xref: /linux/net/ipv4/ip_output.c (revision d39d0ed196aa1685bb24771e92f78633c66ac9cb)
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 <asm/system.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/inetpeer.h>
77 #include <linux/igmp.h>
78 #include <linux/netfilter_ipv4.h>
79 #include <linux/netfilter_bridge.h>
80 #include <linux/mroute.h>
81 #include <linux/netlink.h>
82 #include <linux/tcp.h>
83 
84 int sysctl_ip_default_ttl __read_mostly = IPDEFTTL;
85 
86 /* Generate a checksum for an outgoing IP datagram. */
87 __inline__ void ip_send_check(struct iphdr *iph)
88 {
89 	iph->check = 0;
90 	iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
91 }
92 EXPORT_SYMBOL(ip_send_check);
93 
94 int __ip_local_out(struct sk_buff *skb)
95 {
96 	struct iphdr *iph = ip_hdr(skb);
97 
98 	iph->tot_len = htons(skb->len);
99 	ip_send_check(iph);
100 	return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT, skb, NULL,
101 		       skb_dst(skb)->dev, dst_output);
102 }
103 
104 int ip_local_out(struct sk_buff *skb)
105 {
106 	int err;
107 
108 	err = __ip_local_out(skb);
109 	if (likely(err == 1))
110 		err = dst_output(skb);
111 
112 	return err;
113 }
114 EXPORT_SYMBOL_GPL(ip_local_out);
115 
116 /* dev_loopback_xmit for use with netfilter. */
117 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
118 {
119 	skb_reset_mac_header(newskb);
120 	__skb_pull(newskb, skb_network_offset(newskb));
121 	newskb->pkt_type = PACKET_LOOPBACK;
122 	newskb->ip_summed = CHECKSUM_UNNECESSARY;
123 	WARN_ON(!skb_dst(newskb));
124 	netif_rx_ni(newskb);
125 	return 0;
126 }
127 
128 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
129 {
130 	int ttl = inet->uc_ttl;
131 
132 	if (ttl < 0)
133 		ttl = dst_metric(dst, RTAX_HOPLIMIT);
134 	return ttl;
135 }
136 
137 /*
138  *		Add an ip header to a skbuff and send it out.
139  *
140  */
141 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
142 			  __be32 saddr, __be32 daddr, struct ip_options *opt)
143 {
144 	struct inet_sock *inet = inet_sk(sk);
145 	struct rtable *rt = skb_rtable(skb);
146 	struct iphdr *iph;
147 
148 	/* Build the IP header. */
149 	skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
150 	skb_reset_network_header(skb);
151 	iph = ip_hdr(skb);
152 	iph->version  = 4;
153 	iph->ihl      = 5;
154 	iph->tos      = inet->tos;
155 	if (ip_dont_fragment(sk, &rt->dst))
156 		iph->frag_off = htons(IP_DF);
157 	else
158 		iph->frag_off = 0;
159 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
160 	iph->daddr    = rt->rt_dst;
161 	iph->saddr    = rt->rt_src;
162 	iph->protocol = sk->sk_protocol;
163 	ip_select_ident(iph, &rt->dst, sk);
164 
165 	if (opt && opt->optlen) {
166 		iph->ihl += opt->optlen>>2;
167 		ip_options_build(skb, opt, daddr, rt, 0);
168 	}
169 
170 	skb->priority = sk->sk_priority;
171 	skb->mark = sk->sk_mark;
172 
173 	/* Send it out. */
174 	return ip_local_out(skb);
175 }
176 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
177 
178 static inline int ip_finish_output2(struct sk_buff *skb)
179 {
180 	struct dst_entry *dst = skb_dst(skb);
181 	struct rtable *rt = (struct rtable *)dst;
182 	struct net_device *dev = dst->dev;
183 	unsigned int hh_len = LL_RESERVED_SPACE(dev);
184 
185 	if (rt->rt_type == RTN_MULTICAST) {
186 		IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
187 	} else if (rt->rt_type == RTN_BROADCAST)
188 		IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
189 
190 	/* Be paranoid, rather than too clever. */
191 	if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
192 		struct sk_buff *skb2;
193 
194 		skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
195 		if (skb2 == NULL) {
196 			kfree_skb(skb);
197 			return -ENOMEM;
198 		}
199 		if (skb->sk)
200 			skb_set_owner_w(skb2, skb->sk);
201 		kfree_skb(skb);
202 		skb = skb2;
203 	}
204 
205 	if (dst->hh)
206 		return neigh_hh_output(dst->hh, skb);
207 	else if (dst->neighbour)
208 		return dst->neighbour->output(skb);
209 
210 	if (net_ratelimit())
211 		printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
212 	kfree_skb(skb);
213 	return -EINVAL;
214 }
215 
216 static inline int ip_skb_dst_mtu(struct sk_buff *skb)
217 {
218 	struct inet_sock *inet = skb->sk ? inet_sk(skb->sk) : NULL;
219 
220 	return (inet && inet->pmtudisc == IP_PMTUDISC_PROBE) ?
221 	       skb_dst(skb)->dev->mtu : dst_mtu(skb_dst(skb));
222 }
223 
224 static int ip_finish_output(struct sk_buff *skb)
225 {
226 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
227 	/* Policy lookup after SNAT yielded a new policy */
228 	if (skb_dst(skb)->xfrm != NULL) {
229 		IPCB(skb)->flags |= IPSKB_REROUTED;
230 		return dst_output(skb);
231 	}
232 #endif
233 	if (skb->len > ip_skb_dst_mtu(skb) && !skb_is_gso(skb))
234 		return ip_fragment(skb, ip_finish_output2);
235 	else
236 		return ip_finish_output2(skb);
237 }
238 
239 int ip_mc_output(struct sk_buff *skb)
240 {
241 	struct sock *sk = skb->sk;
242 	struct rtable *rt = skb_rtable(skb);
243 	struct net_device *dev = rt->dst.dev;
244 
245 	/*
246 	 *	If the indicated interface is up and running, send the packet.
247 	 */
248 	IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
249 
250 	skb->dev = dev;
251 	skb->protocol = htons(ETH_P_IP);
252 
253 	/*
254 	 *	Multicasts are looped back for other local users
255 	 */
256 
257 	if (rt->rt_flags&RTCF_MULTICAST) {
258 		if (sk_mc_loop(sk)
259 #ifdef CONFIG_IP_MROUTE
260 		/* Small optimization: do not loopback not local frames,
261 		   which returned after forwarding; they will be  dropped
262 		   by ip_mr_input in any case.
263 		   Note, that local frames are looped back to be delivered
264 		   to local recipients.
265 
266 		   This check is duplicated in ip_mr_input at the moment.
267 		 */
268 		    &&
269 		    ((rt->rt_flags & RTCF_LOCAL) ||
270 		     !(IPCB(skb)->flags & IPSKB_FORWARDED))
271 #endif
272 		   ) {
273 			struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
274 			if (newskb)
275 				NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
276 					newskb, NULL, newskb->dev,
277 					ip_dev_loopback_xmit);
278 		}
279 
280 		/* Multicasts with ttl 0 must not go beyond the host */
281 
282 		if (ip_hdr(skb)->ttl == 0) {
283 			kfree_skb(skb);
284 			return 0;
285 		}
286 	}
287 
288 	if (rt->rt_flags&RTCF_BROADCAST) {
289 		struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
290 		if (newskb)
291 			NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
292 				NULL, newskb->dev, ip_dev_loopback_xmit);
293 	}
294 
295 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
296 			    skb->dev, ip_finish_output,
297 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
298 }
299 
300 int ip_output(struct sk_buff *skb)
301 {
302 	struct net_device *dev = skb_dst(skb)->dev;
303 
304 	IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
305 
306 	skb->dev = dev;
307 	skb->protocol = htons(ETH_P_IP);
308 
309 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
310 			    ip_finish_output,
311 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
312 }
313 
314 int ip_queue_xmit(struct sk_buff *skb)
315 {
316 	struct sock *sk = skb->sk;
317 	struct inet_sock *inet = inet_sk(sk);
318 	struct ip_options *opt = inet->opt;
319 	struct rtable *rt;
320 	struct iphdr *iph;
321 	int res;
322 
323 	/* Skip all of this if the packet is already routed,
324 	 * f.e. by something like SCTP.
325 	 */
326 	rcu_read_lock();
327 	rt = skb_rtable(skb);
328 	if (rt != NULL)
329 		goto packet_routed;
330 
331 	/* Make sure we can route this packet. */
332 	rt = (struct rtable *)__sk_dst_check(sk, 0);
333 	if (rt == NULL) {
334 		__be32 daddr;
335 
336 		/* Use correct destination address if we have options. */
337 		daddr = inet->inet_daddr;
338 		if(opt && opt->srr)
339 			daddr = opt->faddr;
340 
341 		{
342 			struct flowi fl = { .oif = sk->sk_bound_dev_if,
343 					    .mark = sk->sk_mark,
344 					    .nl_u = { .ip4_u =
345 						      { .daddr = daddr,
346 							.saddr = inet->inet_saddr,
347 							.tos = RT_CONN_FLAGS(sk) } },
348 					    .proto = sk->sk_protocol,
349 					    .flags = inet_sk_flowi_flags(sk),
350 					    .uli_u = { .ports =
351 						       { .sport = inet->inet_sport,
352 							 .dport = inet->inet_dport } } };
353 
354 			/* If this fails, retransmit mechanism of transport layer will
355 			 * keep trying until route appears or the connection times
356 			 * itself out.
357 			 */
358 			security_sk_classify_flow(sk, &fl);
359 			if (ip_route_output_flow(sock_net(sk), &rt, &fl, sk, 0))
360 				goto no_route;
361 		}
362 		sk_setup_caps(sk, &rt->dst);
363 	}
364 	skb_dst_set_noref(skb, &rt->dst);
365 
366 packet_routed:
367 	if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
368 		goto no_route;
369 
370 	/* OK, we know where to send it, allocate and build IP header. */
371 	skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
372 	skb_reset_network_header(skb);
373 	iph = ip_hdr(skb);
374 	*((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
375 	if (ip_dont_fragment(sk, &rt->dst) && !skb->local_df)
376 		iph->frag_off = htons(IP_DF);
377 	else
378 		iph->frag_off = 0;
379 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
380 	iph->protocol = sk->sk_protocol;
381 	iph->saddr    = rt->rt_src;
382 	iph->daddr    = rt->rt_dst;
383 	/* Transport layer set skb->h.foo itself. */
384 
385 	if (opt && opt->optlen) {
386 		iph->ihl += opt->optlen >> 2;
387 		ip_options_build(skb, opt, inet->inet_daddr, rt, 0);
388 	}
389 
390 	ip_select_ident_more(iph, &rt->dst, sk,
391 			     (skb_shinfo(skb)->gso_segs ?: 1) - 1);
392 
393 	skb->priority = sk->sk_priority;
394 	skb->mark = sk->sk_mark;
395 
396 	res = ip_local_out(skb);
397 	rcu_read_unlock();
398 	return res;
399 
400 no_route:
401 	rcu_read_unlock();
402 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
403 	kfree_skb(skb);
404 	return -EHOSTUNREACH;
405 }
406 EXPORT_SYMBOL(ip_queue_xmit);
407 
408 
409 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
410 {
411 	to->pkt_type = from->pkt_type;
412 	to->priority = from->priority;
413 	to->protocol = from->protocol;
414 	skb_dst_drop(to);
415 	skb_dst_copy(to, from);
416 	to->dev = from->dev;
417 	to->mark = from->mark;
418 
419 	/* Copy the flags to each fragment. */
420 	IPCB(to)->flags = IPCB(from)->flags;
421 
422 #ifdef CONFIG_NET_SCHED
423 	to->tc_index = from->tc_index;
424 #endif
425 	nf_copy(to, from);
426 #if defined(CONFIG_NETFILTER_XT_TARGET_TRACE) || \
427     defined(CONFIG_NETFILTER_XT_TARGET_TRACE_MODULE)
428 	to->nf_trace = from->nf_trace;
429 #endif
430 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
431 	to->ipvs_property = from->ipvs_property;
432 #endif
433 	skb_copy_secmark(to, from);
434 }
435 
436 /*
437  *	This IP datagram is too large to be sent in one piece.  Break it up into
438  *	smaller pieces (each of size equal to IP header plus
439  *	a block of the data of the original IP data part) that will yet fit in a
440  *	single device frame, and queue such a frame for sending.
441  */
442 
443 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
444 {
445 	struct iphdr *iph;
446 	int ptr;
447 	struct net_device *dev;
448 	struct sk_buff *skb2;
449 	unsigned int mtu, hlen, left, len, ll_rs;
450 	int offset;
451 	__be16 not_last_frag;
452 	struct rtable *rt = skb_rtable(skb);
453 	int err = 0;
454 
455 	dev = rt->dst.dev;
456 
457 	/*
458 	 *	Point into the IP datagram header.
459 	 */
460 
461 	iph = ip_hdr(skb);
462 
463 	if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
464 		IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
465 		icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
466 			  htonl(ip_skb_dst_mtu(skb)));
467 		kfree_skb(skb);
468 		return -EMSGSIZE;
469 	}
470 
471 	/*
472 	 *	Setup starting values.
473 	 */
474 
475 	hlen = iph->ihl * 4;
476 	mtu = dst_mtu(&rt->dst) - hlen;	/* Size of data space */
477 #ifdef CONFIG_BRIDGE_NETFILTER
478 	if (skb->nf_bridge)
479 		mtu -= nf_bridge_mtu_reduction(skb);
480 #endif
481 	IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
482 
483 	/* When frag_list is given, use it. First, check its validity:
484 	 * some transformers could create wrong frag_list or break existing
485 	 * one, it is not prohibited. In this case fall back to copying.
486 	 *
487 	 * LATER: this step can be merged to real generation of fragments,
488 	 * we can switch to copy when see the first bad fragment.
489 	 */
490 	if (skb_has_frags(skb)) {
491 		struct sk_buff *frag;
492 		int first_len = skb_pagelen(skb);
493 		int truesizes = 0;
494 
495 		if (first_len - hlen > mtu ||
496 		    ((first_len - hlen) & 7) ||
497 		    (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
498 		    skb_cloned(skb))
499 			goto slow_path;
500 
501 		skb_walk_frags(skb, frag) {
502 			/* Correct geometry. */
503 			if (frag->len > mtu ||
504 			    ((frag->len & 7) && frag->next) ||
505 			    skb_headroom(frag) < hlen)
506 			    goto slow_path;
507 
508 			/* Partially cloned skb? */
509 			if (skb_shared(frag))
510 				goto slow_path;
511 
512 			BUG_ON(frag->sk);
513 			if (skb->sk) {
514 				frag->sk = skb->sk;
515 				frag->destructor = sock_wfree;
516 			}
517 			truesizes += frag->truesize;
518 		}
519 
520 		/* Everything is OK. Generate! */
521 
522 		err = 0;
523 		offset = 0;
524 		frag = skb_shinfo(skb)->frag_list;
525 		skb_frag_list_init(skb);
526 		skb->data_len = first_len - skb_headlen(skb);
527 		skb->truesize -= truesizes;
528 		skb->len = first_len;
529 		iph->tot_len = htons(first_len);
530 		iph->frag_off = htons(IP_MF);
531 		ip_send_check(iph);
532 
533 		for (;;) {
534 			/* Prepare header of the next frame,
535 			 * before previous one went down. */
536 			if (frag) {
537 				frag->ip_summed = CHECKSUM_NONE;
538 				skb_reset_transport_header(frag);
539 				__skb_push(frag, hlen);
540 				skb_reset_network_header(frag);
541 				memcpy(skb_network_header(frag), iph, hlen);
542 				iph = ip_hdr(frag);
543 				iph->tot_len = htons(frag->len);
544 				ip_copy_metadata(frag, skb);
545 				if (offset == 0)
546 					ip_options_fragment(frag);
547 				offset += skb->len - hlen;
548 				iph->frag_off = htons(offset>>3);
549 				if (frag->next != NULL)
550 					iph->frag_off |= htons(IP_MF);
551 				/* Ready, complete checksum */
552 				ip_send_check(iph);
553 			}
554 
555 			err = output(skb);
556 
557 			if (!err)
558 				IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
559 			if (err || !frag)
560 				break;
561 
562 			skb = frag;
563 			frag = skb->next;
564 			skb->next = NULL;
565 		}
566 
567 		if (err == 0) {
568 			IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
569 			return 0;
570 		}
571 
572 		while (frag) {
573 			skb = frag->next;
574 			kfree_skb(frag);
575 			frag = skb;
576 		}
577 		IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
578 		return err;
579 	}
580 
581 slow_path:
582 	left = skb->len - hlen;		/* Space per frame */
583 	ptr = hlen;		/* Where to start from */
584 
585 	/* for bridged IP traffic encapsulated inside f.e. a vlan header,
586 	 * we need to make room for the encapsulating header
587 	 */
588 	ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
589 
590 	/*
591 	 *	Fragment the datagram.
592 	 */
593 
594 	offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
595 	not_last_frag = iph->frag_off & htons(IP_MF);
596 
597 	/*
598 	 *	Keep copying data until we run out.
599 	 */
600 
601 	while (left > 0) {
602 		len = left;
603 		/* IF: it doesn't fit, use 'mtu' - the data space left */
604 		if (len > mtu)
605 			len = mtu;
606 		/* IF: we are not sending upto and including the packet end
607 		   then align the next start on an eight byte boundary */
608 		if (len < left)	{
609 			len &= ~7;
610 		}
611 		/*
612 		 *	Allocate buffer.
613 		 */
614 
615 		if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
616 			NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
617 			err = -ENOMEM;
618 			goto fail;
619 		}
620 
621 		/*
622 		 *	Set up data on packet
623 		 */
624 
625 		ip_copy_metadata(skb2, skb);
626 		skb_reserve(skb2, ll_rs);
627 		skb_put(skb2, len + hlen);
628 		skb_reset_network_header(skb2);
629 		skb2->transport_header = skb2->network_header + hlen;
630 
631 		/*
632 		 *	Charge the memory for the fragment to any owner
633 		 *	it might possess
634 		 */
635 
636 		if (skb->sk)
637 			skb_set_owner_w(skb2, skb->sk);
638 
639 		/*
640 		 *	Copy the packet header into the new buffer.
641 		 */
642 
643 		skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
644 
645 		/*
646 		 *	Copy a block of the IP datagram.
647 		 */
648 		if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
649 			BUG();
650 		left -= len;
651 
652 		/*
653 		 *	Fill in the new header fields.
654 		 */
655 		iph = ip_hdr(skb2);
656 		iph->frag_off = htons((offset >> 3));
657 
658 		/* ANK: dirty, but effective trick. Upgrade options only if
659 		 * the segment to be fragmented was THE FIRST (otherwise,
660 		 * options are already fixed) and make it ONCE
661 		 * on the initial skb, so that all the following fragments
662 		 * will inherit fixed options.
663 		 */
664 		if (offset == 0)
665 			ip_options_fragment(skb);
666 
667 		/*
668 		 *	Added AC : If we are fragmenting a fragment that's not the
669 		 *		   last fragment then keep MF on each bit
670 		 */
671 		if (left > 0 || not_last_frag)
672 			iph->frag_off |= htons(IP_MF);
673 		ptr += len;
674 		offset += len;
675 
676 		/*
677 		 *	Put this fragment into the sending queue.
678 		 */
679 		iph->tot_len = htons(len + hlen);
680 
681 		ip_send_check(iph);
682 
683 		err = output(skb2);
684 		if (err)
685 			goto fail;
686 
687 		IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
688 	}
689 	kfree_skb(skb);
690 	IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
691 	return err;
692 
693 fail:
694 	kfree_skb(skb);
695 	IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
696 	return err;
697 }
698 EXPORT_SYMBOL(ip_fragment);
699 
700 int
701 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
702 {
703 	struct iovec *iov = from;
704 
705 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
706 		if (memcpy_fromiovecend(to, iov, offset, len) < 0)
707 			return -EFAULT;
708 	} else {
709 		__wsum csum = 0;
710 		if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
711 			return -EFAULT;
712 		skb->csum = csum_block_add(skb->csum, csum, odd);
713 	}
714 	return 0;
715 }
716 EXPORT_SYMBOL(ip_generic_getfrag);
717 
718 static inline __wsum
719 csum_page(struct page *page, int offset, int copy)
720 {
721 	char *kaddr;
722 	__wsum csum;
723 	kaddr = kmap(page);
724 	csum = csum_partial(kaddr + offset, copy, 0);
725 	kunmap(page);
726 	return csum;
727 }
728 
729 static inline int ip_ufo_append_data(struct sock *sk,
730 			int getfrag(void *from, char *to, int offset, int len,
731 			       int odd, struct sk_buff *skb),
732 			void *from, int length, int hh_len, int fragheaderlen,
733 			int transhdrlen, int mtu, unsigned int flags)
734 {
735 	struct sk_buff *skb;
736 	int err;
737 
738 	/* There is support for UDP fragmentation offload by network
739 	 * device, so create one single skb packet containing complete
740 	 * udp datagram
741 	 */
742 	if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL) {
743 		skb = sock_alloc_send_skb(sk,
744 			hh_len + fragheaderlen + transhdrlen + 20,
745 			(flags & MSG_DONTWAIT), &err);
746 
747 		if (skb == NULL)
748 			return err;
749 
750 		/* reserve space for Hardware header */
751 		skb_reserve(skb, hh_len);
752 
753 		/* create space for UDP/IP header */
754 		skb_put(skb, fragheaderlen + transhdrlen);
755 
756 		/* initialize network header pointer */
757 		skb_reset_network_header(skb);
758 
759 		/* initialize protocol header pointer */
760 		skb->transport_header = skb->network_header + fragheaderlen;
761 
762 		skb->ip_summed = CHECKSUM_PARTIAL;
763 		skb->csum = 0;
764 		sk->sk_sndmsg_off = 0;
765 
766 		/* specify the length of each IP datagram fragment */
767 		skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
768 		skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
769 		__skb_queue_tail(&sk->sk_write_queue, skb);
770 	}
771 
772 	return skb_append_datato_frags(sk, skb, getfrag, from,
773 				       (length - transhdrlen));
774 }
775 
776 /*
777  *	ip_append_data() and ip_append_page() can make one large IP datagram
778  *	from many pieces of data. Each pieces will be holded on the socket
779  *	until ip_push_pending_frames() is called. Each piece can be a page
780  *	or non-page data.
781  *
782  *	Not only UDP, other transport protocols - e.g. raw sockets - can use
783  *	this interface potentially.
784  *
785  *	LATER: length must be adjusted by pad at tail, when it is required.
786  */
787 int ip_append_data(struct sock *sk,
788 		   int getfrag(void *from, char *to, int offset, int len,
789 			       int odd, struct sk_buff *skb),
790 		   void *from, int length, int transhdrlen,
791 		   struct ipcm_cookie *ipc, struct rtable **rtp,
792 		   unsigned int flags)
793 {
794 	struct inet_sock *inet = inet_sk(sk);
795 	struct sk_buff *skb;
796 
797 	struct ip_options *opt = NULL;
798 	int hh_len;
799 	int exthdrlen;
800 	int mtu;
801 	int copy;
802 	int err;
803 	int offset = 0;
804 	unsigned int maxfraglen, fragheaderlen;
805 	int csummode = CHECKSUM_NONE;
806 	struct rtable *rt;
807 
808 	if (flags&MSG_PROBE)
809 		return 0;
810 
811 	if (skb_queue_empty(&sk->sk_write_queue)) {
812 		/*
813 		 * setup for corking.
814 		 */
815 		opt = ipc->opt;
816 		if (opt) {
817 			if (inet->cork.opt == NULL) {
818 				inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
819 				if (unlikely(inet->cork.opt == NULL))
820 					return -ENOBUFS;
821 			}
822 			memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
823 			inet->cork.flags |= IPCORK_OPT;
824 			inet->cork.addr = ipc->addr;
825 		}
826 		rt = *rtp;
827 		if (unlikely(!rt))
828 			return -EFAULT;
829 		/*
830 		 * We steal reference to this route, caller should not release it
831 		 */
832 		*rtp = NULL;
833 		inet->cork.fragsize = mtu = inet->pmtudisc == IP_PMTUDISC_PROBE ?
834 					    rt->dst.dev->mtu :
835 					    dst_mtu(rt->dst.path);
836 		inet->cork.dst = &rt->dst;
837 		inet->cork.length = 0;
838 		sk->sk_sndmsg_page = NULL;
839 		sk->sk_sndmsg_off = 0;
840 		if ((exthdrlen = rt->dst.header_len) != 0) {
841 			length += exthdrlen;
842 			transhdrlen += exthdrlen;
843 		}
844 	} else {
845 		rt = (struct rtable *)inet->cork.dst;
846 		if (inet->cork.flags & IPCORK_OPT)
847 			opt = inet->cork.opt;
848 
849 		transhdrlen = 0;
850 		exthdrlen = 0;
851 		mtu = inet->cork.fragsize;
852 	}
853 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
854 
855 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
856 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
857 
858 	if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
859 		ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport,
860 			       mtu-exthdrlen);
861 		return -EMSGSIZE;
862 	}
863 
864 	/*
865 	 * transhdrlen > 0 means that this is the first fragment and we wish
866 	 * it won't be fragmented in the future.
867 	 */
868 	if (transhdrlen &&
869 	    length + fragheaderlen <= mtu &&
870 	    rt->dst.dev->features & NETIF_F_V4_CSUM &&
871 	    !exthdrlen)
872 		csummode = CHECKSUM_PARTIAL;
873 
874 	skb = skb_peek_tail(&sk->sk_write_queue);
875 
876 	inet->cork.length += length;
877 	if (((length > mtu) || (skb && skb_is_gso(skb))) &&
878 	    (sk->sk_protocol == IPPROTO_UDP) &&
879 	    (rt->dst.dev->features & NETIF_F_UFO)) {
880 		err = ip_ufo_append_data(sk, getfrag, from, length, hh_len,
881 					 fragheaderlen, transhdrlen, mtu,
882 					 flags);
883 		if (err)
884 			goto error;
885 		return 0;
886 	}
887 
888 	/* So, what's going on in the loop below?
889 	 *
890 	 * We use calculated fragment length to generate chained skb,
891 	 * each of segments is IP fragment ready for sending to network after
892 	 * adding appropriate IP header.
893 	 */
894 
895 	if (!skb)
896 		goto alloc_new_skb;
897 
898 	while (length > 0) {
899 		/* Check if the remaining data fits into current packet. */
900 		copy = mtu - skb->len;
901 		if (copy < length)
902 			copy = maxfraglen - skb->len;
903 		if (copy <= 0) {
904 			char *data;
905 			unsigned int datalen;
906 			unsigned int fraglen;
907 			unsigned int fraggap;
908 			unsigned int alloclen;
909 			struct sk_buff *skb_prev;
910 alloc_new_skb:
911 			skb_prev = skb;
912 			if (skb_prev)
913 				fraggap = skb_prev->len - maxfraglen;
914 			else
915 				fraggap = 0;
916 
917 			/*
918 			 * If remaining data exceeds the mtu,
919 			 * we know we need more fragment(s).
920 			 */
921 			datalen = length + fraggap;
922 			if (datalen > mtu - fragheaderlen)
923 				datalen = maxfraglen - fragheaderlen;
924 			fraglen = datalen + fragheaderlen;
925 
926 			if ((flags & MSG_MORE) &&
927 			    !(rt->dst.dev->features&NETIF_F_SG))
928 				alloclen = mtu;
929 			else
930 				alloclen = datalen + fragheaderlen;
931 
932 			/* The last fragment gets additional space at tail.
933 			 * Note, with MSG_MORE we overallocate on fragments,
934 			 * because we have no idea what fragment will be
935 			 * the last.
936 			 */
937 			if (datalen == length + fraggap)
938 				alloclen += rt->dst.trailer_len;
939 
940 			if (transhdrlen) {
941 				skb = sock_alloc_send_skb(sk,
942 						alloclen + hh_len + 15,
943 						(flags & MSG_DONTWAIT), &err);
944 			} else {
945 				skb = NULL;
946 				if (atomic_read(&sk->sk_wmem_alloc) <=
947 				    2 * sk->sk_sndbuf)
948 					skb = sock_wmalloc(sk,
949 							   alloclen + hh_len + 15, 1,
950 							   sk->sk_allocation);
951 				if (unlikely(skb == NULL))
952 					err = -ENOBUFS;
953 				else
954 					/* only the initial fragment is
955 					   time stamped */
956 					ipc->shtx.flags = 0;
957 			}
958 			if (skb == NULL)
959 				goto error;
960 
961 			/*
962 			 *	Fill in the control structures
963 			 */
964 			skb->ip_summed = csummode;
965 			skb->csum = 0;
966 			skb_reserve(skb, hh_len);
967 			*skb_tx(skb) = ipc->shtx;
968 
969 			/*
970 			 *	Find where to start putting bytes.
971 			 */
972 			data = skb_put(skb, fraglen);
973 			skb_set_network_header(skb, exthdrlen);
974 			skb->transport_header = (skb->network_header +
975 						 fragheaderlen);
976 			data += fragheaderlen;
977 
978 			if (fraggap) {
979 				skb->csum = skb_copy_and_csum_bits(
980 					skb_prev, maxfraglen,
981 					data + transhdrlen, fraggap, 0);
982 				skb_prev->csum = csum_sub(skb_prev->csum,
983 							  skb->csum);
984 				data += fraggap;
985 				pskb_trim_unique(skb_prev, maxfraglen);
986 			}
987 
988 			copy = datalen - transhdrlen - fraggap;
989 			if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
990 				err = -EFAULT;
991 				kfree_skb(skb);
992 				goto error;
993 			}
994 
995 			offset += copy;
996 			length -= datalen - fraggap;
997 			transhdrlen = 0;
998 			exthdrlen = 0;
999 			csummode = CHECKSUM_NONE;
1000 
1001 			/*
1002 			 * Put the packet on the pending queue.
1003 			 */
1004 			__skb_queue_tail(&sk->sk_write_queue, skb);
1005 			continue;
1006 		}
1007 
1008 		if (copy > length)
1009 			copy = length;
1010 
1011 		if (!(rt->dst.dev->features&NETIF_F_SG)) {
1012 			unsigned int off;
1013 
1014 			off = skb->len;
1015 			if (getfrag(from, skb_put(skb, copy),
1016 					offset, copy, off, skb) < 0) {
1017 				__skb_trim(skb, off);
1018 				err = -EFAULT;
1019 				goto error;
1020 			}
1021 		} else {
1022 			int i = skb_shinfo(skb)->nr_frags;
1023 			skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
1024 			struct page *page = sk->sk_sndmsg_page;
1025 			int off = sk->sk_sndmsg_off;
1026 			unsigned int left;
1027 
1028 			if (page && (left = PAGE_SIZE - off) > 0) {
1029 				if (copy >= left)
1030 					copy = left;
1031 				if (page != frag->page) {
1032 					if (i == MAX_SKB_FRAGS) {
1033 						err = -EMSGSIZE;
1034 						goto error;
1035 					}
1036 					get_page(page);
1037 					skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
1038 					frag = &skb_shinfo(skb)->frags[i];
1039 				}
1040 			} else if (i < MAX_SKB_FRAGS) {
1041 				if (copy > PAGE_SIZE)
1042 					copy = PAGE_SIZE;
1043 				page = alloc_pages(sk->sk_allocation, 0);
1044 				if (page == NULL)  {
1045 					err = -ENOMEM;
1046 					goto error;
1047 				}
1048 				sk->sk_sndmsg_page = page;
1049 				sk->sk_sndmsg_off = 0;
1050 
1051 				skb_fill_page_desc(skb, i, page, 0, 0);
1052 				frag = &skb_shinfo(skb)->frags[i];
1053 			} else {
1054 				err = -EMSGSIZE;
1055 				goto error;
1056 			}
1057 			if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1058 				err = -EFAULT;
1059 				goto error;
1060 			}
1061 			sk->sk_sndmsg_off += copy;
1062 			frag->size += copy;
1063 			skb->len += copy;
1064 			skb->data_len += copy;
1065 			skb->truesize += copy;
1066 			atomic_add(copy, &sk->sk_wmem_alloc);
1067 		}
1068 		offset += copy;
1069 		length -= copy;
1070 	}
1071 
1072 	return 0;
1073 
1074 error:
1075 	inet->cork.length -= length;
1076 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1077 	return err;
1078 }
1079 
1080 ssize_t	ip_append_page(struct sock *sk, struct page *page,
1081 		       int offset, size_t size, int flags)
1082 {
1083 	struct inet_sock *inet = inet_sk(sk);
1084 	struct sk_buff *skb;
1085 	struct rtable *rt;
1086 	struct ip_options *opt = NULL;
1087 	int hh_len;
1088 	int mtu;
1089 	int len;
1090 	int err;
1091 	unsigned int maxfraglen, fragheaderlen, fraggap;
1092 
1093 	if (inet->hdrincl)
1094 		return -EPERM;
1095 
1096 	if (flags&MSG_PROBE)
1097 		return 0;
1098 
1099 	if (skb_queue_empty(&sk->sk_write_queue))
1100 		return -EINVAL;
1101 
1102 	rt = (struct rtable *)inet->cork.dst;
1103 	if (inet->cork.flags & IPCORK_OPT)
1104 		opt = inet->cork.opt;
1105 
1106 	if (!(rt->dst.dev->features&NETIF_F_SG))
1107 		return -EOPNOTSUPP;
1108 
1109 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1110 	mtu = inet->cork.fragsize;
1111 
1112 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1113 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1114 
1115 	if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1116 		ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport, mtu);
1117 		return -EMSGSIZE;
1118 	}
1119 
1120 	if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1121 		return -EINVAL;
1122 
1123 	inet->cork.length += size;
1124 	if ((size + skb->len > mtu) &&
1125 	    (sk->sk_protocol == IPPROTO_UDP) &&
1126 	    (rt->dst.dev->features & NETIF_F_UFO)) {
1127 		skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1128 		skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1129 	}
1130 
1131 
1132 	while (size > 0) {
1133 		int i;
1134 
1135 		if (skb_is_gso(skb))
1136 			len = size;
1137 		else {
1138 
1139 			/* Check if the remaining data fits into current packet. */
1140 			len = mtu - skb->len;
1141 			if (len < size)
1142 				len = maxfraglen - skb->len;
1143 		}
1144 		if (len <= 0) {
1145 			struct sk_buff *skb_prev;
1146 			int alloclen;
1147 
1148 			skb_prev = skb;
1149 			fraggap = skb_prev->len - maxfraglen;
1150 
1151 			alloclen = fragheaderlen + hh_len + fraggap + 15;
1152 			skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1153 			if (unlikely(!skb)) {
1154 				err = -ENOBUFS;
1155 				goto error;
1156 			}
1157 
1158 			/*
1159 			 *	Fill in the control structures
1160 			 */
1161 			skb->ip_summed = CHECKSUM_NONE;
1162 			skb->csum = 0;
1163 			skb_reserve(skb, hh_len);
1164 
1165 			/*
1166 			 *	Find where to start putting bytes.
1167 			 */
1168 			skb_put(skb, fragheaderlen + fraggap);
1169 			skb_reset_network_header(skb);
1170 			skb->transport_header = (skb->network_header +
1171 						 fragheaderlen);
1172 			if (fraggap) {
1173 				skb->csum = skb_copy_and_csum_bits(skb_prev,
1174 								   maxfraglen,
1175 						    skb_transport_header(skb),
1176 								   fraggap, 0);
1177 				skb_prev->csum = csum_sub(skb_prev->csum,
1178 							  skb->csum);
1179 				pskb_trim_unique(skb_prev, maxfraglen);
1180 			}
1181 
1182 			/*
1183 			 * Put the packet on the pending queue.
1184 			 */
1185 			__skb_queue_tail(&sk->sk_write_queue, skb);
1186 			continue;
1187 		}
1188 
1189 		i = skb_shinfo(skb)->nr_frags;
1190 		if (len > size)
1191 			len = size;
1192 		if (skb_can_coalesce(skb, i, page, offset)) {
1193 			skb_shinfo(skb)->frags[i-1].size += len;
1194 		} else if (i < MAX_SKB_FRAGS) {
1195 			get_page(page);
1196 			skb_fill_page_desc(skb, i, page, offset, len);
1197 		} else {
1198 			err = -EMSGSIZE;
1199 			goto error;
1200 		}
1201 
1202 		if (skb->ip_summed == CHECKSUM_NONE) {
1203 			__wsum csum;
1204 			csum = csum_page(page, offset, len);
1205 			skb->csum = csum_block_add(skb->csum, csum, skb->len);
1206 		}
1207 
1208 		skb->len += len;
1209 		skb->data_len += len;
1210 		skb->truesize += len;
1211 		atomic_add(len, &sk->sk_wmem_alloc);
1212 		offset += len;
1213 		size -= len;
1214 	}
1215 	return 0;
1216 
1217 error:
1218 	inet->cork.length -= size;
1219 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1220 	return err;
1221 }
1222 
1223 static void ip_cork_release(struct inet_sock *inet)
1224 {
1225 	inet->cork.flags &= ~IPCORK_OPT;
1226 	kfree(inet->cork.opt);
1227 	inet->cork.opt = NULL;
1228 	dst_release(inet->cork.dst);
1229 	inet->cork.dst = NULL;
1230 }
1231 
1232 /*
1233  *	Combined all pending IP fragments on the socket as one IP datagram
1234  *	and push them out.
1235  */
1236 int ip_push_pending_frames(struct sock *sk)
1237 {
1238 	struct sk_buff *skb, *tmp_skb;
1239 	struct sk_buff **tail_skb;
1240 	struct inet_sock *inet = inet_sk(sk);
1241 	struct net *net = sock_net(sk);
1242 	struct ip_options *opt = NULL;
1243 	struct rtable *rt = (struct rtable *)inet->cork.dst;
1244 	struct iphdr *iph;
1245 	__be16 df = 0;
1246 	__u8 ttl;
1247 	int err = 0;
1248 
1249 	if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1250 		goto out;
1251 	tail_skb = &(skb_shinfo(skb)->frag_list);
1252 
1253 	/* move skb->data to ip header from ext header */
1254 	if (skb->data < skb_network_header(skb))
1255 		__skb_pull(skb, skb_network_offset(skb));
1256 	while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1257 		__skb_pull(tmp_skb, skb_network_header_len(skb));
1258 		*tail_skb = tmp_skb;
1259 		tail_skb = &(tmp_skb->next);
1260 		skb->len += tmp_skb->len;
1261 		skb->data_len += tmp_skb->len;
1262 		skb->truesize += tmp_skb->truesize;
1263 		tmp_skb->destructor = NULL;
1264 		tmp_skb->sk = NULL;
1265 	}
1266 
1267 	/* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1268 	 * to fragment the frame generated here. No matter, what transforms
1269 	 * how transforms change size of the packet, it will come out.
1270 	 */
1271 	if (inet->pmtudisc < IP_PMTUDISC_DO)
1272 		skb->local_df = 1;
1273 
1274 	/* DF bit is set when we want to see DF on outgoing frames.
1275 	 * If local_df is set too, we still allow to fragment this frame
1276 	 * locally. */
1277 	if (inet->pmtudisc >= IP_PMTUDISC_DO ||
1278 	    (skb->len <= dst_mtu(&rt->dst) &&
1279 	     ip_dont_fragment(sk, &rt->dst)))
1280 		df = htons(IP_DF);
1281 
1282 	if (inet->cork.flags & IPCORK_OPT)
1283 		opt = inet->cork.opt;
1284 
1285 	if (rt->rt_type == RTN_MULTICAST)
1286 		ttl = inet->mc_ttl;
1287 	else
1288 		ttl = ip_select_ttl(inet, &rt->dst);
1289 
1290 	iph = (struct iphdr *)skb->data;
1291 	iph->version = 4;
1292 	iph->ihl = 5;
1293 	if (opt) {
1294 		iph->ihl += opt->optlen>>2;
1295 		ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1296 	}
1297 	iph->tos = inet->tos;
1298 	iph->frag_off = df;
1299 	ip_select_ident(iph, &rt->dst, sk);
1300 	iph->ttl = ttl;
1301 	iph->protocol = sk->sk_protocol;
1302 	iph->saddr = rt->rt_src;
1303 	iph->daddr = rt->rt_dst;
1304 
1305 	skb->priority = sk->sk_priority;
1306 	skb->mark = sk->sk_mark;
1307 	/*
1308 	 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1309 	 * on dst refcount
1310 	 */
1311 	inet->cork.dst = NULL;
1312 	skb_dst_set(skb, &rt->dst);
1313 
1314 	if (iph->protocol == IPPROTO_ICMP)
1315 		icmp_out_count(net, ((struct icmphdr *)
1316 			skb_transport_header(skb))->type);
1317 
1318 	/* Netfilter gets whole the not fragmented skb. */
1319 	err = ip_local_out(skb);
1320 	if (err) {
1321 		if (err > 0)
1322 			err = net_xmit_errno(err);
1323 		if (err)
1324 			goto error;
1325 	}
1326 
1327 out:
1328 	ip_cork_release(inet);
1329 	return err;
1330 
1331 error:
1332 	IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1333 	goto out;
1334 }
1335 
1336 /*
1337  *	Throw away all pending data on the socket.
1338  */
1339 void ip_flush_pending_frames(struct sock *sk)
1340 {
1341 	struct sk_buff *skb;
1342 
1343 	while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1344 		kfree_skb(skb);
1345 
1346 	ip_cork_release(inet_sk(sk));
1347 }
1348 
1349 
1350 /*
1351  *	Fetch data from kernel space and fill in checksum if needed.
1352  */
1353 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1354 			      int len, int odd, struct sk_buff *skb)
1355 {
1356 	__wsum csum;
1357 
1358 	csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1359 	skb->csum = csum_block_add(skb->csum, csum, odd);
1360 	return 0;
1361 }
1362 
1363 /*
1364  *	Generic function to send a packet as reply to another packet.
1365  *	Used to send TCP resets so far. ICMP should use this function too.
1366  *
1367  *	Should run single threaded per socket because it uses the sock
1368  *     	structure to pass arguments.
1369  */
1370 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1371 		   unsigned int len)
1372 {
1373 	struct inet_sock *inet = inet_sk(sk);
1374 	struct {
1375 		struct ip_options	opt;
1376 		char			data[40];
1377 	} replyopts;
1378 	struct ipcm_cookie ipc;
1379 	__be32 daddr;
1380 	struct rtable *rt = skb_rtable(skb);
1381 
1382 	if (ip_options_echo(&replyopts.opt, skb))
1383 		return;
1384 
1385 	daddr = ipc.addr = rt->rt_src;
1386 	ipc.opt = NULL;
1387 	ipc.shtx.flags = 0;
1388 
1389 	if (replyopts.opt.optlen) {
1390 		ipc.opt = &replyopts.opt;
1391 
1392 		if (ipc.opt->srr)
1393 			daddr = replyopts.opt.faddr;
1394 	}
1395 
1396 	{
1397 		struct flowi fl = { .oif = arg->bound_dev_if,
1398 				    .nl_u = { .ip4_u =
1399 					      { .daddr = daddr,
1400 						.saddr = rt->rt_spec_dst,
1401 						.tos = RT_TOS(ip_hdr(skb)->tos) } },
1402 				    /* Not quite clean, but right. */
1403 				    .uli_u = { .ports =
1404 					       { .sport = tcp_hdr(skb)->dest,
1405 						 .dport = tcp_hdr(skb)->source } },
1406 				    .proto = sk->sk_protocol,
1407 				    .flags = ip_reply_arg_flowi_flags(arg) };
1408 		security_skb_classify_flow(skb, &fl);
1409 		if (ip_route_output_key(sock_net(sk), &rt, &fl))
1410 			return;
1411 	}
1412 
1413 	/* And let IP do all the hard work.
1414 
1415 	   This chunk is not reenterable, hence spinlock.
1416 	   Note that it uses the fact, that this function is called
1417 	   with locally disabled BH and that sk cannot be already spinlocked.
1418 	 */
1419 	bh_lock_sock(sk);
1420 	inet->tos = ip_hdr(skb)->tos;
1421 	sk->sk_priority = skb->priority;
1422 	sk->sk_protocol = ip_hdr(skb)->protocol;
1423 	sk->sk_bound_dev_if = arg->bound_dev_if;
1424 	ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1425 		       &ipc, &rt, MSG_DONTWAIT);
1426 	if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1427 		if (arg->csumoffset >= 0)
1428 			*((__sum16 *)skb_transport_header(skb) +
1429 			  arg->csumoffset) = csum_fold(csum_add(skb->csum,
1430 								arg->csum));
1431 		skb->ip_summed = CHECKSUM_NONE;
1432 		ip_push_pending_frames(sk);
1433 	}
1434 
1435 	bh_unlock_sock(sk);
1436 
1437 	ip_rt_put(rt);
1438 }
1439 
1440 void __init ip_init(void)
1441 {
1442 	ip_rt_init();
1443 	inet_initpeers();
1444 
1445 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1446 	igmp_mc_proc_init();
1447 #endif
1448 }
1449