xref: /linux/net/ipv4/ip_output.c (revision 59058f2af9ca17d6c3113f6bbf93f6389fd4d0ea)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * INET		An implementation of the TCP/IP protocol suite for the LINUX
4  *		operating system.  INET is implemented using the  BSD Socket
5  *		interface as the means of communication with the user level.
6  *
7  *		The Internet Protocol (IP) output module.
8  *
9  * Authors:	Ross Biro
10  *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11  *		Donald Becker, <becker@super.org>
12  *		Alan Cox, <Alan.Cox@linux.org>
13  *		Richard Underwood
14  *		Stefan Becker, <stefanb@yello.ping.de>
15  *		Jorge Cwik, <jorge@laser.satlink.net>
16  *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
17  *		Hirokazu Takahashi, <taka@valinux.co.jp>
18  *
19  *	See ip_input.c for original log
20  *
21  *	Fixes:
22  *		Alan Cox	:	Missing nonblock feature in ip_build_xmit.
23  *		Mike Kilburn	:	htons() missing in ip_build_xmit.
24  *		Bradford Johnson:	Fix faulty handling of some frames when
25  *					no route is found.
26  *		Alexander Demenshin:	Missing sk/skb free in ip_queue_xmit
27  *					(in case if packet not accepted by
28  *					output firewall rules)
29  *		Mike McLagan	:	Routing by source
30  *		Alexey Kuznetsov:	use new route cache
31  *		Andi Kleen:		Fix broken PMTU recovery and remove
32  *					some redundant tests.
33  *	Vitaly E. Lavrov	:	Transparent proxy revived after year coma.
34  *		Andi Kleen	: 	Replace ip_reply with ip_send_reply.
35  *		Andi Kleen	:	Split fast and slow ip_build_xmit path
36  *					for decreased register pressure on x86
37  *					and more readability.
38  *		Marc Boucher	:	When call_out_firewall returns FW_QUEUE,
39  *					silently drop skb instead of failing with -EPERM.
40  *		Detlev Wengorz	:	Copy protocol for fragments.
41  *		Hirokazu Takahashi:	HW checksumming for outgoing UDP
42  *					datagrams.
43  *		Hirokazu Takahashi:	sendfile() on UDP works now.
44  */
45 
46 #include <linux/uaccess.h>
47 #include <linux/module.h>
48 #include <linux/types.h>
49 #include <linux/kernel.h>
50 #include <linux/mm.h>
51 #include <linux/string.h>
52 #include <linux/errno.h>
53 #include <linux/highmem.h>
54 #include <linux/slab.h>
55 
56 #include <linux/socket.h>
57 #include <linux/sockios.h>
58 #include <linux/in.h>
59 #include <linux/inet.h>
60 #include <linux/netdevice.h>
61 #include <linux/etherdevice.h>
62 #include <linux/proc_fs.h>
63 #include <linux/stat.h>
64 #include <linux/init.h>
65 
66 #include <net/snmp.h>
67 #include <net/ip.h>
68 #include <net/protocol.h>
69 #include <net/route.h>
70 #include <net/xfrm.h>
71 #include <linux/skbuff.h>
72 #include <net/sock.h>
73 #include <net/arp.h>
74 #include <net/icmp.h>
75 #include <net/checksum.h>
76 #include <net/gso.h>
77 #include <net/inetpeer.h>
78 #include <net/inet_ecn.h>
79 #include <net/lwtunnel.h>
80 #include <linux/bpf-cgroup.h>
81 #include <linux/igmp.h>
82 #include <linux/netfilter_ipv4.h>
83 #include <linux/netfilter_bridge.h>
84 #include <linux/netlink.h>
85 #include <linux/tcp.h>
86 
87 static int
88 ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
89 	    unsigned int mtu,
90 	    int (*output)(struct net *, struct sock *, struct sk_buff *));
91 
92 /* Generate a checksum for an outgoing IP datagram. */
93 void ip_send_check(struct iphdr *iph)
94 {
95 	iph->check = 0;
96 	iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
97 }
98 EXPORT_SYMBOL(ip_send_check);
99 
100 int __ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
101 {
102 	struct iphdr *iph = ip_hdr(skb);
103 
104 	IP_INC_STATS(net, IPSTATS_MIB_OUTREQUESTS);
105 
106 	iph_set_totlen(iph, skb->len);
107 	ip_send_check(iph);
108 
109 	/* if egress device is enslaved to an L3 master device pass the
110 	 * skb to its handler for processing
111 	 */
112 	skb = l3mdev_ip_out(sk, skb);
113 	if (unlikely(!skb))
114 		return 0;
115 
116 	skb->protocol = htons(ETH_P_IP);
117 
118 	return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
119 		       net, sk, skb, NULL, skb_dst(skb)->dev,
120 		       dst_output);
121 }
122 
123 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
124 {
125 	int err;
126 
127 	err = __ip_local_out(net, sk, skb);
128 	if (likely(err == 1))
129 		err = dst_output(net, sk, skb);
130 
131 	return err;
132 }
133 EXPORT_SYMBOL_GPL(ip_local_out);
134 
135 static inline int ip_select_ttl(const struct inet_sock *inet,
136 				const struct dst_entry *dst)
137 {
138 	int ttl = READ_ONCE(inet->uc_ttl);
139 
140 	if (ttl < 0)
141 		ttl = ip4_dst_hoplimit(dst);
142 	return ttl;
143 }
144 
145 /*
146  *		Add an ip header to a skbuff and send it out.
147  *
148  */
149 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
150 			  __be32 saddr, __be32 daddr, struct ip_options_rcu *opt,
151 			  u8 tos)
152 {
153 	const struct inet_sock *inet = inet_sk(sk);
154 	struct rtable *rt = skb_rtable(skb);
155 	struct net *net = sock_net(sk);
156 	struct iphdr *iph;
157 
158 	/* Build the IP header. */
159 	skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
160 	skb_reset_network_header(skb);
161 	iph = ip_hdr(skb);
162 	iph->version  = 4;
163 	iph->ihl      = 5;
164 	iph->tos      = tos;
165 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
166 	iph->daddr    = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
167 	iph->saddr    = saddr;
168 	iph->protocol = sk->sk_protocol;
169 	/* Do not bother generating IPID for small packets (eg SYNACK) */
170 	if (skb->len <= IPV4_MIN_MTU || ip_dont_fragment(sk, &rt->dst)) {
171 		iph->frag_off = htons(IP_DF);
172 		iph->id = 0;
173 	} else {
174 		iph->frag_off = 0;
175 		/* TCP packets here are SYNACK with fat IPv4/TCP options.
176 		 * Avoid using the hashed IP ident generator.
177 		 */
178 		if (sk->sk_protocol == IPPROTO_TCP)
179 			iph->id = (__force __be16)get_random_u16();
180 		else
181 			__ip_select_ident(net, iph, 1);
182 	}
183 
184 	if (opt && opt->opt.optlen) {
185 		iph->ihl += opt->opt.optlen>>2;
186 		ip_options_build(skb, &opt->opt, daddr, rt);
187 	}
188 
189 	skb->priority = READ_ONCE(sk->sk_priority);
190 	if (!skb->mark)
191 		skb->mark = READ_ONCE(sk->sk_mark);
192 
193 	/* Send it out. */
194 	return ip_local_out(net, skb->sk, skb);
195 }
196 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
197 
198 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
199 {
200 	struct dst_entry *dst = skb_dst(skb);
201 	struct rtable *rt = (struct rtable *)dst;
202 	struct net_device *dev = dst->dev;
203 	unsigned int hh_len = LL_RESERVED_SPACE(dev);
204 	struct neighbour *neigh;
205 	bool is_v6gw = false;
206 
207 	if (rt->rt_type == RTN_MULTICAST) {
208 		IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
209 	} else if (rt->rt_type == RTN_BROADCAST)
210 		IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
211 
212 	/* OUTOCTETS should be counted after fragment */
213 	IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
214 
215 	if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
216 		skb = skb_expand_head(skb, hh_len);
217 		if (!skb)
218 			return -ENOMEM;
219 	}
220 
221 	if (lwtunnel_xmit_redirect(dst->lwtstate)) {
222 		int res = lwtunnel_xmit(skb);
223 
224 		if (res != LWTUNNEL_XMIT_CONTINUE)
225 			return res;
226 	}
227 
228 	rcu_read_lock();
229 	neigh = ip_neigh_for_gw(rt, skb, &is_v6gw);
230 	if (!IS_ERR(neigh)) {
231 		int res;
232 
233 		sock_confirm_neigh(skb, neigh);
234 		/* if crossing protocols, can not use the cached header */
235 		res = neigh_output(neigh, skb, is_v6gw);
236 		rcu_read_unlock();
237 		return res;
238 	}
239 	rcu_read_unlock();
240 
241 	net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
242 			    __func__);
243 	kfree_skb_reason(skb, SKB_DROP_REASON_NEIGH_CREATEFAIL);
244 	return PTR_ERR(neigh);
245 }
246 
247 static int ip_finish_output_gso(struct net *net, struct sock *sk,
248 				struct sk_buff *skb, unsigned int mtu)
249 {
250 	struct sk_buff *segs, *nskb;
251 	netdev_features_t features;
252 	int ret = 0;
253 
254 	/* common case: seglen is <= mtu
255 	 */
256 	if (skb_gso_validate_network_len(skb, mtu))
257 		return ip_finish_output2(net, sk, skb);
258 
259 	/* Slowpath -  GSO segment length exceeds the egress MTU.
260 	 *
261 	 * This can happen in several cases:
262 	 *  - Forwarding of a TCP GRO skb, when DF flag is not set.
263 	 *  - Forwarding of an skb that arrived on a virtualization interface
264 	 *    (virtio-net/vhost/tap) with TSO/GSO size set by other network
265 	 *    stack.
266 	 *  - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
267 	 *    interface with a smaller MTU.
268 	 *  - Arriving GRO skb (or GSO skb in a virtualized environment) that is
269 	 *    bridged to a NETIF_F_TSO tunnel stacked over an interface with an
270 	 *    insufficient MTU.
271 	 */
272 	features = netif_skb_features(skb);
273 	BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_GSO_CB_OFFSET);
274 	segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
275 	if (IS_ERR_OR_NULL(segs)) {
276 		kfree_skb(skb);
277 		return -ENOMEM;
278 	}
279 
280 	consume_skb(skb);
281 
282 	skb_list_walk_safe(segs, segs, nskb) {
283 		int err;
284 
285 		skb_mark_not_on_list(segs);
286 		err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
287 
288 		if (err && ret == 0)
289 			ret = err;
290 	}
291 
292 	return ret;
293 }
294 
295 static int __ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
296 {
297 	unsigned int mtu;
298 
299 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
300 	/* Policy lookup after SNAT yielded a new policy */
301 	if (skb_dst(skb)->xfrm) {
302 		IPCB(skb)->flags |= IPSKB_REROUTED;
303 		return dst_output(net, sk, skb);
304 	}
305 #endif
306 	mtu = ip_skb_dst_mtu(sk, skb);
307 	if (skb_is_gso(skb))
308 		return ip_finish_output_gso(net, sk, skb, mtu);
309 
310 	if (skb->len > mtu || IPCB(skb)->frag_max_size)
311 		return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
312 
313 	return ip_finish_output2(net, sk, skb);
314 }
315 
316 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
317 {
318 	int ret;
319 
320 	ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
321 	switch (ret) {
322 	case NET_XMIT_SUCCESS:
323 		return __ip_finish_output(net, sk, skb);
324 	case NET_XMIT_CN:
325 		return __ip_finish_output(net, sk, skb) ? : ret;
326 	default:
327 		kfree_skb_reason(skb, SKB_DROP_REASON_BPF_CGROUP_EGRESS);
328 		return ret;
329 	}
330 }
331 
332 static int ip_mc_finish_output(struct net *net, struct sock *sk,
333 			       struct sk_buff *skb)
334 {
335 	struct rtable *new_rt;
336 	bool do_cn = false;
337 	int ret, err;
338 
339 	ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
340 	switch (ret) {
341 	case NET_XMIT_CN:
342 		do_cn = true;
343 		fallthrough;
344 	case NET_XMIT_SUCCESS:
345 		break;
346 	default:
347 		kfree_skb_reason(skb, SKB_DROP_REASON_BPF_CGROUP_EGRESS);
348 		return ret;
349 	}
350 
351 	/* Reset rt_iif so that inet_iif() will return skb->skb_iif. Setting
352 	 * this to non-zero causes ipi_ifindex in in_pktinfo to be overwritten,
353 	 * see ipv4_pktinfo_prepare().
354 	 */
355 	new_rt = rt_dst_clone(net->loopback_dev, skb_rtable(skb));
356 	if (new_rt) {
357 		new_rt->rt_iif = 0;
358 		skb_dst_drop(skb);
359 		skb_dst_set(skb, &new_rt->dst);
360 	}
361 
362 	err = dev_loopback_xmit(net, sk, skb);
363 	return (do_cn && err) ? ret : err;
364 }
365 
366 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
367 {
368 	struct rtable *rt = skb_rtable(skb);
369 	struct net_device *dev = rt->dst.dev;
370 
371 	/*
372 	 *	If the indicated interface is up and running, send the packet.
373 	 */
374 	skb->dev = dev;
375 	skb->protocol = htons(ETH_P_IP);
376 
377 	/*
378 	 *	Multicasts are looped back for other local users
379 	 */
380 
381 	if (rt->rt_flags&RTCF_MULTICAST) {
382 		if (sk_mc_loop(sk)
383 #ifdef CONFIG_IP_MROUTE
384 		/* Small optimization: do not loopback not local frames,
385 		   which returned after forwarding; they will be  dropped
386 		   by ip_mr_input in any case.
387 		   Note, that local frames are looped back to be delivered
388 		   to local recipients.
389 
390 		   This check is duplicated in ip_mr_input at the moment.
391 		 */
392 		    &&
393 		    ((rt->rt_flags & RTCF_LOCAL) ||
394 		     !(IPCB(skb)->flags & IPSKB_FORWARDED))
395 #endif
396 		   ) {
397 			struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
398 			if (newskb)
399 				NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
400 					net, sk, newskb, NULL, newskb->dev,
401 					ip_mc_finish_output);
402 		}
403 
404 		/* Multicasts with ttl 0 must not go beyond the host */
405 
406 		if (ip_hdr(skb)->ttl == 0) {
407 			kfree_skb(skb);
408 			return 0;
409 		}
410 	}
411 
412 	if (rt->rt_flags&RTCF_BROADCAST) {
413 		struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
414 		if (newskb)
415 			NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
416 				net, sk, newskb, NULL, newskb->dev,
417 				ip_mc_finish_output);
418 	}
419 
420 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
421 			    net, sk, skb, NULL, skb->dev,
422 			    ip_finish_output,
423 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
424 }
425 
426 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
427 {
428 	struct net_device *dev = skb_dst(skb)->dev, *indev = skb->dev;
429 
430 	skb->dev = dev;
431 	skb->protocol = htons(ETH_P_IP);
432 
433 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
434 			    net, sk, skb, indev, dev,
435 			    ip_finish_output,
436 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
437 }
438 EXPORT_SYMBOL(ip_output);
439 
440 /*
441  * copy saddr and daddr, possibly using 64bit load/stores
442  * Equivalent to :
443  *   iph->saddr = fl4->saddr;
444  *   iph->daddr = fl4->daddr;
445  */
446 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
447 {
448 	BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
449 		     offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
450 
451 	iph->saddr = fl4->saddr;
452 	iph->daddr = fl4->daddr;
453 }
454 
455 /* Note: skb->sk can be different from sk, in case of tunnels */
456 int __ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl,
457 		    __u8 tos)
458 {
459 	struct inet_sock *inet = inet_sk(sk);
460 	struct net *net = sock_net(sk);
461 	struct ip_options_rcu *inet_opt;
462 	struct flowi4 *fl4;
463 	struct rtable *rt;
464 	struct iphdr *iph;
465 	int res;
466 
467 	/* Skip all of this if the packet is already routed,
468 	 * f.e. by something like SCTP.
469 	 */
470 	rcu_read_lock();
471 	inet_opt = rcu_dereference(inet->inet_opt);
472 	fl4 = &fl->u.ip4;
473 	rt = skb_rtable(skb);
474 	if (rt)
475 		goto packet_routed;
476 
477 	/* Make sure we can route this packet. */
478 	rt = (struct rtable *)__sk_dst_check(sk, 0);
479 	if (!rt) {
480 		__be32 daddr;
481 
482 		/* Use correct destination address if we have options. */
483 		daddr = inet->inet_daddr;
484 		if (inet_opt && inet_opt->opt.srr)
485 			daddr = inet_opt->opt.faddr;
486 
487 		/* If this fails, retransmit mechanism of transport layer will
488 		 * keep trying until route appears or the connection times
489 		 * itself out.
490 		 */
491 		rt = ip_route_output_ports(net, fl4, sk,
492 					   daddr, inet->inet_saddr,
493 					   inet->inet_dport,
494 					   inet->inet_sport,
495 					   sk->sk_protocol,
496 					   RT_CONN_FLAGS_TOS(sk, tos),
497 					   sk->sk_bound_dev_if);
498 		if (IS_ERR(rt))
499 			goto no_route;
500 		sk_setup_caps(sk, &rt->dst);
501 	}
502 	skb_dst_set_noref(skb, &rt->dst);
503 
504 packet_routed:
505 	if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
506 		goto no_route;
507 
508 	/* OK, we know where to send it, allocate and build IP header. */
509 	skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
510 	skb_reset_network_header(skb);
511 	iph = ip_hdr(skb);
512 	*((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (tos & 0xff));
513 	if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
514 		iph->frag_off = htons(IP_DF);
515 	else
516 		iph->frag_off = 0;
517 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
518 	iph->protocol = sk->sk_protocol;
519 	ip_copy_addrs(iph, fl4);
520 
521 	/* Transport layer set skb->h.foo itself. */
522 
523 	if (inet_opt && inet_opt->opt.optlen) {
524 		iph->ihl += inet_opt->opt.optlen >> 2;
525 		ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt);
526 	}
527 
528 	ip_select_ident_segs(net, skb, sk,
529 			     skb_shinfo(skb)->gso_segs ?: 1);
530 
531 	/* TODO : should we use skb->sk here instead of sk ? */
532 	skb->priority = READ_ONCE(sk->sk_priority);
533 	skb->mark = READ_ONCE(sk->sk_mark);
534 
535 	res = ip_local_out(net, sk, skb);
536 	rcu_read_unlock();
537 	return res;
538 
539 no_route:
540 	rcu_read_unlock();
541 	IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
542 	kfree_skb_reason(skb, SKB_DROP_REASON_IP_OUTNOROUTES);
543 	return -EHOSTUNREACH;
544 }
545 EXPORT_SYMBOL(__ip_queue_xmit);
546 
547 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
548 {
549 	return __ip_queue_xmit(sk, skb, fl, READ_ONCE(inet_sk(sk)->tos));
550 }
551 EXPORT_SYMBOL(ip_queue_xmit);
552 
553 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
554 {
555 	to->pkt_type = from->pkt_type;
556 	to->priority = from->priority;
557 	to->protocol = from->protocol;
558 	to->skb_iif = from->skb_iif;
559 	skb_dst_drop(to);
560 	skb_dst_copy(to, from);
561 	to->dev = from->dev;
562 	to->mark = from->mark;
563 
564 	skb_copy_hash(to, from);
565 
566 #ifdef CONFIG_NET_SCHED
567 	to->tc_index = from->tc_index;
568 #endif
569 	nf_copy(to, from);
570 	skb_ext_copy(to, from);
571 #if IS_ENABLED(CONFIG_IP_VS)
572 	to->ipvs_property = from->ipvs_property;
573 #endif
574 	skb_copy_secmark(to, from);
575 }
576 
577 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
578 		       unsigned int mtu,
579 		       int (*output)(struct net *, struct sock *, struct sk_buff *))
580 {
581 	struct iphdr *iph = ip_hdr(skb);
582 
583 	if ((iph->frag_off & htons(IP_DF)) == 0)
584 		return ip_do_fragment(net, sk, skb, output);
585 
586 	if (unlikely(!skb->ignore_df ||
587 		     (IPCB(skb)->frag_max_size &&
588 		      IPCB(skb)->frag_max_size > mtu))) {
589 		IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
590 		icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
591 			  htonl(mtu));
592 		kfree_skb(skb);
593 		return -EMSGSIZE;
594 	}
595 
596 	return ip_do_fragment(net, sk, skb, output);
597 }
598 
599 void ip_fraglist_init(struct sk_buff *skb, struct iphdr *iph,
600 		      unsigned int hlen, struct ip_fraglist_iter *iter)
601 {
602 	unsigned int first_len = skb_pagelen(skb);
603 
604 	iter->frag = skb_shinfo(skb)->frag_list;
605 	skb_frag_list_init(skb);
606 
607 	iter->offset = 0;
608 	iter->iph = iph;
609 	iter->hlen = hlen;
610 
611 	skb->data_len = first_len - skb_headlen(skb);
612 	skb->len = first_len;
613 	iph->tot_len = htons(first_len);
614 	iph->frag_off = htons(IP_MF);
615 	ip_send_check(iph);
616 }
617 EXPORT_SYMBOL(ip_fraglist_init);
618 
619 void ip_fraglist_prepare(struct sk_buff *skb, struct ip_fraglist_iter *iter)
620 {
621 	unsigned int hlen = iter->hlen;
622 	struct iphdr *iph = iter->iph;
623 	struct sk_buff *frag;
624 
625 	frag = iter->frag;
626 	frag->ip_summed = CHECKSUM_NONE;
627 	skb_reset_transport_header(frag);
628 	__skb_push(frag, hlen);
629 	skb_reset_network_header(frag);
630 	memcpy(skb_network_header(frag), iph, hlen);
631 	iter->iph = ip_hdr(frag);
632 	iph = iter->iph;
633 	iph->tot_len = htons(frag->len);
634 	ip_copy_metadata(frag, skb);
635 	iter->offset += skb->len - hlen;
636 	iph->frag_off = htons(iter->offset >> 3);
637 	if (frag->next)
638 		iph->frag_off |= htons(IP_MF);
639 	/* Ready, complete checksum */
640 	ip_send_check(iph);
641 }
642 EXPORT_SYMBOL(ip_fraglist_prepare);
643 
644 void ip_frag_init(struct sk_buff *skb, unsigned int hlen,
645 		  unsigned int ll_rs, unsigned int mtu, bool DF,
646 		  struct ip_frag_state *state)
647 {
648 	struct iphdr *iph = ip_hdr(skb);
649 
650 	state->DF = DF;
651 	state->hlen = hlen;
652 	state->ll_rs = ll_rs;
653 	state->mtu = mtu;
654 
655 	state->left = skb->len - hlen;	/* Space per frame */
656 	state->ptr = hlen;		/* Where to start from */
657 
658 	state->offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
659 	state->not_last_frag = iph->frag_off & htons(IP_MF);
660 }
661 EXPORT_SYMBOL(ip_frag_init);
662 
663 static void ip_frag_ipcb(struct sk_buff *from, struct sk_buff *to,
664 			 bool first_frag)
665 {
666 	/* Copy the flags to each fragment. */
667 	IPCB(to)->flags = IPCB(from)->flags;
668 
669 	/* ANK: dirty, but effective trick. Upgrade options only if
670 	 * the segment to be fragmented was THE FIRST (otherwise,
671 	 * options are already fixed) and make it ONCE
672 	 * on the initial skb, so that all the following fragments
673 	 * will inherit fixed options.
674 	 */
675 	if (first_frag)
676 		ip_options_fragment(from);
677 }
678 
679 struct sk_buff *ip_frag_next(struct sk_buff *skb, struct ip_frag_state *state)
680 {
681 	unsigned int len = state->left;
682 	struct sk_buff *skb2;
683 	struct iphdr *iph;
684 
685 	/* IF: it doesn't fit, use 'mtu' - the data space left */
686 	if (len > state->mtu)
687 		len = state->mtu;
688 	/* IF: we are not sending up to and including the packet end
689 	   then align the next start on an eight byte boundary */
690 	if (len < state->left)	{
691 		len &= ~7;
692 	}
693 
694 	/* Allocate buffer */
695 	skb2 = alloc_skb(len + state->hlen + state->ll_rs, GFP_ATOMIC);
696 	if (!skb2)
697 		return ERR_PTR(-ENOMEM);
698 
699 	/*
700 	 *	Set up data on packet
701 	 */
702 
703 	ip_copy_metadata(skb2, skb);
704 	skb_reserve(skb2, state->ll_rs);
705 	skb_put(skb2, len + state->hlen);
706 	skb_reset_network_header(skb2);
707 	skb2->transport_header = skb2->network_header + state->hlen;
708 
709 	/*
710 	 *	Charge the memory for the fragment to any owner
711 	 *	it might possess
712 	 */
713 
714 	if (skb->sk)
715 		skb_set_owner_w(skb2, skb->sk);
716 
717 	/*
718 	 *	Copy the packet header into the new buffer.
719 	 */
720 
721 	skb_copy_from_linear_data(skb, skb_network_header(skb2), state->hlen);
722 
723 	/*
724 	 *	Copy a block of the IP datagram.
725 	 */
726 	if (skb_copy_bits(skb, state->ptr, skb_transport_header(skb2), len))
727 		BUG();
728 	state->left -= len;
729 
730 	/*
731 	 *	Fill in the new header fields.
732 	 */
733 	iph = ip_hdr(skb2);
734 	iph->frag_off = htons((state->offset >> 3));
735 	if (state->DF)
736 		iph->frag_off |= htons(IP_DF);
737 
738 	/*
739 	 *	Added AC : If we are fragmenting a fragment that's not the
740 	 *		   last fragment then keep MF on each bit
741 	 */
742 	if (state->left > 0 || state->not_last_frag)
743 		iph->frag_off |= htons(IP_MF);
744 	state->ptr += len;
745 	state->offset += len;
746 
747 	iph->tot_len = htons(len + state->hlen);
748 
749 	ip_send_check(iph);
750 
751 	return skb2;
752 }
753 EXPORT_SYMBOL(ip_frag_next);
754 
755 /*
756  *	This IP datagram is too large to be sent in one piece.  Break it up into
757  *	smaller pieces (each of size equal to IP header plus
758  *	a block of the data of the original IP data part) that will yet fit in a
759  *	single device frame, and queue such a frame for sending.
760  */
761 
762 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
763 		   int (*output)(struct net *, struct sock *, struct sk_buff *))
764 {
765 	struct iphdr *iph;
766 	struct sk_buff *skb2;
767 	bool mono_delivery_time = skb->mono_delivery_time;
768 	struct rtable *rt = skb_rtable(skb);
769 	unsigned int mtu, hlen, ll_rs;
770 	struct ip_fraglist_iter iter;
771 	ktime_t tstamp = skb->tstamp;
772 	struct ip_frag_state state;
773 	int err = 0;
774 
775 	/* for offloaded checksums cleanup checksum before fragmentation */
776 	if (skb->ip_summed == CHECKSUM_PARTIAL &&
777 	    (err = skb_checksum_help(skb)))
778 		goto fail;
779 
780 	/*
781 	 *	Point into the IP datagram header.
782 	 */
783 
784 	iph = ip_hdr(skb);
785 
786 	mtu = ip_skb_dst_mtu(sk, skb);
787 	if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
788 		mtu = IPCB(skb)->frag_max_size;
789 
790 	/*
791 	 *	Setup starting values.
792 	 */
793 
794 	hlen = iph->ihl * 4;
795 	mtu = mtu - hlen;	/* Size of data space */
796 	IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
797 	ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
798 
799 	/* When frag_list is given, use it. First, check its validity:
800 	 * some transformers could create wrong frag_list or break existing
801 	 * one, it is not prohibited. In this case fall back to copying.
802 	 *
803 	 * LATER: this step can be merged to real generation of fragments,
804 	 * we can switch to copy when see the first bad fragment.
805 	 */
806 	if (skb_has_frag_list(skb)) {
807 		struct sk_buff *frag, *frag2;
808 		unsigned int first_len = skb_pagelen(skb);
809 
810 		if (first_len - hlen > mtu ||
811 		    ((first_len - hlen) & 7) ||
812 		    ip_is_fragment(iph) ||
813 		    skb_cloned(skb) ||
814 		    skb_headroom(skb) < ll_rs)
815 			goto slow_path;
816 
817 		skb_walk_frags(skb, frag) {
818 			/* Correct geometry. */
819 			if (frag->len > mtu ||
820 			    ((frag->len & 7) && frag->next) ||
821 			    skb_headroom(frag) < hlen + ll_rs)
822 				goto slow_path_clean;
823 
824 			/* Partially cloned skb? */
825 			if (skb_shared(frag))
826 				goto slow_path_clean;
827 
828 			BUG_ON(frag->sk);
829 			if (skb->sk) {
830 				frag->sk = skb->sk;
831 				frag->destructor = sock_wfree;
832 			}
833 			skb->truesize -= frag->truesize;
834 		}
835 
836 		/* Everything is OK. Generate! */
837 		ip_fraglist_init(skb, iph, hlen, &iter);
838 
839 		for (;;) {
840 			/* Prepare header of the next frame,
841 			 * before previous one went down. */
842 			if (iter.frag) {
843 				bool first_frag = (iter.offset == 0);
844 
845 				IPCB(iter.frag)->flags = IPCB(skb)->flags;
846 				ip_fraglist_prepare(skb, &iter);
847 				if (first_frag && IPCB(skb)->opt.optlen) {
848 					/* ipcb->opt is not populated for frags
849 					 * coming from __ip_make_skb(),
850 					 * ip_options_fragment() needs optlen
851 					 */
852 					IPCB(iter.frag)->opt.optlen =
853 						IPCB(skb)->opt.optlen;
854 					ip_options_fragment(iter.frag);
855 					ip_send_check(iter.iph);
856 				}
857 			}
858 
859 			skb_set_delivery_time(skb, tstamp, mono_delivery_time);
860 			err = output(net, sk, skb);
861 
862 			if (!err)
863 				IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
864 			if (err || !iter.frag)
865 				break;
866 
867 			skb = ip_fraglist_next(&iter);
868 		}
869 
870 		if (err == 0) {
871 			IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
872 			return 0;
873 		}
874 
875 		kfree_skb_list(iter.frag);
876 
877 		IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
878 		return err;
879 
880 slow_path_clean:
881 		skb_walk_frags(skb, frag2) {
882 			if (frag2 == frag)
883 				break;
884 			frag2->sk = NULL;
885 			frag2->destructor = NULL;
886 			skb->truesize += frag2->truesize;
887 		}
888 	}
889 
890 slow_path:
891 	/*
892 	 *	Fragment the datagram.
893 	 */
894 
895 	ip_frag_init(skb, hlen, ll_rs, mtu, IPCB(skb)->flags & IPSKB_FRAG_PMTU,
896 		     &state);
897 
898 	/*
899 	 *	Keep copying data until we run out.
900 	 */
901 
902 	while (state.left > 0) {
903 		bool first_frag = (state.offset == 0);
904 
905 		skb2 = ip_frag_next(skb, &state);
906 		if (IS_ERR(skb2)) {
907 			err = PTR_ERR(skb2);
908 			goto fail;
909 		}
910 		ip_frag_ipcb(skb, skb2, first_frag);
911 
912 		/*
913 		 *	Put this fragment into the sending queue.
914 		 */
915 		skb_set_delivery_time(skb2, tstamp, mono_delivery_time);
916 		err = output(net, sk, skb2);
917 		if (err)
918 			goto fail;
919 
920 		IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
921 	}
922 	consume_skb(skb);
923 	IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
924 	return err;
925 
926 fail:
927 	kfree_skb(skb);
928 	IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
929 	return err;
930 }
931 EXPORT_SYMBOL(ip_do_fragment);
932 
933 int
934 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
935 {
936 	struct msghdr *msg = from;
937 
938 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
939 		if (!copy_from_iter_full(to, len, &msg->msg_iter))
940 			return -EFAULT;
941 	} else {
942 		__wsum csum = 0;
943 		if (!csum_and_copy_from_iter_full(to, len, &csum, &msg->msg_iter))
944 			return -EFAULT;
945 		skb->csum = csum_block_add(skb->csum, csum, odd);
946 	}
947 	return 0;
948 }
949 EXPORT_SYMBOL(ip_generic_getfrag);
950 
951 static int __ip_append_data(struct sock *sk,
952 			    struct flowi4 *fl4,
953 			    struct sk_buff_head *queue,
954 			    struct inet_cork *cork,
955 			    struct page_frag *pfrag,
956 			    int getfrag(void *from, char *to, int offset,
957 					int len, int odd, struct sk_buff *skb),
958 			    void *from, int length, int transhdrlen,
959 			    unsigned int flags)
960 {
961 	struct inet_sock *inet = inet_sk(sk);
962 	struct ubuf_info *uarg = NULL;
963 	struct sk_buff *skb;
964 	struct ip_options *opt = cork->opt;
965 	int hh_len;
966 	int exthdrlen;
967 	int mtu;
968 	int copy;
969 	int err;
970 	int offset = 0;
971 	bool zc = false;
972 	unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
973 	int csummode = CHECKSUM_NONE;
974 	struct rtable *rt = (struct rtable *)cork->dst;
975 	unsigned int wmem_alloc_delta = 0;
976 	bool paged, extra_uref = false;
977 	u32 tskey = 0;
978 
979 	skb = skb_peek_tail(queue);
980 
981 	exthdrlen = !skb ? rt->dst.header_len : 0;
982 	mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
983 	paged = !!cork->gso_size;
984 
985 	if (cork->tx_flags & SKBTX_ANY_TSTAMP &&
986 	    READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_OPT_ID)
987 		tskey = atomic_inc_return(&sk->sk_tskey) - 1;
988 
989 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
990 
991 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
992 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
993 	maxnonfragsize = ip_sk_ignore_df(sk) ? IP_MAX_MTU : mtu;
994 
995 	if (cork->length + length > maxnonfragsize - fragheaderlen) {
996 		ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
997 			       mtu - (opt ? opt->optlen : 0));
998 		return -EMSGSIZE;
999 	}
1000 
1001 	/*
1002 	 * transhdrlen > 0 means that this is the first fragment and we wish
1003 	 * it won't be fragmented in the future.
1004 	 */
1005 	if (transhdrlen &&
1006 	    length + fragheaderlen <= mtu &&
1007 	    rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
1008 	    (!(flags & MSG_MORE) || cork->gso_size) &&
1009 	    (!exthdrlen || (rt->dst.dev->features & NETIF_F_HW_ESP_TX_CSUM)))
1010 		csummode = CHECKSUM_PARTIAL;
1011 
1012 	if ((flags & MSG_ZEROCOPY) && length) {
1013 		struct msghdr *msg = from;
1014 
1015 		if (getfrag == ip_generic_getfrag && msg->msg_ubuf) {
1016 			if (skb_zcopy(skb) && msg->msg_ubuf != skb_zcopy(skb))
1017 				return -EINVAL;
1018 
1019 			/* Leave uarg NULL if can't zerocopy, callers should
1020 			 * be able to handle it.
1021 			 */
1022 			if ((rt->dst.dev->features & NETIF_F_SG) &&
1023 			    csummode == CHECKSUM_PARTIAL) {
1024 				paged = true;
1025 				zc = true;
1026 				uarg = msg->msg_ubuf;
1027 			}
1028 		} else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1029 			uarg = msg_zerocopy_realloc(sk, length, skb_zcopy(skb));
1030 			if (!uarg)
1031 				return -ENOBUFS;
1032 			extra_uref = !skb_zcopy(skb);	/* only ref on new uarg */
1033 			if (rt->dst.dev->features & NETIF_F_SG &&
1034 			    csummode == CHECKSUM_PARTIAL) {
1035 				paged = true;
1036 				zc = true;
1037 			} else {
1038 				uarg_to_msgzc(uarg)->zerocopy = 0;
1039 				skb_zcopy_set(skb, uarg, &extra_uref);
1040 			}
1041 		}
1042 	} else if ((flags & MSG_SPLICE_PAGES) && length) {
1043 		if (inet_test_bit(HDRINCL, sk))
1044 			return -EPERM;
1045 		if (rt->dst.dev->features & NETIF_F_SG &&
1046 		    getfrag == ip_generic_getfrag)
1047 			/* We need an empty buffer to attach stuff to */
1048 			paged = true;
1049 		else
1050 			flags &= ~MSG_SPLICE_PAGES;
1051 	}
1052 
1053 	cork->length += length;
1054 
1055 	/* So, what's going on in the loop below?
1056 	 *
1057 	 * We use calculated fragment length to generate chained skb,
1058 	 * each of segments is IP fragment ready for sending to network after
1059 	 * adding appropriate IP header.
1060 	 */
1061 
1062 	if (!skb)
1063 		goto alloc_new_skb;
1064 
1065 	while (length > 0) {
1066 		/* Check if the remaining data fits into current packet. */
1067 		copy = mtu - skb->len;
1068 		if (copy < length)
1069 			copy = maxfraglen - skb->len;
1070 		if (copy <= 0) {
1071 			char *data;
1072 			unsigned int datalen;
1073 			unsigned int fraglen;
1074 			unsigned int fraggap;
1075 			unsigned int alloclen, alloc_extra;
1076 			unsigned int pagedlen;
1077 			struct sk_buff *skb_prev;
1078 alloc_new_skb:
1079 			skb_prev = skb;
1080 			if (skb_prev)
1081 				fraggap = skb_prev->len - maxfraglen;
1082 			else
1083 				fraggap = 0;
1084 
1085 			/*
1086 			 * If remaining data exceeds the mtu,
1087 			 * we know we need more fragment(s).
1088 			 */
1089 			datalen = length + fraggap;
1090 			if (datalen > mtu - fragheaderlen)
1091 				datalen = maxfraglen - fragheaderlen;
1092 			fraglen = datalen + fragheaderlen;
1093 			pagedlen = 0;
1094 
1095 			alloc_extra = hh_len + 15;
1096 			alloc_extra += exthdrlen;
1097 
1098 			/* The last fragment gets additional space at tail.
1099 			 * Note, with MSG_MORE we overallocate on fragments,
1100 			 * because we have no idea what fragment will be
1101 			 * the last.
1102 			 */
1103 			if (datalen == length + fraggap)
1104 				alloc_extra += rt->dst.trailer_len;
1105 
1106 			if ((flags & MSG_MORE) &&
1107 			    !(rt->dst.dev->features&NETIF_F_SG))
1108 				alloclen = mtu;
1109 			else if (!paged &&
1110 				 (fraglen + alloc_extra < SKB_MAX_ALLOC ||
1111 				  !(rt->dst.dev->features & NETIF_F_SG)))
1112 				alloclen = fraglen;
1113 			else {
1114 				alloclen = fragheaderlen + transhdrlen;
1115 				pagedlen = datalen - transhdrlen;
1116 			}
1117 
1118 			alloclen += alloc_extra;
1119 
1120 			if (transhdrlen) {
1121 				skb = sock_alloc_send_skb(sk, alloclen,
1122 						(flags & MSG_DONTWAIT), &err);
1123 			} else {
1124 				skb = NULL;
1125 				if (refcount_read(&sk->sk_wmem_alloc) + wmem_alloc_delta <=
1126 				    2 * sk->sk_sndbuf)
1127 					skb = alloc_skb(alloclen,
1128 							sk->sk_allocation);
1129 				if (unlikely(!skb))
1130 					err = -ENOBUFS;
1131 			}
1132 			if (!skb)
1133 				goto error;
1134 
1135 			/*
1136 			 *	Fill in the control structures
1137 			 */
1138 			skb->ip_summed = csummode;
1139 			skb->csum = 0;
1140 			skb_reserve(skb, hh_len);
1141 
1142 			/*
1143 			 *	Find where to start putting bytes.
1144 			 */
1145 			data = skb_put(skb, fraglen + exthdrlen - pagedlen);
1146 			skb_set_network_header(skb, exthdrlen);
1147 			skb->transport_header = (skb->network_header +
1148 						 fragheaderlen);
1149 			data += fragheaderlen + exthdrlen;
1150 
1151 			if (fraggap) {
1152 				skb->csum = skb_copy_and_csum_bits(
1153 					skb_prev, maxfraglen,
1154 					data + transhdrlen, fraggap);
1155 				skb_prev->csum = csum_sub(skb_prev->csum,
1156 							  skb->csum);
1157 				data += fraggap;
1158 				pskb_trim_unique(skb_prev, maxfraglen);
1159 			}
1160 
1161 			copy = datalen - transhdrlen - fraggap - pagedlen;
1162 			/* [!] NOTE: copy will be negative if pagedlen>0
1163 			 * because then the equation reduces to -fraggap.
1164 			 */
1165 			if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1166 				err = -EFAULT;
1167 				kfree_skb(skb);
1168 				goto error;
1169 			} else if (flags & MSG_SPLICE_PAGES) {
1170 				copy = 0;
1171 			}
1172 
1173 			offset += copy;
1174 			length -= copy + transhdrlen;
1175 			transhdrlen = 0;
1176 			exthdrlen = 0;
1177 			csummode = CHECKSUM_NONE;
1178 
1179 			/* only the initial fragment is time stamped */
1180 			skb_shinfo(skb)->tx_flags = cork->tx_flags;
1181 			cork->tx_flags = 0;
1182 			skb_shinfo(skb)->tskey = tskey;
1183 			tskey = 0;
1184 			skb_zcopy_set(skb, uarg, &extra_uref);
1185 
1186 			if ((flags & MSG_CONFIRM) && !skb_prev)
1187 				skb_set_dst_pending_confirm(skb, 1);
1188 
1189 			/*
1190 			 * Put the packet on the pending queue.
1191 			 */
1192 			if (!skb->destructor) {
1193 				skb->destructor = sock_wfree;
1194 				skb->sk = sk;
1195 				wmem_alloc_delta += skb->truesize;
1196 			}
1197 			__skb_queue_tail(queue, skb);
1198 			continue;
1199 		}
1200 
1201 		if (copy > length)
1202 			copy = length;
1203 
1204 		if (!(rt->dst.dev->features&NETIF_F_SG) &&
1205 		    skb_tailroom(skb) >= copy) {
1206 			unsigned int off;
1207 
1208 			off = skb->len;
1209 			if (getfrag(from, skb_put(skb, copy),
1210 					offset, copy, off, skb) < 0) {
1211 				__skb_trim(skb, off);
1212 				err = -EFAULT;
1213 				goto error;
1214 			}
1215 		} else if (flags & MSG_SPLICE_PAGES) {
1216 			struct msghdr *msg = from;
1217 
1218 			err = -EIO;
1219 			if (WARN_ON_ONCE(copy > msg->msg_iter.count))
1220 				goto error;
1221 
1222 			err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
1223 						   sk->sk_allocation);
1224 			if (err < 0)
1225 				goto error;
1226 			copy = err;
1227 			wmem_alloc_delta += copy;
1228 		} else if (!zc) {
1229 			int i = skb_shinfo(skb)->nr_frags;
1230 
1231 			err = -ENOMEM;
1232 			if (!sk_page_frag_refill(sk, pfrag))
1233 				goto error;
1234 
1235 			skb_zcopy_downgrade_managed(skb);
1236 			if (!skb_can_coalesce(skb, i, pfrag->page,
1237 					      pfrag->offset)) {
1238 				err = -EMSGSIZE;
1239 				if (i == MAX_SKB_FRAGS)
1240 					goto error;
1241 
1242 				__skb_fill_page_desc(skb, i, pfrag->page,
1243 						     pfrag->offset, 0);
1244 				skb_shinfo(skb)->nr_frags = ++i;
1245 				get_page(pfrag->page);
1246 			}
1247 			copy = min_t(int, copy, pfrag->size - pfrag->offset);
1248 			if (getfrag(from,
1249 				    page_address(pfrag->page) + pfrag->offset,
1250 				    offset, copy, skb->len, skb) < 0)
1251 				goto error_efault;
1252 
1253 			pfrag->offset += copy;
1254 			skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1255 			skb_len_add(skb, copy);
1256 			wmem_alloc_delta += copy;
1257 		} else {
1258 			err = skb_zerocopy_iter_dgram(skb, from, copy);
1259 			if (err < 0)
1260 				goto error;
1261 		}
1262 		offset += copy;
1263 		length -= copy;
1264 	}
1265 
1266 	if (wmem_alloc_delta)
1267 		refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1268 	return 0;
1269 
1270 error_efault:
1271 	err = -EFAULT;
1272 error:
1273 	net_zcopy_put_abort(uarg, extra_uref);
1274 	cork->length -= length;
1275 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1276 	refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1277 	return err;
1278 }
1279 
1280 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1281 			 struct ipcm_cookie *ipc, struct rtable **rtp)
1282 {
1283 	struct ip_options_rcu *opt;
1284 	struct rtable *rt;
1285 
1286 	rt = *rtp;
1287 	if (unlikely(!rt))
1288 		return -EFAULT;
1289 
1290 	cork->fragsize = ip_sk_use_pmtu(sk) ?
1291 			 dst_mtu(&rt->dst) : READ_ONCE(rt->dst.dev->mtu);
1292 
1293 	if (!inetdev_valid_mtu(cork->fragsize))
1294 		return -ENETUNREACH;
1295 
1296 	/*
1297 	 * setup for corking.
1298 	 */
1299 	opt = ipc->opt;
1300 	if (opt) {
1301 		if (!cork->opt) {
1302 			cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1303 					    sk->sk_allocation);
1304 			if (unlikely(!cork->opt))
1305 				return -ENOBUFS;
1306 		}
1307 		memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1308 		cork->flags |= IPCORK_OPT;
1309 		cork->addr = ipc->addr;
1310 	}
1311 
1312 	cork->gso_size = ipc->gso_size;
1313 
1314 	cork->dst = &rt->dst;
1315 	/* We stole this route, caller should not release it. */
1316 	*rtp = NULL;
1317 
1318 	cork->length = 0;
1319 	cork->ttl = ipc->ttl;
1320 	cork->tos = ipc->tos;
1321 	cork->mark = ipc->sockc.mark;
1322 	cork->priority = ipc->priority;
1323 	cork->transmit_time = ipc->sockc.transmit_time;
1324 	cork->tx_flags = 0;
1325 	sock_tx_timestamp(sk, ipc->sockc.tsflags, &cork->tx_flags);
1326 
1327 	return 0;
1328 }
1329 
1330 /*
1331  *	ip_append_data() can make one large IP datagram from many pieces of
1332  *	data.  Each piece will be held on the socket until
1333  *	ip_push_pending_frames() is called. Each piece can be a page or
1334  *	non-page data.
1335  *
1336  *	Not only UDP, other transport protocols - e.g. raw sockets - can use
1337  *	this interface potentially.
1338  *
1339  *	LATER: length must be adjusted by pad at tail, when it is required.
1340  */
1341 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1342 		   int getfrag(void *from, char *to, int offset, int len,
1343 			       int odd, struct sk_buff *skb),
1344 		   void *from, int length, int transhdrlen,
1345 		   struct ipcm_cookie *ipc, struct rtable **rtp,
1346 		   unsigned int flags)
1347 {
1348 	struct inet_sock *inet = inet_sk(sk);
1349 	int err;
1350 
1351 	if (flags&MSG_PROBE)
1352 		return 0;
1353 
1354 	if (skb_queue_empty(&sk->sk_write_queue)) {
1355 		err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1356 		if (err)
1357 			return err;
1358 	} else {
1359 		transhdrlen = 0;
1360 	}
1361 
1362 	return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1363 				sk_page_frag(sk), getfrag,
1364 				from, length, transhdrlen, flags);
1365 }
1366 
1367 static void ip_cork_release(struct inet_cork *cork)
1368 {
1369 	cork->flags &= ~IPCORK_OPT;
1370 	kfree(cork->opt);
1371 	cork->opt = NULL;
1372 	dst_release(cork->dst);
1373 	cork->dst = NULL;
1374 }
1375 
1376 /*
1377  *	Combined all pending IP fragments on the socket as one IP datagram
1378  *	and push them out.
1379  */
1380 struct sk_buff *__ip_make_skb(struct sock *sk,
1381 			      struct flowi4 *fl4,
1382 			      struct sk_buff_head *queue,
1383 			      struct inet_cork *cork)
1384 {
1385 	struct sk_buff *skb, *tmp_skb;
1386 	struct sk_buff **tail_skb;
1387 	struct inet_sock *inet = inet_sk(sk);
1388 	struct net *net = sock_net(sk);
1389 	struct ip_options *opt = NULL;
1390 	struct rtable *rt = (struct rtable *)cork->dst;
1391 	struct iphdr *iph;
1392 	u8 pmtudisc, ttl;
1393 	__be16 df = 0;
1394 
1395 	skb = __skb_dequeue(queue);
1396 	if (!skb)
1397 		goto out;
1398 	tail_skb = &(skb_shinfo(skb)->frag_list);
1399 
1400 	/* move skb->data to ip header from ext header */
1401 	if (skb->data < skb_network_header(skb))
1402 		__skb_pull(skb, skb_network_offset(skb));
1403 	while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1404 		__skb_pull(tmp_skb, skb_network_header_len(skb));
1405 		*tail_skb = tmp_skb;
1406 		tail_skb = &(tmp_skb->next);
1407 		skb->len += tmp_skb->len;
1408 		skb->data_len += tmp_skb->len;
1409 		skb->truesize += tmp_skb->truesize;
1410 		tmp_skb->destructor = NULL;
1411 		tmp_skb->sk = NULL;
1412 	}
1413 
1414 	/* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1415 	 * to fragment the frame generated here. No matter, what transforms
1416 	 * how transforms change size of the packet, it will come out.
1417 	 */
1418 	skb->ignore_df = ip_sk_ignore_df(sk);
1419 
1420 	/* DF bit is set when we want to see DF on outgoing frames.
1421 	 * If ignore_df is set too, we still allow to fragment this frame
1422 	 * locally. */
1423 	pmtudisc = READ_ONCE(inet->pmtudisc);
1424 	if (pmtudisc == IP_PMTUDISC_DO ||
1425 	    pmtudisc == IP_PMTUDISC_PROBE ||
1426 	    (skb->len <= dst_mtu(&rt->dst) &&
1427 	     ip_dont_fragment(sk, &rt->dst)))
1428 		df = htons(IP_DF);
1429 
1430 	if (cork->flags & IPCORK_OPT)
1431 		opt = cork->opt;
1432 
1433 	if (cork->ttl != 0)
1434 		ttl = cork->ttl;
1435 	else if (rt->rt_type == RTN_MULTICAST)
1436 		ttl = READ_ONCE(inet->mc_ttl);
1437 	else
1438 		ttl = ip_select_ttl(inet, &rt->dst);
1439 
1440 	iph = ip_hdr(skb);
1441 	iph->version = 4;
1442 	iph->ihl = 5;
1443 	iph->tos = (cork->tos != -1) ? cork->tos : READ_ONCE(inet->tos);
1444 	iph->frag_off = df;
1445 	iph->ttl = ttl;
1446 	iph->protocol = sk->sk_protocol;
1447 	ip_copy_addrs(iph, fl4);
1448 	ip_select_ident(net, skb, sk);
1449 
1450 	if (opt) {
1451 		iph->ihl += opt->optlen >> 2;
1452 		ip_options_build(skb, opt, cork->addr, rt);
1453 	}
1454 
1455 	skb->priority = (cork->tos != -1) ? cork->priority: READ_ONCE(sk->sk_priority);
1456 	skb->mark = cork->mark;
1457 	skb->tstamp = cork->transmit_time;
1458 	/*
1459 	 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1460 	 * on dst refcount
1461 	 */
1462 	cork->dst = NULL;
1463 	skb_dst_set(skb, &rt->dst);
1464 
1465 	if (iph->protocol == IPPROTO_ICMP) {
1466 		u8 icmp_type;
1467 
1468 		/* For such sockets, transhdrlen is zero when do ip_append_data(),
1469 		 * so icmphdr does not in skb linear region and can not get icmp_type
1470 		 * by icmp_hdr(skb)->type.
1471 		 */
1472 		if (sk->sk_type == SOCK_RAW &&
1473 		    !inet_test_bit(HDRINCL, sk))
1474 			icmp_type = fl4->fl4_icmp_type;
1475 		else
1476 			icmp_type = icmp_hdr(skb)->type;
1477 		icmp_out_count(net, icmp_type);
1478 	}
1479 
1480 	ip_cork_release(cork);
1481 out:
1482 	return skb;
1483 }
1484 
1485 int ip_send_skb(struct net *net, struct sk_buff *skb)
1486 {
1487 	int err;
1488 
1489 	err = ip_local_out(net, skb->sk, skb);
1490 	if (err) {
1491 		if (err > 0)
1492 			err = net_xmit_errno(err);
1493 		if (err)
1494 			IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1495 	}
1496 
1497 	return err;
1498 }
1499 
1500 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1501 {
1502 	struct sk_buff *skb;
1503 
1504 	skb = ip_finish_skb(sk, fl4);
1505 	if (!skb)
1506 		return 0;
1507 
1508 	/* Netfilter gets whole the not fragmented skb. */
1509 	return ip_send_skb(sock_net(sk), skb);
1510 }
1511 
1512 /*
1513  *	Throw away all pending data on the socket.
1514  */
1515 static void __ip_flush_pending_frames(struct sock *sk,
1516 				      struct sk_buff_head *queue,
1517 				      struct inet_cork *cork)
1518 {
1519 	struct sk_buff *skb;
1520 
1521 	while ((skb = __skb_dequeue_tail(queue)) != NULL)
1522 		kfree_skb(skb);
1523 
1524 	ip_cork_release(cork);
1525 }
1526 
1527 void ip_flush_pending_frames(struct sock *sk)
1528 {
1529 	__ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1530 }
1531 
1532 struct sk_buff *ip_make_skb(struct sock *sk,
1533 			    struct flowi4 *fl4,
1534 			    int getfrag(void *from, char *to, int offset,
1535 					int len, int odd, struct sk_buff *skb),
1536 			    void *from, int length, int transhdrlen,
1537 			    struct ipcm_cookie *ipc, struct rtable **rtp,
1538 			    struct inet_cork *cork, unsigned int flags)
1539 {
1540 	struct sk_buff_head queue;
1541 	int err;
1542 
1543 	if (flags & MSG_PROBE)
1544 		return NULL;
1545 
1546 	__skb_queue_head_init(&queue);
1547 
1548 	cork->flags = 0;
1549 	cork->addr = 0;
1550 	cork->opt = NULL;
1551 	err = ip_setup_cork(sk, cork, ipc, rtp);
1552 	if (err)
1553 		return ERR_PTR(err);
1554 
1555 	err = __ip_append_data(sk, fl4, &queue, cork,
1556 			       &current->task_frag, getfrag,
1557 			       from, length, transhdrlen, flags);
1558 	if (err) {
1559 		__ip_flush_pending_frames(sk, &queue, cork);
1560 		return ERR_PTR(err);
1561 	}
1562 
1563 	return __ip_make_skb(sk, fl4, &queue, cork);
1564 }
1565 
1566 /*
1567  *	Fetch data from kernel space and fill in checksum if needed.
1568  */
1569 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1570 			      int len, int odd, struct sk_buff *skb)
1571 {
1572 	__wsum csum;
1573 
1574 	csum = csum_partial_copy_nocheck(dptr+offset, to, len);
1575 	skb->csum = csum_block_add(skb->csum, csum, odd);
1576 	return 0;
1577 }
1578 
1579 /*
1580  *	Generic function to send a packet as reply to another packet.
1581  *	Used to send some TCP resets/acks so far.
1582  */
1583 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1584 			   const struct ip_options *sopt,
1585 			   __be32 daddr, __be32 saddr,
1586 			   const struct ip_reply_arg *arg,
1587 			   unsigned int len, u64 transmit_time, u32 txhash)
1588 {
1589 	struct ip_options_data replyopts;
1590 	struct ipcm_cookie ipc;
1591 	struct flowi4 fl4;
1592 	struct rtable *rt = skb_rtable(skb);
1593 	struct net *net = sock_net(sk);
1594 	struct sk_buff *nskb;
1595 	int err;
1596 	int oif;
1597 
1598 	if (__ip_options_echo(net, &replyopts.opt.opt, skb, sopt))
1599 		return;
1600 
1601 	ipcm_init(&ipc);
1602 	ipc.addr = daddr;
1603 	ipc.sockc.transmit_time = transmit_time;
1604 
1605 	if (replyopts.opt.opt.optlen) {
1606 		ipc.opt = &replyopts.opt;
1607 
1608 		if (replyopts.opt.opt.srr)
1609 			daddr = replyopts.opt.opt.faddr;
1610 	}
1611 
1612 	oif = arg->bound_dev_if;
1613 	if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1614 		oif = skb->skb_iif;
1615 
1616 	flowi4_init_output(&fl4, oif,
1617 			   IP4_REPLY_MARK(net, skb->mark) ?: sk->sk_mark,
1618 			   RT_TOS(arg->tos),
1619 			   RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1620 			   ip_reply_arg_flowi_flags(arg),
1621 			   daddr, saddr,
1622 			   tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1623 			   arg->uid);
1624 	security_skb_classify_flow(skb, flowi4_to_flowi_common(&fl4));
1625 	rt = ip_route_output_flow(net, &fl4, sk);
1626 	if (IS_ERR(rt))
1627 		return;
1628 
1629 	inet_sk(sk)->tos = arg->tos & ~INET_ECN_MASK;
1630 
1631 	sk->sk_protocol = ip_hdr(skb)->protocol;
1632 	sk->sk_bound_dev_if = arg->bound_dev_if;
1633 	sk->sk_sndbuf = READ_ONCE(sysctl_wmem_default);
1634 	ipc.sockc.mark = fl4.flowi4_mark;
1635 	err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1636 			     len, 0, &ipc, &rt, MSG_DONTWAIT);
1637 	if (unlikely(err)) {
1638 		ip_flush_pending_frames(sk);
1639 		goto out;
1640 	}
1641 
1642 	nskb = skb_peek(&sk->sk_write_queue);
1643 	if (nskb) {
1644 		if (arg->csumoffset >= 0)
1645 			*((__sum16 *)skb_transport_header(nskb) +
1646 			  arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1647 								arg->csum));
1648 		nskb->ip_summed = CHECKSUM_NONE;
1649 		nskb->mono_delivery_time = !!transmit_time;
1650 		if (txhash)
1651 			skb_set_hash(nskb, txhash, PKT_HASH_TYPE_L4);
1652 		ip_push_pending_frames(sk, &fl4);
1653 	}
1654 out:
1655 	ip_rt_put(rt);
1656 }
1657 
1658 void __init ip_init(void)
1659 {
1660 	ip_rt_init();
1661 	inet_initpeers();
1662 
1663 #if defined(CONFIG_IP_MULTICAST)
1664 	igmp_mc_init();
1665 #endif
1666 }
1667