xref: /openbmc/linux/net/ipv4/ip_output.c (revision b9ccfda2)
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 /* 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 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
117 {
118 	int ttl = inet->uc_ttl;
119 
120 	if (ttl < 0)
121 		ttl = ip4_dst_hoplimit(dst);
122 	return ttl;
123 }
124 
125 /*
126  *		Add an ip header to a skbuff and send it out.
127  *
128  */
129 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
130 			  __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
131 {
132 	struct inet_sock *inet = inet_sk(sk);
133 	struct rtable *rt = skb_rtable(skb);
134 	struct iphdr *iph;
135 
136 	/* Build the IP header. */
137 	skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
138 	skb_reset_network_header(skb);
139 	iph = ip_hdr(skb);
140 	iph->version  = 4;
141 	iph->ihl      = 5;
142 	iph->tos      = inet->tos;
143 	if (ip_dont_fragment(sk, &rt->dst))
144 		iph->frag_off = htons(IP_DF);
145 	else
146 		iph->frag_off = 0;
147 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
148 	iph->daddr    = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
149 	iph->saddr    = saddr;
150 	iph->protocol = sk->sk_protocol;
151 	ip_select_ident(iph, &rt->dst, sk);
152 
153 	if (opt && opt->opt.optlen) {
154 		iph->ihl += opt->opt.optlen>>2;
155 		ip_options_build(skb, &opt->opt, daddr, rt, 0);
156 	}
157 
158 	skb->priority = sk->sk_priority;
159 	skb->mark = sk->sk_mark;
160 
161 	/* Send it out. */
162 	return ip_local_out(skb);
163 }
164 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
165 
166 static inline int ip_finish_output2(struct sk_buff *skb)
167 {
168 	struct dst_entry *dst = skb_dst(skb);
169 	struct rtable *rt = (struct rtable *)dst;
170 	struct net_device *dev = dst->dev;
171 	unsigned int hh_len = LL_RESERVED_SPACE(dev);
172 	struct neighbour *neigh;
173 	u32 nexthop;
174 
175 	if (rt->rt_type == RTN_MULTICAST) {
176 		IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
177 	} else if (rt->rt_type == RTN_BROADCAST)
178 		IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
179 
180 	/* Be paranoid, rather than too clever. */
181 	if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
182 		struct sk_buff *skb2;
183 
184 		skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
185 		if (skb2 == NULL) {
186 			kfree_skb(skb);
187 			return -ENOMEM;
188 		}
189 		if (skb->sk)
190 			skb_set_owner_w(skb2, skb->sk);
191 		consume_skb(skb);
192 		skb = skb2;
193 	}
194 
195 	rcu_read_lock_bh();
196 	nexthop = rt->rt_gateway ? rt->rt_gateway : ip_hdr(skb)->daddr;
197 	neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
198 	if (unlikely(!neigh))
199 		neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
200 	if (neigh) {
201 		int res = dst_neigh_output(dst, neigh, skb);
202 
203 		rcu_read_unlock_bh();
204 		return res;
205 	}
206 	rcu_read_unlock_bh();
207 
208 	net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
209 			    __func__);
210 	kfree_skb(skb);
211 	return -EINVAL;
212 }
213 
214 static inline int ip_skb_dst_mtu(struct sk_buff *skb)
215 {
216 	struct inet_sock *inet = skb->sk ? inet_sk(skb->sk) : NULL;
217 
218 	return (inet && inet->pmtudisc == IP_PMTUDISC_PROBE) ?
219 	       skb_dst(skb)->dev->mtu : dst_mtu(skb_dst(skb));
220 }
221 
222 static int ip_finish_output(struct sk_buff *skb)
223 {
224 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
225 	/* Policy lookup after SNAT yielded a new policy */
226 	if (skb_dst(skb)->xfrm != NULL) {
227 		IPCB(skb)->flags |= IPSKB_REROUTED;
228 		return dst_output(skb);
229 	}
230 #endif
231 	if (skb->len > ip_skb_dst_mtu(skb) && !skb_is_gso(skb))
232 		return ip_fragment(skb, ip_finish_output2);
233 	else
234 		return ip_finish_output2(skb);
235 }
236 
237 int ip_mc_output(struct sk_buff *skb)
238 {
239 	struct sock *sk = skb->sk;
240 	struct rtable *rt = skb_rtable(skb);
241 	struct net_device *dev = rt->dst.dev;
242 
243 	/*
244 	 *	If the indicated interface is up and running, send the packet.
245 	 */
246 	IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
247 
248 	skb->dev = dev;
249 	skb->protocol = htons(ETH_P_IP);
250 
251 	/*
252 	 *	Multicasts are looped back for other local users
253 	 */
254 
255 	if (rt->rt_flags&RTCF_MULTICAST) {
256 		if (sk_mc_loop(sk)
257 #ifdef CONFIG_IP_MROUTE
258 		/* Small optimization: do not loopback not local frames,
259 		   which returned after forwarding; they will be  dropped
260 		   by ip_mr_input in any case.
261 		   Note, that local frames are looped back to be delivered
262 		   to local recipients.
263 
264 		   This check is duplicated in ip_mr_input at the moment.
265 		 */
266 		    &&
267 		    ((rt->rt_flags & RTCF_LOCAL) ||
268 		     !(IPCB(skb)->flags & IPSKB_FORWARDED))
269 #endif
270 		   ) {
271 			struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
272 			if (newskb)
273 				NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
274 					newskb, NULL, newskb->dev,
275 					dev_loopback_xmit);
276 		}
277 
278 		/* Multicasts with ttl 0 must not go beyond the host */
279 
280 		if (ip_hdr(skb)->ttl == 0) {
281 			kfree_skb(skb);
282 			return 0;
283 		}
284 	}
285 
286 	if (rt->rt_flags&RTCF_BROADCAST) {
287 		struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
288 		if (newskb)
289 			NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
290 				NULL, newskb->dev, dev_loopback_xmit);
291 	}
292 
293 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
294 			    skb->dev, ip_finish_output,
295 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
296 }
297 
298 int ip_output(struct sk_buff *skb)
299 {
300 	struct net_device *dev = skb_dst(skb)->dev;
301 
302 	IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
303 
304 	skb->dev = dev;
305 	skb->protocol = htons(ETH_P_IP);
306 
307 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
308 			    ip_finish_output,
309 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
310 }
311 
312 /*
313  * copy saddr and daddr, possibly using 64bit load/stores
314  * Equivalent to :
315  *   iph->saddr = fl4->saddr;
316  *   iph->daddr = fl4->daddr;
317  */
318 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
319 {
320 	BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
321 		     offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
322 	memcpy(&iph->saddr, &fl4->saddr,
323 	       sizeof(fl4->saddr) + sizeof(fl4->daddr));
324 }
325 
326 int ip_queue_xmit(struct sk_buff *skb, struct flowi *fl)
327 {
328 	struct sock *sk = skb->sk;
329 	struct inet_sock *inet = inet_sk(sk);
330 	struct ip_options_rcu *inet_opt;
331 	struct flowi4 *fl4;
332 	struct rtable *rt;
333 	struct iphdr *iph;
334 	int res;
335 
336 	/* Skip all of this if the packet is already routed,
337 	 * f.e. by something like SCTP.
338 	 */
339 	rcu_read_lock();
340 	inet_opt = rcu_dereference(inet->inet_opt);
341 	fl4 = &fl->u.ip4;
342 	rt = skb_rtable(skb);
343 	if (rt != NULL)
344 		goto packet_routed;
345 
346 	/* Make sure we can route this packet. */
347 	rt = (struct rtable *)__sk_dst_check(sk, 0);
348 	if (rt == NULL) {
349 		__be32 daddr;
350 
351 		/* Use correct destination address if we have options. */
352 		daddr = inet->inet_daddr;
353 		if (inet_opt && inet_opt->opt.srr)
354 			daddr = inet_opt->opt.faddr;
355 
356 		/* If this fails, retransmit mechanism of transport layer will
357 		 * keep trying until route appears or the connection times
358 		 * itself out.
359 		 */
360 		rt = ip_route_output_ports(sock_net(sk), fl4, sk,
361 					   daddr, inet->inet_saddr,
362 					   inet->inet_dport,
363 					   inet->inet_sport,
364 					   sk->sk_protocol,
365 					   RT_CONN_FLAGS(sk),
366 					   sk->sk_bound_dev_if);
367 		if (IS_ERR(rt))
368 			goto no_route;
369 		sk_setup_caps(sk, &rt->dst);
370 	}
371 	skb_dst_set_noref(skb, &rt->dst);
372 
373 packet_routed:
374 	if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_gateway)
375 		goto no_route;
376 
377 	/* OK, we know where to send it, allocate and build IP header. */
378 	skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
379 	skb_reset_network_header(skb);
380 	iph = ip_hdr(skb);
381 	*((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
382 	if (ip_dont_fragment(sk, &rt->dst) && !skb->local_df)
383 		iph->frag_off = htons(IP_DF);
384 	else
385 		iph->frag_off = 0;
386 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
387 	iph->protocol = sk->sk_protocol;
388 	ip_copy_addrs(iph, fl4);
389 
390 	/* Transport layer set skb->h.foo itself. */
391 
392 	if (inet_opt && inet_opt->opt.optlen) {
393 		iph->ihl += inet_opt->opt.optlen >> 2;
394 		ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
395 	}
396 
397 	ip_select_ident_more(iph, &rt->dst, sk,
398 			     (skb_shinfo(skb)->gso_segs ?: 1) - 1);
399 
400 	skb->priority = sk->sk_priority;
401 	skb->mark = sk->sk_mark;
402 
403 	res = ip_local_out(skb);
404 	rcu_read_unlock();
405 	return res;
406 
407 no_route:
408 	rcu_read_unlock();
409 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
410 	kfree_skb(skb);
411 	return -EHOSTUNREACH;
412 }
413 EXPORT_SYMBOL(ip_queue_xmit);
414 
415 
416 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
417 {
418 	to->pkt_type = from->pkt_type;
419 	to->priority = from->priority;
420 	to->protocol = from->protocol;
421 	skb_dst_drop(to);
422 	skb_dst_copy(to, from);
423 	to->dev = from->dev;
424 	to->mark = from->mark;
425 
426 	/* Copy the flags to each fragment. */
427 	IPCB(to)->flags = IPCB(from)->flags;
428 
429 #ifdef CONFIG_NET_SCHED
430 	to->tc_index = from->tc_index;
431 #endif
432 	nf_copy(to, from);
433 #if defined(CONFIG_NETFILTER_XT_TARGET_TRACE) || \
434     defined(CONFIG_NETFILTER_XT_TARGET_TRACE_MODULE)
435 	to->nf_trace = from->nf_trace;
436 #endif
437 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
438 	to->ipvs_property = from->ipvs_property;
439 #endif
440 	skb_copy_secmark(to, from);
441 }
442 
443 /*
444  *	This IP datagram is too large to be sent in one piece.  Break it up into
445  *	smaller pieces (each of size equal to IP header plus
446  *	a block of the data of the original IP data part) that will yet fit in a
447  *	single device frame, and queue such a frame for sending.
448  */
449 
450 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
451 {
452 	struct iphdr *iph;
453 	int ptr;
454 	struct net_device *dev;
455 	struct sk_buff *skb2;
456 	unsigned int mtu, hlen, left, len, ll_rs;
457 	int offset;
458 	__be16 not_last_frag;
459 	struct rtable *rt = skb_rtable(skb);
460 	int err = 0;
461 
462 	dev = rt->dst.dev;
463 
464 	/*
465 	 *	Point into the IP datagram header.
466 	 */
467 
468 	iph = ip_hdr(skb);
469 
470 	if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
471 		IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
472 		icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
473 			  htonl(ip_skb_dst_mtu(skb)));
474 		kfree_skb(skb);
475 		return -EMSGSIZE;
476 	}
477 
478 	/*
479 	 *	Setup starting values.
480 	 */
481 
482 	hlen = iph->ihl * 4;
483 	mtu = dst_mtu(&rt->dst) - hlen;	/* Size of data space */
484 #ifdef CONFIG_BRIDGE_NETFILTER
485 	if (skb->nf_bridge)
486 		mtu -= nf_bridge_mtu_reduction(skb);
487 #endif
488 	IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
489 
490 	/* When frag_list is given, use it. First, check its validity:
491 	 * some transformers could create wrong frag_list or break existing
492 	 * one, it is not prohibited. In this case fall back to copying.
493 	 *
494 	 * LATER: this step can be merged to real generation of fragments,
495 	 * we can switch to copy when see the first bad fragment.
496 	 */
497 	if (skb_has_frag_list(skb)) {
498 		struct sk_buff *frag, *frag2;
499 		int first_len = skb_pagelen(skb);
500 
501 		if (first_len - hlen > mtu ||
502 		    ((first_len - hlen) & 7) ||
503 		    ip_is_fragment(iph) ||
504 		    skb_cloned(skb))
505 			goto slow_path;
506 
507 		skb_walk_frags(skb, frag) {
508 			/* Correct geometry. */
509 			if (frag->len > mtu ||
510 			    ((frag->len & 7) && frag->next) ||
511 			    skb_headroom(frag) < hlen)
512 				goto slow_path_clean;
513 
514 			/* Partially cloned skb? */
515 			if (skb_shared(frag))
516 				goto slow_path_clean;
517 
518 			BUG_ON(frag->sk);
519 			if (skb->sk) {
520 				frag->sk = skb->sk;
521 				frag->destructor = sock_wfree;
522 			}
523 			skb->truesize -= frag->truesize;
524 		}
525 
526 		/* Everything is OK. Generate! */
527 
528 		err = 0;
529 		offset = 0;
530 		frag = skb_shinfo(skb)->frag_list;
531 		skb_frag_list_init(skb);
532 		skb->data_len = first_len - skb_headlen(skb);
533 		skb->len = first_len;
534 		iph->tot_len = htons(first_len);
535 		iph->frag_off = htons(IP_MF);
536 		ip_send_check(iph);
537 
538 		for (;;) {
539 			/* Prepare header of the next frame,
540 			 * before previous one went down. */
541 			if (frag) {
542 				frag->ip_summed = CHECKSUM_NONE;
543 				skb_reset_transport_header(frag);
544 				__skb_push(frag, hlen);
545 				skb_reset_network_header(frag);
546 				memcpy(skb_network_header(frag), iph, hlen);
547 				iph = ip_hdr(frag);
548 				iph->tot_len = htons(frag->len);
549 				ip_copy_metadata(frag, skb);
550 				if (offset == 0)
551 					ip_options_fragment(frag);
552 				offset += skb->len - hlen;
553 				iph->frag_off = htons(offset>>3);
554 				if (frag->next != NULL)
555 					iph->frag_off |= htons(IP_MF);
556 				/* Ready, complete checksum */
557 				ip_send_check(iph);
558 			}
559 
560 			err = output(skb);
561 
562 			if (!err)
563 				IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
564 			if (err || !frag)
565 				break;
566 
567 			skb = frag;
568 			frag = skb->next;
569 			skb->next = NULL;
570 		}
571 
572 		if (err == 0) {
573 			IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
574 			return 0;
575 		}
576 
577 		while (frag) {
578 			skb = frag->next;
579 			kfree_skb(frag);
580 			frag = skb;
581 		}
582 		IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
583 		return err;
584 
585 slow_path_clean:
586 		skb_walk_frags(skb, frag2) {
587 			if (frag2 == frag)
588 				break;
589 			frag2->sk = NULL;
590 			frag2->destructor = NULL;
591 			skb->truesize += frag2->truesize;
592 		}
593 	}
594 
595 slow_path:
596 	left = skb->len - hlen;		/* Space per frame */
597 	ptr = hlen;		/* Where to start from */
598 
599 	/* for bridged IP traffic encapsulated inside f.e. a vlan header,
600 	 * we need to make room for the encapsulating header
601 	 */
602 	ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
603 
604 	/*
605 	 *	Fragment the datagram.
606 	 */
607 
608 	offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
609 	not_last_frag = iph->frag_off & htons(IP_MF);
610 
611 	/*
612 	 *	Keep copying data until we run out.
613 	 */
614 
615 	while (left > 0) {
616 		len = left;
617 		/* IF: it doesn't fit, use 'mtu' - the data space left */
618 		if (len > mtu)
619 			len = mtu;
620 		/* IF: we are not sending up to and including the packet end
621 		   then align the next start on an eight byte boundary */
622 		if (len < left)	{
623 			len &= ~7;
624 		}
625 		/*
626 		 *	Allocate buffer.
627 		 */
628 
629 		if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
630 			NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
631 			err = -ENOMEM;
632 			goto fail;
633 		}
634 
635 		/*
636 		 *	Set up data on packet
637 		 */
638 
639 		ip_copy_metadata(skb2, skb);
640 		skb_reserve(skb2, ll_rs);
641 		skb_put(skb2, len + hlen);
642 		skb_reset_network_header(skb2);
643 		skb2->transport_header = skb2->network_header + hlen;
644 
645 		/*
646 		 *	Charge the memory for the fragment to any owner
647 		 *	it might possess
648 		 */
649 
650 		if (skb->sk)
651 			skb_set_owner_w(skb2, skb->sk);
652 
653 		/*
654 		 *	Copy the packet header into the new buffer.
655 		 */
656 
657 		skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
658 
659 		/*
660 		 *	Copy a block of the IP datagram.
661 		 */
662 		if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
663 			BUG();
664 		left -= len;
665 
666 		/*
667 		 *	Fill in the new header fields.
668 		 */
669 		iph = ip_hdr(skb2);
670 		iph->frag_off = htons((offset >> 3));
671 
672 		/* ANK: dirty, but effective trick. Upgrade options only if
673 		 * the segment to be fragmented was THE FIRST (otherwise,
674 		 * options are already fixed) and make it ONCE
675 		 * on the initial skb, so that all the following fragments
676 		 * will inherit fixed options.
677 		 */
678 		if (offset == 0)
679 			ip_options_fragment(skb);
680 
681 		/*
682 		 *	Added AC : If we are fragmenting a fragment that's not the
683 		 *		   last fragment then keep MF on each bit
684 		 */
685 		if (left > 0 || not_last_frag)
686 			iph->frag_off |= htons(IP_MF);
687 		ptr += len;
688 		offset += len;
689 
690 		/*
691 		 *	Put this fragment into the sending queue.
692 		 */
693 		iph->tot_len = htons(len + hlen);
694 
695 		ip_send_check(iph);
696 
697 		err = output(skb2);
698 		if (err)
699 			goto fail;
700 
701 		IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
702 	}
703 	consume_skb(skb);
704 	IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
705 	return err;
706 
707 fail:
708 	kfree_skb(skb);
709 	IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
710 	return err;
711 }
712 EXPORT_SYMBOL(ip_fragment);
713 
714 int
715 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
716 {
717 	struct iovec *iov = from;
718 
719 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
720 		if (memcpy_fromiovecend(to, iov, offset, len) < 0)
721 			return -EFAULT;
722 	} else {
723 		__wsum csum = 0;
724 		if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
725 			return -EFAULT;
726 		skb->csum = csum_block_add(skb->csum, csum, odd);
727 	}
728 	return 0;
729 }
730 EXPORT_SYMBOL(ip_generic_getfrag);
731 
732 static inline __wsum
733 csum_page(struct page *page, int offset, int copy)
734 {
735 	char *kaddr;
736 	__wsum csum;
737 	kaddr = kmap(page);
738 	csum = csum_partial(kaddr + offset, copy, 0);
739 	kunmap(page);
740 	return csum;
741 }
742 
743 static inline int ip_ufo_append_data(struct sock *sk,
744 			struct sk_buff_head *queue,
745 			int getfrag(void *from, char *to, int offset, int len,
746 			       int odd, struct sk_buff *skb),
747 			void *from, int length, int hh_len, int fragheaderlen,
748 			int transhdrlen, int maxfraglen, unsigned int flags)
749 {
750 	struct sk_buff *skb;
751 	int err;
752 
753 	/* There is support for UDP fragmentation offload by network
754 	 * device, so create one single skb packet containing complete
755 	 * udp datagram
756 	 */
757 	if ((skb = skb_peek_tail(queue)) == NULL) {
758 		skb = sock_alloc_send_skb(sk,
759 			hh_len + fragheaderlen + transhdrlen + 20,
760 			(flags & MSG_DONTWAIT), &err);
761 
762 		if (skb == NULL)
763 			return err;
764 
765 		/* reserve space for Hardware header */
766 		skb_reserve(skb, hh_len);
767 
768 		/* create space for UDP/IP header */
769 		skb_put(skb, fragheaderlen + transhdrlen);
770 
771 		/* initialize network header pointer */
772 		skb_reset_network_header(skb);
773 
774 		/* initialize protocol header pointer */
775 		skb->transport_header = skb->network_header + fragheaderlen;
776 
777 		skb->ip_summed = CHECKSUM_PARTIAL;
778 		skb->csum = 0;
779 
780 		/* specify the length of each IP datagram fragment */
781 		skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
782 		skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
783 		__skb_queue_tail(queue, skb);
784 	}
785 
786 	return skb_append_datato_frags(sk, skb, getfrag, from,
787 				       (length - transhdrlen));
788 }
789 
790 static int __ip_append_data(struct sock *sk,
791 			    struct flowi4 *fl4,
792 			    struct sk_buff_head *queue,
793 			    struct inet_cork *cork,
794 			    int getfrag(void *from, char *to, int offset,
795 					int len, int odd, struct sk_buff *skb),
796 			    void *from, int length, int transhdrlen,
797 			    unsigned int flags)
798 {
799 	struct inet_sock *inet = inet_sk(sk);
800 	struct sk_buff *skb;
801 
802 	struct ip_options *opt = cork->opt;
803 	int hh_len;
804 	int exthdrlen;
805 	int mtu;
806 	int copy;
807 	int err;
808 	int offset = 0;
809 	unsigned int maxfraglen, fragheaderlen;
810 	int csummode = CHECKSUM_NONE;
811 	struct rtable *rt = (struct rtable *)cork->dst;
812 
813 	skb = skb_peek_tail(queue);
814 
815 	exthdrlen = !skb ? rt->dst.header_len : 0;
816 	mtu = cork->fragsize;
817 
818 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
819 
820 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
821 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
822 
823 	if (cork->length + length > 0xFFFF - fragheaderlen) {
824 		ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
825 			       mtu-exthdrlen);
826 		return -EMSGSIZE;
827 	}
828 
829 	/*
830 	 * transhdrlen > 0 means that this is the first fragment and we wish
831 	 * it won't be fragmented in the future.
832 	 */
833 	if (transhdrlen &&
834 	    length + fragheaderlen <= mtu &&
835 	    rt->dst.dev->features & NETIF_F_V4_CSUM &&
836 	    !exthdrlen)
837 		csummode = CHECKSUM_PARTIAL;
838 
839 	cork->length += length;
840 	if (((length > mtu) || (skb && skb_is_gso(skb))) &&
841 	    (sk->sk_protocol == IPPROTO_UDP) &&
842 	    (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len) {
843 		err = ip_ufo_append_data(sk, queue, getfrag, from, length,
844 					 hh_len, fragheaderlen, transhdrlen,
845 					 maxfraglen, flags);
846 		if (err)
847 			goto error;
848 		return 0;
849 	}
850 
851 	/* So, what's going on in the loop below?
852 	 *
853 	 * We use calculated fragment length to generate chained skb,
854 	 * each of segments is IP fragment ready for sending to network after
855 	 * adding appropriate IP header.
856 	 */
857 
858 	if (!skb)
859 		goto alloc_new_skb;
860 
861 	while (length > 0) {
862 		/* Check if the remaining data fits into current packet. */
863 		copy = mtu - skb->len;
864 		if (copy < length)
865 			copy = maxfraglen - skb->len;
866 		if (copy <= 0) {
867 			char *data;
868 			unsigned int datalen;
869 			unsigned int fraglen;
870 			unsigned int fraggap;
871 			unsigned int alloclen;
872 			struct sk_buff *skb_prev;
873 alloc_new_skb:
874 			skb_prev = skb;
875 			if (skb_prev)
876 				fraggap = skb_prev->len - maxfraglen;
877 			else
878 				fraggap = 0;
879 
880 			/*
881 			 * If remaining data exceeds the mtu,
882 			 * we know we need more fragment(s).
883 			 */
884 			datalen = length + fraggap;
885 			if (datalen > mtu - fragheaderlen)
886 				datalen = maxfraglen - fragheaderlen;
887 			fraglen = datalen + fragheaderlen;
888 
889 			if ((flags & MSG_MORE) &&
890 			    !(rt->dst.dev->features&NETIF_F_SG))
891 				alloclen = mtu;
892 			else
893 				alloclen = fraglen;
894 
895 			alloclen += exthdrlen;
896 
897 			/* The last fragment gets additional space at tail.
898 			 * Note, with MSG_MORE we overallocate on fragments,
899 			 * because we have no idea what fragment will be
900 			 * the last.
901 			 */
902 			if (datalen == length + fraggap)
903 				alloclen += rt->dst.trailer_len;
904 
905 			if (transhdrlen) {
906 				skb = sock_alloc_send_skb(sk,
907 						alloclen + hh_len + 15,
908 						(flags & MSG_DONTWAIT), &err);
909 			} else {
910 				skb = NULL;
911 				if (atomic_read(&sk->sk_wmem_alloc) <=
912 				    2 * sk->sk_sndbuf)
913 					skb = sock_wmalloc(sk,
914 							   alloclen + hh_len + 15, 1,
915 							   sk->sk_allocation);
916 				if (unlikely(skb == NULL))
917 					err = -ENOBUFS;
918 				else
919 					/* only the initial fragment is
920 					   time stamped */
921 					cork->tx_flags = 0;
922 			}
923 			if (skb == NULL)
924 				goto error;
925 
926 			/*
927 			 *	Fill in the control structures
928 			 */
929 			skb->ip_summed = csummode;
930 			skb->csum = 0;
931 			skb_reserve(skb, hh_len);
932 			skb_shinfo(skb)->tx_flags = cork->tx_flags;
933 
934 			/*
935 			 *	Find where to start putting bytes.
936 			 */
937 			data = skb_put(skb, fraglen + exthdrlen);
938 			skb_set_network_header(skb, exthdrlen);
939 			skb->transport_header = (skb->network_header +
940 						 fragheaderlen);
941 			data += fragheaderlen + 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 				pskb_trim_unique(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(queue, skb);
970 			continue;
971 		}
972 
973 		if (copy > length)
974 			copy = length;
975 
976 		if (!(rt->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 = cork->page;
990 			int off = cork->off;
991 			unsigned int left;
992 
993 			if (page && (left = PAGE_SIZE - off) > 0) {
994 				if (copy >= left)
995 					copy = left;
996 				if (page != skb_frag_page(frag)) {
997 					if (i == MAX_SKB_FRAGS) {
998 						err = -EMSGSIZE;
999 						goto error;
1000 					}
1001 					skb_fill_page_desc(skb, i, page, off, 0);
1002 					skb_frag_ref(skb, i);
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 				cork->page = page;
1014 				cork->off = 0;
1015 
1016 				skb_fill_page_desc(skb, i, page, 0, 0);
1017 				frag = &skb_shinfo(skb)->frags[i];
1018 			} else {
1019 				err = -EMSGSIZE;
1020 				goto error;
1021 			}
1022 			if (getfrag(from, skb_frag_address(frag)+skb_frag_size(frag),
1023 				    offset, copy, skb->len, skb) < 0) {
1024 				err = -EFAULT;
1025 				goto error;
1026 			}
1027 			cork->off += copy;
1028 			skb_frag_size_add(frag, copy);
1029 			skb->len += copy;
1030 			skb->data_len += copy;
1031 			skb->truesize += copy;
1032 			atomic_add(copy, &sk->sk_wmem_alloc);
1033 		}
1034 		offset += copy;
1035 		length -= copy;
1036 	}
1037 
1038 	return 0;
1039 
1040 error:
1041 	cork->length -= length;
1042 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1043 	return err;
1044 }
1045 
1046 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1047 			 struct ipcm_cookie *ipc, struct rtable **rtp)
1048 {
1049 	struct inet_sock *inet = inet_sk(sk);
1050 	struct ip_options_rcu *opt;
1051 	struct rtable *rt;
1052 
1053 	/*
1054 	 * setup for corking.
1055 	 */
1056 	opt = ipc->opt;
1057 	if (opt) {
1058 		if (cork->opt == NULL) {
1059 			cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1060 					    sk->sk_allocation);
1061 			if (unlikely(cork->opt == NULL))
1062 				return -ENOBUFS;
1063 		}
1064 		memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1065 		cork->flags |= IPCORK_OPT;
1066 		cork->addr = ipc->addr;
1067 	}
1068 	rt = *rtp;
1069 	if (unlikely(!rt))
1070 		return -EFAULT;
1071 	/*
1072 	 * We steal reference to this route, caller should not release it
1073 	 */
1074 	*rtp = NULL;
1075 	cork->fragsize = inet->pmtudisc == IP_PMTUDISC_PROBE ?
1076 			 rt->dst.dev->mtu : dst_mtu(&rt->dst);
1077 	cork->dst = &rt->dst;
1078 	cork->length = 0;
1079 	cork->tx_flags = ipc->tx_flags;
1080 	cork->page = NULL;
1081 	cork->off = 0;
1082 
1083 	return 0;
1084 }
1085 
1086 /*
1087  *	ip_append_data() and ip_append_page() can make one large IP datagram
1088  *	from many pieces of data. Each pieces will be holded on the socket
1089  *	until ip_push_pending_frames() is called. Each piece can be a page
1090  *	or non-page data.
1091  *
1092  *	Not only UDP, other transport protocols - e.g. raw sockets - can use
1093  *	this interface potentially.
1094  *
1095  *	LATER: length must be adjusted by pad at tail, when it is required.
1096  */
1097 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1098 		   int getfrag(void *from, char *to, int offset, int len,
1099 			       int odd, struct sk_buff *skb),
1100 		   void *from, int length, int transhdrlen,
1101 		   struct ipcm_cookie *ipc, struct rtable **rtp,
1102 		   unsigned int flags)
1103 {
1104 	struct inet_sock *inet = inet_sk(sk);
1105 	int err;
1106 
1107 	if (flags&MSG_PROBE)
1108 		return 0;
1109 
1110 	if (skb_queue_empty(&sk->sk_write_queue)) {
1111 		err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1112 		if (err)
1113 			return err;
1114 	} else {
1115 		transhdrlen = 0;
1116 	}
1117 
1118 	return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base, getfrag,
1119 				from, length, transhdrlen, flags);
1120 }
1121 
1122 ssize_t	ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1123 		       int offset, size_t size, int flags)
1124 {
1125 	struct inet_sock *inet = inet_sk(sk);
1126 	struct sk_buff *skb;
1127 	struct rtable *rt;
1128 	struct ip_options *opt = NULL;
1129 	struct inet_cork *cork;
1130 	int hh_len;
1131 	int mtu;
1132 	int len;
1133 	int err;
1134 	unsigned int maxfraglen, fragheaderlen, fraggap;
1135 
1136 	if (inet->hdrincl)
1137 		return -EPERM;
1138 
1139 	if (flags&MSG_PROBE)
1140 		return 0;
1141 
1142 	if (skb_queue_empty(&sk->sk_write_queue))
1143 		return -EINVAL;
1144 
1145 	cork = &inet->cork.base;
1146 	rt = (struct rtable *)cork->dst;
1147 	if (cork->flags & IPCORK_OPT)
1148 		opt = cork->opt;
1149 
1150 	if (!(rt->dst.dev->features&NETIF_F_SG))
1151 		return -EOPNOTSUPP;
1152 
1153 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1154 	mtu = cork->fragsize;
1155 
1156 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1157 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1158 
1159 	if (cork->length + size > 0xFFFF - fragheaderlen) {
1160 		ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport, mtu);
1161 		return -EMSGSIZE;
1162 	}
1163 
1164 	if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1165 		return -EINVAL;
1166 
1167 	cork->length += size;
1168 	if ((size + skb->len > mtu) &&
1169 	    (sk->sk_protocol == IPPROTO_UDP) &&
1170 	    (rt->dst.dev->features & NETIF_F_UFO)) {
1171 		skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1172 		skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1173 	}
1174 
1175 
1176 	while (size > 0) {
1177 		int i;
1178 
1179 		if (skb_is_gso(skb))
1180 			len = size;
1181 		else {
1182 
1183 			/* Check if the remaining data fits into current packet. */
1184 			len = mtu - skb->len;
1185 			if (len < size)
1186 				len = maxfraglen - skb->len;
1187 		}
1188 		if (len <= 0) {
1189 			struct sk_buff *skb_prev;
1190 			int alloclen;
1191 
1192 			skb_prev = skb;
1193 			fraggap = skb_prev->len - maxfraglen;
1194 
1195 			alloclen = fragheaderlen + hh_len + fraggap + 15;
1196 			skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1197 			if (unlikely(!skb)) {
1198 				err = -ENOBUFS;
1199 				goto error;
1200 			}
1201 
1202 			/*
1203 			 *	Fill in the control structures
1204 			 */
1205 			skb->ip_summed = CHECKSUM_NONE;
1206 			skb->csum = 0;
1207 			skb_reserve(skb, hh_len);
1208 
1209 			/*
1210 			 *	Find where to start putting bytes.
1211 			 */
1212 			skb_put(skb, fragheaderlen + fraggap);
1213 			skb_reset_network_header(skb);
1214 			skb->transport_header = (skb->network_header +
1215 						 fragheaderlen);
1216 			if (fraggap) {
1217 				skb->csum = skb_copy_and_csum_bits(skb_prev,
1218 								   maxfraglen,
1219 						    skb_transport_header(skb),
1220 								   fraggap, 0);
1221 				skb_prev->csum = csum_sub(skb_prev->csum,
1222 							  skb->csum);
1223 				pskb_trim_unique(skb_prev, maxfraglen);
1224 			}
1225 
1226 			/*
1227 			 * Put the packet on the pending queue.
1228 			 */
1229 			__skb_queue_tail(&sk->sk_write_queue, skb);
1230 			continue;
1231 		}
1232 
1233 		i = skb_shinfo(skb)->nr_frags;
1234 		if (len > size)
1235 			len = size;
1236 		if (skb_can_coalesce(skb, i, page, offset)) {
1237 			skb_frag_size_add(&skb_shinfo(skb)->frags[i-1], len);
1238 		} else if (i < MAX_SKB_FRAGS) {
1239 			get_page(page);
1240 			skb_fill_page_desc(skb, i, page, offset, len);
1241 		} else {
1242 			err = -EMSGSIZE;
1243 			goto error;
1244 		}
1245 
1246 		if (skb->ip_summed == CHECKSUM_NONE) {
1247 			__wsum csum;
1248 			csum = csum_page(page, offset, len);
1249 			skb->csum = csum_block_add(skb->csum, csum, skb->len);
1250 		}
1251 
1252 		skb->len += len;
1253 		skb->data_len += len;
1254 		skb->truesize += len;
1255 		atomic_add(len, &sk->sk_wmem_alloc);
1256 		offset += len;
1257 		size -= len;
1258 	}
1259 	return 0;
1260 
1261 error:
1262 	cork->length -= size;
1263 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1264 	return err;
1265 }
1266 
1267 static void ip_cork_release(struct inet_cork *cork)
1268 {
1269 	cork->flags &= ~IPCORK_OPT;
1270 	kfree(cork->opt);
1271 	cork->opt = NULL;
1272 	dst_release(cork->dst);
1273 	cork->dst = NULL;
1274 }
1275 
1276 /*
1277  *	Combined all pending IP fragments on the socket as one IP datagram
1278  *	and push them out.
1279  */
1280 struct sk_buff *__ip_make_skb(struct sock *sk,
1281 			      struct flowi4 *fl4,
1282 			      struct sk_buff_head *queue,
1283 			      struct inet_cork *cork)
1284 {
1285 	struct sk_buff *skb, *tmp_skb;
1286 	struct sk_buff **tail_skb;
1287 	struct inet_sock *inet = inet_sk(sk);
1288 	struct net *net = sock_net(sk);
1289 	struct ip_options *opt = NULL;
1290 	struct rtable *rt = (struct rtable *)cork->dst;
1291 	struct iphdr *iph;
1292 	__be16 df = 0;
1293 	__u8 ttl;
1294 
1295 	if ((skb = __skb_dequeue(queue)) == NULL)
1296 		goto out;
1297 	tail_skb = &(skb_shinfo(skb)->frag_list);
1298 
1299 	/* move skb->data to ip header from ext header */
1300 	if (skb->data < skb_network_header(skb))
1301 		__skb_pull(skb, skb_network_offset(skb));
1302 	while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1303 		__skb_pull(tmp_skb, skb_network_header_len(skb));
1304 		*tail_skb = tmp_skb;
1305 		tail_skb = &(tmp_skb->next);
1306 		skb->len += tmp_skb->len;
1307 		skb->data_len += tmp_skb->len;
1308 		skb->truesize += tmp_skb->truesize;
1309 		tmp_skb->destructor = NULL;
1310 		tmp_skb->sk = NULL;
1311 	}
1312 
1313 	/* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1314 	 * to fragment the frame generated here. No matter, what transforms
1315 	 * how transforms change size of the packet, it will come out.
1316 	 */
1317 	if (inet->pmtudisc < IP_PMTUDISC_DO)
1318 		skb->local_df = 1;
1319 
1320 	/* DF bit is set when we want to see DF on outgoing frames.
1321 	 * If local_df is set too, we still allow to fragment this frame
1322 	 * locally. */
1323 	if (inet->pmtudisc >= IP_PMTUDISC_DO ||
1324 	    (skb->len <= dst_mtu(&rt->dst) &&
1325 	     ip_dont_fragment(sk, &rt->dst)))
1326 		df = htons(IP_DF);
1327 
1328 	if (cork->flags & IPCORK_OPT)
1329 		opt = cork->opt;
1330 
1331 	if (rt->rt_type == RTN_MULTICAST)
1332 		ttl = inet->mc_ttl;
1333 	else
1334 		ttl = ip_select_ttl(inet, &rt->dst);
1335 
1336 	iph = (struct iphdr *)skb->data;
1337 	iph->version = 4;
1338 	iph->ihl = 5;
1339 	iph->tos = inet->tos;
1340 	iph->frag_off = df;
1341 	ip_select_ident(iph, &rt->dst, sk);
1342 	iph->ttl = ttl;
1343 	iph->protocol = sk->sk_protocol;
1344 	ip_copy_addrs(iph, fl4);
1345 
1346 	if (opt) {
1347 		iph->ihl += opt->optlen>>2;
1348 		ip_options_build(skb, opt, cork->addr, rt, 0);
1349 	}
1350 
1351 	skb->priority = sk->sk_priority;
1352 	skb->mark = sk->sk_mark;
1353 	/*
1354 	 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1355 	 * on dst refcount
1356 	 */
1357 	cork->dst = NULL;
1358 	skb_dst_set(skb, &rt->dst);
1359 
1360 	if (iph->protocol == IPPROTO_ICMP)
1361 		icmp_out_count(net, ((struct icmphdr *)
1362 			skb_transport_header(skb))->type);
1363 
1364 	ip_cork_release(cork);
1365 out:
1366 	return skb;
1367 }
1368 
1369 int ip_send_skb(struct sk_buff *skb)
1370 {
1371 	struct net *net = sock_net(skb->sk);
1372 	int err;
1373 
1374 	err = ip_local_out(skb);
1375 	if (err) {
1376 		if (err > 0)
1377 			err = net_xmit_errno(err);
1378 		if (err)
1379 			IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1380 	}
1381 
1382 	return err;
1383 }
1384 
1385 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1386 {
1387 	struct sk_buff *skb;
1388 
1389 	skb = ip_finish_skb(sk, fl4);
1390 	if (!skb)
1391 		return 0;
1392 
1393 	/* Netfilter gets whole the not fragmented skb. */
1394 	return ip_send_skb(skb);
1395 }
1396 
1397 /*
1398  *	Throw away all pending data on the socket.
1399  */
1400 static void __ip_flush_pending_frames(struct sock *sk,
1401 				      struct sk_buff_head *queue,
1402 				      struct inet_cork *cork)
1403 {
1404 	struct sk_buff *skb;
1405 
1406 	while ((skb = __skb_dequeue_tail(queue)) != NULL)
1407 		kfree_skb(skb);
1408 
1409 	ip_cork_release(cork);
1410 }
1411 
1412 void ip_flush_pending_frames(struct sock *sk)
1413 {
1414 	__ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1415 }
1416 
1417 struct sk_buff *ip_make_skb(struct sock *sk,
1418 			    struct flowi4 *fl4,
1419 			    int getfrag(void *from, char *to, int offset,
1420 					int len, int odd, struct sk_buff *skb),
1421 			    void *from, int length, int transhdrlen,
1422 			    struct ipcm_cookie *ipc, struct rtable **rtp,
1423 			    unsigned int flags)
1424 {
1425 	struct inet_cork cork;
1426 	struct sk_buff_head queue;
1427 	int err;
1428 
1429 	if (flags & MSG_PROBE)
1430 		return NULL;
1431 
1432 	__skb_queue_head_init(&queue);
1433 
1434 	cork.flags = 0;
1435 	cork.addr = 0;
1436 	cork.opt = NULL;
1437 	err = ip_setup_cork(sk, &cork, ipc, rtp);
1438 	if (err)
1439 		return ERR_PTR(err);
1440 
1441 	err = __ip_append_data(sk, fl4, &queue, &cork, getfrag,
1442 			       from, length, transhdrlen, flags);
1443 	if (err) {
1444 		__ip_flush_pending_frames(sk, &queue, &cork);
1445 		return ERR_PTR(err);
1446 	}
1447 
1448 	return __ip_make_skb(sk, fl4, &queue, &cork);
1449 }
1450 
1451 /*
1452  *	Fetch data from kernel space and fill in checksum if needed.
1453  */
1454 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1455 			      int len, int odd, struct sk_buff *skb)
1456 {
1457 	__wsum csum;
1458 
1459 	csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1460 	skb->csum = csum_block_add(skb->csum, csum, odd);
1461 	return 0;
1462 }
1463 
1464 /*
1465  *	Generic function to send a packet as reply to another packet.
1466  *	Used to send some TCP resets/acks so far.
1467  *
1468  *	Use a fake percpu inet socket to avoid false sharing and contention.
1469  */
1470 static DEFINE_PER_CPU(struct inet_sock, unicast_sock) = {
1471 	.sk = {
1472 		.__sk_common = {
1473 			.skc_refcnt = ATOMIC_INIT(1),
1474 		},
1475 		.sk_wmem_alloc	= ATOMIC_INIT(1),
1476 		.sk_allocation	= GFP_ATOMIC,
1477 		.sk_flags	= (1UL << SOCK_USE_WRITE_QUEUE),
1478 	},
1479 	.pmtudisc	= IP_PMTUDISC_WANT,
1480 	.uc_ttl		= -1,
1481 };
1482 
1483 void ip_send_unicast_reply(struct net *net, struct sk_buff *skb, __be32 daddr,
1484 			   __be32 saddr, const struct ip_reply_arg *arg,
1485 			   unsigned int len)
1486 {
1487 	struct ip_options_data replyopts;
1488 	struct ipcm_cookie ipc;
1489 	struct flowi4 fl4;
1490 	struct rtable *rt = skb_rtable(skb);
1491 	struct sk_buff *nskb;
1492 	struct sock *sk;
1493 	struct inet_sock *inet;
1494 
1495 	if (ip_options_echo(&replyopts.opt.opt, skb))
1496 		return;
1497 
1498 	ipc.addr = daddr;
1499 	ipc.opt = NULL;
1500 	ipc.tx_flags = 0;
1501 
1502 	if (replyopts.opt.opt.optlen) {
1503 		ipc.opt = &replyopts.opt;
1504 
1505 		if (replyopts.opt.opt.srr)
1506 			daddr = replyopts.opt.opt.faddr;
1507 	}
1508 
1509 	flowi4_init_output(&fl4, arg->bound_dev_if, 0,
1510 			   RT_TOS(arg->tos),
1511 			   RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1512 			   ip_reply_arg_flowi_flags(arg),
1513 			   daddr, saddr,
1514 			   tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1515 	security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1516 	rt = ip_route_output_key(net, &fl4);
1517 	if (IS_ERR(rt))
1518 		return;
1519 
1520 	inet = &get_cpu_var(unicast_sock);
1521 
1522 	inet->tos = arg->tos;
1523 	sk = &inet->sk;
1524 	sk->sk_priority = skb->priority;
1525 	sk->sk_protocol = ip_hdr(skb)->protocol;
1526 	sk->sk_bound_dev_if = arg->bound_dev_if;
1527 	sock_net_set(sk, net);
1528 	__skb_queue_head_init(&sk->sk_write_queue);
1529 	sk->sk_sndbuf = sysctl_wmem_default;
1530 	ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1531 		       &ipc, &rt, MSG_DONTWAIT);
1532 	nskb = skb_peek(&sk->sk_write_queue);
1533 	if (nskb) {
1534 		if (arg->csumoffset >= 0)
1535 			*((__sum16 *)skb_transport_header(nskb) +
1536 			  arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1537 								arg->csum));
1538 		nskb->ip_summed = CHECKSUM_NONE;
1539 		skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1540 		ip_push_pending_frames(sk, &fl4);
1541 	}
1542 
1543 	put_cpu_var(unicast_sock);
1544 
1545 	ip_rt_put(rt);
1546 }
1547 
1548 void __init ip_init(void)
1549 {
1550 	ip_rt_init();
1551 	inet_initpeers();
1552 
1553 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1554 	igmp_mc_proc_init();
1555 #endif
1556 }
1557