xref: /linux/net/ipv4/ip_output.c (revision 1b69c6d0ae90b7f1a4f61d5c8209d5cb7a55f849)
1 /*
2  * INET		An implementation of the TCP/IP protocol suite for the LINUX
3  *		operating system.  INET is implemented using the  BSD Socket
4  *		interface as the means of communication with the user level.
5  *
6  *		The Internet Protocol (IP) output module.
7  *
8  * Authors:	Ross Biro
9  *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10  *		Donald Becker, <becker@super.org>
11  *		Alan Cox, <Alan.Cox@linux.org>
12  *		Richard Underwood
13  *		Stefan Becker, <stefanb@yello.ping.de>
14  *		Jorge Cwik, <jorge@laser.satlink.net>
15  *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
16  *		Hirokazu Takahashi, <taka@valinux.co.jp>
17  *
18  *	See ip_input.c for original log
19  *
20  *	Fixes:
21  *		Alan Cox	:	Missing nonblock feature in ip_build_xmit.
22  *		Mike Kilburn	:	htons() missing in ip_build_xmit.
23  *		Bradford Johnson:	Fix faulty handling of some frames when
24  *					no route is found.
25  *		Alexander Demenshin:	Missing sk/skb free in ip_queue_xmit
26  *					(in case if packet not accepted by
27  *					output firewall rules)
28  *		Mike McLagan	:	Routing by source
29  *		Alexey Kuznetsov:	use new route cache
30  *		Andi Kleen:		Fix broken PMTU recovery and remove
31  *					some redundant tests.
32  *	Vitaly E. Lavrov	:	Transparent proxy revived after year coma.
33  *		Andi Kleen	: 	Replace ip_reply with ip_send_reply.
34  *		Andi Kleen	:	Split fast and slow ip_build_xmit path
35  *					for decreased register pressure on x86
36  *					and more readibility.
37  *		Marc Boucher	:	When call_out_firewall returns FW_QUEUE,
38  *					silently drop skb instead of failing with -EPERM.
39  *		Detlev Wengorz	:	Copy protocol for fragments.
40  *		Hirokazu Takahashi:	HW checksumming for outgoing UDP
41  *					datagrams.
42  *		Hirokazu Takahashi:	sendfile() on UDP works now.
43  */
44 
45 #include <asm/uaccess.h>
46 #include <linux/module.h>
47 #include <linux/types.h>
48 #include <linux/kernel.h>
49 #include <linux/mm.h>
50 #include <linux/string.h>
51 #include <linux/errno.h>
52 #include <linux/highmem.h>
53 #include <linux/slab.h>
54 
55 #include <linux/socket.h>
56 #include <linux/sockios.h>
57 #include <linux/in.h>
58 #include <linux/inet.h>
59 #include <linux/netdevice.h>
60 #include <linux/etherdevice.h>
61 #include <linux/proc_fs.h>
62 #include <linux/stat.h>
63 #include <linux/init.h>
64 
65 #include <net/snmp.h>
66 #include <net/ip.h>
67 #include <net/protocol.h>
68 #include <net/route.h>
69 #include <net/xfrm.h>
70 #include <linux/skbuff.h>
71 #include <net/sock.h>
72 #include <net/arp.h>
73 #include <net/icmp.h>
74 #include <net/checksum.h>
75 #include <net/inetpeer.h>
76 #include <linux/igmp.h>
77 #include <linux/netfilter_ipv4.h>
78 #include <linux/netfilter_bridge.h>
79 #include <linux/mroute.h>
80 #include <linux/netlink.h>
81 #include <linux/tcp.h>
82 
83 int sysctl_ip_default_ttl __read_mostly = IPDEFTTL;
84 EXPORT_SYMBOL(sysctl_ip_default_ttl);
85 
86 static int ip_fragment(struct sock *sk, struct sk_buff *skb,
87 		       unsigned int mtu,
88 		       int (*output)(struct sock *, struct sk_buff *));
89 
90 /* Generate a checksum for an outgoing IP datagram. */
91 void ip_send_check(struct iphdr *iph)
92 {
93 	iph->check = 0;
94 	iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
95 }
96 EXPORT_SYMBOL(ip_send_check);
97 
98 static int __ip_local_out_sk(struct sock *sk, struct sk_buff *skb)
99 {
100 	struct net *net = dev_net(skb_dst(skb)->dev);
101 	struct iphdr *iph = ip_hdr(skb);
102 
103 	iph->tot_len = htons(skb->len);
104 	ip_send_check(iph);
105 	return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
106 		       net, sk, skb, NULL, skb_dst(skb)->dev,
107 		       dst_output_okfn);
108 }
109 
110 int __ip_local_out(struct sk_buff *skb)
111 {
112 	return __ip_local_out_sk(skb->sk, skb);
113 }
114 
115 int ip_local_out_sk(struct sock *sk, struct sk_buff *skb)
116 {
117 	int err;
118 
119 	err = __ip_local_out(skb);
120 	if (likely(err == 1))
121 		err = dst_output(sk, skb);
122 
123 	return err;
124 }
125 EXPORT_SYMBOL_GPL(ip_local_out_sk);
126 
127 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
128 {
129 	int ttl = inet->uc_ttl;
130 
131 	if (ttl < 0)
132 		ttl = ip4_dst_hoplimit(dst);
133 	return ttl;
134 }
135 
136 /*
137  *		Add an ip header to a skbuff and send it out.
138  *
139  */
140 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
141 			  __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
142 {
143 	struct inet_sock *inet = inet_sk(sk);
144 	struct rtable *rt = skb_rtable(skb);
145 	struct iphdr *iph;
146 
147 	/* Build the IP header. */
148 	skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
149 	skb_reset_network_header(skb);
150 	iph = ip_hdr(skb);
151 	iph->version  = 4;
152 	iph->ihl      = 5;
153 	iph->tos      = inet->tos;
154 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
155 	iph->daddr    = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
156 	iph->saddr    = saddr;
157 	iph->protocol = sk->sk_protocol;
158 	if (ip_dont_fragment(sk, &rt->dst)) {
159 		iph->frag_off = htons(IP_DF);
160 		iph->id = 0;
161 	} else {
162 		iph->frag_off = 0;
163 		__ip_select_ident(sock_net(sk), iph, 1);
164 	}
165 
166 	if (opt && opt->opt.optlen) {
167 		iph->ihl += opt->opt.optlen>>2;
168 		ip_options_build(skb, &opt->opt, daddr, rt, 0);
169 	}
170 
171 	skb->priority = sk->sk_priority;
172 	skb->mark = sk->sk_mark;
173 
174 	/* Send it out. */
175 	return ip_local_out(skb);
176 }
177 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
178 
179 static int ip_finish_output2(struct sock *sk, struct sk_buff *skb)
180 {
181 	struct dst_entry *dst = skb_dst(skb);
182 	struct rtable *rt = (struct rtable *)dst;
183 	struct net_device *dev = dst->dev;
184 	struct net *net = dev_net(dev);
185 	unsigned int hh_len = LL_RESERVED_SPACE(dev);
186 	struct neighbour *neigh;
187 	u32 nexthop;
188 
189 	if (rt->rt_type == RTN_MULTICAST) {
190 		IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
191 	} else if (rt->rt_type == RTN_BROADCAST)
192 		IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
193 
194 	/* Be paranoid, rather than too clever. */
195 	if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
196 		struct sk_buff *skb2;
197 
198 		skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
199 		if (!skb2) {
200 			kfree_skb(skb);
201 			return -ENOMEM;
202 		}
203 		if (skb->sk)
204 			skb_set_owner_w(skb2, skb->sk);
205 		consume_skb(skb);
206 		skb = skb2;
207 	}
208 
209 	rcu_read_lock_bh();
210 	nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
211 	neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
212 	if (unlikely(!neigh))
213 		neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
214 	if (!IS_ERR(neigh)) {
215 		int res = dst_neigh_output(dst, neigh, skb);
216 
217 		rcu_read_unlock_bh();
218 		return res;
219 	}
220 	rcu_read_unlock_bh();
221 
222 	net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
223 			    __func__);
224 	kfree_skb(skb);
225 	return -EINVAL;
226 }
227 
228 static int ip_finish_output_gso(struct sock *sk, struct sk_buff *skb,
229 				unsigned int mtu)
230 {
231 	netdev_features_t features;
232 	struct sk_buff *segs;
233 	int ret = 0;
234 
235 	/* common case: locally created skb or seglen is <= mtu */
236 	if (((IPCB(skb)->flags & IPSKB_FORWARDED) == 0) ||
237 	      skb_gso_network_seglen(skb) <= mtu)
238 		return ip_finish_output2(sk, skb);
239 
240 	/* Slowpath -  GSO segment length is exceeding the dst MTU.
241 	 *
242 	 * This can happen in two cases:
243 	 * 1) TCP GRO packet, DF bit not set
244 	 * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
245 	 * from host network stack.
246 	 */
247 	features = netif_skb_features(skb);
248 	segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
249 	if (IS_ERR_OR_NULL(segs)) {
250 		kfree_skb(skb);
251 		return -ENOMEM;
252 	}
253 
254 	consume_skb(skb);
255 
256 	do {
257 		struct sk_buff *nskb = segs->next;
258 		int err;
259 
260 		segs->next = NULL;
261 		err = ip_fragment(sk, segs, mtu, ip_finish_output2);
262 
263 		if (err && ret == 0)
264 			ret = err;
265 		segs = nskb;
266 	} while (segs);
267 
268 	return ret;
269 }
270 
271 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
272 {
273 	unsigned int mtu;
274 
275 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
276 	/* Policy lookup after SNAT yielded a new policy */
277 	if (skb_dst(skb)->xfrm) {
278 		IPCB(skb)->flags |= IPSKB_REROUTED;
279 		return dst_output(sk, skb);
280 	}
281 #endif
282 	mtu = ip_skb_dst_mtu(skb);
283 	if (skb_is_gso(skb))
284 		return ip_finish_output_gso(sk, skb, mtu);
285 
286 	if (skb->len > mtu || (IPCB(skb)->flags & IPSKB_FRAG_PMTU))
287 		return ip_fragment(sk, skb, mtu, ip_finish_output2);
288 
289 	return ip_finish_output2(sk, skb);
290 }
291 
292 int ip_mc_output(struct sock *sk, struct sk_buff *skb)
293 {
294 	struct rtable *rt = skb_rtable(skb);
295 	struct net_device *dev = rt->dst.dev;
296 	struct net *net = dev_net(dev);
297 
298 	/*
299 	 *	If the indicated interface is up and running, send the packet.
300 	 */
301 	IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
302 
303 	skb->dev = dev;
304 	skb->protocol = htons(ETH_P_IP);
305 
306 	/*
307 	 *	Multicasts are looped back for other local users
308 	 */
309 
310 	if (rt->rt_flags&RTCF_MULTICAST) {
311 		if (sk_mc_loop(sk)
312 #ifdef CONFIG_IP_MROUTE
313 		/* Small optimization: do not loopback not local frames,
314 		   which returned after forwarding; they will be  dropped
315 		   by ip_mr_input in any case.
316 		   Note, that local frames are looped back to be delivered
317 		   to local recipients.
318 
319 		   This check is duplicated in ip_mr_input at the moment.
320 		 */
321 		    &&
322 		    ((rt->rt_flags & RTCF_LOCAL) ||
323 		     !(IPCB(skb)->flags & IPSKB_FORWARDED))
324 #endif
325 		   ) {
326 			struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
327 			if (newskb)
328 				NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
329 					net, sk, newskb, NULL, newskb->dev,
330 					dev_loopback_xmit);
331 		}
332 
333 		/* Multicasts with ttl 0 must not go beyond the host */
334 
335 		if (ip_hdr(skb)->ttl == 0) {
336 			kfree_skb(skb);
337 			return 0;
338 		}
339 	}
340 
341 	if (rt->rt_flags&RTCF_BROADCAST) {
342 		struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
343 		if (newskb)
344 			NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
345 				net, sk, newskb, NULL, newskb->dev,
346 				dev_loopback_xmit);
347 	}
348 
349 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
350 			    net, sk, skb, NULL, skb->dev,
351 			    ip_finish_output,
352 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
353 }
354 
355 int ip_output(struct sock *sk, struct sk_buff *skb)
356 {
357 	struct net_device *dev = skb_dst(skb)->dev;
358 	struct net *net = dev_net(dev);
359 
360 	IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
361 
362 	skb->dev = dev;
363 	skb->protocol = htons(ETH_P_IP);
364 
365 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
366 			    net, sk, skb, NULL, dev,
367 			    ip_finish_output,
368 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
369 }
370 
371 /*
372  * copy saddr and daddr, possibly using 64bit load/stores
373  * Equivalent to :
374  *   iph->saddr = fl4->saddr;
375  *   iph->daddr = fl4->daddr;
376  */
377 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
378 {
379 	BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
380 		     offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
381 	memcpy(&iph->saddr, &fl4->saddr,
382 	       sizeof(fl4->saddr) + sizeof(fl4->daddr));
383 }
384 
385 /* Note: skb->sk can be different from sk, in case of tunnels */
386 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
387 {
388 	struct inet_sock *inet = inet_sk(sk);
389 	struct ip_options_rcu *inet_opt;
390 	struct flowi4 *fl4;
391 	struct rtable *rt;
392 	struct iphdr *iph;
393 	int res;
394 
395 	/* Skip all of this if the packet is already routed,
396 	 * f.e. by something like SCTP.
397 	 */
398 	rcu_read_lock();
399 	inet_opt = rcu_dereference(inet->inet_opt);
400 	fl4 = &fl->u.ip4;
401 	rt = skb_rtable(skb);
402 	if (rt)
403 		goto packet_routed;
404 
405 	/* Make sure we can route this packet. */
406 	rt = (struct rtable *)__sk_dst_check(sk, 0);
407 	if (!rt) {
408 		__be32 daddr;
409 
410 		/* Use correct destination address if we have options. */
411 		daddr = inet->inet_daddr;
412 		if (inet_opt && inet_opt->opt.srr)
413 			daddr = inet_opt->opt.faddr;
414 
415 		/* If this fails, retransmit mechanism of transport layer will
416 		 * keep trying until route appears or the connection times
417 		 * itself out.
418 		 */
419 		rt = ip_route_output_ports(sock_net(sk), fl4, sk,
420 					   daddr, inet->inet_saddr,
421 					   inet->inet_dport,
422 					   inet->inet_sport,
423 					   sk->sk_protocol,
424 					   RT_CONN_FLAGS(sk),
425 					   sk->sk_bound_dev_if);
426 		if (IS_ERR(rt))
427 			goto no_route;
428 		sk_setup_caps(sk, &rt->dst);
429 	}
430 	skb_dst_set_noref(skb, &rt->dst);
431 
432 packet_routed:
433 	if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
434 		goto no_route;
435 
436 	/* OK, we know where to send it, allocate and build IP header. */
437 	skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
438 	skb_reset_network_header(skb);
439 	iph = ip_hdr(skb);
440 	*((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
441 	if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
442 		iph->frag_off = htons(IP_DF);
443 	else
444 		iph->frag_off = 0;
445 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
446 	iph->protocol = sk->sk_protocol;
447 	ip_copy_addrs(iph, fl4);
448 
449 	/* Transport layer set skb->h.foo itself. */
450 
451 	if (inet_opt && inet_opt->opt.optlen) {
452 		iph->ihl += inet_opt->opt.optlen >> 2;
453 		ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
454 	}
455 
456 	ip_select_ident_segs(sock_net(sk), skb, sk,
457 			     skb_shinfo(skb)->gso_segs ?: 1);
458 
459 	/* TODO : should we use skb->sk here instead of sk ? */
460 	skb->priority = sk->sk_priority;
461 	skb->mark = sk->sk_mark;
462 
463 	res = ip_local_out(skb);
464 	rcu_read_unlock();
465 	return res;
466 
467 no_route:
468 	rcu_read_unlock();
469 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
470 	kfree_skb(skb);
471 	return -EHOSTUNREACH;
472 }
473 EXPORT_SYMBOL(ip_queue_xmit);
474 
475 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
476 {
477 	to->pkt_type = from->pkt_type;
478 	to->priority = from->priority;
479 	to->protocol = from->protocol;
480 	skb_dst_drop(to);
481 	skb_dst_copy(to, from);
482 	to->dev = from->dev;
483 	to->mark = from->mark;
484 
485 	/* Copy the flags to each fragment. */
486 	IPCB(to)->flags = IPCB(from)->flags;
487 
488 #ifdef CONFIG_NET_SCHED
489 	to->tc_index = from->tc_index;
490 #endif
491 	nf_copy(to, from);
492 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
493 	to->ipvs_property = from->ipvs_property;
494 #endif
495 	skb_copy_secmark(to, from);
496 }
497 
498 static int ip_fragment(struct sock *sk, struct sk_buff *skb,
499 		       unsigned int mtu,
500 		       int (*output)(struct sock *, struct sk_buff *))
501 {
502 	struct iphdr *iph = ip_hdr(skb);
503 
504 	if ((iph->frag_off & htons(IP_DF)) == 0)
505 		return ip_do_fragment(sk, skb, output);
506 
507 	if (unlikely(!skb->ignore_df ||
508 		     (IPCB(skb)->frag_max_size &&
509 		      IPCB(skb)->frag_max_size > mtu))) {
510 		struct net *net = dev_net(skb_rtable(skb)->dst.dev);
511 
512 		IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
513 		icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
514 			  htonl(mtu));
515 		kfree_skb(skb);
516 		return -EMSGSIZE;
517 	}
518 
519 	return ip_do_fragment(sk, skb, output);
520 }
521 
522 /*
523  *	This IP datagram is too large to be sent in one piece.  Break it up into
524  *	smaller pieces (each of size equal to IP header plus
525  *	a block of the data of the original IP data part) that will yet fit in a
526  *	single device frame, and queue such a frame for sending.
527  */
528 
529 int ip_do_fragment(struct sock *sk, struct sk_buff *skb,
530 		   int (*output)(struct sock *, struct sk_buff *))
531 {
532 	struct iphdr *iph;
533 	int ptr;
534 	struct net_device *dev;
535 	struct sk_buff *skb2;
536 	unsigned int mtu, hlen, left, len, ll_rs;
537 	int offset;
538 	__be16 not_last_frag;
539 	struct rtable *rt = skb_rtable(skb);
540 	struct net *net;
541 	int err = 0;
542 
543 	dev = rt->dst.dev;
544 	net = dev_net(dev);
545 
546 	/*
547 	 *	Point into the IP datagram header.
548 	 */
549 
550 	iph = ip_hdr(skb);
551 
552 	mtu = ip_skb_dst_mtu(skb);
553 	if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
554 		mtu = IPCB(skb)->frag_max_size;
555 
556 	/*
557 	 *	Setup starting values.
558 	 */
559 
560 	hlen = iph->ihl * 4;
561 	mtu = mtu - hlen;	/* Size of data space */
562 	IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
563 
564 	/* When frag_list is given, use it. First, check its validity:
565 	 * some transformers could create wrong frag_list or break existing
566 	 * one, it is not prohibited. In this case fall back to copying.
567 	 *
568 	 * LATER: this step can be merged to real generation of fragments,
569 	 * we can switch to copy when see the first bad fragment.
570 	 */
571 	if (skb_has_frag_list(skb)) {
572 		struct sk_buff *frag, *frag2;
573 		int first_len = skb_pagelen(skb);
574 
575 		if (first_len - hlen > mtu ||
576 		    ((first_len - hlen) & 7) ||
577 		    ip_is_fragment(iph) ||
578 		    skb_cloned(skb))
579 			goto slow_path;
580 
581 		skb_walk_frags(skb, frag) {
582 			/* Correct geometry. */
583 			if (frag->len > mtu ||
584 			    ((frag->len & 7) && frag->next) ||
585 			    skb_headroom(frag) < hlen)
586 				goto slow_path_clean;
587 
588 			/* Partially cloned skb? */
589 			if (skb_shared(frag))
590 				goto slow_path_clean;
591 
592 			BUG_ON(frag->sk);
593 			if (skb->sk) {
594 				frag->sk = skb->sk;
595 				frag->destructor = sock_wfree;
596 			}
597 			skb->truesize -= frag->truesize;
598 		}
599 
600 		/* Everything is OK. Generate! */
601 
602 		err = 0;
603 		offset = 0;
604 		frag = skb_shinfo(skb)->frag_list;
605 		skb_frag_list_init(skb);
606 		skb->data_len = first_len - skb_headlen(skb);
607 		skb->len = first_len;
608 		iph->tot_len = htons(first_len);
609 		iph->frag_off = htons(IP_MF);
610 		ip_send_check(iph);
611 
612 		for (;;) {
613 			/* Prepare header of the next frame,
614 			 * before previous one went down. */
615 			if (frag) {
616 				frag->ip_summed = CHECKSUM_NONE;
617 				skb_reset_transport_header(frag);
618 				__skb_push(frag, hlen);
619 				skb_reset_network_header(frag);
620 				memcpy(skb_network_header(frag), iph, hlen);
621 				iph = ip_hdr(frag);
622 				iph->tot_len = htons(frag->len);
623 				ip_copy_metadata(frag, skb);
624 				if (offset == 0)
625 					ip_options_fragment(frag);
626 				offset += skb->len - hlen;
627 				iph->frag_off = htons(offset>>3);
628 				if (frag->next)
629 					iph->frag_off |= htons(IP_MF);
630 				/* Ready, complete checksum */
631 				ip_send_check(iph);
632 			}
633 
634 			err = output(sk, skb);
635 
636 			if (!err)
637 				IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
638 			if (err || !frag)
639 				break;
640 
641 			skb = frag;
642 			frag = skb->next;
643 			skb->next = NULL;
644 		}
645 
646 		if (err == 0) {
647 			IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
648 			return 0;
649 		}
650 
651 		while (frag) {
652 			skb = frag->next;
653 			kfree_skb(frag);
654 			frag = skb;
655 		}
656 		IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
657 		return err;
658 
659 slow_path_clean:
660 		skb_walk_frags(skb, frag2) {
661 			if (frag2 == frag)
662 				break;
663 			frag2->sk = NULL;
664 			frag2->destructor = NULL;
665 			skb->truesize += frag2->truesize;
666 		}
667 	}
668 
669 slow_path:
670 	/* for offloaded checksums cleanup checksum before fragmentation */
671 	if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
672 		goto fail;
673 	iph = ip_hdr(skb);
674 
675 	left = skb->len - hlen;		/* Space per frame */
676 	ptr = hlen;		/* Where to start from */
677 
678 	ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
679 
680 	/*
681 	 *	Fragment the datagram.
682 	 */
683 
684 	offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
685 	not_last_frag = iph->frag_off & htons(IP_MF);
686 
687 	/*
688 	 *	Keep copying data until we run out.
689 	 */
690 
691 	while (left > 0) {
692 		len = left;
693 		/* IF: it doesn't fit, use 'mtu' - the data space left */
694 		if (len > mtu)
695 			len = mtu;
696 		/* IF: we are not sending up to and including the packet end
697 		   then align the next start on an eight byte boundary */
698 		if (len < left)	{
699 			len &= ~7;
700 		}
701 
702 		/* Allocate buffer */
703 		skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
704 		if (!skb2) {
705 			err = -ENOMEM;
706 			goto fail;
707 		}
708 
709 		/*
710 		 *	Set up data on packet
711 		 */
712 
713 		ip_copy_metadata(skb2, skb);
714 		skb_reserve(skb2, ll_rs);
715 		skb_put(skb2, len + hlen);
716 		skb_reset_network_header(skb2);
717 		skb2->transport_header = skb2->network_header + hlen;
718 
719 		/*
720 		 *	Charge the memory for the fragment to any owner
721 		 *	it might possess
722 		 */
723 
724 		if (skb->sk)
725 			skb_set_owner_w(skb2, skb->sk);
726 
727 		/*
728 		 *	Copy the packet header into the new buffer.
729 		 */
730 
731 		skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
732 
733 		/*
734 		 *	Copy a block of the IP datagram.
735 		 */
736 		if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
737 			BUG();
738 		left -= len;
739 
740 		/*
741 		 *	Fill in the new header fields.
742 		 */
743 		iph = ip_hdr(skb2);
744 		iph->frag_off = htons((offset >> 3));
745 
746 		if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
747 			iph->frag_off |= htons(IP_DF);
748 
749 		/* ANK: dirty, but effective trick. Upgrade options only if
750 		 * the segment to be fragmented was THE FIRST (otherwise,
751 		 * options are already fixed) and make it ONCE
752 		 * on the initial skb, so that all the following fragments
753 		 * will inherit fixed options.
754 		 */
755 		if (offset == 0)
756 			ip_options_fragment(skb);
757 
758 		/*
759 		 *	Added AC : If we are fragmenting a fragment that's not the
760 		 *		   last fragment then keep MF on each bit
761 		 */
762 		if (left > 0 || not_last_frag)
763 			iph->frag_off |= htons(IP_MF);
764 		ptr += len;
765 		offset += len;
766 
767 		/*
768 		 *	Put this fragment into the sending queue.
769 		 */
770 		iph->tot_len = htons(len + hlen);
771 
772 		ip_send_check(iph);
773 
774 		err = output(sk, skb2);
775 		if (err)
776 			goto fail;
777 
778 		IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
779 	}
780 	consume_skb(skb);
781 	IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
782 	return err;
783 
784 fail:
785 	kfree_skb(skb);
786 	IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
787 	return err;
788 }
789 EXPORT_SYMBOL(ip_do_fragment);
790 
791 int
792 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
793 {
794 	struct msghdr *msg = from;
795 
796 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
797 		if (copy_from_iter(to, len, &msg->msg_iter) != len)
798 			return -EFAULT;
799 	} else {
800 		__wsum csum = 0;
801 		if (csum_and_copy_from_iter(to, len, &csum, &msg->msg_iter) != len)
802 			return -EFAULT;
803 		skb->csum = csum_block_add(skb->csum, csum, odd);
804 	}
805 	return 0;
806 }
807 EXPORT_SYMBOL(ip_generic_getfrag);
808 
809 static inline __wsum
810 csum_page(struct page *page, int offset, int copy)
811 {
812 	char *kaddr;
813 	__wsum csum;
814 	kaddr = kmap(page);
815 	csum = csum_partial(kaddr + offset, copy, 0);
816 	kunmap(page);
817 	return csum;
818 }
819 
820 static inline int ip_ufo_append_data(struct sock *sk,
821 			struct sk_buff_head *queue,
822 			int getfrag(void *from, char *to, int offset, int len,
823 			       int odd, struct sk_buff *skb),
824 			void *from, int length, int hh_len, int fragheaderlen,
825 			int transhdrlen, int maxfraglen, unsigned int flags)
826 {
827 	struct sk_buff *skb;
828 	int err;
829 
830 	/* There is support for UDP fragmentation offload by network
831 	 * device, so create one single skb packet containing complete
832 	 * udp datagram
833 	 */
834 	skb = skb_peek_tail(queue);
835 	if (!skb) {
836 		skb = sock_alloc_send_skb(sk,
837 			hh_len + fragheaderlen + transhdrlen + 20,
838 			(flags & MSG_DONTWAIT), &err);
839 
840 		if (!skb)
841 			return err;
842 
843 		/* reserve space for Hardware header */
844 		skb_reserve(skb, hh_len);
845 
846 		/* create space for UDP/IP header */
847 		skb_put(skb, fragheaderlen + transhdrlen);
848 
849 		/* initialize network header pointer */
850 		skb_reset_network_header(skb);
851 
852 		/* initialize protocol header pointer */
853 		skb->transport_header = skb->network_header + fragheaderlen;
854 
855 		skb->csum = 0;
856 
857 		__skb_queue_tail(queue, skb);
858 	} else if (skb_is_gso(skb)) {
859 		goto append;
860 	}
861 
862 	skb->ip_summed = CHECKSUM_PARTIAL;
863 	/* specify the length of each IP datagram fragment */
864 	skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
865 	skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
866 
867 append:
868 	return skb_append_datato_frags(sk, skb, getfrag, from,
869 				       (length - transhdrlen));
870 }
871 
872 static int __ip_append_data(struct sock *sk,
873 			    struct flowi4 *fl4,
874 			    struct sk_buff_head *queue,
875 			    struct inet_cork *cork,
876 			    struct page_frag *pfrag,
877 			    int getfrag(void *from, char *to, int offset,
878 					int len, int odd, struct sk_buff *skb),
879 			    void *from, int length, int transhdrlen,
880 			    unsigned int flags)
881 {
882 	struct inet_sock *inet = inet_sk(sk);
883 	struct sk_buff *skb;
884 
885 	struct ip_options *opt = cork->opt;
886 	int hh_len;
887 	int exthdrlen;
888 	int mtu;
889 	int copy;
890 	int err;
891 	int offset = 0;
892 	unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
893 	int csummode = CHECKSUM_NONE;
894 	struct rtable *rt = (struct rtable *)cork->dst;
895 	u32 tskey = 0;
896 
897 	skb = skb_peek_tail(queue);
898 
899 	exthdrlen = !skb ? rt->dst.header_len : 0;
900 	mtu = cork->fragsize;
901 	if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
902 	    sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
903 		tskey = sk->sk_tskey++;
904 
905 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
906 
907 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
908 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
909 	maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
910 
911 	if (cork->length + length > maxnonfragsize - fragheaderlen) {
912 		ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
913 			       mtu - (opt ? opt->optlen : 0));
914 		return -EMSGSIZE;
915 	}
916 
917 	/*
918 	 * transhdrlen > 0 means that this is the first fragment and we wish
919 	 * it won't be fragmented in the future.
920 	 */
921 	if (transhdrlen &&
922 	    length + fragheaderlen <= mtu &&
923 	    rt->dst.dev->features & NETIF_F_V4_CSUM &&
924 	    !exthdrlen)
925 		csummode = CHECKSUM_PARTIAL;
926 
927 	cork->length += length;
928 	if (((length > mtu) || (skb && skb_is_gso(skb))) &&
929 	    (sk->sk_protocol == IPPROTO_UDP) &&
930 	    (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len &&
931 	    (sk->sk_type == SOCK_DGRAM)) {
932 		err = ip_ufo_append_data(sk, queue, getfrag, from, length,
933 					 hh_len, fragheaderlen, transhdrlen,
934 					 maxfraglen, flags);
935 		if (err)
936 			goto error;
937 		return 0;
938 	}
939 
940 	/* So, what's going on in the loop below?
941 	 *
942 	 * We use calculated fragment length to generate chained skb,
943 	 * each of segments is IP fragment ready for sending to network after
944 	 * adding appropriate IP header.
945 	 */
946 
947 	if (!skb)
948 		goto alloc_new_skb;
949 
950 	while (length > 0) {
951 		/* Check if the remaining data fits into current packet. */
952 		copy = mtu - skb->len;
953 		if (copy < length)
954 			copy = maxfraglen - skb->len;
955 		if (copy <= 0) {
956 			char *data;
957 			unsigned int datalen;
958 			unsigned int fraglen;
959 			unsigned int fraggap;
960 			unsigned int alloclen;
961 			struct sk_buff *skb_prev;
962 alloc_new_skb:
963 			skb_prev = skb;
964 			if (skb_prev)
965 				fraggap = skb_prev->len - maxfraglen;
966 			else
967 				fraggap = 0;
968 
969 			/*
970 			 * If remaining data exceeds the mtu,
971 			 * we know we need more fragment(s).
972 			 */
973 			datalen = length + fraggap;
974 			if (datalen > mtu - fragheaderlen)
975 				datalen = maxfraglen - fragheaderlen;
976 			fraglen = datalen + fragheaderlen;
977 
978 			if ((flags & MSG_MORE) &&
979 			    !(rt->dst.dev->features&NETIF_F_SG))
980 				alloclen = mtu;
981 			else
982 				alloclen = fraglen;
983 
984 			alloclen += exthdrlen;
985 
986 			/* The last fragment gets additional space at tail.
987 			 * Note, with MSG_MORE we overallocate on fragments,
988 			 * because we have no idea what fragment will be
989 			 * the last.
990 			 */
991 			if (datalen == length + fraggap)
992 				alloclen += rt->dst.trailer_len;
993 
994 			if (transhdrlen) {
995 				skb = sock_alloc_send_skb(sk,
996 						alloclen + hh_len + 15,
997 						(flags & MSG_DONTWAIT), &err);
998 			} else {
999 				skb = NULL;
1000 				if (atomic_read(&sk->sk_wmem_alloc) <=
1001 				    2 * sk->sk_sndbuf)
1002 					skb = sock_wmalloc(sk,
1003 							   alloclen + hh_len + 15, 1,
1004 							   sk->sk_allocation);
1005 				if (unlikely(!skb))
1006 					err = -ENOBUFS;
1007 			}
1008 			if (!skb)
1009 				goto error;
1010 
1011 			/*
1012 			 *	Fill in the control structures
1013 			 */
1014 			skb->ip_summed = csummode;
1015 			skb->csum = 0;
1016 			skb_reserve(skb, hh_len);
1017 
1018 			/* only the initial fragment is time stamped */
1019 			skb_shinfo(skb)->tx_flags = cork->tx_flags;
1020 			cork->tx_flags = 0;
1021 			skb_shinfo(skb)->tskey = tskey;
1022 			tskey = 0;
1023 
1024 			/*
1025 			 *	Find where to start putting bytes.
1026 			 */
1027 			data = skb_put(skb, fraglen + exthdrlen);
1028 			skb_set_network_header(skb, exthdrlen);
1029 			skb->transport_header = (skb->network_header +
1030 						 fragheaderlen);
1031 			data += fragheaderlen + exthdrlen;
1032 
1033 			if (fraggap) {
1034 				skb->csum = skb_copy_and_csum_bits(
1035 					skb_prev, maxfraglen,
1036 					data + transhdrlen, fraggap, 0);
1037 				skb_prev->csum = csum_sub(skb_prev->csum,
1038 							  skb->csum);
1039 				data += fraggap;
1040 				pskb_trim_unique(skb_prev, maxfraglen);
1041 			}
1042 
1043 			copy = datalen - transhdrlen - fraggap;
1044 			if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1045 				err = -EFAULT;
1046 				kfree_skb(skb);
1047 				goto error;
1048 			}
1049 
1050 			offset += copy;
1051 			length -= datalen - fraggap;
1052 			transhdrlen = 0;
1053 			exthdrlen = 0;
1054 			csummode = CHECKSUM_NONE;
1055 
1056 			/*
1057 			 * Put the packet on the pending queue.
1058 			 */
1059 			__skb_queue_tail(queue, skb);
1060 			continue;
1061 		}
1062 
1063 		if (copy > length)
1064 			copy = length;
1065 
1066 		if (!(rt->dst.dev->features&NETIF_F_SG)) {
1067 			unsigned int off;
1068 
1069 			off = skb->len;
1070 			if (getfrag(from, skb_put(skb, copy),
1071 					offset, copy, off, skb) < 0) {
1072 				__skb_trim(skb, off);
1073 				err = -EFAULT;
1074 				goto error;
1075 			}
1076 		} else {
1077 			int i = skb_shinfo(skb)->nr_frags;
1078 
1079 			err = -ENOMEM;
1080 			if (!sk_page_frag_refill(sk, pfrag))
1081 				goto error;
1082 
1083 			if (!skb_can_coalesce(skb, i, pfrag->page,
1084 					      pfrag->offset)) {
1085 				err = -EMSGSIZE;
1086 				if (i == MAX_SKB_FRAGS)
1087 					goto error;
1088 
1089 				__skb_fill_page_desc(skb, i, pfrag->page,
1090 						     pfrag->offset, 0);
1091 				skb_shinfo(skb)->nr_frags = ++i;
1092 				get_page(pfrag->page);
1093 			}
1094 			copy = min_t(int, copy, pfrag->size - pfrag->offset);
1095 			if (getfrag(from,
1096 				    page_address(pfrag->page) + pfrag->offset,
1097 				    offset, copy, skb->len, skb) < 0)
1098 				goto error_efault;
1099 
1100 			pfrag->offset += copy;
1101 			skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1102 			skb->len += copy;
1103 			skb->data_len += copy;
1104 			skb->truesize += copy;
1105 			atomic_add(copy, &sk->sk_wmem_alloc);
1106 		}
1107 		offset += copy;
1108 		length -= copy;
1109 	}
1110 
1111 	return 0;
1112 
1113 error_efault:
1114 	err = -EFAULT;
1115 error:
1116 	cork->length -= length;
1117 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1118 	return err;
1119 }
1120 
1121 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1122 			 struct ipcm_cookie *ipc, struct rtable **rtp)
1123 {
1124 	struct ip_options_rcu *opt;
1125 	struct rtable *rt;
1126 
1127 	/*
1128 	 * setup for corking.
1129 	 */
1130 	opt = ipc->opt;
1131 	if (opt) {
1132 		if (!cork->opt) {
1133 			cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1134 					    sk->sk_allocation);
1135 			if (unlikely(!cork->opt))
1136 				return -ENOBUFS;
1137 		}
1138 		memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1139 		cork->flags |= IPCORK_OPT;
1140 		cork->addr = ipc->addr;
1141 	}
1142 	rt = *rtp;
1143 	if (unlikely(!rt))
1144 		return -EFAULT;
1145 	/*
1146 	 * We steal reference to this route, caller should not release it
1147 	 */
1148 	*rtp = NULL;
1149 	cork->fragsize = ip_sk_use_pmtu(sk) ?
1150 			 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1151 	cork->dst = &rt->dst;
1152 	cork->length = 0;
1153 	cork->ttl = ipc->ttl;
1154 	cork->tos = ipc->tos;
1155 	cork->priority = ipc->priority;
1156 	cork->tx_flags = ipc->tx_flags;
1157 
1158 	return 0;
1159 }
1160 
1161 /*
1162  *	ip_append_data() and ip_append_page() can make one large IP datagram
1163  *	from many pieces of data. Each pieces will be holded on the socket
1164  *	until ip_push_pending_frames() is called. Each piece can be a page
1165  *	or non-page data.
1166  *
1167  *	Not only UDP, other transport protocols - e.g. raw sockets - can use
1168  *	this interface potentially.
1169  *
1170  *	LATER: length must be adjusted by pad at tail, when it is required.
1171  */
1172 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1173 		   int getfrag(void *from, char *to, int offset, int len,
1174 			       int odd, struct sk_buff *skb),
1175 		   void *from, int length, int transhdrlen,
1176 		   struct ipcm_cookie *ipc, struct rtable **rtp,
1177 		   unsigned int flags)
1178 {
1179 	struct inet_sock *inet = inet_sk(sk);
1180 	int err;
1181 
1182 	if (flags&MSG_PROBE)
1183 		return 0;
1184 
1185 	if (skb_queue_empty(&sk->sk_write_queue)) {
1186 		err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1187 		if (err)
1188 			return err;
1189 	} else {
1190 		transhdrlen = 0;
1191 	}
1192 
1193 	return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1194 				sk_page_frag(sk), getfrag,
1195 				from, length, transhdrlen, flags);
1196 }
1197 
1198 ssize_t	ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1199 		       int offset, size_t size, int flags)
1200 {
1201 	struct inet_sock *inet = inet_sk(sk);
1202 	struct sk_buff *skb;
1203 	struct rtable *rt;
1204 	struct ip_options *opt = NULL;
1205 	struct inet_cork *cork;
1206 	int hh_len;
1207 	int mtu;
1208 	int len;
1209 	int err;
1210 	unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1211 
1212 	if (inet->hdrincl)
1213 		return -EPERM;
1214 
1215 	if (flags&MSG_PROBE)
1216 		return 0;
1217 
1218 	if (skb_queue_empty(&sk->sk_write_queue))
1219 		return -EINVAL;
1220 
1221 	cork = &inet->cork.base;
1222 	rt = (struct rtable *)cork->dst;
1223 	if (cork->flags & IPCORK_OPT)
1224 		opt = cork->opt;
1225 
1226 	if (!(rt->dst.dev->features&NETIF_F_SG))
1227 		return -EOPNOTSUPP;
1228 
1229 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1230 	mtu = cork->fragsize;
1231 
1232 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1233 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1234 	maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1235 
1236 	if (cork->length + size > maxnonfragsize - fragheaderlen) {
1237 		ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1238 			       mtu - (opt ? opt->optlen : 0));
1239 		return -EMSGSIZE;
1240 	}
1241 
1242 	skb = skb_peek_tail(&sk->sk_write_queue);
1243 	if (!skb)
1244 		return -EINVAL;
1245 
1246 	cork->length += size;
1247 	if ((size + skb->len > mtu) &&
1248 	    (sk->sk_protocol == IPPROTO_UDP) &&
1249 	    (rt->dst.dev->features & NETIF_F_UFO)) {
1250 		skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1251 		skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1252 	}
1253 
1254 	while (size > 0) {
1255 		if (skb_is_gso(skb)) {
1256 			len = size;
1257 		} else {
1258 
1259 			/* Check if the remaining data fits into current packet. */
1260 			len = mtu - skb->len;
1261 			if (len < size)
1262 				len = maxfraglen - skb->len;
1263 		}
1264 		if (len <= 0) {
1265 			struct sk_buff *skb_prev;
1266 			int alloclen;
1267 
1268 			skb_prev = skb;
1269 			fraggap = skb_prev->len - maxfraglen;
1270 
1271 			alloclen = fragheaderlen + hh_len + fraggap + 15;
1272 			skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1273 			if (unlikely(!skb)) {
1274 				err = -ENOBUFS;
1275 				goto error;
1276 			}
1277 
1278 			/*
1279 			 *	Fill in the control structures
1280 			 */
1281 			skb->ip_summed = CHECKSUM_NONE;
1282 			skb->csum = 0;
1283 			skb_reserve(skb, hh_len);
1284 
1285 			/*
1286 			 *	Find where to start putting bytes.
1287 			 */
1288 			skb_put(skb, fragheaderlen + fraggap);
1289 			skb_reset_network_header(skb);
1290 			skb->transport_header = (skb->network_header +
1291 						 fragheaderlen);
1292 			if (fraggap) {
1293 				skb->csum = skb_copy_and_csum_bits(skb_prev,
1294 								   maxfraglen,
1295 						    skb_transport_header(skb),
1296 								   fraggap, 0);
1297 				skb_prev->csum = csum_sub(skb_prev->csum,
1298 							  skb->csum);
1299 				pskb_trim_unique(skb_prev, maxfraglen);
1300 			}
1301 
1302 			/*
1303 			 * Put the packet on the pending queue.
1304 			 */
1305 			__skb_queue_tail(&sk->sk_write_queue, skb);
1306 			continue;
1307 		}
1308 
1309 		if (len > size)
1310 			len = size;
1311 
1312 		if (skb_append_pagefrags(skb, page, offset, len)) {
1313 			err = -EMSGSIZE;
1314 			goto error;
1315 		}
1316 
1317 		if (skb->ip_summed == CHECKSUM_NONE) {
1318 			__wsum csum;
1319 			csum = csum_page(page, offset, len);
1320 			skb->csum = csum_block_add(skb->csum, csum, skb->len);
1321 		}
1322 
1323 		skb->len += len;
1324 		skb->data_len += len;
1325 		skb->truesize += len;
1326 		atomic_add(len, &sk->sk_wmem_alloc);
1327 		offset += len;
1328 		size -= len;
1329 	}
1330 	return 0;
1331 
1332 error:
1333 	cork->length -= size;
1334 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1335 	return err;
1336 }
1337 
1338 static void ip_cork_release(struct inet_cork *cork)
1339 {
1340 	cork->flags &= ~IPCORK_OPT;
1341 	kfree(cork->opt);
1342 	cork->opt = NULL;
1343 	dst_release(cork->dst);
1344 	cork->dst = NULL;
1345 }
1346 
1347 /*
1348  *	Combined all pending IP fragments on the socket as one IP datagram
1349  *	and push them out.
1350  */
1351 struct sk_buff *__ip_make_skb(struct sock *sk,
1352 			      struct flowi4 *fl4,
1353 			      struct sk_buff_head *queue,
1354 			      struct inet_cork *cork)
1355 {
1356 	struct sk_buff *skb, *tmp_skb;
1357 	struct sk_buff **tail_skb;
1358 	struct inet_sock *inet = inet_sk(sk);
1359 	struct net *net = sock_net(sk);
1360 	struct ip_options *opt = NULL;
1361 	struct rtable *rt = (struct rtable *)cork->dst;
1362 	struct iphdr *iph;
1363 	__be16 df = 0;
1364 	__u8 ttl;
1365 
1366 	skb = __skb_dequeue(queue);
1367 	if (!skb)
1368 		goto out;
1369 	tail_skb = &(skb_shinfo(skb)->frag_list);
1370 
1371 	/* move skb->data to ip header from ext header */
1372 	if (skb->data < skb_network_header(skb))
1373 		__skb_pull(skb, skb_network_offset(skb));
1374 	while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1375 		__skb_pull(tmp_skb, skb_network_header_len(skb));
1376 		*tail_skb = tmp_skb;
1377 		tail_skb = &(tmp_skb->next);
1378 		skb->len += tmp_skb->len;
1379 		skb->data_len += tmp_skb->len;
1380 		skb->truesize += tmp_skb->truesize;
1381 		tmp_skb->destructor = NULL;
1382 		tmp_skb->sk = NULL;
1383 	}
1384 
1385 	/* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1386 	 * to fragment the frame generated here. No matter, what transforms
1387 	 * how transforms change size of the packet, it will come out.
1388 	 */
1389 	skb->ignore_df = ip_sk_ignore_df(sk);
1390 
1391 	/* DF bit is set when we want to see DF on outgoing frames.
1392 	 * If ignore_df is set too, we still allow to fragment this frame
1393 	 * locally. */
1394 	if (inet->pmtudisc == IP_PMTUDISC_DO ||
1395 	    inet->pmtudisc == IP_PMTUDISC_PROBE ||
1396 	    (skb->len <= dst_mtu(&rt->dst) &&
1397 	     ip_dont_fragment(sk, &rt->dst)))
1398 		df = htons(IP_DF);
1399 
1400 	if (cork->flags & IPCORK_OPT)
1401 		opt = cork->opt;
1402 
1403 	if (cork->ttl != 0)
1404 		ttl = cork->ttl;
1405 	else if (rt->rt_type == RTN_MULTICAST)
1406 		ttl = inet->mc_ttl;
1407 	else
1408 		ttl = ip_select_ttl(inet, &rt->dst);
1409 
1410 	iph = ip_hdr(skb);
1411 	iph->version = 4;
1412 	iph->ihl = 5;
1413 	iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1414 	iph->frag_off = df;
1415 	iph->ttl = ttl;
1416 	iph->protocol = sk->sk_protocol;
1417 	ip_copy_addrs(iph, fl4);
1418 	ip_select_ident(net, skb, sk);
1419 
1420 	if (opt) {
1421 		iph->ihl += opt->optlen>>2;
1422 		ip_options_build(skb, opt, cork->addr, rt, 0);
1423 	}
1424 
1425 	skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1426 	skb->mark = sk->sk_mark;
1427 	/*
1428 	 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1429 	 * on dst refcount
1430 	 */
1431 	cork->dst = NULL;
1432 	skb_dst_set(skb, &rt->dst);
1433 
1434 	if (iph->protocol == IPPROTO_ICMP)
1435 		icmp_out_count(net, ((struct icmphdr *)
1436 			skb_transport_header(skb))->type);
1437 
1438 	ip_cork_release(cork);
1439 out:
1440 	return skb;
1441 }
1442 
1443 int ip_send_skb(struct net *net, struct sk_buff *skb)
1444 {
1445 	int err;
1446 
1447 	err = ip_local_out(skb);
1448 	if (err) {
1449 		if (err > 0)
1450 			err = net_xmit_errno(err);
1451 		if (err)
1452 			IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1453 	}
1454 
1455 	return err;
1456 }
1457 
1458 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1459 {
1460 	struct sk_buff *skb;
1461 
1462 	skb = ip_finish_skb(sk, fl4);
1463 	if (!skb)
1464 		return 0;
1465 
1466 	/* Netfilter gets whole the not fragmented skb. */
1467 	return ip_send_skb(sock_net(sk), skb);
1468 }
1469 
1470 /*
1471  *	Throw away all pending data on the socket.
1472  */
1473 static void __ip_flush_pending_frames(struct sock *sk,
1474 				      struct sk_buff_head *queue,
1475 				      struct inet_cork *cork)
1476 {
1477 	struct sk_buff *skb;
1478 
1479 	while ((skb = __skb_dequeue_tail(queue)) != NULL)
1480 		kfree_skb(skb);
1481 
1482 	ip_cork_release(cork);
1483 }
1484 
1485 void ip_flush_pending_frames(struct sock *sk)
1486 {
1487 	__ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1488 }
1489 
1490 struct sk_buff *ip_make_skb(struct sock *sk,
1491 			    struct flowi4 *fl4,
1492 			    int getfrag(void *from, char *to, int offset,
1493 					int len, int odd, struct sk_buff *skb),
1494 			    void *from, int length, int transhdrlen,
1495 			    struct ipcm_cookie *ipc, struct rtable **rtp,
1496 			    unsigned int flags)
1497 {
1498 	struct inet_cork cork;
1499 	struct sk_buff_head queue;
1500 	int err;
1501 
1502 	if (flags & MSG_PROBE)
1503 		return NULL;
1504 
1505 	__skb_queue_head_init(&queue);
1506 
1507 	cork.flags = 0;
1508 	cork.addr = 0;
1509 	cork.opt = NULL;
1510 	err = ip_setup_cork(sk, &cork, ipc, rtp);
1511 	if (err)
1512 		return ERR_PTR(err);
1513 
1514 	err = __ip_append_data(sk, fl4, &queue, &cork,
1515 			       &current->task_frag, getfrag,
1516 			       from, length, transhdrlen, flags);
1517 	if (err) {
1518 		__ip_flush_pending_frames(sk, &queue, &cork);
1519 		return ERR_PTR(err);
1520 	}
1521 
1522 	return __ip_make_skb(sk, fl4, &queue, &cork);
1523 }
1524 
1525 /*
1526  *	Fetch data from kernel space and fill in checksum if needed.
1527  */
1528 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1529 			      int len, int odd, struct sk_buff *skb)
1530 {
1531 	__wsum csum;
1532 
1533 	csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1534 	skb->csum = csum_block_add(skb->csum, csum, odd);
1535 	return 0;
1536 }
1537 
1538 /*
1539  *	Generic function to send a packet as reply to another packet.
1540  *	Used to send some TCP resets/acks so far.
1541  */
1542 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1543 			   const struct ip_options *sopt,
1544 			   __be32 daddr, __be32 saddr,
1545 			   const struct ip_reply_arg *arg,
1546 			   unsigned int len)
1547 {
1548 	struct ip_options_data replyopts;
1549 	struct ipcm_cookie ipc;
1550 	struct flowi4 fl4;
1551 	struct rtable *rt = skb_rtable(skb);
1552 	struct net *net = sock_net(sk);
1553 	struct sk_buff *nskb;
1554 	int err;
1555 	int oif;
1556 
1557 	if (__ip_options_echo(&replyopts.opt.opt, skb, sopt))
1558 		return;
1559 
1560 	ipc.addr = daddr;
1561 	ipc.opt = NULL;
1562 	ipc.tx_flags = 0;
1563 	ipc.ttl = 0;
1564 	ipc.tos = -1;
1565 
1566 	if (replyopts.opt.opt.optlen) {
1567 		ipc.opt = &replyopts.opt;
1568 
1569 		if (replyopts.opt.opt.srr)
1570 			daddr = replyopts.opt.opt.faddr;
1571 	}
1572 
1573 	oif = arg->bound_dev_if;
1574 	if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1575 		oif = skb->skb_iif;
1576 
1577 	flowi4_init_output(&fl4, oif,
1578 			   IP4_REPLY_MARK(net, skb->mark),
1579 			   RT_TOS(arg->tos),
1580 			   RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1581 			   ip_reply_arg_flowi_flags(arg),
1582 			   daddr, saddr,
1583 			   tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1584 	security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1585 	rt = ip_route_output_key(net, &fl4);
1586 	if (IS_ERR(rt))
1587 		return;
1588 
1589 	inet_sk(sk)->tos = arg->tos;
1590 
1591 	sk->sk_priority = skb->priority;
1592 	sk->sk_protocol = ip_hdr(skb)->protocol;
1593 	sk->sk_bound_dev_if = arg->bound_dev_if;
1594 	sk->sk_sndbuf = sysctl_wmem_default;
1595 	err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1596 			     len, 0, &ipc, &rt, MSG_DONTWAIT);
1597 	if (unlikely(err)) {
1598 		ip_flush_pending_frames(sk);
1599 		goto out;
1600 	}
1601 
1602 	nskb = skb_peek(&sk->sk_write_queue);
1603 	if (nskb) {
1604 		if (arg->csumoffset >= 0)
1605 			*((__sum16 *)skb_transport_header(nskb) +
1606 			  arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1607 								arg->csum));
1608 		nskb->ip_summed = CHECKSUM_NONE;
1609 		skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1610 		ip_push_pending_frames(sk, &fl4);
1611 	}
1612 out:
1613 	ip_rt_put(rt);
1614 }
1615 
1616 void __init ip_init(void)
1617 {
1618 	ip_rt_init();
1619 	inet_initpeers();
1620 
1621 #if defined(CONFIG_IP_MULTICAST)
1622 	igmp_mc_init();
1623 #endif
1624 }
1625