xref: /linux/drivers/net/veth.c (revision 23c48a124b469cee2eb0c75e6d22d366d1caa118)
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 
40 struct 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 
53 struct veth_rq_stats {
54 	struct veth_stats	vs;
55 	struct u64_stats_sync	syncp;
56 };
57 
58 struct 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 
70 struct 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 
78 struct 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 
87 struct 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 
94 static 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 
106 static 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 
113 static struct {
114 	const char string[ETH_GSTRING_LEN];
115 } ethtool_stats_keys[] = {
116 	{ "peer_ifindex" },
117 };
118 
119 static 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 
129 static void veth_get_drvinfo(struct net_device *dev, struct ethtool_drvinfo *info)
130 {
131 	strlcpy(info->driver, DRV_NAME, sizeof(info->driver));
132 	strlcpy(info->version, DRV_VERSION, sizeof(info->version));
133 }
134 
135 static 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, &ethtool_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 
157 static 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 
169 static 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_irq(&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_irq(&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_irq(&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_irq(&rq_stats->syncp, start));
212 	}
213 }
214 
215 static 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 
224 static int veth_set_channels(struct net_device *dev,
225 			     struct ethtool_channels *ch);
226 
227 static 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 
241 static bool veth_is_xdp_frame(void *ptr)
242 {
243 	return (unsigned long)ptr & VETH_XDP_FLAG;
244 }
245 
246 static struct xdp_frame *veth_ptr_to_xdp(void *ptr)
247 {
248 	return (void *)((unsigned long)ptr & ~VETH_XDP_FLAG);
249 }
250 
251 static void *veth_xdp_to_ptr(struct xdp_frame *xdp)
252 {
253 	return (void *)((unsigned long)xdp | VETH_XDP_FLAG);
254 }
255 
256 static 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 
264 static 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 
275 static 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 
285 static 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  */
303 static 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 
312 static 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 {
346 drop:
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 
358 static 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 
366 static 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_irq(&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_irq(&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 
397 static 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 */
432 static void veth_set_multicast_list(struct net_device *dev)
433 {
434 }
435 
436 static int veth_select_rxq(struct net_device *dev)
437 {
438 	return smp_processor_id() % dev->real_num_rx_queues;
439 }
440 
441 static 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 
449 static 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 
500 out:
501 	rcu_read_unlock();
502 
503 	return ret;
504 }
505 
506 static 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 
521 static 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 
545 static 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);
564 out:
565 	rcu_read_unlock();
566 }
567 
568 static 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 
584 static 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;
645 err_xdp:
646 	rcu_read_unlock();
647 	xdp_return_frame(frame);
648 xdp_xmit:
649 	return NULL;
650 }
651 
652 /* frames array contains VETH_XDP_BATCH at most */
653 static 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 
683 static 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 
696 static 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;
790 drop:
791 	consume_skb(skb);
792 	*pskb = NULL;
793 
794 	return -ENOMEM;
795 }
796 
797 static 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);
891 out:
892 	return skb;
893 drop:
894 	stats->rx_drops++;
895 xdp_drop:
896 	rcu_read_unlock();
897 	kfree_skb(skb);
898 	return NULL;
899 err_xdp:
900 	rcu_read_unlock();
901 	xdp_return_buff(&xdp);
902 xdp_xmit:
903 	return NULL;
904 }
905 
906 static 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 
964 static 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 (done < budget && napi_complete_done(napi, done)) {
978 		/* Write rx_notify_masked before reading ptr_ring */
979 		smp_store_mb(rq->rx_notify_masked, false);
980 		if (unlikely(!__ptr_ring_empty(&rq->xdp_ring))) {
981 			if (napi_schedule_prep(&rq->xdp_napi)) {
982 				WRITE_ONCE(rq->rx_notify_masked, true);
983 				__napi_schedule(&rq->xdp_napi);
984 			}
985 		}
986 	}
987 
988 	if (stats.xdp_tx > 0)
989 		veth_xdp_flush(rq, &bq);
990 	if (stats.xdp_redirect > 0)
991 		xdp_do_flush();
992 	xdp_clear_return_frame_no_direct();
993 
994 	return done;
995 }
996 
997 static int __veth_napi_enable_range(struct net_device *dev, int start, int end)
998 {
999 	struct veth_priv *priv = netdev_priv(dev);
1000 	int err, i;
1001 
1002 	for (i = start; i < end; i++) {
1003 		struct veth_rq *rq = &priv->rq[i];
1004 
1005 		err = ptr_ring_init(&rq->xdp_ring, VETH_RING_SIZE, GFP_KERNEL);
1006 		if (err)
1007 			goto err_xdp_ring;
1008 	}
1009 
1010 	for (i = start; i < end; i++) {
1011 		struct veth_rq *rq = &priv->rq[i];
1012 
1013 		napi_enable(&rq->xdp_napi);
1014 		rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1015 	}
1016 
1017 	return 0;
1018 
1019 err_xdp_ring:
1020 	for (i--; i >= start; i--)
1021 		ptr_ring_cleanup(&priv->rq[i].xdp_ring, veth_ptr_free);
1022 
1023 	return err;
1024 }
1025 
1026 static int __veth_napi_enable(struct net_device *dev)
1027 {
1028 	return __veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1029 }
1030 
1031 static 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 
1053 static void veth_napi_del(struct net_device *dev)
1054 {
1055 	veth_napi_del_range(dev, 0, dev->real_num_rx_queues);
1056 }
1057 
1058 static bool veth_gro_requested(const struct net_device *dev)
1059 {
1060 	return !!(dev->wanted_features & NETIF_F_GRO);
1061 }
1062 
1063 static int veth_enable_xdp_range(struct net_device *dev, int start, int end,
1064 				 bool napi_already_on)
1065 {
1066 	struct veth_priv *priv = netdev_priv(dev);
1067 	int err, i;
1068 
1069 	for (i = start; i < end; i++) {
1070 		struct veth_rq *rq = &priv->rq[i];
1071 
1072 		if (!napi_already_on)
1073 			netif_napi_add(dev, &rq->xdp_napi, veth_poll, NAPI_POLL_WEIGHT);
1074 		err = xdp_rxq_info_reg(&rq->xdp_rxq, dev, i, rq->xdp_napi.napi_id);
1075 		if (err < 0)
1076 			goto err_rxq_reg;
1077 
1078 		err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
1079 						 MEM_TYPE_PAGE_SHARED,
1080 						 NULL);
1081 		if (err < 0)
1082 			goto err_reg_mem;
1083 
1084 		/* Save original mem info as it can be overwritten */
1085 		rq->xdp_mem = rq->xdp_rxq.mem;
1086 	}
1087 	return 0;
1088 
1089 err_reg_mem:
1090 	xdp_rxq_info_unreg(&priv->rq[i].xdp_rxq);
1091 err_rxq_reg:
1092 	for (i--; i >= start; i--) {
1093 		struct veth_rq *rq = &priv->rq[i];
1094 
1095 		xdp_rxq_info_unreg(&rq->xdp_rxq);
1096 		if (!napi_already_on)
1097 			netif_napi_del(&rq->xdp_napi);
1098 	}
1099 
1100 	return err;
1101 }
1102 
1103 static void veth_disable_xdp_range(struct net_device *dev, int start, int end,
1104 				   bool delete_napi)
1105 {
1106 	struct veth_priv *priv = netdev_priv(dev);
1107 	int i;
1108 
1109 	for (i = start; i < end; i++) {
1110 		struct veth_rq *rq = &priv->rq[i];
1111 
1112 		rq->xdp_rxq.mem = rq->xdp_mem;
1113 		xdp_rxq_info_unreg(&rq->xdp_rxq);
1114 
1115 		if (delete_napi)
1116 			netif_napi_del(&rq->xdp_napi);
1117 	}
1118 }
1119 
1120 static int veth_enable_xdp(struct net_device *dev)
1121 {
1122 	bool napi_already_on = veth_gro_requested(dev) && (dev->flags & IFF_UP);
1123 	struct veth_priv *priv = netdev_priv(dev);
1124 	int err, i;
1125 
1126 	if (!xdp_rxq_info_is_reg(&priv->rq[0].xdp_rxq)) {
1127 		err = veth_enable_xdp_range(dev, 0, dev->real_num_rx_queues, napi_already_on);
1128 		if (err)
1129 			return err;
1130 
1131 		if (!napi_already_on) {
1132 			err = __veth_napi_enable(dev);
1133 			if (err) {
1134 				veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, true);
1135 				return err;
1136 			}
1137 
1138 			if (!veth_gro_requested(dev)) {
1139 				/* user-space did not require GRO, but adding XDP
1140 				 * is supposed to get GRO working
1141 				 */
1142 				dev->features |= NETIF_F_GRO;
1143 				netdev_features_change(dev);
1144 			}
1145 		}
1146 	}
1147 
1148 	for (i = 0; i < dev->real_num_rx_queues; i++) {
1149 		rcu_assign_pointer(priv->rq[i].xdp_prog, priv->_xdp_prog);
1150 		rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1151 	}
1152 
1153 	return 0;
1154 }
1155 
1156 static void veth_disable_xdp(struct net_device *dev)
1157 {
1158 	struct veth_priv *priv = netdev_priv(dev);
1159 	int i;
1160 
1161 	for (i = 0; i < dev->real_num_rx_queues; i++)
1162 		rcu_assign_pointer(priv->rq[i].xdp_prog, NULL);
1163 
1164 	if (!netif_running(dev) || !veth_gro_requested(dev)) {
1165 		veth_napi_del(dev);
1166 
1167 		/* if user-space did not require GRO, since adding XDP
1168 		 * enabled it, clear it now
1169 		 */
1170 		if (!veth_gro_requested(dev) && netif_running(dev)) {
1171 			dev->features &= ~NETIF_F_GRO;
1172 			netdev_features_change(dev);
1173 		}
1174 	}
1175 
1176 	veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, false);
1177 }
1178 
1179 static int veth_napi_enable_range(struct net_device *dev, int start, int end)
1180 {
1181 	struct veth_priv *priv = netdev_priv(dev);
1182 	int err, i;
1183 
1184 	for (i = start; i < end; i++) {
1185 		struct veth_rq *rq = &priv->rq[i];
1186 
1187 		netif_napi_add(dev, &rq->xdp_napi, veth_poll, NAPI_POLL_WEIGHT);
1188 	}
1189 
1190 	err = __veth_napi_enable_range(dev, start, end);
1191 	if (err) {
1192 		for (i = start; i < end; i++) {
1193 			struct veth_rq *rq = &priv->rq[i];
1194 
1195 			netif_napi_del(&rq->xdp_napi);
1196 		}
1197 		return err;
1198 	}
1199 	return err;
1200 }
1201 
1202 static int veth_napi_enable(struct net_device *dev)
1203 {
1204 	return veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1205 }
1206 
1207 static void veth_disable_range_safe(struct net_device *dev, int start, int end)
1208 {
1209 	struct veth_priv *priv = netdev_priv(dev);
1210 
1211 	if (start >= end)
1212 		return;
1213 
1214 	if (priv->_xdp_prog) {
1215 		veth_napi_del_range(dev, start, end);
1216 		veth_disable_xdp_range(dev, start, end, false);
1217 	} else if (veth_gro_requested(dev)) {
1218 		veth_napi_del_range(dev, start, end);
1219 	}
1220 }
1221 
1222 static int veth_enable_range_safe(struct net_device *dev, int start, int end)
1223 {
1224 	struct veth_priv *priv = netdev_priv(dev);
1225 	int err;
1226 
1227 	if (start >= end)
1228 		return 0;
1229 
1230 	if (priv->_xdp_prog) {
1231 		/* these channels are freshly initialized, napi is not on there even
1232 		 * when GRO is requeste
1233 		 */
1234 		err = veth_enable_xdp_range(dev, start, end, false);
1235 		if (err)
1236 			return err;
1237 
1238 		err = __veth_napi_enable_range(dev, start, end);
1239 		if (err) {
1240 			/* on error always delete the newly added napis */
1241 			veth_disable_xdp_range(dev, start, end, true);
1242 			return err;
1243 		}
1244 	} else if (veth_gro_requested(dev)) {
1245 		return veth_napi_enable_range(dev, start, end);
1246 	}
1247 	return 0;
1248 }
1249 
1250 static int veth_set_channels(struct net_device *dev,
1251 			     struct ethtool_channels *ch)
1252 {
1253 	struct veth_priv *priv = netdev_priv(dev);
1254 	unsigned int old_rx_count, new_rx_count;
1255 	struct veth_priv *peer_priv;
1256 	struct net_device *peer;
1257 	int err;
1258 
1259 	/* sanity check. Upper bounds are already enforced by the caller */
1260 	if (!ch->rx_count || !ch->tx_count)
1261 		return -EINVAL;
1262 
1263 	/* avoid braking XDP, if that is enabled */
1264 	peer = rtnl_dereference(priv->peer);
1265 	peer_priv = peer ? netdev_priv(peer) : NULL;
1266 	if (priv->_xdp_prog && peer && ch->rx_count < peer->real_num_tx_queues)
1267 		return -EINVAL;
1268 
1269 	if (peer && peer_priv && peer_priv->_xdp_prog && ch->tx_count > peer->real_num_rx_queues)
1270 		return -EINVAL;
1271 
1272 	old_rx_count = dev->real_num_rx_queues;
1273 	new_rx_count = ch->rx_count;
1274 	if (netif_running(dev)) {
1275 		/* turn device off */
1276 		netif_carrier_off(dev);
1277 		if (peer)
1278 			netif_carrier_off(peer);
1279 
1280 		/* try to allocate new resurces, as needed*/
1281 		err = veth_enable_range_safe(dev, old_rx_count, new_rx_count);
1282 		if (err)
1283 			goto out;
1284 	}
1285 
1286 	err = netif_set_real_num_rx_queues(dev, ch->rx_count);
1287 	if (err)
1288 		goto revert;
1289 
1290 	err = netif_set_real_num_tx_queues(dev, ch->tx_count);
1291 	if (err) {
1292 		int err2 = netif_set_real_num_rx_queues(dev, old_rx_count);
1293 
1294 		/* this error condition could happen only if rx and tx change
1295 		 * in opposite directions (e.g. tx nr raises, rx nr decreases)
1296 		 * and we can't do anything to fully restore the original
1297 		 * status
1298 		 */
1299 		if (err2)
1300 			pr_warn("Can't restore rx queues config %d -> %d %d",
1301 				new_rx_count, old_rx_count, err2);
1302 		else
1303 			goto revert;
1304 	}
1305 
1306 out:
1307 	if (netif_running(dev)) {
1308 		/* note that we need to swap the arguments WRT the enable part
1309 		 * to identify the range we have to disable
1310 		 */
1311 		veth_disable_range_safe(dev, new_rx_count, old_rx_count);
1312 		netif_carrier_on(dev);
1313 		if (peer)
1314 			netif_carrier_on(peer);
1315 	}
1316 	return err;
1317 
1318 revert:
1319 	new_rx_count = old_rx_count;
1320 	old_rx_count = ch->rx_count;
1321 	goto out;
1322 }
1323 
1324 static int veth_open(struct net_device *dev)
1325 {
1326 	struct veth_priv *priv = netdev_priv(dev);
1327 	struct net_device *peer = rtnl_dereference(priv->peer);
1328 	int err;
1329 
1330 	if (!peer)
1331 		return -ENOTCONN;
1332 
1333 	if (priv->_xdp_prog) {
1334 		err = veth_enable_xdp(dev);
1335 		if (err)
1336 			return err;
1337 	} else if (veth_gro_requested(dev)) {
1338 		err = veth_napi_enable(dev);
1339 		if (err)
1340 			return err;
1341 	}
1342 
1343 	if (peer->flags & IFF_UP) {
1344 		netif_carrier_on(dev);
1345 		netif_carrier_on(peer);
1346 	}
1347 
1348 	return 0;
1349 }
1350 
1351 static int veth_close(struct net_device *dev)
1352 {
1353 	struct veth_priv *priv = netdev_priv(dev);
1354 	struct net_device *peer = rtnl_dereference(priv->peer);
1355 
1356 	netif_carrier_off(dev);
1357 	if (peer)
1358 		netif_carrier_off(peer);
1359 
1360 	if (priv->_xdp_prog)
1361 		veth_disable_xdp(dev);
1362 	else if (veth_gro_requested(dev))
1363 		veth_napi_del(dev);
1364 
1365 	return 0;
1366 }
1367 
1368 static int is_valid_veth_mtu(int mtu)
1369 {
1370 	return mtu >= ETH_MIN_MTU && mtu <= ETH_MAX_MTU;
1371 }
1372 
1373 static int veth_alloc_queues(struct net_device *dev)
1374 {
1375 	struct veth_priv *priv = netdev_priv(dev);
1376 	int i;
1377 
1378 	priv->rq = kcalloc(dev->num_rx_queues, sizeof(*priv->rq), GFP_KERNEL_ACCOUNT);
1379 	if (!priv->rq)
1380 		return -ENOMEM;
1381 
1382 	for (i = 0; i < dev->num_rx_queues; i++) {
1383 		priv->rq[i].dev = dev;
1384 		u64_stats_init(&priv->rq[i].stats.syncp);
1385 	}
1386 
1387 	return 0;
1388 }
1389 
1390 static void veth_free_queues(struct net_device *dev)
1391 {
1392 	struct veth_priv *priv = netdev_priv(dev);
1393 
1394 	kfree(priv->rq);
1395 }
1396 
1397 static int veth_dev_init(struct net_device *dev)
1398 {
1399 	int err;
1400 
1401 	dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats);
1402 	if (!dev->lstats)
1403 		return -ENOMEM;
1404 
1405 	err = veth_alloc_queues(dev);
1406 	if (err) {
1407 		free_percpu(dev->lstats);
1408 		return err;
1409 	}
1410 
1411 	return 0;
1412 }
1413 
1414 static void veth_dev_free(struct net_device *dev)
1415 {
1416 	veth_free_queues(dev);
1417 	free_percpu(dev->lstats);
1418 }
1419 
1420 #ifdef CONFIG_NET_POLL_CONTROLLER
1421 static void veth_poll_controller(struct net_device *dev)
1422 {
1423 	/* veth only receives frames when its peer sends one
1424 	 * Since it has nothing to do with disabling irqs, we are guaranteed
1425 	 * never to have pending data when we poll for it so
1426 	 * there is nothing to do here.
1427 	 *
1428 	 * We need this though so netpoll recognizes us as an interface that
1429 	 * supports polling, which enables bridge devices in virt setups to
1430 	 * still use netconsole
1431 	 */
1432 }
1433 #endif	/* CONFIG_NET_POLL_CONTROLLER */
1434 
1435 static int veth_get_iflink(const struct net_device *dev)
1436 {
1437 	struct veth_priv *priv = netdev_priv(dev);
1438 	struct net_device *peer;
1439 	int iflink;
1440 
1441 	rcu_read_lock();
1442 	peer = rcu_dereference(priv->peer);
1443 	iflink = peer ? peer->ifindex : 0;
1444 	rcu_read_unlock();
1445 
1446 	return iflink;
1447 }
1448 
1449 static netdev_features_t veth_fix_features(struct net_device *dev,
1450 					   netdev_features_t features)
1451 {
1452 	struct veth_priv *priv = netdev_priv(dev);
1453 	struct net_device *peer;
1454 
1455 	peer = rtnl_dereference(priv->peer);
1456 	if (peer) {
1457 		struct veth_priv *peer_priv = netdev_priv(peer);
1458 
1459 		if (peer_priv->_xdp_prog)
1460 			features &= ~NETIF_F_GSO_SOFTWARE;
1461 	}
1462 	if (priv->_xdp_prog)
1463 		features |= NETIF_F_GRO;
1464 
1465 	return features;
1466 }
1467 
1468 static int veth_set_features(struct net_device *dev,
1469 			     netdev_features_t features)
1470 {
1471 	netdev_features_t changed = features ^ dev->features;
1472 	struct veth_priv *priv = netdev_priv(dev);
1473 	int err;
1474 
1475 	if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog)
1476 		return 0;
1477 
1478 	if (features & NETIF_F_GRO) {
1479 		err = veth_napi_enable(dev);
1480 		if (err)
1481 			return err;
1482 	} else {
1483 		veth_napi_del(dev);
1484 	}
1485 	return 0;
1486 }
1487 
1488 static void veth_set_rx_headroom(struct net_device *dev, int new_hr)
1489 {
1490 	struct veth_priv *peer_priv, *priv = netdev_priv(dev);
1491 	struct net_device *peer;
1492 
1493 	if (new_hr < 0)
1494 		new_hr = 0;
1495 
1496 	rcu_read_lock();
1497 	peer = rcu_dereference(priv->peer);
1498 	if (unlikely(!peer))
1499 		goto out;
1500 
1501 	peer_priv = netdev_priv(peer);
1502 	priv->requested_headroom = new_hr;
1503 	new_hr = max(priv->requested_headroom, peer_priv->requested_headroom);
1504 	dev->needed_headroom = new_hr;
1505 	peer->needed_headroom = new_hr;
1506 
1507 out:
1508 	rcu_read_unlock();
1509 }
1510 
1511 static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1512 			struct netlink_ext_ack *extack)
1513 {
1514 	struct veth_priv *priv = netdev_priv(dev);
1515 	struct bpf_prog *old_prog;
1516 	struct net_device *peer;
1517 	unsigned int max_mtu;
1518 	int err;
1519 
1520 	old_prog = priv->_xdp_prog;
1521 	priv->_xdp_prog = prog;
1522 	peer = rtnl_dereference(priv->peer);
1523 
1524 	if (prog) {
1525 		if (!peer) {
1526 			NL_SET_ERR_MSG_MOD(extack, "Cannot set XDP when peer is detached");
1527 			err = -ENOTCONN;
1528 			goto err;
1529 		}
1530 
1531 		max_mtu = SKB_WITH_OVERHEAD(PAGE_SIZE - VETH_XDP_HEADROOM) -
1532 			  peer->hard_header_len;
1533 		/* Allow increasing the max_mtu if the program supports
1534 		 * XDP fragments.
1535 		 */
1536 		if (prog->aux->xdp_has_frags)
1537 			max_mtu += PAGE_SIZE * MAX_SKB_FRAGS;
1538 
1539 		if (peer->mtu > max_mtu) {
1540 			NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP");
1541 			err = -ERANGE;
1542 			goto err;
1543 		}
1544 
1545 		if (dev->real_num_rx_queues < peer->real_num_tx_queues) {
1546 			NL_SET_ERR_MSG_MOD(extack, "XDP expects number of rx queues not less than peer tx queues");
1547 			err = -ENOSPC;
1548 			goto err;
1549 		}
1550 
1551 		if (dev->flags & IFF_UP) {
1552 			err = veth_enable_xdp(dev);
1553 			if (err) {
1554 				NL_SET_ERR_MSG_MOD(extack, "Setup for XDP failed");
1555 				goto err;
1556 			}
1557 		}
1558 
1559 		if (!old_prog) {
1560 			peer->hw_features &= ~NETIF_F_GSO_SOFTWARE;
1561 			peer->max_mtu = max_mtu;
1562 		}
1563 	}
1564 
1565 	if (old_prog) {
1566 		if (!prog) {
1567 			if (dev->flags & IFF_UP)
1568 				veth_disable_xdp(dev);
1569 
1570 			if (peer) {
1571 				peer->hw_features |= NETIF_F_GSO_SOFTWARE;
1572 				peer->max_mtu = ETH_MAX_MTU;
1573 			}
1574 		}
1575 		bpf_prog_put(old_prog);
1576 	}
1577 
1578 	if ((!!old_prog ^ !!prog) && peer)
1579 		netdev_update_features(peer);
1580 
1581 	return 0;
1582 err:
1583 	priv->_xdp_prog = old_prog;
1584 
1585 	return err;
1586 }
1587 
1588 static int veth_xdp(struct net_device *dev, struct netdev_bpf *xdp)
1589 {
1590 	switch (xdp->command) {
1591 	case XDP_SETUP_PROG:
1592 		return veth_xdp_set(dev, xdp->prog, xdp->extack);
1593 	default:
1594 		return -EINVAL;
1595 	}
1596 }
1597 
1598 static const struct net_device_ops veth_netdev_ops = {
1599 	.ndo_init            = veth_dev_init,
1600 	.ndo_open            = veth_open,
1601 	.ndo_stop            = veth_close,
1602 	.ndo_start_xmit      = veth_xmit,
1603 	.ndo_get_stats64     = veth_get_stats64,
1604 	.ndo_set_rx_mode     = veth_set_multicast_list,
1605 	.ndo_set_mac_address = eth_mac_addr,
1606 #ifdef CONFIG_NET_POLL_CONTROLLER
1607 	.ndo_poll_controller	= veth_poll_controller,
1608 #endif
1609 	.ndo_get_iflink		= veth_get_iflink,
1610 	.ndo_fix_features	= veth_fix_features,
1611 	.ndo_set_features	= veth_set_features,
1612 	.ndo_features_check	= passthru_features_check,
1613 	.ndo_set_rx_headroom	= veth_set_rx_headroom,
1614 	.ndo_bpf		= veth_xdp,
1615 	.ndo_xdp_xmit		= veth_ndo_xdp_xmit,
1616 	.ndo_get_peer_dev	= veth_peer_dev,
1617 };
1618 
1619 #define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HW_CSUM | \
1620 		       NETIF_F_RXCSUM | NETIF_F_SCTP_CRC | NETIF_F_HIGHDMA | \
1621 		       NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \
1622 		       NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX | \
1623 		       NETIF_F_HW_VLAN_STAG_TX | NETIF_F_HW_VLAN_STAG_RX )
1624 
1625 static void veth_setup(struct net_device *dev)
1626 {
1627 	ether_setup(dev);
1628 
1629 	dev->priv_flags &= ~IFF_TX_SKB_SHARING;
1630 	dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
1631 	dev->priv_flags |= IFF_NO_QUEUE;
1632 	dev->priv_flags |= IFF_PHONY_HEADROOM;
1633 
1634 	dev->netdev_ops = &veth_netdev_ops;
1635 	dev->ethtool_ops = &veth_ethtool_ops;
1636 	dev->features |= NETIF_F_LLTX;
1637 	dev->features |= VETH_FEATURES;
1638 	dev->vlan_features = dev->features &
1639 			     ~(NETIF_F_HW_VLAN_CTAG_TX |
1640 			       NETIF_F_HW_VLAN_STAG_TX |
1641 			       NETIF_F_HW_VLAN_CTAG_RX |
1642 			       NETIF_F_HW_VLAN_STAG_RX);
1643 	dev->needs_free_netdev = true;
1644 	dev->priv_destructor = veth_dev_free;
1645 	dev->max_mtu = ETH_MAX_MTU;
1646 
1647 	dev->hw_features = VETH_FEATURES;
1648 	dev->hw_enc_features = VETH_FEATURES;
1649 	dev->mpls_features = NETIF_F_HW_CSUM | NETIF_F_GSO_SOFTWARE;
1650 	netif_set_tso_max_size(dev, GSO_MAX_SIZE);
1651 }
1652 
1653 /*
1654  * netlink interface
1655  */
1656 
1657 static int veth_validate(struct nlattr *tb[], struct nlattr *data[],
1658 			 struct netlink_ext_ack *extack)
1659 {
1660 	if (tb[IFLA_ADDRESS]) {
1661 		if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
1662 			return -EINVAL;
1663 		if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
1664 			return -EADDRNOTAVAIL;
1665 	}
1666 	if (tb[IFLA_MTU]) {
1667 		if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU])))
1668 			return -EINVAL;
1669 	}
1670 	return 0;
1671 }
1672 
1673 static struct rtnl_link_ops veth_link_ops;
1674 
1675 static void veth_disable_gro(struct net_device *dev)
1676 {
1677 	dev->features &= ~NETIF_F_GRO;
1678 	dev->wanted_features &= ~NETIF_F_GRO;
1679 	netdev_update_features(dev);
1680 }
1681 
1682 static int veth_init_queues(struct net_device *dev, struct nlattr *tb[])
1683 {
1684 	int err;
1685 
1686 	if (!tb[IFLA_NUM_TX_QUEUES] && dev->num_tx_queues > 1) {
1687 		err = netif_set_real_num_tx_queues(dev, 1);
1688 		if (err)
1689 			return err;
1690 	}
1691 	if (!tb[IFLA_NUM_RX_QUEUES] && dev->num_rx_queues > 1) {
1692 		err = netif_set_real_num_rx_queues(dev, 1);
1693 		if (err)
1694 			return err;
1695 	}
1696 	return 0;
1697 }
1698 
1699 static int veth_newlink(struct net *src_net, struct net_device *dev,
1700 			struct nlattr *tb[], struct nlattr *data[],
1701 			struct netlink_ext_ack *extack)
1702 {
1703 	int err;
1704 	struct net_device *peer;
1705 	struct veth_priv *priv;
1706 	char ifname[IFNAMSIZ];
1707 	struct nlattr *peer_tb[IFLA_MAX + 1], **tbp;
1708 	unsigned char name_assign_type;
1709 	struct ifinfomsg *ifmp;
1710 	struct net *net;
1711 
1712 	/*
1713 	 * create and register peer first
1714 	 */
1715 	if (data != NULL && data[VETH_INFO_PEER] != NULL) {
1716 		struct nlattr *nla_peer;
1717 
1718 		nla_peer = data[VETH_INFO_PEER];
1719 		ifmp = nla_data(nla_peer);
1720 		err = rtnl_nla_parse_ifla(peer_tb,
1721 					  nla_data(nla_peer) + sizeof(struct ifinfomsg),
1722 					  nla_len(nla_peer) - sizeof(struct ifinfomsg),
1723 					  NULL);
1724 		if (err < 0)
1725 			return err;
1726 
1727 		err = veth_validate(peer_tb, NULL, extack);
1728 		if (err < 0)
1729 			return err;
1730 
1731 		tbp = peer_tb;
1732 	} else {
1733 		ifmp = NULL;
1734 		tbp = tb;
1735 	}
1736 
1737 	if (ifmp && tbp[IFLA_IFNAME]) {
1738 		nla_strscpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ);
1739 		name_assign_type = NET_NAME_USER;
1740 	} else {
1741 		snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d");
1742 		name_assign_type = NET_NAME_ENUM;
1743 	}
1744 
1745 	net = rtnl_link_get_net(src_net, tbp);
1746 	if (IS_ERR(net))
1747 		return PTR_ERR(net);
1748 
1749 	peer = rtnl_create_link(net, ifname, name_assign_type,
1750 				&veth_link_ops, tbp, extack);
1751 	if (IS_ERR(peer)) {
1752 		put_net(net);
1753 		return PTR_ERR(peer);
1754 	}
1755 
1756 	if (!ifmp || !tbp[IFLA_ADDRESS])
1757 		eth_hw_addr_random(peer);
1758 
1759 	if (ifmp && (dev->ifindex != 0))
1760 		peer->ifindex = ifmp->ifi_index;
1761 
1762 	netif_inherit_tso_max(peer, dev);
1763 
1764 	err = register_netdevice(peer);
1765 	put_net(net);
1766 	net = NULL;
1767 	if (err < 0)
1768 		goto err_register_peer;
1769 
1770 	/* keep GRO disabled by default to be consistent with the established
1771 	 * veth behavior
1772 	 */
1773 	veth_disable_gro(peer);
1774 	netif_carrier_off(peer);
1775 
1776 	err = rtnl_configure_link(peer, ifmp);
1777 	if (err < 0)
1778 		goto err_configure_peer;
1779 
1780 	/*
1781 	 * register dev last
1782 	 *
1783 	 * note, that since we've registered new device the dev's name
1784 	 * should be re-allocated
1785 	 */
1786 
1787 	if (tb[IFLA_ADDRESS] == NULL)
1788 		eth_hw_addr_random(dev);
1789 
1790 	if (tb[IFLA_IFNAME])
1791 		nla_strscpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ);
1792 	else
1793 		snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d");
1794 
1795 	err = register_netdevice(dev);
1796 	if (err < 0)
1797 		goto err_register_dev;
1798 
1799 	netif_carrier_off(dev);
1800 
1801 	/*
1802 	 * tie the deviced together
1803 	 */
1804 
1805 	priv = netdev_priv(dev);
1806 	rcu_assign_pointer(priv->peer, peer);
1807 	err = veth_init_queues(dev, tb);
1808 	if (err)
1809 		goto err_queues;
1810 
1811 	priv = netdev_priv(peer);
1812 	rcu_assign_pointer(priv->peer, dev);
1813 	err = veth_init_queues(peer, tb);
1814 	if (err)
1815 		goto err_queues;
1816 
1817 	veth_disable_gro(dev);
1818 	return 0;
1819 
1820 err_queues:
1821 	unregister_netdevice(dev);
1822 err_register_dev:
1823 	/* nothing to do */
1824 err_configure_peer:
1825 	unregister_netdevice(peer);
1826 	return err;
1827 
1828 err_register_peer:
1829 	free_netdev(peer);
1830 	return err;
1831 }
1832 
1833 static void veth_dellink(struct net_device *dev, struct list_head *head)
1834 {
1835 	struct veth_priv *priv;
1836 	struct net_device *peer;
1837 
1838 	priv = netdev_priv(dev);
1839 	peer = rtnl_dereference(priv->peer);
1840 
1841 	/* Note : dellink() is called from default_device_exit_batch(),
1842 	 * before a rcu_synchronize() point. The devices are guaranteed
1843 	 * not being freed before one RCU grace period.
1844 	 */
1845 	RCU_INIT_POINTER(priv->peer, NULL);
1846 	unregister_netdevice_queue(dev, head);
1847 
1848 	if (peer) {
1849 		priv = netdev_priv(peer);
1850 		RCU_INIT_POINTER(priv->peer, NULL);
1851 		unregister_netdevice_queue(peer, head);
1852 	}
1853 }
1854 
1855 static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = {
1856 	[VETH_INFO_PEER]	= { .len = sizeof(struct ifinfomsg) },
1857 };
1858 
1859 static struct net *veth_get_link_net(const struct net_device *dev)
1860 {
1861 	struct veth_priv *priv = netdev_priv(dev);
1862 	struct net_device *peer = rtnl_dereference(priv->peer);
1863 
1864 	return peer ? dev_net(peer) : dev_net(dev);
1865 }
1866 
1867 static unsigned int veth_get_num_queues(void)
1868 {
1869 	/* enforce the same queue limit as rtnl_create_link */
1870 	int queues = num_possible_cpus();
1871 
1872 	if (queues > 4096)
1873 		queues = 4096;
1874 	return queues;
1875 }
1876 
1877 static struct rtnl_link_ops veth_link_ops = {
1878 	.kind		= DRV_NAME,
1879 	.priv_size	= sizeof(struct veth_priv),
1880 	.setup		= veth_setup,
1881 	.validate	= veth_validate,
1882 	.newlink	= veth_newlink,
1883 	.dellink	= veth_dellink,
1884 	.policy		= veth_policy,
1885 	.maxtype	= VETH_INFO_MAX,
1886 	.get_link_net	= veth_get_link_net,
1887 	.get_num_tx_queues	= veth_get_num_queues,
1888 	.get_num_rx_queues	= veth_get_num_queues,
1889 };
1890 
1891 /*
1892  * init/fini
1893  */
1894 
1895 static __init int veth_init(void)
1896 {
1897 	return rtnl_link_register(&veth_link_ops);
1898 }
1899 
1900 static __exit void veth_exit(void)
1901 {
1902 	rtnl_link_unregister(&veth_link_ops);
1903 }
1904 
1905 module_init(veth_init);
1906 module_exit(veth_exit);
1907 
1908 MODULE_DESCRIPTION("Virtual Ethernet Tunnel");
1909 MODULE_LICENSE("GPL v2");
1910 MODULE_ALIAS_RTNL_LINK(DRV_NAME);
1911