[ARM] Support register switch in nommu mode
[linux-2.6/verdex.git] / net / ipv4 / ip_output.c
blob3324fbfe528a048b279c7b7f44e25810a750dbca
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.
6 * The Internet Protocol (IP) output module.
8 * Version: $Id: ip_output.c,v 1.100 2002/02/01 22:01:03 davem Exp $
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>
20 * See ip_input.c for original log
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.
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 #include <linux/config.h>
58 #include <linux/socket.h>
59 #include <linux/sockios.h>
60 #include <linux/in.h>
61 #include <linux/inet.h>
62 #include <linux/netdevice.h>
63 #include <linux/etherdevice.h>
64 #include <linux/proc_fs.h>
65 #include <linux/stat.h>
66 #include <linux/init.h>
68 #include <net/snmp.h>
69 #include <net/ip.h>
70 #include <net/protocol.h>
71 #include <net/route.h>
72 #include <net/xfrm.h>
73 #include <linux/skbuff.h>
74 #include <net/sock.h>
75 #include <net/arp.h>
76 #include <net/icmp.h>
77 #include <net/checksum.h>
78 #include <net/inetpeer.h>
79 #include <net/checksum.h>
80 #include <linux/igmp.h>
81 #include <linux/netfilter_ipv4.h>
82 #include <linux/netfilter_bridge.h>
83 #include <linux/mroute.h>
84 #include <linux/netlink.h>
85 #include <linux/tcp.h>
87 int sysctl_ip_default_ttl = IPDEFTTL;
89 static int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*));
91 /* Generate a checksum for an outgoing IP datagram. */
92 __inline__ void ip_send_check(struct iphdr *iph)
94 iph->check = 0;
95 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
98 /* dev_loopback_xmit for use with netfilter. */
99 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
101 newskb->mac.raw = newskb->data;
102 __skb_pull(newskb, newskb->nh.raw - newskb->data);
103 newskb->pkt_type = PACKET_LOOPBACK;
104 newskb->ip_summed = CHECKSUM_UNNECESSARY;
105 BUG_TRAP(newskb->dst);
106 netif_rx(newskb);
107 return 0;
110 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
112 int ttl = inet->uc_ttl;
114 if (ttl < 0)
115 ttl = dst_metric(dst, RTAX_HOPLIMIT);
116 return ttl;
120 * Add an ip header to a skbuff and send it out.
123 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
124 u32 saddr, u32 daddr, struct ip_options *opt)
126 struct inet_sock *inet = inet_sk(sk);
127 struct rtable *rt = (struct rtable *)skb->dst;
128 struct iphdr *iph;
130 /* Build the IP header. */
131 if (opt)
132 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr) + opt->optlen);
133 else
134 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr));
136 iph->version = 4;
137 iph->ihl = 5;
138 iph->tos = inet->tos;
139 if (ip_dont_fragment(sk, &rt->u.dst))
140 iph->frag_off = htons(IP_DF);
141 else
142 iph->frag_off = 0;
143 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
144 iph->daddr = rt->rt_dst;
145 iph->saddr = rt->rt_src;
146 iph->protocol = sk->sk_protocol;
147 iph->tot_len = htons(skb->len);
148 ip_select_ident(iph, &rt->u.dst, sk);
149 skb->nh.iph = iph;
151 if (opt && opt->optlen) {
152 iph->ihl += opt->optlen>>2;
153 ip_options_build(skb, opt, daddr, rt, 0);
155 ip_send_check(iph);
157 skb->priority = sk->sk_priority;
159 /* Send it out. */
160 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
161 dst_output);
164 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
166 static inline int ip_finish_output2(struct sk_buff *skb)
168 struct dst_entry *dst = skb->dst;
169 struct hh_cache *hh = dst->hh;
170 struct net_device *dev = dst->dev;
171 int hh_len = LL_RESERVED_SPACE(dev);
173 /* Be paranoid, rather than too clever. */
174 if (unlikely(skb_headroom(skb) < hh_len && dev->hard_header)) {
175 struct sk_buff *skb2;
177 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
178 if (skb2 == NULL) {
179 kfree_skb(skb);
180 return -ENOMEM;
182 if (skb->sk)
183 skb_set_owner_w(skb2, skb->sk);
184 kfree_skb(skb);
185 skb = skb2;
188 if (hh) {
189 int hh_alen;
191 read_lock_bh(&hh->hh_lock);
192 hh_alen = HH_DATA_ALIGN(hh->hh_len);
193 memcpy(skb->data - hh_alen, hh->hh_data, hh_alen);
194 read_unlock_bh(&hh->hh_lock);
195 skb_push(skb, hh->hh_len);
196 return hh->hh_output(skb);
197 } else if (dst->neighbour)
198 return dst->neighbour->output(skb);
200 if (net_ratelimit())
201 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
202 kfree_skb(skb);
203 return -EINVAL;
206 static inline int ip_finish_output(struct sk_buff *skb)
208 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
209 /* Policy lookup after SNAT yielded a new policy */
210 if (skb->dst->xfrm != NULL)
211 return xfrm4_output_finish(skb);
212 #endif
213 if (skb->len > dst_mtu(skb->dst) &&
214 !(skb_shinfo(skb)->ufo_size || skb_shinfo(skb)->tso_size))
215 return ip_fragment(skb, ip_finish_output2);
216 else
217 return ip_finish_output2(skb);
220 int ip_mc_output(struct sk_buff *skb)
222 struct sock *sk = skb->sk;
223 struct rtable *rt = (struct rtable*)skb->dst;
224 struct net_device *dev = rt->u.dst.dev;
227 * If the indicated interface is up and running, send the packet.
229 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
231 skb->dev = dev;
232 skb->protocol = htons(ETH_P_IP);
235 * Multicasts are looped back for other local users
238 if (rt->rt_flags&RTCF_MULTICAST) {
239 if ((!sk || inet_sk(sk)->mc_loop)
240 #ifdef CONFIG_IP_MROUTE
241 /* Small optimization: do not loopback not local frames,
242 which returned after forwarding; they will be dropped
243 by ip_mr_input in any case.
244 Note, that local frames are looped back to be delivered
245 to local recipients.
247 This check is duplicated in ip_mr_input at the moment.
249 && ((rt->rt_flags&RTCF_LOCAL) || !(IPCB(skb)->flags&IPSKB_FORWARDED))
250 #endif
252 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
253 if (newskb)
254 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
255 newskb->dev,
256 ip_dev_loopback_xmit);
259 /* Multicasts with ttl 0 must not go beyond the host */
261 if (skb->nh.iph->ttl == 0) {
262 kfree_skb(skb);
263 return 0;
267 if (rt->rt_flags&RTCF_BROADCAST) {
268 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
269 if (newskb)
270 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
271 newskb->dev, ip_dev_loopback_xmit);
274 return NF_HOOK(PF_INET, NF_IP_POST_ROUTING, skb, NULL, skb->dev,
275 ip_finish_output);
278 int ip_output(struct sk_buff *skb)
280 struct net_device *dev = skb->dst->dev;
282 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
284 skb->dev = dev;
285 skb->protocol = htons(ETH_P_IP);
287 return NF_HOOK(PF_INET, NF_IP_POST_ROUTING, skb, NULL, dev,
288 ip_finish_output);
291 int ip_queue_xmit(struct sk_buff *skb, int ipfragok)
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;
299 /* Skip all of this if the packet is already routed,
300 * f.e. by something like SCTP.
302 rt = (struct rtable *) skb->dst;
303 if (rt != NULL)
304 goto packet_routed;
306 /* Make sure we can route this packet. */
307 rt = (struct rtable *)__sk_dst_check(sk, 0);
308 if (rt == NULL) {
309 u32 daddr;
311 /* Use correct destination address if we have options. */
312 daddr = inet->daddr;
313 if(opt && opt->srr)
314 daddr = opt->faddr;
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 } } };
327 /* If this fails, retransmit mechanism of transport layer will
328 * keep trying until route appears or the connection times
329 * itself out.
331 if (ip_route_output_flow(&rt, &fl, sk, 0))
332 goto no_route;
334 sk_setup_caps(sk, &rt->u.dst);
336 skb->dst = dst_clone(&rt->u.dst);
338 packet_routed:
339 if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
340 goto no_route;
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. */
357 if (opt && opt->optlen) {
358 iph->ihl += opt->optlen >> 2;
359 ip_options_build(skb, opt, inet->daddr, rt, 0);
362 ip_select_ident_more(iph, &rt->u.dst, sk,
363 (skb_shinfo(skb)->tso_segs ?: 1) - 1);
365 /* Add an IP checksum. */
366 ip_send_check(iph);
368 skb->priority = sk->sk_priority;
370 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
371 dst_output);
373 no_route:
374 IP_INC_STATS(IPSTATS_MIB_OUTNOROUTES);
375 kfree_skb(skb);
376 return -EHOSTUNREACH;
380 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
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;
389 /* Copy the flags to each fragment. */
390 IPCB(to)->flags = IPCB(from)->flags;
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
414 * This IP datagram is too large to be sent in one piece. Break it up into
415 * smaller pieces (each of size equal to IP header plus
416 * a block of the data of the original IP data part) that will yet fit in a
417 * single device frame, and queue such a frame for sending.
420 static int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*))
422 struct iphdr *iph;
423 int raw = 0;
424 int ptr;
425 struct net_device *dev;
426 struct sk_buff *skb2;
427 unsigned int mtu, hlen, left, len, ll_rs;
428 int offset;
429 __be16 not_last_frag;
430 struct rtable *rt = (struct rtable*)skb->dst;
431 int err = 0;
433 dev = rt->u.dst.dev;
436 * Point into the IP datagram header.
439 iph = skb->nh.iph;
441 if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
442 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
443 htonl(dst_mtu(&rt->u.dst)));
444 kfree_skb(skb);
445 return -EMSGSIZE;
449 * Setup starting values.
452 hlen = iph->ihl * 4;
453 mtu = dst_mtu(&rt->u.dst) - hlen; /* Size of data space */
454 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
456 /* When frag_list is given, use it. First, check its validity:
457 * some transformers could create wrong frag_list or break existing
458 * one, it is not prohibited. In this case fall back to copying.
460 * LATER: this step can be merged to real generation of fragments,
461 * we can switch to copy when see the first bad fragment.
463 if (skb_shinfo(skb)->frag_list) {
464 struct sk_buff *frag;
465 int first_len = skb_pagelen(skb);
467 if (first_len - hlen > mtu ||
468 ((first_len - hlen) & 7) ||
469 (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
470 skb_cloned(skb))
471 goto slow_path;
473 for (frag = skb_shinfo(skb)->frag_list; frag; frag = frag->next) {
474 /* Correct geometry. */
475 if (frag->len > mtu ||
476 ((frag->len & 7) && frag->next) ||
477 skb_headroom(frag) < hlen)
478 goto slow_path;
480 /* Partially cloned skb? */
481 if (skb_shared(frag))
482 goto slow_path;
484 BUG_ON(frag->sk);
485 if (skb->sk) {
486 sock_hold(skb->sk);
487 frag->sk = skb->sk;
488 frag->destructor = sock_wfree;
489 skb->truesize -= frag->truesize;
493 /* Everything is OK. Generate! */
495 err = 0;
496 offset = 0;
497 frag = skb_shinfo(skb)->frag_list;
498 skb_shinfo(skb)->frag_list = NULL;
499 skb->data_len = first_len - skb_headlen(skb);
500 skb->len = first_len;
501 iph->tot_len = htons(first_len);
502 iph->frag_off = htons(IP_MF);
503 ip_send_check(iph);
505 for (;;) {
506 /* Prepare header of the next frame,
507 * before previous one went down. */
508 if (frag) {
509 frag->ip_summed = CHECKSUM_NONE;
510 frag->h.raw = frag->data;
511 frag->nh.raw = __skb_push(frag, hlen);
512 memcpy(frag->nh.raw, iph, hlen);
513 iph = frag->nh.iph;
514 iph->tot_len = htons(frag->len);
515 ip_copy_metadata(frag, skb);
516 if (offset == 0)
517 ip_options_fragment(frag);
518 offset += skb->len - hlen;
519 iph->frag_off = htons(offset>>3);
520 if (frag->next != NULL)
521 iph->frag_off |= htons(IP_MF);
522 /* Ready, complete checksum */
523 ip_send_check(iph);
526 err = output(skb);
528 if (err || !frag)
529 break;
531 skb = frag;
532 frag = skb->next;
533 skb->next = NULL;
536 if (err == 0) {
537 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
538 return 0;
541 while (frag) {
542 skb = frag->next;
543 kfree_skb(frag);
544 frag = skb;
546 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
547 return err;
550 slow_path:
551 left = skb->len - hlen; /* Space per frame */
552 ptr = raw + hlen; /* Where to start from */
554 #ifdef CONFIG_BRIDGE_NETFILTER
555 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
556 * we need to make room for the encapsulating header */
557 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, nf_bridge_pad(skb));
558 mtu -= nf_bridge_pad(skb);
559 #else
560 ll_rs = LL_RESERVED_SPACE(rt->u.dst.dev);
561 #endif
563 * Fragment the datagram.
566 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
567 not_last_frag = iph->frag_off & htons(IP_MF);
570 * Keep copying data until we run out.
573 while(left > 0) {
574 len = left;
575 /* IF: it doesn't fit, use 'mtu' - the data space left */
576 if (len > mtu)
577 len = mtu;
578 /* IF: we are not sending upto and including the packet end
579 then align the next start on an eight byte boundary */
580 if (len < left) {
581 len &= ~7;
584 * Allocate buffer.
587 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
588 NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
589 err = -ENOMEM;
590 goto fail;
594 * Set up data on packet
597 ip_copy_metadata(skb2, skb);
598 skb_reserve(skb2, ll_rs);
599 skb_put(skb2, len + hlen);
600 skb2->nh.raw = skb2->data;
601 skb2->h.raw = skb2->data + hlen;
604 * Charge the memory for the fragment to any owner
605 * it might possess
608 if (skb->sk)
609 skb_set_owner_w(skb2, skb->sk);
612 * Copy the packet header into the new buffer.
615 memcpy(skb2->nh.raw, skb->data, hlen);
618 * Copy a block of the IP datagram.
620 if (skb_copy_bits(skb, ptr, skb2->h.raw, len))
621 BUG();
622 left -= len;
625 * Fill in the new header fields.
627 iph = skb2->nh.iph;
628 iph->frag_off = htons((offset >> 3));
630 /* ANK: dirty, but effective trick. Upgrade options only if
631 * the segment to be fragmented was THE FIRST (otherwise,
632 * options are already fixed) and make it ONCE
633 * on the initial skb, so that all the following fragments
634 * will inherit fixed options.
636 if (offset == 0)
637 ip_options_fragment(skb);
640 * Added AC : If we are fragmenting a fragment that's not the
641 * last fragment then keep MF on each bit
643 if (left > 0 || not_last_frag)
644 iph->frag_off |= htons(IP_MF);
645 ptr += len;
646 offset += len;
649 * Put this fragment into the sending queue.
652 IP_INC_STATS(IPSTATS_MIB_FRAGCREATES);
654 iph->tot_len = htons(len + hlen);
656 ip_send_check(iph);
658 err = output(skb2);
659 if (err)
660 goto fail;
662 kfree_skb(skb);
663 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
664 return err;
666 fail:
667 kfree_skb(skb);
668 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
669 return err;
673 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
675 struct iovec *iov = from;
677 if (skb->ip_summed == CHECKSUM_HW) {
678 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
679 return -EFAULT;
680 } else {
681 unsigned int csum = 0;
682 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
683 return -EFAULT;
684 skb->csum = csum_block_add(skb->csum, csum, odd);
686 return 0;
689 static inline unsigned int
690 csum_page(struct page *page, int offset, int copy)
692 char *kaddr;
693 unsigned int csum;
694 kaddr = kmap(page);
695 csum = csum_partial(kaddr + offset, copy, 0);
696 kunmap(page);
697 return csum;
700 static inline int ip_ufo_append_data(struct sock *sk,
701 int getfrag(void *from, char *to, int offset, int len,
702 int odd, struct sk_buff *skb),
703 void *from, int length, int hh_len, int fragheaderlen,
704 int transhdrlen, int mtu,unsigned int flags)
706 struct sk_buff *skb;
707 int err;
709 /* There is support for UDP fragmentation offload by network
710 * device, so create one single skb packet containing complete
711 * udp datagram
713 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL) {
714 skb = sock_alloc_send_skb(sk,
715 hh_len + fragheaderlen + transhdrlen + 20,
716 (flags & MSG_DONTWAIT), &err);
718 if (skb == NULL)
719 return err;
721 /* reserve space for Hardware header */
722 skb_reserve(skb, hh_len);
724 /* create space for UDP/IP header */
725 skb_put(skb,fragheaderlen + transhdrlen);
727 /* initialize network header pointer */
728 skb->nh.raw = skb->data;
730 /* initialize protocol header pointer */
731 skb->h.raw = skb->data + fragheaderlen;
733 skb->ip_summed = CHECKSUM_HW;
734 skb->csum = 0;
735 sk->sk_sndmsg_off = 0;
738 err = skb_append_datato_frags(sk,skb, getfrag, from,
739 (length - transhdrlen));
740 if (!err) {
741 /* specify the length of each IP datagram fragment*/
742 skb_shinfo(skb)->ufo_size = (mtu - fragheaderlen);
743 __skb_queue_tail(&sk->sk_write_queue, skb);
745 return 0;
747 /* There is not enough support do UFO ,
748 * so follow normal path
750 kfree_skb(skb);
751 return err;
755 * ip_append_data() and ip_append_page() can make one large IP datagram
756 * from many pieces of data. Each pieces will be holded on the socket
757 * until ip_push_pending_frames() is called. Each piece can be a page
758 * or non-page data.
760 * Not only UDP, other transport protocols - e.g. raw sockets - can use
761 * this interface potentially.
763 * LATER: length must be adjusted by pad at tail, when it is required.
765 int ip_append_data(struct sock *sk,
766 int getfrag(void *from, char *to, int offset, int len,
767 int odd, struct sk_buff *skb),
768 void *from, int length, int transhdrlen,
769 struct ipcm_cookie *ipc, struct rtable *rt,
770 unsigned int flags)
772 struct inet_sock *inet = inet_sk(sk);
773 struct sk_buff *skb;
775 struct ip_options *opt = NULL;
776 int hh_len;
777 int exthdrlen;
778 int mtu;
779 int copy;
780 int err;
781 int offset = 0;
782 unsigned int maxfraglen, fragheaderlen;
783 int csummode = CHECKSUM_NONE;
785 if (flags&MSG_PROBE)
786 return 0;
788 if (skb_queue_empty(&sk->sk_write_queue)) {
790 * setup for corking.
792 opt = ipc->opt;
793 if (opt) {
794 if (inet->cork.opt == NULL) {
795 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
796 if (unlikely(inet->cork.opt == NULL))
797 return -ENOBUFS;
799 memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
800 inet->cork.flags |= IPCORK_OPT;
801 inet->cork.addr = ipc->addr;
803 dst_hold(&rt->u.dst);
804 inet->cork.fragsize = mtu = dst_mtu(rt->u.dst.path);
805 inet->cork.rt = rt;
806 inet->cork.length = 0;
807 sk->sk_sndmsg_page = NULL;
808 sk->sk_sndmsg_off = 0;
809 if ((exthdrlen = rt->u.dst.header_len) != 0) {
810 length += exthdrlen;
811 transhdrlen += exthdrlen;
813 } else {
814 rt = inet->cork.rt;
815 if (inet->cork.flags & IPCORK_OPT)
816 opt = inet->cork.opt;
818 transhdrlen = 0;
819 exthdrlen = 0;
820 mtu = inet->cork.fragsize;
822 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
824 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
825 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
827 if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
828 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu-exthdrlen);
829 return -EMSGSIZE;
833 * transhdrlen > 0 means that this is the first fragment and we wish
834 * it won't be fragmented in the future.
836 if (transhdrlen &&
837 length + fragheaderlen <= mtu &&
838 rt->u.dst.dev->features&(NETIF_F_IP_CSUM|NETIF_F_NO_CSUM|NETIF_F_HW_CSUM) &&
839 !exthdrlen)
840 csummode = CHECKSUM_HW;
842 inet->cork.length += length;
843 if (((length > mtu) && (sk->sk_protocol == IPPROTO_UDP)) &&
844 (rt->u.dst.dev->features & NETIF_F_UFO)) {
846 if(ip_ufo_append_data(sk, getfrag, from, length, hh_len,
847 fragheaderlen, transhdrlen, mtu, flags))
848 goto error;
850 return 0;
853 /* So, what's going on in the loop below?
855 * We use calculated fragment length to generate chained skb,
856 * each of segments is IP fragment ready for sending to network after
857 * adding appropriate IP header.
860 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
861 goto alloc_new_skb;
863 while (length > 0) {
864 /* Check if the remaining data fits into current packet. */
865 copy = mtu - skb->len;
866 if (copy < length)
867 copy = maxfraglen - skb->len;
868 if (copy <= 0) {
869 char *data;
870 unsigned int datalen;
871 unsigned int fraglen;
872 unsigned int fraggap;
873 unsigned int alloclen;
874 struct sk_buff *skb_prev;
875 alloc_new_skb:
876 skb_prev = skb;
877 if (skb_prev)
878 fraggap = skb_prev->len - maxfraglen;
879 else
880 fraggap = 0;
883 * If remaining data exceeds the mtu,
884 * we know we need more fragment(s).
886 datalen = length + fraggap;
887 if (datalen > mtu - fragheaderlen)
888 datalen = maxfraglen - fragheaderlen;
889 fraglen = datalen + fragheaderlen;
891 if ((flags & MSG_MORE) &&
892 !(rt->u.dst.dev->features&NETIF_F_SG))
893 alloclen = mtu;
894 else
895 alloclen = datalen + fragheaderlen;
897 /* The last fragment gets additional space at tail.
898 * Note, with MSG_MORE we overallocate on fragments,
899 * because we have no idea what fragment will be
900 * the last.
902 if (datalen == length)
903 alloclen += rt->u.dst.trailer_len;
905 if (transhdrlen) {
906 skb = sock_alloc_send_skb(sk,
907 alloclen + hh_len + 15,
908 (flags & MSG_DONTWAIT), &err);
909 } else {
910 skb = NULL;
911 if (atomic_read(&sk->sk_wmem_alloc) <=
912 2 * sk->sk_sndbuf)
913 skb = sock_wmalloc(sk,
914 alloclen + hh_len + 15, 1,
915 sk->sk_allocation);
916 if (unlikely(skb == NULL))
917 err = -ENOBUFS;
919 if (skb == NULL)
920 goto error;
923 * Fill in the control structures
925 skb->ip_summed = csummode;
926 skb->csum = 0;
927 skb_reserve(skb, hh_len);
930 * Find where to start putting bytes.
932 data = skb_put(skb, fraglen);
933 skb->nh.raw = data + exthdrlen;
934 data += fragheaderlen;
935 skb->h.raw = data + exthdrlen;
937 if (fraggap) {
938 skb->csum = skb_copy_and_csum_bits(
939 skb_prev, maxfraglen,
940 data + transhdrlen, fraggap, 0);
941 skb_prev->csum = csum_sub(skb_prev->csum,
942 skb->csum);
943 data += fraggap;
944 skb_trim(skb_prev, maxfraglen);
947 copy = datalen - transhdrlen - fraggap;
948 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
949 err = -EFAULT;
950 kfree_skb(skb);
951 goto error;
954 offset += copy;
955 length -= datalen - fraggap;
956 transhdrlen = 0;
957 exthdrlen = 0;
958 csummode = CHECKSUM_NONE;
961 * Put the packet on the pending queue.
963 __skb_queue_tail(&sk->sk_write_queue, skb);
964 continue;
967 if (copy > length)
968 copy = length;
970 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
971 unsigned int off;
973 off = skb->len;
974 if (getfrag(from, skb_put(skb, copy),
975 offset, copy, off, skb) < 0) {
976 __skb_trim(skb, off);
977 err = -EFAULT;
978 goto error;
980 } else {
981 int i = skb_shinfo(skb)->nr_frags;
982 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
983 struct page *page = sk->sk_sndmsg_page;
984 int off = sk->sk_sndmsg_off;
985 unsigned int left;
987 if (page && (left = PAGE_SIZE - off) > 0) {
988 if (copy >= left)
989 copy = left;
990 if (page != frag->page) {
991 if (i == MAX_SKB_FRAGS) {
992 err = -EMSGSIZE;
993 goto error;
995 get_page(page);
996 skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
997 frag = &skb_shinfo(skb)->frags[i];
999 } else if (i < MAX_SKB_FRAGS) {
1000 if (copy > PAGE_SIZE)
1001 copy = PAGE_SIZE;
1002 page = alloc_pages(sk->sk_allocation, 0);
1003 if (page == NULL) {
1004 err = -ENOMEM;
1005 goto error;
1007 sk->sk_sndmsg_page = page;
1008 sk->sk_sndmsg_off = 0;
1010 skb_fill_page_desc(skb, i, page, 0, 0);
1011 frag = &skb_shinfo(skb)->frags[i];
1012 skb->truesize += PAGE_SIZE;
1013 atomic_add(PAGE_SIZE, &sk->sk_wmem_alloc);
1014 } else {
1015 err = -EMSGSIZE;
1016 goto error;
1018 if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1019 err = -EFAULT;
1020 goto error;
1022 sk->sk_sndmsg_off += copy;
1023 frag->size += copy;
1024 skb->len += copy;
1025 skb->data_len += copy;
1027 offset += copy;
1028 length -= copy;
1031 return 0;
1033 error:
1034 inet->cork.length -= length;
1035 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1036 return err;
1039 ssize_t ip_append_page(struct sock *sk, struct page *page,
1040 int offset, size_t size, int flags)
1042 struct inet_sock *inet = inet_sk(sk);
1043 struct sk_buff *skb;
1044 struct rtable *rt;
1045 struct ip_options *opt = NULL;
1046 int hh_len;
1047 int mtu;
1048 int len;
1049 int err;
1050 unsigned int maxfraglen, fragheaderlen, fraggap;
1052 if (inet->hdrincl)
1053 return -EPERM;
1055 if (flags&MSG_PROBE)
1056 return 0;
1058 if (skb_queue_empty(&sk->sk_write_queue))
1059 return -EINVAL;
1061 rt = inet->cork.rt;
1062 if (inet->cork.flags & IPCORK_OPT)
1063 opt = inet->cork.opt;
1065 if (!(rt->u.dst.dev->features&NETIF_F_SG))
1066 return -EOPNOTSUPP;
1068 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1069 mtu = inet->cork.fragsize;
1071 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1072 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1074 if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1075 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu);
1076 return -EMSGSIZE;
1079 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1080 return -EINVAL;
1082 inet->cork.length += size;
1083 if ((sk->sk_protocol == IPPROTO_UDP) &&
1084 (rt->u.dst.dev->features & NETIF_F_UFO))
1085 skb_shinfo(skb)->ufo_size = (mtu - fragheaderlen);
1088 while (size > 0) {
1089 int i;
1091 if (skb_shinfo(skb)->ufo_size)
1092 len = size;
1093 else {
1095 /* Check if the remaining data fits into current packet. */
1096 len = mtu - skb->len;
1097 if (len < size)
1098 len = maxfraglen - skb->len;
1100 if (len <= 0) {
1101 struct sk_buff *skb_prev;
1102 char *data;
1103 struct iphdr *iph;
1104 int alloclen;
1106 skb_prev = skb;
1107 fraggap = skb_prev->len - maxfraglen;
1109 alloclen = fragheaderlen + hh_len + fraggap + 15;
1110 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1111 if (unlikely(!skb)) {
1112 err = -ENOBUFS;
1113 goto error;
1117 * Fill in the control structures
1119 skb->ip_summed = CHECKSUM_NONE;
1120 skb->csum = 0;
1121 skb_reserve(skb, hh_len);
1124 * Find where to start putting bytes.
1126 data = skb_put(skb, fragheaderlen + fraggap);
1127 skb->nh.iph = iph = (struct iphdr *)data;
1128 data += fragheaderlen;
1129 skb->h.raw = data;
1131 if (fraggap) {
1132 skb->csum = skb_copy_and_csum_bits(
1133 skb_prev, maxfraglen,
1134 data, fraggap, 0);
1135 skb_prev->csum = csum_sub(skb_prev->csum,
1136 skb->csum);
1137 skb_trim(skb_prev, maxfraglen);
1141 * Put the packet on the pending queue.
1143 __skb_queue_tail(&sk->sk_write_queue, skb);
1144 continue;
1147 i = skb_shinfo(skb)->nr_frags;
1148 if (len > size)
1149 len = size;
1150 if (skb_can_coalesce(skb, i, page, offset)) {
1151 skb_shinfo(skb)->frags[i-1].size += len;
1152 } else if (i < MAX_SKB_FRAGS) {
1153 get_page(page);
1154 skb_fill_page_desc(skb, i, page, offset, len);
1155 } else {
1156 err = -EMSGSIZE;
1157 goto error;
1160 if (skb->ip_summed == CHECKSUM_NONE) {
1161 unsigned int csum;
1162 csum = csum_page(page, offset, len);
1163 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1166 skb->len += len;
1167 skb->data_len += len;
1168 offset += len;
1169 size -= len;
1171 return 0;
1173 error:
1174 inet->cork.length -= size;
1175 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1176 return err;
1180 * Combined all pending IP fragments on the socket as one IP datagram
1181 * and push them out.
1183 int ip_push_pending_frames(struct sock *sk)
1185 struct sk_buff *skb, *tmp_skb;
1186 struct sk_buff **tail_skb;
1187 struct inet_sock *inet = inet_sk(sk);
1188 struct ip_options *opt = NULL;
1189 struct rtable *rt = inet->cork.rt;
1190 struct iphdr *iph;
1191 __be16 df = 0;
1192 __u8 ttl;
1193 int err = 0;
1195 if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1196 goto out;
1197 tail_skb = &(skb_shinfo(skb)->frag_list);
1199 /* move skb->data to ip header from ext header */
1200 if (skb->data < skb->nh.raw)
1201 __skb_pull(skb, skb->nh.raw - skb->data);
1202 while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1203 __skb_pull(tmp_skb, skb->h.raw - skb->nh.raw);
1204 *tail_skb = tmp_skb;
1205 tail_skb = &(tmp_skb->next);
1206 skb->len += tmp_skb->len;
1207 skb->data_len += tmp_skb->len;
1208 skb->truesize += tmp_skb->truesize;
1209 __sock_put(tmp_skb->sk);
1210 tmp_skb->destructor = NULL;
1211 tmp_skb->sk = NULL;
1214 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1215 * to fragment the frame generated here. No matter, what transforms
1216 * how transforms change size of the packet, it will come out.
1218 if (inet->pmtudisc != IP_PMTUDISC_DO)
1219 skb->local_df = 1;
1221 /* DF bit is set when we want to see DF on outgoing frames.
1222 * If local_df is set too, we still allow to fragment this frame
1223 * locally. */
1224 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1225 (skb->len <= dst_mtu(&rt->u.dst) &&
1226 ip_dont_fragment(sk, &rt->u.dst)))
1227 df = htons(IP_DF);
1229 if (inet->cork.flags & IPCORK_OPT)
1230 opt = inet->cork.opt;
1232 if (rt->rt_type == RTN_MULTICAST)
1233 ttl = inet->mc_ttl;
1234 else
1235 ttl = ip_select_ttl(inet, &rt->u.dst);
1237 iph = (struct iphdr *)skb->data;
1238 iph->version = 4;
1239 iph->ihl = 5;
1240 if (opt) {
1241 iph->ihl += opt->optlen>>2;
1242 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1244 iph->tos = inet->tos;
1245 iph->tot_len = htons(skb->len);
1246 iph->frag_off = df;
1247 if (!df) {
1248 __ip_select_ident(iph, &rt->u.dst, 0);
1249 } else {
1250 iph->id = htons(inet->id++);
1252 iph->ttl = ttl;
1253 iph->protocol = sk->sk_protocol;
1254 iph->saddr = rt->rt_src;
1255 iph->daddr = rt->rt_dst;
1256 ip_send_check(iph);
1258 skb->priority = sk->sk_priority;
1259 skb->dst = dst_clone(&rt->u.dst);
1261 /* Netfilter gets whole the not fragmented skb. */
1262 err = NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL,
1263 skb->dst->dev, dst_output);
1264 if (err) {
1265 if (err > 0)
1266 err = inet->recverr ? net_xmit_errno(err) : 0;
1267 if (err)
1268 goto error;
1271 out:
1272 inet->cork.flags &= ~IPCORK_OPT;
1273 kfree(inet->cork.opt);
1274 inet->cork.opt = NULL;
1275 if (inet->cork.rt) {
1276 ip_rt_put(inet->cork.rt);
1277 inet->cork.rt = NULL;
1279 return err;
1281 error:
1282 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1283 goto out;
1287 * Throw away all pending data on the socket.
1289 void ip_flush_pending_frames(struct sock *sk)
1291 struct inet_sock *inet = inet_sk(sk);
1292 struct sk_buff *skb;
1294 while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1295 kfree_skb(skb);
1297 inet->cork.flags &= ~IPCORK_OPT;
1298 kfree(inet->cork.opt);
1299 inet->cork.opt = NULL;
1300 if (inet->cork.rt) {
1301 ip_rt_put(inet->cork.rt);
1302 inet->cork.rt = NULL;
1308 * Fetch data from kernel space and fill in checksum if needed.
1310 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1311 int len, int odd, struct sk_buff *skb)
1313 unsigned int csum;
1315 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1316 skb->csum = csum_block_add(skb->csum, csum, odd);
1317 return 0;
1321 * Generic function to send a packet as reply to another packet.
1322 * Used to send TCP resets so far. ICMP should use this function too.
1324 * Should run single threaded per socket because it uses the sock
1325 * structure to pass arguments.
1327 * LATER: switch from ip_build_xmit to ip_append_*
1329 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1330 unsigned int len)
1332 struct inet_sock *inet = inet_sk(sk);
1333 struct {
1334 struct ip_options opt;
1335 char data[40];
1336 } replyopts;
1337 struct ipcm_cookie ipc;
1338 u32 daddr;
1339 struct rtable *rt = (struct rtable*)skb->dst;
1341 if (ip_options_echo(&replyopts.opt, skb))
1342 return;
1344 daddr = ipc.addr = rt->rt_src;
1345 ipc.opt = NULL;
1347 if (replyopts.opt.optlen) {
1348 ipc.opt = &replyopts.opt;
1350 if (ipc.opt->srr)
1351 daddr = replyopts.opt.faddr;
1355 struct flowi fl = { .nl_u = { .ip4_u =
1356 { .daddr = daddr,
1357 .saddr = rt->rt_spec_dst,
1358 .tos = RT_TOS(skb->nh.iph->tos) } },
1359 /* Not quite clean, but right. */
1360 .uli_u = { .ports =
1361 { .sport = skb->h.th->dest,
1362 .dport = skb->h.th->source } },
1363 .proto = sk->sk_protocol };
1364 if (ip_route_output_key(&rt, &fl))
1365 return;
1368 /* And let IP do all the hard work.
1370 This chunk is not reenterable, hence spinlock.
1371 Note that it uses the fact, that this function is called
1372 with locally disabled BH and that sk cannot be already spinlocked.
1374 bh_lock_sock(sk);
1375 inet->tos = skb->nh.iph->tos;
1376 sk->sk_priority = skb->priority;
1377 sk->sk_protocol = skb->nh.iph->protocol;
1378 ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1379 &ipc, rt, MSG_DONTWAIT);
1380 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1381 if (arg->csumoffset >= 0)
1382 *((u16 *)skb->h.raw + arg->csumoffset) = csum_fold(csum_add(skb->csum, arg->csum));
1383 skb->ip_summed = CHECKSUM_NONE;
1384 ip_push_pending_frames(sk);
1387 bh_unlock_sock(sk);
1389 ip_rt_put(rt);
1392 void __init ip_init(void)
1394 ip_rt_init();
1395 inet_initpeers();
1397 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1398 igmp_mc_proc_init();
1399 #endif
1402 EXPORT_SYMBOL(ip_generic_getfrag);
1403 EXPORT_SYMBOL(ip_queue_xmit);
1404 EXPORT_SYMBOL(ip_send_check);