xref: /openbmc/linux/drivers/net/veth.c (revision 41415b8a)
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)) {
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 struct sk_buff *veth_build_skb(void *head, int headroom, int len,
437 				      int buflen)
438 {
439 	struct sk_buff *skb;
440 
441 	skb = build_skb(head, buflen);
442 	if (!skb)
443 		return NULL;
444 
445 	skb_reserve(skb, headroom);
446 	skb_put(skb, len);
447 
448 	return skb;
449 }
450 
451 static int veth_select_rxq(struct net_device *dev)
452 {
453 	return smp_processor_id() % dev->real_num_rx_queues;
454 }
455 
456 static struct net_device *veth_peer_dev(struct net_device *dev)
457 {
458 	struct veth_priv *priv = netdev_priv(dev);
459 
460 	/* Callers must be under RCU read side. */
461 	return rcu_dereference(priv->peer);
462 }
463 
464 static int veth_xdp_xmit(struct net_device *dev, int n,
465 			 struct xdp_frame **frames,
466 			 u32 flags, bool ndo_xmit)
467 {
468 	struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
469 	int i, ret = -ENXIO, nxmit = 0;
470 	struct net_device *rcv;
471 	unsigned int max_len;
472 	struct veth_rq *rq;
473 
474 	if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
475 		return -EINVAL;
476 
477 	rcu_read_lock();
478 	rcv = rcu_dereference(priv->peer);
479 	if (unlikely(!rcv))
480 		goto out;
481 
482 	rcv_priv = netdev_priv(rcv);
483 	rq = &rcv_priv->rq[veth_select_rxq(rcv)];
484 	/* The napi pointer is set if NAPI is enabled, which ensures that
485 	 * xdp_ring is initialized on receive side and the peer device is up.
486 	 */
487 	if (!rcu_access_pointer(rq->napi))
488 		goto out;
489 
490 	max_len = rcv->mtu + rcv->hard_header_len + VLAN_HLEN;
491 
492 	spin_lock(&rq->xdp_ring.producer_lock);
493 	for (i = 0; i < n; i++) {
494 		struct xdp_frame *frame = frames[i];
495 		void *ptr = veth_xdp_to_ptr(frame);
496 
497 		if (unlikely(frame->len > max_len ||
498 			     __ptr_ring_produce(&rq->xdp_ring, ptr)))
499 			break;
500 		nxmit++;
501 	}
502 	spin_unlock(&rq->xdp_ring.producer_lock);
503 
504 	if (flags & XDP_XMIT_FLUSH)
505 		__veth_xdp_flush(rq);
506 
507 	ret = nxmit;
508 	if (ndo_xmit) {
509 		u64_stats_update_begin(&rq->stats.syncp);
510 		rq->stats.vs.peer_tq_xdp_xmit += nxmit;
511 		rq->stats.vs.peer_tq_xdp_xmit_err += n - nxmit;
512 		u64_stats_update_end(&rq->stats.syncp);
513 	}
514 
515 out:
516 	rcu_read_unlock();
517 
518 	return ret;
519 }
520 
521 static int veth_ndo_xdp_xmit(struct net_device *dev, int n,
522 			     struct xdp_frame **frames, u32 flags)
523 {
524 	int err;
525 
526 	err = veth_xdp_xmit(dev, n, frames, flags, true);
527 	if (err < 0) {
528 		struct veth_priv *priv = netdev_priv(dev);
529 
530 		atomic64_add(n, &priv->dropped);
531 	}
532 
533 	return err;
534 }
535 
536 static void veth_xdp_flush_bq(struct veth_rq *rq, struct veth_xdp_tx_bq *bq)
537 {
538 	int sent, i, err = 0, drops;
539 
540 	sent = veth_xdp_xmit(rq->dev, bq->count, bq->q, 0, false);
541 	if (sent < 0) {
542 		err = sent;
543 		sent = 0;
544 	}
545 
546 	for (i = sent; unlikely(i < bq->count); i++)
547 		xdp_return_frame(bq->q[i]);
548 
549 	drops = bq->count - sent;
550 	trace_xdp_bulk_tx(rq->dev, sent, drops, err);
551 
552 	u64_stats_update_begin(&rq->stats.syncp);
553 	rq->stats.vs.xdp_tx += sent;
554 	rq->stats.vs.xdp_tx_err += drops;
555 	u64_stats_update_end(&rq->stats.syncp);
556 
557 	bq->count = 0;
558 }
559 
560 static void veth_xdp_flush(struct veth_rq *rq, struct veth_xdp_tx_bq *bq)
561 {
562 	struct veth_priv *rcv_priv, *priv = netdev_priv(rq->dev);
563 	struct net_device *rcv;
564 	struct veth_rq *rcv_rq;
565 
566 	rcu_read_lock();
567 	veth_xdp_flush_bq(rq, bq);
568 	rcv = rcu_dereference(priv->peer);
569 	if (unlikely(!rcv))
570 		goto out;
571 
572 	rcv_priv = netdev_priv(rcv);
573 	rcv_rq = &rcv_priv->rq[veth_select_rxq(rcv)];
574 	/* xdp_ring is initialized on receive side? */
575 	if (unlikely(!rcu_access_pointer(rcv_rq->xdp_prog)))
576 		goto out;
577 
578 	__veth_xdp_flush(rcv_rq);
579 out:
580 	rcu_read_unlock();
581 }
582 
583 static int veth_xdp_tx(struct veth_rq *rq, struct xdp_buff *xdp,
584 		       struct veth_xdp_tx_bq *bq)
585 {
586 	struct xdp_frame *frame = xdp_convert_buff_to_frame(xdp);
587 
588 	if (unlikely(!frame))
589 		return -EOVERFLOW;
590 
591 	if (unlikely(bq->count == VETH_XDP_TX_BULK_SIZE))
592 		veth_xdp_flush_bq(rq, bq);
593 
594 	bq->q[bq->count++] = frame;
595 
596 	return 0;
597 }
598 
599 static struct xdp_frame *veth_xdp_rcv_one(struct veth_rq *rq,
600 					  struct xdp_frame *frame,
601 					  struct veth_xdp_tx_bq *bq,
602 					  struct veth_stats *stats)
603 {
604 	struct xdp_frame orig_frame;
605 	struct bpf_prog *xdp_prog;
606 
607 	rcu_read_lock();
608 	xdp_prog = rcu_dereference(rq->xdp_prog);
609 	if (likely(xdp_prog)) {
610 		struct xdp_buff xdp;
611 		u32 act;
612 
613 		xdp_convert_frame_to_buff(frame, &xdp);
614 		xdp.rxq = &rq->xdp_rxq;
615 
616 		act = bpf_prog_run_xdp(xdp_prog, &xdp);
617 
618 		switch (act) {
619 		case XDP_PASS:
620 			if (xdp_update_frame_from_buff(&xdp, frame))
621 				goto err_xdp;
622 			break;
623 		case XDP_TX:
624 			orig_frame = *frame;
625 			xdp.rxq->mem = frame->mem;
626 			if (unlikely(veth_xdp_tx(rq, &xdp, bq) < 0)) {
627 				trace_xdp_exception(rq->dev, xdp_prog, act);
628 				frame = &orig_frame;
629 				stats->rx_drops++;
630 				goto err_xdp;
631 			}
632 			stats->xdp_tx++;
633 			rcu_read_unlock();
634 			goto xdp_xmit;
635 		case XDP_REDIRECT:
636 			orig_frame = *frame;
637 			xdp.rxq->mem = frame->mem;
638 			if (xdp_do_redirect(rq->dev, &xdp, xdp_prog)) {
639 				frame = &orig_frame;
640 				stats->rx_drops++;
641 				goto err_xdp;
642 			}
643 			stats->xdp_redirect++;
644 			rcu_read_unlock();
645 			goto xdp_xmit;
646 		default:
647 			bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act);
648 			fallthrough;
649 		case XDP_ABORTED:
650 			trace_xdp_exception(rq->dev, xdp_prog, act);
651 			fallthrough;
652 		case XDP_DROP:
653 			stats->xdp_drops++;
654 			goto err_xdp;
655 		}
656 	}
657 	rcu_read_unlock();
658 
659 	return frame;
660 err_xdp:
661 	rcu_read_unlock();
662 	xdp_return_frame(frame);
663 xdp_xmit:
664 	return NULL;
665 }
666 
667 /* frames array contains VETH_XDP_BATCH at most */
668 static void veth_xdp_rcv_bulk_skb(struct veth_rq *rq, void **frames,
669 				  int n_xdpf, struct veth_xdp_tx_bq *bq,
670 				  struct veth_stats *stats)
671 {
672 	void *skbs[VETH_XDP_BATCH];
673 	int i;
674 
675 	if (xdp_alloc_skb_bulk(skbs, n_xdpf,
676 			       GFP_ATOMIC | __GFP_ZERO) < 0) {
677 		for (i = 0; i < n_xdpf; i++)
678 			xdp_return_frame(frames[i]);
679 		stats->rx_drops += n_xdpf;
680 
681 		return;
682 	}
683 
684 	for (i = 0; i < n_xdpf; i++) {
685 		struct sk_buff *skb = skbs[i];
686 
687 		skb = __xdp_build_skb_from_frame(frames[i], skb,
688 						 rq->dev);
689 		if (!skb) {
690 			xdp_return_frame(frames[i]);
691 			stats->rx_drops++;
692 			continue;
693 		}
694 		napi_gro_receive(&rq->xdp_napi, skb);
695 	}
696 }
697 
698 static struct sk_buff *veth_xdp_rcv_skb(struct veth_rq *rq,
699 					struct sk_buff *skb,
700 					struct veth_xdp_tx_bq *bq,
701 					struct veth_stats *stats)
702 {
703 	u32 pktlen, headroom, act, metalen, frame_sz;
704 	void *orig_data, *orig_data_end;
705 	struct bpf_prog *xdp_prog;
706 	int mac_len, delta, off;
707 	struct xdp_buff xdp;
708 
709 	skb_prepare_for_gro(skb);
710 
711 	rcu_read_lock();
712 	xdp_prog = rcu_dereference(rq->xdp_prog);
713 	if (unlikely(!xdp_prog)) {
714 		rcu_read_unlock();
715 		goto out;
716 	}
717 
718 	mac_len = skb->data - skb_mac_header(skb);
719 	pktlen = skb->len + mac_len;
720 	headroom = skb_headroom(skb) - mac_len;
721 
722 	if (skb_shared(skb) || skb_head_is_locked(skb) ||
723 	    skb_is_nonlinear(skb) || headroom < XDP_PACKET_HEADROOM) {
724 		struct sk_buff *nskb;
725 		int size, head_off;
726 		void *head, *start;
727 		struct page *page;
728 
729 		size = SKB_DATA_ALIGN(VETH_XDP_HEADROOM + pktlen) +
730 		       SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
731 		if (size > PAGE_SIZE)
732 			goto drop;
733 
734 		page = alloc_page(GFP_ATOMIC | __GFP_NOWARN);
735 		if (!page)
736 			goto drop;
737 
738 		head = page_address(page);
739 		start = head + VETH_XDP_HEADROOM;
740 		if (skb_copy_bits(skb, -mac_len, start, pktlen)) {
741 			page_frag_free(head);
742 			goto drop;
743 		}
744 
745 		nskb = veth_build_skb(head, VETH_XDP_HEADROOM + mac_len,
746 				      skb->len, PAGE_SIZE);
747 		if (!nskb) {
748 			page_frag_free(head);
749 			goto drop;
750 		}
751 
752 		skb_copy_header(nskb, skb);
753 		head_off = skb_headroom(nskb) - skb_headroom(skb);
754 		skb_headers_offset_update(nskb, head_off);
755 		consume_skb(skb);
756 		skb = nskb;
757 	}
758 
759 	/* SKB "head" area always have tailroom for skb_shared_info */
760 	frame_sz = skb_end_pointer(skb) - skb->head;
761 	frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
762 	xdp_init_buff(&xdp, frame_sz, &rq->xdp_rxq);
763 	xdp_prepare_buff(&xdp, skb->head, skb->mac_header, pktlen, true);
764 
765 	orig_data = xdp.data;
766 	orig_data_end = xdp.data_end;
767 
768 	act = bpf_prog_run_xdp(xdp_prog, &xdp);
769 
770 	switch (act) {
771 	case XDP_PASS:
772 		break;
773 	case XDP_TX:
774 		get_page(virt_to_page(xdp.data));
775 		consume_skb(skb);
776 		xdp.rxq->mem = rq->xdp_mem;
777 		if (unlikely(veth_xdp_tx(rq, &xdp, bq) < 0)) {
778 			trace_xdp_exception(rq->dev, xdp_prog, act);
779 			stats->rx_drops++;
780 			goto err_xdp;
781 		}
782 		stats->xdp_tx++;
783 		rcu_read_unlock();
784 		goto xdp_xmit;
785 	case XDP_REDIRECT:
786 		get_page(virt_to_page(xdp.data));
787 		consume_skb(skb);
788 		xdp.rxq->mem = rq->xdp_mem;
789 		if (xdp_do_redirect(rq->dev, &xdp, xdp_prog)) {
790 			stats->rx_drops++;
791 			goto err_xdp;
792 		}
793 		stats->xdp_redirect++;
794 		rcu_read_unlock();
795 		goto xdp_xmit;
796 	default:
797 		bpf_warn_invalid_xdp_action(rq->dev, xdp_prog, act);
798 		fallthrough;
799 	case XDP_ABORTED:
800 		trace_xdp_exception(rq->dev, xdp_prog, act);
801 		fallthrough;
802 	case XDP_DROP:
803 		stats->xdp_drops++;
804 		goto xdp_drop;
805 	}
806 	rcu_read_unlock();
807 
808 	/* check if bpf_xdp_adjust_head was used */
809 	delta = orig_data - xdp.data;
810 	off = mac_len + delta;
811 	if (off > 0)
812 		__skb_push(skb, off);
813 	else if (off < 0)
814 		__skb_pull(skb, -off);
815 	skb->mac_header -= delta;
816 
817 	/* check if bpf_xdp_adjust_tail was used */
818 	off = xdp.data_end - orig_data_end;
819 	if (off != 0)
820 		__skb_put(skb, off); /* positive on grow, negative on shrink */
821 	skb->protocol = eth_type_trans(skb, rq->dev);
822 
823 	metalen = xdp.data - xdp.data_meta;
824 	if (metalen)
825 		skb_metadata_set(skb, metalen);
826 out:
827 	return skb;
828 drop:
829 	stats->rx_drops++;
830 xdp_drop:
831 	rcu_read_unlock();
832 	kfree_skb(skb);
833 	return NULL;
834 err_xdp:
835 	rcu_read_unlock();
836 	page_frag_free(xdp.data);
837 xdp_xmit:
838 	return NULL;
839 }
840 
841 static int veth_xdp_rcv(struct veth_rq *rq, int budget,
842 			struct veth_xdp_tx_bq *bq,
843 			struct veth_stats *stats)
844 {
845 	int i, done = 0, n_xdpf = 0;
846 	void *xdpf[VETH_XDP_BATCH];
847 
848 	for (i = 0; i < budget; i++) {
849 		void *ptr = __ptr_ring_consume(&rq->xdp_ring);
850 
851 		if (!ptr)
852 			break;
853 
854 		if (veth_is_xdp_frame(ptr)) {
855 			/* ndo_xdp_xmit */
856 			struct xdp_frame *frame = veth_ptr_to_xdp(ptr);
857 
858 			stats->xdp_bytes += frame->len;
859 			frame = veth_xdp_rcv_one(rq, frame, bq, stats);
860 			if (frame) {
861 				/* XDP_PASS */
862 				xdpf[n_xdpf++] = frame;
863 				if (n_xdpf == VETH_XDP_BATCH) {
864 					veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf,
865 							      bq, stats);
866 					n_xdpf = 0;
867 				}
868 			}
869 		} else {
870 			/* ndo_start_xmit */
871 			struct sk_buff *skb = ptr;
872 
873 			stats->xdp_bytes += skb->len;
874 			skb = veth_xdp_rcv_skb(rq, skb, bq, stats);
875 			if (skb) {
876 				if (skb_shared(skb) || skb_unclone(skb, GFP_ATOMIC))
877 					netif_receive_skb(skb);
878 				else
879 					napi_gro_receive(&rq->xdp_napi, skb);
880 			}
881 		}
882 		done++;
883 	}
884 
885 	if (n_xdpf)
886 		veth_xdp_rcv_bulk_skb(rq, xdpf, n_xdpf, bq, stats);
887 
888 	u64_stats_update_begin(&rq->stats.syncp);
889 	rq->stats.vs.xdp_redirect += stats->xdp_redirect;
890 	rq->stats.vs.xdp_bytes += stats->xdp_bytes;
891 	rq->stats.vs.xdp_drops += stats->xdp_drops;
892 	rq->stats.vs.rx_drops += stats->rx_drops;
893 	rq->stats.vs.xdp_packets += done;
894 	u64_stats_update_end(&rq->stats.syncp);
895 
896 	return done;
897 }
898 
899 static int veth_poll(struct napi_struct *napi, int budget)
900 {
901 	struct veth_rq *rq =
902 		container_of(napi, struct veth_rq, xdp_napi);
903 	struct veth_stats stats = {};
904 	struct veth_xdp_tx_bq bq;
905 	int done;
906 
907 	bq.count = 0;
908 
909 	xdp_set_return_frame_no_direct();
910 	done = veth_xdp_rcv(rq, budget, &bq, &stats);
911 
912 	if (done < budget && napi_complete_done(napi, done)) {
913 		/* Write rx_notify_masked before reading ptr_ring */
914 		smp_store_mb(rq->rx_notify_masked, false);
915 		if (unlikely(!__ptr_ring_empty(&rq->xdp_ring))) {
916 			if (napi_schedule_prep(&rq->xdp_napi)) {
917 				WRITE_ONCE(rq->rx_notify_masked, true);
918 				__napi_schedule(&rq->xdp_napi);
919 			}
920 		}
921 	}
922 
923 	if (stats.xdp_tx > 0)
924 		veth_xdp_flush(rq, &bq);
925 	if (stats.xdp_redirect > 0)
926 		xdp_do_flush();
927 	xdp_clear_return_frame_no_direct();
928 
929 	return done;
930 }
931 
932 static int __veth_napi_enable_range(struct net_device *dev, int start, int end)
933 {
934 	struct veth_priv *priv = netdev_priv(dev);
935 	int err, i;
936 
937 	for (i = start; i < end; i++) {
938 		struct veth_rq *rq = &priv->rq[i];
939 
940 		err = ptr_ring_init(&rq->xdp_ring, VETH_RING_SIZE, GFP_KERNEL);
941 		if (err)
942 			goto err_xdp_ring;
943 	}
944 
945 	for (i = start; i < end; i++) {
946 		struct veth_rq *rq = &priv->rq[i];
947 
948 		napi_enable(&rq->xdp_napi);
949 		rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
950 	}
951 
952 	return 0;
953 
954 err_xdp_ring:
955 	for (i--; i >= start; i--)
956 		ptr_ring_cleanup(&priv->rq[i].xdp_ring, veth_ptr_free);
957 
958 	return err;
959 }
960 
961 static int __veth_napi_enable(struct net_device *dev)
962 {
963 	return __veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
964 }
965 
966 static void veth_napi_del_range(struct net_device *dev, int start, int end)
967 {
968 	struct veth_priv *priv = netdev_priv(dev);
969 	int i;
970 
971 	for (i = start; i < end; i++) {
972 		struct veth_rq *rq = &priv->rq[i];
973 
974 		rcu_assign_pointer(priv->rq[i].napi, NULL);
975 		napi_disable(&rq->xdp_napi);
976 		__netif_napi_del(&rq->xdp_napi);
977 	}
978 	synchronize_net();
979 
980 	for (i = start; i < end; i++) {
981 		struct veth_rq *rq = &priv->rq[i];
982 
983 		rq->rx_notify_masked = false;
984 		ptr_ring_cleanup(&rq->xdp_ring, veth_ptr_free);
985 	}
986 }
987 
988 static void veth_napi_del(struct net_device *dev)
989 {
990 	veth_napi_del_range(dev, 0, dev->real_num_rx_queues);
991 }
992 
993 static bool veth_gro_requested(const struct net_device *dev)
994 {
995 	return !!(dev->wanted_features & NETIF_F_GRO);
996 }
997 
998 static int veth_enable_xdp_range(struct net_device *dev, int start, int end,
999 				 bool napi_already_on)
1000 {
1001 	struct veth_priv *priv = netdev_priv(dev);
1002 	int err, i;
1003 
1004 	for (i = start; i < end; i++) {
1005 		struct veth_rq *rq = &priv->rq[i];
1006 
1007 		if (!napi_already_on)
1008 			netif_napi_add(dev, &rq->xdp_napi, veth_poll, NAPI_POLL_WEIGHT);
1009 		err = xdp_rxq_info_reg(&rq->xdp_rxq, dev, i, rq->xdp_napi.napi_id);
1010 		if (err < 0)
1011 			goto err_rxq_reg;
1012 
1013 		err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
1014 						 MEM_TYPE_PAGE_SHARED,
1015 						 NULL);
1016 		if (err < 0)
1017 			goto err_reg_mem;
1018 
1019 		/* Save original mem info as it can be overwritten */
1020 		rq->xdp_mem = rq->xdp_rxq.mem;
1021 	}
1022 	return 0;
1023 
1024 err_reg_mem:
1025 	xdp_rxq_info_unreg(&priv->rq[i].xdp_rxq);
1026 err_rxq_reg:
1027 	for (i--; i >= start; i--) {
1028 		struct veth_rq *rq = &priv->rq[i];
1029 
1030 		xdp_rxq_info_unreg(&rq->xdp_rxq);
1031 		if (!napi_already_on)
1032 			netif_napi_del(&rq->xdp_napi);
1033 	}
1034 
1035 	return err;
1036 }
1037 
1038 static void veth_disable_xdp_range(struct net_device *dev, int start, int end,
1039 				   bool delete_napi)
1040 {
1041 	struct veth_priv *priv = netdev_priv(dev);
1042 	int i;
1043 
1044 	for (i = start; i < end; i++) {
1045 		struct veth_rq *rq = &priv->rq[i];
1046 
1047 		rq->xdp_rxq.mem = rq->xdp_mem;
1048 		xdp_rxq_info_unreg(&rq->xdp_rxq);
1049 
1050 		if (delete_napi)
1051 			netif_napi_del(&rq->xdp_napi);
1052 	}
1053 }
1054 
1055 static int veth_enable_xdp(struct net_device *dev)
1056 {
1057 	bool napi_already_on = veth_gro_requested(dev) && (dev->flags & IFF_UP);
1058 	struct veth_priv *priv = netdev_priv(dev);
1059 	int err, i;
1060 
1061 	if (!xdp_rxq_info_is_reg(&priv->rq[0].xdp_rxq)) {
1062 		err = veth_enable_xdp_range(dev, 0, dev->real_num_rx_queues, napi_already_on);
1063 		if (err)
1064 			return err;
1065 
1066 		if (!napi_already_on) {
1067 			err = __veth_napi_enable(dev);
1068 			if (err) {
1069 				veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, true);
1070 				return err;
1071 			}
1072 
1073 			if (!veth_gro_requested(dev)) {
1074 				/* user-space did not require GRO, but adding XDP
1075 				 * is supposed to get GRO working
1076 				 */
1077 				dev->features |= NETIF_F_GRO;
1078 				netdev_features_change(dev);
1079 			}
1080 		}
1081 	}
1082 
1083 	for (i = 0; i < dev->real_num_rx_queues; i++) {
1084 		rcu_assign_pointer(priv->rq[i].xdp_prog, priv->_xdp_prog);
1085 		rcu_assign_pointer(priv->rq[i].napi, &priv->rq[i].xdp_napi);
1086 	}
1087 
1088 	return 0;
1089 }
1090 
1091 static void veth_disable_xdp(struct net_device *dev)
1092 {
1093 	struct veth_priv *priv = netdev_priv(dev);
1094 	int i;
1095 
1096 	for (i = 0; i < dev->real_num_rx_queues; i++)
1097 		rcu_assign_pointer(priv->rq[i].xdp_prog, NULL);
1098 
1099 	if (!netif_running(dev) || !veth_gro_requested(dev)) {
1100 		veth_napi_del(dev);
1101 
1102 		/* if user-space did not require GRO, since adding XDP
1103 		 * enabled it, clear it now
1104 		 */
1105 		if (!veth_gro_requested(dev) && netif_running(dev)) {
1106 			dev->features &= ~NETIF_F_GRO;
1107 			netdev_features_change(dev);
1108 		}
1109 	}
1110 
1111 	veth_disable_xdp_range(dev, 0, dev->real_num_rx_queues, false);
1112 }
1113 
1114 static int veth_napi_enable_range(struct net_device *dev, int start, int end)
1115 {
1116 	struct veth_priv *priv = netdev_priv(dev);
1117 	int err, i;
1118 
1119 	for (i = start; i < end; i++) {
1120 		struct veth_rq *rq = &priv->rq[i];
1121 
1122 		netif_napi_add(dev, &rq->xdp_napi, veth_poll, NAPI_POLL_WEIGHT);
1123 	}
1124 
1125 	err = __veth_napi_enable_range(dev, start, end);
1126 	if (err) {
1127 		for (i = start; i < end; i++) {
1128 			struct veth_rq *rq = &priv->rq[i];
1129 
1130 			netif_napi_del(&rq->xdp_napi);
1131 		}
1132 		return err;
1133 	}
1134 	return err;
1135 }
1136 
1137 static int veth_napi_enable(struct net_device *dev)
1138 {
1139 	return veth_napi_enable_range(dev, 0, dev->real_num_rx_queues);
1140 }
1141 
1142 static void veth_disable_range_safe(struct net_device *dev, int start, int end)
1143 {
1144 	struct veth_priv *priv = netdev_priv(dev);
1145 
1146 	if (start >= end)
1147 		return;
1148 
1149 	if (priv->_xdp_prog) {
1150 		veth_napi_del_range(dev, start, end);
1151 		veth_disable_xdp_range(dev, start, end, false);
1152 	} else if (veth_gro_requested(dev)) {
1153 		veth_napi_del_range(dev, start, end);
1154 	}
1155 }
1156 
1157 static int veth_enable_range_safe(struct net_device *dev, int start, int end)
1158 {
1159 	struct veth_priv *priv = netdev_priv(dev);
1160 	int err;
1161 
1162 	if (start >= end)
1163 		return 0;
1164 
1165 	if (priv->_xdp_prog) {
1166 		/* these channels are freshly initialized, napi is not on there even
1167 		 * when GRO is requeste
1168 		 */
1169 		err = veth_enable_xdp_range(dev, start, end, false);
1170 		if (err)
1171 			return err;
1172 
1173 		err = __veth_napi_enable_range(dev, start, end);
1174 		if (err) {
1175 			/* on error always delete the newly added napis */
1176 			veth_disable_xdp_range(dev, start, end, true);
1177 			return err;
1178 		}
1179 	} else if (veth_gro_requested(dev)) {
1180 		return veth_napi_enable_range(dev, start, end);
1181 	}
1182 	return 0;
1183 }
1184 
1185 static int veth_set_channels(struct net_device *dev,
1186 			     struct ethtool_channels *ch)
1187 {
1188 	struct veth_priv *priv = netdev_priv(dev);
1189 	unsigned int old_rx_count, new_rx_count;
1190 	struct veth_priv *peer_priv;
1191 	struct net_device *peer;
1192 	int err;
1193 
1194 	/* sanity check. Upper bounds are already enforced by the caller */
1195 	if (!ch->rx_count || !ch->tx_count)
1196 		return -EINVAL;
1197 
1198 	/* avoid braking XDP, if that is enabled */
1199 	peer = rtnl_dereference(priv->peer);
1200 	peer_priv = peer ? netdev_priv(peer) : NULL;
1201 	if (priv->_xdp_prog && peer && ch->rx_count < peer->real_num_tx_queues)
1202 		return -EINVAL;
1203 
1204 	if (peer && peer_priv && peer_priv->_xdp_prog && ch->tx_count > peer->real_num_rx_queues)
1205 		return -EINVAL;
1206 
1207 	old_rx_count = dev->real_num_rx_queues;
1208 	new_rx_count = ch->rx_count;
1209 	if (netif_running(dev)) {
1210 		/* turn device off */
1211 		netif_carrier_off(dev);
1212 		if (peer)
1213 			netif_carrier_off(peer);
1214 
1215 		/* try to allocate new resurces, as needed*/
1216 		err = veth_enable_range_safe(dev, old_rx_count, new_rx_count);
1217 		if (err)
1218 			goto out;
1219 	}
1220 
1221 	err = netif_set_real_num_rx_queues(dev, ch->rx_count);
1222 	if (err)
1223 		goto revert;
1224 
1225 	err = netif_set_real_num_tx_queues(dev, ch->tx_count);
1226 	if (err) {
1227 		int err2 = netif_set_real_num_rx_queues(dev, old_rx_count);
1228 
1229 		/* this error condition could happen only if rx and tx change
1230 		 * in opposite directions (e.g. tx nr raises, rx nr decreases)
1231 		 * and we can't do anything to fully restore the original
1232 		 * status
1233 		 */
1234 		if (err2)
1235 			pr_warn("Can't restore rx queues config %d -> %d %d",
1236 				new_rx_count, old_rx_count, err2);
1237 		else
1238 			goto revert;
1239 	}
1240 
1241 out:
1242 	if (netif_running(dev)) {
1243 		/* note that we need to swap the arguments WRT the enable part
1244 		 * to identify the range we have to disable
1245 		 */
1246 		veth_disable_range_safe(dev, new_rx_count, old_rx_count);
1247 		netif_carrier_on(dev);
1248 		if (peer)
1249 			netif_carrier_on(peer);
1250 	}
1251 	return err;
1252 
1253 revert:
1254 	new_rx_count = old_rx_count;
1255 	old_rx_count = ch->rx_count;
1256 	goto out;
1257 }
1258 
1259 static int veth_open(struct net_device *dev)
1260 {
1261 	struct veth_priv *priv = netdev_priv(dev);
1262 	struct net_device *peer = rtnl_dereference(priv->peer);
1263 	int err;
1264 
1265 	if (!peer)
1266 		return -ENOTCONN;
1267 
1268 	if (priv->_xdp_prog) {
1269 		err = veth_enable_xdp(dev);
1270 		if (err)
1271 			return err;
1272 	} else if (veth_gro_requested(dev)) {
1273 		err = veth_napi_enable(dev);
1274 		if (err)
1275 			return err;
1276 	}
1277 
1278 	if (peer->flags & IFF_UP) {
1279 		netif_carrier_on(dev);
1280 		netif_carrier_on(peer);
1281 	}
1282 
1283 	return 0;
1284 }
1285 
1286 static int veth_close(struct net_device *dev)
1287 {
1288 	struct veth_priv *priv = netdev_priv(dev);
1289 	struct net_device *peer = rtnl_dereference(priv->peer);
1290 
1291 	netif_carrier_off(dev);
1292 	if (peer)
1293 		netif_carrier_off(peer);
1294 
1295 	if (priv->_xdp_prog)
1296 		veth_disable_xdp(dev);
1297 	else if (veth_gro_requested(dev))
1298 		veth_napi_del(dev);
1299 
1300 	return 0;
1301 }
1302 
1303 static int is_valid_veth_mtu(int mtu)
1304 {
1305 	return mtu >= ETH_MIN_MTU && mtu <= ETH_MAX_MTU;
1306 }
1307 
1308 static int veth_alloc_queues(struct net_device *dev)
1309 {
1310 	struct veth_priv *priv = netdev_priv(dev);
1311 	int i;
1312 
1313 	priv->rq = kcalloc(dev->num_rx_queues, sizeof(*priv->rq), GFP_KERNEL);
1314 	if (!priv->rq)
1315 		return -ENOMEM;
1316 
1317 	for (i = 0; i < dev->num_rx_queues; i++) {
1318 		priv->rq[i].dev = dev;
1319 		u64_stats_init(&priv->rq[i].stats.syncp);
1320 	}
1321 
1322 	return 0;
1323 }
1324 
1325 static void veth_free_queues(struct net_device *dev)
1326 {
1327 	struct veth_priv *priv = netdev_priv(dev);
1328 
1329 	kfree(priv->rq);
1330 }
1331 
1332 static int veth_dev_init(struct net_device *dev)
1333 {
1334 	int err;
1335 
1336 	dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats);
1337 	if (!dev->lstats)
1338 		return -ENOMEM;
1339 
1340 	err = veth_alloc_queues(dev);
1341 	if (err) {
1342 		free_percpu(dev->lstats);
1343 		return err;
1344 	}
1345 
1346 	return 0;
1347 }
1348 
1349 static void veth_dev_free(struct net_device *dev)
1350 {
1351 	veth_free_queues(dev);
1352 	free_percpu(dev->lstats);
1353 }
1354 
1355 #ifdef CONFIG_NET_POLL_CONTROLLER
1356 static void veth_poll_controller(struct net_device *dev)
1357 {
1358 	/* veth only receives frames when its peer sends one
1359 	 * Since it has nothing to do with disabling irqs, we are guaranteed
1360 	 * never to have pending data when we poll for it so
1361 	 * there is nothing to do here.
1362 	 *
1363 	 * We need this though so netpoll recognizes us as an interface that
1364 	 * supports polling, which enables bridge devices in virt setups to
1365 	 * still use netconsole
1366 	 */
1367 }
1368 #endif	/* CONFIG_NET_POLL_CONTROLLER */
1369 
1370 static int veth_get_iflink(const struct net_device *dev)
1371 {
1372 	struct veth_priv *priv = netdev_priv(dev);
1373 	struct net_device *peer;
1374 	int iflink;
1375 
1376 	rcu_read_lock();
1377 	peer = rcu_dereference(priv->peer);
1378 	iflink = peer ? peer->ifindex : 0;
1379 	rcu_read_unlock();
1380 
1381 	return iflink;
1382 }
1383 
1384 static netdev_features_t veth_fix_features(struct net_device *dev,
1385 					   netdev_features_t features)
1386 {
1387 	struct veth_priv *priv = netdev_priv(dev);
1388 	struct net_device *peer;
1389 
1390 	peer = rtnl_dereference(priv->peer);
1391 	if (peer) {
1392 		struct veth_priv *peer_priv = netdev_priv(peer);
1393 
1394 		if (peer_priv->_xdp_prog)
1395 			features &= ~NETIF_F_GSO_SOFTWARE;
1396 	}
1397 	if (priv->_xdp_prog)
1398 		features |= NETIF_F_GRO;
1399 
1400 	return features;
1401 }
1402 
1403 static int veth_set_features(struct net_device *dev,
1404 			     netdev_features_t features)
1405 {
1406 	netdev_features_t changed = features ^ dev->features;
1407 	struct veth_priv *priv = netdev_priv(dev);
1408 	int err;
1409 
1410 	if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog)
1411 		return 0;
1412 
1413 	if (features & NETIF_F_GRO) {
1414 		err = veth_napi_enable(dev);
1415 		if (err)
1416 			return err;
1417 	} else {
1418 		veth_napi_del(dev);
1419 	}
1420 	return 0;
1421 }
1422 
1423 static void veth_set_rx_headroom(struct net_device *dev, int new_hr)
1424 {
1425 	struct veth_priv *peer_priv, *priv = netdev_priv(dev);
1426 	struct net_device *peer;
1427 
1428 	if (new_hr < 0)
1429 		new_hr = 0;
1430 
1431 	rcu_read_lock();
1432 	peer = rcu_dereference(priv->peer);
1433 	if (unlikely(!peer))
1434 		goto out;
1435 
1436 	peer_priv = netdev_priv(peer);
1437 	priv->requested_headroom = new_hr;
1438 	new_hr = max(priv->requested_headroom, peer_priv->requested_headroom);
1439 	dev->needed_headroom = new_hr;
1440 	peer->needed_headroom = new_hr;
1441 
1442 out:
1443 	rcu_read_unlock();
1444 }
1445 
1446 static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1447 			struct netlink_ext_ack *extack)
1448 {
1449 	struct veth_priv *priv = netdev_priv(dev);
1450 	struct bpf_prog *old_prog;
1451 	struct net_device *peer;
1452 	unsigned int max_mtu;
1453 	int err;
1454 
1455 	old_prog = priv->_xdp_prog;
1456 	priv->_xdp_prog = prog;
1457 	peer = rtnl_dereference(priv->peer);
1458 
1459 	if (prog) {
1460 		if (!peer) {
1461 			NL_SET_ERR_MSG_MOD(extack, "Cannot set XDP when peer is detached");
1462 			err = -ENOTCONN;
1463 			goto err;
1464 		}
1465 
1466 		max_mtu = PAGE_SIZE - VETH_XDP_HEADROOM -
1467 			  peer->hard_header_len -
1468 			  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1469 		if (peer->mtu > max_mtu) {
1470 			NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP");
1471 			err = -ERANGE;
1472 			goto err;
1473 		}
1474 
1475 		if (dev->real_num_rx_queues < peer->real_num_tx_queues) {
1476 			NL_SET_ERR_MSG_MOD(extack, "XDP expects number of rx queues not less than peer tx queues");
1477 			err = -ENOSPC;
1478 			goto err;
1479 		}
1480 
1481 		if (dev->flags & IFF_UP) {
1482 			err = veth_enable_xdp(dev);
1483 			if (err) {
1484 				NL_SET_ERR_MSG_MOD(extack, "Setup for XDP failed");
1485 				goto err;
1486 			}
1487 		}
1488 
1489 		if (!old_prog) {
1490 			peer->hw_features &= ~NETIF_F_GSO_SOFTWARE;
1491 			peer->max_mtu = max_mtu;
1492 		}
1493 	}
1494 
1495 	if (old_prog) {
1496 		if (!prog) {
1497 			if (dev->flags & IFF_UP)
1498 				veth_disable_xdp(dev);
1499 
1500 			if (peer) {
1501 				peer->hw_features |= NETIF_F_GSO_SOFTWARE;
1502 				peer->max_mtu = ETH_MAX_MTU;
1503 			}
1504 		}
1505 		bpf_prog_put(old_prog);
1506 	}
1507 
1508 	if ((!!old_prog ^ !!prog) && peer)
1509 		netdev_update_features(peer);
1510 
1511 	return 0;
1512 err:
1513 	priv->_xdp_prog = old_prog;
1514 
1515 	return err;
1516 }
1517 
1518 static int veth_xdp(struct net_device *dev, struct netdev_bpf *xdp)
1519 {
1520 	switch (xdp->command) {
1521 	case XDP_SETUP_PROG:
1522 		return veth_xdp_set(dev, xdp->prog, xdp->extack);
1523 	default:
1524 		return -EINVAL;
1525 	}
1526 }
1527 
1528 static const struct net_device_ops veth_netdev_ops = {
1529 	.ndo_init            = veth_dev_init,
1530 	.ndo_open            = veth_open,
1531 	.ndo_stop            = veth_close,
1532 	.ndo_start_xmit      = veth_xmit,
1533 	.ndo_get_stats64     = veth_get_stats64,
1534 	.ndo_set_rx_mode     = veth_set_multicast_list,
1535 	.ndo_set_mac_address = eth_mac_addr,
1536 #ifdef CONFIG_NET_POLL_CONTROLLER
1537 	.ndo_poll_controller	= veth_poll_controller,
1538 #endif
1539 	.ndo_get_iflink		= veth_get_iflink,
1540 	.ndo_fix_features	= veth_fix_features,
1541 	.ndo_set_features	= veth_set_features,
1542 	.ndo_features_check	= passthru_features_check,
1543 	.ndo_set_rx_headroom	= veth_set_rx_headroom,
1544 	.ndo_bpf		= veth_xdp,
1545 	.ndo_xdp_xmit		= veth_ndo_xdp_xmit,
1546 	.ndo_get_peer_dev	= veth_peer_dev,
1547 };
1548 
1549 #define VETH_FEATURES (NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_HW_CSUM | \
1550 		       NETIF_F_RXCSUM | NETIF_F_SCTP_CRC | NETIF_F_HIGHDMA | \
1551 		       NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL | \
1552 		       NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX | \
1553 		       NETIF_F_HW_VLAN_STAG_TX | NETIF_F_HW_VLAN_STAG_RX )
1554 
1555 static void veth_setup(struct net_device *dev)
1556 {
1557 	ether_setup(dev);
1558 
1559 	dev->priv_flags &= ~IFF_TX_SKB_SHARING;
1560 	dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
1561 	dev->priv_flags |= IFF_NO_QUEUE;
1562 	dev->priv_flags |= IFF_PHONY_HEADROOM;
1563 
1564 	dev->netdev_ops = &veth_netdev_ops;
1565 	dev->ethtool_ops = &veth_ethtool_ops;
1566 	dev->features |= NETIF_F_LLTX;
1567 	dev->features |= VETH_FEATURES;
1568 	dev->vlan_features = dev->features &
1569 			     ~(NETIF_F_HW_VLAN_CTAG_TX |
1570 			       NETIF_F_HW_VLAN_STAG_TX |
1571 			       NETIF_F_HW_VLAN_CTAG_RX |
1572 			       NETIF_F_HW_VLAN_STAG_RX);
1573 	dev->needs_free_netdev = true;
1574 	dev->priv_destructor = veth_dev_free;
1575 	dev->max_mtu = ETH_MAX_MTU;
1576 
1577 	dev->hw_features = VETH_FEATURES;
1578 	dev->hw_enc_features = VETH_FEATURES;
1579 	dev->mpls_features = NETIF_F_HW_CSUM | NETIF_F_GSO_SOFTWARE;
1580 }
1581 
1582 /*
1583  * netlink interface
1584  */
1585 
1586 static int veth_validate(struct nlattr *tb[], struct nlattr *data[],
1587 			 struct netlink_ext_ack *extack)
1588 {
1589 	if (tb[IFLA_ADDRESS]) {
1590 		if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
1591 			return -EINVAL;
1592 		if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
1593 			return -EADDRNOTAVAIL;
1594 	}
1595 	if (tb[IFLA_MTU]) {
1596 		if (!is_valid_veth_mtu(nla_get_u32(tb[IFLA_MTU])))
1597 			return -EINVAL;
1598 	}
1599 	return 0;
1600 }
1601 
1602 static struct rtnl_link_ops veth_link_ops;
1603 
1604 static void veth_disable_gro(struct net_device *dev)
1605 {
1606 	dev->features &= ~NETIF_F_GRO;
1607 	dev->wanted_features &= ~NETIF_F_GRO;
1608 	netdev_update_features(dev);
1609 }
1610 
1611 static int veth_init_queues(struct net_device *dev, struct nlattr *tb[])
1612 {
1613 	int err;
1614 
1615 	if (!tb[IFLA_NUM_TX_QUEUES] && dev->num_tx_queues > 1) {
1616 		err = netif_set_real_num_tx_queues(dev, 1);
1617 		if (err)
1618 			return err;
1619 	}
1620 	if (!tb[IFLA_NUM_RX_QUEUES] && dev->num_rx_queues > 1) {
1621 		err = netif_set_real_num_rx_queues(dev, 1);
1622 		if (err)
1623 			return err;
1624 	}
1625 	return 0;
1626 }
1627 
1628 static int veth_newlink(struct net *src_net, struct net_device *dev,
1629 			struct nlattr *tb[], struct nlattr *data[],
1630 			struct netlink_ext_ack *extack)
1631 {
1632 	int err;
1633 	struct net_device *peer;
1634 	struct veth_priv *priv;
1635 	char ifname[IFNAMSIZ];
1636 	struct nlattr *peer_tb[IFLA_MAX + 1], **tbp;
1637 	unsigned char name_assign_type;
1638 	struct ifinfomsg *ifmp;
1639 	struct net *net;
1640 
1641 	/*
1642 	 * create and register peer first
1643 	 */
1644 	if (data != NULL && data[VETH_INFO_PEER] != NULL) {
1645 		struct nlattr *nla_peer;
1646 
1647 		nla_peer = data[VETH_INFO_PEER];
1648 		ifmp = nla_data(nla_peer);
1649 		err = rtnl_nla_parse_ifla(peer_tb,
1650 					  nla_data(nla_peer) + sizeof(struct ifinfomsg),
1651 					  nla_len(nla_peer) - sizeof(struct ifinfomsg),
1652 					  NULL);
1653 		if (err < 0)
1654 			return err;
1655 
1656 		err = veth_validate(peer_tb, NULL, extack);
1657 		if (err < 0)
1658 			return err;
1659 
1660 		tbp = peer_tb;
1661 	} else {
1662 		ifmp = NULL;
1663 		tbp = tb;
1664 	}
1665 
1666 	if (ifmp && tbp[IFLA_IFNAME]) {
1667 		nla_strscpy(ifname, tbp[IFLA_IFNAME], IFNAMSIZ);
1668 		name_assign_type = NET_NAME_USER;
1669 	} else {
1670 		snprintf(ifname, IFNAMSIZ, DRV_NAME "%%d");
1671 		name_assign_type = NET_NAME_ENUM;
1672 	}
1673 
1674 	net = rtnl_link_get_net(src_net, tbp);
1675 	if (IS_ERR(net))
1676 		return PTR_ERR(net);
1677 
1678 	peer = rtnl_create_link(net, ifname, name_assign_type,
1679 				&veth_link_ops, tbp, extack);
1680 	if (IS_ERR(peer)) {
1681 		put_net(net);
1682 		return PTR_ERR(peer);
1683 	}
1684 
1685 	if (!ifmp || !tbp[IFLA_ADDRESS])
1686 		eth_hw_addr_random(peer);
1687 
1688 	if (ifmp && (dev->ifindex != 0))
1689 		peer->ifindex = ifmp->ifi_index;
1690 
1691 	netif_set_gso_max_size(peer, dev->gso_max_size);
1692 	netif_set_gso_max_segs(peer, dev->gso_max_segs);
1693 
1694 	err = register_netdevice(peer);
1695 	put_net(net);
1696 	net = NULL;
1697 	if (err < 0)
1698 		goto err_register_peer;
1699 
1700 	/* keep GRO disabled by default to be consistent with the established
1701 	 * veth behavior
1702 	 */
1703 	veth_disable_gro(peer);
1704 	netif_carrier_off(peer);
1705 
1706 	err = rtnl_configure_link(peer, ifmp);
1707 	if (err < 0)
1708 		goto err_configure_peer;
1709 
1710 	/*
1711 	 * register dev last
1712 	 *
1713 	 * note, that since we've registered new device the dev's name
1714 	 * should be re-allocated
1715 	 */
1716 
1717 	if (tb[IFLA_ADDRESS] == NULL)
1718 		eth_hw_addr_random(dev);
1719 
1720 	if (tb[IFLA_IFNAME])
1721 		nla_strscpy(dev->name, tb[IFLA_IFNAME], IFNAMSIZ);
1722 	else
1723 		snprintf(dev->name, IFNAMSIZ, DRV_NAME "%%d");
1724 
1725 	err = register_netdevice(dev);
1726 	if (err < 0)
1727 		goto err_register_dev;
1728 
1729 	netif_carrier_off(dev);
1730 
1731 	/*
1732 	 * tie the deviced together
1733 	 */
1734 
1735 	priv = netdev_priv(dev);
1736 	rcu_assign_pointer(priv->peer, peer);
1737 	err = veth_init_queues(dev, tb);
1738 	if (err)
1739 		goto err_queues;
1740 
1741 	priv = netdev_priv(peer);
1742 	rcu_assign_pointer(priv->peer, dev);
1743 	err = veth_init_queues(peer, tb);
1744 	if (err)
1745 		goto err_queues;
1746 
1747 	veth_disable_gro(dev);
1748 	return 0;
1749 
1750 err_queues:
1751 	unregister_netdevice(dev);
1752 err_register_dev:
1753 	/* nothing to do */
1754 err_configure_peer:
1755 	unregister_netdevice(peer);
1756 	return err;
1757 
1758 err_register_peer:
1759 	free_netdev(peer);
1760 	return err;
1761 }
1762 
1763 static void veth_dellink(struct net_device *dev, struct list_head *head)
1764 {
1765 	struct veth_priv *priv;
1766 	struct net_device *peer;
1767 
1768 	priv = netdev_priv(dev);
1769 	peer = rtnl_dereference(priv->peer);
1770 
1771 	/* Note : dellink() is called from default_device_exit_batch(),
1772 	 * before a rcu_synchronize() point. The devices are guaranteed
1773 	 * not being freed before one RCU grace period.
1774 	 */
1775 	RCU_INIT_POINTER(priv->peer, NULL);
1776 	unregister_netdevice_queue(dev, head);
1777 
1778 	if (peer) {
1779 		priv = netdev_priv(peer);
1780 		RCU_INIT_POINTER(priv->peer, NULL);
1781 		unregister_netdevice_queue(peer, head);
1782 	}
1783 }
1784 
1785 static const struct nla_policy veth_policy[VETH_INFO_MAX + 1] = {
1786 	[VETH_INFO_PEER]	= { .len = sizeof(struct ifinfomsg) },
1787 };
1788 
1789 static struct net *veth_get_link_net(const struct net_device *dev)
1790 {
1791 	struct veth_priv *priv = netdev_priv(dev);
1792 	struct net_device *peer = rtnl_dereference(priv->peer);
1793 
1794 	return peer ? dev_net(peer) : dev_net(dev);
1795 }
1796 
1797 static unsigned int veth_get_num_queues(void)
1798 {
1799 	/* enforce the same queue limit as rtnl_create_link */
1800 	int queues = num_possible_cpus();
1801 
1802 	if (queues > 4096)
1803 		queues = 4096;
1804 	return queues;
1805 }
1806 
1807 static struct rtnl_link_ops veth_link_ops = {
1808 	.kind		= DRV_NAME,
1809 	.priv_size	= sizeof(struct veth_priv),
1810 	.setup		= veth_setup,
1811 	.validate	= veth_validate,
1812 	.newlink	= veth_newlink,
1813 	.dellink	= veth_dellink,
1814 	.policy		= veth_policy,
1815 	.maxtype	= VETH_INFO_MAX,
1816 	.get_link_net	= veth_get_link_net,
1817 	.get_num_tx_queues	= veth_get_num_queues,
1818 	.get_num_rx_queues	= veth_get_num_queues,
1819 };
1820 
1821 /*
1822  * init/fini
1823  */
1824 
1825 static __init int veth_init(void)
1826 {
1827 	return rtnl_link_register(&veth_link_ops);
1828 }
1829 
1830 static __exit void veth_exit(void)
1831 {
1832 	rtnl_link_unregister(&veth_link_ops);
1833 }
1834 
1835 module_init(veth_init);
1836 module_exit(veth_exit);
1837 
1838 MODULE_DESCRIPTION("Virtual Ethernet Tunnel");
1839 MODULE_LICENSE("GPL v2");
1840 MODULE_ALIAS_RTNL_LINK(DRV_NAME);
1841