
625 break;
626 nxmit++;
627 }
628 ret = nxmit;
629
630 if (flags & XDP_XMIT_FLUSH) {
631 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
632 kicks = 1;
633 }
634out:
635 u64_stats_update_begin(&sq->stats.syncp);
636 sq->stats.bytes += bytes;
637 sq->stats.packets += packets;
638 sq->stats.xdp_tx += n;
639 sq->stats.xdp_tx_drops += n - nxmit;
640 sq->stats.kicks += kicks;
641 u64_stats_update_end(&sq->stats.syncp);
642
643 virtnet_xdp_put_sq(vi, sq);
644 return ret;
645}
646
647static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
648{
649 return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
650}
651
652/* We copy the packet for XDP in the following cases:
653 *
654 * 1) Packet is scattered across multiple rx buffers.
655 * 2) Headroom space is insufficient.
656 *
657 * This is inefficient but it's a temporary condition that
658 * we hit right after XDP is enabled and until queue is refilled
659 * with large buffers with sufficient headroom - so it should affect
660 * at most queue size packets.
661 * Afterwards, the conditions to enable
662 * XDP should preclude the underlying device from sending packets
663 * across multiple buffers (num_buf > 1), and we make sure buffers
664 * have enough headroom.
665 */
666static struct page *xdp_linearize_page(struct receive_queue *rq,
667 u16 *num_buf,
668 struct page *p,
669 int offset,
670 int page_off,
671 unsigned int *len)
672{
673 struct page *page = alloc_page(GFP_ATOMIC);
674
675 if (!page)
676 return NULL;
677
678 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
679 page_off += *len;
680
681 while (--*num_buf) {
682 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
683 unsigned int buflen;
684 void *buf;
685 int off;
686
687 buf = virtqueue_get_buf(rq->vq, &buflen);
688 if (unlikely(!buf))
689 goto err_buf;
690
691 p = virt_to_head_page(buf);
692 off = buf - page_address(p);
693
694 /* guard against a misconfigured or uncooperative backend that
695 * is sending packet larger than the MTU.
696 */
697 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
698 put_page(p);
699 goto err_buf;
700 }
701
702 memcpy(page_address(page) + page_off,
703 page_address(p) + off, buflen);
704 page_off += buflen;
705 put_page(p);
706 }
707
708 /* Headroom does not contribute to packet length */
709 *len = page_off - VIRTIO_XDP_HEADROOM;
710 return page;
711err_buf:
712 __free_pages(page, 0);
713 return NULL;
714}
715
716static struct sk_buff *receive_small(struct net_device *dev,
717 struct virtnet_info *vi,
718 struct receive_queue *rq,
719 void *buf, void *ctx,
720 unsigned int len,
721 unsigned int *xdp_xmit,
722 struct virtnet_rq_stats *stats)
723{
724 struct sk_buff *skb;
725 struct bpf_prog *xdp_prog;
726 unsigned int xdp_headroom = (unsigned long)ctx;
727 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
728 unsigned int headroom = vi->hdr_len + header_offset;
729 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
730 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
731 struct page *page = virt_to_head_page(buf);
732 unsigned int delta = 0;
733 struct page *xdp_page;
734 int err;
735 unsigned int metasize = 0;
736
737 len -= vi->hdr_len;
738 stats->bytes += len;
739
740 if (unlikely(len > GOOD_PACKET_LEN)) {
741 pr_debug("%s: rx error: len %u exceeds max size %d\n",
742 dev->name, len, GOOD_PACKET_LEN);
743 dev->stats.rx_length_errors++;
744 goto err_len;
745 }
746 rcu_read_lock();
747 xdp_prog = rcu_dereference(rq->xdp_prog);
748 if (xdp_prog) {
749 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
750 struct xdp_frame *xdpf;
751 struct xdp_buff xdp;
752 void *orig_data;
753 u32 act;
754
755 if (unlikely(hdr->hdr.gso_type))
756 goto err_xdp;
757
758 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
759 int offset = buf - page_address(page) + header_offset;
760 unsigned int tlen = len + vi->hdr_len;
761 u16 num_buf = 1;
762
763 xdp_headroom = virtnet_get_headroom(vi);
764 header_offset = VIRTNET_RX_PAD + xdp_headroom;
765 headroom = vi->hdr_len + header_offset;
766 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
767 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
768 xdp_page = xdp_linearize_page(rq, &num_buf, page,
769 offset, header_offset,
770 &tlen);
771 if (!xdp_page)
772 goto err_xdp;
773
774 buf = page_address(xdp_page);
775 put_page(page);
776 page = xdp_page;
777 }
778
779 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
780 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
781 xdp_headroom, len, true);
782 orig_data = xdp.data;
783 act = bpf_prog_run_xdp(xdp_prog, &xdp);
784 stats->xdp_packets++;
785
786 switch (act) {
787 case XDP_PASS:
788 /* Recalculate length in case bpf program changed it */
789 delta = orig_data - xdp.data;
790 len = xdp.data_end - xdp.data;
791 metasize = xdp.data - xdp.data_meta;
792 break;
793 case XDP_TX:
794 stats->xdp_tx++;
795 xdpf = xdp_convert_buff_to_frame(&xdp);
796 if (unlikely(!xdpf))
797 goto err_xdp;
798 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
799 if (unlikely(!err)) {
800 xdp_return_frame_rx_napi(xdpf);
801 } else if (unlikely(err < 0)) {
802 trace_xdp_exception(vi->dev, xdp_prog, act);
803 goto err_xdp;
804 }
805 *xdp_xmit |= VIRTIO_XDP_TX;
806 rcu_read_unlock();
807 goto xdp_xmit;
808 case XDP_REDIRECT:
809 stats->xdp_redirects++;
810 err = xdp_do_redirect(dev, &xdp, xdp_prog);
811 if (err)
812 goto err_xdp;
813 *xdp_xmit |= VIRTIO_XDP_REDIR;
814 rcu_read_unlock();
815 goto xdp_xmit;
816 default:
817 bpf_warn_invalid_xdp_action(act);
818 fallthrough;
819 case XDP_ABORTED:
820 trace_xdp_exception(vi->dev, xdp_prog, act);
821 goto err_xdp;
822 case XDP_DROP:
823 goto err_xdp;
824 }
825 }
826 rcu_read_unlock();
827
828 skb = build_skb(buf, buflen);
829 if (!skb) {
830 put_page(page);
831 goto err;
832 }
833 skb_reserve(skb, headroom - delta);
834 skb_put(skb, len);
835 if (!xdp_prog) {
836 buf += header_offset;
837 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
838 } /* keep zeroed vnet hdr since XDP is loaded */
839
840 if (metasize)
841 skb_metadata_set(skb, metasize);
842
843err:
844 return skb;
845
846err_xdp:
847 rcu_read_unlock();
848 stats->xdp_drops++;
849err_len:
850 stats->drops++;
851 put_page(page);
852xdp_xmit:
853 return NULL;
854}
855
856static struct sk_buff *receive_big(struct net_device *dev,
857 struct virtnet_info *vi,
858 struct receive_queue *rq,
859 void *buf,
860 unsigned int len,
861 struct virtnet_rq_stats *stats)
862{
863 struct page *page = buf;
864 struct sk_buff *skb =
865 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0, 0);
866
867 stats->bytes += len - vi->hdr_len;
868 if (unlikely(!skb))
869 goto err;
870
871 return skb;
872
873err:
874 stats->drops++;
875 give_pages(rq, page);
876 return NULL;
877}
878
879static struct sk_buff *receive_mergeable(struct net_device *dev,
880 struct virtnet_info *vi,
881 struct receive_queue *rq,
882 void *buf,
883 void *ctx,
884 unsigned int len,
885 unsigned int *xdp_xmit,
886 struct virtnet_rq_stats *stats)
887{
888 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
889 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
890 struct page *page = virt_to_head_page(buf);
891 int offset = buf - page_address(page);
892 struct sk_buff *head_skb, *curr_skb;
893 struct bpf_prog *xdp_prog;
894 unsigned int truesize = mergeable_ctx_to_truesize(ctx);
895 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
896 unsigned int metasize = 0;
897 unsigned int frame_sz;
898 int err;
899
900 head_skb = NULL;
901 stats->bytes += len - vi->hdr_len;
902
903 if (unlikely(len > truesize)) {
904 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
905 dev->name, len, (unsigned long)ctx);
906 dev->stats.rx_length_errors++;
907 goto err_skb;
908 }
909 rcu_read_lock();
910 xdp_prog = rcu_dereference(rq->xdp_prog);
911 if (xdp_prog) {
912 struct xdp_frame *xdpf;
913 struct page *xdp_page;
914 struct xdp_buff xdp;
915 void *data;
916 u32 act;
917
918 /* Transient failure which in theory could occur if
919 * in-flight packets from before XDP was enabled reach
920 * the receive path after XDP is loaded.
921 */
922 if (unlikely(hdr->hdr.gso_type))
923 goto err_xdp;
924
925 /* Buffers with headroom use PAGE_SIZE as alloc size,
926 * see add_recvbuf_mergeable() + get_mergeable_buf_len()
927 */
928 frame_sz = headroom ? PAGE_SIZE : truesize;
929
930 /* This happens when rx buffer size is underestimated
931 * or headroom is not enough because of the buffer
932 * was refilled before XDP is set. This should only
933 * happen for the first several packets, so we don't
934 * care much about its performance.
935 */
936 if (unlikely(num_buf > 1 ||
937 headroom < virtnet_get_headroom(vi))) {
938 /* linearize data for XDP */
939 xdp_page = xdp_linearize_page(rq, &num_buf,
940 page, offset,
941 VIRTIO_XDP_HEADROOM,
942 &len);
943 frame_sz = PAGE_SIZE;
944
945 if (!xdp_page)
946 goto err_xdp;
947 offset = VIRTIO_XDP_HEADROOM;
948 } else {
949 xdp_page = page;
950 }
951
952 /* Allow consuming headroom but reserve enough space to push
953 * the descriptor on if we get an XDP_TX return code.
954 */
955 data = page_address(xdp_page) + offset;
956 xdp_init_buff(&xdp, frame_sz - vi->hdr_len, &rq->xdp_rxq);
957 xdp_prepare_buff(&xdp, data - VIRTIO_XDP_HEADROOM + vi->hdr_len,
958 VIRTIO_XDP_HEADROOM, len - vi->hdr_len, true);
959
960 act = bpf_prog_run_xdp(xdp_prog, &xdp);
961 stats->xdp_packets++;
962
963 switch (act) {
964 case XDP_PASS:
965 metasize = xdp.data - xdp.data_meta;
966
967 /* recalculate offset to account for any header
968 * adjustments and minus the metasize to copy the
969 * metadata in page_to_skb(). Note other cases do not
970 * build an skb and avoid using offset
971 */
972 offset = xdp.data - page_address(xdp_page) -
973 vi->hdr_len - metasize;
974
975 /* recalculate len if xdp.data, xdp.data_end or
976 * xdp.data_meta were adjusted
977 */
978 len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
979 /* We can only create skb based on xdp_page. */
980 if (unlikely(xdp_page != page)) {
981 rcu_read_unlock();
982 put_page(page);
983 head_skb = page_to_skb(vi, rq, xdp_page, offset,
984 len, PAGE_SIZE, false,
985 metasize, true);
986 return head_skb;
987 }
988 break;
989 case XDP_TX:
990 stats->xdp_tx++;
991 xdpf = xdp_convert_buff_to_frame(&xdp);
992 if (unlikely(!xdpf))
993 goto err_xdp;
994 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
995 if (unlikely(!err)) {
996 xdp_return_frame_rx_napi(xdpf);
997 } else if (unlikely(err < 0)) {
998 trace_xdp_exception(vi->dev, xdp_prog, act);
999 if (unlikely(xdp_page != page))
1000 put_page(xdp_page);
1001 goto err_xdp;
1002 }
1003 *xdp_xmit |= VIRTIO_XDP_TX;
1004 if (unlikely(xdp_page != page))
1005 put_page(page);
1006 rcu_read_unlock();
1007 goto xdp_xmit;
1008 case XDP_REDIRECT:
1009 stats->xdp_redirects++;
1010 err = xdp_do_redirect(dev, &xdp, xdp_prog);
1011 if (err) {
1012 if (unlikely(xdp_page != page))
1013 put_page(xdp_page);
1014 goto err_xdp;
1015 }
1016 *xdp_xmit |= VIRTIO_XDP_REDIR;
1017 if (unlikely(xdp_page != page))
1018 put_page(page);
1019 rcu_read_unlock();
1020 goto xdp_xmit;
1021 default:
1022 bpf_warn_invalid_xdp_action(act);
1023 fallthrough;
1024 case XDP_ABORTED:
1025 trace_xdp_exception(vi->dev, xdp_prog, act);
1026 fallthrough;
1027 case XDP_DROP:
1028 if (unlikely(xdp_page != page))
1029 __free_pages(xdp_page, 0);
1030 goto err_xdp;
1031 }
1032 }
1033 rcu_read_unlock();
1034
1035 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
1036 metasize, !!headroom);
1037 curr_skb = head_skb;
1038
1039 if (unlikely(!curr_skb))
1040 goto err_skb;
1041 while (--num_buf) {
1042 int num_skb_frags;
1043
1044 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1045 if (unlikely(!buf)) {
1046 pr_debug("%s: rx error: %d buffers out of %d missing\n",
1047 dev->name, num_buf,
1048 virtio16_to_cpu(vi->vdev,
1049 hdr->num_buffers));
1050 dev->stats.rx_length_errors++;
1051 goto err_buf;
1052 }
1053
1054 stats->bytes += len;
1055 page = virt_to_head_page(buf);
1056
1057 truesize = mergeable_ctx_to_truesize(ctx);
1058 if (unlikely(len > truesize)) {
1059 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1060 dev->name, len, (unsigned long)ctx);
1061 dev->stats.rx_length_errors++;
1062 goto err_skb;
1063 }
1064
1065 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1066 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1067 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1068
1069 if (unlikely(!nskb))
1070 goto err_skb;
1071 if (curr_skb == head_skb)
1072 skb_shinfo(curr_skb)->frag_list = nskb;
1073 else
1074 curr_skb->next = nskb;
1075 curr_skb = nskb;
1076 head_skb->truesize += nskb->truesize;
1077 num_skb_frags = 0;
1078 }
1079 if (curr_skb != head_skb) {
1080 head_skb->data_len += len;
1081 head_skb->len += len;
1082 head_skb->truesize += truesize;
1083 }
1084 offset = buf - page_address(page);
1085 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1086 put_page(page);
1087 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1088 len, truesize);
1089 } else {
1090 skb_add_rx_frag(curr_skb, num_skb_frags, page,
1091 offset, len, truesize);
1092 }
1093 }
1094
1095 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1096 return head_skb;
1097
1098err_xdp:
1099 rcu_read_unlock();
1100 stats->xdp_drops++;
1101err_skb:
1102 put_page(page);
1103 while (num_buf-- > 1) {
1104 buf = virtqueue_get_buf(rq->vq, &len);
1105 if (unlikely(!buf)) {
1106 pr_debug("%s: rx error: %d buffers missing\n",
1107 dev->name, num_buf);
1108 dev->stats.rx_length_errors++;
1109 break;
1110 }
1111 stats->bytes += len;
1112 page = virt_to_head_page(buf);
1113 put_page(page);
1114 }
1115err_buf:
1116 stats->drops++;
1117 dev_kfree_skb(head_skb);
1118xdp_xmit:
1119 return NULL;
1120}
1121
1122static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1123 void *buf, unsigned int len, void **ctx,
1124 unsigned int *xdp_xmit,
1125 struct virtnet_rq_stats *stats)
1126{
1127 struct net_device *dev = vi->dev;
1128 struct sk_buff *skb;
1129 struct virtio_net_hdr_mrg_rxbuf *hdr;
1130
1131 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1132 pr_debug("%s: short packet %i\n", dev->name, len);
1133 dev->stats.rx_length_errors++;
1134 if (vi->mergeable_rx_bufs) {
1135 put_page(virt_to_head_page(buf));
1136 } else if (vi->big_packets) {
1137 give_pages(rq, buf);
1138 } else {
1139 put_page(virt_to_head_page(buf));
1140 }
1141 return;
1142 }
1143
1144 if (vi->mergeable_rx_bufs)
1145 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1146 stats);
1147 else if (vi->big_packets)
1148 skb = receive_big(dev, vi, rq, buf, len, stats);
1149 else
1150 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1151
1152 if (unlikely(!skb))
1153 return;
1154
1155 hdr = skb_vnet_hdr(skb);
1156
1157 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1158 skb->ip_summed = CHECKSUM_UNNECESSARY;
1159
1160 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1161 virtio_is_little_endian(vi->vdev))) {
1162 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1163 dev->name, hdr->hdr.gso_type,
1164 hdr->hdr.gso_size);
1165 goto frame_err;
1166 }
1167
1168 skb_record_rx_queue(skb, vq2rxq(rq->vq));
1169 skb->protocol = eth_type_trans(skb, dev);
1170 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1171 ntohs(skb->protocol), skb->len, skb->pkt_type);
1172
1173 napi_gro_receive(&rq->napi, skb);
1174 return;
1175
1176frame_err:
1177 dev->stats.rx_frame_errors++;
1178 dev_kfree_skb(skb);
1179}
1180
1181/* Unlike mergeable buffers, all buffers are allocated to the
1182 * same size, except for the headroom. For this reason we do
1183 * not need to use mergeable_len_to_ctx here - it is enough
1184 * to store the headroom as the context ignoring the truesize.
1185 */
1186static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1187 gfp_t gfp)
1188{
1189 struct page_frag *alloc_frag = &rq->alloc_frag;
1190 char *buf;
1191 unsigned int xdp_headroom = virtnet_get_headroom(vi);
1192 void *ctx = (void *)(unsigned long)xdp_headroom;
1193 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1194 int err;
1195
1196 len = SKB_DATA_ALIGN(len) +
1197 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1198 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1199 return -ENOMEM;
1200
1201 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1202 get_page(alloc_frag->page);
1203 alloc_frag->offset += len;
1204 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1205 vi->hdr_len + GOOD_PACKET_LEN);
1206 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1207 if (err < 0)
1208 put_page(virt_to_head_page(buf));
1209 return err;
1210}
1211
1212static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1213 gfp_t gfp)
1214{
1215 struct page *first, *list = NULL;
1216 char *p;
1217 int i, err, offset;
1218
1219 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1220
1221 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1222 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1223 first = get_a_page(rq, gfp);
1224 if (!first) {
1225 if (list)
1226 give_pages(rq, list);
1227 return -ENOMEM;
1228 }
1229 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1230
1231 /* chain new page in list head to match sg */
1232 first->private = (unsigned long)list;
1233 list = first;
1234 }
1235
1236 first = get_a_page(rq, gfp);
1237 if (!first) {
1238 give_pages(rq, list);
1239 return -ENOMEM;
1240 }
1241 p = page_address(first);
1242
1243 /* rq->sg[0], rq->sg[1] share the same page */
1244 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1245 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1246
1247 /* rq->sg[1] for data packet, from offset */
1248 offset = sizeof(struct padded_vnet_hdr);
1249 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1250
1251 /* chain first in list head */
1252 first->private = (unsigned long)list;
1253 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1254 first, gfp);
1255 if (err < 0)
1256 give_pages(rq, first);
1257
1258 return err;
1259}
1260
1261static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1262 struct ewma_pkt_len *avg_pkt_len,
1263 unsigned int room)
1264{
1265 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1266 unsigned int len;
1267
1268 if (room)
1269 return PAGE_SIZE - room;
1270
1271 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1272 rq->min_buf_len, PAGE_SIZE - hdr_len);
1273
1274 return ALIGN(len, L1_CACHE_BYTES);
1275}
1276
1277static int add_recvbuf_mergeable(struct virtnet_info *vi,
1278 struct receive_queue *rq, gfp_t gfp)
1279{
1280 struct page_frag *alloc_frag = &rq->alloc_frag;
1281 unsigned int headroom = virtnet_get_headroom(vi);
1282 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1283 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1284 char *buf;
1285 void *ctx;
1286 int err;
1287 unsigned int len, hole;
1288
1289 /* Extra tailroom is needed to satisfy XDP's assumption. This
1290 * means rx frags coalescing won't work, but consider we've
1291 * disabled GSO for XDP, it won't be a big issue.
1292 */
1293 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1294 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1295 return -ENOMEM;
1296
1297 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1298 buf += headroom; /* advance address leaving hole at front of pkt */
1299 get_page(alloc_frag->page);
1300 alloc_frag->offset += len + room;
1301 hole = alloc_frag->size - alloc_frag->offset;
1302 if (hole < len + room) {
1303 /* To avoid internal fragmentation, if there is very likely not
1304 * enough space for another buffer, add the remaining space to
1305 * the current buffer.
1306 */
1307 len += hole;
1308 alloc_frag->offset += hole;
1309 }
1310
1311 sg_init_one(rq->sg, buf, len);
1312 ctx = mergeable_len_to_ctx(len, headroom);
1313 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1314 if (err < 0)
1315 put_page(virt_to_head_page(buf));
1316
1317 return err;
1318}
1319
1320/*
1321 * Returns false if we couldn't fill entirely (OOM).
1322 *
1323 * Normally run in the receive path, but can also be run from ndo_open
1324 * before we're receiving packets, or from refill_work which is
1325 * careful to disable receiving (using napi_disable).
1326 */
1327static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1328 gfp_t gfp)
1329{
1330 int err;
1331 bool oom;
1332
1333 do {
1334 if (vi->mergeable_rx_bufs)
1335 err = add_recvbuf_mergeable(vi, rq, gfp);
1336 else if (vi->big_packets)
1337 err = add_recvbuf_big(vi, rq, gfp);
1338 else
1339 err = add_recvbuf_small(vi, rq, gfp);
1340
1341 oom = err == -ENOMEM;
1342 if (err)
1343 break;
1344 } while (rq->vq->num_free);
1345 if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1346 unsigned long flags;
1347
1348 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1349 rq->stats.kicks++;
1350 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1351 }
1352
1353 return !oom;
1354}
1355
1356static void skb_recv_done(struct virtqueue *rvq)
1357{
1358 struct virtnet_info *vi = rvq->vdev->priv;
1359 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1360
1361 virtqueue_napi_schedule(&rq->napi, rvq);
1362}
1363
1364static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1365{
1366 napi_enable(napi);
1367
1368 /* If all buffers were filled by other side before we napi_enabled, we
1369 * won't get another interrupt, so process any outstanding packets now.
1370 * Call local_bh_enable after to trigger softIRQ processing.
1371 */
1372 local_bh_disable();
1373 virtqueue_napi_schedule(napi, vq);
1374 local_bh_enable();
1375}
1376
1377static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1378 struct virtqueue *vq,
1379 struct napi_struct *napi)
1380{
1381 if (!napi->weight)
1382 return;
1383
1384 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1385 * enable the feature if this is likely affine with the transmit path.
1386 */
1387 if (!vi->affinity_hint_set) {
1388 napi->weight = 0;
1389 return;
1390 }
1391
1392 return virtnet_napi_enable(vq, napi);
1393}
1394
1395static void virtnet_napi_tx_disable(struct napi_struct *napi)
1396{
1397 if (napi->weight)
1398 napi_disable(napi);
1399}
1400
1401static void refill_work(struct work_struct *work)
1402{
1403 struct virtnet_info *vi =
1404 container_of(work, struct virtnet_info, refill.work);
1405 bool still_empty;
1406 int i;
1407
1408 for (i = 0; i < vi->curr_queue_pairs; i++) {
1409 struct receive_queue *rq = &vi->rq[i];
1410
1411 napi_disable(&rq->napi);
1412 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1413 virtnet_napi_enable(rq->vq, &rq->napi);
1414
1415 /* In theory, this can happen: if we don't get any buffers in
1416 * we will *never* try to fill again.
1417 */
1418 if (still_empty)
1419 schedule_delayed_work(&vi->refill, HZ/2);
1420 }
1421}
1422
1423static int virtnet_receive(struct receive_queue *rq, int budget,
1424 unsigned int *xdp_xmit)
1425{
1426 struct virtnet_info *vi = rq->vq->vdev->priv;
1427 struct virtnet_rq_stats stats = {};
1428 unsigned int len;
1429 void *buf;
1430 int i;
1431
1432 if (!vi->big_packets || vi->mergeable_rx_bufs) {
1433 void *ctx;
1434
1435 while (stats.packets < budget &&
1436 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1437 receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1438 stats.packets++;
1439 }
1440 } else {
1441 while (stats.packets < budget &&
1442 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1443 receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1444 stats.packets++;
1445 }
1446 }
1447
1448 if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1449 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1450 schedule_delayed_work(&vi->refill, 0);
1451 }
1452
1453 u64_stats_update_begin(&rq->stats.syncp);
1454 for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1455 size_t offset = virtnet_rq_stats_desc[i].offset;
1456 u64 *item;
1457
1458 item = (u64 *)((u8 *)&rq->stats + offset);
1459 *item += *(u64 *)((u8 *)&stats + offset);
1460 }
1461 u64_stats_update_end(&rq->stats.syncp);
1462
1463 return stats.packets;
1464}
1465
1466static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1467{
1468 unsigned int len;
1469 unsigned int packets = 0;
1470 unsigned int bytes = 0;
1471 void *ptr;
1472
1473 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1474 if (likely(!is_xdp_frame(ptr))) {
1475 struct sk_buff *skb = ptr;
1476
1477 pr_debug("Sent skb %p\n", skb);
1478
1479 bytes += skb->len;
1480 napi_consume_skb(skb, in_napi);
1481 } else {
1482 struct xdp_frame *frame = ptr_to_xdp(ptr);
1483
1484 bytes += frame->len;
1485 xdp_return_frame(frame);
1486 }
1487 packets++;
1488 }
1489
1490 /* Avoid overhead when no packets have been processed
1491 * happens when called speculatively from start_xmit.
1492 */
1493 if (!packets)
1494 return;
1495
1496 u64_stats_update_begin(&sq->stats.syncp);
1497 sq->stats.bytes += bytes;
1498 sq->stats.packets += packets;
1499 u64_stats_update_end(&sq->stats.syncp);
1500}
1501
1502static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1503{
1504 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1505 return false;
1506 else if (q < vi->curr_queue_pairs)
1507 return true;
1508 else
1509 return false;
1510}
1511
1512static void virtnet_poll_cleantx(struct receive_queue *rq)
1513{
1514 struct virtnet_info *vi = rq->vq->vdev->priv;
1515 unsigned int index = vq2rxq(rq->vq);
1516 struct send_queue *sq = &vi->sq[index];
1517 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1518
1519 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1520 return;
1521
1522 if (__netif_tx_trylock(txq)) {
1523 do {
1524 virtqueue_disable_cb(sq->vq);
1525 free_old_xmit_skbs(sq, true);
1526 } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1527
1528 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1529 netif_tx_wake_queue(txq);
1530
1531 __netif_tx_unlock(txq);
1532 }
1533}
1534
1535static int virtnet_poll(struct napi_struct *napi, int budget)
1536{
1537 struct receive_queue *rq =
1538 container_of(napi, struct receive_queue, napi);
1539 struct virtnet_info *vi = rq->vq->vdev->priv;
1540 struct send_queue *sq;
1541 unsigned int received;
1542 unsigned int xdp_xmit = 0;
1543
1544 virtnet_poll_cleantx(rq);
1545
1546 received = virtnet_receive(rq, budget, &xdp_xmit);
1547
1548 /* Out of packets? */
1549 if (received < budget)
1550 virtqueue_napi_complete(napi, rq->vq, received);
1551
1552 if (xdp_xmit & VIRTIO_XDP_REDIR)
1553 xdp_do_flush();
1554
1555 if (xdp_xmit & VIRTIO_XDP_TX) {
1556 sq = virtnet_xdp_get_sq(vi);
1557 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1558 u64_stats_update_begin(&sq->stats.syncp);
1559 sq->stats.kicks++;
1560 u64_stats_update_end(&sq->stats.syncp);
1561 }
1562 virtnet_xdp_put_sq(vi, sq);
1563 }
1564
1565 return received;
1566}
1567
1568static int virtnet_open(struct net_device *dev)
1569{
1570 struct virtnet_info *vi = netdev_priv(dev);
1571 int i, err;
1572
1573 for (i = 0; i < vi->max_queue_pairs; i++) {
1574 if (i < vi->curr_queue_pairs)
1575 /* Make sure we have some buffers: if oom use wq. */
1576 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1577 schedule_delayed_work(&vi->refill, 0);
1578
1579 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id);
1580 if (err < 0)
1581 return err;
1582
1583 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1584 MEM_TYPE_PAGE_SHARED, NULL);
1585 if (err < 0) {
1586 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1587 return err;
1588 }
1589
1590 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1591 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1592 }
1593
1594 return 0;
1595}
1596
1597static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1598{
1599 struct send_queue *sq = container_of(napi, struct send_queue, napi);
1600 struct virtnet_info *vi = sq->vq->vdev->priv;
1601 unsigned int index = vq2txq(sq->vq);
1602 struct netdev_queue *txq;
1603 int opaque;
1604 bool done;
1605
1606 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1607 /* We don't need to enable cb for XDP */
1608 napi_complete_done(napi, 0);
1609 return 0;
1610 }
1611
1612 txq = netdev_get_tx_queue(vi->dev, index);
1613 __netif_tx_lock(txq, raw_smp_processor_id());
1614 virtqueue_disable_cb(sq->vq);
1615 free_old_xmit_skbs(sq, true);
1616
1617 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1618 netif_tx_wake_queue(txq);
1619
1620 opaque = virtqueue_enable_cb_prepare(sq->vq);
1621
1622 done = napi_complete_done(napi, 0);
1623
1624 if (!done)
1625 virtqueue_disable_cb(sq->vq);
1626
1627 __netif_tx_unlock(txq);
1628
1629 if (done) {
1630 if (unlikely(virtqueue_poll(sq->vq, opaque))) {
1631 if (napi_schedule_prep(napi)) {
1632 __netif_tx_lock(txq, raw_smp_processor_id());
1633 virtqueue_disable_cb(sq->vq);
1634 __netif_tx_unlock(txq);
1635 __napi_schedule(napi);
1636 }
1637 }
1638 }
1639
1640 return 0;
1641}
1642
1643static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1644{
1645 struct virtio_net_hdr_mrg_rxbuf *hdr;
1646 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1647 struct virtnet_info *vi = sq->vq->vdev->priv;
1648 int num_sg;
1649 unsigned hdr_len = vi->hdr_len;
1650 bool can_push;
1651
1652 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1653
1654 can_push = vi->any_header_sg &&
1655 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1656 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1657 /* Even if we can, don't push here yet as this would skew
1658 * csum_start offset below. */
1659 if (can_push)
1660 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1661 else
1662 hdr = skb_vnet_hdr(skb);
1663
1664 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1665 virtio_is_little_endian(vi->vdev), false,
1666 0))
1667 return -EPROTO;
1668
1669 if (vi->mergeable_rx_bufs)
1670 hdr->num_buffers = 0;
1671
1672 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1673 if (can_push) {
1674 __skb_push(skb, hdr_len);
1675 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1676 if (unlikely(num_sg < 0))
1677 return num_sg;
1678 /* Pull header back to avoid skew in tx bytes calculations. */
1679 __skb_pull(skb, hdr_len);
1680 } else {
1681 sg_set_buf(sq->sg, hdr, hdr_len);
1682 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1683 if (unlikely(num_sg < 0))
1684 return num_sg;
1685 num_sg++;
1686 }
1687 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1688}
1689
1690static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1691{
1692 struct virtnet_info *vi = netdev_priv(dev);
1693 int qnum = skb_get_queue_mapping(skb);
1694 struct send_queue *sq = &vi->sq[qnum];
1695 int err;
1696 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1697 bool kick = !netdev_xmit_more();
1698 bool use_napi = sq->napi.weight;
1699
1700 /* Free up any pending old buffers before queueing new ones. */
1701 do {
1702 if (use_napi)
1703 virtqueue_disable_cb(sq->vq);
1704
1705 free_old_xmit_skbs(sq, false);
1706
1707 } while (use_napi && kick &&
1708 unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1709
1710 /* timestamp packet in software */
1711 skb_tx_timestamp(skb);
1712
1713 /* Try to transmit */
1714 err = xmit_skb(sq, skb);
1715
1716 /* This should not happen! */
1717 if (unlikely(err)) {
1718 dev->stats.tx_fifo_errors++;
1719 if (net_ratelimit())
1720 dev_warn(&dev->dev,
1721 "Unexpected TXQ (%d) queue failure: %d\n",
1722 qnum, err);
1723 dev->stats.tx_dropped++;
1724 dev_kfree_skb_any(skb);
1725 return NETDEV_TX_OK;
1726 }
1727
1728 /* Don't wait up for transmitted skbs to be freed. */
1729 if (!use_napi) {
1730 skb_orphan(skb);
1731 nf_reset_ct(skb);
1732 }
1733
1734 /* If running out of space, stop queue to avoid getting packets that we
1735 * are then unable to transmit.
1736 * An alternative would be to force queuing layer to requeue the skb by
1737 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1738 * returned in a normal path of operation: it means that driver is not
1739 * maintaining the TX queue stop/start state properly, and causes
1740 * the stack to do a non-trivial amount of useless work.
1741 * Since most packets only take 1 or 2 ring slots, stopping the queue
1742 * early means 16 slots are typically wasted.
1743 */
1744 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1745 netif_stop_subqueue(dev, qnum);
1746 if (!use_napi &&
1747 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1748 /* More just got used, free them then recheck. */
1749 free_old_xmit_skbs(sq, false);
1750 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1751 netif_start_subqueue(dev, qnum);
1752 virtqueue_disable_cb(sq->vq);
1753 }
1754 }
1755 }
1756
1757 if (kick || netif_xmit_stopped(txq)) {
1758 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1759 u64_stats_update_begin(&sq->stats.syncp);
1760 sq->stats.kicks++;
1761 u64_stats_update_end(&sq->stats.syncp);
1762 }
1763 }
1764
1765 return NETDEV_TX_OK;
1766}
1767
1768/*
1769 * Send command via the control virtqueue and check status. Commands
1770 * supported by the hypervisor, as indicated by feature bits, should
1771 * never fail unless improperly formatted.
1772 */
1773static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1774 struct scatterlist *out)
1775{
1776 struct scatterlist *sgs[4], hdr, stat;
1777 unsigned out_num = 0, tmp;
1778 int ret;
1779
1780 /* Caller should know better */
1781 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1782
1783 vi->ctrl->status = ~0;
1784 vi->ctrl->hdr.class = class;
1785 vi->ctrl->hdr.cmd = cmd;
1786 /* Add header */
1787 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1788 sgs[out_num++] = &hdr;
1789
1790 if (out)
1791 sgs[out_num++] = out;
1792
1793 /* Add return status. */
1794 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1795 sgs[out_num] = &stat;
1796
1797 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1798 ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1799 if (ret < 0) {
1800 dev_warn(&vi->vdev->dev,
1801 "Failed to add sgs for command vq: %d\n.", ret);
1802 return false;
1803 }
1804
1805 if (unlikely(!virtqueue_kick(vi->cvq)))
1806 return vi->ctrl->status == VIRTIO_NET_OK;
1807
1808 /* Spin for a response, the kick causes an ioport write, trapping
1809 * into the hypervisor, so the request should be handled immediately.
1810 */
1811 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1812 !virtqueue_is_broken(vi->cvq))
1813 cpu_relax();
1814
1815 return vi->ctrl->status == VIRTIO_NET_OK;
1816}
1817
1818static int virtnet_set_mac_address(struct net_device *dev, void *p)
1819{
1820 struct virtnet_info *vi = netdev_priv(dev);
1821 struct virtio_device *vdev = vi->vdev;
1822 int ret;
1823 struct sockaddr *addr;
1824 struct scatterlist sg;
1825
1826 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1827 return -EOPNOTSUPP;
1828
1829 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1830 if (!addr)
1831 return -ENOMEM;
1832
1833 ret = eth_prepare_mac_addr_change(dev, addr);
1834 if (ret)
1835 goto out;
1836
1837 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1838 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1839 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1840 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1841 dev_warn(&vdev->dev,
1842 "Failed to set mac address by vq command.\n");
1843 ret = -EINVAL;
1844 goto out;
1845 }
1846 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1847 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1848 unsigned int i;
1849
1850 /* Naturally, this has an atomicity problem. */
1851 for (i = 0; i < dev->addr_len; i++)
1852 virtio_cwrite8(vdev,
1853 offsetof(struct virtio_net_config, mac) +
1854 i, addr->sa_data[i]);
1855 }
1856
1857 eth_commit_mac_addr_change(dev, p);
1858 ret = 0;
1859
1860out:
1861 kfree(addr);
1862 return ret;
1863}
1864
1865static void virtnet_stats(struct net_device *dev,
1866 struct rtnl_link_stats64 *tot)
1867{
1868 struct virtnet_info *vi = netdev_priv(dev);
1869 unsigned int start;
1870 int i;
1871
1872 for (i = 0; i < vi->max_queue_pairs; i++) {
1873 u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1874 struct receive_queue *rq = &vi->rq[i];
1875 struct send_queue *sq = &vi->sq[i];
1876
1877 do {
1878 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1879 tpackets = sq->stats.packets;
1880 tbytes = sq->stats.bytes;
1881 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1882
1883 do {
1884 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1885 rpackets = rq->stats.packets;
1886 rbytes = rq->stats.bytes;
1887 rdrops = rq->stats.drops;
1888 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1889
1890 tot->rx_packets += rpackets;
1891 tot->tx_packets += tpackets;
1892 tot->rx_bytes += rbytes;
1893 tot->tx_bytes += tbytes;
1894 tot->rx_dropped += rdrops;
1895 }
1896
1897 tot->tx_dropped = dev->stats.tx_dropped;
1898 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1899 tot->rx_length_errors = dev->stats.rx_length_errors;
1900 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1901}
1902
1903static void virtnet_ack_link_announce(struct virtnet_info *vi)
1904{
1905 rtnl_lock();
1906 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1907 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1908 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1909 rtnl_unlock();
1910}
1911
1912static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1913{
1914 struct scatterlist sg;
1915 struct net_device *dev = vi->dev;
1916
1917 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1918 return 0;
1919
1920 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1921 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1922
1923 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1924 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1925 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1926 queue_pairs);
1927 return -EINVAL;
1928 } else {
1929 vi->curr_queue_pairs = queue_pairs;
1930 /* virtnet_open() will refill when device is going to up. */
1931 if (dev->flags & IFF_UP)
1932 schedule_delayed_work(&vi->refill, 0);
1933 }
1934
1935 return 0;
1936}
1937
1938static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1939{
1940 int err;
1941
1942 rtnl_lock();
1943 err = _virtnet_set_queues(vi, queue_pairs);
1944 rtnl_unlock();
1945 return err;
1946}
1947
1948static int virtnet_close(struct net_device *dev)
1949{
1950 struct virtnet_info *vi = netdev_priv(dev);
1951 int i;
1952
1953 /* Make sure refill_work doesn't re-enable napi! */
1954 cancel_delayed_work_sync(&vi->refill);
1955
1956 for (i = 0; i < vi->max_queue_pairs; i++) {
1957 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1958 napi_disable(&vi->rq[i].napi);
1959 virtnet_napi_tx_disable(&vi->sq[i].napi);
1960 }
1961
1962 return 0;
1963}
1964
1965static void virtnet_set_rx_mode(struct net_device *dev)
1966{
1967 struct virtnet_info *vi = netdev_priv(dev);
1968 struct scatterlist sg[2];
1969 struct virtio_net_ctrl_mac *mac_data;
1970 struct netdev_hw_addr *ha;
1971 int uc_count;
1972 int mc_count;
1973 void *buf;
1974 int i;
1975
1976 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1977 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1978 return;
1979
1980 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1981 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1982
1983 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1984
1985 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1986 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1987 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1988 vi->ctrl->promisc ? "en" : "dis");
1989
1990 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1991
1992 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1993 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1994 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1995 vi->ctrl->allmulti ? "en" : "dis");
1996
1997 uc_count = netdev_uc_count(dev);
1998 mc_count = netdev_mc_count(dev);
1999 /* MAC filter - use one buffer for both lists */
2000 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
2001 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
2002 mac_data = buf;
2003 if (!buf)
2004 return;
2005
2006 sg_init_table(sg, 2);
2007
2008 /* Store the unicast list and count in the front of the buffer */
2009 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
2010 i = 0;
2011 netdev_for_each_uc_addr(ha, dev)
2012 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2013
2014 sg_set_buf(&sg[0], mac_data,
2015 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
2016
2017 /* multicast list and count fill the end */
2018 mac_data = (void *)&mac_data->macs[uc_count][0];
2019
2020 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
2021 i = 0;
2022 netdev_for_each_mc_addr(ha, dev)
2023 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2024
2025 sg_set_buf(&sg[1], mac_data,
2026 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
2027
2028 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2029 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
2030 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
2031
2032 kfree(buf);
2033}
2034
2035static int virtnet_vlan_rx_add_vid(struct net_device *dev,
2036 __be16 proto, u16 vid)
2037{
2038 struct virtnet_info *vi = netdev_priv(dev);
2039 struct scatterlist sg;
2040
2041 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2042 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2043
2044 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2045 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2046 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2047 return 0;
2048}
2049
2050static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2051 __be16 proto, u16 vid)
2052{
2053 struct virtnet_info *vi = netdev_priv(dev);
2054 struct scatterlist sg;
2055
2056 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2057 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2058
2059 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2060 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2061 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2062 return 0;
2063}
2064
2065static void virtnet_clean_affinity(struct virtnet_info *vi)
2066{
2067 int i;
2068
2069 if (vi->affinity_hint_set) {
2070 for (i = 0; i < vi->max_queue_pairs; i++) {
2071 virtqueue_set_affinity(vi->rq[i].vq, NULL);
2072 virtqueue_set_affinity(vi->sq[i].vq, NULL);
2073 }
2074
2075 vi->affinity_hint_set = false;
2076 }
2077}
2078
2079static void virtnet_set_affinity(struct virtnet_info *vi)
2080{
2081 cpumask_var_t mask;
2082 int stragglers;
2083 int group_size;
2084 int i, j, cpu;
2085 int num_cpu;
2086 int stride;
2087
2088 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2089 virtnet_clean_affinity(vi);
2090 return;
2091 }
2092
2093 num_cpu = num_online_cpus();
2094 stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2095 stragglers = num_cpu >= vi->curr_queue_pairs ?
2096 num_cpu % vi->curr_queue_pairs :
2097 0;
2098 cpu = cpumask_next(-1, cpu_online_mask);
2099
2100 for (i = 0; i < vi->curr_queue_pairs; i++) {
2101 group_size = stride + (i < stragglers ? 1 : 0);
2102
2103 for (j = 0; j < group_size; j++) {
2104 cpumask_set_cpu(cpu, mask);
2105 cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2106 nr_cpu_ids, false);
2107 }
2108 virtqueue_set_affinity(vi->rq[i].vq, mask);
2109 virtqueue_set_affinity(vi->sq[i].vq, mask);
2110 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2111 cpumask_clear(mask);
2112 }
2113
2114 vi->affinity_hint_set = true;
2115 free_cpumask_var(mask);
2116}
2117
2118static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2119{
2120 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2121 node);
2122 virtnet_set_affinity(vi);
2123 return 0;
2124}
2125
2126static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2127{
2128 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2129 node_dead);
2130 virtnet_set_affinity(vi);
2131 return 0;
2132}
2133
2134static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2135{
2136 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2137 node);
2138
2139 virtnet_clean_affinity(vi);
2140 return 0;
2141}
2142
2143static enum cpuhp_state virtionet_online;
2144
2145static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2146{
2147 int ret;
2148
2149 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2150 if (ret)
2151 return ret;
2152 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2153 &vi->node_dead);
2154 if (!ret)
2155 return ret;
2156 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2157 return ret;
2158}
2159
2160static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2161{
2162 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2163 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2164 &vi->node_dead);
2165}
2166
2167static void virtnet_get_ringparam(struct net_device *dev,
2168 struct ethtool_ringparam *ring)
2169{
2170 struct virtnet_info *vi = netdev_priv(dev);
2171
2172 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2173 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2174 ring->rx_pending = ring->rx_max_pending;
2175 ring->tx_pending = ring->tx_max_pending;
2176}
2177
2178
2179static void virtnet_get_drvinfo(struct net_device *dev,
2180 struct ethtool_drvinfo *info)
2181{
2182 struct virtnet_info *vi = netdev_priv(dev);
2183 struct virtio_device *vdev = vi->vdev;
2184
2185 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2186 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2187 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2188
2189}
2190
2191/* TODO: Eliminate OOO packets during switching */
2192static int virtnet_set_channels(struct net_device *dev,
2193 struct ethtool_channels *channels)
2194{
2195 struct virtnet_info *vi = netdev_priv(dev);
2196 u16 queue_pairs = channels->combined_count;
2197 int err;
2198
2199 /* We don't support separate rx/tx channels.
2200 * We don't allow setting 'other' channels.
2201 */
2202 if (channels->rx_count || channels->tx_count || channels->other_count)
2203 return -EINVAL;
2204
2205 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2206 return -EINVAL;
2207
2208 /* For now we don't support modifying channels while XDP is loaded
2209 * also when XDP is loaded all RX queues have XDP programs so we only
2210 * need to check a single RX queue.
2211 */
2212 if (vi->rq[0].xdp_prog)
2213 return -EINVAL;
2214
2215 get_online_cpus();
2216 err = _virtnet_set_queues(vi, queue_pairs);
2217 if (err) {
2218 put_online_cpus();
2219 goto err;
2220 }
2221 virtnet_set_affinity(vi);
2222 put_online_cpus();
2223
2224 netif_set_real_num_tx_queues(dev, queue_pairs);
2225 netif_set_real_num_rx_queues(dev, queue_pairs);
2226 err:
2227 return err;
2228}
2229
2230static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2231{
2232 struct virtnet_info *vi = netdev_priv(dev);
2233 unsigned int i, j;
2234 u8 *p = data;
2235
2236 switch (stringset) {
2237 case ETH_SS_STATS:
2238 for (i = 0; i < vi->curr_queue_pairs; i++) {
2239 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
2240 ethtool_sprintf(&p, "rx_queue_%u_%s", i,
2241 virtnet_rq_stats_desc[j].desc);
2242 }
2243
2244 for (i = 0; i < vi->curr_queue_pairs; i++) {
2245 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
2246 ethtool_sprintf(&p, "tx_queue_%u_%s", i,
2247 virtnet_sq_stats_desc[j].desc);
2248 }
2249 break;
2250 }
2251}
2252
2253static int virtnet_get_sset_count(struct net_device *dev, int sset)
2254{
2255 struct virtnet_info *vi = netdev_priv(dev);
2256
2257 switch (sset) {
2258 case ETH_SS_STATS:
2259 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2260 VIRTNET_SQ_STATS_LEN);
2261 default:
2262 return -EOPNOTSUPP;
2263 }
2264}
2265
2266static void virtnet_get_ethtool_stats(struct net_device *dev,
2267 struct ethtool_stats *stats, u64 *data)
2268{
2269 struct virtnet_info *vi = netdev_priv(dev);
2270 unsigned int idx = 0, start, i, j;
2271 const u8 *stats_base;
2272 size_t offset;
2273
2274 for (i = 0; i < vi->curr_queue_pairs; i++) {
2275 struct receive_queue *rq = &vi->rq[i];
2276
2277 stats_base = (u8 *)&rq->stats;
2278 do {
2279 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2280 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2281 offset = virtnet_rq_stats_desc[j].offset;
2282 data[idx + j] = *(u64 *)(stats_base + offset);
2283 }
2284 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2285 idx += VIRTNET_RQ_STATS_LEN;
2286 }
2287
2288 for (i = 0; i < vi->curr_queue_pairs; i++) {
2289 struct send_queue *sq = &vi->sq[i];
2290
2291 stats_base = (u8 *)&sq->stats;
2292 do {
2293 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2294 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2295 offset = virtnet_sq_stats_desc[j].offset;
2296 data[idx + j] = *(u64 *)(stats_base + offset);
2297 }
2298 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2299 idx += VIRTNET_SQ_STATS_LEN;
2300 }
2301}
2302
2303static void virtnet_get_channels(struct net_device *dev,
2304 struct ethtool_channels *channels)
2305{
2306 struct virtnet_info *vi = netdev_priv(dev);
2307
2308 channels->combined_count = vi->curr_queue_pairs;
2309 channels->max_combined = vi->max_queue_pairs;
2310 channels->max_other = 0;
2311 channels->rx_count = 0;
2312 channels->tx_count = 0;
2313 channels->other_count = 0;
2314}
2315
2316static int virtnet_set_link_ksettings(struct net_device *dev,
2317 const struct ethtool_link_ksettings *cmd)
2318{
2319 struct virtnet_info *vi = netdev_priv(dev);
2320
2321 return ethtool_virtdev_set_link_ksettings(dev, cmd,
2322 &vi->speed, &vi->duplex);
2323}
2324
2325static int virtnet_get_link_ksettings(struct net_device *dev,
2326 struct ethtool_link_ksettings *cmd)
2327{
2328 struct virtnet_info *vi = netdev_priv(dev);
2329
2330 cmd->base.speed = vi->speed;
2331 cmd->base.duplex = vi->duplex;
2332 cmd->base.port = PORT_OTHER;
2333
2334 return 0;
2335}
2336
2337static int virtnet_set_coalesce(struct net_device *dev,
2338 struct ethtool_coalesce *ec)
2339{
2340 struct virtnet_info *vi = netdev_priv(dev);
2341 int i, napi_weight;
2342
2343 if (ec->tx_max_coalesced_frames > 1 ||
2344 ec->rx_max_coalesced_frames != 1)
2345 return -EINVAL;
2346
2347 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2348 if (napi_weight ^ vi->sq[0].napi.weight) {
2349 if (dev->flags & IFF_UP)
2350 return -EBUSY;
2351 for (i = 0; i < vi->max_queue_pairs; i++)
2352 vi->sq[i].napi.weight = napi_weight;
2353 }
2354
2355 return 0;
2356}
2357
2358static int virtnet_get_coalesce(struct net_device *dev,
2359 struct ethtool_coalesce *ec)
2360{
2361 struct ethtool_coalesce ec_default = {
2362 .cmd = ETHTOOL_GCOALESCE,
2363 .rx_max_coalesced_frames = 1,
2364 };
2365 struct virtnet_info *vi = netdev_priv(dev);
2366
2367 memcpy(ec, &ec_default, sizeof(ec_default));
2368
2369 if (vi->sq[0].napi.weight)
2370 ec->tx_max_coalesced_frames = 1;
2371
2372 return 0;
2373}
2374
2375static void virtnet_init_settings(struct net_device *dev)
2376{
2377 struct virtnet_info *vi = netdev_priv(dev);
2378
2379 vi->speed = SPEED_UNKNOWN;
2380 vi->duplex = DUPLEX_UNKNOWN;
2381}
2382
2383static void virtnet_update_settings(struct virtnet_info *vi)
2384{
2385 u32 speed;
2386 u8 duplex;
2387
2388 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2389 return;
2390
2391 virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
2392
2393 if (ethtool_validate_speed(speed))
2394 vi->speed = speed;
2395
2396 virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
2397
2398 if (ethtool_validate_duplex(duplex))
2399 vi->duplex = duplex;
2400}
2401
2402static const struct ethtool_ops virtnet_ethtool_ops = {
2403 .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES,
2404 .get_drvinfo = virtnet_get_drvinfo,
2405 .get_link = ethtool_op_get_link,
2406 .get_ringparam = virtnet_get_ringparam,
2407 .get_strings = virtnet_get_strings,
2408 .get_sset_count = virtnet_get_sset_count,
2409 .get_ethtool_stats = virtnet_get_ethtool_stats,
2410 .set_channels = virtnet_set_channels,
2411 .get_channels = virtnet_get_channels,
2412 .get_ts_info = ethtool_op_get_ts_info,
2413 .get_link_ksettings = virtnet_get_link_ksettings,
2414 .set_link_ksettings = virtnet_set_link_ksettings,
2415 .set_coalesce = virtnet_set_coalesce,
2416 .get_coalesce = virtnet_get_coalesce,
2417};
2418
2419static void virtnet_freeze_down(struct virtio_device *vdev)
2420{
2421 struct virtnet_info *vi = vdev->priv;
2422 int i;
2423
2424 /* Make sure no work handler is accessing the device */
2425 flush_work(&vi->config_work);
2426
2427 netif_tx_lock_bh(vi->dev);
2428 netif_device_detach(vi->dev);
2429 netif_tx_unlock_bh(vi->dev);
2430 cancel_delayed_work_sync(&vi->refill);
2431
2432 if (netif_running(vi->dev)) {
2433 for (i = 0; i < vi->max_queue_pairs; i++) {
2434 napi_disable(&vi->rq[i].napi);
2435 virtnet_napi_tx_disable(&vi->sq[i].napi);
2436 }
2437 }
2438}
2439
2440static int init_vqs(struct virtnet_info *vi);
2441
2442static int virtnet_restore_up(struct virtio_device *vdev)
2443{
2444 struct virtnet_info *vi = vdev->priv;
2445 int err, i;
2446
2447 err = init_vqs(vi);
2448 if (err)
2449 return err;
2450
2451 virtio_device_ready(vdev);
2452
2453 if (netif_running(vi->dev)) {
2454 for (i = 0; i < vi->curr_queue_pairs; i++)
2455 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2456 schedule_delayed_work(&vi->refill, 0);
2457
2458 for (i = 0; i < vi->max_queue_pairs; i++) {
2459 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2460 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2461 &vi->sq[i].napi);
2462 }
2463 }
2464
2465 netif_tx_lock_bh(vi->dev);
2466 netif_device_attach(vi->dev);
2467 netif_tx_unlock_bh(vi->dev);
2468 return err;
2469}
2470
2471static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2472{
2473 struct scatterlist sg;
2474 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2475
2476 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2477
2478 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2479 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2480 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
2481 return -EINVAL;
2482 }
2483
2484 return 0;
2485}
2486
2487static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2488{
2489 u64 offloads = 0;
2490
2491 if (!vi->guest_offloads)
2492 return 0;
2493
2494 return virtnet_set_guest_offloads(vi, offloads);
2495}
2496
2497static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2498{
2499 u64 offloads = vi->guest_offloads;
2500
2501 if (!vi->guest_offloads)
2502 return 0;
2503
2504 return virtnet_set_guest_offloads(vi, offloads);
2505}
2506
2507static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2508 struct netlink_ext_ack *extack)
2509{
2510 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2511 struct virtnet_info *vi = netdev_priv(dev);
2512 struct bpf_prog *old_prog;
2513 u16 xdp_qp = 0, curr_qp;
2514 int i, err;
2515
2516 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2517 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2518 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2519 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2520 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2521 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2522 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
2523 return -EOPNOTSUPP;
2524 }
2525
2526 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2527 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2528 return -EINVAL;
2529 }
2530
2531 if (dev->mtu > max_sz) {
2532 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2533 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2534 return -EINVAL;
2535 }
2536
2537 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2538 if (prog)
2539 xdp_qp = nr_cpu_ids;
2540
2541 /* XDP requires extra queues for XDP_TX */
2542 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2543 netdev_warn(dev, "XDP request %i queues but max is %i. XDP_TX and XDP_REDIRECT will operate in a slower locked tx mode.\n",
2544 curr_qp + xdp_qp, vi->max_queue_pairs);
2545 xdp_qp = 0;
2546 }
2547
2548 old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2549 if (!prog && !old_prog)
2550 return 0;
2551
2552 if (prog)
2553 bpf_prog_add(prog, vi->max_queue_pairs - 1);
2554
2555 /* Make sure NAPI is not using any XDP TX queues for RX. */
2556 if (netif_running(dev)) {
2557 for (i = 0; i < vi->max_queue_pairs; i++) {
2558 napi_disable(&vi->rq[i].napi);
2559 virtnet_napi_tx_disable(&vi->sq[i].napi);
2560 }
2561 }
2562
2563 if (!prog) {
2564 for (i = 0; i < vi->max_queue_pairs; i++) {
2565 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2566 if (i == 0)
2567 virtnet_restore_guest_offloads(vi);
2568 }
2569 synchronize_net();
2570 }
2571
2572 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2573 if (err)
2574 goto err;
2575 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2576 vi->xdp_queue_pairs = xdp_qp;
2577
2578 if (prog) {
2579 vi->xdp_enabled = true;
2580 for (i = 0; i < vi->max_queue_pairs; i++) {
2581 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2582 if (i == 0 && !old_prog)
2583 virtnet_clear_guest_offloads(vi);
2584 }
2585 } else {
2586 vi->xdp_enabled = false;
2587 }
2588
2589 for (i = 0; i < vi->max_queue_pairs; i++) {
2590 if (old_prog)
2591 bpf_prog_put(old_prog);
2592 if (netif_running(dev)) {
2593 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2594 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2595 &vi->sq[i].napi);
2596 }
2597 }
2598
2599 return 0;
2600
2601err:
2602 if (!prog) {
2603 virtnet_clear_guest_offloads(vi);
2604 for (i = 0; i < vi->max_queue_pairs; i++)
2605 rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2606 }
2607
2608 if (netif_running(dev)) {
2609 for (i = 0; i < vi->max_queue_pairs; i++) {
2610 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2611 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2612 &vi->sq[i].napi);
2613 }
2614 }
2615 if (prog)
2616 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2617 return err;
2618}
2619
2620static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2621{
2622 switch (xdp->command) {
2623 case XDP_SETUP_PROG:
2624 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2625 default:
2626 return -EINVAL;
2627 }
2628}
2629
2630static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2631 size_t len)
2632{
2633 struct virtnet_info *vi = netdev_priv(dev);
2634 int ret;
2635
2636 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2637 return -EOPNOTSUPP;
2638
2639 ret = snprintf(buf, len, "sby");
2640 if (ret >= len)
2641 return -EOPNOTSUPP;
2642
2643 return 0;
2644}
2645
2646static int virtnet_set_features(struct net_device *dev,
2647 netdev_features_t features)
2648{
2649 struct virtnet_info *vi = netdev_priv(dev);
2650 u64 offloads;
2651 int err;
2652
2653 if ((dev->features ^ features) & NETIF_F_GRO_HW) {
2654 if (vi->xdp_enabled)
2655 return -EBUSY;
2656
2657 if (features & NETIF_F_GRO_HW)
2658 offloads = vi->guest_offloads_capable;
2659 else
2660 offloads = vi->guest_offloads_capable &
2661 ~GUEST_OFFLOAD_GRO_HW_MASK;
2662
2663 err = virtnet_set_guest_offloads(vi, offloads);
2664 if (err)
2665 return err;
2666 vi->guest_offloads = offloads;
2667 }
2668
2669 return 0;
2670}
2671
2672static const struct net_device_ops virtnet_netdev = {
2673 .ndo_open = virtnet_open,
2674 .ndo_stop = virtnet_close,
2675 .ndo_start_xmit = start_xmit,
2676 .ndo_validate_addr = eth_validate_addr,
2677 .ndo_set_mac_address = virtnet_set_mac_address,
2678 .ndo_set_rx_mode = virtnet_set_rx_mode,
2679 .ndo_get_stats64 = virtnet_stats,
2680 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2681 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2682 .ndo_bpf = virtnet_xdp,
2683 .ndo_xdp_xmit = virtnet_xdp_xmit,
2684 .ndo_features_check = passthru_features_check,
2685 .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2686 .ndo_set_features = virtnet_set_features,
2687};
2688
2689static void virtnet_config_changed_work(struct work_struct *work)
2690{
2691 struct virtnet_info *vi =
2692 container_of(work, struct virtnet_info, config_work);
2693 u16 v;
2694
2695 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2696 struct virtio_net_config, status, &v) < 0)
2697 return;
2698
2699 if (v & VIRTIO_NET_S_ANNOUNCE) {
2700 netdev_notify_peers(vi->dev);
2701 virtnet_ack_link_announce(vi);
2702 }
2703
2704 /* Ignore unknown (future) status bits */
2705 v &= VIRTIO_NET_S_LINK_UP;
2706
2707 if (vi->status == v)
2708 return;
2709
2710 vi->status = v;
2711
2712 if (vi->status & VIRTIO_NET_S_LINK_UP) {
2713 virtnet_update_settings(vi);
2714 netif_carrier_on(vi->dev);
2715 netif_tx_wake_all_queues(vi->dev);
2716 } else {
2717 netif_carrier_off(vi->dev);
2718 netif_tx_stop_all_queues(vi->dev);
2719 }
2720}
2721
2722static void virtnet_config_changed(struct virtio_device *vdev)
2723{
2724 struct virtnet_info *vi = vdev->priv;
2725
2726 schedule_work(&vi->config_work);
2727}
2728
2729static void virtnet_free_queues(struct virtnet_info *vi)
2730{
2731 int i;
2732
2733 for (i = 0; i < vi->max_queue_pairs; i++) {
2734 __netif_napi_del(&vi->rq[i].napi);
2735 __netif_napi_del(&vi->sq[i].napi);
2736 }
2737
2738 /* We called __netif_napi_del(),
2739 * we need to respect an RCU grace period before freeing vi->rq
2740 */
2741 synchronize_net();
2742
2743 kfree(vi->rq);
2744 kfree(vi->sq);
2745 kfree(vi->ctrl);
2746}
2747
2748static void _free_receive_bufs(struct virtnet_info *vi)
2749{
2750 struct bpf_prog *old_prog;
2751 int i;
2752
2753 for (i = 0; i < vi->max_queue_pairs; i++) {
2754 while (vi->rq[i].pages)
2755 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2756
2757 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2758 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2759 if (old_prog)
2760 bpf_prog_put(old_prog);
2761 }
2762}
2763
2764static void free_receive_bufs(struct virtnet_info *vi)
2765{
2766 rtnl_lock();
2767 _free_receive_bufs(vi);
2768 rtnl_unlock();
2769}
2770
2771static void free_receive_page_frags(struct virtnet_info *vi)
2772{
2773 int i;
2774 for (i = 0; i < vi->max_queue_pairs; i++)
2775 if (vi->rq[i].alloc_frag.page)
2776 put_page(vi->rq[i].alloc_frag.page);
2777}
2778
2779static void free_unused_bufs(struct virtnet_info *vi)
2780{
2781 void *buf;
2782 int i;
2783
2784 for (i = 0; i < vi->max_queue_pairs; i++) {
2785 struct virtqueue *vq = vi->sq[i].vq;
2786 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2787 if (!is_xdp_frame(buf))
2788 dev_kfree_skb(buf);
2789 else
2790 xdp_return_frame(ptr_to_xdp(buf));
2791 }
2792 }
2793
2794 for (i = 0; i < vi->max_queue_pairs; i++) {
2795 struct virtqueue *vq = vi->rq[i].vq;
2796
2797 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2798 if (vi->mergeable_rx_bufs) {
2799 put_page(virt_to_head_page(buf));
2800 } else if (vi->big_packets) {
2801 give_pages(&vi->rq[i], buf);
2802 } else {
2803 put_page(virt_to_head_page(buf));
2804 }
2805 }
2806 }
2807}
2808
2809static void virtnet_del_vqs(struct virtnet_info *vi)
2810{
2811 struct virtio_device *vdev = vi->vdev;
2812
2813 virtnet_clean_affinity(vi);
2814
2815 vdev->config->del_vqs(vdev);
2816
2817 virtnet_free_queues(vi);
2818}
2819
2820/* How large should a single buffer be so a queue full of these can fit at
2821 * least one full packet?
2822 * Logic below assumes the mergeable buffer header is used.
2823 */
2824static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2825{
2826 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2827 unsigned int rq_size = virtqueue_get_vring_size(vq);
2828 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2829 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2830 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2831
2832 return max(max(min_buf_len, hdr_len) - hdr_len,
2833 (unsigned int)GOOD_PACKET_LEN);
2834}
2835
2836static int virtnet_find_vqs(struct virtnet_info *vi)
2837{
2838 vq_callback_t **callbacks;
2839 struct virtqueue **vqs;
2840 int ret = -ENOMEM;
2841 int i, total_vqs;
2842 const char **names;
2843 bool *ctx;
2844
2845 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2846 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2847 * possible control vq.
2848 */
2849 total_vqs = vi->max_queue_pairs * 2 +
2850 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2851
2852 /* Allocate space for find_vqs parameters */
2853 vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2854 if (!vqs)
2855 goto err_vq;
2856 callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2857 if (!callbacks)
2858 goto err_callback;
2859 names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2860 if (!names)
2861 goto err_names;
2862 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2863 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2864 if (!ctx)
2865 goto err_ctx;
2866 } else {
2867 ctx = NULL;
2868 }
2869
2870 /* Parameters for control virtqueue, if any */
2871 if (vi->has_cvq) {
2872 callbacks[total_vqs - 1] = NULL;
2873 names[total_vqs - 1] = "control";
2874 }
2875
2876 /* Allocate/initialize parameters for send/receive virtqueues */
2877 for (i = 0; i < vi->max_queue_pairs; i++) {
2878 callbacks[rxq2vq(i)] = skb_recv_done;
2879 callbacks[txq2vq(i)] = skb_xmit_done;
2880 sprintf(vi->rq[i].name, "input.%d", i);
2881 sprintf(vi->sq[i].name, "output.%d", i);
2882 names[rxq2vq(i)] = vi->rq[i].name;
2883 names[txq2vq(i)] = vi->sq[i].name;
2884 if (ctx)
2885 ctx[rxq2vq(i)] = true;
2886 }
2887
2888 ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
2889 names, ctx, NULL);
2890 if (ret)
2891 goto err_find;
2892
2893 if (vi->has_cvq) {
2894 vi->cvq = vqs[total_vqs - 1];
2895 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2896 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2897 }
2898
2899 for (i = 0; i < vi->max_queue_pairs; i++) {
2900 vi->rq[i].vq = vqs[rxq2vq(i)];
2901 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2902 vi->sq[i].vq = vqs[txq2vq(i)];
2903 }
2904
2905 /* run here: ret == 0. */
2906
2907
2908err_find:
2909 kfree(ctx);
2910err_ctx:
2911 kfree(names);
2912err_names:
2913 kfree(callbacks);
2914err_callback:
2915 kfree(vqs);
2916err_vq:
2917 return ret;
2918}
2919
2920static int virtnet_alloc_queues(struct virtnet_info *vi)
2921{
2922 int i;
2923
2924 if (vi->has_cvq) {
2925 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2926 if (!vi->ctrl)
2927 goto err_ctrl;
2928 } else {
2929 vi->ctrl = NULL;
2930 }
2931 vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2932 if (!vi->sq)
2933 goto err_sq;
2934 vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2935 if (!vi->rq)
2936 goto err_rq;
2937
2938 INIT_DELAYED_WORK(&vi->refill, refill_work);
2939 for (i = 0; i < vi->max_queue_pairs; i++) {
2940 vi->rq[i].pages = NULL;
2941 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2942 napi_weight);
2943 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2944 napi_tx ? napi_weight : 0);
2945
2946 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2947 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2948 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2949
2950 u64_stats_init(&vi->rq[i].stats.syncp);
2951 u64_stats_init(&vi->sq[i].stats.syncp);
2952 }
2953
2954 return 0;
2955
2956err_rq:
2957 kfree(vi->sq);
2958err_sq:
2959 kfree(vi->ctrl);
2960err_ctrl:
2961 return -ENOMEM;
2962}
2963
2964static int init_vqs(struct virtnet_info *vi)
2965{
2966 int ret;
2967
2968 /* Allocate send & receive queues */
2969 ret = virtnet_alloc_queues(vi);
2970 if (ret)
2971 goto err;
2972
2973 ret = virtnet_find_vqs(vi);
2974 if (ret)
2975 goto err_free;
2976
2977 get_online_cpus();
2978 virtnet_set_affinity(vi);
2979 put_online_cpus();
2980
2981 return 0;
2982
2983err_free:
2984 virtnet_free_queues(vi);
2985err:
2986 return ret;
2987}
2988
2989#ifdef CONFIG_SYSFS
2990static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2991 char *buf)
2992{
2993 struct virtnet_info *vi = netdev_priv(queue->dev);
2994 unsigned int queue_index = get_netdev_rx_queue_index(queue);
2995 unsigned int headroom = virtnet_get_headroom(vi);
2996 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2997 struct ewma_pkt_len *avg;
2998
2999 BUG_ON(queue_index >= vi->max_queue_pairs);
3000 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
3001 return sprintf(buf, "%u\n",
3002 get_mergeable_buf_len(&vi->rq[queue_index], avg,
3003 SKB_DATA_ALIGN(headroom + tailroom)));
3004}
3005
3006static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
3007 __ATTR_RO(mergeable_rx_buffer_size);
3008
3009static struct attribute *virtio_net_mrg_rx_attrs[] = {
3010 &mergeable_rx_buffer_size_attribute.attr,
3011 NULL
3012};
3013
3014static const struct attribute_group virtio_net_mrg_rx_group = {
3015 .name = "virtio_net",
3016 .attrs = virtio_net_mrg_rx_attrs
3017};
3018#endif
3019
3020static bool virtnet_fail_on_feature(struct virtio_device *vdev,
3021 unsigned int fbit,
3022 const char *fname, const char *dname)
3023{
3024 if (!virtio_has_feature(vdev, fbit))
3025 return false;
3026
3027 dev_err(&vdev->dev, "device advertises feature %s but not %s",
3028 fname, dname);
3029
3030 return true;
3031}
3032
3033#define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
3034 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
3035
3036static bool virtnet_validate_features(struct virtio_device *vdev)
3037{
3038 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
3039 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
3040 "VIRTIO_NET_F_CTRL_VQ") ||
3041 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
3042 "VIRTIO_NET_F_CTRL_VQ") ||
3043 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
3044 "VIRTIO_NET_F_CTRL_VQ") ||
3045 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
3046 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
3047 "VIRTIO_NET_F_CTRL_VQ"))) {
3048 return false;
3049 }
3050
3051 return true;
3052}
3053
3054#define MIN_MTU ETH_MIN_MTU
3055#define MAX_MTU ETH_MAX_MTU
3056
3057static int virtnet_validate(struct virtio_device *vdev)
3058{
3059 if (!vdev->config->get) {
3060 dev_err(&vdev->dev, "%s failure: config access disabled\n",
3061 __func__);
3062 return -EINVAL;
3063 }
3064
3065 if (!virtnet_validate_features(vdev))
3066 return -EINVAL;
3067
3068 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3069 int mtu = virtio_cread16(vdev,
3070 offsetof(struct virtio_net_config,
3071 mtu));
3072 if (mtu < MIN_MTU)
3073 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
3074 }
3075
3076 return 0;
3077}
3078
3079static int virtnet_probe(struct virtio_device *vdev)
3080{
3081 int i, err = -ENOMEM;
3082 struct net_device *dev;
3083 struct virtnet_info *vi;
3084 u16 max_queue_pairs;
3085 int mtu;
3086
3087 /* Find if host supports multiqueue virtio_net device */
3088 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
3089 struct virtio_net_config,
3090 max_virtqueue_pairs, &max_queue_pairs);
3091
3092 /* We need at least 2 queue's */
3093 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3094 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3095 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3096 max_queue_pairs = 1;
3097
3098 /* Allocate ourselves a network device with room for our info */
3099 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3100 if (!dev)
3101 return -ENOMEM;
3102
3103 /* Set up network device as normal. */
3104 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
3105 IFF_TX_SKB_NO_LINEAR;
3106 dev->netdev_ops = &virtnet_netdev;
3107 dev->features = NETIF_F_HIGHDMA;
3108
3109 dev->ethtool_ops = &virtnet_ethtool_ops;
3110 SET_NETDEV_DEV(dev, &vdev->dev);
3111
3112 /* Do we support "hardware" checksums? */
3113 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3114 /* This opens up the world of extra features. */
3115 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3116 if (csum)
3117 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3118
3119 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3120 dev->hw_features |= NETIF_F_TSO
3121 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3122 }
3123 /* Individual feature bits: what can host handle? */
3124 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3125 dev->hw_features |= NETIF_F_TSO;
3126 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3127 dev->hw_features |= NETIF_F_TSO6;
3128 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3129 dev->hw_features |= NETIF_F_TSO_ECN;
3130
3131 dev->features |= NETIF_F_GSO_ROBUST;
3132
3133 if (gso)
3134 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3135 /* (!csum && gso) case will be fixed by register_netdev() */
3136 }
3137 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3138 dev->features |= NETIF_F_RXCSUM;
3139 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3140 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3141 dev->features |= NETIF_F_GRO_HW;
3142 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3143 dev->hw_features |= NETIF_F_GRO_HW;
3144
3145 dev->vlan_features = dev->features;
3146
3147 /* MTU range: 68 - 65535 */
3148 dev->min_mtu = MIN_MTU;
3149 dev->max_mtu = MAX_MTU;
3150
3151 /* Configuration may specify what MAC to use. Otherwise random. */
3152 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
3153 virtio_cread_bytes(vdev,
3154 offsetof(struct virtio_net_config, mac),
3155 dev->dev_addr, dev->addr_len);
3156 else
3157 eth_hw_addr_random(dev);
3158
3159 /* Set up our device-specific information */
3160 vi = netdev_priv(dev);
3161 vi->dev = dev;
3162 vi->vdev = vdev;
3163 vdev->priv = vi;
3164
3165 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3166
3167 /* If we can receive ANY GSO packets, we must allocate large ones. */
3168 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3169 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3170 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3171 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3172 vi->big_packets = true;
3173
3174 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3175 vi->mergeable_rx_bufs = true;
3176
3177 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3178 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3179 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3180 else
3181 vi->hdr_len = sizeof(struct virtio_net_hdr);
3182
3183 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3184 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3185 vi->any_header_sg = true;
3186
3187 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3188 vi->has_cvq = true;
3189
3190 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3191 mtu = virtio_cread16(vdev,
3192 offsetof(struct virtio_net_config,
3193 mtu));
3194 if (mtu < dev->min_mtu) {
3195 /* Should never trigger: MTU was previously validated
3196 * in virtnet_validate.
3197 */
3198 dev_err(&vdev->dev,
3199 "device MTU appears to have changed it is now %d < %d",
3200 mtu, dev->min_mtu);
3201 err = -EINVAL;
3202 goto free;
3203 }
3204
3205 dev->mtu = mtu;
3206 dev->max_mtu = mtu;
3207
3208 /* TODO: size buffers correctly in this case. */
3209 if (dev->mtu > ETH_DATA_LEN)
3210 vi->big_packets = true;
3211 }
3212
3213 if (vi->any_header_sg)
3214 dev->needed_headroom = vi->hdr_len;
3215
3216 /* Enable multiqueue by default */
3217 if (num_online_cpus() >= max_queue_pairs)
3218 vi->curr_queue_pairs = max_queue_pairs;
3219 else
3220 vi->curr_queue_pairs = num_online_cpus();
3221 vi->max_queue_pairs = max_queue_pairs;
3222
3223 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3224 err = init_vqs(vi);
3225 if (err)
3226 goto free;
3227
3228#ifdef CONFIG_SYSFS
3229 if (vi->mergeable_rx_bufs)
3230 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3231#endif
3232 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3233 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3234
3235 virtnet_init_settings(dev);
3236
3237 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3238 vi->failover = net_failover_create(vi->dev);
3239 if (IS_ERR(vi->failover)) {
3240 err = PTR_ERR(vi->failover);
3241 goto free_vqs;
3242 }
3243 }
3244
3245 err = register_netdev(dev);
3246 if (err) {
3247 pr_debug("virtio_net: registering device failed\n");
3248 goto free_failover;
3249 }
3250
3251 virtio_device_ready(vdev);
3252
3253 err = virtnet_cpu_notif_add(vi);
3254 if (err) {
3255 pr_debug("virtio_net: registering cpu notifier failed\n");
3256 goto free_unregister_netdev;
3257 }
3258
3259 virtnet_set_queues(vi, vi->curr_queue_pairs);
3260
3261 /* Assume link up if device can't report link status,
3262 otherwise get link status from config. */
3263 netif_carrier_off(dev);
3264 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3265 schedule_work(&vi->config_work);
3266 } else {
3267 vi->status = VIRTIO_NET_S_LINK_UP;
3268 virtnet_update_settings(vi);
3269 netif_carrier_on(dev);
3270 }
3271
3272 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3273 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3274 set_bit(guest_offloads[i], &vi->guest_offloads);
3275 vi->guest_offloads_capable = vi->guest_offloads;
3276
3277 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3278 dev->name, max_queue_pairs);
3279
3280 return 0;
3281
3282free_unregister_netdev:
3283 vi->vdev->config->reset(vdev);
3284
3285 unregister_netdev(dev);
3286free_failover:
3287 net_failover_destroy(vi->failover);
3288free_vqs:
3289 cancel_delayed_work_sync(&vi->refill);
3290 free_receive_page_frags(vi);
3291 virtnet_del_vqs(vi);
3292free:
3293 free_netdev(dev);
3294 return err;
3295}
3296
3297static void remove_vq_common(struct virtnet_info *vi)
3298{
3299 vi->vdev->config->reset(vi->vdev);
3300
3301 /* Free unused buffers in both send and recv, if any. */
3302 free_unused_bufs(vi);
3303
3304 free_receive_bufs(vi);
3305
3306 free_receive_page_frags(vi);
3307
3308 virtnet_del_vqs(vi);
3309}
3310
3311static void virtnet_remove(struct virtio_device *vdev)
3312{
3313 struct virtnet_info *vi = vdev->priv;
3314
3315 virtnet_cpu_notif_remove(vi);
3316
3317 /* Make sure no work handler is accessing the device. */
3318 flush_work(&vi->config_work);
3319
3320 unregister_netdev(vi->dev);
3321
3322 net_failover_destroy(vi->failover);
3323
3324 remove_vq_common(vi);
3325
3326 free_netdev(vi->dev);
3327}
3328
3329static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3330{
3331 struct virtnet_info *vi = vdev->priv;
3332
3333 virtnet_cpu_notif_remove(vi);
3334 virtnet_freeze_down(vdev);
3335 remove_vq_common(vi);
3336
3337 return 0;
3338}
3339
3340static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3341{
3342 struct virtnet_info *vi = vdev->priv;
3343 int err;
3344
3345 err = virtnet_restore_up(vdev);
3346 if (err)
3347 return err;
3348 virtnet_set_queues(vi, vi->curr_queue_pairs);
3349
3350 err = virtnet_cpu_notif_add(vi);
3351 if (err) {
3352 virtnet_freeze_down(vdev);
3353 remove_vq_common(vi);
3354 return err;
3355 }
3356
3357 return 0;
3358}
3359
3360static struct virtio_device_id id_table[] = {
3361 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3362 { 0 },
3363};
3364
3365#define VIRTNET_FEATURES \
3366 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3367 VIRTIO_NET_F_MAC, \
3368 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3369 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3370 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3371 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3372 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3373 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3374 VIRTIO_NET_F_CTRL_MAC_ADDR, \
3375 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3376 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3377
3378static unsigned int features[] = {
3379 VIRTNET_FEATURES,
3380};
3381
3382static unsigned int features_legacy[] = {
3383 VIRTNET_FEATURES,
3384 VIRTIO_NET_F_GSO,
3385 VIRTIO_F_ANY_LAYOUT,
3386};
3387
3388static struct virtio_driver virtio_net_driver = {
3389 .feature_table = features,
3390 .feature_table_size = ARRAY_SIZE(features),
3391 .feature_table_legacy = features_legacy,
3392 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3393 .driver.name = KBUILD_MODNAME,
3394 .driver.owner = THIS_MODULE,
3395 .id_table = id_table,
3396 .validate = virtnet_validate,
3397 .probe = virtnet_probe,
3398 .remove = virtnet_remove,
3399 .config_changed = virtnet_config_changed,
3400#ifdef CONFIG_PM_SLEEP
3401 .freeze = virtnet_freeze,
3402 .restore = virtnet_restore,
3403#endif
3404};
3405
3406static __init int virtio_net_driver_init(void)
3407{
3408 int ret;
3409
3410 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3411 virtnet_cpu_online,
3412 virtnet_cpu_down_prep);
3413 if (ret < 0)
3414 goto out;
3415 virtionet_online = ret;
3416 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3417 NULL, virtnet_cpu_dead);
3418 if (ret)
3419 goto err_dead;
3420
3421 ret = register_virtio_driver(&virtio_net_driver);
3422 if (ret)
3423 goto err_virtio;
3424 return 0;
3425err_virtio:
3426 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3427err_dead:
3428 cpuhp_remove_multi_state(virtionet_online);
3429out:
3430 return ret;
3431}
3432module_init(virtio_net_driver_init);
3433
3434static __exit void virtio_net_driver_exit(void)
3435{
3436 unregister_virtio_driver(&virtio_net_driver);
3437 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3438 cpuhp_remove_multi_state(virtionet_online);
3439}
3440module_exit(virtio_net_driver_exit);
3441
3442MODULE_DEVICE_TABLE(virtio, id_table);
3443MODULE_DESCRIPTION("Virtio network driver");
3444MODULE_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 <net/route.h>
23#include <net/xdp.h>
24#include <net/net_failover.h>
25
26static int napi_weight = NAPI_POLL_WEIGHT;
27module_param(napi_weight, int, 0444);
28
29static bool csum = true, gso = true, napi_tx = true;
30module_param(csum, bool, 0444);
31module_param(gso, bool, 0444);
32module_param(napi_tx, bool, 0644);
33
34/* FIXME: MTU in config. */
35#define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
36#define GOOD_COPY_LEN 128
37
38#define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
39
40/* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
41#define VIRTIO_XDP_HEADROOM 256
42
43/* Separating two types of XDP xmit */
44#define VIRTIO_XDP_TX BIT(0)
45#define VIRTIO_XDP_REDIR BIT(1)
46
47#define VIRTIO_XDP_FLAG BIT(0)
48
49/* RX packet size EWMA. The average packet size is used to determine the packet
50 * buffer size when refilling RX rings. As the entire RX ring may be refilled
51 * at once, the weight is chosen so that the EWMA will be insensitive to short-
52 * term, transient changes in packet size.
53 */
54DECLARE_EWMA(pkt_len, 0, 64)
55
56#define VIRTNET_DRIVER_VERSION "1.0.0"
57
58static const unsigned long guest_offloads[] = {
59 VIRTIO_NET_F_GUEST_TSO4,
60 VIRTIO_NET_F_GUEST_TSO6,
61 VIRTIO_NET_F_GUEST_ECN,
62 VIRTIO_NET_F_GUEST_UFO,
63 VIRTIO_NET_F_GUEST_CSUM
64};
65
66struct virtnet_stat_desc {
67 char desc[ETH_GSTRING_LEN];
68 size_t offset;
69};
70
71struct virtnet_sq_stats {
72 struct u64_stats_sync syncp;
73 u64 packets;
74 u64 bytes;
75 u64 xdp_tx;
76 u64 xdp_tx_drops;
77 u64 kicks;
78};
79
80struct virtnet_rq_stats {
81 struct u64_stats_sync syncp;
82 u64 packets;
83 u64 bytes;
84 u64 drops;
85 u64 xdp_packets;
86 u64 xdp_tx;
87 u64 xdp_redirects;
88 u64 xdp_drops;
89 u64 kicks;
90};
91
92#define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m)
93#define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m)
94
95static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
96 { "packets", VIRTNET_SQ_STAT(packets) },
97 { "bytes", VIRTNET_SQ_STAT(bytes) },
98 { "xdp_tx", VIRTNET_SQ_STAT(xdp_tx) },
99 { "xdp_tx_drops", VIRTNET_SQ_STAT(xdp_tx_drops) },
100 { "kicks", VIRTNET_SQ_STAT(kicks) },
101};
102
103static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
104 { "packets", VIRTNET_RQ_STAT(packets) },
105 { "bytes", VIRTNET_RQ_STAT(bytes) },
106 { "drops", VIRTNET_RQ_STAT(drops) },
107 { "xdp_packets", VIRTNET_RQ_STAT(xdp_packets) },
108 { "xdp_tx", VIRTNET_RQ_STAT(xdp_tx) },
109 { "xdp_redirects", VIRTNET_RQ_STAT(xdp_redirects) },
110 { "xdp_drops", VIRTNET_RQ_STAT(xdp_drops) },
111 { "kicks", VIRTNET_RQ_STAT(kicks) },
112};
113
114#define VIRTNET_SQ_STATS_LEN ARRAY_SIZE(virtnet_sq_stats_desc)
115#define VIRTNET_RQ_STATS_LEN ARRAY_SIZE(virtnet_rq_stats_desc)
116
117/* Internal representation of a send virtqueue */
118struct send_queue {
119 /* Virtqueue associated with this send _queue */
120 struct virtqueue *vq;
121
122 /* TX: fragments + linear part + virtio header */
123 struct scatterlist sg[MAX_SKB_FRAGS + 2];
124
125 /* Name of the send queue: output.$index */
126 char name[40];
127
128 struct virtnet_sq_stats stats;
129
130 struct napi_struct napi;
131};
132
133/* Internal representation of a receive virtqueue */
134struct receive_queue {
135 /* Virtqueue associated with this receive_queue */
136 struct virtqueue *vq;
137
138 struct napi_struct napi;
139
140 struct bpf_prog __rcu *xdp_prog;
141
142 struct virtnet_rq_stats stats;
143
144 /* Chain pages by the private ptr. */
145 struct page *pages;
146
147 /* Average packet length for mergeable receive buffers. */
148 struct ewma_pkt_len mrg_avg_pkt_len;
149
150 /* Page frag for packet buffer allocation. */
151 struct page_frag alloc_frag;
152
153 /* RX: fragments + linear part + virtio header */
154 struct scatterlist sg[MAX_SKB_FRAGS + 2];
155
156 /* Min single buffer size for mergeable buffers case. */
157 unsigned int min_buf_len;
158
159 /* Name of this receive queue: input.$index */
160 char name[40];
161
162 struct xdp_rxq_info xdp_rxq;
163};
164
165/* Control VQ buffers: protected by the rtnl lock */
166struct control_buf {
167 struct virtio_net_ctrl_hdr hdr;
168 virtio_net_ctrl_ack status;
169 struct virtio_net_ctrl_mq mq;
170 u8 promisc;
171 u8 allmulti;
172 __virtio16 vid;
173 __virtio64 offloads;
174};
175
176struct virtnet_info {
177 struct virtio_device *vdev;
178 struct virtqueue *cvq;
179 struct net_device *dev;
180 struct send_queue *sq;
181 struct receive_queue *rq;
182 unsigned int status;
183
184 /* Max # of queue pairs supported by the device */
185 u16 max_queue_pairs;
186
187 /* # of queue pairs currently used by the driver */
188 u16 curr_queue_pairs;
189
190 /* # of XDP queue pairs currently used by the driver */
191 u16 xdp_queue_pairs;
192
193 /* I like... big packets and I cannot lie! */
194 bool big_packets;
195
196 /* Host will merge rx buffers for big packets (shake it! shake it!) */
197 bool mergeable_rx_bufs;
198
199 /* Has control virtqueue */
200 bool has_cvq;
201
202 /* Host can handle any s/g split between our header and packet data */
203 bool any_header_sg;
204
205 /* Packet virtio header size */
206 u8 hdr_len;
207
208 /* Work struct for refilling if we run low on memory. */
209 struct delayed_work refill;
210
211 /* Work struct for config space updates */
212 struct work_struct config_work;
213
214 /* Does the affinity hint is set for virtqueues? */
215 bool affinity_hint_set;
216
217 /* CPU hotplug instances for online & dead */
218 struct hlist_node node;
219 struct hlist_node node_dead;
220
221 struct control_buf *ctrl;
222
223 /* Ethtool settings */
224 u8 duplex;
225 u32 speed;
226
227 unsigned long guest_offloads;
228 unsigned long guest_offloads_capable;
229
230 /* failover when STANDBY feature enabled */
231 struct failover *failover;
232};
233
234struct padded_vnet_hdr {
235 struct virtio_net_hdr_mrg_rxbuf hdr;
236 /*
237 * hdr is in a separate sg buffer, and data sg buffer shares same page
238 * with this header sg. This padding makes next sg 16 byte aligned
239 * after the header.
240 */
241 char padding[4];
242};
243
244static bool is_xdp_frame(void *ptr)
245{
246 return (unsigned long)ptr & VIRTIO_XDP_FLAG;
247}
248
249static void *xdp_to_ptr(struct xdp_frame *ptr)
250{
251 return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
252}
253
254static struct xdp_frame *ptr_to_xdp(void *ptr)
255{
256 return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
257}
258
259/* Converting between virtqueue no. and kernel tx/rx queue no.
260 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
261 */
262static int vq2txq(struct virtqueue *vq)
263{
264 return (vq->index - 1) / 2;
265}
266
267static int txq2vq(int txq)
268{
269 return txq * 2 + 1;
270}
271
272static int vq2rxq(struct virtqueue *vq)
273{
274 return vq->index / 2;
275}
276
277static int rxq2vq(int rxq)
278{
279 return rxq * 2;
280}
281
282static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
283{
284 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
285}
286
287/*
288 * private is used to chain pages for big packets, put the whole
289 * most recent used list in the beginning for reuse
290 */
291static void give_pages(struct receive_queue *rq, struct page *page)
292{
293 struct page *end;
294
295 /* Find end of list, sew whole thing into vi->rq.pages. */
296 for (end = page; end->private; end = (struct page *)end->private);
297 end->private = (unsigned long)rq->pages;
298 rq->pages = page;
299}
300
301static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
302{
303 struct page *p = rq->pages;
304
305 if (p) {
306 rq->pages = (struct page *)p->private;
307 /* clear private here, it is used to chain pages */
308 p->private = 0;
309 } else
310 p = alloc_page(gfp_mask);
311 return p;
312}
313
314static void virtqueue_napi_schedule(struct napi_struct *napi,
315 struct virtqueue *vq)
316{
317 if (napi_schedule_prep(napi)) {
318 virtqueue_disable_cb(vq);
319 __napi_schedule(napi);
320 }
321}
322
323static void virtqueue_napi_complete(struct napi_struct *napi,
324 struct virtqueue *vq, int processed)
325{
326 int opaque;
327
328 opaque = virtqueue_enable_cb_prepare(vq);
329 if (napi_complete_done(napi, processed)) {
330 if (unlikely(virtqueue_poll(vq, opaque)))
331 virtqueue_napi_schedule(napi, vq);
332 } else {
333 virtqueue_disable_cb(vq);
334 }
335}
336
337static void skb_xmit_done(struct virtqueue *vq)
338{
339 struct virtnet_info *vi = vq->vdev->priv;
340 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
341
342 /* Suppress further interrupts. */
343 virtqueue_disable_cb(vq);
344
345 if (napi->weight)
346 virtqueue_napi_schedule(napi, vq);
347 else
348 /* We were probably waiting for more output buffers. */
349 netif_wake_subqueue(vi->dev, vq2txq(vq));
350}
351
352#define MRG_CTX_HEADER_SHIFT 22
353static void *mergeable_len_to_ctx(unsigned int truesize,
354 unsigned int headroom)
355{
356 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
357}
358
359static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
360{
361 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
362}
363
364static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
365{
366 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
367}
368
369/* Called from bottom half context */
370static struct sk_buff *page_to_skb(struct virtnet_info *vi,
371 struct receive_queue *rq,
372 struct page *page, unsigned int offset,
373 unsigned int len, unsigned int truesize,
374 bool hdr_valid)
375{
376 struct sk_buff *skb;
377 struct virtio_net_hdr_mrg_rxbuf *hdr;
378 unsigned int copy, hdr_len, hdr_padded_len;
379 char *p;
380
381 p = page_address(page) + offset;
382
383 /* copy small packet so we can reuse these pages for small data */
384 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
385 if (unlikely(!skb))
386 return NULL;
387
388 hdr = skb_vnet_hdr(skb);
389
390 hdr_len = vi->hdr_len;
391 if (vi->mergeable_rx_bufs)
392 hdr_padded_len = sizeof(*hdr);
393 else
394 hdr_padded_len = sizeof(struct padded_vnet_hdr);
395
396 if (hdr_valid)
397 memcpy(hdr, p, hdr_len);
398
399 len -= hdr_len;
400 offset += hdr_padded_len;
401 p += hdr_padded_len;
402
403 copy = len;
404 if (copy > skb_tailroom(skb))
405 copy = skb_tailroom(skb);
406 skb_put_data(skb, p, copy);
407
408 len -= copy;
409 offset += copy;
410
411 if (vi->mergeable_rx_bufs) {
412 if (len)
413 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
414 else
415 put_page(page);
416 return skb;
417 }
418
419 /*
420 * Verify that we can indeed put this data into a skb.
421 * This is here to handle cases when the device erroneously
422 * tries to receive more than is possible. This is usually
423 * the case of a broken device.
424 */
425 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
426 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
427 dev_kfree_skb(skb);
428 return NULL;
429 }
430 BUG_ON(offset >= PAGE_SIZE);
431 while (len) {
432 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
433 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
434 frag_size, truesize);
435 len -= frag_size;
436 page = (struct page *)page->private;
437 offset = 0;
438 }
439
440 if (page)
441 give_pages(rq, page);
442
443 return skb;
444}
445
446static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
447 struct send_queue *sq,
448 struct xdp_frame *xdpf)
449{
450 struct virtio_net_hdr_mrg_rxbuf *hdr;
451 int err;
452
453 /* virtqueue want to use data area in-front of packet */
454 if (unlikely(xdpf->metasize > 0))
455 return -EOPNOTSUPP;
456
457 if (unlikely(xdpf->headroom < vi->hdr_len))
458 return -EOVERFLOW;
459
460 /* Make room for virtqueue hdr (also change xdpf->headroom?) */
461 xdpf->data -= vi->hdr_len;
462 /* Zero header and leave csum up to XDP layers */
463 hdr = xdpf->data;
464 memset(hdr, 0, vi->hdr_len);
465 xdpf->len += vi->hdr_len;
466
467 sg_init_one(sq->sg, xdpf->data, xdpf->len);
468
469 err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
470 GFP_ATOMIC);
471 if (unlikely(err))
472 return -ENOSPC; /* Caller handle free/refcnt */
473
474 return 0;
475}
476
477static struct send_queue *virtnet_xdp_sq(struct virtnet_info *vi)
478{
479 unsigned int qp;
480
481 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
482 return &vi->sq[qp];
483}
484
485static int virtnet_xdp_xmit(struct net_device *dev,
486 int n, struct xdp_frame **frames, u32 flags)
487{
488 struct virtnet_info *vi = netdev_priv(dev);
489 struct receive_queue *rq = vi->rq;
490 struct bpf_prog *xdp_prog;
491 struct send_queue *sq;
492 unsigned int len;
493 int packets = 0;
494 int bytes = 0;
495 int drops = 0;
496 int kicks = 0;
497 int ret, err;
498 void *ptr;
499 int i;
500
501 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
502 * indicate XDP resources have been successfully allocated.
503 */
504 xdp_prog = rcu_dereference(rq->xdp_prog);
505 if (!xdp_prog)
506 return -ENXIO;
507
508 sq = virtnet_xdp_sq(vi);
509
510 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
511 ret = -EINVAL;
512 drops = n;
513 goto out;
514 }
515
516 /* Free up any pending old buffers before queueing new ones. */
517 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
518 if (likely(is_xdp_frame(ptr))) {
519 struct xdp_frame *frame = ptr_to_xdp(ptr);
520
521 bytes += frame->len;
522 xdp_return_frame(frame);
523 } else {
524 struct sk_buff *skb = ptr;
525
526 bytes += skb->len;
527 napi_consume_skb(skb, false);
528 }
529 packets++;
530 }
531
532 for (i = 0; i < n; i++) {
533 struct xdp_frame *xdpf = frames[i];
534
535 err = __virtnet_xdp_xmit_one(vi, sq, xdpf);
536 if (err) {
537 xdp_return_frame_rx_napi(xdpf);
538 drops++;
539 }
540 }
541 ret = n - drops;
542
543 if (flags & XDP_XMIT_FLUSH) {
544 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
545 kicks = 1;
546 }
547out:
548 u64_stats_update_begin(&sq->stats.syncp);
549 sq->stats.bytes += bytes;
550 sq->stats.packets += packets;
551 sq->stats.xdp_tx += n;
552 sq->stats.xdp_tx_drops += drops;
553 sq->stats.kicks += kicks;
554 u64_stats_update_end(&sq->stats.syncp);
555
556 return ret;
557}
558
559static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
560{
561 return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
562}
563
564/* We copy the packet for XDP in the following cases:
565 *
566 * 1) Packet is scattered across multiple rx buffers.
567 * 2) Headroom space is insufficient.
568 *
569 * This is inefficient but it's a temporary condition that
570 * we hit right after XDP is enabled and until queue is refilled
571 * with large buffers with sufficient headroom - so it should affect
572 * at most queue size packets.
573 * Afterwards, the conditions to enable
574 * XDP should preclude the underlying device from sending packets
575 * across multiple buffers (num_buf > 1), and we make sure buffers
576 * have enough headroom.
577 */
578static struct page *xdp_linearize_page(struct receive_queue *rq,
579 u16 *num_buf,
580 struct page *p,
581 int offset,
582 int page_off,
583 unsigned int *len)
584{
585 struct page *page = alloc_page(GFP_ATOMIC);
586
587 if (!page)
588 return NULL;
589
590 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
591 page_off += *len;
592
593 while (--*num_buf) {
594 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
595 unsigned int buflen;
596 void *buf;
597 int off;
598
599 buf = virtqueue_get_buf(rq->vq, &buflen);
600 if (unlikely(!buf))
601 goto err_buf;
602
603 p = virt_to_head_page(buf);
604 off = buf - page_address(p);
605
606 /* guard against a misconfigured or uncooperative backend that
607 * is sending packet larger than the MTU.
608 */
609 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
610 put_page(p);
611 goto err_buf;
612 }
613
614 memcpy(page_address(page) + page_off,
615 page_address(p) + off, buflen);
616 page_off += buflen;
617 put_page(p);
618 }
619
620 /* Headroom does not contribute to packet length */
621 *len = page_off - VIRTIO_XDP_HEADROOM;
622 return page;
623err_buf:
624 __free_pages(page, 0);
625 return NULL;
626}
627
628static struct sk_buff *receive_small(struct net_device *dev,
629 struct virtnet_info *vi,
630 struct receive_queue *rq,
631 void *buf, void *ctx,
632 unsigned int len,
633 unsigned int *xdp_xmit,
634 struct virtnet_rq_stats *stats)
635{
636 struct sk_buff *skb;
637 struct bpf_prog *xdp_prog;
638 unsigned int xdp_headroom = (unsigned long)ctx;
639 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
640 unsigned int headroom = vi->hdr_len + header_offset;
641 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
642 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
643 struct page *page = virt_to_head_page(buf);
644 unsigned int delta = 0;
645 struct page *xdp_page;
646 int err;
647
648 len -= vi->hdr_len;
649 stats->bytes += len;
650
651 rcu_read_lock();
652 xdp_prog = rcu_dereference(rq->xdp_prog);
653 if (xdp_prog) {
654 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
655 struct xdp_frame *xdpf;
656 struct xdp_buff xdp;
657 void *orig_data;
658 u32 act;
659
660 if (unlikely(hdr->hdr.gso_type))
661 goto err_xdp;
662
663 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
664 int offset = buf - page_address(page) + header_offset;
665 unsigned int tlen = len + vi->hdr_len;
666 u16 num_buf = 1;
667
668 xdp_headroom = virtnet_get_headroom(vi);
669 header_offset = VIRTNET_RX_PAD + xdp_headroom;
670 headroom = vi->hdr_len + header_offset;
671 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
672 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
673 xdp_page = xdp_linearize_page(rq, &num_buf, page,
674 offset, header_offset,
675 &tlen);
676 if (!xdp_page)
677 goto err_xdp;
678
679 buf = page_address(xdp_page);
680 put_page(page);
681 page = xdp_page;
682 }
683
684 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
685 xdp.data = xdp.data_hard_start + xdp_headroom;
686 xdp_set_data_meta_invalid(&xdp);
687 xdp.data_end = xdp.data + len;
688 xdp.rxq = &rq->xdp_rxq;
689 orig_data = xdp.data;
690 act = bpf_prog_run_xdp(xdp_prog, &xdp);
691 stats->xdp_packets++;
692
693 switch (act) {
694 case XDP_PASS:
695 /* Recalculate length in case bpf program changed it */
696 delta = orig_data - xdp.data;
697 len = xdp.data_end - xdp.data;
698 break;
699 case XDP_TX:
700 stats->xdp_tx++;
701 xdpf = convert_to_xdp_frame(&xdp);
702 if (unlikely(!xdpf))
703 goto err_xdp;
704 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
705 if (unlikely(err < 0)) {
706 trace_xdp_exception(vi->dev, xdp_prog, act);
707 goto err_xdp;
708 }
709 *xdp_xmit |= VIRTIO_XDP_TX;
710 rcu_read_unlock();
711 goto xdp_xmit;
712 case XDP_REDIRECT:
713 stats->xdp_redirects++;
714 err = xdp_do_redirect(dev, &xdp, xdp_prog);
715 if (err)
716 goto err_xdp;
717 *xdp_xmit |= VIRTIO_XDP_REDIR;
718 rcu_read_unlock();
719 goto xdp_xmit;
720 default:
721 bpf_warn_invalid_xdp_action(act);
722 /* fall through */
723 case XDP_ABORTED:
724 trace_xdp_exception(vi->dev, xdp_prog, act);
725 case XDP_DROP:
726 goto err_xdp;
727 }
728 }
729 rcu_read_unlock();
730
731 skb = build_skb(buf, buflen);
732 if (!skb) {
733 put_page(page);
734 goto err;
735 }
736 skb_reserve(skb, headroom - delta);
737 skb_put(skb, len);
738 if (!delta) {
739 buf += header_offset;
740 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
741 } /* keep zeroed vnet hdr since packet was changed by bpf */
742
743err:
744 return skb;
745
746err_xdp:
747 rcu_read_unlock();
748 stats->xdp_drops++;
749 stats->drops++;
750 put_page(page);
751xdp_xmit:
752 return NULL;
753}
754
755static struct sk_buff *receive_big(struct net_device *dev,
756 struct virtnet_info *vi,
757 struct receive_queue *rq,
758 void *buf,
759 unsigned int len,
760 struct virtnet_rq_stats *stats)
761{
762 struct page *page = buf;
763 struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
764 PAGE_SIZE, true);
765
766 stats->bytes += len - vi->hdr_len;
767 if (unlikely(!skb))
768 goto err;
769
770 return skb;
771
772err:
773 stats->drops++;
774 give_pages(rq, page);
775 return NULL;
776}
777
778static struct sk_buff *receive_mergeable(struct net_device *dev,
779 struct virtnet_info *vi,
780 struct receive_queue *rq,
781 void *buf,
782 void *ctx,
783 unsigned int len,
784 unsigned int *xdp_xmit,
785 struct virtnet_rq_stats *stats)
786{
787 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
788 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
789 struct page *page = virt_to_head_page(buf);
790 int offset = buf - page_address(page);
791 struct sk_buff *head_skb, *curr_skb;
792 struct bpf_prog *xdp_prog;
793 unsigned int truesize;
794 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
795 int err;
796
797 head_skb = NULL;
798 stats->bytes += len - vi->hdr_len;
799
800 rcu_read_lock();
801 xdp_prog = rcu_dereference(rq->xdp_prog);
802 if (xdp_prog) {
803 struct xdp_frame *xdpf;
804 struct page *xdp_page;
805 struct xdp_buff xdp;
806 void *data;
807 u32 act;
808
809 /* Transient failure which in theory could occur if
810 * in-flight packets from before XDP was enabled reach
811 * the receive path after XDP is loaded.
812 */
813 if (unlikely(hdr->hdr.gso_type))
814 goto err_xdp;
815
816 /* This happens when rx buffer size is underestimated
817 * or headroom is not enough because of the buffer
818 * was refilled before XDP is set. This should only
819 * happen for the first several packets, so we don't
820 * care much about its performance.
821 */
822 if (unlikely(num_buf > 1 ||
823 headroom < virtnet_get_headroom(vi))) {
824 /* linearize data for XDP */
825 xdp_page = xdp_linearize_page(rq, &num_buf,
826 page, offset,
827 VIRTIO_XDP_HEADROOM,
828 &len);
829 if (!xdp_page)
830 goto err_xdp;
831 offset = VIRTIO_XDP_HEADROOM;
832 } else {
833 xdp_page = page;
834 }
835
836 /* Allow consuming headroom but reserve enough space to push
837 * the descriptor on if we get an XDP_TX return code.
838 */
839 data = page_address(xdp_page) + offset;
840 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
841 xdp.data = data + vi->hdr_len;
842 xdp_set_data_meta_invalid(&xdp);
843 xdp.data_end = xdp.data + (len - vi->hdr_len);
844 xdp.rxq = &rq->xdp_rxq;
845
846 act = bpf_prog_run_xdp(xdp_prog, &xdp);
847 stats->xdp_packets++;
848
849 switch (act) {
850 case XDP_PASS:
851 /* recalculate offset to account for any header
852 * adjustments. Note other cases do not build an
853 * skb and avoid using offset
854 */
855 offset = xdp.data -
856 page_address(xdp_page) - vi->hdr_len;
857
858 /* recalculate len if xdp.data or xdp.data_end were
859 * adjusted
860 */
861 len = xdp.data_end - xdp.data + vi->hdr_len;
862 /* We can only create skb based on xdp_page. */
863 if (unlikely(xdp_page != page)) {
864 rcu_read_unlock();
865 put_page(page);
866 head_skb = page_to_skb(vi, rq, xdp_page,
867 offset, len,
868 PAGE_SIZE, false);
869 return head_skb;
870 }
871 break;
872 case XDP_TX:
873 stats->xdp_tx++;
874 xdpf = convert_to_xdp_frame(&xdp);
875 if (unlikely(!xdpf))
876 goto err_xdp;
877 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
878 if (unlikely(err < 0)) {
879 trace_xdp_exception(vi->dev, xdp_prog, act);
880 if (unlikely(xdp_page != page))
881 put_page(xdp_page);
882 goto err_xdp;
883 }
884 *xdp_xmit |= VIRTIO_XDP_TX;
885 if (unlikely(xdp_page != page))
886 put_page(page);
887 rcu_read_unlock();
888 goto xdp_xmit;
889 case XDP_REDIRECT:
890 stats->xdp_redirects++;
891 err = xdp_do_redirect(dev, &xdp, xdp_prog);
892 if (err) {
893 if (unlikely(xdp_page != page))
894 put_page(xdp_page);
895 goto err_xdp;
896 }
897 *xdp_xmit |= VIRTIO_XDP_REDIR;
898 if (unlikely(xdp_page != page))
899 put_page(page);
900 rcu_read_unlock();
901 goto xdp_xmit;
902 default:
903 bpf_warn_invalid_xdp_action(act);
904 /* fall through */
905 case XDP_ABORTED:
906 trace_xdp_exception(vi->dev, xdp_prog, act);
907 /* fall through */
908 case XDP_DROP:
909 if (unlikely(xdp_page != page))
910 __free_pages(xdp_page, 0);
911 goto err_xdp;
912 }
913 }
914 rcu_read_unlock();
915
916 truesize = mergeable_ctx_to_truesize(ctx);
917 if (unlikely(len > truesize)) {
918 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
919 dev->name, len, (unsigned long)ctx);
920 dev->stats.rx_length_errors++;
921 goto err_skb;
922 }
923
924 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog);
925 curr_skb = head_skb;
926
927 if (unlikely(!curr_skb))
928 goto err_skb;
929 while (--num_buf) {
930 int num_skb_frags;
931
932 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
933 if (unlikely(!buf)) {
934 pr_debug("%s: rx error: %d buffers out of %d missing\n",
935 dev->name, num_buf,
936 virtio16_to_cpu(vi->vdev,
937 hdr->num_buffers));
938 dev->stats.rx_length_errors++;
939 goto err_buf;
940 }
941
942 stats->bytes += len;
943 page = virt_to_head_page(buf);
944
945 truesize = mergeable_ctx_to_truesize(ctx);
946 if (unlikely(len > truesize)) {
947 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
948 dev->name, len, (unsigned long)ctx);
949 dev->stats.rx_length_errors++;
950 goto err_skb;
951 }
952
953 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
954 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
955 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
956
957 if (unlikely(!nskb))
958 goto err_skb;
959 if (curr_skb == head_skb)
960 skb_shinfo(curr_skb)->frag_list = nskb;
961 else
962 curr_skb->next = nskb;
963 curr_skb = nskb;
964 head_skb->truesize += nskb->truesize;
965 num_skb_frags = 0;
966 }
967 if (curr_skb != head_skb) {
968 head_skb->data_len += len;
969 head_skb->len += len;
970 head_skb->truesize += truesize;
971 }
972 offset = buf - page_address(page);
973 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
974 put_page(page);
975 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
976 len, truesize);
977 } else {
978 skb_add_rx_frag(curr_skb, num_skb_frags, page,
979 offset, len, truesize);
980 }
981 }
982
983 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
984 return head_skb;
985
986err_xdp:
987 rcu_read_unlock();
988 stats->xdp_drops++;
989err_skb:
990 put_page(page);
991 while (num_buf-- > 1) {
992 buf = virtqueue_get_buf(rq->vq, &len);
993 if (unlikely(!buf)) {
994 pr_debug("%s: rx error: %d buffers missing\n",
995 dev->name, num_buf);
996 dev->stats.rx_length_errors++;
997 break;
998 }
999 stats->bytes += len;
1000 page = virt_to_head_page(buf);
1001 put_page(page);
1002 }
1003err_buf:
1004 stats->drops++;
1005 dev_kfree_skb(head_skb);
1006xdp_xmit:
1007 return NULL;
1008}
1009
1010static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1011 void *buf, unsigned int len, void **ctx,
1012 unsigned int *xdp_xmit,
1013 struct virtnet_rq_stats *stats)
1014{
1015 struct net_device *dev = vi->dev;
1016 struct sk_buff *skb;
1017 struct virtio_net_hdr_mrg_rxbuf *hdr;
1018
1019 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1020 pr_debug("%s: short packet %i\n", dev->name, len);
1021 dev->stats.rx_length_errors++;
1022 if (vi->mergeable_rx_bufs) {
1023 put_page(virt_to_head_page(buf));
1024 } else if (vi->big_packets) {
1025 give_pages(rq, buf);
1026 } else {
1027 put_page(virt_to_head_page(buf));
1028 }
1029 return;
1030 }
1031
1032 if (vi->mergeable_rx_bufs)
1033 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1034 stats);
1035 else if (vi->big_packets)
1036 skb = receive_big(dev, vi, rq, buf, len, stats);
1037 else
1038 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1039
1040 if (unlikely(!skb))
1041 return;
1042
1043 hdr = skb_vnet_hdr(skb);
1044
1045 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1046 skb->ip_summed = CHECKSUM_UNNECESSARY;
1047
1048 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1049 virtio_is_little_endian(vi->vdev))) {
1050 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1051 dev->name, hdr->hdr.gso_type,
1052 hdr->hdr.gso_size);
1053 goto frame_err;
1054 }
1055
1056 skb_record_rx_queue(skb, vq2rxq(rq->vq));
1057 skb->protocol = eth_type_trans(skb, dev);
1058 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1059 ntohs(skb->protocol), skb->len, skb->pkt_type);
1060
1061 napi_gro_receive(&rq->napi, skb);
1062 return;
1063
1064frame_err:
1065 dev->stats.rx_frame_errors++;
1066 dev_kfree_skb(skb);
1067}
1068
1069/* Unlike mergeable buffers, all buffers are allocated to the
1070 * same size, except for the headroom. For this reason we do
1071 * not need to use mergeable_len_to_ctx here - it is enough
1072 * to store the headroom as the context ignoring the truesize.
1073 */
1074static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1075 gfp_t gfp)
1076{
1077 struct page_frag *alloc_frag = &rq->alloc_frag;
1078 char *buf;
1079 unsigned int xdp_headroom = virtnet_get_headroom(vi);
1080 void *ctx = (void *)(unsigned long)xdp_headroom;
1081 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1082 int err;
1083
1084 len = SKB_DATA_ALIGN(len) +
1085 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1086 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1087 return -ENOMEM;
1088
1089 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1090 get_page(alloc_frag->page);
1091 alloc_frag->offset += len;
1092 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1093 vi->hdr_len + GOOD_PACKET_LEN);
1094 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1095 if (err < 0)
1096 put_page(virt_to_head_page(buf));
1097 return err;
1098}
1099
1100static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1101 gfp_t gfp)
1102{
1103 struct page *first, *list = NULL;
1104 char *p;
1105 int i, err, offset;
1106
1107 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1108
1109 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1110 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1111 first = get_a_page(rq, gfp);
1112 if (!first) {
1113 if (list)
1114 give_pages(rq, list);
1115 return -ENOMEM;
1116 }
1117 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1118
1119 /* chain new page in list head to match sg */
1120 first->private = (unsigned long)list;
1121 list = first;
1122 }
1123
1124 first = get_a_page(rq, gfp);
1125 if (!first) {
1126 give_pages(rq, list);
1127 return -ENOMEM;
1128 }
1129 p = page_address(first);
1130
1131 /* rq->sg[0], rq->sg[1] share the same page */
1132 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1133 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1134
1135 /* rq->sg[1] for data packet, from offset */
1136 offset = sizeof(struct padded_vnet_hdr);
1137 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1138
1139 /* chain first in list head */
1140 first->private = (unsigned long)list;
1141 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1142 first, gfp);
1143 if (err < 0)
1144 give_pages(rq, first);
1145
1146 return err;
1147}
1148
1149static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1150 struct ewma_pkt_len *avg_pkt_len,
1151 unsigned int room)
1152{
1153 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1154 unsigned int len;
1155
1156 if (room)
1157 return PAGE_SIZE - room;
1158
1159 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1160 rq->min_buf_len, PAGE_SIZE - hdr_len);
1161
1162 return ALIGN(len, L1_CACHE_BYTES);
1163}
1164
1165static int add_recvbuf_mergeable(struct virtnet_info *vi,
1166 struct receive_queue *rq, gfp_t gfp)
1167{
1168 struct page_frag *alloc_frag = &rq->alloc_frag;
1169 unsigned int headroom = virtnet_get_headroom(vi);
1170 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1171 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1172 char *buf;
1173 void *ctx;
1174 int err;
1175 unsigned int len, hole;
1176
1177 /* Extra tailroom is needed to satisfy XDP's assumption. This
1178 * means rx frags coalescing won't work, but consider we've
1179 * disabled GSO for XDP, it won't be a big issue.
1180 */
1181 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1182 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1183 return -ENOMEM;
1184
1185 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1186 buf += headroom; /* advance address leaving hole at front of pkt */
1187 get_page(alloc_frag->page);
1188 alloc_frag->offset += len + room;
1189 hole = alloc_frag->size - alloc_frag->offset;
1190 if (hole < len + room) {
1191 /* To avoid internal fragmentation, if there is very likely not
1192 * enough space for another buffer, add the remaining space to
1193 * the current buffer.
1194 */
1195 len += hole;
1196 alloc_frag->offset += hole;
1197 }
1198
1199 sg_init_one(rq->sg, buf, len);
1200 ctx = mergeable_len_to_ctx(len, headroom);
1201 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1202 if (err < 0)
1203 put_page(virt_to_head_page(buf));
1204
1205 return err;
1206}
1207
1208/*
1209 * Returns false if we couldn't fill entirely (OOM).
1210 *
1211 * Normally run in the receive path, but can also be run from ndo_open
1212 * before we're receiving packets, or from refill_work which is
1213 * careful to disable receiving (using napi_disable).
1214 */
1215static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1216 gfp_t gfp)
1217{
1218 int err;
1219 bool oom;
1220
1221 do {
1222 if (vi->mergeable_rx_bufs)
1223 err = add_recvbuf_mergeable(vi, rq, gfp);
1224 else if (vi->big_packets)
1225 err = add_recvbuf_big(vi, rq, gfp);
1226 else
1227 err = add_recvbuf_small(vi, rq, gfp);
1228
1229 oom = err == -ENOMEM;
1230 if (err)
1231 break;
1232 } while (rq->vq->num_free);
1233 if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1234 u64_stats_update_begin(&rq->stats.syncp);
1235 rq->stats.kicks++;
1236 u64_stats_update_end(&rq->stats.syncp);
1237 }
1238
1239 return !oom;
1240}
1241
1242static void skb_recv_done(struct virtqueue *rvq)
1243{
1244 struct virtnet_info *vi = rvq->vdev->priv;
1245 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1246
1247 virtqueue_napi_schedule(&rq->napi, rvq);
1248}
1249
1250static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1251{
1252 napi_enable(napi);
1253
1254 /* If all buffers were filled by other side before we napi_enabled, we
1255 * won't get another interrupt, so process any outstanding packets now.
1256 * Call local_bh_enable after to trigger softIRQ processing.
1257 */
1258 local_bh_disable();
1259 virtqueue_napi_schedule(napi, vq);
1260 local_bh_enable();
1261}
1262
1263static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1264 struct virtqueue *vq,
1265 struct napi_struct *napi)
1266{
1267 if (!napi->weight)
1268 return;
1269
1270 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1271 * enable the feature if this is likely affine with the transmit path.
1272 */
1273 if (!vi->affinity_hint_set) {
1274 napi->weight = 0;
1275 return;
1276 }
1277
1278 return virtnet_napi_enable(vq, napi);
1279}
1280
1281static void virtnet_napi_tx_disable(struct napi_struct *napi)
1282{
1283 if (napi->weight)
1284 napi_disable(napi);
1285}
1286
1287static void refill_work(struct work_struct *work)
1288{
1289 struct virtnet_info *vi =
1290 container_of(work, struct virtnet_info, refill.work);
1291 bool still_empty;
1292 int i;
1293
1294 for (i = 0; i < vi->curr_queue_pairs; i++) {
1295 struct receive_queue *rq = &vi->rq[i];
1296
1297 napi_disable(&rq->napi);
1298 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1299 virtnet_napi_enable(rq->vq, &rq->napi);
1300
1301 /* In theory, this can happen: if we don't get any buffers in
1302 * we will *never* try to fill again.
1303 */
1304 if (still_empty)
1305 schedule_delayed_work(&vi->refill, HZ/2);
1306 }
1307}
1308
1309static int virtnet_receive(struct receive_queue *rq, int budget,
1310 unsigned int *xdp_xmit)
1311{
1312 struct virtnet_info *vi = rq->vq->vdev->priv;
1313 struct virtnet_rq_stats stats = {};
1314 unsigned int len;
1315 void *buf;
1316 int i;
1317
1318 if (!vi->big_packets || vi->mergeable_rx_bufs) {
1319 void *ctx;
1320
1321 while (stats.packets < budget &&
1322 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1323 receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1324 stats.packets++;
1325 }
1326 } else {
1327 while (stats.packets < budget &&
1328 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1329 receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1330 stats.packets++;
1331 }
1332 }
1333
1334 if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1335 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1336 schedule_delayed_work(&vi->refill, 0);
1337 }
1338
1339 u64_stats_update_begin(&rq->stats.syncp);
1340 for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1341 size_t offset = virtnet_rq_stats_desc[i].offset;
1342 u64 *item;
1343
1344 item = (u64 *)((u8 *)&rq->stats + offset);
1345 *item += *(u64 *)((u8 *)&stats + offset);
1346 }
1347 u64_stats_update_end(&rq->stats.syncp);
1348
1349 return stats.packets;
1350}
1351
1352static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1353{
1354 unsigned int len;
1355 unsigned int packets = 0;
1356 unsigned int bytes = 0;
1357 void *ptr;
1358
1359 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1360 if (likely(!is_xdp_frame(ptr))) {
1361 struct sk_buff *skb = ptr;
1362
1363 pr_debug("Sent skb %p\n", skb);
1364
1365 bytes += skb->len;
1366 napi_consume_skb(skb, in_napi);
1367 } else {
1368 struct xdp_frame *frame = ptr_to_xdp(ptr);
1369
1370 bytes += frame->len;
1371 xdp_return_frame(frame);
1372 }
1373 packets++;
1374 }
1375
1376 /* Avoid overhead when no packets have been processed
1377 * happens when called speculatively from start_xmit.
1378 */
1379 if (!packets)
1380 return;
1381
1382 u64_stats_update_begin(&sq->stats.syncp);
1383 sq->stats.bytes += bytes;
1384 sq->stats.packets += packets;
1385 u64_stats_update_end(&sq->stats.syncp);
1386}
1387
1388static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1389{
1390 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1391 return false;
1392 else if (q < vi->curr_queue_pairs)
1393 return true;
1394 else
1395 return false;
1396}
1397
1398static void virtnet_poll_cleantx(struct receive_queue *rq)
1399{
1400 struct virtnet_info *vi = rq->vq->vdev->priv;
1401 unsigned int index = vq2rxq(rq->vq);
1402 struct send_queue *sq = &vi->sq[index];
1403 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1404
1405 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1406 return;
1407
1408 if (__netif_tx_trylock(txq)) {
1409 free_old_xmit_skbs(sq, true);
1410 __netif_tx_unlock(txq);
1411 }
1412
1413 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1414 netif_tx_wake_queue(txq);
1415}
1416
1417static int virtnet_poll(struct napi_struct *napi, int budget)
1418{
1419 struct receive_queue *rq =
1420 container_of(napi, struct receive_queue, napi);
1421 struct virtnet_info *vi = rq->vq->vdev->priv;
1422 struct send_queue *sq;
1423 unsigned int received;
1424 unsigned int xdp_xmit = 0;
1425
1426 virtnet_poll_cleantx(rq);
1427
1428 received = virtnet_receive(rq, budget, &xdp_xmit);
1429
1430 /* Out of packets? */
1431 if (received < budget)
1432 virtqueue_napi_complete(napi, rq->vq, received);
1433
1434 if (xdp_xmit & VIRTIO_XDP_REDIR)
1435 xdp_do_flush_map();
1436
1437 if (xdp_xmit & VIRTIO_XDP_TX) {
1438 sq = virtnet_xdp_sq(vi);
1439 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1440 u64_stats_update_begin(&sq->stats.syncp);
1441 sq->stats.kicks++;
1442 u64_stats_update_end(&sq->stats.syncp);
1443 }
1444 }
1445
1446 return received;
1447}
1448
1449static int virtnet_open(struct net_device *dev)
1450{
1451 struct virtnet_info *vi = netdev_priv(dev);
1452 int i, err;
1453
1454 for (i = 0; i < vi->max_queue_pairs; i++) {
1455 if (i < vi->curr_queue_pairs)
1456 /* Make sure we have some buffers: if oom use wq. */
1457 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1458 schedule_delayed_work(&vi->refill, 0);
1459
1460 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i);
1461 if (err < 0)
1462 return err;
1463
1464 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1465 MEM_TYPE_PAGE_SHARED, NULL);
1466 if (err < 0) {
1467 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1468 return err;
1469 }
1470
1471 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1472 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1473 }
1474
1475 return 0;
1476}
1477
1478static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1479{
1480 struct send_queue *sq = container_of(napi, struct send_queue, napi);
1481 struct virtnet_info *vi = sq->vq->vdev->priv;
1482 unsigned int index = vq2txq(sq->vq);
1483 struct netdev_queue *txq;
1484
1485 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1486 /* We don't need to enable cb for XDP */
1487 napi_complete_done(napi, 0);
1488 return 0;
1489 }
1490
1491 txq = netdev_get_tx_queue(vi->dev, index);
1492 __netif_tx_lock(txq, raw_smp_processor_id());
1493 free_old_xmit_skbs(sq, true);
1494 __netif_tx_unlock(txq);
1495
1496 virtqueue_napi_complete(napi, sq->vq, 0);
1497
1498 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1499 netif_tx_wake_queue(txq);
1500
1501 return 0;
1502}
1503
1504static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1505{
1506 struct virtio_net_hdr_mrg_rxbuf *hdr;
1507 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1508 struct virtnet_info *vi = sq->vq->vdev->priv;
1509 int num_sg;
1510 unsigned hdr_len = vi->hdr_len;
1511 bool can_push;
1512
1513 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1514
1515 can_push = vi->any_header_sg &&
1516 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1517 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1518 /* Even if we can, don't push here yet as this would skew
1519 * csum_start offset below. */
1520 if (can_push)
1521 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1522 else
1523 hdr = skb_vnet_hdr(skb);
1524
1525 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1526 virtio_is_little_endian(vi->vdev), false,
1527 0))
1528 BUG();
1529
1530 if (vi->mergeable_rx_bufs)
1531 hdr->num_buffers = 0;
1532
1533 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1534 if (can_push) {
1535 __skb_push(skb, hdr_len);
1536 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1537 if (unlikely(num_sg < 0))
1538 return num_sg;
1539 /* Pull header back to avoid skew in tx bytes calculations. */
1540 __skb_pull(skb, hdr_len);
1541 } else {
1542 sg_set_buf(sq->sg, hdr, hdr_len);
1543 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1544 if (unlikely(num_sg < 0))
1545 return num_sg;
1546 num_sg++;
1547 }
1548 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1549}
1550
1551static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1552{
1553 struct virtnet_info *vi = netdev_priv(dev);
1554 int qnum = skb_get_queue_mapping(skb);
1555 struct send_queue *sq = &vi->sq[qnum];
1556 int err;
1557 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1558 bool kick = !netdev_xmit_more();
1559 bool use_napi = sq->napi.weight;
1560
1561 /* Free up any pending old buffers before queueing new ones. */
1562 free_old_xmit_skbs(sq, false);
1563
1564 if (use_napi && kick)
1565 virtqueue_enable_cb_delayed(sq->vq);
1566
1567 /* timestamp packet in software */
1568 skb_tx_timestamp(skb);
1569
1570 /* Try to transmit */
1571 err = xmit_skb(sq, skb);
1572
1573 /* This should not happen! */
1574 if (unlikely(err)) {
1575 dev->stats.tx_fifo_errors++;
1576 if (net_ratelimit())
1577 dev_warn(&dev->dev,
1578 "Unexpected TXQ (%d) queue failure: %d\n",
1579 qnum, err);
1580 dev->stats.tx_dropped++;
1581 dev_kfree_skb_any(skb);
1582 return NETDEV_TX_OK;
1583 }
1584
1585 /* Don't wait up for transmitted skbs to be freed. */
1586 if (!use_napi) {
1587 skb_orphan(skb);
1588 nf_reset_ct(skb);
1589 }
1590
1591 /* If running out of space, stop queue to avoid getting packets that we
1592 * are then unable to transmit.
1593 * An alternative would be to force queuing layer to requeue the skb by
1594 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1595 * returned in a normal path of operation: it means that driver is not
1596 * maintaining the TX queue stop/start state properly, and causes
1597 * the stack to do a non-trivial amount of useless work.
1598 * Since most packets only take 1 or 2 ring slots, stopping the queue
1599 * early means 16 slots are typically wasted.
1600 */
1601 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1602 netif_stop_subqueue(dev, qnum);
1603 if (!use_napi &&
1604 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1605 /* More just got used, free them then recheck. */
1606 free_old_xmit_skbs(sq, false);
1607 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1608 netif_start_subqueue(dev, qnum);
1609 virtqueue_disable_cb(sq->vq);
1610 }
1611 }
1612 }
1613
1614 if (kick || netif_xmit_stopped(txq)) {
1615 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1616 u64_stats_update_begin(&sq->stats.syncp);
1617 sq->stats.kicks++;
1618 u64_stats_update_end(&sq->stats.syncp);
1619 }
1620 }
1621
1622 return NETDEV_TX_OK;
1623}
1624
1625/*
1626 * Send command via the control virtqueue and check status. Commands
1627 * supported by the hypervisor, as indicated by feature bits, should
1628 * never fail unless improperly formatted.
1629 */
1630static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1631 struct scatterlist *out)
1632{
1633 struct scatterlist *sgs[4], hdr, stat;
1634 unsigned out_num = 0, tmp;
1635
1636 /* Caller should know better */
1637 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1638
1639 vi->ctrl->status = ~0;
1640 vi->ctrl->hdr.class = class;
1641 vi->ctrl->hdr.cmd = cmd;
1642 /* Add header */
1643 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1644 sgs[out_num++] = &hdr;
1645
1646 if (out)
1647 sgs[out_num++] = out;
1648
1649 /* Add return status. */
1650 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1651 sgs[out_num] = &stat;
1652
1653 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1654 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1655
1656 if (unlikely(!virtqueue_kick(vi->cvq)))
1657 return vi->ctrl->status == VIRTIO_NET_OK;
1658
1659 /* Spin for a response, the kick causes an ioport write, trapping
1660 * into the hypervisor, so the request should be handled immediately.
1661 */
1662 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1663 !virtqueue_is_broken(vi->cvq))
1664 cpu_relax();
1665
1666 return vi->ctrl->status == VIRTIO_NET_OK;
1667}
1668
1669static int virtnet_set_mac_address(struct net_device *dev, void *p)
1670{
1671 struct virtnet_info *vi = netdev_priv(dev);
1672 struct virtio_device *vdev = vi->vdev;
1673 int ret;
1674 struct sockaddr *addr;
1675 struct scatterlist sg;
1676
1677 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1678 return -EOPNOTSUPP;
1679
1680 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1681 if (!addr)
1682 return -ENOMEM;
1683
1684 ret = eth_prepare_mac_addr_change(dev, addr);
1685 if (ret)
1686 goto out;
1687
1688 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1689 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1690 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1691 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1692 dev_warn(&vdev->dev,
1693 "Failed to set mac address by vq command.\n");
1694 ret = -EINVAL;
1695 goto out;
1696 }
1697 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1698 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1699 unsigned int i;
1700
1701 /* Naturally, this has an atomicity problem. */
1702 for (i = 0; i < dev->addr_len; i++)
1703 virtio_cwrite8(vdev,
1704 offsetof(struct virtio_net_config, mac) +
1705 i, addr->sa_data[i]);
1706 }
1707
1708 eth_commit_mac_addr_change(dev, p);
1709 ret = 0;
1710
1711out:
1712 kfree(addr);
1713 return ret;
1714}
1715
1716static void virtnet_stats(struct net_device *dev,
1717 struct rtnl_link_stats64 *tot)
1718{
1719 struct virtnet_info *vi = netdev_priv(dev);
1720 unsigned int start;
1721 int i;
1722
1723 for (i = 0; i < vi->max_queue_pairs; i++) {
1724 u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1725 struct receive_queue *rq = &vi->rq[i];
1726 struct send_queue *sq = &vi->sq[i];
1727
1728 do {
1729 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1730 tpackets = sq->stats.packets;
1731 tbytes = sq->stats.bytes;
1732 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1733
1734 do {
1735 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1736 rpackets = rq->stats.packets;
1737 rbytes = rq->stats.bytes;
1738 rdrops = rq->stats.drops;
1739 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1740
1741 tot->rx_packets += rpackets;
1742 tot->tx_packets += tpackets;
1743 tot->rx_bytes += rbytes;
1744 tot->tx_bytes += tbytes;
1745 tot->rx_dropped += rdrops;
1746 }
1747
1748 tot->tx_dropped = dev->stats.tx_dropped;
1749 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1750 tot->rx_length_errors = dev->stats.rx_length_errors;
1751 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1752}
1753
1754static void virtnet_ack_link_announce(struct virtnet_info *vi)
1755{
1756 rtnl_lock();
1757 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1758 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1759 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1760 rtnl_unlock();
1761}
1762
1763static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1764{
1765 struct scatterlist sg;
1766 struct net_device *dev = vi->dev;
1767
1768 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1769 return 0;
1770
1771 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1772 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1773
1774 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1775 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1776 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1777 queue_pairs);
1778 return -EINVAL;
1779 } else {
1780 vi->curr_queue_pairs = queue_pairs;
1781 /* virtnet_open() will refill when device is going to up. */
1782 if (dev->flags & IFF_UP)
1783 schedule_delayed_work(&vi->refill, 0);
1784 }
1785
1786 return 0;
1787}
1788
1789static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1790{
1791 int err;
1792
1793 rtnl_lock();
1794 err = _virtnet_set_queues(vi, queue_pairs);
1795 rtnl_unlock();
1796 return err;
1797}
1798
1799static int virtnet_close(struct net_device *dev)
1800{
1801 struct virtnet_info *vi = netdev_priv(dev);
1802 int i;
1803
1804 /* Make sure refill_work doesn't re-enable napi! */
1805 cancel_delayed_work_sync(&vi->refill);
1806
1807 for (i = 0; i < vi->max_queue_pairs; i++) {
1808 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1809 napi_disable(&vi->rq[i].napi);
1810 virtnet_napi_tx_disable(&vi->sq[i].napi);
1811 }
1812
1813 return 0;
1814}
1815
1816static void virtnet_set_rx_mode(struct net_device *dev)
1817{
1818 struct virtnet_info *vi = netdev_priv(dev);
1819 struct scatterlist sg[2];
1820 struct virtio_net_ctrl_mac *mac_data;
1821 struct netdev_hw_addr *ha;
1822 int uc_count;
1823 int mc_count;
1824 void *buf;
1825 int i;
1826
1827 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1828 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1829 return;
1830
1831 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1832 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1833
1834 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1835
1836 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1837 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1838 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1839 vi->ctrl->promisc ? "en" : "dis");
1840
1841 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1842
1843 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1844 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1845 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1846 vi->ctrl->allmulti ? "en" : "dis");
1847
1848 uc_count = netdev_uc_count(dev);
1849 mc_count = netdev_mc_count(dev);
1850 /* MAC filter - use one buffer for both lists */
1851 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1852 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1853 mac_data = buf;
1854 if (!buf)
1855 return;
1856
1857 sg_init_table(sg, 2);
1858
1859 /* Store the unicast list and count in the front of the buffer */
1860 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1861 i = 0;
1862 netdev_for_each_uc_addr(ha, dev)
1863 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1864
1865 sg_set_buf(&sg[0], mac_data,
1866 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1867
1868 /* multicast list and count fill the end */
1869 mac_data = (void *)&mac_data->macs[uc_count][0];
1870
1871 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1872 i = 0;
1873 netdev_for_each_mc_addr(ha, dev)
1874 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1875
1876 sg_set_buf(&sg[1], mac_data,
1877 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1878
1879 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1880 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1881 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1882
1883 kfree(buf);
1884}
1885
1886static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1887 __be16 proto, u16 vid)
1888{
1889 struct virtnet_info *vi = netdev_priv(dev);
1890 struct scatterlist sg;
1891
1892 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1893 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1894
1895 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1896 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1897 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1898 return 0;
1899}
1900
1901static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1902 __be16 proto, u16 vid)
1903{
1904 struct virtnet_info *vi = netdev_priv(dev);
1905 struct scatterlist sg;
1906
1907 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1908 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1909
1910 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1911 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1912 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1913 return 0;
1914}
1915
1916static void virtnet_clean_affinity(struct virtnet_info *vi)
1917{
1918 int i;
1919
1920 if (vi->affinity_hint_set) {
1921 for (i = 0; i < vi->max_queue_pairs; i++) {
1922 virtqueue_set_affinity(vi->rq[i].vq, NULL);
1923 virtqueue_set_affinity(vi->sq[i].vq, NULL);
1924 }
1925
1926 vi->affinity_hint_set = false;
1927 }
1928}
1929
1930static void virtnet_set_affinity(struct virtnet_info *vi)
1931{
1932 cpumask_var_t mask;
1933 int stragglers;
1934 int group_size;
1935 int i, j, cpu;
1936 int num_cpu;
1937 int stride;
1938
1939 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
1940 virtnet_clean_affinity(vi);
1941 return;
1942 }
1943
1944 num_cpu = num_online_cpus();
1945 stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
1946 stragglers = num_cpu >= vi->curr_queue_pairs ?
1947 num_cpu % vi->curr_queue_pairs :
1948 0;
1949 cpu = cpumask_next(-1, cpu_online_mask);
1950
1951 for (i = 0; i < vi->curr_queue_pairs; i++) {
1952 group_size = stride + (i < stragglers ? 1 : 0);
1953
1954 for (j = 0; j < group_size; j++) {
1955 cpumask_set_cpu(cpu, mask);
1956 cpu = cpumask_next_wrap(cpu, cpu_online_mask,
1957 nr_cpu_ids, false);
1958 }
1959 virtqueue_set_affinity(vi->rq[i].vq, mask);
1960 virtqueue_set_affinity(vi->sq[i].vq, mask);
1961 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, false);
1962 cpumask_clear(mask);
1963 }
1964
1965 vi->affinity_hint_set = true;
1966 free_cpumask_var(mask);
1967}
1968
1969static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1970{
1971 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1972 node);
1973 virtnet_set_affinity(vi);
1974 return 0;
1975}
1976
1977static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1978{
1979 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1980 node_dead);
1981 virtnet_set_affinity(vi);
1982 return 0;
1983}
1984
1985static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1986{
1987 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1988 node);
1989
1990 virtnet_clean_affinity(vi);
1991 return 0;
1992}
1993
1994static enum cpuhp_state virtionet_online;
1995
1996static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1997{
1998 int ret;
1999
2000 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2001 if (ret)
2002 return ret;
2003 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2004 &vi->node_dead);
2005 if (!ret)
2006 return ret;
2007 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2008 return ret;
2009}
2010
2011static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2012{
2013 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2014 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2015 &vi->node_dead);
2016}
2017
2018static void virtnet_get_ringparam(struct net_device *dev,
2019 struct ethtool_ringparam *ring)
2020{
2021 struct virtnet_info *vi = netdev_priv(dev);
2022
2023 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2024 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2025 ring->rx_pending = ring->rx_max_pending;
2026 ring->tx_pending = ring->tx_max_pending;
2027}
2028
2029
2030static void virtnet_get_drvinfo(struct net_device *dev,
2031 struct ethtool_drvinfo *info)
2032{
2033 struct virtnet_info *vi = netdev_priv(dev);
2034 struct virtio_device *vdev = vi->vdev;
2035
2036 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2037 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2038 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2039
2040}
2041
2042/* TODO: Eliminate OOO packets during switching */
2043static int virtnet_set_channels(struct net_device *dev,
2044 struct ethtool_channels *channels)
2045{
2046 struct virtnet_info *vi = netdev_priv(dev);
2047 u16 queue_pairs = channels->combined_count;
2048 int err;
2049
2050 /* We don't support separate rx/tx channels.
2051 * We don't allow setting 'other' channels.
2052 */
2053 if (channels->rx_count || channels->tx_count || channels->other_count)
2054 return -EINVAL;
2055
2056 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2057 return -EINVAL;
2058
2059 /* For now we don't support modifying channels while XDP is loaded
2060 * also when XDP is loaded all RX queues have XDP programs so we only
2061 * need to check a single RX queue.
2062 */
2063 if (vi->rq[0].xdp_prog)
2064 return -EINVAL;
2065
2066 get_online_cpus();
2067 err = _virtnet_set_queues(vi, queue_pairs);
2068 if (!err) {
2069 netif_set_real_num_tx_queues(dev, queue_pairs);
2070 netif_set_real_num_rx_queues(dev, queue_pairs);
2071
2072 virtnet_set_affinity(vi);
2073 }
2074 put_online_cpus();
2075
2076 return err;
2077}
2078
2079static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2080{
2081 struct virtnet_info *vi = netdev_priv(dev);
2082 char *p = (char *)data;
2083 unsigned int i, j;
2084
2085 switch (stringset) {
2086 case ETH_SS_STATS:
2087 for (i = 0; i < vi->curr_queue_pairs; i++) {
2088 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2089 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s",
2090 i, virtnet_rq_stats_desc[j].desc);
2091 p += ETH_GSTRING_LEN;
2092 }
2093 }
2094
2095 for (i = 0; i < vi->curr_queue_pairs; i++) {
2096 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2097 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s",
2098 i, virtnet_sq_stats_desc[j].desc);
2099 p += ETH_GSTRING_LEN;
2100 }
2101 }
2102 break;
2103 }
2104}
2105
2106static int virtnet_get_sset_count(struct net_device *dev, int sset)
2107{
2108 struct virtnet_info *vi = netdev_priv(dev);
2109
2110 switch (sset) {
2111 case ETH_SS_STATS:
2112 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2113 VIRTNET_SQ_STATS_LEN);
2114 default:
2115 return -EOPNOTSUPP;
2116 }
2117}
2118
2119static void virtnet_get_ethtool_stats(struct net_device *dev,
2120 struct ethtool_stats *stats, u64 *data)
2121{
2122 struct virtnet_info *vi = netdev_priv(dev);
2123 unsigned int idx = 0, start, i, j;
2124 const u8 *stats_base;
2125 size_t offset;
2126
2127 for (i = 0; i < vi->curr_queue_pairs; i++) {
2128 struct receive_queue *rq = &vi->rq[i];
2129
2130 stats_base = (u8 *)&rq->stats;
2131 do {
2132 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2133 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2134 offset = virtnet_rq_stats_desc[j].offset;
2135 data[idx + j] = *(u64 *)(stats_base + offset);
2136 }
2137 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2138 idx += VIRTNET_RQ_STATS_LEN;
2139 }
2140
2141 for (i = 0; i < vi->curr_queue_pairs; i++) {
2142 struct send_queue *sq = &vi->sq[i];
2143
2144 stats_base = (u8 *)&sq->stats;
2145 do {
2146 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2147 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2148 offset = virtnet_sq_stats_desc[j].offset;
2149 data[idx + j] = *(u64 *)(stats_base + offset);
2150 }
2151 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2152 idx += VIRTNET_SQ_STATS_LEN;
2153 }
2154}
2155
2156static void virtnet_get_channels(struct net_device *dev,
2157 struct ethtool_channels *channels)
2158{
2159 struct virtnet_info *vi = netdev_priv(dev);
2160
2161 channels->combined_count = vi->curr_queue_pairs;
2162 channels->max_combined = vi->max_queue_pairs;
2163 channels->max_other = 0;
2164 channels->rx_count = 0;
2165 channels->tx_count = 0;
2166 channels->other_count = 0;
2167}
2168
2169/* Check if the user is trying to change anything besides speed/duplex */
2170static bool
2171virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
2172{
2173 struct ethtool_link_ksettings diff1 = *cmd;
2174 struct ethtool_link_ksettings diff2 = {};
2175
2176 /* cmd is always set so we need to clear it, validate the port type
2177 * and also without autonegotiation we can ignore advertising
2178 */
2179 diff1.base.speed = 0;
2180 diff2.base.port = PORT_OTHER;
2181 ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
2182 diff1.base.duplex = 0;
2183 diff1.base.cmd = 0;
2184 diff1.base.link_mode_masks_nwords = 0;
2185
2186 return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
2187 bitmap_empty(diff1.link_modes.supported,
2188 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2189 bitmap_empty(diff1.link_modes.advertising,
2190 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2191 bitmap_empty(diff1.link_modes.lp_advertising,
2192 __ETHTOOL_LINK_MODE_MASK_NBITS);
2193}
2194
2195static int virtnet_set_link_ksettings(struct net_device *dev,
2196 const struct ethtool_link_ksettings *cmd)
2197{
2198 struct virtnet_info *vi = netdev_priv(dev);
2199 u32 speed;
2200
2201 speed = cmd->base.speed;
2202 /* don't allow custom speed and duplex */
2203 if (!ethtool_validate_speed(speed) ||
2204 !ethtool_validate_duplex(cmd->base.duplex) ||
2205 !virtnet_validate_ethtool_cmd(cmd))
2206 return -EINVAL;
2207 vi->speed = speed;
2208 vi->duplex = cmd->base.duplex;
2209
2210 return 0;
2211}
2212
2213static int virtnet_get_link_ksettings(struct net_device *dev,
2214 struct ethtool_link_ksettings *cmd)
2215{
2216 struct virtnet_info *vi = netdev_priv(dev);
2217
2218 cmd->base.speed = vi->speed;
2219 cmd->base.duplex = vi->duplex;
2220 cmd->base.port = PORT_OTHER;
2221
2222 return 0;
2223}
2224
2225static int virtnet_set_coalesce(struct net_device *dev,
2226 struct ethtool_coalesce *ec)
2227{
2228 struct ethtool_coalesce ec_default = {
2229 .cmd = ETHTOOL_SCOALESCE,
2230 .rx_max_coalesced_frames = 1,
2231 };
2232 struct virtnet_info *vi = netdev_priv(dev);
2233 int i, napi_weight;
2234
2235 if (ec->tx_max_coalesced_frames > 1)
2236 return -EINVAL;
2237
2238 ec_default.tx_max_coalesced_frames = ec->tx_max_coalesced_frames;
2239 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2240
2241 /* disallow changes to fields not explicitly tested above */
2242 if (memcmp(ec, &ec_default, sizeof(ec_default)))
2243 return -EINVAL;
2244
2245 if (napi_weight ^ vi->sq[0].napi.weight) {
2246 if (dev->flags & IFF_UP)
2247 return -EBUSY;
2248 for (i = 0; i < vi->max_queue_pairs; i++)
2249 vi->sq[i].napi.weight = napi_weight;
2250 }
2251
2252 return 0;
2253}
2254
2255static int virtnet_get_coalesce(struct net_device *dev,
2256 struct ethtool_coalesce *ec)
2257{
2258 struct ethtool_coalesce ec_default = {
2259 .cmd = ETHTOOL_GCOALESCE,
2260 .rx_max_coalesced_frames = 1,
2261 };
2262 struct virtnet_info *vi = netdev_priv(dev);
2263
2264 memcpy(ec, &ec_default, sizeof(ec_default));
2265
2266 if (vi->sq[0].napi.weight)
2267 ec->tx_max_coalesced_frames = 1;
2268
2269 return 0;
2270}
2271
2272static void virtnet_init_settings(struct net_device *dev)
2273{
2274 struct virtnet_info *vi = netdev_priv(dev);
2275
2276 vi->speed = SPEED_UNKNOWN;
2277 vi->duplex = DUPLEX_UNKNOWN;
2278}
2279
2280static void virtnet_update_settings(struct virtnet_info *vi)
2281{
2282 u32 speed;
2283 u8 duplex;
2284
2285 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2286 return;
2287
2288 speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config,
2289 speed));
2290 if (ethtool_validate_speed(speed))
2291 vi->speed = speed;
2292 duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config,
2293 duplex));
2294 if (ethtool_validate_duplex(duplex))
2295 vi->duplex = duplex;
2296}
2297
2298static const struct ethtool_ops virtnet_ethtool_ops = {
2299 .get_drvinfo = virtnet_get_drvinfo,
2300 .get_link = ethtool_op_get_link,
2301 .get_ringparam = virtnet_get_ringparam,
2302 .get_strings = virtnet_get_strings,
2303 .get_sset_count = virtnet_get_sset_count,
2304 .get_ethtool_stats = virtnet_get_ethtool_stats,
2305 .set_channels = virtnet_set_channels,
2306 .get_channels = virtnet_get_channels,
2307 .get_ts_info = ethtool_op_get_ts_info,
2308 .get_link_ksettings = virtnet_get_link_ksettings,
2309 .set_link_ksettings = virtnet_set_link_ksettings,
2310 .set_coalesce = virtnet_set_coalesce,
2311 .get_coalesce = virtnet_get_coalesce,
2312};
2313
2314static void virtnet_freeze_down(struct virtio_device *vdev)
2315{
2316 struct virtnet_info *vi = vdev->priv;
2317 int i;
2318
2319 /* Make sure no work handler is accessing the device */
2320 flush_work(&vi->config_work);
2321
2322 netif_tx_lock_bh(vi->dev);
2323 netif_device_detach(vi->dev);
2324 netif_tx_unlock_bh(vi->dev);
2325 cancel_delayed_work_sync(&vi->refill);
2326
2327 if (netif_running(vi->dev)) {
2328 for (i = 0; i < vi->max_queue_pairs; i++) {
2329 napi_disable(&vi->rq[i].napi);
2330 virtnet_napi_tx_disable(&vi->sq[i].napi);
2331 }
2332 }
2333}
2334
2335static int init_vqs(struct virtnet_info *vi);
2336
2337static int virtnet_restore_up(struct virtio_device *vdev)
2338{
2339 struct virtnet_info *vi = vdev->priv;
2340 int err, i;
2341
2342 err = init_vqs(vi);
2343 if (err)
2344 return err;
2345
2346 virtio_device_ready(vdev);
2347
2348 if (netif_running(vi->dev)) {
2349 for (i = 0; i < vi->curr_queue_pairs; i++)
2350 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2351 schedule_delayed_work(&vi->refill, 0);
2352
2353 for (i = 0; i < vi->max_queue_pairs; i++) {
2354 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2355 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2356 &vi->sq[i].napi);
2357 }
2358 }
2359
2360 netif_tx_lock_bh(vi->dev);
2361 netif_device_attach(vi->dev);
2362 netif_tx_unlock_bh(vi->dev);
2363 return err;
2364}
2365
2366static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2367{
2368 struct scatterlist sg;
2369 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2370
2371 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2372
2373 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2374 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2375 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
2376 return -EINVAL;
2377 }
2378
2379 return 0;
2380}
2381
2382static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2383{
2384 u64 offloads = 0;
2385
2386 if (!vi->guest_offloads)
2387 return 0;
2388
2389 return virtnet_set_guest_offloads(vi, offloads);
2390}
2391
2392static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2393{
2394 u64 offloads = vi->guest_offloads;
2395
2396 if (!vi->guest_offloads)
2397 return 0;
2398
2399 return virtnet_set_guest_offloads(vi, offloads);
2400}
2401
2402static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2403 struct netlink_ext_ack *extack)
2404{
2405 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2406 struct virtnet_info *vi = netdev_priv(dev);
2407 struct bpf_prog *old_prog;
2408 u16 xdp_qp = 0, curr_qp;
2409 int i, err;
2410
2411 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2412 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2413 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2414 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2415 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2416 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2417 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
2418 return -EOPNOTSUPP;
2419 }
2420
2421 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2422 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2423 return -EINVAL;
2424 }
2425
2426 if (dev->mtu > max_sz) {
2427 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2428 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2429 return -EINVAL;
2430 }
2431
2432 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2433 if (prog)
2434 xdp_qp = nr_cpu_ids;
2435
2436 /* XDP requires extra queues for XDP_TX */
2437 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2438 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2439 netdev_warn(dev, "request %i queues but max is %i\n",
2440 curr_qp + xdp_qp, vi->max_queue_pairs);
2441 return -ENOMEM;
2442 }
2443
2444 old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2445 if (!prog && !old_prog)
2446 return 0;
2447
2448 if (prog) {
2449 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
2450 if (IS_ERR(prog))
2451 return PTR_ERR(prog);
2452 }
2453
2454 /* Make sure NAPI is not using any XDP TX queues for RX. */
2455 if (netif_running(dev)) {
2456 for (i = 0; i < vi->max_queue_pairs; i++) {
2457 napi_disable(&vi->rq[i].napi);
2458 virtnet_napi_tx_disable(&vi->sq[i].napi);
2459 }
2460 }
2461
2462 if (!prog) {
2463 for (i = 0; i < vi->max_queue_pairs; i++) {
2464 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2465 if (i == 0)
2466 virtnet_restore_guest_offloads(vi);
2467 }
2468 synchronize_net();
2469 }
2470
2471 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2472 if (err)
2473 goto err;
2474 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2475 vi->xdp_queue_pairs = xdp_qp;
2476
2477 if (prog) {
2478 for (i = 0; i < vi->max_queue_pairs; i++) {
2479 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2480 if (i == 0 && !old_prog)
2481 virtnet_clear_guest_offloads(vi);
2482 }
2483 }
2484
2485 for (i = 0; i < vi->max_queue_pairs; i++) {
2486 if (old_prog)
2487 bpf_prog_put(old_prog);
2488 if (netif_running(dev)) {
2489 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2490 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2491 &vi->sq[i].napi);
2492 }
2493 }
2494
2495 return 0;
2496
2497err:
2498 if (!prog) {
2499 virtnet_clear_guest_offloads(vi);
2500 for (i = 0; i < vi->max_queue_pairs; i++)
2501 rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2502 }
2503
2504 if (netif_running(dev)) {
2505 for (i = 0; i < vi->max_queue_pairs; i++) {
2506 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2507 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2508 &vi->sq[i].napi);
2509 }
2510 }
2511 if (prog)
2512 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2513 return err;
2514}
2515
2516static u32 virtnet_xdp_query(struct net_device *dev)
2517{
2518 struct virtnet_info *vi = netdev_priv(dev);
2519 const struct bpf_prog *xdp_prog;
2520 int i;
2521
2522 for (i = 0; i < vi->max_queue_pairs; i++) {
2523 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2524 if (xdp_prog)
2525 return xdp_prog->aux->id;
2526 }
2527 return 0;
2528}
2529
2530static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2531{
2532 switch (xdp->command) {
2533 case XDP_SETUP_PROG:
2534 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2535 case XDP_QUERY_PROG:
2536 xdp->prog_id = virtnet_xdp_query(dev);
2537 return 0;
2538 default:
2539 return -EINVAL;
2540 }
2541}
2542
2543static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2544 size_t len)
2545{
2546 struct virtnet_info *vi = netdev_priv(dev);
2547 int ret;
2548
2549 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2550 return -EOPNOTSUPP;
2551
2552 ret = snprintf(buf, len, "sby");
2553 if (ret >= len)
2554 return -EOPNOTSUPP;
2555
2556 return 0;
2557}
2558
2559static int virtnet_set_features(struct net_device *dev,
2560 netdev_features_t features)
2561{
2562 struct virtnet_info *vi = netdev_priv(dev);
2563 u64 offloads;
2564 int err;
2565
2566 if ((dev->features ^ features) & NETIF_F_LRO) {
2567 if (vi->xdp_queue_pairs)
2568 return -EBUSY;
2569
2570 if (features & NETIF_F_LRO)
2571 offloads = vi->guest_offloads_capable;
2572 else
2573 offloads = 0;
2574
2575 err = virtnet_set_guest_offloads(vi, offloads);
2576 if (err)
2577 return err;
2578 vi->guest_offloads = offloads;
2579 }
2580
2581 return 0;
2582}
2583
2584static const struct net_device_ops virtnet_netdev = {
2585 .ndo_open = virtnet_open,
2586 .ndo_stop = virtnet_close,
2587 .ndo_start_xmit = start_xmit,
2588 .ndo_validate_addr = eth_validate_addr,
2589 .ndo_set_mac_address = virtnet_set_mac_address,
2590 .ndo_set_rx_mode = virtnet_set_rx_mode,
2591 .ndo_get_stats64 = virtnet_stats,
2592 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2593 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2594 .ndo_bpf = virtnet_xdp,
2595 .ndo_xdp_xmit = virtnet_xdp_xmit,
2596 .ndo_features_check = passthru_features_check,
2597 .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2598 .ndo_set_features = virtnet_set_features,
2599};
2600
2601static void virtnet_config_changed_work(struct work_struct *work)
2602{
2603 struct virtnet_info *vi =
2604 container_of(work, struct virtnet_info, config_work);
2605 u16 v;
2606
2607 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2608 struct virtio_net_config, status, &v) < 0)
2609 return;
2610
2611 if (v & VIRTIO_NET_S_ANNOUNCE) {
2612 netdev_notify_peers(vi->dev);
2613 virtnet_ack_link_announce(vi);
2614 }
2615
2616 /* Ignore unknown (future) status bits */
2617 v &= VIRTIO_NET_S_LINK_UP;
2618
2619 if (vi->status == v)
2620 return;
2621
2622 vi->status = v;
2623
2624 if (vi->status & VIRTIO_NET_S_LINK_UP) {
2625 virtnet_update_settings(vi);
2626 netif_carrier_on(vi->dev);
2627 netif_tx_wake_all_queues(vi->dev);
2628 } else {
2629 netif_carrier_off(vi->dev);
2630 netif_tx_stop_all_queues(vi->dev);
2631 }
2632}
2633
2634static void virtnet_config_changed(struct virtio_device *vdev)
2635{
2636 struct virtnet_info *vi = vdev->priv;
2637
2638 schedule_work(&vi->config_work);
2639}
2640
2641static void virtnet_free_queues(struct virtnet_info *vi)
2642{
2643 int i;
2644
2645 for (i = 0; i < vi->max_queue_pairs; i++) {
2646 napi_hash_del(&vi->rq[i].napi);
2647 netif_napi_del(&vi->rq[i].napi);
2648 netif_napi_del(&vi->sq[i].napi);
2649 }
2650
2651 /* We called napi_hash_del() before netif_napi_del(),
2652 * we need to respect an RCU grace period before freeing vi->rq
2653 */
2654 synchronize_net();
2655
2656 kfree(vi->rq);
2657 kfree(vi->sq);
2658 kfree(vi->ctrl);
2659}
2660
2661static void _free_receive_bufs(struct virtnet_info *vi)
2662{
2663 struct bpf_prog *old_prog;
2664 int i;
2665
2666 for (i = 0; i < vi->max_queue_pairs; i++) {
2667 while (vi->rq[i].pages)
2668 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2669
2670 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2671 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2672 if (old_prog)
2673 bpf_prog_put(old_prog);
2674 }
2675}
2676
2677static void free_receive_bufs(struct virtnet_info *vi)
2678{
2679 rtnl_lock();
2680 _free_receive_bufs(vi);
2681 rtnl_unlock();
2682}
2683
2684static void free_receive_page_frags(struct virtnet_info *vi)
2685{
2686 int i;
2687 for (i = 0; i < vi->max_queue_pairs; i++)
2688 if (vi->rq[i].alloc_frag.page)
2689 put_page(vi->rq[i].alloc_frag.page);
2690}
2691
2692static void free_unused_bufs(struct virtnet_info *vi)
2693{
2694 void *buf;
2695 int i;
2696
2697 for (i = 0; i < vi->max_queue_pairs; i++) {
2698 struct virtqueue *vq = vi->sq[i].vq;
2699 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2700 if (!is_xdp_frame(buf))
2701 dev_kfree_skb(buf);
2702 else
2703 xdp_return_frame(ptr_to_xdp(buf));
2704 }
2705 }
2706
2707 for (i = 0; i < vi->max_queue_pairs; i++) {
2708 struct virtqueue *vq = vi->rq[i].vq;
2709
2710 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2711 if (vi->mergeable_rx_bufs) {
2712 put_page(virt_to_head_page(buf));
2713 } else if (vi->big_packets) {
2714 give_pages(&vi->rq[i], buf);
2715 } else {
2716 put_page(virt_to_head_page(buf));
2717 }
2718 }
2719 }
2720}
2721
2722static void virtnet_del_vqs(struct virtnet_info *vi)
2723{
2724 struct virtio_device *vdev = vi->vdev;
2725
2726 virtnet_clean_affinity(vi);
2727
2728 vdev->config->del_vqs(vdev);
2729
2730 virtnet_free_queues(vi);
2731}
2732
2733/* How large should a single buffer be so a queue full of these can fit at
2734 * least one full packet?
2735 * Logic below assumes the mergeable buffer header is used.
2736 */
2737static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2738{
2739 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2740 unsigned int rq_size = virtqueue_get_vring_size(vq);
2741 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2742 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2743 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2744
2745 return max(max(min_buf_len, hdr_len) - hdr_len,
2746 (unsigned int)GOOD_PACKET_LEN);
2747}
2748
2749static int virtnet_find_vqs(struct virtnet_info *vi)
2750{
2751 vq_callback_t **callbacks;
2752 struct virtqueue **vqs;
2753 int ret = -ENOMEM;
2754 int i, total_vqs;
2755 const char **names;
2756 bool *ctx;
2757
2758 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2759 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2760 * possible control vq.
2761 */
2762 total_vqs = vi->max_queue_pairs * 2 +
2763 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2764
2765 /* Allocate space for find_vqs parameters */
2766 vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2767 if (!vqs)
2768 goto err_vq;
2769 callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2770 if (!callbacks)
2771 goto err_callback;
2772 names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2773 if (!names)
2774 goto err_names;
2775 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2776 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2777 if (!ctx)
2778 goto err_ctx;
2779 } else {
2780 ctx = NULL;
2781 }
2782
2783 /* Parameters for control virtqueue, if any */
2784 if (vi->has_cvq) {
2785 callbacks[total_vqs - 1] = NULL;
2786 names[total_vqs - 1] = "control";
2787 }
2788
2789 /* Allocate/initialize parameters for send/receive virtqueues */
2790 for (i = 0; i < vi->max_queue_pairs; i++) {
2791 callbacks[rxq2vq(i)] = skb_recv_done;
2792 callbacks[txq2vq(i)] = skb_xmit_done;
2793 sprintf(vi->rq[i].name, "input.%d", i);
2794 sprintf(vi->sq[i].name, "output.%d", i);
2795 names[rxq2vq(i)] = vi->rq[i].name;
2796 names[txq2vq(i)] = vi->sq[i].name;
2797 if (ctx)
2798 ctx[rxq2vq(i)] = true;
2799 }
2800
2801 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2802 names, ctx, NULL);
2803 if (ret)
2804 goto err_find;
2805
2806 if (vi->has_cvq) {
2807 vi->cvq = vqs[total_vqs - 1];
2808 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2809 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2810 }
2811
2812 for (i = 0; i < vi->max_queue_pairs; i++) {
2813 vi->rq[i].vq = vqs[rxq2vq(i)];
2814 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2815 vi->sq[i].vq = vqs[txq2vq(i)];
2816 }
2817
2818 /* run here: ret == 0. */
2819
2820
2821err_find:
2822 kfree(ctx);
2823err_ctx:
2824 kfree(names);
2825err_names:
2826 kfree(callbacks);
2827err_callback:
2828 kfree(vqs);
2829err_vq:
2830 return ret;
2831}
2832
2833static int virtnet_alloc_queues(struct virtnet_info *vi)
2834{
2835 int i;
2836
2837 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2838 if (!vi->ctrl)
2839 goto err_ctrl;
2840 vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2841 if (!vi->sq)
2842 goto err_sq;
2843 vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2844 if (!vi->rq)
2845 goto err_rq;
2846
2847 INIT_DELAYED_WORK(&vi->refill, refill_work);
2848 for (i = 0; i < vi->max_queue_pairs; i++) {
2849 vi->rq[i].pages = NULL;
2850 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2851 napi_weight);
2852 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2853 napi_tx ? napi_weight : 0);
2854
2855 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2856 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2857 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2858
2859 u64_stats_init(&vi->rq[i].stats.syncp);
2860 u64_stats_init(&vi->sq[i].stats.syncp);
2861 }
2862
2863 return 0;
2864
2865err_rq:
2866 kfree(vi->sq);
2867err_sq:
2868 kfree(vi->ctrl);
2869err_ctrl:
2870 return -ENOMEM;
2871}
2872
2873static int init_vqs(struct virtnet_info *vi)
2874{
2875 int ret;
2876
2877 /* Allocate send & receive queues */
2878 ret = virtnet_alloc_queues(vi);
2879 if (ret)
2880 goto err;
2881
2882 ret = virtnet_find_vqs(vi);
2883 if (ret)
2884 goto err_free;
2885
2886 get_online_cpus();
2887 virtnet_set_affinity(vi);
2888 put_online_cpus();
2889
2890 return 0;
2891
2892err_free:
2893 virtnet_free_queues(vi);
2894err:
2895 return ret;
2896}
2897
2898#ifdef CONFIG_SYSFS
2899static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2900 char *buf)
2901{
2902 struct virtnet_info *vi = netdev_priv(queue->dev);
2903 unsigned int queue_index = get_netdev_rx_queue_index(queue);
2904 unsigned int headroom = virtnet_get_headroom(vi);
2905 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2906 struct ewma_pkt_len *avg;
2907
2908 BUG_ON(queue_index >= vi->max_queue_pairs);
2909 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2910 return sprintf(buf, "%u\n",
2911 get_mergeable_buf_len(&vi->rq[queue_index], avg,
2912 SKB_DATA_ALIGN(headroom + tailroom)));
2913}
2914
2915static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2916 __ATTR_RO(mergeable_rx_buffer_size);
2917
2918static struct attribute *virtio_net_mrg_rx_attrs[] = {
2919 &mergeable_rx_buffer_size_attribute.attr,
2920 NULL
2921};
2922
2923static const struct attribute_group virtio_net_mrg_rx_group = {
2924 .name = "virtio_net",
2925 .attrs = virtio_net_mrg_rx_attrs
2926};
2927#endif
2928
2929static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2930 unsigned int fbit,
2931 const char *fname, const char *dname)
2932{
2933 if (!virtio_has_feature(vdev, fbit))
2934 return false;
2935
2936 dev_err(&vdev->dev, "device advertises feature %s but not %s",
2937 fname, dname);
2938
2939 return true;
2940}
2941
2942#define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2943 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2944
2945static bool virtnet_validate_features(struct virtio_device *vdev)
2946{
2947 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2948 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2949 "VIRTIO_NET_F_CTRL_VQ") ||
2950 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2951 "VIRTIO_NET_F_CTRL_VQ") ||
2952 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2953 "VIRTIO_NET_F_CTRL_VQ") ||
2954 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2955 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2956 "VIRTIO_NET_F_CTRL_VQ"))) {
2957 return false;
2958 }
2959
2960 return true;
2961}
2962
2963#define MIN_MTU ETH_MIN_MTU
2964#define MAX_MTU ETH_MAX_MTU
2965
2966static int virtnet_validate(struct virtio_device *vdev)
2967{
2968 if (!vdev->config->get) {
2969 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2970 __func__);
2971 return -EINVAL;
2972 }
2973
2974 if (!virtnet_validate_features(vdev))
2975 return -EINVAL;
2976
2977 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2978 int mtu = virtio_cread16(vdev,
2979 offsetof(struct virtio_net_config,
2980 mtu));
2981 if (mtu < MIN_MTU)
2982 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2983 }
2984
2985 return 0;
2986}
2987
2988static int virtnet_probe(struct virtio_device *vdev)
2989{
2990 int i, err = -ENOMEM;
2991 struct net_device *dev;
2992 struct virtnet_info *vi;
2993 u16 max_queue_pairs;
2994 int mtu;
2995
2996 /* Find if host supports multiqueue virtio_net device */
2997 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2998 struct virtio_net_config,
2999 max_virtqueue_pairs, &max_queue_pairs);
3000
3001 /* We need at least 2 queue's */
3002 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3003 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3004 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3005 max_queue_pairs = 1;
3006
3007 /* Allocate ourselves a network device with room for our info */
3008 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3009 if (!dev)
3010 return -ENOMEM;
3011
3012 /* Set up network device as normal. */
3013 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
3014 dev->netdev_ops = &virtnet_netdev;
3015 dev->features = NETIF_F_HIGHDMA;
3016
3017 dev->ethtool_ops = &virtnet_ethtool_ops;
3018 SET_NETDEV_DEV(dev, &vdev->dev);
3019
3020 /* Do we support "hardware" checksums? */
3021 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3022 /* This opens up the world of extra features. */
3023 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3024 if (csum)
3025 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3026
3027 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3028 dev->hw_features |= NETIF_F_TSO
3029 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3030 }
3031 /* Individual feature bits: what can host handle? */
3032 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3033 dev->hw_features |= NETIF_F_TSO;
3034 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3035 dev->hw_features |= NETIF_F_TSO6;
3036 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3037 dev->hw_features |= NETIF_F_TSO_ECN;
3038
3039 dev->features |= NETIF_F_GSO_ROBUST;
3040
3041 if (gso)
3042 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3043 /* (!csum && gso) case will be fixed by register_netdev() */
3044 }
3045 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3046 dev->features |= NETIF_F_RXCSUM;
3047 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3048 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3049 dev->features |= NETIF_F_LRO;
3050 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3051 dev->hw_features |= NETIF_F_LRO;
3052
3053 dev->vlan_features = dev->features;
3054
3055 /* MTU range: 68 - 65535 */
3056 dev->min_mtu = MIN_MTU;
3057 dev->max_mtu = MAX_MTU;
3058
3059 /* Configuration may specify what MAC to use. Otherwise random. */
3060 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
3061 virtio_cread_bytes(vdev,
3062 offsetof(struct virtio_net_config, mac),
3063 dev->dev_addr, dev->addr_len);
3064 else
3065 eth_hw_addr_random(dev);
3066
3067 /* Set up our device-specific information */
3068 vi = netdev_priv(dev);
3069 vi->dev = dev;
3070 vi->vdev = vdev;
3071 vdev->priv = vi;
3072
3073 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3074
3075 /* If we can receive ANY GSO packets, we must allocate large ones. */
3076 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3077 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3078 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3079 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3080 vi->big_packets = true;
3081
3082 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3083 vi->mergeable_rx_bufs = true;
3084
3085 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3086 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3087 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3088 else
3089 vi->hdr_len = sizeof(struct virtio_net_hdr);
3090
3091 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3092 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3093 vi->any_header_sg = true;
3094
3095 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3096 vi->has_cvq = true;
3097
3098 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3099 mtu = virtio_cread16(vdev,
3100 offsetof(struct virtio_net_config,
3101 mtu));
3102 if (mtu < dev->min_mtu) {
3103 /* Should never trigger: MTU was previously validated
3104 * in virtnet_validate.
3105 */
3106 dev_err(&vdev->dev,
3107 "device MTU appears to have changed it is now %d < %d",
3108 mtu, dev->min_mtu);
3109 goto free;
3110 }
3111
3112 dev->mtu = mtu;
3113 dev->max_mtu = mtu;
3114
3115 /* TODO: size buffers correctly in this case. */
3116 if (dev->mtu > ETH_DATA_LEN)
3117 vi->big_packets = true;
3118 }
3119
3120 if (vi->any_header_sg)
3121 dev->needed_headroom = vi->hdr_len;
3122
3123 /* Enable multiqueue by default */
3124 if (num_online_cpus() >= max_queue_pairs)
3125 vi->curr_queue_pairs = max_queue_pairs;
3126 else
3127 vi->curr_queue_pairs = num_online_cpus();
3128 vi->max_queue_pairs = max_queue_pairs;
3129
3130 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3131 err = init_vqs(vi);
3132 if (err)
3133 goto free;
3134
3135#ifdef CONFIG_SYSFS
3136 if (vi->mergeable_rx_bufs)
3137 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3138#endif
3139 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3140 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3141
3142 virtnet_init_settings(dev);
3143
3144 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3145 vi->failover = net_failover_create(vi->dev);
3146 if (IS_ERR(vi->failover)) {
3147 err = PTR_ERR(vi->failover);
3148 goto free_vqs;
3149 }
3150 }
3151
3152 err = register_netdev(dev);
3153 if (err) {
3154 pr_debug("virtio_net: registering device failed\n");
3155 goto free_failover;
3156 }
3157
3158 virtio_device_ready(vdev);
3159
3160 err = virtnet_cpu_notif_add(vi);
3161 if (err) {
3162 pr_debug("virtio_net: registering cpu notifier failed\n");
3163 goto free_unregister_netdev;
3164 }
3165
3166 virtnet_set_queues(vi, vi->curr_queue_pairs);
3167
3168 /* Assume link up if device can't report link status,
3169 otherwise get link status from config. */
3170 netif_carrier_off(dev);
3171 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3172 schedule_work(&vi->config_work);
3173 } else {
3174 vi->status = VIRTIO_NET_S_LINK_UP;
3175 virtnet_update_settings(vi);
3176 netif_carrier_on(dev);
3177 }
3178
3179 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3180 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3181 set_bit(guest_offloads[i], &vi->guest_offloads);
3182 vi->guest_offloads_capable = vi->guest_offloads;
3183
3184 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3185 dev->name, max_queue_pairs);
3186
3187 return 0;
3188
3189free_unregister_netdev:
3190 vi->vdev->config->reset(vdev);
3191
3192 unregister_netdev(dev);
3193free_failover:
3194 net_failover_destroy(vi->failover);
3195free_vqs:
3196 cancel_delayed_work_sync(&vi->refill);
3197 free_receive_page_frags(vi);
3198 virtnet_del_vqs(vi);
3199free:
3200 free_netdev(dev);
3201 return err;
3202}
3203
3204static void remove_vq_common(struct virtnet_info *vi)
3205{
3206 vi->vdev->config->reset(vi->vdev);
3207
3208 /* Free unused buffers in both send and recv, if any. */
3209 free_unused_bufs(vi);
3210
3211 free_receive_bufs(vi);
3212
3213 free_receive_page_frags(vi);
3214
3215 virtnet_del_vqs(vi);
3216}
3217
3218static void virtnet_remove(struct virtio_device *vdev)
3219{
3220 struct virtnet_info *vi = vdev->priv;
3221
3222 virtnet_cpu_notif_remove(vi);
3223
3224 /* Make sure no work handler is accessing the device. */
3225 flush_work(&vi->config_work);
3226
3227 unregister_netdev(vi->dev);
3228
3229 net_failover_destroy(vi->failover);
3230
3231 remove_vq_common(vi);
3232
3233 free_netdev(vi->dev);
3234}
3235
3236static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3237{
3238 struct virtnet_info *vi = vdev->priv;
3239
3240 virtnet_cpu_notif_remove(vi);
3241 virtnet_freeze_down(vdev);
3242 remove_vq_common(vi);
3243
3244 return 0;
3245}
3246
3247static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3248{
3249 struct virtnet_info *vi = vdev->priv;
3250 int err;
3251
3252 err = virtnet_restore_up(vdev);
3253 if (err)
3254 return err;
3255 virtnet_set_queues(vi, vi->curr_queue_pairs);
3256
3257 err = virtnet_cpu_notif_add(vi);
3258 if (err)
3259 return err;
3260
3261 return 0;
3262}
3263
3264static struct virtio_device_id id_table[] = {
3265 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3266 { 0 },
3267};
3268
3269#define VIRTNET_FEATURES \
3270 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3271 VIRTIO_NET_F_MAC, \
3272 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3273 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3274 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3275 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3276 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3277 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3278 VIRTIO_NET_F_CTRL_MAC_ADDR, \
3279 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3280 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3281
3282static unsigned int features[] = {
3283 VIRTNET_FEATURES,
3284};
3285
3286static unsigned int features_legacy[] = {
3287 VIRTNET_FEATURES,
3288 VIRTIO_NET_F_GSO,
3289 VIRTIO_F_ANY_LAYOUT,
3290};
3291
3292static struct virtio_driver virtio_net_driver = {
3293 .feature_table = features,
3294 .feature_table_size = ARRAY_SIZE(features),
3295 .feature_table_legacy = features_legacy,
3296 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3297 .driver.name = KBUILD_MODNAME,
3298 .driver.owner = THIS_MODULE,
3299 .id_table = id_table,
3300 .validate = virtnet_validate,
3301 .probe = virtnet_probe,
3302 .remove = virtnet_remove,
3303 .config_changed = virtnet_config_changed,
3304#ifdef CONFIG_PM_SLEEP
3305 .freeze = virtnet_freeze,
3306 .restore = virtnet_restore,
3307#endif
3308};
3309
3310static __init int virtio_net_driver_init(void)
3311{
3312 int ret;
3313
3314 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3315 virtnet_cpu_online,
3316 virtnet_cpu_down_prep);
3317 if (ret < 0)
3318 goto out;
3319 virtionet_online = ret;
3320 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3321 NULL, virtnet_cpu_dead);
3322 if (ret)
3323 goto err_dead;
3324
3325 ret = register_virtio_driver(&virtio_net_driver);
3326 if (ret)
3327 goto err_virtio;
3328 return 0;
3329err_virtio:
3330 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3331err_dead:
3332 cpuhp_remove_multi_state(virtionet_online);
3333out:
3334 return ret;
3335}
3336module_init(virtio_net_driver_init);
3337
3338static __exit void virtio_net_driver_exit(void)
3339{
3340 unregister_virtio_driver(&virtio_net_driver);
3341 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3342 cpuhp_remove_multi_state(virtionet_online);
3343}
3344module_exit(virtio_net_driver_exit);
3345
3346MODULE_DEVICE_TABLE(virtio, id_table);
3347MODULE_DESCRIPTION("Virtio network driver");
3348MODULE_LICENSE("GPL");