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);
92 EXPORT_SYMBOL(ip_send_check);
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(NFPROTO_IPV4, NF_INET_LOCAL_OUT, skb, NULL,
101 skb_dst(skb)->dev, dst_output);
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 WARN_ON(!skb_dst(newskb));
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 = skb_rtable(skb);
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->dst))
156 iph->frag_off = htons(IP_DF);
159 iph->ttl = ip_select_ttl(inet, &rt->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->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;
171 skb->mark = sk->sk_mark;
174 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,
345 .fl4_src = inet->inet_saddr,
346 .fl4_tos = RT_CONN_FLAGS(sk),
347 .proto = sk->sk_protocol,
348 .flags = inet_sk_flowi_flags(sk),
349 .fl_ip_sport = inet->inet_sport,
350 .fl_ip_dport = inet->inet_dport };
352 /* If this fails, retransmit mechanism of transport layer will
353 * keep trying until route appears or the connection times
356 security_sk_classify_flow(sk, &fl);
357 if (ip_route_output_flow(sock_net(sk), &rt, &fl, sk, 0))
360 sk_setup_caps(sk, &rt->dst);
362 skb_dst_set_noref(skb, &rt->dst);
365 if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
368 /* OK, we know where to send it, allocate and build IP header. */
369 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
370 skb_reset_network_header(skb);
372 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
373 if (ip_dont_fragment(sk, &rt->dst) && !skb->local_df)
374 iph->frag_off = htons(IP_DF);
377 iph->ttl = ip_select_ttl(inet, &rt->dst);
378 iph->protocol = sk->sk_protocol;
379 iph->saddr = rt->rt_src;
380 iph->daddr = rt->rt_dst;
381 /* Transport layer set skb->h.foo itself. */
383 if (opt && opt->optlen) {
384 iph->ihl += opt->optlen >> 2;
385 ip_options_build(skb, opt, inet->inet_daddr, rt, 0);
388 ip_select_ident_more(iph, &rt->dst, sk,
389 (skb_shinfo(skb)->gso_segs ?: 1) - 1);
391 skb->priority = sk->sk_priority;
392 skb->mark = sk->sk_mark;
394 res = ip_local_out(skb);
400 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
402 return -EHOSTUNREACH;
404 EXPORT_SYMBOL(ip_queue_xmit);
407 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
409 to->pkt_type = from->pkt_type;
410 to->priority = from->priority;
411 to->protocol = from->protocol;
413 skb_dst_copy(to, from);
415 to->mark = from->mark;
417 /* Copy the flags to each fragment. */
418 IPCB(to)->flags = IPCB(from)->flags;
420 #ifdef CONFIG_NET_SCHED
421 to->tc_index = from->tc_index;
424 #if defined(CONFIG_NETFILTER_XT_TARGET_TRACE) || \
425 defined(CONFIG_NETFILTER_XT_TARGET_TRACE_MODULE)
426 to->nf_trace = from->nf_trace;
428 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
429 to->ipvs_property = from->ipvs_property;
431 skb_copy_secmark(to, from);
435 * This IP datagram is too large to be sent in one piece. Break it up into
436 * smaller pieces (each of size equal to IP header plus
437 * a block of the data of the original IP data part) that will yet fit in a
438 * single device frame, and queue such a frame for sending.
441 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
445 struct net_device *dev;
446 struct sk_buff *skb2;
447 unsigned int mtu, hlen, left, len, ll_rs;
449 __be16 not_last_frag;
450 struct rtable *rt = skb_rtable(skb);
456 * Point into the IP datagram header.
461 if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
462 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
463 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
464 htonl(ip_skb_dst_mtu(skb)));
470 * Setup starting values.
474 mtu = dst_mtu(&rt->dst) - hlen; /* Size of data space */
475 #ifdef CONFIG_BRIDGE_NETFILTER
477 mtu -= nf_bridge_mtu_reduction(skb);
479 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
481 /* When frag_list is given, use it. First, check its validity:
482 * some transformers could create wrong frag_list or break existing
483 * one, it is not prohibited. In this case fall back to copying.
485 * LATER: this step can be merged to real generation of fragments,
486 * we can switch to copy when see the first bad fragment.
488 if (skb_has_frag_list(skb)) {
489 struct sk_buff *frag, *frag2;
490 int first_len = skb_pagelen(skb);
492 if (first_len - hlen > mtu ||
493 ((first_len - hlen) & 7) ||
494 (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
498 skb_walk_frags(skb, frag) {
499 /* Correct geometry. */
500 if (frag->len > mtu ||
501 ((frag->len & 7) && frag->next) ||
502 skb_headroom(frag) < hlen)
503 goto slow_path_clean;
505 /* Partially cloned skb? */
506 if (skb_shared(frag))
507 goto slow_path_clean;
512 frag->destructor = sock_wfree;
514 skb->truesize -= frag->truesize;
517 /* Everything is OK. Generate! */
521 frag = skb_shinfo(skb)->frag_list;
522 skb_frag_list_init(skb);
523 skb->data_len = first_len - skb_headlen(skb);
524 skb->len = first_len;
525 iph->tot_len = htons(first_len);
526 iph->frag_off = htons(IP_MF);
530 /* Prepare header of the next frame,
531 * before previous one went down. */
533 frag->ip_summed = CHECKSUM_NONE;
534 skb_reset_transport_header(frag);
535 __skb_push(frag, hlen);
536 skb_reset_network_header(frag);
537 memcpy(skb_network_header(frag), iph, hlen);
539 iph->tot_len = htons(frag->len);
540 ip_copy_metadata(frag, skb);
542 ip_options_fragment(frag);
543 offset += skb->len - hlen;
544 iph->frag_off = htons(offset>>3);
545 if (frag->next != NULL)
546 iph->frag_off |= htons(IP_MF);
547 /* Ready, complete checksum */
554 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
564 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
573 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
577 skb_walk_frags(skb, frag2) {
581 frag2->destructor = NULL;
582 skb->truesize += frag2->truesize;
587 left = skb->len - hlen; /* Space per frame */
588 ptr = hlen; /* Where to start from */
590 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
591 * we need to make room for the encapsulating header
593 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
596 * Fragment the datagram.
599 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
600 not_last_frag = iph->frag_off & htons(IP_MF);
603 * Keep copying data until we run out.
608 /* IF: it doesn't fit, use 'mtu' - the data space left */
611 /* IF: we are not sending upto and including the packet end
612 then align the next start on an eight byte boundary */
620 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
621 NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
627 * Set up data on packet
630 ip_copy_metadata(skb2, skb);
631 skb_reserve(skb2, ll_rs);
632 skb_put(skb2, len + hlen);
633 skb_reset_network_header(skb2);
634 skb2->transport_header = skb2->network_header + hlen;
637 * Charge the memory for the fragment to any owner
642 skb_set_owner_w(skb2, skb->sk);
645 * Copy the packet header into the new buffer.
648 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
651 * Copy a block of the IP datagram.
653 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
658 * Fill in the new header fields.
661 iph->frag_off = htons((offset >> 3));
663 /* ANK: dirty, but effective trick. Upgrade options only if
664 * the segment to be fragmented was THE FIRST (otherwise,
665 * options are already fixed) and make it ONCE
666 * on the initial skb, so that all the following fragments
667 * will inherit fixed options.
670 ip_options_fragment(skb);
673 * Added AC : If we are fragmenting a fragment that's not the
674 * last fragment then keep MF on each bit
676 if (left > 0 || not_last_frag)
677 iph->frag_off |= htons(IP_MF);
682 * Put this fragment into the sending queue.
684 iph->tot_len = htons(len + hlen);
692 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
695 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
700 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
703 EXPORT_SYMBOL(ip_fragment);
706 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
708 struct iovec *iov = from;
710 if (skb->ip_summed == CHECKSUM_PARTIAL) {
711 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
715 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
717 skb->csum = csum_block_add(skb->csum, csum, odd);
721 EXPORT_SYMBOL(ip_generic_getfrag);
724 csum_page(struct page *page, int offset, int copy)
729 csum = csum_partial(kaddr + offset, copy, 0);
734 static inline int ip_ufo_append_data(struct sock *sk,
735 int getfrag(void *from, char *to, int offset, int len,
736 int odd, struct sk_buff *skb),
737 void *from, int length, int hh_len, int fragheaderlen,
738 int transhdrlen, int mtu, unsigned int flags)
743 /* There is support for UDP fragmentation offload by network
744 * device, so create one single skb packet containing complete
747 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL) {
748 skb = sock_alloc_send_skb(sk,
749 hh_len + fragheaderlen + transhdrlen + 20,
750 (flags & MSG_DONTWAIT), &err);
755 /* reserve space for Hardware header */
756 skb_reserve(skb, hh_len);
758 /* create space for UDP/IP header */
759 skb_put(skb, fragheaderlen + transhdrlen);
761 /* initialize network header pointer */
762 skb_reset_network_header(skb);
764 /* initialize protocol header pointer */
765 skb->transport_header = skb->network_header + fragheaderlen;
767 skb->ip_summed = CHECKSUM_PARTIAL;
769 sk->sk_sndmsg_off = 0;
771 /* specify the length of each IP datagram fragment */
772 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
773 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
774 __skb_queue_tail(&sk->sk_write_queue, skb);
777 return skb_append_datato_frags(sk, skb, getfrag, from,
778 (length - transhdrlen));
782 * ip_append_data() and ip_append_page() can make one large IP datagram
783 * from many pieces of data. Each pieces will be holded on the socket
784 * until ip_push_pending_frames() is called. Each piece can be a page
787 * Not only UDP, other transport protocols - e.g. raw sockets - can use
788 * this interface potentially.
790 * LATER: length must be adjusted by pad at tail, when it is required.
792 int ip_append_data(struct sock *sk,
793 int getfrag(void *from, char *to, int offset, int len,
794 int odd, struct sk_buff *skb),
795 void *from, int length, int transhdrlen,
796 struct ipcm_cookie *ipc, struct rtable **rtp,
799 struct inet_sock *inet = inet_sk(sk);
802 struct ip_options *opt = NULL;
809 unsigned int maxfraglen, fragheaderlen;
810 int csummode = CHECKSUM_NONE;
816 if (skb_queue_empty(&sk->sk_write_queue)) {
822 if (inet->cork.opt == NULL) {
823 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
824 if (unlikely(inet->cork.opt == NULL))
827 memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
828 inet->cork.flags |= IPCORK_OPT;
829 inet->cork.addr = ipc->addr;
835 * We steal reference to this route, caller should not release it
838 inet->cork.fragsize = mtu = inet->pmtudisc == IP_PMTUDISC_PROBE ?
840 dst_mtu(rt->dst.path);
841 inet->cork.dst = &rt->dst;
842 inet->cork.length = 0;
843 sk->sk_sndmsg_page = NULL;
844 sk->sk_sndmsg_off = 0;
845 exthdrlen = rt->dst.header_len;
847 transhdrlen += exthdrlen;
849 rt = (struct rtable *)inet->cork.dst;
850 if (inet->cork.flags & IPCORK_OPT)
851 opt = inet->cork.opt;
855 mtu = inet->cork.fragsize;
857 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
859 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
860 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
862 if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
863 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport,
869 * transhdrlen > 0 means that this is the first fragment and we wish
870 * it won't be fragmented in the future.
873 length + fragheaderlen <= mtu &&
874 rt->dst.dev->features & NETIF_F_V4_CSUM &&
876 csummode = CHECKSUM_PARTIAL;
878 skb = skb_peek_tail(&sk->sk_write_queue);
880 inet->cork.length += length;
881 if (((length > mtu) || (skb && skb_is_gso(skb))) &&
882 (sk->sk_protocol == IPPROTO_UDP) &&
883 (rt->dst.dev->features & NETIF_F_UFO)) {
884 err = ip_ufo_append_data(sk, getfrag, from, length, hh_len,
885 fragheaderlen, transhdrlen, mtu,
892 /* So, what's going on in the loop below?
894 * We use calculated fragment length to generate chained skb,
895 * each of segments is IP fragment ready for sending to network after
896 * adding appropriate IP header.
903 /* Check if the remaining data fits into current packet. */
904 copy = mtu - skb->len;
906 copy = maxfraglen - skb->len;
909 unsigned int datalen;
910 unsigned int fraglen;
911 unsigned int fraggap;
912 unsigned int alloclen;
913 struct sk_buff *skb_prev;
917 fraggap = skb_prev->len - maxfraglen;
922 * If remaining data exceeds the mtu,
923 * we know we need more fragment(s).
925 datalen = length + fraggap;
926 if (datalen > mtu - fragheaderlen)
927 datalen = maxfraglen - fragheaderlen;
928 fraglen = datalen + fragheaderlen;
930 if ((flags & MSG_MORE) &&
931 !(rt->dst.dev->features&NETIF_F_SG))
936 /* The last fragment gets additional space at tail.
937 * Note, with MSG_MORE we overallocate on fragments,
938 * because we have no idea what fragment will be
941 if (datalen == length + fraggap) {
942 alloclen += rt->dst.trailer_len;
943 /* make sure mtu is not reached */
944 if (datalen > mtu - fragheaderlen - rt->dst.trailer_len)
945 datalen -= ALIGN(rt->dst.trailer_len, 8);
948 skb = sock_alloc_send_skb(sk,
949 alloclen + hh_len + 15,
950 (flags & MSG_DONTWAIT), &err);
953 if (atomic_read(&sk->sk_wmem_alloc) <=
955 skb = sock_wmalloc(sk,
956 alloclen + hh_len + 15, 1,
958 if (unlikely(skb == NULL))
961 /* only the initial fragment is
969 * Fill in the control structures
971 skb->ip_summed = csummode;
973 skb_reserve(skb, hh_len);
974 skb_shinfo(skb)->tx_flags = ipc->tx_flags;
977 * Find where to start putting bytes.
979 data = skb_put(skb, fraglen);
980 skb_set_network_header(skb, exthdrlen);
981 skb->transport_header = (skb->network_header +
983 data += fragheaderlen;
986 skb->csum = skb_copy_and_csum_bits(
987 skb_prev, maxfraglen,
988 data + transhdrlen, fraggap, 0);
989 skb_prev->csum = csum_sub(skb_prev->csum,
992 pskb_trim_unique(skb_prev, maxfraglen);
995 copy = datalen - transhdrlen - fraggap;
996 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1003 length -= datalen - fraggap;
1006 csummode = CHECKSUM_NONE;
1009 * Put the packet on the pending queue.
1011 __skb_queue_tail(&sk->sk_write_queue, skb);
1018 if (!(rt->dst.dev->features&NETIF_F_SG)) {
1022 if (getfrag(from, skb_put(skb, copy),
1023 offset, copy, off, skb) < 0) {
1024 __skb_trim(skb, off);
1029 int i = skb_shinfo(skb)->nr_frags;
1030 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
1031 struct page *page = sk->sk_sndmsg_page;
1032 int off = sk->sk_sndmsg_off;
1035 if (page && (left = PAGE_SIZE - off) > 0) {
1038 if (page != frag->page) {
1039 if (i == MAX_SKB_FRAGS) {
1044 skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
1045 frag = &skb_shinfo(skb)->frags[i];
1047 } else if (i < MAX_SKB_FRAGS) {
1048 if (copy > PAGE_SIZE)
1050 page = alloc_pages(sk->sk_allocation, 0);
1055 sk->sk_sndmsg_page = page;
1056 sk->sk_sndmsg_off = 0;
1058 skb_fill_page_desc(skb, i, page, 0, 0);
1059 frag = &skb_shinfo(skb)->frags[i];
1064 if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1068 sk->sk_sndmsg_off += copy;
1071 skb->data_len += copy;
1072 skb->truesize += copy;
1073 atomic_add(copy, &sk->sk_wmem_alloc);
1082 inet->cork.length -= length;
1083 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1087 ssize_t ip_append_page(struct sock *sk, struct page *page,
1088 int offset, size_t size, int flags)
1090 struct inet_sock *inet = inet_sk(sk);
1091 struct sk_buff *skb;
1093 struct ip_options *opt = NULL;
1098 unsigned int maxfraglen, fragheaderlen, fraggap;
1103 if (flags&MSG_PROBE)
1106 if (skb_queue_empty(&sk->sk_write_queue))
1109 rt = (struct rtable *)inet->cork.dst;
1110 if (inet->cork.flags & IPCORK_OPT)
1111 opt = inet->cork.opt;
1113 if (!(rt->dst.dev->features&NETIF_F_SG))
1116 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1117 mtu = inet->cork.fragsize;
1119 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1120 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1122 if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1123 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport, mtu);
1127 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1130 inet->cork.length += size;
1131 if ((size + skb->len > mtu) &&
1132 (sk->sk_protocol == IPPROTO_UDP) &&
1133 (rt->dst.dev->features & NETIF_F_UFO)) {
1134 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1135 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1142 if (skb_is_gso(skb))
1146 /* Check if the remaining data fits into current packet. */
1147 len = mtu - skb->len;
1149 len = maxfraglen - skb->len;
1152 struct sk_buff *skb_prev;
1156 fraggap = skb_prev->len - maxfraglen;
1158 alloclen = fragheaderlen + hh_len + fraggap + 15;
1159 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1160 if (unlikely(!skb)) {
1166 * Fill in the control structures
1168 skb->ip_summed = CHECKSUM_NONE;
1170 skb_reserve(skb, hh_len);
1173 * Find where to start putting bytes.
1175 skb_put(skb, fragheaderlen + fraggap);
1176 skb_reset_network_header(skb);
1177 skb->transport_header = (skb->network_header +
1180 skb->csum = skb_copy_and_csum_bits(skb_prev,
1182 skb_transport_header(skb),
1184 skb_prev->csum = csum_sub(skb_prev->csum,
1186 pskb_trim_unique(skb_prev, maxfraglen);
1190 * Put the packet on the pending queue.
1192 __skb_queue_tail(&sk->sk_write_queue, skb);
1196 i = skb_shinfo(skb)->nr_frags;
1199 if (skb_can_coalesce(skb, i, page, offset)) {
1200 skb_shinfo(skb)->frags[i-1].size += len;
1201 } else if (i < MAX_SKB_FRAGS) {
1203 skb_fill_page_desc(skb, i, page, offset, len);
1209 if (skb->ip_summed == CHECKSUM_NONE) {
1211 csum = csum_page(page, offset, len);
1212 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1216 skb->data_len += len;
1217 skb->truesize += len;
1218 atomic_add(len, &sk->sk_wmem_alloc);
1225 inet->cork.length -= size;
1226 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1230 static void ip_cork_release(struct inet_sock *inet)
1232 inet->cork.flags &= ~IPCORK_OPT;
1233 kfree(inet->cork.opt);
1234 inet->cork.opt = NULL;
1235 dst_release(inet->cork.dst);
1236 inet->cork.dst = NULL;
1240 * Combined all pending IP fragments on the socket as one IP datagram
1241 * and push them out.
1243 int ip_push_pending_frames(struct sock *sk)
1245 struct sk_buff *skb, *tmp_skb;
1246 struct sk_buff **tail_skb;
1247 struct inet_sock *inet = inet_sk(sk);
1248 struct net *net = sock_net(sk);
1249 struct ip_options *opt = NULL;
1250 struct rtable *rt = (struct rtable *)inet->cork.dst;
1256 if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1258 tail_skb = &(skb_shinfo(skb)->frag_list);
1260 /* move skb->data to ip header from ext header */
1261 if (skb->data < skb_network_header(skb))
1262 __skb_pull(skb, skb_network_offset(skb));
1263 while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1264 __skb_pull(tmp_skb, skb_network_header_len(skb));
1265 *tail_skb = tmp_skb;
1266 tail_skb = &(tmp_skb->next);
1267 skb->len += tmp_skb->len;
1268 skb->data_len += tmp_skb->len;
1269 skb->truesize += tmp_skb->truesize;
1270 tmp_skb->destructor = NULL;
1274 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1275 * to fragment the frame generated here. No matter, what transforms
1276 * how transforms change size of the packet, it will come out.
1278 if (inet->pmtudisc < IP_PMTUDISC_DO)
1281 /* DF bit is set when we want to see DF on outgoing frames.
1282 * If local_df is set too, we still allow to fragment this frame
1284 if (inet->pmtudisc >= IP_PMTUDISC_DO ||
1285 (skb->len <= dst_mtu(&rt->dst) &&
1286 ip_dont_fragment(sk, &rt->dst)))
1289 if (inet->cork.flags & IPCORK_OPT)
1290 opt = inet->cork.opt;
1292 if (rt->rt_type == RTN_MULTICAST)
1295 ttl = ip_select_ttl(inet, &rt->dst);
1297 iph = (struct iphdr *)skb->data;
1301 iph->ihl += opt->optlen>>2;
1302 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1304 iph->tos = inet->tos;
1306 ip_select_ident(iph, &rt->dst, sk);
1308 iph->protocol = sk->sk_protocol;
1309 iph->saddr = rt->rt_src;
1310 iph->daddr = rt->rt_dst;
1312 skb->priority = sk->sk_priority;
1313 skb->mark = sk->sk_mark;
1315 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1318 inet->cork.dst = NULL;
1319 skb_dst_set(skb, &rt->dst);
1321 if (iph->protocol == IPPROTO_ICMP)
1322 icmp_out_count(net, ((struct icmphdr *)
1323 skb_transport_header(skb))->type);
1325 /* Netfilter gets whole the not fragmented skb. */
1326 err = ip_local_out(skb);
1329 err = net_xmit_errno(err);
1335 ip_cork_release(inet);
1339 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1344 * Throw away all pending data on the socket.
1346 void ip_flush_pending_frames(struct sock *sk)
1348 struct sk_buff *skb;
1350 while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1353 ip_cork_release(inet_sk(sk));
1358 * Fetch data from kernel space and fill in checksum if needed.
1360 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1361 int len, int odd, struct sk_buff *skb)
1365 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1366 skb->csum = csum_block_add(skb->csum, csum, odd);
1371 * Generic function to send a packet as reply to another packet.
1372 * Used to send TCP resets so far. ICMP should use this function too.
1374 * Should run single threaded per socket because it uses the sock
1375 * structure to pass arguments.
1377 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1380 struct inet_sock *inet = inet_sk(sk);
1382 struct ip_options opt;
1385 struct ipcm_cookie ipc;
1387 struct rtable *rt = skb_rtable(skb);
1389 if (ip_options_echo(&replyopts.opt, skb))
1392 daddr = ipc.addr = rt->rt_src;
1396 if (replyopts.opt.optlen) {
1397 ipc.opt = &replyopts.opt;
1400 daddr = replyopts.opt.faddr;
1404 struct flowi fl = { .oif = arg->bound_dev_if,
1406 .fl4_src = rt->rt_spec_dst,
1407 .fl4_tos = RT_TOS(ip_hdr(skb)->tos),
1408 .fl_ip_sport = tcp_hdr(skb)->dest,
1409 .fl_ip_dport = tcp_hdr(skb)->source,
1410 .proto = sk->sk_protocol,
1411 .flags = ip_reply_arg_flowi_flags(arg) };
1412 security_skb_classify_flow(skb, &fl);
1413 if (ip_route_output_key(sock_net(sk), &rt, &fl))
1417 /* And let IP do all the hard work.
1419 This chunk is not reenterable, hence spinlock.
1420 Note that it uses the fact, that this function is called
1421 with locally disabled BH and that sk cannot be already spinlocked.
1424 inet->tos = ip_hdr(skb)->tos;
1425 sk->sk_priority = skb->priority;
1426 sk->sk_protocol = ip_hdr(skb)->protocol;
1427 sk->sk_bound_dev_if = arg->bound_dev_if;
1428 ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1429 &ipc, &rt, MSG_DONTWAIT);
1430 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1431 if (arg->csumoffset >= 0)
1432 *((__sum16 *)skb_transport_header(skb) +
1433 arg->csumoffset) = csum_fold(csum_add(skb->csum,
1435 skb->ip_summed = CHECKSUM_NONE;
1436 ip_push_pending_frames(sk);
1444 void __init ip_init(void)
1449 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1450 igmp_mc_proc_init();