?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
xfs: add full xfs_dqblk verifier
[linux/fpc-iii.git] / drivers / net / virtio_net.c
blob770422e953f75e1e9a1bb3a317d83f8070e6ae36
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
36 static int napi_weight = NAPI_POLL_WEIGHT;
37 module_param(napi_weight, int, 0444);
38
39 static bool csum = true, gso = true, napi_tx;
40 module_param(csum, bool, 0444);
41 module_param(gso, bool, 0444);
42 module_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 */
58 DECLARE_EWMA(pkt_len, 0, 64)
59
60 #define VIRTNET_DRIVER_VERSION "1.0.0"
61
62 static 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
69 struct virtnet_stat_desc {
70 char desc[ETH_GSTRING_LEN];
71 size_t offset;
72 };
73
74 struct virtnet_sq_stats {
75 struct u64_stats_sync syncp;
76 u64 packets;
77 u64 bytes;
78 };
79
80 struct 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
89 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
90 { "packets", VIRTNET_SQ_STAT(packets) },
91 { "bytes", VIRTNET_SQ_STAT(bytes) },
92 };
93
94 static 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 */
103 struct 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.$index */
111 char name[40];
112
113 struct virtnet_sq_stats stats;
114
115 struct napi_struct napi;
116 };
117
118 /* Internal representation of a receive virtqueue */
119 struct receive_queue {
120 /* Virtqueue associated with this receive_queue */
121 struct virtqueue *vq;
122
123 struct napi_struct napi;
124
125 struct bpf_prog __rcu *xdp_prog;
126
127 struct virtnet_rq_stats stats;
128
129 /* Chain pages by the private ptr. */
130 struct page *pages;
131
132 /* Average packet length for mergeable receive buffers. */
133 struct ewma_pkt_len mrg_avg_pkt_len;
134
135 /* Page frag for packet buffer allocation. */
136 struct page_frag alloc_frag;
137
138 /* RX: fragments + linear part + virtio header */
139 struct scatterlist sg[MAX_SKB_FRAGS + 2];
140
141 /* Min single buffer size for mergeable buffers case. */
142 unsigned int min_buf_len;
143
144 /* Name of this receive queue: input.$index */
145 char name[40];
146
147 struct xdp_rxq_info xdp_rxq;
148 };
149
150 /* Control VQ buffers: protected by the rtnl lock */
151 struct 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
161 struct 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
215 struct 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 */
228 static int vq2txq(struct virtqueue *vq)
229 {
230 return (vq->index - 1) / 2;
231 }
232
233 static int txq2vq(int txq)
234 {
235 return txq * 2 + 1;
236 }
237
238 static int vq2rxq(struct virtqueue *vq)
239 {
240 return vq->index / 2;
241 }
242
243 static int rxq2vq(int rxq)
244 {
245 return rxq * 2;
246 }
247
248 static 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 */
257 static 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
267 static 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
280 static 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
289 static 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
303 static 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
319 static 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
325 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
326 {
327 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
328 }
329
330 static 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 */
336 static 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
410 static 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
422 static 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
456 static 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
476 static 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 */
495 static 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;
540 err_buf:
541 __free_pages(page, 0);
542 return NULL;
543 }
544
545 static 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
650 err:
651 return skb;
652
653 err_xdp:
654 rcu_read_unlock();
655 dev->stats.rx_dropped++;
656 put_page(page);
657 xdp_xmit:
658 return NULL;
659 }
660
661 static 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
675 err:
676 dev->stats.rx_dropped++;
677 give_pages(rq, page);
678 return NULL;
679 }
680
681 static 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 /* This happens when rx buffer size is underestimated
711 * or headroom is not enough because of the buffer
712 * was refilled before XDP is set. This should only
713 * happen for the first several packets, so we don't
714 * care much about its performance.
715 */
716 if (unlikely(num_buf > 1 ||
717 headroom < virtnet_get_headroom(vi))) {
718 /* linearize data for XDP */
719 xdp_page = xdp_linearize_page(rq, &num_buf,
720 page, offset,
721 VIRTIO_XDP_HEADROOM,
722 &len);
723 if (!xdp_page)
724 goto err_xdp;
725 offset = VIRTIO_XDP_HEADROOM;
726 } else {
727 xdp_page = page;
728 }
729
730 /* Transient failure which in theory could occur if
731 * in-flight packets from before XDP was enabled reach
732 * the receive path after XDP is loaded. In practice I
733 * was not able to create this condition.
734 */
735 if (unlikely(hdr->hdr.gso_type))
736 goto err_xdp;
737
738 /* Allow consuming headroom but reserve enough space to push
739 * the descriptor on if we get an XDP_TX return code.
740 */
741 data = page_address(xdp_page) + offset;
742 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
743 xdp.data = data + vi->hdr_len;
744 xdp_set_data_meta_invalid(&xdp);
745 xdp.data_end = xdp.data + (len - vi->hdr_len);
746 xdp.rxq = &rq->xdp_rxq;
747
748 act = bpf_prog_run_xdp(xdp_prog, &xdp);
749
750 switch (act) {
751 case XDP_PASS:
752 /* recalculate offset to account for any header
753 * adjustments. Note other cases do not build an
754 * skb and avoid using offset
755 */
756 offset = xdp.data -
757 page_address(xdp_page) - vi->hdr_len;
758
759 /* We can only create skb based on xdp_page. */
760 if (unlikely(xdp_page != page)) {
761 rcu_read_unlock();
762 put_page(page);
763 head_skb = page_to_skb(vi, rq, xdp_page,
764 offset, len, PAGE_SIZE);
765 return head_skb;
766 }
767 break;
768 case XDP_TX:
769 sent = __virtnet_xdp_xmit(vi, &xdp);
770 if (unlikely(!sent)) {
771 trace_xdp_exception(vi->dev, xdp_prog, act);
772 if (unlikely(xdp_page != page))
773 put_page(xdp_page);
774 goto err_xdp;
775 }
776 *xdp_xmit = true;
777 if (unlikely(xdp_page != page))
778 goto err_xdp;
779 rcu_read_unlock();
780 goto xdp_xmit;
781 case XDP_REDIRECT:
782 err = xdp_do_redirect(dev, &xdp, xdp_prog);
783 if (err) {
784 if (unlikely(xdp_page != page))
785 put_page(xdp_page);
786 goto err_xdp;
787 }
788 *xdp_xmit = true;
789 if (unlikely(xdp_page != page))
790 goto err_xdp;
791 rcu_read_unlock();
792 goto xdp_xmit;
793 default:
794 bpf_warn_invalid_xdp_action(act);
795 case XDP_ABORTED:
796 trace_xdp_exception(vi->dev, xdp_prog, act);
797 case XDP_DROP:
798 if (unlikely(xdp_page != page))
799 __free_pages(xdp_page, 0);
800 goto err_xdp;
801 }
802 }
803 rcu_read_unlock();
804
805 truesize = mergeable_ctx_to_truesize(ctx);
806 if (unlikely(len > truesize)) {
807 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
808 dev->name, len, (unsigned long)ctx);
809 dev->stats.rx_length_errors++;
810 goto err_skb;
811 }
812
813 head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
814 curr_skb = head_skb;
815
816 if (unlikely(!curr_skb))
817 goto err_skb;
818 while (--num_buf) {
819 int num_skb_frags;
820
821 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
822 if (unlikely(!buf)) {
823 pr_debug("%s: rx error: %d buffers out of %d missing\n",
824 dev->name, num_buf,
825 virtio16_to_cpu(vi->vdev,
826 hdr->num_buffers));
827 dev->stats.rx_length_errors++;
828 goto err_buf;
829 }
830
831 page = virt_to_head_page(buf);
832
833 truesize = mergeable_ctx_to_truesize(ctx);
834 if (unlikely(len > truesize)) {
835 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
836 dev->name, len, (unsigned long)ctx);
837 dev->stats.rx_length_errors++;
838 goto err_skb;
839 }
840
841 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
842 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
843 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
844
845 if (unlikely(!nskb))
846 goto err_skb;
847 if (curr_skb == head_skb)
848 skb_shinfo(curr_skb)->frag_list = nskb;
849 else
850 curr_skb->next = nskb;
851 curr_skb = nskb;
852 head_skb->truesize += nskb->truesize;
853 num_skb_frags = 0;
854 }
855 if (curr_skb != head_skb) {
856 head_skb->data_len += len;
857 head_skb->len += len;
858 head_skb->truesize += truesize;
859 }
860 offset = buf - page_address(page);
861 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
862 put_page(page);
863 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
864 len, truesize);
865 } else {
866 skb_add_rx_frag(curr_skb, num_skb_frags, page,
867 offset, len, truesize);
868 }
869 }
870
871 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
872 return head_skb;
873
874 err_xdp:
875 rcu_read_unlock();
876 err_skb:
877 put_page(page);
878 while (--num_buf) {
879 buf = virtqueue_get_buf(rq->vq, &len);
880 if (unlikely(!buf)) {
881 pr_debug("%s: rx error: %d buffers missing\n",
882 dev->name, num_buf);
883 dev->stats.rx_length_errors++;
884 break;
885 }
886 page = virt_to_head_page(buf);
887 put_page(page);
888 }
889 err_buf:
890 dev->stats.rx_dropped++;
891 dev_kfree_skb(head_skb);
892 xdp_xmit:
893 return NULL;
894 }
895
896 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
897 void *buf, unsigned int len, void **ctx, bool *xdp_xmit)
898 {
899 struct net_device *dev = vi->dev;
900 struct sk_buff *skb;
901 struct virtio_net_hdr_mrg_rxbuf *hdr;
902 int ret;
903
904 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
905 pr_debug("%s: short packet %i\n", dev->name, len);
906 dev->stats.rx_length_errors++;
907 if (vi->mergeable_rx_bufs) {
908 put_page(virt_to_head_page(buf));
909 } else if (vi->big_packets) {
910 give_pages(rq, buf);
911 } else {
912 put_page(virt_to_head_page(buf));
913 }
914 return 0;
915 }
916
917 if (vi->mergeable_rx_bufs)
918 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit);
919 else if (vi->big_packets)
920 skb = receive_big(dev, vi, rq, buf, len);
921 else
922 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit);
923
924 if (unlikely(!skb))
925 return 0;
926
927 hdr = skb_vnet_hdr(skb);
928
929 ret = skb->len;
930
931 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
932 skb->ip_summed = CHECKSUM_UNNECESSARY;
933
934 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
935 virtio_is_little_endian(vi->vdev))) {
936 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
937 dev->name, hdr->hdr.gso_type,
938 hdr->hdr.gso_size);
939 goto frame_err;
940 }
941
942 skb->protocol = eth_type_trans(skb, dev);
943 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
944 ntohs(skb->protocol), skb->len, skb->pkt_type);
945
946 napi_gro_receive(&rq->napi, skb);
947 return ret;
948
949 frame_err:
950 dev->stats.rx_frame_errors++;
951 dev_kfree_skb(skb);
952 return 0;
953 }
954
955 /* Unlike mergeable buffers, all buffers are allocated to the
956 * same size, except for the headroom. For this reason we do
957 * not need to use mergeable_len_to_ctx here - it is enough
958 * to store the headroom as the context ignoring the truesize.
959 */
960 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
961 gfp_t gfp)
962 {
963 struct page_frag *alloc_frag = &rq->alloc_frag;
964 char *buf;
965 unsigned int xdp_headroom = virtnet_get_headroom(vi);
966 void *ctx = (void *)(unsigned long)xdp_headroom;
967 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
968 int err;
969
970 len = SKB_DATA_ALIGN(len) +
971 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
972 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
973 return -ENOMEM;
974
975 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
976 get_page(alloc_frag->page);
977 alloc_frag->offset += len;
978 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
979 vi->hdr_len + GOOD_PACKET_LEN);
980 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
981 if (err < 0)
982 put_page(virt_to_head_page(buf));
983 return err;
984 }
985
986 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
987 gfp_t gfp)
988 {
989 struct page *first, *list = NULL;
990 char *p;
991 int i, err, offset;
992
993 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
994
995 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
996 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
997 first = get_a_page(rq, gfp);
998 if (!first) {
999 if (list)
1000 give_pages(rq, list);
1001 return -ENOMEM;
1002 }
1003 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1004
1005 /* chain new page in list head to match sg */
1006 first->private = (unsigned long)list;
1007 list = first;
1008 }
1009
1010 first = get_a_page(rq, gfp);
1011 if (!first) {
1012 give_pages(rq, list);
1013 return -ENOMEM;
1014 }
1015 p = page_address(first);
1016
1017 /* rq->sg[0], rq->sg[1] share the same page */
1018 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1019 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1020
1021 /* rq->sg[1] for data packet, from offset */
1022 offset = sizeof(struct padded_vnet_hdr);
1023 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1024
1025 /* chain first in list head */
1026 first->private = (unsigned long)list;
1027 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1028 first, gfp);
1029 if (err < 0)
1030 give_pages(rq, first);
1031
1032 return err;
1033 }
1034
1035 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1036 struct ewma_pkt_len *avg_pkt_len,
1037 unsigned int room)
1038 {
1039 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1040 unsigned int len;
1041
1042 if (room)
1043 return PAGE_SIZE - room;
1044
1045 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1046 rq->min_buf_len, PAGE_SIZE - hdr_len);
1047
1048 return ALIGN(len, L1_CACHE_BYTES);
1049 }
1050
1051 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1052 struct receive_queue *rq, gfp_t gfp)
1053 {
1054 struct page_frag *alloc_frag = &rq->alloc_frag;
1055 unsigned int headroom = virtnet_get_headroom(vi);
1056 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1057 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1058 char *buf;
1059 void *ctx;
1060 int err;
1061 unsigned int len, hole;
1062
1063 /* Extra tailroom is needed to satisfy XDP's assumption. This
1064 * means rx frags coalescing won't work, but consider we've
1065 * disabled GSO for XDP, it won't be a big issue.
1066 */
1067 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1068 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1069 return -ENOMEM;
1070
1071 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1072 buf += headroom; /* advance address leaving hole at front of pkt */
1073 get_page(alloc_frag->page);
1074 alloc_frag->offset += len + room;
1075 hole = alloc_frag->size - alloc_frag->offset;
1076 if (hole < len + room) {
1077 /* To avoid internal fragmentation, if there is very likely not
1078 * enough space for another buffer, add the remaining space to
1079 * the current buffer.
1080 */
1081 len += hole;
1082 alloc_frag->offset += hole;
1083 }
1084
1085 sg_init_one(rq->sg, buf, len);
1086 ctx = mergeable_len_to_ctx(len, headroom);
1087 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1088 if (err < 0)
1089 put_page(virt_to_head_page(buf));
1090
1091 return err;
1092 }
1093
1094 /*
1095 * Returns false if we couldn't fill entirely (OOM).
1096 *
1097 * Normally run in the receive path, but can also be run from ndo_open
1098 * before we're receiving packets, or from refill_work which is
1099 * careful to disable receiving (using napi_disable).
1100 */
1101 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1102 gfp_t gfp)
1103 {
1104 int err;
1105 bool oom;
1106
1107 do {
1108 if (vi->mergeable_rx_bufs)
1109 err = add_recvbuf_mergeable(vi, rq, gfp);
1110 else if (vi->big_packets)
1111 err = add_recvbuf_big(vi, rq, gfp);
1112 else
1113 err = add_recvbuf_small(vi, rq, gfp);
1114
1115 oom = err == -ENOMEM;
1116 if (err)
1117 break;
1118 } while (rq->vq->num_free);
1119 virtqueue_kick(rq->vq);
1120 return !oom;
1121 }
1122
1123 static void skb_recv_done(struct virtqueue *rvq)
1124 {
1125 struct virtnet_info *vi = rvq->vdev->priv;
1126 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1127
1128 virtqueue_napi_schedule(&rq->napi, rvq);
1129 }
1130
1131 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1132 {
1133 napi_enable(napi);
1134
1135 /* If all buffers were filled by other side before we napi_enabled, we
1136 * won't get another interrupt, so process any outstanding packets now.
1137 * Call local_bh_enable after to trigger softIRQ processing.
1138 */
1139 local_bh_disable();
1140 virtqueue_napi_schedule(napi, vq);
1141 local_bh_enable();
1142 }
1143
1144 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1145 struct virtqueue *vq,
1146 struct napi_struct *napi)
1147 {
1148 if (!napi->weight)
1149 return;
1150
1151 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1152 * enable the feature if this is likely affine with the transmit path.
1153 */
1154 if (!vi->affinity_hint_set) {
1155 napi->weight = 0;
1156 return;
1157 }
1158
1159 return virtnet_napi_enable(vq, napi);
1160 }
1161
1162 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1163 {
1164 if (napi->weight)
1165 napi_disable(napi);
1166 }
1167
1168 static void refill_work(struct work_struct *work)
1169 {
1170 struct virtnet_info *vi =
1171 container_of(work, struct virtnet_info, refill.work);
1172 bool still_empty;
1173 int i;
1174
1175 for (i = 0; i < vi->curr_queue_pairs; i++) {
1176 struct receive_queue *rq = &vi->rq[i];
1177
1178 napi_disable(&rq->napi);
1179 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1180 virtnet_napi_enable(rq->vq, &rq->napi);
1181
1182 /* In theory, this can happen: if we don't get any buffers in
1183 * we will *never* try to fill again.
1184 */
1185 if (still_empty)
1186 schedule_delayed_work(&vi->refill, HZ/2);
1187 }
1188 }
1189
1190 static int virtnet_receive(struct receive_queue *rq, int budget, bool *xdp_xmit)
1191 {
1192 struct virtnet_info *vi = rq->vq->vdev->priv;
1193 unsigned int len, received = 0, bytes = 0;
1194 void *buf;
1195
1196 if (!vi->big_packets || vi->mergeable_rx_bufs) {
1197 void *ctx;
1198
1199 while (received < budget &&
1200 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1201 bytes += receive_buf(vi, rq, buf, len, ctx, xdp_xmit);
1202 received++;
1203 }
1204 } else {
1205 while (received < budget &&
1206 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1207 bytes += receive_buf(vi, rq, buf, len, NULL, xdp_xmit);
1208 received++;
1209 }
1210 }
1211
1212 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1213 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1214 schedule_delayed_work(&vi->refill, 0);
1215 }
1216
1217 u64_stats_update_begin(&rq->stats.syncp);
1218 rq->stats.bytes += bytes;
1219 rq->stats.packets += received;
1220 u64_stats_update_end(&rq->stats.syncp);
1221
1222 return received;
1223 }
1224
1225 static void free_old_xmit_skbs(struct send_queue *sq)
1226 {
1227 struct sk_buff *skb;
1228 unsigned int len;
1229 unsigned int packets = 0;
1230 unsigned int bytes = 0;
1231
1232 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1233 pr_debug("Sent skb %p\n", skb);
1234
1235 bytes += skb->len;
1236 packets++;
1237
1238 dev_consume_skb_any(skb);
1239 }
1240
1241 /* Avoid overhead when no packets have been processed
1242 * happens when called speculatively from start_xmit.
1243 */
1244 if (!packets)
1245 return;
1246
1247 u64_stats_update_begin(&sq->stats.syncp);
1248 sq->stats.bytes += bytes;
1249 sq->stats.packets += packets;
1250 u64_stats_update_end(&sq->stats.syncp);
1251 }
1252
1253 static void virtnet_poll_cleantx(struct receive_queue *rq)
1254 {
1255 struct virtnet_info *vi = rq->vq->vdev->priv;
1256 unsigned int index = vq2rxq(rq->vq);
1257 struct send_queue *sq = &vi->sq[index];
1258 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1259
1260 if (!sq->napi.weight)
1261 return;
1262
1263 if (__netif_tx_trylock(txq)) {
1264 free_old_xmit_skbs(sq);
1265 __netif_tx_unlock(txq);
1266 }
1267
1268 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1269 netif_tx_wake_queue(txq);
1270 }
1271
1272 static int virtnet_poll(struct napi_struct *napi, int budget)
1273 {
1274 struct receive_queue *rq =
1275 container_of(napi, struct receive_queue, napi);
1276 struct virtnet_info *vi = rq->vq->vdev->priv;
1277 struct send_queue *sq;
1278 unsigned int received, qp;
1279 bool xdp_xmit = false;
1280
1281 virtnet_poll_cleantx(rq);
1282
1283 received = virtnet_receive(rq, budget, &xdp_xmit);
1284
1285 /* Out of packets? */
1286 if (received < budget)
1287 virtqueue_napi_complete(napi, rq->vq, received);
1288
1289 if (xdp_xmit) {
1290 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs +
1291 smp_processor_id();
1292 sq = &vi->sq[qp];
1293 virtqueue_kick(sq->vq);
1294 xdp_do_flush_map();
1295 }
1296
1297 return received;
1298 }
1299
1300 static int virtnet_open(struct net_device *dev)
1301 {
1302 struct virtnet_info *vi = netdev_priv(dev);
1303 int i, err;
1304
1305 for (i = 0; i < vi->max_queue_pairs; i++) {
1306 if (i < vi->curr_queue_pairs)
1307 /* Make sure we have some buffers: if oom use wq. */
1308 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1309 schedule_delayed_work(&vi->refill, 0);
1310
1311 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i);
1312 if (err < 0)
1313 return err;
1314
1315 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1316 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1317 }
1318
1319 return 0;
1320 }
1321
1322 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1323 {
1324 struct send_queue *sq = container_of(napi, struct send_queue, napi);
1325 struct virtnet_info *vi = sq->vq->vdev->priv;
1326 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
1327
1328 __netif_tx_lock(txq, raw_smp_processor_id());
1329 free_old_xmit_skbs(sq);
1330 __netif_tx_unlock(txq);
1331
1332 virtqueue_napi_complete(napi, sq->vq, 0);
1333
1334 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1335 netif_tx_wake_queue(txq);
1336
1337 return 0;
1338 }
1339
1340 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1341 {
1342 struct virtio_net_hdr_mrg_rxbuf *hdr;
1343 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1344 struct virtnet_info *vi = sq->vq->vdev->priv;
1345 int num_sg;
1346 unsigned hdr_len = vi->hdr_len;
1347 bool can_push;
1348
1349 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1350
1351 can_push = vi->any_header_sg &&
1352 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1353 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1354 /* Even if we can, don't push here yet as this would skew
1355 * csum_start offset below. */
1356 if (can_push)
1357 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1358 else
1359 hdr = skb_vnet_hdr(skb);
1360
1361 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1362 virtio_is_little_endian(vi->vdev), false))
1363 BUG();
1364
1365 if (vi->mergeable_rx_bufs)
1366 hdr->num_buffers = 0;
1367
1368 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1369 if (can_push) {
1370 __skb_push(skb, hdr_len);
1371 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1372 if (unlikely(num_sg < 0))
1373 return num_sg;
1374 /* Pull header back to avoid skew in tx bytes calculations. */
1375 __skb_pull(skb, hdr_len);
1376 } else {
1377 sg_set_buf(sq->sg, hdr, hdr_len);
1378 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1379 if (unlikely(num_sg < 0))
1380 return num_sg;
1381 num_sg++;
1382 }
1383 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1384 }
1385
1386 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1387 {
1388 struct virtnet_info *vi = netdev_priv(dev);
1389 int qnum = skb_get_queue_mapping(skb);
1390 struct send_queue *sq = &vi->sq[qnum];
1391 int err;
1392 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1393 bool kick = !skb->xmit_more;
1394 bool use_napi = sq->napi.weight;
1395
1396 /* Free up any pending old buffers before queueing new ones. */
1397 free_old_xmit_skbs(sq);
1398
1399 if (use_napi && kick)
1400 virtqueue_enable_cb_delayed(sq->vq);
1401
1402 /* timestamp packet in software */
1403 skb_tx_timestamp(skb);
1404
1405 /* Try to transmit */
1406 err = xmit_skb(sq, skb);
1407
1408 /* This should not happen! */
1409 if (unlikely(err)) {
1410 dev->stats.tx_fifo_errors++;
1411 if (net_ratelimit())
1412 dev_warn(&dev->dev,
1413 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1414 dev->stats.tx_dropped++;
1415 dev_kfree_skb_any(skb);
1416 return NETDEV_TX_OK;
1417 }
1418
1419 /* Don't wait up for transmitted skbs to be freed. */
1420 if (!use_napi) {
1421 skb_orphan(skb);
1422 nf_reset(skb);
1423 }
1424
1425 /* If running out of space, stop queue to avoid getting packets that we
1426 * are then unable to transmit.
1427 * An alternative would be to force queuing layer to requeue the skb by
1428 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1429 * returned in a normal path of operation: it means that driver is not
1430 * maintaining the TX queue stop/start state properly, and causes
1431 * the stack to do a non-trivial amount of useless work.
1432 * Since most packets only take 1 or 2 ring slots, stopping the queue
1433 * early means 16 slots are typically wasted.
1434 */
1435 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1436 netif_stop_subqueue(dev, qnum);
1437 if (!use_napi &&
1438 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1439 /* More just got used, free them then recheck. */
1440 free_old_xmit_skbs(sq);
1441 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1442 netif_start_subqueue(dev, qnum);
1443 virtqueue_disable_cb(sq->vq);
1444 }
1445 }
1446 }
1447
1448 if (kick || netif_xmit_stopped(txq))
1449 virtqueue_kick(sq->vq);
1450
1451 return NETDEV_TX_OK;
1452 }
1453
1454 /*
1455 * Send command via the control virtqueue and check status. Commands
1456 * supported by the hypervisor, as indicated by feature bits, should
1457 * never fail unless improperly formatted.
1458 */
1459 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1460 struct scatterlist *out)
1461 {
1462 struct scatterlist *sgs[4], hdr, stat;
1463 unsigned out_num = 0, tmp;
1464
1465 /* Caller should know better */
1466 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1467
1468 vi->ctrl->status = ~0;
1469 vi->ctrl->hdr.class = class;
1470 vi->ctrl->hdr.cmd = cmd;
1471 /* Add header */
1472 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1473 sgs[out_num++] = &hdr;
1474
1475 if (out)
1476 sgs[out_num++] = out;
1477
1478 /* Add return status. */
1479 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1480 sgs[out_num] = &stat;
1481
1482 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1483 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1484
1485 if (unlikely(!virtqueue_kick(vi->cvq)))
1486 return vi->ctrl->status == VIRTIO_NET_OK;
1487
1488 /* Spin for a response, the kick causes an ioport write, trapping
1489 * into the hypervisor, so the request should be handled immediately.
1490 */
1491 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1492 !virtqueue_is_broken(vi->cvq))
1493 cpu_relax();
1494
1495 return vi->ctrl->status == VIRTIO_NET_OK;
1496 }
1497
1498 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1499 {
1500 struct virtnet_info *vi = netdev_priv(dev);
1501 struct virtio_device *vdev = vi->vdev;
1502 int ret;
1503 struct sockaddr *addr;
1504 struct scatterlist sg;
1505
1506 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1507 if (!addr)
1508 return -ENOMEM;
1509
1510 ret = eth_prepare_mac_addr_change(dev, addr);
1511 if (ret)
1512 goto out;
1513
1514 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1515 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1516 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1517 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1518 dev_warn(&vdev->dev,
1519 "Failed to set mac address by vq command.\n");
1520 ret = -EINVAL;
1521 goto out;
1522 }
1523 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1524 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1525 unsigned int i;
1526
1527 /* Naturally, this has an atomicity problem. */
1528 for (i = 0; i < dev->addr_len; i++)
1529 virtio_cwrite8(vdev,
1530 offsetof(struct virtio_net_config, mac) +
1531 i, addr->sa_data[i]);
1532 }
1533
1534 eth_commit_mac_addr_change(dev, p);
1535 ret = 0;
1536
1537 out:
1538 kfree(addr);
1539 return ret;
1540 }
1541
1542 static void virtnet_stats(struct net_device *dev,
1543 struct rtnl_link_stats64 *tot)
1544 {
1545 struct virtnet_info *vi = netdev_priv(dev);
1546 unsigned int start;
1547 int i;
1548
1549 for (i = 0; i < vi->max_queue_pairs; i++) {
1550 u64 tpackets, tbytes, rpackets, rbytes;
1551 struct receive_queue *rq = &vi->rq[i];
1552 struct send_queue *sq = &vi->sq[i];
1553
1554 do {
1555 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1556 tpackets = sq->stats.packets;
1557 tbytes = sq->stats.bytes;
1558 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1559
1560 do {
1561 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1562 rpackets = rq->stats.packets;
1563 rbytes = rq->stats.bytes;
1564 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1565
1566 tot->rx_packets += rpackets;
1567 tot->tx_packets += tpackets;
1568 tot->rx_bytes += rbytes;
1569 tot->tx_bytes += tbytes;
1570 }
1571
1572 tot->tx_dropped = dev->stats.tx_dropped;
1573 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1574 tot->rx_dropped = dev->stats.rx_dropped;
1575 tot->rx_length_errors = dev->stats.rx_length_errors;
1576 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1577 }
1578
1579 #ifdef CONFIG_NET_POLL_CONTROLLER
1580 static void virtnet_netpoll(struct net_device *dev)
1581 {
1582 struct virtnet_info *vi = netdev_priv(dev);
1583 int i;
1584
1585 for (i = 0; i < vi->curr_queue_pairs; i++)
1586 napi_schedule(&vi->rq[i].napi);
1587 }
1588 #endif
1589
1590 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1591 {
1592 rtnl_lock();
1593 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1594 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1595 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1596 rtnl_unlock();
1597 }
1598
1599 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1600 {
1601 struct scatterlist sg;
1602 struct net_device *dev = vi->dev;
1603
1604 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1605 return 0;
1606
1607 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1608 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1609
1610 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1611 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1612 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1613 queue_pairs);
1614 return -EINVAL;
1615 } else {
1616 vi->curr_queue_pairs = queue_pairs;
1617 /* virtnet_open() will refill when device is going to up. */
1618 if (dev->flags & IFF_UP)
1619 schedule_delayed_work(&vi->refill, 0);
1620 }
1621
1622 return 0;
1623 }
1624
1625 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1626 {
1627 int err;
1628
1629 rtnl_lock();
1630 err = _virtnet_set_queues(vi, queue_pairs);
1631 rtnl_unlock();
1632 return err;
1633 }
1634
1635 static int virtnet_close(struct net_device *dev)
1636 {
1637 struct virtnet_info *vi = netdev_priv(dev);
1638 int i;
1639
1640 /* Make sure refill_work doesn't re-enable napi! */
1641 cancel_delayed_work_sync(&vi->refill);
1642
1643 for (i = 0; i < vi->max_queue_pairs; i++) {
1644 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1645 napi_disable(&vi->rq[i].napi);
1646 virtnet_napi_tx_disable(&vi->sq[i].napi);
1647 }
1648
1649 return 0;
1650 }
1651
1652 static void virtnet_set_rx_mode(struct net_device *dev)
1653 {
1654 struct virtnet_info *vi = netdev_priv(dev);
1655 struct scatterlist sg[2];
1656 struct virtio_net_ctrl_mac *mac_data;
1657 struct netdev_hw_addr *ha;
1658 int uc_count;
1659 int mc_count;
1660 void *buf;
1661 int i;
1662
1663 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1664 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1665 return;
1666
1667 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1668 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1669
1670 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1671
1672 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1673 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1674 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1675 vi->ctrl->promisc ? "en" : "dis");
1676
1677 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1678
1679 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1680 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1681 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1682 vi->ctrl->allmulti ? "en" : "dis");
1683
1684 uc_count = netdev_uc_count(dev);
1685 mc_count = netdev_mc_count(dev);
1686 /* MAC filter - use one buffer for both lists */
1687 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1688 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1689 mac_data = buf;
1690 if (!buf)
1691 return;
1692
1693 sg_init_table(sg, 2);
1694
1695 /* Store the unicast list and count in the front of the buffer */
1696 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1697 i = 0;
1698 netdev_for_each_uc_addr(ha, dev)
1699 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1700
1701 sg_set_buf(&sg[0], mac_data,
1702 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1703
1704 /* multicast list and count fill the end */
1705 mac_data = (void *)&mac_data->macs[uc_count][0];
1706
1707 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1708 i = 0;
1709 netdev_for_each_mc_addr(ha, dev)
1710 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1711
1712 sg_set_buf(&sg[1], mac_data,
1713 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1714
1715 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1716 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1717 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1718
1719 kfree(buf);
1720 }
1721
1722 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1723 __be16 proto, u16 vid)
1724 {
1725 struct virtnet_info *vi = netdev_priv(dev);
1726 struct scatterlist sg;
1727
1728 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1729 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1730
1731 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1732 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1733 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1734 return 0;
1735 }
1736
1737 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1738 __be16 proto, u16 vid)
1739 {
1740 struct virtnet_info *vi = netdev_priv(dev);
1741 struct scatterlist sg;
1742
1743 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1744 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1745
1746 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1747 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1748 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1749 return 0;
1750 }
1751
1752 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1753 {
1754 int i;
1755
1756 if (vi->affinity_hint_set) {
1757 for (i = 0; i < vi->max_queue_pairs; i++) {
1758 virtqueue_set_affinity(vi->rq[i].vq, -1);
1759 virtqueue_set_affinity(vi->sq[i].vq, -1);
1760 }
1761
1762 vi->affinity_hint_set = false;
1763 }
1764 }
1765
1766 static void virtnet_set_affinity(struct virtnet_info *vi)
1767 {
1768 int i;
1769 int cpu;
1770
1771 /* In multiqueue mode, when the number of cpu is equal to the number of
1772 * queue pairs, we let the queue pairs to be private to one cpu by
1773 * setting the affinity hint to eliminate the contention.
1774 */
1775 if (vi->curr_queue_pairs == 1 ||
1776 vi->max_queue_pairs != num_online_cpus()) {
1777 virtnet_clean_affinity(vi, -1);
1778 return;
1779 }
1780
1781 i = 0;
1782 for_each_online_cpu(cpu) {
1783 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1784 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1785 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1786 i++;
1787 }
1788
1789 vi->affinity_hint_set = true;
1790 }
1791
1792 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1793 {
1794 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1795 node);
1796 virtnet_set_affinity(vi);
1797 return 0;
1798 }
1799
1800 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1801 {
1802 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1803 node_dead);
1804 virtnet_set_affinity(vi);
1805 return 0;
1806 }
1807
1808 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1809 {
1810 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1811 node);
1812
1813 virtnet_clean_affinity(vi, cpu);
1814 return 0;
1815 }
1816
1817 static enum cpuhp_state virtionet_online;
1818
1819 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1820 {
1821 int ret;
1822
1823 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1824 if (ret)
1825 return ret;
1826 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1827 &vi->node_dead);
1828 if (!ret)
1829 return ret;
1830 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1831 return ret;
1832 }
1833
1834 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1835 {
1836 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1837 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1838 &vi->node_dead);
1839 }
1840
1841 static void virtnet_get_ringparam(struct net_device *dev,
1842 struct ethtool_ringparam *ring)
1843 {
1844 struct virtnet_info *vi = netdev_priv(dev);
1845
1846 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1847 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1848 ring->rx_pending = ring->rx_max_pending;
1849 ring->tx_pending = ring->tx_max_pending;
1850 }
1851
1852
1853 static void virtnet_get_drvinfo(struct net_device *dev,
1854 struct ethtool_drvinfo *info)
1855 {
1856 struct virtnet_info *vi = netdev_priv(dev);
1857 struct virtio_device *vdev = vi->vdev;
1858
1859 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1860 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1861 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1862
1863 }
1864
1865 /* TODO: Eliminate OOO packets during switching */
1866 static int virtnet_set_channels(struct net_device *dev,
1867 struct ethtool_channels *channels)
1868 {
1869 struct virtnet_info *vi = netdev_priv(dev);
1870 u16 queue_pairs = channels->combined_count;
1871 int err;
1872
1873 /* We don't support separate rx/tx channels.
1874 * We don't allow setting 'other' channels.
1875 */
1876 if (channels->rx_count || channels->tx_count || channels->other_count)
1877 return -EINVAL;
1878
1879 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1880 return -EINVAL;
1881
1882 /* For now we don't support modifying channels while XDP is loaded
1883 * also when XDP is loaded all RX queues have XDP programs so we only
1884 * need to check a single RX queue.
1885 */
1886 if (vi->rq[0].xdp_prog)
1887 return -EINVAL;
1888
1889 get_online_cpus();
1890 err = _virtnet_set_queues(vi, queue_pairs);
1891 if (!err) {
1892 netif_set_real_num_tx_queues(dev, queue_pairs);
1893 netif_set_real_num_rx_queues(dev, queue_pairs);
1894
1895 virtnet_set_affinity(vi);
1896 }
1897 put_online_cpus();
1898
1899 return err;
1900 }
1901
1902 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
1903 {
1904 struct virtnet_info *vi = netdev_priv(dev);
1905 char *p = (char *)data;
1906 unsigned int i, j;
1907
1908 switch (stringset) {
1909 case ETH_SS_STATS:
1910 for (i = 0; i < vi->curr_queue_pairs; i++) {
1911 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
1912 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s",
1913 i, virtnet_rq_stats_desc[j].desc);
1914 p += ETH_GSTRING_LEN;
1915 }
1916 }
1917
1918 for (i = 0; i < vi->curr_queue_pairs; i++) {
1919 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
1920 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s",
1921 i, virtnet_sq_stats_desc[j].desc);
1922 p += ETH_GSTRING_LEN;
1923 }
1924 }
1925 break;
1926 }
1927 }
1928
1929 static int virtnet_get_sset_count(struct net_device *dev, int sset)
1930 {
1931 struct virtnet_info *vi = netdev_priv(dev);
1932
1933 switch (sset) {
1934 case ETH_SS_STATS:
1935 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
1936 VIRTNET_SQ_STATS_LEN);
1937 default:
1938 return -EOPNOTSUPP;
1939 }
1940 }
1941
1942 static void virtnet_get_ethtool_stats(struct net_device *dev,
1943 struct ethtool_stats *stats, u64 *data)
1944 {
1945 struct virtnet_info *vi = netdev_priv(dev);
1946 unsigned int idx = 0, start, i, j;
1947 const u8 *stats_base;
1948 size_t offset;
1949
1950 for (i = 0; i < vi->curr_queue_pairs; i++) {
1951 struct receive_queue *rq = &vi->rq[i];
1952
1953 stats_base = (u8 *)&rq->stats;
1954 do {
1955 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1956 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
1957 offset = virtnet_rq_stats_desc[j].offset;
1958 data[idx + j] = *(u64 *)(stats_base + offset);
1959 }
1960 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1961 idx += VIRTNET_RQ_STATS_LEN;
1962 }
1963
1964 for (i = 0; i < vi->curr_queue_pairs; i++) {
1965 struct send_queue *sq = &vi->sq[i];
1966
1967 stats_base = (u8 *)&sq->stats;
1968 do {
1969 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1970 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
1971 offset = virtnet_sq_stats_desc[j].offset;
1972 data[idx + j] = *(u64 *)(stats_base + offset);
1973 }
1974 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1975 idx += VIRTNET_SQ_STATS_LEN;
1976 }
1977 }
1978
1979 static void virtnet_get_channels(struct net_device *dev,
1980 struct ethtool_channels *channels)
1981 {
1982 struct virtnet_info *vi = netdev_priv(dev);
1983
1984 channels->combined_count = vi->curr_queue_pairs;
1985 channels->max_combined = vi->max_queue_pairs;
1986 channels->max_other = 0;
1987 channels->rx_count = 0;
1988 channels->tx_count = 0;
1989 channels->other_count = 0;
1990 }
1991
1992 /* Check if the user is trying to change anything besides speed/duplex */
1993 static bool
1994 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
1995 {
1996 struct ethtool_link_ksettings diff1 = *cmd;
1997 struct ethtool_link_ksettings diff2 = {};
1998
1999 /* cmd is always set so we need to clear it, validate the port type
2000 * and also without autonegotiation we can ignore advertising
2001 */
2002 diff1.base.speed = 0;
2003 diff2.base.port = PORT_OTHER;
2004 ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
2005 diff1.base.duplex = 0;
2006 diff1.base.cmd = 0;
2007 diff1.base.link_mode_masks_nwords = 0;
2008
2009 return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
2010 bitmap_empty(diff1.link_modes.supported,
2011 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2012 bitmap_empty(diff1.link_modes.advertising,
2013 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2014 bitmap_empty(diff1.link_modes.lp_advertising,
2015 __ETHTOOL_LINK_MODE_MASK_NBITS);
2016 }
2017
2018 static int virtnet_set_link_ksettings(struct net_device *dev,
2019 const struct ethtool_link_ksettings *cmd)
2020 {
2021 struct virtnet_info *vi = netdev_priv(dev);
2022 u32 speed;
2023
2024 speed = cmd->base.speed;
2025 /* don't allow custom speed and duplex */
2026 if (!ethtool_validate_speed(speed) ||
2027 !ethtool_validate_duplex(cmd->base.duplex) ||
2028 !virtnet_validate_ethtool_cmd(cmd))
2029 return -EINVAL;
2030 vi->speed = speed;
2031 vi->duplex = cmd->base.duplex;
2032
2033 return 0;
2034 }
2035
2036 static int virtnet_get_link_ksettings(struct net_device *dev,
2037 struct ethtool_link_ksettings *cmd)
2038 {
2039 struct virtnet_info *vi = netdev_priv(dev);
2040
2041 cmd->base.speed = vi->speed;
2042 cmd->base.duplex = vi->duplex;
2043 cmd->base.port = PORT_OTHER;
2044
2045 return 0;
2046 }
2047
2048 static void virtnet_init_settings(struct net_device *dev)
2049 {
2050 struct virtnet_info *vi = netdev_priv(dev);
2051
2052 vi->speed = SPEED_UNKNOWN;
2053 vi->duplex = DUPLEX_UNKNOWN;
2054 }
2055
2056 static void virtnet_update_settings(struct virtnet_info *vi)
2057 {
2058 u32 speed;
2059 u8 duplex;
2060
2061 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2062 return;
2063
2064 speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config,
2065 speed));
2066 if (ethtool_validate_speed(speed))
2067 vi->speed = speed;
2068 duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config,
2069 duplex));
2070 if (ethtool_validate_duplex(duplex))
2071 vi->duplex = duplex;
2072 }
2073
2074 static const struct ethtool_ops virtnet_ethtool_ops = {
2075 .get_drvinfo = virtnet_get_drvinfo,
2076 .get_link = ethtool_op_get_link,
2077 .get_ringparam = virtnet_get_ringparam,
2078 .get_strings = virtnet_get_strings,
2079 .get_sset_count = virtnet_get_sset_count,
2080 .get_ethtool_stats = virtnet_get_ethtool_stats,
2081 .set_channels = virtnet_set_channels,
2082 .get_channels = virtnet_get_channels,
2083 .get_ts_info = ethtool_op_get_ts_info,
2084 .get_link_ksettings = virtnet_get_link_ksettings,
2085 .set_link_ksettings = virtnet_set_link_ksettings,
2086 };
2087
2088 static void virtnet_freeze_down(struct virtio_device *vdev)
2089 {
2090 struct virtnet_info *vi = vdev->priv;
2091 int i;
2092
2093 /* Make sure no work handler is accessing the device */
2094 flush_work(&vi->config_work);
2095
2096 netif_device_detach(vi->dev);
2097 netif_tx_disable(vi->dev);
2098 cancel_delayed_work_sync(&vi->refill);
2099
2100 if (netif_running(vi->dev)) {
2101 for (i = 0; i < vi->max_queue_pairs; i++) {
2102 napi_disable(&vi->rq[i].napi);
2103 virtnet_napi_tx_disable(&vi->sq[i].napi);
2104 }
2105 }
2106 }
2107
2108 static int init_vqs(struct virtnet_info *vi);
2109
2110 static int virtnet_restore_up(struct virtio_device *vdev)
2111 {
2112 struct virtnet_info *vi = vdev->priv;
2113 int err, i;
2114
2115 err = init_vqs(vi);
2116 if (err)
2117 return err;
2118
2119 virtio_device_ready(vdev);
2120
2121 if (netif_running(vi->dev)) {
2122 for (i = 0; i < vi->curr_queue_pairs; i++)
2123 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2124 schedule_delayed_work(&vi->refill, 0);
2125
2126 for (i = 0; i < vi->max_queue_pairs; i++) {
2127 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2128 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2129 &vi->sq[i].napi);
2130 }
2131 }
2132
2133 netif_device_attach(vi->dev);
2134 return err;
2135 }
2136
2137 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2138 {
2139 struct scatterlist sg;
2140 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2141
2142 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2143
2144 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2145 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2146 dev_warn(&vi->dev->dev, "Fail to set guest offload. \n");
2147 return -EINVAL;
2148 }
2149
2150 return 0;
2151 }
2152
2153 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2154 {
2155 u64 offloads = 0;
2156
2157 if (!vi->guest_offloads)
2158 return 0;
2159
2160 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
2161 offloads = 1ULL << VIRTIO_NET_F_GUEST_CSUM;
2162
2163 return virtnet_set_guest_offloads(vi, offloads);
2164 }
2165
2166 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2167 {
2168 u64 offloads = vi->guest_offloads;
2169
2170 if (!vi->guest_offloads)
2171 return 0;
2172 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
2173 offloads |= 1ULL << VIRTIO_NET_F_GUEST_CSUM;
2174
2175 return virtnet_set_guest_offloads(vi, offloads);
2176 }
2177
2178 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2179 struct netlink_ext_ack *extack)
2180 {
2181 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2182 struct virtnet_info *vi = netdev_priv(dev);
2183 struct bpf_prog *old_prog;
2184 u16 xdp_qp = 0, curr_qp;
2185 int i, err;
2186
2187 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2188 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2189 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2190 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2191 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO))) {
2192 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first");
2193 return -EOPNOTSUPP;
2194 }
2195
2196 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2197 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2198 return -EINVAL;
2199 }
2200
2201 if (dev->mtu > max_sz) {
2202 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2203 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2204 return -EINVAL;
2205 }
2206
2207 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2208 if (prog)
2209 xdp_qp = nr_cpu_ids;
2210
2211 /* XDP requires extra queues for XDP_TX */
2212 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2213 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2214 netdev_warn(dev, "request %i queues but max is %i\n",
2215 curr_qp + xdp_qp, vi->max_queue_pairs);
2216 return -ENOMEM;
2217 }
2218
2219 if (prog) {
2220 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
2221 if (IS_ERR(prog))
2222 return PTR_ERR(prog);
2223 }
2224
2225 /* Make sure NAPI is not using any XDP TX queues for RX. */
2226 if (netif_running(dev))
2227 for (i = 0; i < vi->max_queue_pairs; i++)
2228 napi_disable(&vi->rq[i].napi);
2229
2230 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2231 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2232 if (err)
2233 goto err;
2234 vi->xdp_queue_pairs = xdp_qp;
2235
2236 for (i = 0; i < vi->max_queue_pairs; i++) {
2237 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2238 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2239 if (i == 0) {
2240 if (!old_prog)
2241 virtnet_clear_guest_offloads(vi);
2242 if (!prog)
2243 virtnet_restore_guest_offloads(vi);
2244 }
2245 if (old_prog)
2246 bpf_prog_put(old_prog);
2247 if (netif_running(dev))
2248 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2249 }
2250
2251 return 0;
2252
2253 err:
2254 for (i = 0; i < vi->max_queue_pairs; i++)
2255 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2256 if (prog)
2257 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2258 return err;
2259 }
2260
2261 static u32 virtnet_xdp_query(struct net_device *dev)
2262 {
2263 struct virtnet_info *vi = netdev_priv(dev);
2264 const struct bpf_prog *xdp_prog;
2265 int i;
2266
2267 for (i = 0; i < vi->max_queue_pairs; i++) {
2268 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2269 if (xdp_prog)
2270 return xdp_prog->aux->id;
2271 }
2272 return 0;
2273 }
2274
2275 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2276 {
2277 switch (xdp->command) {
2278 case XDP_SETUP_PROG:
2279 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2280 case XDP_QUERY_PROG:
2281 xdp->prog_id = virtnet_xdp_query(dev);
2282 xdp->prog_attached = !!xdp->prog_id;
2283 return 0;
2284 default:
2285 return -EINVAL;
2286 }
2287 }
2288
2289 static const struct net_device_ops virtnet_netdev = {
2290 .ndo_open = virtnet_open,
2291 .ndo_stop = virtnet_close,
2292 .ndo_start_xmit = start_xmit,
2293 .ndo_validate_addr = eth_validate_addr,
2294 .ndo_set_mac_address = virtnet_set_mac_address,
2295 .ndo_set_rx_mode = virtnet_set_rx_mode,
2296 .ndo_get_stats64 = virtnet_stats,
2297 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2298 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2299 #ifdef CONFIG_NET_POLL_CONTROLLER
2300 .ndo_poll_controller = virtnet_netpoll,
2301 #endif
2302 .ndo_bpf = virtnet_xdp,
2303 .ndo_xdp_xmit = virtnet_xdp_xmit,
2304 .ndo_xdp_flush = virtnet_xdp_flush,
2305 .ndo_features_check = passthru_features_check,
2306 };
2307
2308 static void virtnet_config_changed_work(struct work_struct *work)
2309 {
2310 struct virtnet_info *vi =
2311 container_of(work, struct virtnet_info, config_work);
2312 u16 v;
2313
2314 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2315 struct virtio_net_config, status, &v) < 0)
2316 return;
2317
2318 if (v & VIRTIO_NET_S_ANNOUNCE) {
2319 netdev_notify_peers(vi->dev);
2320 virtnet_ack_link_announce(vi);
2321 }
2322
2323 /* Ignore unknown (future) status bits */
2324 v &= VIRTIO_NET_S_LINK_UP;
2325
2326 if (vi->status == v)
2327 return;
2328
2329 vi->status = v;
2330
2331 if (vi->status & VIRTIO_NET_S_LINK_UP) {
2332 virtnet_update_settings(vi);
2333 netif_carrier_on(vi->dev);
2334 netif_tx_wake_all_queues(vi->dev);
2335 } else {
2336 netif_carrier_off(vi->dev);
2337 netif_tx_stop_all_queues(vi->dev);
2338 }
2339 }
2340
2341 static void virtnet_config_changed(struct virtio_device *vdev)
2342 {
2343 struct virtnet_info *vi = vdev->priv;
2344
2345 schedule_work(&vi->config_work);
2346 }
2347
2348 static void virtnet_free_queues(struct virtnet_info *vi)
2349 {
2350 int i;
2351
2352 for (i = 0; i < vi->max_queue_pairs; i++) {
2353 napi_hash_del(&vi->rq[i].napi);
2354 netif_napi_del(&vi->rq[i].napi);
2355 netif_napi_del(&vi->sq[i].napi);
2356 }
2357
2358 /* We called napi_hash_del() before netif_napi_del(),
2359 * we need to respect an RCU grace period before freeing vi->rq
2360 */
2361 synchronize_net();
2362
2363 kfree(vi->rq);
2364 kfree(vi->sq);
2365 kfree(vi->ctrl);
2366 }
2367
2368 static void _free_receive_bufs(struct virtnet_info *vi)
2369 {
2370 struct bpf_prog *old_prog;
2371 int i;
2372
2373 for (i = 0; i < vi->max_queue_pairs; i++) {
2374 while (vi->rq[i].pages)
2375 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2376
2377 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2378 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2379 if (old_prog)
2380 bpf_prog_put(old_prog);
2381 }
2382 }
2383
2384 static void free_receive_bufs(struct virtnet_info *vi)
2385 {
2386 rtnl_lock();
2387 _free_receive_bufs(vi);
2388 rtnl_unlock();
2389 }
2390
2391 static void free_receive_page_frags(struct virtnet_info *vi)
2392 {
2393 int i;
2394 for (i = 0; i < vi->max_queue_pairs; i++)
2395 if (vi->rq[i].alloc_frag.page)
2396 put_page(vi->rq[i].alloc_frag.page);
2397 }
2398
2399 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
2400 {
2401 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
2402 return false;
2403 else if (q < vi->curr_queue_pairs)
2404 return true;
2405 else
2406 return false;
2407 }
2408
2409 static void free_unused_bufs(struct virtnet_info *vi)
2410 {
2411 void *buf;
2412 int i;
2413
2414 for (i = 0; i < vi->max_queue_pairs; i++) {
2415 struct virtqueue *vq = vi->sq[i].vq;
2416 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2417 if (!is_xdp_raw_buffer_queue(vi, i))
2418 dev_kfree_skb(buf);
2419 else
2420 put_page(virt_to_head_page(buf));
2421 }
2422 }
2423
2424 for (i = 0; i < vi->max_queue_pairs; i++) {
2425 struct virtqueue *vq = vi->rq[i].vq;
2426
2427 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2428 if (vi->mergeable_rx_bufs) {
2429 put_page(virt_to_head_page(buf));
2430 } else if (vi->big_packets) {
2431 give_pages(&vi->rq[i], buf);
2432 } else {
2433 put_page(virt_to_head_page(buf));
2434 }
2435 }
2436 }
2437 }
2438
2439 static void virtnet_del_vqs(struct virtnet_info *vi)
2440 {
2441 struct virtio_device *vdev = vi->vdev;
2442
2443 virtnet_clean_affinity(vi, -1);
2444
2445 vdev->config->del_vqs(vdev);
2446
2447 virtnet_free_queues(vi);
2448 }
2449
2450 /* How large should a single buffer be so a queue full of these can fit at
2451 * least one full packet?
2452 * Logic below assumes the mergeable buffer header is used.
2453 */
2454 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2455 {
2456 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2457 unsigned int rq_size = virtqueue_get_vring_size(vq);
2458 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2459 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2460 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2461
2462 return max(max(min_buf_len, hdr_len) - hdr_len,
2463 (unsigned int)GOOD_PACKET_LEN);
2464 }
2465
2466 static int virtnet_find_vqs(struct virtnet_info *vi)
2467 {
2468 vq_callback_t **callbacks;
2469 struct virtqueue **vqs;
2470 int ret = -ENOMEM;
2471 int i, total_vqs;
2472 const char **names;
2473 bool *ctx;
2474
2475 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2476 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2477 * possible control vq.
2478 */
2479 total_vqs = vi->max_queue_pairs * 2 +
2480 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2481
2482 /* Allocate space for find_vqs parameters */
2483 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
2484 if (!vqs)
2485 goto err_vq;
2486 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
2487 if (!callbacks)
2488 goto err_callback;
2489 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
2490 if (!names)
2491 goto err_names;
2492 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2493 ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL);
2494 if (!ctx)
2495 goto err_ctx;
2496 } else {
2497 ctx = NULL;
2498 }
2499
2500 /* Parameters for control virtqueue, if any */
2501 if (vi->has_cvq) {
2502 callbacks[total_vqs - 1] = NULL;
2503 names[total_vqs - 1] = "control";
2504 }
2505
2506 /* Allocate/initialize parameters for send/receive virtqueues */
2507 for (i = 0; i < vi->max_queue_pairs; i++) {
2508 callbacks[rxq2vq(i)] = skb_recv_done;
2509 callbacks[txq2vq(i)] = skb_xmit_done;
2510 sprintf(vi->rq[i].name, "input.%d", i);
2511 sprintf(vi->sq[i].name, "output.%d", i);
2512 names[rxq2vq(i)] = vi->rq[i].name;
2513 names[txq2vq(i)] = vi->sq[i].name;
2514 if (ctx)
2515 ctx[rxq2vq(i)] = true;
2516 }
2517
2518 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2519 names, ctx, NULL);
2520 if (ret)
2521 goto err_find;
2522
2523 if (vi->has_cvq) {
2524 vi->cvq = vqs[total_vqs - 1];
2525 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2526 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2527 }
2528
2529 for (i = 0; i < vi->max_queue_pairs; i++) {
2530 vi->rq[i].vq = vqs[rxq2vq(i)];
2531 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2532 vi->sq[i].vq = vqs[txq2vq(i)];
2533 }
2534
2535 kfree(names);
2536 kfree(callbacks);
2537 kfree(vqs);
2538 kfree(ctx);
2539
2540 return 0;
2541
2542 err_find:
2543 kfree(ctx);
2544 err_ctx:
2545 kfree(names);
2546 err_names:
2547 kfree(callbacks);
2548 err_callback:
2549 kfree(vqs);
2550 err_vq:
2551 return ret;
2552 }
2553
2554 static int virtnet_alloc_queues(struct virtnet_info *vi)
2555 {
2556 int i;
2557
2558 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2559 if (!vi->ctrl)
2560 goto err_ctrl;
2561 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2562 if (!vi->sq)
2563 goto err_sq;
2564 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2565 if (!vi->rq)
2566 goto err_rq;
2567
2568 INIT_DELAYED_WORK(&vi->refill, refill_work);
2569 for (i = 0; i < vi->max_queue_pairs; i++) {
2570 vi->rq[i].pages = NULL;
2571 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2572 napi_weight);
2573 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2574 napi_tx ? napi_weight : 0);
2575
2576 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2577 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2578 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2579
2580 u64_stats_init(&vi->rq[i].stats.syncp);
2581 u64_stats_init(&vi->sq[i].stats.syncp);
2582 }
2583
2584 return 0;
2585
2586 err_rq:
2587 kfree(vi->sq);
2588 err_sq:
2589 kfree(vi->ctrl);
2590 err_ctrl:
2591 return -ENOMEM;
2592 }
2593
2594 static int init_vqs(struct virtnet_info *vi)
2595 {
2596 int ret;
2597
2598 /* Allocate send & receive queues */
2599 ret = virtnet_alloc_queues(vi);
2600 if (ret)
2601 goto err;
2602
2603 ret = virtnet_find_vqs(vi);
2604 if (ret)
2605 goto err_free;
2606
2607 get_online_cpus();
2608 virtnet_set_affinity(vi);
2609 put_online_cpus();
2610
2611 return 0;
2612
2613 err_free:
2614 virtnet_free_queues(vi);
2615 err:
2616 return ret;
2617 }
2618
2619 #ifdef CONFIG_SYSFS
2620 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2621 char *buf)
2622 {
2623 struct virtnet_info *vi = netdev_priv(queue->dev);
2624 unsigned int queue_index = get_netdev_rx_queue_index(queue);
2625 unsigned int headroom = virtnet_get_headroom(vi);
2626 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2627 struct ewma_pkt_len *avg;
2628
2629 BUG_ON(queue_index >= vi->max_queue_pairs);
2630 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2631 return sprintf(buf, "%u\n",
2632 get_mergeable_buf_len(&vi->rq[queue_index], avg,
2633 SKB_DATA_ALIGN(headroom + tailroom)));
2634 }
2635
2636 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2637 __ATTR_RO(mergeable_rx_buffer_size);
2638
2639 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2640 &mergeable_rx_buffer_size_attribute.attr,
2641 NULL
2642 };
2643
2644 static const struct attribute_group virtio_net_mrg_rx_group = {
2645 .name = "virtio_net",
2646 .attrs = virtio_net_mrg_rx_attrs
2647 };
2648 #endif
2649
2650 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2651 unsigned int fbit,
2652 const char *fname, const char *dname)
2653 {
2654 if (!virtio_has_feature(vdev, fbit))
2655 return false;
2656
2657 dev_err(&vdev->dev, "device advertises feature %s but not %s",
2658 fname, dname);
2659
2660 return true;
2661 }
2662
2663 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2664 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2665
2666 static bool virtnet_validate_features(struct virtio_device *vdev)
2667 {
2668 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2669 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2670 "VIRTIO_NET_F_CTRL_VQ") ||
2671 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2672 "VIRTIO_NET_F_CTRL_VQ") ||
2673 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2674 "VIRTIO_NET_F_CTRL_VQ") ||
2675 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2676 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2677 "VIRTIO_NET_F_CTRL_VQ"))) {
2678 return false;
2679 }
2680
2681 return true;
2682 }
2683
2684 #define MIN_MTU ETH_MIN_MTU
2685 #define MAX_MTU ETH_MAX_MTU
2686
2687 static int virtnet_validate(struct virtio_device *vdev)
2688 {
2689 if (!vdev->config->get) {
2690 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2691 __func__);
2692 return -EINVAL;
2693 }
2694
2695 if (!virtnet_validate_features(vdev))
2696 return -EINVAL;
2697
2698 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2699 int mtu = virtio_cread16(vdev,
2700 offsetof(struct virtio_net_config,
2701 mtu));
2702 if (mtu < MIN_MTU)
2703 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2704 }
2705
2706 return 0;
2707 }
2708
2709 static int virtnet_probe(struct virtio_device *vdev)
2710 {
2711 int i, err = -ENOMEM;
2712 struct net_device *dev;
2713 struct virtnet_info *vi;
2714 u16 max_queue_pairs;
2715 int mtu;
2716
2717 /* Find if host supports multiqueue virtio_net device */
2718 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2719 struct virtio_net_config,
2720 max_virtqueue_pairs, &max_queue_pairs);
2721
2722 /* We need at least 2 queue's */
2723 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2724 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2725 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2726 max_queue_pairs = 1;
2727
2728 /* Allocate ourselves a network device with room for our info */
2729 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2730 if (!dev)
2731 return -ENOMEM;
2732
2733 /* Set up network device as normal. */
2734 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2735 dev->netdev_ops = &virtnet_netdev;
2736 dev->features = NETIF_F_HIGHDMA;
2737
2738 dev->ethtool_ops = &virtnet_ethtool_ops;
2739 SET_NETDEV_DEV(dev, &vdev->dev);
2740
2741 /* Do we support "hardware" checksums? */
2742 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2743 /* This opens up the world of extra features. */
2744 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2745 if (csum)
2746 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2747
2748 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2749 dev->hw_features |= NETIF_F_TSO
2750 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2751 }
2752 /* Individual feature bits: what can host handle? */
2753 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2754 dev->hw_features |= NETIF_F_TSO;
2755 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2756 dev->hw_features |= NETIF_F_TSO6;
2757 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2758 dev->hw_features |= NETIF_F_TSO_ECN;
2759
2760 dev->features |= NETIF_F_GSO_ROBUST;
2761
2762 if (gso)
2763 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
2764 /* (!csum && gso) case will be fixed by register_netdev() */
2765 }
2766 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2767 dev->features |= NETIF_F_RXCSUM;
2768
2769 dev->vlan_features = dev->features;
2770
2771 /* MTU range: 68 - 65535 */
2772 dev->min_mtu = MIN_MTU;
2773 dev->max_mtu = MAX_MTU;
2774
2775 /* Configuration may specify what MAC to use. Otherwise random. */
2776 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2777 virtio_cread_bytes(vdev,
2778 offsetof(struct virtio_net_config, mac),
2779 dev->dev_addr, dev->addr_len);
2780 else
2781 eth_hw_addr_random(dev);
2782
2783 /* Set up our device-specific information */
2784 vi = netdev_priv(dev);
2785 vi->dev = dev;
2786 vi->vdev = vdev;
2787 vdev->priv = vi;
2788
2789 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2790
2791 /* If we can receive ANY GSO packets, we must allocate large ones. */
2792 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2793 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2794 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2795 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2796 vi->big_packets = true;
2797
2798 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2799 vi->mergeable_rx_bufs = true;
2800
2801 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2802 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2803 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2804 else
2805 vi->hdr_len = sizeof(struct virtio_net_hdr);
2806
2807 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2808 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2809 vi->any_header_sg = true;
2810
2811 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2812 vi->has_cvq = true;
2813
2814 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2815 mtu = virtio_cread16(vdev,
2816 offsetof(struct virtio_net_config,
2817 mtu));
2818 if (mtu < dev->min_mtu) {
2819 /* Should never trigger: MTU was previously validated
2820 * in virtnet_validate.
2821 */
2822 dev_err(&vdev->dev, "device MTU appears to have changed "
2823 "it is now %d < %d", mtu, dev->min_mtu);
2824 goto free;
2825 }
2826
2827 dev->mtu = mtu;
2828 dev->max_mtu = mtu;
2829
2830 /* TODO: size buffers correctly in this case. */
2831 if (dev->mtu > ETH_DATA_LEN)
2832 vi->big_packets = true;
2833 }
2834
2835 if (vi->any_header_sg)
2836 dev->needed_headroom = vi->hdr_len;
2837
2838 /* Enable multiqueue by default */
2839 if (num_online_cpus() >= max_queue_pairs)
2840 vi->curr_queue_pairs = max_queue_pairs;
2841 else
2842 vi->curr_queue_pairs = num_online_cpus();
2843 vi->max_queue_pairs = max_queue_pairs;
2844
2845 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2846 err = init_vqs(vi);
2847 if (err)
2848 goto free;
2849
2850 #ifdef CONFIG_SYSFS
2851 if (vi->mergeable_rx_bufs)
2852 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2853 #endif
2854 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2855 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2856
2857 virtnet_init_settings(dev);
2858
2859 err = register_netdev(dev);
2860 if (err) {
2861 pr_debug("virtio_net: registering device failed\n");
2862 goto free_vqs;
2863 }
2864
2865 virtio_device_ready(vdev);
2866
2867 err = virtnet_cpu_notif_add(vi);
2868 if (err) {
2869 pr_debug("virtio_net: registering cpu notifier failed\n");
2870 goto free_unregister_netdev;
2871 }
2872
2873 virtnet_set_queues(vi, vi->curr_queue_pairs);
2874
2875 /* Assume link up if device can't report link status,
2876 otherwise get link status from config. */
2877 netif_carrier_off(dev);
2878 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2879 schedule_work(&vi->config_work);
2880 } else {
2881 vi->status = VIRTIO_NET_S_LINK_UP;
2882 virtnet_update_settings(vi);
2883 netif_carrier_on(dev);
2884 }
2885
2886 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
2887 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
2888 set_bit(guest_offloads[i], &vi->guest_offloads);
2889
2890 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2891 dev->name, max_queue_pairs);
2892
2893 return 0;
2894
2895 free_unregister_netdev:
2896 vi->vdev->config->reset(vdev);
2897
2898 unregister_netdev(dev);
2899 free_vqs:
2900 cancel_delayed_work_sync(&vi->refill);
2901 free_receive_page_frags(vi);
2902 virtnet_del_vqs(vi);
2903 free:
2904 free_netdev(dev);
2905 return err;
2906 }
2907
2908 static void remove_vq_common(struct virtnet_info *vi)
2909 {
2910 vi->vdev->config->reset(vi->vdev);
2911
2912 /* Free unused buffers in both send and recv, if any. */
2913 free_unused_bufs(vi);
2914
2915 free_receive_bufs(vi);
2916
2917 free_receive_page_frags(vi);
2918
2919 virtnet_del_vqs(vi);
2920 }
2921
2922 static void virtnet_remove(struct virtio_device *vdev)
2923 {
2924 struct virtnet_info *vi = vdev->priv;
2925
2926 virtnet_cpu_notif_remove(vi);
2927
2928 /* Make sure no work handler is accessing the device. */
2929 flush_work(&vi->config_work);
2930
2931 unregister_netdev(vi->dev);
2932
2933 remove_vq_common(vi);
2934
2935 free_netdev(vi->dev);
2936 }
2937
2938 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
2939 {
2940 struct virtnet_info *vi = vdev->priv;
2941
2942 virtnet_cpu_notif_remove(vi);
2943 virtnet_freeze_down(vdev);
2944 remove_vq_common(vi);
2945
2946 return 0;
2947 }
2948
2949 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
2950 {
2951 struct virtnet_info *vi = vdev->priv;
2952 int err;
2953
2954 err = virtnet_restore_up(vdev);
2955 if (err)
2956 return err;
2957 virtnet_set_queues(vi, vi->curr_queue_pairs);
2958
2959 err = virtnet_cpu_notif_add(vi);
2960 if (err)
2961 return err;
2962
2963 return 0;
2964 }
2965
2966 static struct virtio_device_id id_table[] = {
2967 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2968 { 0 },
2969 };
2970
2971 #define VIRTNET_FEATURES \
2972 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2973 VIRTIO_NET_F_MAC, \
2974 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2975 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2976 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2977 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2978 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2979 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2980 VIRTIO_NET_F_CTRL_MAC_ADDR, \
2981 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
2982 VIRTIO_NET_F_SPEED_DUPLEX
2983
2984 static unsigned int features[] = {
2985 VIRTNET_FEATURES,
2986 };
2987
2988 static unsigned int features_legacy[] = {
2989 VIRTNET_FEATURES,
2990 VIRTIO_NET_F_GSO,
2991 VIRTIO_F_ANY_LAYOUT,
2992 };
2993
2994 static struct virtio_driver virtio_net_driver = {
2995 .feature_table = features,
2996 .feature_table_size = ARRAY_SIZE(features),
2997 .feature_table_legacy = features_legacy,
2998 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2999 .driver.name = KBUILD_MODNAME,
3000 .driver.owner = THIS_MODULE,
3001 .id_table = id_table,
3002 .validate = virtnet_validate,
3003 .probe = virtnet_probe,
3004 .remove = virtnet_remove,
3005 .config_changed = virtnet_config_changed,
3006 #ifdef CONFIG_PM_SLEEP
3007 .freeze = virtnet_freeze,
3008 .restore = virtnet_restore,
3009 #endif
3010 };
3011
3012 static __init int virtio_net_driver_init(void)
3013 {
3014 int ret;
3015
3016 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3017 virtnet_cpu_online,
3018 virtnet_cpu_down_prep);
3019 if (ret < 0)
3020 goto out;
3021 virtionet_online = ret;
3022 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3023 NULL, virtnet_cpu_dead);
3024 if (ret)
3025 goto err_dead;
3026
3027 ret = register_virtio_driver(&virtio_net_driver);
3028 if (ret)
3029 goto err_virtio;
3030 return 0;
3031 err_virtio:
3032 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3033 err_dead:
3034 cpuhp_remove_multi_state(virtionet_online);
3035 out:
3036 return ret;
3037 }
3038 module_init(virtio_net_driver_init);
3039
3040 static __exit void virtio_net_driver_exit(void)
3041 {
3042 unregister_virtio_driver(&virtio_net_driver);
3043 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3044 cpuhp_remove_multi_state(virtionet_online);
3045 }
3046 module_exit(virtio_net_driver_exit);
3047
3048 MODULE_DEVICE_TABLE(virtio, id_table);
3049 MODULE_DESCRIPTION("Virtio network driver");
3050 MODULE_LICENSE("GPL");
Linux with added FPC-III support