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>
53 #include <linux/string.h>
54 #include <linux/errno.h>
55 #include <linux/highmem.h>
57 #include <linux/socket.h>
58 #include <linux/sockios.h>
60 #include <linux/inet.h>
61 #include <linux/netdevice.h>
62 #include <linux/etherdevice.h>
63 #include <linux/proc_fs.h>
64 #include <linux/stat.h>
65 #include <linux/init.h>
69 #include <net/protocol.h>
70 #include <net/route.h>
72 #include <linux/skbuff.h>
76 #include <net/checksum.h>
77 #include <net/inetpeer.h>
78 #include <net/checksum.h>
79 #include <linux/igmp.h>
80 #include <linux/netfilter_ipv4.h>
81 #include <linux/netfilter_bridge.h>
82 #include <linux/mroute.h>
83 #include <linux/netlink.h>
84 #include <linux/tcp.h>
86 int sysctl_ip_default_ttl __read_mostly = IPDEFTTL;
88 /* Generate a checksum for an outgoing IP datagram. */
89 __inline__ void ip_send_check(struct iphdr *iph)
92 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
95 /* dev_loopback_xmit for use with netfilter. */
96 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
98 skb_reset_mac_header(newskb);
99 __skb_pull(newskb, skb_network_offset(newskb));
100 newskb->pkt_type = PACKET_LOOPBACK;
101 newskb->ip_summed = CHECKSUM_UNNECESSARY;
102 BUG_TRAP(newskb->dst);
107 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
109 int ttl = inet->uc_ttl;
112 ttl = dst_metric(dst, RTAX_HOPLIMIT);
117 * Add an ip header to a skbuff and send it out.
120 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
121 __be32 saddr, __be32 daddr, struct ip_options *opt)
123 struct inet_sock *inet = inet_sk(sk);
124 struct rtable *rt = (struct rtable *)skb->dst;
127 /* Build the IP header. */
128 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
129 skb_reset_network_header(skb);
133 iph->tos = inet->tos;
134 if (ip_dont_fragment(sk, &rt->u.dst))
135 iph->frag_off = htons(IP_DF);
138 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
139 iph->daddr = rt->rt_dst;
140 iph->saddr = rt->rt_src;
141 iph->protocol = sk->sk_protocol;
142 iph->tot_len = htons(skb->len);
143 ip_select_ident(iph, &rt->u.dst, sk);
145 if (opt && opt->optlen) {
146 iph->ihl += opt->optlen>>2;
147 ip_options_build(skb, opt, daddr, rt, 0);
151 skb->priority = sk->sk_priority;
154 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
158 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
160 static inline int ip_finish_output2(struct sk_buff *skb)
162 struct dst_entry *dst = skb->dst;
163 struct net_device *dev = dst->dev;
164 int hh_len = LL_RESERVED_SPACE(dev);
166 /* Be paranoid, rather than too clever. */
167 if (unlikely(skb_headroom(skb) < hh_len && dev->hard_header)) {
168 struct sk_buff *skb2;
170 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
176 skb_set_owner_w(skb2, skb->sk);
182 return neigh_hh_output(dst->hh, skb);
183 else if (dst->neighbour)
184 return dst->neighbour->output(skb);
187 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
192 static inline int ip_finish_output(struct sk_buff *skb)
194 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
195 /* Policy lookup after SNAT yielded a new policy */
196 if (skb->dst->xfrm != NULL) {
197 IPCB(skb)->flags |= IPSKB_REROUTED;
198 return dst_output(skb);
201 if (skb->len > dst_mtu(skb->dst) && !skb_is_gso(skb))
202 return ip_fragment(skb, ip_finish_output2);
204 return ip_finish_output2(skb);
207 int ip_mc_output(struct sk_buff *skb)
209 struct sock *sk = skb->sk;
210 struct rtable *rt = (struct rtable*)skb->dst;
211 struct net_device *dev = rt->u.dst.dev;
214 * If the indicated interface is up and running, send the packet.
216 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
219 skb->protocol = htons(ETH_P_IP);
222 * Multicasts are looped back for other local users
225 if (rt->rt_flags&RTCF_MULTICAST) {
226 if ((!sk || inet_sk(sk)->mc_loop)
227 #ifdef CONFIG_IP_MROUTE
228 /* Small optimization: do not loopback not local frames,
229 which returned after forwarding; they will be dropped
230 by ip_mr_input in any case.
231 Note, that local frames are looped back to be delivered
234 This check is duplicated in ip_mr_input at the moment.
236 && ((rt->rt_flags&RTCF_LOCAL) || !(IPCB(skb)->flags&IPSKB_FORWARDED))
239 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
241 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
243 ip_dev_loopback_xmit);
246 /* Multicasts with ttl 0 must not go beyond the host */
248 if (ip_hdr(skb)->ttl == 0) {
254 if (rt->rt_flags&RTCF_BROADCAST) {
255 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
257 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
258 newskb->dev, ip_dev_loopback_xmit);
261 return NF_HOOK_COND(PF_INET, NF_IP_POST_ROUTING, skb, NULL, skb->dev,
263 !(IPCB(skb)->flags & IPSKB_REROUTED));
266 int ip_output(struct sk_buff *skb)
268 struct net_device *dev = skb->dst->dev;
270 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
273 skb->protocol = htons(ETH_P_IP);
275 return NF_HOOK_COND(PF_INET, NF_IP_POST_ROUTING, skb, NULL, dev,
277 !(IPCB(skb)->flags & IPSKB_REROUTED));
280 int ip_queue_xmit(struct sk_buff *skb, int ipfragok)
282 struct sock *sk = skb->sk;
283 struct inet_sock *inet = inet_sk(sk);
284 struct ip_options *opt = inet->opt;
288 /* Skip all of this if the packet is already routed,
289 * f.e. by something like SCTP.
291 rt = (struct rtable *) skb->dst;
295 /* Make sure we can route this packet. */
296 rt = (struct rtable *)__sk_dst_check(sk, 0);
300 /* Use correct destination address if we have options. */
306 struct flowi fl = { .oif = sk->sk_bound_dev_if,
309 .saddr = inet->saddr,
310 .tos = RT_CONN_FLAGS(sk) } },
311 .proto = sk->sk_protocol,
313 { .sport = inet->sport,
314 .dport = inet->dport } } };
316 /* If this fails, retransmit mechanism of transport layer will
317 * keep trying until route appears or the connection times
320 security_sk_classify_flow(sk, &fl);
321 if (ip_route_output_flow(&rt, &fl, sk, 0))
324 sk_setup_caps(sk, &rt->u.dst);
326 skb->dst = dst_clone(&rt->u.dst);
329 if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
332 /* OK, we know where to send it, allocate and build IP header. */
333 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
334 skb_reset_network_header(skb);
336 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
337 iph->tot_len = htons(skb->len);
338 if (ip_dont_fragment(sk, &rt->u.dst) && !ipfragok)
339 iph->frag_off = htons(IP_DF);
342 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
343 iph->protocol = sk->sk_protocol;
344 iph->saddr = rt->rt_src;
345 iph->daddr = rt->rt_dst;
346 /* Transport layer set skb->h.foo itself. */
348 if (opt && opt->optlen) {
349 iph->ihl += opt->optlen >> 2;
350 ip_options_build(skb, opt, inet->daddr, rt, 0);
353 ip_select_ident_more(iph, &rt->u.dst, sk,
354 (skb_shinfo(skb)->gso_segs ?: 1) - 1);
356 /* Add an IP checksum. */
359 skb->priority = sk->sk_priority;
361 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
365 IP_INC_STATS(IPSTATS_MIB_OUTNOROUTES);
367 return -EHOSTUNREACH;
371 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
373 to->pkt_type = from->pkt_type;
374 to->priority = from->priority;
375 to->protocol = from->protocol;
376 dst_release(to->dst);
377 to->dst = dst_clone(from->dst);
379 to->mark = from->mark;
381 /* Copy the flags to each fragment. */
382 IPCB(to)->flags = IPCB(from)->flags;
384 #ifdef CONFIG_NET_SCHED
385 to->tc_index = from->tc_index;
387 #ifdef CONFIG_NETFILTER
388 /* Connection association is same as pre-frag packet */
389 nf_conntrack_put(to->nfct);
390 to->nfct = from->nfct;
391 nf_conntrack_get(to->nfct);
392 to->nfctinfo = from->nfctinfo;
393 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
394 to->ipvs_property = from->ipvs_property;
396 #ifdef CONFIG_BRIDGE_NETFILTER
397 nf_bridge_put(to->nf_bridge);
398 to->nf_bridge = from->nf_bridge;
399 nf_bridge_get(to->nf_bridge);
402 skb_copy_secmark(to, from);
406 * This IP datagram is too large to be sent in one piece. Break it up into
407 * smaller pieces (each of size equal to IP header plus
408 * a block of the data of the original IP data part) that will yet fit in a
409 * single device frame, and queue such a frame for sending.
412 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*))
417 struct net_device *dev;
418 struct sk_buff *skb2;
419 unsigned int mtu, hlen, left, len, ll_rs, pad;
421 __be16 not_last_frag;
422 struct rtable *rt = (struct rtable*)skb->dst;
428 * Point into the IP datagram header.
433 if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
434 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
435 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
436 htonl(dst_mtu(&rt->u.dst)));
442 * Setup starting values.
446 mtu = dst_mtu(&rt->u.dst) - hlen; /* Size of data space */
447 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
449 /* When frag_list is given, use it. First, check its validity:
450 * some transformers could create wrong frag_list or break existing
451 * one, it is not prohibited. In this case fall back to copying.
453 * LATER: this step can be merged to real generation of fragments,
454 * we can switch to copy when see the first bad fragment.
456 if (skb_shinfo(skb)->frag_list) {
457 struct sk_buff *frag;
458 int first_len = skb_pagelen(skb);
460 if (first_len - hlen > mtu ||
461 ((first_len - hlen) & 7) ||
462 (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
466 for (frag = skb_shinfo(skb)->frag_list; frag; frag = frag->next) {
467 /* Correct geometry. */
468 if (frag->len > mtu ||
469 ((frag->len & 7) && frag->next) ||
470 skb_headroom(frag) < hlen)
473 /* Partially cloned skb? */
474 if (skb_shared(frag))
481 frag->destructor = sock_wfree;
482 skb->truesize -= frag->truesize;
486 /* Everything is OK. Generate! */
490 frag = skb_shinfo(skb)->frag_list;
491 skb_shinfo(skb)->frag_list = NULL;
492 skb->data_len = first_len - skb_headlen(skb);
493 skb->len = first_len;
494 iph->tot_len = htons(first_len);
495 iph->frag_off = htons(IP_MF);
499 /* Prepare header of the next frame,
500 * before previous one went down. */
502 frag->ip_summed = CHECKSUM_NONE;
503 skb_reset_transport_header(frag);
504 __skb_push(frag, hlen);
505 skb_reset_network_header(frag);
506 memcpy(skb_network_header(frag), iph, hlen);
508 iph->tot_len = htons(frag->len);
509 ip_copy_metadata(frag, skb);
511 ip_options_fragment(frag);
512 offset += skb->len - hlen;
513 iph->frag_off = htons(offset>>3);
514 if (frag->next != NULL)
515 iph->frag_off |= htons(IP_MF);
516 /* Ready, complete checksum */
523 IP_INC_STATS(IPSTATS_MIB_FRAGCREATES);
533 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
542 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
547 left = skb->len - hlen; /* Space per frame */
548 ptr = raw + hlen; /* Where to start from */
550 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
551 * we need to make room for the encapsulating header
553 pad = nf_bridge_pad(skb);
554 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, pad);
558 * Fragment the datagram.
561 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
562 not_last_frag = iph->frag_off & htons(IP_MF);
565 * Keep copying data until we run out.
570 /* IF: it doesn't fit, use 'mtu' - the data space left */
573 /* IF: we are not sending upto and including the packet end
574 then align the next start on an eight byte boundary */
582 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
583 NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
589 * Set up data on packet
592 ip_copy_metadata(skb2, skb);
593 skb_reserve(skb2, ll_rs);
594 skb_put(skb2, len + hlen);
595 skb_reset_network_header(skb2);
596 skb2->h.raw = skb2->nh.raw + hlen;
599 * Charge the memory for the fragment to any owner
604 skb_set_owner_w(skb2, skb->sk);
607 * Copy the packet header into the new buffer.
610 memcpy(skb_network_header(skb2), skb->data, hlen);
613 * Copy a block of the IP datagram.
615 if (skb_copy_bits(skb, ptr, skb2->h.raw, len))
620 * Fill in the new header fields.
623 iph->frag_off = htons((offset >> 3));
625 /* ANK: dirty, but effective trick. Upgrade options only if
626 * the segment to be fragmented was THE FIRST (otherwise,
627 * options are already fixed) and make it ONCE
628 * on the initial skb, so that all the following fragments
629 * will inherit fixed options.
632 ip_options_fragment(skb);
635 * Added AC : If we are fragmenting a fragment that's not the
636 * last fragment then keep MF on each bit
638 if (left > 0 || not_last_frag)
639 iph->frag_off |= htons(IP_MF);
644 * Put this fragment into the sending queue.
646 iph->tot_len = htons(len + hlen);
654 IP_INC_STATS(IPSTATS_MIB_FRAGCREATES);
657 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
662 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
666 EXPORT_SYMBOL(ip_fragment);
669 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
671 struct iovec *iov = from;
673 if (skb->ip_summed == CHECKSUM_PARTIAL) {
674 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
678 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
680 skb->csum = csum_block_add(skb->csum, csum, odd);
686 csum_page(struct page *page, int offset, int copy)
691 csum = csum_partial(kaddr + offset, copy, 0);
696 static inline int ip_ufo_append_data(struct sock *sk,
697 int getfrag(void *from, char *to, int offset, int len,
698 int odd, struct sk_buff *skb),
699 void *from, int length, int hh_len, int fragheaderlen,
700 int transhdrlen, int mtu,unsigned int flags)
705 /* There is support for UDP fragmentation offload by network
706 * device, so create one single skb packet containing complete
709 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL) {
710 skb = sock_alloc_send_skb(sk,
711 hh_len + fragheaderlen + transhdrlen + 20,
712 (flags & MSG_DONTWAIT), &err);
717 /* reserve space for Hardware header */
718 skb_reserve(skb, hh_len);
720 /* create space for UDP/IP header */
721 skb_put(skb,fragheaderlen + transhdrlen);
723 /* initialize network header pointer */
724 skb_reset_network_header(skb);
726 /* initialize protocol header pointer */
727 skb->h.raw = skb->nh.raw + fragheaderlen;
729 skb->ip_summed = CHECKSUM_PARTIAL;
731 sk->sk_sndmsg_off = 0;
734 err = skb_append_datato_frags(sk,skb, getfrag, from,
735 (length - transhdrlen));
737 /* specify the length of each IP datagram fragment*/
738 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
739 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
740 __skb_queue_tail(&sk->sk_write_queue, skb);
744 /* There is not enough support do UFO ,
745 * so follow normal path
752 * ip_append_data() and ip_append_page() can make one large IP datagram
753 * from many pieces of data. Each pieces will be holded on the socket
754 * until ip_push_pending_frames() is called. Each piece can be a page
757 * Not only UDP, other transport protocols - e.g. raw sockets - can use
758 * this interface potentially.
760 * LATER: length must be adjusted by pad at tail, when it is required.
762 int ip_append_data(struct sock *sk,
763 int getfrag(void *from, char *to, int offset, int len,
764 int odd, struct sk_buff *skb),
765 void *from, int length, int transhdrlen,
766 struct ipcm_cookie *ipc, struct rtable *rt,
769 struct inet_sock *inet = inet_sk(sk);
772 struct ip_options *opt = NULL;
779 unsigned int maxfraglen, fragheaderlen;
780 int csummode = CHECKSUM_NONE;
785 if (skb_queue_empty(&sk->sk_write_queue)) {
791 if (inet->cork.opt == NULL) {
792 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
793 if (unlikely(inet->cork.opt == NULL))
796 memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
797 inet->cork.flags |= IPCORK_OPT;
798 inet->cork.addr = ipc->addr;
800 dst_hold(&rt->u.dst);
801 inet->cork.fragsize = mtu = dst_mtu(rt->u.dst.path);
803 inet->cork.length = 0;
804 sk->sk_sndmsg_page = NULL;
805 sk->sk_sndmsg_off = 0;
806 if ((exthdrlen = rt->u.dst.header_len) != 0) {
808 transhdrlen += exthdrlen;
812 if (inet->cork.flags & IPCORK_OPT)
813 opt = inet->cork.opt;
817 mtu = inet->cork.fragsize;
819 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
821 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
822 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
824 if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
825 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu-exthdrlen);
830 * transhdrlen > 0 means that this is the first fragment and we wish
831 * it won't be fragmented in the future.
834 length + fragheaderlen <= mtu &&
835 rt->u.dst.dev->features & NETIF_F_ALL_CSUM &&
837 csummode = CHECKSUM_PARTIAL;
839 inet->cork.length += length;
840 if (((length > mtu) && (sk->sk_protocol == IPPROTO_UDP)) &&
841 (rt->u.dst.dev->features & NETIF_F_UFO)) {
843 err = ip_ufo_append_data(sk, getfrag, from, length, hh_len,
844 fragheaderlen, transhdrlen, mtu,
851 /* So, what's going on in the loop below?
853 * We use calculated fragment length to generate chained skb,
854 * each of segments is IP fragment ready for sending to network after
855 * adding appropriate IP header.
858 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
862 /* Check if the remaining data fits into current packet. */
863 copy = mtu - skb->len;
865 copy = maxfraglen - skb->len;
868 unsigned int datalen;
869 unsigned int fraglen;
870 unsigned int fraggap;
871 unsigned int alloclen;
872 struct sk_buff *skb_prev;
876 fraggap = skb_prev->len - maxfraglen;
881 * If remaining data exceeds the mtu,
882 * we know we need more fragment(s).
884 datalen = length + fraggap;
885 if (datalen > mtu - fragheaderlen)
886 datalen = maxfraglen - fragheaderlen;
887 fraglen = datalen + fragheaderlen;
889 if ((flags & MSG_MORE) &&
890 !(rt->u.dst.dev->features&NETIF_F_SG))
893 alloclen = datalen + fragheaderlen;
895 /* The last fragment gets additional space at tail.
896 * Note, with MSG_MORE we overallocate on fragments,
897 * because we have no idea what fragment will be
900 if (datalen == length + fraggap)
901 alloclen += rt->u.dst.trailer_len;
904 skb = sock_alloc_send_skb(sk,
905 alloclen + hh_len + 15,
906 (flags & MSG_DONTWAIT), &err);
909 if (atomic_read(&sk->sk_wmem_alloc) <=
911 skb = sock_wmalloc(sk,
912 alloclen + hh_len + 15, 1,
914 if (unlikely(skb == NULL))
921 * Fill in the control structures
923 skb->ip_summed = csummode;
925 skb_reserve(skb, hh_len);
928 * Find where to start putting bytes.
930 data = skb_put(skb, fraglen);
931 skb_set_network_header(skb, exthdrlen);
932 skb->h.raw = skb->nh.raw + fragheaderlen;
933 data += fragheaderlen;
936 skb->csum = skb_copy_and_csum_bits(
937 skb_prev, maxfraglen,
938 data + transhdrlen, fraggap, 0);
939 skb_prev->csum = csum_sub(skb_prev->csum,
942 pskb_trim_unique(skb_prev, maxfraglen);
945 copy = datalen - transhdrlen - fraggap;
946 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
953 length -= datalen - fraggap;
956 csummode = CHECKSUM_NONE;
959 * Put the packet on the pending queue.
961 __skb_queue_tail(&sk->sk_write_queue, skb);
968 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
972 if (getfrag(from, skb_put(skb, copy),
973 offset, copy, off, skb) < 0) {
974 __skb_trim(skb, off);
979 int i = skb_shinfo(skb)->nr_frags;
980 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
981 struct page *page = sk->sk_sndmsg_page;
982 int off = sk->sk_sndmsg_off;
985 if (page && (left = PAGE_SIZE - off) > 0) {
988 if (page != frag->page) {
989 if (i == MAX_SKB_FRAGS) {
994 skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
995 frag = &skb_shinfo(skb)->frags[i];
997 } else if (i < MAX_SKB_FRAGS) {
998 if (copy > PAGE_SIZE)
1000 page = alloc_pages(sk->sk_allocation, 0);
1005 sk->sk_sndmsg_page = page;
1006 sk->sk_sndmsg_off = 0;
1008 skb_fill_page_desc(skb, i, page, 0, 0);
1009 frag = &skb_shinfo(skb)->frags[i];
1010 skb->truesize += PAGE_SIZE;
1011 atomic_add(PAGE_SIZE, &sk->sk_wmem_alloc);
1016 if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1020 sk->sk_sndmsg_off += copy;
1023 skb->data_len += copy;
1032 inet->cork.length -= length;
1033 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1037 ssize_t ip_append_page(struct sock *sk, struct page *page,
1038 int offset, size_t size, int flags)
1040 struct inet_sock *inet = inet_sk(sk);
1041 struct sk_buff *skb;
1043 struct ip_options *opt = NULL;
1048 unsigned int maxfraglen, fragheaderlen, fraggap;
1053 if (flags&MSG_PROBE)
1056 if (skb_queue_empty(&sk->sk_write_queue))
1060 if (inet->cork.flags & IPCORK_OPT)
1061 opt = inet->cork.opt;
1063 if (!(rt->u.dst.dev->features&NETIF_F_SG))
1066 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1067 mtu = inet->cork.fragsize;
1069 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1070 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1072 if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1073 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu);
1077 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1080 inet->cork.length += size;
1081 if ((sk->sk_protocol == IPPROTO_UDP) &&
1082 (rt->u.dst.dev->features & NETIF_F_UFO)) {
1083 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1084 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1091 if (skb_is_gso(skb))
1095 /* Check if the remaining data fits into current packet. */
1096 len = mtu - skb->len;
1098 len = maxfraglen - skb->len;
1101 struct sk_buff *skb_prev;
1105 fraggap = skb_prev->len - maxfraglen;
1107 alloclen = fragheaderlen + hh_len + fraggap + 15;
1108 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1109 if (unlikely(!skb)) {
1115 * Fill in the control structures
1117 skb->ip_summed = CHECKSUM_NONE;
1119 skb_reserve(skb, hh_len);
1122 * Find where to start putting bytes.
1124 skb_put(skb, fragheaderlen + fraggap);
1125 skb_reset_network_header(skb);
1126 skb->h.raw = skb->nh.raw + fragheaderlen;
1129 skb->csum = skb_copy_and_csum_bits(skb_prev,
1133 skb_prev->csum = csum_sub(skb_prev->csum,
1135 pskb_trim_unique(skb_prev, maxfraglen);
1139 * Put the packet on the pending queue.
1141 __skb_queue_tail(&sk->sk_write_queue, skb);
1145 i = skb_shinfo(skb)->nr_frags;
1148 if (skb_can_coalesce(skb, i, page, offset)) {
1149 skb_shinfo(skb)->frags[i-1].size += len;
1150 } else if (i < MAX_SKB_FRAGS) {
1152 skb_fill_page_desc(skb, i, page, offset, len);
1158 if (skb->ip_summed == CHECKSUM_NONE) {
1160 csum = csum_page(page, offset, len);
1161 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1165 skb->data_len += len;
1172 inet->cork.length -= size;
1173 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1178 * Combined all pending IP fragments on the socket as one IP datagram
1179 * and push them out.
1181 int ip_push_pending_frames(struct sock *sk)
1183 struct sk_buff *skb, *tmp_skb;
1184 struct sk_buff **tail_skb;
1185 struct inet_sock *inet = inet_sk(sk);
1186 struct ip_options *opt = NULL;
1187 struct rtable *rt = inet->cork.rt;
1193 if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1195 tail_skb = &(skb_shinfo(skb)->frag_list);
1197 /* move skb->data to ip header from ext header */
1198 if (skb->data < skb_network_header(skb))
1199 __skb_pull(skb, skb_network_offset(skb));
1200 while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1201 __skb_pull(tmp_skb, skb->h.raw - skb->nh.raw);
1202 *tail_skb = tmp_skb;
1203 tail_skb = &(tmp_skb->next);
1204 skb->len += tmp_skb->len;
1205 skb->data_len += tmp_skb->len;
1206 skb->truesize += tmp_skb->truesize;
1207 __sock_put(tmp_skb->sk);
1208 tmp_skb->destructor = NULL;
1212 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1213 * to fragment the frame generated here. No matter, what transforms
1214 * how transforms change size of the packet, it will come out.
1216 if (inet->pmtudisc != IP_PMTUDISC_DO)
1219 /* DF bit is set when we want to see DF on outgoing frames.
1220 * If local_df is set too, we still allow to fragment this frame
1222 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1223 (skb->len <= dst_mtu(&rt->u.dst) &&
1224 ip_dont_fragment(sk, &rt->u.dst)))
1227 if (inet->cork.flags & IPCORK_OPT)
1228 opt = inet->cork.opt;
1230 if (rt->rt_type == RTN_MULTICAST)
1233 ttl = ip_select_ttl(inet, &rt->u.dst);
1235 iph = (struct iphdr *)skb->data;
1239 iph->ihl += opt->optlen>>2;
1240 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1242 iph->tos = inet->tos;
1243 iph->tot_len = htons(skb->len);
1245 ip_select_ident(iph, &rt->u.dst, sk);
1247 iph->protocol = sk->sk_protocol;
1248 iph->saddr = rt->rt_src;
1249 iph->daddr = rt->rt_dst;
1252 skb->priority = sk->sk_priority;
1253 skb->dst = dst_clone(&rt->u.dst);
1255 /* Netfilter gets whole the not fragmented skb. */
1256 err = NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL,
1257 skb->dst->dev, dst_output);
1260 err = inet->recverr ? net_xmit_errno(err) : 0;
1266 inet->cork.flags &= ~IPCORK_OPT;
1267 kfree(inet->cork.opt);
1268 inet->cork.opt = NULL;
1269 if (inet->cork.rt) {
1270 ip_rt_put(inet->cork.rt);
1271 inet->cork.rt = NULL;
1276 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1281 * Throw away all pending data on the socket.
1283 void ip_flush_pending_frames(struct sock *sk)
1285 struct inet_sock *inet = inet_sk(sk);
1286 struct sk_buff *skb;
1288 while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1291 inet->cork.flags &= ~IPCORK_OPT;
1292 kfree(inet->cork.opt);
1293 inet->cork.opt = NULL;
1294 if (inet->cork.rt) {
1295 ip_rt_put(inet->cork.rt);
1296 inet->cork.rt = NULL;
1302 * Fetch data from kernel space and fill in checksum if needed.
1304 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1305 int len, int odd, struct sk_buff *skb)
1309 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1310 skb->csum = csum_block_add(skb->csum, csum, odd);
1315 * Generic function to send a packet as reply to another packet.
1316 * Used to send TCP resets so far. ICMP should use this function too.
1318 * Should run single threaded per socket because it uses the sock
1319 * structure to pass arguments.
1321 * LATER: switch from ip_build_xmit to ip_append_*
1323 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1326 struct inet_sock *inet = inet_sk(sk);
1328 struct ip_options opt;
1331 struct ipcm_cookie ipc;
1333 struct rtable *rt = (struct rtable*)skb->dst;
1335 if (ip_options_echo(&replyopts.opt, skb))
1338 daddr = ipc.addr = rt->rt_src;
1341 if (replyopts.opt.optlen) {
1342 ipc.opt = &replyopts.opt;
1345 daddr = replyopts.opt.faddr;
1349 struct flowi fl = { .nl_u = { .ip4_u =
1351 .saddr = rt->rt_spec_dst,
1352 .tos = RT_TOS(ip_hdr(skb)->tos) } },
1353 /* Not quite clean, but right. */
1355 { .sport = tcp_hdr(skb)->dest,
1356 .dport = tcp_hdr(skb)->source } },
1357 .proto = sk->sk_protocol };
1358 security_skb_classify_flow(skb, &fl);
1359 if (ip_route_output_key(&rt, &fl))
1363 /* And let IP do all the hard work.
1365 This chunk is not reenterable, hence spinlock.
1366 Note that it uses the fact, that this function is called
1367 with locally disabled BH and that sk cannot be already spinlocked.
1370 inet->tos = ip_hdr(skb)->tos;
1371 sk->sk_priority = skb->priority;
1372 sk->sk_protocol = ip_hdr(skb)->protocol;
1373 ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1374 &ipc, rt, MSG_DONTWAIT);
1375 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1376 if (arg->csumoffset >= 0)
1377 *((__sum16 *)skb->h.raw + arg->csumoffset) = csum_fold(csum_add(skb->csum, arg->csum));
1378 skb->ip_summed = CHECKSUM_NONE;
1379 ip_push_pending_frames(sk);
1387 void __init ip_init(void)
1392 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1393 igmp_mc_proc_init();
1397 EXPORT_SYMBOL(ip_generic_getfrag);
1398 EXPORT_SYMBOL(ip_queue_xmit);
1399 EXPORT_SYMBOL(ip_send_check);