xref: /linux/net/ipv4/ip_output.c (revision f3d9478b2ce468c3115b02ecae7e975990697f15)
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  * Version:	$Id: ip_output.c,v 1.100 2002/02/01 22:01:03 davem Exp $
9  *
10  * Authors:	Ross Biro
11  *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *		Donald Becker, <becker@super.org>
13  *		Alan Cox, <Alan.Cox@linux.org>
14  *		Richard Underwood
15  *		Stefan Becker, <stefanb@yello.ping.de>
16  *		Jorge Cwik, <jorge@laser.satlink.net>
17  *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18  *		Hirokazu Takahashi, <taka@valinux.co.jp>
19  *
20  *	See ip_input.c for original log
21  *
22  *	Fixes:
23  *		Alan Cox	:	Missing nonblock feature in ip_build_xmit.
24  *		Mike Kilburn	:	htons() missing in ip_build_xmit.
25  *		Bradford Johnson:	Fix faulty handling of some frames when
26  *					no route is found.
27  *		Alexander Demenshin:	Missing sk/skb free in ip_queue_xmit
28  *					(in case if packet not accepted by
29  *					output firewall rules)
30  *		Mike McLagan	:	Routing by source
31  *		Alexey Kuznetsov:	use new route cache
32  *		Andi Kleen:		Fix broken PMTU recovery and remove
33  *					some redundant tests.
34  *	Vitaly E. Lavrov	:	Transparent proxy revived after year coma.
35  *		Andi Kleen	: 	Replace ip_reply with ip_send_reply.
36  *		Andi Kleen	:	Split fast and slow ip_build_xmit path
37  *					for decreased register pressure on x86
38  *					and more readibility.
39  *		Marc Boucher	:	When call_out_firewall returns FW_QUEUE,
40  *					silently drop skb instead of failing with -EPERM.
41  *		Detlev Wengorz	:	Copy protocol for fragments.
42  *		Hirokazu Takahashi:	HW checksumming for outgoing UDP
43  *					datagrams.
44  *		Hirokazu Takahashi:	sendfile() on UDP works now.
45  */
46 
47 #include <asm/uaccess.h>
48 #include <asm/system.h>
49 #include <linux/module.h>
50 #include <linux/types.h>
51 #include <linux/kernel.h>
52 #include <linux/sched.h>
53 #include <linux/mm.h>
54 #include <linux/string.h>
55 #include <linux/errno.h>
56 #include <linux/config.h>
57 
58 #include <linux/socket.h>
59 #include <linux/sockios.h>
60 #include <linux/in.h>
61 #include <linux/inet.h>
62 #include <linux/netdevice.h>
63 #include <linux/etherdevice.h>
64 #include <linux/proc_fs.h>
65 #include <linux/stat.h>
66 #include <linux/init.h>
67 
68 #include <net/snmp.h>
69 #include <net/ip.h>
70 #include <net/protocol.h>
71 #include <net/route.h>
72 #include <net/xfrm.h>
73 #include <linux/skbuff.h>
74 #include <net/sock.h>
75 #include <net/arp.h>
76 #include <net/icmp.h>
77 #include <net/checksum.h>
78 #include <net/inetpeer.h>
79 #include <net/checksum.h>
80 #include <linux/igmp.h>
81 #include <linux/netfilter_ipv4.h>
82 #include <linux/netfilter_bridge.h>
83 #include <linux/mroute.h>
84 #include <linux/netlink.h>
85 #include <linux/tcp.h>
86 
87 int sysctl_ip_default_ttl = IPDEFTTL;
88 
89 /* Generate a checksum for an outgoing IP datagram. */
90 __inline__ void ip_send_check(struct iphdr *iph)
91 {
92 	iph->check = 0;
93 	iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
94 }
95 
96 /* dev_loopback_xmit for use with netfilter. */
97 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
98 {
99 	newskb->mac.raw = newskb->data;
100 	__skb_pull(newskb, newskb->nh.raw - newskb->data);
101 	newskb->pkt_type = PACKET_LOOPBACK;
102 	newskb->ip_summed = CHECKSUM_UNNECESSARY;
103 	BUG_TRAP(newskb->dst);
104 	netif_rx(newskb);
105 	return 0;
106 }
107 
108 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
109 {
110 	int ttl = inet->uc_ttl;
111 
112 	if (ttl < 0)
113 		ttl = dst_metric(dst, RTAX_HOPLIMIT);
114 	return ttl;
115 }
116 
117 /*
118  *		Add an ip header to a skbuff and send it out.
119  *
120  */
121 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
122 			  u32 saddr, u32 daddr, struct ip_options *opt)
123 {
124 	struct inet_sock *inet = inet_sk(sk);
125 	struct rtable *rt = (struct rtable *)skb->dst;
126 	struct iphdr *iph;
127 
128 	/* Build the IP header. */
129 	if (opt)
130 		iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr) + opt->optlen);
131 	else
132 		iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr));
133 
134 	iph->version  = 4;
135 	iph->ihl      = 5;
136 	iph->tos      = inet->tos;
137 	if (ip_dont_fragment(sk, &rt->u.dst))
138 		iph->frag_off = htons(IP_DF);
139 	else
140 		iph->frag_off = 0;
141 	iph->ttl      = ip_select_ttl(inet, &rt->u.dst);
142 	iph->daddr    = rt->rt_dst;
143 	iph->saddr    = rt->rt_src;
144 	iph->protocol = sk->sk_protocol;
145 	iph->tot_len  = htons(skb->len);
146 	ip_select_ident(iph, &rt->u.dst, sk);
147 	skb->nh.iph   = iph;
148 
149 	if (opt && opt->optlen) {
150 		iph->ihl += opt->optlen>>2;
151 		ip_options_build(skb, opt, daddr, rt, 0);
152 	}
153 	ip_send_check(iph);
154 
155 	skb->priority = sk->sk_priority;
156 
157 	/* Send it out. */
158 	return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
159 		       dst_output);
160 }
161 
162 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
163 
164 static inline int ip_finish_output2(struct sk_buff *skb)
165 {
166 	struct dst_entry *dst = skb->dst;
167 	struct hh_cache *hh = dst->hh;
168 	struct net_device *dev = dst->dev;
169 	int hh_len = LL_RESERVED_SPACE(dev);
170 
171 	/* Be paranoid, rather than too clever. */
172 	if (unlikely(skb_headroom(skb) < hh_len && dev->hard_header)) {
173 		struct sk_buff *skb2;
174 
175 		skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
176 		if (skb2 == NULL) {
177 			kfree_skb(skb);
178 			return -ENOMEM;
179 		}
180 		if (skb->sk)
181 			skb_set_owner_w(skb2, skb->sk);
182 		kfree_skb(skb);
183 		skb = skb2;
184 	}
185 
186 	if (hh) {
187 		int hh_alen;
188 
189 		read_lock_bh(&hh->hh_lock);
190 		hh_alen = HH_DATA_ALIGN(hh->hh_len);
191   		memcpy(skb->data - hh_alen, hh->hh_data, hh_alen);
192 		read_unlock_bh(&hh->hh_lock);
193 	        skb_push(skb, hh->hh_len);
194 		return hh->hh_output(skb);
195 	} else if (dst->neighbour)
196 		return dst->neighbour->output(skb);
197 
198 	if (net_ratelimit())
199 		printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
200 	kfree_skb(skb);
201 	return -EINVAL;
202 }
203 
204 static inline int ip_finish_output(struct sk_buff *skb)
205 {
206 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
207 	/* Policy lookup after SNAT yielded a new policy */
208 	if (skb->dst->xfrm != NULL) {
209 		IPCB(skb)->flags |= IPSKB_REROUTED;
210 		return dst_output(skb);
211 	}
212 #endif
213 	if (skb->len > dst_mtu(skb->dst) &&
214 	    !(skb_shinfo(skb)->ufo_size || skb_shinfo(skb)->tso_size))
215 		return ip_fragment(skb, ip_finish_output2);
216 	else
217 		return ip_finish_output2(skb);
218 }
219 
220 int ip_mc_output(struct sk_buff *skb)
221 {
222 	struct sock *sk = skb->sk;
223 	struct rtable *rt = (struct rtable*)skb->dst;
224 	struct net_device *dev = rt->u.dst.dev;
225 
226 	/*
227 	 *	If the indicated interface is up and running, send the packet.
228 	 */
229 	IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
230 
231 	skb->dev = dev;
232 	skb->protocol = htons(ETH_P_IP);
233 
234 	/*
235 	 *	Multicasts are looped back for other local users
236 	 */
237 
238 	if (rt->rt_flags&RTCF_MULTICAST) {
239 		if ((!sk || inet_sk(sk)->mc_loop)
240 #ifdef CONFIG_IP_MROUTE
241 		/* Small optimization: do not loopback not local frames,
242 		   which returned after forwarding; they will be  dropped
243 		   by ip_mr_input in any case.
244 		   Note, that local frames are looped back to be delivered
245 		   to local recipients.
246 
247 		   This check is duplicated in ip_mr_input at the moment.
248 		 */
249 		    && ((rt->rt_flags&RTCF_LOCAL) || !(IPCB(skb)->flags&IPSKB_FORWARDED))
250 #endif
251 		) {
252 			struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
253 			if (newskb)
254 				NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
255 					newskb->dev,
256 					ip_dev_loopback_xmit);
257 		}
258 
259 		/* Multicasts with ttl 0 must not go beyond the host */
260 
261 		if (skb->nh.iph->ttl == 0) {
262 			kfree_skb(skb);
263 			return 0;
264 		}
265 	}
266 
267 	if (rt->rt_flags&RTCF_BROADCAST) {
268 		struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
269 		if (newskb)
270 			NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
271 				newskb->dev, ip_dev_loopback_xmit);
272 	}
273 
274 	return NF_HOOK_COND(PF_INET, NF_IP_POST_ROUTING, skb, NULL, skb->dev,
275 			    ip_finish_output,
276 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
277 }
278 
279 int ip_output(struct sk_buff *skb)
280 {
281 	struct net_device *dev = skb->dst->dev;
282 
283 	IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
284 
285 	skb->dev = dev;
286 	skb->protocol = htons(ETH_P_IP);
287 
288 	return NF_HOOK_COND(PF_INET, NF_IP_POST_ROUTING, skb, NULL, dev,
289 		            ip_finish_output,
290 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
291 }
292 
293 int ip_queue_xmit(struct sk_buff *skb, int ipfragok)
294 {
295 	struct sock *sk = skb->sk;
296 	struct inet_sock *inet = inet_sk(sk);
297 	struct ip_options *opt = inet->opt;
298 	struct rtable *rt;
299 	struct iphdr *iph;
300 
301 	/* Skip all of this if the packet is already routed,
302 	 * f.e. by something like SCTP.
303 	 */
304 	rt = (struct rtable *) skb->dst;
305 	if (rt != NULL)
306 		goto packet_routed;
307 
308 	/* Make sure we can route this packet. */
309 	rt = (struct rtable *)__sk_dst_check(sk, 0);
310 	if (rt == NULL) {
311 		u32 daddr;
312 
313 		/* Use correct destination address if we have options. */
314 		daddr = inet->daddr;
315 		if(opt && opt->srr)
316 			daddr = opt->faddr;
317 
318 		{
319 			struct flowi fl = { .oif = sk->sk_bound_dev_if,
320 					    .nl_u = { .ip4_u =
321 						      { .daddr = daddr,
322 							.saddr = inet->saddr,
323 							.tos = RT_CONN_FLAGS(sk) } },
324 					    .proto = sk->sk_protocol,
325 					    .uli_u = { .ports =
326 						       { .sport = inet->sport,
327 							 .dport = inet->dport } } };
328 
329 			/* If this fails, retransmit mechanism of transport layer will
330 			 * keep trying until route appears or the connection times
331 			 * itself out.
332 			 */
333 			if (ip_route_output_flow(&rt, &fl, sk, 0))
334 				goto no_route;
335 		}
336 		sk_setup_caps(sk, &rt->u.dst);
337 	}
338 	skb->dst = dst_clone(&rt->u.dst);
339 
340 packet_routed:
341 	if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
342 		goto no_route;
343 
344 	/* OK, we know where to send it, allocate and build IP header. */
345 	iph = (struct iphdr *) skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
346 	*((__u16 *)iph)	= htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
347 	iph->tot_len = htons(skb->len);
348 	if (ip_dont_fragment(sk, &rt->u.dst) && !ipfragok)
349 		iph->frag_off = htons(IP_DF);
350 	else
351 		iph->frag_off = 0;
352 	iph->ttl      = ip_select_ttl(inet, &rt->u.dst);
353 	iph->protocol = sk->sk_protocol;
354 	iph->saddr    = rt->rt_src;
355 	iph->daddr    = rt->rt_dst;
356 	skb->nh.iph   = iph;
357 	/* Transport layer set skb->h.foo itself. */
358 
359 	if (opt && opt->optlen) {
360 		iph->ihl += opt->optlen >> 2;
361 		ip_options_build(skb, opt, inet->daddr, rt, 0);
362 	}
363 
364 	ip_select_ident_more(iph, &rt->u.dst, sk,
365 			     (skb_shinfo(skb)->tso_segs ?: 1) - 1);
366 
367 	/* Add an IP checksum. */
368 	ip_send_check(iph);
369 
370 	skb->priority = sk->sk_priority;
371 
372 	return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
373 		       dst_output);
374 
375 no_route:
376 	IP_INC_STATS(IPSTATS_MIB_OUTNOROUTES);
377 	kfree_skb(skb);
378 	return -EHOSTUNREACH;
379 }
380 
381 
382 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
383 {
384 	to->pkt_type = from->pkt_type;
385 	to->priority = from->priority;
386 	to->protocol = from->protocol;
387 	dst_release(to->dst);
388 	to->dst = dst_clone(from->dst);
389 	to->dev = from->dev;
390 
391 	/* Copy the flags to each fragment. */
392 	IPCB(to)->flags = IPCB(from)->flags;
393 
394 #ifdef CONFIG_NET_SCHED
395 	to->tc_index = from->tc_index;
396 #endif
397 #ifdef CONFIG_NETFILTER
398 	to->nfmark = from->nfmark;
399 	/* Connection association is same as pre-frag packet */
400 	nf_conntrack_put(to->nfct);
401 	to->nfct = from->nfct;
402 	nf_conntrack_get(to->nfct);
403 	to->nfctinfo = from->nfctinfo;
404 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
405 	to->ipvs_property = from->ipvs_property;
406 #endif
407 #ifdef CONFIG_BRIDGE_NETFILTER
408 	nf_bridge_put(to->nf_bridge);
409 	to->nf_bridge = from->nf_bridge;
410 	nf_bridge_get(to->nf_bridge);
411 #endif
412 #endif
413 	skb_copy_secmark(to, from);
414 }
415 
416 /*
417  *	This IP datagram is too large to be sent in one piece.  Break it up into
418  *	smaller pieces (each of size equal to IP header plus
419  *	a block of the data of the original IP data part) that will yet fit in a
420  *	single device frame, and queue such a frame for sending.
421  */
422 
423 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*))
424 {
425 	struct iphdr *iph;
426 	int raw = 0;
427 	int ptr;
428 	struct net_device *dev;
429 	struct sk_buff *skb2;
430 	unsigned int mtu, hlen, left, len, ll_rs;
431 	int offset;
432 	__be16 not_last_frag;
433 	struct rtable *rt = (struct rtable*)skb->dst;
434 	int err = 0;
435 
436 	dev = rt->u.dst.dev;
437 
438 	/*
439 	 *	Point into the IP datagram header.
440 	 */
441 
442 	iph = skb->nh.iph;
443 
444 	if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
445 		icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
446 			  htonl(dst_mtu(&rt->u.dst)));
447 		kfree_skb(skb);
448 		return -EMSGSIZE;
449 	}
450 
451 	/*
452 	 *	Setup starting values.
453 	 */
454 
455 	hlen = iph->ihl * 4;
456 	mtu = dst_mtu(&rt->u.dst) - hlen;	/* Size of data space */
457 	IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
458 
459 	/* When frag_list is given, use it. First, check its validity:
460 	 * some transformers could create wrong frag_list or break existing
461 	 * one, it is not prohibited. In this case fall back to copying.
462 	 *
463 	 * LATER: this step can be merged to real generation of fragments,
464 	 * we can switch to copy when see the first bad fragment.
465 	 */
466 	if (skb_shinfo(skb)->frag_list) {
467 		struct sk_buff *frag;
468 		int first_len = skb_pagelen(skb);
469 
470 		if (first_len - hlen > mtu ||
471 		    ((first_len - hlen) & 7) ||
472 		    (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
473 		    skb_cloned(skb))
474 			goto slow_path;
475 
476 		for (frag = skb_shinfo(skb)->frag_list; frag; frag = frag->next) {
477 			/* Correct geometry. */
478 			if (frag->len > mtu ||
479 			    ((frag->len & 7) && frag->next) ||
480 			    skb_headroom(frag) < hlen)
481 			    goto slow_path;
482 
483 			/* Partially cloned skb? */
484 			if (skb_shared(frag))
485 				goto slow_path;
486 
487 			BUG_ON(frag->sk);
488 			if (skb->sk) {
489 				sock_hold(skb->sk);
490 				frag->sk = skb->sk;
491 				frag->destructor = sock_wfree;
492 				skb->truesize -= frag->truesize;
493 			}
494 		}
495 
496 		/* Everything is OK. Generate! */
497 
498 		err = 0;
499 		offset = 0;
500 		frag = skb_shinfo(skb)->frag_list;
501 		skb_shinfo(skb)->frag_list = NULL;
502 		skb->data_len = first_len - skb_headlen(skb);
503 		skb->len = first_len;
504 		iph->tot_len = htons(first_len);
505 		iph->frag_off = htons(IP_MF);
506 		ip_send_check(iph);
507 
508 		for (;;) {
509 			/* Prepare header of the next frame,
510 			 * before previous one went down. */
511 			if (frag) {
512 				frag->ip_summed = CHECKSUM_NONE;
513 				frag->h.raw = frag->data;
514 				frag->nh.raw = __skb_push(frag, hlen);
515 				memcpy(frag->nh.raw, iph, hlen);
516 				iph = frag->nh.iph;
517 				iph->tot_len = htons(frag->len);
518 				ip_copy_metadata(frag, skb);
519 				if (offset == 0)
520 					ip_options_fragment(frag);
521 				offset += skb->len - hlen;
522 				iph->frag_off = htons(offset>>3);
523 				if (frag->next != NULL)
524 					iph->frag_off |= htons(IP_MF);
525 				/* Ready, complete checksum */
526 				ip_send_check(iph);
527 			}
528 
529 			err = output(skb);
530 
531 			if (err || !frag)
532 				break;
533 
534 			skb = frag;
535 			frag = skb->next;
536 			skb->next = NULL;
537 		}
538 
539 		if (err == 0) {
540 			IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
541 			return 0;
542 		}
543 
544 		while (frag) {
545 			skb = frag->next;
546 			kfree_skb(frag);
547 			frag = skb;
548 		}
549 		IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
550 		return err;
551 	}
552 
553 slow_path:
554 	left = skb->len - hlen;		/* Space per frame */
555 	ptr = raw + hlen;		/* Where to start from */
556 
557 #ifdef CONFIG_BRIDGE_NETFILTER
558 	/* for bridged IP traffic encapsulated inside f.e. a vlan header,
559 	 * we need to make room for the encapsulating header */
560 	ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, nf_bridge_pad(skb));
561 	mtu -= nf_bridge_pad(skb);
562 #else
563 	ll_rs = LL_RESERVED_SPACE(rt->u.dst.dev);
564 #endif
565 	/*
566 	 *	Fragment the datagram.
567 	 */
568 
569 	offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
570 	not_last_frag = iph->frag_off & htons(IP_MF);
571 
572 	/*
573 	 *	Keep copying data until we run out.
574 	 */
575 
576 	while(left > 0)	{
577 		len = left;
578 		/* IF: it doesn't fit, use 'mtu' - the data space left */
579 		if (len > mtu)
580 			len = mtu;
581 		/* IF: we are not sending upto and including the packet end
582 		   then align the next start on an eight byte boundary */
583 		if (len < left)	{
584 			len &= ~7;
585 		}
586 		/*
587 		 *	Allocate buffer.
588 		 */
589 
590 		if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
591 			NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
592 			err = -ENOMEM;
593 			goto fail;
594 		}
595 
596 		/*
597 		 *	Set up data on packet
598 		 */
599 
600 		ip_copy_metadata(skb2, skb);
601 		skb_reserve(skb2, ll_rs);
602 		skb_put(skb2, len + hlen);
603 		skb2->nh.raw = skb2->data;
604 		skb2->h.raw = skb2->data + hlen;
605 
606 		/*
607 		 *	Charge the memory for the fragment to any owner
608 		 *	it might possess
609 		 */
610 
611 		if (skb->sk)
612 			skb_set_owner_w(skb2, skb->sk);
613 
614 		/*
615 		 *	Copy the packet header into the new buffer.
616 		 */
617 
618 		memcpy(skb2->nh.raw, skb->data, hlen);
619 
620 		/*
621 		 *	Copy a block of the IP datagram.
622 		 */
623 		if (skb_copy_bits(skb, ptr, skb2->h.raw, len))
624 			BUG();
625 		left -= len;
626 
627 		/*
628 		 *	Fill in the new header fields.
629 		 */
630 		iph = skb2->nh.iph;
631 		iph->frag_off = htons((offset >> 3));
632 
633 		/* ANK: dirty, but effective trick. Upgrade options only if
634 		 * the segment to be fragmented was THE FIRST (otherwise,
635 		 * options are already fixed) and make it ONCE
636 		 * on the initial skb, so that all the following fragments
637 		 * will inherit fixed options.
638 		 */
639 		if (offset == 0)
640 			ip_options_fragment(skb);
641 
642 		/*
643 		 *	Added AC : If we are fragmenting a fragment that's not the
644 		 *		   last fragment then keep MF on each bit
645 		 */
646 		if (left > 0 || not_last_frag)
647 			iph->frag_off |= htons(IP_MF);
648 		ptr += len;
649 		offset += len;
650 
651 		/*
652 		 *	Put this fragment into the sending queue.
653 		 */
654 
655 		IP_INC_STATS(IPSTATS_MIB_FRAGCREATES);
656 
657 		iph->tot_len = htons(len + hlen);
658 
659 		ip_send_check(iph);
660 
661 		err = output(skb2);
662 		if (err)
663 			goto fail;
664 	}
665 	kfree_skb(skb);
666 	IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
667 	return err;
668 
669 fail:
670 	kfree_skb(skb);
671 	IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
672 	return err;
673 }
674 
675 EXPORT_SYMBOL(ip_fragment);
676 
677 int
678 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
679 {
680 	struct iovec *iov = from;
681 
682 	if (skb->ip_summed == CHECKSUM_HW) {
683 		if (memcpy_fromiovecend(to, iov, offset, len) < 0)
684 			return -EFAULT;
685 	} else {
686 		unsigned int csum = 0;
687 		if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
688 			return -EFAULT;
689 		skb->csum = csum_block_add(skb->csum, csum, odd);
690 	}
691 	return 0;
692 }
693 
694 static inline unsigned int
695 csum_page(struct page *page, int offset, int copy)
696 {
697 	char *kaddr;
698 	unsigned int csum;
699 	kaddr = kmap(page);
700 	csum = csum_partial(kaddr + offset, copy, 0);
701 	kunmap(page);
702 	return csum;
703 }
704 
705 static inline int ip_ufo_append_data(struct sock *sk,
706 			int getfrag(void *from, char *to, int offset, int len,
707 			       int odd, struct sk_buff *skb),
708 			void *from, int length, int hh_len, int fragheaderlen,
709 			int transhdrlen, int mtu,unsigned int flags)
710 {
711 	struct sk_buff *skb;
712 	int err;
713 
714 	/* There is support for UDP fragmentation offload by network
715 	 * device, so create one single skb packet containing complete
716 	 * udp datagram
717 	 */
718 	if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL) {
719 		skb = sock_alloc_send_skb(sk,
720 			hh_len + fragheaderlen + transhdrlen + 20,
721 			(flags & MSG_DONTWAIT), &err);
722 
723 		if (skb == NULL)
724 			return err;
725 
726 		/* reserve space for Hardware header */
727 		skb_reserve(skb, hh_len);
728 
729 		/* create space for UDP/IP header */
730 		skb_put(skb,fragheaderlen + transhdrlen);
731 
732 		/* initialize network header pointer */
733 		skb->nh.raw = skb->data;
734 
735 		/* initialize protocol header pointer */
736 		skb->h.raw = skb->data + fragheaderlen;
737 
738 		skb->ip_summed = CHECKSUM_HW;
739 		skb->csum = 0;
740 		sk->sk_sndmsg_off = 0;
741 	}
742 
743 	err = skb_append_datato_frags(sk,skb, getfrag, from,
744 			       (length - transhdrlen));
745 	if (!err) {
746 		/* specify the length of each IP datagram fragment*/
747 		skb_shinfo(skb)->ufo_size = (mtu - fragheaderlen);
748 		__skb_queue_tail(&sk->sk_write_queue, skb);
749 
750 		return 0;
751 	}
752 	/* There is not enough support do UFO ,
753 	 * so follow normal path
754 	 */
755 	kfree_skb(skb);
756 	return err;
757 }
758 
759 /*
760  *	ip_append_data() and ip_append_page() can make one large IP datagram
761  *	from many pieces of data. Each pieces will be holded on the socket
762  *	until ip_push_pending_frames() is called. Each piece can be a page
763  *	or non-page data.
764  *
765  *	Not only UDP, other transport protocols - e.g. raw sockets - can use
766  *	this interface potentially.
767  *
768  *	LATER: length must be adjusted by pad at tail, when it is required.
769  */
770 int ip_append_data(struct sock *sk,
771 		   int getfrag(void *from, char *to, int offset, int len,
772 			       int odd, struct sk_buff *skb),
773 		   void *from, int length, int transhdrlen,
774 		   struct ipcm_cookie *ipc, struct rtable *rt,
775 		   unsigned int flags)
776 {
777 	struct inet_sock *inet = inet_sk(sk);
778 	struct sk_buff *skb;
779 
780 	struct ip_options *opt = NULL;
781 	int hh_len;
782 	int exthdrlen;
783 	int mtu;
784 	int copy;
785 	int err;
786 	int offset = 0;
787 	unsigned int maxfraglen, fragheaderlen;
788 	int csummode = CHECKSUM_NONE;
789 
790 	if (flags&MSG_PROBE)
791 		return 0;
792 
793 	if (skb_queue_empty(&sk->sk_write_queue)) {
794 		/*
795 		 * setup for corking.
796 		 */
797 		opt = ipc->opt;
798 		if (opt) {
799 			if (inet->cork.opt == NULL) {
800 				inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
801 				if (unlikely(inet->cork.opt == NULL))
802 					return -ENOBUFS;
803 			}
804 			memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
805 			inet->cork.flags |= IPCORK_OPT;
806 			inet->cork.addr = ipc->addr;
807 		}
808 		dst_hold(&rt->u.dst);
809 		inet->cork.fragsize = mtu = dst_mtu(rt->u.dst.path);
810 		inet->cork.rt = rt;
811 		inet->cork.length = 0;
812 		sk->sk_sndmsg_page = NULL;
813 		sk->sk_sndmsg_off = 0;
814 		if ((exthdrlen = rt->u.dst.header_len) != 0) {
815 			length += exthdrlen;
816 			transhdrlen += exthdrlen;
817 		}
818 	} else {
819 		rt = inet->cork.rt;
820 		if (inet->cork.flags & IPCORK_OPT)
821 			opt = inet->cork.opt;
822 
823 		transhdrlen = 0;
824 		exthdrlen = 0;
825 		mtu = inet->cork.fragsize;
826 	}
827 	hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
828 
829 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
830 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
831 
832 	if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
833 		ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu-exthdrlen);
834 		return -EMSGSIZE;
835 	}
836 
837 	/*
838 	 * transhdrlen > 0 means that this is the first fragment and we wish
839 	 * it won't be fragmented in the future.
840 	 */
841 	if (transhdrlen &&
842 	    length + fragheaderlen <= mtu &&
843 	    rt->u.dst.dev->features & NETIF_F_ALL_CSUM &&
844 	    !exthdrlen)
845 		csummode = CHECKSUM_HW;
846 
847 	inet->cork.length += length;
848 	if (((length > mtu) && (sk->sk_protocol == IPPROTO_UDP)) &&
849 			(rt->u.dst.dev->features & NETIF_F_UFO)) {
850 
851 		err = ip_ufo_append_data(sk, getfrag, from, length, hh_len,
852 					 fragheaderlen, transhdrlen, mtu,
853 					 flags);
854 		if (err)
855 			goto error;
856 		return 0;
857 	}
858 
859 	/* So, what's going on in the loop below?
860 	 *
861 	 * We use calculated fragment length to generate chained skb,
862 	 * each of segments is IP fragment ready for sending to network after
863 	 * adding appropriate IP header.
864 	 */
865 
866 	if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
867 		goto alloc_new_skb;
868 
869 	while (length > 0) {
870 		/* Check if the remaining data fits into current packet. */
871 		copy = mtu - skb->len;
872 		if (copy < length)
873 			copy = maxfraglen - skb->len;
874 		if (copy <= 0) {
875 			char *data;
876 			unsigned int datalen;
877 			unsigned int fraglen;
878 			unsigned int fraggap;
879 			unsigned int alloclen;
880 			struct sk_buff *skb_prev;
881 alloc_new_skb:
882 			skb_prev = skb;
883 			if (skb_prev)
884 				fraggap = skb_prev->len - maxfraglen;
885 			else
886 				fraggap = 0;
887 
888 			/*
889 			 * If remaining data exceeds the mtu,
890 			 * we know we need more fragment(s).
891 			 */
892 			datalen = length + fraggap;
893 			if (datalen > mtu - fragheaderlen)
894 				datalen = maxfraglen - fragheaderlen;
895 			fraglen = datalen + fragheaderlen;
896 
897 			if ((flags & MSG_MORE) &&
898 			    !(rt->u.dst.dev->features&NETIF_F_SG))
899 				alloclen = mtu;
900 			else
901 				alloclen = datalen + fragheaderlen;
902 
903 			/* The last fragment gets additional space at tail.
904 			 * Note, with MSG_MORE we overallocate on fragments,
905 			 * because we have no idea what fragment will be
906 			 * the last.
907 			 */
908 			if (datalen == length + fraggap)
909 				alloclen += rt->u.dst.trailer_len;
910 
911 			if (transhdrlen) {
912 				skb = sock_alloc_send_skb(sk,
913 						alloclen + hh_len + 15,
914 						(flags & MSG_DONTWAIT), &err);
915 			} else {
916 				skb = NULL;
917 				if (atomic_read(&sk->sk_wmem_alloc) <=
918 				    2 * sk->sk_sndbuf)
919 					skb = sock_wmalloc(sk,
920 							   alloclen + hh_len + 15, 1,
921 							   sk->sk_allocation);
922 				if (unlikely(skb == NULL))
923 					err = -ENOBUFS;
924 			}
925 			if (skb == NULL)
926 				goto error;
927 
928 			/*
929 			 *	Fill in the control structures
930 			 */
931 			skb->ip_summed = csummode;
932 			skb->csum = 0;
933 			skb_reserve(skb, hh_len);
934 
935 			/*
936 			 *	Find where to start putting bytes.
937 			 */
938 			data = skb_put(skb, fraglen);
939 			skb->nh.raw = data + exthdrlen;
940 			data += fragheaderlen;
941 			skb->h.raw = data + exthdrlen;
942 
943 			if (fraggap) {
944 				skb->csum = skb_copy_and_csum_bits(
945 					skb_prev, maxfraglen,
946 					data + transhdrlen, fraggap, 0);
947 				skb_prev->csum = csum_sub(skb_prev->csum,
948 							  skb->csum);
949 				data += fraggap;
950 				skb_trim(skb_prev, maxfraglen);
951 			}
952 
953 			copy = datalen - transhdrlen - fraggap;
954 			if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
955 				err = -EFAULT;
956 				kfree_skb(skb);
957 				goto error;
958 			}
959 
960 			offset += copy;
961 			length -= datalen - fraggap;
962 			transhdrlen = 0;
963 			exthdrlen = 0;
964 			csummode = CHECKSUM_NONE;
965 
966 			/*
967 			 * Put the packet on the pending queue.
968 			 */
969 			__skb_queue_tail(&sk->sk_write_queue, skb);
970 			continue;
971 		}
972 
973 		if (copy > length)
974 			copy = length;
975 
976 		if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
977 			unsigned int off;
978 
979 			off = skb->len;
980 			if (getfrag(from, skb_put(skb, copy),
981 					offset, copy, off, skb) < 0) {
982 				__skb_trim(skb, off);
983 				err = -EFAULT;
984 				goto error;
985 			}
986 		} else {
987 			int i = skb_shinfo(skb)->nr_frags;
988 			skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
989 			struct page *page = sk->sk_sndmsg_page;
990 			int off = sk->sk_sndmsg_off;
991 			unsigned int left;
992 
993 			if (page && (left = PAGE_SIZE - off) > 0) {
994 				if (copy >= left)
995 					copy = left;
996 				if (page != frag->page) {
997 					if (i == MAX_SKB_FRAGS) {
998 						err = -EMSGSIZE;
999 						goto error;
1000 					}
1001 					get_page(page);
1002 	 				skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
1003 					frag = &skb_shinfo(skb)->frags[i];
1004 				}
1005 			} else if (i < MAX_SKB_FRAGS) {
1006 				if (copy > PAGE_SIZE)
1007 					copy = PAGE_SIZE;
1008 				page = alloc_pages(sk->sk_allocation, 0);
1009 				if (page == NULL)  {
1010 					err = -ENOMEM;
1011 					goto error;
1012 				}
1013 				sk->sk_sndmsg_page = page;
1014 				sk->sk_sndmsg_off = 0;
1015 
1016 				skb_fill_page_desc(skb, i, page, 0, 0);
1017 				frag = &skb_shinfo(skb)->frags[i];
1018 				skb->truesize += PAGE_SIZE;
1019 				atomic_add(PAGE_SIZE, &sk->sk_wmem_alloc);
1020 			} else {
1021 				err = -EMSGSIZE;
1022 				goto error;
1023 			}
1024 			if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1025 				err = -EFAULT;
1026 				goto error;
1027 			}
1028 			sk->sk_sndmsg_off += copy;
1029 			frag->size += copy;
1030 			skb->len += copy;
1031 			skb->data_len += copy;
1032 		}
1033 		offset += copy;
1034 		length -= copy;
1035 	}
1036 
1037 	return 0;
1038 
1039 error:
1040 	inet->cork.length -= length;
1041 	IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1042 	return err;
1043 }
1044 
1045 ssize_t	ip_append_page(struct sock *sk, struct page *page,
1046 		       int offset, size_t size, int flags)
1047 {
1048 	struct inet_sock *inet = inet_sk(sk);
1049 	struct sk_buff *skb;
1050 	struct rtable *rt;
1051 	struct ip_options *opt = NULL;
1052 	int hh_len;
1053 	int mtu;
1054 	int len;
1055 	int err;
1056 	unsigned int maxfraglen, fragheaderlen, fraggap;
1057 
1058 	if (inet->hdrincl)
1059 		return -EPERM;
1060 
1061 	if (flags&MSG_PROBE)
1062 		return 0;
1063 
1064 	if (skb_queue_empty(&sk->sk_write_queue))
1065 		return -EINVAL;
1066 
1067 	rt = inet->cork.rt;
1068 	if (inet->cork.flags & IPCORK_OPT)
1069 		opt = inet->cork.opt;
1070 
1071 	if (!(rt->u.dst.dev->features&NETIF_F_SG))
1072 		return -EOPNOTSUPP;
1073 
1074 	hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1075 	mtu = inet->cork.fragsize;
1076 
1077 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1078 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1079 
1080 	if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1081 		ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu);
1082 		return -EMSGSIZE;
1083 	}
1084 
1085 	if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1086 		return -EINVAL;
1087 
1088 	inet->cork.length += size;
1089 	if ((sk->sk_protocol == IPPROTO_UDP) &&
1090 	    (rt->u.dst.dev->features & NETIF_F_UFO))
1091 		skb_shinfo(skb)->ufo_size = (mtu - fragheaderlen);
1092 
1093 
1094 	while (size > 0) {
1095 		int i;
1096 
1097 		if (skb_shinfo(skb)->ufo_size)
1098 			len = size;
1099 		else {
1100 
1101 			/* Check if the remaining data fits into current packet. */
1102 			len = mtu - skb->len;
1103 			if (len < size)
1104 				len = maxfraglen - skb->len;
1105 		}
1106 		if (len <= 0) {
1107 			struct sk_buff *skb_prev;
1108 			char *data;
1109 			struct iphdr *iph;
1110 			int alloclen;
1111 
1112 			skb_prev = skb;
1113 			fraggap = skb_prev->len - maxfraglen;
1114 
1115 			alloclen = fragheaderlen + hh_len + fraggap + 15;
1116 			skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1117 			if (unlikely(!skb)) {
1118 				err = -ENOBUFS;
1119 				goto error;
1120 			}
1121 
1122 			/*
1123 			 *	Fill in the control structures
1124 			 */
1125 			skb->ip_summed = CHECKSUM_NONE;
1126 			skb->csum = 0;
1127 			skb_reserve(skb, hh_len);
1128 
1129 			/*
1130 			 *	Find where to start putting bytes.
1131 			 */
1132 			data = skb_put(skb, fragheaderlen + fraggap);
1133 			skb->nh.iph = iph = (struct iphdr *)data;
1134 			data += fragheaderlen;
1135 			skb->h.raw = data;
1136 
1137 			if (fraggap) {
1138 				skb->csum = skb_copy_and_csum_bits(
1139 					skb_prev, maxfraglen,
1140 					data, fraggap, 0);
1141 				skb_prev->csum = csum_sub(skb_prev->csum,
1142 							  skb->csum);
1143 				skb_trim(skb_prev, maxfraglen);
1144 			}
1145 
1146 			/*
1147 			 * Put the packet on the pending queue.
1148 			 */
1149 			__skb_queue_tail(&sk->sk_write_queue, skb);
1150 			continue;
1151 		}
1152 
1153 		i = skb_shinfo(skb)->nr_frags;
1154 		if (len > size)
1155 			len = size;
1156 		if (skb_can_coalesce(skb, i, page, offset)) {
1157 			skb_shinfo(skb)->frags[i-1].size += len;
1158 		} else if (i < MAX_SKB_FRAGS) {
1159 			get_page(page);
1160 			skb_fill_page_desc(skb, i, page, offset, len);
1161 		} else {
1162 			err = -EMSGSIZE;
1163 			goto error;
1164 		}
1165 
1166 		if (skb->ip_summed == CHECKSUM_NONE) {
1167 			unsigned int csum;
1168 			csum = csum_page(page, offset, len);
1169 			skb->csum = csum_block_add(skb->csum, csum, skb->len);
1170 		}
1171 
1172 		skb->len += len;
1173 		skb->data_len += len;
1174 		offset += len;
1175 		size -= len;
1176 	}
1177 	return 0;
1178 
1179 error:
1180 	inet->cork.length -= size;
1181 	IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1182 	return err;
1183 }
1184 
1185 /*
1186  *	Combined all pending IP fragments on the socket as one IP datagram
1187  *	and push them out.
1188  */
1189 int ip_push_pending_frames(struct sock *sk)
1190 {
1191 	struct sk_buff *skb, *tmp_skb;
1192 	struct sk_buff **tail_skb;
1193 	struct inet_sock *inet = inet_sk(sk);
1194 	struct ip_options *opt = NULL;
1195 	struct rtable *rt = inet->cork.rt;
1196 	struct iphdr *iph;
1197 	__be16 df = 0;
1198 	__u8 ttl;
1199 	int err = 0;
1200 
1201 	if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1202 		goto out;
1203 	tail_skb = &(skb_shinfo(skb)->frag_list);
1204 
1205 	/* move skb->data to ip header from ext header */
1206 	if (skb->data < skb->nh.raw)
1207 		__skb_pull(skb, skb->nh.raw - skb->data);
1208 	while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1209 		__skb_pull(tmp_skb, skb->h.raw - skb->nh.raw);
1210 		*tail_skb = tmp_skb;
1211 		tail_skb = &(tmp_skb->next);
1212 		skb->len += tmp_skb->len;
1213 		skb->data_len += tmp_skb->len;
1214 		skb->truesize += tmp_skb->truesize;
1215 		__sock_put(tmp_skb->sk);
1216 		tmp_skb->destructor = NULL;
1217 		tmp_skb->sk = NULL;
1218 	}
1219 
1220 	/* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1221 	 * to fragment the frame generated here. No matter, what transforms
1222 	 * how transforms change size of the packet, it will come out.
1223 	 */
1224 	if (inet->pmtudisc != IP_PMTUDISC_DO)
1225 		skb->local_df = 1;
1226 
1227 	/* DF bit is set when we want to see DF on outgoing frames.
1228 	 * If local_df is set too, we still allow to fragment this frame
1229 	 * locally. */
1230 	if (inet->pmtudisc == IP_PMTUDISC_DO ||
1231 	    (skb->len <= dst_mtu(&rt->u.dst) &&
1232 	     ip_dont_fragment(sk, &rt->u.dst)))
1233 		df = htons(IP_DF);
1234 
1235 	if (inet->cork.flags & IPCORK_OPT)
1236 		opt = inet->cork.opt;
1237 
1238 	if (rt->rt_type == RTN_MULTICAST)
1239 		ttl = inet->mc_ttl;
1240 	else
1241 		ttl = ip_select_ttl(inet, &rt->u.dst);
1242 
1243 	iph = (struct iphdr *)skb->data;
1244 	iph->version = 4;
1245 	iph->ihl = 5;
1246 	if (opt) {
1247 		iph->ihl += opt->optlen>>2;
1248 		ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1249 	}
1250 	iph->tos = inet->tos;
1251 	iph->tot_len = htons(skb->len);
1252 	iph->frag_off = df;
1253 	ip_select_ident(iph, &rt->u.dst, sk);
1254 	iph->ttl = ttl;
1255 	iph->protocol = sk->sk_protocol;
1256 	iph->saddr = rt->rt_src;
1257 	iph->daddr = rt->rt_dst;
1258 	ip_send_check(iph);
1259 
1260 	skb->priority = sk->sk_priority;
1261 	skb->dst = dst_clone(&rt->u.dst);
1262 
1263 	/* Netfilter gets whole the not fragmented skb. */
1264 	err = NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL,
1265 		      skb->dst->dev, dst_output);
1266 	if (err) {
1267 		if (err > 0)
1268 			err = inet->recverr ? net_xmit_errno(err) : 0;
1269 		if (err)
1270 			goto error;
1271 	}
1272 
1273 out:
1274 	inet->cork.flags &= ~IPCORK_OPT;
1275 	kfree(inet->cork.opt);
1276 	inet->cork.opt = NULL;
1277 	if (inet->cork.rt) {
1278 		ip_rt_put(inet->cork.rt);
1279 		inet->cork.rt = NULL;
1280 	}
1281 	return err;
1282 
1283 error:
1284 	IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1285 	goto out;
1286 }
1287 
1288 /*
1289  *	Throw away all pending data on the socket.
1290  */
1291 void ip_flush_pending_frames(struct sock *sk)
1292 {
1293 	struct inet_sock *inet = inet_sk(sk);
1294 	struct sk_buff *skb;
1295 
1296 	while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1297 		kfree_skb(skb);
1298 
1299 	inet->cork.flags &= ~IPCORK_OPT;
1300 	kfree(inet->cork.opt);
1301 	inet->cork.opt = NULL;
1302 	if (inet->cork.rt) {
1303 		ip_rt_put(inet->cork.rt);
1304 		inet->cork.rt = NULL;
1305 	}
1306 }
1307 
1308 
1309 /*
1310  *	Fetch data from kernel space and fill in checksum if needed.
1311  */
1312 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1313 			      int len, int odd, struct sk_buff *skb)
1314 {
1315 	unsigned int csum;
1316 
1317 	csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1318 	skb->csum = csum_block_add(skb->csum, csum, odd);
1319 	return 0;
1320 }
1321 
1322 /*
1323  *	Generic function to send a packet as reply to another packet.
1324  *	Used to send TCP resets so far. ICMP should use this function too.
1325  *
1326  *	Should run single threaded per socket because it uses the sock
1327  *     	structure to pass arguments.
1328  *
1329  *	LATER: switch from ip_build_xmit to ip_append_*
1330  */
1331 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1332 		   unsigned int len)
1333 {
1334 	struct inet_sock *inet = inet_sk(sk);
1335 	struct {
1336 		struct ip_options	opt;
1337 		char			data[40];
1338 	} replyopts;
1339 	struct ipcm_cookie ipc;
1340 	u32 daddr;
1341 	struct rtable *rt = (struct rtable*)skb->dst;
1342 
1343 	if (ip_options_echo(&replyopts.opt, skb))
1344 		return;
1345 
1346 	daddr = ipc.addr = rt->rt_src;
1347 	ipc.opt = NULL;
1348 
1349 	if (replyopts.opt.optlen) {
1350 		ipc.opt = &replyopts.opt;
1351 
1352 		if (ipc.opt->srr)
1353 			daddr = replyopts.opt.faddr;
1354 	}
1355 
1356 	{
1357 		struct flowi fl = { .nl_u = { .ip4_u =
1358 					      { .daddr = daddr,
1359 						.saddr = rt->rt_spec_dst,
1360 						.tos = RT_TOS(skb->nh.iph->tos) } },
1361 				    /* Not quite clean, but right. */
1362 				    .uli_u = { .ports =
1363 					       { .sport = skb->h.th->dest,
1364 					         .dport = skb->h.th->source } },
1365 				    .proto = sk->sk_protocol };
1366 		if (ip_route_output_key(&rt, &fl))
1367 			return;
1368 	}
1369 
1370 	/* And let IP do all the hard work.
1371 
1372 	   This chunk is not reenterable, hence spinlock.
1373 	   Note that it uses the fact, that this function is called
1374 	   with locally disabled BH and that sk cannot be already spinlocked.
1375 	 */
1376 	bh_lock_sock(sk);
1377 	inet->tos = skb->nh.iph->tos;
1378 	sk->sk_priority = skb->priority;
1379 	sk->sk_protocol = skb->nh.iph->protocol;
1380 	ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1381 		       &ipc, rt, MSG_DONTWAIT);
1382 	if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1383 		if (arg->csumoffset >= 0)
1384 			*((u16 *)skb->h.raw + arg->csumoffset) = csum_fold(csum_add(skb->csum, arg->csum));
1385 		skb->ip_summed = CHECKSUM_NONE;
1386 		ip_push_pending_frames(sk);
1387 	}
1388 
1389 	bh_unlock_sock(sk);
1390 
1391 	ip_rt_put(rt);
1392 }
1393 
1394 void __init ip_init(void)
1395 {
1396 	ip_rt_init();
1397 	inet_initpeers();
1398 
1399 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1400 	igmp_mc_proc_init();
1401 #endif
1402 }
1403 
1404 EXPORT_SYMBOL(ip_generic_getfrag);
1405 EXPORT_SYMBOL(ip_queue_xmit);
1406 EXPORT_SYMBOL(ip_send_check);
1407