xref: /openbmc/linux/net/ipv4/ip_output.c (revision 2d96b44f)
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 <net/lwtunnel.h>
77 #include <linux/igmp.h>
78 #include <linux/netfilter_ipv4.h>
79 #include <linux/netfilter_bridge.h>
80 #include <linux/netlink.h>
81 #include <linux/tcp.h>
82 
83 static int
84 ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
85 	    unsigned int mtu,
86 	    int (*output)(struct net *, struct sock *, struct sk_buff *));
87 
88 /* Generate a checksum for an outgoing IP datagram. */
89 void ip_send_check(struct iphdr *iph)
90 {
91 	iph->check = 0;
92 	iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
93 }
94 EXPORT_SYMBOL(ip_send_check);
95 
96 int __ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
97 {
98 	struct iphdr *iph = ip_hdr(skb);
99 
100 	iph->tot_len = htons(skb->len);
101 	ip_send_check(iph);
102 
103 	/* if egress device is enslaved to an L3 master device pass the
104 	 * skb to its handler for processing
105 	 */
106 	skb = l3mdev_ip_out(sk, skb);
107 	if (unlikely(!skb))
108 		return 0;
109 
110 	return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
111 		       net, sk, skb, NULL, skb_dst(skb)->dev,
112 		       dst_output);
113 }
114 
115 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
116 {
117 	int err;
118 
119 	err = __ip_local_out(net, sk, skb);
120 	if (likely(err == 1))
121 		err = dst_output(net, sk, skb);
122 
123 	return err;
124 }
125 EXPORT_SYMBOL_GPL(ip_local_out);
126 
127 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
128 {
129 	int ttl = inet->uc_ttl;
130 
131 	if (ttl < 0)
132 		ttl = ip4_dst_hoplimit(dst);
133 	return ttl;
134 }
135 
136 /*
137  *		Add an ip header to a skbuff and send it out.
138  *
139  */
140 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
141 			  __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
142 {
143 	struct inet_sock *inet = inet_sk(sk);
144 	struct rtable *rt = skb_rtable(skb);
145 	struct net *net = sock_net(sk);
146 	struct iphdr *iph;
147 
148 	/* Build the IP header. */
149 	skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
150 	skb_reset_network_header(skb);
151 	iph = ip_hdr(skb);
152 	iph->version  = 4;
153 	iph->ihl      = 5;
154 	iph->tos      = inet->tos;
155 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
156 	iph->daddr    = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
157 	iph->saddr    = saddr;
158 	iph->protocol = sk->sk_protocol;
159 	if (ip_dont_fragment(sk, &rt->dst)) {
160 		iph->frag_off = htons(IP_DF);
161 		iph->id = 0;
162 	} else {
163 		iph->frag_off = 0;
164 		__ip_select_ident(net, iph, 1);
165 	}
166 
167 	if (opt && opt->opt.optlen) {
168 		iph->ihl += opt->opt.optlen>>2;
169 		ip_options_build(skb, &opt->opt, daddr, rt, 0);
170 	}
171 
172 	skb->priority = sk->sk_priority;
173 	skb->mark = sk->sk_mark;
174 
175 	/* Send it out. */
176 	return ip_local_out(net, skb->sk, skb);
177 }
178 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
179 
180 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
181 {
182 	struct dst_entry *dst = skb_dst(skb);
183 	struct rtable *rt = (struct rtable *)dst;
184 	struct net_device *dev = dst->dev;
185 	unsigned int hh_len = LL_RESERVED_SPACE(dev);
186 	struct neighbour *neigh;
187 	u32 nexthop;
188 
189 	if (rt->rt_type == RTN_MULTICAST) {
190 		IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
191 	} else if (rt->rt_type == RTN_BROADCAST)
192 		IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
193 
194 	/* Be paranoid, rather than too clever. */
195 	if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
196 		struct sk_buff *skb2;
197 
198 		skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
199 		if (!skb2) {
200 			kfree_skb(skb);
201 			return -ENOMEM;
202 		}
203 		if (skb->sk)
204 			skb_set_owner_w(skb2, skb->sk);
205 		consume_skb(skb);
206 		skb = skb2;
207 	}
208 
209 	if (lwtunnel_xmit_redirect(dst->lwtstate)) {
210 		int res = lwtunnel_xmit(skb);
211 
212 		if (res < 0 || res == LWTUNNEL_XMIT_DONE)
213 			return res;
214 	}
215 
216 	rcu_read_lock_bh();
217 	nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
218 	neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
219 	if (unlikely(!neigh))
220 		neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
221 	if (!IS_ERR(neigh)) {
222 		int res = dst_neigh_output(dst, neigh, skb);
223 
224 		rcu_read_unlock_bh();
225 		return res;
226 	}
227 	rcu_read_unlock_bh();
228 
229 	net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
230 			    __func__);
231 	kfree_skb(skb);
232 	return -EINVAL;
233 }
234 
235 static int ip_finish_output_gso(struct net *net, struct sock *sk,
236 				struct sk_buff *skb, unsigned int mtu)
237 {
238 	netdev_features_t features;
239 	struct sk_buff *segs;
240 	int ret = 0;
241 
242 	/* common case: fragmentation of segments is not allowed,
243 	 * or seglen is <= mtu
244 	 */
245 	if (((IPCB(skb)->flags & IPSKB_FRAG_SEGS) == 0) ||
246 	      skb_gso_validate_mtu(skb, mtu))
247 		return ip_finish_output2(net, sk, skb);
248 
249 	/* Slowpath -  GSO segment length is exceeding the dst MTU.
250 	 *
251 	 * This can happen in two cases:
252 	 * 1) TCP GRO packet, DF bit not set
253 	 * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
254 	 * from host network stack.
255 	 */
256 	features = netif_skb_features(skb);
257 	BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_SGO_CB_OFFSET);
258 	segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
259 	if (IS_ERR_OR_NULL(segs)) {
260 		kfree_skb(skb);
261 		return -ENOMEM;
262 	}
263 
264 	consume_skb(skb);
265 
266 	do {
267 		struct sk_buff *nskb = segs->next;
268 		int err;
269 
270 		segs->next = NULL;
271 		err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
272 
273 		if (err && ret == 0)
274 			ret = err;
275 		segs = nskb;
276 	} while (segs);
277 
278 	return ret;
279 }
280 
281 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
282 {
283 	unsigned int mtu;
284 
285 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
286 	/* Policy lookup after SNAT yielded a new policy */
287 	if (skb_dst(skb)->xfrm) {
288 		IPCB(skb)->flags |= IPSKB_REROUTED;
289 		return dst_output(net, sk, skb);
290 	}
291 #endif
292 	mtu = ip_skb_dst_mtu(sk, skb);
293 	if (skb_is_gso(skb))
294 		return ip_finish_output_gso(net, sk, skb, mtu);
295 
296 	if (skb->len > mtu || (IPCB(skb)->flags & IPSKB_FRAG_PMTU))
297 		return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
298 
299 	return ip_finish_output2(net, sk, skb);
300 }
301 
302 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
303 {
304 	struct rtable *rt = skb_rtable(skb);
305 	struct net_device *dev = rt->dst.dev;
306 
307 	/*
308 	 *	If the indicated interface is up and running, send the packet.
309 	 */
310 	IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
311 
312 	skb->dev = dev;
313 	skb->protocol = htons(ETH_P_IP);
314 
315 	/*
316 	 *	Multicasts are looped back for other local users
317 	 */
318 
319 	if (rt->rt_flags&RTCF_MULTICAST) {
320 		if (sk_mc_loop(sk)
321 #ifdef CONFIG_IP_MROUTE
322 		/* Small optimization: do not loopback not local frames,
323 		   which returned after forwarding; they will be  dropped
324 		   by ip_mr_input in any case.
325 		   Note, that local frames are looped back to be delivered
326 		   to local recipients.
327 
328 		   This check is duplicated in ip_mr_input at the moment.
329 		 */
330 		    &&
331 		    ((rt->rt_flags & RTCF_LOCAL) ||
332 		     !(IPCB(skb)->flags & IPSKB_FORWARDED))
333 #endif
334 		   ) {
335 			struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
336 			if (newskb)
337 				NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
338 					net, sk, newskb, NULL, newskb->dev,
339 					dev_loopback_xmit);
340 		}
341 
342 		/* Multicasts with ttl 0 must not go beyond the host */
343 
344 		if (ip_hdr(skb)->ttl == 0) {
345 			kfree_skb(skb);
346 			return 0;
347 		}
348 	}
349 
350 	if (rt->rt_flags&RTCF_BROADCAST) {
351 		struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
352 		if (newskb)
353 			NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
354 				net, sk, newskb, NULL, newskb->dev,
355 				dev_loopback_xmit);
356 	}
357 
358 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
359 			    net, sk, skb, NULL, skb->dev,
360 			    ip_finish_output,
361 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
362 }
363 
364 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
365 {
366 	struct net_device *dev = skb_dst(skb)->dev;
367 
368 	IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
369 
370 	skb->dev = dev;
371 	skb->protocol = htons(ETH_P_IP);
372 
373 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
374 			    net, sk, skb, NULL, dev,
375 			    ip_finish_output,
376 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
377 }
378 
379 /*
380  * copy saddr and daddr, possibly using 64bit load/stores
381  * Equivalent to :
382  *   iph->saddr = fl4->saddr;
383  *   iph->daddr = fl4->daddr;
384  */
385 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
386 {
387 	BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
388 		     offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
389 	memcpy(&iph->saddr, &fl4->saddr,
390 	       sizeof(fl4->saddr) + sizeof(fl4->daddr));
391 }
392 
393 /* Note: skb->sk can be different from sk, in case of tunnels */
394 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
395 {
396 	struct inet_sock *inet = inet_sk(sk);
397 	struct net *net = sock_net(sk);
398 	struct ip_options_rcu *inet_opt;
399 	struct flowi4 *fl4;
400 	struct rtable *rt;
401 	struct iphdr *iph;
402 	int res;
403 
404 	/* Skip all of this if the packet is already routed,
405 	 * f.e. by something like SCTP.
406 	 */
407 	rcu_read_lock();
408 	inet_opt = rcu_dereference(inet->inet_opt);
409 	fl4 = &fl->u.ip4;
410 	rt = skb_rtable(skb);
411 	if (rt)
412 		goto packet_routed;
413 
414 	/* Make sure we can route this packet. */
415 	rt = (struct rtable *)__sk_dst_check(sk, 0);
416 	if (!rt) {
417 		__be32 daddr;
418 
419 		/* Use correct destination address if we have options. */
420 		daddr = inet->inet_daddr;
421 		if (inet_opt && inet_opt->opt.srr)
422 			daddr = inet_opt->opt.faddr;
423 
424 		/* If this fails, retransmit mechanism of transport layer will
425 		 * keep trying until route appears or the connection times
426 		 * itself out.
427 		 */
428 		rt = ip_route_output_ports(net, fl4, sk,
429 					   daddr, inet->inet_saddr,
430 					   inet->inet_dport,
431 					   inet->inet_sport,
432 					   sk->sk_protocol,
433 					   RT_CONN_FLAGS(sk),
434 					   sk->sk_bound_dev_if);
435 		if (IS_ERR(rt))
436 			goto no_route;
437 		sk_setup_caps(sk, &rt->dst);
438 	}
439 	skb_dst_set_noref(skb, &rt->dst);
440 
441 packet_routed:
442 	if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
443 		goto no_route;
444 
445 	/* OK, we know where to send it, allocate and build IP header. */
446 	skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
447 	skb_reset_network_header(skb);
448 	iph = ip_hdr(skb);
449 	*((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
450 	if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
451 		iph->frag_off = htons(IP_DF);
452 	else
453 		iph->frag_off = 0;
454 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
455 	iph->protocol = sk->sk_protocol;
456 	ip_copy_addrs(iph, fl4);
457 
458 	/* Transport layer set skb->h.foo itself. */
459 
460 	if (inet_opt && inet_opt->opt.optlen) {
461 		iph->ihl += inet_opt->opt.optlen >> 2;
462 		ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
463 	}
464 
465 	ip_select_ident_segs(net, skb, sk,
466 			     skb_shinfo(skb)->gso_segs ?: 1);
467 
468 	/* TODO : should we use skb->sk here instead of sk ? */
469 	skb->priority = sk->sk_priority;
470 	skb->mark = sk->sk_mark;
471 
472 	res = ip_local_out(net, sk, skb);
473 	rcu_read_unlock();
474 	return res;
475 
476 no_route:
477 	rcu_read_unlock();
478 	IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
479 	kfree_skb(skb);
480 	return -EHOSTUNREACH;
481 }
482 EXPORT_SYMBOL(ip_queue_xmit);
483 
484 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
485 {
486 	to->pkt_type = from->pkt_type;
487 	to->priority = from->priority;
488 	to->protocol = from->protocol;
489 	skb_dst_drop(to);
490 	skb_dst_copy(to, from);
491 	to->dev = from->dev;
492 	to->mark = from->mark;
493 
494 	/* Copy the flags to each fragment. */
495 	IPCB(to)->flags = IPCB(from)->flags;
496 
497 #ifdef CONFIG_NET_SCHED
498 	to->tc_index = from->tc_index;
499 #endif
500 	nf_copy(to, from);
501 #if IS_ENABLED(CONFIG_IP_VS)
502 	to->ipvs_property = from->ipvs_property;
503 #endif
504 	skb_copy_secmark(to, from);
505 }
506 
507 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
508 		       unsigned int mtu,
509 		       int (*output)(struct net *, struct sock *, struct sk_buff *))
510 {
511 	struct iphdr *iph = ip_hdr(skb);
512 
513 	if ((iph->frag_off & htons(IP_DF)) == 0)
514 		return ip_do_fragment(net, sk, skb, output);
515 
516 	if (unlikely(!skb->ignore_df ||
517 		     (IPCB(skb)->frag_max_size &&
518 		      IPCB(skb)->frag_max_size > mtu))) {
519 		IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
520 		icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
521 			  htonl(mtu));
522 		kfree_skb(skb);
523 		return -EMSGSIZE;
524 	}
525 
526 	return ip_do_fragment(net, sk, skb, output);
527 }
528 
529 /*
530  *	This IP datagram is too large to be sent in one piece.  Break it up into
531  *	smaller pieces (each of size equal to IP header plus
532  *	a block of the data of the original IP data part) that will yet fit in a
533  *	single device frame, and queue such a frame for sending.
534  */
535 
536 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
537 		   int (*output)(struct net *, struct sock *, struct sk_buff *))
538 {
539 	struct iphdr *iph;
540 	int ptr;
541 	struct net_device *dev;
542 	struct sk_buff *skb2;
543 	unsigned int mtu, hlen, left, len, ll_rs;
544 	int offset;
545 	__be16 not_last_frag;
546 	struct rtable *rt = skb_rtable(skb);
547 	int err = 0;
548 
549 	dev = rt->dst.dev;
550 
551 	/* for offloaded checksums cleanup checksum before fragmentation */
552 	if (skb->ip_summed == CHECKSUM_PARTIAL &&
553 	    (err = skb_checksum_help(skb)))
554 		goto fail;
555 
556 	/*
557 	 *	Point into the IP datagram header.
558 	 */
559 
560 	iph = ip_hdr(skb);
561 
562 	mtu = ip_skb_dst_mtu(sk, skb);
563 	if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
564 		mtu = IPCB(skb)->frag_max_size;
565 
566 	/*
567 	 *	Setup starting values.
568 	 */
569 
570 	hlen = iph->ihl * 4;
571 	mtu = mtu - hlen;	/* Size of data space */
572 	IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
573 
574 	/* When frag_list is given, use it. First, check its validity:
575 	 * some transformers could create wrong frag_list or break existing
576 	 * one, it is not prohibited. In this case fall back to copying.
577 	 *
578 	 * LATER: this step can be merged to real generation of fragments,
579 	 * we can switch to copy when see the first bad fragment.
580 	 */
581 	if (skb_has_frag_list(skb)) {
582 		struct sk_buff *frag, *frag2;
583 		int first_len = skb_pagelen(skb);
584 
585 		if (first_len - hlen > mtu ||
586 		    ((first_len - hlen) & 7) ||
587 		    ip_is_fragment(iph) ||
588 		    skb_cloned(skb))
589 			goto slow_path;
590 
591 		skb_walk_frags(skb, frag) {
592 			/* Correct geometry. */
593 			if (frag->len > mtu ||
594 			    ((frag->len & 7) && frag->next) ||
595 			    skb_headroom(frag) < hlen)
596 				goto slow_path_clean;
597 
598 			/* Partially cloned skb? */
599 			if (skb_shared(frag))
600 				goto slow_path_clean;
601 
602 			BUG_ON(frag->sk);
603 			if (skb->sk) {
604 				frag->sk = skb->sk;
605 				frag->destructor = sock_wfree;
606 			}
607 			skb->truesize -= frag->truesize;
608 		}
609 
610 		/* Everything is OK. Generate! */
611 
612 		err = 0;
613 		offset = 0;
614 		frag = skb_shinfo(skb)->frag_list;
615 		skb_frag_list_init(skb);
616 		skb->data_len = first_len - skb_headlen(skb);
617 		skb->len = first_len;
618 		iph->tot_len = htons(first_len);
619 		iph->frag_off = htons(IP_MF);
620 		ip_send_check(iph);
621 
622 		for (;;) {
623 			/* Prepare header of the next frame,
624 			 * before previous one went down. */
625 			if (frag) {
626 				frag->ip_summed = CHECKSUM_NONE;
627 				skb_reset_transport_header(frag);
628 				__skb_push(frag, hlen);
629 				skb_reset_network_header(frag);
630 				memcpy(skb_network_header(frag), iph, hlen);
631 				iph = ip_hdr(frag);
632 				iph->tot_len = htons(frag->len);
633 				ip_copy_metadata(frag, skb);
634 				if (offset == 0)
635 					ip_options_fragment(frag);
636 				offset += skb->len - hlen;
637 				iph->frag_off = htons(offset>>3);
638 				if (frag->next)
639 					iph->frag_off |= htons(IP_MF);
640 				/* Ready, complete checksum */
641 				ip_send_check(iph);
642 			}
643 
644 			err = output(net, sk, skb);
645 
646 			if (!err)
647 				IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
648 			if (err || !frag)
649 				break;
650 
651 			skb = frag;
652 			frag = skb->next;
653 			skb->next = NULL;
654 		}
655 
656 		if (err == 0) {
657 			IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
658 			return 0;
659 		}
660 
661 		while (frag) {
662 			skb = frag->next;
663 			kfree_skb(frag);
664 			frag = skb;
665 		}
666 		IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
667 		return err;
668 
669 slow_path_clean:
670 		skb_walk_frags(skb, frag2) {
671 			if (frag2 == frag)
672 				break;
673 			frag2->sk = NULL;
674 			frag2->destructor = NULL;
675 			skb->truesize += frag2->truesize;
676 		}
677 	}
678 
679 slow_path:
680 	iph = ip_hdr(skb);
681 
682 	left = skb->len - hlen;		/* Space per frame */
683 	ptr = hlen;		/* Where to start from */
684 
685 	ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
686 
687 	/*
688 	 *	Fragment the datagram.
689 	 */
690 
691 	offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
692 	not_last_frag = iph->frag_off & htons(IP_MF);
693 
694 	/*
695 	 *	Keep copying data until we run out.
696 	 */
697 
698 	while (left > 0) {
699 		len = left;
700 		/* IF: it doesn't fit, use 'mtu' - the data space left */
701 		if (len > mtu)
702 			len = mtu;
703 		/* IF: we are not sending up to and including the packet end
704 		   then align the next start on an eight byte boundary */
705 		if (len < left)	{
706 			len &= ~7;
707 		}
708 
709 		/* Allocate buffer */
710 		skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
711 		if (!skb2) {
712 			err = -ENOMEM;
713 			goto fail;
714 		}
715 
716 		/*
717 		 *	Set up data on packet
718 		 */
719 
720 		ip_copy_metadata(skb2, skb);
721 		skb_reserve(skb2, ll_rs);
722 		skb_put(skb2, len + hlen);
723 		skb_reset_network_header(skb2);
724 		skb2->transport_header = skb2->network_header + hlen;
725 
726 		/*
727 		 *	Charge the memory for the fragment to any owner
728 		 *	it might possess
729 		 */
730 
731 		if (skb->sk)
732 			skb_set_owner_w(skb2, skb->sk);
733 
734 		/*
735 		 *	Copy the packet header into the new buffer.
736 		 */
737 
738 		skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
739 
740 		/*
741 		 *	Copy a block of the IP datagram.
742 		 */
743 		if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
744 			BUG();
745 		left -= len;
746 
747 		/*
748 		 *	Fill in the new header fields.
749 		 */
750 		iph = ip_hdr(skb2);
751 		iph->frag_off = htons((offset >> 3));
752 
753 		if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
754 			iph->frag_off |= htons(IP_DF);
755 
756 		/* ANK: dirty, but effective trick. Upgrade options only if
757 		 * the segment to be fragmented was THE FIRST (otherwise,
758 		 * options are already fixed) and make it ONCE
759 		 * on the initial skb, so that all the following fragments
760 		 * will inherit fixed options.
761 		 */
762 		if (offset == 0)
763 			ip_options_fragment(skb);
764 
765 		/*
766 		 *	Added AC : If we are fragmenting a fragment that's not the
767 		 *		   last fragment then keep MF on each bit
768 		 */
769 		if (left > 0 || not_last_frag)
770 			iph->frag_off |= htons(IP_MF);
771 		ptr += len;
772 		offset += len;
773 
774 		/*
775 		 *	Put this fragment into the sending queue.
776 		 */
777 		iph->tot_len = htons(len + hlen);
778 
779 		ip_send_check(iph);
780 
781 		err = output(net, sk, skb2);
782 		if (err)
783 			goto fail;
784 
785 		IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
786 	}
787 	consume_skb(skb);
788 	IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
789 	return err;
790 
791 fail:
792 	kfree_skb(skb);
793 	IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
794 	return err;
795 }
796 EXPORT_SYMBOL(ip_do_fragment);
797 
798 int
799 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
800 {
801 	struct msghdr *msg = from;
802 
803 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
804 		if (copy_from_iter(to, len, &msg->msg_iter) != len)
805 			return -EFAULT;
806 	} else {
807 		__wsum csum = 0;
808 		if (csum_and_copy_from_iter(to, len, &csum, &msg->msg_iter) != len)
809 			return -EFAULT;
810 		skb->csum = csum_block_add(skb->csum, csum, odd);
811 	}
812 	return 0;
813 }
814 EXPORT_SYMBOL(ip_generic_getfrag);
815 
816 static inline __wsum
817 csum_page(struct page *page, int offset, int copy)
818 {
819 	char *kaddr;
820 	__wsum csum;
821 	kaddr = kmap(page);
822 	csum = csum_partial(kaddr + offset, copy, 0);
823 	kunmap(page);
824 	return csum;
825 }
826 
827 static inline int ip_ufo_append_data(struct sock *sk,
828 			struct sk_buff_head *queue,
829 			int getfrag(void *from, char *to, int offset, int len,
830 			       int odd, struct sk_buff *skb),
831 			void *from, int length, int hh_len, int fragheaderlen,
832 			int transhdrlen, int maxfraglen, unsigned int flags)
833 {
834 	struct sk_buff *skb;
835 	int err;
836 
837 	/* There is support for UDP fragmentation offload by network
838 	 * device, so create one single skb packet containing complete
839 	 * udp datagram
840 	 */
841 	skb = skb_peek_tail(queue);
842 	if (!skb) {
843 		skb = sock_alloc_send_skb(sk,
844 			hh_len + fragheaderlen + transhdrlen + 20,
845 			(flags & MSG_DONTWAIT), &err);
846 
847 		if (!skb)
848 			return err;
849 
850 		/* reserve space for Hardware header */
851 		skb_reserve(skb, hh_len);
852 
853 		/* create space for UDP/IP header */
854 		skb_put(skb, fragheaderlen + transhdrlen);
855 
856 		/* initialize network header pointer */
857 		skb_reset_network_header(skb);
858 
859 		/* initialize protocol header pointer */
860 		skb->transport_header = skb->network_header + fragheaderlen;
861 
862 		skb->csum = 0;
863 
864 		__skb_queue_tail(queue, skb);
865 	} else if (skb_is_gso(skb)) {
866 		goto append;
867 	}
868 
869 	skb->ip_summed = CHECKSUM_PARTIAL;
870 	/* specify the length of each IP datagram fragment */
871 	skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
872 	skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
873 
874 append:
875 	return skb_append_datato_frags(sk, skb, getfrag, from,
876 				       (length - transhdrlen));
877 }
878 
879 static int __ip_append_data(struct sock *sk,
880 			    struct flowi4 *fl4,
881 			    struct sk_buff_head *queue,
882 			    struct inet_cork *cork,
883 			    struct page_frag *pfrag,
884 			    int getfrag(void *from, char *to, int offset,
885 					int len, int odd, struct sk_buff *skb),
886 			    void *from, int length, int transhdrlen,
887 			    unsigned int flags)
888 {
889 	struct inet_sock *inet = inet_sk(sk);
890 	struct sk_buff *skb;
891 
892 	struct ip_options *opt = cork->opt;
893 	int hh_len;
894 	int exthdrlen;
895 	int mtu;
896 	int copy;
897 	int err;
898 	int offset = 0;
899 	unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
900 	int csummode = CHECKSUM_NONE;
901 	struct rtable *rt = (struct rtable *)cork->dst;
902 	u32 tskey = 0;
903 
904 	skb = skb_peek_tail(queue);
905 
906 	exthdrlen = !skb ? rt->dst.header_len : 0;
907 	mtu = cork->fragsize;
908 	if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
909 	    sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
910 		tskey = sk->sk_tskey++;
911 
912 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
913 
914 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
915 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
916 	maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
917 
918 	if (cork->length + length > maxnonfragsize - fragheaderlen) {
919 		ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
920 			       mtu - (opt ? opt->optlen : 0));
921 		return -EMSGSIZE;
922 	}
923 
924 	/*
925 	 * transhdrlen > 0 means that this is the first fragment and we wish
926 	 * it won't be fragmented in the future.
927 	 */
928 	if (transhdrlen &&
929 	    length + fragheaderlen <= mtu &&
930 	    rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
931 	    !(flags & MSG_MORE) &&
932 	    !exthdrlen)
933 		csummode = CHECKSUM_PARTIAL;
934 
935 	cork->length += length;
936 	if (((length > mtu) || (skb && skb_is_gso(skb))) &&
937 	    (sk->sk_protocol == IPPROTO_UDP) &&
938 	    (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len &&
939 	    (sk->sk_type == SOCK_DGRAM) && !sk->sk_no_check_tx) {
940 		err = ip_ufo_append_data(sk, queue, getfrag, from, length,
941 					 hh_len, fragheaderlen, transhdrlen,
942 					 maxfraglen, flags);
943 		if (err)
944 			goto error;
945 		return 0;
946 	}
947 
948 	/* So, what's going on in the loop below?
949 	 *
950 	 * We use calculated fragment length to generate chained skb,
951 	 * each of segments is IP fragment ready for sending to network after
952 	 * adding appropriate IP header.
953 	 */
954 
955 	if (!skb)
956 		goto alloc_new_skb;
957 
958 	while (length > 0) {
959 		/* Check if the remaining data fits into current packet. */
960 		copy = mtu - skb->len;
961 		if (copy < length)
962 			copy = maxfraglen - skb->len;
963 		if (copy <= 0) {
964 			char *data;
965 			unsigned int datalen;
966 			unsigned int fraglen;
967 			unsigned int fraggap;
968 			unsigned int alloclen;
969 			struct sk_buff *skb_prev;
970 alloc_new_skb:
971 			skb_prev = skb;
972 			if (skb_prev)
973 				fraggap = skb_prev->len - maxfraglen;
974 			else
975 				fraggap = 0;
976 
977 			/*
978 			 * If remaining data exceeds the mtu,
979 			 * we know we need more fragment(s).
980 			 */
981 			datalen = length + fraggap;
982 			if (datalen > mtu - fragheaderlen)
983 				datalen = maxfraglen - fragheaderlen;
984 			fraglen = datalen + fragheaderlen;
985 
986 			if ((flags & MSG_MORE) &&
987 			    !(rt->dst.dev->features&NETIF_F_SG))
988 				alloclen = mtu;
989 			else
990 				alloclen = fraglen;
991 
992 			alloclen += exthdrlen;
993 
994 			/* The last fragment gets additional space at tail.
995 			 * Note, with MSG_MORE we overallocate on fragments,
996 			 * because we have no idea what fragment will be
997 			 * the last.
998 			 */
999 			if (datalen == length + fraggap)
1000 				alloclen += rt->dst.trailer_len;
1001 
1002 			if (transhdrlen) {
1003 				skb = sock_alloc_send_skb(sk,
1004 						alloclen + hh_len + 15,
1005 						(flags & MSG_DONTWAIT), &err);
1006 			} else {
1007 				skb = NULL;
1008 				if (atomic_read(&sk->sk_wmem_alloc) <=
1009 				    2 * sk->sk_sndbuf)
1010 					skb = sock_wmalloc(sk,
1011 							   alloclen + hh_len + 15, 1,
1012 							   sk->sk_allocation);
1013 				if (unlikely(!skb))
1014 					err = -ENOBUFS;
1015 			}
1016 			if (!skb)
1017 				goto error;
1018 
1019 			/*
1020 			 *	Fill in the control structures
1021 			 */
1022 			skb->ip_summed = csummode;
1023 			skb->csum = 0;
1024 			skb_reserve(skb, hh_len);
1025 
1026 			/* only the initial fragment is time stamped */
1027 			skb_shinfo(skb)->tx_flags = cork->tx_flags;
1028 			cork->tx_flags = 0;
1029 			skb_shinfo(skb)->tskey = tskey;
1030 			tskey = 0;
1031 
1032 			/*
1033 			 *	Find where to start putting bytes.
1034 			 */
1035 			data = skb_put(skb, fraglen + exthdrlen);
1036 			skb_set_network_header(skb, exthdrlen);
1037 			skb->transport_header = (skb->network_header +
1038 						 fragheaderlen);
1039 			data += fragheaderlen + exthdrlen;
1040 
1041 			if (fraggap) {
1042 				skb->csum = skb_copy_and_csum_bits(
1043 					skb_prev, maxfraglen,
1044 					data + transhdrlen, fraggap, 0);
1045 				skb_prev->csum = csum_sub(skb_prev->csum,
1046 							  skb->csum);
1047 				data += fraggap;
1048 				pskb_trim_unique(skb_prev, maxfraglen);
1049 			}
1050 
1051 			copy = datalen - transhdrlen - fraggap;
1052 			if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1053 				err = -EFAULT;
1054 				kfree_skb(skb);
1055 				goto error;
1056 			}
1057 
1058 			offset += copy;
1059 			length -= datalen - fraggap;
1060 			transhdrlen = 0;
1061 			exthdrlen = 0;
1062 			csummode = CHECKSUM_NONE;
1063 
1064 			/*
1065 			 * Put the packet on the pending queue.
1066 			 */
1067 			__skb_queue_tail(queue, skb);
1068 			continue;
1069 		}
1070 
1071 		if (copy > length)
1072 			copy = length;
1073 
1074 		if (!(rt->dst.dev->features&NETIF_F_SG)) {
1075 			unsigned int off;
1076 
1077 			off = skb->len;
1078 			if (getfrag(from, skb_put(skb, copy),
1079 					offset, copy, off, skb) < 0) {
1080 				__skb_trim(skb, off);
1081 				err = -EFAULT;
1082 				goto error;
1083 			}
1084 		} else {
1085 			int i = skb_shinfo(skb)->nr_frags;
1086 
1087 			err = -ENOMEM;
1088 			if (!sk_page_frag_refill(sk, pfrag))
1089 				goto error;
1090 
1091 			if (!skb_can_coalesce(skb, i, pfrag->page,
1092 					      pfrag->offset)) {
1093 				err = -EMSGSIZE;
1094 				if (i == MAX_SKB_FRAGS)
1095 					goto error;
1096 
1097 				__skb_fill_page_desc(skb, i, pfrag->page,
1098 						     pfrag->offset, 0);
1099 				skb_shinfo(skb)->nr_frags = ++i;
1100 				get_page(pfrag->page);
1101 			}
1102 			copy = min_t(int, copy, pfrag->size - pfrag->offset);
1103 			if (getfrag(from,
1104 				    page_address(pfrag->page) + pfrag->offset,
1105 				    offset, copy, skb->len, skb) < 0)
1106 				goto error_efault;
1107 
1108 			pfrag->offset += copy;
1109 			skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1110 			skb->len += copy;
1111 			skb->data_len += copy;
1112 			skb->truesize += copy;
1113 			atomic_add(copy, &sk->sk_wmem_alloc);
1114 		}
1115 		offset += copy;
1116 		length -= copy;
1117 	}
1118 
1119 	return 0;
1120 
1121 error_efault:
1122 	err = -EFAULT;
1123 error:
1124 	cork->length -= length;
1125 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1126 	return err;
1127 }
1128 
1129 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1130 			 struct ipcm_cookie *ipc, struct rtable **rtp)
1131 {
1132 	struct ip_options_rcu *opt;
1133 	struct rtable *rt;
1134 
1135 	/*
1136 	 * setup for corking.
1137 	 */
1138 	opt = ipc->opt;
1139 	if (opt) {
1140 		if (!cork->opt) {
1141 			cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1142 					    sk->sk_allocation);
1143 			if (unlikely(!cork->opt))
1144 				return -ENOBUFS;
1145 		}
1146 		memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1147 		cork->flags |= IPCORK_OPT;
1148 		cork->addr = ipc->addr;
1149 	}
1150 	rt = *rtp;
1151 	if (unlikely(!rt))
1152 		return -EFAULT;
1153 	/*
1154 	 * We steal reference to this route, caller should not release it
1155 	 */
1156 	*rtp = NULL;
1157 	cork->fragsize = ip_sk_use_pmtu(sk) ?
1158 			 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1159 	cork->dst = &rt->dst;
1160 	cork->length = 0;
1161 	cork->ttl = ipc->ttl;
1162 	cork->tos = ipc->tos;
1163 	cork->priority = ipc->priority;
1164 	cork->tx_flags = ipc->tx_flags;
1165 
1166 	return 0;
1167 }
1168 
1169 /*
1170  *	ip_append_data() and ip_append_page() can make one large IP datagram
1171  *	from many pieces of data. Each pieces will be holded on the socket
1172  *	until ip_push_pending_frames() is called. Each piece can be a page
1173  *	or non-page data.
1174  *
1175  *	Not only UDP, other transport protocols - e.g. raw sockets - can use
1176  *	this interface potentially.
1177  *
1178  *	LATER: length must be adjusted by pad at tail, when it is required.
1179  */
1180 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1181 		   int getfrag(void *from, char *to, int offset, int len,
1182 			       int odd, struct sk_buff *skb),
1183 		   void *from, int length, int transhdrlen,
1184 		   struct ipcm_cookie *ipc, struct rtable **rtp,
1185 		   unsigned int flags)
1186 {
1187 	struct inet_sock *inet = inet_sk(sk);
1188 	int err;
1189 
1190 	if (flags&MSG_PROBE)
1191 		return 0;
1192 
1193 	if (skb_queue_empty(&sk->sk_write_queue)) {
1194 		err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1195 		if (err)
1196 			return err;
1197 	} else {
1198 		transhdrlen = 0;
1199 	}
1200 
1201 	return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1202 				sk_page_frag(sk), getfrag,
1203 				from, length, transhdrlen, flags);
1204 }
1205 
1206 ssize_t	ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1207 		       int offset, size_t size, int flags)
1208 {
1209 	struct inet_sock *inet = inet_sk(sk);
1210 	struct sk_buff *skb;
1211 	struct rtable *rt;
1212 	struct ip_options *opt = NULL;
1213 	struct inet_cork *cork;
1214 	int hh_len;
1215 	int mtu;
1216 	int len;
1217 	int err;
1218 	unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1219 
1220 	if (inet->hdrincl)
1221 		return -EPERM;
1222 
1223 	if (flags&MSG_PROBE)
1224 		return 0;
1225 
1226 	if (skb_queue_empty(&sk->sk_write_queue))
1227 		return -EINVAL;
1228 
1229 	cork = &inet->cork.base;
1230 	rt = (struct rtable *)cork->dst;
1231 	if (cork->flags & IPCORK_OPT)
1232 		opt = cork->opt;
1233 
1234 	if (!(rt->dst.dev->features&NETIF_F_SG))
1235 		return -EOPNOTSUPP;
1236 
1237 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1238 	mtu = cork->fragsize;
1239 
1240 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1241 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1242 	maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1243 
1244 	if (cork->length + size > maxnonfragsize - fragheaderlen) {
1245 		ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1246 			       mtu - (opt ? opt->optlen : 0));
1247 		return -EMSGSIZE;
1248 	}
1249 
1250 	skb = skb_peek_tail(&sk->sk_write_queue);
1251 	if (!skb)
1252 		return -EINVAL;
1253 
1254 	if ((size + skb->len > mtu) &&
1255 	    (sk->sk_protocol == IPPROTO_UDP) &&
1256 	    (rt->dst.dev->features & NETIF_F_UFO)) {
1257 		if (skb->ip_summed != CHECKSUM_PARTIAL)
1258 			return -EOPNOTSUPP;
1259 
1260 		skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1261 		skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1262 	}
1263 	cork->length += size;
1264 
1265 	while (size > 0) {
1266 		if (skb_is_gso(skb)) {
1267 			len = size;
1268 		} else {
1269 
1270 			/* Check if the remaining data fits into current packet. */
1271 			len = mtu - skb->len;
1272 			if (len < size)
1273 				len = maxfraglen - skb->len;
1274 		}
1275 		if (len <= 0) {
1276 			struct sk_buff *skb_prev;
1277 			int alloclen;
1278 
1279 			skb_prev = skb;
1280 			fraggap = skb_prev->len - maxfraglen;
1281 
1282 			alloclen = fragheaderlen + hh_len + fraggap + 15;
1283 			skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1284 			if (unlikely(!skb)) {
1285 				err = -ENOBUFS;
1286 				goto error;
1287 			}
1288 
1289 			/*
1290 			 *	Fill in the control structures
1291 			 */
1292 			skb->ip_summed = CHECKSUM_NONE;
1293 			skb->csum = 0;
1294 			skb_reserve(skb, hh_len);
1295 
1296 			/*
1297 			 *	Find where to start putting bytes.
1298 			 */
1299 			skb_put(skb, fragheaderlen + fraggap);
1300 			skb_reset_network_header(skb);
1301 			skb->transport_header = (skb->network_header +
1302 						 fragheaderlen);
1303 			if (fraggap) {
1304 				skb->csum = skb_copy_and_csum_bits(skb_prev,
1305 								   maxfraglen,
1306 						    skb_transport_header(skb),
1307 								   fraggap, 0);
1308 				skb_prev->csum = csum_sub(skb_prev->csum,
1309 							  skb->csum);
1310 				pskb_trim_unique(skb_prev, maxfraglen);
1311 			}
1312 
1313 			/*
1314 			 * Put the packet on the pending queue.
1315 			 */
1316 			__skb_queue_tail(&sk->sk_write_queue, skb);
1317 			continue;
1318 		}
1319 
1320 		if (len > size)
1321 			len = size;
1322 
1323 		if (skb_append_pagefrags(skb, page, offset, len)) {
1324 			err = -EMSGSIZE;
1325 			goto error;
1326 		}
1327 
1328 		if (skb->ip_summed == CHECKSUM_NONE) {
1329 			__wsum csum;
1330 			csum = csum_page(page, offset, len);
1331 			skb->csum = csum_block_add(skb->csum, csum, skb->len);
1332 		}
1333 
1334 		skb->len += len;
1335 		skb->data_len += len;
1336 		skb->truesize += len;
1337 		atomic_add(len, &sk->sk_wmem_alloc);
1338 		offset += len;
1339 		size -= len;
1340 	}
1341 	return 0;
1342 
1343 error:
1344 	cork->length -= size;
1345 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1346 	return err;
1347 }
1348 
1349 static void ip_cork_release(struct inet_cork *cork)
1350 {
1351 	cork->flags &= ~IPCORK_OPT;
1352 	kfree(cork->opt);
1353 	cork->opt = NULL;
1354 	dst_release(cork->dst);
1355 	cork->dst = NULL;
1356 }
1357 
1358 /*
1359  *	Combined all pending IP fragments on the socket as one IP datagram
1360  *	and push them out.
1361  */
1362 struct sk_buff *__ip_make_skb(struct sock *sk,
1363 			      struct flowi4 *fl4,
1364 			      struct sk_buff_head *queue,
1365 			      struct inet_cork *cork)
1366 {
1367 	struct sk_buff *skb, *tmp_skb;
1368 	struct sk_buff **tail_skb;
1369 	struct inet_sock *inet = inet_sk(sk);
1370 	struct net *net = sock_net(sk);
1371 	struct ip_options *opt = NULL;
1372 	struct rtable *rt = (struct rtable *)cork->dst;
1373 	struct iphdr *iph;
1374 	__be16 df = 0;
1375 	__u8 ttl;
1376 
1377 	skb = __skb_dequeue(queue);
1378 	if (!skb)
1379 		goto out;
1380 	tail_skb = &(skb_shinfo(skb)->frag_list);
1381 
1382 	/* move skb->data to ip header from ext header */
1383 	if (skb->data < skb_network_header(skb))
1384 		__skb_pull(skb, skb_network_offset(skb));
1385 	while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1386 		__skb_pull(tmp_skb, skb_network_header_len(skb));
1387 		*tail_skb = tmp_skb;
1388 		tail_skb = &(tmp_skb->next);
1389 		skb->len += tmp_skb->len;
1390 		skb->data_len += tmp_skb->len;
1391 		skb->truesize += tmp_skb->truesize;
1392 		tmp_skb->destructor = NULL;
1393 		tmp_skb->sk = NULL;
1394 	}
1395 
1396 	/* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1397 	 * to fragment the frame generated here. No matter, what transforms
1398 	 * how transforms change size of the packet, it will come out.
1399 	 */
1400 	skb->ignore_df = ip_sk_ignore_df(sk);
1401 
1402 	/* DF bit is set when we want to see DF on outgoing frames.
1403 	 * If ignore_df is set too, we still allow to fragment this frame
1404 	 * locally. */
1405 	if (inet->pmtudisc == IP_PMTUDISC_DO ||
1406 	    inet->pmtudisc == IP_PMTUDISC_PROBE ||
1407 	    (skb->len <= dst_mtu(&rt->dst) &&
1408 	     ip_dont_fragment(sk, &rt->dst)))
1409 		df = htons(IP_DF);
1410 
1411 	if (cork->flags & IPCORK_OPT)
1412 		opt = cork->opt;
1413 
1414 	if (cork->ttl != 0)
1415 		ttl = cork->ttl;
1416 	else if (rt->rt_type == RTN_MULTICAST)
1417 		ttl = inet->mc_ttl;
1418 	else
1419 		ttl = ip_select_ttl(inet, &rt->dst);
1420 
1421 	iph = ip_hdr(skb);
1422 	iph->version = 4;
1423 	iph->ihl = 5;
1424 	iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1425 	iph->frag_off = df;
1426 	iph->ttl = ttl;
1427 	iph->protocol = sk->sk_protocol;
1428 	ip_copy_addrs(iph, fl4);
1429 	ip_select_ident(net, skb, sk);
1430 
1431 	if (opt) {
1432 		iph->ihl += opt->optlen>>2;
1433 		ip_options_build(skb, opt, cork->addr, rt, 0);
1434 	}
1435 
1436 	skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1437 	skb->mark = sk->sk_mark;
1438 	/*
1439 	 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1440 	 * on dst refcount
1441 	 */
1442 	cork->dst = NULL;
1443 	skb_dst_set(skb, &rt->dst);
1444 
1445 	if (iph->protocol == IPPROTO_ICMP)
1446 		icmp_out_count(net, ((struct icmphdr *)
1447 			skb_transport_header(skb))->type);
1448 
1449 	ip_cork_release(cork);
1450 out:
1451 	return skb;
1452 }
1453 
1454 int ip_send_skb(struct net *net, struct sk_buff *skb)
1455 {
1456 	int err;
1457 
1458 	err = ip_local_out(net, skb->sk, skb);
1459 	if (err) {
1460 		if (err > 0)
1461 			err = net_xmit_errno(err);
1462 		if (err)
1463 			IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1464 	}
1465 
1466 	return err;
1467 }
1468 
1469 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1470 {
1471 	struct sk_buff *skb;
1472 
1473 	skb = ip_finish_skb(sk, fl4);
1474 	if (!skb)
1475 		return 0;
1476 
1477 	/* Netfilter gets whole the not fragmented skb. */
1478 	return ip_send_skb(sock_net(sk), skb);
1479 }
1480 
1481 /*
1482  *	Throw away all pending data on the socket.
1483  */
1484 static void __ip_flush_pending_frames(struct sock *sk,
1485 				      struct sk_buff_head *queue,
1486 				      struct inet_cork *cork)
1487 {
1488 	struct sk_buff *skb;
1489 
1490 	while ((skb = __skb_dequeue_tail(queue)) != NULL)
1491 		kfree_skb(skb);
1492 
1493 	ip_cork_release(cork);
1494 }
1495 
1496 void ip_flush_pending_frames(struct sock *sk)
1497 {
1498 	__ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1499 }
1500 
1501 struct sk_buff *ip_make_skb(struct sock *sk,
1502 			    struct flowi4 *fl4,
1503 			    int getfrag(void *from, char *to, int offset,
1504 					int len, int odd, struct sk_buff *skb),
1505 			    void *from, int length, int transhdrlen,
1506 			    struct ipcm_cookie *ipc, struct rtable **rtp,
1507 			    unsigned int flags)
1508 {
1509 	struct inet_cork cork;
1510 	struct sk_buff_head queue;
1511 	int err;
1512 
1513 	if (flags & MSG_PROBE)
1514 		return NULL;
1515 
1516 	__skb_queue_head_init(&queue);
1517 
1518 	cork.flags = 0;
1519 	cork.addr = 0;
1520 	cork.opt = NULL;
1521 	err = ip_setup_cork(sk, &cork, ipc, rtp);
1522 	if (err)
1523 		return ERR_PTR(err);
1524 
1525 	err = __ip_append_data(sk, fl4, &queue, &cork,
1526 			       &current->task_frag, getfrag,
1527 			       from, length, transhdrlen, flags);
1528 	if (err) {
1529 		__ip_flush_pending_frames(sk, &queue, &cork);
1530 		return ERR_PTR(err);
1531 	}
1532 
1533 	return __ip_make_skb(sk, fl4, &queue, &cork);
1534 }
1535 
1536 /*
1537  *	Fetch data from kernel space and fill in checksum if needed.
1538  */
1539 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1540 			      int len, int odd, struct sk_buff *skb)
1541 {
1542 	__wsum csum;
1543 
1544 	csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1545 	skb->csum = csum_block_add(skb->csum, csum, odd);
1546 	return 0;
1547 }
1548 
1549 /*
1550  *	Generic function to send a packet as reply to another packet.
1551  *	Used to send some TCP resets/acks so far.
1552  */
1553 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1554 			   const struct ip_options *sopt,
1555 			   __be32 daddr, __be32 saddr,
1556 			   const struct ip_reply_arg *arg,
1557 			   unsigned int len)
1558 {
1559 	struct ip_options_data replyopts;
1560 	struct ipcm_cookie ipc;
1561 	struct flowi4 fl4;
1562 	struct rtable *rt = skb_rtable(skb);
1563 	struct net *net = sock_net(sk);
1564 	struct sk_buff *nskb;
1565 	int err;
1566 	int oif;
1567 
1568 	if (__ip_options_echo(&replyopts.opt.opt, skb, sopt))
1569 		return;
1570 
1571 	ipc.addr = daddr;
1572 	ipc.opt = NULL;
1573 	ipc.tx_flags = 0;
1574 	ipc.ttl = 0;
1575 	ipc.tos = -1;
1576 
1577 	if (replyopts.opt.opt.optlen) {
1578 		ipc.opt = &replyopts.opt;
1579 
1580 		if (replyopts.opt.opt.srr)
1581 			daddr = replyopts.opt.opt.faddr;
1582 	}
1583 
1584 	oif = arg->bound_dev_if;
1585 	oif = oif ? : skb->skb_iif;
1586 
1587 	flowi4_init_output(&fl4, oif,
1588 			   IP4_REPLY_MARK(net, skb->mark),
1589 			   RT_TOS(arg->tos),
1590 			   RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1591 			   ip_reply_arg_flowi_flags(arg),
1592 			   daddr, saddr,
1593 			   tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1594 	security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1595 	rt = ip_route_output_key(net, &fl4);
1596 	if (IS_ERR(rt))
1597 		return;
1598 
1599 	inet_sk(sk)->tos = arg->tos;
1600 
1601 	sk->sk_priority = skb->priority;
1602 	sk->sk_protocol = ip_hdr(skb)->protocol;
1603 	sk->sk_bound_dev_if = arg->bound_dev_if;
1604 	sk->sk_sndbuf = sysctl_wmem_default;
1605 	err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1606 			     len, 0, &ipc, &rt, MSG_DONTWAIT);
1607 	if (unlikely(err)) {
1608 		ip_flush_pending_frames(sk);
1609 		goto out;
1610 	}
1611 
1612 	nskb = skb_peek(&sk->sk_write_queue);
1613 	if (nskb) {
1614 		if (arg->csumoffset >= 0)
1615 			*((__sum16 *)skb_transport_header(nskb) +
1616 			  arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1617 								arg->csum));
1618 		nskb->ip_summed = CHECKSUM_NONE;
1619 		ip_push_pending_frames(sk, &fl4);
1620 	}
1621 out:
1622 	ip_rt_put(rt);
1623 }
1624 
1625 void __init ip_init(void)
1626 {
1627 	ip_rt_init();
1628 	inet_initpeers();
1629 
1630 #if defined(CONFIG_IP_MULTICAST)
1631 	igmp_mc_init();
1632 #endif
1633 }
1634