Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * drivers/net/veth.c
4 *
5 * Copyright (C) 2007 OpenVZ http://openvz.org, SWsoft Inc
6 *
7 * Author: Pavel Emelianov <xemul@openvz.org>
8 * Ethtool interface from: Eric W. Biederman <ebiederm@xmission.com>
9 *
10 */
11
12#include <linux/netdevice.h>
13#include <linux/slab.h>
14#include <linux/ethtool.h>
15#include <linux/etherdevice.h>
16#include <linux/u64_stats_sync.h>
17
18#include <net/rtnetlink.h>
19#include <net/dst.h>
20#include <net/xfrm.h>
21#include <net/xdp.h>
22#include <linux/veth.h>
23#include <linux/module.h>
24#include <linux/bpf.h>
25#include <linux/filter.h>
26#include <linux/ptr_ring.h>
27#include <linux/bpf_trace.h>
28#include <linux/net_tstamp.h>
29#include <linux/skbuff_ref.h>
30#include <net/page_pool/helpers.h>
31
32#define DRV_NAME "veth"
33#define DRV_VERSION "1.0"
34
35#define VETH_XDP_FLAG BIT(0)
36#define VETH_RING_SIZE 256
37#define VETH_XDP_HEADROOM (XDP_PACKET_HEADROOM + NET_IP_ALIGN)
38
39#define VETH_XDP_TX_BULK_SIZE 16
40#define VETH_XDP_BATCH 16
41
42struct veth_stats {
43 u64 rx_drops;
44 /* xdp */
45 u64 xdp_packets;
46 u64 xdp_bytes;
47 u64 xdp_redirect;
48 u64 xdp_drops;
49 u64 xdp_tx;
50 u64 xdp_tx_err;
51 u64 peer_tq_xdp_xmit;
52 u64 peer_tq_xdp_xmit_err;
53};
54
55struct veth_rq_stats {
56 struct veth_stats vs;
57 struct u64_stats_sync syncp;
58};
59
60struct veth_rq {
61 struct napi_struct xdp_napi;
62 struct napi_struct __rcu *napi; /* points to xdp_napi when the latter is initialized */
63 struct net_device *dev;
64 struct bpf_prog __rcu *xdp_prog;
65 struct xdp_mem_info xdp_mem;
66 struct veth_rq_stats stats;
67 bool rx_notify_masked;
68 struct ptr_ring xdp_ring;
69 struct xdp_rxq_info xdp_rxq;
70 struct page_pool *page_pool;
71};
72
73struct veth_priv {
74 struct net_device __rcu *peer;
75 atomic64_t dropped;
76 struct bpf_prog *_xdp_prog;
77 struct veth_rq *rq;
78 unsigned int requested_headroom;
79};
80
81struct veth_xdp_tx_bq {
82 struct xdp_frame *q[VETH_XDP_TX_BULK_SIZE];
83 unsigned int count;
84};
85
86/*
87 * ethtool interface
88 */
89
90struct veth_q_stat_desc {
91 char desc[ETH_GSTRING_LEN];
92 size_t offset;
93};
94
95#define VETH_RQ_STAT(m) offsetof(struct veth_stats, m)
96
97static const struct veth_q_stat_desc veth_rq_stats_desc[] = {
98 { "xdp_packets", VETH_RQ_STAT(xdp_packets) },
99 { "xdp_bytes", VETH_RQ_STAT(xdp_bytes) },
100 { "drops", VETH_RQ_STAT(rx_drops) },
101 { "xdp_redirect", VETH_RQ_STAT(xdp_redirect) },
102 { "xdp_drops", VETH_RQ_STAT(xdp_drops) },
103 { "xdp_tx", VETH_RQ_STAT(xdp_tx) },
104 { "xdp_tx_errors", VETH_RQ_STAT(xdp_tx_err) },
105};
106
107#define VETH_RQ_STATS_LEN ARRAY_SIZE(veth_rq_stats_desc)
108
109static const struct veth_q_stat_desc veth_tq_stats_desc[] = {
110 { "xdp_xmit", VETH_RQ_STAT(peer_tq_xdp_xmit) },
111 { "xdp_xmit_errors", VETH_RQ_STAT(peer_tq_xdp_xmit_err) },
112};
113
114#define VETH_TQ_STATS_LEN ARRAY_SIZE(veth_tq_stats_desc)
115
116static struct {
117 const char string[ETH_GSTRING_LEN];
118} ethtool_stats_keys[] = {
119 { "peer_ifindex" },
120};
121
122struct veth_xdp_buff {
123 struct xdp_buff xdp;
124 struct sk_buff *skb;
125};
126
127static int veth_get_link_ksettings(struct net_device *dev,
128 struct ethtool_link_ksettings *cmd)
129{
130 cmd->base.speed = SPEED_10000;
131 cmd->base.duplex = DUPLEX_FULL;
132 cmd->base.port = PORT_TP;
133 cmd->base.autoneg = AUTONEG_DISABLE;
134 return 0;
135}
136
137static void veth_get_drvinfo(struct net_device *dev, struct ethtool_drvinfo *info)
138{
139 strscpy(info->driver, DRV_NAME, sizeof(info->driver));
140 strscpy(info->version, DRV_VERSION, sizeof(info->version));
141}
142
143static void veth_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
144{
145 u8 *p = buf;
146 int i, j;
147
148 switch(stringset) {
149 case ETH_SS_STATS:
150 memcpy(p, ðtool_stats_keys, sizeof(ethtool_stats_keys));
151 p += sizeof(ethtool_stats_keys);
152 for (i = 0; i < dev->real_num_rx_queues; i++)
153 for (j = 0; j < VETH_RQ_STATS_LEN; j++)
154 ethtool_sprintf(&p, "rx_queue_%u_%.18s",
155 i, veth_rq_stats_desc[j].desc);
156
157 for (i = 0; i < dev->real_num_tx_queues; i++)
158 for (j = 0; j < VETH_TQ_STATS_LEN; j++)
159 ethtool_sprintf(&p, "tx_queue_%u_%.18s",
160 i, veth_tq_stats_desc[j].desc);
161
162 page_pool_ethtool_stats_get_strings(p);
163 break;
164 }
165}
166
167static int veth_get_sset_count(struct net_device *dev, int sset)
168{
169 switch (sset) {
170 case ETH_SS_STATS:
171 return ARRAY_SIZE(ethtool_stats_keys) +
172 VETH_RQ_STATS_LEN * dev->real_num_rx_queues +
173 VETH_TQ_STATS_LEN * dev->real_num_tx_queues +
174 page_pool_ethtool_stats_get_count();
175 default:
176 return -EOPNOTSUPP;
177 }
178}
179
180static void veth_get_page_pool_stats(struct net_device *dev, u64 *data)
181{
182#ifdef CONFIG_PAGE_POOL_STATS
183 struct veth_priv *priv = netdev_priv(dev);
184 struct page_pool_stats pp_stats = {};
185 int i;
186
187 for (i = 0; i < dev->real_num_rx_queues; i++) {
188 if (!priv->rq[i].page_pool)
189 continue;
190 page_pool_get_stats(priv->rq[i].page_pool, &pp_stats);
191 }
192 page_pool_ethtool_stats_get(data, &pp_stats);
193#endif /* CONFIG_PAGE_POOL_STATS */
194}
195
196static void veth_get_ethtool_stats(struct net_device *dev,
197 struct ethtool_stats *stats, u64 *data)
198{
199 struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
200 struct net_device *peer = rtnl_dereference(priv->peer);
201 int i, j, idx, pp_idx;
202
203 data[0] = peer ? peer->ifindex : 0;
204 idx = 1;
205 for (i = 0; i < dev->real_num_rx_queues; i++) {
206 const struct veth_rq_stats *rq_stats = &priv->rq[i].stats;
207 const void *stats_base = (void *)&rq_stats->vs;
208 unsigned int start;
209 size_t offset;
210
211 do {
212 start = u64_stats_fetch_begin(&rq_stats->syncp);
213 for (j = 0; j < VETH_RQ_STATS_LEN; j++) {
214 offset = veth_rq_stats_desc[j].offset;
215 data[idx + j] = *(u64 *)(stats_base + offset);
216 }
217 } while (u64_stats_fetch_retry(&rq_stats->syncp, start));
218 idx += VETH_RQ_STATS_LEN;
219 }
220 pp_idx = idx;
221
222 if (!peer)
223 goto page_pool_stats;
224
225 rcv_priv = netdev_priv(peer);
226 for (i = 0; i < peer->real_num_rx_queues; i++) {
227 const struct veth_rq_stats *rq_stats = &rcv_priv->rq[i].stats;
228 const void *base = (void *)&rq_stats->vs;
229 unsigned int start, tx_idx = idx;
230 size_t offset;
231
232 tx_idx += (i % dev->real_num_tx_queues) * VETH_TQ_STATS_LEN;
233 do {
234 start = u64_stats_fetch_begin(&rq_stats->syncp);
235 for (j = 0; j < VETH_TQ_STATS_LEN; j++) {
236 offset = veth_tq_stats_desc[j].offset;
237 data[tx_idx + j] += *(u64 *)(base + offset);
238 }
239 } while (u64_stats_fetch_retry(&rq_stats->syncp, start));
240 }
241 pp_idx = idx + dev->real_num_tx_queues * VETH_TQ_STATS_LEN;
242
243page_pool_stats:
244 veth_get_page_pool_stats(dev, &data[pp_idx]);
245}
246
247static void veth_get_channels(struct net_device *dev,
248 struct ethtool_channels *channels)
249{
250 channels->tx_count = dev->real_num_tx_queues;
251 channels->rx_count = dev->real_num_rx_queues;
252 channels->max_tx = dev->num_tx_queues;
253 channels->max_rx = dev->num_rx_queues;
254}
255
256static int veth_set_channels(struct net_device *dev,
257 struct ethtool_channels *ch);
258
259static const struct ethtool_ops veth_ethtool_ops = {
260 .get_drvinfo = veth_get_drvinfo,
261 .get_link = ethtool_op_get_link,
262 .get_strings = veth_get_strings,
263 .get_sset_count = veth_get_sset_count,
264 .get_ethtool_stats = veth_get_ethtool_stats,
265 .get_link_ksettings = veth_get_link_ksettings,
266 .get_ts_info = ethtool_op_get_ts_info,
267 .get_channels = veth_get_channels,
268 .set_channels = veth_set_channels,
269};
270
271/* general routines */
272
273static bool veth_is_xdp_frame(void *ptr)
274{
275 return (unsigned long)ptr & VETH_XDP_FLAG;
276}
277
278static struct xdp_frame *veth_ptr_to_xdp(void *ptr)
279{
280 return (void *)((unsigned long)ptr & ~VETH_XDP_FLAG);
281}
282
283static void *veth_xdp_to_ptr(struct xdp_frame *xdp)
284{
285 return (void *)((unsigned long)xdp | VETH_XDP_FLAG);
286}
287
288static void veth_ptr_free(void *ptr)
289{
290 if (veth_is_xdp_frame(ptr))
291 xdp_return_frame(veth_ptr_to_xdp(ptr));
292 else
293 kfree_skb(ptr);
294}
295
296static void __veth_xdp_flush(struct veth_rq *rq)
297{
298 /* Write ptr_ring before reading rx_notify_masked */
299 smp_mb();
300 if (!READ_ONCE(rq->rx_notify_masked) &&
301 napi_schedule_prep(&rq->xdp_napi)) {
302 WRITE_ONCE(rq->rx_notify_masked, true);
303 __napi_schedule(&rq->xdp_napi);
304 }
305}
306
307static int veth_xdp_rx(struct veth_rq *rq, struct sk_buff *skb)
308{
309 if (unlikely(ptr_ring_produce(&rq->xdp_ring, skb))) {
310 dev_kfree_skb_any(skb);
311 return NET_RX_DROP;
312 }
313
314 return NET_RX_SUCCESS;
315}
316
317static int veth_forward_skb(struct net_device *dev, struct sk_buff *skb,
318 struct veth_rq *rq, bool xdp)
319{
320 return __dev_forward_skb(dev, skb) ?: xdp ?
321 veth_xdp_rx(rq, skb) :
322 __netif_rx(skb);
323}
324
325/* return true if the specified skb has chances of GRO aggregation
326 * Don't strive for accuracy, but try to avoid GRO overhead in the most
327 * common scenarios.
328 * When XDP is enabled, all traffic is considered eligible, as the xmit
329 * device has TSO off.
330 * When TSO is enabled on the xmit device, we are likely interested only
331 * in UDP aggregation, explicitly check for that if the skb is suspected
332 * - the sock_wfree destructor is used by UDP, ICMP and XDP sockets -
333 * to belong to locally generated UDP traffic.
334 */
335static bool veth_skb_is_eligible_for_gro(const struct net_device *dev,
336 const struct net_device *rcv,
337 const struct sk_buff *skb)
338{
339 return !(dev->features & NETIF_F_ALL_TSO) ||
340 (skb->destructor == sock_wfree &&
341 rcv->features & (NETIF_F_GRO_FRAGLIST | NETIF_F_GRO_UDP_FWD));
342}
343
344static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev)
345{
346 struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
347 struct veth_rq *rq = NULL;
348 int ret = NETDEV_TX_OK;
349 struct net_device *rcv;
350 int length = skb->len;
351 bool use_napi = false;
352 int rxq;
353
354 rcu_read_lock();
355 rcv = rcu_dereference(priv->peer);
356 if (unlikely(!rcv) || !pskb_may_pull(skb, ETH_HLEN)) {
357 kfree_skb(skb);
358 goto drop;
359 }
360
361 rcv_priv = netdev_priv(rcv);
362 rxq = skb_get_queue_mapping(skb);
363 if (rxq < rcv->real_num_rx_queues) {
364 rq = &rcv_priv->rq[rxq];
365
366 /* The napi pointer is available when an XDP program is
367 * attached or when GRO is enabled
368 * Don't bother with napi/GRO if the skb can't be aggregated
369 */
370 use_napi = rcu_access_pointer(rq->napi) &&
371 veth_skb_is_eligible_for_gro(dev, rcv, skb);
372 }
373
374 skb_tx_timestamp(skb);
375 if (likely(veth_forward_skb(rcv, skb, rq, use_napi) == NET_RX_SUCCESS)) {
376 if (!use_napi)
377 dev_sw_netstats_tx_add(dev, 1, length);
378 else
379 __veth_xdp_flush(rq);
380 } else {
381drop:
382 atomic64_inc(&priv->dropped);
383 ret = NET_XMIT_DROP;
384 }
385
386 rcu_read_unlock();
387
388 return ret;
389}
390
391static void veth_stats_rx(struct veth_stats *result, struct net_device *dev)
392{
393 struct veth_priv *priv = netdev_priv(dev);
394 int i;
395
396 result->peer_tq_xdp_xmit_err = 0;
397 result->xdp_packets = 0;
398 result->xdp_tx_err = 0;
399 result->xdp_bytes = 0;
400 result->rx_drops = 0;
401 for (i = 0; i < dev->num_rx_queues; i++) {
402 u64 packets, bytes, drops, xdp_tx_err, peer_tq_xdp_xmit_err;
403 struct veth_rq_stats *stats = &priv->rq[i].stats;
404 unsigned int start;
405
406 do {
407 start = u64_stats_fetch_begin(&stats->syncp);
408 peer_tq_xdp_xmit_err = stats->vs.peer_tq_xdp_xmit_err;
409 xdp_tx_err = stats->vs.xdp_tx_err;
410 packets = stats->vs.xdp_packets;
411 bytes = stats->vs.xdp_bytes;
412 drops = stats->vs.rx_drops;
413 } while (u64_stats_fetch_retry(&stats->syncp, start));
414 result->peer_tq_xdp_xmit_err += peer_tq_xdp_xmit_err;
415 result->xdp_tx_err += xdp_tx_err;
416 result->xdp_packets += packets;
417 result->xdp_bytes += bytes;
418 result->rx_drops += drops;
419 }
420}
421
422static void veth_get_stats64(struct net_device *dev,
423 struct rtnl_link_stats64 *tot)
424{
425 struct veth_priv *priv = netdev_priv(dev);
426 struct net_device *peer;
427 struct veth_stats rx;
428
429 tot->tx_dropped = atomic64_read(&priv->dropped);
430 dev_fetch_sw_netstats(tot, dev->tstats);
431
432 veth_stats_rx(&rx, dev);
433 tot->tx_dropped += rx.xdp_tx_err;
434 tot->rx_dropped = rx.rx_drops + rx.peer_tq_xdp_xmit_err;
435 tot->rx_bytes += rx.xdp_bytes;
436 tot->rx_packets += rx.xdp_packets;
437
438 rcu_read_lock();
439 peer = rcu_dereference(priv->peer);
440 if (peer) {
441 struct rtnl_link_stats64 tot_peer = {};
442
443 dev_fetch_sw_netstats(&tot_peer, peer->tstats);
444 tot->rx_bytes += tot_peer.tx_bytes;
445 tot->rx_packets += tot_peer.tx_packets;
446
447 veth_stats_rx(&rx, peer);
448 tot->tx_dropped += rx.peer_tq_xdp_xmit_err;
449 tot->rx_dropped += rx.xdp_tx_err;
450 tot->tx_bytes += rx.xdp_bytes;
451 tot->tx_packets += rx.xdp_packets;
452 }
453 rcu_read_unlock();
454}
455
456/* fake multicast ability */
457static void veth_set_multicast_list(struct net_device *dev)
458{
459}
460
461static int veth_select_rxq(struct net_device *dev)
462{
463 return smp_processor_id() % dev->real_num_rx_queues;
464}
465
466static struct net_device *veth_peer_dev(struct net_device *dev)
467{
468 struct veth_priv *priv = netdev_priv(dev);
469
470 /* Callers must be under RCU read side. */
471 return rcu_dereference(priv->peer);
472}
473
474static int veth_xdp_xmit(struct net_device *dev, int n,
475 struct xdp_frame **frames,
476 u32 flags, bool ndo_xmit)
477{
478 struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
479 int i, ret = -ENXIO, nxmit = 0;
480 struct net_device *rcv;
481 unsigned int max_len;
482 struct veth_rq *rq;
483
484 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
485 return -EINVAL;
486
487 rcu_read_lock();
488 rcv = rcu_dereference(priv->peer);
489 if (unlikely(!rcv))
490 goto out;
491
492 rcv_priv = netdev_priv(rcv);
493 rq = &rcv_priv->rq[veth_select_rxq(rcv)];
494 /* The napi pointer is set if NAPI is enabled, which ensures that
495 * xdp_ring is initialized on receive side and the peer device is up.
496 */
497 if (!rcu_access_pointer(rq->napi))
498 goto out;
499
500 max_len = rcv->mtu + rcv->hard_header_len + VLAN_HLEN;
501
502 spin_lock(&rq->xdp_ring.producer_lock);
503 for (i = 0; i < n; i++) {
504 struct xdp_frame *frame = frames[i];
505 void *ptr = veth_xdp_to_ptr(frame);
506
507 if (unlikely(xdp_get_frame_len(frame) > max_len ||
508 __ptr_ring_produce(&rq->xdp_ring, ptr)))
509 break;
510 nxmit++;
511 }
512 spin_unlock(&rq->xdp_ring.producer_lock);
513
514 if (flags & XDP_XMIT_FLUSH)
515 __veth_xdp_flush(rq);
516
517 ret = nxmit;
518 if (ndo_xmit) {
519 u64_stats_update_begin(&rq->stats.syncp);
520 rq->stats.vs.peer_tq_xdp_xmit += nxmit;
521 rq->stats.vs.peer_tq_xdp_xmit_err += n - nxmit;
522 u64_stats_update_end(&rq->stats.syncp);
523 }
524
525out:
526 rcu_read_unlock();
527
528 return ret;
529}
530
531static int veth_ndo_xdp_xmit(struct net_device *dev, int n,
532 struct xdp_frame **frames, u32 flags)
533{
534 int err;
535
536 err = veth_xdp_xmit(dev, n, frames, flags, true);
537 if (err < 0) {
538 struct veth_priv *priv = netdev_priv(dev);
539
540 atomic64_add(n, &priv->dropped);
541 }
542
543 return err;
544}
545
546static void veth_xdp_flush_bq(struct veth_rq *rq, struct veth_xdp_tx_bq *bq)
547{
548 int sent, i, err = 0, drops;
549
550 sent = veth_xdp_xmit(rq->dev, bq->count, bq->q, 0, false);
551 if (sent < 0) {
552 err = sent;
553 sent = 0;
554 }
555
556 for (i = sent; unlikely(i < bq->count); i++)
557 xdp_return_frame(bq->q[i]);
558
559 drops = bq->count - sent;
560 trace_xdp_bulk_tx(rq->dev, sent, drops, err);
561
562 u64_stats_update_begin(&rq->stats.syncp);
563 rq->stats.vs.xdp_tx += sent;
564 rq->stats.vs.xdp_tx_err += drops;
565 u64_stats_update_end(&rq->stats.syncp);
566
567 bq->count = 0;
568}
569
570static void veth_xdp_flush(struct veth_rq *rq, struct veth_xdp_tx_bq *bq)
571{
572 struct veth_priv *rcv_priv, *priv = netdev_priv(rq->dev);
573 struct net_device *rcv;
574 struct veth_rq *rcv_rq;
575
576 rcu_read_lock();
577 veth_xdp_flush_bq(rq, bq);
578 rcv = rcu_dereference(priv->peer);
579 if (unlikely(!rcv))
580 goto out;
581
582 rcv_priv = netdev_priv(rcv);
583 rcv_rq = &rcv_priv->rq[veth_select_rxq(rcv)];
584 /* xdp_ring is initialized on receive side? */
585 if (unlikely(!rcu_access_pointer(rcv_rq->xdp_prog)))
586 goto out;
587
588 __veth_xdp_flush(rcv_rq);
589out:
590 rcu_read_unlock();
591}
592
593static int veth_xdp_tx(struct veth_rq *rq, struct xdp_buff *xdp,
594 struct veth_xdp_tx_bq *bq)
595{
596 struct xdp_frame *frame = xdp_convert_buff_to_frame(xdp);
597
598 if (unlikely(!frame))
599 return -EOVERFLOW;
600
601 if (unlikely(bq->count == VETH_XDP_TX_BULK_SIZE))
602 veth_xdp_flush_bq(rq, bq);
603
604 bq->q[bq->count++] = frame;
605
606 return 0;
607}
608
609static struct xdp_frame *veth_xdp_rcv_one(struct veth_rq *rq,
610 struct xdp_frame *frame,
611 struct veth_xdp_tx_bq *bq,
612 struct veth_stats *stats)
613{
614 struct xdp_frame orig_frame;
615 struct bpf_prog *xdp_prog;
616
617 rcu_read_lock();
618 xdp_prog = rcu_dereference(rq->xdp_prog);
619 if (likely(xdp_prog)) {
620 struct veth_xdp_buff vxbuf;
621 struct xdp_buff *xdp = &vxbuf.xdp;
622 u32 act;
623
624 xdp_convert_frame_to_buff(frame, xdp);
625 xdp->rxq = &rq->xdp_rxq;
626 vxbuf.skb = NULL;
627
628 act = bpf_prog_run_xdp(xdp_prog, xdp);
629
630 switch (act) {
631 case XDP_PASS:
632 if (xdp_update_frame_from_buff(xdp, frame))
633 goto err_xdp;
634 break;
635 case XDP_TX:
636 orig_frame = *frame;
637 xdp->rxq->mem = frame->mem;
638 if (unlikely(veth_xdp_tx(rq, xdp, bq) < 0)) {
639 trace_xdp_exception(rq->dev, xdp_prog, act);
640 frame = &orig_frame;
641 stats->rx_drops++;
642 goto err_xdp;
643 }
644 stats->xdp_tx++;
645 rcu_read_unlock();
646 goto xdp_xmit;
647 case XDP_REDIRECT:
648 orig_frame = *frame;
649 xdp->rxq->mem = frame->mem;
650 if (xdp_do_redirect(rq->dev, xdp, xdp_prog)) {
651 frame = &orig_frame;
652 stats->rx_drops++;
653 goto err_xdp;
654 }
655 stats->xdp_redirect++;
656 rcu_read_unlock();
657 goto xdp_xmit;
658 default:
659 bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act);
660 fallthrough;
661 case XDP_ABORTED:
662 trace_xdp_exception(rq->dev, xdp_prog, act);
663 fallthrough;
664 case XDP_DROP:
665 stats->xdp_drops++;
666 goto err_xdp;
667 }
668 }
669 rcu_read_unlock();
670
671 return frame;
672err_xdp:
673 rcu_read_unlock();
674 xdp_return_frame(frame);
675xdp_xmit:
676 return NULL;
677}
678
679/* frames array contains VETH_XDP_BATCH at most */
680static void veth_xdp_rcv_bulk_skb(struct veth_rq *rq, void **frames,
681 int n_xdpf, struct veth_xdp_tx_bq *bq,
682 struct veth_stats *stats)
683{
684 void *skbs[VETH_XDP_BATCH];
685 int i;
686
687 if (xdp_alloc_skb_bulk(skbs, n_xdpf,
688 GFP_ATOMIC | __GFP_ZERO) < 0) {
689 for (i = 0; i < n_xdpf; i++)
690 xdp_return_frame(frames[i]);
691 stats->rx_drops += n_xdpf;
692
693 return;
694 }
695
696 for (i = 0; i < n_xdpf; i++) {
697 struct sk_buff *skb = skbs[i];
698
699 skb = __xdp_build_skb_from_frame(frames[i], skb,
700 rq->dev);
701 if (!skb) {
702 xdp_return_frame(frames[i]);
703 stats->rx_drops++;
704 continue;
705 }
706 napi_gro_receive(&rq->xdp_napi, skb);
707 }
708}
709
710static void veth_xdp_get(struct xdp_buff *xdp)
711{
712 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
713 int i;
714
715 get_page(virt_to_page(xdp->data));
716 if (likely(!xdp_buff_has_frags(xdp)))
717 return;
718
719 for (i = 0; i < sinfo->nr_frags; i++)
720 __skb_frag_ref(&sinfo->frags[i]);
721}
722
723static int veth_convert_skb_to_xdp_buff(struct veth_rq *rq,
724 struct xdp_buff *xdp,
725 struct sk_buff **pskb)
726{
727 struct sk_buff *skb = *pskb;
728 u32 frame_sz;
729
730 if (skb_shared(skb) || skb_head_is_locked(skb) ||
731 skb_shinfo(skb)->nr_frags ||
732 skb_headroom(skb) < XDP_PACKET_HEADROOM) {
733 if (skb_pp_cow_data(rq->page_pool, pskb, XDP_PACKET_HEADROOM))
734 goto drop;
735
736 skb = *pskb;
737 }
738
739 /* SKB "head" area always have tailroom for skb_shared_info */
740 frame_sz = skb_end_pointer(skb) - skb->head;
741 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
742 xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq);
743 xdp_prepare_buff(xdp, skb->head, skb_headroom(skb),
744 skb_headlen(skb), true);
745
746 if (skb_is_nonlinear(skb)) {
747 skb_shinfo(skb)->xdp_frags_size = skb->data_len;
748 xdp_buff_set_frags_flag(xdp);
749 } else {
750 xdp_buff_clear_frags_flag(xdp);
751 }
752 *pskb = skb;
753
754 return 0;
755drop:
756 consume_skb(skb);
757 *pskb = NULL;
758
759 return -ENOMEM;
760}
761
762static struct sk_buff *veth_xdp_rcv_skb(struct veth_rq *rq,
763 struct sk_buff *skb,
764 struct veth_xdp_tx_bq *bq,
765 struct veth_stats *stats)
766{
767 void *orig_data, *orig_data_end;
768 struct bpf_prog *xdp_prog;
769 struct veth_xdp_buff vxbuf;
770 struct xdp_buff *xdp = &vxbuf.xdp;
771 u32 act, metalen;
772 int off;
773
774 skb_prepare_for_gro(skb);
775
776 rcu_read_lock();
777 xdp_prog = rcu_dereference(rq->xdp_prog);
778 if (unlikely(!xdp_prog)) {
779 rcu_read_unlock();
780 goto out;
781 }
782
783 __skb_push(skb, skb->data - skb_mac_header(skb));
784 if (veth_convert_skb_to_xdp_buff(rq, xdp, &skb))
785 goto drop;
786 vxbuf.skb = skb;
787
788 orig_data = xdp->data;
789 orig_data_end = xdp->data_end;
790
791 act = bpf_prog_run_xdp(xdp_prog, xdp);
792
793 switch (act) {
794 case XDP_PASS:
795 break;
796 case XDP_TX:
797 veth_xdp_get(xdp);
798 consume_skb(skb);
799 xdp->rxq->mem = rq->xdp_mem;
800 if (unlikely(veth_xdp_tx(rq, xdp, bq) < 0)) {
801 trace_xdp_exception(rq->dev, xdp_prog, act);
802 stats->rx_drops++;
803 goto err_xdp;
804 }
805 stats->xdp_tx++;
806 rcu_read_unlock();
807 goto xdp_xmit;
808 case XDP_REDIRECT:
809 veth_xdp_get(xdp);
810 consume_skb(skb);
811 xdp->rxq->mem = rq->xdp_mem;
812 if (xdp_do_redirect(rq->dev, xdp, xdp_prog)) {
813 stats->rx_drops++;
814 goto err_xdp;
815 }
816 stats->xdp_redirect++;
817 rcu_read_unlock();
818 goto xdp_xmit;
819 default:
820 bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act);
821 fallthrough;
822 case XDP_ABORTED:
823 trace_xdp_exception(rq->dev, xdp_prog, act);
824 fallthrough;
825 case XDP_DROP:
826 stats->xdp_drops++;
827 goto xdp_drop;
828 }
829 rcu_read_unlock();
830
831 /* check if bpf_xdp_adjust_head was used */
832 off = orig_data - xdp->data;
833 if (off > 0)
834 __skb_push(skb, off);
835 else if (off < 0)
836 __skb_pull(skb, -off);
837
838 skb_reset_mac_header(skb);
839
840 /* check if bpf_xdp_adjust_tail was used */
841 off = xdp->data_end - orig_data_end;
842 if (off != 0)
843 __skb_put(skb, off); /* positive on grow, negative on shrink */
844
845 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers
846 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here.
847 */
848 if (xdp_buff_has_frags(xdp))
849 skb->data_len = skb_shinfo(skb)->xdp_frags_size;
850 else
851 skb->data_len = 0;
852
853 skb->protocol = eth_type_trans(skb, rq->dev);
854
855 metalen = xdp->data - xdp->data_meta;
856 if (metalen)
857 skb_metadata_set(skb, metalen);
858out:
859 return skb;
860drop:
861 stats->rx_drops++;
862xdp_drop:
863 rcu_read_unlock();
864 kfree_skb(skb);
865 return NULL;
866err_xdp:
867 rcu_read_unlock();
868 xdp_return_buff(xdp);
869xdp_xmit:
870 return NULL;
871}
872
873static int veth_xdp_rcv(struct veth_rq *rq, int budget,
874 struct veth_xdp_tx_bq *bq,
875 struct veth_stats *stats)
876{
877 int i, done = 0, n_xdpf = 0;
878 void *xdpf[VETH_XDP_BATCH];
879
880 for (i = 0; i < budget; i++) {
881 void *ptr = __ptr_ring_consume(&rq->xdp_ring);
882
883 if (!ptr)
884 break;
885
886 if (veth_is_xdp_frame(ptr)) {
887 /* ndo_xdp_xmit */
888 struct xdp_frame *frame = veth_ptr_to_xdp(ptr);
889
890 stats->xdp_bytes += xdp_get_frame_len(frame);
891 frame = veth_xdp_rcv_one(rq, frame, bq, stats);
892 if (frame) {
893 /* XDP_PASS */
894 xdpf[n_xdpf++] = frame;
895 if (n_xdpf == VETH_XDP_BATCH) {
896 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf,
897 bq, stats);
898 n_xdpf = 0;
899 }
900 }
901 } else {
902 /* ndo_start_xmit */
903 struct sk_buff *skb = ptr;
904
905 stats->xdp_bytes += skb->len;
906 skb = veth_xdp_rcv_skb(rq, skb, bq, stats);
907 if (skb) {
908 if (skb_shared(skb) || skb_unclone(skb, GFP_ATOMIC))
909 netif_receive_skb(skb);
910 else
911 napi_gro_receive(&rq->xdp_napi, skb);
912 }
913 }
914 done++;
915 }
916
917 if (n_xdpf)
918 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, bq, stats);
919
920 u64_stats_update_begin(&rq->stats.syncp);
921 rq->stats.vs.xdp_redirect += stats->xdp_redirect;
922 rq->stats.vs.xdp_bytes += stats->xdp_bytes;
923 rq->stats.vs.xdp_drops += stats->xdp_drops;
924 rq->stats.vs.rx_drops += stats->rx_drops;
925 rq->stats.vs.xdp_packets += done;
926 u64_stats_update_end(&rq->stats.syncp);
927
928 return done;
929}
930
931static int veth_poll(struct napi_struct *napi, int budget)
932{
933 struct veth_rq *rq =
934 container_of(napi, struct veth_rq, xdp_napi);
935 struct veth_stats stats = {};
936 struct veth_xdp_tx_bq bq;
937 int done;
938
939 bq.count = 0;
940
941 xdp_set_return_frame_no_direct();
942 done = veth_xdp_rcv(rq, budget, &bq, &stats);
943
944 if (stats.xdp_redirect > 0)
945 xdp_do_flush();
946
947 if (done < budget && napi_complete_done(napi, done)) {
948 /* Write rx_notify_masked before reading ptr_ring */
949 smp_store_mb(rq->rx_notify_masked, false);
950 if (unlikely(!__ptr_ring_empty(&rq->xdp_ring))) {
951 if (napi_schedule_prep(&rq->xdp_napi)) {
952 WRITE_ONCE(rq->rx_notify_masked, true);
953 __napi_schedule(&rq->xdp_napi);
954 }
955 }
956 }
957
958 if (stats.xdp_tx > 0)
959 veth_xdp_flush(rq, &bq);
960 xdp_clear_return_frame_no_direct();
961
962 return done;
963}
964
965static int veth_create_page_pool(struct veth_rq *rq)
966{
967 struct page_pool_params pp_params = {
968 .order = 0,
969 .pool_size = VETH_RING_SIZE,
970 .nid = NUMA_NO_NODE,
971 .dev = &rq->dev->dev,
972 };
973
974 rq->page_pool = page_pool_create(&pp_params);
975 if (IS_ERR(rq->page_pool)) {
976 int err = PTR_ERR(rq->page_pool);
977
978 rq->page_pool = NULL;
979 return err;
980 }
981
982 return 0;
983}
984
985static int __veth_napi_enable_range(struct net_device *dev, int start, int end)
986{
987 struct veth_priv *priv = netdev_priv(dev);
988 int err, i;
989
990 for (i = start; i < end; i++) {
991 err = veth_create_page_pool(&priv->rq[i]);
992 if (err)
993 goto err_page_pool;
994 }
995
996 for (i = start; i < end; i++) {
997 struct veth_rq *rq = &priv->rq[i];
998
999 err = ptr_ring_init(&rq->xdp_ring, VETH_RING_SIZE, GFP_KERNEL);
1000 if (err)
1001 goto err_xdp_ring;
1002 }
1003
1004 for (i = start; i < end; i++) {
1005 struct veth_rq *rq = &priv->rq[i];
1006
1007 napi_enable(&rq->xdp_napi);
1008 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1009 }
1010
1011 return 0;
1012
1013err_xdp_ring:
1014 for (i--; i >= start; i--)
1015 ptr_ring_cleanup(&priv->rq[i].xdp_ring, veth_ptr_free);
1016 i = end;
1017err_page_pool:
1018 for (i--; i >= start; i--) {
1019 page_pool_destroy(priv->rq[i].page_pool);
1020 priv->rq[i].page_pool = NULL;
1021 }
1022
1023 return err;
1024}
1025
1026static int __veth_napi_enable(struct net_device *dev)
1027{
1028 return __veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1029}
1030
1031static void veth_napi_del_range(struct net_device *dev, int start, int end)
1032{
1033 struct veth_priv *priv = netdev_priv(dev);
1034 int i;
1035
1036 for (i = start; i < end; i++) {
1037 struct veth_rq *rq = &priv->rq[i];
1038
1039 rcu_assign_pointer(priv->rq[i].napi, NULL);
1040 napi_disable(&rq->xdp_napi);
1041 __netif_napi_del(&rq->xdp_napi);
1042 }
1043 synchronize_net();
1044
1045 for (i = start; i < end; i++) {
1046 struct veth_rq *rq = &priv->rq[i];
1047
1048 rq->rx_notify_masked = false;
1049 ptr_ring_cleanup(&rq->xdp_ring, veth_ptr_free);
1050 }
1051
1052 for (i = start; i < end; i++) {
1053 page_pool_destroy(priv->rq[i].page_pool);
1054 priv->rq[i].page_pool = NULL;
1055 }
1056}
1057
1058static void veth_napi_del(struct net_device *dev)
1059{
1060 veth_napi_del_range(dev, 0, dev->real_num_rx_queues);
1061}
1062
1063static bool veth_gro_requested(const struct net_device *dev)
1064{
1065 return !!(dev->wanted_features & NETIF_F_GRO);
1066}
1067
1068static int veth_enable_xdp_range(struct net_device *dev, int start, int end,
1069 bool napi_already_on)
1070{
1071 struct veth_priv *priv = netdev_priv(dev);
1072 int err, i;
1073
1074 for (i = start; i < end; i++) {
1075 struct veth_rq *rq = &priv->rq[i];
1076
1077 if (!napi_already_on)
1078 netif_napi_add(dev, &rq->xdp_napi, veth_poll);
1079 err = xdp_rxq_info_reg(&rq->xdp_rxq, dev, i, rq->xdp_napi.napi_id);
1080 if (err < 0)
1081 goto err_rxq_reg;
1082
1083 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
1084 MEM_TYPE_PAGE_SHARED,
1085 NULL);
1086 if (err < 0)
1087 goto err_reg_mem;
1088
1089 /* Save original mem info as it can be overwritten */
1090 rq->xdp_mem = rq->xdp_rxq.mem;
1091 }
1092 return 0;
1093
1094err_reg_mem:
1095 xdp_rxq_info_unreg(&priv->rq[i].xdp_rxq);
1096err_rxq_reg:
1097 for (i--; i >= start; i--) {
1098 struct veth_rq *rq = &priv->rq[i];
1099
1100 xdp_rxq_info_unreg(&rq->xdp_rxq);
1101 if (!napi_already_on)
1102 netif_napi_del(&rq->xdp_napi);
1103 }
1104
1105 return err;
1106}
1107
1108static void veth_disable_xdp_range(struct net_device *dev, int start, int end,
1109 bool delete_napi)
1110{
1111 struct veth_priv *priv = netdev_priv(dev);
1112 int i;
1113
1114 for (i = start; i < end; i++) {
1115 struct veth_rq *rq = &priv->rq[i];
1116
1117 rq->xdp_rxq.mem = rq->xdp_mem;
1118 xdp_rxq_info_unreg(&rq->xdp_rxq);
1119
1120 if (delete_napi)
1121 netif_napi_del(&rq->xdp_napi);
1122 }
1123}
1124
1125static int veth_enable_xdp(struct net_device *dev)
1126{
1127 bool napi_already_on = veth_gro_requested(dev) && (dev->flags & IFF_UP);
1128 struct veth_priv *priv = netdev_priv(dev);
1129 int err, i;
1130
1131 if (!xdp_rxq_info_is_reg(&priv->rq[0].xdp_rxq)) {
1132 err = veth_enable_xdp_range(dev, 0, dev->real_num_rx_queues, napi_already_on);
1133 if (err)
1134 return err;
1135
1136 if (!napi_already_on) {
1137 err = __veth_napi_enable(dev);
1138 if (err) {
1139 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, true);
1140 return err;
1141 }
1142 }
1143 }
1144
1145 for (i = 0; i < dev->real_num_rx_queues; i++) {
1146 rcu_assign_pointer(priv->rq[i].xdp_prog, priv->_xdp_prog);
1147 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1148 }
1149
1150 return 0;
1151}
1152
1153static void veth_disable_xdp(struct net_device *dev)
1154{
1155 struct veth_priv *priv = netdev_priv(dev);
1156 int i;
1157
1158 for (i = 0; i < dev->real_num_rx_queues; i++)
1159 rcu_assign_pointer(priv->rq[i].xdp_prog, NULL);
1160
1161 if (!netif_running(dev) || !veth_gro_requested(dev))
1162 veth_napi_del(dev);
1163
1164 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, false);
1165}
1166
1167static int veth_napi_enable_range(struct net_device *dev, int start, int end)
1168{
1169 struct veth_priv *priv = netdev_priv(dev);
1170 int err, i;
1171
1172 for (i = start; i < end; i++) {
1173 struct veth_rq *rq = &priv->rq[i];
1174
1175 netif_napi_add(dev, &rq->xdp_napi, veth_poll);
1176 }
1177
1178 err = __veth_napi_enable_range(dev, start, end);
1179 if (err) {
1180 for (i = start; i < end; i++) {
1181 struct veth_rq *rq = &priv->rq[i];
1182
1183 netif_napi_del(&rq->xdp_napi);
1184 }
1185 return err;
1186 }
1187 return err;
1188}
1189
1190static int veth_napi_enable(struct net_device *dev)
1191{
1192 return veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1193}
1194
1195static void veth_disable_range_safe(struct net_device *dev, int start, int end)
1196{
1197 struct veth_priv *priv = netdev_priv(dev);
1198
1199 if (start >= end)
1200 return;
1201
1202 if (priv->_xdp_prog) {
1203 veth_napi_del_range(dev, start, end);
1204 veth_disable_xdp_range(dev, start, end, false);
1205 } else if (veth_gro_requested(dev)) {
1206 veth_napi_del_range(dev, start, end);
1207 }
1208}
1209
1210static int veth_enable_range_safe(struct net_device *dev, int start, int end)
1211{
1212 struct veth_priv *priv = netdev_priv(dev);
1213 int err;
1214
1215 if (start >= end)
1216 return 0;
1217
1218 if (priv->_xdp_prog) {
1219 /* these channels are freshly initialized, napi is not on there even
1220 * when GRO is requeste
1221 */
1222 err = veth_enable_xdp_range(dev, start, end, false);
1223 if (err)
1224 return err;
1225
1226 err = __veth_napi_enable_range(dev, start, end);
1227 if (err) {
1228 /* on error always delete the newly added napis */
1229 veth_disable_xdp_range(dev, start, end, true);
1230 return err;
1231 }
1232 } else if (veth_gro_requested(dev)) {
1233 return veth_napi_enable_range(dev, start, end);
1234 }
1235 return 0;
1236}
1237
1238static void veth_set_xdp_features(struct net_device *dev)
1239{
1240 struct veth_priv *priv = netdev_priv(dev);
1241 struct net_device *peer;
1242
1243 peer = rtnl_dereference(priv->peer);
1244 if (peer && peer->real_num_tx_queues <= dev->real_num_rx_queues) {
1245 struct veth_priv *priv_peer = netdev_priv(peer);
1246 xdp_features_t val = NETDEV_XDP_ACT_BASIC |
1247 NETDEV_XDP_ACT_REDIRECT |
1248 NETDEV_XDP_ACT_RX_SG;
1249
1250 if (priv_peer->_xdp_prog || veth_gro_requested(peer))
1251 val |= NETDEV_XDP_ACT_NDO_XMIT |
1252 NETDEV_XDP_ACT_NDO_XMIT_SG;
1253 xdp_set_features_flag(dev, val);
1254 } else {
1255 xdp_clear_features_flag(dev);
1256 }
1257}
1258
1259static int veth_set_channels(struct net_device *dev,
1260 struct ethtool_channels *ch)
1261{
1262 struct veth_priv *priv = netdev_priv(dev);
1263 unsigned int old_rx_count, new_rx_count;
1264 struct veth_priv *peer_priv;
1265 struct net_device *peer;
1266 int err;
1267
1268 /* sanity check. Upper bounds are already enforced by the caller */
1269 if (!ch->rx_count || !ch->tx_count)
1270 return -EINVAL;
1271
1272 /* avoid braking XDP, if that is enabled */
1273 peer = rtnl_dereference(priv->peer);
1274 peer_priv = peer ? netdev_priv(peer) : NULL;
1275 if (priv->_xdp_prog && peer && ch->rx_count < peer->real_num_tx_queues)
1276 return -EINVAL;
1277
1278 if (peer && peer_priv && peer_priv->_xdp_prog && ch->tx_count > peer->real_num_rx_queues)
1279 return -EINVAL;
1280
1281 old_rx_count = dev->real_num_rx_queues;
1282 new_rx_count = ch->rx_count;
1283 if (netif_running(dev)) {
1284 /* turn device off */
1285 netif_carrier_off(dev);
1286 if (peer)
1287 netif_carrier_off(peer);
1288
1289 /* try to allocate new resurces, as needed*/
1290 err = veth_enable_range_safe(dev, old_rx_count, new_rx_count);
1291 if (err)
1292 goto out;
1293 }
1294
1295 err = netif_set_real_num_rx_queues(dev, ch->rx_count);
1296 if (err)
1297 goto revert;
1298
1299 err = netif_set_real_num_tx_queues(dev, ch->tx_count);
1300 if (err) {
1301 int err2 = netif_set_real_num_rx_queues(dev, old_rx_count);
1302
1303 /* this error condition could happen only if rx and tx change
1304 * in opposite directions (e.g. tx nr raises, rx nr decreases)
1305 * and we can't do anything to fully restore the original
1306 * status
1307 */
1308 if (err2)
1309 pr_warn("Can't restore rx queues config %d -> %d %d",
1310 new_rx_count, old_rx_count, err2);
1311 else
1312 goto revert;
1313 }
1314
1315out:
1316 if (netif_running(dev)) {
1317 /* note that we need to swap the arguments WRT the enable part
1318 * to identify the range we have to disable
1319 */
1320 veth_disable_range_safe(dev, new_rx_count, old_rx_count);
1321 netif_carrier_on(dev);
1322 if (peer)
1323 netif_carrier_on(peer);
1324 }
1325
1326 /* update XDP supported features */
1327 veth_set_xdp_features(dev);
1328 if (peer)
1329 veth_set_xdp_features(peer);
1330
1331 return err;
1332
1333revert:
1334 new_rx_count = old_rx_count;
1335 old_rx_count = ch->rx_count;
1336 goto out;
1337}
1338
1339static int veth_open(struct net_device *dev)
1340{
1341 struct veth_priv *priv = netdev_priv(dev);
1342 struct net_device *peer = rtnl_dereference(priv->peer);
1343 int err;
1344
1345 if (!peer)
1346 return -ENOTCONN;
1347
1348 if (priv->_xdp_prog) {
1349 err = veth_enable_xdp(dev);
1350 if (err)
1351 return err;
1352 } else if (veth_gro_requested(dev)) {
1353 err = veth_napi_enable(dev);
1354 if (err)
1355 return err;
1356 }
1357
1358 if (peer->flags & IFF_UP) {
1359 netif_carrier_on(dev);
1360 netif_carrier_on(peer);
1361 }
1362
1363 veth_set_xdp_features(dev);
1364
1365 return 0;
1366}
1367
1368static int veth_close(struct net_device *dev)
1369{
1370 struct veth_priv *priv = netdev_priv(dev);
1371 struct net_device *peer = rtnl_dereference(priv->peer);
1372
1373 netif_carrier_off(dev);
1374 if (peer)
1375 netif_carrier_off(peer);
1376
1377 if (priv->_xdp_prog)
1378 veth_disable_xdp(dev);
1379 else if (veth_gro_requested(dev))
1380 veth_napi_del(dev);
1381
1382 return 0;
1383}
1384
1385static int is_valid_veth_mtu(int mtu)
1386{
1387 return mtu >= ETH_MIN_MTU && mtu <= ETH_MAX_MTU;
1388}
1389
1390static int veth_alloc_queues(struct net_device *dev)
1391{
1392 struct veth_priv *priv = netdev_priv(dev);
1393 int i;
1394
1395 priv->rq = kvcalloc(dev->num_rx_queues, sizeof(*priv->rq),
1396 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
1397 if (!priv->rq)
1398 return -ENOMEM;
1399
1400 for (i = 0; i < dev->num_rx_queues; i++) {
1401 priv->rq[i].dev = dev;
1402 u64_stats_init(&priv->rq[i].stats.syncp);
1403 }
1404
1405 return 0;
1406}
1407
1408static void veth_free_queues(struct net_device *dev)
1409{
1410 struct veth_priv *priv = netdev_priv(dev);
1411
1412 kvfree(priv->rq);
1413}
1414
1415static int veth_dev_init(struct net_device *dev)
1416{
1417 netdev_lockdep_set_classes(dev);
1418 return veth_alloc_queues(dev);
1419}
1420
1421static void veth_dev_free(struct net_device *dev)
1422{
1423 veth_free_queues(dev);
1424}
1425
1426#ifdef CONFIG_NET_POLL_CONTROLLER
1427static void veth_poll_controller(struct net_device *dev)
1428{
1429 /* veth only receives frames when its peer sends one
1430 * Since it has nothing to do with disabling irqs, we are guaranteed
1431 * never to have pending data when we poll for it so
1432 * there is nothing to do here.
1433 *
1434 * We need this though so netpoll recognizes us as an interface that
1435 * supports polling, which enables bridge devices in virt setups to
1436 * still use netconsole
1437 */
1438}
1439#endif /* CONFIG_NET_POLL_CONTROLLER */
1440
1441static int veth_get_iflink(const struct net_device *dev)
1442{
1443 struct veth_priv *priv = netdev_priv(dev);
1444 struct net_device *peer;
1445 int iflink;
1446
1447 rcu_read_lock();
1448 peer = rcu_dereference(priv->peer);
1449 iflink = peer ? READ_ONCE(peer->ifindex) : 0;
1450 rcu_read_unlock();
1451
1452 return iflink;
1453}
1454
1455static netdev_features_t veth_fix_features(struct net_device *dev,
1456 netdev_features_t features)
1457{
1458 struct veth_priv *priv = netdev_priv(dev);
1459 struct net_device *peer;
1460
1461 peer = rtnl_dereference(priv->peer);
1462 if (peer) {
1463 struct veth_priv *peer_priv = netdev_priv(peer);
1464
1465 if (peer_priv->_xdp_prog)
1466 features &= ~NETIF_F_GSO_SOFTWARE;
1467 }
1468
1469 return features;
1470}
1471
1472static int veth_set_features(struct net_device *dev,
1473 netdev_features_t features)
1474{
1475 netdev_features_t changed = features ^ dev->features;
1476 struct veth_priv *priv = netdev_priv(dev);
1477 struct net_device *peer;
1478 int err;
1479
1480 if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog)
1481 return 0;
1482
1483 peer = rtnl_dereference(priv->peer);
1484 if (features & NETIF_F_GRO) {
1485 err = veth_napi_enable(dev);
1486 if (err)
1487 return err;
1488
1489 if (peer)
1490 xdp_features_set_redirect_target(peer, true);
1491 } else {
1492 if (peer)
1493 xdp_features_clear_redirect_target(peer);
1494 veth_napi_del(dev);
1495 }
1496 return 0;
1497}
1498
1499static void veth_set_rx_headroom(struct net_device *dev, int new_hr)
1500{
1501 struct veth_priv *peer_priv, *priv = netdev_priv(dev);
1502 struct net_device *peer;
1503
1504 if (new_hr < 0)
1505 new_hr = 0;
1506
1507 rcu_read_lock();
1508 peer = rcu_dereference(priv->peer);
1509 if (unlikely(!peer))
1510 goto out;
1511
1512 peer_priv = netdev_priv(peer);
1513 priv->requested_headroom = new_hr;
1514 new_hr = max(priv->requested_headroom, peer_priv->requested_headroom);
1515 dev->needed_headroom = new_hr;
1516 peer->needed_headroom = new_hr;
1517
1518out:
1519 rcu_read_unlock();
1520}
1521
1522static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1523 struct netlink_ext_ack *extack)
1524{
1525 struct veth_priv *priv = netdev_priv(dev);
1526 struct bpf_prog *old_prog;
1527 struct net_device *peer;
1528 unsigned int max_mtu;
1529 int err;
1530
1531 old_prog = priv->_xdp_prog;
1532 priv->_xdp_prog = prog;
1533 peer = rtnl_dereference(priv->peer);
1534
1535 if (prog) {
1536 if (!peer) {
1537 NL_SET_ERR_MSG_MOD(extack, "Cannot set XDP when peer is detached");
1538 err = -ENOTCONN;
1539 goto err;
1540 }
1541
1542 max_mtu = SKB_WITH_OVERHEAD(PAGE_SIZE - VETH_XDP_HEADROOM) -
1543 peer->hard_header_len;
1544 /* Allow increasing the max_mtu if the program supports
1545 * XDP fragments.
1546 */
1547 if (prog->aux->xdp_has_frags)
1548 max_mtu += PAGE_SIZE * MAX_SKB_FRAGS;
1549
1550 if (peer->mtu > max_mtu) {
1551 NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP");
1552 err = -ERANGE;
1553 goto err;
1554 }
1555
1556 if (dev->real_num_rx_queues < peer->real_num_tx_queues) {
1557 NL_SET_ERR_MSG_MOD(extack, "XDP expects number of rx queues not less than peer tx queues");
1558 err = -ENOSPC;
1559 goto err;
1560 }
1561
1562 if (dev->flags & IFF_UP) {
1563 err = veth_enable_xdp(dev);
1564 if (err) {
1565 NL_SET_ERR_MSG_MOD(extack, "Setup for XDP failed");
1566 goto err;
1567 }
1568 }
1569
1570 if (!old_prog) {
1571 peer->hw_features &= ~NETIF_F_GSO_SOFTWARE;
1572 peer->max_mtu = max_mtu;
1573 }
1574
1575 xdp_features_set_redirect_target(peer, true);
1576 }
1577
1578 if (old_prog) {
1579 if (!prog) {
1580 if (peer && !veth_gro_requested(dev))
1581 xdp_features_clear_redirect_target(peer);
1582
1583 if (dev->flags & IFF_UP)
1584 veth_disable_xdp(dev);
1585
1586 if (peer) {
1587 peer->hw_features |= NETIF_F_GSO_SOFTWARE;
1588 peer->max_mtu = ETH_MAX_MTU;
1589 }
1590 }
1591 bpf_prog_put(old_prog);
1592 }
1593
1594 if ((!!old_prog ^ !!prog) && peer)
1595 netdev_update_features(peer);
1596
1597 return 0;
1598err:
1599 priv->_xdp_prog = old_prog;
1600
1601 return err;
1602}
1603
1604static int veth_xdp(struct net_device *dev, struct netdev_bpf *xdp)
1605{
1606 switch (xdp->command) {
1607 case XDP_SETUP_PROG:
1608 return veth_xdp_set(dev, xdp->prog, xdp->extack);
1609 default:
1610 return -EINVAL;
1611 }
1612}
1613
1614static int veth_xdp_rx_timestamp(const struct xdp_md *ctx, u64 *timestamp)
1615{
1616 struct veth_xdp_buff *_ctx = (void *)ctx;
1617
1618 if (!_ctx->skb)
1619 return -ENODATA;
1620
1621 *timestamp = skb_hwtstamps(_ctx->skb)->hwtstamp;
1622 return 0;
1623}
1624
1625static int veth_xdp_rx_hash(const struct xdp_md *ctx, u32 *hash,
1626 enum xdp_rss_hash_type *rss_type)
1627{
1628 struct veth_xdp_buff *_ctx = (void *)ctx;
1629 struct sk_buff *skb = _ctx->skb;
1630
1631 if (!skb)
1632 return -ENODATA;
1633
1634 *hash = skb_get_hash(skb);
1635 *rss_type = skb->l4_hash ? XDP_RSS_TYPE_L4_ANY : XDP_RSS_TYPE_NONE;
1636
1637 return 0;
1638}
1639
1640static int veth_xdp_rx_vlan_tag(const struct xdp_md *ctx, __be16 *vlan_proto,
1641 u16 *vlan_tci)
1642{
1643 const struct veth_xdp_buff *_ctx = (void *)ctx;
1644 const struct sk_buff *skb = _ctx->skb;
1645 int err;
1646
1647 if (!skb)
1648 return -ENODATA;
1649
1650 err = __vlan_hwaccel_get_tag(skb, vlan_tci);
1651 if (err)
1652 return err;
1653
1654 *vlan_proto = skb->vlan_proto;
1655 return err;
1656}
1657
1658static const struct net_device_ops veth_netdev_ops = {
1659 .ndo_init = veth_dev_init,
1660 .ndo_open = veth_open,
1661 .ndo_stop = veth_close,
1662 .ndo_start_xmit = veth_xmit,
1663 .ndo_get_stats64 = veth_get_stats64,
1664 .ndo_set_rx_mode = veth_set_multicast_list,
1665 .ndo_set_mac_address = eth_mac_addr,
1666#ifdef CONFIG_NET_POLL_CONTROLLER
1667 .ndo_poll_controller = veth_poll_controller,
1668#endif
1669 .ndo_get_iflink = veth_get_iflink,
1670 .ndo_fix_features = veth_fix_features,
1671 .ndo_set_features = veth_set_features,
1672 .ndo_features_check = passthru_features_check,
1673 .ndo_set_rx_headroom = veth_set_rx_headroom,
1674 .ndo_bpf = veth_xdp,
1675 .ndo_xdp_xmit = veth_ndo_xdp_xmit,
1676 .ndo_get_peer_dev = veth_peer_dev,
1677};
1678
1679static const struct xdp_metadata_ops veth_xdp_metadata_ops = {
1680 .xmo_rx_timestamp = veth_xdp_rx_timestamp,
1681 .xmo_rx_hash = veth_xdp_rx_hash,
1682 .xmo_rx_vlan_tag = veth_xdp_rx_vlan_tag,
1683};
1684
1685#define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HW_CSUM | \
1686 NETIF_F_RXCSUM | NETIF_F_SCTP_CRC | NETIF_F_HIGHDMA | \
1687 NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \
1688 NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX | \
1689 NETIF_F_HW_VLAN_STAG_TX | NETIF_F_HW_VLAN_STAG_RX )
1690
1691static void veth_setup(struct net_device *dev)
1692{
1693 ether_setup(dev);
1694
1695 dev->priv_flags &= ~IFF_TX_SKB_SHARING;
1696 dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
1697 dev->priv_flags |= IFF_NO_QUEUE;
1698 dev->priv_flags |= IFF_PHONY_HEADROOM;
1699 dev->priv_flags |= IFF_DISABLE_NETPOLL;
1700 dev->lltx = true;
1701
1702 dev->netdev_ops = &veth_netdev_ops;
1703 dev->xdp_metadata_ops = &veth_xdp_metadata_ops;
1704 dev->ethtool_ops = &veth_ethtool_ops;
1705 dev->features |= VETH_FEATURES;
1706 dev->vlan_features = dev->features &
1707 ~(NETIF_F_HW_VLAN_CTAG_TX |
1708 NETIF_F_HW_VLAN_STAG_TX |
1709 NETIF_F_HW_VLAN_CTAG_RX |
1710 NETIF_F_HW_VLAN_STAG_RX);
1711 dev->needs_free_netdev = true;
1712 dev->priv_destructor = veth_dev_free;
1713 dev->pcpu_stat_type = NETDEV_PCPU_STAT_TSTATS;
1714 dev->max_mtu = ETH_MAX_MTU;
1715
1716 dev->hw_features = VETH_FEATURES;
1717 dev->hw_enc_features = VETH_FEATURES;
1718 dev->mpls_features = NETIF_F_HW_CSUM | NETIF_F_GSO_SOFTWARE;
1719 netif_set_tso_max_size(dev, GSO_MAX_SIZE);
1720}
1721
1722/*
1723 * netlink interface
1724 */
1725
1726static int veth_validate(struct nlattr *tb[], struct nlattr *data[],
1727 struct netlink_ext_ack *extack)
1728{
1729 if (tb[IFLA_ADDRESS]) {
1730 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
1731 return -EINVAL;
1732 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
1733 return -EADDRNOTAVAIL;
1734 }
1735 if (tb[IFLA_MTU]) {
1736 if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU])))
1737 return -EINVAL;
1738 }
1739 return 0;
1740}
1741
1742static struct rtnl_link_ops veth_link_ops;
1743
1744static void veth_disable_gro(struct net_device *dev)
1745{
1746 dev->features &= ~NETIF_F_GRO;
1747 dev->wanted_features &= ~NETIF_F_GRO;
1748 netdev_update_features(dev);
1749}
1750
1751static int veth_init_queues(struct net_device *dev, struct nlattr *tb[])
1752{
1753 int err;
1754
1755 if (!tb[IFLA_NUM_TX_QUEUES] && dev->num_tx_queues > 1) {
1756 err = netif_set_real_num_tx_queues(dev, 1);
1757 if (err)
1758 return err;
1759 }
1760 if (!tb[IFLA_NUM_RX_QUEUES] && dev->num_rx_queues > 1) {
1761 err = netif_set_real_num_rx_queues(dev, 1);
1762 if (err)
1763 return err;
1764 }
1765 return 0;
1766}
1767
1768static int veth_newlink(struct net *peer_net, struct net_device *dev,
1769 struct nlattr *tb[], struct nlattr *data[],
1770 struct netlink_ext_ack *extack)
1771{
1772 int err;
1773 struct net_device *peer;
1774 struct veth_priv *priv;
1775 char ifname[IFNAMSIZ];
1776 struct nlattr *peer_tb[IFLA_MAX + 1], **tbp;
1777 unsigned char name_assign_type;
1778 struct ifinfomsg *ifmp;
1779
1780 /*
1781 * create and register peer first
1782 */
1783 if (data && data[VETH_INFO_PEER]) {
1784 struct nlattr *nla_peer = data[VETH_INFO_PEER];
1785
1786 ifmp = nla_data(nla_peer);
1787 rtnl_nla_parse_ifinfomsg(peer_tb, nla_peer, extack);
1788 tbp = peer_tb;
1789 } else {
1790 ifmp = NULL;
1791 tbp = tb;
1792 }
1793
1794 if (ifmp && tbp[IFLA_IFNAME]) {
1795 nla_strscpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ);
1796 name_assign_type = NET_NAME_USER;
1797 } else {
1798 snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d");
1799 name_assign_type = NET_NAME_ENUM;
1800 }
1801
1802 peer = rtnl_create_link(peer_net, ifname, name_assign_type,
1803 &veth_link_ops, tbp, extack);
1804 if (IS_ERR(peer))
1805 return PTR_ERR(peer);
1806
1807 if (!ifmp || !tbp[IFLA_ADDRESS])
1808 eth_hw_addr_random(peer);
1809
1810 if (ifmp && (dev->ifindex != 0))
1811 peer->ifindex = ifmp->ifi_index;
1812
1813 netif_inherit_tso_max(peer, dev);
1814
1815 err = register_netdevice(peer);
1816 if (err < 0)
1817 goto err_register_peer;
1818
1819 /* keep GRO disabled by default to be consistent with the established
1820 * veth behavior
1821 */
1822 veth_disable_gro(peer);
1823 netif_carrier_off(peer);
1824
1825 err = rtnl_configure_link(peer, ifmp, 0, NULL);
1826 if (err < 0)
1827 goto err_configure_peer;
1828
1829 /*
1830 * register dev last
1831 *
1832 * note, that since we've registered new device the dev's name
1833 * should be re-allocated
1834 */
1835
1836 if (tb[IFLA_ADDRESS] == NULL)
1837 eth_hw_addr_random(dev);
1838
1839 if (tb[IFLA_IFNAME])
1840 nla_strscpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ);
1841 else
1842 snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d");
1843
1844 err = register_netdevice(dev);
1845 if (err < 0)
1846 goto err_register_dev;
1847
1848 netif_carrier_off(dev);
1849
1850 /*
1851 * tie the deviced together
1852 */
1853
1854 priv = netdev_priv(dev);
1855 rcu_assign_pointer(priv->peer, peer);
1856 err = veth_init_queues(dev, tb);
1857 if (err)
1858 goto err_queues;
1859
1860 priv = netdev_priv(peer);
1861 rcu_assign_pointer(priv->peer, dev);
1862 err = veth_init_queues(peer, tb);
1863 if (err)
1864 goto err_queues;
1865
1866 veth_disable_gro(dev);
1867 /* update XDP supported features */
1868 veth_set_xdp_features(dev);
1869 veth_set_xdp_features(peer);
1870
1871 return 0;
1872
1873err_queues:
1874 unregister_netdevice(dev);
1875err_register_dev:
1876 /* nothing to do */
1877err_configure_peer:
1878 unregister_netdevice(peer);
1879 return err;
1880
1881err_register_peer:
1882 free_netdev(peer);
1883 return err;
1884}
1885
1886static void veth_dellink(struct net_device *dev, struct list_head *head)
1887{
1888 struct veth_priv *priv;
1889 struct net_device *peer;
1890
1891 priv = netdev_priv(dev);
1892 peer = rtnl_dereference(priv->peer);
1893
1894 /* Note : dellink() is called from default_device_exit_batch(),
1895 * before a rcu_synchronize() point. The devices are guaranteed
1896 * not being freed before one RCU grace period.
1897 */
1898 RCU_INIT_POINTER(priv->peer, NULL);
1899 unregister_netdevice_queue(dev, head);
1900
1901 if (peer) {
1902 priv = netdev_priv(peer);
1903 RCU_INIT_POINTER(priv->peer, NULL);
1904 unregister_netdevice_queue(peer, head);
1905 }
1906}
1907
1908static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = {
1909 [VETH_INFO_PEER] = { .len = sizeof(struct ifinfomsg) },
1910};
1911
1912static struct net *veth_get_link_net(const struct net_device *dev)
1913{
1914 struct veth_priv *priv = netdev_priv(dev);
1915 struct net_device *peer = rtnl_dereference(priv->peer);
1916
1917 return peer ? dev_net(peer) : dev_net(dev);
1918}
1919
1920static unsigned int veth_get_num_queues(void)
1921{
1922 /* enforce the same queue limit as rtnl_create_link */
1923 int queues = num_possible_cpus();
1924
1925 if (queues > 4096)
1926 queues = 4096;
1927 return queues;
1928}
1929
1930static struct rtnl_link_ops veth_link_ops = {
1931 .kind = DRV_NAME,
1932 .priv_size = sizeof(struct veth_priv),
1933 .setup = veth_setup,
1934 .validate = veth_validate,
1935 .newlink = veth_newlink,
1936 .dellink = veth_dellink,
1937 .policy = veth_policy,
1938 .peer_type = VETH_INFO_PEER,
1939 .maxtype = VETH_INFO_MAX,
1940 .get_link_net = veth_get_link_net,
1941 .get_num_tx_queues = veth_get_num_queues,
1942 .get_num_rx_queues = veth_get_num_queues,
1943};
1944
1945/*
1946 * init/fini
1947 */
1948
1949static __init int veth_init(void)
1950{
1951 return rtnl_link_register(&veth_link_ops);
1952}
1953
1954static __exit void veth_exit(void)
1955{
1956 rtnl_link_unregister(&veth_link_ops);
1957}
1958
1959module_init(veth_init);
1960module_exit(veth_exit);
1961
1962MODULE_DESCRIPTION("Virtual Ethernet Tunnel");
1963MODULE_LICENSE("GPL v2");
1964MODULE_ALIAS_RTNL_LINK(DRV_NAME);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * drivers/net/veth.c
4 *
5 * Copyright (C) 2007 OpenVZ http://openvz.org, SWsoft Inc
6 *
7 * Author: Pavel Emelianov <xemul@openvz.org>
8 * Ethtool interface from: Eric W. Biederman <ebiederm@xmission.com>
9 *
10 */
11
12#include <linux/netdevice.h>
13#include <linux/slab.h>
14#include <linux/ethtool.h>
15#include <linux/etherdevice.h>
16#include <linux/u64_stats_sync.h>
17
18#include <net/rtnetlink.h>
19#include <net/dst.h>
20#include <net/xfrm.h>
21#include <net/xdp.h>
22#include <linux/veth.h>
23#include <linux/module.h>
24#include <linux/bpf.h>
25#include <linux/filter.h>
26#include <linux/ptr_ring.h>
27#include <linux/bpf_trace.h>
28#include <linux/net_tstamp.h>
29
30#define DRV_NAME "veth"
31#define DRV_VERSION "1.0"
32
33#define VETH_XDP_FLAG BIT(0)
34#define VETH_RING_SIZE 256
35#define VETH_XDP_HEADROOM (XDP_PACKET_HEADROOM + NET_IP_ALIGN)
36
37#define VETH_XDP_TX_BULK_SIZE 16
38#define VETH_XDP_BATCH 16
39
40struct veth_stats {
41 u64 rx_drops;
42 /* xdp */
43 u64 xdp_packets;
44 u64 xdp_bytes;
45 u64 xdp_redirect;
46 u64 xdp_drops;
47 u64 xdp_tx;
48 u64 xdp_tx_err;
49 u64 peer_tq_xdp_xmit;
50 u64 peer_tq_xdp_xmit_err;
51};
52
53struct veth_rq_stats {
54 struct veth_stats vs;
55 struct u64_stats_sync syncp;
56};
57
58struct veth_rq {
59 struct napi_struct xdp_napi;
60 struct napi_struct __rcu *napi; /* points to xdp_napi when the latter is initialized */
61 struct net_device *dev;
62 struct bpf_prog __rcu *xdp_prog;
63 struct xdp_mem_info xdp_mem;
64 struct veth_rq_stats stats;
65 bool rx_notify_masked;
66 struct ptr_ring xdp_ring;
67 struct xdp_rxq_info xdp_rxq;
68};
69
70struct veth_priv {
71 struct net_device __rcu *peer;
72 atomic64_t dropped;
73 struct bpf_prog *_xdp_prog;
74 struct veth_rq *rq;
75 unsigned int requested_headroom;
76};
77
78struct veth_xdp_tx_bq {
79 struct xdp_frame *q[VETH_XDP_TX_BULK_SIZE];
80 unsigned int count;
81};
82
83/*
84 * ethtool interface
85 */
86
87struct veth_q_stat_desc {
88 char desc[ETH_GSTRING_LEN];
89 size_t offset;
90};
91
92#define VETH_RQ_STAT(m) offsetof(struct veth_stats, m)
93
94static const struct veth_q_stat_desc veth_rq_stats_desc[] = {
95 { "xdp_packets", VETH_RQ_STAT(xdp_packets) },
96 { "xdp_bytes", VETH_RQ_STAT(xdp_bytes) },
97 { "drops", VETH_RQ_STAT(rx_drops) },
98 { "xdp_redirect", VETH_RQ_STAT(xdp_redirect) },
99 { "xdp_drops", VETH_RQ_STAT(xdp_drops) },
100 { "xdp_tx", VETH_RQ_STAT(xdp_tx) },
101 { "xdp_tx_errors", VETH_RQ_STAT(xdp_tx_err) },
102};
103
104#define VETH_RQ_STATS_LEN ARRAY_SIZE(veth_rq_stats_desc)
105
106static const struct veth_q_stat_desc veth_tq_stats_desc[] = {
107 { "xdp_xmit", VETH_RQ_STAT(peer_tq_xdp_xmit) },
108 { "xdp_xmit_errors", VETH_RQ_STAT(peer_tq_xdp_xmit_err) },
109};
110
111#define VETH_TQ_STATS_LEN ARRAY_SIZE(veth_tq_stats_desc)
112
113static struct {
114 const char string[ETH_GSTRING_LEN];
115} ethtool_stats_keys[] = {
116 { "peer_ifindex" },
117};
118
119static int veth_get_link_ksettings(struct net_device *dev,
120 struct ethtool_link_ksettings *cmd)
121{
122 cmd->base.speed = SPEED_10000;
123 cmd->base.duplex = DUPLEX_FULL;
124 cmd->base.port = PORT_TP;
125 cmd->base.autoneg = AUTONEG_DISABLE;
126 return 0;
127}
128
129static void veth_get_drvinfo(struct net_device *dev, struct ethtool_drvinfo *info)
130{
131 strscpy(info->driver, DRV_NAME, sizeof(info->driver));
132 strscpy(info->version, DRV_VERSION, sizeof(info->version));
133}
134
135static void veth_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
136{
137 u8 *p = buf;
138 int i, j;
139
140 switch(stringset) {
141 case ETH_SS_STATS:
142 memcpy(p, ðtool_stats_keys, sizeof(ethtool_stats_keys));
143 p += sizeof(ethtool_stats_keys);
144 for (i = 0; i < dev->real_num_rx_queues; i++)
145 for (j = 0; j < VETH_RQ_STATS_LEN; j++)
146 ethtool_sprintf(&p, "rx_queue_%u_%.18s",
147 i, veth_rq_stats_desc[j].desc);
148
149 for (i = 0; i < dev->real_num_tx_queues; i++)
150 for (j = 0; j < VETH_TQ_STATS_LEN; j++)
151 ethtool_sprintf(&p, "tx_queue_%u_%.18s",
152 i, veth_tq_stats_desc[j].desc);
153 break;
154 }
155}
156
157static int veth_get_sset_count(struct net_device *dev, int sset)
158{
159 switch (sset) {
160 case ETH_SS_STATS:
161 return ARRAY_SIZE(ethtool_stats_keys) +
162 VETH_RQ_STATS_LEN * dev->real_num_rx_queues +
163 VETH_TQ_STATS_LEN * dev->real_num_tx_queues;
164 default:
165 return -EOPNOTSUPP;
166 }
167}
168
169static void veth_get_ethtool_stats(struct net_device *dev,
170 struct ethtool_stats *stats, u64 *data)
171{
172 struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
173 struct net_device *peer = rtnl_dereference(priv->peer);
174 int i, j, idx;
175
176 data[0] = peer ? peer->ifindex : 0;
177 idx = 1;
178 for (i = 0; i < dev->real_num_rx_queues; i++) {
179 const struct veth_rq_stats *rq_stats = &priv->rq[i].stats;
180 const void *stats_base = (void *)&rq_stats->vs;
181 unsigned int start;
182 size_t offset;
183
184 do {
185 start = u64_stats_fetch_begin(&rq_stats->syncp);
186 for (j = 0; j < VETH_RQ_STATS_LEN; j++) {
187 offset = veth_rq_stats_desc[j].offset;
188 data[idx + j] = *(u64 *)(stats_base + offset);
189 }
190 } while (u64_stats_fetch_retry(&rq_stats->syncp, start));
191 idx += VETH_RQ_STATS_LEN;
192 }
193
194 if (!peer)
195 return;
196
197 rcv_priv = netdev_priv(peer);
198 for (i = 0; i < peer->real_num_rx_queues; i++) {
199 const struct veth_rq_stats *rq_stats = &rcv_priv->rq[i].stats;
200 const void *base = (void *)&rq_stats->vs;
201 unsigned int start, tx_idx = idx;
202 size_t offset;
203
204 tx_idx += (i % dev->real_num_tx_queues) * VETH_TQ_STATS_LEN;
205 do {
206 start = u64_stats_fetch_begin(&rq_stats->syncp);
207 for (j = 0; j < VETH_TQ_STATS_LEN; j++) {
208 offset = veth_tq_stats_desc[j].offset;
209 data[tx_idx + j] += *(u64 *)(base + offset);
210 }
211 } while (u64_stats_fetch_retry(&rq_stats->syncp, start));
212 }
213}
214
215static void veth_get_channels(struct net_device *dev,
216 struct ethtool_channels *channels)
217{
218 channels->tx_count = dev->real_num_tx_queues;
219 channels->rx_count = dev->real_num_rx_queues;
220 channels->max_tx = dev->num_tx_queues;
221 channels->max_rx = dev->num_rx_queues;
222}
223
224static int veth_set_channels(struct net_device *dev,
225 struct ethtool_channels *ch);
226
227static const struct ethtool_ops veth_ethtool_ops = {
228 .get_drvinfo = veth_get_drvinfo,
229 .get_link = ethtool_op_get_link,
230 .get_strings = veth_get_strings,
231 .get_sset_count = veth_get_sset_count,
232 .get_ethtool_stats = veth_get_ethtool_stats,
233 .get_link_ksettings = veth_get_link_ksettings,
234 .get_ts_info = ethtool_op_get_ts_info,
235 .get_channels = veth_get_channels,
236 .set_channels = veth_set_channels,
237};
238
239/* general routines */
240
241static bool veth_is_xdp_frame(void *ptr)
242{
243 return (unsigned long)ptr & VETH_XDP_FLAG;
244}
245
246static struct xdp_frame *veth_ptr_to_xdp(void *ptr)
247{
248 return (void *)((unsigned long)ptr & ~VETH_XDP_FLAG);
249}
250
251static void *veth_xdp_to_ptr(struct xdp_frame *xdp)
252{
253 return (void *)((unsigned long)xdp | VETH_XDP_FLAG);
254}
255
256static void veth_ptr_free(void *ptr)
257{
258 if (veth_is_xdp_frame(ptr))
259 xdp_return_frame(veth_ptr_to_xdp(ptr));
260 else
261 kfree_skb(ptr);
262}
263
264static void __veth_xdp_flush(struct veth_rq *rq)
265{
266 /* Write ptr_ring before reading rx_notify_masked */
267 smp_mb();
268 if (!READ_ONCE(rq->rx_notify_masked) &&
269 napi_schedule_prep(&rq->xdp_napi)) {
270 WRITE_ONCE(rq->rx_notify_masked, true);
271 __napi_schedule(&rq->xdp_napi);
272 }
273}
274
275static int veth_xdp_rx(struct veth_rq *rq, struct sk_buff *skb)
276{
277 if (unlikely(ptr_ring_produce(&rq->xdp_ring, skb))) {
278 dev_kfree_skb_any(skb);
279 return NET_RX_DROP;
280 }
281
282 return NET_RX_SUCCESS;
283}
284
285static int veth_forward_skb(struct net_device *dev, struct sk_buff *skb,
286 struct veth_rq *rq, bool xdp)
287{
288 return __dev_forward_skb(dev, skb) ?: xdp ?
289 veth_xdp_rx(rq, skb) :
290 __netif_rx(skb);
291}
292
293/* return true if the specified skb has chances of GRO aggregation
294 * Don't strive for accuracy, but try to avoid GRO overhead in the most
295 * common scenarios.
296 * When XDP is enabled, all traffic is considered eligible, as the xmit
297 * device has TSO off.
298 * When TSO is enabled on the xmit device, we are likely interested only
299 * in UDP aggregation, explicitly check for that if the skb is suspected
300 * - the sock_wfree destructor is used by UDP, ICMP and XDP sockets -
301 * to belong to locally generated UDP traffic.
302 */
303static bool veth_skb_is_eligible_for_gro(const struct net_device *dev,
304 const struct net_device *rcv,
305 const struct sk_buff *skb)
306{
307 return !(dev->features & NETIF_F_ALL_TSO) ||
308 (skb->destructor == sock_wfree &&
309 rcv->features & (NETIF_F_GRO_FRAGLIST | NETIF_F_GRO_UDP_FWD));
310}
311
312static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev)
313{
314 struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
315 struct veth_rq *rq = NULL;
316 struct net_device *rcv;
317 int length = skb->len;
318 bool use_napi = false;
319 int rxq;
320
321 rcu_read_lock();
322 rcv = rcu_dereference(priv->peer);
323 if (unlikely(!rcv) || !pskb_may_pull(skb, ETH_HLEN)) {
324 kfree_skb(skb);
325 goto drop;
326 }
327
328 rcv_priv = netdev_priv(rcv);
329 rxq = skb_get_queue_mapping(skb);
330 if (rxq < rcv->real_num_rx_queues) {
331 rq = &rcv_priv->rq[rxq];
332
333 /* The napi pointer is available when an XDP program is
334 * attached or when GRO is enabled
335 * Don't bother with napi/GRO if the skb can't be aggregated
336 */
337 use_napi = rcu_access_pointer(rq->napi) &&
338 veth_skb_is_eligible_for_gro(dev, rcv, skb);
339 }
340
341 skb_tx_timestamp(skb);
342 if (likely(veth_forward_skb(rcv, skb, rq, use_napi) == NET_RX_SUCCESS)) {
343 if (!use_napi)
344 dev_lstats_add(dev, length);
345 } else {
346drop:
347 atomic64_inc(&priv->dropped);
348 }
349
350 if (use_napi)
351 __veth_xdp_flush(rq);
352
353 rcu_read_unlock();
354
355 return NETDEV_TX_OK;
356}
357
358static u64 veth_stats_tx(struct net_device *dev, u64 *packets, u64 *bytes)
359{
360 struct veth_priv *priv = netdev_priv(dev);
361
362 dev_lstats_read(dev, packets, bytes);
363 return atomic64_read(&priv->dropped);
364}
365
366static void veth_stats_rx(struct veth_stats *result, struct net_device *dev)
367{
368 struct veth_priv *priv = netdev_priv(dev);
369 int i;
370
371 result->peer_tq_xdp_xmit_err = 0;
372 result->xdp_packets = 0;
373 result->xdp_tx_err = 0;
374 result->xdp_bytes = 0;
375 result->rx_drops = 0;
376 for (i = 0; i < dev->num_rx_queues; i++) {
377 u64 packets, bytes, drops, xdp_tx_err, peer_tq_xdp_xmit_err;
378 struct veth_rq_stats *stats = &priv->rq[i].stats;
379 unsigned int start;
380
381 do {
382 start = u64_stats_fetch_begin(&stats->syncp);
383 peer_tq_xdp_xmit_err = stats->vs.peer_tq_xdp_xmit_err;
384 xdp_tx_err = stats->vs.xdp_tx_err;
385 packets = stats->vs.xdp_packets;
386 bytes = stats->vs.xdp_bytes;
387 drops = stats->vs.rx_drops;
388 } while (u64_stats_fetch_retry(&stats->syncp, start));
389 result->peer_tq_xdp_xmit_err += peer_tq_xdp_xmit_err;
390 result->xdp_tx_err += xdp_tx_err;
391 result->xdp_packets += packets;
392 result->xdp_bytes += bytes;
393 result->rx_drops += drops;
394 }
395}
396
397static void veth_get_stats64(struct net_device *dev,
398 struct rtnl_link_stats64 *tot)
399{
400 struct veth_priv *priv = netdev_priv(dev);
401 struct net_device *peer;
402 struct veth_stats rx;
403 u64 packets, bytes;
404
405 tot->tx_dropped = veth_stats_tx(dev, &packets, &bytes);
406 tot->tx_bytes = bytes;
407 tot->tx_packets = packets;
408
409 veth_stats_rx(&rx, dev);
410 tot->tx_dropped += rx.xdp_tx_err;
411 tot->rx_dropped = rx.rx_drops + rx.peer_tq_xdp_xmit_err;
412 tot->rx_bytes = rx.xdp_bytes;
413 tot->rx_packets = rx.xdp_packets;
414
415 rcu_read_lock();
416 peer = rcu_dereference(priv->peer);
417 if (peer) {
418 veth_stats_tx(peer, &packets, &bytes);
419 tot->rx_bytes += bytes;
420 tot->rx_packets += packets;
421
422 veth_stats_rx(&rx, peer);
423 tot->tx_dropped += rx.peer_tq_xdp_xmit_err;
424 tot->rx_dropped += rx.xdp_tx_err;
425 tot->tx_bytes += rx.xdp_bytes;
426 tot->tx_packets += rx.xdp_packets;
427 }
428 rcu_read_unlock();
429}
430
431/* fake multicast ability */
432static void veth_set_multicast_list(struct net_device *dev)
433{
434}
435
436static int veth_select_rxq(struct net_device *dev)
437{
438 return smp_processor_id() % dev->real_num_rx_queues;
439}
440
441static struct net_device *veth_peer_dev(struct net_device *dev)
442{
443 struct veth_priv *priv = netdev_priv(dev);
444
445 /* Callers must be under RCU read side. */
446 return rcu_dereference(priv->peer);
447}
448
449static int veth_xdp_xmit(struct net_device *dev, int n,
450 struct xdp_frame **frames,
451 u32 flags, bool ndo_xmit)
452{
453 struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
454 int i, ret = -ENXIO, nxmit = 0;
455 struct net_device *rcv;
456 unsigned int max_len;
457 struct veth_rq *rq;
458
459 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
460 return -EINVAL;
461
462 rcu_read_lock();
463 rcv = rcu_dereference(priv->peer);
464 if (unlikely(!rcv))
465 goto out;
466
467 rcv_priv = netdev_priv(rcv);
468 rq = &rcv_priv->rq[veth_select_rxq(rcv)];
469 /* The napi pointer is set if NAPI is enabled, which ensures that
470 * xdp_ring is initialized on receive side and the peer device is up.
471 */
472 if (!rcu_access_pointer(rq->napi))
473 goto out;
474
475 max_len = rcv->mtu + rcv->hard_header_len + VLAN_HLEN;
476
477 spin_lock(&rq->xdp_ring.producer_lock);
478 for (i = 0; i < n; i++) {
479 struct xdp_frame *frame = frames[i];
480 void *ptr = veth_xdp_to_ptr(frame);
481
482 if (unlikely(xdp_get_frame_len(frame) > max_len ||
483 __ptr_ring_produce(&rq->xdp_ring, ptr)))
484 break;
485 nxmit++;
486 }
487 spin_unlock(&rq->xdp_ring.producer_lock);
488
489 if (flags & XDP_XMIT_FLUSH)
490 __veth_xdp_flush(rq);
491
492 ret = nxmit;
493 if (ndo_xmit) {
494 u64_stats_update_begin(&rq->stats.syncp);
495 rq->stats.vs.peer_tq_xdp_xmit += nxmit;
496 rq->stats.vs.peer_tq_xdp_xmit_err += n - nxmit;
497 u64_stats_update_end(&rq->stats.syncp);
498 }
499
500out:
501 rcu_read_unlock();
502
503 return ret;
504}
505
506static int veth_ndo_xdp_xmit(struct net_device *dev, int n,
507 struct xdp_frame **frames, u32 flags)
508{
509 int err;
510
511 err = veth_xdp_xmit(dev, n, frames, flags, true);
512 if (err < 0) {
513 struct veth_priv *priv = netdev_priv(dev);
514
515 atomic64_add(n, &priv->dropped);
516 }
517
518 return err;
519}
520
521static void veth_xdp_flush_bq(struct veth_rq *rq, struct veth_xdp_tx_bq *bq)
522{
523 int sent, i, err = 0, drops;
524
525 sent = veth_xdp_xmit(rq->dev, bq->count, bq->q, 0, false);
526 if (sent < 0) {
527 err = sent;
528 sent = 0;
529 }
530
531 for (i = sent; unlikely(i < bq->count); i++)
532 xdp_return_frame(bq->q[i]);
533
534 drops = bq->count - sent;
535 trace_xdp_bulk_tx(rq->dev, sent, drops, err);
536
537 u64_stats_update_begin(&rq->stats.syncp);
538 rq->stats.vs.xdp_tx += sent;
539 rq->stats.vs.xdp_tx_err += drops;
540 u64_stats_update_end(&rq->stats.syncp);
541
542 bq->count = 0;
543}
544
545static void veth_xdp_flush(struct veth_rq *rq, struct veth_xdp_tx_bq *bq)
546{
547 struct veth_priv *rcv_priv, *priv = netdev_priv(rq->dev);
548 struct net_device *rcv;
549 struct veth_rq *rcv_rq;
550
551 rcu_read_lock();
552 veth_xdp_flush_bq(rq, bq);
553 rcv = rcu_dereference(priv->peer);
554 if (unlikely(!rcv))
555 goto out;
556
557 rcv_priv = netdev_priv(rcv);
558 rcv_rq = &rcv_priv->rq[veth_select_rxq(rcv)];
559 /* xdp_ring is initialized on receive side? */
560 if (unlikely(!rcu_access_pointer(rcv_rq->xdp_prog)))
561 goto out;
562
563 __veth_xdp_flush(rcv_rq);
564out:
565 rcu_read_unlock();
566}
567
568static int veth_xdp_tx(struct veth_rq *rq, struct xdp_buff *xdp,
569 struct veth_xdp_tx_bq *bq)
570{
571 struct xdp_frame *frame = xdp_convert_buff_to_frame(xdp);
572
573 if (unlikely(!frame))
574 return -EOVERFLOW;
575
576 if (unlikely(bq->count == VETH_XDP_TX_BULK_SIZE))
577 veth_xdp_flush_bq(rq, bq);
578
579 bq->q[bq->count++] = frame;
580
581 return 0;
582}
583
584static struct xdp_frame *veth_xdp_rcv_one(struct veth_rq *rq,
585 struct xdp_frame *frame,
586 struct veth_xdp_tx_bq *bq,
587 struct veth_stats *stats)
588{
589 struct xdp_frame orig_frame;
590 struct bpf_prog *xdp_prog;
591
592 rcu_read_lock();
593 xdp_prog = rcu_dereference(rq->xdp_prog);
594 if (likely(xdp_prog)) {
595 struct xdp_buff xdp;
596 u32 act;
597
598 xdp_convert_frame_to_buff(frame, &xdp);
599 xdp.rxq = &rq->xdp_rxq;
600
601 act = bpf_prog_run_xdp(xdp_prog, &xdp);
602
603 switch (act) {
604 case XDP_PASS:
605 if (xdp_update_frame_from_buff(&xdp, frame))
606 goto err_xdp;
607 break;
608 case XDP_TX:
609 orig_frame = *frame;
610 xdp.rxq->mem = frame->mem;
611 if (unlikely(veth_xdp_tx(rq, &xdp, bq) < 0)) {
612 trace_xdp_exception(rq->dev, xdp_prog, act);
613 frame = &orig_frame;
614 stats->rx_drops++;
615 goto err_xdp;
616 }
617 stats->xdp_tx++;
618 rcu_read_unlock();
619 goto xdp_xmit;
620 case XDP_REDIRECT:
621 orig_frame = *frame;
622 xdp.rxq->mem = frame->mem;
623 if (xdp_do_redirect(rq->dev, &xdp, xdp_prog)) {
624 frame = &orig_frame;
625 stats->rx_drops++;
626 goto err_xdp;
627 }
628 stats->xdp_redirect++;
629 rcu_read_unlock();
630 goto xdp_xmit;
631 default:
632 bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act);
633 fallthrough;
634 case XDP_ABORTED:
635 trace_xdp_exception(rq->dev, xdp_prog, act);
636 fallthrough;
637 case XDP_DROP:
638 stats->xdp_drops++;
639 goto err_xdp;
640 }
641 }
642 rcu_read_unlock();
643
644 return frame;
645err_xdp:
646 rcu_read_unlock();
647 xdp_return_frame(frame);
648xdp_xmit:
649 return NULL;
650}
651
652/* frames array contains VETH_XDP_BATCH at most */
653static void veth_xdp_rcv_bulk_skb(struct veth_rq *rq, void **frames,
654 int n_xdpf, struct veth_xdp_tx_bq *bq,
655 struct veth_stats *stats)
656{
657 void *skbs[VETH_XDP_BATCH];
658 int i;
659
660 if (xdp_alloc_skb_bulk(skbs, n_xdpf,
661 GFP_ATOMIC | __GFP_ZERO) < 0) {
662 for (i = 0; i < n_xdpf; i++)
663 xdp_return_frame(frames[i]);
664 stats->rx_drops += n_xdpf;
665
666 return;
667 }
668
669 for (i = 0; i < n_xdpf; i++) {
670 struct sk_buff *skb = skbs[i];
671
672 skb = __xdp_build_skb_from_frame(frames[i], skb,
673 rq->dev);
674 if (!skb) {
675 xdp_return_frame(frames[i]);
676 stats->rx_drops++;
677 continue;
678 }
679 napi_gro_receive(&rq->xdp_napi, skb);
680 }
681}
682
683static void veth_xdp_get(struct xdp_buff *xdp)
684{
685 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
686 int i;
687
688 get_page(virt_to_page(xdp->data));
689 if (likely(!xdp_buff_has_frags(xdp)))
690 return;
691
692 for (i = 0; i < sinfo->nr_frags; i++)
693 __skb_frag_ref(&sinfo->frags[i]);
694}
695
696static int veth_convert_skb_to_xdp_buff(struct veth_rq *rq,
697 struct xdp_buff *xdp,
698 struct sk_buff **pskb)
699{
700 struct sk_buff *skb = *pskb;
701 u32 frame_sz;
702
703 if (skb_shared(skb) || skb_head_is_locked(skb) ||
704 skb_shinfo(skb)->nr_frags) {
705 u32 size, len, max_head_size, off;
706 struct sk_buff *nskb;
707 struct page *page;
708 int i, head_off;
709
710 /* We need a private copy of the skb and data buffers since
711 * the ebpf program can modify it. We segment the original skb
712 * into order-0 pages without linearize it.
713 *
714 * Make sure we have enough space for linear and paged area
715 */
716 max_head_size = SKB_WITH_OVERHEAD(PAGE_SIZE -
717 VETH_XDP_HEADROOM);
718 if (skb->len > PAGE_SIZE * MAX_SKB_FRAGS + max_head_size)
719 goto drop;
720
721 /* Allocate skb head */
722 page = alloc_page(GFP_ATOMIC | __GFP_NOWARN);
723 if (!page)
724 goto drop;
725
726 nskb = build_skb(page_address(page), PAGE_SIZE);
727 if (!nskb) {
728 put_page(page);
729 goto drop;
730 }
731
732 skb_reserve(nskb, VETH_XDP_HEADROOM);
733 size = min_t(u32, skb->len, max_head_size);
734 if (skb_copy_bits(skb, 0, nskb->data, size)) {
735 consume_skb(nskb);
736 goto drop;
737 }
738 skb_put(nskb, size);
739
740 skb_copy_header(nskb, skb);
741 head_off = skb_headroom(nskb) - skb_headroom(skb);
742 skb_headers_offset_update(nskb, head_off);
743
744 /* Allocate paged area of new skb */
745 off = size;
746 len = skb->len - off;
747
748 for (i = 0; i < MAX_SKB_FRAGS && off < skb->len; i++) {
749 page = alloc_page(GFP_ATOMIC | __GFP_NOWARN);
750 if (!page) {
751 consume_skb(nskb);
752 goto drop;
753 }
754
755 size = min_t(u32, len, PAGE_SIZE);
756 skb_add_rx_frag(nskb, i, page, 0, size, PAGE_SIZE);
757 if (skb_copy_bits(skb, off, page_address(page),
758 size)) {
759 consume_skb(nskb);
760 goto drop;
761 }
762
763 len -= size;
764 off += size;
765 }
766
767 consume_skb(skb);
768 skb = nskb;
769 } else if (skb_headroom(skb) < XDP_PACKET_HEADROOM &&
770 pskb_expand_head(skb, VETH_XDP_HEADROOM, 0, GFP_ATOMIC)) {
771 goto drop;
772 }
773
774 /* SKB "head" area always have tailroom for skb_shared_info */
775 frame_sz = skb_end_pointer(skb) - skb->head;
776 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
777 xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq);
778 xdp_prepare_buff(xdp, skb->head, skb_headroom(skb),
779 skb_headlen(skb), true);
780
781 if (skb_is_nonlinear(skb)) {
782 skb_shinfo(skb)->xdp_frags_size = skb->data_len;
783 xdp_buff_set_frags_flag(xdp);
784 } else {
785 xdp_buff_clear_frags_flag(xdp);
786 }
787 *pskb = skb;
788
789 return 0;
790drop:
791 consume_skb(skb);
792 *pskb = NULL;
793
794 return -ENOMEM;
795}
796
797static struct sk_buff *veth_xdp_rcv_skb(struct veth_rq *rq,
798 struct sk_buff *skb,
799 struct veth_xdp_tx_bq *bq,
800 struct veth_stats *stats)
801{
802 void *orig_data, *orig_data_end;
803 struct bpf_prog *xdp_prog;
804 struct xdp_buff xdp;
805 u32 act, metalen;
806 int off;
807
808 skb_prepare_for_gro(skb);
809
810 rcu_read_lock();
811 xdp_prog = rcu_dereference(rq->xdp_prog);
812 if (unlikely(!xdp_prog)) {
813 rcu_read_unlock();
814 goto out;
815 }
816
817 __skb_push(skb, skb->data - skb_mac_header(skb));
818 if (veth_convert_skb_to_xdp_buff(rq, &xdp, &skb))
819 goto drop;
820
821 orig_data = xdp.data;
822 orig_data_end = xdp.data_end;
823
824 act = bpf_prog_run_xdp(xdp_prog, &xdp);
825
826 switch (act) {
827 case XDP_PASS:
828 break;
829 case XDP_TX:
830 veth_xdp_get(&xdp);
831 consume_skb(skb);
832 xdp.rxq->mem = rq->xdp_mem;
833 if (unlikely(veth_xdp_tx(rq, &xdp, bq) < 0)) {
834 trace_xdp_exception(rq->dev, xdp_prog, act);
835 stats->rx_drops++;
836 goto err_xdp;
837 }
838 stats->xdp_tx++;
839 rcu_read_unlock();
840 goto xdp_xmit;
841 case XDP_REDIRECT:
842 veth_xdp_get(&xdp);
843 consume_skb(skb);
844 xdp.rxq->mem = rq->xdp_mem;
845 if (xdp_do_redirect(rq->dev, &xdp, xdp_prog)) {
846 stats->rx_drops++;
847 goto err_xdp;
848 }
849 stats->xdp_redirect++;
850 rcu_read_unlock();
851 goto xdp_xmit;
852 default:
853 bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act);
854 fallthrough;
855 case XDP_ABORTED:
856 trace_xdp_exception(rq->dev, xdp_prog, act);
857 fallthrough;
858 case XDP_DROP:
859 stats->xdp_drops++;
860 goto xdp_drop;
861 }
862 rcu_read_unlock();
863
864 /* check if bpf_xdp_adjust_head was used */
865 off = orig_data - xdp.data;
866 if (off > 0)
867 __skb_push(skb, off);
868 else if (off < 0)
869 __skb_pull(skb, -off);
870
871 skb_reset_mac_header(skb);
872
873 /* check if bpf_xdp_adjust_tail was used */
874 off = xdp.data_end - orig_data_end;
875 if (off != 0)
876 __skb_put(skb, off); /* positive on grow, negative on shrink */
877
878 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers
879 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here.
880 */
881 if (xdp_buff_has_frags(&xdp))
882 skb->data_len = skb_shinfo(skb)->xdp_frags_size;
883 else
884 skb->data_len = 0;
885
886 skb->protocol = eth_type_trans(skb, rq->dev);
887
888 metalen = xdp.data - xdp.data_meta;
889 if (metalen)
890 skb_metadata_set(skb, metalen);
891out:
892 return skb;
893drop:
894 stats->rx_drops++;
895xdp_drop:
896 rcu_read_unlock();
897 kfree_skb(skb);
898 return NULL;
899err_xdp:
900 rcu_read_unlock();
901 xdp_return_buff(&xdp);
902xdp_xmit:
903 return NULL;
904}
905
906static int veth_xdp_rcv(struct veth_rq *rq, int budget,
907 struct veth_xdp_tx_bq *bq,
908 struct veth_stats *stats)
909{
910 int i, done = 0, n_xdpf = 0;
911 void *xdpf[VETH_XDP_BATCH];
912
913 for (i = 0; i < budget; i++) {
914 void *ptr = __ptr_ring_consume(&rq->xdp_ring);
915
916 if (!ptr)
917 break;
918
919 if (veth_is_xdp_frame(ptr)) {
920 /* ndo_xdp_xmit */
921 struct xdp_frame *frame = veth_ptr_to_xdp(ptr);
922
923 stats->xdp_bytes += xdp_get_frame_len(frame);
924 frame = veth_xdp_rcv_one(rq, frame, bq, stats);
925 if (frame) {
926 /* XDP_PASS */
927 xdpf[n_xdpf++] = frame;
928 if (n_xdpf == VETH_XDP_BATCH) {
929 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf,
930 bq, stats);
931 n_xdpf = 0;
932 }
933 }
934 } else {
935 /* ndo_start_xmit */
936 struct sk_buff *skb = ptr;
937
938 stats->xdp_bytes += skb->len;
939 skb = veth_xdp_rcv_skb(rq, skb, bq, stats);
940 if (skb) {
941 if (skb_shared(skb) || skb_unclone(skb, GFP_ATOMIC))
942 netif_receive_skb(skb);
943 else
944 napi_gro_receive(&rq->xdp_napi, skb);
945 }
946 }
947 done++;
948 }
949
950 if (n_xdpf)
951 veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, bq, stats);
952
953 u64_stats_update_begin(&rq->stats.syncp);
954 rq->stats.vs.xdp_redirect += stats->xdp_redirect;
955 rq->stats.vs.xdp_bytes += stats->xdp_bytes;
956 rq->stats.vs.xdp_drops += stats->xdp_drops;
957 rq->stats.vs.rx_drops += stats->rx_drops;
958 rq->stats.vs.xdp_packets += done;
959 u64_stats_update_end(&rq->stats.syncp);
960
961 return done;
962}
963
964static int veth_poll(struct napi_struct *napi, int budget)
965{
966 struct veth_rq *rq =
967 container_of(napi, struct veth_rq, xdp_napi);
968 struct veth_stats stats = {};
969 struct veth_xdp_tx_bq bq;
970 int done;
971
972 bq.count = 0;
973
974 xdp_set_return_frame_no_direct();
975 done = veth_xdp_rcv(rq, budget, &bq, &stats);
976
977 if (stats.xdp_redirect > 0)
978 xdp_do_flush();
979
980 if (done < budget && napi_complete_done(napi, done)) {
981 /* Write rx_notify_masked before reading ptr_ring */
982 smp_store_mb(rq->rx_notify_masked, false);
983 if (unlikely(!__ptr_ring_empty(&rq->xdp_ring))) {
984 if (napi_schedule_prep(&rq->xdp_napi)) {
985 WRITE_ONCE(rq->rx_notify_masked, true);
986 __napi_schedule(&rq->xdp_napi);
987 }
988 }
989 }
990
991 if (stats.xdp_tx > 0)
992 veth_xdp_flush(rq, &bq);
993 xdp_clear_return_frame_no_direct();
994
995 return done;
996}
997
998static int __veth_napi_enable_range(struct net_device *dev, int start, int end)
999{
1000 struct veth_priv *priv = netdev_priv(dev);
1001 int err, i;
1002
1003 for (i = start; i < end; i++) {
1004 struct veth_rq *rq = &priv->rq[i];
1005
1006 err = ptr_ring_init(&rq->xdp_ring, VETH_RING_SIZE, GFP_KERNEL);
1007 if (err)
1008 goto err_xdp_ring;
1009 }
1010
1011 for (i = start; i < end; i++) {
1012 struct veth_rq *rq = &priv->rq[i];
1013
1014 napi_enable(&rq->xdp_napi);
1015 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1016 }
1017
1018 return 0;
1019
1020err_xdp_ring:
1021 for (i--; i >= start; i--)
1022 ptr_ring_cleanup(&priv->rq[i].xdp_ring, veth_ptr_free);
1023
1024 return err;
1025}
1026
1027static int __veth_napi_enable(struct net_device *dev)
1028{
1029 return __veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1030}
1031
1032static void veth_napi_del_range(struct net_device *dev, int start, int end)
1033{
1034 struct veth_priv *priv = netdev_priv(dev);
1035 int i;
1036
1037 for (i = start; i < end; i++) {
1038 struct veth_rq *rq = &priv->rq[i];
1039
1040 rcu_assign_pointer(priv->rq[i].napi, NULL);
1041 napi_disable(&rq->xdp_napi);
1042 __netif_napi_del(&rq->xdp_napi);
1043 }
1044 synchronize_net();
1045
1046 for (i = start; i < end; i++) {
1047 struct veth_rq *rq = &priv->rq[i];
1048
1049 rq->rx_notify_masked = false;
1050 ptr_ring_cleanup(&rq->xdp_ring, veth_ptr_free);
1051 }
1052}
1053
1054static void veth_napi_del(struct net_device *dev)
1055{
1056 veth_napi_del_range(dev, 0, dev->real_num_rx_queues);
1057}
1058
1059static bool veth_gro_requested(const struct net_device *dev)
1060{
1061 return !!(dev->wanted_features & NETIF_F_GRO);
1062}
1063
1064static int veth_enable_xdp_range(struct net_device *dev, int start, int end,
1065 bool napi_already_on)
1066{
1067 struct veth_priv *priv = netdev_priv(dev);
1068 int err, i;
1069
1070 for (i = start; i < end; i++) {
1071 struct veth_rq *rq = &priv->rq[i];
1072
1073 if (!napi_already_on)
1074 netif_napi_add(dev, &rq->xdp_napi, veth_poll);
1075 err = xdp_rxq_info_reg(&rq->xdp_rxq, dev, i, rq->xdp_napi.napi_id);
1076 if (err < 0)
1077 goto err_rxq_reg;
1078
1079 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
1080 MEM_TYPE_PAGE_SHARED,
1081 NULL);
1082 if (err < 0)
1083 goto err_reg_mem;
1084
1085 /* Save original mem info as it can be overwritten */
1086 rq->xdp_mem = rq->xdp_rxq.mem;
1087 }
1088 return 0;
1089
1090err_reg_mem:
1091 xdp_rxq_info_unreg(&priv->rq[i].xdp_rxq);
1092err_rxq_reg:
1093 for (i--; i >= start; i--) {
1094 struct veth_rq *rq = &priv->rq[i];
1095
1096 xdp_rxq_info_unreg(&rq->xdp_rxq);
1097 if (!napi_already_on)
1098 netif_napi_del(&rq->xdp_napi);
1099 }
1100
1101 return err;
1102}
1103
1104static void veth_disable_xdp_range(struct net_device *dev, int start, int end,
1105 bool delete_napi)
1106{
1107 struct veth_priv *priv = netdev_priv(dev);
1108 int i;
1109
1110 for (i = start; i < end; i++) {
1111 struct veth_rq *rq = &priv->rq[i];
1112
1113 rq->xdp_rxq.mem = rq->xdp_mem;
1114 xdp_rxq_info_unreg(&rq->xdp_rxq);
1115
1116 if (delete_napi)
1117 netif_napi_del(&rq->xdp_napi);
1118 }
1119}
1120
1121static int veth_enable_xdp(struct net_device *dev)
1122{
1123 bool napi_already_on = veth_gro_requested(dev) && (dev->flags & IFF_UP);
1124 struct veth_priv *priv = netdev_priv(dev);
1125 int err, i;
1126
1127 if (!xdp_rxq_info_is_reg(&priv->rq[0].xdp_rxq)) {
1128 err = veth_enable_xdp_range(dev, 0, dev->real_num_rx_queues, napi_already_on);
1129 if (err)
1130 return err;
1131
1132 if (!napi_already_on) {
1133 err = __veth_napi_enable(dev);
1134 if (err) {
1135 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, true);
1136 return err;
1137 }
1138
1139 if (!veth_gro_requested(dev)) {
1140 /* user-space did not require GRO, but adding XDP
1141 * is supposed to get GRO working
1142 */
1143 dev->features |= NETIF_F_GRO;
1144 netdev_features_change(dev);
1145 }
1146 }
1147 }
1148
1149 for (i = 0; i < dev->real_num_rx_queues; i++) {
1150 rcu_assign_pointer(priv->rq[i].xdp_prog, priv->_xdp_prog);
1151 rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1152 }
1153
1154 return 0;
1155}
1156
1157static void veth_disable_xdp(struct net_device *dev)
1158{
1159 struct veth_priv *priv = netdev_priv(dev);
1160 int i;
1161
1162 for (i = 0; i < dev->real_num_rx_queues; i++)
1163 rcu_assign_pointer(priv->rq[i].xdp_prog, NULL);
1164
1165 if (!netif_running(dev) || !veth_gro_requested(dev)) {
1166 veth_napi_del(dev);
1167
1168 /* if user-space did not require GRO, since adding XDP
1169 * enabled it, clear it now
1170 */
1171 if (!veth_gro_requested(dev) && netif_running(dev)) {
1172 dev->features &= ~NETIF_F_GRO;
1173 netdev_features_change(dev);
1174 }
1175 }
1176
1177 veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, false);
1178}
1179
1180static int veth_napi_enable_range(struct net_device *dev, int start, int end)
1181{
1182 struct veth_priv *priv = netdev_priv(dev);
1183 int err, i;
1184
1185 for (i = start; i < end; i++) {
1186 struct veth_rq *rq = &priv->rq[i];
1187
1188 netif_napi_add(dev, &rq->xdp_napi, veth_poll);
1189 }
1190
1191 err = __veth_napi_enable_range(dev, start, end);
1192 if (err) {
1193 for (i = start; i < end; i++) {
1194 struct veth_rq *rq = &priv->rq[i];
1195
1196 netif_napi_del(&rq->xdp_napi);
1197 }
1198 return err;
1199 }
1200 return err;
1201}
1202
1203static int veth_napi_enable(struct net_device *dev)
1204{
1205 return veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1206}
1207
1208static void veth_disable_range_safe(struct net_device *dev, int start, int end)
1209{
1210 struct veth_priv *priv = netdev_priv(dev);
1211
1212 if (start >= end)
1213 return;
1214
1215 if (priv->_xdp_prog) {
1216 veth_napi_del_range(dev, start, end);
1217 veth_disable_xdp_range(dev, start, end, false);
1218 } else if (veth_gro_requested(dev)) {
1219 veth_napi_del_range(dev, start, end);
1220 }
1221}
1222
1223static int veth_enable_range_safe(struct net_device *dev, int start, int end)
1224{
1225 struct veth_priv *priv = netdev_priv(dev);
1226 int err;
1227
1228 if (start >= end)
1229 return 0;
1230
1231 if (priv->_xdp_prog) {
1232 /* these channels are freshly initialized, napi is not on there even
1233 * when GRO is requeste
1234 */
1235 err = veth_enable_xdp_range(dev, start, end, false);
1236 if (err)
1237 return err;
1238
1239 err = __veth_napi_enable_range(dev, start, end);
1240 if (err) {
1241 /* on error always delete the newly added napis */
1242 veth_disable_xdp_range(dev, start, end, true);
1243 return err;
1244 }
1245 } else if (veth_gro_requested(dev)) {
1246 return veth_napi_enable_range(dev, start, end);
1247 }
1248 return 0;
1249}
1250
1251static int veth_set_channels(struct net_device *dev,
1252 struct ethtool_channels *ch)
1253{
1254 struct veth_priv *priv = netdev_priv(dev);
1255 unsigned int old_rx_count, new_rx_count;
1256 struct veth_priv *peer_priv;
1257 struct net_device *peer;
1258 int err;
1259
1260 /* sanity check. Upper bounds are already enforced by the caller */
1261 if (!ch->rx_count || !ch->tx_count)
1262 return -EINVAL;
1263
1264 /* avoid braking XDP, if that is enabled */
1265 peer = rtnl_dereference(priv->peer);
1266 peer_priv = peer ? netdev_priv(peer) : NULL;
1267 if (priv->_xdp_prog && peer && ch->rx_count < peer->real_num_tx_queues)
1268 return -EINVAL;
1269
1270 if (peer && peer_priv && peer_priv->_xdp_prog && ch->tx_count > peer->real_num_rx_queues)
1271 return -EINVAL;
1272
1273 old_rx_count = dev->real_num_rx_queues;
1274 new_rx_count = ch->rx_count;
1275 if (netif_running(dev)) {
1276 /* turn device off */
1277 netif_carrier_off(dev);
1278 if (peer)
1279 netif_carrier_off(peer);
1280
1281 /* try to allocate new resurces, as needed*/
1282 err = veth_enable_range_safe(dev, old_rx_count, new_rx_count);
1283 if (err)
1284 goto out;
1285 }
1286
1287 err = netif_set_real_num_rx_queues(dev, ch->rx_count);
1288 if (err)
1289 goto revert;
1290
1291 err = netif_set_real_num_tx_queues(dev, ch->tx_count);
1292 if (err) {
1293 int err2 = netif_set_real_num_rx_queues(dev, old_rx_count);
1294
1295 /* this error condition could happen only if rx and tx change
1296 * in opposite directions (e.g. tx nr raises, rx nr decreases)
1297 * and we can't do anything to fully restore the original
1298 * status
1299 */
1300 if (err2)
1301 pr_warn("Can't restore rx queues config %d -> %d %d",
1302 new_rx_count, old_rx_count, err2);
1303 else
1304 goto revert;
1305 }
1306
1307out:
1308 if (netif_running(dev)) {
1309 /* note that we need to swap the arguments WRT the enable part
1310 * to identify the range we have to disable
1311 */
1312 veth_disable_range_safe(dev, new_rx_count, old_rx_count);
1313 netif_carrier_on(dev);
1314 if (peer)
1315 netif_carrier_on(peer);
1316 }
1317 return err;
1318
1319revert:
1320 new_rx_count = old_rx_count;
1321 old_rx_count = ch->rx_count;
1322 goto out;
1323}
1324
1325static int veth_open(struct net_device *dev)
1326{
1327 struct veth_priv *priv = netdev_priv(dev);
1328 struct net_device *peer = rtnl_dereference(priv->peer);
1329 int err;
1330
1331 if (!peer)
1332 return -ENOTCONN;
1333
1334 if (priv->_xdp_prog) {
1335 err = veth_enable_xdp(dev);
1336 if (err)
1337 return err;
1338 } else if (veth_gro_requested(dev)) {
1339 err = veth_napi_enable(dev);
1340 if (err)
1341 return err;
1342 }
1343
1344 if (peer->flags & IFF_UP) {
1345 netif_carrier_on(dev);
1346 netif_carrier_on(peer);
1347 }
1348
1349 return 0;
1350}
1351
1352static int veth_close(struct net_device *dev)
1353{
1354 struct veth_priv *priv = netdev_priv(dev);
1355 struct net_device *peer = rtnl_dereference(priv->peer);
1356
1357 netif_carrier_off(dev);
1358 if (peer)
1359 netif_carrier_off(peer);
1360
1361 if (priv->_xdp_prog)
1362 veth_disable_xdp(dev);
1363 else if (veth_gro_requested(dev))
1364 veth_napi_del(dev);
1365
1366 return 0;
1367}
1368
1369static int is_valid_veth_mtu(int mtu)
1370{
1371 return mtu >= ETH_MIN_MTU && mtu <= ETH_MAX_MTU;
1372}
1373
1374static int veth_alloc_queues(struct net_device *dev)
1375{
1376 struct veth_priv *priv = netdev_priv(dev);
1377 int i;
1378
1379 priv->rq = kcalloc(dev->num_rx_queues, sizeof(*priv->rq), GFP_KERNEL_ACCOUNT);
1380 if (!priv->rq)
1381 return -ENOMEM;
1382
1383 for (i = 0; i < dev->num_rx_queues; i++) {
1384 priv->rq[i].dev = dev;
1385 u64_stats_init(&priv->rq[i].stats.syncp);
1386 }
1387
1388 return 0;
1389}
1390
1391static void veth_free_queues(struct net_device *dev)
1392{
1393 struct veth_priv *priv = netdev_priv(dev);
1394
1395 kfree(priv->rq);
1396}
1397
1398static int veth_dev_init(struct net_device *dev)
1399{
1400 int err;
1401
1402 dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats);
1403 if (!dev->lstats)
1404 return -ENOMEM;
1405
1406 err = veth_alloc_queues(dev);
1407 if (err) {
1408 free_percpu(dev->lstats);
1409 return err;
1410 }
1411
1412 return 0;
1413}
1414
1415static void veth_dev_free(struct net_device *dev)
1416{
1417 veth_free_queues(dev);
1418 free_percpu(dev->lstats);
1419}
1420
1421#ifdef CONFIG_NET_POLL_CONTROLLER
1422static void veth_poll_controller(struct net_device *dev)
1423{
1424 /* veth only receives frames when its peer sends one
1425 * Since it has nothing to do with disabling irqs, we are guaranteed
1426 * never to have pending data when we poll for it so
1427 * there is nothing to do here.
1428 *
1429 * We need this though so netpoll recognizes us as an interface that
1430 * supports polling, which enables bridge devices in virt setups to
1431 * still use netconsole
1432 */
1433}
1434#endif /* CONFIG_NET_POLL_CONTROLLER */
1435
1436static int veth_get_iflink(const struct net_device *dev)
1437{
1438 struct veth_priv *priv = netdev_priv(dev);
1439 struct net_device *peer;
1440 int iflink;
1441
1442 rcu_read_lock();
1443 peer = rcu_dereference(priv->peer);
1444 iflink = peer ? peer->ifindex : 0;
1445 rcu_read_unlock();
1446
1447 return iflink;
1448}
1449
1450static netdev_features_t veth_fix_features(struct net_device *dev,
1451 netdev_features_t features)
1452{
1453 struct veth_priv *priv = netdev_priv(dev);
1454 struct net_device *peer;
1455
1456 peer = rtnl_dereference(priv->peer);
1457 if (peer) {
1458 struct veth_priv *peer_priv = netdev_priv(peer);
1459
1460 if (peer_priv->_xdp_prog)
1461 features &= ~NETIF_F_GSO_SOFTWARE;
1462 }
1463 if (priv->_xdp_prog)
1464 features |= NETIF_F_GRO;
1465
1466 return features;
1467}
1468
1469static int veth_set_features(struct net_device *dev,
1470 netdev_features_t features)
1471{
1472 netdev_features_t changed = features ^ dev->features;
1473 struct veth_priv *priv = netdev_priv(dev);
1474 int err;
1475
1476 if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog)
1477 return 0;
1478
1479 if (features & NETIF_F_GRO) {
1480 err = veth_napi_enable(dev);
1481 if (err)
1482 return err;
1483 } else {
1484 veth_napi_del(dev);
1485 }
1486 return 0;
1487}
1488
1489static void veth_set_rx_headroom(struct net_device *dev, int new_hr)
1490{
1491 struct veth_priv *peer_priv, *priv = netdev_priv(dev);
1492 struct net_device *peer;
1493
1494 if (new_hr < 0)
1495 new_hr = 0;
1496
1497 rcu_read_lock();
1498 peer = rcu_dereference(priv->peer);
1499 if (unlikely(!peer))
1500 goto out;
1501
1502 peer_priv = netdev_priv(peer);
1503 priv->requested_headroom = new_hr;
1504 new_hr = max(priv->requested_headroom, peer_priv->requested_headroom);
1505 dev->needed_headroom = new_hr;
1506 peer->needed_headroom = new_hr;
1507
1508out:
1509 rcu_read_unlock();
1510}
1511
1512static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1513 struct netlink_ext_ack *extack)
1514{
1515 struct veth_priv *priv = netdev_priv(dev);
1516 struct bpf_prog *old_prog;
1517 struct net_device *peer;
1518 unsigned int max_mtu;
1519 int err;
1520
1521 old_prog = priv->_xdp_prog;
1522 priv->_xdp_prog = prog;
1523 peer = rtnl_dereference(priv->peer);
1524
1525 if (prog) {
1526 if (!peer) {
1527 NL_SET_ERR_MSG_MOD(extack, "Cannot set XDP when peer is detached");
1528 err = -ENOTCONN;
1529 goto err;
1530 }
1531
1532 max_mtu = SKB_WITH_OVERHEAD(PAGE_SIZE - VETH_XDP_HEADROOM) -
1533 peer->hard_header_len;
1534 /* Allow increasing the max_mtu if the program supports
1535 * XDP fragments.
1536 */
1537 if (prog->aux->xdp_has_frags)
1538 max_mtu += PAGE_SIZE * MAX_SKB_FRAGS;
1539
1540 if (peer->mtu > max_mtu) {
1541 NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP");
1542 err = -ERANGE;
1543 goto err;
1544 }
1545
1546 if (dev->real_num_rx_queues < peer->real_num_tx_queues) {
1547 NL_SET_ERR_MSG_MOD(extack, "XDP expects number of rx queues not less than peer tx queues");
1548 err = -ENOSPC;
1549 goto err;
1550 }
1551
1552 if (dev->flags & IFF_UP) {
1553 err = veth_enable_xdp(dev);
1554 if (err) {
1555 NL_SET_ERR_MSG_MOD(extack, "Setup for XDP failed");
1556 goto err;
1557 }
1558 }
1559
1560 if (!old_prog) {
1561 peer->hw_features &= ~NETIF_F_GSO_SOFTWARE;
1562 peer->max_mtu = max_mtu;
1563 }
1564 }
1565
1566 if (old_prog) {
1567 if (!prog) {
1568 if (dev->flags & IFF_UP)
1569 veth_disable_xdp(dev);
1570
1571 if (peer) {
1572 peer->hw_features |= NETIF_F_GSO_SOFTWARE;
1573 peer->max_mtu = ETH_MAX_MTU;
1574 }
1575 }
1576 bpf_prog_put(old_prog);
1577 }
1578
1579 if ((!!old_prog ^ !!prog) && peer)
1580 netdev_update_features(peer);
1581
1582 return 0;
1583err:
1584 priv->_xdp_prog = old_prog;
1585
1586 return err;
1587}
1588
1589static int veth_xdp(struct net_device *dev, struct netdev_bpf *xdp)
1590{
1591 switch (xdp->command) {
1592 case XDP_SETUP_PROG:
1593 return veth_xdp_set(dev, xdp->prog, xdp->extack);
1594 default:
1595 return -EINVAL;
1596 }
1597}
1598
1599static const struct net_device_ops veth_netdev_ops = {
1600 .ndo_init = veth_dev_init,
1601 .ndo_open = veth_open,
1602 .ndo_stop = veth_close,
1603 .ndo_start_xmit = veth_xmit,
1604 .ndo_get_stats64 = veth_get_stats64,
1605 .ndo_set_rx_mode = veth_set_multicast_list,
1606 .ndo_set_mac_address = eth_mac_addr,
1607#ifdef CONFIG_NET_POLL_CONTROLLER
1608 .ndo_poll_controller = veth_poll_controller,
1609#endif
1610 .ndo_get_iflink = veth_get_iflink,
1611 .ndo_fix_features = veth_fix_features,
1612 .ndo_set_features = veth_set_features,
1613 .ndo_features_check = passthru_features_check,
1614 .ndo_set_rx_headroom = veth_set_rx_headroom,
1615 .ndo_bpf = veth_xdp,
1616 .ndo_xdp_xmit = veth_ndo_xdp_xmit,
1617 .ndo_get_peer_dev = veth_peer_dev,
1618};
1619
1620#define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HW_CSUM | \
1621 NETIF_F_RXCSUM | NETIF_F_SCTP_CRC | NETIF_F_HIGHDMA | \
1622 NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \
1623 NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX | \
1624 NETIF_F_HW_VLAN_STAG_TX | NETIF_F_HW_VLAN_STAG_RX )
1625
1626static void veth_setup(struct net_device *dev)
1627{
1628 ether_setup(dev);
1629
1630 dev->priv_flags &= ~IFF_TX_SKB_SHARING;
1631 dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
1632 dev->priv_flags |= IFF_NO_QUEUE;
1633 dev->priv_flags |= IFF_PHONY_HEADROOM;
1634
1635 dev->netdev_ops = &veth_netdev_ops;
1636 dev->ethtool_ops = &veth_ethtool_ops;
1637 dev->features |= NETIF_F_LLTX;
1638 dev->features |= VETH_FEATURES;
1639 dev->vlan_features = dev->features &
1640 ~(NETIF_F_HW_VLAN_CTAG_TX |
1641 NETIF_F_HW_VLAN_STAG_TX |
1642 NETIF_F_HW_VLAN_CTAG_RX |
1643 NETIF_F_HW_VLAN_STAG_RX);
1644 dev->needs_free_netdev = true;
1645 dev->priv_destructor = veth_dev_free;
1646 dev->max_mtu = ETH_MAX_MTU;
1647
1648 dev->hw_features = VETH_FEATURES;
1649 dev->hw_enc_features = VETH_FEATURES;
1650 dev->mpls_features = NETIF_F_HW_CSUM | NETIF_F_GSO_SOFTWARE;
1651 netif_set_tso_max_size(dev, GSO_MAX_SIZE);
1652}
1653
1654/*
1655 * netlink interface
1656 */
1657
1658static int veth_validate(struct nlattr *tb[], struct nlattr *data[],
1659 struct netlink_ext_ack *extack)
1660{
1661 if (tb[IFLA_ADDRESS]) {
1662 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
1663 return -EINVAL;
1664 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
1665 return -EADDRNOTAVAIL;
1666 }
1667 if (tb[IFLA_MTU]) {
1668 if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU])))
1669 return -EINVAL;
1670 }
1671 return 0;
1672}
1673
1674static struct rtnl_link_ops veth_link_ops;
1675
1676static void veth_disable_gro(struct net_device *dev)
1677{
1678 dev->features &= ~NETIF_F_GRO;
1679 dev->wanted_features &= ~NETIF_F_GRO;
1680 netdev_update_features(dev);
1681}
1682
1683static int veth_init_queues(struct net_device *dev, struct nlattr *tb[])
1684{
1685 int err;
1686
1687 if (!tb[IFLA_NUM_TX_QUEUES] && dev->num_tx_queues > 1) {
1688 err = netif_set_real_num_tx_queues(dev, 1);
1689 if (err)
1690 return err;
1691 }
1692 if (!tb[IFLA_NUM_RX_QUEUES] && dev->num_rx_queues > 1) {
1693 err = netif_set_real_num_rx_queues(dev, 1);
1694 if (err)
1695 return err;
1696 }
1697 return 0;
1698}
1699
1700static int veth_newlink(struct net *src_net, struct net_device *dev,
1701 struct nlattr *tb[], struct nlattr *data[],
1702 struct netlink_ext_ack *extack)
1703{
1704 int err;
1705 struct net_device *peer;
1706 struct veth_priv *priv;
1707 char ifname[IFNAMSIZ];
1708 struct nlattr *peer_tb[IFLA_MAX + 1], **tbp;
1709 unsigned char name_assign_type;
1710 struct ifinfomsg *ifmp;
1711 struct net *net;
1712
1713 /*
1714 * create and register peer first
1715 */
1716 if (data != NULL && data[VETH_INFO_PEER] != NULL) {
1717 struct nlattr *nla_peer;
1718
1719 nla_peer = data[VETH_INFO_PEER];
1720 ifmp = nla_data(nla_peer);
1721 err = rtnl_nla_parse_ifla(peer_tb,
1722 nla_data(nla_peer) + sizeof(struct ifinfomsg),
1723 nla_len(nla_peer) - sizeof(struct ifinfomsg),
1724 NULL);
1725 if (err < 0)
1726 return err;
1727
1728 err = veth_validate(peer_tb, NULL, extack);
1729 if (err < 0)
1730 return err;
1731
1732 tbp = peer_tb;
1733 } else {
1734 ifmp = NULL;
1735 tbp = tb;
1736 }
1737
1738 if (ifmp && tbp[IFLA_IFNAME]) {
1739 nla_strscpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ);
1740 name_assign_type = NET_NAME_USER;
1741 } else {
1742 snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d");
1743 name_assign_type = NET_NAME_ENUM;
1744 }
1745
1746 net = rtnl_link_get_net(src_net, tbp);
1747 if (IS_ERR(net))
1748 return PTR_ERR(net);
1749
1750 peer = rtnl_create_link(net, ifname, name_assign_type,
1751 &veth_link_ops, tbp, extack);
1752 if (IS_ERR(peer)) {
1753 put_net(net);
1754 return PTR_ERR(peer);
1755 }
1756
1757 if (!ifmp || !tbp[IFLA_ADDRESS])
1758 eth_hw_addr_random(peer);
1759
1760 if (ifmp && (dev->ifindex != 0))
1761 peer->ifindex = ifmp->ifi_index;
1762
1763 netif_inherit_tso_max(peer, dev);
1764
1765 err = register_netdevice(peer);
1766 put_net(net);
1767 net = NULL;
1768 if (err < 0)
1769 goto err_register_peer;
1770
1771 /* keep GRO disabled by default to be consistent with the established
1772 * veth behavior
1773 */
1774 veth_disable_gro(peer);
1775 netif_carrier_off(peer);
1776
1777 err = rtnl_configure_link(peer, ifmp, 0, NULL);
1778 if (err < 0)
1779 goto err_configure_peer;
1780
1781 /*
1782 * register dev last
1783 *
1784 * note, that since we've registered new device the dev's name
1785 * should be re-allocated
1786 */
1787
1788 if (tb[IFLA_ADDRESS] == NULL)
1789 eth_hw_addr_random(dev);
1790
1791 if (tb[IFLA_IFNAME])
1792 nla_strscpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ);
1793 else
1794 snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d");
1795
1796 err = register_netdevice(dev);
1797 if (err < 0)
1798 goto err_register_dev;
1799
1800 netif_carrier_off(dev);
1801
1802 /*
1803 * tie the deviced together
1804 */
1805
1806 priv = netdev_priv(dev);
1807 rcu_assign_pointer(priv->peer, peer);
1808 err = veth_init_queues(dev, tb);
1809 if (err)
1810 goto err_queues;
1811
1812 priv = netdev_priv(peer);
1813 rcu_assign_pointer(priv->peer, dev);
1814 err = veth_init_queues(peer, tb);
1815 if (err)
1816 goto err_queues;
1817
1818 veth_disable_gro(dev);
1819 return 0;
1820
1821err_queues:
1822 unregister_netdevice(dev);
1823err_register_dev:
1824 /* nothing to do */
1825err_configure_peer:
1826 unregister_netdevice(peer);
1827 return err;
1828
1829err_register_peer:
1830 free_netdev(peer);
1831 return err;
1832}
1833
1834static void veth_dellink(struct net_device *dev, struct list_head *head)
1835{
1836 struct veth_priv *priv;
1837 struct net_device *peer;
1838
1839 priv = netdev_priv(dev);
1840 peer = rtnl_dereference(priv->peer);
1841
1842 /* Note : dellink() is called from default_device_exit_batch(),
1843 * before a rcu_synchronize() point. The devices are guaranteed
1844 * not being freed before one RCU grace period.
1845 */
1846 RCU_INIT_POINTER(priv->peer, NULL);
1847 unregister_netdevice_queue(dev, head);
1848
1849 if (peer) {
1850 priv = netdev_priv(peer);
1851 RCU_INIT_POINTER(priv->peer, NULL);
1852 unregister_netdevice_queue(peer, head);
1853 }
1854}
1855
1856static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = {
1857 [VETH_INFO_PEER] = { .len = sizeof(struct ifinfomsg) },
1858};
1859
1860static struct net *veth_get_link_net(const struct net_device *dev)
1861{
1862 struct veth_priv *priv = netdev_priv(dev);
1863 struct net_device *peer = rtnl_dereference(priv->peer);
1864
1865 return peer ? dev_net(peer) : dev_net(dev);
1866}
1867
1868static unsigned int veth_get_num_queues(void)
1869{
1870 /* enforce the same queue limit as rtnl_create_link */
1871 int queues = num_possible_cpus();
1872
1873 if (queues > 4096)
1874 queues = 4096;
1875 return queues;
1876}
1877
1878static struct rtnl_link_ops veth_link_ops = {
1879 .kind = DRV_NAME,
1880 .priv_size = sizeof(struct veth_priv),
1881 .setup = veth_setup,
1882 .validate = veth_validate,
1883 .newlink = veth_newlink,
1884 .dellink = veth_dellink,
1885 .policy = veth_policy,
1886 .maxtype = VETH_INFO_MAX,
1887 .get_link_net = veth_get_link_net,
1888 .get_num_tx_queues = veth_get_num_queues,
1889 .get_num_rx_queues = veth_get_num_queues,
1890};
1891
1892/*
1893 * init/fini
1894 */
1895
1896static __init int veth_init(void)
1897{
1898 return rtnl_link_register(&veth_link_ops);
1899}
1900
1901static __exit void veth_exit(void)
1902{
1903 rtnl_link_unregister(&veth_link_ops);
1904}
1905
1906module_init(veth_init);
1907module_exit(veth_exit);
1908
1909MODULE_DESCRIPTION("Virtual Ethernet Tunnel");
1910MODULE_LICENSE("GPL v2");
1911MODULE_ALIAS_RTNL_LINK(DRV_NAME);