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