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>
55 #include <linux/socket.h>
56 #include <linux/sockios.h>
58 #include <linux/inet.h>
59 #include <linux/netdevice.h>
60 #include <linux/etherdevice.h>
61 #include <linux/proc_fs.h>
62 #include <linux/stat.h>
63 #include <linux/init.h>
67 #include <net/protocol.h>
68 #include <net/route.h>
70 #include <linux/skbuff.h>
74 #include <net/checksum.h>
75 #include <net/inetpeer.h>
76 #include <linux/igmp.h>
77 #include <linux/netfilter_ipv4.h>
78 #include <linux/netfilter_bridge.h>
79 #include <linux/mroute.h>
80 #include <linux/netlink.h>
81 #include <linux/tcp.h>
83 int sysctl_ip_default_ttl __read_mostly = IPDEFTTL;
85 /* Generate a checksum for an outgoing IP datagram. */
86 __inline__ void ip_send_check(struct iphdr *iph)
89 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
92 int __ip_local_out(struct sk_buff *skb)
94 struct iphdr *iph = ip_hdr(skb);
96 iph->tot_len = htons(skb->len);
98 return nf_hook(PF_INET, NF_INET_LOCAL_OUT, skb, NULL, skb_dst(skb)->dev,
102 int ip_local_out(struct sk_buff *skb)
106 err = __ip_local_out(skb);
107 if (likely(err == 1))
108 err = dst_output(skb);
112 EXPORT_SYMBOL_GPL(ip_local_out);
114 /* dev_loopback_xmit for use with netfilter. */
115 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
117 skb_reset_mac_header(newskb);
118 __skb_pull(newskb, skb_network_offset(newskb));
119 newskb->pkt_type = PACKET_LOOPBACK;
120 newskb->ip_summed = CHECKSUM_UNNECESSARY;
121 WARN_ON(!skb_dst(newskb));
126 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
128 int ttl = inet->uc_ttl;
131 ttl = dst_metric(dst, RTAX_HOPLIMIT);
136 * Add an ip header to a skbuff and send it out.
139 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
140 __be32 saddr, __be32 daddr, struct ip_options *opt)
142 struct inet_sock *inet = inet_sk(sk);
143 struct rtable *rt = skb_rtable(skb);
146 /* Build the IP header. */
147 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
148 skb_reset_network_header(skb);
152 iph->tos = inet->tos;
153 if (ip_dont_fragment(sk, &rt->u.dst))
154 iph->frag_off = htons(IP_DF);
157 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
158 iph->daddr = rt->rt_dst;
159 iph->saddr = rt->rt_src;
160 iph->protocol = sk->sk_protocol;
161 ip_select_ident(iph, &rt->u.dst, sk);
163 if (opt && opt->optlen) {
164 iph->ihl += opt->optlen>>2;
165 ip_options_build(skb, opt, daddr, rt, 0);
168 skb->priority = sk->sk_priority;
169 skb->mark = sk->sk_mark;
172 return ip_local_out(skb);
175 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
177 static inline int ip_finish_output2(struct sk_buff *skb)
179 struct dst_entry *dst = skb_dst(skb);
180 struct rtable *rt = (struct rtable *)dst;
181 struct net_device *dev = dst->dev;
182 unsigned int hh_len = LL_RESERVED_SPACE(dev);
184 if (rt->rt_type == RTN_MULTICAST) {
185 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
186 } else if (rt->rt_type == RTN_BROADCAST)
187 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
189 /* Be paranoid, rather than too clever. */
190 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
191 struct sk_buff *skb2;
193 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
199 skb_set_owner_w(skb2, skb->sk);
205 return neigh_hh_output(dst->hh, skb);
206 else if (dst->neighbour)
207 return dst->neighbour->output(skb);
210 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
215 static inline int ip_skb_dst_mtu(struct sk_buff *skb)
217 struct inet_sock *inet = skb->sk ? inet_sk(skb->sk) : NULL;
219 return (inet && inet->pmtudisc == IP_PMTUDISC_PROBE) ?
220 skb_dst(skb)->dev->mtu : dst_mtu(skb_dst(skb));
223 static int ip_finish_output(struct sk_buff *skb)
225 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
226 /* Policy lookup after SNAT yielded a new policy */
227 if (skb_dst(skb)->xfrm != NULL) {
228 IPCB(skb)->flags |= IPSKB_REROUTED;
229 return dst_output(skb);
232 if (skb->len > ip_skb_dst_mtu(skb) && !skb_is_gso(skb))
233 return ip_fragment(skb, ip_finish_output2);
235 return ip_finish_output2(skb);
238 int ip_mc_output(struct sk_buff *skb)
240 struct sock *sk = skb->sk;
241 struct rtable *rt = skb_rtable(skb);
242 struct net_device *dev = rt->u.dst.dev;
245 * If the indicated interface is up and running, send the packet.
247 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
250 skb->protocol = htons(ETH_P_IP);
253 * Multicasts are looped back for other local users
256 if (rt->rt_flags&RTCF_MULTICAST) {
257 if ((!sk || inet_sk(sk)->mc_loop)
258 #ifdef CONFIG_IP_MROUTE
259 /* Small optimization: do not loopback not local frames,
260 which returned after forwarding; they will be dropped
261 by ip_mr_input in any case.
262 Note, that local frames are looped back to be delivered
265 This check is duplicated in ip_mr_input at the moment.
268 ((rt->rt_flags & RTCF_LOCAL) ||
269 !(IPCB(skb)->flags & IPSKB_FORWARDED))
272 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
274 NF_HOOK(PF_INET, NF_INET_POST_ROUTING, newskb,
276 ip_dev_loopback_xmit);
279 /* Multicasts with ttl 0 must not go beyond the host */
281 if (ip_hdr(skb)->ttl == 0) {
287 if (rt->rt_flags&RTCF_BROADCAST) {
288 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
290 NF_HOOK(PF_INET, NF_INET_POST_ROUTING, newskb, NULL,
291 newskb->dev, ip_dev_loopback_xmit);
294 return NF_HOOK_COND(PF_INET, NF_INET_POST_ROUTING, skb, NULL, skb->dev,
296 !(IPCB(skb)->flags & IPSKB_REROUTED));
299 int ip_output(struct sk_buff *skb)
301 struct net_device *dev = skb_dst(skb)->dev;
303 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
306 skb->protocol = htons(ETH_P_IP);
308 return NF_HOOK_COND(PF_INET, NF_INET_POST_ROUTING, skb, NULL, dev,
310 !(IPCB(skb)->flags & IPSKB_REROUTED));
313 int ip_queue_xmit(struct sk_buff *skb, int ipfragok)
315 struct sock *sk = skb->sk;
316 struct inet_sock *inet = inet_sk(sk);
317 struct ip_options *opt = inet->opt;
321 /* Skip all of this if the packet is already routed,
322 * f.e. by something like SCTP.
324 rt = skb_rtable(skb);
328 /* Make sure we can route this packet. */
329 rt = (struct rtable *)__sk_dst_check(sk, 0);
333 /* Use correct destination address if we have options. */
334 daddr = inet->inet_daddr;
339 struct flowi fl = { .oif = sk->sk_bound_dev_if,
343 .saddr = inet->inet_saddr,
344 .tos = RT_CONN_FLAGS(sk) } },
345 .proto = sk->sk_protocol,
346 .flags = inet_sk_flowi_flags(sk),
348 { .sport = inet->inet_sport,
349 .dport = inet->inet_dport } } };
351 /* If this fails, retransmit mechanism of transport layer will
352 * keep trying until route appears or the connection times
355 security_sk_classify_flow(sk, &fl);
356 if (ip_route_output_flow(sock_net(sk), &rt, &fl, sk, 0))
359 sk_setup_caps(sk, &rt->u.dst);
361 skb_dst_set(skb, dst_clone(&rt->u.dst));
364 if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
367 /* OK, we know where to send it, allocate and build IP header. */
368 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
369 skb_reset_network_header(skb);
371 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
372 if (ip_dont_fragment(sk, &rt->u.dst) && !ipfragok)
373 iph->frag_off = htons(IP_DF);
376 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
377 iph->protocol = sk->sk_protocol;
378 iph->saddr = rt->rt_src;
379 iph->daddr = rt->rt_dst;
380 /* Transport layer set skb->h.foo itself. */
382 if (opt && opt->optlen) {
383 iph->ihl += opt->optlen >> 2;
384 ip_options_build(skb, opt, inet->inet_daddr, rt, 0);
387 ip_select_ident_more(iph, &rt->u.dst, sk,
388 (skb_shinfo(skb)->gso_segs ?: 1) - 1);
390 skb->priority = sk->sk_priority;
391 skb->mark = sk->sk_mark;
393 return ip_local_out(skb);
396 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
398 return -EHOSTUNREACH;
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_set(to, dst_clone(skb_dst(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 *))
441 struct net_device *dev;
442 struct sk_buff *skb2;
443 unsigned int mtu, hlen, left, len, ll_rs, pad;
445 __be16 not_last_frag;
446 struct rtable *rt = skb_rtable(skb);
452 * Point into the IP datagram header.
457 if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
458 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
459 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
460 htonl(ip_skb_dst_mtu(skb)));
466 * Setup starting values.
470 mtu = dst_mtu(&rt->u.dst) - hlen; /* Size of data space */
471 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
473 /* When frag_list is given, use it. First, check its validity:
474 * some transformers could create wrong frag_list or break existing
475 * one, it is not prohibited. In this case fall back to copying.
477 * LATER: this step can be merged to real generation of fragments,
478 * we can switch to copy when see the first bad fragment.
480 if (skb_has_frags(skb)) {
481 struct sk_buff *frag;
482 int first_len = skb_pagelen(skb);
485 if (first_len - hlen > mtu ||
486 ((first_len - hlen) & 7) ||
487 (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
491 skb_walk_frags(skb, frag) {
492 /* Correct geometry. */
493 if (frag->len > mtu ||
494 ((frag->len & 7) && frag->next) ||
495 skb_headroom(frag) < hlen)
498 /* Partially cloned skb? */
499 if (skb_shared(frag))
505 frag->destructor = sock_wfree;
507 truesizes += frag->truesize;
510 /* Everything is OK. Generate! */
514 frag = skb_shinfo(skb)->frag_list;
515 skb_frag_list_init(skb);
516 skb->data_len = first_len - skb_headlen(skb);
517 skb->truesize -= truesizes;
518 skb->len = first_len;
519 iph->tot_len = htons(first_len);
520 iph->frag_off = htons(IP_MF);
524 /* Prepare header of the next frame,
525 * before previous one went down. */
527 frag->ip_summed = CHECKSUM_NONE;
528 skb_reset_transport_header(frag);
529 __skb_push(frag, hlen);
530 skb_reset_network_header(frag);
531 memcpy(skb_network_header(frag), iph, hlen);
533 iph->tot_len = htons(frag->len);
534 ip_copy_metadata(frag, skb);
536 ip_options_fragment(frag);
537 offset += skb->len - hlen;
538 iph->frag_off = htons(offset>>3);
539 if (frag->next != NULL)
540 iph->frag_off |= htons(IP_MF);
541 /* Ready, complete checksum */
548 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
558 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
567 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
572 left = skb->len - hlen; /* Space per frame */
573 ptr = raw + hlen; /* Where to start from */
575 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
576 * we need to make room for the encapsulating header
578 pad = nf_bridge_pad(skb);
579 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, pad);
583 * Fragment the datagram.
586 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
587 not_last_frag = iph->frag_off & htons(IP_MF);
590 * Keep copying data until we run out.
595 /* IF: it doesn't fit, use 'mtu' - the data space left */
598 /* IF: we are not sending upto and including the packet end
599 then align the next start on an eight byte boundary */
607 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
608 NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
614 * Set up data on packet
617 ip_copy_metadata(skb2, skb);
618 skb_reserve(skb2, ll_rs);
619 skb_put(skb2, len + hlen);
620 skb_reset_network_header(skb2);
621 skb2->transport_header = skb2->network_header + hlen;
624 * Charge the memory for the fragment to any owner
629 skb_set_owner_w(skb2, skb->sk);
632 * Copy the packet header into the new buffer.
635 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
638 * Copy a block of the IP datagram.
640 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
645 * Fill in the new header fields.
648 iph->frag_off = htons((offset >> 3));
650 /* ANK: dirty, but effective trick. Upgrade options only if
651 * the segment to be fragmented was THE FIRST (otherwise,
652 * options are already fixed) and make it ONCE
653 * on the initial skb, so that all the following fragments
654 * will inherit fixed options.
657 ip_options_fragment(skb);
660 * Added AC : If we are fragmenting a fragment that's not the
661 * last fragment then keep MF on each bit
663 if (left > 0 || not_last_frag)
664 iph->frag_off |= htons(IP_MF);
669 * Put this fragment into the sending queue.
671 iph->tot_len = htons(len + hlen);
679 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
682 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
687 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
691 EXPORT_SYMBOL(ip_fragment);
694 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
696 struct iovec *iov = from;
698 if (skb->ip_summed == CHECKSUM_PARTIAL) {
699 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
703 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
705 skb->csum = csum_block_add(skb->csum, csum, odd);
711 csum_page(struct page *page, int offset, int copy)
716 csum = csum_partial(kaddr + offset, copy, 0);
721 static inline int ip_ufo_append_data(struct sock *sk,
722 int getfrag(void *from, char *to, int offset, int len,
723 int odd, struct sk_buff *skb),
724 void *from, int length, int hh_len, int fragheaderlen,
725 int transhdrlen, int mtu, unsigned int flags)
730 /* There is support for UDP fragmentation offload by network
731 * device, so create one single skb packet containing complete
734 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL) {
735 skb = sock_alloc_send_skb(sk,
736 hh_len + fragheaderlen + transhdrlen + 20,
737 (flags & MSG_DONTWAIT), &err);
742 /* reserve space for Hardware header */
743 skb_reserve(skb, hh_len);
745 /* create space for UDP/IP header */
746 skb_put(skb, fragheaderlen + transhdrlen);
748 /* initialize network header pointer */
749 skb_reset_network_header(skb);
751 /* initialize protocol header pointer */
752 skb->transport_header = skb->network_header + fragheaderlen;
754 skb->ip_summed = CHECKSUM_PARTIAL;
756 sk->sk_sndmsg_off = 0;
758 /* specify the length of each IP datagram fragment */
759 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
760 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
761 __skb_queue_tail(&sk->sk_write_queue, skb);
764 return skb_append_datato_frags(sk, skb, getfrag, from,
765 (length - transhdrlen));
769 * ip_append_data() and ip_append_page() can make one large IP datagram
770 * from many pieces of data. Each pieces will be holded on the socket
771 * until ip_push_pending_frames() is called. Each piece can be a page
774 * Not only UDP, other transport protocols - e.g. raw sockets - can use
775 * this interface potentially.
777 * LATER: length must be adjusted by pad at tail, when it is required.
779 int ip_append_data(struct sock *sk,
780 int getfrag(void *from, char *to, int offset, int len,
781 int odd, struct sk_buff *skb),
782 void *from, int length, int transhdrlen,
783 struct ipcm_cookie *ipc, struct rtable **rtp,
786 struct inet_sock *inet = inet_sk(sk);
789 struct ip_options *opt = NULL;
796 unsigned int maxfraglen, fragheaderlen;
797 int csummode = CHECKSUM_NONE;
803 if (skb_queue_empty(&sk->sk_write_queue)) {
809 if (inet->cork.opt == NULL) {
810 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
811 if (unlikely(inet->cork.opt == NULL))
814 memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
815 inet->cork.flags |= IPCORK_OPT;
816 inet->cork.addr = ipc->addr;
822 * We steal reference to this route, caller should not release it
825 inet->cork.fragsize = mtu = inet->pmtudisc == IP_PMTUDISC_PROBE ?
827 dst_mtu(rt->u.dst.path);
828 inet->cork.dst = &rt->u.dst;
829 inet->cork.length = 0;
830 sk->sk_sndmsg_page = NULL;
831 sk->sk_sndmsg_off = 0;
832 if ((exthdrlen = rt->u.dst.header_len) != 0) {
834 transhdrlen += exthdrlen;
837 rt = (struct rtable *)inet->cork.dst;
838 if (inet->cork.flags & IPCORK_OPT)
839 opt = inet->cork.opt;
843 mtu = inet->cork.fragsize;
845 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
847 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
848 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
850 if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
851 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport,
857 * transhdrlen > 0 means that this is the first fragment and we wish
858 * it won't be fragmented in the future.
861 length + fragheaderlen <= mtu &&
862 rt->u.dst.dev->features & NETIF_F_V4_CSUM &&
864 csummode = CHECKSUM_PARTIAL;
866 inet->cork.length += length;
867 if (((length> mtu) || !skb_queue_empty(&sk->sk_write_queue)) &&
868 (sk->sk_protocol == IPPROTO_UDP) &&
869 (rt->u.dst.dev->features & NETIF_F_UFO)) {
870 err = ip_ufo_append_data(sk, getfrag, from, length, hh_len,
871 fragheaderlen, transhdrlen, mtu,
878 /* So, what's going on in the loop below?
880 * We use calculated fragment length to generate chained skb,
881 * each of segments is IP fragment ready for sending to network after
882 * adding appropriate IP header.
885 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
889 /* Check if the remaining data fits into current packet. */
890 copy = mtu - skb->len;
892 copy = maxfraglen - skb->len;
895 unsigned int datalen;
896 unsigned int fraglen;
897 unsigned int fraggap;
898 unsigned int alloclen;
899 struct sk_buff *skb_prev;
903 fraggap = skb_prev->len - maxfraglen;
908 * If remaining data exceeds the mtu,
909 * we know we need more fragment(s).
911 datalen = length + fraggap;
912 if (datalen > mtu - fragheaderlen)
913 datalen = maxfraglen - fragheaderlen;
914 fraglen = datalen + fragheaderlen;
916 if ((flags & MSG_MORE) &&
917 !(rt->u.dst.dev->features&NETIF_F_SG))
920 alloclen = datalen + fragheaderlen;
922 /* The last fragment gets additional space at tail.
923 * Note, with MSG_MORE we overallocate on fragments,
924 * because we have no idea what fragment will be
927 if (datalen == length + fraggap)
928 alloclen += rt->u.dst.trailer_len;
931 skb = sock_alloc_send_skb(sk,
932 alloclen + hh_len + 15,
933 (flags & MSG_DONTWAIT), &err);
936 if (atomic_read(&sk->sk_wmem_alloc) <=
938 skb = sock_wmalloc(sk,
939 alloclen + hh_len + 15, 1,
941 if (unlikely(skb == NULL))
944 /* only the initial fragment is
952 * Fill in the control structures
954 skb->ip_summed = csummode;
956 skb_reserve(skb, hh_len);
957 *skb_tx(skb) = ipc->shtx;
960 * Find where to start putting bytes.
962 data = skb_put(skb, fraglen);
963 skb_set_network_header(skb, exthdrlen);
964 skb->transport_header = (skb->network_header +
966 data += fragheaderlen;
969 skb->csum = skb_copy_and_csum_bits(
970 skb_prev, maxfraglen,
971 data + transhdrlen, fraggap, 0);
972 skb_prev->csum = csum_sub(skb_prev->csum,
975 pskb_trim_unique(skb_prev, maxfraglen);
978 copy = datalen - transhdrlen - fraggap;
979 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
986 length -= datalen - fraggap;
989 csummode = CHECKSUM_NONE;
992 * Put the packet on the pending queue.
994 __skb_queue_tail(&sk->sk_write_queue, skb);
1001 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
1005 if (getfrag(from, skb_put(skb, copy),
1006 offset, copy, off, skb) < 0) {
1007 __skb_trim(skb, off);
1012 int i = skb_shinfo(skb)->nr_frags;
1013 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
1014 struct page *page = sk->sk_sndmsg_page;
1015 int off = sk->sk_sndmsg_off;
1018 if (page && (left = PAGE_SIZE - off) > 0) {
1021 if (page != frag->page) {
1022 if (i == MAX_SKB_FRAGS) {
1027 skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
1028 frag = &skb_shinfo(skb)->frags[i];
1030 } else if (i < MAX_SKB_FRAGS) {
1031 if (copy > PAGE_SIZE)
1033 page = alloc_pages(sk->sk_allocation, 0);
1038 sk->sk_sndmsg_page = page;
1039 sk->sk_sndmsg_off = 0;
1041 skb_fill_page_desc(skb, i, page, 0, 0);
1042 frag = &skb_shinfo(skb)->frags[i];
1047 if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1051 sk->sk_sndmsg_off += copy;
1054 skb->data_len += copy;
1055 skb->truesize += copy;
1056 atomic_add(copy, &sk->sk_wmem_alloc);
1065 inet->cork.length -= length;
1066 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1070 ssize_t ip_append_page(struct sock *sk, struct page *page,
1071 int offset, size_t size, int flags)
1073 struct inet_sock *inet = inet_sk(sk);
1074 struct sk_buff *skb;
1076 struct ip_options *opt = NULL;
1081 unsigned int maxfraglen, fragheaderlen, fraggap;
1086 if (flags&MSG_PROBE)
1089 if (skb_queue_empty(&sk->sk_write_queue))
1092 rt = (struct rtable *)inet->cork.dst;
1093 if (inet->cork.flags & IPCORK_OPT)
1094 opt = inet->cork.opt;
1096 if (!(rt->u.dst.dev->features&NETIF_F_SG))
1099 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1100 mtu = inet->cork.fragsize;
1102 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1103 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1105 if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1106 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport, mtu);
1110 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1113 inet->cork.length += size;
1114 if ((sk->sk_protocol == IPPROTO_UDP) &&
1115 (rt->u.dst.dev->features & NETIF_F_UFO)) {
1116 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1117 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1124 if (skb_is_gso(skb))
1128 /* Check if the remaining data fits into current packet. */
1129 len = mtu - skb->len;
1131 len = maxfraglen - skb->len;
1134 struct sk_buff *skb_prev;
1138 fraggap = skb_prev->len - maxfraglen;
1140 alloclen = fragheaderlen + hh_len + fraggap + 15;
1141 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1142 if (unlikely(!skb)) {
1148 * Fill in the control structures
1150 skb->ip_summed = CHECKSUM_NONE;
1152 skb_reserve(skb, hh_len);
1155 * Find where to start putting bytes.
1157 skb_put(skb, fragheaderlen + fraggap);
1158 skb_reset_network_header(skb);
1159 skb->transport_header = (skb->network_header +
1162 skb->csum = skb_copy_and_csum_bits(skb_prev,
1164 skb_transport_header(skb),
1166 skb_prev->csum = csum_sub(skb_prev->csum,
1168 pskb_trim_unique(skb_prev, maxfraglen);
1172 * Put the packet on the pending queue.
1174 __skb_queue_tail(&sk->sk_write_queue, skb);
1178 i = skb_shinfo(skb)->nr_frags;
1181 if (skb_can_coalesce(skb, i, page, offset)) {
1182 skb_shinfo(skb)->frags[i-1].size += len;
1183 } else if (i < MAX_SKB_FRAGS) {
1185 skb_fill_page_desc(skb, i, page, offset, len);
1191 if (skb->ip_summed == CHECKSUM_NONE) {
1193 csum = csum_page(page, offset, len);
1194 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1198 skb->data_len += len;
1199 skb->truesize += len;
1200 atomic_add(len, &sk->sk_wmem_alloc);
1207 inet->cork.length -= size;
1208 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1212 static void ip_cork_release(struct inet_sock *inet)
1214 inet->cork.flags &= ~IPCORK_OPT;
1215 kfree(inet->cork.opt);
1216 inet->cork.opt = NULL;
1217 dst_release(inet->cork.dst);
1218 inet->cork.dst = NULL;
1222 * Combined all pending IP fragments on the socket as one IP datagram
1223 * and push them out.
1225 int ip_push_pending_frames(struct sock *sk)
1227 struct sk_buff *skb, *tmp_skb;
1228 struct sk_buff **tail_skb;
1229 struct inet_sock *inet = inet_sk(sk);
1230 struct net *net = sock_net(sk);
1231 struct ip_options *opt = NULL;
1232 struct rtable *rt = (struct rtable *)inet->cork.dst;
1238 if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1240 tail_skb = &(skb_shinfo(skb)->frag_list);
1242 /* move skb->data to ip header from ext header */
1243 if (skb->data < skb_network_header(skb))
1244 __skb_pull(skb, skb_network_offset(skb));
1245 while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1246 __skb_pull(tmp_skb, skb_network_header_len(skb));
1247 *tail_skb = tmp_skb;
1248 tail_skb = &(tmp_skb->next);
1249 skb->len += tmp_skb->len;
1250 skb->data_len += tmp_skb->len;
1251 skb->truesize += tmp_skb->truesize;
1252 tmp_skb->destructor = NULL;
1256 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1257 * to fragment the frame generated here. No matter, what transforms
1258 * how transforms change size of the packet, it will come out.
1260 if (inet->pmtudisc < IP_PMTUDISC_DO)
1263 /* DF bit is set when we want to see DF on outgoing frames.
1264 * If local_df is set too, we still allow to fragment this frame
1266 if (inet->pmtudisc >= IP_PMTUDISC_DO ||
1267 (skb->len <= dst_mtu(&rt->u.dst) &&
1268 ip_dont_fragment(sk, &rt->u.dst)))
1271 if (inet->cork.flags & IPCORK_OPT)
1272 opt = inet->cork.opt;
1274 if (rt->rt_type == RTN_MULTICAST)
1277 ttl = ip_select_ttl(inet, &rt->u.dst);
1279 iph = (struct iphdr *)skb->data;
1283 iph->ihl += opt->optlen>>2;
1284 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1286 iph->tos = inet->tos;
1288 ip_select_ident(iph, &rt->u.dst, sk);
1290 iph->protocol = sk->sk_protocol;
1291 iph->saddr = rt->rt_src;
1292 iph->daddr = rt->rt_dst;
1294 skb->priority = sk->sk_priority;
1295 skb->mark = sk->sk_mark;
1297 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1300 inet->cork.dst = NULL;
1301 skb_dst_set(skb, &rt->u.dst);
1303 if (iph->protocol == IPPROTO_ICMP)
1304 icmp_out_count(net, ((struct icmphdr *)
1305 skb_transport_header(skb))->type);
1307 /* Netfilter gets whole the not fragmented skb. */
1308 err = ip_local_out(skb);
1311 err = net_xmit_errno(err);
1317 ip_cork_release(inet);
1321 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1326 * Throw away all pending data on the socket.
1328 void ip_flush_pending_frames(struct sock *sk)
1330 struct sk_buff *skb;
1332 while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1335 ip_cork_release(inet_sk(sk));
1340 * Fetch data from kernel space and fill in checksum if needed.
1342 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1343 int len, int odd, struct sk_buff *skb)
1347 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1348 skb->csum = csum_block_add(skb->csum, csum, odd);
1353 * Generic function to send a packet as reply to another packet.
1354 * Used to send TCP resets so far. ICMP should use this function too.
1356 * Should run single threaded per socket because it uses the sock
1357 * structure to pass arguments.
1359 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1362 struct inet_sock *inet = inet_sk(sk);
1364 struct ip_options opt;
1367 struct ipcm_cookie ipc;
1369 struct rtable *rt = skb_rtable(skb);
1371 if (ip_options_echo(&replyopts.opt, skb))
1374 daddr = ipc.addr = rt->rt_src;
1378 if (replyopts.opt.optlen) {
1379 ipc.opt = &replyopts.opt;
1382 daddr = replyopts.opt.faddr;
1386 struct flowi fl = { .oif = arg->bound_dev_if,
1389 .saddr = rt->rt_spec_dst,
1390 .tos = RT_TOS(ip_hdr(skb)->tos) } },
1391 /* Not quite clean, but right. */
1393 { .sport = tcp_hdr(skb)->dest,
1394 .dport = tcp_hdr(skb)->source } },
1395 .proto = sk->sk_protocol,
1396 .flags = ip_reply_arg_flowi_flags(arg) };
1397 security_skb_classify_flow(skb, &fl);
1398 if (ip_route_output_key(sock_net(sk), &rt, &fl))
1402 /* And let IP do all the hard work.
1404 This chunk is not reenterable, hence spinlock.
1405 Note that it uses the fact, that this function is called
1406 with locally disabled BH and that sk cannot be already spinlocked.
1409 inet->tos = ip_hdr(skb)->tos;
1410 sk->sk_priority = skb->priority;
1411 sk->sk_protocol = ip_hdr(skb)->protocol;
1412 sk->sk_bound_dev_if = arg->bound_dev_if;
1413 ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1414 &ipc, &rt, MSG_DONTWAIT);
1415 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1416 if (arg->csumoffset >= 0)
1417 *((__sum16 *)skb_transport_header(skb) +
1418 arg->csumoffset) = csum_fold(csum_add(skb->csum,
1420 skb->ip_summed = CHECKSUM_NONE;
1421 ip_push_pending_frames(sk);
1429 void __init ip_init(void)
1434 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1435 igmp_mc_proc_init();
1439 EXPORT_SYMBOL(ip_generic_getfrag);
1440 EXPORT_SYMBOL(ip_queue_xmit);
1441 EXPORT_SYMBOL(ip_send_check);