xref: /openbmc/linux/net/ipv4/ip_output.c (revision db7193a5)
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 	iph_set_totlen(iph, skb->len);
105 	ip_send_check(iph);
106 
107 	/* if egress device is enslaved to an L3 master device pass the
108 	 * skb to its handler for processing
109 	 */
110 	skb = l3mdev_ip_out(sk, skb);
111 	if (unlikely(!skb))
112 		return 0;
113 
114 	skb->protocol = htons(ETH_P_IP);
115 
116 	return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
117 		       net, sk, skb, NULL, skb_dst(skb)->dev,
118 		       dst_output);
119 }
120 
121 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
122 {
123 	int err;
124 
125 	err = __ip_local_out(net, sk, skb);
126 	if (likely(err == 1))
127 		err = dst_output(net, sk, skb);
128 
129 	return err;
130 }
131 EXPORT_SYMBOL_GPL(ip_local_out);
132 
133 static inline int ip_select_ttl(const struct inet_sock *inet,
134 				const struct dst_entry *dst)
135 {
136 	int ttl = inet->uc_ttl;
137 
138 	if (ttl < 0)
139 		ttl = ip4_dst_hoplimit(dst);
140 	return ttl;
141 }
142 
143 /*
144  *		Add an ip header to a skbuff and send it out.
145  *
146  */
147 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
148 			  __be32 saddr, __be32 daddr, struct ip_options_rcu *opt,
149 			  u8 tos)
150 {
151 	const struct inet_sock *inet = inet_sk(sk);
152 	struct rtable *rt = skb_rtable(skb);
153 	struct net *net = sock_net(sk);
154 	struct iphdr *iph;
155 
156 	/* Build the IP header. */
157 	skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
158 	skb_reset_network_header(skb);
159 	iph = ip_hdr(skb);
160 	iph->version  = 4;
161 	iph->ihl      = 5;
162 	iph->tos      = tos;
163 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
164 	iph->daddr    = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
165 	iph->saddr    = saddr;
166 	iph->protocol = sk->sk_protocol;
167 	/* Do not bother generating IPID for small packets (eg SYNACK) */
168 	if (skb->len <= IPV4_MIN_MTU || ip_dont_fragment(sk, &rt->dst)) {
169 		iph->frag_off = htons(IP_DF);
170 		iph->id = 0;
171 	} else {
172 		iph->frag_off = 0;
173 		/* TCP packets here are SYNACK with fat IPv4/TCP options.
174 		 * Avoid using the hashed IP ident generator.
175 		 */
176 		if (sk->sk_protocol == IPPROTO_TCP)
177 			iph->id = (__force __be16)get_random_u16();
178 		else
179 			__ip_select_ident(net, iph, 1);
180 	}
181 
182 	if (opt && opt->opt.optlen) {
183 		iph->ihl += opt->opt.optlen>>2;
184 		ip_options_build(skb, &opt->opt, daddr, rt);
185 	}
186 
187 	skb->priority = READ_ONCE(sk->sk_priority);
188 	if (!skb->mark)
189 		skb->mark = READ_ONCE(sk->sk_mark);
190 
191 	/* Send it out. */
192 	return ip_local_out(net, skb->sk, skb);
193 }
194 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
195 
196 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
197 {
198 	struct dst_entry *dst = skb_dst(skb);
199 	struct rtable *rt = (struct rtable *)dst;
200 	struct net_device *dev = dst->dev;
201 	unsigned int hh_len = LL_RESERVED_SPACE(dev);
202 	struct neighbour *neigh;
203 	bool is_v6gw = false;
204 
205 	if (rt->rt_type == RTN_MULTICAST) {
206 		IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
207 	} else if (rt->rt_type == RTN_BROADCAST)
208 		IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
209 
210 	if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
211 		skb = skb_expand_head(skb, hh_len);
212 		if (!skb)
213 			return -ENOMEM;
214 	}
215 
216 	if (lwtunnel_xmit_redirect(dst->lwtstate)) {
217 		int res = lwtunnel_xmit(skb);
218 
219 		if (res < 0 || res == LWTUNNEL_XMIT_DONE)
220 			return res;
221 	}
222 
223 	rcu_read_lock();
224 	neigh = ip_neigh_for_gw(rt, skb, &is_v6gw);
225 	if (!IS_ERR(neigh)) {
226 		int res;
227 
228 		sock_confirm_neigh(skb, neigh);
229 		/* if crossing protocols, can not use the cached header */
230 		res = neigh_output(neigh, skb, is_v6gw);
231 		rcu_read_unlock();
232 		return res;
233 	}
234 	rcu_read_unlock();
235 
236 	net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
237 			    __func__);
238 	kfree_skb_reason(skb, SKB_DROP_REASON_NEIGH_CREATEFAIL);
239 	return -EINVAL;
240 }
241 
242 static int ip_finish_output_gso(struct net *net, struct sock *sk,
243 				struct sk_buff *skb, unsigned int mtu)
244 {
245 	struct sk_buff *segs, *nskb;
246 	netdev_features_t features;
247 	int ret = 0;
248 
249 	/* common case: seglen is <= mtu
250 	 */
251 	if (skb_gso_validate_network_len(skb, mtu))
252 		return ip_finish_output2(net, sk, skb);
253 
254 	/* Slowpath -  GSO segment length exceeds the egress MTU.
255 	 *
256 	 * This can happen in several cases:
257 	 *  - Forwarding of a TCP GRO skb, when DF flag is not set.
258 	 *  - Forwarding of an skb that arrived on a virtualization interface
259 	 *    (virtio-net/vhost/tap) with TSO/GSO size set by other network
260 	 *    stack.
261 	 *  - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
262 	 *    interface with a smaller MTU.
263 	 *  - Arriving GRO skb (or GSO skb in a virtualized environment) that is
264 	 *    bridged to a NETIF_F_TSO tunnel stacked over an interface with an
265 	 *    insufficient MTU.
266 	 */
267 	features = netif_skb_features(skb);
268 	BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_GSO_CB_OFFSET);
269 	segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
270 	if (IS_ERR_OR_NULL(segs)) {
271 		kfree_skb(skb);
272 		return -ENOMEM;
273 	}
274 
275 	consume_skb(skb);
276 
277 	skb_list_walk_safe(segs, segs, nskb) {
278 		int err;
279 
280 		skb_mark_not_on_list(segs);
281 		err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
282 
283 		if (err && ret == 0)
284 			ret = err;
285 	}
286 
287 	return ret;
288 }
289 
290 static int __ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
291 {
292 	unsigned int mtu;
293 
294 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
295 	/* Policy lookup after SNAT yielded a new policy */
296 	if (skb_dst(skb)->xfrm) {
297 		IPCB(skb)->flags |= IPSKB_REROUTED;
298 		return dst_output(net, sk, skb);
299 	}
300 #endif
301 	mtu = ip_skb_dst_mtu(sk, skb);
302 	if (skb_is_gso(skb))
303 		return ip_finish_output_gso(net, sk, skb, mtu);
304 
305 	if (skb->len > mtu || IPCB(skb)->frag_max_size)
306 		return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
307 
308 	return ip_finish_output2(net, sk, skb);
309 }
310 
311 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
312 {
313 	int ret;
314 
315 	ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
316 	switch (ret) {
317 	case NET_XMIT_SUCCESS:
318 		return __ip_finish_output(net, sk, skb);
319 	case NET_XMIT_CN:
320 		return __ip_finish_output(net, sk, skb) ? : ret;
321 	default:
322 		kfree_skb_reason(skb, SKB_DROP_REASON_BPF_CGROUP_EGRESS);
323 		return ret;
324 	}
325 }
326 
327 static int ip_mc_finish_output(struct net *net, struct sock *sk,
328 			       struct sk_buff *skb)
329 {
330 	struct rtable *new_rt;
331 	bool do_cn = false;
332 	int ret, err;
333 
334 	ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
335 	switch (ret) {
336 	case NET_XMIT_CN:
337 		do_cn = true;
338 		fallthrough;
339 	case NET_XMIT_SUCCESS:
340 		break;
341 	default:
342 		kfree_skb_reason(skb, SKB_DROP_REASON_BPF_CGROUP_EGRESS);
343 		return ret;
344 	}
345 
346 	/* Reset rt_iif so that inet_iif() will return skb->skb_iif. Setting
347 	 * this to non-zero causes ipi_ifindex in in_pktinfo to be overwritten,
348 	 * see ipv4_pktinfo_prepare().
349 	 */
350 	new_rt = rt_dst_clone(net->loopback_dev, skb_rtable(skb));
351 	if (new_rt) {
352 		new_rt->rt_iif = 0;
353 		skb_dst_drop(skb);
354 		skb_dst_set(skb, &new_rt->dst);
355 	}
356 
357 	err = dev_loopback_xmit(net, sk, skb);
358 	return (do_cn && err) ? ret : err;
359 }
360 
361 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
362 {
363 	struct rtable *rt = skb_rtable(skb);
364 	struct net_device *dev = rt->dst.dev;
365 
366 	/*
367 	 *	If the indicated interface is up and running, send the packet.
368 	 */
369 	IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
370 
371 	skb->dev = dev;
372 	skb->protocol = htons(ETH_P_IP);
373 
374 	/*
375 	 *	Multicasts are looped back for other local users
376 	 */
377 
378 	if (rt->rt_flags&RTCF_MULTICAST) {
379 		if (sk_mc_loop(sk)
380 #ifdef CONFIG_IP_MROUTE
381 		/* Small optimization: do not loopback not local frames,
382 		   which returned after forwarding; they will be  dropped
383 		   by ip_mr_input in any case.
384 		   Note, that local frames are looped back to be delivered
385 		   to local recipients.
386 
387 		   This check is duplicated in ip_mr_input at the moment.
388 		 */
389 		    &&
390 		    ((rt->rt_flags & RTCF_LOCAL) ||
391 		     !(IPCB(skb)->flags & IPSKB_FORWARDED))
392 #endif
393 		   ) {
394 			struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
395 			if (newskb)
396 				NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
397 					net, sk, newskb, NULL, newskb->dev,
398 					ip_mc_finish_output);
399 		}
400 
401 		/* Multicasts with ttl 0 must not go beyond the host */
402 
403 		if (ip_hdr(skb)->ttl == 0) {
404 			kfree_skb(skb);
405 			return 0;
406 		}
407 	}
408 
409 	if (rt->rt_flags&RTCF_BROADCAST) {
410 		struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
411 		if (newskb)
412 			NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
413 				net, sk, newskb, NULL, newskb->dev,
414 				ip_mc_finish_output);
415 	}
416 
417 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
418 			    net, sk, skb, NULL, skb->dev,
419 			    ip_finish_output,
420 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
421 }
422 
423 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
424 {
425 	struct net_device *dev = skb_dst(skb)->dev, *indev = skb->dev;
426 
427 	IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
428 
429 	skb->dev = dev;
430 	skb->protocol = htons(ETH_P_IP);
431 
432 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
433 			    net, sk, skb, indev, dev,
434 			    ip_finish_output,
435 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
436 }
437 EXPORT_SYMBOL(ip_output);
438 
439 /*
440  * copy saddr and daddr, possibly using 64bit load/stores
441  * Equivalent to :
442  *   iph->saddr = fl4->saddr;
443  *   iph->daddr = fl4->daddr;
444  */
445 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
446 {
447 	BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
448 		     offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
449 
450 	iph->saddr = fl4->saddr;
451 	iph->daddr = fl4->daddr;
452 }
453 
454 /* Note: skb->sk can be different from sk, in case of tunnels */
455 int __ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl,
456 		    __u8 tos)
457 {
458 	struct inet_sock *inet = inet_sk(sk);
459 	struct net *net = sock_net(sk);
460 	struct ip_options_rcu *inet_opt;
461 	struct flowi4 *fl4;
462 	struct rtable *rt;
463 	struct iphdr *iph;
464 	int res;
465 
466 	/* Skip all of this if the packet is already routed,
467 	 * f.e. by something like SCTP.
468 	 */
469 	rcu_read_lock();
470 	inet_opt = rcu_dereference(inet->inet_opt);
471 	fl4 = &fl->u.ip4;
472 	rt = skb_rtable(skb);
473 	if (rt)
474 		goto packet_routed;
475 
476 	/* Make sure we can route this packet. */
477 	rt = (struct rtable *)__sk_dst_check(sk, 0);
478 	if (!rt) {
479 		__be32 daddr;
480 
481 		/* Use correct destination address if we have options. */
482 		daddr = inet->inet_daddr;
483 		if (inet_opt && inet_opt->opt.srr)
484 			daddr = inet_opt->opt.faddr;
485 
486 		/* If this fails, retransmit mechanism of transport layer will
487 		 * keep trying until route appears or the connection times
488 		 * itself out.
489 		 */
490 		rt = ip_route_output_ports(net, fl4, sk,
491 					   daddr, inet->inet_saddr,
492 					   inet->inet_dport,
493 					   inet->inet_sport,
494 					   sk->sk_protocol,
495 					   RT_CONN_FLAGS_TOS(sk, tos),
496 					   sk->sk_bound_dev_if);
497 		if (IS_ERR(rt))
498 			goto no_route;
499 		sk_setup_caps(sk, &rt->dst);
500 	}
501 	skb_dst_set_noref(skb, &rt->dst);
502 
503 packet_routed:
504 	if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
505 		goto no_route;
506 
507 	/* OK, we know where to send it, allocate and build IP header. */
508 	skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
509 	skb_reset_network_header(skb);
510 	iph = ip_hdr(skb);
511 	*((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (tos & 0xff));
512 	if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
513 		iph->frag_off = htons(IP_DF);
514 	else
515 		iph->frag_off = 0;
516 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
517 	iph->protocol = sk->sk_protocol;
518 	ip_copy_addrs(iph, fl4);
519 
520 	/* Transport layer set skb->h.foo itself. */
521 
522 	if (inet_opt && inet_opt->opt.optlen) {
523 		iph->ihl += inet_opt->opt.optlen >> 2;
524 		ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt);
525 	}
526 
527 	ip_select_ident_segs(net, skb, sk,
528 			     skb_shinfo(skb)->gso_segs ?: 1);
529 
530 	/* TODO : should we use skb->sk here instead of sk ? */
531 	skb->priority = READ_ONCE(sk->sk_priority);
532 	skb->mark = READ_ONCE(sk->sk_mark);
533 
534 	res = ip_local_out(net, sk, skb);
535 	rcu_read_unlock();
536 	return res;
537 
538 no_route:
539 	rcu_read_unlock();
540 	IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
541 	kfree_skb_reason(skb, SKB_DROP_REASON_IP_OUTNOROUTES);
542 	return -EHOSTUNREACH;
543 }
544 EXPORT_SYMBOL(__ip_queue_xmit);
545 
546 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
547 {
548 	return __ip_queue_xmit(sk, skb, fl, inet_sk(sk)->tos);
549 }
550 EXPORT_SYMBOL(ip_queue_xmit);
551 
552 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
553 {
554 	to->pkt_type = from->pkt_type;
555 	to->priority = from->priority;
556 	to->protocol = from->protocol;
557 	to->skb_iif = from->skb_iif;
558 	skb_dst_drop(to);
559 	skb_dst_copy(to, from);
560 	to->dev = from->dev;
561 	to->mark = from->mark;
562 
563 	skb_copy_hash(to, from);
564 
565 #ifdef CONFIG_NET_SCHED
566 	to->tc_index = from->tc_index;
567 #endif
568 	nf_copy(to, from);
569 	skb_ext_copy(to, from);
570 #if IS_ENABLED(CONFIG_IP_VS)
571 	to->ipvs_property = from->ipvs_property;
572 #endif
573 	skb_copy_secmark(to, from);
574 }
575 
576 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
577 		       unsigned int mtu,
578 		       int (*output)(struct net *, struct sock *, struct sk_buff *))
579 {
580 	struct iphdr *iph = ip_hdr(skb);
581 
582 	if ((iph->frag_off & htons(IP_DF)) == 0)
583 		return ip_do_fragment(net, sk, skb, output);
584 
585 	if (unlikely(!skb->ignore_df ||
586 		     (IPCB(skb)->frag_max_size &&
587 		      IPCB(skb)->frag_max_size > mtu))) {
588 		IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
589 		icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
590 			  htonl(mtu));
591 		kfree_skb(skb);
592 		return -EMSGSIZE;
593 	}
594 
595 	return ip_do_fragment(net, sk, skb, output);
596 }
597 
598 void ip_fraglist_init(struct sk_buff *skb, struct iphdr *iph,
599 		      unsigned int hlen, struct ip_fraglist_iter *iter)
600 {
601 	unsigned int first_len = skb_pagelen(skb);
602 
603 	iter->frag = skb_shinfo(skb)->frag_list;
604 	skb_frag_list_init(skb);
605 
606 	iter->offset = 0;
607 	iter->iph = iph;
608 	iter->hlen = hlen;
609 
610 	skb->data_len = first_len - skb_headlen(skb);
611 	skb->len = first_len;
612 	iph->tot_len = htons(first_len);
613 	iph->frag_off = htons(IP_MF);
614 	ip_send_check(iph);
615 }
616 EXPORT_SYMBOL(ip_fraglist_init);
617 
618 void ip_fraglist_prepare(struct sk_buff *skb, struct ip_fraglist_iter *iter)
619 {
620 	unsigned int hlen = iter->hlen;
621 	struct iphdr *iph = iter->iph;
622 	struct sk_buff *frag;
623 
624 	frag = iter->frag;
625 	frag->ip_summed = CHECKSUM_NONE;
626 	skb_reset_transport_header(frag);
627 	__skb_push(frag, hlen);
628 	skb_reset_network_header(frag);
629 	memcpy(skb_network_header(frag), iph, hlen);
630 	iter->iph = ip_hdr(frag);
631 	iph = iter->iph;
632 	iph->tot_len = htons(frag->len);
633 	ip_copy_metadata(frag, skb);
634 	iter->offset += skb->len - hlen;
635 	iph->frag_off = htons(iter->offset >> 3);
636 	if (frag->next)
637 		iph->frag_off |= htons(IP_MF);
638 	/* Ready, complete checksum */
639 	ip_send_check(iph);
640 }
641 EXPORT_SYMBOL(ip_fraglist_prepare);
642 
643 void ip_frag_init(struct sk_buff *skb, unsigned int hlen,
644 		  unsigned int ll_rs, unsigned int mtu, bool DF,
645 		  struct ip_frag_state *state)
646 {
647 	struct iphdr *iph = ip_hdr(skb);
648 
649 	state->DF = DF;
650 	state->hlen = hlen;
651 	state->ll_rs = ll_rs;
652 	state->mtu = mtu;
653 
654 	state->left = skb->len - hlen;	/* Space per frame */
655 	state->ptr = hlen;		/* Where to start from */
656 
657 	state->offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
658 	state->not_last_frag = iph->frag_off & htons(IP_MF);
659 }
660 EXPORT_SYMBOL(ip_frag_init);
661 
662 static void ip_frag_ipcb(struct sk_buff *from, struct sk_buff *to,
663 			 bool first_frag)
664 {
665 	/* Copy the flags to each fragment. */
666 	IPCB(to)->flags = IPCB(from)->flags;
667 
668 	/* ANK: dirty, but effective trick. Upgrade options only if
669 	 * the segment to be fragmented was THE FIRST (otherwise,
670 	 * options are already fixed) and make it ONCE
671 	 * on the initial skb, so that all the following fragments
672 	 * will inherit fixed options.
673 	 */
674 	if (first_frag)
675 		ip_options_fragment(from);
676 }
677 
678 struct sk_buff *ip_frag_next(struct sk_buff *skb, struct ip_frag_state *state)
679 {
680 	unsigned int len = state->left;
681 	struct sk_buff *skb2;
682 	struct iphdr *iph;
683 
684 	/* IF: it doesn't fit, use 'mtu' - the data space left */
685 	if (len > state->mtu)
686 		len = state->mtu;
687 	/* IF: we are not sending up to and including the packet end
688 	   then align the next start on an eight byte boundary */
689 	if (len < state->left)	{
690 		len &= ~7;
691 	}
692 
693 	/* Allocate buffer */
694 	skb2 = alloc_skb(len + state->hlen + state->ll_rs, GFP_ATOMIC);
695 	if (!skb2)
696 		return ERR_PTR(-ENOMEM);
697 
698 	/*
699 	 *	Set up data on packet
700 	 */
701 
702 	ip_copy_metadata(skb2, skb);
703 	skb_reserve(skb2, state->ll_rs);
704 	skb_put(skb2, len + state->hlen);
705 	skb_reset_network_header(skb2);
706 	skb2->transport_header = skb2->network_header + state->hlen;
707 
708 	/*
709 	 *	Charge the memory for the fragment to any owner
710 	 *	it might possess
711 	 */
712 
713 	if (skb->sk)
714 		skb_set_owner_w(skb2, skb->sk);
715 
716 	/*
717 	 *	Copy the packet header into the new buffer.
718 	 */
719 
720 	skb_copy_from_linear_data(skb, skb_network_header(skb2), state->hlen);
721 
722 	/*
723 	 *	Copy a block of the IP datagram.
724 	 */
725 	if (skb_copy_bits(skb, state->ptr, skb_transport_header(skb2), len))
726 		BUG();
727 	state->left -= len;
728 
729 	/*
730 	 *	Fill in the new header fields.
731 	 */
732 	iph = ip_hdr(skb2);
733 	iph->frag_off = htons((state->offset >> 3));
734 	if (state->DF)
735 		iph->frag_off |= htons(IP_DF);
736 
737 	/*
738 	 *	Added AC : If we are fragmenting a fragment that's not the
739 	 *		   last fragment then keep MF on each bit
740 	 */
741 	if (state->left > 0 || state->not_last_frag)
742 		iph->frag_off |= htons(IP_MF);
743 	state->ptr += len;
744 	state->offset += len;
745 
746 	iph->tot_len = htons(len + state->hlen);
747 
748 	ip_send_check(iph);
749 
750 	return skb2;
751 }
752 EXPORT_SYMBOL(ip_frag_next);
753 
754 /*
755  *	This IP datagram is too large to be sent in one piece.  Break it up into
756  *	smaller pieces (each of size equal to IP header plus
757  *	a block of the data of the original IP data part) that will yet fit in a
758  *	single device frame, and queue such a frame for sending.
759  */
760 
761 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
762 		   int (*output)(struct net *, struct sock *, struct sk_buff *))
763 {
764 	struct iphdr *iph;
765 	struct sk_buff *skb2;
766 	bool mono_delivery_time = skb->mono_delivery_time;
767 	struct rtable *rt = skb_rtable(skb);
768 	unsigned int mtu, hlen, ll_rs;
769 	struct ip_fraglist_iter iter;
770 	ktime_t tstamp = skb->tstamp;
771 	struct ip_frag_state state;
772 	int err = 0;
773 
774 	/* for offloaded checksums cleanup checksum before fragmentation */
775 	if (skb->ip_summed == CHECKSUM_PARTIAL &&
776 	    (err = skb_checksum_help(skb)))
777 		goto fail;
778 
779 	/*
780 	 *	Point into the IP datagram header.
781 	 */
782 
783 	iph = ip_hdr(skb);
784 
785 	mtu = ip_skb_dst_mtu(sk, skb);
786 	if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
787 		mtu = IPCB(skb)->frag_max_size;
788 
789 	/*
790 	 *	Setup starting values.
791 	 */
792 
793 	hlen = iph->ihl * 4;
794 	mtu = mtu - hlen;	/* Size of data space */
795 	IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
796 	ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
797 
798 	/* When frag_list is given, use it. First, check its validity:
799 	 * some transformers could create wrong frag_list or break existing
800 	 * one, it is not prohibited. In this case fall back to copying.
801 	 *
802 	 * LATER: this step can be merged to real generation of fragments,
803 	 * we can switch to copy when see the first bad fragment.
804 	 */
805 	if (skb_has_frag_list(skb)) {
806 		struct sk_buff *frag, *frag2;
807 		unsigned int first_len = skb_pagelen(skb);
808 
809 		if (first_len - hlen > mtu ||
810 		    ((first_len - hlen) & 7) ||
811 		    ip_is_fragment(iph) ||
812 		    skb_cloned(skb) ||
813 		    skb_headroom(skb) < ll_rs)
814 			goto slow_path;
815 
816 		skb_walk_frags(skb, frag) {
817 			/* Correct geometry. */
818 			if (frag->len > mtu ||
819 			    ((frag->len & 7) && frag->next) ||
820 			    skb_headroom(frag) < hlen + ll_rs)
821 				goto slow_path_clean;
822 
823 			/* Partially cloned skb? */
824 			if (skb_shared(frag))
825 				goto slow_path_clean;
826 
827 			BUG_ON(frag->sk);
828 			if (skb->sk) {
829 				frag->sk = skb->sk;
830 				frag->destructor = sock_wfree;
831 			}
832 			skb->truesize -= frag->truesize;
833 		}
834 
835 		/* Everything is OK. Generate! */
836 		ip_fraglist_init(skb, iph, hlen, &iter);
837 
838 		for (;;) {
839 			/* Prepare header of the next frame,
840 			 * before previous one went down. */
841 			if (iter.frag) {
842 				bool first_frag = (iter.offset == 0);
843 
844 				IPCB(iter.frag)->flags = IPCB(skb)->flags;
845 				ip_fraglist_prepare(skb, &iter);
846 				if (first_frag && IPCB(skb)->opt.optlen) {
847 					/* ipcb->opt is not populated for frags
848 					 * coming from __ip_make_skb(),
849 					 * ip_options_fragment() needs optlen
850 					 */
851 					IPCB(iter.frag)->opt.optlen =
852 						IPCB(skb)->opt.optlen;
853 					ip_options_fragment(iter.frag);
854 					ip_send_check(iter.iph);
855 				}
856 			}
857 
858 			skb_set_delivery_time(skb, tstamp, mono_delivery_time);
859 			err = output(net, sk, skb);
860 
861 			if (!err)
862 				IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
863 			if (err || !iter.frag)
864 				break;
865 
866 			skb = ip_fraglist_next(&iter);
867 		}
868 
869 		if (err == 0) {
870 			IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
871 			return 0;
872 		}
873 
874 		kfree_skb_list(iter.frag);
875 
876 		IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
877 		return err;
878 
879 slow_path_clean:
880 		skb_walk_frags(skb, frag2) {
881 			if (frag2 == frag)
882 				break;
883 			frag2->sk = NULL;
884 			frag2->destructor = NULL;
885 			skb->truesize += frag2->truesize;
886 		}
887 	}
888 
889 slow_path:
890 	/*
891 	 *	Fragment the datagram.
892 	 */
893 
894 	ip_frag_init(skb, hlen, ll_rs, mtu, IPCB(skb)->flags & IPSKB_FRAG_PMTU,
895 		     &state);
896 
897 	/*
898 	 *	Keep copying data until we run out.
899 	 */
900 
901 	while (state.left > 0) {
902 		bool first_frag = (state.offset == 0);
903 
904 		skb2 = ip_frag_next(skb, &state);
905 		if (IS_ERR(skb2)) {
906 			err = PTR_ERR(skb2);
907 			goto fail;
908 		}
909 		ip_frag_ipcb(skb, skb2, first_frag);
910 
911 		/*
912 		 *	Put this fragment into the sending queue.
913 		 */
914 		skb_set_delivery_time(skb2, tstamp, mono_delivery_time);
915 		err = output(net, sk, skb2);
916 		if (err)
917 			goto fail;
918 
919 		IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
920 	}
921 	consume_skb(skb);
922 	IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
923 	return err;
924 
925 fail:
926 	kfree_skb(skb);
927 	IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
928 	return err;
929 }
930 EXPORT_SYMBOL(ip_do_fragment);
931 
932 int
933 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
934 {
935 	struct msghdr *msg = from;
936 
937 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
938 		if (!copy_from_iter_full(to, len, &msg->msg_iter))
939 			return -EFAULT;
940 	} else {
941 		__wsum csum = 0;
942 		if (!csum_and_copy_from_iter_full(to, len, &csum, &msg->msg_iter))
943 			return -EFAULT;
944 		skb->csum = csum_block_add(skb->csum, csum, odd);
945 	}
946 	return 0;
947 }
948 EXPORT_SYMBOL(ip_generic_getfrag);
949 
950 static int __ip_append_data(struct sock *sk,
951 			    struct flowi4 *fl4,
952 			    struct sk_buff_head *queue,
953 			    struct inet_cork *cork,
954 			    struct page_frag *pfrag,
955 			    int getfrag(void *from, char *to, int offset,
956 					int len, int odd, struct sk_buff *skb),
957 			    void *from, int length, int transhdrlen,
958 			    unsigned int flags)
959 {
960 	struct inet_sock *inet = inet_sk(sk);
961 	struct ubuf_info *uarg = NULL;
962 	struct sk_buff *skb;
963 	struct ip_options *opt = cork->opt;
964 	int hh_len;
965 	int exthdrlen;
966 	int mtu;
967 	int copy;
968 	int err;
969 	int offset = 0;
970 	bool zc = false;
971 	unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
972 	int csummode = CHECKSUM_NONE;
973 	struct rtable *rt = (struct rtable *)cork->dst;
974 	unsigned int wmem_alloc_delta = 0;
975 	bool paged, extra_uref = false;
976 	u32 tskey = 0;
977 
978 	skb = skb_peek_tail(queue);
979 
980 	exthdrlen = !skb ? rt->dst.header_len : 0;
981 	mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
982 	paged = !!cork->gso_size;
983 
984 	if (cork->tx_flags & SKBTX_ANY_TSTAMP &&
985 	    sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
986 		tskey = atomic_inc_return(&sk->sk_tskey) - 1;
987 
988 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
989 
990 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
991 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
992 	maxnonfragsize = ip_sk_ignore_df(sk) ? IP_MAX_MTU : mtu;
993 
994 	if (cork->length + length > maxnonfragsize - fragheaderlen) {
995 		ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
996 			       mtu - (opt ? opt->optlen : 0));
997 		return -EMSGSIZE;
998 	}
999 
1000 	/*
1001 	 * transhdrlen > 0 means that this is the first fragment and we wish
1002 	 * it won't be fragmented in the future.
1003 	 */
1004 	if (transhdrlen &&
1005 	    length + fragheaderlen <= mtu &&
1006 	    rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
1007 	    (!(flags & MSG_MORE) || cork->gso_size) &&
1008 	    (!exthdrlen || (rt->dst.dev->features & NETIF_F_HW_ESP_TX_CSUM)))
1009 		csummode = CHECKSUM_PARTIAL;
1010 
1011 	if ((flags & MSG_ZEROCOPY) && length) {
1012 		struct msghdr *msg = from;
1013 
1014 		if (getfrag == ip_generic_getfrag && msg->msg_ubuf) {
1015 			if (skb_zcopy(skb) && msg->msg_ubuf != skb_zcopy(skb))
1016 				return -EINVAL;
1017 
1018 			/* Leave uarg NULL if can't zerocopy, callers should
1019 			 * be able to handle it.
1020 			 */
1021 			if ((rt->dst.dev->features & NETIF_F_SG) &&
1022 			    csummode == CHECKSUM_PARTIAL) {
1023 				paged = true;
1024 				zc = true;
1025 				uarg = msg->msg_ubuf;
1026 			}
1027 		} else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1028 			uarg = msg_zerocopy_realloc(sk, length, skb_zcopy(skb));
1029 			if (!uarg)
1030 				return -ENOBUFS;
1031 			extra_uref = !skb_zcopy(skb);	/* only ref on new uarg */
1032 			if (rt->dst.dev->features & NETIF_F_SG &&
1033 			    csummode == CHECKSUM_PARTIAL) {
1034 				paged = true;
1035 				zc = true;
1036 			} else {
1037 				uarg_to_msgzc(uarg)->zerocopy = 0;
1038 				skb_zcopy_set(skb, uarg, &extra_uref);
1039 			}
1040 		}
1041 	} else if ((flags & MSG_SPLICE_PAGES) && length) {
1042 		if (inet->hdrincl)
1043 			return -EPERM;
1044 		if (rt->dst.dev->features & NETIF_F_SG &&
1045 		    getfrag == ip_generic_getfrag)
1046 			/* We need an empty buffer to attach stuff to */
1047 			paged = true;
1048 		else
1049 			flags &= ~MSG_SPLICE_PAGES;
1050 	}
1051 
1052 	cork->length += length;
1053 
1054 	/* So, what's going on in the loop below?
1055 	 *
1056 	 * We use calculated fragment length to generate chained skb,
1057 	 * each of segments is IP fragment ready for sending to network after
1058 	 * adding appropriate IP header.
1059 	 */
1060 
1061 	if (!skb)
1062 		goto alloc_new_skb;
1063 
1064 	while (length > 0) {
1065 		/* Check if the remaining data fits into current packet. */
1066 		copy = mtu - skb->len;
1067 		if (copy < length)
1068 			copy = maxfraglen - skb->len;
1069 		if (copy <= 0) {
1070 			char *data;
1071 			unsigned int datalen;
1072 			unsigned int fraglen;
1073 			unsigned int fraggap;
1074 			unsigned int alloclen, alloc_extra;
1075 			unsigned int pagedlen;
1076 			struct sk_buff *skb_prev;
1077 alloc_new_skb:
1078 			skb_prev = skb;
1079 			if (skb_prev)
1080 				fraggap = skb_prev->len - maxfraglen;
1081 			else
1082 				fraggap = 0;
1083 
1084 			/*
1085 			 * If remaining data exceeds the mtu,
1086 			 * we know we need more fragment(s).
1087 			 */
1088 			datalen = length + fraggap;
1089 			if (datalen > mtu - fragheaderlen)
1090 				datalen = maxfraglen - fragheaderlen;
1091 			fraglen = datalen + fragheaderlen;
1092 			pagedlen = 0;
1093 
1094 			alloc_extra = hh_len + 15;
1095 			alloc_extra += exthdrlen;
1096 
1097 			/* The last fragment gets additional space at tail.
1098 			 * Note, with MSG_MORE we overallocate on fragments,
1099 			 * because we have no idea what fragment will be
1100 			 * the last.
1101 			 */
1102 			if (datalen == length + fraggap)
1103 				alloc_extra += rt->dst.trailer_len;
1104 
1105 			if ((flags & MSG_MORE) &&
1106 			    !(rt->dst.dev->features&NETIF_F_SG))
1107 				alloclen = mtu;
1108 			else if (!paged &&
1109 				 (fraglen + alloc_extra < SKB_MAX_ALLOC ||
1110 				  !(rt->dst.dev->features & NETIF_F_SG)))
1111 				alloclen = fraglen;
1112 			else {
1113 				alloclen = fragheaderlen + transhdrlen;
1114 				pagedlen = datalen - transhdrlen;
1115 			}
1116 
1117 			alloclen += alloc_extra;
1118 
1119 			if (transhdrlen) {
1120 				skb = sock_alloc_send_skb(sk, alloclen,
1121 						(flags & MSG_DONTWAIT), &err);
1122 			} else {
1123 				skb = NULL;
1124 				if (refcount_read(&sk->sk_wmem_alloc) + wmem_alloc_delta <=
1125 				    2 * sk->sk_sndbuf)
1126 					skb = alloc_skb(alloclen,
1127 							sk->sk_allocation);
1128 				if (unlikely(!skb))
1129 					err = -ENOBUFS;
1130 			}
1131 			if (!skb)
1132 				goto error;
1133 
1134 			/*
1135 			 *	Fill in the control structures
1136 			 */
1137 			skb->ip_summed = csummode;
1138 			skb->csum = 0;
1139 			skb_reserve(skb, hh_len);
1140 
1141 			/*
1142 			 *	Find where to start putting bytes.
1143 			 */
1144 			data = skb_put(skb, fraglen + exthdrlen - pagedlen);
1145 			skb_set_network_header(skb, exthdrlen);
1146 			skb->transport_header = (skb->network_header +
1147 						 fragheaderlen);
1148 			data += fragheaderlen + exthdrlen;
1149 
1150 			if (fraggap) {
1151 				skb->csum = skb_copy_and_csum_bits(
1152 					skb_prev, maxfraglen,
1153 					data + transhdrlen, fraggap);
1154 				skb_prev->csum = csum_sub(skb_prev->csum,
1155 							  skb->csum);
1156 				data += fraggap;
1157 				pskb_trim_unique(skb_prev, maxfraglen);
1158 			}
1159 
1160 			copy = datalen - transhdrlen - fraggap - pagedlen;
1161 			/* [!] NOTE: copy will be negative if pagedlen>0
1162 			 * because then the equation reduces to -fraggap.
1163 			 */
1164 			if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1165 				err = -EFAULT;
1166 				kfree_skb(skb);
1167 				goto error;
1168 			} else if (flags & MSG_SPLICE_PAGES) {
1169 				copy = 0;
1170 			}
1171 
1172 			offset += copy;
1173 			length -= copy + transhdrlen;
1174 			transhdrlen = 0;
1175 			exthdrlen = 0;
1176 			csummode = CHECKSUM_NONE;
1177 
1178 			/* only the initial fragment is time stamped */
1179 			skb_shinfo(skb)->tx_flags = cork->tx_flags;
1180 			cork->tx_flags = 0;
1181 			skb_shinfo(skb)->tskey = tskey;
1182 			tskey = 0;
1183 			skb_zcopy_set(skb, uarg, &extra_uref);
1184 
1185 			if ((flags & MSG_CONFIRM) && !skb_prev)
1186 				skb_set_dst_pending_confirm(skb, 1);
1187 
1188 			/*
1189 			 * Put the packet on the pending queue.
1190 			 */
1191 			if (!skb->destructor) {
1192 				skb->destructor = sock_wfree;
1193 				skb->sk = sk;
1194 				wmem_alloc_delta += skb->truesize;
1195 			}
1196 			__skb_queue_tail(queue, skb);
1197 			continue;
1198 		}
1199 
1200 		if (copy > length)
1201 			copy = length;
1202 
1203 		if (!(rt->dst.dev->features&NETIF_F_SG) &&
1204 		    skb_tailroom(skb) >= copy) {
1205 			unsigned int off;
1206 
1207 			off = skb->len;
1208 			if (getfrag(from, skb_put(skb, copy),
1209 					offset, copy, off, skb) < 0) {
1210 				__skb_trim(skb, off);
1211 				err = -EFAULT;
1212 				goto error;
1213 			}
1214 		} else if (flags & MSG_SPLICE_PAGES) {
1215 			struct msghdr *msg = from;
1216 
1217 			err = -EIO;
1218 			if (WARN_ON_ONCE(copy > msg->msg_iter.count))
1219 				goto error;
1220 
1221 			err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
1222 						   sk->sk_allocation);
1223 			if (err < 0)
1224 				goto error;
1225 			copy = err;
1226 			wmem_alloc_delta += copy;
1227 		} else if (!zc) {
1228 			int i = skb_shinfo(skb)->nr_frags;
1229 
1230 			err = -ENOMEM;
1231 			if (!sk_page_frag_refill(sk, pfrag))
1232 				goto error;
1233 
1234 			skb_zcopy_downgrade_managed(skb);
1235 			if (!skb_can_coalesce(skb, i, pfrag->page,
1236 					      pfrag->offset)) {
1237 				err = -EMSGSIZE;
1238 				if (i == MAX_SKB_FRAGS)
1239 					goto error;
1240 
1241 				__skb_fill_page_desc(skb, i, pfrag->page,
1242 						     pfrag->offset, 0);
1243 				skb_shinfo(skb)->nr_frags = ++i;
1244 				get_page(pfrag->page);
1245 			}
1246 			copy = min_t(int, copy, pfrag->size - pfrag->offset);
1247 			if (getfrag(from,
1248 				    page_address(pfrag->page) + pfrag->offset,
1249 				    offset, copy, skb->len, skb) < 0)
1250 				goto error_efault;
1251 
1252 			pfrag->offset += copy;
1253 			skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1254 			skb_len_add(skb, copy);
1255 			wmem_alloc_delta += copy;
1256 		} else {
1257 			err = skb_zerocopy_iter_dgram(skb, from, copy);
1258 			if (err < 0)
1259 				goto error;
1260 		}
1261 		offset += copy;
1262 		length -= copy;
1263 	}
1264 
1265 	if (wmem_alloc_delta)
1266 		refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1267 	return 0;
1268 
1269 error_efault:
1270 	err = -EFAULT;
1271 error:
1272 	net_zcopy_put_abort(uarg, extra_uref);
1273 	cork->length -= length;
1274 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1275 	refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1276 	return err;
1277 }
1278 
1279 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1280 			 struct ipcm_cookie *ipc, struct rtable **rtp)
1281 {
1282 	struct ip_options_rcu *opt;
1283 	struct rtable *rt;
1284 
1285 	rt = *rtp;
1286 	if (unlikely(!rt))
1287 		return -EFAULT;
1288 
1289 	/*
1290 	 * setup for corking.
1291 	 */
1292 	opt = ipc->opt;
1293 	if (opt) {
1294 		if (!cork->opt) {
1295 			cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1296 					    sk->sk_allocation);
1297 			if (unlikely(!cork->opt))
1298 				return -ENOBUFS;
1299 		}
1300 		memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1301 		cork->flags |= IPCORK_OPT;
1302 		cork->addr = ipc->addr;
1303 	}
1304 
1305 	cork->fragsize = ip_sk_use_pmtu(sk) ?
1306 			 dst_mtu(&rt->dst) : READ_ONCE(rt->dst.dev->mtu);
1307 
1308 	if (!inetdev_valid_mtu(cork->fragsize))
1309 		return -ENETUNREACH;
1310 
1311 	cork->gso_size = ipc->gso_size;
1312 
1313 	cork->dst = &rt->dst;
1314 	/* We stole this route, caller should not release it. */
1315 	*rtp = NULL;
1316 
1317 	cork->length = 0;
1318 	cork->ttl = ipc->ttl;
1319 	cork->tos = ipc->tos;
1320 	cork->mark = ipc->sockc.mark;
1321 	cork->priority = ipc->priority;
1322 	cork->transmit_time = ipc->sockc.transmit_time;
1323 	cork->tx_flags = 0;
1324 	sock_tx_timestamp(sk, ipc->sockc.tsflags, &cork->tx_flags);
1325 
1326 	return 0;
1327 }
1328 
1329 /*
1330  *	ip_append_data() can make one large IP datagram from many pieces of
1331  *	data.  Each piece will be held on the socket until
1332  *	ip_push_pending_frames() is called. Each piece can be a page or
1333  *	non-page data.
1334  *
1335  *	Not only UDP, other transport protocols - e.g. raw sockets - can use
1336  *	this interface potentially.
1337  *
1338  *	LATER: length must be adjusted by pad at tail, when it is required.
1339  */
1340 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1341 		   int getfrag(void *from, char *to, int offset, int len,
1342 			       int odd, struct sk_buff *skb),
1343 		   void *from, int length, int transhdrlen,
1344 		   struct ipcm_cookie *ipc, struct rtable **rtp,
1345 		   unsigned int flags)
1346 {
1347 	struct inet_sock *inet = inet_sk(sk);
1348 	int err;
1349 
1350 	if (flags&MSG_PROBE)
1351 		return 0;
1352 
1353 	if (skb_queue_empty(&sk->sk_write_queue)) {
1354 		err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1355 		if (err)
1356 			return err;
1357 	} else {
1358 		transhdrlen = 0;
1359 	}
1360 
1361 	return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1362 				sk_page_frag(sk), getfrag,
1363 				from, length, transhdrlen, flags);
1364 }
1365 
1366 static void ip_cork_release(struct inet_cork *cork)
1367 {
1368 	cork->flags &= ~IPCORK_OPT;
1369 	kfree(cork->opt);
1370 	cork->opt = NULL;
1371 	dst_release(cork->dst);
1372 	cork->dst = NULL;
1373 }
1374 
1375 /*
1376  *	Combined all pending IP fragments on the socket as one IP datagram
1377  *	and push them out.
1378  */
1379 struct sk_buff *__ip_make_skb(struct sock *sk,
1380 			      struct flowi4 *fl4,
1381 			      struct sk_buff_head *queue,
1382 			      struct inet_cork *cork)
1383 {
1384 	struct sk_buff *skb, *tmp_skb;
1385 	struct sk_buff **tail_skb;
1386 	struct inet_sock *inet = inet_sk(sk);
1387 	struct net *net = sock_net(sk);
1388 	struct ip_options *opt = NULL;
1389 	struct rtable *rt = (struct rtable *)cork->dst;
1390 	struct iphdr *iph;
1391 	__be16 df = 0;
1392 	__u8 ttl;
1393 
1394 	skb = __skb_dequeue(queue);
1395 	if (!skb)
1396 		goto out;
1397 	tail_skb = &(skb_shinfo(skb)->frag_list);
1398 
1399 	/* move skb->data to ip header from ext header */
1400 	if (skb->data < skb_network_header(skb))
1401 		__skb_pull(skb, skb_network_offset(skb));
1402 	while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1403 		__skb_pull(tmp_skb, skb_network_header_len(skb));
1404 		*tail_skb = tmp_skb;
1405 		tail_skb = &(tmp_skb->next);
1406 		skb->len += tmp_skb->len;
1407 		skb->data_len += tmp_skb->len;
1408 		skb->truesize += tmp_skb->truesize;
1409 		tmp_skb->destructor = NULL;
1410 		tmp_skb->sk = NULL;
1411 	}
1412 
1413 	/* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1414 	 * to fragment the frame generated here. No matter, what transforms
1415 	 * how transforms change size of the packet, it will come out.
1416 	 */
1417 	skb->ignore_df = ip_sk_ignore_df(sk);
1418 
1419 	/* DF bit is set when we want to see DF on outgoing frames.
1420 	 * If ignore_df is set too, we still allow to fragment this frame
1421 	 * locally. */
1422 	if (inet->pmtudisc == IP_PMTUDISC_DO ||
1423 	    inet->pmtudisc == IP_PMTUDISC_PROBE ||
1424 	    (skb->len <= dst_mtu(&rt->dst) &&
1425 	     ip_dont_fragment(sk, &rt->dst)))
1426 		df = htons(IP_DF);
1427 
1428 	if (cork->flags & IPCORK_OPT)
1429 		opt = cork->opt;
1430 
1431 	if (cork->ttl != 0)
1432 		ttl = cork->ttl;
1433 	else if (rt->rt_type == RTN_MULTICAST)
1434 		ttl = inet->mc_ttl;
1435 	else
1436 		ttl = ip_select_ttl(inet, &rt->dst);
1437 
1438 	iph = ip_hdr(skb);
1439 	iph->version = 4;
1440 	iph->ihl = 5;
1441 	iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1442 	iph->frag_off = df;
1443 	iph->ttl = ttl;
1444 	iph->protocol = sk->sk_protocol;
1445 	ip_copy_addrs(iph, fl4);
1446 	ip_select_ident(net, skb, sk);
1447 
1448 	if (opt) {
1449 		iph->ihl += opt->optlen >> 2;
1450 		ip_options_build(skb, opt, cork->addr, rt);
1451 	}
1452 
1453 	skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1454 	skb->mark = cork->mark;
1455 	skb->tstamp = cork->transmit_time;
1456 	/*
1457 	 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1458 	 * on dst refcount
1459 	 */
1460 	cork->dst = NULL;
1461 	skb_dst_set(skb, &rt->dst);
1462 
1463 	if (iph->protocol == IPPROTO_ICMP) {
1464 		u8 icmp_type;
1465 
1466 		/* For such sockets, transhdrlen is zero when do ip_append_data(),
1467 		 * so icmphdr does not in skb linear region and can not get icmp_type
1468 		 * by icmp_hdr(skb)->type.
1469 		 */
1470 		if (sk->sk_type == SOCK_RAW && !inet_sk(sk)->hdrincl)
1471 			icmp_type = fl4->fl4_icmp_type;
1472 		else
1473 			icmp_type = icmp_hdr(skb)->type;
1474 		icmp_out_count(net, icmp_type);
1475 	}
1476 
1477 	ip_cork_release(cork);
1478 out:
1479 	return skb;
1480 }
1481 
1482 int ip_send_skb(struct net *net, struct sk_buff *skb)
1483 {
1484 	int err;
1485 
1486 	err = ip_local_out(net, skb->sk, skb);
1487 	if (err) {
1488 		if (err > 0)
1489 			err = net_xmit_errno(err);
1490 		if (err)
1491 			IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1492 	}
1493 
1494 	return err;
1495 }
1496 
1497 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1498 {
1499 	struct sk_buff *skb;
1500 
1501 	skb = ip_finish_skb(sk, fl4);
1502 	if (!skb)
1503 		return 0;
1504 
1505 	/* Netfilter gets whole the not fragmented skb. */
1506 	return ip_send_skb(sock_net(sk), skb);
1507 }
1508 
1509 /*
1510  *	Throw away all pending data on the socket.
1511  */
1512 static void __ip_flush_pending_frames(struct sock *sk,
1513 				      struct sk_buff_head *queue,
1514 				      struct inet_cork *cork)
1515 {
1516 	struct sk_buff *skb;
1517 
1518 	while ((skb = __skb_dequeue_tail(queue)) != NULL)
1519 		kfree_skb(skb);
1520 
1521 	ip_cork_release(cork);
1522 }
1523 
1524 void ip_flush_pending_frames(struct sock *sk)
1525 {
1526 	__ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1527 }
1528 
1529 struct sk_buff *ip_make_skb(struct sock *sk,
1530 			    struct flowi4 *fl4,
1531 			    int getfrag(void *from, char *to, int offset,
1532 					int len, int odd, struct sk_buff *skb),
1533 			    void *from, int length, int transhdrlen,
1534 			    struct ipcm_cookie *ipc, struct rtable **rtp,
1535 			    struct inet_cork *cork, unsigned int flags)
1536 {
1537 	struct sk_buff_head queue;
1538 	int err;
1539 
1540 	if (flags & MSG_PROBE)
1541 		return NULL;
1542 
1543 	__skb_queue_head_init(&queue);
1544 
1545 	cork->flags = 0;
1546 	cork->addr = 0;
1547 	cork->opt = NULL;
1548 	err = ip_setup_cork(sk, cork, ipc, rtp);
1549 	if (err)
1550 		return ERR_PTR(err);
1551 
1552 	err = __ip_append_data(sk, fl4, &queue, cork,
1553 			       &current->task_frag, getfrag,
1554 			       from, length, transhdrlen, flags);
1555 	if (err) {
1556 		__ip_flush_pending_frames(sk, &queue, cork);
1557 		return ERR_PTR(err);
1558 	}
1559 
1560 	return __ip_make_skb(sk, fl4, &queue, cork);
1561 }
1562 
1563 /*
1564  *	Fetch data from kernel space and fill in checksum if needed.
1565  */
1566 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1567 			      int len, int odd, struct sk_buff *skb)
1568 {
1569 	__wsum csum;
1570 
1571 	csum = csum_partial_copy_nocheck(dptr+offset, to, len);
1572 	skb->csum = csum_block_add(skb->csum, csum, odd);
1573 	return 0;
1574 }
1575 
1576 /*
1577  *	Generic function to send a packet as reply to another packet.
1578  *	Used to send some TCP resets/acks so far.
1579  */
1580 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1581 			   const struct ip_options *sopt,
1582 			   __be32 daddr, __be32 saddr,
1583 			   const struct ip_reply_arg *arg,
1584 			   unsigned int len, u64 transmit_time, u32 txhash)
1585 {
1586 	struct ip_options_data replyopts;
1587 	struct ipcm_cookie ipc;
1588 	struct flowi4 fl4;
1589 	struct rtable *rt = skb_rtable(skb);
1590 	struct net *net = sock_net(sk);
1591 	struct sk_buff *nskb;
1592 	int err;
1593 	int oif;
1594 
1595 	if (__ip_options_echo(net, &replyopts.opt.opt, skb, sopt))
1596 		return;
1597 
1598 	ipcm_init(&ipc);
1599 	ipc.addr = daddr;
1600 	ipc.sockc.transmit_time = transmit_time;
1601 
1602 	if (replyopts.opt.opt.optlen) {
1603 		ipc.opt = &replyopts.opt;
1604 
1605 		if (replyopts.opt.opt.srr)
1606 			daddr = replyopts.opt.opt.faddr;
1607 	}
1608 
1609 	oif = arg->bound_dev_if;
1610 	if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1611 		oif = skb->skb_iif;
1612 
1613 	flowi4_init_output(&fl4, oif,
1614 			   IP4_REPLY_MARK(net, skb->mark) ?: sk->sk_mark,
1615 			   RT_TOS(arg->tos),
1616 			   RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1617 			   ip_reply_arg_flowi_flags(arg),
1618 			   daddr, saddr,
1619 			   tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1620 			   arg->uid);
1621 	security_skb_classify_flow(skb, flowi4_to_flowi_common(&fl4));
1622 	rt = ip_route_output_flow(net, &fl4, sk);
1623 	if (IS_ERR(rt))
1624 		return;
1625 
1626 	inet_sk(sk)->tos = arg->tos & ~INET_ECN_MASK;
1627 
1628 	sk->sk_protocol = ip_hdr(skb)->protocol;
1629 	sk->sk_bound_dev_if = arg->bound_dev_if;
1630 	sk->sk_sndbuf = READ_ONCE(sysctl_wmem_default);
1631 	ipc.sockc.mark = fl4.flowi4_mark;
1632 	err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1633 			     len, 0, &ipc, &rt, MSG_DONTWAIT);
1634 	if (unlikely(err)) {
1635 		ip_flush_pending_frames(sk);
1636 		goto out;
1637 	}
1638 
1639 	nskb = skb_peek(&sk->sk_write_queue);
1640 	if (nskb) {
1641 		if (arg->csumoffset >= 0)
1642 			*((__sum16 *)skb_transport_header(nskb) +
1643 			  arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1644 								arg->csum));
1645 		nskb->ip_summed = CHECKSUM_NONE;
1646 		nskb->mono_delivery_time = !!transmit_time;
1647 		if (txhash)
1648 			skb_set_hash(nskb, txhash, PKT_HASH_TYPE_L4);
1649 		ip_push_pending_frames(sk, &fl4);
1650 	}
1651 out:
1652 	ip_rt_put(rt);
1653 }
1654 
1655 void __init ip_init(void)
1656 {
1657 	ip_rt_init();
1658 	inet_initpeers();
1659 
1660 #if defined(CONFIG_IP_MULTICAST)
1661 	igmp_mc_init();
1662 #endif
1663 }
1664