xref: /openbmc/linux/net/ipv4/ip_output.c (revision 4161b450)
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 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_sk(struct sock *sk, struct sk_buff *skb)
105 {
106 	int err;
107 
108 	err = __ip_local_out(skb);
109 	if (likely(err == 1))
110 		err = dst_output_sk(sk, skb);
111 
112 	return err;
113 }
114 EXPORT_SYMBOL_GPL(ip_local_out_sk);
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(skb, 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 = (__force u32) rt_nexthop(rt, 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 (!IS_ERR(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 int ip_finish_output_gso(struct sk_buff *skb)
215 {
216 	netdev_features_t features;
217 	struct sk_buff *segs;
218 	int ret = 0;
219 
220 	/* common case: locally created skb or seglen is <= mtu */
221 	if (((IPCB(skb)->flags & IPSKB_FORWARDED) == 0) ||
222 	      skb_gso_network_seglen(skb) <= ip_skb_dst_mtu(skb))
223 		return ip_finish_output2(skb);
224 
225 	/* Slowpath -  GSO segment length is exceeding the dst MTU.
226 	 *
227 	 * This can happen in two cases:
228 	 * 1) TCP GRO packet, DF bit not set
229 	 * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
230 	 * from host network stack.
231 	 */
232 	features = netif_skb_features(skb);
233 	segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
234 	if (IS_ERR_OR_NULL(segs)) {
235 		kfree_skb(skb);
236 		return -ENOMEM;
237 	}
238 
239 	consume_skb(skb);
240 
241 	do {
242 		struct sk_buff *nskb = segs->next;
243 		int err;
244 
245 		segs->next = NULL;
246 		err = ip_fragment(segs, ip_finish_output2);
247 
248 		if (err && ret == 0)
249 			ret = err;
250 		segs = nskb;
251 	} while (segs);
252 
253 	return ret;
254 }
255 
256 static int ip_finish_output(struct sk_buff *skb)
257 {
258 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
259 	/* Policy lookup after SNAT yielded a new policy */
260 	if (skb_dst(skb)->xfrm != NULL) {
261 		IPCB(skb)->flags |= IPSKB_REROUTED;
262 		return dst_output(skb);
263 	}
264 #endif
265 	if (skb_is_gso(skb))
266 		return ip_finish_output_gso(skb);
267 
268 	if (skb->len > ip_skb_dst_mtu(skb))
269 		return ip_fragment(skb, ip_finish_output2);
270 
271 	return ip_finish_output2(skb);
272 }
273 
274 int ip_mc_output(struct sock *sk, struct sk_buff *skb)
275 {
276 	struct rtable *rt = skb_rtable(skb);
277 	struct net_device *dev = rt->dst.dev;
278 
279 	/*
280 	 *	If the indicated interface is up and running, send the packet.
281 	 */
282 	IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
283 
284 	skb->dev = dev;
285 	skb->protocol = htons(ETH_P_IP);
286 
287 	/*
288 	 *	Multicasts are looped back for other local users
289 	 */
290 
291 	if (rt->rt_flags&RTCF_MULTICAST) {
292 		if (sk_mc_loop(sk)
293 #ifdef CONFIG_IP_MROUTE
294 		/* Small optimization: do not loopback not local frames,
295 		   which returned after forwarding; they will be  dropped
296 		   by ip_mr_input in any case.
297 		   Note, that local frames are looped back to be delivered
298 		   to local recipients.
299 
300 		   This check is duplicated in ip_mr_input at the moment.
301 		 */
302 		    &&
303 		    ((rt->rt_flags & RTCF_LOCAL) ||
304 		     !(IPCB(skb)->flags & IPSKB_FORWARDED))
305 #endif
306 		   ) {
307 			struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
308 			if (newskb)
309 				NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
310 					newskb, NULL, newskb->dev,
311 					dev_loopback_xmit);
312 		}
313 
314 		/* Multicasts with ttl 0 must not go beyond the host */
315 
316 		if (ip_hdr(skb)->ttl == 0) {
317 			kfree_skb(skb);
318 			return 0;
319 		}
320 	}
321 
322 	if (rt->rt_flags&RTCF_BROADCAST) {
323 		struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
324 		if (newskb)
325 			NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
326 				NULL, newskb->dev, dev_loopback_xmit);
327 	}
328 
329 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
330 			    skb->dev, ip_finish_output,
331 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
332 }
333 
334 int ip_output(struct sock *sk, struct sk_buff *skb)
335 {
336 	struct net_device *dev = skb_dst(skb)->dev;
337 
338 	IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
339 
340 	skb->dev = dev;
341 	skb->protocol = htons(ETH_P_IP);
342 
343 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
344 			    ip_finish_output,
345 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
346 }
347 
348 /*
349  * copy saddr and daddr, possibly using 64bit load/stores
350  * Equivalent to :
351  *   iph->saddr = fl4->saddr;
352  *   iph->daddr = fl4->daddr;
353  */
354 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
355 {
356 	BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
357 		     offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
358 	memcpy(&iph->saddr, &fl4->saddr,
359 	       sizeof(fl4->saddr) + sizeof(fl4->daddr));
360 }
361 
362 /* Note: skb->sk can be different from sk, in case of tunnels */
363 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
364 {
365 	struct inet_sock *inet = inet_sk(sk);
366 	struct ip_options_rcu *inet_opt;
367 	struct flowi4 *fl4;
368 	struct rtable *rt;
369 	struct iphdr *iph;
370 	int res;
371 
372 	/* Skip all of this if the packet is already routed,
373 	 * f.e. by something like SCTP.
374 	 */
375 	rcu_read_lock();
376 	inet_opt = rcu_dereference(inet->inet_opt);
377 	fl4 = &fl->u.ip4;
378 	rt = skb_rtable(skb);
379 	if (rt != NULL)
380 		goto packet_routed;
381 
382 	/* Make sure we can route this packet. */
383 	rt = (struct rtable *)__sk_dst_check(sk, 0);
384 	if (rt == NULL) {
385 		__be32 daddr;
386 
387 		/* Use correct destination address if we have options. */
388 		daddr = inet->inet_daddr;
389 		if (inet_opt && inet_opt->opt.srr)
390 			daddr = inet_opt->opt.faddr;
391 
392 		/* If this fails, retransmit mechanism of transport layer will
393 		 * keep trying until route appears or the connection times
394 		 * itself out.
395 		 */
396 		rt = ip_route_output_ports(sock_net(sk), fl4, sk,
397 					   daddr, inet->inet_saddr,
398 					   inet->inet_dport,
399 					   inet->inet_sport,
400 					   sk->sk_protocol,
401 					   RT_CONN_FLAGS(sk),
402 					   sk->sk_bound_dev_if);
403 		if (IS_ERR(rt))
404 			goto no_route;
405 		sk_setup_caps(sk, &rt->dst);
406 	}
407 	skb_dst_set_noref(skb, &rt->dst);
408 
409 packet_routed:
410 	if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
411 		goto no_route;
412 
413 	/* OK, we know where to send it, allocate and build IP header. */
414 	skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
415 	skb_reset_network_header(skb);
416 	iph = ip_hdr(skb);
417 	*((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
418 	if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
419 		iph->frag_off = htons(IP_DF);
420 	else
421 		iph->frag_off = 0;
422 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
423 	iph->protocol = sk->sk_protocol;
424 	ip_copy_addrs(iph, fl4);
425 
426 	/* Transport layer set skb->h.foo itself. */
427 
428 	if (inet_opt && inet_opt->opt.optlen) {
429 		iph->ihl += inet_opt->opt.optlen >> 2;
430 		ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
431 	}
432 
433 	ip_select_ident_segs(skb, sk, skb_shinfo(skb)->gso_segs ?: 1);
434 
435 	/* TODO : should we use skb->sk here instead of sk ? */
436 	skb->priority = sk->sk_priority;
437 	skb->mark = sk->sk_mark;
438 
439 	res = ip_local_out(skb);
440 	rcu_read_unlock();
441 	return res;
442 
443 no_route:
444 	rcu_read_unlock();
445 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
446 	kfree_skb(skb);
447 	return -EHOSTUNREACH;
448 }
449 EXPORT_SYMBOL(ip_queue_xmit);
450 
451 
452 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
453 {
454 	to->pkt_type = from->pkt_type;
455 	to->priority = from->priority;
456 	to->protocol = from->protocol;
457 	skb_dst_drop(to);
458 	skb_dst_copy(to, from);
459 	to->dev = from->dev;
460 	to->mark = from->mark;
461 
462 	/* Copy the flags to each fragment. */
463 	IPCB(to)->flags = IPCB(from)->flags;
464 
465 #ifdef CONFIG_NET_SCHED
466 	to->tc_index = from->tc_index;
467 #endif
468 	nf_copy(to, from);
469 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
470 	to->ipvs_property = from->ipvs_property;
471 #endif
472 	skb_copy_secmark(to, from);
473 }
474 
475 /*
476  *	This IP datagram is too large to be sent in one piece.  Break it up into
477  *	smaller pieces (each of size equal to IP header plus
478  *	a block of the data of the original IP data part) that will yet fit in a
479  *	single device frame, and queue such a frame for sending.
480  */
481 
482 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
483 {
484 	struct iphdr *iph;
485 	int ptr;
486 	struct net_device *dev;
487 	struct sk_buff *skb2;
488 	unsigned int mtu, hlen, left, len, ll_rs;
489 	int offset;
490 	__be16 not_last_frag;
491 	struct rtable *rt = skb_rtable(skb);
492 	int err = 0;
493 
494 	dev = rt->dst.dev;
495 
496 	/*
497 	 *	Point into the IP datagram header.
498 	 */
499 
500 	iph = ip_hdr(skb);
501 
502 	mtu = ip_skb_dst_mtu(skb);
503 	if (unlikely(((iph->frag_off & htons(IP_DF)) && !skb->ignore_df) ||
504 		     (IPCB(skb)->frag_max_size &&
505 		      IPCB(skb)->frag_max_size > mtu))) {
506 		IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
507 		icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
508 			  htonl(mtu));
509 		kfree_skb(skb);
510 		return -EMSGSIZE;
511 	}
512 
513 	/*
514 	 *	Setup starting values.
515 	 */
516 
517 	hlen = iph->ihl * 4;
518 	mtu = mtu - hlen;	/* Size of data space */
519 #if IS_ENABLED(CONFIG_BRIDGE_NETFILTER)
520 	if (skb->nf_bridge)
521 		mtu -= nf_bridge_mtu_reduction(skb);
522 #endif
523 	IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
524 
525 	/* When frag_list is given, use it. First, check its validity:
526 	 * some transformers could create wrong frag_list or break existing
527 	 * one, it is not prohibited. In this case fall back to copying.
528 	 *
529 	 * LATER: this step can be merged to real generation of fragments,
530 	 * we can switch to copy when see the first bad fragment.
531 	 */
532 	if (skb_has_frag_list(skb)) {
533 		struct sk_buff *frag, *frag2;
534 		int first_len = skb_pagelen(skb);
535 
536 		if (first_len - hlen > mtu ||
537 		    ((first_len - hlen) & 7) ||
538 		    ip_is_fragment(iph) ||
539 		    skb_cloned(skb))
540 			goto slow_path;
541 
542 		skb_walk_frags(skb, frag) {
543 			/* Correct geometry. */
544 			if (frag->len > mtu ||
545 			    ((frag->len & 7) && frag->next) ||
546 			    skb_headroom(frag) < hlen)
547 				goto slow_path_clean;
548 
549 			/* Partially cloned skb? */
550 			if (skb_shared(frag))
551 				goto slow_path_clean;
552 
553 			BUG_ON(frag->sk);
554 			if (skb->sk) {
555 				frag->sk = skb->sk;
556 				frag->destructor = sock_wfree;
557 			}
558 			skb->truesize -= frag->truesize;
559 		}
560 
561 		/* Everything is OK. Generate! */
562 
563 		err = 0;
564 		offset = 0;
565 		frag = skb_shinfo(skb)->frag_list;
566 		skb_frag_list_init(skb);
567 		skb->data_len = first_len - skb_headlen(skb);
568 		skb->len = first_len;
569 		iph->tot_len = htons(first_len);
570 		iph->frag_off = htons(IP_MF);
571 		ip_send_check(iph);
572 
573 		for (;;) {
574 			/* Prepare header of the next frame,
575 			 * before previous one went down. */
576 			if (frag) {
577 				frag->ip_summed = CHECKSUM_NONE;
578 				skb_reset_transport_header(frag);
579 				__skb_push(frag, hlen);
580 				skb_reset_network_header(frag);
581 				memcpy(skb_network_header(frag), iph, hlen);
582 				iph = ip_hdr(frag);
583 				iph->tot_len = htons(frag->len);
584 				ip_copy_metadata(frag, skb);
585 				if (offset == 0)
586 					ip_options_fragment(frag);
587 				offset += skb->len - hlen;
588 				iph->frag_off = htons(offset>>3);
589 				if (frag->next != NULL)
590 					iph->frag_off |= htons(IP_MF);
591 				/* Ready, complete checksum */
592 				ip_send_check(iph);
593 			}
594 
595 			err = output(skb);
596 
597 			if (!err)
598 				IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
599 			if (err || !frag)
600 				break;
601 
602 			skb = frag;
603 			frag = skb->next;
604 			skb->next = NULL;
605 		}
606 
607 		if (err == 0) {
608 			IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
609 			return 0;
610 		}
611 
612 		while (frag) {
613 			skb = frag->next;
614 			kfree_skb(frag);
615 			frag = skb;
616 		}
617 		IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
618 		return err;
619 
620 slow_path_clean:
621 		skb_walk_frags(skb, frag2) {
622 			if (frag2 == frag)
623 				break;
624 			frag2->sk = NULL;
625 			frag2->destructor = NULL;
626 			skb->truesize += frag2->truesize;
627 		}
628 	}
629 
630 slow_path:
631 	/* for offloaded checksums cleanup checksum before fragmentation */
632 	if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
633 		goto fail;
634 	iph = ip_hdr(skb);
635 
636 	left = skb->len - hlen;		/* Space per frame */
637 	ptr = hlen;		/* Where to start from */
638 
639 	/* for bridged IP traffic encapsulated inside f.e. a vlan header,
640 	 * we need to make room for the encapsulating header
641 	 */
642 	ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
643 
644 	/*
645 	 *	Fragment the datagram.
646 	 */
647 
648 	offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
649 	not_last_frag = iph->frag_off & htons(IP_MF);
650 
651 	/*
652 	 *	Keep copying data until we run out.
653 	 */
654 
655 	while (left > 0) {
656 		len = left;
657 		/* IF: it doesn't fit, use 'mtu' - the data space left */
658 		if (len > mtu)
659 			len = mtu;
660 		/* IF: we are not sending up to and including the packet end
661 		   then align the next start on an eight byte boundary */
662 		if (len < left)	{
663 			len &= ~7;
664 		}
665 
666 		/* Allocate buffer */
667 		skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
668 		if (!skb2) {
669 			err = -ENOMEM;
670 			goto fail;
671 		}
672 
673 		/*
674 		 *	Set up data on packet
675 		 */
676 
677 		ip_copy_metadata(skb2, skb);
678 		skb_reserve(skb2, ll_rs);
679 		skb_put(skb2, len + hlen);
680 		skb_reset_network_header(skb2);
681 		skb2->transport_header = skb2->network_header + hlen;
682 
683 		/*
684 		 *	Charge the memory for the fragment to any owner
685 		 *	it might possess
686 		 */
687 
688 		if (skb->sk)
689 			skb_set_owner_w(skb2, skb->sk);
690 
691 		/*
692 		 *	Copy the packet header into the new buffer.
693 		 */
694 
695 		skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
696 
697 		/*
698 		 *	Copy a block of the IP datagram.
699 		 */
700 		if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
701 			BUG();
702 		left -= len;
703 
704 		/*
705 		 *	Fill in the new header fields.
706 		 */
707 		iph = ip_hdr(skb2);
708 		iph->frag_off = htons((offset >> 3));
709 
710 		/* ANK: dirty, but effective trick. Upgrade options only if
711 		 * the segment to be fragmented was THE FIRST (otherwise,
712 		 * options are already fixed) and make it ONCE
713 		 * on the initial skb, so that all the following fragments
714 		 * will inherit fixed options.
715 		 */
716 		if (offset == 0)
717 			ip_options_fragment(skb);
718 
719 		/*
720 		 *	Added AC : If we are fragmenting a fragment that's not the
721 		 *		   last fragment then keep MF on each bit
722 		 */
723 		if (left > 0 || not_last_frag)
724 			iph->frag_off |= htons(IP_MF);
725 		ptr += len;
726 		offset += len;
727 
728 		/*
729 		 *	Put this fragment into the sending queue.
730 		 */
731 		iph->tot_len = htons(len + hlen);
732 
733 		ip_send_check(iph);
734 
735 		err = output(skb2);
736 		if (err)
737 			goto fail;
738 
739 		IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
740 	}
741 	consume_skb(skb);
742 	IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
743 	return err;
744 
745 fail:
746 	kfree_skb(skb);
747 	IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
748 	return err;
749 }
750 EXPORT_SYMBOL(ip_fragment);
751 
752 int
753 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
754 {
755 	struct msghdr *msg = from;
756 
757 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
758 		/* XXX: stripping const */
759 		if (memcpy_fromiovecend(to, (struct iovec *)msg->msg_iter.iov, offset, len) < 0)
760 			return -EFAULT;
761 	} else {
762 		__wsum csum = 0;
763 		/* XXX: stripping const */
764 		if (csum_partial_copy_fromiovecend(to, (struct iovec *)msg->msg_iter.iov, offset, len, &csum) < 0)
765 			return -EFAULT;
766 		skb->csum = csum_block_add(skb->csum, csum, odd);
767 	}
768 	return 0;
769 }
770 EXPORT_SYMBOL(ip_generic_getfrag);
771 
772 static inline __wsum
773 csum_page(struct page *page, int offset, int copy)
774 {
775 	char *kaddr;
776 	__wsum csum;
777 	kaddr = kmap(page);
778 	csum = csum_partial(kaddr + offset, copy, 0);
779 	kunmap(page);
780 	return csum;
781 }
782 
783 static inline int ip_ufo_append_data(struct sock *sk,
784 			struct sk_buff_head *queue,
785 			int getfrag(void *from, char *to, int offset, int len,
786 			       int odd, struct sk_buff *skb),
787 			void *from, int length, int hh_len, int fragheaderlen,
788 			int transhdrlen, int maxfraglen, unsigned int flags)
789 {
790 	struct sk_buff *skb;
791 	int err;
792 
793 	/* There is support for UDP fragmentation offload by network
794 	 * device, so create one single skb packet containing complete
795 	 * udp datagram
796 	 */
797 	if ((skb = skb_peek_tail(queue)) == NULL) {
798 		skb = sock_alloc_send_skb(sk,
799 			hh_len + fragheaderlen + transhdrlen + 20,
800 			(flags & MSG_DONTWAIT), &err);
801 
802 		if (skb == NULL)
803 			return err;
804 
805 		/* reserve space for Hardware header */
806 		skb_reserve(skb, hh_len);
807 
808 		/* create space for UDP/IP header */
809 		skb_put(skb, fragheaderlen + transhdrlen);
810 
811 		/* initialize network header pointer */
812 		skb_reset_network_header(skb);
813 
814 		/* initialize protocol header pointer */
815 		skb->transport_header = skb->network_header + fragheaderlen;
816 
817 		skb->csum = 0;
818 
819 
820 		__skb_queue_tail(queue, skb);
821 	} else if (skb_is_gso(skb)) {
822 		goto append;
823 	}
824 
825 	skb->ip_summed = CHECKSUM_PARTIAL;
826 	/* specify the length of each IP datagram fragment */
827 	skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
828 	skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
829 
830 append:
831 	return skb_append_datato_frags(sk, skb, getfrag, from,
832 				       (length - transhdrlen));
833 }
834 
835 static int __ip_append_data(struct sock *sk,
836 			    struct flowi4 *fl4,
837 			    struct sk_buff_head *queue,
838 			    struct inet_cork *cork,
839 			    struct page_frag *pfrag,
840 			    int getfrag(void *from, char *to, int offset,
841 					int len, int odd, struct sk_buff *skb),
842 			    void *from, int length, int transhdrlen,
843 			    unsigned int flags)
844 {
845 	struct inet_sock *inet = inet_sk(sk);
846 	struct sk_buff *skb;
847 
848 	struct ip_options *opt = cork->opt;
849 	int hh_len;
850 	int exthdrlen;
851 	int mtu;
852 	int copy;
853 	int err;
854 	int offset = 0;
855 	unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
856 	int csummode = CHECKSUM_NONE;
857 	struct rtable *rt = (struct rtable *)cork->dst;
858 	u32 tskey = 0;
859 
860 	skb = skb_peek_tail(queue);
861 
862 	exthdrlen = !skb ? rt->dst.header_len : 0;
863 	mtu = cork->fragsize;
864 	if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
865 	    sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
866 		tskey = sk->sk_tskey++;
867 
868 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
869 
870 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
871 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
872 	maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
873 
874 	if (cork->length + length > maxnonfragsize - fragheaderlen) {
875 		ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
876 			       mtu - (opt ? opt->optlen : 0));
877 		return -EMSGSIZE;
878 	}
879 
880 	/*
881 	 * transhdrlen > 0 means that this is the first fragment and we wish
882 	 * it won't be fragmented in the future.
883 	 */
884 	if (transhdrlen &&
885 	    length + fragheaderlen <= mtu &&
886 	    rt->dst.dev->features & NETIF_F_V4_CSUM &&
887 	    !exthdrlen)
888 		csummode = CHECKSUM_PARTIAL;
889 
890 	cork->length += length;
891 	if (((length > mtu) || (skb && skb_is_gso(skb))) &&
892 	    (sk->sk_protocol == IPPROTO_UDP) &&
893 	    (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len) {
894 		err = ip_ufo_append_data(sk, queue, getfrag, from, length,
895 					 hh_len, fragheaderlen, transhdrlen,
896 					 maxfraglen, flags);
897 		if (err)
898 			goto error;
899 		return 0;
900 	}
901 
902 	/* So, what's going on in the loop below?
903 	 *
904 	 * We use calculated fragment length to generate chained skb,
905 	 * each of segments is IP fragment ready for sending to network after
906 	 * adding appropriate IP header.
907 	 */
908 
909 	if (!skb)
910 		goto alloc_new_skb;
911 
912 	while (length > 0) {
913 		/* Check if the remaining data fits into current packet. */
914 		copy = mtu - skb->len;
915 		if (copy < length)
916 			copy = maxfraglen - skb->len;
917 		if (copy <= 0) {
918 			char *data;
919 			unsigned int datalen;
920 			unsigned int fraglen;
921 			unsigned int fraggap;
922 			unsigned int alloclen;
923 			struct sk_buff *skb_prev;
924 alloc_new_skb:
925 			skb_prev = skb;
926 			if (skb_prev)
927 				fraggap = skb_prev->len - maxfraglen;
928 			else
929 				fraggap = 0;
930 
931 			/*
932 			 * If remaining data exceeds the mtu,
933 			 * we know we need more fragment(s).
934 			 */
935 			datalen = length + fraggap;
936 			if (datalen > mtu - fragheaderlen)
937 				datalen = maxfraglen - fragheaderlen;
938 			fraglen = datalen + fragheaderlen;
939 
940 			if ((flags & MSG_MORE) &&
941 			    !(rt->dst.dev->features&NETIF_F_SG))
942 				alloclen = mtu;
943 			else
944 				alloclen = fraglen;
945 
946 			alloclen += exthdrlen;
947 
948 			/* The last fragment gets additional space at tail.
949 			 * Note, with MSG_MORE we overallocate on fragments,
950 			 * because we have no idea what fragment will be
951 			 * the last.
952 			 */
953 			if (datalen == length + fraggap)
954 				alloclen += rt->dst.trailer_len;
955 
956 			if (transhdrlen) {
957 				skb = sock_alloc_send_skb(sk,
958 						alloclen + hh_len + 15,
959 						(flags & MSG_DONTWAIT), &err);
960 			} else {
961 				skb = NULL;
962 				if (atomic_read(&sk->sk_wmem_alloc) <=
963 				    2 * sk->sk_sndbuf)
964 					skb = sock_wmalloc(sk,
965 							   alloclen + hh_len + 15, 1,
966 							   sk->sk_allocation);
967 				if (unlikely(skb == NULL))
968 					err = -ENOBUFS;
969 			}
970 			if (skb == NULL)
971 				goto error;
972 
973 			/*
974 			 *	Fill in the control structures
975 			 */
976 			skb->ip_summed = csummode;
977 			skb->csum = 0;
978 			skb_reserve(skb, hh_len);
979 
980 			/* only the initial fragment is time stamped */
981 			skb_shinfo(skb)->tx_flags = cork->tx_flags;
982 			cork->tx_flags = 0;
983 			skb_shinfo(skb)->tskey = tskey;
984 			tskey = 0;
985 
986 			/*
987 			 *	Find where to start putting bytes.
988 			 */
989 			data = skb_put(skb, fraglen + exthdrlen);
990 			skb_set_network_header(skb, exthdrlen);
991 			skb->transport_header = (skb->network_header +
992 						 fragheaderlen);
993 			data += fragheaderlen + exthdrlen;
994 
995 			if (fraggap) {
996 				skb->csum = skb_copy_and_csum_bits(
997 					skb_prev, maxfraglen,
998 					data + transhdrlen, fraggap, 0);
999 				skb_prev->csum = csum_sub(skb_prev->csum,
1000 							  skb->csum);
1001 				data += fraggap;
1002 				pskb_trim_unique(skb_prev, maxfraglen);
1003 			}
1004 
1005 			copy = datalen - transhdrlen - fraggap;
1006 			if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1007 				err = -EFAULT;
1008 				kfree_skb(skb);
1009 				goto error;
1010 			}
1011 
1012 			offset += copy;
1013 			length -= datalen - fraggap;
1014 			transhdrlen = 0;
1015 			exthdrlen = 0;
1016 			csummode = CHECKSUM_NONE;
1017 
1018 			/*
1019 			 * Put the packet on the pending queue.
1020 			 */
1021 			__skb_queue_tail(queue, skb);
1022 			continue;
1023 		}
1024 
1025 		if (copy > length)
1026 			copy = length;
1027 
1028 		if (!(rt->dst.dev->features&NETIF_F_SG)) {
1029 			unsigned int off;
1030 
1031 			off = skb->len;
1032 			if (getfrag(from, skb_put(skb, copy),
1033 					offset, copy, off, skb) < 0) {
1034 				__skb_trim(skb, off);
1035 				err = -EFAULT;
1036 				goto error;
1037 			}
1038 		} else {
1039 			int i = skb_shinfo(skb)->nr_frags;
1040 
1041 			err = -ENOMEM;
1042 			if (!sk_page_frag_refill(sk, pfrag))
1043 				goto error;
1044 
1045 			if (!skb_can_coalesce(skb, i, pfrag->page,
1046 					      pfrag->offset)) {
1047 				err = -EMSGSIZE;
1048 				if (i == MAX_SKB_FRAGS)
1049 					goto error;
1050 
1051 				__skb_fill_page_desc(skb, i, pfrag->page,
1052 						     pfrag->offset, 0);
1053 				skb_shinfo(skb)->nr_frags = ++i;
1054 				get_page(pfrag->page);
1055 			}
1056 			copy = min_t(int, copy, pfrag->size - pfrag->offset);
1057 			if (getfrag(from,
1058 				    page_address(pfrag->page) + pfrag->offset,
1059 				    offset, copy, skb->len, skb) < 0)
1060 				goto error_efault;
1061 
1062 			pfrag->offset += copy;
1063 			skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1064 			skb->len += copy;
1065 			skb->data_len += copy;
1066 			skb->truesize += copy;
1067 			atomic_add(copy, &sk->sk_wmem_alloc);
1068 		}
1069 		offset += copy;
1070 		length -= copy;
1071 	}
1072 
1073 	return 0;
1074 
1075 error_efault:
1076 	err = -EFAULT;
1077 error:
1078 	cork->length -= length;
1079 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1080 	return err;
1081 }
1082 
1083 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1084 			 struct ipcm_cookie *ipc, struct rtable **rtp)
1085 {
1086 	struct ip_options_rcu *opt;
1087 	struct rtable *rt;
1088 
1089 	/*
1090 	 * setup for corking.
1091 	 */
1092 	opt = ipc->opt;
1093 	if (opt) {
1094 		if (cork->opt == NULL) {
1095 			cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1096 					    sk->sk_allocation);
1097 			if (unlikely(cork->opt == NULL))
1098 				return -ENOBUFS;
1099 		}
1100 		memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1101 		cork->flags |= IPCORK_OPT;
1102 		cork->addr = ipc->addr;
1103 	}
1104 	rt = *rtp;
1105 	if (unlikely(!rt))
1106 		return -EFAULT;
1107 	/*
1108 	 * We steal reference to this route, caller should not release it
1109 	 */
1110 	*rtp = NULL;
1111 	cork->fragsize = ip_sk_use_pmtu(sk) ?
1112 			 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1113 	cork->dst = &rt->dst;
1114 	cork->length = 0;
1115 	cork->ttl = ipc->ttl;
1116 	cork->tos = ipc->tos;
1117 	cork->priority = ipc->priority;
1118 	cork->tx_flags = ipc->tx_flags;
1119 
1120 	return 0;
1121 }
1122 
1123 /*
1124  *	ip_append_data() and ip_append_page() can make one large IP datagram
1125  *	from many pieces of data. Each pieces will be holded on the socket
1126  *	until ip_push_pending_frames() is called. Each piece can be a page
1127  *	or non-page data.
1128  *
1129  *	Not only UDP, other transport protocols - e.g. raw sockets - can use
1130  *	this interface potentially.
1131  *
1132  *	LATER: length must be adjusted by pad at tail, when it is required.
1133  */
1134 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1135 		   int getfrag(void *from, char *to, int offset, int len,
1136 			       int odd, struct sk_buff *skb),
1137 		   void *from, int length, int transhdrlen,
1138 		   struct ipcm_cookie *ipc, struct rtable **rtp,
1139 		   unsigned int flags)
1140 {
1141 	struct inet_sock *inet = inet_sk(sk);
1142 	int err;
1143 
1144 	if (flags&MSG_PROBE)
1145 		return 0;
1146 
1147 	if (skb_queue_empty(&sk->sk_write_queue)) {
1148 		err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1149 		if (err)
1150 			return err;
1151 	} else {
1152 		transhdrlen = 0;
1153 	}
1154 
1155 	return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1156 				sk_page_frag(sk), getfrag,
1157 				from, length, transhdrlen, flags);
1158 }
1159 
1160 ssize_t	ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1161 		       int offset, size_t size, int flags)
1162 {
1163 	struct inet_sock *inet = inet_sk(sk);
1164 	struct sk_buff *skb;
1165 	struct rtable *rt;
1166 	struct ip_options *opt = NULL;
1167 	struct inet_cork *cork;
1168 	int hh_len;
1169 	int mtu;
1170 	int len;
1171 	int err;
1172 	unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1173 
1174 	if (inet->hdrincl)
1175 		return -EPERM;
1176 
1177 	if (flags&MSG_PROBE)
1178 		return 0;
1179 
1180 	if (skb_queue_empty(&sk->sk_write_queue))
1181 		return -EINVAL;
1182 
1183 	cork = &inet->cork.base;
1184 	rt = (struct rtable *)cork->dst;
1185 	if (cork->flags & IPCORK_OPT)
1186 		opt = cork->opt;
1187 
1188 	if (!(rt->dst.dev->features&NETIF_F_SG))
1189 		return -EOPNOTSUPP;
1190 
1191 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1192 	mtu = cork->fragsize;
1193 
1194 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1195 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1196 	maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1197 
1198 	if (cork->length + size > maxnonfragsize - fragheaderlen) {
1199 		ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1200 			       mtu - (opt ? opt->optlen : 0));
1201 		return -EMSGSIZE;
1202 	}
1203 
1204 	if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1205 		return -EINVAL;
1206 
1207 	cork->length += size;
1208 	if ((size + skb->len > mtu) &&
1209 	    (sk->sk_protocol == IPPROTO_UDP) &&
1210 	    (rt->dst.dev->features & NETIF_F_UFO)) {
1211 		skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1212 		skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1213 	}
1214 
1215 
1216 	while (size > 0) {
1217 		int i;
1218 
1219 		if (skb_is_gso(skb))
1220 			len = size;
1221 		else {
1222 
1223 			/* Check if the remaining data fits into current packet. */
1224 			len = mtu - skb->len;
1225 			if (len < size)
1226 				len = maxfraglen - skb->len;
1227 		}
1228 		if (len <= 0) {
1229 			struct sk_buff *skb_prev;
1230 			int alloclen;
1231 
1232 			skb_prev = skb;
1233 			fraggap = skb_prev->len - maxfraglen;
1234 
1235 			alloclen = fragheaderlen + hh_len + fraggap + 15;
1236 			skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1237 			if (unlikely(!skb)) {
1238 				err = -ENOBUFS;
1239 				goto error;
1240 			}
1241 
1242 			/*
1243 			 *	Fill in the control structures
1244 			 */
1245 			skb->ip_summed = CHECKSUM_NONE;
1246 			skb->csum = 0;
1247 			skb_reserve(skb, hh_len);
1248 
1249 			/*
1250 			 *	Find where to start putting bytes.
1251 			 */
1252 			skb_put(skb, fragheaderlen + fraggap);
1253 			skb_reset_network_header(skb);
1254 			skb->transport_header = (skb->network_header +
1255 						 fragheaderlen);
1256 			if (fraggap) {
1257 				skb->csum = skb_copy_and_csum_bits(skb_prev,
1258 								   maxfraglen,
1259 						    skb_transport_header(skb),
1260 								   fraggap, 0);
1261 				skb_prev->csum = csum_sub(skb_prev->csum,
1262 							  skb->csum);
1263 				pskb_trim_unique(skb_prev, maxfraglen);
1264 			}
1265 
1266 			/*
1267 			 * Put the packet on the pending queue.
1268 			 */
1269 			__skb_queue_tail(&sk->sk_write_queue, skb);
1270 			continue;
1271 		}
1272 
1273 		i = skb_shinfo(skb)->nr_frags;
1274 		if (len > size)
1275 			len = size;
1276 		if (skb_can_coalesce(skb, i, page, offset)) {
1277 			skb_frag_size_add(&skb_shinfo(skb)->frags[i-1], len);
1278 		} else if (i < MAX_SKB_FRAGS) {
1279 			get_page(page);
1280 			skb_fill_page_desc(skb, i, page, offset, len);
1281 		} else {
1282 			err = -EMSGSIZE;
1283 			goto error;
1284 		}
1285 
1286 		if (skb->ip_summed == CHECKSUM_NONE) {
1287 			__wsum csum;
1288 			csum = csum_page(page, offset, len);
1289 			skb->csum = csum_block_add(skb->csum, csum, skb->len);
1290 		}
1291 
1292 		skb->len += len;
1293 		skb->data_len += len;
1294 		skb->truesize += len;
1295 		atomic_add(len, &sk->sk_wmem_alloc);
1296 		offset += len;
1297 		size -= len;
1298 	}
1299 	return 0;
1300 
1301 error:
1302 	cork->length -= size;
1303 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1304 	return err;
1305 }
1306 
1307 static void ip_cork_release(struct inet_cork *cork)
1308 {
1309 	cork->flags &= ~IPCORK_OPT;
1310 	kfree(cork->opt);
1311 	cork->opt = NULL;
1312 	dst_release(cork->dst);
1313 	cork->dst = NULL;
1314 }
1315 
1316 /*
1317  *	Combined all pending IP fragments on the socket as one IP datagram
1318  *	and push them out.
1319  */
1320 struct sk_buff *__ip_make_skb(struct sock *sk,
1321 			      struct flowi4 *fl4,
1322 			      struct sk_buff_head *queue,
1323 			      struct inet_cork *cork)
1324 {
1325 	struct sk_buff *skb, *tmp_skb;
1326 	struct sk_buff **tail_skb;
1327 	struct inet_sock *inet = inet_sk(sk);
1328 	struct net *net = sock_net(sk);
1329 	struct ip_options *opt = NULL;
1330 	struct rtable *rt = (struct rtable *)cork->dst;
1331 	struct iphdr *iph;
1332 	__be16 df = 0;
1333 	__u8 ttl;
1334 
1335 	if ((skb = __skb_dequeue(queue)) == NULL)
1336 		goto out;
1337 	tail_skb = &(skb_shinfo(skb)->frag_list);
1338 
1339 	/* move skb->data to ip header from ext header */
1340 	if (skb->data < skb_network_header(skb))
1341 		__skb_pull(skb, skb_network_offset(skb));
1342 	while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1343 		__skb_pull(tmp_skb, skb_network_header_len(skb));
1344 		*tail_skb = tmp_skb;
1345 		tail_skb = &(tmp_skb->next);
1346 		skb->len += tmp_skb->len;
1347 		skb->data_len += tmp_skb->len;
1348 		skb->truesize += tmp_skb->truesize;
1349 		tmp_skb->destructor = NULL;
1350 		tmp_skb->sk = NULL;
1351 	}
1352 
1353 	/* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1354 	 * to fragment the frame generated here. No matter, what transforms
1355 	 * how transforms change size of the packet, it will come out.
1356 	 */
1357 	skb->ignore_df = ip_sk_ignore_df(sk);
1358 
1359 	/* DF bit is set when we want to see DF on outgoing frames.
1360 	 * If ignore_df is set too, we still allow to fragment this frame
1361 	 * locally. */
1362 	if (inet->pmtudisc == IP_PMTUDISC_DO ||
1363 	    inet->pmtudisc == IP_PMTUDISC_PROBE ||
1364 	    (skb->len <= dst_mtu(&rt->dst) &&
1365 	     ip_dont_fragment(sk, &rt->dst)))
1366 		df = htons(IP_DF);
1367 
1368 	if (cork->flags & IPCORK_OPT)
1369 		opt = cork->opt;
1370 
1371 	if (cork->ttl != 0)
1372 		ttl = cork->ttl;
1373 	else if (rt->rt_type == RTN_MULTICAST)
1374 		ttl = inet->mc_ttl;
1375 	else
1376 		ttl = ip_select_ttl(inet, &rt->dst);
1377 
1378 	iph = ip_hdr(skb);
1379 	iph->version = 4;
1380 	iph->ihl = 5;
1381 	iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1382 	iph->frag_off = df;
1383 	iph->ttl = ttl;
1384 	iph->protocol = sk->sk_protocol;
1385 	ip_copy_addrs(iph, fl4);
1386 	ip_select_ident(skb, sk);
1387 
1388 	if (opt) {
1389 		iph->ihl += opt->optlen>>2;
1390 		ip_options_build(skb, opt, cork->addr, rt, 0);
1391 	}
1392 
1393 	skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1394 	skb->mark = sk->sk_mark;
1395 	/*
1396 	 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1397 	 * on dst refcount
1398 	 */
1399 	cork->dst = NULL;
1400 	skb_dst_set(skb, &rt->dst);
1401 
1402 	if (iph->protocol == IPPROTO_ICMP)
1403 		icmp_out_count(net, ((struct icmphdr *)
1404 			skb_transport_header(skb))->type);
1405 
1406 	ip_cork_release(cork);
1407 out:
1408 	return skb;
1409 }
1410 
1411 int ip_send_skb(struct net *net, struct sk_buff *skb)
1412 {
1413 	int err;
1414 
1415 	err = ip_local_out(skb);
1416 	if (err) {
1417 		if (err > 0)
1418 			err = net_xmit_errno(err);
1419 		if (err)
1420 			IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1421 	}
1422 
1423 	return err;
1424 }
1425 
1426 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1427 {
1428 	struct sk_buff *skb;
1429 
1430 	skb = ip_finish_skb(sk, fl4);
1431 	if (!skb)
1432 		return 0;
1433 
1434 	/* Netfilter gets whole the not fragmented skb. */
1435 	return ip_send_skb(sock_net(sk), skb);
1436 }
1437 
1438 /*
1439  *	Throw away all pending data on the socket.
1440  */
1441 static void __ip_flush_pending_frames(struct sock *sk,
1442 				      struct sk_buff_head *queue,
1443 				      struct inet_cork *cork)
1444 {
1445 	struct sk_buff *skb;
1446 
1447 	while ((skb = __skb_dequeue_tail(queue)) != NULL)
1448 		kfree_skb(skb);
1449 
1450 	ip_cork_release(cork);
1451 }
1452 
1453 void ip_flush_pending_frames(struct sock *sk)
1454 {
1455 	__ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1456 }
1457 
1458 struct sk_buff *ip_make_skb(struct sock *sk,
1459 			    struct flowi4 *fl4,
1460 			    int getfrag(void *from, char *to, int offset,
1461 					int len, int odd, struct sk_buff *skb),
1462 			    void *from, int length, int transhdrlen,
1463 			    struct ipcm_cookie *ipc, struct rtable **rtp,
1464 			    unsigned int flags)
1465 {
1466 	struct inet_cork cork;
1467 	struct sk_buff_head queue;
1468 	int err;
1469 
1470 	if (flags & MSG_PROBE)
1471 		return NULL;
1472 
1473 	__skb_queue_head_init(&queue);
1474 
1475 	cork.flags = 0;
1476 	cork.addr = 0;
1477 	cork.opt = NULL;
1478 	err = ip_setup_cork(sk, &cork, ipc, rtp);
1479 	if (err)
1480 		return ERR_PTR(err);
1481 
1482 	err = __ip_append_data(sk, fl4, &queue, &cork,
1483 			       &current->task_frag, getfrag,
1484 			       from, length, transhdrlen, flags);
1485 	if (err) {
1486 		__ip_flush_pending_frames(sk, &queue, &cork);
1487 		return ERR_PTR(err);
1488 	}
1489 
1490 	return __ip_make_skb(sk, fl4, &queue, &cork);
1491 }
1492 
1493 /*
1494  *	Fetch data from kernel space and fill in checksum if needed.
1495  */
1496 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1497 			      int len, int odd, struct sk_buff *skb)
1498 {
1499 	__wsum csum;
1500 
1501 	csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1502 	skb->csum = csum_block_add(skb->csum, csum, odd);
1503 	return 0;
1504 }
1505 
1506 /*
1507  *	Generic function to send a packet as reply to another packet.
1508  *	Used to send some TCP resets/acks so far.
1509  *
1510  *	Use a fake percpu inet socket to avoid false sharing and contention.
1511  */
1512 static DEFINE_PER_CPU(struct inet_sock, unicast_sock) = {
1513 	.sk = {
1514 		.__sk_common = {
1515 			.skc_refcnt = ATOMIC_INIT(1),
1516 		},
1517 		.sk_wmem_alloc	= ATOMIC_INIT(1),
1518 		.sk_allocation	= GFP_ATOMIC,
1519 		.sk_flags	= (1UL << SOCK_USE_WRITE_QUEUE),
1520 	},
1521 	.pmtudisc	= IP_PMTUDISC_WANT,
1522 	.uc_ttl		= -1,
1523 };
1524 
1525 void ip_send_unicast_reply(struct net *net, struct sk_buff *skb,
1526 			   const struct ip_options *sopt,
1527 			   __be32 daddr, __be32 saddr,
1528 			   const struct ip_reply_arg *arg,
1529 			   unsigned int len)
1530 {
1531 	struct ip_options_data replyopts;
1532 	struct ipcm_cookie ipc;
1533 	struct flowi4 fl4;
1534 	struct rtable *rt = skb_rtable(skb);
1535 	struct sk_buff *nskb;
1536 	struct sock *sk;
1537 	struct inet_sock *inet;
1538 	int err;
1539 
1540 	if (__ip_options_echo(&replyopts.opt.opt, skb, sopt))
1541 		return;
1542 
1543 	ipc.addr = daddr;
1544 	ipc.opt = NULL;
1545 	ipc.tx_flags = 0;
1546 	ipc.ttl = 0;
1547 	ipc.tos = -1;
1548 
1549 	if (replyopts.opt.opt.optlen) {
1550 		ipc.opt = &replyopts.opt;
1551 
1552 		if (replyopts.opt.opt.srr)
1553 			daddr = replyopts.opt.opt.faddr;
1554 	}
1555 
1556 	flowi4_init_output(&fl4, arg->bound_dev_if,
1557 			   IP4_REPLY_MARK(net, skb->mark),
1558 			   RT_TOS(arg->tos),
1559 			   RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1560 			   ip_reply_arg_flowi_flags(arg),
1561 			   daddr, saddr,
1562 			   tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1563 	security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1564 	rt = ip_route_output_key(net, &fl4);
1565 	if (IS_ERR(rt))
1566 		return;
1567 
1568 	inet = &get_cpu_var(unicast_sock);
1569 
1570 	inet->tos = arg->tos;
1571 	sk = &inet->sk;
1572 	sk->sk_priority = skb->priority;
1573 	sk->sk_protocol = ip_hdr(skb)->protocol;
1574 	sk->sk_bound_dev_if = arg->bound_dev_if;
1575 	sock_net_set(sk, net);
1576 	__skb_queue_head_init(&sk->sk_write_queue);
1577 	sk->sk_sndbuf = sysctl_wmem_default;
1578 	err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1579 			     len, 0, &ipc, &rt, MSG_DONTWAIT);
1580 	if (unlikely(err)) {
1581 		ip_flush_pending_frames(sk);
1582 		goto out;
1583 	}
1584 
1585 	nskb = skb_peek(&sk->sk_write_queue);
1586 	if (nskb) {
1587 		if (arg->csumoffset >= 0)
1588 			*((__sum16 *)skb_transport_header(nskb) +
1589 			  arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1590 								arg->csum));
1591 		nskb->ip_summed = CHECKSUM_NONE;
1592 		skb_orphan(nskb);
1593 		skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1594 		ip_push_pending_frames(sk, &fl4);
1595 	}
1596 out:
1597 	put_cpu_var(unicast_sock);
1598 
1599 	ip_rt_put(rt);
1600 }
1601 
1602 void __init ip_init(void)
1603 {
1604 	ip_rt_init();
1605 	inet_initpeers();
1606 
1607 #if defined(CONFIG_IP_MULTICAST)
1608 	igmp_mc_init();
1609 #endif
1610 }
1611