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 $
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 <net/checksum.h>
82 #include <linux/igmp.h>
83 #include <linux/netfilter_ipv4.h>
84 #include <linux/netfilter_bridge.h>
85 #include <linux/mroute.h>
86 #include <linux/netlink.h>
89 * Shall we try to damage output packets if routing dev changes?
92 int sysctl_ip_dynaddr;
93 int sysctl_ip_default_ttl = IPDEFTTL;
95 /* Generate a checksum for an outgoing IP datagram. */
96 __inline__ void ip_send_check(struct iphdr *iph)
99 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
102 /* dev_loopback_xmit for use with netfilter. */
103 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
105 newskb->mac.raw = newskb->data;
106 __skb_pull(newskb, newskb->nh.raw - newskb->data);
107 newskb->pkt_type = PACKET_LOOPBACK;
108 newskb->ip_summed = CHECKSUM_UNNECESSARY;
109 BUG_TRAP(newskb->dst);
111 #ifdef CONFIG_NETFILTER_DEBUG
112 nf_debug_ip_loopback_xmit(newskb);
118 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
120 int ttl = inet->uc_ttl;
123 ttl = dst_metric(dst, RTAX_HOPLIMIT);
128 * Add an ip header to a skbuff and send it out.
131 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
132 u32 saddr, u32 daddr, struct ip_options *opt)
134 struct inet_sock *inet = inet_sk(sk);
135 struct rtable *rt = (struct rtable *)skb->dst;
138 /* Build the IP header. */
140 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr) + opt->optlen);
142 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr));
146 iph->tos = inet->tos;
147 if (ip_dont_fragment(sk, &rt->u.dst))
148 iph->frag_off = htons(IP_DF);
151 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
152 iph->daddr = rt->rt_dst;
153 iph->saddr = rt->rt_src;
154 iph->protocol = sk->sk_protocol;
155 iph->tot_len = htons(skb->len);
156 ip_select_ident(iph, &rt->u.dst, sk);
159 if (opt && opt->optlen) {
160 iph->ihl += opt->optlen>>2;
161 ip_options_build(skb, opt, daddr, rt, 0);
165 skb->priority = sk->sk_priority;
168 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
172 static inline int ip_finish_output2(struct sk_buff *skb)
174 struct dst_entry *dst = skb->dst;
175 struct hh_cache *hh = dst->hh;
176 struct net_device *dev = dst->dev;
177 int hh_len = LL_RESERVED_SPACE(dev);
179 /* Be paranoid, rather than too clever. */
180 if (unlikely(skb_headroom(skb) < hh_len && dev->hard_header)) {
181 struct sk_buff *skb2;
183 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
189 skb_set_owner_w(skb2, skb->sk);
194 #ifdef CONFIG_NETFILTER_DEBUG
195 nf_debug_ip_finish_output2(skb);
196 #endif /*CONFIG_NETFILTER_DEBUG*/
201 read_lock_bh(&hh->hh_lock);
202 hh_alen = HH_DATA_ALIGN(hh->hh_len);
203 memcpy(skb->data - hh_alen, hh->hh_data, hh_alen);
204 read_unlock_bh(&hh->hh_lock);
205 skb_push(skb, hh->hh_len);
206 return hh->hh_output(skb);
207 } else if (dst->neighbour)
208 return dst->neighbour->output(skb);
211 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
216 int ip_finish_output(struct sk_buff *skb)
218 struct net_device *dev = skb->dst->dev;
221 skb->protocol = htons(ETH_P_IP);
223 return NF_HOOK(PF_INET, NF_IP_POST_ROUTING, skb, NULL, dev,
227 int ip_mc_output(struct sk_buff *skb)
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_mtu(&rt->u.dst))
282 return ip_fragment(skb, ip_finish_output);
284 return ip_finish_output(skb);
287 int ip_output(struct sk_buff *skb)
289 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
291 if (skb->len > dst_mtu(skb->dst) && !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_sock *inet = inet_sk(sk);
301 struct ip_options *opt = inet->opt;
305 /* Skip all of this if the packet is already routed,
306 * f.e. by something like SCTP.
308 rt = (struct rtable *) skb->dst;
312 /* Make sure we can route this packet. */
313 rt = (struct rtable *)__sk_dst_check(sk, 0);
317 /* Use correct destination address if we have options. */
323 struct flowi fl = { .oif = sk->sk_bound_dev_if,
326 .saddr = inet->saddr,
327 .tos = RT_CONN_FLAGS(sk) } },
328 .proto = sk->sk_protocol,
330 { .sport = inet->sport,
331 .dport = inet->dport } } };
333 /* If this fails, retransmit mechanism of transport layer will
334 * keep trying until route appears or the connection times
337 if (ip_route_output_flow(&rt, &fl, sk, 0))
340 __sk_dst_set(sk, &rt->u.dst);
341 tcp_v4_setup_caps(sk, &rt->u.dst);
343 skb->dst = dst_clone(&rt->u.dst);
346 if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
349 /* OK, we know where to send it, allocate and build IP header. */
350 iph = (struct iphdr *) skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
351 *((__u16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
352 iph->tot_len = htons(skb->len);
353 if (ip_dont_fragment(sk, &rt->u.dst) && !ipfragok)
354 iph->frag_off = htons(IP_DF);
357 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
358 iph->protocol = sk->sk_protocol;
359 iph->saddr = rt->rt_src;
360 iph->daddr = rt->rt_dst;
362 /* Transport layer set skb->h.foo itself. */
364 if (opt && opt->optlen) {
365 iph->ihl += opt->optlen >> 2;
366 ip_options_build(skb, opt, inet->daddr, rt, 0);
369 ip_select_ident_more(iph, &rt->u.dst, sk, skb_shinfo(skb)->tso_segs);
371 /* Add an IP checksum. */
374 skb->priority = sk->sk_priority;
376 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
380 IP_INC_STATS(IPSTATS_MIB_OUTNOROUTES);
382 return -EHOSTUNREACH;
386 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
388 to->pkt_type = from->pkt_type;
389 to->priority = from->priority;
390 to->protocol = from->protocol;
391 to->security = from->security;
392 dst_release(to->dst);
393 to->dst = dst_clone(from->dst);
396 /* Copy the flags to each fragment. */
397 IPCB(to)->flags = IPCB(from)->flags;
399 #ifdef CONFIG_NET_SCHED
400 to->tc_index = from->tc_index;
402 #ifdef CONFIG_NETFILTER
403 to->nfmark = from->nfmark;
404 to->nfcache = from->nfcache;
405 /* Connection association is same as pre-frag packet */
406 nf_conntrack_put(to->nfct);
407 to->nfct = from->nfct;
408 nf_conntrack_get(to->nfct);
409 to->nfctinfo = from->nfctinfo;
410 #ifdef CONFIG_BRIDGE_NETFILTER
411 nf_bridge_put(to->nf_bridge);
412 to->nf_bridge = from->nf_bridge;
413 nf_bridge_get(to->nf_bridge);
415 #ifdef CONFIG_NETFILTER_DEBUG
416 to->nf_debug = from->nf_debug;
422 * This IP datagram is too large to be sent in one piece. Break it up into
423 * smaller pieces (each of size equal to IP header plus
424 * a block of the data of the original IP data part) that will yet fit in a
425 * single device frame, and queue such a frame for sending.
428 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*))
433 struct net_device *dev;
434 struct sk_buff *skb2;
435 unsigned int mtu, hlen, left, len, ll_rs;
438 struct rtable *rt = (struct rtable*)skb->dst;
444 * Point into the IP datagram header.
449 if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
450 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
451 htonl(dst_mtu(&rt->u.dst)));
457 * Setup starting values.
461 mtu = dst_mtu(&rt->u.dst) - hlen; /* Size of data space */
463 /* When frag_list is given, use it. First, check its validity:
464 * some transformers could create wrong frag_list or break existing
465 * one, it is not prohibited. In this case fall back to copying.
467 * LATER: this step can be merged to real generation of fragments,
468 * we can switch to copy when see the first bad fragment.
470 if (skb_shinfo(skb)->frag_list) {
471 struct sk_buff *frag;
472 int first_len = skb_pagelen(skb);
474 if (first_len - hlen > mtu ||
475 ((first_len - hlen) & 7) ||
476 (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
480 for (frag = skb_shinfo(skb)->frag_list; frag; frag = frag->next) {
481 /* Correct geometry. */
482 if (frag->len > mtu ||
483 ((frag->len & 7) && frag->next) ||
484 skb_headroom(frag) < hlen)
487 /* Partially cloned skb? */
488 if (skb_shared(frag))
495 frag->destructor = sock_wfree;
496 skb->truesize -= frag->truesize;
500 /* Everything is OK. Generate! */
504 frag = skb_shinfo(skb)->frag_list;
505 skb_shinfo(skb)->frag_list = NULL;
506 skb->data_len = first_len - skb_headlen(skb);
507 skb->len = first_len;
508 iph->tot_len = htons(first_len);
509 iph->frag_off = htons(IP_MF);
513 /* Prepare header of the next frame,
514 * before previous one went down. */
516 frag->ip_summed = CHECKSUM_NONE;
517 frag->h.raw = frag->data;
518 frag->nh.raw = __skb_push(frag, hlen);
519 memcpy(frag->nh.raw, iph, hlen);
521 iph->tot_len = htons(frag->len);
522 ip_copy_metadata(frag, skb);
524 ip_options_fragment(frag);
525 offset += skb->len - hlen;
526 iph->frag_off = htons(offset>>3);
527 if (frag->next != NULL)
528 iph->frag_off |= htons(IP_MF);
529 /* Ready, complete checksum */
544 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
553 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
558 left = skb->len - hlen; /* Space per frame */
559 ptr = raw + hlen; /* Where to start from */
561 #ifdef CONFIG_BRIDGE_NETFILTER
562 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
563 * we need to make room for the encapsulating header */
564 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, nf_bridge_pad(skb));
565 mtu -= nf_bridge_pad(skb);
567 ll_rs = LL_RESERVED_SPACE(rt->u.dst.dev);
570 * Fragment the datagram.
573 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
574 not_last_frag = iph->frag_off & htons(IP_MF);
577 * Keep copying data until we run out.
582 /* IF: it doesn't fit, use 'mtu' - the data space left */
585 /* IF: we are not sending upto and including the packet end
586 then align the next start on an eight byte boundary */
594 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
595 NETDEBUG(printk(KERN_INFO "IP: frag: no memory for new fragment!\n"));
601 * Set up data on packet
604 ip_copy_metadata(skb2, skb);
605 skb_reserve(skb2, ll_rs);
606 skb_put(skb2, len + hlen);
607 skb2->nh.raw = skb2->data;
608 skb2->h.raw = skb2->data + hlen;
611 * Charge the memory for the fragment to any owner
616 skb_set_owner_w(skb2, skb->sk);
619 * Copy the packet header into the new buffer.
622 memcpy(skb2->nh.raw, skb->data, hlen);
625 * Copy a block of the IP datagram.
627 if (skb_copy_bits(skb, ptr, skb2->h.raw, len))
632 * Fill in the new header fields.
635 iph->frag_off = htons((offset >> 3));
637 /* ANK: dirty, but effective trick. Upgrade options only if
638 * the segment to be fragmented was THE FIRST (otherwise,
639 * options are already fixed) and make it ONCE
640 * on the initial skb, so that all the following fragments
641 * will inherit fixed options.
644 ip_options_fragment(skb);
647 * Added AC : If we are fragmenting a fragment that's not the
648 * last fragment then keep MF on each bit
650 if (left > 0 || not_last_frag)
651 iph->frag_off |= htons(IP_MF);
656 * Put this fragment into the sending queue.
659 IP_INC_STATS(IPSTATS_MIB_FRAGCREATES);
661 iph->tot_len = htons(len + hlen);
670 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
675 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
680 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
682 struct iovec *iov = from;
684 if (skb->ip_summed == CHECKSUM_HW) {
685 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
688 unsigned int csum = 0;
689 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
691 skb->csum = csum_block_add(skb->csum, csum, odd);
696 static inline unsigned int
697 csum_page(struct page *page, int offset, int copy)
702 csum = csum_partial(kaddr + offset, copy, 0);
708 * ip_append_data() and ip_append_page() can make one large IP datagram
709 * from many pieces of data. Each pieces will be holded on the socket
710 * until ip_push_pending_frames() is called. Each piece can be a page
713 * Not only UDP, other transport protocols - e.g. raw sockets - can use
714 * this interface potentially.
716 * LATER: length must be adjusted by pad at tail, when it is required.
718 int ip_append_data(struct sock *sk,
719 int getfrag(void *from, char *to, int offset, int len,
720 int odd, struct sk_buff *skb),
721 void *from, int length, int transhdrlen,
722 struct ipcm_cookie *ipc, struct rtable *rt,
725 struct inet_sock *inet = inet_sk(sk);
728 struct ip_options *opt = NULL;
735 unsigned int maxfraglen, fragheaderlen;
736 int csummode = CHECKSUM_NONE;
741 if (skb_queue_empty(&sk->sk_write_queue)) {
747 if (inet->cork.opt == NULL) {
748 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
749 if (unlikely(inet->cork.opt == NULL))
752 memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
753 inet->cork.flags |= IPCORK_OPT;
754 inet->cork.addr = ipc->addr;
756 dst_hold(&rt->u.dst);
757 inet->cork.fragsize = mtu = dst_mtu(rt->u.dst.path);
759 inet->cork.length = 0;
760 sk->sk_sndmsg_page = NULL;
761 sk->sk_sndmsg_off = 0;
762 if ((exthdrlen = rt->u.dst.header_len) != 0) {
764 transhdrlen += exthdrlen;
768 if (inet->cork.flags & IPCORK_OPT)
769 opt = inet->cork.opt;
773 mtu = inet->cork.fragsize;
775 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
777 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
778 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
780 if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
781 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu-exthdrlen);
786 * transhdrlen > 0 means that this is the first fragment and we wish
787 * it won't be fragmented in the future.
790 length + fragheaderlen <= mtu &&
791 rt->u.dst.dev->features&(NETIF_F_IP_CSUM|NETIF_F_NO_CSUM|NETIF_F_HW_CSUM) &&
793 csummode = CHECKSUM_HW;
795 inet->cork.length += length;
797 /* So, what's going on in the loop below?
799 * We use calculated fragment length to generate chained skb,
800 * each of segments is IP fragment ready for sending to network after
801 * adding appropriate IP header.
804 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
808 /* Check if the remaining data fits into current packet. */
809 copy = mtu - skb->len;
811 copy = maxfraglen - skb->len;
814 unsigned int datalen;
815 unsigned int fraglen;
816 unsigned int fraggap;
817 unsigned int alloclen;
818 struct sk_buff *skb_prev;
822 fraggap = skb_prev->len - maxfraglen;
827 * If remaining data exceeds the mtu,
828 * we know we need more fragment(s).
830 datalen = length + fraggap;
831 if (datalen > mtu - fragheaderlen)
832 datalen = maxfraglen - fragheaderlen;
833 fraglen = datalen + fragheaderlen;
835 if ((flags & MSG_MORE) &&
836 !(rt->u.dst.dev->features&NETIF_F_SG))
839 alloclen = datalen + fragheaderlen;
841 /* The last fragment gets additional space at tail.
842 * Note, with MSG_MORE we overallocate on fragments,
843 * because we have no idea what fragment will be
846 if (datalen == length)
847 alloclen += rt->u.dst.trailer_len;
850 skb = sock_alloc_send_skb(sk,
851 alloclen + hh_len + 15,
852 (flags & MSG_DONTWAIT), &err);
855 if (atomic_read(&sk->sk_wmem_alloc) <=
857 skb = sock_wmalloc(sk,
858 alloclen + hh_len + 15, 1,
860 if (unlikely(skb == NULL))
867 * Fill in the control structures
869 skb->ip_summed = csummode;
871 skb_reserve(skb, hh_len);
874 * Find where to start putting bytes.
876 data = skb_put(skb, fraglen);
877 skb->nh.raw = data + exthdrlen;
878 data += fragheaderlen;
879 skb->h.raw = data + exthdrlen;
882 skb->csum = skb_copy_and_csum_bits(
883 skb_prev, maxfraglen,
884 data + transhdrlen, fraggap, 0);
885 skb_prev->csum = csum_sub(skb_prev->csum,
888 skb_trim(skb_prev, maxfraglen);
891 copy = datalen - transhdrlen - fraggap;
892 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
899 length -= datalen - fraggap;
902 csummode = CHECKSUM_NONE;
905 * Put the packet on the pending queue.
907 __skb_queue_tail(&sk->sk_write_queue, skb);
914 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
918 if (getfrag(from, skb_put(skb, copy),
919 offset, copy, off, skb) < 0) {
920 __skb_trim(skb, off);
925 int i = skb_shinfo(skb)->nr_frags;
926 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
927 struct page *page = sk->sk_sndmsg_page;
928 int off = sk->sk_sndmsg_off;
931 if (page && (left = PAGE_SIZE - off) > 0) {
934 if (page != frag->page) {
935 if (i == MAX_SKB_FRAGS) {
940 skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
941 frag = &skb_shinfo(skb)->frags[i];
943 } else if (i < MAX_SKB_FRAGS) {
944 if (copy > PAGE_SIZE)
946 page = alloc_pages(sk->sk_allocation, 0);
951 sk->sk_sndmsg_page = page;
952 sk->sk_sndmsg_off = 0;
954 skb_fill_page_desc(skb, i, page, 0, 0);
955 frag = &skb_shinfo(skb)->frags[i];
956 skb->truesize += PAGE_SIZE;
957 atomic_add(PAGE_SIZE, &sk->sk_wmem_alloc);
962 if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
966 sk->sk_sndmsg_off += copy;
969 skb->data_len += copy;
978 inet->cork.length -= length;
979 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
983 ssize_t ip_append_page(struct sock *sk, struct page *page,
984 int offset, size_t size, int flags)
986 struct inet_sock *inet = inet_sk(sk);
989 struct ip_options *opt = NULL;
994 unsigned int maxfraglen, fragheaderlen, fraggap;
1002 if (skb_queue_empty(&sk->sk_write_queue))
1006 if (inet->cork.flags & IPCORK_OPT)
1007 opt = inet->cork.opt;
1009 if (!(rt->u.dst.dev->features&NETIF_F_SG))
1012 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1013 mtu = inet->cork.fragsize;
1015 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1016 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1018 if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1019 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu);
1023 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1026 inet->cork.length += size;
1031 /* Check if the remaining data fits into current packet. */
1032 len = mtu - skb->len;
1034 len = maxfraglen - skb->len;
1036 struct sk_buff *skb_prev;
1043 fraggap = skb_prev->len - maxfraglen;
1047 alloclen = fragheaderlen + hh_len + fraggap + 15;
1048 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1049 if (unlikely(!skb)) {
1055 * Fill in the control structures
1057 skb->ip_summed = CHECKSUM_NONE;
1059 skb_reserve(skb, hh_len);
1062 * Find where to start putting bytes.
1064 data = skb_put(skb, fragheaderlen + fraggap);
1065 skb->nh.iph = iph = (struct iphdr *)data;
1066 data += fragheaderlen;
1070 skb->csum = skb_copy_and_csum_bits(
1071 skb_prev, maxfraglen,
1073 skb_prev->csum = csum_sub(skb_prev->csum,
1075 skb_trim(skb_prev, maxfraglen);
1079 * Put the packet on the pending queue.
1081 __skb_queue_tail(&sk->sk_write_queue, skb);
1085 i = skb_shinfo(skb)->nr_frags;
1088 if (skb_can_coalesce(skb, i, page, offset)) {
1089 skb_shinfo(skb)->frags[i-1].size += len;
1090 } else if (i < MAX_SKB_FRAGS) {
1092 skb_fill_page_desc(skb, i, page, offset, len);
1098 if (skb->ip_summed == CHECKSUM_NONE) {
1100 csum = csum_page(page, offset, len);
1101 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1105 skb->data_len += len;
1112 inet->cork.length -= size;
1113 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1118 * Combined all pending IP fragments on the socket as one IP datagram
1119 * and push them out.
1121 int ip_push_pending_frames(struct sock *sk)
1123 struct sk_buff *skb, *tmp_skb;
1124 struct sk_buff **tail_skb;
1125 struct inet_sock *inet = inet_sk(sk);
1126 struct ip_options *opt = NULL;
1127 struct rtable *rt = inet->cork.rt;
1133 if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1135 tail_skb = &(skb_shinfo(skb)->frag_list);
1137 /* move skb->data to ip header from ext header */
1138 if (skb->data < skb->nh.raw)
1139 __skb_pull(skb, skb->nh.raw - skb->data);
1140 while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1141 __skb_pull(tmp_skb, skb->h.raw - skb->nh.raw);
1142 *tail_skb = tmp_skb;
1143 tail_skb = &(tmp_skb->next);
1144 skb->len += tmp_skb->len;
1145 skb->data_len += tmp_skb->len;
1146 skb->truesize += tmp_skb->truesize;
1147 __sock_put(tmp_skb->sk);
1148 tmp_skb->destructor = NULL;
1152 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1153 * to fragment the frame generated here. No matter, what transforms
1154 * how transforms change size of the packet, it will come out.
1156 if (inet->pmtudisc != IP_PMTUDISC_DO)
1159 /* DF bit is set when we want to see DF on outgoing frames.
1160 * If local_df is set too, we still allow to fragment this frame
1162 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1163 (skb->len <= dst_mtu(&rt->u.dst) &&
1164 ip_dont_fragment(sk, &rt->u.dst)))
1167 if (inet->cork.flags & IPCORK_OPT)
1168 opt = inet->cork.opt;
1170 if (rt->rt_type == RTN_MULTICAST)
1173 ttl = ip_select_ttl(inet, &rt->u.dst);
1175 iph = (struct iphdr *)skb->data;
1179 iph->ihl += opt->optlen>>2;
1180 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1182 iph->tos = inet->tos;
1183 iph->tot_len = htons(skb->len);
1186 __ip_select_ident(iph, &rt->u.dst, 0);
1188 iph->id = htons(inet->id++);
1191 iph->protocol = sk->sk_protocol;
1192 iph->saddr = rt->rt_src;
1193 iph->daddr = rt->rt_dst;
1196 skb->priority = sk->sk_priority;
1197 skb->dst = dst_clone(&rt->u.dst);
1199 /* Netfilter gets whole the not fragmented skb. */
1200 err = NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL,
1201 skb->dst->dev, dst_output);
1204 err = inet->recverr ? net_xmit_errno(err) : 0;
1210 inet->cork.flags &= ~IPCORK_OPT;
1211 if (inet->cork.opt) {
1212 kfree(inet->cork.opt);
1213 inet->cork.opt = NULL;
1215 if (inet->cork.rt) {
1216 ip_rt_put(inet->cork.rt);
1217 inet->cork.rt = NULL;
1222 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1227 * Throw away all pending data on the socket.
1229 void ip_flush_pending_frames(struct sock *sk)
1231 struct inet_sock *inet = inet_sk(sk);
1232 struct sk_buff *skb;
1234 while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1237 inet->cork.flags &= ~IPCORK_OPT;
1238 if (inet->cork.opt) {
1239 kfree(inet->cork.opt);
1240 inet->cork.opt = NULL;
1242 if (inet->cork.rt) {
1243 ip_rt_put(inet->cork.rt);
1244 inet->cork.rt = NULL;
1250 * Fetch data from kernel space and fill in checksum if needed.
1252 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1253 int len, int odd, struct sk_buff *skb)
1257 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1258 skb->csum = csum_block_add(skb->csum, csum, odd);
1263 * Generic function to send a packet as reply to another packet.
1264 * Used to send TCP resets so far. ICMP should use this function too.
1266 * Should run single threaded per socket because it uses the sock
1267 * structure to pass arguments.
1269 * LATER: switch from ip_build_xmit to ip_append_*
1271 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1274 struct inet_sock *inet = inet_sk(sk);
1276 struct ip_options opt;
1279 struct ipcm_cookie ipc;
1281 struct rtable *rt = (struct rtable*)skb->dst;
1283 if (ip_options_echo(&replyopts.opt, skb))
1286 daddr = ipc.addr = rt->rt_src;
1289 if (replyopts.opt.optlen) {
1290 ipc.opt = &replyopts.opt;
1293 daddr = replyopts.opt.faddr;
1297 struct flowi fl = { .nl_u = { .ip4_u =
1299 .saddr = rt->rt_spec_dst,
1300 .tos = RT_TOS(skb->nh.iph->tos) } },
1301 /* Not quite clean, but right. */
1303 { .sport = skb->h.th->dest,
1304 .dport = skb->h.th->source } },
1305 .proto = sk->sk_protocol };
1306 if (ip_route_output_key(&rt, &fl))
1310 /* And let IP do all the hard work.
1312 This chunk is not reenterable, hence spinlock.
1313 Note that it uses the fact, that this function is called
1314 with locally disabled BH and that sk cannot be already spinlocked.
1317 inet->tos = skb->nh.iph->tos;
1318 sk->sk_priority = skb->priority;
1319 sk->sk_protocol = skb->nh.iph->protocol;
1320 ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1321 &ipc, rt, MSG_DONTWAIT);
1322 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1323 if (arg->csumoffset >= 0)
1324 *((u16 *)skb->h.raw + arg->csumoffset) = csum_fold(csum_add(skb->csum, arg->csum));
1325 skb->ip_summed = CHECKSUM_NONE;
1326 ip_push_pending_frames(sk);
1335 * IP protocol layer initialiser
1338 static struct packet_type ip_packet_type = {
1339 .type = __constant_htons(ETH_P_IP),
1344 * IP registers the packet type and then calls the subprotocol initialisers
1347 void __init ip_init(void)
1349 dev_add_pack(&ip_packet_type);
1354 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1355 igmp_mc_proc_init();
1359 EXPORT_SYMBOL(ip_finish_output);
1360 EXPORT_SYMBOL(ip_fragment);
1361 EXPORT_SYMBOL(ip_generic_getfrag);
1362 EXPORT_SYMBOL(ip_queue_xmit);
1363 EXPORT_SYMBOL(ip_send_check);
1365 #ifdef CONFIG_SYSCTL
1366 EXPORT_SYMBOL(sysctl_ip_default_ttl);