xref: /openbmc/linux/net/ipv4/ip_output.c (revision 161f4089)
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(struct sk_buff *skb)
105 {
106 	int err;
107 
108 	err = __ip_local_out(skb);
109 	if (likely(err == 1))
110 		err = dst_output(skb);
111 
112 	return err;
113 }
114 EXPORT_SYMBOL_GPL(ip_local_out);
115 
116 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
117 {
118 	int ttl = inet->uc_ttl;
119 
120 	if (ttl < 0)
121 		ttl = ip4_dst_hoplimit(dst);
122 	return ttl;
123 }
124 
125 /*
126  *		Add an ip header to a skbuff and send it out.
127  *
128  */
129 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
130 			  __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
131 {
132 	struct inet_sock *inet = inet_sk(sk);
133 	struct rtable *rt = skb_rtable(skb);
134 	struct iphdr *iph;
135 
136 	/* Build the IP header. */
137 	skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
138 	skb_reset_network_header(skb);
139 	iph = ip_hdr(skb);
140 	iph->version  = 4;
141 	iph->ihl      = 5;
142 	iph->tos      = inet->tos;
143 	if (ip_dont_fragment(sk, &rt->dst))
144 		iph->frag_off = htons(IP_DF);
145 	else
146 		iph->frag_off = 0;
147 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
148 	iph->daddr    = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
149 	iph->saddr    = saddr;
150 	iph->protocol = sk->sk_protocol;
151 	ip_select_ident(skb, &rt->dst, sk);
152 
153 	if (opt && opt->opt.optlen) {
154 		iph->ihl += opt->opt.optlen>>2;
155 		ip_options_build(skb, &opt->opt, daddr, rt, 0);
156 	}
157 
158 	skb->priority = sk->sk_priority;
159 	skb->mark = sk->sk_mark;
160 
161 	/* Send it out. */
162 	return ip_local_out(skb);
163 }
164 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
165 
166 static inline int ip_finish_output2(struct sk_buff *skb)
167 {
168 	struct dst_entry *dst = skb_dst(skb);
169 	struct rtable *rt = (struct rtable *)dst;
170 	struct net_device *dev = dst->dev;
171 	unsigned int hh_len = LL_RESERVED_SPACE(dev);
172 	struct neighbour *neigh;
173 	u32 nexthop;
174 
175 	if (rt->rt_type == RTN_MULTICAST) {
176 		IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
177 	} else if (rt->rt_type == RTN_BROADCAST)
178 		IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
179 
180 	/* Be paranoid, rather than too clever. */
181 	if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
182 		struct sk_buff *skb2;
183 
184 		skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
185 		if (skb2 == NULL) {
186 			kfree_skb(skb);
187 			return -ENOMEM;
188 		}
189 		if (skb->sk)
190 			skb_set_owner_w(skb2, skb->sk);
191 		consume_skb(skb);
192 		skb = skb2;
193 	}
194 
195 	rcu_read_lock_bh();
196 	nexthop = (__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(struct sk_buff *skb)
215 {
216 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
217 	/* Policy lookup after SNAT yielded a new policy */
218 	if (skb_dst(skb)->xfrm != NULL) {
219 		IPCB(skb)->flags |= IPSKB_REROUTED;
220 		return dst_output(skb);
221 	}
222 #endif
223 	if (skb->len > ip_skb_dst_mtu(skb) && !skb_is_gso(skb))
224 		return ip_fragment(skb, ip_finish_output2);
225 	else
226 		return ip_finish_output2(skb);
227 }
228 
229 int ip_mc_output(struct sk_buff *skb)
230 {
231 	struct sock *sk = skb->sk;
232 	struct rtable *rt = skb_rtable(skb);
233 	struct net_device *dev = rt->dst.dev;
234 
235 	/*
236 	 *	If the indicated interface is up and running, send the packet.
237 	 */
238 	IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
239 
240 	skb->dev = dev;
241 	skb->protocol = htons(ETH_P_IP);
242 
243 	/*
244 	 *	Multicasts are looped back for other local users
245 	 */
246 
247 	if (rt->rt_flags&RTCF_MULTICAST) {
248 		if (sk_mc_loop(sk)
249 #ifdef CONFIG_IP_MROUTE
250 		/* Small optimization: do not loopback not local frames,
251 		   which returned after forwarding; they will be  dropped
252 		   by ip_mr_input in any case.
253 		   Note, that local frames are looped back to be delivered
254 		   to local recipients.
255 
256 		   This check is duplicated in ip_mr_input at the moment.
257 		 */
258 		    &&
259 		    ((rt->rt_flags & RTCF_LOCAL) ||
260 		     !(IPCB(skb)->flags & IPSKB_FORWARDED))
261 #endif
262 		   ) {
263 			struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
264 			if (newskb)
265 				NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
266 					newskb, NULL, newskb->dev,
267 					dev_loopback_xmit);
268 		}
269 
270 		/* Multicasts with ttl 0 must not go beyond the host */
271 
272 		if (ip_hdr(skb)->ttl == 0) {
273 			kfree_skb(skb);
274 			return 0;
275 		}
276 	}
277 
278 	if (rt->rt_flags&RTCF_BROADCAST) {
279 		struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
280 		if (newskb)
281 			NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
282 				NULL, newskb->dev, dev_loopback_xmit);
283 	}
284 
285 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
286 			    skb->dev, ip_finish_output,
287 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
288 }
289 
290 int ip_output(struct sk_buff *skb)
291 {
292 	struct net_device *dev = skb_dst(skb)->dev;
293 
294 	IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
295 
296 	skb->dev = dev;
297 	skb->protocol = htons(ETH_P_IP);
298 
299 	return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
300 			    ip_finish_output,
301 			    !(IPCB(skb)->flags & IPSKB_REROUTED));
302 }
303 
304 /*
305  * copy saddr and daddr, possibly using 64bit load/stores
306  * Equivalent to :
307  *   iph->saddr = fl4->saddr;
308  *   iph->daddr = fl4->daddr;
309  */
310 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
311 {
312 	BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
313 		     offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
314 	memcpy(&iph->saddr, &fl4->saddr,
315 	       sizeof(fl4->saddr) + sizeof(fl4->daddr));
316 }
317 
318 int ip_queue_xmit(struct sk_buff *skb, struct flowi *fl)
319 {
320 	struct sock *sk = skb->sk;
321 	struct inet_sock *inet = inet_sk(sk);
322 	struct ip_options_rcu *inet_opt;
323 	struct flowi4 *fl4;
324 	struct rtable *rt;
325 	struct iphdr *iph;
326 	int res;
327 
328 	/* Skip all of this if the packet is already routed,
329 	 * f.e. by something like SCTP.
330 	 */
331 	rcu_read_lock();
332 	inet_opt = rcu_dereference(inet->inet_opt);
333 	fl4 = &fl->u.ip4;
334 	rt = skb_rtable(skb);
335 	if (rt != NULL)
336 		goto packet_routed;
337 
338 	/* Make sure we can route this packet. */
339 	rt = (struct rtable *)__sk_dst_check(sk, 0);
340 	if (rt == NULL) {
341 		__be32 daddr;
342 
343 		/* Use correct destination address if we have options. */
344 		daddr = inet->inet_daddr;
345 		if (inet_opt && inet_opt->opt.srr)
346 			daddr = inet_opt->opt.faddr;
347 
348 		/* If this fails, retransmit mechanism of transport layer will
349 		 * keep trying until route appears or the connection times
350 		 * itself out.
351 		 */
352 		rt = ip_route_output_ports(sock_net(sk), fl4, sk,
353 					   daddr, inet->inet_saddr,
354 					   inet->inet_dport,
355 					   inet->inet_sport,
356 					   sk->sk_protocol,
357 					   RT_CONN_FLAGS(sk),
358 					   sk->sk_bound_dev_if);
359 		if (IS_ERR(rt))
360 			goto no_route;
361 		sk_setup_caps(sk, &rt->dst);
362 	}
363 	skb_dst_set_noref(skb, &rt->dst);
364 
365 packet_routed:
366 	if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
367 		goto no_route;
368 
369 	/* OK, we know where to send it, allocate and build IP header. */
370 	skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
371 	skb_reset_network_header(skb);
372 	iph = ip_hdr(skb);
373 	*((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
374 	if (ip_dont_fragment(sk, &rt->dst) && !skb->local_df)
375 		iph->frag_off = htons(IP_DF);
376 	else
377 		iph->frag_off = 0;
378 	iph->ttl      = ip_select_ttl(inet, &rt->dst);
379 	iph->protocol = sk->sk_protocol;
380 	ip_copy_addrs(iph, fl4);
381 
382 	/* Transport layer set skb->h.foo itself. */
383 
384 	if (inet_opt && inet_opt->opt.optlen) {
385 		iph->ihl += inet_opt->opt.optlen >> 2;
386 		ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
387 	}
388 
389 	ip_select_ident_more(skb, &rt->dst, sk,
390 			     (skb_shinfo(skb)->gso_segs ?: 1) - 1);
391 
392 	skb->priority = sk->sk_priority;
393 	skb->mark = sk->sk_mark;
394 
395 	res = ip_local_out(skb);
396 	rcu_read_unlock();
397 	return res;
398 
399 no_route:
400 	rcu_read_unlock();
401 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
402 	kfree_skb(skb);
403 	return -EHOSTUNREACH;
404 }
405 EXPORT_SYMBOL(ip_queue_xmit);
406 
407 
408 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
409 {
410 	to->pkt_type = from->pkt_type;
411 	to->priority = from->priority;
412 	to->protocol = from->protocol;
413 	skb_dst_drop(to);
414 	skb_dst_copy(to, from);
415 	to->dev = from->dev;
416 	to->mark = from->mark;
417 
418 	/* Copy the flags to each fragment. */
419 	IPCB(to)->flags = IPCB(from)->flags;
420 
421 #ifdef CONFIG_NET_SCHED
422 	to->tc_index = from->tc_index;
423 #endif
424 	nf_copy(to, from);
425 #if IS_ENABLED(CONFIG_NETFILTER_XT_TARGET_TRACE)
426 	to->nf_trace = from->nf_trace;
427 #endif
428 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
429 	to->ipvs_property = from->ipvs_property;
430 #endif
431 	skb_copy_secmark(to, from);
432 }
433 
434 /*
435  *	This IP datagram is too large to be sent in one piece.  Break it up into
436  *	smaller pieces (each of size equal to IP header plus
437  *	a block of the data of the original IP data part) that will yet fit in a
438  *	single device frame, and queue such a frame for sending.
439  */
440 
441 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
442 {
443 	struct iphdr *iph;
444 	int ptr;
445 	struct net_device *dev;
446 	struct sk_buff *skb2;
447 	unsigned int mtu, hlen, left, len, ll_rs;
448 	int offset;
449 	__be16 not_last_frag;
450 	struct rtable *rt = skb_rtable(skb);
451 	int err = 0;
452 
453 	dev = rt->dst.dev;
454 
455 	/*
456 	 *	Point into the IP datagram header.
457 	 */
458 
459 	iph = ip_hdr(skb);
460 
461 	if (unlikely(((iph->frag_off & htons(IP_DF)) && !skb->local_df) ||
462 		     (IPCB(skb)->frag_max_size &&
463 		      IPCB(skb)->frag_max_size > dst_mtu(&rt->dst)))) {
464 		IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
465 		icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
466 			  htonl(ip_skb_dst_mtu(skb)));
467 		kfree_skb(skb);
468 		return -EMSGSIZE;
469 	}
470 
471 	/*
472 	 *	Setup starting values.
473 	 */
474 
475 	hlen = iph->ihl * 4;
476 	mtu = dst_mtu(&rt->dst) - hlen;	/* Size of data space */
477 #ifdef CONFIG_BRIDGE_NETFILTER
478 	if (skb->nf_bridge)
479 		mtu -= nf_bridge_mtu_reduction(skb);
480 #endif
481 	IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
482 
483 	/* When frag_list is given, use it. First, check its validity:
484 	 * some transformers could create wrong frag_list or break existing
485 	 * one, it is not prohibited. In this case fall back to copying.
486 	 *
487 	 * LATER: this step can be merged to real generation of fragments,
488 	 * we can switch to copy when see the first bad fragment.
489 	 */
490 	if (skb_has_frag_list(skb)) {
491 		struct sk_buff *frag, *frag2;
492 		int first_len = skb_pagelen(skb);
493 
494 		if (first_len - hlen > mtu ||
495 		    ((first_len - hlen) & 7) ||
496 		    ip_is_fragment(iph) ||
497 		    skb_cloned(skb))
498 			goto slow_path;
499 
500 		skb_walk_frags(skb, frag) {
501 			/* Correct geometry. */
502 			if (frag->len > mtu ||
503 			    ((frag->len & 7) && frag->next) ||
504 			    skb_headroom(frag) < hlen)
505 				goto slow_path_clean;
506 
507 			/* Partially cloned skb? */
508 			if (skb_shared(frag))
509 				goto slow_path_clean;
510 
511 			BUG_ON(frag->sk);
512 			if (skb->sk) {
513 				frag->sk = skb->sk;
514 				frag->destructor = sock_wfree;
515 			}
516 			skb->truesize -= frag->truesize;
517 		}
518 
519 		/* Everything is OK. Generate! */
520 
521 		err = 0;
522 		offset = 0;
523 		frag = skb_shinfo(skb)->frag_list;
524 		skb_frag_list_init(skb);
525 		skb->data_len = first_len - skb_headlen(skb);
526 		skb->len = first_len;
527 		iph->tot_len = htons(first_len);
528 		iph->frag_off = htons(IP_MF);
529 		ip_send_check(iph);
530 
531 		for (;;) {
532 			/* Prepare header of the next frame,
533 			 * before previous one went down. */
534 			if (frag) {
535 				frag->ip_summed = CHECKSUM_NONE;
536 				skb_reset_transport_header(frag);
537 				__skb_push(frag, hlen);
538 				skb_reset_network_header(frag);
539 				memcpy(skb_network_header(frag), iph, hlen);
540 				iph = ip_hdr(frag);
541 				iph->tot_len = htons(frag->len);
542 				ip_copy_metadata(frag, skb);
543 				if (offset == 0)
544 					ip_options_fragment(frag);
545 				offset += skb->len - hlen;
546 				iph->frag_off = htons(offset>>3);
547 				if (frag->next != NULL)
548 					iph->frag_off |= htons(IP_MF);
549 				/* Ready, complete checksum */
550 				ip_send_check(iph);
551 			}
552 
553 			err = output(skb);
554 
555 			if (!err)
556 				IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
557 			if (err || !frag)
558 				break;
559 
560 			skb = frag;
561 			frag = skb->next;
562 			skb->next = NULL;
563 		}
564 
565 		if (err == 0) {
566 			IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
567 			return 0;
568 		}
569 
570 		while (frag) {
571 			skb = frag->next;
572 			kfree_skb(frag);
573 			frag = skb;
574 		}
575 		IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
576 		return err;
577 
578 slow_path_clean:
579 		skb_walk_frags(skb, frag2) {
580 			if (frag2 == frag)
581 				break;
582 			frag2->sk = NULL;
583 			frag2->destructor = NULL;
584 			skb->truesize += frag2->truesize;
585 		}
586 	}
587 
588 slow_path:
589 	/* for offloaded checksums cleanup checksum before fragmentation */
590 	if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
591 		goto fail;
592 	iph = ip_hdr(skb);
593 
594 	left = skb->len - hlen;		/* Space per frame */
595 	ptr = hlen;		/* Where to start from */
596 
597 	/* for bridged IP traffic encapsulated inside f.e. a vlan header,
598 	 * we need to make room for the encapsulating header
599 	 */
600 	ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
601 
602 	/*
603 	 *	Fragment the datagram.
604 	 */
605 
606 	offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
607 	not_last_frag = iph->frag_off & htons(IP_MF);
608 
609 	/*
610 	 *	Keep copying data until we run out.
611 	 */
612 
613 	while (left > 0) {
614 		len = left;
615 		/* IF: it doesn't fit, use 'mtu' - the data space left */
616 		if (len > mtu)
617 			len = mtu;
618 		/* IF: we are not sending up to and including the packet end
619 		   then align the next start on an eight byte boundary */
620 		if (len < left)	{
621 			len &= ~7;
622 		}
623 		/*
624 		 *	Allocate buffer.
625 		 */
626 
627 		if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
628 			NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
629 			err = -ENOMEM;
630 			goto fail;
631 		}
632 
633 		/*
634 		 *	Set up data on packet
635 		 */
636 
637 		ip_copy_metadata(skb2, skb);
638 		skb_reserve(skb2, ll_rs);
639 		skb_put(skb2, len + hlen);
640 		skb_reset_network_header(skb2);
641 		skb2->transport_header = skb2->network_header + hlen;
642 
643 		/*
644 		 *	Charge the memory for the fragment to any owner
645 		 *	it might possess
646 		 */
647 
648 		if (skb->sk)
649 			skb_set_owner_w(skb2, skb->sk);
650 
651 		/*
652 		 *	Copy the packet header into the new buffer.
653 		 */
654 
655 		skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
656 
657 		/*
658 		 *	Copy a block of the IP datagram.
659 		 */
660 		if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
661 			BUG();
662 		left -= len;
663 
664 		/*
665 		 *	Fill in the new header fields.
666 		 */
667 		iph = ip_hdr(skb2);
668 		iph->frag_off = htons((offset >> 3));
669 
670 		/* ANK: dirty, but effective trick. Upgrade options only if
671 		 * the segment to be fragmented was THE FIRST (otherwise,
672 		 * options are already fixed) and make it ONCE
673 		 * on the initial skb, so that all the following fragments
674 		 * will inherit fixed options.
675 		 */
676 		if (offset == 0)
677 			ip_options_fragment(skb);
678 
679 		/*
680 		 *	Added AC : If we are fragmenting a fragment that's not the
681 		 *		   last fragment then keep MF on each bit
682 		 */
683 		if (left > 0 || not_last_frag)
684 			iph->frag_off |= htons(IP_MF);
685 		ptr += len;
686 		offset += len;
687 
688 		/*
689 		 *	Put this fragment into the sending queue.
690 		 */
691 		iph->tot_len = htons(len + hlen);
692 
693 		ip_send_check(iph);
694 
695 		err = output(skb2);
696 		if (err)
697 			goto fail;
698 
699 		IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
700 	}
701 	consume_skb(skb);
702 	IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
703 	return err;
704 
705 fail:
706 	kfree_skb(skb);
707 	IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
708 	return err;
709 }
710 EXPORT_SYMBOL(ip_fragment);
711 
712 int
713 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
714 {
715 	struct iovec *iov = from;
716 
717 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
718 		if (memcpy_fromiovecend(to, iov, offset, len) < 0)
719 			return -EFAULT;
720 	} else {
721 		__wsum csum = 0;
722 		if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
723 			return -EFAULT;
724 		skb->csum = csum_block_add(skb->csum, csum, odd);
725 	}
726 	return 0;
727 }
728 EXPORT_SYMBOL(ip_generic_getfrag);
729 
730 static inline __wsum
731 csum_page(struct page *page, int offset, int copy)
732 {
733 	char *kaddr;
734 	__wsum csum;
735 	kaddr = kmap(page);
736 	csum = csum_partial(kaddr + offset, copy, 0);
737 	kunmap(page);
738 	return csum;
739 }
740 
741 static inline int ip_ufo_append_data(struct sock *sk,
742 			struct sk_buff_head *queue,
743 			int getfrag(void *from, char *to, int offset, int len,
744 			       int odd, struct sk_buff *skb),
745 			void *from, int length, int hh_len, int fragheaderlen,
746 			int transhdrlen, int maxfraglen, unsigned int flags)
747 {
748 	struct sk_buff *skb;
749 	int err;
750 
751 	/* There is support for UDP fragmentation offload by network
752 	 * device, so create one single skb packet containing complete
753 	 * udp datagram
754 	 */
755 	if ((skb = skb_peek_tail(queue)) == NULL) {
756 		skb = sock_alloc_send_skb(sk,
757 			hh_len + fragheaderlen + transhdrlen + 20,
758 			(flags & MSG_DONTWAIT), &err);
759 
760 		if (skb == NULL)
761 			return err;
762 
763 		/* reserve space for Hardware header */
764 		skb_reserve(skb, hh_len);
765 
766 		/* create space for UDP/IP header */
767 		skb_put(skb, fragheaderlen + transhdrlen);
768 
769 		/* initialize network header pointer */
770 		skb_reset_network_header(skb);
771 
772 		/* initialize protocol header pointer */
773 		skb->transport_header = skb->network_header + fragheaderlen;
774 
775 		skb->csum = 0;
776 
777 
778 		__skb_queue_tail(queue, skb);
779 	} else if (skb_is_gso(skb)) {
780 		goto append;
781 	}
782 
783 	skb->ip_summed = CHECKSUM_PARTIAL;
784 	/* specify the length of each IP datagram fragment */
785 	skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
786 	skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
787 
788 append:
789 	return skb_append_datato_frags(sk, skb, getfrag, from,
790 				       (length - transhdrlen));
791 }
792 
793 static int __ip_append_data(struct sock *sk,
794 			    struct flowi4 *fl4,
795 			    struct sk_buff_head *queue,
796 			    struct inet_cork *cork,
797 			    struct page_frag *pfrag,
798 			    int getfrag(void *from, char *to, int offset,
799 					int len, int odd, struct sk_buff *skb),
800 			    void *from, int length, int transhdrlen,
801 			    unsigned int flags)
802 {
803 	struct inet_sock *inet = inet_sk(sk);
804 	struct sk_buff *skb;
805 
806 	struct ip_options *opt = cork->opt;
807 	int hh_len;
808 	int exthdrlen;
809 	int mtu;
810 	int copy;
811 	int err;
812 	int offset = 0;
813 	unsigned int maxfraglen, fragheaderlen;
814 	int csummode = CHECKSUM_NONE;
815 	struct rtable *rt = (struct rtable *)cork->dst;
816 
817 	skb = skb_peek_tail(queue);
818 
819 	exthdrlen = !skb ? rt->dst.header_len : 0;
820 	mtu = cork->fragsize;
821 
822 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
823 
824 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
825 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
826 
827 	if (cork->length + length > 0xFFFF - fragheaderlen) {
828 		ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
829 			       mtu-exthdrlen);
830 		return -EMSGSIZE;
831 	}
832 
833 	/*
834 	 * transhdrlen > 0 means that this is the first fragment and we wish
835 	 * it won't be fragmented in the future.
836 	 */
837 	if (transhdrlen &&
838 	    length + fragheaderlen <= mtu &&
839 	    rt->dst.dev->features & NETIF_F_V4_CSUM &&
840 	    !exthdrlen)
841 		csummode = CHECKSUM_PARTIAL;
842 
843 	cork->length += length;
844 	if (((length > mtu) || (skb && skb_is_gso(skb))) &&
845 	    (sk->sk_protocol == IPPROTO_UDP) &&
846 	    (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len) {
847 		err = ip_ufo_append_data(sk, queue, getfrag, from, length,
848 					 hh_len, fragheaderlen, transhdrlen,
849 					 maxfraglen, flags);
850 		if (err)
851 			goto error;
852 		return 0;
853 	}
854 
855 	/* So, what's going on in the loop below?
856 	 *
857 	 * We use calculated fragment length to generate chained skb,
858 	 * each of segments is IP fragment ready for sending to network after
859 	 * adding appropriate IP header.
860 	 */
861 
862 	if (!skb)
863 		goto alloc_new_skb;
864 
865 	while (length > 0) {
866 		/* Check if the remaining data fits into current packet. */
867 		copy = mtu - skb->len;
868 		if (copy < length)
869 			copy = maxfraglen - skb->len;
870 		if (copy <= 0) {
871 			char *data;
872 			unsigned int datalen;
873 			unsigned int fraglen;
874 			unsigned int fraggap;
875 			unsigned int alloclen;
876 			struct sk_buff *skb_prev;
877 alloc_new_skb:
878 			skb_prev = skb;
879 			if (skb_prev)
880 				fraggap = skb_prev->len - maxfraglen;
881 			else
882 				fraggap = 0;
883 
884 			/*
885 			 * If remaining data exceeds the mtu,
886 			 * we know we need more fragment(s).
887 			 */
888 			datalen = length + fraggap;
889 			if (datalen > mtu - fragheaderlen)
890 				datalen = maxfraglen - fragheaderlen;
891 			fraglen = datalen + fragheaderlen;
892 
893 			if ((flags & MSG_MORE) &&
894 			    !(rt->dst.dev->features&NETIF_F_SG))
895 				alloclen = mtu;
896 			else
897 				alloclen = fraglen;
898 
899 			alloclen += exthdrlen;
900 
901 			/* The last fragment gets additional space at tail.
902 			 * Note, with MSG_MORE we overallocate on fragments,
903 			 * because we have no idea what fragment will be
904 			 * the last.
905 			 */
906 			if (datalen == length + fraggap)
907 				alloclen += rt->dst.trailer_len;
908 
909 			if (transhdrlen) {
910 				skb = sock_alloc_send_skb(sk,
911 						alloclen + hh_len + 15,
912 						(flags & MSG_DONTWAIT), &err);
913 			} else {
914 				skb = NULL;
915 				if (atomic_read(&sk->sk_wmem_alloc) <=
916 				    2 * sk->sk_sndbuf)
917 					skb = sock_wmalloc(sk,
918 							   alloclen + hh_len + 15, 1,
919 							   sk->sk_allocation);
920 				if (unlikely(skb == NULL))
921 					err = -ENOBUFS;
922 				else
923 					/* only the initial fragment is
924 					   time stamped */
925 					cork->tx_flags = 0;
926 			}
927 			if (skb == NULL)
928 				goto error;
929 
930 			/*
931 			 *	Fill in the control structures
932 			 */
933 			skb->ip_summed = csummode;
934 			skb->csum = 0;
935 			skb_reserve(skb, hh_len);
936 			skb_shinfo(skb)->tx_flags = cork->tx_flags;
937 
938 			/*
939 			 *	Find where to start putting bytes.
940 			 */
941 			data = skb_put(skb, fraglen + exthdrlen);
942 			skb_set_network_header(skb, exthdrlen);
943 			skb->transport_header = (skb->network_header +
944 						 fragheaderlen);
945 			data += fragheaderlen + exthdrlen;
946 
947 			if (fraggap) {
948 				skb->csum = skb_copy_and_csum_bits(
949 					skb_prev, maxfraglen,
950 					data + transhdrlen, fraggap, 0);
951 				skb_prev->csum = csum_sub(skb_prev->csum,
952 							  skb->csum);
953 				data += fraggap;
954 				pskb_trim_unique(skb_prev, maxfraglen);
955 			}
956 
957 			copy = datalen - transhdrlen - fraggap;
958 			if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
959 				err = -EFAULT;
960 				kfree_skb(skb);
961 				goto error;
962 			}
963 
964 			offset += copy;
965 			length -= datalen - fraggap;
966 			transhdrlen = 0;
967 			exthdrlen = 0;
968 			csummode = CHECKSUM_NONE;
969 
970 			/*
971 			 * Put the packet on the pending queue.
972 			 */
973 			__skb_queue_tail(queue, skb);
974 			continue;
975 		}
976 
977 		if (copy > length)
978 			copy = length;
979 
980 		if (!(rt->dst.dev->features&NETIF_F_SG)) {
981 			unsigned int off;
982 
983 			off = skb->len;
984 			if (getfrag(from, skb_put(skb, copy),
985 					offset, copy, off, skb) < 0) {
986 				__skb_trim(skb, off);
987 				err = -EFAULT;
988 				goto error;
989 			}
990 		} else {
991 			int i = skb_shinfo(skb)->nr_frags;
992 
993 			err = -ENOMEM;
994 			if (!sk_page_frag_refill(sk, pfrag))
995 				goto error;
996 
997 			if (!skb_can_coalesce(skb, i, pfrag->page,
998 					      pfrag->offset)) {
999 				err = -EMSGSIZE;
1000 				if (i == MAX_SKB_FRAGS)
1001 					goto error;
1002 
1003 				__skb_fill_page_desc(skb, i, pfrag->page,
1004 						     pfrag->offset, 0);
1005 				skb_shinfo(skb)->nr_frags = ++i;
1006 				get_page(pfrag->page);
1007 			}
1008 			copy = min_t(int, copy, pfrag->size - pfrag->offset);
1009 			if (getfrag(from,
1010 				    page_address(pfrag->page) + pfrag->offset,
1011 				    offset, copy, skb->len, skb) < 0)
1012 				goto error_efault;
1013 
1014 			pfrag->offset += copy;
1015 			skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1016 			skb->len += copy;
1017 			skb->data_len += copy;
1018 			skb->truesize += copy;
1019 			atomic_add(copy, &sk->sk_wmem_alloc);
1020 		}
1021 		offset += copy;
1022 		length -= copy;
1023 	}
1024 
1025 	return 0;
1026 
1027 error_efault:
1028 	err = -EFAULT;
1029 error:
1030 	cork->length -= length;
1031 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1032 	return err;
1033 }
1034 
1035 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1036 			 struct ipcm_cookie *ipc, struct rtable **rtp)
1037 {
1038 	struct inet_sock *inet = inet_sk(sk);
1039 	struct ip_options_rcu *opt;
1040 	struct rtable *rt;
1041 
1042 	/*
1043 	 * setup for corking.
1044 	 */
1045 	opt = ipc->opt;
1046 	if (opt) {
1047 		if (cork->opt == NULL) {
1048 			cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1049 					    sk->sk_allocation);
1050 			if (unlikely(cork->opt == NULL))
1051 				return -ENOBUFS;
1052 		}
1053 		memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1054 		cork->flags |= IPCORK_OPT;
1055 		cork->addr = ipc->addr;
1056 	}
1057 	rt = *rtp;
1058 	if (unlikely(!rt))
1059 		return -EFAULT;
1060 	/*
1061 	 * We steal reference to this route, caller should not release it
1062 	 */
1063 	*rtp = NULL;
1064 	cork->fragsize = inet->pmtudisc == IP_PMTUDISC_PROBE ?
1065 			 rt->dst.dev->mtu : dst_mtu(&rt->dst);
1066 	cork->dst = &rt->dst;
1067 	cork->length = 0;
1068 	cork->tx_flags = ipc->tx_flags;
1069 
1070 	return 0;
1071 }
1072 
1073 /*
1074  *	ip_append_data() and ip_append_page() can make one large IP datagram
1075  *	from many pieces of data. Each pieces will be holded on the socket
1076  *	until ip_push_pending_frames() is called. Each piece can be a page
1077  *	or non-page data.
1078  *
1079  *	Not only UDP, other transport protocols - e.g. raw sockets - can use
1080  *	this interface potentially.
1081  *
1082  *	LATER: length must be adjusted by pad at tail, when it is required.
1083  */
1084 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1085 		   int getfrag(void *from, char *to, int offset, int len,
1086 			       int odd, struct sk_buff *skb),
1087 		   void *from, int length, int transhdrlen,
1088 		   struct ipcm_cookie *ipc, struct rtable **rtp,
1089 		   unsigned int flags)
1090 {
1091 	struct inet_sock *inet = inet_sk(sk);
1092 	int err;
1093 
1094 	if (flags&MSG_PROBE)
1095 		return 0;
1096 
1097 	if (skb_queue_empty(&sk->sk_write_queue)) {
1098 		err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1099 		if (err)
1100 			return err;
1101 	} else {
1102 		transhdrlen = 0;
1103 	}
1104 
1105 	return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1106 				sk_page_frag(sk), getfrag,
1107 				from, length, transhdrlen, flags);
1108 }
1109 
1110 ssize_t	ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1111 		       int offset, size_t size, int flags)
1112 {
1113 	struct inet_sock *inet = inet_sk(sk);
1114 	struct sk_buff *skb;
1115 	struct rtable *rt;
1116 	struct ip_options *opt = NULL;
1117 	struct inet_cork *cork;
1118 	int hh_len;
1119 	int mtu;
1120 	int len;
1121 	int err;
1122 	unsigned int maxfraglen, fragheaderlen, fraggap;
1123 
1124 	if (inet->hdrincl)
1125 		return -EPERM;
1126 
1127 	if (flags&MSG_PROBE)
1128 		return 0;
1129 
1130 	if (skb_queue_empty(&sk->sk_write_queue))
1131 		return -EINVAL;
1132 
1133 	cork = &inet->cork.base;
1134 	rt = (struct rtable *)cork->dst;
1135 	if (cork->flags & IPCORK_OPT)
1136 		opt = cork->opt;
1137 
1138 	if (!(rt->dst.dev->features&NETIF_F_SG))
1139 		return -EOPNOTSUPP;
1140 
1141 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1142 	mtu = cork->fragsize;
1143 
1144 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1145 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1146 
1147 	if (cork->length + size > 0xFFFF - fragheaderlen) {
1148 		ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport, mtu);
1149 		return -EMSGSIZE;
1150 	}
1151 
1152 	if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1153 		return -EINVAL;
1154 
1155 	cork->length += size;
1156 	if ((size + skb->len > mtu) &&
1157 	    (sk->sk_protocol == IPPROTO_UDP) &&
1158 	    (rt->dst.dev->features & NETIF_F_UFO)) {
1159 		skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1160 		skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1161 	}
1162 
1163 
1164 	while (size > 0) {
1165 		int i;
1166 
1167 		if (skb_is_gso(skb))
1168 			len = size;
1169 		else {
1170 
1171 			/* Check if the remaining data fits into current packet. */
1172 			len = mtu - skb->len;
1173 			if (len < size)
1174 				len = maxfraglen - skb->len;
1175 		}
1176 		if (len <= 0) {
1177 			struct sk_buff *skb_prev;
1178 			int alloclen;
1179 
1180 			skb_prev = skb;
1181 			fraggap = skb_prev->len - maxfraglen;
1182 
1183 			alloclen = fragheaderlen + hh_len + fraggap + 15;
1184 			skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1185 			if (unlikely(!skb)) {
1186 				err = -ENOBUFS;
1187 				goto error;
1188 			}
1189 
1190 			/*
1191 			 *	Fill in the control structures
1192 			 */
1193 			skb->ip_summed = CHECKSUM_NONE;
1194 			skb->csum = 0;
1195 			skb_reserve(skb, hh_len);
1196 
1197 			/*
1198 			 *	Find where to start putting bytes.
1199 			 */
1200 			skb_put(skb, fragheaderlen + fraggap);
1201 			skb_reset_network_header(skb);
1202 			skb->transport_header = (skb->network_header +
1203 						 fragheaderlen);
1204 			if (fraggap) {
1205 				skb->csum = skb_copy_and_csum_bits(skb_prev,
1206 								   maxfraglen,
1207 						    skb_transport_header(skb),
1208 								   fraggap, 0);
1209 				skb_prev->csum = csum_sub(skb_prev->csum,
1210 							  skb->csum);
1211 				pskb_trim_unique(skb_prev, maxfraglen);
1212 			}
1213 
1214 			/*
1215 			 * Put the packet on the pending queue.
1216 			 */
1217 			__skb_queue_tail(&sk->sk_write_queue, skb);
1218 			continue;
1219 		}
1220 
1221 		i = skb_shinfo(skb)->nr_frags;
1222 		if (len > size)
1223 			len = size;
1224 		if (skb_can_coalesce(skb, i, page, offset)) {
1225 			skb_frag_size_add(&skb_shinfo(skb)->frags[i-1], len);
1226 		} else if (i < MAX_SKB_FRAGS) {
1227 			get_page(page);
1228 			skb_fill_page_desc(skb, i, page, offset, len);
1229 		} else {
1230 			err = -EMSGSIZE;
1231 			goto error;
1232 		}
1233 
1234 		if (skb->ip_summed == CHECKSUM_NONE) {
1235 			__wsum csum;
1236 			csum = csum_page(page, offset, len);
1237 			skb->csum = csum_block_add(skb->csum, csum, skb->len);
1238 		}
1239 
1240 		skb->len += len;
1241 		skb->data_len += len;
1242 		skb->truesize += len;
1243 		atomic_add(len, &sk->sk_wmem_alloc);
1244 		offset += len;
1245 		size -= len;
1246 	}
1247 	return 0;
1248 
1249 error:
1250 	cork->length -= size;
1251 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1252 	return err;
1253 }
1254 
1255 static void ip_cork_release(struct inet_cork *cork)
1256 {
1257 	cork->flags &= ~IPCORK_OPT;
1258 	kfree(cork->opt);
1259 	cork->opt = NULL;
1260 	dst_release(cork->dst);
1261 	cork->dst = NULL;
1262 }
1263 
1264 /*
1265  *	Combined all pending IP fragments on the socket as one IP datagram
1266  *	and push them out.
1267  */
1268 struct sk_buff *__ip_make_skb(struct sock *sk,
1269 			      struct flowi4 *fl4,
1270 			      struct sk_buff_head *queue,
1271 			      struct inet_cork *cork)
1272 {
1273 	struct sk_buff *skb, *tmp_skb;
1274 	struct sk_buff **tail_skb;
1275 	struct inet_sock *inet = inet_sk(sk);
1276 	struct net *net = sock_net(sk);
1277 	struct ip_options *opt = NULL;
1278 	struct rtable *rt = (struct rtable *)cork->dst;
1279 	struct iphdr *iph;
1280 	__be16 df = 0;
1281 	__u8 ttl;
1282 
1283 	if ((skb = __skb_dequeue(queue)) == NULL)
1284 		goto out;
1285 	tail_skb = &(skb_shinfo(skb)->frag_list);
1286 
1287 	/* move skb->data to ip header from ext header */
1288 	if (skb->data < skb_network_header(skb))
1289 		__skb_pull(skb, skb_network_offset(skb));
1290 	while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1291 		__skb_pull(tmp_skb, skb_network_header_len(skb));
1292 		*tail_skb = tmp_skb;
1293 		tail_skb = &(tmp_skb->next);
1294 		skb->len += tmp_skb->len;
1295 		skb->data_len += tmp_skb->len;
1296 		skb->truesize += tmp_skb->truesize;
1297 		tmp_skb->destructor = NULL;
1298 		tmp_skb->sk = NULL;
1299 	}
1300 
1301 	/* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1302 	 * to fragment the frame generated here. No matter, what transforms
1303 	 * how transforms change size of the packet, it will come out.
1304 	 */
1305 	if (inet->pmtudisc < IP_PMTUDISC_DO)
1306 		skb->local_df = 1;
1307 
1308 	/* DF bit is set when we want to see DF on outgoing frames.
1309 	 * If local_df is set too, we still allow to fragment this frame
1310 	 * locally. */
1311 	if (inet->pmtudisc >= IP_PMTUDISC_DO ||
1312 	    (skb->len <= dst_mtu(&rt->dst) &&
1313 	     ip_dont_fragment(sk, &rt->dst)))
1314 		df = htons(IP_DF);
1315 
1316 	if (cork->flags & IPCORK_OPT)
1317 		opt = cork->opt;
1318 
1319 	if (rt->rt_type == RTN_MULTICAST)
1320 		ttl = inet->mc_ttl;
1321 	else
1322 		ttl = ip_select_ttl(inet, &rt->dst);
1323 
1324 	iph = ip_hdr(skb);
1325 	iph->version = 4;
1326 	iph->ihl = 5;
1327 	iph->tos = inet->tos;
1328 	iph->frag_off = df;
1329 	iph->ttl = ttl;
1330 	iph->protocol = sk->sk_protocol;
1331 	ip_copy_addrs(iph, fl4);
1332 	ip_select_ident(skb, &rt->dst, sk);
1333 
1334 	if (opt) {
1335 		iph->ihl += opt->optlen>>2;
1336 		ip_options_build(skb, opt, cork->addr, rt, 0);
1337 	}
1338 
1339 	skb->priority = sk->sk_priority;
1340 	skb->mark = sk->sk_mark;
1341 	/*
1342 	 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1343 	 * on dst refcount
1344 	 */
1345 	cork->dst = NULL;
1346 	skb_dst_set(skb, &rt->dst);
1347 
1348 	if (iph->protocol == IPPROTO_ICMP)
1349 		icmp_out_count(net, ((struct icmphdr *)
1350 			skb_transport_header(skb))->type);
1351 
1352 	ip_cork_release(cork);
1353 out:
1354 	return skb;
1355 }
1356 
1357 int ip_send_skb(struct net *net, struct sk_buff *skb)
1358 {
1359 	int err;
1360 
1361 	err = ip_local_out(skb);
1362 	if (err) {
1363 		if (err > 0)
1364 			err = net_xmit_errno(err);
1365 		if (err)
1366 			IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1367 	}
1368 
1369 	return err;
1370 }
1371 
1372 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1373 {
1374 	struct sk_buff *skb;
1375 
1376 	skb = ip_finish_skb(sk, fl4);
1377 	if (!skb)
1378 		return 0;
1379 
1380 	/* Netfilter gets whole the not fragmented skb. */
1381 	return ip_send_skb(sock_net(sk), skb);
1382 }
1383 
1384 /*
1385  *	Throw away all pending data on the socket.
1386  */
1387 static void __ip_flush_pending_frames(struct sock *sk,
1388 				      struct sk_buff_head *queue,
1389 				      struct inet_cork *cork)
1390 {
1391 	struct sk_buff *skb;
1392 
1393 	while ((skb = __skb_dequeue_tail(queue)) != NULL)
1394 		kfree_skb(skb);
1395 
1396 	ip_cork_release(cork);
1397 }
1398 
1399 void ip_flush_pending_frames(struct sock *sk)
1400 {
1401 	__ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1402 }
1403 
1404 struct sk_buff *ip_make_skb(struct sock *sk,
1405 			    struct flowi4 *fl4,
1406 			    int getfrag(void *from, char *to, int offset,
1407 					int len, int odd, struct sk_buff *skb),
1408 			    void *from, int length, int transhdrlen,
1409 			    struct ipcm_cookie *ipc, struct rtable **rtp,
1410 			    unsigned int flags)
1411 {
1412 	struct inet_cork cork;
1413 	struct sk_buff_head queue;
1414 	int err;
1415 
1416 	if (flags & MSG_PROBE)
1417 		return NULL;
1418 
1419 	__skb_queue_head_init(&queue);
1420 
1421 	cork.flags = 0;
1422 	cork.addr = 0;
1423 	cork.opt = NULL;
1424 	err = ip_setup_cork(sk, &cork, ipc, rtp);
1425 	if (err)
1426 		return ERR_PTR(err);
1427 
1428 	err = __ip_append_data(sk, fl4, &queue, &cork,
1429 			       &current->task_frag, getfrag,
1430 			       from, length, transhdrlen, flags);
1431 	if (err) {
1432 		__ip_flush_pending_frames(sk, &queue, &cork);
1433 		return ERR_PTR(err);
1434 	}
1435 
1436 	return __ip_make_skb(sk, fl4, &queue, &cork);
1437 }
1438 
1439 /*
1440  *	Fetch data from kernel space and fill in checksum if needed.
1441  */
1442 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1443 			      int len, int odd, struct sk_buff *skb)
1444 {
1445 	__wsum csum;
1446 
1447 	csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1448 	skb->csum = csum_block_add(skb->csum, csum, odd);
1449 	return 0;
1450 }
1451 
1452 /*
1453  *	Generic function to send a packet as reply to another packet.
1454  *	Used to send some TCP resets/acks so far.
1455  *
1456  *	Use a fake percpu inet socket to avoid false sharing and contention.
1457  */
1458 static DEFINE_PER_CPU(struct inet_sock, unicast_sock) = {
1459 	.sk = {
1460 		.__sk_common = {
1461 			.skc_refcnt = ATOMIC_INIT(1),
1462 		},
1463 		.sk_wmem_alloc	= ATOMIC_INIT(1),
1464 		.sk_allocation	= GFP_ATOMIC,
1465 		.sk_flags	= (1UL << SOCK_USE_WRITE_QUEUE),
1466 	},
1467 	.pmtudisc	= IP_PMTUDISC_WANT,
1468 	.uc_ttl		= -1,
1469 };
1470 
1471 void ip_send_unicast_reply(struct net *net, struct sk_buff *skb, __be32 daddr,
1472 			   __be32 saddr, const struct ip_reply_arg *arg,
1473 			   unsigned int len)
1474 {
1475 	struct ip_options_data replyopts;
1476 	struct ipcm_cookie ipc;
1477 	struct flowi4 fl4;
1478 	struct rtable *rt = skb_rtable(skb);
1479 	struct sk_buff *nskb;
1480 	struct sock *sk;
1481 	struct inet_sock *inet;
1482 
1483 	if (ip_options_echo(&replyopts.opt.opt, skb))
1484 		return;
1485 
1486 	ipc.addr = daddr;
1487 	ipc.opt = NULL;
1488 	ipc.tx_flags = 0;
1489 
1490 	if (replyopts.opt.opt.optlen) {
1491 		ipc.opt = &replyopts.opt;
1492 
1493 		if (replyopts.opt.opt.srr)
1494 			daddr = replyopts.opt.opt.faddr;
1495 	}
1496 
1497 	flowi4_init_output(&fl4, arg->bound_dev_if, 0,
1498 			   RT_TOS(arg->tos),
1499 			   RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1500 			   ip_reply_arg_flowi_flags(arg),
1501 			   daddr, saddr,
1502 			   tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1503 	security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1504 	rt = ip_route_output_key(net, &fl4);
1505 	if (IS_ERR(rt))
1506 		return;
1507 
1508 	inet = &get_cpu_var(unicast_sock);
1509 
1510 	inet->tos = arg->tos;
1511 	sk = &inet->sk;
1512 	sk->sk_priority = skb->priority;
1513 	sk->sk_protocol = ip_hdr(skb)->protocol;
1514 	sk->sk_bound_dev_if = arg->bound_dev_if;
1515 	sock_net_set(sk, net);
1516 	__skb_queue_head_init(&sk->sk_write_queue);
1517 	sk->sk_sndbuf = sysctl_wmem_default;
1518 	ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1519 		       &ipc, &rt, MSG_DONTWAIT);
1520 	nskb = skb_peek(&sk->sk_write_queue);
1521 	if (nskb) {
1522 		if (arg->csumoffset >= 0)
1523 			*((__sum16 *)skb_transport_header(nskb) +
1524 			  arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1525 								arg->csum));
1526 		nskb->ip_summed = CHECKSUM_NONE;
1527 		skb_orphan(nskb);
1528 		skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1529 		ip_push_pending_frames(sk, &fl4);
1530 	}
1531 
1532 	put_cpu_var(unicast_sock);
1533 
1534 	ip_rt_put(rt);
1535 }
1536 
1537 void __init ip_init(void)
1538 {
1539 	ip_rt_init();
1540 	inet_initpeers();
1541 
1542 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1543 	igmp_mc_proc_init();
1544 #endif
1545 }
1546