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, <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 *skb)
228 struct sock *sk = skb->sk;
229 struct rtable *rt = (struct rtable*)skb->dst;
230 struct net_device *dev = rt->u.dst.dev;
233 * If the indicated interface is up and running, send the packet.
235 IP_INC_STATS(IpOutRequests);
238 skb->protocol = htons(ETH_P_IP);
241 * Multicasts are looped back for other local users
244 if (rt->rt_flags&RTCF_MULTICAST) {
245 if ((!sk || inet_sk(sk)->mc_loop)
246 #ifdef CONFIG_IP_MROUTE
247 /* Small optimization: do not loopback not local frames,
248 which returned after forwarding; they will be dropped
249 by ip_mr_input in any case.
250 Note, that local frames are looped back to be delivered
253 This check is duplicated in ip_mr_input at the moment.
255 && ((rt->rt_flags&RTCF_LOCAL) || !(IPCB(skb)->flags&IPSKB_FORWARDED))
258 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
260 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
262 ip_dev_loopback_xmit);
265 /* Multicasts with ttl 0 must not go beyond the host */
267 if (skb->nh.iph->ttl == 0) {
273 if (rt->rt_flags&RTCF_BROADCAST) {
274 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
276 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
277 newskb->dev, ip_dev_loopback_xmit);
280 if (skb->len > dst_pmtu(&rt->u.dst) || skb_shinfo(skb)->frag_list)
281 return ip_fragment(skb, ip_finish_output);
283 return ip_finish_output(skb);
286 int ip_output(struct sk_buff *skb)
288 IP_INC_STATS(IpOutRequests);
290 if ((skb->len > dst_pmtu(skb->dst) || skb_shinfo(skb)->frag_list) &&
291 !skb_shinfo(skb)->tso_size)
292 return ip_fragment(skb, ip_finish_output);
294 return ip_finish_output(skb);
297 int ip_queue_xmit(struct sk_buff *skb, int ipfragok)
299 struct sock *sk = skb->sk;
300 struct inet_opt *inet = inet_sk(sk);
301 struct ip_options *opt = inet->opt;
306 /* Skip all of this if the packet is already routed,
307 * f.e. by something like SCTP.
309 rt = (struct rtable *) skb->dst;
313 /* Make sure we can route this packet. */
314 rt = (struct rtable *)__sk_dst_check(sk, 0);
318 /* Use correct destination address if we have options. */
324 struct flowi fl = { .oif = sk->sk_bound_dev_if,
327 .saddr = inet->saddr,
328 .tos = RT_CONN_FLAGS(sk) } },
329 .proto = sk->sk_protocol,
331 { .sport = inet->sport,
332 .dport = inet->dport } } };
334 /* If this fails, retransmit mechanism of transport layer will
335 * keep trying until route appears or the connection times
338 if (ip_route_output_flow(&rt, &fl, sk, 0))
341 __sk_dst_set(sk, &rt->u.dst);
342 tcp_v4_setup_caps(sk, &rt->u.dst);
344 skb->dst = dst_clone(&rt->u.dst);
347 if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
350 /* OK, we know where to send it, allocate and build IP header. */
351 iph = (struct iphdr *) skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
352 *((__u16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
353 iph->tot_len = htons(skb->len);
354 if (ip_dont_fragment(sk, &rt->u.dst) && !ipfragok)
355 iph->frag_off = htons(IP_DF);
358 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
359 iph->protocol = sk->sk_protocol;
360 iph->saddr = rt->rt_src;
361 iph->daddr = rt->rt_dst;
363 /* Transport layer set skb->h.foo itself. */
365 if(opt && opt->optlen) {
366 iph->ihl += opt->optlen >> 2;
367 ip_options_build(skb, opt, inet->daddr, rt, 0);
370 mtu = dst_pmtu(&rt->u.dst);
371 if (skb->len > mtu && (sk->sk_route_caps & NETIF_F_TSO)) {
374 /* Hack zone: all this must be done by TCP. */
375 hlen = ((skb->h.raw - skb->data) + (skb->h.th->doff << 2));
376 skb_shinfo(skb)->tso_size = mtu - hlen;
377 skb_shinfo(skb)->tso_segs =
378 (skb->len - hlen + skb_shinfo(skb)->tso_size - 1)/
379 skb_shinfo(skb)->tso_size - 1;
382 ip_select_ident_more(iph, &rt->u.dst, sk, skb_shinfo(skb)->tso_segs);
384 /* Add an IP checksum. */
387 skb->priority = sk->sk_priority;
389 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
393 IP_INC_STATS(IpOutNoRoutes);
395 return -EHOSTUNREACH;
399 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
401 to->pkt_type = from->pkt_type;
402 to->priority = from->priority;
403 to->protocol = from->protocol;
404 to->security = from->security;
405 to->dst = dst_clone(from->dst);
408 /* Copy the flags to each fragment. */
409 IPCB(to)->flags = IPCB(from)->flags;
411 #ifdef CONFIG_NET_SCHED
412 to->tc_index = from->tc_index;
414 #ifdef CONFIG_NETFILTER
415 to->nfmark = from->nfmark;
416 to->nfcache = from->nfcache;
417 /* Connection association is same as pre-frag packet */
418 nf_conntrack_put(to->nfct);
419 to->nfct = from->nfct;
420 nf_conntrack_get(to->nfct);
421 #ifdef CONFIG_BRIDGE_NETFILTER
422 nf_bridge_put(to->nf_bridge);
423 to->nf_bridge = from->nf_bridge;
424 nf_bridge_get(to->nf_bridge);
426 #ifdef CONFIG_NETFILTER_DEBUG
427 to->nf_debug = from->nf_debug;
433 * This IP datagram is too large to be sent in one piece. Break it up into
434 * smaller pieces (each of size equal to IP header plus
435 * a block of the data of the original IP data part) that will yet fit in a
436 * single device frame, and queue such a frame for sending.
439 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*))
444 struct net_device *dev;
445 struct sk_buff *skb2;
446 unsigned int mtu, hlen, left, len, ll_rs;
449 struct rtable *rt = (struct rtable*)skb->dst;
455 * Point into the IP datagram header.
460 if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
461 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
462 htonl(dst_pmtu(&rt->u.dst)));
468 * Setup starting values.
472 mtu = dst_pmtu(&rt->u.dst) - hlen; /* Size of data space */
474 /* When frag_list is given, use it. First, check its validity:
475 * some transformers could create wrong frag_list or break existing
476 * one, it is not prohibited. In this case fall back to copying.
478 * LATER: this step can be merged to real generation of fragments,
479 * we can switch to copy when see the first bad fragment.
481 if (skb_shinfo(skb)->frag_list) {
482 struct sk_buff *frag;
483 int first_len = skb_pagelen(skb);
485 if (first_len - hlen > mtu ||
486 ((first_len - hlen) & 7) ||
487 (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
491 for (frag = skb_shinfo(skb)->frag_list; frag; frag = frag->next) {
492 /* Correct geometry. */
493 if (frag->len > mtu ||
494 ((frag->len & 7) && frag->next) ||
495 skb_headroom(frag) < hlen)
498 /* Correct socket ownership. */
499 if (frag->sk == NULL && skb->sk)
502 /* Partially cloned skb? */
503 if (skb_shared(frag))
507 /* Everything is OK. Generate! */
511 frag = skb_shinfo(skb)->frag_list;
512 skb_shinfo(skb)->frag_list = 0;
513 skb->data_len = first_len - skb_headlen(skb);
514 skb->len = first_len;
515 iph->tot_len = htons(first_len);
516 iph->frag_off |= htons(IP_MF);
520 /* Prepare header of the next frame,
521 * before previous one went down. */
523 frag->h.raw = frag->data;
524 frag->nh.raw = __skb_push(frag, hlen);
525 memcpy(frag->nh.raw, iph, hlen);
527 iph->tot_len = htons(frag->len);
528 ip_copy_metadata(frag, skb);
530 ip_options_fragment(frag);
531 offset += skb->len - hlen;
532 iph->frag_off = htons(offset>>3);
533 if (frag->next != NULL)
534 iph->frag_off |= htons(IP_MF);
535 /* Ready, complete checksum */
550 IP_INC_STATS(IpFragOKs);
559 IP_INC_STATS(IpFragFails);
564 left = skb->len - hlen; /* Space per frame */
565 ptr = raw + hlen; /* Where to start from */
567 #ifdef CONFIG_BRIDGE_NETFILTER
568 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
569 * we need to make room for the encapsulating header */
570 ll_rs = LL_RESERVED_SPACE(rt->u.dst.dev + nf_bridge_pad(skb));
571 mtu -= nf_bridge_pad(skb);
573 ll_rs = LL_RESERVED_SPACE(rt->u.dst.dev);
576 * Fragment the datagram.
579 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
580 not_last_frag = iph->frag_off & htons(IP_MF);
583 * Keep copying data until we run out.
588 /* IF: it doesn't fit, use 'mtu' - the data space left */
591 /* IF: we are not sending upto and including the packet end
592 then align the next start on an eight byte boundary */
600 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
601 NETDEBUG(printk(KERN_INFO "IP: frag: no memory for new fragment!\n"));
607 * Set up data on packet
610 ip_copy_metadata(skb2, skb);
611 skb_reserve(skb2, ll_rs);
612 skb_put(skb2, len + hlen);
613 skb2->nh.raw = skb2->data;
614 skb2->h.raw = skb2->data + hlen;
617 * Charge the memory for the fragment to any owner
622 skb_set_owner_w(skb2, skb->sk);
625 * Copy the packet header into the new buffer.
628 memcpy(skb2->nh.raw, skb->data, hlen);
631 * Copy a block of the IP datagram.
633 if (skb_copy_bits(skb, ptr, skb2->h.raw, len))
638 * Fill in the new header fields.
641 iph->frag_off = htons((offset >> 3));
643 /* ANK: dirty, but effective trick. Upgrade options only if
644 * the segment to be fragmented was THE FIRST (otherwise,
645 * options are already fixed) and make it ONCE
646 * on the initial skb, so that all the following fragments
647 * will inherit fixed options.
650 ip_options_fragment(skb);
653 * Added AC : If we are fragmenting a fragment that's not the
654 * last fragment then keep MF on each bit
656 if (left > 0 || not_last_frag)
657 iph->frag_off |= htons(IP_MF);
662 * Put this fragment into the sending queue.
665 IP_INC_STATS(IpFragCreates);
667 iph->tot_len = htons(len + hlen);
676 IP_INC_STATS(IpFragOKs);
681 IP_INC_STATS(IpFragFails);
686 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
688 struct iovec *iov = from;
690 if (skb->ip_summed == CHECKSUM_HW) {
691 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
694 unsigned int csum = 0;
695 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
697 skb->csum = csum_block_add(skb->csum, csum, odd);
703 skb_can_coalesce(struct sk_buff *skb, int i, struct page *page, int off)
706 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
707 return page == frag->page &&
708 off == frag->page_offset+frag->size;
713 static inline unsigned int
714 csum_page(struct page *page, int offset, int copy)
719 csum = csum_partial(kaddr + offset, copy, 0);
725 * ip_append_data() and ip_append_page() can make one large IP datagram
726 * from many pieces of data. Each pieces will be holded on the socket
727 * until ip_push_pending_frames() is called. Eache pieces can be a page
730 * Not only UDP, other transport protocols - e.g. raw sockets - can use
731 * this interface potentially.
733 * LATER: length must be adjusted by pad at tail, when it is required.
735 int ip_append_data(struct sock *sk,
736 int getfrag(void *from, char *to, int offset, int len,
737 int odd, struct sk_buff *skb),
738 void *from, int length, int transhdrlen,
739 struct ipcm_cookie *ipc, struct rtable *rt,
742 struct inet_opt *inet = inet_sk(sk);
745 struct ip_options *opt = NULL;
752 unsigned int maxfraglen, fragheaderlen;
753 int csummode = CHECKSUM_NONE;
758 if (skb_queue_empty(&sk->sk_write_queue)) {
764 if (inet->cork.opt == NULL) {
765 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
766 if (unlikely(inet->cork.opt == NULL))
769 memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
770 inet->cork.flags |= IPCORK_OPT;
771 inet->cork.addr = ipc->addr;
773 dst_hold(&rt->u.dst);
774 inet->cork.fragsize = mtu = dst_pmtu(&rt->u.dst);
776 inet->cork.length = 0;
777 inet->sndmsg_page = NULL;
778 inet->sndmsg_off = 0;
779 if ((exthdrlen = rt->u.dst.header_len) != 0) {
781 transhdrlen += exthdrlen;
785 if (inet->cork.flags & IPCORK_OPT)
786 opt = inet->cork.opt;
790 mtu = inet->cork.fragsize;
792 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
794 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
795 maxfraglen = ((mtu-fragheaderlen) & ~7) + fragheaderlen;
797 if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
798 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu-exthdrlen);
803 * transhdrlen > 0 means that this is the first fragment and we wish
804 * it won't be fragmented in the future.
807 length + fragheaderlen <= maxfraglen &&
808 rt->u.dst.dev->features&(NETIF_F_IP_CSUM|NETIF_F_NO_CSUM|NETIF_F_HW_CSUM) &&
810 csummode = CHECKSUM_HW;
812 inet->cork.length += length;
814 /* So, what's going on in the loop below?
816 * We use calculated fragment length to generate chained skb,
817 * each of segments is IP fragment ready for sending to network after
818 * adding appropriate IP header.
822 * If mtu-fragheaderlen is not 0 modulo 8, we generate additional
823 * small fragment of length (mtu-fragheaderlen)%8, even though
824 * it is not necessary. Not a big bug, but needs a fix.
827 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
831 if ((copy = maxfraglen - skb->len) <= 0) {
833 unsigned int datalen;
834 unsigned int fraglen;
835 unsigned int alloclen;
839 datalen = maxfraglen - fragheaderlen;
840 if (datalen > length)
843 fraglen = datalen + fragheaderlen;
844 if ((flags & MSG_MORE) &&
845 !(rt->u.dst.dev->features&NETIF_F_SG))
846 alloclen = maxfraglen;
848 alloclen = datalen + fragheaderlen;
850 /* The last fragment gets additional space at tail.
851 * Note, with MSG_MORE we overallocate on fragments,
852 * because we have no idea what fragment will be
855 if (datalen == length)
856 alloclen += rt->u.dst.trailer_len;
859 skb = sock_alloc_send_skb(sk,
860 alloclen + hh_len + 15,
861 (flags & MSG_DONTWAIT), &err);
864 if (atomic_read(&sk->sk_wmem_alloc) <=
866 skb = sock_wmalloc(sk,
867 alloclen + hh_len + 15, 1,
869 if (unlikely(skb == NULL))
876 * Fill in the control structures
878 skb->ip_summed = csummode;
880 skb_reserve(skb, hh_len);
883 * Find where to start putting bytes.
885 data = skb_put(skb, fraglen);
886 skb->nh.raw = data + exthdrlen;
887 data += fragheaderlen;
888 skb->h.raw = data + exthdrlen;
890 copy = datalen - transhdrlen;
891 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, 0, skb) < 0) {
901 csummode = CHECKSUM_NONE;
904 * Put the packet on the pending queue.
906 __skb_queue_tail(&sk->sk_write_queue, skb);
913 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
917 if (getfrag(from, skb_put(skb, copy),
918 offset, copy, off, skb) < 0) {
919 __skb_trim(skb, off);
924 int i = skb_shinfo(skb)->nr_frags;
925 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
926 struct page *page = inet->sndmsg_page;
927 int off = inet->sndmsg_off;
930 if (page && (left = PAGE_SIZE - off) > 0) {
933 if (page != frag->page) {
934 if (i == MAX_SKB_FRAGS) {
939 skb_fill_page_desc(skb, i, page, inet->sndmsg_off, 0);
940 frag = &skb_shinfo(skb)->frags[i];
942 } else if (i < MAX_SKB_FRAGS) {
943 if (copy > PAGE_SIZE)
945 page = alloc_pages(sk->sk_allocation, 0);
950 inet->sndmsg_page = page;
951 inet->sndmsg_off = 0;
953 skb_fill_page_desc(skb, i, page, 0, 0);
954 frag = &skb_shinfo(skb)->frags[i];
955 skb->truesize += PAGE_SIZE;
956 atomic_add(PAGE_SIZE, &sk->sk_wmem_alloc);
961 if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
965 inet->sndmsg_off += copy;
968 skb->data_len += copy;
977 inet->cork.length -= length;
978 IP_INC_STATS(IpOutDiscards);
982 ssize_t ip_append_page(struct sock *sk, struct page *page,
983 int offset, size_t size, int flags)
985 struct inet_opt *inet = inet_sk(sk);
988 struct ip_options *opt = NULL;
993 unsigned int maxfraglen, fragheaderlen;
1001 if (skb_queue_empty(&sk->sk_write_queue))
1005 if (inet->cork.flags & IPCORK_OPT)
1006 opt = inet->cork.opt;
1008 if (!(rt->u.dst.dev->features&NETIF_F_SG))
1011 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1012 mtu = inet->cork.fragsize;
1014 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1015 maxfraglen = ((mtu-fragheaderlen) & ~7) + fragheaderlen;
1017 if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1018 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu);
1022 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1025 inet->cork.length += size;
1029 if ((len = maxfraglen - skb->len) <= 0) {
1034 skb = sock_wmalloc(sk, fragheaderlen + hh_len + 15, 1,
1036 if (unlikely(!skb)) {
1042 * Fill in the control structures
1044 skb->ip_summed = CHECKSUM_NONE;
1046 skb_reserve(skb, hh_len);
1049 * Find where to start putting bytes.
1051 data = skb_put(skb, fragheaderlen);
1052 skb->nh.iph = iph = (struct iphdr *)data;
1053 data += fragheaderlen;
1057 * Put the packet on the pending queue.
1059 __skb_queue_tail(&sk->sk_write_queue, skb);
1063 i = skb_shinfo(skb)->nr_frags;
1066 if (skb_can_coalesce(skb, i, page, offset)) {
1067 skb_shinfo(skb)->frags[i-1].size += len;
1068 } else if (i < MAX_SKB_FRAGS) {
1070 skb_fill_page_desc(skb, i, page, offset, len);
1076 if (skb->ip_summed == CHECKSUM_NONE) {
1078 csum = csum_page(page, offset, len);
1079 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1083 skb->data_len += len;
1090 inet->cork.length -= size;
1091 IP_INC_STATS(IpOutDiscards);
1096 * Combined all pending IP fragments on the socket as one IP datagram
1097 * and push them out.
1099 int ip_push_pending_frames(struct sock *sk)
1101 struct sk_buff *skb, *tmp_skb;
1102 struct sk_buff **tail_skb;
1103 struct inet_opt *inet = inet_sk(sk);
1104 struct ip_options *opt = NULL;
1105 struct rtable *rt = inet->cork.rt;
1111 if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1113 tail_skb = &(skb_shinfo(skb)->frag_list);
1115 /* move skb->data to ip header from ext header */
1116 if (skb->data < skb->nh.raw)
1117 __skb_pull(skb, skb->nh.raw - skb->data);
1118 while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1119 __skb_pull(tmp_skb, skb->h.raw - skb->nh.raw);
1120 *tail_skb = tmp_skb;
1121 tail_skb = &(tmp_skb->next);
1122 skb->len += tmp_skb->len;
1123 skb->data_len += tmp_skb->len;
1124 #if 0 /* Logically correct, but useless work, ip_fragment() will have to undo */
1125 skb->truesize += tmp_skb->truesize;
1126 __sock_put(tmp_skb->sk);
1127 tmp_skb->destructor = NULL;
1132 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1133 * to fragment the frame generated here. No matter, what transforms
1134 * how transforms change size of the packet, it will come out.
1136 if (inet->pmtudisc != IP_PMTUDISC_DO)
1139 /* DF bit is set when we want to see DF on outgoing frames.
1140 * If local_df is set too, we still allow to fragment this frame
1142 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1143 (!skb_shinfo(skb)->frag_list && ip_dont_fragment(sk, &rt->u.dst)))
1146 if (inet->cork.flags & IPCORK_OPT)
1147 opt = inet->cork.opt;
1149 if (rt->rt_type == RTN_MULTICAST)
1152 ttl = ip_select_ttl(inet, &rt->u.dst);
1154 iph = (struct iphdr *)skb->data;
1158 iph->ihl += opt->optlen>>2;
1159 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1161 iph->tos = inet->tos;
1162 iph->tot_len = htons(skb->len);
1165 __ip_select_ident(iph, &rt->u.dst, 0);
1167 iph->id = htons(inet->id++);
1170 iph->protocol = sk->sk_protocol;
1171 iph->saddr = rt->rt_src;
1172 iph->daddr = rt->rt_dst;
1175 skb->priority = sk->sk_priority;
1176 skb->dst = dst_clone(&rt->u.dst);
1178 /* Netfilter gets whole the not fragmented skb. */
1179 err = NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL,
1180 skb->dst->dev, dst_output);
1183 err = inet->recverr ? net_xmit_errno(err) : 0;
1189 inet->cork.flags &= ~IPCORK_OPT;
1190 if (inet->cork.opt) {
1191 kfree(inet->cork.opt);
1192 inet->cork.opt = NULL;
1194 if (inet->cork.rt) {
1195 ip_rt_put(inet->cork.rt);
1196 inet->cork.rt = NULL;
1201 IP_INC_STATS(IpOutDiscards);
1206 * Throw away all pending data on the socket.
1208 void ip_flush_pending_frames(struct sock *sk)
1210 struct inet_opt *inet = inet_sk(sk);
1211 struct sk_buff *skb;
1213 while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1216 inet->cork.flags &= ~IPCORK_OPT;
1217 if (inet->cork.opt) {
1218 kfree(inet->cork.opt);
1219 inet->cork.opt = NULL;
1221 if (inet->cork.rt) {
1222 ip_rt_put(inet->cork.rt);
1223 inet->cork.rt = NULL;
1229 * Fetch data from kernel space and fill in checksum if needed.
1231 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1232 int len, int odd, struct sk_buff *skb)
1236 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1237 skb->csum = csum_block_add(skb->csum, csum, odd);
1242 * Generic function to send a packet as reply to another packet.
1243 * Used to send TCP resets so far. ICMP should use this function too.
1245 * Should run single threaded per socket because it uses the sock
1246 * structure to pass arguments.
1248 * LATER: switch from ip_build_xmit to ip_append_*
1250 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1253 struct inet_opt *inet = inet_sk(sk);
1255 struct ip_options opt;
1258 struct ipcm_cookie ipc;
1260 struct rtable *rt = (struct rtable*)skb->dst;
1262 if (ip_options_echo(&replyopts.opt, skb))
1265 daddr = ipc.addr = rt->rt_src;
1268 if (replyopts.opt.optlen) {
1269 ipc.opt = &replyopts.opt;
1272 daddr = replyopts.opt.faddr;
1276 struct flowi fl = { .nl_u = { .ip4_u =
1278 .saddr = rt->rt_spec_dst,
1279 .tos = RT_TOS(skb->nh.iph->tos) } },
1280 /* Not quite clean, but right. */
1282 { .sport = skb->h.th->dest,
1283 .dport = skb->h.th->source } },
1284 .proto = sk->sk_protocol };
1285 if (ip_route_output_key(&rt, &fl))
1289 /* And let IP do all the hard work.
1291 This chunk is not reenterable, hence spinlock.
1292 Note that it uses the fact, that this function is called
1293 with locally disabled BH and that sk cannot be already spinlocked.
1296 inet->tos = skb->nh.iph->tos;
1297 sk->sk_priority = skb->priority;
1298 sk->sk_protocol = skb->nh.iph->protocol;
1299 ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1300 &ipc, rt, MSG_DONTWAIT);
1301 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1302 if (arg->csumoffset >= 0)
1303 *((u16 *)skb->h.raw + arg->csumoffset) = csum_fold(csum_add(skb->csum, arg->csum));
1304 skb->ip_summed = CHECKSUM_NONE;
1305 ip_push_pending_frames(sk);
1314 * IP protocol layer initialiser
1317 static struct packet_type ip_packet_type = {
1318 .type = __constant_htons(ETH_P_IP),
1323 * IP registers the packet type and then calls the subprotocol initialisers
1326 void __init ip_init(void)
1328 dev_add_pack(&ip_packet_type);
1333 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1334 igmp_mc_proc_init();
1338 EXPORT_SYMBOL(ip_finish_output);
1339 EXPORT_SYMBOL(ip_fragment);
1340 EXPORT_SYMBOL(ip_generic_getfrag);
1341 EXPORT_SYMBOL(ip_queue_xmit);
1342 EXPORT_SYMBOL(ip_send_check);
1344 #ifdef CONFIG_SYSCTL
1345 EXPORT_SYMBOL(sysctl_ip_default_ttl);