Merge ../linus
[linux-2.6] / net / ipv4 / ip_output.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              The Internet Protocol (IP) output module.
7  *
8  * Version:     $Id: ip_output.c,v 1.100 2002/02/01 22:01:03 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Donald Becker, <becker@super.org>
13  *              Alan Cox, <Alan.Cox@linux.org>
14  *              Richard Underwood
15  *              Stefan Becker, <stefanb@yello.ping.de>
16  *              Jorge Cwik, <jorge@laser.satlink.net>
17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18  *              Hirokazu Takahashi, <taka@valinux.co.jp>
19  *
20  *      See ip_input.c for original log
21  *
22  *      Fixes:
23  *              Alan Cox        :       Missing nonblock feature in ip_build_xmit.
24  *              Mike Kilburn    :       htons() missing in ip_build_xmit.
25  *              Bradford Johnson:       Fix faulty handling of some frames when 
26  *                                      no route is found.
27  *              Alexander Demenshin:    Missing sk/skb free in ip_queue_xmit
28  *                                      (in case if packet not accepted by
29  *                                      output firewall rules)
30  *              Mike McLagan    :       Routing by source
31  *              Alexey Kuznetsov:       use new route cache
32  *              Andi Kleen:             Fix broken PMTU recovery and remove
33  *                                      some redundant tests.
34  *      Vitaly E. Lavrov        :       Transparent proxy revived after year coma.
35  *              Andi Kleen      :       Replace ip_reply with ip_send_reply.
36  *              Andi Kleen      :       Split fast and slow ip_build_xmit path 
37  *                                      for decreased register pressure on x86 
38  *                                      and more readibility. 
39  *              Marc Boucher    :       When call_out_firewall returns FW_QUEUE,
40  *                                      silently drop skb instead of failing with -EPERM.
41  *              Detlev Wengorz  :       Copy protocol for fragments.
42  *              Hirokazu Takahashi:     HW checksumming for outgoing UDP
43  *                                      datagrams.
44  *              Hirokazu Takahashi:     sendfile() on UDP works now.
45  */
46
47 #include <asm/uaccess.h>
48 #include <asm/system.h>
49 #include <linux/module.h>
50 #include <linux/types.h>
51 #include <linux/kernel.h>
52 #include <linux/sched.h>
53 #include <linux/mm.h>
54 #include <linux/string.h>
55 #include <linux/errno.h>
56
57 #include <linux/socket.h>
58 #include <linux/sockios.h>
59 #include <linux/in.h>
60 #include <linux/inet.h>
61 #include <linux/netdevice.h>
62 #include <linux/etherdevice.h>
63 #include <linux/proc_fs.h>
64 #include <linux/stat.h>
65 #include <linux/init.h>
66
67 #include <net/snmp.h>
68 #include <net/ip.h>
69 #include <net/protocol.h>
70 #include <net/route.h>
71 #include <net/xfrm.h>
72 #include <linux/skbuff.h>
73 #include <net/sock.h>
74 #include <net/arp.h>
75 #include <net/icmp.h>
76 #include <net/checksum.h>
77 #include <net/inetpeer.h>
78 #include <net/checksum.h>
79 #include <linux/igmp.h>
80 #include <linux/netfilter_ipv4.h>
81 #include <linux/netfilter_bridge.h>
82 #include <linux/mroute.h>
83 #include <linux/netlink.h>
84 #include <linux/tcp.h>
85
86 int sysctl_ip_default_ttl = IPDEFTTL;
87
88 /* Generate a checksum for an outgoing IP datagram. */
89 __inline__ 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
95 /* dev_loopback_xmit for use with netfilter. */
96 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
97 {
98         newskb->mac.raw = newskb->data;
99         __skb_pull(newskb, newskb->nh.raw - newskb->data);
100         newskb->pkt_type = PACKET_LOOPBACK;
101         newskb->ip_summed = CHECKSUM_UNNECESSARY;
102         BUG_TRAP(newskb->dst);
103         netif_rx(newskb);
104         return 0;
105 }
106
107 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
108 {
109         int ttl = inet->uc_ttl;
110
111         if (ttl < 0)
112                 ttl = dst_metric(dst, RTAX_HOPLIMIT);
113         return ttl;
114 }
115
116 /* 
117  *              Add an ip header to a skbuff and send it out.
118  *
119  */
120 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
121                           u32 saddr, u32 daddr, struct ip_options *opt)
122 {
123         struct inet_sock *inet = inet_sk(sk);
124         struct rtable *rt = (struct rtable *)skb->dst;
125         struct iphdr *iph;
126
127         /* Build the IP header. */
128         if (opt)
129                 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr) + opt->optlen);
130         else
131                 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr));
132
133         iph->version  = 4;
134         iph->ihl      = 5;
135         iph->tos      = inet->tos;
136         if (ip_dont_fragment(sk, &rt->u.dst))
137                 iph->frag_off = htons(IP_DF);
138         else
139                 iph->frag_off = 0;
140         iph->ttl      = ip_select_ttl(inet, &rt->u.dst);
141         iph->daddr    = rt->rt_dst;
142         iph->saddr    = rt->rt_src;
143         iph->protocol = sk->sk_protocol;
144         iph->tot_len  = htons(skb->len);
145         ip_select_ident(iph, &rt->u.dst, sk);
146         skb->nh.iph   = iph;
147
148         if (opt && opt->optlen) {
149                 iph->ihl += opt->optlen>>2;
150                 ip_options_build(skb, opt, daddr, rt, 0);
151         }
152         ip_send_check(iph);
153
154         skb->priority = sk->sk_priority;
155
156         /* Send it out. */
157         return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
158                        dst_output);
159 }
160
161 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
162
163 static inline int ip_finish_output2(struct sk_buff *skb)
164 {
165         struct dst_entry *dst = skb->dst;
166         struct hh_cache *hh = dst->hh;
167         struct net_device *dev = dst->dev;
168         int hh_len = LL_RESERVED_SPACE(dev);
169
170         /* Be paranoid, rather than too clever. */
171         if (unlikely(skb_headroom(skb) < hh_len && dev->hard_header)) {
172                 struct sk_buff *skb2;
173
174                 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
175                 if (skb2 == NULL) {
176                         kfree_skb(skb);
177                         return -ENOMEM;
178                 }
179                 if (skb->sk)
180                         skb_set_owner_w(skb2, skb->sk);
181                 kfree_skb(skb);
182                 skb = skb2;
183         }
184
185         if (hh) {
186                 int hh_alen;
187
188                 read_lock_bh(&hh->hh_lock);
189                 hh_alen = HH_DATA_ALIGN(hh->hh_len);
190                 memcpy(skb->data - hh_alen, hh->hh_data, hh_alen);
191                 read_unlock_bh(&hh->hh_lock);
192                 skb_push(skb, hh->hh_len);
193                 return hh->hh_output(skb);
194         } else if (dst->neighbour)
195                 return dst->neighbour->output(skb);
196
197         if (net_ratelimit())
198                 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
199         kfree_skb(skb);
200         return -EINVAL;
201 }
202
203 static inline int ip_finish_output(struct sk_buff *skb)
204 {
205 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
206         /* Policy lookup after SNAT yielded a new policy */
207         if (skb->dst->xfrm != NULL) {
208                 IPCB(skb)->flags |= IPSKB_REROUTED;
209                 return dst_output(skb);
210         }
211 #endif
212         if (skb->len > dst_mtu(skb->dst) && !skb_is_gso(skb))
213                 return ip_fragment(skb, ip_finish_output2);
214         else
215                 return ip_finish_output2(skb);
216 }
217
218 int ip_mc_output(struct sk_buff *skb)
219 {
220         struct sock *sk = skb->sk;
221         struct rtable *rt = (struct rtable*)skb->dst;
222         struct net_device *dev = rt->u.dst.dev;
223
224         /*
225          *      If the indicated interface is up and running, send the packet.
226          */
227         IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
228
229         skb->dev = dev;
230         skb->protocol = htons(ETH_P_IP);
231
232         /*
233          *      Multicasts are looped back for other local users
234          */
235
236         if (rt->rt_flags&RTCF_MULTICAST) {
237                 if ((!sk || inet_sk(sk)->mc_loop)
238 #ifdef CONFIG_IP_MROUTE
239                 /* Small optimization: do not loopback not local frames,
240                    which returned after forwarding; they will be  dropped
241                    by ip_mr_input in any case.
242                    Note, that local frames are looped back to be delivered
243                    to local recipients.
244
245                    This check is duplicated in ip_mr_input at the moment.
246                  */
247                     && ((rt->rt_flags&RTCF_LOCAL) || !(IPCB(skb)->flags&IPSKB_FORWARDED))
248 #endif
249                 ) {
250                         struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
251                         if (newskb)
252                                 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
253                                         newskb->dev, 
254                                         ip_dev_loopback_xmit);
255                 }
256
257                 /* Multicasts with ttl 0 must not go beyond the host */
258
259                 if (skb->nh.iph->ttl == 0) {
260                         kfree_skb(skb);
261                         return 0;
262                 }
263         }
264
265         if (rt->rt_flags&RTCF_BROADCAST) {
266                 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
267                 if (newskb)
268                         NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
269                                 newskb->dev, ip_dev_loopback_xmit);
270         }
271
272         return NF_HOOK_COND(PF_INET, NF_IP_POST_ROUTING, skb, NULL, skb->dev,
273                             ip_finish_output,
274                             !(IPCB(skb)->flags & IPSKB_REROUTED));
275 }
276
277 int ip_output(struct sk_buff *skb)
278 {
279         struct net_device *dev = skb->dst->dev;
280
281         IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
282
283         skb->dev = dev;
284         skb->protocol = htons(ETH_P_IP);
285
286         return NF_HOOK_COND(PF_INET, NF_IP_POST_ROUTING, skb, NULL, dev,
287                             ip_finish_output,
288                             !(IPCB(skb)->flags & IPSKB_REROUTED));
289 }
290
291 int ip_queue_xmit(struct sk_buff *skb, int ipfragok)
292 {
293         struct sock *sk = skb->sk;
294         struct inet_sock *inet = inet_sk(sk);
295         struct ip_options *opt = inet->opt;
296         struct rtable *rt;
297         struct iphdr *iph;
298
299         /* Skip all of this if the packet is already routed,
300          * f.e. by something like SCTP.
301          */
302         rt = (struct rtable *) skb->dst;
303         if (rt != NULL)
304                 goto packet_routed;
305
306         /* Make sure we can route this packet. */
307         rt = (struct rtable *)__sk_dst_check(sk, 0);
308         if (rt == NULL) {
309                 u32 daddr;
310
311                 /* Use correct destination address if we have options. */
312                 daddr = inet->daddr;
313                 if(opt && opt->srr)
314                         daddr = opt->faddr;
315
316                 {
317                         struct flowi fl = { .oif = sk->sk_bound_dev_if,
318                                             .nl_u = { .ip4_u =
319                                                       { .daddr = daddr,
320                                                         .saddr = inet->saddr,
321                                                         .tos = RT_CONN_FLAGS(sk) } },
322                                             .proto = sk->sk_protocol,
323                                             .uli_u = { .ports =
324                                                        { .sport = inet->sport,
325                                                          .dport = inet->dport } } };
326
327                         /* If this fails, retransmit mechanism of transport layer will
328                          * keep trying until route appears or the connection times
329                          * itself out.
330                          */
331                         if (ip_route_output_flow(&rt, &fl, sk, 0))
332                                 goto no_route;
333                 }
334                 sk_setup_caps(sk, &rt->u.dst);
335         }
336         skb->dst = dst_clone(&rt->u.dst);
337
338 packet_routed:
339         if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
340                 goto no_route;
341
342         /* OK, we know where to send it, allocate and build IP header. */
343         iph = (struct iphdr *) skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
344         *((__u16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
345         iph->tot_len = htons(skb->len);
346         if (ip_dont_fragment(sk, &rt->u.dst) && !ipfragok)
347                 iph->frag_off = htons(IP_DF);
348         else
349                 iph->frag_off = 0;
350         iph->ttl      = ip_select_ttl(inet, &rt->u.dst);
351         iph->protocol = sk->sk_protocol;
352         iph->saddr    = rt->rt_src;
353         iph->daddr    = rt->rt_dst;
354         skb->nh.iph   = iph;
355         /* Transport layer set skb->h.foo itself. */
356
357         if (opt && opt->optlen) {
358                 iph->ihl += opt->optlen >> 2;
359                 ip_options_build(skb, opt, inet->daddr, rt, 0);
360         }
361
362         ip_select_ident_more(iph, &rt->u.dst, sk,
363                              (skb_shinfo(skb)->gso_segs ?: 1) - 1);
364
365         /* Add an IP checksum. */
366         ip_send_check(iph);
367
368         skb->priority = sk->sk_priority;
369
370         return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
371                        dst_output);
372
373 no_route:
374         IP_INC_STATS(IPSTATS_MIB_OUTNOROUTES);
375         kfree_skb(skb);
376         return -EHOSTUNREACH;
377 }
378
379
380 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
381 {
382         to->pkt_type = from->pkt_type;
383         to->priority = from->priority;
384         to->protocol = from->protocol;
385         dst_release(to->dst);
386         to->dst = dst_clone(from->dst);
387         to->dev = from->dev;
388
389         /* Copy the flags to each fragment. */
390         IPCB(to)->flags = IPCB(from)->flags;
391
392 #ifdef CONFIG_NET_SCHED
393         to->tc_index = from->tc_index;
394 #endif
395 #ifdef CONFIG_NETFILTER
396         to->nfmark = from->nfmark;
397         /* Connection association is same as pre-frag packet */
398         nf_conntrack_put(to->nfct);
399         to->nfct = from->nfct;
400         nf_conntrack_get(to->nfct);
401         to->nfctinfo = from->nfctinfo;
402 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
403         to->ipvs_property = from->ipvs_property;
404 #endif
405 #ifdef CONFIG_BRIDGE_NETFILTER
406         nf_bridge_put(to->nf_bridge);
407         to->nf_bridge = from->nf_bridge;
408         nf_bridge_get(to->nf_bridge);
409 #endif
410 #endif
411         skb_copy_secmark(to, from);
412 }
413
414 /*
415  *      This IP datagram is too large to be sent in one piece.  Break it up into
416  *      smaller pieces (each of size equal to IP header plus
417  *      a block of the data of the original IP data part) that will yet fit in a
418  *      single device frame, and queue such a frame for sending.
419  */
420
421 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*))
422 {
423         struct iphdr *iph;
424         int raw = 0;
425         int ptr;
426         struct net_device *dev;
427         struct sk_buff *skb2;
428         unsigned int mtu, hlen, left, len, ll_rs;
429         int offset;
430         __be16 not_last_frag;
431         struct rtable *rt = (struct rtable*)skb->dst;
432         int err = 0;
433
434         dev = rt->u.dst.dev;
435
436         /*
437          *      Point into the IP datagram header.
438          */
439
440         iph = skb->nh.iph;
441
442         if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
443                 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
444                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
445                           htonl(dst_mtu(&rt->u.dst)));
446                 kfree_skb(skb);
447                 return -EMSGSIZE;
448         }
449
450         /*
451          *      Setup starting values.
452          */
453
454         hlen = iph->ihl * 4;
455         mtu = dst_mtu(&rt->u.dst) - hlen;       /* Size of data space */
456         IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
457
458         /* When frag_list is given, use it. First, check its validity:
459          * some transformers could create wrong frag_list or break existing
460          * one, it is not prohibited. In this case fall back to copying.
461          *
462          * LATER: this step can be merged to real generation of fragments,
463          * we can switch to copy when see the first bad fragment.
464          */
465         if (skb_shinfo(skb)->frag_list) {
466                 struct sk_buff *frag;
467                 int first_len = skb_pagelen(skb);
468
469                 if (first_len - hlen > mtu ||
470                     ((first_len - hlen) & 7) ||
471                     (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
472                     skb_cloned(skb))
473                         goto slow_path;
474
475                 for (frag = skb_shinfo(skb)->frag_list; frag; frag = frag->next) {
476                         /* Correct geometry. */
477                         if (frag->len > mtu ||
478                             ((frag->len & 7) && frag->next) ||
479                             skb_headroom(frag) < hlen)
480                             goto slow_path;
481
482                         /* Partially cloned skb? */
483                         if (skb_shared(frag))
484                                 goto slow_path;
485
486                         BUG_ON(frag->sk);
487                         if (skb->sk) {
488                                 sock_hold(skb->sk);
489                                 frag->sk = skb->sk;
490                                 frag->destructor = sock_wfree;
491                                 skb->truesize -= frag->truesize;
492                         }
493                 }
494
495                 /* Everything is OK. Generate! */
496
497                 err = 0;
498                 offset = 0;
499                 frag = skb_shinfo(skb)->frag_list;
500                 skb_shinfo(skb)->frag_list = NULL;
501                 skb->data_len = first_len - skb_headlen(skb);
502                 skb->len = first_len;
503                 iph->tot_len = htons(first_len);
504                 iph->frag_off = htons(IP_MF);
505                 ip_send_check(iph);
506
507                 for (;;) {
508                         /* Prepare header of the next frame,
509                          * before previous one went down. */
510                         if (frag) {
511                                 frag->ip_summed = CHECKSUM_NONE;
512                                 frag->h.raw = frag->data;
513                                 frag->nh.raw = __skb_push(frag, hlen);
514                                 memcpy(frag->nh.raw, iph, hlen);
515                                 iph = frag->nh.iph;
516                                 iph->tot_len = htons(frag->len);
517                                 ip_copy_metadata(frag, skb);
518                                 if (offset == 0)
519                                         ip_options_fragment(frag);
520                                 offset += skb->len - hlen;
521                                 iph->frag_off = htons(offset>>3);
522                                 if (frag->next != NULL)
523                                         iph->frag_off |= htons(IP_MF);
524                                 /* Ready, complete checksum */
525                                 ip_send_check(iph);
526                         }
527
528                         err = output(skb);
529
530                         if (!err)
531                                 IP_INC_STATS(IPSTATS_MIB_FRAGCREATES);
532                         if (err || !frag)
533                                 break;
534
535                         skb = frag;
536                         frag = skb->next;
537                         skb->next = NULL;
538                 }
539
540                 if (err == 0) {
541                         IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
542                         return 0;
543                 }
544
545                 while (frag) {
546                         skb = frag->next;
547                         kfree_skb(frag);
548                         frag = skb;
549                 }
550                 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
551                 return err;
552         }
553
554 slow_path:
555         left = skb->len - hlen;         /* Space per frame */
556         ptr = raw + hlen;               /* Where to start from */
557
558 #ifdef CONFIG_BRIDGE_NETFILTER
559         /* for bridged IP traffic encapsulated inside f.e. a vlan header,
560          * we need to make room for the encapsulating header */
561         ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, nf_bridge_pad(skb));
562         mtu -= nf_bridge_pad(skb);
563 #else
564         ll_rs = LL_RESERVED_SPACE(rt->u.dst.dev);
565 #endif
566         /*
567          *      Fragment the datagram.
568          */
569
570         offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
571         not_last_frag = iph->frag_off & htons(IP_MF);
572
573         /*
574          *      Keep copying data until we run out.
575          */
576
577         while(left > 0) {
578                 len = left;
579                 /* IF: it doesn't fit, use 'mtu' - the data space left */
580                 if (len > mtu)
581                         len = mtu;
582                 /* IF: we are not sending upto and including the packet end
583                    then align the next start on an eight byte boundary */
584                 if (len < left) {
585                         len &= ~7;
586                 }
587                 /*
588                  *      Allocate buffer.
589                  */
590
591                 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
592                         NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
593                         err = -ENOMEM;
594                         goto fail;
595                 }
596
597                 /*
598                  *      Set up data on packet
599                  */
600
601                 ip_copy_metadata(skb2, skb);
602                 skb_reserve(skb2, ll_rs);
603                 skb_put(skb2, len + hlen);
604                 skb2->nh.raw = skb2->data;
605                 skb2->h.raw = skb2->data + hlen;
606
607                 /*
608                  *      Charge the memory for the fragment to any owner
609                  *      it might possess
610                  */
611
612                 if (skb->sk)
613                         skb_set_owner_w(skb2, skb->sk);
614
615                 /*
616                  *      Copy the packet header into the new buffer.
617                  */
618
619                 memcpy(skb2->nh.raw, skb->data, hlen);
620
621                 /*
622                  *      Copy a block of the IP datagram.
623                  */
624                 if (skb_copy_bits(skb, ptr, skb2->h.raw, len))
625                         BUG();
626                 left -= len;
627
628                 /*
629                  *      Fill in the new header fields.
630                  */
631                 iph = skb2->nh.iph;
632                 iph->frag_off = htons((offset >> 3));
633
634                 /* ANK: dirty, but effective trick. Upgrade options only if
635                  * the segment to be fragmented was THE FIRST (otherwise,
636                  * options are already fixed) and make it ONCE
637                  * on the initial skb, so that all the following fragments
638                  * will inherit fixed options.
639                  */
640                 if (offset == 0)
641                         ip_options_fragment(skb);
642
643                 /*
644                  *      Added AC : If we are fragmenting a fragment that's not the
645                  *                 last fragment then keep MF on each bit
646                  */
647                 if (left > 0 || not_last_frag)
648                         iph->frag_off |= htons(IP_MF);
649                 ptr += len;
650                 offset += len;
651
652                 /*
653                  *      Put this fragment into the sending queue.
654                  */
655                 iph->tot_len = htons(len + hlen);
656
657                 ip_send_check(iph);
658
659                 err = output(skb2);
660                 if (err)
661                         goto fail;
662
663                 IP_INC_STATS(IPSTATS_MIB_FRAGCREATES);
664         }
665         kfree_skb(skb);
666         IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
667         return err;
668
669 fail:
670         kfree_skb(skb); 
671         IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
672         return err;
673 }
674
675 EXPORT_SYMBOL(ip_fragment);
676
677 int
678 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
679 {
680         struct iovec *iov = from;
681
682         if (skb->ip_summed == CHECKSUM_HW) {
683                 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
684                         return -EFAULT;
685         } else {
686                 unsigned int csum = 0;
687                 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
688                         return -EFAULT;
689                 skb->csum = csum_block_add(skb->csum, csum, odd);
690         }
691         return 0;
692 }
693
694 static inline unsigned int
695 csum_page(struct page *page, int offset, int copy)
696 {
697         char *kaddr;
698         unsigned int csum;
699         kaddr = kmap(page);
700         csum = csum_partial(kaddr + offset, copy, 0);
701         kunmap(page);
702         return csum;
703 }
704
705 static inline int ip_ufo_append_data(struct sock *sk,
706                         int getfrag(void *from, char *to, int offset, int len,
707                                int odd, struct sk_buff *skb),
708                         void *from, int length, int hh_len, int fragheaderlen,
709                         int transhdrlen, int mtu,unsigned int flags)
710 {
711         struct sk_buff *skb;
712         int err;
713
714         /* There is support for UDP fragmentation offload by network
715          * device, so create one single skb packet containing complete
716          * udp datagram
717          */
718         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL) {
719                 skb = sock_alloc_send_skb(sk,
720                         hh_len + fragheaderlen + transhdrlen + 20,
721                         (flags & MSG_DONTWAIT), &err);
722
723                 if (skb == NULL)
724                         return err;
725
726                 /* reserve space for Hardware header */
727                 skb_reserve(skb, hh_len);
728
729                 /* create space for UDP/IP header */
730                 skb_put(skb,fragheaderlen + transhdrlen);
731
732                 /* initialize network header pointer */
733                 skb->nh.raw = skb->data;
734
735                 /* initialize protocol header pointer */
736                 skb->h.raw = skb->data + fragheaderlen;
737
738                 skb->ip_summed = CHECKSUM_HW;
739                 skb->csum = 0;
740                 sk->sk_sndmsg_off = 0;
741         }
742
743         err = skb_append_datato_frags(sk,skb, getfrag, from,
744                                (length - transhdrlen));
745         if (!err) {
746                 /* specify the length of each IP datagram fragment*/
747                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
748                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
749                 __skb_queue_tail(&sk->sk_write_queue, skb);
750
751                 return 0;
752         }
753         /* There is not enough support do UFO ,
754          * so follow normal path
755          */
756         kfree_skb(skb);
757         return err;
758 }
759
760 /*
761  *      ip_append_data() and ip_append_page() can make one large IP datagram
762  *      from many pieces of data. Each pieces will be holded on the socket
763  *      until ip_push_pending_frames() is called. Each piece can be a page
764  *      or non-page data.
765  *      
766  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
767  *      this interface potentially.
768  *
769  *      LATER: length must be adjusted by pad at tail, when it is required.
770  */
771 int ip_append_data(struct sock *sk,
772                    int getfrag(void *from, char *to, int offset, int len,
773                                int odd, struct sk_buff *skb),
774                    void *from, int length, int transhdrlen,
775                    struct ipcm_cookie *ipc, struct rtable *rt,
776                    unsigned int flags)
777 {
778         struct inet_sock *inet = inet_sk(sk);
779         struct sk_buff *skb;
780
781         struct ip_options *opt = NULL;
782         int hh_len;
783         int exthdrlen;
784         int mtu;
785         int copy;
786         int err;
787         int offset = 0;
788         unsigned int maxfraglen, fragheaderlen;
789         int csummode = CHECKSUM_NONE;
790
791         if (flags&MSG_PROBE)
792                 return 0;
793
794         if (skb_queue_empty(&sk->sk_write_queue)) {
795                 /*
796                  * setup for corking.
797                  */
798                 opt = ipc->opt;
799                 if (opt) {
800                         if (inet->cork.opt == NULL) {
801                                 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
802                                 if (unlikely(inet->cork.opt == NULL))
803                                         return -ENOBUFS;
804                         }
805                         memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
806                         inet->cork.flags |= IPCORK_OPT;
807                         inet->cork.addr = ipc->addr;
808                 }
809                 dst_hold(&rt->u.dst);
810                 inet->cork.fragsize = mtu = dst_mtu(rt->u.dst.path);
811                 inet->cork.rt = rt;
812                 inet->cork.length = 0;
813                 sk->sk_sndmsg_page = NULL;
814                 sk->sk_sndmsg_off = 0;
815                 if ((exthdrlen = rt->u.dst.header_len) != 0) {
816                         length += exthdrlen;
817                         transhdrlen += exthdrlen;
818                 }
819         } else {
820                 rt = inet->cork.rt;
821                 if (inet->cork.flags & IPCORK_OPT)
822                         opt = inet->cork.opt;
823
824                 transhdrlen = 0;
825                 exthdrlen = 0;
826                 mtu = inet->cork.fragsize;
827         }
828         hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
829
830         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
831         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
832
833         if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
834                 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu-exthdrlen);
835                 return -EMSGSIZE;
836         }
837
838         /*
839          * transhdrlen > 0 means that this is the first fragment and we wish
840          * it won't be fragmented in the future.
841          */
842         if (transhdrlen &&
843             length + fragheaderlen <= mtu &&
844             rt->u.dst.dev->features & NETIF_F_ALL_CSUM &&
845             !exthdrlen)
846                 csummode = CHECKSUM_HW;
847
848         inet->cork.length += length;
849         if (((length > mtu) && (sk->sk_protocol == IPPROTO_UDP)) &&
850                         (rt->u.dst.dev->features & NETIF_F_UFO)) {
851
852                 err = ip_ufo_append_data(sk, getfrag, from, length, hh_len,
853                                          fragheaderlen, transhdrlen, mtu,
854                                          flags);
855                 if (err)
856                         goto error;
857                 return 0;
858         }
859
860         /* So, what's going on in the loop below?
861          *
862          * We use calculated fragment length to generate chained skb,
863          * each of segments is IP fragment ready for sending to network after
864          * adding appropriate IP header.
865          */
866
867         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
868                 goto alloc_new_skb;
869
870         while (length > 0) {
871                 /* Check if the remaining data fits into current packet. */
872                 copy = mtu - skb->len;
873                 if (copy < length)
874                         copy = maxfraglen - skb->len;
875                 if (copy <= 0) {
876                         char *data;
877                         unsigned int datalen;
878                         unsigned int fraglen;
879                         unsigned int fraggap;
880                         unsigned int alloclen;
881                         struct sk_buff *skb_prev;
882 alloc_new_skb:
883                         skb_prev = skb;
884                         if (skb_prev)
885                                 fraggap = skb_prev->len - maxfraglen;
886                         else
887                                 fraggap = 0;
888
889                         /*
890                          * If remaining data exceeds the mtu,
891                          * we know we need more fragment(s).
892                          */
893                         datalen = length + fraggap;
894                         if (datalen > mtu - fragheaderlen)
895                                 datalen = maxfraglen - fragheaderlen;
896                         fraglen = datalen + fragheaderlen;
897
898                         if ((flags & MSG_MORE) && 
899                             !(rt->u.dst.dev->features&NETIF_F_SG))
900                                 alloclen = mtu;
901                         else
902                                 alloclen = datalen + fragheaderlen;
903
904                         /* The last fragment gets additional space at tail.
905                          * Note, with MSG_MORE we overallocate on fragments,
906                          * because we have no idea what fragment will be
907                          * the last.
908                          */
909                         if (datalen == length + fraggap)
910                                 alloclen += rt->u.dst.trailer_len;
911
912                         if (transhdrlen) {
913                                 skb = sock_alloc_send_skb(sk, 
914                                                 alloclen + hh_len + 15,
915                                                 (flags & MSG_DONTWAIT), &err);
916                         } else {
917                                 skb = NULL;
918                                 if (atomic_read(&sk->sk_wmem_alloc) <=
919                                     2 * sk->sk_sndbuf)
920                                         skb = sock_wmalloc(sk, 
921                                                            alloclen + hh_len + 15, 1,
922                                                            sk->sk_allocation);
923                                 if (unlikely(skb == NULL))
924                                         err = -ENOBUFS;
925                         }
926                         if (skb == NULL)
927                                 goto error;
928
929                         /*
930                          *      Fill in the control structures
931                          */
932                         skb->ip_summed = csummode;
933                         skb->csum = 0;
934                         skb_reserve(skb, hh_len);
935
936                         /*
937                          *      Find where to start putting bytes.
938                          */
939                         data = skb_put(skb, fraglen);
940                         skb->nh.raw = data + exthdrlen;
941                         data += fragheaderlen;
942                         skb->h.raw = data + exthdrlen;
943
944                         if (fraggap) {
945                                 skb->csum = skb_copy_and_csum_bits(
946                                         skb_prev, maxfraglen,
947                                         data + transhdrlen, fraggap, 0);
948                                 skb_prev->csum = csum_sub(skb_prev->csum,
949                                                           skb->csum);
950                                 data += fraggap;
951                                 pskb_trim_unique(skb_prev, maxfraglen);
952                         }
953
954                         copy = datalen - transhdrlen - fraggap;
955                         if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
956                                 err = -EFAULT;
957                                 kfree_skb(skb);
958                                 goto error;
959                         }
960
961                         offset += copy;
962                         length -= datalen - fraggap;
963                         transhdrlen = 0;
964                         exthdrlen = 0;
965                         csummode = CHECKSUM_NONE;
966
967                         /*
968                          * Put the packet on the pending queue.
969                          */
970                         __skb_queue_tail(&sk->sk_write_queue, skb);
971                         continue;
972                 }
973
974                 if (copy > length)
975                         copy = length;
976
977                 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
978                         unsigned int off;
979
980                         off = skb->len;
981                         if (getfrag(from, skb_put(skb, copy), 
982                                         offset, copy, off, skb) < 0) {
983                                 __skb_trim(skb, off);
984                                 err = -EFAULT;
985                                 goto error;
986                         }
987                 } else {
988                         int i = skb_shinfo(skb)->nr_frags;
989                         skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
990                         struct page *page = sk->sk_sndmsg_page;
991                         int off = sk->sk_sndmsg_off;
992                         unsigned int left;
993
994                         if (page && (left = PAGE_SIZE - off) > 0) {
995                                 if (copy >= left)
996                                         copy = left;
997                                 if (page != frag->page) {
998                                         if (i == MAX_SKB_FRAGS) {
999                                                 err = -EMSGSIZE;
1000                                                 goto error;
1001                                         }
1002                                         get_page(page);
1003                                         skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
1004                                         frag = &skb_shinfo(skb)->frags[i];
1005                                 }
1006                         } else if (i < MAX_SKB_FRAGS) {
1007                                 if (copy > PAGE_SIZE)
1008                                         copy = PAGE_SIZE;
1009                                 page = alloc_pages(sk->sk_allocation, 0);
1010                                 if (page == NULL)  {
1011                                         err = -ENOMEM;
1012                                         goto error;
1013                                 }
1014                                 sk->sk_sndmsg_page = page;
1015                                 sk->sk_sndmsg_off = 0;
1016
1017                                 skb_fill_page_desc(skb, i, page, 0, 0);
1018                                 frag = &skb_shinfo(skb)->frags[i];
1019                                 skb->truesize += PAGE_SIZE;
1020                                 atomic_add(PAGE_SIZE, &sk->sk_wmem_alloc);
1021                         } else {
1022                                 err = -EMSGSIZE;
1023                                 goto error;
1024                         }
1025                         if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1026                                 err = -EFAULT;
1027                                 goto error;
1028                         }
1029                         sk->sk_sndmsg_off += copy;
1030                         frag->size += copy;
1031                         skb->len += copy;
1032                         skb->data_len += copy;
1033                 }
1034                 offset += copy;
1035                 length -= copy;
1036         }
1037
1038         return 0;
1039
1040 error:
1041         inet->cork.length -= length;
1042         IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1043         return err; 
1044 }
1045
1046 ssize_t ip_append_page(struct sock *sk, struct page *page,
1047                        int offset, size_t size, int flags)
1048 {
1049         struct inet_sock *inet = inet_sk(sk);
1050         struct sk_buff *skb;
1051         struct rtable *rt;
1052         struct ip_options *opt = NULL;
1053         int hh_len;
1054         int mtu;
1055         int len;
1056         int err;
1057         unsigned int maxfraglen, fragheaderlen, fraggap;
1058
1059         if (inet->hdrincl)
1060                 return -EPERM;
1061
1062         if (flags&MSG_PROBE)
1063                 return 0;
1064
1065         if (skb_queue_empty(&sk->sk_write_queue))
1066                 return -EINVAL;
1067
1068         rt = inet->cork.rt;
1069         if (inet->cork.flags & IPCORK_OPT)
1070                 opt = inet->cork.opt;
1071
1072         if (!(rt->u.dst.dev->features&NETIF_F_SG))
1073                 return -EOPNOTSUPP;
1074
1075         hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1076         mtu = inet->cork.fragsize;
1077
1078         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1079         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1080
1081         if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1082                 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu);
1083                 return -EMSGSIZE;
1084         }
1085
1086         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1087                 return -EINVAL;
1088
1089         inet->cork.length += size;
1090         if ((sk->sk_protocol == IPPROTO_UDP) &&
1091             (rt->u.dst.dev->features & NETIF_F_UFO)) {
1092                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1093                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1094         }
1095
1096
1097         while (size > 0) {
1098                 int i;
1099
1100                 if (skb_is_gso(skb))
1101                         len = size;
1102                 else {
1103
1104                         /* Check if the remaining data fits into current packet. */
1105                         len = mtu - skb->len;
1106                         if (len < size)
1107                                 len = maxfraglen - skb->len;
1108                 }
1109                 if (len <= 0) {
1110                         struct sk_buff *skb_prev;
1111                         char *data;
1112                         struct iphdr *iph;
1113                         int alloclen;
1114
1115                         skb_prev = skb;
1116                         fraggap = skb_prev->len - maxfraglen;
1117
1118                         alloclen = fragheaderlen + hh_len + fraggap + 15;
1119                         skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1120                         if (unlikely(!skb)) {
1121                                 err = -ENOBUFS;
1122                                 goto error;
1123                         }
1124
1125                         /*
1126                          *      Fill in the control structures
1127                          */
1128                         skb->ip_summed = CHECKSUM_NONE;
1129                         skb->csum = 0;
1130                         skb_reserve(skb, hh_len);
1131
1132                         /*
1133                          *      Find where to start putting bytes.
1134                          */
1135                         data = skb_put(skb, fragheaderlen + fraggap);
1136                         skb->nh.iph = iph = (struct iphdr *)data;
1137                         data += fragheaderlen;
1138                         skb->h.raw = data;
1139
1140                         if (fraggap) {
1141                                 skb->csum = skb_copy_and_csum_bits(
1142                                         skb_prev, maxfraglen,
1143                                         data, fraggap, 0);
1144                                 skb_prev->csum = csum_sub(skb_prev->csum,
1145                                                           skb->csum);
1146                                 pskb_trim_unique(skb_prev, maxfraglen);
1147                         }
1148
1149                         /*
1150                          * Put the packet on the pending queue.
1151                          */
1152                         __skb_queue_tail(&sk->sk_write_queue, skb);
1153                         continue;
1154                 }
1155
1156                 i = skb_shinfo(skb)->nr_frags;
1157                 if (len > size)
1158                         len = size;
1159                 if (skb_can_coalesce(skb, i, page, offset)) {
1160                         skb_shinfo(skb)->frags[i-1].size += len;
1161                 } else if (i < MAX_SKB_FRAGS) {
1162                         get_page(page);
1163                         skb_fill_page_desc(skb, i, page, offset, len);
1164                 } else {
1165                         err = -EMSGSIZE;
1166                         goto error;
1167                 }
1168
1169                 if (skb->ip_summed == CHECKSUM_NONE) {
1170                         unsigned int csum;
1171                         csum = csum_page(page, offset, len);
1172                         skb->csum = csum_block_add(skb->csum, csum, skb->len);
1173                 }
1174
1175                 skb->len += len;
1176                 skb->data_len += len;
1177                 offset += len;
1178                 size -= len;
1179         }
1180         return 0;
1181
1182 error:
1183         inet->cork.length -= size;
1184         IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1185         return err;
1186 }
1187
1188 /*
1189  *      Combined all pending IP fragments on the socket as one IP datagram
1190  *      and push them out.
1191  */
1192 int ip_push_pending_frames(struct sock *sk)
1193 {
1194         struct sk_buff *skb, *tmp_skb;
1195         struct sk_buff **tail_skb;
1196         struct inet_sock *inet = inet_sk(sk);
1197         struct ip_options *opt = NULL;
1198         struct rtable *rt = inet->cork.rt;
1199         struct iphdr *iph;
1200         __be16 df = 0;
1201         __u8 ttl;
1202         int err = 0;
1203
1204         if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1205                 goto out;
1206         tail_skb = &(skb_shinfo(skb)->frag_list);
1207
1208         /* move skb->data to ip header from ext header */
1209         if (skb->data < skb->nh.raw)
1210                 __skb_pull(skb, skb->nh.raw - skb->data);
1211         while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1212                 __skb_pull(tmp_skb, skb->h.raw - skb->nh.raw);
1213                 *tail_skb = tmp_skb;
1214                 tail_skb = &(tmp_skb->next);
1215                 skb->len += tmp_skb->len;
1216                 skb->data_len += tmp_skb->len;
1217                 skb->truesize += tmp_skb->truesize;
1218                 __sock_put(tmp_skb->sk);
1219                 tmp_skb->destructor = NULL;
1220                 tmp_skb->sk = NULL;
1221         }
1222
1223         /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1224          * to fragment the frame generated here. No matter, what transforms
1225          * how transforms change size of the packet, it will come out.
1226          */
1227         if (inet->pmtudisc != IP_PMTUDISC_DO)
1228                 skb->local_df = 1;
1229
1230         /* DF bit is set when we want to see DF on outgoing frames.
1231          * If local_df is set too, we still allow to fragment this frame
1232          * locally. */
1233         if (inet->pmtudisc == IP_PMTUDISC_DO ||
1234             (skb->len <= dst_mtu(&rt->u.dst) &&
1235              ip_dont_fragment(sk, &rt->u.dst)))
1236                 df = htons(IP_DF);
1237
1238         if (inet->cork.flags & IPCORK_OPT)
1239                 opt = inet->cork.opt;
1240
1241         if (rt->rt_type == RTN_MULTICAST)
1242                 ttl = inet->mc_ttl;
1243         else
1244                 ttl = ip_select_ttl(inet, &rt->u.dst);
1245
1246         iph = (struct iphdr *)skb->data;
1247         iph->version = 4;
1248         iph->ihl = 5;
1249         if (opt) {
1250                 iph->ihl += opt->optlen>>2;
1251                 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1252         }
1253         iph->tos = inet->tos;
1254         iph->tot_len = htons(skb->len);
1255         iph->frag_off = df;
1256         ip_select_ident(iph, &rt->u.dst, sk);
1257         iph->ttl = ttl;
1258         iph->protocol = sk->sk_protocol;
1259         iph->saddr = rt->rt_src;
1260         iph->daddr = rt->rt_dst;
1261         ip_send_check(iph);
1262
1263         skb->priority = sk->sk_priority;
1264         skb->dst = dst_clone(&rt->u.dst);
1265
1266         /* Netfilter gets whole the not fragmented skb. */
1267         err = NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, 
1268                       skb->dst->dev, dst_output);
1269         if (err) {
1270                 if (err > 0)
1271                         err = inet->recverr ? net_xmit_errno(err) : 0;
1272                 if (err)
1273                         goto error;
1274         }
1275
1276 out:
1277         inet->cork.flags &= ~IPCORK_OPT;
1278         kfree(inet->cork.opt);
1279         inet->cork.opt = NULL;
1280         if (inet->cork.rt) {
1281                 ip_rt_put(inet->cork.rt);
1282                 inet->cork.rt = NULL;
1283         }
1284         return err;
1285
1286 error:
1287         IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1288         goto out;
1289 }
1290
1291 /*
1292  *      Throw away all pending data on the socket.
1293  */
1294 void ip_flush_pending_frames(struct sock *sk)
1295 {
1296         struct inet_sock *inet = inet_sk(sk);
1297         struct sk_buff *skb;
1298
1299         while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1300                 kfree_skb(skb);
1301
1302         inet->cork.flags &= ~IPCORK_OPT;
1303         kfree(inet->cork.opt);
1304         inet->cork.opt = NULL;
1305         if (inet->cork.rt) {
1306                 ip_rt_put(inet->cork.rt);
1307                 inet->cork.rt = NULL;
1308         }
1309 }
1310
1311
1312 /*
1313  *      Fetch data from kernel space and fill in checksum if needed.
1314  */
1315 static int ip_reply_glue_bits(void *dptr, char *to, int offset, 
1316                               int len, int odd, struct sk_buff *skb)
1317 {
1318         unsigned int csum;
1319
1320         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1321         skb->csum = csum_block_add(skb->csum, csum, odd);
1322         return 0;  
1323 }
1324
1325 /* 
1326  *      Generic function to send a packet as reply to another packet.
1327  *      Used to send TCP resets so far. ICMP should use this function too.
1328  *
1329  *      Should run single threaded per socket because it uses the sock 
1330  *      structure to pass arguments.
1331  *
1332  *      LATER: switch from ip_build_xmit to ip_append_*
1333  */
1334 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1335                    unsigned int len)
1336 {
1337         struct inet_sock *inet = inet_sk(sk);
1338         struct {
1339                 struct ip_options       opt;
1340                 char                    data[40];
1341         } replyopts;
1342         struct ipcm_cookie ipc;
1343         u32 daddr;
1344         struct rtable *rt = (struct rtable*)skb->dst;
1345
1346         if (ip_options_echo(&replyopts.opt, skb))
1347                 return;
1348
1349         daddr = ipc.addr = rt->rt_src;
1350         ipc.opt = NULL;
1351
1352         if (replyopts.opt.optlen) {
1353                 ipc.opt = &replyopts.opt;
1354
1355                 if (ipc.opt->srr)
1356                         daddr = replyopts.opt.faddr;
1357         }
1358
1359         {
1360                 struct flowi fl = { .nl_u = { .ip4_u =
1361                                               { .daddr = daddr,
1362                                                 .saddr = rt->rt_spec_dst,
1363                                                 .tos = RT_TOS(skb->nh.iph->tos) } },
1364                                     /* Not quite clean, but right. */
1365                                     .uli_u = { .ports =
1366                                                { .sport = skb->h.th->dest,
1367                                                  .dport = skb->h.th->source } },
1368                                     .proto = sk->sk_protocol };
1369                 if (ip_route_output_key(&rt, &fl))
1370                         return;
1371         }
1372
1373         /* And let IP do all the hard work.
1374
1375            This chunk is not reenterable, hence spinlock.
1376            Note that it uses the fact, that this function is called
1377            with locally disabled BH and that sk cannot be already spinlocked.
1378          */
1379         bh_lock_sock(sk);
1380         inet->tos = skb->nh.iph->tos;
1381         sk->sk_priority = skb->priority;
1382         sk->sk_protocol = skb->nh.iph->protocol;
1383         ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1384                        &ipc, rt, MSG_DONTWAIT);
1385         if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1386                 if (arg->csumoffset >= 0)
1387                         *((u16 *)skb->h.raw + arg->csumoffset) = csum_fold(csum_add(skb->csum, arg->csum));
1388                 skb->ip_summed = CHECKSUM_NONE;
1389                 ip_push_pending_frames(sk);
1390         }
1391
1392         bh_unlock_sock(sk);
1393
1394         ip_rt_put(rt);
1395 }
1396
1397 void __init ip_init(void)
1398 {
1399         ip_rt_init();
1400         inet_initpeers();
1401
1402 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1403         igmp_mc_proc_init();
1404 #endif
1405 }
1406
1407 EXPORT_SYMBOL(ip_generic_getfrag);
1408 EXPORT_SYMBOL(ip_queue_xmit);
1409 EXPORT_SYMBOL(ip_send_check);