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 <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*/
203 read_lock_bh(&hh->hh_lock);
204 hh_alen = HH_DATA_ALIGN(hh->hh_len);
205 memcpy(skb->data - hh_alen, hh->hh_data, hh_alen);
206 read_unlock_bh(&hh->hh_lock);
207 skb_push(skb, hh->hh_len);
208 return hh->hh_output(skb);
209 } else if (dst->neighbour)
210 return dst->neighbour->output(skb);
213 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
218 int ip_finish_output(struct sk_buff *skb)
220 struct net_device *dev = skb->dst->dev;
223 skb->protocol = htons(ETH_P_IP);
225 return NF_HOOK(PF_INET, NF_IP_POST_ROUTING, skb, NULL, dev,
229 int ip_mc_output(struct sk_buff *skb)
231 struct sock *sk = skb->sk;
232 struct rtable *rt = (struct rtable*)skb->dst;
233 struct net_device *dev = rt->u.dst.dev;
236 * If the indicated interface is up and running, send the packet.
238 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
241 skb->protocol = htons(ETH_P_IP);
244 * Multicasts are looped back for other local users
247 if (rt->rt_flags&RTCF_MULTICAST) {
248 if ((!sk || inet_sk(sk)->mc_loop)
249 #ifdef CONFIG_IP_MROUTE
250 /* Small optimization: do not loopback not local frames,
251 which returned after forwarding; they will be dropped
252 by ip_mr_input in any case.
253 Note, that local frames are looped back to be delivered
256 This check is duplicated in ip_mr_input at the moment.
258 && ((rt->rt_flags&RTCF_LOCAL) || !(IPCB(skb)->flags&IPSKB_FORWARDED))
261 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
263 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
265 ip_dev_loopback_xmit);
268 /* Multicasts with ttl 0 must not go beyond the host */
270 if (skb->nh.iph->ttl == 0) {
276 if (rt->rt_flags&RTCF_BROADCAST) {
277 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
279 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
280 newskb->dev, ip_dev_loopback_xmit);
283 if (skb->len > dst_mtu(&rt->u.dst))
284 return ip_fragment(skb, ip_finish_output);
286 return ip_finish_output(skb);
289 int ip_output(struct sk_buff *skb)
291 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
293 if (skb->len > dst_mtu(skb->dst) && !skb_shinfo(skb)->tso_size)
294 return ip_fragment(skb, ip_finish_output);
296 return ip_finish_output(skb);
299 int ip_queue_xmit(struct sk_buff *skb, int ipfragok)
301 struct sock *sk = skb->sk;
302 struct inet_sock *inet = inet_sk(sk);
303 struct ip_options *opt = inet->opt;
307 /* Skip all of this if the packet is already routed,
308 * f.e. by something like SCTP.
310 rt = (struct rtable *) skb->dst;
314 /* Make sure we can route this packet. */
315 rt = (struct rtable *)__sk_dst_check(sk, 0);
319 /* Use correct destination address if we have options. */
325 struct flowi fl = { .oif = sk->sk_bound_dev_if,
328 .saddr = inet->saddr,
329 .tos = RT_CONN_FLAGS(sk) } },
330 .proto = sk->sk_protocol,
332 { .sport = inet->sport,
333 .dport = inet->dport } } };
335 /* If this fails, retransmit mechanism of transport layer will
336 * keep trying until route appears or the connection times
339 if (ip_route_output_flow(&rt, &fl, sk, 0))
342 __sk_dst_set(sk, &rt->u.dst);
343 tcp_v4_setup_caps(sk, &rt->u.dst);
345 skb->dst = dst_clone(&rt->u.dst);
348 if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
351 /* OK, we know where to send it, allocate and build IP header. */
352 iph = (struct iphdr *) skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
353 *((__u16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
354 iph->tot_len = htons(skb->len);
355 if (ip_dont_fragment(sk, &rt->u.dst) && !ipfragok)
356 iph->frag_off = htons(IP_DF);
359 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
360 iph->protocol = sk->sk_protocol;
361 iph->saddr = rt->rt_src;
362 iph->daddr = rt->rt_dst;
364 /* Transport layer set skb->h.foo itself. */
366 if (opt && opt->optlen) {
367 iph->ihl += opt->optlen >> 2;
368 ip_options_build(skb, opt, inet->daddr, rt, 0);
371 ip_select_ident_more(iph, &rt->u.dst, sk, skb_shinfo(skb)->tso_segs);
373 /* Add an IP checksum. */
376 skb->priority = sk->sk_priority;
378 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
382 IP_INC_STATS(IPSTATS_MIB_OUTNOROUTES);
384 return -EHOSTUNREACH;
388 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
390 to->pkt_type = from->pkt_type;
391 to->priority = from->priority;
392 to->protocol = from->protocol;
393 to->security = from->security;
394 dst_release(to->dst);
395 to->dst = dst_clone(from->dst);
398 /* Copy the flags to each fragment. */
399 IPCB(to)->flags = IPCB(from)->flags;
401 #ifdef CONFIG_NET_SCHED
402 to->tc_index = from->tc_index;
404 #ifdef CONFIG_NETFILTER
405 to->nfmark = from->nfmark;
406 to->nfcache = from->nfcache;
407 /* Connection association is same as pre-frag packet */
408 nf_conntrack_put(to->nfct);
409 to->nfct = from->nfct;
410 nf_conntrack_get(to->nfct);
411 to->nfctinfo = from->nfctinfo;
412 #ifdef CONFIG_BRIDGE_NETFILTER
413 nf_bridge_put(to->nf_bridge);
414 to->nf_bridge = from->nf_bridge;
415 nf_bridge_get(to->nf_bridge);
417 #ifdef CONFIG_NETFILTER_DEBUG
418 to->nf_debug = from->nf_debug;
424 * This IP datagram is too large to be sent in one piece. Break it up into
425 * smaller pieces (each of size equal to IP header plus
426 * a block of the data of the original IP data part) that will yet fit in a
427 * single device frame, and queue such a frame for sending.
430 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*))
435 struct net_device *dev;
436 struct sk_buff *skb2;
437 unsigned int mtu, hlen, left, len, ll_rs;
440 struct rtable *rt = (struct rtable*)skb->dst;
446 * Point into the IP datagram header.
451 if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
452 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
453 htonl(dst_mtu(&rt->u.dst)));
459 * Setup starting values.
463 mtu = dst_mtu(&rt->u.dst) - hlen; /* Size of data space */
465 /* When frag_list is given, use it. First, check its validity:
466 * some transformers could create wrong frag_list or break existing
467 * one, it is not prohibited. In this case fall back to copying.
469 * LATER: this step can be merged to real generation of fragments,
470 * we can switch to copy when see the first bad fragment.
472 if (skb_shinfo(skb)->frag_list) {
473 struct sk_buff *frag;
474 int first_len = skb_pagelen(skb);
476 if (first_len - hlen > mtu ||
477 ((first_len - hlen) & 7) ||
478 (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
482 for (frag = skb_shinfo(skb)->frag_list; frag; frag = frag->next) {
483 /* Correct geometry. */
484 if (frag->len > mtu ||
485 ((frag->len & 7) && frag->next) ||
486 skb_headroom(frag) < hlen)
489 /* Partially cloned skb? */
490 if (skb_shared(frag))
494 /* Everything is OK. Generate! */
498 frag = skb_shinfo(skb)->frag_list;
499 skb_shinfo(skb)->frag_list = NULL;
500 skb->data_len = first_len - skb_headlen(skb);
501 skb->len = first_len;
502 iph->tot_len = htons(first_len);
503 iph->frag_off = htons(IP_MF);
507 /* Prepare header of the next frame,
508 * before previous one went down. */
510 frag->ip_summed = CHECKSUM_NONE;
511 frag->h.raw = frag->data;
512 frag->nh.raw = __skb_push(frag, hlen);
513 memcpy(frag->nh.raw, iph, hlen);
515 iph->tot_len = htons(frag->len);
516 ip_copy_metadata(frag, skb);
518 ip_options_fragment(frag);
519 offset += skb->len - hlen;
520 iph->frag_off = htons(offset>>3);
521 if (frag->next != NULL)
522 iph->frag_off |= htons(IP_MF);
523 /* Ready, complete checksum */
538 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
547 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
552 left = skb->len - hlen; /* Space per frame */
553 ptr = raw + hlen; /* Where to start from */
555 #ifdef CONFIG_BRIDGE_NETFILTER
556 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
557 * we need to make room for the encapsulating header */
558 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, nf_bridge_pad(skb));
559 mtu -= nf_bridge_pad(skb);
561 ll_rs = LL_RESERVED_SPACE(rt->u.dst.dev);
564 * Fragment the datagram.
567 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
568 not_last_frag = iph->frag_off & htons(IP_MF);
571 * Keep copying data until we run out.
576 /* IF: it doesn't fit, use 'mtu' - the data space left */
579 /* IF: we are not sending upto and including the packet end
580 then align the next start on an eight byte boundary */
588 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
589 NETDEBUG(printk(KERN_INFO "IP: frag: no memory for new fragment!\n"));
595 * Set up data on packet
598 ip_copy_metadata(skb2, skb);
599 skb_reserve(skb2, ll_rs);
600 skb_put(skb2, len + hlen);
601 skb2->nh.raw = skb2->data;
602 skb2->h.raw = skb2->data + hlen;
605 * Charge the memory for the fragment to any owner
610 skb_set_owner_w(skb2, skb->sk);
613 * Copy the packet header into the new buffer.
616 memcpy(skb2->nh.raw, skb->data, hlen);
619 * Copy a block of the IP datagram.
621 if (skb_copy_bits(skb, ptr, skb2->h.raw, len))
626 * Fill in the new header fields.
629 iph->frag_off = htons((offset >> 3));
631 /* ANK: dirty, but effective trick. Upgrade options only if
632 * the segment to be fragmented was THE FIRST (otherwise,
633 * options are already fixed) and make it ONCE
634 * on the initial skb, so that all the following fragments
635 * will inherit fixed options.
638 ip_options_fragment(skb);
641 * Added AC : If we are fragmenting a fragment that's not the
642 * last fragment then keep MF on each bit
644 if (left > 0 || not_last_frag)
645 iph->frag_off |= htons(IP_MF);
650 * Put this fragment into the sending queue.
653 IP_INC_STATS(IPSTATS_MIB_FRAGCREATES);
655 iph->tot_len = htons(len + hlen);
664 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
669 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
674 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
676 struct iovec *iov = from;
678 if (skb->ip_summed == CHECKSUM_HW) {
679 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
682 unsigned int csum = 0;
683 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
685 skb->csum = csum_block_add(skb->csum, csum, odd);
690 static inline unsigned int
691 csum_page(struct page *page, int offset, int copy)
696 csum = csum_partial(kaddr + offset, copy, 0);
702 * ip_append_data() and ip_append_page() can make one large IP datagram
703 * from many pieces of data. Each pieces will be holded on the socket
704 * until ip_push_pending_frames() is called. Each piece can be a page
707 * Not only UDP, other transport protocols - e.g. raw sockets - can use
708 * this interface potentially.
710 * LATER: length must be adjusted by pad at tail, when it is required.
712 int ip_append_data(struct sock *sk,
713 int getfrag(void *from, char *to, int offset, int len,
714 int odd, struct sk_buff *skb),
715 void *from, int length, int transhdrlen,
716 struct ipcm_cookie *ipc, struct rtable *rt,
719 struct inet_sock *inet = inet_sk(sk);
722 struct ip_options *opt = NULL;
729 unsigned int maxfraglen, fragheaderlen;
730 int csummode = CHECKSUM_NONE;
735 if (skb_queue_empty(&sk->sk_write_queue)) {
741 if (inet->cork.opt == NULL) {
742 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
743 if (unlikely(inet->cork.opt == NULL))
746 memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
747 inet->cork.flags |= IPCORK_OPT;
748 inet->cork.addr = ipc->addr;
750 dst_hold(&rt->u.dst);
751 inet->cork.fragsize = mtu = dst_mtu(rt->u.dst.path);
753 inet->cork.length = 0;
754 sk->sk_sndmsg_page = NULL;
755 sk->sk_sndmsg_off = 0;
756 if ((exthdrlen = rt->u.dst.header_len) != 0) {
758 transhdrlen += exthdrlen;
762 if (inet->cork.flags & IPCORK_OPT)
763 opt = inet->cork.opt;
767 mtu = inet->cork.fragsize;
769 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
771 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
772 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
774 if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
775 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu-exthdrlen);
780 * transhdrlen > 0 means that this is the first fragment and we wish
781 * it won't be fragmented in the future.
784 length + fragheaderlen <= mtu &&
785 rt->u.dst.dev->features&(NETIF_F_IP_CSUM|NETIF_F_NO_CSUM|NETIF_F_HW_CSUM) &&
787 csummode = CHECKSUM_HW;
789 inet->cork.length += length;
791 /* So, what's going on in the loop below?
793 * We use calculated fragment length to generate chained skb,
794 * each of segments is IP fragment ready for sending to network after
795 * adding appropriate IP header.
798 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
802 /* Check if the remaining data fits into current packet. */
803 copy = mtu - skb->len;
805 copy = maxfraglen - skb->len;
808 unsigned int datalen;
809 unsigned int fraglen;
810 unsigned int fraggap;
811 unsigned int alloclen;
812 struct sk_buff *skb_prev;
816 fraggap = skb_prev->len - maxfraglen;
821 * If remaining data exceeds the mtu,
822 * we know we need more fragment(s).
824 datalen = length + fraggap;
825 if (datalen > mtu - fragheaderlen)
826 datalen = maxfraglen - fragheaderlen;
827 fraglen = datalen + fragheaderlen;
829 if ((flags & MSG_MORE) &&
830 !(rt->u.dst.dev->features&NETIF_F_SG))
833 alloclen = datalen + fragheaderlen;
835 /* The last fragment gets additional space at tail.
836 * Note, with MSG_MORE we overallocate on fragments,
837 * because we have no idea what fragment will be
840 if (datalen == length)
841 alloclen += rt->u.dst.trailer_len;
844 skb = sock_alloc_send_skb(sk,
845 alloclen + hh_len + 15,
846 (flags & MSG_DONTWAIT), &err);
849 if (atomic_read(&sk->sk_wmem_alloc) <=
851 skb = sock_wmalloc(sk,
852 alloclen + hh_len + 15, 1,
854 if (unlikely(skb == NULL))
861 * Fill in the control structures
863 skb->ip_summed = csummode;
865 skb_reserve(skb, hh_len);
868 * Find where to start putting bytes.
870 data = skb_put(skb, fraglen);
871 skb->nh.raw = data + exthdrlen;
872 data += fragheaderlen;
873 skb->h.raw = data + exthdrlen;
876 skb->csum = skb_copy_and_csum_bits(
877 skb_prev, maxfraglen,
878 data + transhdrlen, fraggap, 0);
879 skb_prev->csum = csum_sub(skb_prev->csum,
882 skb_trim(skb_prev, maxfraglen);
885 copy = datalen - transhdrlen - fraggap;
886 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
893 length -= datalen - fraggap;
896 csummode = CHECKSUM_NONE;
899 * Put the packet on the pending queue.
901 __skb_queue_tail(&sk->sk_write_queue, skb);
908 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
912 if (getfrag(from, skb_put(skb, copy),
913 offset, copy, off, skb) < 0) {
914 __skb_trim(skb, off);
919 int i = skb_shinfo(skb)->nr_frags;
920 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
921 struct page *page = sk->sk_sndmsg_page;
922 int off = sk->sk_sndmsg_off;
925 if (page && (left = PAGE_SIZE - off) > 0) {
928 if (page != frag->page) {
929 if (i == MAX_SKB_FRAGS) {
934 skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
935 frag = &skb_shinfo(skb)->frags[i];
937 } else if (i < MAX_SKB_FRAGS) {
938 if (copy > PAGE_SIZE)
940 page = alloc_pages(sk->sk_allocation, 0);
945 sk->sk_sndmsg_page = page;
946 sk->sk_sndmsg_off = 0;
948 skb_fill_page_desc(skb, i, page, 0, 0);
949 frag = &skb_shinfo(skb)->frags[i];
950 skb->truesize += PAGE_SIZE;
951 atomic_add(PAGE_SIZE, &sk->sk_wmem_alloc);
956 if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
960 sk->sk_sndmsg_off += copy;
963 skb->data_len += copy;
972 inet->cork.length -= length;
973 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
977 ssize_t ip_append_page(struct sock *sk, struct page *page,
978 int offset, size_t size, int flags)
980 struct inet_sock *inet = inet_sk(sk);
983 struct ip_options *opt = NULL;
988 unsigned int maxfraglen, fragheaderlen, fraggap;
996 if (skb_queue_empty(&sk->sk_write_queue))
1000 if (inet->cork.flags & IPCORK_OPT)
1001 opt = inet->cork.opt;
1003 if (!(rt->u.dst.dev->features&NETIF_F_SG))
1006 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1007 mtu = inet->cork.fragsize;
1009 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1010 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1012 if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1013 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu);
1017 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1020 inet->cork.length += size;
1025 /* Check if the remaining data fits into current packet. */
1026 len = mtu - skb->len;
1028 len = maxfraglen - skb->len;
1030 struct sk_buff *skb_prev;
1037 fraggap = skb_prev->len - maxfraglen;
1041 alloclen = fragheaderlen + hh_len + fraggap + 15;
1042 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1043 if (unlikely(!skb)) {
1049 * Fill in the control structures
1051 skb->ip_summed = CHECKSUM_NONE;
1053 skb_reserve(skb, hh_len);
1056 * Find where to start putting bytes.
1058 data = skb_put(skb, fragheaderlen + fraggap);
1059 skb->nh.iph = iph = (struct iphdr *)data;
1060 data += fragheaderlen;
1064 skb->csum = skb_copy_and_csum_bits(
1065 skb_prev, maxfraglen,
1067 skb_prev->csum = csum_sub(skb_prev->csum,
1069 skb_trim(skb_prev, maxfraglen);
1073 * Put the packet on the pending queue.
1075 __skb_queue_tail(&sk->sk_write_queue, skb);
1079 i = skb_shinfo(skb)->nr_frags;
1082 if (skb_can_coalesce(skb, i, page, offset)) {
1083 skb_shinfo(skb)->frags[i-1].size += len;
1084 } else if (i < MAX_SKB_FRAGS) {
1086 skb_fill_page_desc(skb, i, page, offset, len);
1092 if (skb->ip_summed == CHECKSUM_NONE) {
1094 csum = csum_page(page, offset, len);
1095 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1099 skb->data_len += len;
1106 inet->cork.length -= size;
1107 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1112 * Combined all pending IP fragments on the socket as one IP datagram
1113 * and push them out.
1115 int ip_push_pending_frames(struct sock *sk)
1117 struct sk_buff *skb, *tmp_skb;
1118 struct sk_buff **tail_skb;
1119 struct inet_sock *inet = inet_sk(sk);
1120 struct ip_options *opt = NULL;
1121 struct rtable *rt = inet->cork.rt;
1127 if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1129 tail_skb = &(skb_shinfo(skb)->frag_list);
1131 /* move skb->data to ip header from ext header */
1132 if (skb->data < skb->nh.raw)
1133 __skb_pull(skb, skb->nh.raw - skb->data);
1134 while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1135 __skb_pull(tmp_skb, skb->h.raw - skb->nh.raw);
1136 *tail_skb = tmp_skb;
1137 tail_skb = &(tmp_skb->next);
1138 skb->len += tmp_skb->len;
1139 skb->data_len += tmp_skb->len;
1140 skb->truesize += tmp_skb->truesize;
1141 __sock_put(tmp_skb->sk);
1142 tmp_skb->destructor = NULL;
1146 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1147 * to fragment the frame generated here. No matter, what transforms
1148 * how transforms change size of the packet, it will come out.
1150 if (inet->pmtudisc != IP_PMTUDISC_DO)
1153 /* DF bit is set when we want to see DF on outgoing frames.
1154 * If local_df is set too, we still allow to fragment this frame
1156 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1157 (skb->len <= dst_mtu(&rt->u.dst) &&
1158 ip_dont_fragment(sk, &rt->u.dst)))
1161 if (inet->cork.flags & IPCORK_OPT)
1162 opt = inet->cork.opt;
1164 if (rt->rt_type == RTN_MULTICAST)
1167 ttl = ip_select_ttl(inet, &rt->u.dst);
1169 iph = (struct iphdr *)skb->data;
1173 iph->ihl += opt->optlen>>2;
1174 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1176 iph->tos = inet->tos;
1177 iph->tot_len = htons(skb->len);
1180 __ip_select_ident(iph, &rt->u.dst, 0);
1182 iph->id = htons(inet->id++);
1185 iph->protocol = sk->sk_protocol;
1186 iph->saddr = rt->rt_src;
1187 iph->daddr = rt->rt_dst;
1190 skb->priority = sk->sk_priority;
1191 skb->dst = dst_clone(&rt->u.dst);
1193 /* Netfilter gets whole the not fragmented skb. */
1194 err = NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL,
1195 skb->dst->dev, dst_output);
1198 err = inet->recverr ? net_xmit_errno(err) : 0;
1204 inet->cork.flags &= ~IPCORK_OPT;
1205 if (inet->cork.opt) {
1206 kfree(inet->cork.opt);
1207 inet->cork.opt = NULL;
1209 if (inet->cork.rt) {
1210 ip_rt_put(inet->cork.rt);
1211 inet->cork.rt = NULL;
1216 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1221 * Throw away all pending data on the socket.
1223 void ip_flush_pending_frames(struct sock *sk)
1225 struct inet_sock *inet = inet_sk(sk);
1226 struct sk_buff *skb;
1228 while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1231 inet->cork.flags &= ~IPCORK_OPT;
1232 if (inet->cork.opt) {
1233 kfree(inet->cork.opt);
1234 inet->cork.opt = NULL;
1236 if (inet->cork.rt) {
1237 ip_rt_put(inet->cork.rt);
1238 inet->cork.rt = NULL;
1244 * Fetch data from kernel space and fill in checksum if needed.
1246 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1247 int len, int odd, struct sk_buff *skb)
1251 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1252 skb->csum = csum_block_add(skb->csum, csum, odd);
1257 * Generic function to send a packet as reply to another packet.
1258 * Used to send TCP resets so far. ICMP should use this function too.
1260 * Should run single threaded per socket because it uses the sock
1261 * structure to pass arguments.
1263 * LATER: switch from ip_build_xmit to ip_append_*
1265 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1268 struct inet_sock *inet = inet_sk(sk);
1270 struct ip_options opt;
1273 struct ipcm_cookie ipc;
1275 struct rtable *rt = (struct rtable*)skb->dst;
1277 if (ip_options_echo(&replyopts.opt, skb))
1280 daddr = ipc.addr = rt->rt_src;
1283 if (replyopts.opt.optlen) {
1284 ipc.opt = &replyopts.opt;
1287 daddr = replyopts.opt.faddr;
1291 struct flowi fl = { .nl_u = { .ip4_u =
1293 .saddr = rt->rt_spec_dst,
1294 .tos = RT_TOS(skb->nh.iph->tos) } },
1295 /* Not quite clean, but right. */
1297 { .sport = skb->h.th->dest,
1298 .dport = skb->h.th->source } },
1299 .proto = sk->sk_protocol };
1300 if (ip_route_output_key(&rt, &fl))
1304 /* And let IP do all the hard work.
1306 This chunk is not reenterable, hence spinlock.
1307 Note that it uses the fact, that this function is called
1308 with locally disabled BH and that sk cannot be already spinlocked.
1311 inet->tos = skb->nh.iph->tos;
1312 sk->sk_priority = skb->priority;
1313 sk->sk_protocol = skb->nh.iph->protocol;
1314 ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1315 &ipc, rt, MSG_DONTWAIT);
1316 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1317 if (arg->csumoffset >= 0)
1318 *((u16 *)skb->h.raw + arg->csumoffset) = csum_fold(csum_add(skb->csum, arg->csum));
1319 skb->ip_summed = CHECKSUM_NONE;
1320 ip_push_pending_frames(sk);
1329 * IP protocol layer initialiser
1332 static struct packet_type ip_packet_type = {
1333 .type = __constant_htons(ETH_P_IP),
1338 * IP registers the packet type and then calls the subprotocol initialisers
1341 void __init ip_init(void)
1343 dev_add_pack(&ip_packet_type);
1348 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1349 igmp_mc_proc_init();
1353 EXPORT_SYMBOL(ip_finish_output);
1354 EXPORT_SYMBOL(ip_fragment);
1355 EXPORT_SYMBOL(ip_generic_getfrag);
1356 EXPORT_SYMBOL(ip_queue_xmit);
1357 EXPORT_SYMBOL(ip_send_check);
1359 #ifdef CONFIG_SYSCTL
1360 EXPORT_SYMBOL(sysctl_ip_default_ttl);