Loading...
1/* A network driver using virtio.
2 *
3 * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, see <http://www.gnu.org/licenses/>.
17 */
18//#define DEBUG
19#include <linux/netdevice.h>
20#include <linux/etherdevice.h>
21#include <linux/ethtool.h>
22#include <linux/module.h>
23#include <linux/virtio.h>
24#include <linux/virtio_net.h>
25#include <linux/bpf.h>
26#include <linux/bpf_trace.h>
27#include <linux/scatterlist.h>
28#include <linux/if_vlan.h>
29#include <linux/slab.h>
30#include <linux/cpu.h>
31#include <linux/average.h>
32#include <linux/filter.h>
33#include <net/route.h>
34#include <net/xdp.h>
35
36static int napi_weight = NAPI_POLL_WEIGHT;
37module_param(napi_weight, int, 0444);
38
39static bool csum = true, gso = true, napi_tx;
40module_param(csum, bool, 0444);
41module_param(gso, bool, 0444);
42module_param(napi_tx, bool, 0644);
43
44/* FIXME: MTU in config. */
45#define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
46#define GOOD_COPY_LEN 128
47
48#define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
49
50/* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
51#define VIRTIO_XDP_HEADROOM 256
52
53/* RX packet size EWMA. The average packet size is used to determine the packet
54 * buffer size when refilling RX rings. As the entire RX ring may be refilled
55 * at once, the weight is chosen so that the EWMA will be insensitive to short-
56 * term, transient changes in packet size.
57 */
58DECLARE_EWMA(pkt_len, 0, 64)
59
60#define VIRTNET_DRIVER_VERSION "1.0.0"
61
62static const unsigned long guest_offloads[] = {
63 VIRTIO_NET_F_GUEST_TSO4,
64 VIRTIO_NET_F_GUEST_TSO6,
65 VIRTIO_NET_F_GUEST_ECN,
66 VIRTIO_NET_F_GUEST_UFO
67};
68
69struct virtnet_stat_desc {
70 char desc[ETH_GSTRING_LEN];
71 size_t offset;
72};
73
74struct virtnet_sq_stats {
75 struct u64_stats_sync syncp;
76 u64 packets;
77 u64 bytes;
78};
79
80struct virtnet_rq_stats {
81 struct u64_stats_sync syncp;
82 u64 packets;
83 u64 bytes;
84};
85
86#define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m)
87#define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m)
88
89static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
90 { "packets", VIRTNET_SQ_STAT(packets) },
91 { "bytes", VIRTNET_SQ_STAT(bytes) },
92};
93
94static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
95 { "packets", VIRTNET_RQ_STAT(packets) },
96 { "bytes", VIRTNET_RQ_STAT(bytes) },
97};
98
99#define VIRTNET_SQ_STATS_LEN ARRAY_SIZE(virtnet_sq_stats_desc)
100#define VIRTNET_RQ_STATS_LEN ARRAY_SIZE(virtnet_rq_stats_desc)
101
102/* Internal representation of a send virtqueue */
103struct send_queue {
104 /* Virtqueue associated with this send _queue */
105 struct virtqueue *vq;
106
107 /* TX: fragments + linear part + virtio header */
108 struct scatterlist sg[MAX_SKB_FRAGS + 2];
109
110 /* Name of the send queue: output.$index */
111 char name[40];
112
113 struct virtnet_sq_stats stats;
114
115 struct napi_struct napi;
116};
117
118/* Internal representation of a receive virtqueue */
119struct receive_queue {
120 /* Virtqueue associated with this receive_queue */
121 struct virtqueue *vq;
122
123 struct napi_struct napi;
124
125 struct bpf_prog __rcu *xdp_prog;
126
127 struct virtnet_rq_stats stats;
128
129 /* Chain pages by the private ptr. */
130 struct page *pages;
131
132 /* Average packet length for mergeable receive buffers. */
133 struct ewma_pkt_len mrg_avg_pkt_len;
134
135 /* Page frag for packet buffer allocation. */
136 struct page_frag alloc_frag;
137
138 /* RX: fragments + linear part + virtio header */
139 struct scatterlist sg[MAX_SKB_FRAGS + 2];
140
141 /* Min single buffer size for mergeable buffers case. */
142 unsigned int min_buf_len;
143
144 /* Name of this receive queue: input.$index */
145 char name[40];
146
147 struct xdp_rxq_info xdp_rxq;
148};
149
150/* Control VQ buffers: protected by the rtnl lock */
151struct control_buf {
152 struct virtio_net_ctrl_hdr hdr;
153 virtio_net_ctrl_ack status;
154 struct virtio_net_ctrl_mq mq;
155 u8 promisc;
156 u8 allmulti;
157 __virtio16 vid;
158 __virtio64 offloads;
159};
160
161struct virtnet_info {
162 struct virtio_device *vdev;
163 struct virtqueue *cvq;
164 struct net_device *dev;
165 struct send_queue *sq;
166 struct receive_queue *rq;
167 unsigned int status;
168
169 /* Max # of queue pairs supported by the device */
170 u16 max_queue_pairs;
171
172 /* # of queue pairs currently used by the driver */
173 u16 curr_queue_pairs;
174
175 /* # of XDP queue pairs currently used by the driver */
176 u16 xdp_queue_pairs;
177
178 /* I like... big packets and I cannot lie! */
179 bool big_packets;
180
181 /* Host will merge rx buffers for big packets (shake it! shake it!) */
182 bool mergeable_rx_bufs;
183
184 /* Has control virtqueue */
185 bool has_cvq;
186
187 /* Host can handle any s/g split between our header and packet data */
188 bool any_header_sg;
189
190 /* Packet virtio header size */
191 u8 hdr_len;
192
193 /* Work struct for refilling if we run low on memory. */
194 struct delayed_work refill;
195
196 /* Work struct for config space updates */
197 struct work_struct config_work;
198
199 /* Does the affinity hint is set for virtqueues? */
200 bool affinity_hint_set;
201
202 /* CPU hotplug instances for online & dead */
203 struct hlist_node node;
204 struct hlist_node node_dead;
205
206 struct control_buf *ctrl;
207
208 /* Ethtool settings */
209 u8 duplex;
210 u32 speed;
211
212 unsigned long guest_offloads;
213};
214
215struct padded_vnet_hdr {
216 struct virtio_net_hdr_mrg_rxbuf hdr;
217 /*
218 * hdr is in a separate sg buffer, and data sg buffer shares same page
219 * with this header sg. This padding makes next sg 16 byte aligned
220 * after the header.
221 */
222 char padding[4];
223};
224
225/* Converting between virtqueue no. and kernel tx/rx queue no.
226 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
227 */
228static int vq2txq(struct virtqueue *vq)
229{
230 return (vq->index - 1) / 2;
231}
232
233static int txq2vq(int txq)
234{
235 return txq * 2 + 1;
236}
237
238static int vq2rxq(struct virtqueue *vq)
239{
240 return vq->index / 2;
241}
242
243static int rxq2vq(int rxq)
244{
245 return rxq * 2;
246}
247
248static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
249{
250 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
251}
252
253/*
254 * private is used to chain pages for big packets, put the whole
255 * most recent used list in the beginning for reuse
256 */
257static void give_pages(struct receive_queue *rq, struct page *page)
258{
259 struct page *end;
260
261 /* Find end of list, sew whole thing into vi->rq.pages. */
262 for (end = page; end->private; end = (struct page *)end->private);
263 end->private = (unsigned long)rq->pages;
264 rq->pages = page;
265}
266
267static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
268{
269 struct page *p = rq->pages;
270
271 if (p) {
272 rq->pages = (struct page *)p->private;
273 /* clear private here, it is used to chain pages */
274 p->private = 0;
275 } else
276 p = alloc_page(gfp_mask);
277 return p;
278}
279
280static void virtqueue_napi_schedule(struct napi_struct *napi,
281 struct virtqueue *vq)
282{
283 if (napi_schedule_prep(napi)) {
284 virtqueue_disable_cb(vq);
285 __napi_schedule(napi);
286 }
287}
288
289static void virtqueue_napi_complete(struct napi_struct *napi,
290 struct virtqueue *vq, int processed)
291{
292 int opaque;
293
294 opaque = virtqueue_enable_cb_prepare(vq);
295 if (napi_complete_done(napi, processed)) {
296 if (unlikely(virtqueue_poll(vq, opaque)))
297 virtqueue_napi_schedule(napi, vq);
298 } else {
299 virtqueue_disable_cb(vq);
300 }
301}
302
303static void skb_xmit_done(struct virtqueue *vq)
304{
305 struct virtnet_info *vi = vq->vdev->priv;
306 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
307
308 /* Suppress further interrupts. */
309 virtqueue_disable_cb(vq);
310
311 if (napi->weight)
312 virtqueue_napi_schedule(napi, vq);
313 else
314 /* We were probably waiting for more output buffers. */
315 netif_wake_subqueue(vi->dev, vq2txq(vq));
316}
317
318#define MRG_CTX_HEADER_SHIFT 22
319static void *mergeable_len_to_ctx(unsigned int truesize,
320 unsigned int headroom)
321{
322 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
323}
324
325static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
326{
327 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
328}
329
330static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
331{
332 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
333}
334
335/* Called from bottom half context */
336static struct sk_buff *page_to_skb(struct virtnet_info *vi,
337 struct receive_queue *rq,
338 struct page *page, unsigned int offset,
339 unsigned int len, unsigned int truesize)
340{
341 struct sk_buff *skb;
342 struct virtio_net_hdr_mrg_rxbuf *hdr;
343 unsigned int copy, hdr_len, hdr_padded_len;
344 char *p;
345
346 p = page_address(page) + offset;
347
348 /* copy small packet so we can reuse these pages for small data */
349 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
350 if (unlikely(!skb))
351 return NULL;
352
353 hdr = skb_vnet_hdr(skb);
354
355 hdr_len = vi->hdr_len;
356 if (vi->mergeable_rx_bufs)
357 hdr_padded_len = sizeof(*hdr);
358 else
359 hdr_padded_len = sizeof(struct padded_vnet_hdr);
360
361 memcpy(hdr, p, hdr_len);
362
363 len -= hdr_len;
364 offset += hdr_padded_len;
365 p += hdr_padded_len;
366
367 copy = len;
368 if (copy > skb_tailroom(skb))
369 copy = skb_tailroom(skb);
370 skb_put_data(skb, p, copy);
371
372 len -= copy;
373 offset += copy;
374
375 if (vi->mergeable_rx_bufs) {
376 if (len)
377 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
378 else
379 put_page(page);
380 return skb;
381 }
382
383 /*
384 * Verify that we can indeed put this data into a skb.
385 * This is here to handle cases when the device erroneously
386 * tries to receive more than is possible. This is usually
387 * the case of a broken device.
388 */
389 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
390 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
391 dev_kfree_skb(skb);
392 return NULL;
393 }
394 BUG_ON(offset >= PAGE_SIZE);
395 while (len) {
396 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
397 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
398 frag_size, truesize);
399 len -= frag_size;
400 page = (struct page *)page->private;
401 offset = 0;
402 }
403
404 if (page)
405 give_pages(rq, page);
406
407 return skb;
408}
409
410static void virtnet_xdp_flush(struct net_device *dev)
411{
412 struct virtnet_info *vi = netdev_priv(dev);
413 struct send_queue *sq;
414 unsigned int qp;
415
416 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
417 sq = &vi->sq[qp];
418
419 virtqueue_kick(sq->vq);
420}
421
422static bool __virtnet_xdp_xmit(struct virtnet_info *vi,
423 struct xdp_buff *xdp)
424{
425 struct virtio_net_hdr_mrg_rxbuf *hdr;
426 unsigned int len;
427 struct send_queue *sq;
428 unsigned int qp;
429 void *xdp_sent;
430 int err;
431
432 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
433 sq = &vi->sq[qp];
434
435 /* Free up any pending old buffers before queueing new ones. */
436 while ((xdp_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) {
437 struct page *sent_page = virt_to_head_page(xdp_sent);
438
439 put_page(sent_page);
440 }
441
442 xdp->data -= vi->hdr_len;
443 /* Zero header and leave csum up to XDP layers */
444 hdr = xdp->data;
445 memset(hdr, 0, vi->hdr_len);
446
447 sg_init_one(sq->sg, xdp->data, xdp->data_end - xdp->data);
448
449 err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp->data, GFP_ATOMIC);
450 if (unlikely(err))
451 return false; /* Caller handle free/refcnt */
452
453 return true;
454}
455
456static int virtnet_xdp_xmit(struct net_device *dev, struct xdp_buff *xdp)
457{
458 struct virtnet_info *vi = netdev_priv(dev);
459 struct receive_queue *rq = vi->rq;
460 struct bpf_prog *xdp_prog;
461 bool sent;
462
463 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
464 * indicate XDP resources have been successfully allocated.
465 */
466 xdp_prog = rcu_dereference(rq->xdp_prog);
467 if (!xdp_prog)
468 return -ENXIO;
469
470 sent = __virtnet_xdp_xmit(vi, xdp);
471 if (!sent)
472 return -ENOSPC;
473 return 0;
474}
475
476static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
477{
478 return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
479}
480
481/* We copy the packet for XDP in the following cases:
482 *
483 * 1) Packet is scattered across multiple rx buffers.
484 * 2) Headroom space is insufficient.
485 *
486 * This is inefficient but it's a temporary condition that
487 * we hit right after XDP is enabled and until queue is refilled
488 * with large buffers with sufficient headroom - so it should affect
489 * at most queue size packets.
490 * Afterwards, the conditions to enable
491 * XDP should preclude the underlying device from sending packets
492 * across multiple buffers (num_buf > 1), and we make sure buffers
493 * have enough headroom.
494 */
495static struct page *xdp_linearize_page(struct receive_queue *rq,
496 u16 *num_buf,
497 struct page *p,
498 int offset,
499 int page_off,
500 unsigned int *len)
501{
502 struct page *page = alloc_page(GFP_ATOMIC);
503
504 if (!page)
505 return NULL;
506
507 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
508 page_off += *len;
509
510 while (--*num_buf) {
511 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
512 unsigned int buflen;
513 void *buf;
514 int off;
515
516 buf = virtqueue_get_buf(rq->vq, &buflen);
517 if (unlikely(!buf))
518 goto err_buf;
519
520 p = virt_to_head_page(buf);
521 off = buf - page_address(p);
522
523 /* guard against a misconfigured or uncooperative backend that
524 * is sending packet larger than the MTU.
525 */
526 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
527 put_page(p);
528 goto err_buf;
529 }
530
531 memcpy(page_address(page) + page_off,
532 page_address(p) + off, buflen);
533 page_off += buflen;
534 put_page(p);
535 }
536
537 /* Headroom does not contribute to packet length */
538 *len = page_off - VIRTIO_XDP_HEADROOM;
539 return page;
540err_buf:
541 __free_pages(page, 0);
542 return NULL;
543}
544
545static struct sk_buff *receive_small(struct net_device *dev,
546 struct virtnet_info *vi,
547 struct receive_queue *rq,
548 void *buf, void *ctx,
549 unsigned int len,
550 bool *xdp_xmit)
551{
552 struct sk_buff *skb;
553 struct bpf_prog *xdp_prog;
554 unsigned int xdp_headroom = (unsigned long)ctx;
555 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
556 unsigned int headroom = vi->hdr_len + header_offset;
557 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
558 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
559 struct page *page = virt_to_head_page(buf);
560 unsigned int delta = 0;
561 struct page *xdp_page;
562 bool sent;
563 int err;
564
565 len -= vi->hdr_len;
566
567 rcu_read_lock();
568 xdp_prog = rcu_dereference(rq->xdp_prog);
569 if (xdp_prog) {
570 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
571 struct xdp_buff xdp;
572 void *orig_data;
573 u32 act;
574
575 if (unlikely(hdr->hdr.gso_type))
576 goto err_xdp;
577
578 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
579 int offset = buf - page_address(page) + header_offset;
580 unsigned int tlen = len + vi->hdr_len;
581 u16 num_buf = 1;
582
583 xdp_headroom = virtnet_get_headroom(vi);
584 header_offset = VIRTNET_RX_PAD + xdp_headroom;
585 headroom = vi->hdr_len + header_offset;
586 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
587 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
588 xdp_page = xdp_linearize_page(rq, &num_buf, page,
589 offset, header_offset,
590 &tlen);
591 if (!xdp_page)
592 goto err_xdp;
593
594 buf = page_address(xdp_page);
595 put_page(page);
596 page = xdp_page;
597 }
598
599 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
600 xdp.data = xdp.data_hard_start + xdp_headroom;
601 xdp_set_data_meta_invalid(&xdp);
602 xdp.data_end = xdp.data + len;
603 xdp.rxq = &rq->xdp_rxq;
604 orig_data = xdp.data;
605 act = bpf_prog_run_xdp(xdp_prog, &xdp);
606
607 switch (act) {
608 case XDP_PASS:
609 /* Recalculate length in case bpf program changed it */
610 delta = orig_data - xdp.data;
611 break;
612 case XDP_TX:
613 sent = __virtnet_xdp_xmit(vi, &xdp);
614 if (unlikely(!sent)) {
615 trace_xdp_exception(vi->dev, xdp_prog, act);
616 goto err_xdp;
617 }
618 *xdp_xmit = true;
619 rcu_read_unlock();
620 goto xdp_xmit;
621 case XDP_REDIRECT:
622 err = xdp_do_redirect(dev, &xdp, xdp_prog);
623 if (err)
624 goto err_xdp;
625 *xdp_xmit = true;
626 rcu_read_unlock();
627 goto xdp_xmit;
628 default:
629 bpf_warn_invalid_xdp_action(act);
630 case XDP_ABORTED:
631 trace_xdp_exception(vi->dev, xdp_prog, act);
632 case XDP_DROP:
633 goto err_xdp;
634 }
635 }
636 rcu_read_unlock();
637
638 skb = build_skb(buf, buflen);
639 if (!skb) {
640 put_page(page);
641 goto err;
642 }
643 skb_reserve(skb, headroom - delta);
644 skb_put(skb, len + delta);
645 if (!delta) {
646 buf += header_offset;
647 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
648 } /* keep zeroed vnet hdr since packet was changed by bpf */
649
650err:
651 return skb;
652
653err_xdp:
654 rcu_read_unlock();
655 dev->stats.rx_dropped++;
656 put_page(page);
657xdp_xmit:
658 return NULL;
659}
660
661static struct sk_buff *receive_big(struct net_device *dev,
662 struct virtnet_info *vi,
663 struct receive_queue *rq,
664 void *buf,
665 unsigned int len)
666{
667 struct page *page = buf;
668 struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
669
670 if (unlikely(!skb))
671 goto err;
672
673 return skb;
674
675err:
676 dev->stats.rx_dropped++;
677 give_pages(rq, page);
678 return NULL;
679}
680
681static struct sk_buff *receive_mergeable(struct net_device *dev,
682 struct virtnet_info *vi,
683 struct receive_queue *rq,
684 void *buf,
685 void *ctx,
686 unsigned int len,
687 bool *xdp_xmit)
688{
689 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
690 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
691 struct page *page = virt_to_head_page(buf);
692 int offset = buf - page_address(page);
693 struct sk_buff *head_skb, *curr_skb;
694 struct bpf_prog *xdp_prog;
695 unsigned int truesize;
696 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
697 bool sent;
698 int err;
699
700 head_skb = NULL;
701
702 rcu_read_lock();
703 xdp_prog = rcu_dereference(rq->xdp_prog);
704 if (xdp_prog) {
705 struct page *xdp_page;
706 struct xdp_buff xdp;
707 void *data;
708 u32 act;
709
710 /* Transient failure which in theory could occur if
711 * in-flight packets from before XDP was enabled reach
712 * the receive path after XDP is loaded.
713 */
714 if (unlikely(hdr->hdr.gso_type))
715 goto err_xdp;
716
717 /* This happens when rx buffer size is underestimated
718 * or headroom is not enough because of the buffer
719 * was refilled before XDP is set. This should only
720 * happen for the first several packets, so we don't
721 * care much about its performance.
722 */
723 if (unlikely(num_buf > 1 ||
724 headroom < virtnet_get_headroom(vi))) {
725 /* linearize data for XDP */
726 xdp_page = xdp_linearize_page(rq, &num_buf,
727 page, offset,
728 VIRTIO_XDP_HEADROOM,
729 &len);
730 if (!xdp_page)
731 goto err_xdp;
732 offset = VIRTIO_XDP_HEADROOM;
733 } else {
734 xdp_page = page;
735 }
736
737 /* Allow consuming headroom but reserve enough space to push
738 * the descriptor on if we get an XDP_TX return code.
739 */
740 data = page_address(xdp_page) + offset;
741 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
742 xdp.data = data + vi->hdr_len;
743 xdp_set_data_meta_invalid(&xdp);
744 xdp.data_end = xdp.data + (len - vi->hdr_len);
745 xdp.rxq = &rq->xdp_rxq;
746
747 act = bpf_prog_run_xdp(xdp_prog, &xdp);
748
749 switch (act) {
750 case XDP_PASS:
751 /* recalculate offset to account for any header
752 * adjustments. Note other cases do not build an
753 * skb and avoid using offset
754 */
755 offset = xdp.data -
756 page_address(xdp_page) - vi->hdr_len;
757
758 /* We can only create skb based on xdp_page. */
759 if (unlikely(xdp_page != page)) {
760 rcu_read_unlock();
761 put_page(page);
762 head_skb = page_to_skb(vi, rq, xdp_page,
763 offset, len, PAGE_SIZE);
764 return head_skb;
765 }
766 break;
767 case XDP_TX:
768 sent = __virtnet_xdp_xmit(vi, &xdp);
769 if (unlikely(!sent)) {
770 trace_xdp_exception(vi->dev, xdp_prog, act);
771 if (unlikely(xdp_page != page))
772 put_page(xdp_page);
773 goto err_xdp;
774 }
775 *xdp_xmit = true;
776 if (unlikely(xdp_page != page))
777 put_page(page);
778 rcu_read_unlock();
779 goto xdp_xmit;
780 case XDP_REDIRECT:
781 err = xdp_do_redirect(dev, &xdp, xdp_prog);
782 if (err) {
783 if (unlikely(xdp_page != page))
784 put_page(xdp_page);
785 goto err_xdp;
786 }
787 *xdp_xmit = true;
788 if (unlikely(xdp_page != page))
789 put_page(page);
790 rcu_read_unlock();
791 goto xdp_xmit;
792 default:
793 bpf_warn_invalid_xdp_action(act);
794 case XDP_ABORTED:
795 trace_xdp_exception(vi->dev, xdp_prog, act);
796 case XDP_DROP:
797 if (unlikely(xdp_page != page))
798 __free_pages(xdp_page, 0);
799 goto err_xdp;
800 }
801 }
802 rcu_read_unlock();
803
804 truesize = mergeable_ctx_to_truesize(ctx);
805 if (unlikely(len > truesize)) {
806 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
807 dev->name, len, (unsigned long)ctx);
808 dev->stats.rx_length_errors++;
809 goto err_skb;
810 }
811
812 head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
813 curr_skb = head_skb;
814
815 if (unlikely(!curr_skb))
816 goto err_skb;
817 while (--num_buf) {
818 int num_skb_frags;
819
820 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
821 if (unlikely(!buf)) {
822 pr_debug("%s: rx error: %d buffers out of %d missing\n",
823 dev->name, num_buf,
824 virtio16_to_cpu(vi->vdev,
825 hdr->num_buffers));
826 dev->stats.rx_length_errors++;
827 goto err_buf;
828 }
829
830 page = virt_to_head_page(buf);
831
832 truesize = mergeable_ctx_to_truesize(ctx);
833 if (unlikely(len > truesize)) {
834 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
835 dev->name, len, (unsigned long)ctx);
836 dev->stats.rx_length_errors++;
837 goto err_skb;
838 }
839
840 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
841 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
842 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
843
844 if (unlikely(!nskb))
845 goto err_skb;
846 if (curr_skb == head_skb)
847 skb_shinfo(curr_skb)->frag_list = nskb;
848 else
849 curr_skb->next = nskb;
850 curr_skb = nskb;
851 head_skb->truesize += nskb->truesize;
852 num_skb_frags = 0;
853 }
854 if (curr_skb != head_skb) {
855 head_skb->data_len += len;
856 head_skb->len += len;
857 head_skb->truesize += truesize;
858 }
859 offset = buf - page_address(page);
860 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
861 put_page(page);
862 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
863 len, truesize);
864 } else {
865 skb_add_rx_frag(curr_skb, num_skb_frags, page,
866 offset, len, truesize);
867 }
868 }
869
870 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
871 return head_skb;
872
873err_xdp:
874 rcu_read_unlock();
875err_skb:
876 put_page(page);
877 while (num_buf-- > 1) {
878 buf = virtqueue_get_buf(rq->vq, &len);
879 if (unlikely(!buf)) {
880 pr_debug("%s: rx error: %d buffers missing\n",
881 dev->name, num_buf);
882 dev->stats.rx_length_errors++;
883 break;
884 }
885 page = virt_to_head_page(buf);
886 put_page(page);
887 }
888err_buf:
889 dev->stats.rx_dropped++;
890 dev_kfree_skb(head_skb);
891xdp_xmit:
892 return NULL;
893}
894
895static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
896 void *buf, unsigned int len, void **ctx, bool *xdp_xmit)
897{
898 struct net_device *dev = vi->dev;
899 struct sk_buff *skb;
900 struct virtio_net_hdr_mrg_rxbuf *hdr;
901 int ret;
902
903 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
904 pr_debug("%s: short packet %i\n", dev->name, len);
905 dev->stats.rx_length_errors++;
906 if (vi->mergeable_rx_bufs) {
907 put_page(virt_to_head_page(buf));
908 } else if (vi->big_packets) {
909 give_pages(rq, buf);
910 } else {
911 put_page(virt_to_head_page(buf));
912 }
913 return 0;
914 }
915
916 if (vi->mergeable_rx_bufs)
917 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit);
918 else if (vi->big_packets)
919 skb = receive_big(dev, vi, rq, buf, len);
920 else
921 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit);
922
923 if (unlikely(!skb))
924 return 0;
925
926 hdr = skb_vnet_hdr(skb);
927
928 ret = skb->len;
929
930 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
931 skb->ip_summed = CHECKSUM_UNNECESSARY;
932
933 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
934 virtio_is_little_endian(vi->vdev))) {
935 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
936 dev->name, hdr->hdr.gso_type,
937 hdr->hdr.gso_size);
938 goto frame_err;
939 }
940
941 skb->protocol = eth_type_trans(skb, dev);
942 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
943 ntohs(skb->protocol), skb->len, skb->pkt_type);
944
945 napi_gro_receive(&rq->napi, skb);
946 return ret;
947
948frame_err:
949 dev->stats.rx_frame_errors++;
950 dev_kfree_skb(skb);
951 return 0;
952}
953
954/* Unlike mergeable buffers, all buffers are allocated to the
955 * same size, except for the headroom. For this reason we do
956 * not need to use mergeable_len_to_ctx here - it is enough
957 * to store the headroom as the context ignoring the truesize.
958 */
959static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
960 gfp_t gfp)
961{
962 struct page_frag *alloc_frag = &rq->alloc_frag;
963 char *buf;
964 unsigned int xdp_headroom = virtnet_get_headroom(vi);
965 void *ctx = (void *)(unsigned long)xdp_headroom;
966 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
967 int err;
968
969 len = SKB_DATA_ALIGN(len) +
970 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
971 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
972 return -ENOMEM;
973
974 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
975 get_page(alloc_frag->page);
976 alloc_frag->offset += len;
977 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
978 vi->hdr_len + GOOD_PACKET_LEN);
979 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
980 if (err < 0)
981 put_page(virt_to_head_page(buf));
982 return err;
983}
984
985static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
986 gfp_t gfp)
987{
988 struct page *first, *list = NULL;
989 char *p;
990 int i, err, offset;
991
992 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
993
994 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
995 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
996 first = get_a_page(rq, gfp);
997 if (!first) {
998 if (list)
999 give_pages(rq, list);
1000 return -ENOMEM;
1001 }
1002 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1003
1004 /* chain new page in list head to match sg */
1005 first->private = (unsigned long)list;
1006 list = first;
1007 }
1008
1009 first = get_a_page(rq, gfp);
1010 if (!first) {
1011 give_pages(rq, list);
1012 return -ENOMEM;
1013 }
1014 p = page_address(first);
1015
1016 /* rq->sg[0], rq->sg[1] share the same page */
1017 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1018 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1019
1020 /* rq->sg[1] for data packet, from offset */
1021 offset = sizeof(struct padded_vnet_hdr);
1022 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1023
1024 /* chain first in list head */
1025 first->private = (unsigned long)list;
1026 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1027 first, gfp);
1028 if (err < 0)
1029 give_pages(rq, first);
1030
1031 return err;
1032}
1033
1034static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1035 struct ewma_pkt_len *avg_pkt_len,
1036 unsigned int room)
1037{
1038 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1039 unsigned int len;
1040
1041 if (room)
1042 return PAGE_SIZE - room;
1043
1044 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1045 rq->min_buf_len, PAGE_SIZE - hdr_len);
1046
1047 return ALIGN(len, L1_CACHE_BYTES);
1048}
1049
1050static int add_recvbuf_mergeable(struct virtnet_info *vi,
1051 struct receive_queue *rq, gfp_t gfp)
1052{
1053 struct page_frag *alloc_frag = &rq->alloc_frag;
1054 unsigned int headroom = virtnet_get_headroom(vi);
1055 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1056 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1057 char *buf;
1058 void *ctx;
1059 int err;
1060 unsigned int len, hole;
1061
1062 /* Extra tailroom is needed to satisfy XDP's assumption. This
1063 * means rx frags coalescing won't work, but consider we've
1064 * disabled GSO for XDP, it won't be a big issue.
1065 */
1066 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1067 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1068 return -ENOMEM;
1069
1070 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1071 buf += headroom; /* advance address leaving hole at front of pkt */
1072 get_page(alloc_frag->page);
1073 alloc_frag->offset += len + room;
1074 hole = alloc_frag->size - alloc_frag->offset;
1075 if (hole < len + room) {
1076 /* To avoid internal fragmentation, if there is very likely not
1077 * enough space for another buffer, add the remaining space to
1078 * the current buffer.
1079 */
1080 len += hole;
1081 alloc_frag->offset += hole;
1082 }
1083
1084 sg_init_one(rq->sg, buf, len);
1085 ctx = mergeable_len_to_ctx(len, headroom);
1086 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1087 if (err < 0)
1088 put_page(virt_to_head_page(buf));
1089
1090 return err;
1091}
1092
1093/*
1094 * Returns false if we couldn't fill entirely (OOM).
1095 *
1096 * Normally run in the receive path, but can also be run from ndo_open
1097 * before we're receiving packets, or from refill_work which is
1098 * careful to disable receiving (using napi_disable).
1099 */
1100static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1101 gfp_t gfp)
1102{
1103 int err;
1104 bool oom;
1105
1106 do {
1107 if (vi->mergeable_rx_bufs)
1108 err = add_recvbuf_mergeable(vi, rq, gfp);
1109 else if (vi->big_packets)
1110 err = add_recvbuf_big(vi, rq, gfp);
1111 else
1112 err = add_recvbuf_small(vi, rq, gfp);
1113
1114 oom = err == -ENOMEM;
1115 if (err)
1116 break;
1117 } while (rq->vq->num_free);
1118 virtqueue_kick(rq->vq);
1119 return !oom;
1120}
1121
1122static void skb_recv_done(struct virtqueue *rvq)
1123{
1124 struct virtnet_info *vi = rvq->vdev->priv;
1125 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1126
1127 virtqueue_napi_schedule(&rq->napi, rvq);
1128}
1129
1130static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1131{
1132 napi_enable(napi);
1133
1134 /* If all buffers were filled by other side before we napi_enabled, we
1135 * won't get another interrupt, so process any outstanding packets now.
1136 * Call local_bh_enable after to trigger softIRQ processing.
1137 */
1138 local_bh_disable();
1139 virtqueue_napi_schedule(napi, vq);
1140 local_bh_enable();
1141}
1142
1143static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1144 struct virtqueue *vq,
1145 struct napi_struct *napi)
1146{
1147 if (!napi->weight)
1148 return;
1149
1150 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1151 * enable the feature if this is likely affine with the transmit path.
1152 */
1153 if (!vi->affinity_hint_set) {
1154 napi->weight = 0;
1155 return;
1156 }
1157
1158 return virtnet_napi_enable(vq, napi);
1159}
1160
1161static void virtnet_napi_tx_disable(struct napi_struct *napi)
1162{
1163 if (napi->weight)
1164 napi_disable(napi);
1165}
1166
1167static void refill_work(struct work_struct *work)
1168{
1169 struct virtnet_info *vi =
1170 container_of(work, struct virtnet_info, refill.work);
1171 bool still_empty;
1172 int i;
1173
1174 for (i = 0; i < vi->curr_queue_pairs; i++) {
1175 struct receive_queue *rq = &vi->rq[i];
1176
1177 napi_disable(&rq->napi);
1178 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1179 virtnet_napi_enable(rq->vq, &rq->napi);
1180
1181 /* In theory, this can happen: if we don't get any buffers in
1182 * we will *never* try to fill again.
1183 */
1184 if (still_empty)
1185 schedule_delayed_work(&vi->refill, HZ/2);
1186 }
1187}
1188
1189static int virtnet_receive(struct receive_queue *rq, int budget, bool *xdp_xmit)
1190{
1191 struct virtnet_info *vi = rq->vq->vdev->priv;
1192 unsigned int len, received = 0, bytes = 0;
1193 void *buf;
1194
1195 if (!vi->big_packets || vi->mergeable_rx_bufs) {
1196 void *ctx;
1197
1198 while (received < budget &&
1199 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1200 bytes += receive_buf(vi, rq, buf, len, ctx, xdp_xmit);
1201 received++;
1202 }
1203 } else {
1204 while (received < budget &&
1205 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1206 bytes += receive_buf(vi, rq, buf, len, NULL, xdp_xmit);
1207 received++;
1208 }
1209 }
1210
1211 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1212 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1213 schedule_delayed_work(&vi->refill, 0);
1214 }
1215
1216 u64_stats_update_begin(&rq->stats.syncp);
1217 rq->stats.bytes += bytes;
1218 rq->stats.packets += received;
1219 u64_stats_update_end(&rq->stats.syncp);
1220
1221 return received;
1222}
1223
1224static void free_old_xmit_skbs(struct send_queue *sq)
1225{
1226 struct sk_buff *skb;
1227 unsigned int len;
1228 unsigned int packets = 0;
1229 unsigned int bytes = 0;
1230
1231 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1232 pr_debug("Sent skb %p\n", skb);
1233
1234 bytes += skb->len;
1235 packets++;
1236
1237 dev_consume_skb_any(skb);
1238 }
1239
1240 /* Avoid overhead when no packets have been processed
1241 * happens when called speculatively from start_xmit.
1242 */
1243 if (!packets)
1244 return;
1245
1246 u64_stats_update_begin(&sq->stats.syncp);
1247 sq->stats.bytes += bytes;
1248 sq->stats.packets += packets;
1249 u64_stats_update_end(&sq->stats.syncp);
1250}
1251
1252static void virtnet_poll_cleantx(struct receive_queue *rq)
1253{
1254 struct virtnet_info *vi = rq->vq->vdev->priv;
1255 unsigned int index = vq2rxq(rq->vq);
1256 struct send_queue *sq = &vi->sq[index];
1257 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1258
1259 if (!sq->napi.weight)
1260 return;
1261
1262 if (__netif_tx_trylock(txq)) {
1263 free_old_xmit_skbs(sq);
1264 __netif_tx_unlock(txq);
1265 }
1266
1267 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1268 netif_tx_wake_queue(txq);
1269}
1270
1271static int virtnet_poll(struct napi_struct *napi, int budget)
1272{
1273 struct receive_queue *rq =
1274 container_of(napi, struct receive_queue, napi);
1275 struct virtnet_info *vi = rq->vq->vdev->priv;
1276 struct send_queue *sq;
1277 unsigned int received, qp;
1278 bool xdp_xmit = false;
1279
1280 virtnet_poll_cleantx(rq);
1281
1282 received = virtnet_receive(rq, budget, &xdp_xmit);
1283
1284 /* Out of packets? */
1285 if (received < budget)
1286 virtqueue_napi_complete(napi, rq->vq, received);
1287
1288 if (xdp_xmit) {
1289 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs +
1290 smp_processor_id();
1291 sq = &vi->sq[qp];
1292 virtqueue_kick(sq->vq);
1293 xdp_do_flush_map();
1294 }
1295
1296 return received;
1297}
1298
1299static int virtnet_open(struct net_device *dev)
1300{
1301 struct virtnet_info *vi = netdev_priv(dev);
1302 int i, err;
1303
1304 for (i = 0; i < vi->max_queue_pairs; i++) {
1305 if (i < vi->curr_queue_pairs)
1306 /* Make sure we have some buffers: if oom use wq. */
1307 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1308 schedule_delayed_work(&vi->refill, 0);
1309
1310 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i);
1311 if (err < 0)
1312 return err;
1313
1314 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1315 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1316 }
1317
1318 return 0;
1319}
1320
1321static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1322{
1323 struct send_queue *sq = container_of(napi, struct send_queue, napi);
1324 struct virtnet_info *vi = sq->vq->vdev->priv;
1325 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
1326
1327 __netif_tx_lock(txq, raw_smp_processor_id());
1328 free_old_xmit_skbs(sq);
1329 __netif_tx_unlock(txq);
1330
1331 virtqueue_napi_complete(napi, sq->vq, 0);
1332
1333 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1334 netif_tx_wake_queue(txq);
1335
1336 return 0;
1337}
1338
1339static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1340{
1341 struct virtio_net_hdr_mrg_rxbuf *hdr;
1342 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1343 struct virtnet_info *vi = sq->vq->vdev->priv;
1344 int num_sg;
1345 unsigned hdr_len = vi->hdr_len;
1346 bool can_push;
1347
1348 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1349
1350 can_push = vi->any_header_sg &&
1351 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1352 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1353 /* Even if we can, don't push here yet as this would skew
1354 * csum_start offset below. */
1355 if (can_push)
1356 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1357 else
1358 hdr = skb_vnet_hdr(skb);
1359
1360 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1361 virtio_is_little_endian(vi->vdev), false))
1362 BUG();
1363
1364 if (vi->mergeable_rx_bufs)
1365 hdr->num_buffers = 0;
1366
1367 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1368 if (can_push) {
1369 __skb_push(skb, hdr_len);
1370 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1371 if (unlikely(num_sg < 0))
1372 return num_sg;
1373 /* Pull header back to avoid skew in tx bytes calculations. */
1374 __skb_pull(skb, hdr_len);
1375 } else {
1376 sg_set_buf(sq->sg, hdr, hdr_len);
1377 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1378 if (unlikely(num_sg < 0))
1379 return num_sg;
1380 num_sg++;
1381 }
1382 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1383}
1384
1385static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1386{
1387 struct virtnet_info *vi = netdev_priv(dev);
1388 int qnum = skb_get_queue_mapping(skb);
1389 struct send_queue *sq = &vi->sq[qnum];
1390 int err;
1391 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1392 bool kick = !skb->xmit_more;
1393 bool use_napi = sq->napi.weight;
1394
1395 /* Free up any pending old buffers before queueing new ones. */
1396 free_old_xmit_skbs(sq);
1397
1398 if (use_napi && kick)
1399 virtqueue_enable_cb_delayed(sq->vq);
1400
1401 /* timestamp packet in software */
1402 skb_tx_timestamp(skb);
1403
1404 /* Try to transmit */
1405 err = xmit_skb(sq, skb);
1406
1407 /* This should not happen! */
1408 if (unlikely(err)) {
1409 dev->stats.tx_fifo_errors++;
1410 if (net_ratelimit())
1411 dev_warn(&dev->dev,
1412 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1413 dev->stats.tx_dropped++;
1414 dev_kfree_skb_any(skb);
1415 return NETDEV_TX_OK;
1416 }
1417
1418 /* Don't wait up for transmitted skbs to be freed. */
1419 if (!use_napi) {
1420 skb_orphan(skb);
1421 nf_reset(skb);
1422 }
1423
1424 /* If running out of space, stop queue to avoid getting packets that we
1425 * are then unable to transmit.
1426 * An alternative would be to force queuing layer to requeue the skb by
1427 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1428 * returned in a normal path of operation: it means that driver is not
1429 * maintaining the TX queue stop/start state properly, and causes
1430 * the stack to do a non-trivial amount of useless work.
1431 * Since most packets only take 1 or 2 ring slots, stopping the queue
1432 * early means 16 slots are typically wasted.
1433 */
1434 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1435 netif_stop_subqueue(dev, qnum);
1436 if (!use_napi &&
1437 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1438 /* More just got used, free them then recheck. */
1439 free_old_xmit_skbs(sq);
1440 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1441 netif_start_subqueue(dev, qnum);
1442 virtqueue_disable_cb(sq->vq);
1443 }
1444 }
1445 }
1446
1447 if (kick || netif_xmit_stopped(txq))
1448 virtqueue_kick(sq->vq);
1449
1450 return NETDEV_TX_OK;
1451}
1452
1453/*
1454 * Send command via the control virtqueue and check status. Commands
1455 * supported by the hypervisor, as indicated by feature bits, should
1456 * never fail unless improperly formatted.
1457 */
1458static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1459 struct scatterlist *out)
1460{
1461 struct scatterlist *sgs[4], hdr, stat;
1462 unsigned out_num = 0, tmp;
1463
1464 /* Caller should know better */
1465 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1466
1467 vi->ctrl->status = ~0;
1468 vi->ctrl->hdr.class = class;
1469 vi->ctrl->hdr.cmd = cmd;
1470 /* Add header */
1471 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1472 sgs[out_num++] = &hdr;
1473
1474 if (out)
1475 sgs[out_num++] = out;
1476
1477 /* Add return status. */
1478 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1479 sgs[out_num] = &stat;
1480
1481 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1482 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1483
1484 if (unlikely(!virtqueue_kick(vi->cvq)))
1485 return vi->ctrl->status == VIRTIO_NET_OK;
1486
1487 /* Spin for a response, the kick causes an ioport write, trapping
1488 * into the hypervisor, so the request should be handled immediately.
1489 */
1490 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1491 !virtqueue_is_broken(vi->cvq))
1492 cpu_relax();
1493
1494 return vi->ctrl->status == VIRTIO_NET_OK;
1495}
1496
1497static int virtnet_set_mac_address(struct net_device *dev, void *p)
1498{
1499 struct virtnet_info *vi = netdev_priv(dev);
1500 struct virtio_device *vdev = vi->vdev;
1501 int ret;
1502 struct sockaddr *addr;
1503 struct scatterlist sg;
1504
1505 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1506 if (!addr)
1507 return -ENOMEM;
1508
1509 ret = eth_prepare_mac_addr_change(dev, addr);
1510 if (ret)
1511 goto out;
1512
1513 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1514 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1515 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1516 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1517 dev_warn(&vdev->dev,
1518 "Failed to set mac address by vq command.\n");
1519 ret = -EINVAL;
1520 goto out;
1521 }
1522 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1523 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1524 unsigned int i;
1525
1526 /* Naturally, this has an atomicity problem. */
1527 for (i = 0; i < dev->addr_len; i++)
1528 virtio_cwrite8(vdev,
1529 offsetof(struct virtio_net_config, mac) +
1530 i, addr->sa_data[i]);
1531 }
1532
1533 eth_commit_mac_addr_change(dev, p);
1534 ret = 0;
1535
1536out:
1537 kfree(addr);
1538 return ret;
1539}
1540
1541static void virtnet_stats(struct net_device *dev,
1542 struct rtnl_link_stats64 *tot)
1543{
1544 struct virtnet_info *vi = netdev_priv(dev);
1545 unsigned int start;
1546 int i;
1547
1548 for (i = 0; i < vi->max_queue_pairs; i++) {
1549 u64 tpackets, tbytes, rpackets, rbytes;
1550 struct receive_queue *rq = &vi->rq[i];
1551 struct send_queue *sq = &vi->sq[i];
1552
1553 do {
1554 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1555 tpackets = sq->stats.packets;
1556 tbytes = sq->stats.bytes;
1557 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1558
1559 do {
1560 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1561 rpackets = rq->stats.packets;
1562 rbytes = rq->stats.bytes;
1563 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1564
1565 tot->rx_packets += rpackets;
1566 tot->tx_packets += tpackets;
1567 tot->rx_bytes += rbytes;
1568 tot->tx_bytes += tbytes;
1569 }
1570
1571 tot->tx_dropped = dev->stats.tx_dropped;
1572 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1573 tot->rx_dropped = dev->stats.rx_dropped;
1574 tot->rx_length_errors = dev->stats.rx_length_errors;
1575 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1576}
1577
1578#ifdef CONFIG_NET_POLL_CONTROLLER
1579static void virtnet_netpoll(struct net_device *dev)
1580{
1581 struct virtnet_info *vi = netdev_priv(dev);
1582 int i;
1583
1584 for (i = 0; i < vi->curr_queue_pairs; i++)
1585 napi_schedule(&vi->rq[i].napi);
1586}
1587#endif
1588
1589static void virtnet_ack_link_announce(struct virtnet_info *vi)
1590{
1591 rtnl_lock();
1592 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1593 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1594 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1595 rtnl_unlock();
1596}
1597
1598static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1599{
1600 struct scatterlist sg;
1601 struct net_device *dev = vi->dev;
1602
1603 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1604 return 0;
1605
1606 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1607 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1608
1609 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1610 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1611 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1612 queue_pairs);
1613 return -EINVAL;
1614 } else {
1615 vi->curr_queue_pairs = queue_pairs;
1616 /* virtnet_open() will refill when device is going to up. */
1617 if (dev->flags & IFF_UP)
1618 schedule_delayed_work(&vi->refill, 0);
1619 }
1620
1621 return 0;
1622}
1623
1624static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1625{
1626 int err;
1627
1628 rtnl_lock();
1629 err = _virtnet_set_queues(vi, queue_pairs);
1630 rtnl_unlock();
1631 return err;
1632}
1633
1634static int virtnet_close(struct net_device *dev)
1635{
1636 struct virtnet_info *vi = netdev_priv(dev);
1637 int i;
1638
1639 /* Make sure refill_work doesn't re-enable napi! */
1640 cancel_delayed_work_sync(&vi->refill);
1641
1642 for (i = 0; i < vi->max_queue_pairs; i++) {
1643 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1644 napi_disable(&vi->rq[i].napi);
1645 virtnet_napi_tx_disable(&vi->sq[i].napi);
1646 }
1647
1648 return 0;
1649}
1650
1651static void virtnet_set_rx_mode(struct net_device *dev)
1652{
1653 struct virtnet_info *vi = netdev_priv(dev);
1654 struct scatterlist sg[2];
1655 struct virtio_net_ctrl_mac *mac_data;
1656 struct netdev_hw_addr *ha;
1657 int uc_count;
1658 int mc_count;
1659 void *buf;
1660 int i;
1661
1662 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1663 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1664 return;
1665
1666 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1667 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1668
1669 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1670
1671 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1672 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1673 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1674 vi->ctrl->promisc ? "en" : "dis");
1675
1676 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1677
1678 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1679 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1680 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1681 vi->ctrl->allmulti ? "en" : "dis");
1682
1683 uc_count = netdev_uc_count(dev);
1684 mc_count = netdev_mc_count(dev);
1685 /* MAC filter - use one buffer for both lists */
1686 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1687 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1688 mac_data = buf;
1689 if (!buf)
1690 return;
1691
1692 sg_init_table(sg, 2);
1693
1694 /* Store the unicast list and count in the front of the buffer */
1695 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1696 i = 0;
1697 netdev_for_each_uc_addr(ha, dev)
1698 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1699
1700 sg_set_buf(&sg[0], mac_data,
1701 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1702
1703 /* multicast list and count fill the end */
1704 mac_data = (void *)&mac_data->macs[uc_count][0];
1705
1706 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1707 i = 0;
1708 netdev_for_each_mc_addr(ha, dev)
1709 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1710
1711 sg_set_buf(&sg[1], mac_data,
1712 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1713
1714 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1715 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1716 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1717
1718 kfree(buf);
1719}
1720
1721static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1722 __be16 proto, u16 vid)
1723{
1724 struct virtnet_info *vi = netdev_priv(dev);
1725 struct scatterlist sg;
1726
1727 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1728 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1729
1730 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1731 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1732 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1733 return 0;
1734}
1735
1736static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1737 __be16 proto, u16 vid)
1738{
1739 struct virtnet_info *vi = netdev_priv(dev);
1740 struct scatterlist sg;
1741
1742 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1743 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1744
1745 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1746 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1747 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1748 return 0;
1749}
1750
1751static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1752{
1753 int i;
1754
1755 if (vi->affinity_hint_set) {
1756 for (i = 0; i < vi->max_queue_pairs; i++) {
1757 virtqueue_set_affinity(vi->rq[i].vq, -1);
1758 virtqueue_set_affinity(vi->sq[i].vq, -1);
1759 }
1760
1761 vi->affinity_hint_set = false;
1762 }
1763}
1764
1765static void virtnet_set_affinity(struct virtnet_info *vi)
1766{
1767 int i;
1768 int cpu;
1769
1770 /* In multiqueue mode, when the number of cpu is equal to the number of
1771 * queue pairs, we let the queue pairs to be private to one cpu by
1772 * setting the affinity hint to eliminate the contention.
1773 */
1774 if (vi->curr_queue_pairs == 1 ||
1775 vi->max_queue_pairs != num_online_cpus()) {
1776 virtnet_clean_affinity(vi, -1);
1777 return;
1778 }
1779
1780 i = 0;
1781 for_each_online_cpu(cpu) {
1782 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1783 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1784 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1785 i++;
1786 }
1787
1788 vi->affinity_hint_set = true;
1789}
1790
1791static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1792{
1793 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1794 node);
1795 virtnet_set_affinity(vi);
1796 return 0;
1797}
1798
1799static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1800{
1801 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1802 node_dead);
1803 virtnet_set_affinity(vi);
1804 return 0;
1805}
1806
1807static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1808{
1809 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1810 node);
1811
1812 virtnet_clean_affinity(vi, cpu);
1813 return 0;
1814}
1815
1816static enum cpuhp_state virtionet_online;
1817
1818static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1819{
1820 int ret;
1821
1822 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1823 if (ret)
1824 return ret;
1825 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1826 &vi->node_dead);
1827 if (!ret)
1828 return ret;
1829 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1830 return ret;
1831}
1832
1833static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1834{
1835 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1836 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1837 &vi->node_dead);
1838}
1839
1840static void virtnet_get_ringparam(struct net_device *dev,
1841 struct ethtool_ringparam *ring)
1842{
1843 struct virtnet_info *vi = netdev_priv(dev);
1844
1845 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1846 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1847 ring->rx_pending = ring->rx_max_pending;
1848 ring->tx_pending = ring->tx_max_pending;
1849}
1850
1851
1852static void virtnet_get_drvinfo(struct net_device *dev,
1853 struct ethtool_drvinfo *info)
1854{
1855 struct virtnet_info *vi = netdev_priv(dev);
1856 struct virtio_device *vdev = vi->vdev;
1857
1858 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1859 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1860 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1861
1862}
1863
1864/* TODO: Eliminate OOO packets during switching */
1865static int virtnet_set_channels(struct net_device *dev,
1866 struct ethtool_channels *channels)
1867{
1868 struct virtnet_info *vi = netdev_priv(dev);
1869 u16 queue_pairs = channels->combined_count;
1870 int err;
1871
1872 /* We don't support separate rx/tx channels.
1873 * We don't allow setting 'other' channels.
1874 */
1875 if (channels->rx_count || channels->tx_count || channels->other_count)
1876 return -EINVAL;
1877
1878 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1879 return -EINVAL;
1880
1881 /* For now we don't support modifying channels while XDP is loaded
1882 * also when XDP is loaded all RX queues have XDP programs so we only
1883 * need to check a single RX queue.
1884 */
1885 if (vi->rq[0].xdp_prog)
1886 return -EINVAL;
1887
1888 get_online_cpus();
1889 err = _virtnet_set_queues(vi, queue_pairs);
1890 if (!err) {
1891 netif_set_real_num_tx_queues(dev, queue_pairs);
1892 netif_set_real_num_rx_queues(dev, queue_pairs);
1893
1894 virtnet_set_affinity(vi);
1895 }
1896 put_online_cpus();
1897
1898 return err;
1899}
1900
1901static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
1902{
1903 struct virtnet_info *vi = netdev_priv(dev);
1904 char *p = (char *)data;
1905 unsigned int i, j;
1906
1907 switch (stringset) {
1908 case ETH_SS_STATS:
1909 for (i = 0; i < vi->curr_queue_pairs; i++) {
1910 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
1911 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s",
1912 i, virtnet_rq_stats_desc[j].desc);
1913 p += ETH_GSTRING_LEN;
1914 }
1915 }
1916
1917 for (i = 0; i < vi->curr_queue_pairs; i++) {
1918 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
1919 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s",
1920 i, virtnet_sq_stats_desc[j].desc);
1921 p += ETH_GSTRING_LEN;
1922 }
1923 }
1924 break;
1925 }
1926}
1927
1928static int virtnet_get_sset_count(struct net_device *dev, int sset)
1929{
1930 struct virtnet_info *vi = netdev_priv(dev);
1931
1932 switch (sset) {
1933 case ETH_SS_STATS:
1934 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
1935 VIRTNET_SQ_STATS_LEN);
1936 default:
1937 return -EOPNOTSUPP;
1938 }
1939}
1940
1941static void virtnet_get_ethtool_stats(struct net_device *dev,
1942 struct ethtool_stats *stats, u64 *data)
1943{
1944 struct virtnet_info *vi = netdev_priv(dev);
1945 unsigned int idx = 0, start, i, j;
1946 const u8 *stats_base;
1947 size_t offset;
1948
1949 for (i = 0; i < vi->curr_queue_pairs; i++) {
1950 struct receive_queue *rq = &vi->rq[i];
1951
1952 stats_base = (u8 *)&rq->stats;
1953 do {
1954 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1955 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
1956 offset = virtnet_rq_stats_desc[j].offset;
1957 data[idx + j] = *(u64 *)(stats_base + offset);
1958 }
1959 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1960 idx += VIRTNET_RQ_STATS_LEN;
1961 }
1962
1963 for (i = 0; i < vi->curr_queue_pairs; i++) {
1964 struct send_queue *sq = &vi->sq[i];
1965
1966 stats_base = (u8 *)&sq->stats;
1967 do {
1968 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1969 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
1970 offset = virtnet_sq_stats_desc[j].offset;
1971 data[idx + j] = *(u64 *)(stats_base + offset);
1972 }
1973 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1974 idx += VIRTNET_SQ_STATS_LEN;
1975 }
1976}
1977
1978static void virtnet_get_channels(struct net_device *dev,
1979 struct ethtool_channels *channels)
1980{
1981 struct virtnet_info *vi = netdev_priv(dev);
1982
1983 channels->combined_count = vi->curr_queue_pairs;
1984 channels->max_combined = vi->max_queue_pairs;
1985 channels->max_other = 0;
1986 channels->rx_count = 0;
1987 channels->tx_count = 0;
1988 channels->other_count = 0;
1989}
1990
1991/* Check if the user is trying to change anything besides speed/duplex */
1992static bool
1993virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
1994{
1995 struct ethtool_link_ksettings diff1 = *cmd;
1996 struct ethtool_link_ksettings diff2 = {};
1997
1998 /* cmd is always set so we need to clear it, validate the port type
1999 * and also without autonegotiation we can ignore advertising
2000 */
2001 diff1.base.speed = 0;
2002 diff2.base.port = PORT_OTHER;
2003 ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
2004 diff1.base.duplex = 0;
2005 diff1.base.cmd = 0;
2006 diff1.base.link_mode_masks_nwords = 0;
2007
2008 return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
2009 bitmap_empty(diff1.link_modes.supported,
2010 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2011 bitmap_empty(diff1.link_modes.advertising,
2012 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2013 bitmap_empty(diff1.link_modes.lp_advertising,
2014 __ETHTOOL_LINK_MODE_MASK_NBITS);
2015}
2016
2017static int virtnet_set_link_ksettings(struct net_device *dev,
2018 const struct ethtool_link_ksettings *cmd)
2019{
2020 struct virtnet_info *vi = netdev_priv(dev);
2021 u32 speed;
2022
2023 speed = cmd->base.speed;
2024 /* don't allow custom speed and duplex */
2025 if (!ethtool_validate_speed(speed) ||
2026 !ethtool_validate_duplex(cmd->base.duplex) ||
2027 !virtnet_validate_ethtool_cmd(cmd))
2028 return -EINVAL;
2029 vi->speed = speed;
2030 vi->duplex = cmd->base.duplex;
2031
2032 return 0;
2033}
2034
2035static int virtnet_get_link_ksettings(struct net_device *dev,
2036 struct ethtool_link_ksettings *cmd)
2037{
2038 struct virtnet_info *vi = netdev_priv(dev);
2039
2040 cmd->base.speed = vi->speed;
2041 cmd->base.duplex = vi->duplex;
2042 cmd->base.port = PORT_OTHER;
2043
2044 return 0;
2045}
2046
2047static void virtnet_init_settings(struct net_device *dev)
2048{
2049 struct virtnet_info *vi = netdev_priv(dev);
2050
2051 vi->speed = SPEED_UNKNOWN;
2052 vi->duplex = DUPLEX_UNKNOWN;
2053}
2054
2055static void virtnet_update_settings(struct virtnet_info *vi)
2056{
2057 u32 speed;
2058 u8 duplex;
2059
2060 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2061 return;
2062
2063 speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config,
2064 speed));
2065 if (ethtool_validate_speed(speed))
2066 vi->speed = speed;
2067 duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config,
2068 duplex));
2069 if (ethtool_validate_duplex(duplex))
2070 vi->duplex = duplex;
2071}
2072
2073static const struct ethtool_ops virtnet_ethtool_ops = {
2074 .get_drvinfo = virtnet_get_drvinfo,
2075 .get_link = ethtool_op_get_link,
2076 .get_ringparam = virtnet_get_ringparam,
2077 .get_strings = virtnet_get_strings,
2078 .get_sset_count = virtnet_get_sset_count,
2079 .get_ethtool_stats = virtnet_get_ethtool_stats,
2080 .set_channels = virtnet_set_channels,
2081 .get_channels = virtnet_get_channels,
2082 .get_ts_info = ethtool_op_get_ts_info,
2083 .get_link_ksettings = virtnet_get_link_ksettings,
2084 .set_link_ksettings = virtnet_set_link_ksettings,
2085};
2086
2087static void virtnet_freeze_down(struct virtio_device *vdev)
2088{
2089 struct virtnet_info *vi = vdev->priv;
2090 int i;
2091
2092 /* Make sure no work handler is accessing the device */
2093 flush_work(&vi->config_work);
2094
2095 netif_device_detach(vi->dev);
2096 netif_tx_disable(vi->dev);
2097 cancel_delayed_work_sync(&vi->refill);
2098
2099 if (netif_running(vi->dev)) {
2100 for (i = 0; i < vi->max_queue_pairs; i++) {
2101 napi_disable(&vi->rq[i].napi);
2102 virtnet_napi_tx_disable(&vi->sq[i].napi);
2103 }
2104 }
2105}
2106
2107static int init_vqs(struct virtnet_info *vi);
2108
2109static int virtnet_restore_up(struct virtio_device *vdev)
2110{
2111 struct virtnet_info *vi = vdev->priv;
2112 int err, i;
2113
2114 err = init_vqs(vi);
2115 if (err)
2116 return err;
2117
2118 virtio_device_ready(vdev);
2119
2120 if (netif_running(vi->dev)) {
2121 for (i = 0; i < vi->curr_queue_pairs; i++)
2122 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2123 schedule_delayed_work(&vi->refill, 0);
2124
2125 for (i = 0; i < vi->max_queue_pairs; i++) {
2126 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2127 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2128 &vi->sq[i].napi);
2129 }
2130 }
2131
2132 netif_device_attach(vi->dev);
2133 return err;
2134}
2135
2136static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2137{
2138 struct scatterlist sg;
2139 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2140
2141 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2142
2143 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2144 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2145 dev_warn(&vi->dev->dev, "Fail to set guest offload. \n");
2146 return -EINVAL;
2147 }
2148
2149 return 0;
2150}
2151
2152static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2153{
2154 u64 offloads = 0;
2155
2156 if (!vi->guest_offloads)
2157 return 0;
2158
2159 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
2160 offloads = 1ULL << VIRTIO_NET_F_GUEST_CSUM;
2161
2162 return virtnet_set_guest_offloads(vi, offloads);
2163}
2164
2165static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2166{
2167 u64 offloads = vi->guest_offloads;
2168
2169 if (!vi->guest_offloads)
2170 return 0;
2171 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
2172 offloads |= 1ULL << VIRTIO_NET_F_GUEST_CSUM;
2173
2174 return virtnet_set_guest_offloads(vi, offloads);
2175}
2176
2177static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2178 struct netlink_ext_ack *extack)
2179{
2180 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2181 struct virtnet_info *vi = netdev_priv(dev);
2182 struct bpf_prog *old_prog;
2183 u16 xdp_qp = 0, curr_qp;
2184 int i, err;
2185
2186 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2187 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2188 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2189 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2190 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO))) {
2191 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first");
2192 return -EOPNOTSUPP;
2193 }
2194
2195 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2196 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2197 return -EINVAL;
2198 }
2199
2200 if (dev->mtu > max_sz) {
2201 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2202 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2203 return -EINVAL;
2204 }
2205
2206 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2207 if (prog)
2208 xdp_qp = nr_cpu_ids;
2209
2210 /* XDP requires extra queues for XDP_TX */
2211 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2212 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2213 netdev_warn(dev, "request %i queues but max is %i\n",
2214 curr_qp + xdp_qp, vi->max_queue_pairs);
2215 return -ENOMEM;
2216 }
2217
2218 if (prog) {
2219 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
2220 if (IS_ERR(prog))
2221 return PTR_ERR(prog);
2222 }
2223
2224 /* Make sure NAPI is not using any XDP TX queues for RX. */
2225 if (netif_running(dev))
2226 for (i = 0; i < vi->max_queue_pairs; i++)
2227 napi_disable(&vi->rq[i].napi);
2228
2229 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2230 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2231 if (err)
2232 goto err;
2233 vi->xdp_queue_pairs = xdp_qp;
2234
2235 for (i = 0; i < vi->max_queue_pairs; i++) {
2236 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2237 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2238 if (i == 0) {
2239 if (!old_prog)
2240 virtnet_clear_guest_offloads(vi);
2241 if (!prog)
2242 virtnet_restore_guest_offloads(vi);
2243 }
2244 if (old_prog)
2245 bpf_prog_put(old_prog);
2246 if (netif_running(dev))
2247 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2248 }
2249
2250 return 0;
2251
2252err:
2253 for (i = 0; i < vi->max_queue_pairs; i++)
2254 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2255 if (prog)
2256 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2257 return err;
2258}
2259
2260static u32 virtnet_xdp_query(struct net_device *dev)
2261{
2262 struct virtnet_info *vi = netdev_priv(dev);
2263 const struct bpf_prog *xdp_prog;
2264 int i;
2265
2266 for (i = 0; i < vi->max_queue_pairs; i++) {
2267 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2268 if (xdp_prog)
2269 return xdp_prog->aux->id;
2270 }
2271 return 0;
2272}
2273
2274static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2275{
2276 switch (xdp->command) {
2277 case XDP_SETUP_PROG:
2278 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2279 case XDP_QUERY_PROG:
2280 xdp->prog_id = virtnet_xdp_query(dev);
2281 xdp->prog_attached = !!xdp->prog_id;
2282 return 0;
2283 default:
2284 return -EINVAL;
2285 }
2286}
2287
2288static const struct net_device_ops virtnet_netdev = {
2289 .ndo_open = virtnet_open,
2290 .ndo_stop = virtnet_close,
2291 .ndo_start_xmit = start_xmit,
2292 .ndo_validate_addr = eth_validate_addr,
2293 .ndo_set_mac_address = virtnet_set_mac_address,
2294 .ndo_set_rx_mode = virtnet_set_rx_mode,
2295 .ndo_get_stats64 = virtnet_stats,
2296 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2297 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2298#ifdef CONFIG_NET_POLL_CONTROLLER
2299 .ndo_poll_controller = virtnet_netpoll,
2300#endif
2301 .ndo_bpf = virtnet_xdp,
2302 .ndo_xdp_xmit = virtnet_xdp_xmit,
2303 .ndo_xdp_flush = virtnet_xdp_flush,
2304 .ndo_features_check = passthru_features_check,
2305};
2306
2307static void virtnet_config_changed_work(struct work_struct *work)
2308{
2309 struct virtnet_info *vi =
2310 container_of(work, struct virtnet_info, config_work);
2311 u16 v;
2312
2313 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2314 struct virtio_net_config, status, &v) < 0)
2315 return;
2316
2317 if (v & VIRTIO_NET_S_ANNOUNCE) {
2318 netdev_notify_peers(vi->dev);
2319 virtnet_ack_link_announce(vi);
2320 }
2321
2322 /* Ignore unknown (future) status bits */
2323 v &= VIRTIO_NET_S_LINK_UP;
2324
2325 if (vi->status == v)
2326 return;
2327
2328 vi->status = v;
2329
2330 if (vi->status & VIRTIO_NET_S_LINK_UP) {
2331 virtnet_update_settings(vi);
2332 netif_carrier_on(vi->dev);
2333 netif_tx_wake_all_queues(vi->dev);
2334 } else {
2335 netif_carrier_off(vi->dev);
2336 netif_tx_stop_all_queues(vi->dev);
2337 }
2338}
2339
2340static void virtnet_config_changed(struct virtio_device *vdev)
2341{
2342 struct virtnet_info *vi = vdev->priv;
2343
2344 schedule_work(&vi->config_work);
2345}
2346
2347static void virtnet_free_queues(struct virtnet_info *vi)
2348{
2349 int i;
2350
2351 for (i = 0; i < vi->max_queue_pairs; i++) {
2352 napi_hash_del(&vi->rq[i].napi);
2353 netif_napi_del(&vi->rq[i].napi);
2354 netif_napi_del(&vi->sq[i].napi);
2355 }
2356
2357 /* We called napi_hash_del() before netif_napi_del(),
2358 * we need to respect an RCU grace period before freeing vi->rq
2359 */
2360 synchronize_net();
2361
2362 kfree(vi->rq);
2363 kfree(vi->sq);
2364 kfree(vi->ctrl);
2365}
2366
2367static void _free_receive_bufs(struct virtnet_info *vi)
2368{
2369 struct bpf_prog *old_prog;
2370 int i;
2371
2372 for (i = 0; i < vi->max_queue_pairs; i++) {
2373 while (vi->rq[i].pages)
2374 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2375
2376 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2377 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2378 if (old_prog)
2379 bpf_prog_put(old_prog);
2380 }
2381}
2382
2383static void free_receive_bufs(struct virtnet_info *vi)
2384{
2385 rtnl_lock();
2386 _free_receive_bufs(vi);
2387 rtnl_unlock();
2388}
2389
2390static void free_receive_page_frags(struct virtnet_info *vi)
2391{
2392 int i;
2393 for (i = 0; i < vi->max_queue_pairs; i++)
2394 if (vi->rq[i].alloc_frag.page)
2395 put_page(vi->rq[i].alloc_frag.page);
2396}
2397
2398static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
2399{
2400 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
2401 return false;
2402 else if (q < vi->curr_queue_pairs)
2403 return true;
2404 else
2405 return false;
2406}
2407
2408static void free_unused_bufs(struct virtnet_info *vi)
2409{
2410 void *buf;
2411 int i;
2412
2413 for (i = 0; i < vi->max_queue_pairs; i++) {
2414 struct virtqueue *vq = vi->sq[i].vq;
2415 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2416 if (!is_xdp_raw_buffer_queue(vi, i))
2417 dev_kfree_skb(buf);
2418 else
2419 put_page(virt_to_head_page(buf));
2420 }
2421 }
2422
2423 for (i = 0; i < vi->max_queue_pairs; i++) {
2424 struct virtqueue *vq = vi->rq[i].vq;
2425
2426 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2427 if (vi->mergeable_rx_bufs) {
2428 put_page(virt_to_head_page(buf));
2429 } else if (vi->big_packets) {
2430 give_pages(&vi->rq[i], buf);
2431 } else {
2432 put_page(virt_to_head_page(buf));
2433 }
2434 }
2435 }
2436}
2437
2438static void virtnet_del_vqs(struct virtnet_info *vi)
2439{
2440 struct virtio_device *vdev = vi->vdev;
2441
2442 virtnet_clean_affinity(vi, -1);
2443
2444 vdev->config->del_vqs(vdev);
2445
2446 virtnet_free_queues(vi);
2447}
2448
2449/* How large should a single buffer be so a queue full of these can fit at
2450 * least one full packet?
2451 * Logic below assumes the mergeable buffer header is used.
2452 */
2453static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2454{
2455 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2456 unsigned int rq_size = virtqueue_get_vring_size(vq);
2457 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2458 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2459 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2460
2461 return max(max(min_buf_len, hdr_len) - hdr_len,
2462 (unsigned int)GOOD_PACKET_LEN);
2463}
2464
2465static int virtnet_find_vqs(struct virtnet_info *vi)
2466{
2467 vq_callback_t **callbacks;
2468 struct virtqueue **vqs;
2469 int ret = -ENOMEM;
2470 int i, total_vqs;
2471 const char **names;
2472 bool *ctx;
2473
2474 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2475 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2476 * possible control vq.
2477 */
2478 total_vqs = vi->max_queue_pairs * 2 +
2479 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2480
2481 /* Allocate space for find_vqs parameters */
2482 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
2483 if (!vqs)
2484 goto err_vq;
2485 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
2486 if (!callbacks)
2487 goto err_callback;
2488 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
2489 if (!names)
2490 goto err_names;
2491 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2492 ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL);
2493 if (!ctx)
2494 goto err_ctx;
2495 } else {
2496 ctx = NULL;
2497 }
2498
2499 /* Parameters for control virtqueue, if any */
2500 if (vi->has_cvq) {
2501 callbacks[total_vqs - 1] = NULL;
2502 names[total_vqs - 1] = "control";
2503 }
2504
2505 /* Allocate/initialize parameters for send/receive virtqueues */
2506 for (i = 0; i < vi->max_queue_pairs; i++) {
2507 callbacks[rxq2vq(i)] = skb_recv_done;
2508 callbacks[txq2vq(i)] = skb_xmit_done;
2509 sprintf(vi->rq[i].name, "input.%d", i);
2510 sprintf(vi->sq[i].name, "output.%d", i);
2511 names[rxq2vq(i)] = vi->rq[i].name;
2512 names[txq2vq(i)] = vi->sq[i].name;
2513 if (ctx)
2514 ctx[rxq2vq(i)] = true;
2515 }
2516
2517 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2518 names, ctx, NULL);
2519 if (ret)
2520 goto err_find;
2521
2522 if (vi->has_cvq) {
2523 vi->cvq = vqs[total_vqs - 1];
2524 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2525 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2526 }
2527
2528 for (i = 0; i < vi->max_queue_pairs; i++) {
2529 vi->rq[i].vq = vqs[rxq2vq(i)];
2530 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2531 vi->sq[i].vq = vqs[txq2vq(i)];
2532 }
2533
2534 kfree(names);
2535 kfree(callbacks);
2536 kfree(vqs);
2537 kfree(ctx);
2538
2539 return 0;
2540
2541err_find:
2542 kfree(ctx);
2543err_ctx:
2544 kfree(names);
2545err_names:
2546 kfree(callbacks);
2547err_callback:
2548 kfree(vqs);
2549err_vq:
2550 return ret;
2551}
2552
2553static int virtnet_alloc_queues(struct virtnet_info *vi)
2554{
2555 int i;
2556
2557 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2558 if (!vi->ctrl)
2559 goto err_ctrl;
2560 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2561 if (!vi->sq)
2562 goto err_sq;
2563 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2564 if (!vi->rq)
2565 goto err_rq;
2566
2567 INIT_DELAYED_WORK(&vi->refill, refill_work);
2568 for (i = 0; i < vi->max_queue_pairs; i++) {
2569 vi->rq[i].pages = NULL;
2570 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2571 napi_weight);
2572 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2573 napi_tx ? napi_weight : 0);
2574
2575 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2576 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2577 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2578
2579 u64_stats_init(&vi->rq[i].stats.syncp);
2580 u64_stats_init(&vi->sq[i].stats.syncp);
2581 }
2582
2583 return 0;
2584
2585err_rq:
2586 kfree(vi->sq);
2587err_sq:
2588 kfree(vi->ctrl);
2589err_ctrl:
2590 return -ENOMEM;
2591}
2592
2593static int init_vqs(struct virtnet_info *vi)
2594{
2595 int ret;
2596
2597 /* Allocate send & receive queues */
2598 ret = virtnet_alloc_queues(vi);
2599 if (ret)
2600 goto err;
2601
2602 ret = virtnet_find_vqs(vi);
2603 if (ret)
2604 goto err_free;
2605
2606 get_online_cpus();
2607 virtnet_set_affinity(vi);
2608 put_online_cpus();
2609
2610 return 0;
2611
2612err_free:
2613 virtnet_free_queues(vi);
2614err:
2615 return ret;
2616}
2617
2618#ifdef CONFIG_SYSFS
2619static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2620 char *buf)
2621{
2622 struct virtnet_info *vi = netdev_priv(queue->dev);
2623 unsigned int queue_index = get_netdev_rx_queue_index(queue);
2624 unsigned int headroom = virtnet_get_headroom(vi);
2625 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2626 struct ewma_pkt_len *avg;
2627
2628 BUG_ON(queue_index >= vi->max_queue_pairs);
2629 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2630 return sprintf(buf, "%u\n",
2631 get_mergeable_buf_len(&vi->rq[queue_index], avg,
2632 SKB_DATA_ALIGN(headroom + tailroom)));
2633}
2634
2635static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2636 __ATTR_RO(mergeable_rx_buffer_size);
2637
2638static struct attribute *virtio_net_mrg_rx_attrs[] = {
2639 &mergeable_rx_buffer_size_attribute.attr,
2640 NULL
2641};
2642
2643static const struct attribute_group virtio_net_mrg_rx_group = {
2644 .name = "virtio_net",
2645 .attrs = virtio_net_mrg_rx_attrs
2646};
2647#endif
2648
2649static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2650 unsigned int fbit,
2651 const char *fname, const char *dname)
2652{
2653 if (!virtio_has_feature(vdev, fbit))
2654 return false;
2655
2656 dev_err(&vdev->dev, "device advertises feature %s but not %s",
2657 fname, dname);
2658
2659 return true;
2660}
2661
2662#define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2663 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2664
2665static bool virtnet_validate_features(struct virtio_device *vdev)
2666{
2667 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2668 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2669 "VIRTIO_NET_F_CTRL_VQ") ||
2670 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2671 "VIRTIO_NET_F_CTRL_VQ") ||
2672 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2673 "VIRTIO_NET_F_CTRL_VQ") ||
2674 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2675 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2676 "VIRTIO_NET_F_CTRL_VQ"))) {
2677 return false;
2678 }
2679
2680 return true;
2681}
2682
2683#define MIN_MTU ETH_MIN_MTU
2684#define MAX_MTU ETH_MAX_MTU
2685
2686static int virtnet_validate(struct virtio_device *vdev)
2687{
2688 if (!vdev->config->get) {
2689 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2690 __func__);
2691 return -EINVAL;
2692 }
2693
2694 if (!virtnet_validate_features(vdev))
2695 return -EINVAL;
2696
2697 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2698 int mtu = virtio_cread16(vdev,
2699 offsetof(struct virtio_net_config,
2700 mtu));
2701 if (mtu < MIN_MTU)
2702 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2703 }
2704
2705 return 0;
2706}
2707
2708static int virtnet_probe(struct virtio_device *vdev)
2709{
2710 int i, err = -ENOMEM;
2711 struct net_device *dev;
2712 struct virtnet_info *vi;
2713 u16 max_queue_pairs;
2714 int mtu;
2715
2716 /* Find if host supports multiqueue virtio_net device */
2717 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2718 struct virtio_net_config,
2719 max_virtqueue_pairs, &max_queue_pairs);
2720
2721 /* We need at least 2 queue's */
2722 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2723 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2724 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2725 max_queue_pairs = 1;
2726
2727 /* Allocate ourselves a network device with room for our info */
2728 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2729 if (!dev)
2730 return -ENOMEM;
2731
2732 /* Set up network device as normal. */
2733 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2734 dev->netdev_ops = &virtnet_netdev;
2735 dev->features = NETIF_F_HIGHDMA;
2736
2737 dev->ethtool_ops = &virtnet_ethtool_ops;
2738 SET_NETDEV_DEV(dev, &vdev->dev);
2739
2740 /* Do we support "hardware" checksums? */
2741 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2742 /* This opens up the world of extra features. */
2743 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2744 if (csum)
2745 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2746
2747 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2748 dev->hw_features |= NETIF_F_TSO
2749 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2750 }
2751 /* Individual feature bits: what can host handle? */
2752 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2753 dev->hw_features |= NETIF_F_TSO;
2754 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2755 dev->hw_features |= NETIF_F_TSO6;
2756 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2757 dev->hw_features |= NETIF_F_TSO_ECN;
2758
2759 dev->features |= NETIF_F_GSO_ROBUST;
2760
2761 if (gso)
2762 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
2763 /* (!csum && gso) case will be fixed by register_netdev() */
2764 }
2765 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2766 dev->features |= NETIF_F_RXCSUM;
2767
2768 dev->vlan_features = dev->features;
2769
2770 /* MTU range: 68 - 65535 */
2771 dev->min_mtu = MIN_MTU;
2772 dev->max_mtu = MAX_MTU;
2773
2774 /* Configuration may specify what MAC to use. Otherwise random. */
2775 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2776 virtio_cread_bytes(vdev,
2777 offsetof(struct virtio_net_config, mac),
2778 dev->dev_addr, dev->addr_len);
2779 else
2780 eth_hw_addr_random(dev);
2781
2782 /* Set up our device-specific information */
2783 vi = netdev_priv(dev);
2784 vi->dev = dev;
2785 vi->vdev = vdev;
2786 vdev->priv = vi;
2787
2788 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2789
2790 /* If we can receive ANY GSO packets, we must allocate large ones. */
2791 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2792 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2793 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2794 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2795 vi->big_packets = true;
2796
2797 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2798 vi->mergeable_rx_bufs = true;
2799
2800 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2801 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2802 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2803 else
2804 vi->hdr_len = sizeof(struct virtio_net_hdr);
2805
2806 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2807 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2808 vi->any_header_sg = true;
2809
2810 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2811 vi->has_cvq = true;
2812
2813 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2814 mtu = virtio_cread16(vdev,
2815 offsetof(struct virtio_net_config,
2816 mtu));
2817 if (mtu < dev->min_mtu) {
2818 /* Should never trigger: MTU was previously validated
2819 * in virtnet_validate.
2820 */
2821 dev_err(&vdev->dev, "device MTU appears to have changed "
2822 "it is now %d < %d", mtu, dev->min_mtu);
2823 goto free;
2824 }
2825
2826 dev->mtu = mtu;
2827 dev->max_mtu = mtu;
2828
2829 /* TODO: size buffers correctly in this case. */
2830 if (dev->mtu > ETH_DATA_LEN)
2831 vi->big_packets = true;
2832 }
2833
2834 if (vi->any_header_sg)
2835 dev->needed_headroom = vi->hdr_len;
2836
2837 /* Enable multiqueue by default */
2838 if (num_online_cpus() >= max_queue_pairs)
2839 vi->curr_queue_pairs = max_queue_pairs;
2840 else
2841 vi->curr_queue_pairs = num_online_cpus();
2842 vi->max_queue_pairs = max_queue_pairs;
2843
2844 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2845 err = init_vqs(vi);
2846 if (err)
2847 goto free;
2848
2849#ifdef CONFIG_SYSFS
2850 if (vi->mergeable_rx_bufs)
2851 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2852#endif
2853 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2854 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2855
2856 virtnet_init_settings(dev);
2857
2858 err = register_netdev(dev);
2859 if (err) {
2860 pr_debug("virtio_net: registering device failed\n");
2861 goto free_vqs;
2862 }
2863
2864 virtio_device_ready(vdev);
2865
2866 err = virtnet_cpu_notif_add(vi);
2867 if (err) {
2868 pr_debug("virtio_net: registering cpu notifier failed\n");
2869 goto free_unregister_netdev;
2870 }
2871
2872 virtnet_set_queues(vi, vi->curr_queue_pairs);
2873
2874 /* Assume link up if device can't report link status,
2875 otherwise get link status from config. */
2876 netif_carrier_off(dev);
2877 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2878 schedule_work(&vi->config_work);
2879 } else {
2880 vi->status = VIRTIO_NET_S_LINK_UP;
2881 virtnet_update_settings(vi);
2882 netif_carrier_on(dev);
2883 }
2884
2885 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
2886 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
2887 set_bit(guest_offloads[i], &vi->guest_offloads);
2888
2889 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2890 dev->name, max_queue_pairs);
2891
2892 return 0;
2893
2894free_unregister_netdev:
2895 vi->vdev->config->reset(vdev);
2896
2897 unregister_netdev(dev);
2898free_vqs:
2899 cancel_delayed_work_sync(&vi->refill);
2900 free_receive_page_frags(vi);
2901 virtnet_del_vqs(vi);
2902free:
2903 free_netdev(dev);
2904 return err;
2905}
2906
2907static void remove_vq_common(struct virtnet_info *vi)
2908{
2909 vi->vdev->config->reset(vi->vdev);
2910
2911 /* Free unused buffers in both send and recv, if any. */
2912 free_unused_bufs(vi);
2913
2914 free_receive_bufs(vi);
2915
2916 free_receive_page_frags(vi);
2917
2918 virtnet_del_vqs(vi);
2919}
2920
2921static void virtnet_remove(struct virtio_device *vdev)
2922{
2923 struct virtnet_info *vi = vdev->priv;
2924
2925 virtnet_cpu_notif_remove(vi);
2926
2927 /* Make sure no work handler is accessing the device. */
2928 flush_work(&vi->config_work);
2929
2930 unregister_netdev(vi->dev);
2931
2932 remove_vq_common(vi);
2933
2934 free_netdev(vi->dev);
2935}
2936
2937static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
2938{
2939 struct virtnet_info *vi = vdev->priv;
2940
2941 virtnet_cpu_notif_remove(vi);
2942 virtnet_freeze_down(vdev);
2943 remove_vq_common(vi);
2944
2945 return 0;
2946}
2947
2948static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
2949{
2950 struct virtnet_info *vi = vdev->priv;
2951 int err;
2952
2953 err = virtnet_restore_up(vdev);
2954 if (err)
2955 return err;
2956 virtnet_set_queues(vi, vi->curr_queue_pairs);
2957
2958 err = virtnet_cpu_notif_add(vi);
2959 if (err)
2960 return err;
2961
2962 return 0;
2963}
2964
2965static struct virtio_device_id id_table[] = {
2966 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2967 { 0 },
2968};
2969
2970#define VIRTNET_FEATURES \
2971 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2972 VIRTIO_NET_F_MAC, \
2973 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2974 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2975 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2976 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2977 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2978 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2979 VIRTIO_NET_F_CTRL_MAC_ADDR, \
2980 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
2981 VIRTIO_NET_F_SPEED_DUPLEX
2982
2983static unsigned int features[] = {
2984 VIRTNET_FEATURES,
2985};
2986
2987static unsigned int features_legacy[] = {
2988 VIRTNET_FEATURES,
2989 VIRTIO_NET_F_GSO,
2990 VIRTIO_F_ANY_LAYOUT,
2991};
2992
2993static struct virtio_driver virtio_net_driver = {
2994 .feature_table = features,
2995 .feature_table_size = ARRAY_SIZE(features),
2996 .feature_table_legacy = features_legacy,
2997 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2998 .driver.name = KBUILD_MODNAME,
2999 .driver.owner = THIS_MODULE,
3000 .id_table = id_table,
3001 .validate = virtnet_validate,
3002 .probe = virtnet_probe,
3003 .remove = virtnet_remove,
3004 .config_changed = virtnet_config_changed,
3005#ifdef CONFIG_PM_SLEEP
3006 .freeze = virtnet_freeze,
3007 .restore = virtnet_restore,
3008#endif
3009};
3010
3011static __init int virtio_net_driver_init(void)
3012{
3013 int ret;
3014
3015 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3016 virtnet_cpu_online,
3017 virtnet_cpu_down_prep);
3018 if (ret < 0)
3019 goto out;
3020 virtionet_online = ret;
3021 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3022 NULL, virtnet_cpu_dead);
3023 if (ret)
3024 goto err_dead;
3025
3026 ret = register_virtio_driver(&virtio_net_driver);
3027 if (ret)
3028 goto err_virtio;
3029 return 0;
3030err_virtio:
3031 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3032err_dead:
3033 cpuhp_remove_multi_state(virtionet_online);
3034out:
3035 return ret;
3036}
3037module_init(virtio_net_driver_init);
3038
3039static __exit void virtio_net_driver_exit(void)
3040{
3041 unregister_virtio_driver(&virtio_net_driver);
3042 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3043 cpuhp_remove_multi_state(virtionet_online);
3044}
3045module_exit(virtio_net_driver_exit);
3046
3047MODULE_DEVICE_TABLE(virtio, id_table);
3048MODULE_DESCRIPTION("Virtio network driver");
3049MODULE_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
66#define GUEST_OFFLOAD_LRO_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 /* I like... big packets and I cannot lie! */
199 bool big_packets;
200
201 /* Host will merge rx buffers for big packets (shake it! shake it!) */
202 bool mergeable_rx_bufs;
203
204 /* Has control virtqueue */
205 bool has_cvq;
206
207 /* Host can handle any s/g split between our header and packet data */
208 bool any_header_sg;
209
210 /* Packet virtio header size */
211 u8 hdr_len;
212
213 /* Work struct for refilling if we run low on memory. */
214 struct delayed_work refill;
215
216 /* Work struct for config space updates */
217 struct work_struct config_work;
218
219 /* Does the affinity hint is set for virtqueues? */
220 bool affinity_hint_set;
221
222 /* CPU hotplug instances for online & dead */
223 struct hlist_node node;
224 struct hlist_node node_dead;
225
226 struct control_buf *ctrl;
227
228 /* Ethtool settings */
229 u8 duplex;
230 u32 speed;
231
232 unsigned long guest_offloads;
233 unsigned long guest_offloads_capable;
234
235 /* failover when STANDBY feature enabled */
236 struct failover *failover;
237};
238
239struct padded_vnet_hdr {
240 struct virtio_net_hdr_mrg_rxbuf hdr;
241 /*
242 * hdr is in a separate sg buffer, and data sg buffer shares same page
243 * with this header sg. This padding makes next sg 16 byte aligned
244 * after the header.
245 */
246 char padding[4];
247};
248
249static bool is_xdp_frame(void *ptr)
250{
251 return (unsigned long)ptr & VIRTIO_XDP_FLAG;
252}
253
254static void *xdp_to_ptr(struct xdp_frame *ptr)
255{
256 return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
257}
258
259static struct xdp_frame *ptr_to_xdp(void *ptr)
260{
261 return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
262}
263
264/* Converting between virtqueue no. and kernel tx/rx queue no.
265 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
266 */
267static int vq2txq(struct virtqueue *vq)
268{
269 return (vq->index - 1) / 2;
270}
271
272static int txq2vq(int txq)
273{
274 return txq * 2 + 1;
275}
276
277static int vq2rxq(struct virtqueue *vq)
278{
279 return vq->index / 2;
280}
281
282static int rxq2vq(int rxq)
283{
284 return rxq * 2;
285}
286
287static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
288{
289 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
290}
291
292/*
293 * private is used to chain pages for big packets, put the whole
294 * most recent used list in the beginning for reuse
295 */
296static void give_pages(struct receive_queue *rq, struct page *page)
297{
298 struct page *end;
299
300 /* Find end of list, sew whole thing into vi->rq.pages. */
301 for (end = page; end->private; end = (struct page *)end->private);
302 end->private = (unsigned long)rq->pages;
303 rq->pages = page;
304}
305
306static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
307{
308 struct page *p = rq->pages;
309
310 if (p) {
311 rq->pages = (struct page *)p->private;
312 /* clear private here, it is used to chain pages */
313 p->private = 0;
314 } else
315 p = alloc_page(gfp_mask);
316 return p;
317}
318
319static void virtqueue_napi_schedule(struct napi_struct *napi,
320 struct virtqueue *vq)
321{
322 if (napi_schedule_prep(napi)) {
323 virtqueue_disable_cb(vq);
324 __napi_schedule(napi);
325 }
326}
327
328static void virtqueue_napi_complete(struct napi_struct *napi,
329 struct virtqueue *vq, int processed)
330{
331 int opaque;
332
333 opaque = virtqueue_enable_cb_prepare(vq);
334 if (napi_complete_done(napi, processed)) {
335 if (unlikely(virtqueue_poll(vq, opaque)))
336 virtqueue_napi_schedule(napi, vq);
337 } else {
338 virtqueue_disable_cb(vq);
339 }
340}
341
342static void skb_xmit_done(struct virtqueue *vq)
343{
344 struct virtnet_info *vi = vq->vdev->priv;
345 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
346
347 /* Suppress further interrupts. */
348 virtqueue_disable_cb(vq);
349
350 if (napi->weight)
351 virtqueue_napi_schedule(napi, vq);
352 else
353 /* We were probably waiting for more output buffers. */
354 netif_wake_subqueue(vi->dev, vq2txq(vq));
355}
356
357#define MRG_CTX_HEADER_SHIFT 22
358static void *mergeable_len_to_ctx(unsigned int truesize,
359 unsigned int headroom)
360{
361 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
362}
363
364static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
365{
366 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
367}
368
369static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
370{
371 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
372}
373
374/* Called from bottom half context */
375static struct sk_buff *page_to_skb(struct virtnet_info *vi,
376 struct receive_queue *rq,
377 struct page *page, unsigned int offset,
378 unsigned int len, unsigned int truesize,
379 bool hdr_valid, unsigned int metasize)
380{
381 struct sk_buff *skb;
382 struct virtio_net_hdr_mrg_rxbuf *hdr;
383 unsigned int copy, hdr_len, hdr_padded_len;
384 char *p;
385
386 p = page_address(page) + offset;
387
388 /* copy small packet so we can reuse these pages for small data */
389 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
390 if (unlikely(!skb))
391 return NULL;
392
393 hdr = skb_vnet_hdr(skb);
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 /* hdr_valid means no XDP, so we can copy the vnet header */
402 if (hdr_valid)
403 memcpy(hdr, p, hdr_len);
404
405 len -= hdr_len;
406 offset += hdr_padded_len;
407 p += hdr_padded_len;
408
409 copy = len;
410 if (copy > skb_tailroom(skb))
411 copy = skb_tailroom(skb);
412 skb_put_data(skb, p, copy);
413
414 if (metasize) {
415 __skb_pull(skb, metasize);
416 skb_metadata_set(skb, metasize);
417 }
418
419 len -= copy;
420 offset += copy;
421
422 if (vi->mergeable_rx_bufs) {
423 if (len)
424 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
425 else
426 put_page(page);
427 return skb;
428 }
429
430 /*
431 * Verify that we can indeed put this data into a skb.
432 * This is here to handle cases when the device erroneously
433 * tries to receive more than is possible. This is usually
434 * the case of a broken device.
435 */
436 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
437 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
438 dev_kfree_skb(skb);
439 return NULL;
440 }
441 BUG_ON(offset >= PAGE_SIZE);
442 while (len) {
443 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
444 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
445 frag_size, truesize);
446 len -= frag_size;
447 page = (struct page *)page->private;
448 offset = 0;
449 }
450
451 if (page)
452 give_pages(rq, page);
453
454 return skb;
455}
456
457static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
458 struct send_queue *sq,
459 struct xdp_frame *xdpf)
460{
461 struct virtio_net_hdr_mrg_rxbuf *hdr;
462 int err;
463
464 if (unlikely(xdpf->headroom < vi->hdr_len))
465 return -EOVERFLOW;
466
467 /* Make room for virtqueue hdr (also change xdpf->headroom?) */
468 xdpf->data -= vi->hdr_len;
469 /* Zero header and leave csum up to XDP layers */
470 hdr = xdpf->data;
471 memset(hdr, 0, vi->hdr_len);
472 xdpf->len += vi->hdr_len;
473
474 sg_init_one(sq->sg, xdpf->data, xdpf->len);
475
476 err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
477 GFP_ATOMIC);
478 if (unlikely(err))
479 return -ENOSPC; /* Caller handle free/refcnt */
480
481 return 0;
482}
483
484static struct send_queue *virtnet_xdp_sq(struct virtnet_info *vi)
485{
486 unsigned int qp;
487
488 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
489 return &vi->sq[qp];
490}
491
492static int virtnet_xdp_xmit(struct net_device *dev,
493 int n, struct xdp_frame **frames, u32 flags)
494{
495 struct virtnet_info *vi = netdev_priv(dev);
496 struct receive_queue *rq = vi->rq;
497 struct bpf_prog *xdp_prog;
498 struct send_queue *sq;
499 unsigned int len;
500 int packets = 0;
501 int bytes = 0;
502 int drops = 0;
503 int kicks = 0;
504 int ret, err;
505 void *ptr;
506 int i;
507
508 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
509 * indicate XDP resources have been successfully allocated.
510 */
511 xdp_prog = rcu_access_pointer(rq->xdp_prog);
512 if (!xdp_prog)
513 return -ENXIO;
514
515 sq = virtnet_xdp_sq(vi);
516
517 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
518 ret = -EINVAL;
519 drops = n;
520 goto out;
521 }
522
523 /* Free up any pending old buffers before queueing new ones. */
524 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
525 if (likely(is_xdp_frame(ptr))) {
526 struct xdp_frame *frame = ptr_to_xdp(ptr);
527
528 bytes += frame->len;
529 xdp_return_frame(frame);
530 } else {
531 struct sk_buff *skb = ptr;
532
533 bytes += skb->len;
534 napi_consume_skb(skb, false);
535 }
536 packets++;
537 }
538
539 for (i = 0; i < n; i++) {
540 struct xdp_frame *xdpf = frames[i];
541
542 err = __virtnet_xdp_xmit_one(vi, sq, xdpf);
543 if (err) {
544 xdp_return_frame_rx_napi(xdpf);
545 drops++;
546 }
547 }
548 ret = n - drops;
549
550 if (flags & XDP_XMIT_FLUSH) {
551 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
552 kicks = 1;
553 }
554out:
555 u64_stats_update_begin(&sq->stats.syncp);
556 sq->stats.bytes += bytes;
557 sq->stats.packets += packets;
558 sq->stats.xdp_tx += n;
559 sq->stats.xdp_tx_drops += drops;
560 sq->stats.kicks += kicks;
561 u64_stats_update_end(&sq->stats.syncp);
562
563 return ret;
564}
565
566static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
567{
568 return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
569}
570
571/* We copy the packet for XDP in the following cases:
572 *
573 * 1) Packet is scattered across multiple rx buffers.
574 * 2) Headroom space is insufficient.
575 *
576 * This is inefficient but it's a temporary condition that
577 * we hit right after XDP is enabled and until queue is refilled
578 * with large buffers with sufficient headroom - so it should affect
579 * at most queue size packets.
580 * Afterwards, the conditions to enable
581 * XDP should preclude the underlying device from sending packets
582 * across multiple buffers (num_buf > 1), and we make sure buffers
583 * have enough headroom.
584 */
585static struct page *xdp_linearize_page(struct receive_queue *rq,
586 u16 *num_buf,
587 struct page *p,
588 int offset,
589 int page_off,
590 unsigned int *len)
591{
592 struct page *page = alloc_page(GFP_ATOMIC);
593
594 if (!page)
595 return NULL;
596
597 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
598 page_off += *len;
599
600 while (--*num_buf) {
601 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
602 unsigned int buflen;
603 void *buf;
604 int off;
605
606 buf = virtqueue_get_buf(rq->vq, &buflen);
607 if (unlikely(!buf))
608 goto err_buf;
609
610 p = virt_to_head_page(buf);
611 off = buf - page_address(p);
612
613 /* guard against a misconfigured or uncooperative backend that
614 * is sending packet larger than the MTU.
615 */
616 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
617 put_page(p);
618 goto err_buf;
619 }
620
621 memcpy(page_address(page) + page_off,
622 page_address(p) + off, buflen);
623 page_off += buflen;
624 put_page(p);
625 }
626
627 /* Headroom does not contribute to packet length */
628 *len = page_off - VIRTIO_XDP_HEADROOM;
629 return page;
630err_buf:
631 __free_pages(page, 0);
632 return NULL;
633}
634
635static struct sk_buff *receive_small(struct net_device *dev,
636 struct virtnet_info *vi,
637 struct receive_queue *rq,
638 void *buf, void *ctx,
639 unsigned int len,
640 unsigned int *xdp_xmit,
641 struct virtnet_rq_stats *stats)
642{
643 struct sk_buff *skb;
644 struct bpf_prog *xdp_prog;
645 unsigned int xdp_headroom = (unsigned long)ctx;
646 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
647 unsigned int headroom = vi->hdr_len + header_offset;
648 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
649 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
650 struct page *page = virt_to_head_page(buf);
651 unsigned int delta = 0;
652 struct page *xdp_page;
653 int err;
654 unsigned int metasize = 0;
655
656 len -= vi->hdr_len;
657 stats->bytes += len;
658
659 rcu_read_lock();
660 xdp_prog = rcu_dereference(rq->xdp_prog);
661 if (xdp_prog) {
662 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
663 struct xdp_frame *xdpf;
664 struct xdp_buff xdp;
665 void *orig_data;
666 u32 act;
667
668 if (unlikely(hdr->hdr.gso_type))
669 goto err_xdp;
670
671 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
672 int offset = buf - page_address(page) + header_offset;
673 unsigned int tlen = len + vi->hdr_len;
674 u16 num_buf = 1;
675
676 xdp_headroom = virtnet_get_headroom(vi);
677 header_offset = VIRTNET_RX_PAD + xdp_headroom;
678 headroom = vi->hdr_len + header_offset;
679 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
680 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
681 xdp_page = xdp_linearize_page(rq, &num_buf, page,
682 offset, header_offset,
683 &tlen);
684 if (!xdp_page)
685 goto err_xdp;
686
687 buf = page_address(xdp_page);
688 put_page(page);
689 page = xdp_page;
690 }
691
692 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
693 xdp.data = xdp.data_hard_start + xdp_headroom;
694 xdp.data_end = xdp.data + len;
695 xdp.data_meta = xdp.data;
696 xdp.rxq = &rq->xdp_rxq;
697 xdp.frame_sz = buflen;
698 orig_data = xdp.data;
699 act = bpf_prog_run_xdp(xdp_prog, &xdp);
700 stats->xdp_packets++;
701
702 switch (act) {
703 case XDP_PASS:
704 /* Recalculate length in case bpf program changed it */
705 delta = orig_data - xdp.data;
706 len = xdp.data_end - xdp.data;
707 metasize = xdp.data - xdp.data_meta;
708 break;
709 case XDP_TX:
710 stats->xdp_tx++;
711 xdpf = xdp_convert_buff_to_frame(&xdp);
712 if (unlikely(!xdpf))
713 goto err_xdp;
714 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
715 if (unlikely(err < 0)) {
716 trace_xdp_exception(vi->dev, xdp_prog, act);
717 goto err_xdp;
718 }
719 *xdp_xmit |= VIRTIO_XDP_TX;
720 rcu_read_unlock();
721 goto xdp_xmit;
722 case XDP_REDIRECT:
723 stats->xdp_redirects++;
724 err = xdp_do_redirect(dev, &xdp, xdp_prog);
725 if (err)
726 goto err_xdp;
727 *xdp_xmit |= VIRTIO_XDP_REDIR;
728 rcu_read_unlock();
729 goto xdp_xmit;
730 default:
731 bpf_warn_invalid_xdp_action(act);
732 fallthrough;
733 case XDP_ABORTED:
734 trace_xdp_exception(vi->dev, xdp_prog, act);
735 case XDP_DROP:
736 goto err_xdp;
737 }
738 }
739 rcu_read_unlock();
740
741 skb = build_skb(buf, buflen);
742 if (!skb) {
743 put_page(page);
744 goto err;
745 }
746 skb_reserve(skb, headroom - delta);
747 skb_put(skb, len);
748 if (!xdp_prog) {
749 buf += header_offset;
750 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
751 } /* keep zeroed vnet hdr since XDP is loaded */
752
753 if (metasize)
754 skb_metadata_set(skb, metasize);
755
756err:
757 return skb;
758
759err_xdp:
760 rcu_read_unlock();
761 stats->xdp_drops++;
762 stats->drops++;
763 put_page(page);
764xdp_xmit:
765 return NULL;
766}
767
768static struct sk_buff *receive_big(struct net_device *dev,
769 struct virtnet_info *vi,
770 struct receive_queue *rq,
771 void *buf,
772 unsigned int len,
773 struct virtnet_rq_stats *stats)
774{
775 struct page *page = buf;
776 struct sk_buff *skb =
777 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0);
778
779 stats->bytes += len - vi->hdr_len;
780 if (unlikely(!skb))
781 goto err;
782
783 return skb;
784
785err:
786 stats->drops++;
787 give_pages(rq, page);
788 return NULL;
789}
790
791static struct sk_buff *receive_mergeable(struct net_device *dev,
792 struct virtnet_info *vi,
793 struct receive_queue *rq,
794 void *buf,
795 void *ctx,
796 unsigned int len,
797 unsigned int *xdp_xmit,
798 struct virtnet_rq_stats *stats)
799{
800 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
801 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
802 struct page *page = virt_to_head_page(buf);
803 int offset = buf - page_address(page);
804 struct sk_buff *head_skb, *curr_skb;
805 struct bpf_prog *xdp_prog;
806 unsigned int truesize = mergeable_ctx_to_truesize(ctx);
807 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
808 unsigned int metasize = 0;
809 unsigned int frame_sz;
810 int err;
811
812 head_skb = NULL;
813 stats->bytes += len - vi->hdr_len;
814
815 rcu_read_lock();
816 xdp_prog = rcu_dereference(rq->xdp_prog);
817 if (xdp_prog) {
818 struct xdp_frame *xdpf;
819 struct page *xdp_page;
820 struct xdp_buff xdp;
821 void *data;
822 u32 act;
823
824 /* Transient failure which in theory could occur if
825 * in-flight packets from before XDP was enabled reach
826 * the receive path after XDP is loaded.
827 */
828 if (unlikely(hdr->hdr.gso_type))
829 goto err_xdp;
830
831 /* Buffers with headroom use PAGE_SIZE as alloc size,
832 * see add_recvbuf_mergeable() + get_mergeable_buf_len()
833 */
834 frame_sz = headroom ? PAGE_SIZE : truesize;
835
836 /* This happens when rx buffer size is underestimated
837 * or headroom is not enough because of the buffer
838 * was refilled before XDP is set. This should only
839 * happen for the first several packets, so we don't
840 * care much about its performance.
841 */
842 if (unlikely(num_buf > 1 ||
843 headroom < virtnet_get_headroom(vi))) {
844 /* linearize data for XDP */
845 xdp_page = xdp_linearize_page(rq, &num_buf,
846 page, offset,
847 VIRTIO_XDP_HEADROOM,
848 &len);
849 frame_sz = PAGE_SIZE;
850
851 if (!xdp_page)
852 goto err_xdp;
853 offset = VIRTIO_XDP_HEADROOM;
854 } else {
855 xdp_page = page;
856 }
857
858 /* Allow consuming headroom but reserve enough space to push
859 * the descriptor on if we get an XDP_TX return code.
860 */
861 data = page_address(xdp_page) + offset;
862 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
863 xdp.data = data + vi->hdr_len;
864 xdp.data_end = xdp.data + (len - vi->hdr_len);
865 xdp.data_meta = xdp.data;
866 xdp.rxq = &rq->xdp_rxq;
867 xdp.frame_sz = frame_sz - vi->hdr_len;
868
869 act = bpf_prog_run_xdp(xdp_prog, &xdp);
870 stats->xdp_packets++;
871
872 switch (act) {
873 case XDP_PASS:
874 metasize = xdp.data - xdp.data_meta;
875
876 /* recalculate offset to account for any header
877 * adjustments and minus the metasize to copy the
878 * metadata in page_to_skb(). Note other cases do not
879 * build an skb and avoid using offset
880 */
881 offset = xdp.data - page_address(xdp_page) -
882 vi->hdr_len - metasize;
883
884 /* recalculate len if xdp.data, xdp.data_end or
885 * xdp.data_meta were adjusted
886 */
887 len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
888 /* We can only create skb based on xdp_page. */
889 if (unlikely(xdp_page != page)) {
890 rcu_read_unlock();
891 put_page(page);
892 head_skb = page_to_skb(vi, rq, xdp_page, offset,
893 len, PAGE_SIZE, false,
894 metasize);
895 return head_skb;
896 }
897 break;
898 case XDP_TX:
899 stats->xdp_tx++;
900 xdpf = xdp_convert_buff_to_frame(&xdp);
901 if (unlikely(!xdpf))
902 goto err_xdp;
903 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
904 if (unlikely(err < 0)) {
905 trace_xdp_exception(vi->dev, xdp_prog, act);
906 if (unlikely(xdp_page != page))
907 put_page(xdp_page);
908 goto err_xdp;
909 }
910 *xdp_xmit |= VIRTIO_XDP_TX;
911 if (unlikely(xdp_page != page))
912 put_page(page);
913 rcu_read_unlock();
914 goto xdp_xmit;
915 case XDP_REDIRECT:
916 stats->xdp_redirects++;
917 err = xdp_do_redirect(dev, &xdp, xdp_prog);
918 if (err) {
919 if (unlikely(xdp_page != page))
920 put_page(xdp_page);
921 goto err_xdp;
922 }
923 *xdp_xmit |= VIRTIO_XDP_REDIR;
924 if (unlikely(xdp_page != page))
925 put_page(page);
926 rcu_read_unlock();
927 goto xdp_xmit;
928 default:
929 bpf_warn_invalid_xdp_action(act);
930 fallthrough;
931 case XDP_ABORTED:
932 trace_xdp_exception(vi->dev, xdp_prog, act);
933 fallthrough;
934 case XDP_DROP:
935 if (unlikely(xdp_page != page))
936 __free_pages(xdp_page, 0);
937 goto err_xdp;
938 }
939 }
940 rcu_read_unlock();
941
942 if (unlikely(len > truesize)) {
943 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
944 dev->name, len, (unsigned long)ctx);
945 dev->stats.rx_length_errors++;
946 goto err_skb;
947 }
948
949 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
950 metasize);
951 curr_skb = head_skb;
952
953 if (unlikely(!curr_skb))
954 goto err_skb;
955 while (--num_buf) {
956 int num_skb_frags;
957
958 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
959 if (unlikely(!buf)) {
960 pr_debug("%s: rx error: %d buffers out of %d missing\n",
961 dev->name, num_buf,
962 virtio16_to_cpu(vi->vdev,
963 hdr->num_buffers));
964 dev->stats.rx_length_errors++;
965 goto err_buf;
966 }
967
968 stats->bytes += len;
969 page = virt_to_head_page(buf);
970
971 truesize = mergeable_ctx_to_truesize(ctx);
972 if (unlikely(len > truesize)) {
973 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
974 dev->name, len, (unsigned long)ctx);
975 dev->stats.rx_length_errors++;
976 goto err_skb;
977 }
978
979 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
980 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
981 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
982
983 if (unlikely(!nskb))
984 goto err_skb;
985 if (curr_skb == head_skb)
986 skb_shinfo(curr_skb)->frag_list = nskb;
987 else
988 curr_skb->next = nskb;
989 curr_skb = nskb;
990 head_skb->truesize += nskb->truesize;
991 num_skb_frags = 0;
992 }
993 if (curr_skb != head_skb) {
994 head_skb->data_len += len;
995 head_skb->len += len;
996 head_skb->truesize += truesize;
997 }
998 offset = buf - page_address(page);
999 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1000 put_page(page);
1001 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1002 len, truesize);
1003 } else {
1004 skb_add_rx_frag(curr_skb, num_skb_frags, page,
1005 offset, len, truesize);
1006 }
1007 }
1008
1009 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1010 return head_skb;
1011
1012err_xdp:
1013 rcu_read_unlock();
1014 stats->xdp_drops++;
1015err_skb:
1016 put_page(page);
1017 while (num_buf-- > 1) {
1018 buf = virtqueue_get_buf(rq->vq, &len);
1019 if (unlikely(!buf)) {
1020 pr_debug("%s: rx error: %d buffers missing\n",
1021 dev->name, num_buf);
1022 dev->stats.rx_length_errors++;
1023 break;
1024 }
1025 stats->bytes += len;
1026 page = virt_to_head_page(buf);
1027 put_page(page);
1028 }
1029err_buf:
1030 stats->drops++;
1031 dev_kfree_skb(head_skb);
1032xdp_xmit:
1033 return NULL;
1034}
1035
1036static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1037 void *buf, unsigned int len, void **ctx,
1038 unsigned int *xdp_xmit,
1039 struct virtnet_rq_stats *stats)
1040{
1041 struct net_device *dev = vi->dev;
1042 struct sk_buff *skb;
1043 struct virtio_net_hdr_mrg_rxbuf *hdr;
1044
1045 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1046 pr_debug("%s: short packet %i\n", dev->name, len);
1047 dev->stats.rx_length_errors++;
1048 if (vi->mergeable_rx_bufs) {
1049 put_page(virt_to_head_page(buf));
1050 } else if (vi->big_packets) {
1051 give_pages(rq, buf);
1052 } else {
1053 put_page(virt_to_head_page(buf));
1054 }
1055 return;
1056 }
1057
1058 if (vi->mergeable_rx_bufs)
1059 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1060 stats);
1061 else if (vi->big_packets)
1062 skb = receive_big(dev, vi, rq, buf, len, stats);
1063 else
1064 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1065
1066 if (unlikely(!skb))
1067 return;
1068
1069 hdr = skb_vnet_hdr(skb);
1070
1071 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1072 skb->ip_summed = CHECKSUM_UNNECESSARY;
1073
1074 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1075 virtio_is_little_endian(vi->vdev))) {
1076 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1077 dev->name, hdr->hdr.gso_type,
1078 hdr->hdr.gso_size);
1079 goto frame_err;
1080 }
1081
1082 skb_record_rx_queue(skb, vq2rxq(rq->vq));
1083 skb->protocol = eth_type_trans(skb, dev);
1084 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1085 ntohs(skb->protocol), skb->len, skb->pkt_type);
1086
1087 napi_gro_receive(&rq->napi, skb);
1088 return;
1089
1090frame_err:
1091 dev->stats.rx_frame_errors++;
1092 dev_kfree_skb(skb);
1093}
1094
1095/* Unlike mergeable buffers, all buffers are allocated to the
1096 * same size, except for the headroom. For this reason we do
1097 * not need to use mergeable_len_to_ctx here - it is enough
1098 * to store the headroom as the context ignoring the truesize.
1099 */
1100static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1101 gfp_t gfp)
1102{
1103 struct page_frag *alloc_frag = &rq->alloc_frag;
1104 char *buf;
1105 unsigned int xdp_headroom = virtnet_get_headroom(vi);
1106 void *ctx = (void *)(unsigned long)xdp_headroom;
1107 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1108 int err;
1109
1110 len = SKB_DATA_ALIGN(len) +
1111 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1112 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1113 return -ENOMEM;
1114
1115 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1116 get_page(alloc_frag->page);
1117 alloc_frag->offset += len;
1118 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1119 vi->hdr_len + GOOD_PACKET_LEN);
1120 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1121 if (err < 0)
1122 put_page(virt_to_head_page(buf));
1123 return err;
1124}
1125
1126static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1127 gfp_t gfp)
1128{
1129 struct page *first, *list = NULL;
1130 char *p;
1131 int i, err, offset;
1132
1133 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1134
1135 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1136 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1137 first = get_a_page(rq, gfp);
1138 if (!first) {
1139 if (list)
1140 give_pages(rq, list);
1141 return -ENOMEM;
1142 }
1143 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1144
1145 /* chain new page in list head to match sg */
1146 first->private = (unsigned long)list;
1147 list = first;
1148 }
1149
1150 first = get_a_page(rq, gfp);
1151 if (!first) {
1152 give_pages(rq, list);
1153 return -ENOMEM;
1154 }
1155 p = page_address(first);
1156
1157 /* rq->sg[0], rq->sg[1] share the same page */
1158 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1159 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1160
1161 /* rq->sg[1] for data packet, from offset */
1162 offset = sizeof(struct padded_vnet_hdr);
1163 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1164
1165 /* chain first in list head */
1166 first->private = (unsigned long)list;
1167 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1168 first, gfp);
1169 if (err < 0)
1170 give_pages(rq, first);
1171
1172 return err;
1173}
1174
1175static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1176 struct ewma_pkt_len *avg_pkt_len,
1177 unsigned int room)
1178{
1179 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1180 unsigned int len;
1181
1182 if (room)
1183 return PAGE_SIZE - room;
1184
1185 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1186 rq->min_buf_len, PAGE_SIZE - hdr_len);
1187
1188 return ALIGN(len, L1_CACHE_BYTES);
1189}
1190
1191static int add_recvbuf_mergeable(struct virtnet_info *vi,
1192 struct receive_queue *rq, gfp_t gfp)
1193{
1194 struct page_frag *alloc_frag = &rq->alloc_frag;
1195 unsigned int headroom = virtnet_get_headroom(vi);
1196 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1197 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1198 char *buf;
1199 void *ctx;
1200 int err;
1201 unsigned int len, hole;
1202
1203 /* Extra tailroom is needed to satisfy XDP's assumption. This
1204 * means rx frags coalescing won't work, but consider we've
1205 * disabled GSO for XDP, it won't be a big issue.
1206 */
1207 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1208 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1209 return -ENOMEM;
1210
1211 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1212 buf += headroom; /* advance address leaving hole at front of pkt */
1213 get_page(alloc_frag->page);
1214 alloc_frag->offset += len + room;
1215 hole = alloc_frag->size - alloc_frag->offset;
1216 if (hole < len + room) {
1217 /* To avoid internal fragmentation, if there is very likely not
1218 * enough space for another buffer, add the remaining space to
1219 * the current buffer.
1220 */
1221 len += hole;
1222 alloc_frag->offset += hole;
1223 }
1224
1225 sg_init_one(rq->sg, buf, len);
1226 ctx = mergeable_len_to_ctx(len, headroom);
1227 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1228 if (err < 0)
1229 put_page(virt_to_head_page(buf));
1230
1231 return err;
1232}
1233
1234/*
1235 * Returns false if we couldn't fill entirely (OOM).
1236 *
1237 * Normally run in the receive path, but can also be run from ndo_open
1238 * before we're receiving packets, or from refill_work which is
1239 * careful to disable receiving (using napi_disable).
1240 */
1241static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1242 gfp_t gfp)
1243{
1244 int err;
1245 bool oom;
1246
1247 do {
1248 if (vi->mergeable_rx_bufs)
1249 err = add_recvbuf_mergeable(vi, rq, gfp);
1250 else if (vi->big_packets)
1251 err = add_recvbuf_big(vi, rq, gfp);
1252 else
1253 err = add_recvbuf_small(vi, rq, gfp);
1254
1255 oom = err == -ENOMEM;
1256 if (err)
1257 break;
1258 } while (rq->vq->num_free);
1259 if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1260 unsigned long flags;
1261
1262 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1263 rq->stats.kicks++;
1264 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1265 }
1266
1267 return !oom;
1268}
1269
1270static void skb_recv_done(struct virtqueue *rvq)
1271{
1272 struct virtnet_info *vi = rvq->vdev->priv;
1273 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1274
1275 virtqueue_napi_schedule(&rq->napi, rvq);
1276}
1277
1278static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1279{
1280 napi_enable(napi);
1281
1282 /* If all buffers were filled by other side before we napi_enabled, we
1283 * won't get another interrupt, so process any outstanding packets now.
1284 * Call local_bh_enable after to trigger softIRQ processing.
1285 */
1286 local_bh_disable();
1287 virtqueue_napi_schedule(napi, vq);
1288 local_bh_enable();
1289}
1290
1291static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1292 struct virtqueue *vq,
1293 struct napi_struct *napi)
1294{
1295 if (!napi->weight)
1296 return;
1297
1298 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1299 * enable the feature if this is likely affine with the transmit path.
1300 */
1301 if (!vi->affinity_hint_set) {
1302 napi->weight = 0;
1303 return;
1304 }
1305
1306 return virtnet_napi_enable(vq, napi);
1307}
1308
1309static void virtnet_napi_tx_disable(struct napi_struct *napi)
1310{
1311 if (napi->weight)
1312 napi_disable(napi);
1313}
1314
1315static void refill_work(struct work_struct *work)
1316{
1317 struct virtnet_info *vi =
1318 container_of(work, struct virtnet_info, refill.work);
1319 bool still_empty;
1320 int i;
1321
1322 for (i = 0; i < vi->curr_queue_pairs; i++) {
1323 struct receive_queue *rq = &vi->rq[i];
1324
1325 napi_disable(&rq->napi);
1326 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1327 virtnet_napi_enable(rq->vq, &rq->napi);
1328
1329 /* In theory, this can happen: if we don't get any buffers in
1330 * we will *never* try to fill again.
1331 */
1332 if (still_empty)
1333 schedule_delayed_work(&vi->refill, HZ/2);
1334 }
1335}
1336
1337static int virtnet_receive(struct receive_queue *rq, int budget,
1338 unsigned int *xdp_xmit)
1339{
1340 struct virtnet_info *vi = rq->vq->vdev->priv;
1341 struct virtnet_rq_stats stats = {};
1342 unsigned int len;
1343 void *buf;
1344 int i;
1345
1346 if (!vi->big_packets || vi->mergeable_rx_bufs) {
1347 void *ctx;
1348
1349 while (stats.packets < budget &&
1350 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1351 receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1352 stats.packets++;
1353 }
1354 } else {
1355 while (stats.packets < budget &&
1356 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1357 receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1358 stats.packets++;
1359 }
1360 }
1361
1362 if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1363 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1364 schedule_delayed_work(&vi->refill, 0);
1365 }
1366
1367 u64_stats_update_begin(&rq->stats.syncp);
1368 for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1369 size_t offset = virtnet_rq_stats_desc[i].offset;
1370 u64 *item;
1371
1372 item = (u64 *)((u8 *)&rq->stats + offset);
1373 *item += *(u64 *)((u8 *)&stats + offset);
1374 }
1375 u64_stats_update_end(&rq->stats.syncp);
1376
1377 return stats.packets;
1378}
1379
1380static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1381{
1382 unsigned int len;
1383 unsigned int packets = 0;
1384 unsigned int bytes = 0;
1385 void *ptr;
1386
1387 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1388 if (likely(!is_xdp_frame(ptr))) {
1389 struct sk_buff *skb = ptr;
1390
1391 pr_debug("Sent skb %p\n", skb);
1392
1393 bytes += skb->len;
1394 napi_consume_skb(skb, in_napi);
1395 } else {
1396 struct xdp_frame *frame = ptr_to_xdp(ptr);
1397
1398 bytes += frame->len;
1399 xdp_return_frame(frame);
1400 }
1401 packets++;
1402 }
1403
1404 /* Avoid overhead when no packets have been processed
1405 * happens when called speculatively from start_xmit.
1406 */
1407 if (!packets)
1408 return;
1409
1410 u64_stats_update_begin(&sq->stats.syncp);
1411 sq->stats.bytes += bytes;
1412 sq->stats.packets += packets;
1413 u64_stats_update_end(&sq->stats.syncp);
1414}
1415
1416static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1417{
1418 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1419 return false;
1420 else if (q < vi->curr_queue_pairs)
1421 return true;
1422 else
1423 return false;
1424}
1425
1426static void virtnet_poll_cleantx(struct receive_queue *rq)
1427{
1428 struct virtnet_info *vi = rq->vq->vdev->priv;
1429 unsigned int index = vq2rxq(rq->vq);
1430 struct send_queue *sq = &vi->sq[index];
1431 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1432
1433 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1434 return;
1435
1436 if (__netif_tx_trylock(txq)) {
1437 free_old_xmit_skbs(sq, true);
1438 __netif_tx_unlock(txq);
1439 }
1440
1441 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1442 netif_tx_wake_queue(txq);
1443}
1444
1445static int virtnet_poll(struct napi_struct *napi, int budget)
1446{
1447 struct receive_queue *rq =
1448 container_of(napi, struct receive_queue, napi);
1449 struct virtnet_info *vi = rq->vq->vdev->priv;
1450 struct send_queue *sq;
1451 unsigned int received;
1452 unsigned int xdp_xmit = 0;
1453
1454 virtnet_poll_cleantx(rq);
1455
1456 received = virtnet_receive(rq, budget, &xdp_xmit);
1457
1458 /* Out of packets? */
1459 if (received < budget)
1460 virtqueue_napi_complete(napi, rq->vq, received);
1461
1462 if (xdp_xmit & VIRTIO_XDP_REDIR)
1463 xdp_do_flush();
1464
1465 if (xdp_xmit & VIRTIO_XDP_TX) {
1466 sq = virtnet_xdp_sq(vi);
1467 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1468 u64_stats_update_begin(&sq->stats.syncp);
1469 sq->stats.kicks++;
1470 u64_stats_update_end(&sq->stats.syncp);
1471 }
1472 }
1473
1474 return received;
1475}
1476
1477static int virtnet_open(struct net_device *dev)
1478{
1479 struct virtnet_info *vi = netdev_priv(dev);
1480 int i, err;
1481
1482 for (i = 0; i < vi->max_queue_pairs; i++) {
1483 if (i < vi->curr_queue_pairs)
1484 /* Make sure we have some buffers: if oom use wq. */
1485 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1486 schedule_delayed_work(&vi->refill, 0);
1487
1488 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i);
1489 if (err < 0)
1490 return err;
1491
1492 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1493 MEM_TYPE_PAGE_SHARED, NULL);
1494 if (err < 0) {
1495 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1496 return err;
1497 }
1498
1499 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1500 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1501 }
1502
1503 return 0;
1504}
1505
1506static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1507{
1508 struct send_queue *sq = container_of(napi, struct send_queue, napi);
1509 struct virtnet_info *vi = sq->vq->vdev->priv;
1510 unsigned int index = vq2txq(sq->vq);
1511 struct netdev_queue *txq;
1512
1513 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1514 /* We don't need to enable cb for XDP */
1515 napi_complete_done(napi, 0);
1516 return 0;
1517 }
1518
1519 txq = netdev_get_tx_queue(vi->dev, index);
1520 __netif_tx_lock(txq, raw_smp_processor_id());
1521 free_old_xmit_skbs(sq, true);
1522 __netif_tx_unlock(txq);
1523
1524 virtqueue_napi_complete(napi, sq->vq, 0);
1525
1526 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1527 netif_tx_wake_queue(txq);
1528
1529 return 0;
1530}
1531
1532static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1533{
1534 struct virtio_net_hdr_mrg_rxbuf *hdr;
1535 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1536 struct virtnet_info *vi = sq->vq->vdev->priv;
1537 int num_sg;
1538 unsigned hdr_len = vi->hdr_len;
1539 bool can_push;
1540
1541 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1542
1543 can_push = vi->any_header_sg &&
1544 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1545 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1546 /* Even if we can, don't push here yet as this would skew
1547 * csum_start offset below. */
1548 if (can_push)
1549 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1550 else
1551 hdr = skb_vnet_hdr(skb);
1552
1553 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1554 virtio_is_little_endian(vi->vdev), false,
1555 0))
1556 BUG();
1557
1558 if (vi->mergeable_rx_bufs)
1559 hdr->num_buffers = 0;
1560
1561 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1562 if (can_push) {
1563 __skb_push(skb, hdr_len);
1564 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1565 if (unlikely(num_sg < 0))
1566 return num_sg;
1567 /* Pull header back to avoid skew in tx bytes calculations. */
1568 __skb_pull(skb, hdr_len);
1569 } else {
1570 sg_set_buf(sq->sg, hdr, hdr_len);
1571 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1572 if (unlikely(num_sg < 0))
1573 return num_sg;
1574 num_sg++;
1575 }
1576 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1577}
1578
1579static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1580{
1581 struct virtnet_info *vi = netdev_priv(dev);
1582 int qnum = skb_get_queue_mapping(skb);
1583 struct send_queue *sq = &vi->sq[qnum];
1584 int err;
1585 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1586 bool kick = !netdev_xmit_more();
1587 bool use_napi = sq->napi.weight;
1588
1589 /* Free up any pending old buffers before queueing new ones. */
1590 free_old_xmit_skbs(sq, false);
1591
1592 if (use_napi && kick)
1593 virtqueue_enable_cb_delayed(sq->vq);
1594
1595 /* timestamp packet in software */
1596 skb_tx_timestamp(skb);
1597
1598 /* Try to transmit */
1599 err = xmit_skb(sq, skb);
1600
1601 /* This should not happen! */
1602 if (unlikely(err)) {
1603 dev->stats.tx_fifo_errors++;
1604 if (net_ratelimit())
1605 dev_warn(&dev->dev,
1606 "Unexpected TXQ (%d) queue failure: %d\n",
1607 qnum, err);
1608 dev->stats.tx_dropped++;
1609 dev_kfree_skb_any(skb);
1610 return NETDEV_TX_OK;
1611 }
1612
1613 /* Don't wait up for transmitted skbs to be freed. */
1614 if (!use_napi) {
1615 skb_orphan(skb);
1616 nf_reset_ct(skb);
1617 }
1618
1619 /* If running out of space, stop queue to avoid getting packets that we
1620 * are then unable to transmit.
1621 * An alternative would be to force queuing layer to requeue the skb by
1622 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1623 * returned in a normal path of operation: it means that driver is not
1624 * maintaining the TX queue stop/start state properly, and causes
1625 * the stack to do a non-trivial amount of useless work.
1626 * Since most packets only take 1 or 2 ring slots, stopping the queue
1627 * early means 16 slots are typically wasted.
1628 */
1629 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1630 netif_stop_subqueue(dev, qnum);
1631 if (!use_napi &&
1632 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1633 /* More just got used, free them then recheck. */
1634 free_old_xmit_skbs(sq, false);
1635 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1636 netif_start_subqueue(dev, qnum);
1637 virtqueue_disable_cb(sq->vq);
1638 }
1639 }
1640 }
1641
1642 if (kick || netif_xmit_stopped(txq)) {
1643 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1644 u64_stats_update_begin(&sq->stats.syncp);
1645 sq->stats.kicks++;
1646 u64_stats_update_end(&sq->stats.syncp);
1647 }
1648 }
1649
1650 return NETDEV_TX_OK;
1651}
1652
1653/*
1654 * Send command via the control virtqueue and check status. Commands
1655 * supported by the hypervisor, as indicated by feature bits, should
1656 * never fail unless improperly formatted.
1657 */
1658static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1659 struct scatterlist *out)
1660{
1661 struct scatterlist *sgs[4], hdr, stat;
1662 unsigned out_num = 0, tmp;
1663
1664 /* Caller should know better */
1665 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1666
1667 vi->ctrl->status = ~0;
1668 vi->ctrl->hdr.class = class;
1669 vi->ctrl->hdr.cmd = cmd;
1670 /* Add header */
1671 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1672 sgs[out_num++] = &hdr;
1673
1674 if (out)
1675 sgs[out_num++] = out;
1676
1677 /* Add return status. */
1678 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1679 sgs[out_num] = &stat;
1680
1681 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1682 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1683
1684 if (unlikely(!virtqueue_kick(vi->cvq)))
1685 return vi->ctrl->status == VIRTIO_NET_OK;
1686
1687 /* Spin for a response, the kick causes an ioport write, trapping
1688 * into the hypervisor, so the request should be handled immediately.
1689 */
1690 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1691 !virtqueue_is_broken(vi->cvq))
1692 cpu_relax();
1693
1694 return vi->ctrl->status == VIRTIO_NET_OK;
1695}
1696
1697static int virtnet_set_mac_address(struct net_device *dev, void *p)
1698{
1699 struct virtnet_info *vi = netdev_priv(dev);
1700 struct virtio_device *vdev = vi->vdev;
1701 int ret;
1702 struct sockaddr *addr;
1703 struct scatterlist sg;
1704
1705 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1706 return -EOPNOTSUPP;
1707
1708 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1709 if (!addr)
1710 return -ENOMEM;
1711
1712 ret = eth_prepare_mac_addr_change(dev, addr);
1713 if (ret)
1714 goto out;
1715
1716 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1717 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1718 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1719 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1720 dev_warn(&vdev->dev,
1721 "Failed to set mac address by vq command.\n");
1722 ret = -EINVAL;
1723 goto out;
1724 }
1725 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1726 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1727 unsigned int i;
1728
1729 /* Naturally, this has an atomicity problem. */
1730 for (i = 0; i < dev->addr_len; i++)
1731 virtio_cwrite8(vdev,
1732 offsetof(struct virtio_net_config, mac) +
1733 i, addr->sa_data[i]);
1734 }
1735
1736 eth_commit_mac_addr_change(dev, p);
1737 ret = 0;
1738
1739out:
1740 kfree(addr);
1741 return ret;
1742}
1743
1744static void virtnet_stats(struct net_device *dev,
1745 struct rtnl_link_stats64 *tot)
1746{
1747 struct virtnet_info *vi = netdev_priv(dev);
1748 unsigned int start;
1749 int i;
1750
1751 for (i = 0; i < vi->max_queue_pairs; i++) {
1752 u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1753 struct receive_queue *rq = &vi->rq[i];
1754 struct send_queue *sq = &vi->sq[i];
1755
1756 do {
1757 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1758 tpackets = sq->stats.packets;
1759 tbytes = sq->stats.bytes;
1760 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1761
1762 do {
1763 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1764 rpackets = rq->stats.packets;
1765 rbytes = rq->stats.bytes;
1766 rdrops = rq->stats.drops;
1767 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1768
1769 tot->rx_packets += rpackets;
1770 tot->tx_packets += tpackets;
1771 tot->rx_bytes += rbytes;
1772 tot->tx_bytes += tbytes;
1773 tot->rx_dropped += rdrops;
1774 }
1775
1776 tot->tx_dropped = dev->stats.tx_dropped;
1777 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1778 tot->rx_length_errors = dev->stats.rx_length_errors;
1779 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1780}
1781
1782static void virtnet_ack_link_announce(struct virtnet_info *vi)
1783{
1784 rtnl_lock();
1785 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1786 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1787 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1788 rtnl_unlock();
1789}
1790
1791static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1792{
1793 struct scatterlist sg;
1794 struct net_device *dev = vi->dev;
1795
1796 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1797 return 0;
1798
1799 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1800 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1801
1802 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1803 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1804 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1805 queue_pairs);
1806 return -EINVAL;
1807 } else {
1808 vi->curr_queue_pairs = queue_pairs;
1809 /* virtnet_open() will refill when device is going to up. */
1810 if (dev->flags & IFF_UP)
1811 schedule_delayed_work(&vi->refill, 0);
1812 }
1813
1814 return 0;
1815}
1816
1817static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1818{
1819 int err;
1820
1821 rtnl_lock();
1822 err = _virtnet_set_queues(vi, queue_pairs);
1823 rtnl_unlock();
1824 return err;
1825}
1826
1827static int virtnet_close(struct net_device *dev)
1828{
1829 struct virtnet_info *vi = netdev_priv(dev);
1830 int i;
1831
1832 /* Make sure refill_work doesn't re-enable napi! */
1833 cancel_delayed_work_sync(&vi->refill);
1834
1835 for (i = 0; i < vi->max_queue_pairs; i++) {
1836 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1837 napi_disable(&vi->rq[i].napi);
1838 virtnet_napi_tx_disable(&vi->sq[i].napi);
1839 }
1840
1841 return 0;
1842}
1843
1844static void virtnet_set_rx_mode(struct net_device *dev)
1845{
1846 struct virtnet_info *vi = netdev_priv(dev);
1847 struct scatterlist sg[2];
1848 struct virtio_net_ctrl_mac *mac_data;
1849 struct netdev_hw_addr *ha;
1850 int uc_count;
1851 int mc_count;
1852 void *buf;
1853 int i;
1854
1855 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1856 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1857 return;
1858
1859 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1860 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1861
1862 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1863
1864 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1865 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1866 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1867 vi->ctrl->promisc ? "en" : "dis");
1868
1869 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1870
1871 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1872 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1873 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1874 vi->ctrl->allmulti ? "en" : "dis");
1875
1876 uc_count = netdev_uc_count(dev);
1877 mc_count = netdev_mc_count(dev);
1878 /* MAC filter - use one buffer for both lists */
1879 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1880 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1881 mac_data = buf;
1882 if (!buf)
1883 return;
1884
1885 sg_init_table(sg, 2);
1886
1887 /* Store the unicast list and count in the front of the buffer */
1888 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1889 i = 0;
1890 netdev_for_each_uc_addr(ha, dev)
1891 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1892
1893 sg_set_buf(&sg[0], mac_data,
1894 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1895
1896 /* multicast list and count fill the end */
1897 mac_data = (void *)&mac_data->macs[uc_count][0];
1898
1899 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1900 i = 0;
1901 netdev_for_each_mc_addr(ha, dev)
1902 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1903
1904 sg_set_buf(&sg[1], mac_data,
1905 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1906
1907 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1908 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1909 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1910
1911 kfree(buf);
1912}
1913
1914static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1915 __be16 proto, u16 vid)
1916{
1917 struct virtnet_info *vi = netdev_priv(dev);
1918 struct scatterlist sg;
1919
1920 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1921 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1922
1923 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1924 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1925 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1926 return 0;
1927}
1928
1929static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1930 __be16 proto, u16 vid)
1931{
1932 struct virtnet_info *vi = netdev_priv(dev);
1933 struct scatterlist sg;
1934
1935 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1936 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1937
1938 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1939 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1940 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1941 return 0;
1942}
1943
1944static void virtnet_clean_affinity(struct virtnet_info *vi)
1945{
1946 int i;
1947
1948 if (vi->affinity_hint_set) {
1949 for (i = 0; i < vi->max_queue_pairs; i++) {
1950 virtqueue_set_affinity(vi->rq[i].vq, NULL);
1951 virtqueue_set_affinity(vi->sq[i].vq, NULL);
1952 }
1953
1954 vi->affinity_hint_set = false;
1955 }
1956}
1957
1958static void virtnet_set_affinity(struct virtnet_info *vi)
1959{
1960 cpumask_var_t mask;
1961 int stragglers;
1962 int group_size;
1963 int i, j, cpu;
1964 int num_cpu;
1965 int stride;
1966
1967 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
1968 virtnet_clean_affinity(vi);
1969 return;
1970 }
1971
1972 num_cpu = num_online_cpus();
1973 stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
1974 stragglers = num_cpu >= vi->curr_queue_pairs ?
1975 num_cpu % vi->curr_queue_pairs :
1976 0;
1977 cpu = cpumask_next(-1, cpu_online_mask);
1978
1979 for (i = 0; i < vi->curr_queue_pairs; i++) {
1980 group_size = stride + (i < stragglers ? 1 : 0);
1981
1982 for (j = 0; j < group_size; j++) {
1983 cpumask_set_cpu(cpu, mask);
1984 cpu = cpumask_next_wrap(cpu, cpu_online_mask,
1985 nr_cpu_ids, false);
1986 }
1987 virtqueue_set_affinity(vi->rq[i].vq, mask);
1988 virtqueue_set_affinity(vi->sq[i].vq, mask);
1989 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, false);
1990 cpumask_clear(mask);
1991 }
1992
1993 vi->affinity_hint_set = true;
1994 free_cpumask_var(mask);
1995}
1996
1997static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1998{
1999 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2000 node);
2001 virtnet_set_affinity(vi);
2002 return 0;
2003}
2004
2005static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2006{
2007 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2008 node_dead);
2009 virtnet_set_affinity(vi);
2010 return 0;
2011}
2012
2013static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2014{
2015 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2016 node);
2017
2018 virtnet_clean_affinity(vi);
2019 return 0;
2020}
2021
2022static enum cpuhp_state virtionet_online;
2023
2024static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2025{
2026 int ret;
2027
2028 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2029 if (ret)
2030 return ret;
2031 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2032 &vi->node_dead);
2033 if (!ret)
2034 return ret;
2035 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2036 return ret;
2037}
2038
2039static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2040{
2041 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2042 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2043 &vi->node_dead);
2044}
2045
2046static void virtnet_get_ringparam(struct net_device *dev,
2047 struct ethtool_ringparam *ring)
2048{
2049 struct virtnet_info *vi = netdev_priv(dev);
2050
2051 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2052 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2053 ring->rx_pending = ring->rx_max_pending;
2054 ring->tx_pending = ring->tx_max_pending;
2055}
2056
2057
2058static void virtnet_get_drvinfo(struct net_device *dev,
2059 struct ethtool_drvinfo *info)
2060{
2061 struct virtnet_info *vi = netdev_priv(dev);
2062 struct virtio_device *vdev = vi->vdev;
2063
2064 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2065 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2066 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2067
2068}
2069
2070/* TODO: Eliminate OOO packets during switching */
2071static int virtnet_set_channels(struct net_device *dev,
2072 struct ethtool_channels *channels)
2073{
2074 struct virtnet_info *vi = netdev_priv(dev);
2075 u16 queue_pairs = channels->combined_count;
2076 int err;
2077
2078 /* We don't support separate rx/tx channels.
2079 * We don't allow setting 'other' channels.
2080 */
2081 if (channels->rx_count || channels->tx_count || channels->other_count)
2082 return -EINVAL;
2083
2084 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2085 return -EINVAL;
2086
2087 /* For now we don't support modifying channels while XDP is loaded
2088 * also when XDP is loaded all RX queues have XDP programs so we only
2089 * need to check a single RX queue.
2090 */
2091 if (vi->rq[0].xdp_prog)
2092 return -EINVAL;
2093
2094 get_online_cpus();
2095 err = _virtnet_set_queues(vi, queue_pairs);
2096 if (!err) {
2097 netif_set_real_num_tx_queues(dev, queue_pairs);
2098 netif_set_real_num_rx_queues(dev, queue_pairs);
2099
2100 virtnet_set_affinity(vi);
2101 }
2102 put_online_cpus();
2103
2104 return err;
2105}
2106
2107static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2108{
2109 struct virtnet_info *vi = netdev_priv(dev);
2110 char *p = (char *)data;
2111 unsigned int i, j;
2112
2113 switch (stringset) {
2114 case ETH_SS_STATS:
2115 for (i = 0; i < vi->curr_queue_pairs; i++) {
2116 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2117 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s",
2118 i, virtnet_rq_stats_desc[j].desc);
2119 p += ETH_GSTRING_LEN;
2120 }
2121 }
2122
2123 for (i = 0; i < vi->curr_queue_pairs; i++) {
2124 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2125 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s",
2126 i, virtnet_sq_stats_desc[j].desc);
2127 p += ETH_GSTRING_LEN;
2128 }
2129 }
2130 break;
2131 }
2132}
2133
2134static int virtnet_get_sset_count(struct net_device *dev, int sset)
2135{
2136 struct virtnet_info *vi = netdev_priv(dev);
2137
2138 switch (sset) {
2139 case ETH_SS_STATS:
2140 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2141 VIRTNET_SQ_STATS_LEN);
2142 default:
2143 return -EOPNOTSUPP;
2144 }
2145}
2146
2147static void virtnet_get_ethtool_stats(struct net_device *dev,
2148 struct ethtool_stats *stats, u64 *data)
2149{
2150 struct virtnet_info *vi = netdev_priv(dev);
2151 unsigned int idx = 0, start, i, j;
2152 const u8 *stats_base;
2153 size_t offset;
2154
2155 for (i = 0; i < vi->curr_queue_pairs; i++) {
2156 struct receive_queue *rq = &vi->rq[i];
2157
2158 stats_base = (u8 *)&rq->stats;
2159 do {
2160 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2161 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2162 offset = virtnet_rq_stats_desc[j].offset;
2163 data[idx + j] = *(u64 *)(stats_base + offset);
2164 }
2165 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2166 idx += VIRTNET_RQ_STATS_LEN;
2167 }
2168
2169 for (i = 0; i < vi->curr_queue_pairs; i++) {
2170 struct send_queue *sq = &vi->sq[i];
2171
2172 stats_base = (u8 *)&sq->stats;
2173 do {
2174 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2175 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2176 offset = virtnet_sq_stats_desc[j].offset;
2177 data[idx + j] = *(u64 *)(stats_base + offset);
2178 }
2179 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2180 idx += VIRTNET_SQ_STATS_LEN;
2181 }
2182}
2183
2184static void virtnet_get_channels(struct net_device *dev,
2185 struct ethtool_channels *channels)
2186{
2187 struct virtnet_info *vi = netdev_priv(dev);
2188
2189 channels->combined_count = vi->curr_queue_pairs;
2190 channels->max_combined = vi->max_queue_pairs;
2191 channels->max_other = 0;
2192 channels->rx_count = 0;
2193 channels->tx_count = 0;
2194 channels->other_count = 0;
2195}
2196
2197static int virtnet_set_link_ksettings(struct net_device *dev,
2198 const struct ethtool_link_ksettings *cmd)
2199{
2200 struct virtnet_info *vi = netdev_priv(dev);
2201
2202 return ethtool_virtdev_set_link_ksettings(dev, cmd,
2203 &vi->speed, &vi->duplex);
2204}
2205
2206static int virtnet_get_link_ksettings(struct net_device *dev,
2207 struct ethtool_link_ksettings *cmd)
2208{
2209 struct virtnet_info *vi = netdev_priv(dev);
2210
2211 cmd->base.speed = vi->speed;
2212 cmd->base.duplex = vi->duplex;
2213 cmd->base.port = PORT_OTHER;
2214
2215 return 0;
2216}
2217
2218static int virtnet_set_coalesce(struct net_device *dev,
2219 struct ethtool_coalesce *ec)
2220{
2221 struct virtnet_info *vi = netdev_priv(dev);
2222 int i, napi_weight;
2223
2224 if (ec->tx_max_coalesced_frames > 1 ||
2225 ec->rx_max_coalesced_frames != 1)
2226 return -EINVAL;
2227
2228 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2229 if (napi_weight ^ vi->sq[0].napi.weight) {
2230 if (dev->flags & IFF_UP)
2231 return -EBUSY;
2232 for (i = 0; i < vi->max_queue_pairs; i++)
2233 vi->sq[i].napi.weight = napi_weight;
2234 }
2235
2236 return 0;
2237}
2238
2239static int virtnet_get_coalesce(struct net_device *dev,
2240 struct ethtool_coalesce *ec)
2241{
2242 struct ethtool_coalesce ec_default = {
2243 .cmd = ETHTOOL_GCOALESCE,
2244 .rx_max_coalesced_frames = 1,
2245 };
2246 struct virtnet_info *vi = netdev_priv(dev);
2247
2248 memcpy(ec, &ec_default, sizeof(ec_default));
2249
2250 if (vi->sq[0].napi.weight)
2251 ec->tx_max_coalesced_frames = 1;
2252
2253 return 0;
2254}
2255
2256static void virtnet_init_settings(struct net_device *dev)
2257{
2258 struct virtnet_info *vi = netdev_priv(dev);
2259
2260 vi->speed = SPEED_UNKNOWN;
2261 vi->duplex = DUPLEX_UNKNOWN;
2262}
2263
2264static void virtnet_update_settings(struct virtnet_info *vi)
2265{
2266 u32 speed;
2267 u8 duplex;
2268
2269 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2270 return;
2271
2272 virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
2273
2274 if (ethtool_validate_speed(speed))
2275 vi->speed = speed;
2276
2277 virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
2278
2279 if (ethtool_validate_duplex(duplex))
2280 vi->duplex = duplex;
2281}
2282
2283static const struct ethtool_ops virtnet_ethtool_ops = {
2284 .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES,
2285 .get_drvinfo = virtnet_get_drvinfo,
2286 .get_link = ethtool_op_get_link,
2287 .get_ringparam = virtnet_get_ringparam,
2288 .get_strings = virtnet_get_strings,
2289 .get_sset_count = virtnet_get_sset_count,
2290 .get_ethtool_stats = virtnet_get_ethtool_stats,
2291 .set_channels = virtnet_set_channels,
2292 .get_channels = virtnet_get_channels,
2293 .get_ts_info = ethtool_op_get_ts_info,
2294 .get_link_ksettings = virtnet_get_link_ksettings,
2295 .set_link_ksettings = virtnet_set_link_ksettings,
2296 .set_coalesce = virtnet_set_coalesce,
2297 .get_coalesce = virtnet_get_coalesce,
2298};
2299
2300static void virtnet_freeze_down(struct virtio_device *vdev)
2301{
2302 struct virtnet_info *vi = vdev->priv;
2303 int i;
2304
2305 /* Make sure no work handler is accessing the device */
2306 flush_work(&vi->config_work);
2307
2308 netif_tx_lock_bh(vi->dev);
2309 netif_device_detach(vi->dev);
2310 netif_tx_unlock_bh(vi->dev);
2311 cancel_delayed_work_sync(&vi->refill);
2312
2313 if (netif_running(vi->dev)) {
2314 for (i = 0; i < vi->max_queue_pairs; i++) {
2315 napi_disable(&vi->rq[i].napi);
2316 virtnet_napi_tx_disable(&vi->sq[i].napi);
2317 }
2318 }
2319}
2320
2321static int init_vqs(struct virtnet_info *vi);
2322
2323static int virtnet_restore_up(struct virtio_device *vdev)
2324{
2325 struct virtnet_info *vi = vdev->priv;
2326 int err, i;
2327
2328 err = init_vqs(vi);
2329 if (err)
2330 return err;
2331
2332 virtio_device_ready(vdev);
2333
2334 if (netif_running(vi->dev)) {
2335 for (i = 0; i < vi->curr_queue_pairs; i++)
2336 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2337 schedule_delayed_work(&vi->refill, 0);
2338
2339 for (i = 0; i < vi->max_queue_pairs; i++) {
2340 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2341 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2342 &vi->sq[i].napi);
2343 }
2344 }
2345
2346 netif_tx_lock_bh(vi->dev);
2347 netif_device_attach(vi->dev);
2348 netif_tx_unlock_bh(vi->dev);
2349 return err;
2350}
2351
2352static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2353{
2354 struct scatterlist sg;
2355 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2356
2357 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2358
2359 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2360 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2361 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
2362 return -EINVAL;
2363 }
2364
2365 return 0;
2366}
2367
2368static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2369{
2370 u64 offloads = 0;
2371
2372 if (!vi->guest_offloads)
2373 return 0;
2374
2375 return virtnet_set_guest_offloads(vi, offloads);
2376}
2377
2378static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2379{
2380 u64 offloads = vi->guest_offloads;
2381
2382 if (!vi->guest_offloads)
2383 return 0;
2384
2385 return virtnet_set_guest_offloads(vi, offloads);
2386}
2387
2388static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2389 struct netlink_ext_ack *extack)
2390{
2391 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2392 struct virtnet_info *vi = netdev_priv(dev);
2393 struct bpf_prog *old_prog;
2394 u16 xdp_qp = 0, curr_qp;
2395 int i, err;
2396
2397 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2398 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2399 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2400 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2401 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2402 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2403 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
2404 return -EOPNOTSUPP;
2405 }
2406
2407 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2408 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2409 return -EINVAL;
2410 }
2411
2412 if (dev->mtu > max_sz) {
2413 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2414 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2415 return -EINVAL;
2416 }
2417
2418 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2419 if (prog)
2420 xdp_qp = nr_cpu_ids;
2421
2422 /* XDP requires extra queues for XDP_TX */
2423 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2424 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2425 netdev_warn(dev, "request %i queues but max is %i\n",
2426 curr_qp + xdp_qp, vi->max_queue_pairs);
2427 return -ENOMEM;
2428 }
2429
2430 old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2431 if (!prog && !old_prog)
2432 return 0;
2433
2434 if (prog)
2435 bpf_prog_add(prog, vi->max_queue_pairs - 1);
2436
2437 /* Make sure NAPI is not using any XDP TX queues for RX. */
2438 if (netif_running(dev)) {
2439 for (i = 0; i < vi->max_queue_pairs; i++) {
2440 napi_disable(&vi->rq[i].napi);
2441 virtnet_napi_tx_disable(&vi->sq[i].napi);
2442 }
2443 }
2444
2445 if (!prog) {
2446 for (i = 0; i < vi->max_queue_pairs; i++) {
2447 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2448 if (i == 0)
2449 virtnet_restore_guest_offloads(vi);
2450 }
2451 synchronize_net();
2452 }
2453
2454 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2455 if (err)
2456 goto err;
2457 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2458 vi->xdp_queue_pairs = xdp_qp;
2459
2460 if (prog) {
2461 for (i = 0; i < vi->max_queue_pairs; i++) {
2462 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2463 if (i == 0 && !old_prog)
2464 virtnet_clear_guest_offloads(vi);
2465 }
2466 }
2467
2468 for (i = 0; i < vi->max_queue_pairs; i++) {
2469 if (old_prog)
2470 bpf_prog_put(old_prog);
2471 if (netif_running(dev)) {
2472 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2473 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2474 &vi->sq[i].napi);
2475 }
2476 }
2477
2478 return 0;
2479
2480err:
2481 if (!prog) {
2482 virtnet_clear_guest_offloads(vi);
2483 for (i = 0; i < vi->max_queue_pairs; i++)
2484 rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2485 }
2486
2487 if (netif_running(dev)) {
2488 for (i = 0; i < vi->max_queue_pairs; i++) {
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 if (prog)
2495 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2496 return err;
2497}
2498
2499static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2500{
2501 switch (xdp->command) {
2502 case XDP_SETUP_PROG:
2503 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2504 default:
2505 return -EINVAL;
2506 }
2507}
2508
2509static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2510 size_t len)
2511{
2512 struct virtnet_info *vi = netdev_priv(dev);
2513 int ret;
2514
2515 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2516 return -EOPNOTSUPP;
2517
2518 ret = snprintf(buf, len, "sby");
2519 if (ret >= len)
2520 return -EOPNOTSUPP;
2521
2522 return 0;
2523}
2524
2525static int virtnet_set_features(struct net_device *dev,
2526 netdev_features_t features)
2527{
2528 struct virtnet_info *vi = netdev_priv(dev);
2529 u64 offloads;
2530 int err;
2531
2532 if ((dev->features ^ features) & NETIF_F_LRO) {
2533 if (vi->xdp_queue_pairs)
2534 return -EBUSY;
2535
2536 if (features & NETIF_F_LRO)
2537 offloads = vi->guest_offloads_capable;
2538 else
2539 offloads = vi->guest_offloads_capable &
2540 ~GUEST_OFFLOAD_LRO_MASK;
2541
2542 err = virtnet_set_guest_offloads(vi, offloads);
2543 if (err)
2544 return err;
2545 vi->guest_offloads = offloads;
2546 }
2547
2548 return 0;
2549}
2550
2551static const struct net_device_ops virtnet_netdev = {
2552 .ndo_open = virtnet_open,
2553 .ndo_stop = virtnet_close,
2554 .ndo_start_xmit = start_xmit,
2555 .ndo_validate_addr = eth_validate_addr,
2556 .ndo_set_mac_address = virtnet_set_mac_address,
2557 .ndo_set_rx_mode = virtnet_set_rx_mode,
2558 .ndo_get_stats64 = virtnet_stats,
2559 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2560 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2561 .ndo_bpf = virtnet_xdp,
2562 .ndo_xdp_xmit = virtnet_xdp_xmit,
2563 .ndo_features_check = passthru_features_check,
2564 .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2565 .ndo_set_features = virtnet_set_features,
2566};
2567
2568static void virtnet_config_changed_work(struct work_struct *work)
2569{
2570 struct virtnet_info *vi =
2571 container_of(work, struct virtnet_info, config_work);
2572 u16 v;
2573
2574 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2575 struct virtio_net_config, status, &v) < 0)
2576 return;
2577
2578 if (v & VIRTIO_NET_S_ANNOUNCE) {
2579 netdev_notify_peers(vi->dev);
2580 virtnet_ack_link_announce(vi);
2581 }
2582
2583 /* Ignore unknown (future) status bits */
2584 v &= VIRTIO_NET_S_LINK_UP;
2585
2586 if (vi->status == v)
2587 return;
2588
2589 vi->status = v;
2590
2591 if (vi->status & VIRTIO_NET_S_LINK_UP) {
2592 virtnet_update_settings(vi);
2593 netif_carrier_on(vi->dev);
2594 netif_tx_wake_all_queues(vi->dev);
2595 } else {
2596 netif_carrier_off(vi->dev);
2597 netif_tx_stop_all_queues(vi->dev);
2598 }
2599}
2600
2601static void virtnet_config_changed(struct virtio_device *vdev)
2602{
2603 struct virtnet_info *vi = vdev->priv;
2604
2605 schedule_work(&vi->config_work);
2606}
2607
2608static void virtnet_free_queues(struct virtnet_info *vi)
2609{
2610 int i;
2611
2612 for (i = 0; i < vi->max_queue_pairs; i++) {
2613 napi_hash_del(&vi->rq[i].napi);
2614 netif_napi_del(&vi->rq[i].napi);
2615 netif_napi_del(&vi->sq[i].napi);
2616 }
2617
2618 /* We called napi_hash_del() before netif_napi_del(),
2619 * we need to respect an RCU grace period before freeing vi->rq
2620 */
2621 synchronize_net();
2622
2623 kfree(vi->rq);
2624 kfree(vi->sq);
2625 kfree(vi->ctrl);
2626}
2627
2628static void _free_receive_bufs(struct virtnet_info *vi)
2629{
2630 struct bpf_prog *old_prog;
2631 int i;
2632
2633 for (i = 0; i < vi->max_queue_pairs; i++) {
2634 while (vi->rq[i].pages)
2635 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2636
2637 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2638 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2639 if (old_prog)
2640 bpf_prog_put(old_prog);
2641 }
2642}
2643
2644static void free_receive_bufs(struct virtnet_info *vi)
2645{
2646 rtnl_lock();
2647 _free_receive_bufs(vi);
2648 rtnl_unlock();
2649}
2650
2651static void free_receive_page_frags(struct virtnet_info *vi)
2652{
2653 int i;
2654 for (i = 0; i < vi->max_queue_pairs; i++)
2655 if (vi->rq[i].alloc_frag.page)
2656 put_page(vi->rq[i].alloc_frag.page);
2657}
2658
2659static void free_unused_bufs(struct virtnet_info *vi)
2660{
2661 void *buf;
2662 int i;
2663
2664 for (i = 0; i < vi->max_queue_pairs; i++) {
2665 struct virtqueue *vq = vi->sq[i].vq;
2666 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2667 if (!is_xdp_frame(buf))
2668 dev_kfree_skb(buf);
2669 else
2670 xdp_return_frame(ptr_to_xdp(buf));
2671 }
2672 }
2673
2674 for (i = 0; i < vi->max_queue_pairs; i++) {
2675 struct virtqueue *vq = vi->rq[i].vq;
2676
2677 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2678 if (vi->mergeable_rx_bufs) {
2679 put_page(virt_to_head_page(buf));
2680 } else if (vi->big_packets) {
2681 give_pages(&vi->rq[i], buf);
2682 } else {
2683 put_page(virt_to_head_page(buf));
2684 }
2685 }
2686 }
2687}
2688
2689static void virtnet_del_vqs(struct virtnet_info *vi)
2690{
2691 struct virtio_device *vdev = vi->vdev;
2692
2693 virtnet_clean_affinity(vi);
2694
2695 vdev->config->del_vqs(vdev);
2696
2697 virtnet_free_queues(vi);
2698}
2699
2700/* How large should a single buffer be so a queue full of these can fit at
2701 * least one full packet?
2702 * Logic below assumes the mergeable buffer header is used.
2703 */
2704static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2705{
2706 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2707 unsigned int rq_size = virtqueue_get_vring_size(vq);
2708 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2709 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2710 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2711
2712 return max(max(min_buf_len, hdr_len) - hdr_len,
2713 (unsigned int)GOOD_PACKET_LEN);
2714}
2715
2716static int virtnet_find_vqs(struct virtnet_info *vi)
2717{
2718 vq_callback_t **callbacks;
2719 struct virtqueue **vqs;
2720 int ret = -ENOMEM;
2721 int i, total_vqs;
2722 const char **names;
2723 bool *ctx;
2724
2725 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2726 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2727 * possible control vq.
2728 */
2729 total_vqs = vi->max_queue_pairs * 2 +
2730 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2731
2732 /* Allocate space for find_vqs parameters */
2733 vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2734 if (!vqs)
2735 goto err_vq;
2736 callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2737 if (!callbacks)
2738 goto err_callback;
2739 names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2740 if (!names)
2741 goto err_names;
2742 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2743 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2744 if (!ctx)
2745 goto err_ctx;
2746 } else {
2747 ctx = NULL;
2748 }
2749
2750 /* Parameters for control virtqueue, if any */
2751 if (vi->has_cvq) {
2752 callbacks[total_vqs - 1] = NULL;
2753 names[total_vqs - 1] = "control";
2754 }
2755
2756 /* Allocate/initialize parameters for send/receive virtqueues */
2757 for (i = 0; i < vi->max_queue_pairs; i++) {
2758 callbacks[rxq2vq(i)] = skb_recv_done;
2759 callbacks[txq2vq(i)] = skb_xmit_done;
2760 sprintf(vi->rq[i].name, "input.%d", i);
2761 sprintf(vi->sq[i].name, "output.%d", i);
2762 names[rxq2vq(i)] = vi->rq[i].name;
2763 names[txq2vq(i)] = vi->sq[i].name;
2764 if (ctx)
2765 ctx[rxq2vq(i)] = true;
2766 }
2767
2768 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2769 names, ctx, NULL);
2770 if (ret)
2771 goto err_find;
2772
2773 if (vi->has_cvq) {
2774 vi->cvq = vqs[total_vqs - 1];
2775 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2776 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2777 }
2778
2779 for (i = 0; i < vi->max_queue_pairs; i++) {
2780 vi->rq[i].vq = vqs[rxq2vq(i)];
2781 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2782 vi->sq[i].vq = vqs[txq2vq(i)];
2783 }
2784
2785 /* run here: ret == 0. */
2786
2787
2788err_find:
2789 kfree(ctx);
2790err_ctx:
2791 kfree(names);
2792err_names:
2793 kfree(callbacks);
2794err_callback:
2795 kfree(vqs);
2796err_vq:
2797 return ret;
2798}
2799
2800static int virtnet_alloc_queues(struct virtnet_info *vi)
2801{
2802 int i;
2803
2804 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2805 if (!vi->ctrl)
2806 goto err_ctrl;
2807 vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2808 if (!vi->sq)
2809 goto err_sq;
2810 vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2811 if (!vi->rq)
2812 goto err_rq;
2813
2814 INIT_DELAYED_WORK(&vi->refill, refill_work);
2815 for (i = 0; i < vi->max_queue_pairs; i++) {
2816 vi->rq[i].pages = NULL;
2817 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2818 napi_weight);
2819 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2820 napi_tx ? napi_weight : 0);
2821
2822 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2823 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2824 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2825
2826 u64_stats_init(&vi->rq[i].stats.syncp);
2827 u64_stats_init(&vi->sq[i].stats.syncp);
2828 }
2829
2830 return 0;
2831
2832err_rq:
2833 kfree(vi->sq);
2834err_sq:
2835 kfree(vi->ctrl);
2836err_ctrl:
2837 return -ENOMEM;
2838}
2839
2840static int init_vqs(struct virtnet_info *vi)
2841{
2842 int ret;
2843
2844 /* Allocate send & receive queues */
2845 ret = virtnet_alloc_queues(vi);
2846 if (ret)
2847 goto err;
2848
2849 ret = virtnet_find_vqs(vi);
2850 if (ret)
2851 goto err_free;
2852
2853 get_online_cpus();
2854 virtnet_set_affinity(vi);
2855 put_online_cpus();
2856
2857 return 0;
2858
2859err_free:
2860 virtnet_free_queues(vi);
2861err:
2862 return ret;
2863}
2864
2865#ifdef CONFIG_SYSFS
2866static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2867 char *buf)
2868{
2869 struct virtnet_info *vi = netdev_priv(queue->dev);
2870 unsigned int queue_index = get_netdev_rx_queue_index(queue);
2871 unsigned int headroom = virtnet_get_headroom(vi);
2872 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2873 struct ewma_pkt_len *avg;
2874
2875 BUG_ON(queue_index >= vi->max_queue_pairs);
2876 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2877 return sprintf(buf, "%u\n",
2878 get_mergeable_buf_len(&vi->rq[queue_index], avg,
2879 SKB_DATA_ALIGN(headroom + tailroom)));
2880}
2881
2882static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2883 __ATTR_RO(mergeable_rx_buffer_size);
2884
2885static struct attribute *virtio_net_mrg_rx_attrs[] = {
2886 &mergeable_rx_buffer_size_attribute.attr,
2887 NULL
2888};
2889
2890static const struct attribute_group virtio_net_mrg_rx_group = {
2891 .name = "virtio_net",
2892 .attrs = virtio_net_mrg_rx_attrs
2893};
2894#endif
2895
2896static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2897 unsigned int fbit,
2898 const char *fname, const char *dname)
2899{
2900 if (!virtio_has_feature(vdev, fbit))
2901 return false;
2902
2903 dev_err(&vdev->dev, "device advertises feature %s but not %s",
2904 fname, dname);
2905
2906 return true;
2907}
2908
2909#define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2910 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2911
2912static bool virtnet_validate_features(struct virtio_device *vdev)
2913{
2914 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2915 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2916 "VIRTIO_NET_F_CTRL_VQ") ||
2917 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2918 "VIRTIO_NET_F_CTRL_VQ") ||
2919 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2920 "VIRTIO_NET_F_CTRL_VQ") ||
2921 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2922 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2923 "VIRTIO_NET_F_CTRL_VQ"))) {
2924 return false;
2925 }
2926
2927 return true;
2928}
2929
2930#define MIN_MTU ETH_MIN_MTU
2931#define MAX_MTU ETH_MAX_MTU
2932
2933static int virtnet_validate(struct virtio_device *vdev)
2934{
2935 if (!vdev->config->get) {
2936 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2937 __func__);
2938 return -EINVAL;
2939 }
2940
2941 if (!virtnet_validate_features(vdev))
2942 return -EINVAL;
2943
2944 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2945 int mtu = virtio_cread16(vdev,
2946 offsetof(struct virtio_net_config,
2947 mtu));
2948 if (mtu < MIN_MTU)
2949 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2950 }
2951
2952 return 0;
2953}
2954
2955static int virtnet_probe(struct virtio_device *vdev)
2956{
2957 int i, err = -ENOMEM;
2958 struct net_device *dev;
2959 struct virtnet_info *vi;
2960 u16 max_queue_pairs;
2961 int mtu;
2962
2963 /* Find if host supports multiqueue virtio_net device */
2964 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2965 struct virtio_net_config,
2966 max_virtqueue_pairs, &max_queue_pairs);
2967
2968 /* We need at least 2 queue's */
2969 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2970 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2971 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2972 max_queue_pairs = 1;
2973
2974 /* Allocate ourselves a network device with room for our info */
2975 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2976 if (!dev)
2977 return -ENOMEM;
2978
2979 /* Set up network device as normal. */
2980 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2981 dev->netdev_ops = &virtnet_netdev;
2982 dev->features = NETIF_F_HIGHDMA;
2983
2984 dev->ethtool_ops = &virtnet_ethtool_ops;
2985 SET_NETDEV_DEV(dev, &vdev->dev);
2986
2987 /* Do we support "hardware" checksums? */
2988 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2989 /* This opens up the world of extra features. */
2990 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2991 if (csum)
2992 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2993
2994 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2995 dev->hw_features |= NETIF_F_TSO
2996 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2997 }
2998 /* Individual feature bits: what can host handle? */
2999 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3000 dev->hw_features |= NETIF_F_TSO;
3001 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3002 dev->hw_features |= NETIF_F_TSO6;
3003 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3004 dev->hw_features |= NETIF_F_TSO_ECN;
3005
3006 dev->features |= NETIF_F_GSO_ROBUST;
3007
3008 if (gso)
3009 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3010 /* (!csum && gso) case will be fixed by register_netdev() */
3011 }
3012 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3013 dev->features |= NETIF_F_RXCSUM;
3014 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3015 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3016 dev->features |= NETIF_F_LRO;
3017 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3018 dev->hw_features |= NETIF_F_LRO;
3019
3020 dev->vlan_features = dev->features;
3021
3022 /* MTU range: 68 - 65535 */
3023 dev->min_mtu = MIN_MTU;
3024 dev->max_mtu = MAX_MTU;
3025
3026 /* Configuration may specify what MAC to use. Otherwise random. */
3027 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
3028 virtio_cread_bytes(vdev,
3029 offsetof(struct virtio_net_config, mac),
3030 dev->dev_addr, dev->addr_len);
3031 else
3032 eth_hw_addr_random(dev);
3033
3034 /* Set up our device-specific information */
3035 vi = netdev_priv(dev);
3036 vi->dev = dev;
3037 vi->vdev = vdev;
3038 vdev->priv = vi;
3039
3040 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3041
3042 /* If we can receive ANY GSO packets, we must allocate large ones. */
3043 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3044 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3045 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3046 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3047 vi->big_packets = true;
3048
3049 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3050 vi->mergeable_rx_bufs = true;
3051
3052 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3053 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3054 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3055 else
3056 vi->hdr_len = sizeof(struct virtio_net_hdr);
3057
3058 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3059 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3060 vi->any_header_sg = true;
3061
3062 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3063 vi->has_cvq = true;
3064
3065 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3066 mtu = virtio_cread16(vdev,
3067 offsetof(struct virtio_net_config,
3068 mtu));
3069 if (mtu < dev->min_mtu) {
3070 /* Should never trigger: MTU was previously validated
3071 * in virtnet_validate.
3072 */
3073 dev_err(&vdev->dev,
3074 "device MTU appears to have changed it is now %d < %d",
3075 mtu, dev->min_mtu);
3076 goto free;
3077 }
3078
3079 dev->mtu = mtu;
3080 dev->max_mtu = mtu;
3081
3082 /* TODO: size buffers correctly in this case. */
3083 if (dev->mtu > ETH_DATA_LEN)
3084 vi->big_packets = true;
3085 }
3086
3087 if (vi->any_header_sg)
3088 dev->needed_headroom = vi->hdr_len;
3089
3090 /* Enable multiqueue by default */
3091 if (num_online_cpus() >= max_queue_pairs)
3092 vi->curr_queue_pairs = max_queue_pairs;
3093 else
3094 vi->curr_queue_pairs = num_online_cpus();
3095 vi->max_queue_pairs = max_queue_pairs;
3096
3097 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3098 err = init_vqs(vi);
3099 if (err)
3100 goto free;
3101
3102#ifdef CONFIG_SYSFS
3103 if (vi->mergeable_rx_bufs)
3104 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3105#endif
3106 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3107 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3108
3109 virtnet_init_settings(dev);
3110
3111 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3112 vi->failover = net_failover_create(vi->dev);
3113 if (IS_ERR(vi->failover)) {
3114 err = PTR_ERR(vi->failover);
3115 goto free_vqs;
3116 }
3117 }
3118
3119 err = register_netdev(dev);
3120 if (err) {
3121 pr_debug("virtio_net: registering device failed\n");
3122 goto free_failover;
3123 }
3124
3125 virtio_device_ready(vdev);
3126
3127 err = virtnet_cpu_notif_add(vi);
3128 if (err) {
3129 pr_debug("virtio_net: registering cpu notifier failed\n");
3130 goto free_unregister_netdev;
3131 }
3132
3133 virtnet_set_queues(vi, vi->curr_queue_pairs);
3134
3135 /* Assume link up if device can't report link status,
3136 otherwise get link status from config. */
3137 netif_carrier_off(dev);
3138 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3139 schedule_work(&vi->config_work);
3140 } else {
3141 vi->status = VIRTIO_NET_S_LINK_UP;
3142 virtnet_update_settings(vi);
3143 netif_carrier_on(dev);
3144 }
3145
3146 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3147 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3148 set_bit(guest_offloads[i], &vi->guest_offloads);
3149 vi->guest_offloads_capable = vi->guest_offloads;
3150
3151 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3152 dev->name, max_queue_pairs);
3153
3154 return 0;
3155
3156free_unregister_netdev:
3157 vi->vdev->config->reset(vdev);
3158
3159 unregister_netdev(dev);
3160free_failover:
3161 net_failover_destroy(vi->failover);
3162free_vqs:
3163 cancel_delayed_work_sync(&vi->refill);
3164 free_receive_page_frags(vi);
3165 virtnet_del_vqs(vi);
3166free:
3167 free_netdev(dev);
3168 return err;
3169}
3170
3171static void remove_vq_common(struct virtnet_info *vi)
3172{
3173 vi->vdev->config->reset(vi->vdev);
3174
3175 /* Free unused buffers in both send and recv, if any. */
3176 free_unused_bufs(vi);
3177
3178 free_receive_bufs(vi);
3179
3180 free_receive_page_frags(vi);
3181
3182 virtnet_del_vqs(vi);
3183}
3184
3185static void virtnet_remove(struct virtio_device *vdev)
3186{
3187 struct virtnet_info *vi = vdev->priv;
3188
3189 virtnet_cpu_notif_remove(vi);
3190
3191 /* Make sure no work handler is accessing the device. */
3192 flush_work(&vi->config_work);
3193
3194 unregister_netdev(vi->dev);
3195
3196 net_failover_destroy(vi->failover);
3197
3198 remove_vq_common(vi);
3199
3200 free_netdev(vi->dev);
3201}
3202
3203static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3204{
3205 struct virtnet_info *vi = vdev->priv;
3206
3207 virtnet_cpu_notif_remove(vi);
3208 virtnet_freeze_down(vdev);
3209 remove_vq_common(vi);
3210
3211 return 0;
3212}
3213
3214static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3215{
3216 struct virtnet_info *vi = vdev->priv;
3217 int err;
3218
3219 err = virtnet_restore_up(vdev);
3220 if (err)
3221 return err;
3222 virtnet_set_queues(vi, vi->curr_queue_pairs);
3223
3224 err = virtnet_cpu_notif_add(vi);
3225 if (err)
3226 return err;
3227
3228 return 0;
3229}
3230
3231static struct virtio_device_id id_table[] = {
3232 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3233 { 0 },
3234};
3235
3236#define VIRTNET_FEATURES \
3237 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3238 VIRTIO_NET_F_MAC, \
3239 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3240 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3241 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3242 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3243 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3244 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3245 VIRTIO_NET_F_CTRL_MAC_ADDR, \
3246 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3247 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3248
3249static unsigned int features[] = {
3250 VIRTNET_FEATURES,
3251};
3252
3253static unsigned int features_legacy[] = {
3254 VIRTNET_FEATURES,
3255 VIRTIO_NET_F_GSO,
3256 VIRTIO_F_ANY_LAYOUT,
3257};
3258
3259static struct virtio_driver virtio_net_driver = {
3260 .feature_table = features,
3261 .feature_table_size = ARRAY_SIZE(features),
3262 .feature_table_legacy = features_legacy,
3263 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3264 .driver.name = KBUILD_MODNAME,
3265 .driver.owner = THIS_MODULE,
3266 .id_table = id_table,
3267 .validate = virtnet_validate,
3268 .probe = virtnet_probe,
3269 .remove = virtnet_remove,
3270 .config_changed = virtnet_config_changed,
3271#ifdef CONFIG_PM_SLEEP
3272 .freeze = virtnet_freeze,
3273 .restore = virtnet_restore,
3274#endif
3275};
3276
3277static __init int virtio_net_driver_init(void)
3278{
3279 int ret;
3280
3281 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3282 virtnet_cpu_online,
3283 virtnet_cpu_down_prep);
3284 if (ret < 0)
3285 goto out;
3286 virtionet_online = ret;
3287 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3288 NULL, virtnet_cpu_dead);
3289 if (ret)
3290 goto err_dead;
3291
3292 ret = register_virtio_driver(&virtio_net_driver);
3293 if (ret)
3294 goto err_virtio;
3295 return 0;
3296err_virtio:
3297 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3298err_dead:
3299 cpuhp_remove_multi_state(virtionet_online);
3300out:
3301 return ret;
3302}
3303module_init(virtio_net_driver_init);
3304
3305static __exit void virtio_net_driver_exit(void)
3306{
3307 unregister_virtio_driver(&virtio_net_driver);
3308 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3309 cpuhp_remove_multi_state(virtionet_online);
3310}
3311module_exit(virtio_net_driver_exit);
3312
3313MODULE_DEVICE_TABLE(virtio, id_table);
3314MODULE_DESCRIPTION("Virtio network driver");
3315MODULE_LICENSE("GPL");