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