udp_offload.c

  1// SPDX-License-Identifier: GPL-2.0-or-later
  2/*
  3 *	IPV4 GSO/GRO offload support
  4 *	Linux INET implementation
  5 *
  6 *	UDPv4 GSO support
  7 */
  8
  9#include <linux/skbuff.h>
 10#include <net/gro.h>
 11#include <net/gso.h>
 12#include <net/udp.h>
 13#include <net/protocol.h>
 14#include <net/inet_common.h>
 15#include <net/udp_tunnel.h>
 16
 17#if IS_ENABLED(CONFIG_NET_UDP_TUNNEL)
 18
 19/*
 20 * Dummy GRO tunnel callback, exists mainly to avoid dangling/NULL
 21 * values for the udp tunnel static call.
 22 */
 23static struct sk_buff *dummy_gro_rcv(struct sock *sk,
 24				     struct list_head *head,
 25				     struct sk_buff *skb)
 26{
 27	NAPI_GRO_CB(skb)->flush = 1;
 28	return NULL;
 29}
 30
 31typedef struct sk_buff *(*udp_tunnel_gro_rcv_t)(struct sock *sk,
 32						struct list_head *head,
 33						struct sk_buff *skb);
 34
 35struct udp_tunnel_type_entry {
 36	udp_tunnel_gro_rcv_t gro_receive;
 37	refcount_t count;
 38};
 39
 40#define UDP_MAX_TUNNEL_TYPES (IS_ENABLED(CONFIG_GENEVE) + \
 41			      IS_ENABLED(CONFIG_VXLAN) * 2 + \
 42			      IS_ENABLED(CONFIG_NET_FOU) * 2 + \
 43			      IS_ENABLED(CONFIG_XFRM) * 2)
 44
 45DEFINE_STATIC_CALL(udp_tunnel_gro_rcv, dummy_gro_rcv);
 46static DEFINE_STATIC_KEY_FALSE(udp_tunnel_static_call);
 47static DEFINE_MUTEX(udp_tunnel_gro_type_lock);
 48static struct udp_tunnel_type_entry udp_tunnel_gro_types[UDP_MAX_TUNNEL_TYPES];
 49static unsigned int udp_tunnel_gro_type_nr;
 50static DEFINE_SPINLOCK(udp_tunnel_gro_lock);
 51
 52void udp_tunnel_update_gro_lookup(struct net *net, struct sock *sk, bool add)
 53{
 54	bool is_ipv6 = sk->sk_family == AF_INET6;
 55	struct udp_sock *tup, *up = udp_sk(sk);
 56	struct udp_tunnel_gro *udp_tunnel_gro;
 57
 58	spin_lock(&udp_tunnel_gro_lock);
 59	udp_tunnel_gro = &net->ipv4.udp_tunnel_gro[is_ipv6];
 60	if (add)
 61		hlist_add_head(&up->tunnel_list, &udp_tunnel_gro->list);
 62	else if (up->tunnel_list.pprev)
 63		hlist_del_init(&up->tunnel_list);
 64
 65	if (udp_tunnel_gro->list.first &&
 66	    !udp_tunnel_gro->list.first->next) {
 67		tup = hlist_entry(udp_tunnel_gro->list.first, struct udp_sock,
 68				  tunnel_list);
 69
 70		rcu_assign_pointer(udp_tunnel_gro->sk, (struct sock *)tup);
 71	} else {
 72		RCU_INIT_POINTER(udp_tunnel_gro->sk, NULL);
 73	}
 74
 75	spin_unlock(&udp_tunnel_gro_lock);
 76}
 77EXPORT_SYMBOL_GPL(udp_tunnel_update_gro_lookup);
 78
 79void udp_tunnel_update_gro_rcv(struct sock *sk, bool add)
 80{
 81	struct udp_tunnel_type_entry *cur = NULL;
 82	struct udp_sock *up = udp_sk(sk);
 83	int i, old_gro_type_nr;
 84
 85	if (!UDP_MAX_TUNNEL_TYPES || !up->gro_receive)
 86		return;
 87
 88	mutex_lock(&udp_tunnel_gro_type_lock);
 89
 90	/* Check if the static call is permanently disabled. */
 91	if (udp_tunnel_gro_type_nr > UDP_MAX_TUNNEL_TYPES)
 92		goto out;
 93
 94	for (i = 0; i < udp_tunnel_gro_type_nr; i++)
 95		if (udp_tunnel_gro_types[i].gro_receive == up->gro_receive)
 96			cur = &udp_tunnel_gro_types[i];
 97
 98	old_gro_type_nr = udp_tunnel_gro_type_nr;
 99	if (add) {
100		/*
101		 * Update the matching entry, if found, or add a new one
102		 * if needed
103		 */
104		if (cur) {
105			refcount_inc(&cur->count);
106			goto out;
107		}
108
109		if (unlikely(udp_tunnel_gro_type_nr == UDP_MAX_TUNNEL_TYPES)) {
110			pr_err_once("Too many UDP tunnel types, please increase UDP_MAX_TUNNEL_TYPES\n");
111			/* Ensure static call will never be enabled */
112			udp_tunnel_gro_type_nr = UDP_MAX_TUNNEL_TYPES + 1;
113		} else {
114			cur = &udp_tunnel_gro_types[udp_tunnel_gro_type_nr++];
115			refcount_set(&cur->count, 1);
116			cur->gro_receive = up->gro_receive;
117		}
118	} else {
119		/*
120		 * The stack cleanups only successfully added tunnel, the
121		 * lookup on removal should never fail.
122		 */
123		if (WARN_ON_ONCE(!cur))
124			goto out;
125
126		if (!refcount_dec_and_test(&cur->count))
127			goto out;
128
129		/* Avoid gaps, so that the enable tunnel has always id 0 */
130		*cur = udp_tunnel_gro_types[--udp_tunnel_gro_type_nr];
131	}
132
133	if (udp_tunnel_gro_type_nr == 1) {
134		static_call_update(udp_tunnel_gro_rcv,
135				   udp_tunnel_gro_types[0].gro_receive);
136		static_branch_enable(&udp_tunnel_static_call);
137	} else if (old_gro_type_nr == 1) {
138		static_branch_disable(&udp_tunnel_static_call);
139		static_call_update(udp_tunnel_gro_rcv, dummy_gro_rcv);
140	}
141
142out:
143	mutex_unlock(&udp_tunnel_gro_type_lock);
144}
145EXPORT_SYMBOL_GPL(udp_tunnel_update_gro_rcv);
146
147static struct sk_buff *udp_tunnel_gro_rcv(struct sock *sk,
148					  struct list_head *head,
149					  struct sk_buff *skb)
150{
151	if (static_branch_likely(&udp_tunnel_static_call)) {
152		if (unlikely(gro_recursion_inc_test(skb))) {
153			NAPI_GRO_CB(skb)->flush |= 1;
154			return NULL;
155		}
156		return static_call(udp_tunnel_gro_rcv)(sk, head, skb);
157	}
158	return call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb);
159}
160
161#else
162
163static struct sk_buff *udp_tunnel_gro_rcv(struct sock *sk,
164					  struct list_head *head,
165					  struct sk_buff *skb)
166{
167	return call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb);
168}
169
170#endif
171
172static struct sk_buff *__skb_udp_tunnel_segment(struct sk_buff *skb,
173	netdev_features_t features,
174	struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
175					     netdev_features_t features),
176	__be16 new_protocol, bool is_ipv6)
177{
178	int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb);
179	bool remcsum, need_csum, offload_csum, gso_partial;
180	struct sk_buff *segs = ERR_PTR(-EINVAL);
181	struct udphdr *uh = udp_hdr(skb);
182	u16 mac_offset = skb->mac_header;
183	__be16 protocol = skb->protocol;
184	u16 mac_len = skb->mac_len;
185	int udp_offset, outer_hlen;
186	__wsum partial;
187	bool need_ipsec;
188
189	if (unlikely(!pskb_may_pull(skb, tnl_hlen)))
190		goto out;
191
192	/* Adjust partial header checksum to negate old length.
193	 * We cannot rely on the value contained in uh->len as it is
194	 * possible that the actual value exceeds the boundaries of the
195	 * 16 bit length field due to the header being added outside of an
196	 * IP or IPv6 frame that was already limited to 64K - 1.
197	 */
198	if (skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)
199		partial = (__force __wsum)uh->len;
200	else
201		partial = (__force __wsum)htonl(skb->len);
202	partial = csum_sub(csum_unfold(uh->check), partial);
203
204	/* setup inner skb. */
205	skb->encapsulation = 0;
206	SKB_GSO_CB(skb)->encap_level = 0;
207	__skb_pull(skb, tnl_hlen);
208	skb_reset_mac_header(skb);
209	skb_set_network_header(skb, skb_inner_network_offset(skb));
210	skb_set_transport_header(skb, skb_inner_transport_offset(skb));
211	skb->mac_len = skb_inner_network_offset(skb);
212	skb->protocol = new_protocol;
213
214	need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM);
215	skb->encap_hdr_csum = need_csum;
216
217	remcsum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_TUNNEL_REMCSUM);
218	skb->remcsum_offload = remcsum;
219
220	need_ipsec = (skb_dst(skb) && dst_xfrm(skb_dst(skb))) || skb_sec_path(skb);
221	/* Try to offload checksum if possible */
222	offload_csum = !!(need_csum &&
223			  !need_ipsec &&
224			  (skb->dev->features &
225			   (is_ipv6 ? (NETIF_F_HW_CSUM | NETIF_F_IPV6_CSUM) :
226				      (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM))));
227
228	features &= skb->dev->hw_enc_features;
229	if (need_csum)
230		features &= ~NETIF_F_SCTP_CRC;
231
232	/* The only checksum offload we care about from here on out is the
233	 * outer one so strip the existing checksum feature flags and
234	 * instead set the flag based on our outer checksum offload value.
235	 */
236	if (remcsum) {
237		features &= ~NETIF_F_CSUM_MASK;
238		if (!need_csum || offload_csum)
239			features |= NETIF_F_HW_CSUM;
240	}
241
242	/* segment inner packet. */
243	segs = gso_inner_segment(skb, features);
244	if (IS_ERR_OR_NULL(segs)) {
245		skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
246				     mac_len);
247		goto out;
248	}
249
250	gso_partial = !!(skb_shinfo(segs)->gso_type & SKB_GSO_PARTIAL);
251
252	outer_hlen = skb_tnl_header_len(skb);
253	udp_offset = outer_hlen - tnl_hlen;
254	skb = segs;
255	do {
256		unsigned int len;
257
258		if (remcsum)
259			skb->ip_summed = CHECKSUM_NONE;
260
261		/* Set up inner headers if we are offloading inner checksum */
262		if (skb->ip_summed == CHECKSUM_PARTIAL) {
263			skb_reset_inner_headers(skb);
264			skb->encapsulation = 1;
265		}
266
267		skb->mac_len = mac_len;
268		skb->protocol = protocol;
269
270		__skb_push(skb, outer_hlen);
271		skb_reset_mac_header(skb);
272		skb_set_network_header(skb, mac_len);
273		skb_set_transport_header(skb, udp_offset);
274		len = skb->len - udp_offset;
275		uh = udp_hdr(skb);
276
277		/* If we are only performing partial GSO the inner header
278		 * will be using a length value equal to only one MSS sized
279		 * segment instead of the entire frame.
280		 */
281		if (gso_partial && skb_is_gso(skb)) {
282			uh->len = htons(skb_shinfo(skb)->gso_size +
283					SKB_GSO_CB(skb)->data_offset +
284					skb->head - (unsigned char *)uh);
285		} else {
286			uh->len = htons(len);
287		}
288
289		if (!need_csum)
290			continue;
291
292		uh->check = ~csum_fold(csum_add(partial,
293				       (__force __wsum)htonl(len)));
294
295		if (skb->encapsulation || !offload_csum) {
296			uh->check = gso_make_checksum(skb, ~uh->check);
297			if (uh->check == 0)
298				uh->check = CSUM_MANGLED_0;
299		} else {
300			skb->ip_summed = CHECKSUM_PARTIAL;
301			skb->csum_start = skb_transport_header(skb) - skb->head;
302			skb->csum_offset = offsetof(struct udphdr, check);
303		}
304	} while ((skb = skb->next));
305out:
306	return segs;
307}
308
309struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
310				       netdev_features_t features,
311				       bool is_ipv6)
312{
313	const struct net_offload __rcu **offloads;
314	__be16 protocol = skb->protocol;
315	const struct net_offload *ops;
316	struct sk_buff *segs = ERR_PTR(-EINVAL);
317	struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
318					     netdev_features_t features);
319
320	rcu_read_lock();
321
322	switch (skb->inner_protocol_type) {
323	case ENCAP_TYPE_ETHER:
324		protocol = skb->inner_protocol;
325		gso_inner_segment = skb_mac_gso_segment;
326		break;
327	case ENCAP_TYPE_IPPROTO:
328		offloads = is_ipv6 ? inet6_offloads : inet_offloads;
329		ops = rcu_dereference(offloads[skb->inner_ipproto]);
330		if (!ops || !ops->callbacks.gso_segment)
331			goto out_unlock;
332		gso_inner_segment = ops->callbacks.gso_segment;
333		break;
334	default:
335		goto out_unlock;
336	}
337
338	segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
339					protocol, is_ipv6);
340
341out_unlock:
342	rcu_read_unlock();
343
344	return segs;
345}
346EXPORT_SYMBOL(skb_udp_tunnel_segment);
347
348static void __udpv4_gso_segment_csum(struct sk_buff *seg,
349				     __be32 *oldip, __be32 *newip,
350				     __be16 *oldport, __be16 *newport)
351{
352	struct udphdr *uh;
353	struct iphdr *iph;
354
355	if (*oldip == *newip && *oldport == *newport)
356		return;
357
358	uh = udp_hdr(seg);
359	iph = ip_hdr(seg);
360
361	if (uh->check) {
362		inet_proto_csum_replace4(&uh->check, seg, *oldip, *newip,
363					 true);
364		inet_proto_csum_replace2(&uh->check, seg, *oldport, *newport,
365					 false);
366		if (!uh->check)
367			uh->check = CSUM_MANGLED_0;
368	}
369	*oldport = *newport;
370
371	csum_replace4(&iph->check, *oldip, *newip);
372	*oldip = *newip;
373}
374
375static struct sk_buff *__udpv4_gso_segment_list_csum(struct sk_buff *segs)
376{
377	struct sk_buff *seg;
378	struct udphdr *uh, *uh2;
379	struct iphdr *iph, *iph2;
380
381	seg = segs;
382	uh = udp_hdr(seg);
383	iph = ip_hdr(seg);
384
385	if ((udp_hdr(seg)->dest == udp_hdr(seg->next)->dest) &&
386	    (udp_hdr(seg)->source == udp_hdr(seg->next)->source) &&
387	    (ip_hdr(seg)->daddr == ip_hdr(seg->next)->daddr) &&
388	    (ip_hdr(seg)->saddr == ip_hdr(seg->next)->saddr))
389		return segs;
390
391	while ((seg = seg->next)) {
392		uh2 = udp_hdr(seg);
393		iph2 = ip_hdr(seg);
394
395		__udpv4_gso_segment_csum(seg,
396					 &iph2->saddr, &iph->saddr,
397					 &uh2->source, &uh->source);
398		__udpv4_gso_segment_csum(seg,
399					 &iph2->daddr, &iph->daddr,
400					 &uh2->dest, &uh->dest);
401	}
402
403	return segs;
404}
405
406static void __udpv6_gso_segment_csum(struct sk_buff *seg,
407				     struct in6_addr *oldip,
408				     const struct in6_addr *newip,
409				     __be16 *oldport, __be16 newport)
410{
411	struct udphdr *uh = udp_hdr(seg);
412
413	if (ipv6_addr_equal(oldip, newip) && *oldport == newport)
414		return;
415
416	if (uh->check) {
417		inet_proto_csum_replace16(&uh->check, seg, oldip->s6_addr32,
418					  newip->s6_addr32, true);
419
420		inet_proto_csum_replace2(&uh->check, seg, *oldport, newport,
421					 false);
422		if (!uh->check)
423			uh->check = CSUM_MANGLED_0;
424	}
425
426	*oldip = *newip;
427	*oldport = newport;
428}
429
430static struct sk_buff *__udpv6_gso_segment_list_csum(struct sk_buff *segs)
431{
432	const struct ipv6hdr *iph;
433	const struct udphdr *uh;
434	struct ipv6hdr *iph2;
435	struct sk_buff *seg;
436	struct udphdr *uh2;
437
438	seg = segs;
439	uh = udp_hdr(seg);
440	iph = ipv6_hdr(seg);
441	uh2 = udp_hdr(seg->next);
442	iph2 = ipv6_hdr(seg->next);
443
444	if (!(*(const u32 *)&uh->source ^ *(const u32 *)&uh2->source) &&
445	    ipv6_addr_equal(&iph->saddr, &iph2->saddr) &&
446	    ipv6_addr_equal(&iph->daddr, &iph2->daddr))
447		return segs;
448
449	while ((seg = seg->next)) {
450		uh2 = udp_hdr(seg);
451		iph2 = ipv6_hdr(seg);
452
453		__udpv6_gso_segment_csum(seg, &iph2->saddr, &iph->saddr,
454					 &uh2->source, uh->source);
455		__udpv6_gso_segment_csum(seg, &iph2->daddr, &iph->daddr,
456					 &uh2->dest, uh->dest);
457	}
458
459	return segs;
460}
461
462static struct sk_buff *__udp_gso_segment_list(struct sk_buff *skb,
463					      netdev_features_t features,
464					      bool is_ipv6)
465{
466	unsigned int mss = skb_shinfo(skb)->gso_size;
467
468	skb = skb_segment_list(skb, features, skb_mac_header_len(skb));
469	if (IS_ERR(skb))
470		return skb;
471
472	udp_hdr(skb)->len = htons(sizeof(struct udphdr) + mss);
473
474	if (is_ipv6)
475		return __udpv6_gso_segment_list_csum(skb);
476	else
477		return __udpv4_gso_segment_list_csum(skb);
478}
479
480struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb,
481				  netdev_features_t features, bool is_ipv6)
482{
483	struct sock *sk = gso_skb->sk;
484	unsigned int sum_truesize = 0;
485	struct sk_buff *segs, *seg;
486	struct udphdr *uh;
487	unsigned int mss;
488	bool copy_dtor;
489	__sum16 check;
490	__be16 newlen;
491	int ret = 0;
492
493	mss = skb_shinfo(gso_skb)->gso_size;
494	if (gso_skb->len <= sizeof(*uh) + mss)
495		return ERR_PTR(-EINVAL);
496
497	if (unlikely(skb_checksum_start(gso_skb) !=
498		     skb_transport_header(gso_skb) &&
499		     !(skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST)))
500		return ERR_PTR(-EINVAL);
501
502	/* We don't know if egress device can segment and checksum the packet
503	 * when IPv6 extension headers are present. Fall back to software GSO.
504	 */
505	if (gso_skb->ip_summed != CHECKSUM_PARTIAL)
506		features &= ~(NETIF_F_GSO_UDP_L4 | NETIF_F_CSUM_MASK);
507
508	if (skb_gso_ok(gso_skb, features | NETIF_F_GSO_ROBUST)) {
509		/* Packet is from an untrusted source, reset gso_segs. */
510		skb_shinfo(gso_skb)->gso_segs = DIV_ROUND_UP(gso_skb->len - sizeof(*uh),
511							     mss);
512		return NULL;
513	}
514
515	if (skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST) {
516		 /* Detect modified geometry and pass those to skb_segment. */
517		if ((skb_pagelen(gso_skb) - sizeof(*uh) == skb_shinfo(gso_skb)->gso_size) &&
518		    !(skb_shinfo(gso_skb)->gso_type & SKB_GSO_DODGY))
519			return __udp_gso_segment_list(gso_skb, features, is_ipv6);
520
521		ret = __skb_linearize(gso_skb);
522		if (ret)
523			return ERR_PTR(ret);
524
525		 /* Setup csum, as fraglist skips this in udp4_gro_receive. */
526		gso_skb->csum_start = skb_transport_header(gso_skb) - gso_skb->head;
527		gso_skb->csum_offset = offsetof(struct udphdr, check);
528		gso_skb->ip_summed = CHECKSUM_PARTIAL;
529
530		uh = udp_hdr(gso_skb);
531		if (is_ipv6)
532			uh->check = ~udp_v6_check(gso_skb->len,
533						  &ipv6_hdr(gso_skb)->saddr,
534						  &ipv6_hdr(gso_skb)->daddr, 0);
535		else
536			uh->check = ~udp_v4_check(gso_skb->len,
537						  ip_hdr(gso_skb)->saddr,
538						  ip_hdr(gso_skb)->daddr, 0);
539	}
540
541	skb_pull(gso_skb, sizeof(*uh));
542
543	/* clear destructor to avoid skb_segment assigning it to tail */
544	copy_dtor = gso_skb->destructor == sock_wfree;
545	if (copy_dtor) {
546		gso_skb->destructor = NULL;
547		gso_skb->sk = NULL;
548	}
549
550	segs = skb_segment(gso_skb, features);
551	if (IS_ERR_OR_NULL(segs)) {
552		if (copy_dtor) {
553			gso_skb->destructor = sock_wfree;
554			gso_skb->sk = sk;
555		}
556		return segs;
557	}
558
559	/* GSO partial and frag_list segmentation only requires splitting
560	 * the frame into an MSS multiple and possibly a remainder, both
561	 * cases return a GSO skb. So update the mss now.
562	 */
563	if (skb_is_gso(segs))
564		mss *= skb_shinfo(segs)->gso_segs;
565
566	seg = segs;
567	uh = udp_hdr(seg);
568
569	/* preserve TX timestamp flags and TS key for first segment */
570	skb_shinfo(seg)->tskey = skb_shinfo(gso_skb)->tskey;
571	skb_shinfo(seg)->tx_flags |=
572			(skb_shinfo(gso_skb)->tx_flags & SKBTX_ANY_TSTAMP);
573
574	/* compute checksum adjustment based on old length versus new */
575	newlen = htons(sizeof(*uh) + mss);
576	check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
577
578	for (;;) {
579		if (copy_dtor) {
580			seg->destructor = sock_wfree;
581			seg->sk = sk;
582			sum_truesize += seg->truesize;
583		}
584
585		if (!seg->next)
586			break;
587
588		uh->len = newlen;
589		uh->check = check;
590
591		if (seg->ip_summed == CHECKSUM_PARTIAL)
592			gso_reset_checksum(seg, ~check);
593		else
594			uh->check = gso_make_checksum(seg, ~check) ? :
595				    CSUM_MANGLED_0;
596
597		seg = seg->next;
598		uh = udp_hdr(seg);
599	}
600
601	/* last packet can be partial gso_size, account for that in checksum */
602	newlen = htons(skb_tail_pointer(seg) - skb_transport_header(seg) +
603		       seg->data_len);
604	check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
605
606	uh->len = newlen;
607	uh->check = check;
608
609	if (seg->ip_summed == CHECKSUM_PARTIAL)
610		gso_reset_checksum(seg, ~check);
611	else
612		uh->check = gso_make_checksum(seg, ~check) ? : CSUM_MANGLED_0;
613
614	/* On the TX path, CHECKSUM_NONE and CHECKSUM_UNNECESSARY have the same
615	 * meaning. However, check for bad offloads in the GSO stack expects the
616	 * latter, if the checksum was calculated in software. To vouch for the
617	 * segment skbs we actually need to set it on the gso_skb.
618	 */
619	if (gso_skb->ip_summed == CHECKSUM_NONE)
620		gso_skb->ip_summed = CHECKSUM_UNNECESSARY;
621
622	/* update refcount for the packet */
623	if (copy_dtor) {
624		int delta = sum_truesize - gso_skb->truesize;
625
626		/* In some pathological cases, delta can be negative.
627		 * We need to either use refcount_add() or refcount_sub_and_test()
628		 */
629		if (likely(delta >= 0))
630			refcount_add(delta, &sk->sk_wmem_alloc);
631		else
632			WARN_ON_ONCE(refcount_sub_and_test(-delta, &sk->sk_wmem_alloc));
633	}
634	return segs;
635}
636EXPORT_SYMBOL_GPL(__udp_gso_segment);
637
638static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
639					 netdev_features_t features)
640{
641	struct sk_buff *segs = ERR_PTR(-EINVAL);
642	unsigned int mss;
643	__wsum csum;
644	struct udphdr *uh;
645	struct iphdr *iph;
646
647	if (skb->encapsulation &&
648	    (skb_shinfo(skb)->gso_type &
649	     (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
650		segs = skb_udp_tunnel_segment(skb, features, false);
651		goto out;
652	}
653
654	if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_UDP | SKB_GSO_UDP_L4)))
655		goto out;
656
657	if (!pskb_may_pull(skb, sizeof(struct udphdr)))
658		goto out;
659
660	if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4)
661		return __udp_gso_segment(skb, features, false);
662
663	mss = skb_shinfo(skb)->gso_size;
664	if (unlikely(skb->len <= mss))
665		goto out;
666
667	/* Do software UFO. Complete and fill in the UDP checksum as
668	 * HW cannot do checksum of UDP packets sent as multiple
669	 * IP fragments.
670	 */
671
672	uh = udp_hdr(skb);
673	iph = ip_hdr(skb);
674
675	uh->check = 0;
676	csum = skb_checksum(skb, 0, skb->len, 0);
677	uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
678	if (uh->check == 0)
679		uh->check = CSUM_MANGLED_0;
680
681	skb->ip_summed = CHECKSUM_UNNECESSARY;
682
683	/* If there is no outer header we can fake a checksum offload
684	 * due to the fact that we have already done the checksum in
685	 * software prior to segmenting the frame.
686	 */
687	if (!skb->encap_hdr_csum)
688		features |= NETIF_F_HW_CSUM;
689
690	/* Fragment the skb. IP headers of the fragments are updated in
691	 * inet_gso_segment()
692	 */
693	segs = skb_segment(skb, features);
694out:
695	return segs;
696}
697
698
699#define UDP_GRO_CNT_MAX 64
700static struct sk_buff *udp_gro_receive_segment(struct list_head *head,
701					       struct sk_buff *skb)
702{
703	struct udphdr *uh = udp_gro_udphdr(skb);
704	struct sk_buff *pp = NULL;
705	struct udphdr *uh2;
706	struct sk_buff *p;
707	unsigned int ulen;
708	int ret = 0;
709	int flush;
710
711	/* requires non zero csum, for symmetry with GSO */
712	if (!uh->check) {
713		NAPI_GRO_CB(skb)->flush = 1;
714		return NULL;
715	}
716
717	/* Do not deal with padded or malicious packets, sorry ! */
718	ulen = ntohs(uh->len);
719	if (ulen <= sizeof(*uh) || ulen != skb_gro_len(skb)) {
720		NAPI_GRO_CB(skb)->flush = 1;
721		return NULL;
722	}
723	/* pull encapsulating udp header */
724	skb_gro_pull(skb, sizeof(struct udphdr));
725
726	list_for_each_entry(p, head, list) {
727		if (!NAPI_GRO_CB(p)->same_flow)
728			continue;
729
730		uh2 = udp_hdr(p);
731
732		/* Match ports only, as csum is always non zero */
733		if ((*(u32 *)&uh->source != *(u32 *)&uh2->source)) {
734			NAPI_GRO_CB(p)->same_flow = 0;
735			continue;
736		}
737
738		if (NAPI_GRO_CB(skb)->is_flist != NAPI_GRO_CB(p)->is_flist) {
739			NAPI_GRO_CB(skb)->flush = 1;
740			return p;
741		}
742
743		flush = gro_receive_network_flush(uh, uh2, p);
744
745		/* Terminate the flow on len mismatch or if it grow "too much".
746		 * Under small packet flood GRO count could elsewhere grow a lot
747		 * leading to excessive truesize values.
748		 * On len mismatch merge the first packet shorter than gso_size,
749		 * otherwise complete the GRO packet.
750		 */
751		if (ulen > ntohs(uh2->len) || flush) {
752			pp = p;
753		} else {
754			if (NAPI_GRO_CB(skb)->is_flist) {
755				if (!pskb_may_pull(skb, skb_gro_offset(skb))) {
756					NAPI_GRO_CB(skb)->flush = 1;
757					return NULL;
758				}
759				if ((skb->ip_summed != p->ip_summed) ||
760				    (skb->csum_level != p->csum_level)) {
761					NAPI_GRO_CB(skb)->flush = 1;
762					return NULL;
763				}
764				skb_set_network_header(skb, skb_gro_receive_network_offset(skb));
765				ret = skb_gro_receive_list(p, skb);
766			} else {
767				skb_gro_postpull_rcsum(skb, uh,
768						       sizeof(struct udphdr));
769
770				ret = skb_gro_receive(p, skb);
771			}
772		}
773
774		if (ret || ulen != ntohs(uh2->len) ||
775		    NAPI_GRO_CB(p)->count >= UDP_GRO_CNT_MAX)
776			pp = p;
777
778		return pp;
779	}
780
781	/* mismatch, but we never need to flush */
782	return NULL;
783}
784
785struct sk_buff *udp_gro_receive(struct list_head *head, struct sk_buff *skb,
786				struct udphdr *uh, struct sock *sk)
787{
788	struct sk_buff *pp = NULL;
789	struct sk_buff *p;
790	struct udphdr *uh2;
791	unsigned int off = skb_gro_offset(skb);
792	int flush = 1;
793
794	/* We can do L4 aggregation only if the packet can't land in a tunnel
795	 * otherwise we could corrupt the inner stream. Detecting such packets
796	 * cannot be foolproof and the aggregation might still happen in some
797	 * cases. Such packets should be caught in udp_unexpected_gso later.
798	 */
799	NAPI_GRO_CB(skb)->is_flist = 0;
800	if (!sk || !udp_sk(sk)->gro_receive) {
801		/* If the packet was locally encapsulated in a UDP tunnel that
802		 * wasn't detected above, do not GRO.
803		 */
804		if (skb->encapsulation)
805			goto out;
806
807		if (skb->dev->features & NETIF_F_GRO_FRAGLIST)
808			NAPI_GRO_CB(skb)->is_flist = sk ? !udp_test_bit(GRO_ENABLED, sk) : 1;
809
810		if ((!sk && (skb->dev->features & NETIF_F_GRO_UDP_FWD)) ||
811		    (sk && udp_test_bit(GRO_ENABLED, sk)) || NAPI_GRO_CB(skb)->is_flist)
812			return call_gro_receive(udp_gro_receive_segment, head, skb);
813
814		/* no GRO, be sure flush the current packet */
815		goto out;
816	}
817
818	if (NAPI_GRO_CB(skb)->encap_mark ||
819	    (uh->check && skb->ip_summed != CHECKSUM_PARTIAL &&
820	     NAPI_GRO_CB(skb)->csum_cnt == 0 &&
821	     !NAPI_GRO_CB(skb)->csum_valid))
822		goto out;
823
824	/* mark that this skb passed once through the tunnel gro layer */
825	NAPI_GRO_CB(skb)->encap_mark = 1;
826
827	flush = 0;
828
829	list_for_each_entry(p, head, list) {
830		if (!NAPI_GRO_CB(p)->same_flow)
831			continue;
832
833		uh2 = (struct udphdr   *)(p->data + off);
834
835		/* Match ports and either checksums are either both zero
836		 * or nonzero.
837		 */
838		if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
839		    (!uh->check ^ !uh2->check)) {
840			NAPI_GRO_CB(p)->same_flow = 0;
841			continue;
842		}
843	}
844
845	skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
846	skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
847	pp = udp_tunnel_gro_rcv(sk, head, skb);
848
849out:
850	skb_gro_flush_final(skb, pp, flush);
851	return pp;
852}
853EXPORT_SYMBOL(udp_gro_receive);
854
855static struct sock *udp4_gro_lookup_skb(struct sk_buff *skb, __be16 sport,
856					__be16 dport)
857{
858	const struct iphdr *iph = skb_gro_network_header(skb);
859	struct net *net = dev_net_rcu(skb->dev);
860	struct sock *sk;
861	int iif, sdif;
862
863	sk = udp_tunnel_sk(net, false);
864	if (sk && dport == htons(sk->sk_num))
865		return sk;
866
867	inet_get_iif_sdif(skb, &iif, &sdif);
868
869	return __udp4_lib_lookup(net, iph->saddr, sport,
870				 iph->daddr, dport, iif,
871				 sdif, net->ipv4.udp_table, NULL);
872}
873
874INDIRECT_CALLABLE_SCOPE
875struct sk_buff *udp4_gro_receive(struct list_head *head, struct sk_buff *skb)
876{
877	struct udphdr *uh = udp_gro_udphdr(skb);
878	struct sock *sk = NULL;
879	struct sk_buff *pp;
880
881	if (unlikely(!uh))
882		goto flush;
883
884	/* Don't bother verifying checksum if we're going to flush anyway. */
885	if (NAPI_GRO_CB(skb)->flush)
886		goto skip;
887
888	if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
889						 inet_gro_compute_pseudo))
890		goto flush;
891	else if (uh->check)
892		skb_gro_checksum_try_convert(skb, IPPROTO_UDP,
893					     inet_gro_compute_pseudo);
894skip:
895	if (static_branch_unlikely(&udp_encap_needed_key))
896		sk = udp4_gro_lookup_skb(skb, uh->source, uh->dest);
897
898	pp = udp_gro_receive(head, skb, uh, sk);
899	return pp;
900
901flush:
902	NAPI_GRO_CB(skb)->flush = 1;
903	return NULL;
904}
905
906static int udp_gro_complete_segment(struct sk_buff *skb)
907{
908	struct udphdr *uh = udp_hdr(skb);
909
910	skb->csum_start = (unsigned char *)uh - skb->head;
911	skb->csum_offset = offsetof(struct udphdr, check);
912	skb->ip_summed = CHECKSUM_PARTIAL;
913
914	skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
915	skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_L4;
916
917	if (skb->encapsulation)
918		skb->inner_transport_header = skb->transport_header;
919
920	return 0;
921}
922
923int udp_gro_complete(struct sk_buff *skb, int nhoff,
924		     udp_lookup_t lookup)
925{
926	__be16 newlen = htons(skb->len - nhoff);
927	struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
928	struct sock *sk;
929	int err;
930
931	uh->len = newlen;
932
933	sk = INDIRECT_CALL_INET(lookup, udp6_lib_lookup_skb,
934				udp4_lib_lookup_skb, skb, uh->source, uh->dest);
935	if (sk && udp_sk(sk)->gro_complete) {
936		skb_shinfo(skb)->gso_type = uh->check ? SKB_GSO_UDP_TUNNEL_CSUM
937					: SKB_GSO_UDP_TUNNEL;
938
939		/* clear the encap mark, so that inner frag_list gro_complete
940		 * can take place
941		 */
942		NAPI_GRO_CB(skb)->encap_mark = 0;
943
944		/* Set encapsulation before calling into inner gro_complete()
945		 * functions to make them set up the inner offsets.
946		 */
947		skb->encapsulation = 1;
948		err = udp_sk(sk)->gro_complete(sk, skb,
949				nhoff + sizeof(struct udphdr));
950	} else {
951		err = udp_gro_complete_segment(skb);
952	}
953
954	if (skb->remcsum_offload)
955		skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
956
957	return err;
958}
959EXPORT_SYMBOL(udp_gro_complete);
960
961INDIRECT_CALLABLE_SCOPE int udp4_gro_complete(struct sk_buff *skb, int nhoff)
962{
963	const u16 offset = NAPI_GRO_CB(skb)->network_offsets[skb->encapsulation];
964	const struct iphdr *iph = (struct iphdr *)(skb->data + offset);
965	struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
966
967	/* do fraglist only if there is no outer UDP encap (or we already processed it) */
968	if (NAPI_GRO_CB(skb)->is_flist && !NAPI_GRO_CB(skb)->encap_mark) {
969		uh->len = htons(skb->len - nhoff);
970
971		skb_shinfo(skb)->gso_type |= (SKB_GSO_FRAGLIST|SKB_GSO_UDP_L4);
972		skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
973
974		__skb_incr_checksum_unnecessary(skb);
975
976		return 0;
977	}
978
979	if (uh->check)
980		uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
981					  iph->daddr, 0);
982
983	return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb);
984}
985
986int __init udpv4_offload_init(void)
987{
988	net_hotdata.udpv4_offload = (struct net_offload) {
989		.callbacks = {
990			.gso_segment = udp4_ufo_fragment,
991			.gro_receive  =	udp4_gro_receive,
992			.gro_complete =	udp4_gro_complete,
993		},
994	};
995
996	return inet_add_offload(&net_hotdata.udpv4_offload, IPPROTO_UDP);
997}