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);
119 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
121 int ttl = inet->uc_ttl;
124 ttl = dst_metric(dst, RTAX_HOPLIMIT);
129 * Add an ip header to a skbuff and send it out.
132 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
133 u32 saddr, u32 daddr, struct ip_options *opt)
135 struct inet_sock *inet = inet_sk(sk);
136 struct rtable *rt = (struct rtable *)skb->dst;
139 /* Build the IP header. */
141 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr) + opt->optlen);
143 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr));
147 iph->tos = inet->tos;
148 if (ip_dont_fragment(sk, &rt->u.dst))
149 iph->frag_off = htons(IP_DF);
152 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
153 iph->daddr = rt->rt_dst;
154 iph->saddr = rt->rt_src;
155 iph->protocol = sk->sk_protocol;
156 iph->tot_len = htons(skb->len);
157 ip_select_ident(iph, &rt->u.dst, sk);
160 if (opt && opt->optlen) {
161 iph->ihl += opt->optlen>>2;
162 ip_options_build(skb, opt, daddr, rt, 0);
166 skb->priority = sk->sk_priority;
169 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
173 static inline int ip_finish_output2(struct sk_buff *skb)
175 struct dst_entry *dst = skb->dst;
176 struct hh_cache *hh = dst->hh;
177 struct net_device *dev = dst->dev;
178 int hh_len = LL_RESERVED_SPACE(dev);
180 /* Be paranoid, rather than too clever. */
181 if (unlikely(skb_headroom(skb) < hh_len && dev->hard_header)) {
182 struct sk_buff *skb2;
184 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
190 skb_set_owner_w(skb2, skb->sk);
195 #ifdef CONFIG_NETFILTER_DEBUG
196 nf_debug_ip_finish_output2(skb);
197 #endif /*CONFIG_NETFILTER_DEBUG*/
199 #ifdef CONFIG_BRIDGE_NETFILTER
200 /* bridge-netfilter defers calling some IP hooks to the bridge layer
201 * and still needs the conntrack reference.
203 if (skb->nf_bridge == NULL)
210 read_lock_bh(&hh->hh_lock);
211 hh_alen = HH_DATA_ALIGN(hh->hh_len);
212 memcpy(skb->data - hh_alen, hh->hh_data, hh_alen);
213 read_unlock_bh(&hh->hh_lock);
214 skb_push(skb, hh->hh_len);
215 return hh->hh_output(skb);
216 } else if (dst->neighbour)
217 return dst->neighbour->output(skb);
220 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
225 int ip_finish_output(struct sk_buff *skb)
227 struct net_device *dev = skb->dst->dev;
230 skb->protocol = htons(ETH_P_IP);
232 return NF_HOOK(PF_INET, NF_IP_POST_ROUTING, skb, NULL, dev,
236 int ip_mc_output(struct sk_buff *skb)
238 struct sock *sk = skb->sk;
239 struct rtable *rt = (struct rtable*)skb->dst;
240 struct net_device *dev = rt->u.dst.dev;
243 * If the indicated interface is up and running, send the packet.
245 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
248 skb->protocol = htons(ETH_P_IP);
251 * Multicasts are looped back for other local users
254 if (rt->rt_flags&RTCF_MULTICAST) {
255 if ((!sk || inet_sk(sk)->mc_loop)
256 #ifdef CONFIG_IP_MROUTE
257 /* Small optimization: do not loopback not local frames,
258 which returned after forwarding; they will be dropped
259 by ip_mr_input in any case.
260 Note, that local frames are looped back to be delivered
263 This check is duplicated in ip_mr_input at the moment.
265 && ((rt->rt_flags&RTCF_LOCAL) || !(IPCB(skb)->flags&IPSKB_FORWARDED))
268 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
270 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
272 ip_dev_loopback_xmit);
275 /* Multicasts with ttl 0 must not go beyond the host */
277 if (skb->nh.iph->ttl == 0) {
283 if (rt->rt_flags&RTCF_BROADCAST) {
284 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
286 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
287 newskb->dev, ip_dev_loopback_xmit);
290 if (skb->len > dst_mtu(&rt->u.dst))
291 return ip_fragment(skb, ip_finish_output);
293 return ip_finish_output(skb);
296 int ip_output(struct sk_buff *skb)
298 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
300 if (skb->len > dst_mtu(skb->dst) && !skb_shinfo(skb)->tso_size)
301 return ip_fragment(skb, ip_finish_output);
303 return ip_finish_output(skb);
306 int ip_queue_xmit(struct sk_buff *skb, int ipfragok)
308 struct sock *sk = skb->sk;
309 struct inet_sock *inet = inet_sk(sk);
310 struct ip_options *opt = inet->opt;
314 /* Skip all of this if the packet is already routed,
315 * f.e. by something like SCTP.
317 rt = (struct rtable *) skb->dst;
321 /* Make sure we can route this packet. */
322 rt = (struct rtable *)__sk_dst_check(sk, 0);
326 /* Use correct destination address if we have options. */
332 struct flowi fl = { .oif = sk->sk_bound_dev_if,
335 .saddr = inet->saddr,
336 .tos = RT_CONN_FLAGS(sk) } },
337 .proto = sk->sk_protocol,
339 { .sport = inet->sport,
340 .dport = inet->dport } } };
342 /* If this fails, retransmit mechanism of transport layer will
343 * keep trying until route appears or the connection times
346 if (ip_route_output_flow(&rt, &fl, sk, 0))
349 __sk_dst_set(sk, &rt->u.dst);
350 tcp_v4_setup_caps(sk, &rt->u.dst);
352 skb->dst = dst_clone(&rt->u.dst);
355 if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
358 /* OK, we know where to send it, allocate and build IP header. */
359 iph = (struct iphdr *) skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
360 *((__u16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
361 iph->tot_len = htons(skb->len);
362 if (ip_dont_fragment(sk, &rt->u.dst) && !ipfragok)
363 iph->frag_off = htons(IP_DF);
366 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
367 iph->protocol = sk->sk_protocol;
368 iph->saddr = rt->rt_src;
369 iph->daddr = rt->rt_dst;
371 /* Transport layer set skb->h.foo itself. */
373 if (opt && opt->optlen) {
374 iph->ihl += opt->optlen >> 2;
375 ip_options_build(skb, opt, inet->daddr, rt, 0);
378 ip_select_ident_more(iph, &rt->u.dst, sk, skb_shinfo(skb)->tso_segs);
380 /* Add an IP checksum. */
383 skb->priority = sk->sk_priority;
385 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
389 IP_INC_STATS(IPSTATS_MIB_OUTNOROUTES);
391 return -EHOSTUNREACH;
395 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
397 to->pkt_type = from->pkt_type;
398 to->priority = from->priority;
399 to->protocol = from->protocol;
400 to->security = from->security;
401 dst_release(to->dst);
402 to->dst = dst_clone(from->dst);
405 /* Copy the flags to each fragment. */
406 IPCB(to)->flags = IPCB(from)->flags;
408 #ifdef CONFIG_NET_SCHED
409 to->tc_index = from->tc_index;
411 #ifdef CONFIG_NETFILTER
412 to->nfmark = from->nfmark;
413 to->nfcache = from->nfcache;
414 /* Connection association is same as pre-frag packet */
415 nf_conntrack_put(to->nfct);
416 to->nfct = from->nfct;
417 nf_conntrack_get(to->nfct);
418 to->nfctinfo = from->nfctinfo;
419 #ifdef CONFIG_BRIDGE_NETFILTER
420 nf_bridge_put(to->nf_bridge);
421 to->nf_bridge = from->nf_bridge;
422 nf_bridge_get(to->nf_bridge);
424 #ifdef CONFIG_NETFILTER_DEBUG
425 to->nf_debug = from->nf_debug;
431 * This IP datagram is too large to be sent in one piece. Break it up into
432 * smaller pieces (each of size equal to IP header plus
433 * a block of the data of the original IP data part) that will yet fit in a
434 * single device frame, and queue such a frame for sending.
437 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*))
442 struct net_device *dev;
443 struct sk_buff *skb2;
444 unsigned int mtu, hlen, left, len, ll_rs;
447 struct rtable *rt = (struct rtable*)skb->dst;
453 * Point into the IP datagram header.
458 if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
459 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
460 htonl(dst_mtu(&rt->u.dst)));
466 * Setup starting values.
470 mtu = dst_mtu(&rt->u.dst) - hlen; /* Size of data space */
472 /* When frag_list is given, use it. First, check its validity:
473 * some transformers could create wrong frag_list or break existing
474 * one, it is not prohibited. In this case fall back to copying.
476 * LATER: this step can be merged to real generation of fragments,
477 * we can switch to copy when see the first bad fragment.
479 if (skb_shinfo(skb)->frag_list) {
480 struct sk_buff *frag;
481 int first_len = skb_pagelen(skb);
483 if (first_len - hlen > mtu ||
484 ((first_len - hlen) & 7) ||
485 (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
489 for (frag = skb_shinfo(skb)->frag_list; frag; frag = frag->next) {
490 /* Correct geometry. */
491 if (frag->len > mtu ||
492 ((frag->len & 7) && frag->next) ||
493 skb_headroom(frag) < hlen)
496 /* Partially cloned skb? */
497 if (skb_shared(frag))
504 frag->destructor = sock_wfree;
505 skb->truesize -= frag->truesize;
509 /* Everything is OK. Generate! */
513 frag = skb_shinfo(skb)->frag_list;
514 skb_shinfo(skb)->frag_list = NULL;
515 skb->data_len = first_len - skb_headlen(skb);
516 skb->len = first_len;
517 iph->tot_len = htons(first_len);
518 iph->frag_off = htons(IP_MF);
522 /* Prepare header of the next frame,
523 * before previous one went down. */
525 frag->ip_summed = CHECKSUM_NONE;
526 frag->h.raw = frag->data;
527 frag->nh.raw = __skb_push(frag, hlen);
528 memcpy(frag->nh.raw, iph, hlen);
530 iph->tot_len = htons(frag->len);
531 ip_copy_metadata(frag, skb);
533 ip_options_fragment(frag);
534 offset += skb->len - hlen;
535 iph->frag_off = htons(offset>>3);
536 if (frag->next != NULL)
537 iph->frag_off |= htons(IP_MF);
538 /* Ready, complete checksum */
553 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
562 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
567 left = skb->len - hlen; /* Space per frame */
568 ptr = raw + hlen; /* Where to start from */
570 #ifdef CONFIG_BRIDGE_NETFILTER
571 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
572 * we need to make room for the encapsulating header */
573 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, nf_bridge_pad(skb));
574 mtu -= nf_bridge_pad(skb);
576 ll_rs = LL_RESERVED_SPACE(rt->u.dst.dev);
579 * Fragment the datagram.
582 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
583 not_last_frag = iph->frag_off & htons(IP_MF);
586 * Keep copying data until we run out.
591 /* IF: it doesn't fit, use 'mtu' - the data space left */
594 /* IF: we are not sending upto and including the packet end
595 then align the next start on an eight byte boundary */
603 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
604 NETDEBUG(printk(KERN_INFO "IP: frag: no memory for new fragment!\n"));
610 * Set up data on packet
613 ip_copy_metadata(skb2, skb);
614 skb_reserve(skb2, ll_rs);
615 skb_put(skb2, len + hlen);
616 skb2->nh.raw = skb2->data;
617 skb2->h.raw = skb2->data + hlen;
620 * Charge the memory for the fragment to any owner
625 skb_set_owner_w(skb2, skb->sk);
628 * Copy the packet header into the new buffer.
631 memcpy(skb2->nh.raw, skb->data, hlen);
634 * Copy a block of the IP datagram.
636 if (skb_copy_bits(skb, ptr, skb2->h.raw, len))
641 * Fill in the new header fields.
644 iph->frag_off = htons((offset >> 3));
646 /* ANK: dirty, but effective trick. Upgrade options only if
647 * the segment to be fragmented was THE FIRST (otherwise,
648 * options are already fixed) and make it ONCE
649 * on the initial skb, so that all the following fragments
650 * will inherit fixed options.
653 ip_options_fragment(skb);
656 * Added AC : If we are fragmenting a fragment that's not the
657 * last fragment then keep MF on each bit
659 if (left > 0 || not_last_frag)
660 iph->frag_off |= htons(IP_MF);
665 * Put this fragment into the sending queue.
668 IP_INC_STATS(IPSTATS_MIB_FRAGCREATES);
670 iph->tot_len = htons(len + hlen);
679 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
684 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
689 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
691 struct iovec *iov = from;
693 if (skb->ip_summed == CHECKSUM_HW) {
694 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
697 unsigned int csum = 0;
698 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
700 skb->csum = csum_block_add(skb->csum, csum, odd);
705 static inline unsigned int
706 csum_page(struct page *page, int offset, int copy)
711 csum = csum_partial(kaddr + offset, copy, 0);
717 * ip_append_data() and ip_append_page() can make one large IP datagram
718 * from many pieces of data. Each pieces will be holded on the socket
719 * until ip_push_pending_frames() is called. Each piece can be a page
722 * Not only UDP, other transport protocols - e.g. raw sockets - can use
723 * this interface potentially.
725 * LATER: length must be adjusted by pad at tail, when it is required.
727 int ip_append_data(struct sock *sk,
728 int getfrag(void *from, char *to, int offset, int len,
729 int odd, struct sk_buff *skb),
730 void *from, int length, int transhdrlen,
731 struct ipcm_cookie *ipc, struct rtable *rt,
734 struct inet_sock *inet = inet_sk(sk);
737 struct ip_options *opt = NULL;
744 unsigned int maxfraglen, fragheaderlen;
745 int csummode = CHECKSUM_NONE;
750 if (skb_queue_empty(&sk->sk_write_queue)) {
756 if (inet->cork.opt == NULL) {
757 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
758 if (unlikely(inet->cork.opt == NULL))
761 memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
762 inet->cork.flags |= IPCORK_OPT;
763 inet->cork.addr = ipc->addr;
765 dst_hold(&rt->u.dst);
766 inet->cork.fragsize = mtu = dst_mtu(rt->u.dst.path);
768 inet->cork.length = 0;
769 sk->sk_sndmsg_page = NULL;
770 sk->sk_sndmsg_off = 0;
771 if ((exthdrlen = rt->u.dst.header_len) != 0) {
773 transhdrlen += exthdrlen;
777 if (inet->cork.flags & IPCORK_OPT)
778 opt = inet->cork.opt;
782 mtu = inet->cork.fragsize;
784 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
786 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
787 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
789 if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
790 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu-exthdrlen);
795 * transhdrlen > 0 means that this is the first fragment and we wish
796 * it won't be fragmented in the future.
799 length + fragheaderlen <= mtu &&
800 rt->u.dst.dev->features&(NETIF_F_IP_CSUM|NETIF_F_NO_CSUM|NETIF_F_HW_CSUM) &&
802 csummode = CHECKSUM_HW;
804 inet->cork.length += length;
806 /* So, what's going on in the loop below?
808 * We use calculated fragment length to generate chained skb,
809 * each of segments is IP fragment ready for sending to network after
810 * adding appropriate IP header.
813 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
817 /* Check if the remaining data fits into current packet. */
818 copy = mtu - skb->len;
820 copy = maxfraglen - skb->len;
823 unsigned int datalen;
824 unsigned int fraglen;
825 unsigned int fraggap;
826 unsigned int alloclen;
827 struct sk_buff *skb_prev;
831 fraggap = skb_prev->len - maxfraglen;
836 * If remaining data exceeds the mtu,
837 * we know we need more fragment(s).
839 datalen = length + fraggap;
840 if (datalen > mtu - fragheaderlen)
841 datalen = maxfraglen - fragheaderlen;
842 fraglen = datalen + fragheaderlen;
844 if ((flags & MSG_MORE) &&
845 !(rt->u.dst.dev->features&NETIF_F_SG))
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;
891 skb->csum = skb_copy_and_csum_bits(
892 skb_prev, maxfraglen,
893 data + transhdrlen, fraggap, 0);
894 skb_prev->csum = csum_sub(skb_prev->csum,
897 skb_trim(skb_prev, maxfraglen);
900 copy = datalen - transhdrlen - fraggap;
901 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
908 length -= datalen - fraggap;
911 csummode = CHECKSUM_NONE;
914 * Put the packet on the pending queue.
916 __skb_queue_tail(&sk->sk_write_queue, skb);
923 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
927 if (getfrag(from, skb_put(skb, copy),
928 offset, copy, off, skb) < 0) {
929 __skb_trim(skb, off);
934 int i = skb_shinfo(skb)->nr_frags;
935 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
936 struct page *page = sk->sk_sndmsg_page;
937 int off = sk->sk_sndmsg_off;
940 if (page && (left = PAGE_SIZE - off) > 0) {
943 if (page != frag->page) {
944 if (i == MAX_SKB_FRAGS) {
949 skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
950 frag = &skb_shinfo(skb)->frags[i];
952 } else if (i < MAX_SKB_FRAGS) {
953 if (copy > PAGE_SIZE)
955 page = alloc_pages(sk->sk_allocation, 0);
960 sk->sk_sndmsg_page = page;
961 sk->sk_sndmsg_off = 0;
963 skb_fill_page_desc(skb, i, page, 0, 0);
964 frag = &skb_shinfo(skb)->frags[i];
965 skb->truesize += PAGE_SIZE;
966 atomic_add(PAGE_SIZE, &sk->sk_wmem_alloc);
971 if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
975 sk->sk_sndmsg_off += copy;
978 skb->data_len += copy;
987 inet->cork.length -= length;
988 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
992 ssize_t ip_append_page(struct sock *sk, struct page *page,
993 int offset, size_t size, int flags)
995 struct inet_sock *inet = inet_sk(sk);
998 struct ip_options *opt = NULL;
1003 unsigned int maxfraglen, fragheaderlen, fraggap;
1008 if (flags&MSG_PROBE)
1011 if (skb_queue_empty(&sk->sk_write_queue))
1015 if (inet->cork.flags & IPCORK_OPT)
1016 opt = inet->cork.opt;
1018 if (!(rt->u.dst.dev->features&NETIF_F_SG))
1021 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1022 mtu = inet->cork.fragsize;
1024 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1025 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1027 if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1028 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu);
1032 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1035 inet->cork.length += size;
1040 /* Check if the remaining data fits into current packet. */
1041 len = mtu - skb->len;
1043 len = maxfraglen - skb->len;
1045 struct sk_buff *skb_prev;
1052 fraggap = skb_prev->len - maxfraglen;
1056 alloclen = fragheaderlen + hh_len + fraggap + 15;
1057 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1058 if (unlikely(!skb)) {
1064 * Fill in the control structures
1066 skb->ip_summed = CHECKSUM_NONE;
1068 skb_reserve(skb, hh_len);
1071 * Find where to start putting bytes.
1073 data = skb_put(skb, fragheaderlen + fraggap);
1074 skb->nh.iph = iph = (struct iphdr *)data;
1075 data += fragheaderlen;
1079 skb->csum = skb_copy_and_csum_bits(
1080 skb_prev, maxfraglen,
1082 skb_prev->csum = csum_sub(skb_prev->csum,
1084 skb_trim(skb_prev, maxfraglen);
1088 * Put the packet on the pending queue.
1090 __skb_queue_tail(&sk->sk_write_queue, skb);
1094 i = skb_shinfo(skb)->nr_frags;
1097 if (skb_can_coalesce(skb, i, page, offset)) {
1098 skb_shinfo(skb)->frags[i-1].size += len;
1099 } else if (i < MAX_SKB_FRAGS) {
1101 skb_fill_page_desc(skb, i, page, offset, len);
1107 if (skb->ip_summed == CHECKSUM_NONE) {
1109 csum = csum_page(page, offset, len);
1110 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1114 skb->data_len += len;
1121 inet->cork.length -= size;
1122 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1127 * Combined all pending IP fragments on the socket as one IP datagram
1128 * and push them out.
1130 int ip_push_pending_frames(struct sock *sk)
1132 struct sk_buff *skb, *tmp_skb;
1133 struct sk_buff **tail_skb;
1134 struct inet_sock *inet = inet_sk(sk);
1135 struct ip_options *opt = NULL;
1136 struct rtable *rt = inet->cork.rt;
1142 if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1144 tail_skb = &(skb_shinfo(skb)->frag_list);
1146 /* move skb->data to ip header from ext header */
1147 if (skb->data < skb->nh.raw)
1148 __skb_pull(skb, skb->nh.raw - skb->data);
1149 while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1150 __skb_pull(tmp_skb, skb->h.raw - skb->nh.raw);
1151 *tail_skb = tmp_skb;
1152 tail_skb = &(tmp_skb->next);
1153 skb->len += tmp_skb->len;
1154 skb->data_len += tmp_skb->len;
1155 skb->truesize += tmp_skb->truesize;
1156 __sock_put(tmp_skb->sk);
1157 tmp_skb->destructor = NULL;
1161 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1162 * to fragment the frame generated here. No matter, what transforms
1163 * how transforms change size of the packet, it will come out.
1165 if (inet->pmtudisc != IP_PMTUDISC_DO)
1168 /* DF bit is set when we want to see DF on outgoing frames.
1169 * If local_df is set too, we still allow to fragment this frame
1171 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1172 (skb->len <= dst_mtu(&rt->u.dst) &&
1173 ip_dont_fragment(sk, &rt->u.dst)))
1176 if (inet->cork.flags & IPCORK_OPT)
1177 opt = inet->cork.opt;
1179 if (rt->rt_type == RTN_MULTICAST)
1182 ttl = ip_select_ttl(inet, &rt->u.dst);
1184 iph = (struct iphdr *)skb->data;
1188 iph->ihl += opt->optlen>>2;
1189 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1191 iph->tos = inet->tos;
1192 iph->tot_len = htons(skb->len);
1195 __ip_select_ident(iph, &rt->u.dst, 0);
1197 iph->id = htons(inet->id++);
1200 iph->protocol = sk->sk_protocol;
1201 iph->saddr = rt->rt_src;
1202 iph->daddr = rt->rt_dst;
1205 skb->priority = sk->sk_priority;
1206 skb->dst = dst_clone(&rt->u.dst);
1208 /* Netfilter gets whole the not fragmented skb. */
1209 err = NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL,
1210 skb->dst->dev, dst_output);
1213 err = inet->recverr ? net_xmit_errno(err) : 0;
1219 inet->cork.flags &= ~IPCORK_OPT;
1220 if (inet->cork.opt) {
1221 kfree(inet->cork.opt);
1222 inet->cork.opt = NULL;
1224 if (inet->cork.rt) {
1225 ip_rt_put(inet->cork.rt);
1226 inet->cork.rt = NULL;
1231 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1236 * Throw away all pending data on the socket.
1238 void ip_flush_pending_frames(struct sock *sk)
1240 struct inet_sock *inet = inet_sk(sk);
1241 struct sk_buff *skb;
1243 while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1246 inet->cork.flags &= ~IPCORK_OPT;
1247 if (inet->cork.opt) {
1248 kfree(inet->cork.opt);
1249 inet->cork.opt = NULL;
1251 if (inet->cork.rt) {
1252 ip_rt_put(inet->cork.rt);
1253 inet->cork.rt = NULL;
1259 * Fetch data from kernel space and fill in checksum if needed.
1261 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1262 int len, int odd, struct sk_buff *skb)
1266 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1267 skb->csum = csum_block_add(skb->csum, csum, odd);
1272 * Generic function to send a packet as reply to another packet.
1273 * Used to send TCP resets so far. ICMP should use this function too.
1275 * Should run single threaded per socket because it uses the sock
1276 * structure to pass arguments.
1278 * LATER: switch from ip_build_xmit to ip_append_*
1280 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1283 struct inet_sock *inet = inet_sk(sk);
1285 struct ip_options opt;
1288 struct ipcm_cookie ipc;
1290 struct rtable *rt = (struct rtable*)skb->dst;
1292 if (ip_options_echo(&replyopts.opt, skb))
1295 daddr = ipc.addr = rt->rt_src;
1298 if (replyopts.opt.optlen) {
1299 ipc.opt = &replyopts.opt;
1302 daddr = replyopts.opt.faddr;
1306 struct flowi fl = { .nl_u = { .ip4_u =
1308 .saddr = rt->rt_spec_dst,
1309 .tos = RT_TOS(skb->nh.iph->tos) } },
1310 /* Not quite clean, but right. */
1312 { .sport = skb->h.th->dest,
1313 .dport = skb->h.th->source } },
1314 .proto = sk->sk_protocol };
1315 if (ip_route_output_key(&rt, &fl))
1319 /* And let IP do all the hard work.
1321 This chunk is not reenterable, hence spinlock.
1322 Note that it uses the fact, that this function is called
1323 with locally disabled BH and that sk cannot be already spinlocked.
1326 inet->tos = skb->nh.iph->tos;
1327 sk->sk_priority = skb->priority;
1328 sk->sk_protocol = skb->nh.iph->protocol;
1329 ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1330 &ipc, rt, MSG_DONTWAIT);
1331 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1332 if (arg->csumoffset >= 0)
1333 *((u16 *)skb->h.raw + arg->csumoffset) = csum_fold(csum_add(skb->csum, arg->csum));
1334 skb->ip_summed = CHECKSUM_NONE;
1335 ip_push_pending_frames(sk);
1344 * IP protocol layer initialiser
1347 static struct packet_type ip_packet_type = {
1348 .type = __constant_htons(ETH_P_IP),
1353 * IP registers the packet type and then calls the subprotocol initialisers
1356 void __init ip_init(void)
1358 dev_add_pack(&ip_packet_type);
1363 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1364 igmp_mc_proc_init();
1368 EXPORT_SYMBOL(ip_finish_output);
1369 EXPORT_SYMBOL(ip_fragment);
1370 EXPORT_SYMBOL(ip_generic_getfrag);
1371 EXPORT_SYMBOL(ip_queue_xmit);
1372 EXPORT_SYMBOL(ip_send_check);
1374 #ifdef CONFIG_SYSCTL
1375 EXPORT_SYMBOL(sysctl_ip_default_ttl);