2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * The Internet Protocol (IP) output module.
8 * Version: $Id: ip_output.c,v 1.100 2002/02/01 22:01:03 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Donald Becker, <becker@super.org>
13 * Alan Cox, <Alan.Cox@linux.org>
15 * Stefan Becker, <stefanb@yello.ping.de>
16 * Jorge Cwik, <jorge@laser.satlink.net>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Hirokazu Takahashi, <taka@valinux.co.jp>
20 * See ip_input.c for original log
23 * Alan Cox : Missing nonblock feature in ip_build_xmit.
24 * Mike Kilburn : htons() missing in ip_build_xmit.
25 * Bradford Johnson: Fix faulty handling of some frames when
27 * Alexander Demenshin: Missing sk/skb free in ip_queue_xmit
28 * (in case if packet not accepted by
29 * output firewall rules)
30 * Mike McLagan : Routing by source
31 * Alexey Kuznetsov: use new route cache
32 * Andi Kleen: Fix broken PMTU recovery and remove
33 * some redundant tests.
34 * Vitaly E. Lavrov : Transparent proxy revived after year coma.
35 * Andi Kleen : Replace ip_reply with ip_send_reply.
36 * Andi Kleen : Split fast and slow ip_build_xmit path
37 * for decreased register pressure on x86
38 * and more readibility.
39 * Marc Boucher : When call_out_firewall returns FW_QUEUE,
40 * silently drop skb instead of failing with -EPERM.
41 * Detlev Wengorz : Copy protocol for fragments.
42 * Hirokazu Takahashi: HW checksumming for outgoing UDP
44 * Hirokazu Takahashi: sendfile() on UDP works now.
47 #include <asm/uaccess.h>
48 #include <asm/system.h>
49 #include <linux/module.h>
50 #include <linux/types.h>
51 #include <linux/kernel.h>
53 #include <linux/string.h>
54 #include <linux/errno.h>
55 #include <linux/highmem.h>
57 #include <linux/socket.h>
58 #include <linux/sockios.h>
60 #include <linux/inet.h>
61 #include <linux/netdevice.h>
62 #include <linux/etherdevice.h>
63 #include <linux/proc_fs.h>
64 #include <linux/stat.h>
65 #include <linux/init.h>
69 #include <net/protocol.h>
70 #include <net/route.h>
72 #include <linux/skbuff.h>
76 #include <net/checksum.h>
77 #include <net/inetpeer.h>
78 #include <linux/igmp.h>
79 #include <linux/netfilter_ipv4.h>
80 #include <linux/netfilter_bridge.h>
81 #include <linux/mroute.h>
82 #include <linux/netlink.h>
83 #include <linux/tcp.h>
85 int sysctl_ip_default_ttl __read_mostly = IPDEFTTL;
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);
94 int __ip_local_out(struct sk_buff *skb)
96 struct iphdr *iph = ip_hdr(skb);
98 iph->tot_len = htons(skb->len);
100 return nf_hook(PF_INET, NF_INET_LOCAL_OUT, skb, NULL, skb->dst->dev,
104 int ip_local_out(struct sk_buff *skb)
108 err = __ip_local_out(skb);
109 if (likely(err == 1))
110 err = dst_output(skb);
114 EXPORT_SYMBOL_GPL(ip_local_out);
116 /* dev_loopback_xmit for use with netfilter. */
117 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
119 skb_reset_mac_header(newskb);
120 __skb_pull(newskb, skb_network_offset(newskb));
121 newskb->pkt_type = PACKET_LOOPBACK;
122 newskb->ip_summed = CHECKSUM_UNNECESSARY;
123 BUG_TRAP(newskb->dst);
128 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
130 int ttl = inet->uc_ttl;
133 ttl = dst_metric(dst, RTAX_HOPLIMIT);
138 * Add an ip header to a skbuff and send it out.
141 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
142 __be32 saddr, __be32 daddr, struct ip_options *opt)
144 struct inet_sock *inet = inet_sk(sk);
145 struct rtable *rt = (struct rtable *)skb->dst;
148 /* Build the IP header. */
149 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
150 skb_reset_network_header(skb);
154 iph->tos = inet->tos;
155 if (ip_dont_fragment(sk, &rt->u.dst))
156 iph->frag_off = htons(IP_DF);
159 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
160 iph->daddr = rt->rt_dst;
161 iph->saddr = rt->rt_src;
162 iph->protocol = sk->sk_protocol;
163 ip_select_ident(iph, &rt->u.dst, sk);
165 if (opt && opt->optlen) {
166 iph->ihl += opt->optlen>>2;
167 ip_options_build(skb, opt, daddr, rt, 0);
170 skb->priority = sk->sk_priority;
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;
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_INC_STATS(IPSTATS_MIB_OUTMCASTPKTS);
187 else if (rt->rt_type == RTN_BROADCAST)
188 IP_INC_STATS(IPSTATS_MIB_OUTBCASTPKTS);
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->dev->mtu : dst_mtu(skb->dst);
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->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 = (struct rtable*)skb->dst;
243 struct net_device *dev = rt->u.dst.dev;
246 * If the indicated interface is up and running, send the packet.
248 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
251 skb->protocol = htons(ETH_P_IP);
254 * Multicasts are looped back for other local users
257 if (rt->rt_flags&RTCF_MULTICAST) {
258 if ((!sk || inet_sk(sk)->mc_loop)
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.
268 && ((rt->rt_flags&RTCF_LOCAL) || !(IPCB(skb)->flags&IPSKB_FORWARDED))
271 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
273 NF_HOOK(PF_INET, NF_INET_POST_ROUTING, newskb,
275 ip_dev_loopback_xmit);
278 /* Multicasts with ttl 0 must not go beyond the host */
280 if (ip_hdr(skb)->ttl == 0) {
286 if (rt->rt_flags&RTCF_BROADCAST) {
287 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
289 NF_HOOK(PF_INET, NF_INET_POST_ROUTING, newskb, NULL,
290 newskb->dev, ip_dev_loopback_xmit);
293 return NF_HOOK_COND(PF_INET, NF_INET_POST_ROUTING, skb, NULL, skb->dev,
295 !(IPCB(skb)->flags & IPSKB_REROUTED));
298 int ip_output(struct sk_buff *skb)
300 struct net_device *dev = skb->dst->dev;
302 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
305 skb->protocol = htons(ETH_P_IP);
307 return NF_HOOK_COND(PF_INET, NF_INET_POST_ROUTING, skb, NULL, dev,
309 !(IPCB(skb)->flags & IPSKB_REROUTED));
312 int ip_queue_xmit(struct sk_buff *skb, int ipfragok)
314 struct sock *sk = skb->sk;
315 struct inet_sock *inet = inet_sk(sk);
316 struct ip_options *opt = inet->opt;
320 /* Skip all of this if the packet is already routed,
321 * f.e. by something like SCTP.
323 rt = (struct rtable *) skb->dst;
327 /* Make sure we can route this packet. */
328 rt = (struct rtable *)__sk_dst_check(sk, 0);
332 /* Use correct destination address if we have options. */
338 struct flowi fl = { .oif = sk->sk_bound_dev_if,
341 .saddr = inet->saddr,
342 .tos = RT_CONN_FLAGS(sk) } },
343 .proto = sk->sk_protocol,
345 { .sport = inet->sport,
346 .dport = inet->dport } } };
348 /* If this fails, retransmit mechanism of transport layer will
349 * keep trying until route appears or the connection times
352 security_sk_classify_flow(sk, &fl);
353 if (ip_route_output_flow(&init_net, &rt, &fl, sk, 0))
356 sk_setup_caps(sk, &rt->u.dst);
358 skb->dst = dst_clone(&rt->u.dst);
361 if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
364 /* OK, we know where to send it, allocate and build IP header. */
365 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
366 skb_reset_network_header(skb);
368 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
369 if (ip_dont_fragment(sk, &rt->u.dst) && !ipfragok)
370 iph->frag_off = htons(IP_DF);
373 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
374 iph->protocol = sk->sk_protocol;
375 iph->saddr = rt->rt_src;
376 iph->daddr = rt->rt_dst;
377 /* Transport layer set skb->h.foo itself. */
379 if (opt && opt->optlen) {
380 iph->ihl += opt->optlen >> 2;
381 ip_options_build(skb, opt, inet->daddr, rt, 0);
384 ip_select_ident_more(iph, &rt->u.dst, sk,
385 (skb_shinfo(skb)->gso_segs ?: 1) - 1);
387 skb->priority = sk->sk_priority;
389 return ip_local_out(skb);
392 IP_INC_STATS(IPSTATS_MIB_OUTNOROUTES);
394 return -EHOSTUNREACH;
398 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
400 to->pkt_type = from->pkt_type;
401 to->priority = from->priority;
402 to->protocol = from->protocol;
403 dst_release(to->dst);
404 to->dst = dst_clone(from->dst);
406 to->mark = from->mark;
408 /* Copy the flags to each fragment. */
409 IPCB(to)->flags = IPCB(from)->flags;
411 #ifdef CONFIG_NET_SCHED
412 to->tc_index = from->tc_index;
415 #if defined(CONFIG_NETFILTER_XT_TARGET_TRACE) || \
416 defined(CONFIG_NETFILTER_XT_TARGET_TRACE_MODULE)
417 to->nf_trace = from->nf_trace;
419 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
420 to->ipvs_property = from->ipvs_property;
422 skb_copy_secmark(to, from);
426 * This IP datagram is too large to be sent in one piece. Break it up into
427 * smaller pieces (each of size equal to IP header plus
428 * a block of the data of the original IP data part) that will yet fit in a
429 * single device frame, and queue such a frame for sending.
432 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*))
437 struct net_device *dev;
438 struct sk_buff *skb2;
439 unsigned int mtu, hlen, left, len, ll_rs, pad;
441 __be16 not_last_frag;
442 struct rtable *rt = (struct rtable*)skb->dst;
448 * Point into the IP datagram header.
453 if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
454 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
455 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
456 htonl(ip_skb_dst_mtu(skb)));
462 * Setup starting values.
466 mtu = dst_mtu(&rt->u.dst) - hlen; /* Size of data space */
467 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
469 /* When frag_list is given, use it. First, check its validity:
470 * some transformers could create wrong frag_list or break existing
471 * one, it is not prohibited. In this case fall back to copying.
473 * LATER: this step can be merged to real generation of fragments,
474 * we can switch to copy when see the first bad fragment.
476 if (skb_shinfo(skb)->frag_list) {
477 struct sk_buff *frag;
478 int first_len = skb_pagelen(skb);
481 if (first_len - hlen > mtu ||
482 ((first_len - hlen) & 7) ||
483 (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
487 for (frag = skb_shinfo(skb)->frag_list; frag; frag = frag->next) {
488 /* Correct geometry. */
489 if (frag->len > mtu ||
490 ((frag->len & 7) && frag->next) ||
491 skb_headroom(frag) < hlen)
494 /* Partially cloned skb? */
495 if (skb_shared(frag))
502 frag->destructor = sock_wfree;
503 truesizes += frag->truesize;
507 /* Everything is OK. Generate! */
511 frag = skb_shinfo(skb)->frag_list;
512 skb_shinfo(skb)->frag_list = NULL;
513 skb->data_len = first_len - skb_headlen(skb);
514 skb->truesize -= truesizes;
515 skb->len = first_len;
516 iph->tot_len = htons(first_len);
517 iph->frag_off = htons(IP_MF);
521 /* Prepare header of the next frame,
522 * before previous one went down. */
524 frag->ip_summed = CHECKSUM_NONE;
525 skb_reset_transport_header(frag);
526 __skb_push(frag, hlen);
527 skb_reset_network_header(frag);
528 memcpy(skb_network_header(frag), iph, hlen);
530 iph->tot_len = htons(frag->len);
531 ip_copy_metadata(frag, skb);
533 ip_options_fragment(frag);
534 offset += skb->len - hlen;
535 iph->frag_off = htons(offset>>3);
536 if (frag->next != NULL)
537 iph->frag_off |= htons(IP_MF);
538 /* Ready, complete checksum */
545 IP_INC_STATS(IPSTATS_MIB_FRAGCREATES);
555 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
564 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
569 left = skb->len - hlen; /* Space per frame */
570 ptr = raw + hlen; /* Where to start from */
572 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
573 * we need to make room for the encapsulating header
575 pad = nf_bridge_pad(skb);
576 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, pad);
580 * Fragment the datagram.
583 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
584 not_last_frag = iph->frag_off & htons(IP_MF);
587 * Keep copying data until we run out.
592 /* IF: it doesn't fit, use 'mtu' - the data space left */
595 /* IF: we are not sending upto and including the packet end
596 then align the next start on an eight byte boundary */
604 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
605 NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
611 * Set up data on packet
614 ip_copy_metadata(skb2, skb);
615 skb_reserve(skb2, ll_rs);
616 skb_put(skb2, len + hlen);
617 skb_reset_network_header(skb2);
618 skb2->transport_header = skb2->network_header + hlen;
621 * Charge the memory for the fragment to any owner
626 skb_set_owner_w(skb2, skb->sk);
629 * Copy the packet header into the new buffer.
632 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
635 * Copy a block of the IP datagram.
637 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
642 * Fill in the new header fields.
645 iph->frag_off = htons((offset >> 3));
647 /* ANK: dirty, but effective trick. Upgrade options only if
648 * the segment to be fragmented was THE FIRST (otherwise,
649 * options are already fixed) and make it ONCE
650 * on the initial skb, so that all the following fragments
651 * will inherit fixed options.
654 ip_options_fragment(skb);
657 * Added AC : If we are fragmenting a fragment that's not the
658 * last fragment then keep MF on each bit
660 if (left > 0 || not_last_frag)
661 iph->frag_off |= htons(IP_MF);
666 * Put this fragment into the sending queue.
668 iph->tot_len = htons(len + hlen);
676 IP_INC_STATS(IPSTATS_MIB_FRAGCREATES);
679 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
684 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
688 EXPORT_SYMBOL(ip_fragment);
691 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
693 struct iovec *iov = from;
695 if (skb->ip_summed == CHECKSUM_PARTIAL) {
696 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
700 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
702 skb->csum = csum_block_add(skb->csum, csum, odd);
708 csum_page(struct page *page, int offset, int copy)
713 csum = csum_partial(kaddr + offset, copy, 0);
718 static inline int ip_ufo_append_data(struct sock *sk,
719 int getfrag(void *from, char *to, int offset, int len,
720 int odd, struct sk_buff *skb),
721 void *from, int length, int hh_len, int fragheaderlen,
722 int transhdrlen, int mtu,unsigned int flags)
727 /* There is support for UDP fragmentation offload by network
728 * device, so create one single skb packet containing complete
731 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL) {
732 skb = sock_alloc_send_skb(sk,
733 hh_len + fragheaderlen + transhdrlen + 20,
734 (flags & MSG_DONTWAIT), &err);
739 /* reserve space for Hardware header */
740 skb_reserve(skb, hh_len);
742 /* create space for UDP/IP header */
743 skb_put(skb,fragheaderlen + transhdrlen);
745 /* initialize network header pointer */
746 skb_reset_network_header(skb);
748 /* initialize protocol header pointer */
749 skb->transport_header = skb->network_header + fragheaderlen;
751 skb->ip_summed = CHECKSUM_PARTIAL;
753 sk->sk_sndmsg_off = 0;
756 err = skb_append_datato_frags(sk,skb, getfrag, from,
757 (length - transhdrlen));
759 /* specify the length of each IP datagram fragment*/
760 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
761 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
762 __skb_queue_tail(&sk->sk_write_queue, skb);
766 /* There is not enough support do UFO ,
767 * so follow normal path
774 * ip_append_data() and ip_append_page() can make one large IP datagram
775 * from many pieces of data. Each pieces will be holded on the socket
776 * until ip_push_pending_frames() is called. Each piece can be a page
779 * Not only UDP, other transport protocols - e.g. raw sockets - can use
780 * this interface potentially.
782 * LATER: length must be adjusted by pad at tail, when it is required.
784 int ip_append_data(struct sock *sk,
785 int getfrag(void *from, char *to, int offset, int len,
786 int odd, struct sk_buff *skb),
787 void *from, int length, int transhdrlen,
788 struct ipcm_cookie *ipc, struct rtable *rt,
791 struct inet_sock *inet = inet_sk(sk);
794 struct ip_options *opt = NULL;
801 unsigned int maxfraglen, fragheaderlen;
802 int csummode = CHECKSUM_NONE;
807 if (skb_queue_empty(&sk->sk_write_queue)) {
813 if (inet->cork.opt == NULL) {
814 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
815 if (unlikely(inet->cork.opt == NULL))
818 memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
819 inet->cork.flags |= IPCORK_OPT;
820 inet->cork.addr = ipc->addr;
822 dst_hold(&rt->u.dst);
823 inet->cork.fragsize = mtu = inet->pmtudisc == IP_PMTUDISC_PROBE ?
825 dst_mtu(rt->u.dst.path);
827 inet->cork.length = 0;
828 sk->sk_sndmsg_page = NULL;
829 sk->sk_sndmsg_off = 0;
830 if ((exthdrlen = rt->u.dst.header_len) != 0) {
832 transhdrlen += exthdrlen;
836 if (inet->cork.flags & IPCORK_OPT)
837 opt = inet->cork.opt;
841 mtu = inet->cork.fragsize;
843 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
845 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
846 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
848 if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
849 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu-exthdrlen);
854 * transhdrlen > 0 means that this is the first fragment and we wish
855 * it won't be fragmented in the future.
858 length + fragheaderlen <= mtu &&
859 rt->u.dst.dev->features & NETIF_F_V4_CSUM &&
861 csummode = CHECKSUM_PARTIAL;
863 inet->cork.length += length;
864 if (((length > mtu) && (sk->sk_protocol == IPPROTO_UDP)) &&
865 (rt->u.dst.dev->features & NETIF_F_UFO)) {
867 err = ip_ufo_append_data(sk, getfrag, from, length, hh_len,
868 fragheaderlen, transhdrlen, mtu,
875 /* So, what's going on in the loop below?
877 * We use calculated fragment length to generate chained skb,
878 * each of segments is IP fragment ready for sending to network after
879 * adding appropriate IP header.
882 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
886 /* Check if the remaining data fits into current packet. */
887 copy = mtu - skb->len;
889 copy = maxfraglen - skb->len;
892 unsigned int datalen;
893 unsigned int fraglen;
894 unsigned int fraggap;
895 unsigned int alloclen;
896 struct sk_buff *skb_prev;
900 fraggap = skb_prev->len - maxfraglen;
905 * If remaining data exceeds the mtu,
906 * we know we need more fragment(s).
908 datalen = length + fraggap;
909 if (datalen > mtu - fragheaderlen)
910 datalen = maxfraglen - fragheaderlen;
911 fraglen = datalen + fragheaderlen;
913 if ((flags & MSG_MORE) &&
914 !(rt->u.dst.dev->features&NETIF_F_SG))
917 alloclen = datalen + fragheaderlen;
919 /* The last fragment gets additional space at tail.
920 * Note, with MSG_MORE we overallocate on fragments,
921 * because we have no idea what fragment will be
924 if (datalen == length + fraggap)
925 alloclen += rt->u.dst.trailer_len;
928 skb = sock_alloc_send_skb(sk,
929 alloclen + hh_len + 15,
930 (flags & MSG_DONTWAIT), &err);
933 if (atomic_read(&sk->sk_wmem_alloc) <=
935 skb = sock_wmalloc(sk,
936 alloclen + hh_len + 15, 1,
938 if (unlikely(skb == NULL))
945 * Fill in the control structures
947 skb->ip_summed = csummode;
949 skb_reserve(skb, hh_len);
952 * Find where to start putting bytes.
954 data = skb_put(skb, fraglen);
955 skb_set_network_header(skb, exthdrlen);
956 skb->transport_header = (skb->network_header +
958 data += fragheaderlen;
961 skb->csum = skb_copy_and_csum_bits(
962 skb_prev, maxfraglen,
963 data + transhdrlen, fraggap, 0);
964 skb_prev->csum = csum_sub(skb_prev->csum,
967 pskb_trim_unique(skb_prev, maxfraglen);
970 copy = datalen - transhdrlen - fraggap;
971 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
978 length -= datalen - fraggap;
981 csummode = CHECKSUM_NONE;
984 * Put the packet on the pending queue.
986 __skb_queue_tail(&sk->sk_write_queue, skb);
993 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
997 if (getfrag(from, skb_put(skb, copy),
998 offset, copy, off, skb) < 0) {
999 __skb_trim(skb, off);
1004 int i = skb_shinfo(skb)->nr_frags;
1005 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
1006 struct page *page = sk->sk_sndmsg_page;
1007 int off = sk->sk_sndmsg_off;
1010 if (page && (left = PAGE_SIZE - off) > 0) {
1013 if (page != frag->page) {
1014 if (i == MAX_SKB_FRAGS) {
1019 skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
1020 frag = &skb_shinfo(skb)->frags[i];
1022 } else if (i < MAX_SKB_FRAGS) {
1023 if (copy > PAGE_SIZE)
1025 page = alloc_pages(sk->sk_allocation, 0);
1030 sk->sk_sndmsg_page = page;
1031 sk->sk_sndmsg_off = 0;
1033 skb_fill_page_desc(skb, i, page, 0, 0);
1034 frag = &skb_shinfo(skb)->frags[i];
1039 if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1043 sk->sk_sndmsg_off += copy;
1046 skb->data_len += copy;
1047 skb->truesize += copy;
1048 atomic_add(copy, &sk->sk_wmem_alloc);
1057 inet->cork.length -= length;
1058 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1062 ssize_t ip_append_page(struct sock *sk, struct page *page,
1063 int offset, size_t size, int flags)
1065 struct inet_sock *inet = inet_sk(sk);
1066 struct sk_buff *skb;
1068 struct ip_options *opt = NULL;
1073 unsigned int maxfraglen, fragheaderlen, fraggap;
1078 if (flags&MSG_PROBE)
1081 if (skb_queue_empty(&sk->sk_write_queue))
1085 if (inet->cork.flags & IPCORK_OPT)
1086 opt = inet->cork.opt;
1088 if (!(rt->u.dst.dev->features&NETIF_F_SG))
1091 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1092 mtu = inet->cork.fragsize;
1094 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1095 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1097 if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1098 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu);
1102 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1105 inet->cork.length += size;
1106 if ((sk->sk_protocol == IPPROTO_UDP) &&
1107 (rt->u.dst.dev->features & NETIF_F_UFO)) {
1108 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1109 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1116 if (skb_is_gso(skb))
1120 /* Check if the remaining data fits into current packet. */
1121 len = mtu - skb->len;
1123 len = maxfraglen - skb->len;
1126 struct sk_buff *skb_prev;
1130 fraggap = skb_prev->len - maxfraglen;
1132 alloclen = fragheaderlen + hh_len + fraggap + 15;
1133 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1134 if (unlikely(!skb)) {
1140 * Fill in the control structures
1142 skb->ip_summed = CHECKSUM_NONE;
1144 skb_reserve(skb, hh_len);
1147 * Find where to start putting bytes.
1149 skb_put(skb, fragheaderlen + fraggap);
1150 skb_reset_network_header(skb);
1151 skb->transport_header = (skb->network_header +
1154 skb->csum = skb_copy_and_csum_bits(skb_prev,
1156 skb_transport_header(skb),
1158 skb_prev->csum = csum_sub(skb_prev->csum,
1160 pskb_trim_unique(skb_prev, maxfraglen);
1164 * Put the packet on the pending queue.
1166 __skb_queue_tail(&sk->sk_write_queue, skb);
1170 i = skb_shinfo(skb)->nr_frags;
1173 if (skb_can_coalesce(skb, i, page, offset)) {
1174 skb_shinfo(skb)->frags[i-1].size += len;
1175 } else if (i < MAX_SKB_FRAGS) {
1177 skb_fill_page_desc(skb, i, page, offset, len);
1183 if (skb->ip_summed == CHECKSUM_NONE) {
1185 csum = csum_page(page, offset, len);
1186 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1190 skb->data_len += len;
1191 skb->truesize += len;
1192 atomic_add(len, &sk->sk_wmem_alloc);
1199 inet->cork.length -= size;
1200 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1204 static void ip_cork_release(struct inet_sock *inet)
1206 inet->cork.flags &= ~IPCORK_OPT;
1207 kfree(inet->cork.opt);
1208 inet->cork.opt = NULL;
1209 if (inet->cork.rt) {
1210 ip_rt_put(inet->cork.rt);
1211 inet->cork.rt = NULL;
1216 * Combined all pending IP fragments on the socket as one IP datagram
1217 * and push them out.
1219 int ip_push_pending_frames(struct sock *sk)
1221 struct sk_buff *skb, *tmp_skb;
1222 struct sk_buff **tail_skb;
1223 struct inet_sock *inet = inet_sk(sk);
1224 struct ip_options *opt = NULL;
1225 struct rtable *rt = inet->cork.rt;
1231 if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1233 tail_skb = &(skb_shinfo(skb)->frag_list);
1235 /* move skb->data to ip header from ext header */
1236 if (skb->data < skb_network_header(skb))
1237 __skb_pull(skb, skb_network_offset(skb));
1238 while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1239 __skb_pull(tmp_skb, skb_network_header_len(skb));
1240 *tail_skb = tmp_skb;
1241 tail_skb = &(tmp_skb->next);
1242 skb->len += tmp_skb->len;
1243 skb->data_len += tmp_skb->len;
1244 skb->truesize += tmp_skb->truesize;
1245 __sock_put(tmp_skb->sk);
1246 tmp_skb->destructor = NULL;
1250 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1251 * to fragment the frame generated here. No matter, what transforms
1252 * how transforms change size of the packet, it will come out.
1254 if (inet->pmtudisc < IP_PMTUDISC_DO)
1257 /* DF bit is set when we want to see DF on outgoing frames.
1258 * If local_df is set too, we still allow to fragment this frame
1260 if (inet->pmtudisc >= IP_PMTUDISC_DO ||
1261 (skb->len <= dst_mtu(&rt->u.dst) &&
1262 ip_dont_fragment(sk, &rt->u.dst)))
1265 if (inet->cork.flags & IPCORK_OPT)
1266 opt = inet->cork.opt;
1268 if (rt->rt_type == RTN_MULTICAST)
1271 ttl = ip_select_ttl(inet, &rt->u.dst);
1273 iph = (struct iphdr *)skb->data;
1277 iph->ihl += opt->optlen>>2;
1278 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1280 iph->tos = inet->tos;
1282 ip_select_ident(iph, &rt->u.dst, sk);
1284 iph->protocol = sk->sk_protocol;
1285 iph->saddr = rt->rt_src;
1286 iph->daddr = rt->rt_dst;
1288 skb->priority = sk->sk_priority;
1289 skb->dst = dst_clone(&rt->u.dst);
1291 if (iph->protocol == IPPROTO_ICMP)
1292 icmp_out_count(((struct icmphdr *)
1293 skb_transport_header(skb))->type);
1295 /* Netfilter gets whole the not fragmented skb. */
1296 err = ip_local_out(skb);
1299 err = inet->recverr ? net_xmit_errno(err) : 0;
1305 ip_cork_release(inet);
1309 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1314 * Throw away all pending data on the socket.
1316 void ip_flush_pending_frames(struct sock *sk)
1318 struct sk_buff *skb;
1320 while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1323 ip_cork_release(inet_sk(sk));
1328 * Fetch data from kernel space and fill in checksum if needed.
1330 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1331 int len, int odd, struct sk_buff *skb)
1335 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1336 skb->csum = csum_block_add(skb->csum, csum, odd);
1341 * Generic function to send a packet as reply to another packet.
1342 * Used to send TCP resets so far. ICMP should use this function too.
1344 * Should run single threaded per socket because it uses the sock
1345 * structure to pass arguments.
1347 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1350 struct inet_sock *inet = inet_sk(sk);
1352 struct ip_options opt;
1355 struct ipcm_cookie ipc;
1357 struct rtable *rt = (struct rtable*)skb->dst;
1359 if (ip_options_echo(&replyopts.opt, skb))
1362 daddr = ipc.addr = rt->rt_src;
1365 if (replyopts.opt.optlen) {
1366 ipc.opt = &replyopts.opt;
1369 daddr = replyopts.opt.faddr;
1373 struct flowi fl = { .oif = arg->bound_dev_if,
1376 .saddr = rt->rt_spec_dst,
1377 .tos = RT_TOS(ip_hdr(skb)->tos) } },
1378 /* Not quite clean, but right. */
1380 { .sport = tcp_hdr(skb)->dest,
1381 .dport = tcp_hdr(skb)->source } },
1382 .proto = sk->sk_protocol };
1383 security_skb_classify_flow(skb, &fl);
1384 if (ip_route_output_key(sk->sk_net, &rt, &fl))
1388 /* And let IP do all the hard work.
1390 This chunk is not reenterable, hence spinlock.
1391 Note that it uses the fact, that this function is called
1392 with locally disabled BH and that sk cannot be already spinlocked.
1395 inet->tos = ip_hdr(skb)->tos;
1396 sk->sk_priority = skb->priority;
1397 sk->sk_protocol = ip_hdr(skb)->protocol;
1398 sk->sk_bound_dev_if = arg->bound_dev_if;
1399 ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1400 &ipc, rt, MSG_DONTWAIT);
1401 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1402 if (arg->csumoffset >= 0)
1403 *((__sum16 *)skb_transport_header(skb) +
1404 arg->csumoffset) = csum_fold(csum_add(skb->csum,
1406 skb->ip_summed = CHECKSUM_NONE;
1407 ip_push_pending_frames(sk);
1415 void __init ip_init(void)
1420 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1421 igmp_mc_proc_init();
1425 EXPORT_SYMBOL(ip_generic_getfrag);
1426 EXPORT_SYMBOL(ip_queue_xmit);
1427 EXPORT_SYMBOL(ip_send_check);