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 1.1.1.1 2005/04/29 01:44:09 echo Exp $
10 * Authors: Ross Biro, <bir7@leland.Stanford.Edu>
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Donald Becker, <becker@super.org>
13 * Alan Cox, <Alan.Cox@linux.org>
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
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
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
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>
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>
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>
70 #include <net/protocol.h>
71 #include <net/route.h>
74 #include <linux/skbuff.h>
79 #include <net/checksum.h>
80 #include <net/inetpeer.h>
81 #include <linux/igmp.h>
82 #include <linux/netfilter_ipv4.h>
83 #include <linux/netfilter_bridge.h>
84 #include <linux/mroute.h>
85 #include <linux/netlink.h>
88 * Shall we try to damage output packets if routing dev changes?
91 int sysctl_ip_dynaddr;
92 int sysctl_ip_default_ttl = IPDEFTTL;
94 /* Generate a checksum for an outgoing IP datagram. */
95 __inline__ void ip_send_check(struct iphdr *iph)
98 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
101 /* dev_loopback_xmit for use with netfilter. */
102 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
104 newskb->mac.raw = newskb->data;
105 __skb_pull(newskb, newskb->nh.raw - newskb->data);
106 newskb->pkt_type = PACKET_LOOPBACK;
107 newskb->ip_summed = CHECKSUM_UNNECESSARY;
108 BUG_TRAP(newskb->dst);
110 #ifdef CONFIG_NETFILTER_DEBUG
111 nf_debug_ip_loopback_xmit(newskb);
117 static inline int ip_select_ttl(struct inet_opt *inet, struct dst_entry *dst)
119 int ttl = inet->uc_ttl;
122 ttl = dst_metric(dst, RTAX_HOPLIMIT);
127 * Add an ip header to a skbuff and send it out.
130 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
131 u32 saddr, u32 daddr, struct ip_options *opt)
133 struct inet_opt *inet = inet_sk(sk);
134 struct rtable *rt = (struct rtable *)skb->dst;
137 /* Build the IP header. */
139 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr) + opt->optlen);
141 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr));
145 iph->tos = inet->tos;
146 if (ip_dont_fragment(sk, &rt->u.dst))
147 iph->frag_off = htons(IP_DF);
150 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
151 iph->daddr = rt->rt_dst;
152 iph->saddr = rt->rt_src;
153 iph->protocol = sk->sk_protocol;
154 iph->tot_len = htons(skb->len);
155 ip_select_ident(iph, &rt->u.dst, sk);
158 if (opt && opt->optlen) {
159 iph->ihl += opt->optlen>>2;
160 ip_options_build(skb, opt, daddr, rt, 0);
164 skb->priority = sk->sk_priority;
167 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
171 static inline int ip_finish_output2(struct sk_buff *skb)
173 struct dst_entry *dst = skb->dst;
174 struct hh_cache *hh = dst->hh;
175 struct net_device *dev = dst->dev;
176 int hh_len = LL_RESERVED_SPACE(dev);
178 /* Be paranoid, rather than too clever. */
179 if (unlikely(skb_headroom(skb) < hh_len && dev->hard_header)) {
180 struct sk_buff *skb2;
182 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
188 skb_set_owner_w(skb2, skb->sk);
193 #ifdef CONFIG_NETFILTER_DEBUG
194 nf_debug_ip_finish_output2(skb);
195 #endif /*CONFIG_NETFILTER_DEBUG*/
200 read_lock_bh(&hh->hh_lock);
201 hh_alen = HH_DATA_ALIGN(hh->hh_len);
202 memcpy(skb->data - hh_alen, hh->hh_data, hh_alen);
203 read_unlock_bh(&hh->hh_lock);
204 skb_push(skb, hh->hh_len);
205 return hh->hh_output(skb);
206 } else if (dst->neighbour)
207 return dst->neighbour->output(skb);
210 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
215 int ip_finish_output(struct sk_buff *skb)
217 struct net_device *dev = skb->dst->dev;
220 skb->protocol = htons(ETH_P_IP);
222 return NF_HOOK(PF_INET, NF_IP_POST_ROUTING, skb, NULL, dev,
226 int ip_mc_output(struct sk_buff **pskb)
228 struct sk_buff *skb = *pskb;
229 struct sock *sk = skb->sk;
230 struct rtable *rt = (struct rtable*)skb->dst;
231 struct net_device *dev = rt->u.dst.dev;
234 * If the indicated interface is up and running, send the packet.
236 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
239 skb->protocol = htons(ETH_P_IP);
242 * Multicasts are looped back for other local users
245 if (rt->rt_flags&RTCF_MULTICAST) {
246 if ((!sk || inet_sk(sk)->mc_loop)
247 #ifdef CONFIG_IP_MROUTE
248 /* Small optimization: do not loopback not local frames,
249 which returned after forwarding; they will be dropped
250 by ip_mr_input in any case.
251 Note, that local frames are looped back to be delivered
254 This check is duplicated in ip_mr_input at the moment.
256 && ((rt->rt_flags&RTCF_LOCAL) || !(IPCB(skb)->flags&IPSKB_FORWARDED))
259 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
261 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
263 ip_dev_loopback_xmit);
266 /* Multicasts with ttl 0 must not go beyond the host */
268 if (skb->nh.iph->ttl == 0) {
274 if (rt->rt_flags&RTCF_BROADCAST) {
275 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
277 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
278 newskb->dev, ip_dev_loopback_xmit);
281 if (skb->len > dst_pmtu(&rt->u.dst) || skb_shinfo(skb)->frag_list)
282 return ip_fragment(skb, ip_finish_output);
284 return ip_finish_output(skb);
287 int ip_output(struct sk_buff **pskb)
289 struct sk_buff *skb = *pskb;
291 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
293 if ((skb->len > dst_pmtu(skb->dst) || skb_shinfo(skb)->frag_list) &&
294 !skb_shinfo(skb)->tso_size)
295 return ip_fragment(skb, ip_finish_output);
297 return ip_finish_output(skb);
300 int ip_queue_xmit(struct sk_buff *skb, int ipfragok)
302 struct sock *sk = skb->sk;
303 struct inet_opt *inet = inet_sk(sk);
304 struct ip_options *opt = inet->opt;
309 /* Skip all of this if the packet is already routed,
310 * f.e. by something like SCTP.
312 rt = (struct rtable *) skb->dst;
316 /* Make sure we can route this packet. */
317 rt = (struct rtable *)__sk_dst_check(sk, 0);
321 /* Use correct destination address if we have options. */
327 struct flowi fl = { .oif = sk->sk_bound_dev_if,
330 .saddr = inet->saddr,
331 .tos = RT_CONN_FLAGS(sk) } },
332 .proto = sk->sk_protocol,
334 { .sport = inet->sport,
335 .dport = inet->dport } } };
337 /* If this fails, retransmit mechanism of transport layer will
338 * keep trying until route appears or the connection times
341 if (ip_route_output_flow(&rt, &fl, sk, 0))
344 __sk_dst_set(sk, &rt->u.dst);
345 tcp_v4_setup_caps(sk, &rt->u.dst);
347 skb->dst = dst_clone(&rt->u.dst);
350 if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
353 /* OK, we know where to send it, allocate and build IP header. */
354 iph = (struct iphdr *) skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
355 *((__u16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
356 iph->tot_len = htons(skb->len);
357 if (ip_dont_fragment(sk, &rt->u.dst) && !ipfragok)
358 iph->frag_off = htons(IP_DF);
361 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
362 iph->protocol = sk->sk_protocol;
363 iph->saddr = rt->rt_src;
364 iph->daddr = rt->rt_dst;
366 /* Transport layer set skb->h.foo itself. */
368 if(opt && opt->optlen) {
369 iph->ihl += opt->optlen >> 2;
370 ip_options_build(skb, opt, inet->daddr, rt, 0);
373 mtu = dst_pmtu(&rt->u.dst);
374 if (skb->len > mtu && (sk->sk_route_caps & NETIF_F_TSO)) {
377 /* Hack zone: all this must be done by TCP. */
378 hlen = ((skb->h.raw - skb->data) + (skb->h.th->doff << 2));
379 skb_shinfo(skb)->tso_size = mtu - hlen;
380 skb_shinfo(skb)->tso_segs =
381 (skb->len - hlen + skb_shinfo(skb)->tso_size - 1)/
382 skb_shinfo(skb)->tso_size - 1;
385 ip_select_ident_more(iph, &rt->u.dst, sk, skb_shinfo(skb)->tso_segs);
387 /* Add an IP checksum. */
390 skb->priority = sk->sk_priority;
392 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
396 IP_INC_STATS(IPSTATS_MIB_OUTNOROUTES);
398 return -EHOSTUNREACH;
402 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
404 to->pkt_type = from->pkt_type;
405 to->priority = from->priority;
406 to->protocol = from->protocol;
407 to->security = from->security;
408 to->dst = dst_clone(from->dst);
411 /* Copy the flags to each fragment. */
412 IPCB(to)->flags = IPCB(from)->flags;
414 #ifdef CONFIG_NET_SCHED
415 to->tc_index = from->tc_index;
417 #ifdef CONFIG_NETFILTER
418 to->nfmark = from->nfmark;
419 to->nfcache = from->nfcache;
420 /* Connection association is same as pre-frag packet */
421 nf_conntrack_put(to->nfct);
422 to->nfct = from->nfct;
423 nf_conntrack_get(to->nfct);
424 #ifdef CONFIG_BRIDGE_NETFILTER
425 nf_bridge_put(to->nf_bridge);
426 to->nf_bridge = from->nf_bridge;
427 nf_bridge_get(to->nf_bridge);
429 #ifdef CONFIG_NETFILTER_DEBUG
430 to->nf_debug = from->nf_debug;
436 * This IP datagram is too large to be sent in one piece. Break it up into
437 * smaller pieces (each of size equal to IP header plus
438 * a block of the data of the original IP data part) that will yet fit in a
439 * single device frame, and queue such a frame for sending.
442 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*))
447 struct net_device *dev;
448 struct sk_buff *skb2;
449 unsigned int mtu, hlen, left, len, ll_rs;
452 struct rtable *rt = (struct rtable*)skb->dst;
458 * Point into the IP datagram header.
463 if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
464 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
465 htonl(dst_pmtu(&rt->u.dst)));
471 * Setup starting values.
475 mtu = dst_pmtu(&rt->u.dst) - hlen; /* Size of data space */
477 /* When frag_list is given, use it. First, check its validity:
478 * some transformers could create wrong frag_list or break existing
479 * one, it is not prohibited. In this case fall back to copying.
481 * LATER: this step can be merged to real generation of fragments,
482 * we can switch to copy when see the first bad fragment.
484 if (skb_shinfo(skb)->frag_list) {
485 struct sk_buff *frag;
486 int first_len = skb_pagelen(skb);
488 if (first_len - hlen > mtu ||
489 ((first_len - hlen) & 7) ||
490 (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
494 for (frag = skb_shinfo(skb)->frag_list; frag; frag = frag->next) {
495 /* Correct geometry. */
496 if (frag->len > mtu ||
497 ((frag->len & 7) && frag->next) ||
498 skb_headroom(frag) < hlen)
501 /* Partially cloned skb? */
502 if (skb_shared(frag))
506 /* Everything is OK. Generate! */
510 frag = skb_shinfo(skb)->frag_list;
511 skb_shinfo(skb)->frag_list = NULL;
512 skb->data_len = first_len - skb_headlen(skb);
513 skb->len = first_len;
514 iph->tot_len = htons(first_len);
515 iph->frag_off |= htons(IP_MF);
519 /* Prepare header of the next frame,
520 * before previous one went down. */
522 frag->h.raw = frag->data;
523 frag->nh.raw = __skb_push(frag, hlen);
524 memcpy(frag->nh.raw, iph, hlen);
526 iph->tot_len = htons(frag->len);
527 ip_copy_metadata(frag, skb);
529 ip_options_fragment(frag);
530 offset += skb->len - hlen;
531 iph->frag_off = htons(offset>>3);
532 if (frag->next != NULL)
533 iph->frag_off |= htons(IP_MF);
534 /* Ready, complete checksum */
549 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
558 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
563 left = skb->len - hlen; /* Space per frame */
564 ptr = raw + hlen; /* Where to start from */
566 #ifdef CONFIG_BRIDGE_NETFILTER
567 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
568 * we need to make room for the encapsulating header */
569 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, nf_bridge_pad(skb));
570 mtu -= nf_bridge_pad(skb);
572 ll_rs = LL_RESERVED_SPACE(rt->u.dst.dev);
575 * Fragment the datagram.
578 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
579 not_last_frag = iph->frag_off & htons(IP_MF);
582 * Keep copying data until we run out.
587 /* IF: it doesn't fit, use 'mtu' - the data space left */
590 /* IF: we are not sending upto and including the packet end
591 then align the next start on an eight byte boundary */
599 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
600 NETDEBUG(printk(KERN_INFO "IP: frag: no memory for new fragment!\n"));
606 * Set up data on packet
609 ip_copy_metadata(skb2, skb);
610 skb_reserve(skb2, ll_rs);
611 skb_put(skb2, len + hlen);
612 skb2->nh.raw = skb2->data;
613 skb2->h.raw = skb2->data + hlen;
616 * Charge the memory for the fragment to any owner
621 skb_set_owner_w(skb2, skb->sk);
624 * Copy the packet header into the new buffer.
627 memcpy(skb2->nh.raw, skb->data, hlen);
630 * Copy a block of the IP datagram.
632 if (skb_copy_bits(skb, ptr, skb2->h.raw, len))
637 * Fill in the new header fields.
640 iph->frag_off = htons((offset >> 3));
642 /* ANK: dirty, but effective trick. Upgrade options only if
643 * the segment to be fragmented was THE FIRST (otherwise,
644 * options are already fixed) and make it ONCE
645 * on the initial skb, so that all the following fragments
646 * will inherit fixed options.
649 ip_options_fragment(skb);
652 * Added AC : If we are fragmenting a fragment that's not the
653 * last fragment then keep MF on each bit
655 if (left > 0 || not_last_frag)
656 iph->frag_off |= htons(IP_MF);
661 * Put this fragment into the sending queue.
664 IP_INC_STATS(IPSTATS_MIB_FRAGCREATES);
666 iph->tot_len = htons(len + hlen);
675 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
680 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
685 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
687 struct iovec *iov = from;
689 if (skb->ip_summed == CHECKSUM_HW) {
690 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
693 unsigned int csum = 0;
694 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
696 skb->csum = csum_block_add(skb->csum, csum, odd);
701 static inline unsigned int
702 csum_page(struct page *page, int offset, int copy)
707 csum = csum_partial(kaddr + offset, copy, 0);
713 * ip_append_data() and ip_append_page() can make one large IP datagram
714 * from many pieces of data. Each pieces will be holded on the socket
715 * until ip_push_pending_frames() is called. Eache pieces can be a page
718 * Not only UDP, other transport protocols - e.g. raw sockets - can use
719 * this interface potentially.
721 * LATER: length must be adjusted by pad at tail, when it is required.
723 int ip_append_data(struct sock *sk,
724 int getfrag(void *from, char *to, int offset, int len,
725 int odd, struct sk_buff *skb),
726 void *from, int length, int transhdrlen,
727 struct ipcm_cookie *ipc, struct rtable *rt,
730 struct inet_opt *inet = inet_sk(sk);
733 struct ip_options *opt = NULL;
740 unsigned int maxfraglen, fragheaderlen;
741 int csummode = CHECKSUM_NONE;
746 if (skb_queue_empty(&sk->sk_write_queue)) {
752 if (inet->cork.opt == NULL) {
753 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
754 if (unlikely(inet->cork.opt == NULL))
757 memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
758 inet->cork.flags |= IPCORK_OPT;
759 inet->cork.addr = ipc->addr;
761 dst_hold(&rt->u.dst);
762 inet->cork.fragsize = mtu = dst_pmtu(&rt->u.dst);
764 inet->cork.length = 0;
765 sk->sk_sndmsg_page = NULL;
766 sk->sk_sndmsg_off = 0;
767 if ((exthdrlen = rt->u.dst.header_len) != 0) {
769 transhdrlen += exthdrlen;
773 if (inet->cork.flags & IPCORK_OPT)
774 opt = inet->cork.opt;
778 mtu = inet->cork.fragsize;
780 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
782 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
783 maxfraglen = ((mtu-fragheaderlen) & ~7) + fragheaderlen;
785 if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
786 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu-exthdrlen);
791 * transhdrlen > 0 means that this is the first fragment and we wish
792 * it won't be fragmented in the future.
795 length + fragheaderlen <= maxfraglen &&
796 rt->u.dst.dev->features&(NETIF_F_IP_CSUM|NETIF_F_NO_CSUM|NETIF_F_HW_CSUM) &&
798 csummode = CHECKSUM_HW;
800 inet->cork.length += length;
802 /* So, what's going on in the loop below?
804 * We use calculated fragment length to generate chained skb,
805 * each of segments is IP fragment ready for sending to network after
806 * adding appropriate IP header.
810 * If mtu-fragheaderlen is not 0 modulo 8, we generate additional
811 * small fragment of length (mtu-fragheaderlen)%8, even though
812 * it is not necessary. Not a big bug, but needs a fix.
815 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
819 if ((copy = maxfraglen - skb->len) <= 0) {
821 unsigned int datalen;
822 unsigned int fraglen;
823 unsigned int alloclen;
827 datalen = maxfraglen - fragheaderlen;
828 if (datalen > length)
831 fraglen = datalen + fragheaderlen;
832 if ((flags & MSG_MORE) &&
833 !(rt->u.dst.dev->features&NETIF_F_SG))
834 alloclen = maxfraglen;
836 alloclen = datalen + fragheaderlen;
838 /* The last fragment gets additional space at tail.
839 * Note, with MSG_MORE we overallocate on fragments,
840 * because we have no idea what fragment will be
843 if (datalen == length)
844 alloclen += rt->u.dst.trailer_len;
847 skb = sock_alloc_send_skb(sk,
848 alloclen + hh_len + 15,
849 (flags & MSG_DONTWAIT), &err);
852 if (atomic_read(&sk->sk_wmem_alloc) <=
854 skb = sock_wmalloc(sk,
855 alloclen + hh_len + 15, 1,
857 if (unlikely(skb == NULL))
864 * Fill in the control structures
866 skb->ip_summed = csummode;
868 skb_reserve(skb, hh_len);
871 * Find where to start putting bytes.
873 data = skb_put(skb, fraglen);
874 skb->nh.raw = data + exthdrlen;
875 data += fragheaderlen;
876 skb->h.raw = data + exthdrlen;
878 copy = datalen - transhdrlen;
879 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, 0, skb) < 0) {
889 csummode = CHECKSUM_NONE;
892 * Put the packet on the pending queue.
894 __skb_queue_tail(&sk->sk_write_queue, skb);
901 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
905 if (getfrag(from, skb_put(skb, copy),
906 offset, copy, off, skb) < 0) {
907 __skb_trim(skb, off);
912 int i = skb_shinfo(skb)->nr_frags;
913 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
914 struct page *page = sk->sk_sndmsg_page;
915 int off = sk->sk_sndmsg_off;
918 if (page && (left = PAGE_SIZE - off) > 0) {
921 if (page != frag->page) {
922 if (i == MAX_SKB_FRAGS) {
927 skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
928 frag = &skb_shinfo(skb)->frags[i];
930 } else if (i < MAX_SKB_FRAGS) {
931 if (copy > PAGE_SIZE)
933 page = alloc_pages(sk->sk_allocation, 0);
938 sk->sk_sndmsg_page = page;
939 sk->sk_sndmsg_off = 0;
941 skb_fill_page_desc(skb, i, page, 0, 0);
942 frag = &skb_shinfo(skb)->frags[i];
943 skb->truesize += PAGE_SIZE;
944 atomic_add(PAGE_SIZE, &sk->sk_wmem_alloc);
949 if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
953 sk->sk_sndmsg_off += copy;
956 skb->data_len += copy;
965 inet->cork.length -= length;
966 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
970 ssize_t ip_append_page(struct sock *sk, struct page *page,
971 int offset, size_t size, int flags)
973 struct inet_opt *inet = inet_sk(sk);
976 struct ip_options *opt = NULL;
981 unsigned int maxfraglen, fragheaderlen;
989 if (skb_queue_empty(&sk->sk_write_queue))
993 if (inet->cork.flags & IPCORK_OPT)
994 opt = inet->cork.opt;
996 if (!(rt->u.dst.dev->features&NETIF_F_SG))
999 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1000 mtu = inet->cork.fragsize;
1002 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1003 maxfraglen = ((mtu-fragheaderlen) & ~7) + fragheaderlen;
1005 if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1006 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu);
1010 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1013 inet->cork.length += size;
1017 if ((len = maxfraglen - skb->len) <= 0) {
1022 skb = sock_wmalloc(sk, fragheaderlen + hh_len + 15, 1,
1024 if (unlikely(!skb)) {
1030 * Fill in the control structures
1032 skb->ip_summed = CHECKSUM_NONE;
1034 skb_reserve(skb, hh_len);
1037 * Find where to start putting bytes.
1039 data = skb_put(skb, fragheaderlen);
1040 skb->nh.iph = iph = (struct iphdr *)data;
1041 data += fragheaderlen;
1045 * Put the packet on the pending queue.
1047 __skb_queue_tail(&sk->sk_write_queue, skb);
1051 i = skb_shinfo(skb)->nr_frags;
1054 if (skb_can_coalesce(skb, i, page, offset)) {
1055 skb_shinfo(skb)->frags[i-1].size += len;
1056 } else if (i < MAX_SKB_FRAGS) {
1058 skb_fill_page_desc(skb, i, page, offset, len);
1064 if (skb->ip_summed == CHECKSUM_NONE) {
1066 csum = csum_page(page, offset, len);
1067 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1071 skb->data_len += len;
1078 inet->cork.length -= size;
1079 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1084 * Combined all pending IP fragments on the socket as one IP datagram
1085 * and push them out.
1087 int ip_push_pending_frames(struct sock *sk)
1089 struct sk_buff *skb, *tmp_skb;
1090 struct sk_buff **tail_skb;
1091 struct inet_opt *inet = inet_sk(sk);
1092 struct ip_options *opt = NULL;
1093 struct rtable *rt = inet->cork.rt;
1099 if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1101 tail_skb = &(skb_shinfo(skb)->frag_list);
1103 /* move skb->data to ip header from ext header */
1104 if (skb->data < skb->nh.raw)
1105 __skb_pull(skb, skb->nh.raw - skb->data);
1106 while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1107 __skb_pull(tmp_skb, skb->h.raw - skb->nh.raw);
1108 *tail_skb = tmp_skb;
1109 tail_skb = &(tmp_skb->next);
1110 skb->len += tmp_skb->len;
1111 skb->data_len += tmp_skb->len;
1112 skb->truesize += tmp_skb->truesize;
1113 __sock_put(tmp_skb->sk);
1114 tmp_skb->destructor = NULL;
1118 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1119 * to fragment the frame generated here. No matter, what transforms
1120 * how transforms change size of the packet, it will come out.
1122 if (inet->pmtudisc != IP_PMTUDISC_DO)
1125 /* DF bit is set when we want to see DF on outgoing frames.
1126 * If local_df is set too, we still allow to fragment this frame
1128 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1129 (!skb_shinfo(skb)->frag_list && ip_dont_fragment(sk, &rt->u.dst)))
1132 if (inet->cork.flags & IPCORK_OPT)
1133 opt = inet->cork.opt;
1135 if (rt->rt_type == RTN_MULTICAST)
1138 ttl = ip_select_ttl(inet, &rt->u.dst);
1140 iph = (struct iphdr *)skb->data;
1144 iph->ihl += opt->optlen>>2;
1145 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1147 iph->tos = inet->tos;
1148 iph->tot_len = htons(skb->len);
1151 __ip_select_ident(iph, &rt->u.dst, 0);
1153 iph->id = htons(inet->id++);
1156 iph->protocol = sk->sk_protocol;
1157 iph->saddr = rt->rt_src;
1158 iph->daddr = rt->rt_dst;
1161 skb->priority = sk->sk_priority;
1162 skb->dst = dst_clone(&rt->u.dst);
1164 /* Netfilter gets whole the not fragmented skb. */
1165 err = NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL,
1166 skb->dst->dev, dst_output);
1169 err = inet->recverr ? net_xmit_errno(err) : 0;
1175 inet->cork.flags &= ~IPCORK_OPT;
1176 if (inet->cork.opt) {
1177 kfree(inet->cork.opt);
1178 inet->cork.opt = NULL;
1180 if (inet->cork.rt) {
1181 ip_rt_put(inet->cork.rt);
1182 inet->cork.rt = NULL;
1187 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1192 * Throw away all pending data on the socket.
1194 void ip_flush_pending_frames(struct sock *sk)
1196 struct inet_opt *inet = inet_sk(sk);
1197 struct sk_buff *skb;
1199 while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1202 inet->cork.flags &= ~IPCORK_OPT;
1203 if (inet->cork.opt) {
1204 kfree(inet->cork.opt);
1205 inet->cork.opt = NULL;
1207 if (inet->cork.rt) {
1208 ip_rt_put(inet->cork.rt);
1209 inet->cork.rt = NULL;
1215 * Fetch data from kernel space and fill in checksum if needed.
1217 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1218 int len, int odd, struct sk_buff *skb)
1222 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1223 skb->csum = csum_block_add(skb->csum, csum, odd);
1228 * Generic function to send a packet as reply to another packet.
1229 * Used to send TCP resets so far. ICMP should use this function too.
1231 * Should run single threaded per socket because it uses the sock
1232 * structure to pass arguments.
1234 * LATER: switch from ip_build_xmit to ip_append_*
1236 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1239 struct inet_opt *inet = inet_sk(sk);
1241 struct ip_options opt;
1244 struct ipcm_cookie ipc;
1246 struct rtable *rt = (struct rtable*)skb->dst;
1248 if (ip_options_echo(&replyopts.opt, skb))
1251 daddr = ipc.addr = rt->rt_src;
1254 if (replyopts.opt.optlen) {
1255 ipc.opt = &replyopts.opt;
1258 daddr = replyopts.opt.faddr;
1262 struct flowi fl = { .nl_u = { .ip4_u =
1264 .saddr = rt->rt_spec_dst,
1265 .tos = RT_TOS(skb->nh.iph->tos) } },
1266 /* Not quite clean, but right. */
1268 { .sport = skb->h.th->dest,
1269 .dport = skb->h.th->source } },
1270 .proto = sk->sk_protocol };
1271 if (ip_route_output_key(&rt, &fl))
1275 /* And let IP do all the hard work.
1277 This chunk is not reenterable, hence spinlock.
1278 Note that it uses the fact, that this function is called
1279 with locally disabled BH and that sk cannot be already spinlocked.
1282 inet->tos = skb->nh.iph->tos;
1283 sk->sk_priority = skb->priority;
1284 sk->sk_protocol = skb->nh.iph->protocol;
1285 ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1286 &ipc, rt, MSG_DONTWAIT);
1287 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1288 if (arg->csumoffset >= 0)
1289 *((u16 *)skb->h.raw + arg->csumoffset) = csum_fold(csum_add(skb->csum, arg->csum));
1290 skb->ip_summed = CHECKSUM_NONE;
1291 ip_push_pending_frames(sk);
1300 * IP protocol layer initialiser
1303 static struct packet_type ip_packet_type = {
1304 .type = __constant_htons(ETH_P_IP),
1309 * IP registers the packet type and then calls the subprotocol initialisers
1312 void __init ip_init(void)
1314 dev_add_pack(&ip_packet_type);
1319 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1320 igmp_mc_proc_init();
1324 EXPORT_SYMBOL(ip_finish_output);
1325 EXPORT_SYMBOL(ip_fragment);
1326 EXPORT_SYMBOL(ip_generic_getfrag);
1327 EXPORT_SYMBOL(ip_queue_xmit);
1328 EXPORT_SYMBOL(ip_send_check);
1330 #ifdef CONFIG_SYSCTL
1331 EXPORT_SYMBOL(sysctl_ip_default_ttl);