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.
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Donald Becker, <becker@super.org>
11 * Alan Cox, <Alan.Cox@linux.org>
13 * Stefan Becker, <stefanb@yello.ping.de>
14 * Jorge Cwik, <jorge@laser.satlink.net>
15 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
16 * Hirokazu Takahashi, <taka@valinux.co.jp>
18 * See ip_input.c for original log
21 * Alan Cox : Missing nonblock feature in ip_build_xmit.
22 * Mike Kilburn : htons() missing in ip_build_xmit.
23 * Bradford Johnson: Fix faulty handling of some frames when
25 * Alexander Demenshin: Missing sk/skb free in ip_queue_xmit
26 * (in case if packet not accepted by
27 * output firewall rules)
28 * Mike McLagan : Routing by source
29 * Alexey Kuznetsov: use new route cache
30 * Andi Kleen: Fix broken PMTU recovery and remove
31 * some redundant tests.
32 * Vitaly E. Lavrov : Transparent proxy revived after year coma.
33 * Andi Kleen : Replace ip_reply with ip_send_reply.
34 * Andi Kleen : Split fast and slow ip_build_xmit path
35 * for decreased register pressure on x86
36 * and more readibility.
37 * Marc Boucher : When call_out_firewall returns FW_QUEUE,
38 * silently drop skb instead of failing with -EPERM.
39 * Detlev Wengorz : Copy protocol for fragments.
40 * Hirokazu Takahashi: HW checksumming for outgoing UDP
42 * Hirokazu Takahashi: sendfile() on UDP works now.
45 #include <asm/uaccess.h>
46 #include <asm/system.h>
47 #include <linux/module.h>
48 #include <linux/types.h>
49 #include <linux/kernel.h>
51 #include <linux/string.h>
52 #include <linux/errno.h>
53 #include <linux/highmem.h>
54 #include <linux/slab.h>
56 #include <linux/socket.h>
57 #include <linux/sockios.h>
59 #include <linux/inet.h>
60 #include <linux/netdevice.h>
61 #include <linux/etherdevice.h>
62 #include <linux/proc_fs.h>
63 #include <linux/stat.h>
64 #include <linux/init.h>
68 #include <net/protocol.h>
69 #include <net/route.h>
71 #include <linux/skbuff.h>
75 #include <net/checksum.h>
76 #include <net/inetpeer.h>
77 #include <linux/igmp.h>
78 #include <linux/netfilter_ipv4.h>
79 #include <linux/netfilter_bridge.h>
80 #include <linux/mroute.h>
81 #include <linux/netlink.h>
82 #include <linux/tcp.h>
84 int sysctl_ip_default_ttl __read_mostly = IPDEFTTL;
85 EXPORT_SYMBOL(sysctl_ip_default_ttl);
87 /* Generate a checksum for an outgoing IP datagram. */
88 __inline__ void ip_send_check(struct iphdr *iph)
91 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
93 EXPORT_SYMBOL(ip_send_check);
95 int __ip_local_out(struct sk_buff *skb)
97 struct iphdr *iph = ip_hdr(skb);
99 iph->tot_len = htons(skb->len);
101 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT, skb, NULL,
102 skb_dst(skb)->dev, dst_output);
105 int ip_local_out(struct sk_buff *skb)
109 err = __ip_local_out(skb);
110 if (likely(err == 1))
111 err = dst_output(skb);
115 EXPORT_SYMBOL_GPL(ip_local_out);
117 /* dev_loopback_xmit for use with netfilter. */
118 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
120 skb_reset_mac_header(newskb);
121 __skb_pull(newskb, skb_network_offset(newskb));
122 newskb->pkt_type = PACKET_LOOPBACK;
123 newskb->ip_summed = CHECKSUM_UNNECESSARY;
124 WARN_ON(!skb_dst(newskb));
129 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
131 int ttl = inet->uc_ttl;
134 ttl = ip4_dst_hoplimit(dst);
139 * Add an ip header to a skbuff and send it out.
142 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
143 __be32 saddr, __be32 daddr, struct ip_options *opt)
145 struct inet_sock *inet = inet_sk(sk);
146 struct rtable *rt = skb_rtable(skb);
149 /* Build the IP header. */
150 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
151 skb_reset_network_header(skb);
155 iph->tos = inet->tos;
156 if (ip_dont_fragment(sk, &rt->dst))
157 iph->frag_off = htons(IP_DF);
160 iph->ttl = ip_select_ttl(inet, &rt->dst);
161 iph->daddr = rt->rt_dst;
162 iph->saddr = rt->rt_src;
163 iph->protocol = sk->sk_protocol;
164 ip_select_ident(iph, &rt->dst, sk);
166 if (opt && opt->optlen) {
167 iph->ihl += opt->optlen>>2;
168 ip_options_build(skb, opt, daddr, rt, 0);
171 skb->priority = sk->sk_priority;
172 skb->mark = sk->sk_mark;
175 return ip_local_out(skb);
177 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
179 static inline int ip_finish_output2(struct sk_buff *skb)
181 struct dst_entry *dst = skb_dst(skb);
182 struct rtable *rt = (struct rtable *)dst;
183 struct net_device *dev = dst->dev;
184 unsigned int hh_len = LL_RESERVED_SPACE(dev);
186 if (rt->rt_type == RTN_MULTICAST) {
187 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
188 } else if (rt->rt_type == RTN_BROADCAST)
189 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
191 /* Be paranoid, rather than too clever. */
192 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
193 struct sk_buff *skb2;
195 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
201 skb_set_owner_w(skb2, skb->sk);
207 return neigh_hh_output(dst->hh, skb);
208 else if (dst->neighbour)
209 return dst->neighbour->output(skb);
212 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
217 static inline int ip_skb_dst_mtu(struct sk_buff *skb)
219 struct inet_sock *inet = skb->sk ? inet_sk(skb->sk) : NULL;
221 return (inet && inet->pmtudisc == IP_PMTUDISC_PROBE) ?
222 skb_dst(skb)->dev->mtu : dst_mtu(skb_dst(skb));
225 static int ip_finish_output(struct sk_buff *skb)
227 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
228 /* Policy lookup after SNAT yielded a new policy */
229 if (skb_dst(skb)->xfrm != NULL) {
230 IPCB(skb)->flags |= IPSKB_REROUTED;
231 return dst_output(skb);
234 if (skb->len > ip_skb_dst_mtu(skb) && !skb_is_gso(skb))
235 return ip_fragment(skb, ip_finish_output2);
237 return ip_finish_output2(skb);
240 int ip_mc_output(struct sk_buff *skb)
242 struct sock *sk = skb->sk;
243 struct rtable *rt = skb_rtable(skb);
244 struct net_device *dev = rt->dst.dev;
247 * If the indicated interface is up and running, send the packet.
249 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
252 skb->protocol = htons(ETH_P_IP);
255 * Multicasts are looped back for other local users
258 if (rt->rt_flags&RTCF_MULTICAST) {
260 #ifdef CONFIG_IP_MROUTE
261 /* Small optimization: do not loopback not local frames,
262 which returned after forwarding; they will be dropped
263 by ip_mr_input in any case.
264 Note, that local frames are looped back to be delivered
267 This check is duplicated in ip_mr_input at the moment.
270 ((rt->rt_flags & RTCF_LOCAL) ||
271 !(IPCB(skb)->flags & IPSKB_FORWARDED))
274 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
276 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
277 newskb, NULL, newskb->dev,
278 ip_dev_loopback_xmit);
281 /* Multicasts with ttl 0 must not go beyond the host */
283 if (ip_hdr(skb)->ttl == 0) {
289 if (rt->rt_flags&RTCF_BROADCAST) {
290 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
292 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
293 NULL, newskb->dev, ip_dev_loopback_xmit);
296 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
297 skb->dev, ip_finish_output,
298 !(IPCB(skb)->flags & IPSKB_REROUTED));
301 int ip_output(struct sk_buff *skb)
303 struct net_device *dev = skb_dst(skb)->dev;
305 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
308 skb->protocol = htons(ETH_P_IP);
310 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
312 !(IPCB(skb)->flags & IPSKB_REROUTED));
315 int ip_queue_xmit(struct sk_buff *skb)
317 struct sock *sk = skb->sk;
318 struct inet_sock *inet = inet_sk(sk);
319 struct ip_options *opt = inet->opt;
324 /* Skip all of this if the packet is already routed,
325 * f.e. by something like SCTP.
328 rt = skb_rtable(skb);
332 /* Make sure we can route this packet. */
333 rt = (struct rtable *)__sk_dst_check(sk, 0);
337 /* Use correct destination address if we have options. */
338 daddr = inet->inet_daddr;
342 /* If this fails, retransmit mechanism of transport layer will
343 * keep trying until route appears or the connection times
346 rt = ip_route_output_ports(sock_net(sk), sk,
347 daddr, inet->inet_saddr,
352 sk->sk_bound_dev_if);
355 sk_setup_caps(sk, &rt->dst);
357 skb_dst_set_noref(skb, &rt->dst);
360 if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
363 /* OK, we know where to send it, allocate and build IP header. */
364 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
365 skb_reset_network_header(skb);
367 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
368 if (ip_dont_fragment(sk, &rt->dst) && !skb->local_df)
369 iph->frag_off = htons(IP_DF);
372 iph->ttl = ip_select_ttl(inet, &rt->dst);
373 iph->protocol = sk->sk_protocol;
374 iph->saddr = rt->rt_src;
375 iph->daddr = rt->rt_dst;
376 /* Transport layer set skb->h.foo itself. */
378 if (opt && opt->optlen) {
379 iph->ihl += opt->optlen >> 2;
380 ip_options_build(skb, opt, inet->inet_daddr, rt, 0);
383 ip_select_ident_more(iph, &rt->dst, sk,
384 (skb_shinfo(skb)->gso_segs ?: 1) - 1);
386 skb->priority = sk->sk_priority;
387 skb->mark = sk->sk_mark;
389 res = ip_local_out(skb);
395 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
397 return -EHOSTUNREACH;
399 EXPORT_SYMBOL(ip_queue_xmit);
402 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
404 to->pkt_type = from->pkt_type;
405 to->priority = from->priority;
406 to->protocol = from->protocol;
408 skb_dst_copy(to, from);
410 to->mark = from->mark;
412 /* Copy the flags to each fragment. */
413 IPCB(to)->flags = IPCB(from)->flags;
415 #ifdef CONFIG_NET_SCHED
416 to->tc_index = from->tc_index;
419 #if defined(CONFIG_NETFILTER_XT_TARGET_TRACE) || \
420 defined(CONFIG_NETFILTER_XT_TARGET_TRACE_MODULE)
421 to->nf_trace = from->nf_trace;
423 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
424 to->ipvs_property = from->ipvs_property;
426 skb_copy_secmark(to, from);
430 * This IP datagram is too large to be sent in one piece. Break it up into
431 * smaller pieces (each of size equal to IP header plus
432 * a block of the data of the original IP data part) that will yet fit in a
433 * single device frame, and queue such a frame for sending.
436 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
440 struct net_device *dev;
441 struct sk_buff *skb2;
442 unsigned int mtu, hlen, left, len, ll_rs;
444 __be16 not_last_frag;
445 struct rtable *rt = skb_rtable(skb);
451 * Point into the IP datagram header.
456 if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
457 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
458 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
459 htonl(ip_skb_dst_mtu(skb)));
465 * Setup starting values.
469 mtu = dst_mtu(&rt->dst) - hlen; /* Size of data space */
470 #ifdef CONFIG_BRIDGE_NETFILTER
472 mtu -= nf_bridge_mtu_reduction(skb);
474 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
476 /* When frag_list is given, use it. First, check its validity:
477 * some transformers could create wrong frag_list or break existing
478 * one, it is not prohibited. In this case fall back to copying.
480 * LATER: this step can be merged to real generation of fragments,
481 * we can switch to copy when see the first bad fragment.
483 if (skb_has_frag_list(skb)) {
484 struct sk_buff *frag, *frag2;
485 int first_len = skb_pagelen(skb);
487 if (first_len - hlen > mtu ||
488 ((first_len - hlen) & 7) ||
489 (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
493 skb_walk_frags(skb, frag) {
494 /* Correct geometry. */
495 if (frag->len > mtu ||
496 ((frag->len & 7) && frag->next) ||
497 skb_headroom(frag) < hlen)
498 goto slow_path_clean;
500 /* Partially cloned skb? */
501 if (skb_shared(frag))
502 goto slow_path_clean;
507 frag->destructor = sock_wfree;
509 skb->truesize -= frag->truesize;
512 /* Everything is OK. Generate! */
516 frag = skb_shinfo(skb)->frag_list;
517 skb_frag_list_init(skb);
518 skb->data_len = first_len - skb_headlen(skb);
519 skb->len = first_len;
520 iph->tot_len = htons(first_len);
521 iph->frag_off = htons(IP_MF);
525 /* Prepare header of the next frame,
526 * before previous one went down. */
528 frag->ip_summed = CHECKSUM_NONE;
529 skb_reset_transport_header(frag);
530 __skb_push(frag, hlen);
531 skb_reset_network_header(frag);
532 memcpy(skb_network_header(frag), iph, hlen);
534 iph->tot_len = htons(frag->len);
535 ip_copy_metadata(frag, skb);
537 ip_options_fragment(frag);
538 offset += skb->len - hlen;
539 iph->frag_off = htons(offset>>3);
540 if (frag->next != NULL)
541 iph->frag_off |= htons(IP_MF);
542 /* Ready, complete checksum */
549 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
559 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
568 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
572 skb_walk_frags(skb, frag2) {
576 frag2->destructor = NULL;
577 skb->truesize += frag2->truesize;
582 left = skb->len - hlen; /* Space per frame */
583 ptr = hlen; /* Where to start from */
585 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
586 * we need to make room for the encapsulating header
588 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
591 * Fragment the datagram.
594 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
595 not_last_frag = iph->frag_off & htons(IP_MF);
598 * Keep copying data until we run out.
603 /* IF: it doesn't fit, use 'mtu' - the data space left */
606 /* IF: we are not sending upto and including the packet end
607 then align the next start on an eight byte boundary */
615 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
616 NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
622 * Set up data on packet
625 ip_copy_metadata(skb2, skb);
626 skb_reserve(skb2, ll_rs);
627 skb_put(skb2, len + hlen);
628 skb_reset_network_header(skb2);
629 skb2->transport_header = skb2->network_header + hlen;
632 * Charge the memory for the fragment to any owner
637 skb_set_owner_w(skb2, skb->sk);
640 * Copy the packet header into the new buffer.
643 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
646 * Copy a block of the IP datagram.
648 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
653 * Fill in the new header fields.
656 iph->frag_off = htons((offset >> 3));
658 /* ANK: dirty, but effective trick. Upgrade options only if
659 * the segment to be fragmented was THE FIRST (otherwise,
660 * options are already fixed) and make it ONCE
661 * on the initial skb, so that all the following fragments
662 * will inherit fixed options.
665 ip_options_fragment(skb);
668 * Added AC : If we are fragmenting a fragment that's not the
669 * last fragment then keep MF on each bit
671 if (left > 0 || not_last_frag)
672 iph->frag_off |= htons(IP_MF);
677 * Put this fragment into the sending queue.
679 iph->tot_len = htons(len + hlen);
687 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
690 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
695 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
698 EXPORT_SYMBOL(ip_fragment);
701 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
703 struct iovec *iov = from;
705 if (skb->ip_summed == CHECKSUM_PARTIAL) {
706 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
710 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
712 skb->csum = csum_block_add(skb->csum, csum, odd);
716 EXPORT_SYMBOL(ip_generic_getfrag);
719 csum_page(struct page *page, int offset, int copy)
724 csum = csum_partial(kaddr + offset, copy, 0);
729 static inline int ip_ufo_append_data(struct sock *sk,
730 struct sk_buff_head *queue,
731 int getfrag(void *from, char *to, int offset, int len,
732 int odd, struct sk_buff *skb),
733 void *from, int length, int hh_len, int fragheaderlen,
734 int transhdrlen, int mtu, unsigned int flags)
739 /* There is support for UDP fragmentation offload by network
740 * device, so create one single skb packet containing complete
743 if ((skb = skb_peek_tail(queue)) == NULL) {
744 skb = sock_alloc_send_skb(sk,
745 hh_len + fragheaderlen + transhdrlen + 20,
746 (flags & MSG_DONTWAIT), &err);
751 /* reserve space for Hardware header */
752 skb_reserve(skb, hh_len);
754 /* create space for UDP/IP header */
755 skb_put(skb, fragheaderlen + transhdrlen);
757 /* initialize network header pointer */
758 skb_reset_network_header(skb);
760 /* initialize protocol header pointer */
761 skb->transport_header = skb->network_header + fragheaderlen;
763 skb->ip_summed = CHECKSUM_PARTIAL;
766 /* specify the length of each IP datagram fragment */
767 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
768 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
769 __skb_queue_tail(queue, skb);
772 return skb_append_datato_frags(sk, skb, getfrag, from,
773 (length - transhdrlen));
776 static int __ip_append_data(struct sock *sk, struct sk_buff_head *queue,
777 struct inet_cork *cork,
778 int getfrag(void *from, char *to, int offset,
779 int len, int odd, struct sk_buff *skb),
780 void *from, int length, int transhdrlen,
783 struct inet_sock *inet = inet_sk(sk);
786 struct ip_options *opt = cork->opt;
793 unsigned int maxfraglen, fragheaderlen;
794 int csummode = CHECKSUM_NONE;
795 struct rtable *rt = (struct rtable *)cork->dst;
797 exthdrlen = transhdrlen ? rt->dst.header_len : 0;
799 transhdrlen += exthdrlen;
800 mtu = cork->fragsize;
802 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
804 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
805 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
807 if (cork->length + length > 0xFFFF - fragheaderlen) {
808 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport,
814 * transhdrlen > 0 means that this is the first fragment and we wish
815 * it won't be fragmented in the future.
818 length + fragheaderlen <= mtu &&
819 rt->dst.dev->features & NETIF_F_V4_CSUM &&
821 csummode = CHECKSUM_PARTIAL;
823 skb = skb_peek_tail(queue);
825 cork->length += length;
826 if (((length > mtu) || (skb && skb_is_gso(skb))) &&
827 (sk->sk_protocol == IPPROTO_UDP) &&
828 (rt->dst.dev->features & NETIF_F_UFO)) {
829 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
830 hh_len, fragheaderlen, transhdrlen,
837 /* So, what's going on in the loop below?
839 * We use calculated fragment length to generate chained skb,
840 * each of segments is IP fragment ready for sending to network after
841 * adding appropriate IP header.
848 /* Check if the remaining data fits into current packet. */
849 copy = mtu - skb->len;
851 copy = maxfraglen - skb->len;
854 unsigned int datalen;
855 unsigned int fraglen;
856 unsigned int fraggap;
857 unsigned int alloclen;
858 struct sk_buff *skb_prev;
862 fraggap = skb_prev->len - maxfraglen;
867 * If remaining data exceeds the mtu,
868 * we know we need more fragment(s).
870 datalen = length + fraggap;
871 if (datalen > mtu - fragheaderlen)
872 datalen = maxfraglen - fragheaderlen;
873 fraglen = datalen + fragheaderlen;
875 if ((flags & MSG_MORE) &&
876 !(rt->dst.dev->features&NETIF_F_SG))
881 /* The last fragment gets additional space at tail.
882 * Note, with MSG_MORE we overallocate on fragments,
883 * because we have no idea what fragment will be
886 if (datalen == length + fraggap) {
887 alloclen += rt->dst.trailer_len;
888 /* make sure mtu is not reached */
889 if (datalen > mtu - fragheaderlen - rt->dst.trailer_len)
890 datalen -= ALIGN(rt->dst.trailer_len, 8);
893 skb = sock_alloc_send_skb(sk,
894 alloclen + hh_len + 15,
895 (flags & MSG_DONTWAIT), &err);
898 if (atomic_read(&sk->sk_wmem_alloc) <=
900 skb = sock_wmalloc(sk,
901 alloclen + hh_len + 15, 1,
903 if (unlikely(skb == NULL))
906 /* only the initial fragment is
914 * Fill in the control structures
916 skb->ip_summed = csummode;
918 skb_reserve(skb, hh_len);
919 skb_shinfo(skb)->tx_flags = cork->tx_flags;
922 * Find where to start putting bytes.
924 data = skb_put(skb, fraglen);
925 skb_set_network_header(skb, exthdrlen);
926 skb->transport_header = (skb->network_header +
928 data += fragheaderlen;
931 skb->csum = skb_copy_and_csum_bits(
932 skb_prev, maxfraglen,
933 data + transhdrlen, fraggap, 0);
934 skb_prev->csum = csum_sub(skb_prev->csum,
937 pskb_trim_unique(skb_prev, maxfraglen);
940 copy = datalen - transhdrlen - fraggap;
941 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
948 length -= datalen - fraggap;
951 csummode = CHECKSUM_NONE;
954 * Put the packet on the pending queue.
956 __skb_queue_tail(queue, skb);
963 if (!(rt->dst.dev->features&NETIF_F_SG)) {
967 if (getfrag(from, skb_put(skb, copy),
968 offset, copy, off, skb) < 0) {
969 __skb_trim(skb, off);
974 int i = skb_shinfo(skb)->nr_frags;
975 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
976 struct page *page = cork->page;
980 if (page && (left = PAGE_SIZE - off) > 0) {
983 if (page != frag->page) {
984 if (i == MAX_SKB_FRAGS) {
989 skb_fill_page_desc(skb, i, page, off, 0);
990 frag = &skb_shinfo(skb)->frags[i];
992 } else if (i < MAX_SKB_FRAGS) {
993 if (copy > PAGE_SIZE)
995 page = alloc_pages(sk->sk_allocation, 0);
1003 skb_fill_page_desc(skb, i, page, 0, 0);
1004 frag = &skb_shinfo(skb)->frags[i];
1009 if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1016 skb->data_len += copy;
1017 skb->truesize += copy;
1018 atomic_add(copy, &sk->sk_wmem_alloc);
1027 cork->length -= length;
1028 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1032 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1033 struct ipcm_cookie *ipc, struct rtable **rtp)
1035 struct inet_sock *inet = inet_sk(sk);
1036 struct ip_options *opt;
1040 * setup for corking.
1044 if (cork->opt == NULL) {
1045 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1047 if (unlikely(cork->opt == NULL))
1050 memcpy(cork->opt, opt, sizeof(struct ip_options) + opt->optlen);
1051 cork->flags |= IPCORK_OPT;
1052 cork->addr = ipc->addr;
1058 * We steal reference to this route, caller should not release it
1061 cork->fragsize = inet->pmtudisc == IP_PMTUDISC_PROBE ?
1062 rt->dst.dev->mtu : dst_mtu(rt->dst.path);
1063 cork->dst = &rt->dst;
1065 cork->tx_flags = ipc->tx_flags;
1073 * ip_append_data() and ip_append_page() can make one large IP datagram
1074 * from many pieces of data. Each pieces will be holded on the socket
1075 * until ip_push_pending_frames() is called. Each piece can be a page
1078 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1079 * this interface potentially.
1081 * LATER: length must be adjusted by pad at tail, when it is required.
1083 int ip_append_data(struct sock *sk,
1084 int getfrag(void *from, char *to, int offset, int len,
1085 int odd, struct sk_buff *skb),
1086 void *from, int length, int transhdrlen,
1087 struct ipcm_cookie *ipc, struct rtable **rtp,
1090 struct inet_sock *inet = inet_sk(sk);
1093 if (flags&MSG_PROBE)
1096 if (skb_queue_empty(&sk->sk_write_queue)) {
1097 err = ip_setup_cork(sk, &inet->cork, ipc, rtp);
1104 return __ip_append_data(sk, &sk->sk_write_queue, &inet->cork, getfrag,
1105 from, length, transhdrlen, flags);
1108 ssize_t ip_append_page(struct sock *sk, struct page *page,
1109 int offset, size_t size, int flags)
1111 struct inet_sock *inet = inet_sk(sk);
1112 struct sk_buff *skb;
1114 struct ip_options *opt = NULL;
1119 unsigned int maxfraglen, fragheaderlen, fraggap;
1124 if (flags&MSG_PROBE)
1127 if (skb_queue_empty(&sk->sk_write_queue))
1130 rt = (struct rtable *)inet->cork.dst;
1131 if (inet->cork.flags & IPCORK_OPT)
1132 opt = inet->cork.opt;
1134 if (!(rt->dst.dev->features&NETIF_F_SG))
1137 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1138 mtu = inet->cork.fragsize;
1140 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1141 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1143 if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1144 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport, mtu);
1148 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1151 inet->cork.length += size;
1152 if ((size + skb->len > mtu) &&
1153 (sk->sk_protocol == IPPROTO_UDP) &&
1154 (rt->dst.dev->features & NETIF_F_UFO)) {
1155 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1156 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1163 if (skb_is_gso(skb))
1167 /* Check if the remaining data fits into current packet. */
1168 len = mtu - skb->len;
1170 len = maxfraglen - skb->len;
1173 struct sk_buff *skb_prev;
1177 fraggap = skb_prev->len - maxfraglen;
1179 alloclen = fragheaderlen + hh_len + fraggap + 15;
1180 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1181 if (unlikely(!skb)) {
1187 * Fill in the control structures
1189 skb->ip_summed = CHECKSUM_NONE;
1191 skb_reserve(skb, hh_len);
1194 * Find where to start putting bytes.
1196 skb_put(skb, fragheaderlen + fraggap);
1197 skb_reset_network_header(skb);
1198 skb->transport_header = (skb->network_header +
1201 skb->csum = skb_copy_and_csum_bits(skb_prev,
1203 skb_transport_header(skb),
1205 skb_prev->csum = csum_sub(skb_prev->csum,
1207 pskb_trim_unique(skb_prev, maxfraglen);
1211 * Put the packet on the pending queue.
1213 __skb_queue_tail(&sk->sk_write_queue, skb);
1217 i = skb_shinfo(skb)->nr_frags;
1220 if (skb_can_coalesce(skb, i, page, offset)) {
1221 skb_shinfo(skb)->frags[i-1].size += len;
1222 } else if (i < MAX_SKB_FRAGS) {
1224 skb_fill_page_desc(skb, i, page, offset, len);
1230 if (skb->ip_summed == CHECKSUM_NONE) {
1232 csum = csum_page(page, offset, len);
1233 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1237 skb->data_len += len;
1238 skb->truesize += len;
1239 atomic_add(len, &sk->sk_wmem_alloc);
1246 inet->cork.length -= size;
1247 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1251 static void ip_cork_release(struct inet_cork *cork)
1253 cork->flags &= ~IPCORK_OPT;
1256 dst_release(cork->dst);
1261 * Combined all pending IP fragments on the socket as one IP datagram
1262 * and push them out.
1264 struct sk_buff *__ip_make_skb(struct sock *sk,
1265 struct sk_buff_head *queue,
1266 struct inet_cork *cork)
1268 struct sk_buff *skb, *tmp_skb;
1269 struct sk_buff **tail_skb;
1270 struct inet_sock *inet = inet_sk(sk);
1271 struct net *net = sock_net(sk);
1272 struct ip_options *opt = NULL;
1273 struct rtable *rt = (struct rtable *)cork->dst;
1278 if ((skb = __skb_dequeue(queue)) == NULL)
1280 tail_skb = &(skb_shinfo(skb)->frag_list);
1282 /* move skb->data to ip header from ext header */
1283 if (skb->data < skb_network_header(skb))
1284 __skb_pull(skb, skb_network_offset(skb));
1285 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1286 __skb_pull(tmp_skb, skb_network_header_len(skb));
1287 *tail_skb = tmp_skb;
1288 tail_skb = &(tmp_skb->next);
1289 skb->len += tmp_skb->len;
1290 skb->data_len += tmp_skb->len;
1291 skb->truesize += tmp_skb->truesize;
1292 tmp_skb->destructor = NULL;
1296 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1297 * to fragment the frame generated here. No matter, what transforms
1298 * how transforms change size of the packet, it will come out.
1300 if (inet->pmtudisc < IP_PMTUDISC_DO)
1303 /* DF bit is set when we want to see DF on outgoing frames.
1304 * If local_df is set too, we still allow to fragment this frame
1306 if (inet->pmtudisc >= IP_PMTUDISC_DO ||
1307 (skb->len <= dst_mtu(&rt->dst) &&
1308 ip_dont_fragment(sk, &rt->dst)))
1311 if (cork->flags & IPCORK_OPT)
1314 if (rt->rt_type == RTN_MULTICAST)
1317 ttl = ip_select_ttl(inet, &rt->dst);
1319 iph = (struct iphdr *)skb->data;
1323 iph->ihl += opt->optlen>>2;
1324 ip_options_build(skb, opt, cork->addr, rt, 0);
1326 iph->tos = inet->tos;
1328 ip_select_ident(iph, &rt->dst, sk);
1330 iph->protocol = sk->sk_protocol;
1331 iph->saddr = rt->rt_src;
1332 iph->daddr = rt->rt_dst;
1334 skb->priority = sk->sk_priority;
1335 skb->mark = sk->sk_mark;
1337 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1341 skb_dst_set(skb, &rt->dst);
1343 if (iph->protocol == IPPROTO_ICMP)
1344 icmp_out_count(net, ((struct icmphdr *)
1345 skb_transport_header(skb))->type);
1347 ip_cork_release(cork);
1352 int ip_send_skb(struct sk_buff *skb)
1354 struct net *net = sock_net(skb->sk);
1357 err = ip_local_out(skb);
1360 err = net_xmit_errno(err);
1362 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1368 int ip_push_pending_frames(struct sock *sk)
1370 struct sk_buff *skb;
1372 skb = ip_finish_skb(sk);
1376 /* Netfilter gets whole the not fragmented skb. */
1377 return ip_send_skb(skb);
1381 * Throw away all pending data on the socket.
1383 static void __ip_flush_pending_frames(struct sock *sk,
1384 struct sk_buff_head *queue,
1385 struct inet_cork *cork)
1387 struct sk_buff *skb;
1389 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1392 ip_cork_release(cork);
1395 void ip_flush_pending_frames(struct sock *sk)
1397 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork);
1400 struct sk_buff *ip_make_skb(struct sock *sk,
1401 int getfrag(void *from, char *to, int offset,
1402 int len, int odd, struct sk_buff *skb),
1403 void *from, int length, int transhdrlen,
1404 struct ipcm_cookie *ipc, struct rtable **rtp,
1407 struct inet_cork cork = {};
1408 struct sk_buff_head queue;
1411 if (flags & MSG_PROBE)
1414 __skb_queue_head_init(&queue);
1416 err = ip_setup_cork(sk, &cork, ipc, rtp);
1418 return ERR_PTR(err);
1420 err = __ip_append_data(sk, &queue, &cork, getfrag,
1421 from, length, transhdrlen, flags);
1423 __ip_flush_pending_frames(sk, &queue, &cork);
1424 return ERR_PTR(err);
1427 return __ip_make_skb(sk, &queue, &cork);
1431 * Fetch data from kernel space and fill in checksum if needed.
1433 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1434 int len, int odd, struct sk_buff *skb)
1438 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1439 skb->csum = csum_block_add(skb->csum, csum, odd);
1444 * Generic function to send a packet as reply to another packet.
1445 * Used to send TCP resets so far. ICMP should use this function too.
1447 * Should run single threaded per socket because it uses the sock
1448 * structure to pass arguments.
1450 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1453 struct inet_sock *inet = inet_sk(sk);
1455 struct ip_options opt;
1458 struct ipcm_cookie ipc;
1460 struct rtable *rt = skb_rtable(skb);
1462 if (ip_options_echo(&replyopts.opt, skb))
1465 daddr = ipc.addr = rt->rt_src;
1469 if (replyopts.opt.optlen) {
1470 ipc.opt = &replyopts.opt;
1473 daddr = replyopts.opt.faddr;
1478 .flowi_oif = arg->bound_dev_if,
1480 .fl4_src = rt->rt_spec_dst,
1481 .fl4_tos = RT_TOS(ip_hdr(skb)->tos),
1482 .fl_ip_sport = tcp_hdr(skb)->dest,
1483 .fl_ip_dport = tcp_hdr(skb)->source,
1484 .flowi_proto = sk->sk_protocol,
1485 .flowi_flags = ip_reply_arg_flowi_flags(arg),
1487 security_skb_classify_flow(skb, &fl);
1488 rt = ip_route_output_key(sock_net(sk), &fl);
1493 /* And let IP do all the hard work.
1495 This chunk is not reenterable, hence spinlock.
1496 Note that it uses the fact, that this function is called
1497 with locally disabled BH and that sk cannot be already spinlocked.
1500 inet->tos = ip_hdr(skb)->tos;
1501 sk->sk_priority = skb->priority;
1502 sk->sk_protocol = ip_hdr(skb)->protocol;
1503 sk->sk_bound_dev_if = arg->bound_dev_if;
1504 ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1505 &ipc, &rt, MSG_DONTWAIT);
1506 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1507 if (arg->csumoffset >= 0)
1508 *((__sum16 *)skb_transport_header(skb) +
1509 arg->csumoffset) = csum_fold(csum_add(skb->csum,
1511 skb->ip_summed = CHECKSUM_NONE;
1512 ip_push_pending_frames(sk);
1520 void __init ip_init(void)
1525 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1526 igmp_mc_proc_init();