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;
86 /* Generate a checksum for an outgoing IP datagram. */
87 __inline__ void ip_send_check(struct iphdr *iph)
90 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
93 int __ip_local_out(struct sk_buff *skb)
95 struct iphdr *iph = ip_hdr(skb);
97 iph->tot_len = htons(skb->len);
99 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT, skb, NULL,
100 skb_dst(skb)->dev, dst_output);
103 int ip_local_out(struct sk_buff *skb)
107 err = __ip_local_out(skb);
108 if (likely(err == 1))
109 err = dst_output(skb);
113 EXPORT_SYMBOL_GPL(ip_local_out);
115 /* dev_loopback_xmit for use with netfilter. */
116 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
118 skb_reset_mac_header(newskb);
119 __skb_pull(newskb, skb_network_offset(newskb));
120 newskb->pkt_type = PACKET_LOOPBACK;
121 newskb->ip_summed = CHECKSUM_UNNECESSARY;
122 WARN_ON(!skb_dst(newskb));
127 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
129 int ttl = inet->uc_ttl;
132 ttl = dst_metric(dst, RTAX_HOPLIMIT);
137 * Add an ip header to a skbuff and send it out.
140 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
141 __be32 saddr, __be32 daddr, struct ip_options *opt)
143 struct inet_sock *inet = inet_sk(sk);
144 struct rtable *rt = skb_rtable(skb);
147 /* Build the IP header. */
148 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
149 skb_reset_network_header(skb);
153 iph->tos = inet->tos;
154 if (ip_dont_fragment(sk, &rt->dst))
155 iph->frag_off = htons(IP_DF);
158 iph->ttl = ip_select_ttl(inet, &rt->dst);
159 iph->daddr = rt->rt_dst;
160 iph->saddr = rt->rt_src;
161 iph->protocol = sk->sk_protocol;
162 ip_select_ident(iph, &rt->dst, sk);
164 if (opt && opt->optlen) {
165 iph->ihl += opt->optlen>>2;
166 ip_options_build(skb, opt, daddr, rt, 0);
169 skb->priority = sk->sk_priority;
170 skb->mark = sk->sk_mark;
173 return ip_local_out(skb);
176 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
178 static inline int ip_finish_output2(struct sk_buff *skb)
180 struct dst_entry *dst = skb_dst(skb);
181 struct rtable *rt = (struct rtable *)dst;
182 struct net_device *dev = dst->dev;
183 unsigned int hh_len = LL_RESERVED_SPACE(dev);
185 if (rt->rt_type == RTN_MULTICAST) {
186 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
187 } else if (rt->rt_type == RTN_BROADCAST)
188 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
190 /* Be paranoid, rather than too clever. */
191 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
192 struct sk_buff *skb2;
194 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
200 skb_set_owner_w(skb2, skb->sk);
206 return neigh_hh_output(dst->hh, skb);
207 else if (dst->neighbour)
208 return dst->neighbour->output(skb);
211 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
216 static inline int ip_skb_dst_mtu(struct sk_buff *skb)
218 struct inet_sock *inet = skb->sk ? inet_sk(skb->sk) : NULL;
220 return (inet && inet->pmtudisc == IP_PMTUDISC_PROBE) ?
221 skb_dst(skb)->dev->mtu : dst_mtu(skb_dst(skb));
224 static int ip_finish_output(struct sk_buff *skb)
226 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
227 /* Policy lookup after SNAT yielded a new policy */
228 if (skb_dst(skb)->xfrm != NULL) {
229 IPCB(skb)->flags |= IPSKB_REROUTED;
230 return dst_output(skb);
233 if (skb->len > ip_skb_dst_mtu(skb) && !skb_is_gso(skb))
234 return ip_fragment(skb, ip_finish_output2);
236 return ip_finish_output2(skb);
239 int ip_mc_output(struct sk_buff *skb)
241 struct sock *sk = skb->sk;
242 struct rtable *rt = skb_rtable(skb);
243 struct net_device *dev = rt->dst.dev;
246 * If the indicated interface is up and running, send the packet.
248 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
251 skb->protocol = htons(ETH_P_IP);
254 * Multicasts are looped back for other local users
257 if (rt->rt_flags&RTCF_MULTICAST) {
259 #ifdef CONFIG_IP_MROUTE
260 /* Small optimization: do not loopback not local frames,
261 which returned after forwarding; they will be dropped
262 by ip_mr_input in any case.
263 Note, that local frames are looped back to be delivered
266 This check is duplicated in ip_mr_input at the moment.
269 ((rt->rt_flags & RTCF_LOCAL) ||
270 !(IPCB(skb)->flags & IPSKB_FORWARDED))
273 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
275 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
276 newskb, NULL, newskb->dev,
277 ip_dev_loopback_xmit);
280 /* Multicasts with ttl 0 must not go beyond the host */
282 if (ip_hdr(skb)->ttl == 0) {
288 if (rt->rt_flags&RTCF_BROADCAST) {
289 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
291 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
292 NULL, newskb->dev, ip_dev_loopback_xmit);
295 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
296 skb->dev, ip_finish_output,
297 !(IPCB(skb)->flags & IPSKB_REROUTED));
300 int ip_output(struct sk_buff *skb)
302 struct net_device *dev = skb_dst(skb)->dev;
304 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
307 skb->protocol = htons(ETH_P_IP);
309 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
311 !(IPCB(skb)->flags & IPSKB_REROUTED));
314 int ip_queue_xmit(struct sk_buff *skb)
316 struct sock *sk = skb->sk;
317 struct inet_sock *inet = inet_sk(sk);
318 struct ip_options *opt = inet->opt;
323 /* Skip all of this if the packet is already routed,
324 * f.e. by something like SCTP.
327 rt = skb_rtable(skb);
331 /* Make sure we can route this packet. */
332 rt = (struct rtable *)__sk_dst_check(sk, 0);
336 /* Use correct destination address if we have options. */
337 daddr = inet->inet_daddr;
342 struct flowi fl = { .oif = sk->sk_bound_dev_if,
346 .saddr = inet->inet_saddr,
347 .tos = RT_CONN_FLAGS(sk) } },
348 .proto = sk->sk_protocol,
349 .flags = inet_sk_flowi_flags(sk),
351 { .sport = inet->inet_sport,
352 .dport = inet->inet_dport } } };
354 /* If this fails, retransmit mechanism of transport layer will
355 * keep trying until route appears or the connection times
358 security_sk_classify_flow(sk, &fl);
359 if (ip_route_output_flow(sock_net(sk), &rt, &fl, sk, 0))
362 sk_setup_caps(sk, &rt->dst);
364 skb_dst_set_noref(skb, &rt->dst);
367 if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
370 /* OK, we know where to send it, allocate and build IP header. */
371 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
372 skb_reset_network_header(skb);
374 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
375 if (ip_dont_fragment(sk, &rt->dst) && !skb->local_df)
376 iph->frag_off = htons(IP_DF);
379 iph->ttl = ip_select_ttl(inet, &rt->dst);
380 iph->protocol = sk->sk_protocol;
381 iph->saddr = rt->rt_src;
382 iph->daddr = rt->rt_dst;
383 /* Transport layer set skb->h.foo itself. */
385 if (opt && opt->optlen) {
386 iph->ihl += opt->optlen >> 2;
387 ip_options_build(skb, opt, inet->inet_daddr, rt, 0);
390 ip_select_ident_more(iph, &rt->dst, sk,
391 (skb_shinfo(skb)->gso_segs ?: 1) - 1);
393 skb->priority = sk->sk_priority;
394 skb->mark = sk->sk_mark;
396 res = ip_local_out(skb);
402 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
404 return -EHOSTUNREACH;
408 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
410 to->pkt_type = from->pkt_type;
411 to->priority = from->priority;
412 to->protocol = from->protocol;
414 skb_dst_copy(to, from);
416 to->mark = from->mark;
418 /* Copy the flags to each fragment. */
419 IPCB(to)->flags = IPCB(from)->flags;
421 #ifdef CONFIG_NET_SCHED
422 to->tc_index = from->tc_index;
425 #if defined(CONFIG_NETFILTER_XT_TARGET_TRACE) || \
426 defined(CONFIG_NETFILTER_XT_TARGET_TRACE_MODULE)
427 to->nf_trace = from->nf_trace;
429 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
430 to->ipvs_property = from->ipvs_property;
432 skb_copy_secmark(to, from);
436 * This IP datagram is too large to be sent in one piece. Break it up into
437 * smaller pieces (each of size equal to IP header plus
438 * a block of the data of the original IP data part) that will yet fit in a
439 * single device frame, and queue such a frame for sending.
442 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
446 struct net_device *dev;
447 struct sk_buff *skb2;
448 unsigned int mtu, hlen, left, len, ll_rs, pad;
450 __be16 not_last_frag;
451 struct rtable *rt = skb_rtable(skb);
457 * Point into the IP datagram header.
462 if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
463 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
464 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
465 htonl(ip_skb_dst_mtu(skb)));
471 * Setup starting values.
475 mtu = dst_mtu(&rt->dst) - hlen; /* Size of data space */
476 #ifdef CONFIG_BRIDGE_NETFILTER
478 mtu -= nf_bridge_mtu_reduction(skb);
480 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
482 /* When frag_list is given, use it. First, check its validity:
483 * some transformers could create wrong frag_list or break existing
484 * one, it is not prohibited. In this case fall back to copying.
486 * LATER: this step can be merged to real generation of fragments,
487 * we can switch to copy when see the first bad fragment.
489 if (skb_has_frags(skb)) {
490 struct sk_buff *frag;
491 int first_len = skb_pagelen(skb);
494 if (first_len - hlen > mtu ||
495 ((first_len - hlen) & 7) ||
496 (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
500 skb_walk_frags(skb, frag) {
501 /* Correct geometry. */
502 if (frag->len > mtu ||
503 ((frag->len & 7) && frag->next) ||
504 skb_headroom(frag) < hlen)
507 /* Partially cloned skb? */
508 if (skb_shared(frag))
514 frag->destructor = sock_wfree;
516 truesizes += frag->truesize;
519 /* Everything is OK. Generate! */
523 frag = skb_shinfo(skb)->frag_list;
524 skb_frag_list_init(skb);
525 skb->data_len = first_len - skb_headlen(skb);
526 skb->truesize -= truesizes;
527 skb->len = first_len;
528 iph->tot_len = htons(first_len);
529 iph->frag_off = htons(IP_MF);
533 /* Prepare header of the next frame,
534 * before previous one went down. */
536 frag->ip_summed = CHECKSUM_NONE;
537 skb_reset_transport_header(frag);
538 __skb_push(frag, hlen);
539 skb_reset_network_header(frag);
540 memcpy(skb_network_header(frag), iph, hlen);
542 iph->tot_len = htons(frag->len);
543 ip_copy_metadata(frag, skb);
545 ip_options_fragment(frag);
546 offset += skb->len - hlen;
547 iph->frag_off = htons(offset>>3);
548 if (frag->next != NULL)
549 iph->frag_off |= htons(IP_MF);
550 /* Ready, complete checksum */
557 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
567 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
576 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
581 left = skb->len - hlen; /* Space per frame */
582 ptr = hlen; /* Where to start from */
584 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
585 * we need to make room for the encapsulating header
587 pad = nf_bridge_pad(skb);
588 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, pad);
592 * Fragment the datagram.
595 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
596 not_last_frag = iph->frag_off & htons(IP_MF);
599 * Keep copying data until we run out.
604 /* IF: it doesn't fit, use 'mtu' - the data space left */
607 /* IF: we are not sending upto and including the packet end
608 then align the next start on an eight byte boundary */
616 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
617 NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
623 * Set up data on packet
626 ip_copy_metadata(skb2, skb);
627 skb_reserve(skb2, ll_rs);
628 skb_put(skb2, len + hlen);
629 skb_reset_network_header(skb2);
630 skb2->transport_header = skb2->network_header + hlen;
633 * Charge the memory for the fragment to any owner
638 skb_set_owner_w(skb2, skb->sk);
641 * Copy the packet header into the new buffer.
644 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
647 * Copy a block of the IP datagram.
649 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
654 * Fill in the new header fields.
657 iph->frag_off = htons((offset >> 3));
659 /* ANK: dirty, but effective trick. Upgrade options only if
660 * the segment to be fragmented was THE FIRST (otherwise,
661 * options are already fixed) and make it ONCE
662 * on the initial skb, so that all the following fragments
663 * will inherit fixed options.
666 ip_options_fragment(skb);
669 * Added AC : If we are fragmenting a fragment that's not the
670 * last fragment then keep MF on each bit
672 if (left > 0 || not_last_frag)
673 iph->frag_off |= htons(IP_MF);
678 * Put this fragment into the sending queue.
680 iph->tot_len = htons(len + hlen);
688 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
691 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
696 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
700 EXPORT_SYMBOL(ip_fragment);
703 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
705 struct iovec *iov = from;
707 if (skb->ip_summed == CHECKSUM_PARTIAL) {
708 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
712 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
714 skb->csum = csum_block_add(skb->csum, csum, odd);
720 csum_page(struct page *page, int offset, int copy)
725 csum = csum_partial(kaddr + offset, copy, 0);
730 static inline int ip_ufo_append_data(struct sock *sk,
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(&sk->sk_write_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;
765 sk->sk_sndmsg_off = 0;
767 /* specify the length of each IP datagram fragment */
768 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
769 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
770 __skb_queue_tail(&sk->sk_write_queue, skb);
773 return skb_append_datato_frags(sk, skb, getfrag, from,
774 (length - transhdrlen));
778 * ip_append_data() and ip_append_page() can make one large IP datagram
779 * from many pieces of data. Each pieces will be holded on the socket
780 * until ip_push_pending_frames() is called. Each piece can be a page
783 * Not only UDP, other transport protocols - e.g. raw sockets - can use
784 * this interface potentially.
786 * LATER: length must be adjusted by pad at tail, when it is required.
788 int ip_append_data(struct sock *sk,
789 int getfrag(void *from, char *to, int offset, int len,
790 int odd, struct sk_buff *skb),
791 void *from, int length, int transhdrlen,
792 struct ipcm_cookie *ipc, struct rtable **rtp,
795 struct inet_sock *inet = inet_sk(sk);
798 struct ip_options *opt = NULL;
805 unsigned int maxfraglen, fragheaderlen;
806 int csummode = CHECKSUM_NONE;
812 if (skb_queue_empty(&sk->sk_write_queue)) {
818 if (inet->cork.opt == NULL) {
819 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
820 if (unlikely(inet->cork.opt == NULL))
823 memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
824 inet->cork.flags |= IPCORK_OPT;
825 inet->cork.addr = ipc->addr;
831 * We steal reference to this route, caller should not release it
834 inet->cork.fragsize = mtu = inet->pmtudisc == IP_PMTUDISC_PROBE ?
836 dst_mtu(rt->dst.path);
837 inet->cork.dst = &rt->dst;
838 inet->cork.length = 0;
839 sk->sk_sndmsg_page = NULL;
840 sk->sk_sndmsg_off = 0;
841 if ((exthdrlen = rt->dst.header_len) != 0) {
843 transhdrlen += exthdrlen;
846 rt = (struct rtable *)inet->cork.dst;
847 if (inet->cork.flags & IPCORK_OPT)
848 opt = inet->cork.opt;
852 mtu = inet->cork.fragsize;
854 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
856 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
857 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
859 if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
860 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport,
866 * transhdrlen > 0 means that this is the first fragment and we wish
867 * it won't be fragmented in the future.
870 length + fragheaderlen <= mtu &&
871 rt->dst.dev->features & NETIF_F_V4_CSUM &&
873 csummode = CHECKSUM_PARTIAL;
875 skb = skb_peek_tail(&sk->sk_write_queue);
877 inet->cork.length += length;
878 if (((length > mtu) || (skb && skb_is_gso(skb))) &&
879 (sk->sk_protocol == IPPROTO_UDP) &&
880 (rt->dst.dev->features & NETIF_F_UFO)) {
881 err = ip_ufo_append_data(sk, getfrag, from, length, hh_len,
882 fragheaderlen, transhdrlen, mtu,
889 /* So, what's going on in the loop below?
891 * We use calculated fragment length to generate chained skb,
892 * each of segments is IP fragment ready for sending to network after
893 * adding appropriate IP header.
900 /* Check if the remaining data fits into current packet. */
901 copy = mtu - skb->len;
903 copy = maxfraglen - skb->len;
906 unsigned int datalen;
907 unsigned int fraglen;
908 unsigned int fraggap;
909 unsigned int alloclen;
910 struct sk_buff *skb_prev;
914 fraggap = skb_prev->len - maxfraglen;
919 * If remaining data exceeds the mtu,
920 * we know we need more fragment(s).
922 datalen = length + fraggap;
923 if (datalen > mtu - fragheaderlen)
924 datalen = maxfraglen - fragheaderlen;
925 fraglen = datalen + fragheaderlen;
927 if ((flags & MSG_MORE) &&
928 !(rt->dst.dev->features&NETIF_F_SG))
931 alloclen = datalen + fragheaderlen;
933 /* The last fragment gets additional space at tail.
934 * Note, with MSG_MORE we overallocate on fragments,
935 * because we have no idea what fragment will be
938 if (datalen == length + fraggap)
939 alloclen += rt->dst.trailer_len;
942 skb = sock_alloc_send_skb(sk,
943 alloclen + hh_len + 15,
944 (flags & MSG_DONTWAIT), &err);
947 if (atomic_read(&sk->sk_wmem_alloc) <=
949 skb = sock_wmalloc(sk,
950 alloclen + hh_len + 15, 1,
952 if (unlikely(skb == NULL))
955 /* only the initial fragment is
963 * Fill in the control structures
965 skb->ip_summed = csummode;
967 skb_reserve(skb, hh_len);
968 *skb_tx(skb) = ipc->shtx;
971 * Find where to start putting bytes.
973 data = skb_put(skb, fraglen);
974 skb_set_network_header(skb, exthdrlen);
975 skb->transport_header = (skb->network_header +
977 data += fragheaderlen;
980 skb->csum = skb_copy_and_csum_bits(
981 skb_prev, maxfraglen,
982 data + transhdrlen, fraggap, 0);
983 skb_prev->csum = csum_sub(skb_prev->csum,
986 pskb_trim_unique(skb_prev, maxfraglen);
989 copy = datalen - transhdrlen - fraggap;
990 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
997 length -= datalen - fraggap;
1000 csummode = CHECKSUM_NONE;
1003 * Put the packet on the pending queue.
1005 __skb_queue_tail(&sk->sk_write_queue, skb);
1012 if (!(rt->dst.dev->features&NETIF_F_SG)) {
1016 if (getfrag(from, skb_put(skb, copy),
1017 offset, copy, off, skb) < 0) {
1018 __skb_trim(skb, off);
1023 int i = skb_shinfo(skb)->nr_frags;
1024 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
1025 struct page *page = sk->sk_sndmsg_page;
1026 int off = sk->sk_sndmsg_off;
1029 if (page && (left = PAGE_SIZE - off) > 0) {
1032 if (page != frag->page) {
1033 if (i == MAX_SKB_FRAGS) {
1038 skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
1039 frag = &skb_shinfo(skb)->frags[i];
1041 } else if (i < MAX_SKB_FRAGS) {
1042 if (copy > PAGE_SIZE)
1044 page = alloc_pages(sk->sk_allocation, 0);
1049 sk->sk_sndmsg_page = page;
1050 sk->sk_sndmsg_off = 0;
1052 skb_fill_page_desc(skb, i, page, 0, 0);
1053 frag = &skb_shinfo(skb)->frags[i];
1058 if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1062 sk->sk_sndmsg_off += copy;
1065 skb->data_len += copy;
1066 skb->truesize += copy;
1067 atomic_add(copy, &sk->sk_wmem_alloc);
1076 inet->cork.length -= length;
1077 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1081 ssize_t ip_append_page(struct sock *sk, struct page *page,
1082 int offset, size_t size, int flags)
1084 struct inet_sock *inet = inet_sk(sk);
1085 struct sk_buff *skb;
1087 struct ip_options *opt = NULL;
1092 unsigned int maxfraglen, fragheaderlen, fraggap;
1097 if (flags&MSG_PROBE)
1100 if (skb_queue_empty(&sk->sk_write_queue))
1103 rt = (struct rtable *)inet->cork.dst;
1104 if (inet->cork.flags & IPCORK_OPT)
1105 opt = inet->cork.opt;
1107 if (!(rt->dst.dev->features&NETIF_F_SG))
1110 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1111 mtu = inet->cork.fragsize;
1113 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1114 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1116 if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1117 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport, mtu);
1121 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1124 inet->cork.length += size;
1125 if ((size + skb->len > mtu) &&
1126 (sk->sk_protocol == IPPROTO_UDP) &&
1127 (rt->dst.dev->features & NETIF_F_UFO)) {
1128 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1129 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1136 if (skb_is_gso(skb))
1140 /* Check if the remaining data fits into current packet. */
1141 len = mtu - skb->len;
1143 len = maxfraglen - skb->len;
1146 struct sk_buff *skb_prev;
1150 fraggap = skb_prev->len - maxfraglen;
1152 alloclen = fragheaderlen + hh_len + fraggap + 15;
1153 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1154 if (unlikely(!skb)) {
1160 * Fill in the control structures
1162 skb->ip_summed = CHECKSUM_NONE;
1164 skb_reserve(skb, hh_len);
1167 * Find where to start putting bytes.
1169 skb_put(skb, fragheaderlen + fraggap);
1170 skb_reset_network_header(skb);
1171 skb->transport_header = (skb->network_header +
1174 skb->csum = skb_copy_and_csum_bits(skb_prev,
1176 skb_transport_header(skb),
1178 skb_prev->csum = csum_sub(skb_prev->csum,
1180 pskb_trim_unique(skb_prev, maxfraglen);
1184 * Put the packet on the pending queue.
1186 __skb_queue_tail(&sk->sk_write_queue, skb);
1190 i = skb_shinfo(skb)->nr_frags;
1193 if (skb_can_coalesce(skb, i, page, offset)) {
1194 skb_shinfo(skb)->frags[i-1].size += len;
1195 } else if (i < MAX_SKB_FRAGS) {
1197 skb_fill_page_desc(skb, i, page, offset, len);
1203 if (skb->ip_summed == CHECKSUM_NONE) {
1205 csum = csum_page(page, offset, len);
1206 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1210 skb->data_len += len;
1211 skb->truesize += len;
1212 atomic_add(len, &sk->sk_wmem_alloc);
1219 inet->cork.length -= size;
1220 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1224 static void ip_cork_release(struct inet_sock *inet)
1226 inet->cork.flags &= ~IPCORK_OPT;
1227 kfree(inet->cork.opt);
1228 inet->cork.opt = NULL;
1229 dst_release(inet->cork.dst);
1230 inet->cork.dst = NULL;
1234 * Combined all pending IP fragments on the socket as one IP datagram
1235 * and push them out.
1237 int ip_push_pending_frames(struct sock *sk)
1239 struct sk_buff *skb, *tmp_skb;
1240 struct sk_buff **tail_skb;
1241 struct inet_sock *inet = inet_sk(sk);
1242 struct net *net = sock_net(sk);
1243 struct ip_options *opt = NULL;
1244 struct rtable *rt = (struct rtable *)inet->cork.dst;
1250 if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1252 tail_skb = &(skb_shinfo(skb)->frag_list);
1254 /* move skb->data to ip header from ext header */
1255 if (skb->data < skb_network_header(skb))
1256 __skb_pull(skb, skb_network_offset(skb));
1257 while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1258 __skb_pull(tmp_skb, skb_network_header_len(skb));
1259 *tail_skb = tmp_skb;
1260 tail_skb = &(tmp_skb->next);
1261 skb->len += tmp_skb->len;
1262 skb->data_len += tmp_skb->len;
1263 skb->truesize += tmp_skb->truesize;
1264 tmp_skb->destructor = NULL;
1268 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1269 * to fragment the frame generated here. No matter, what transforms
1270 * how transforms change size of the packet, it will come out.
1272 if (inet->pmtudisc < IP_PMTUDISC_DO)
1275 /* DF bit is set when we want to see DF on outgoing frames.
1276 * If local_df is set too, we still allow to fragment this frame
1278 if (inet->pmtudisc >= IP_PMTUDISC_DO ||
1279 (skb->len <= dst_mtu(&rt->dst) &&
1280 ip_dont_fragment(sk, &rt->dst)))
1283 if (inet->cork.flags & IPCORK_OPT)
1284 opt = inet->cork.opt;
1286 if (rt->rt_type == RTN_MULTICAST)
1289 ttl = ip_select_ttl(inet, &rt->dst);
1291 iph = (struct iphdr *)skb->data;
1295 iph->ihl += opt->optlen>>2;
1296 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1298 iph->tos = inet->tos;
1300 ip_select_ident(iph, &rt->dst, sk);
1302 iph->protocol = sk->sk_protocol;
1303 iph->saddr = rt->rt_src;
1304 iph->daddr = rt->rt_dst;
1306 skb->priority = sk->sk_priority;
1307 skb->mark = sk->sk_mark;
1309 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1312 inet->cork.dst = NULL;
1313 skb_dst_set(skb, &rt->dst);
1315 if (iph->protocol == IPPROTO_ICMP)
1316 icmp_out_count(net, ((struct icmphdr *)
1317 skb_transport_header(skb))->type);
1319 /* Netfilter gets whole the not fragmented skb. */
1320 err = ip_local_out(skb);
1323 err = net_xmit_errno(err);
1329 ip_cork_release(inet);
1333 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1338 * Throw away all pending data on the socket.
1340 void ip_flush_pending_frames(struct sock *sk)
1342 struct sk_buff *skb;
1344 while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1347 ip_cork_release(inet_sk(sk));
1352 * Fetch data from kernel space and fill in checksum if needed.
1354 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1355 int len, int odd, struct sk_buff *skb)
1359 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1360 skb->csum = csum_block_add(skb->csum, csum, odd);
1365 * Generic function to send a packet as reply to another packet.
1366 * Used to send TCP resets so far. ICMP should use this function too.
1368 * Should run single threaded per socket because it uses the sock
1369 * structure to pass arguments.
1371 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1374 struct inet_sock *inet = inet_sk(sk);
1376 struct ip_options opt;
1379 struct ipcm_cookie ipc;
1381 struct rtable *rt = skb_rtable(skb);
1383 if (ip_options_echo(&replyopts.opt, skb))
1386 daddr = ipc.addr = rt->rt_src;
1390 if (replyopts.opt.optlen) {
1391 ipc.opt = &replyopts.opt;
1394 daddr = replyopts.opt.faddr;
1398 struct flowi fl = { .oif = arg->bound_dev_if,
1401 .saddr = rt->rt_spec_dst,
1402 .tos = RT_TOS(ip_hdr(skb)->tos) } },
1403 /* Not quite clean, but right. */
1405 { .sport = tcp_hdr(skb)->dest,
1406 .dport = tcp_hdr(skb)->source } },
1407 .proto = sk->sk_protocol,
1408 .flags = ip_reply_arg_flowi_flags(arg) };
1409 security_skb_classify_flow(skb, &fl);
1410 if (ip_route_output_key(sock_net(sk), &rt, &fl))
1414 /* And let IP do all the hard work.
1416 This chunk is not reenterable, hence spinlock.
1417 Note that it uses the fact, that this function is called
1418 with locally disabled BH and that sk cannot be already spinlocked.
1421 inet->tos = ip_hdr(skb)->tos;
1422 sk->sk_priority = skb->priority;
1423 sk->sk_protocol = ip_hdr(skb)->protocol;
1424 sk->sk_bound_dev_if = arg->bound_dev_if;
1425 ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1426 &ipc, &rt, MSG_DONTWAIT);
1427 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1428 if (arg->csumoffset >= 0)
1429 *((__sum16 *)skb_transport_header(skb) +
1430 arg->csumoffset) = csum_fold(csum_add(skb->csum,
1432 skb->ip_summed = CHECKSUM_NONE;
1433 ip_push_pending_frames(sk);
1441 void __init ip_init(void)
1446 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1447 igmp_mc_proc_init();
1451 EXPORT_SYMBOL(ip_generic_getfrag);
1452 EXPORT_SYMBOL(ip_queue_xmit);
1453 EXPORT_SYMBOL(ip_send_check);