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 <linux/module.h>
47 #include <linux/types.h>
48 #include <linux/kernel.h>
50 #include <linux/string.h>
51 #include <linux/errno.h>
52 #include <linux/highmem.h>
53 #include <linux/slab.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;
84 EXPORT_SYMBOL(sysctl_ip_default_ttl);
87 ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
89 int (*output)(struct net *, struct sock *, struct sk_buff *));
91 /* Generate a checksum for an outgoing IP datagram. */
92 void ip_send_check(struct iphdr *iph)
95 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
97 EXPORT_SYMBOL(ip_send_check);
99 int __ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
101 struct iphdr *iph = ip_hdr(skb);
103 iph->tot_len = htons(skb->len);
106 skb->protocol = htons(ETH_P_IP);
108 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
109 net, sk, skb, NULL, skb_dst(skb)->dev,
113 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
117 err = __ip_local_out(net, sk, skb);
118 if (likely(err == 1))
119 err = dst_output(net, sk, skb);
123 EXPORT_SYMBOL_GPL(ip_local_out);
125 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
127 int ttl = inet->uc_ttl;
130 ttl = ip4_dst_hoplimit(dst);
135 * Add an ip header to a skbuff and send it out.
138 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
139 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
141 struct inet_sock *inet = inet_sk(sk);
142 struct rtable *rt = skb_rtable(skb);
143 struct net *net = sock_net(sk);
146 /* Build the IP header. */
147 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
148 skb_reset_network_header(skb);
152 iph->tos = inet->tos;
153 iph->ttl = ip_select_ttl(inet, &rt->dst);
154 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
156 iph->protocol = sk->sk_protocol;
157 if (ip_dont_fragment(sk, &rt->dst)) {
158 iph->frag_off = htons(IP_DF);
162 __ip_select_ident(net, iph, 1);
165 if (opt && opt->opt.optlen) {
166 iph->ihl += opt->opt.optlen>>2;
167 ip_options_build(skb, &opt->opt, daddr, rt, 0);
170 skb->priority = sk->sk_priority;
171 skb->mark = sk->sk_mark;
174 return ip_local_out(net, skb->sk, skb);
176 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
178 static int ip_finish_output2(struct net *net, struct sock *sk, 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);
184 struct neighbour *neigh;
187 if (rt->rt_type == RTN_MULTICAST) {
188 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
189 } else if (rt->rt_type == RTN_BROADCAST)
190 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
192 /* Be paranoid, rather than too clever. */
193 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
194 struct sk_buff *skb2;
196 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
202 skb_set_owner_w(skb2, skb->sk);
208 nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
209 neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
210 if (unlikely(!neigh))
211 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
212 if (!IS_ERR(neigh)) {
213 int res = dst_neigh_output(dst, neigh, skb);
215 rcu_read_unlock_bh();
218 rcu_read_unlock_bh();
220 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
226 static int ip_finish_output_gso(struct net *net, struct sock *sk,
227 struct sk_buff *skb, unsigned int mtu)
229 netdev_features_t features;
230 struct sk_buff *segs;
233 /* common case: locally created skb or seglen is <= mtu */
234 if (((IPCB(skb)->flags & IPSKB_FORWARDED) == 0) ||
235 skb_gso_network_seglen(skb) <= mtu)
236 return ip_finish_output2(net, sk, skb);
238 /* Slowpath - GSO segment length is exceeding the dst MTU.
240 * This can happen in two cases:
241 * 1) TCP GRO packet, DF bit not set
242 * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
243 * from host network stack.
245 features = netif_skb_features(skb);
246 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_SGO_CB_OFFSET);
247 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
248 if (IS_ERR_OR_NULL(segs)) {
256 struct sk_buff *nskb = segs->next;
260 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
270 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
274 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
275 /* Policy lookup after SNAT yielded a new policy */
276 if (skb_dst(skb)->xfrm) {
277 IPCB(skb)->flags |= IPSKB_REROUTED;
278 return dst_output(net, sk, skb);
281 mtu = ip_skb_dst_mtu(skb);
283 return ip_finish_output_gso(net, sk, skb, mtu);
285 if (skb->len > mtu || (IPCB(skb)->flags & IPSKB_FRAG_PMTU))
286 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
288 return ip_finish_output2(net, sk, skb);
291 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
293 struct rtable *rt = skb_rtable(skb);
294 struct net_device *dev = rt->dst.dev;
297 * If the indicated interface is up and running, send the packet.
299 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
302 skb->protocol = htons(ETH_P_IP);
305 * Multicasts are looped back for other local users
308 if (rt->rt_flags&RTCF_MULTICAST) {
310 #ifdef CONFIG_IP_MROUTE
311 /* Small optimization: do not loopback not local frames,
312 which returned after forwarding; they will be dropped
313 by ip_mr_input in any case.
314 Note, that local frames are looped back to be delivered
317 This check is duplicated in ip_mr_input at the moment.
320 ((rt->rt_flags & RTCF_LOCAL) ||
321 !(IPCB(skb)->flags & IPSKB_FORWARDED))
324 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
326 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
327 net, sk, newskb, NULL, newskb->dev,
331 /* Multicasts with ttl 0 must not go beyond the host */
333 if (ip_hdr(skb)->ttl == 0) {
339 if (rt->rt_flags&RTCF_BROADCAST) {
340 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
342 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
343 net, sk, newskb, NULL, newskb->dev,
347 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
348 net, sk, skb, NULL, skb->dev,
350 !(IPCB(skb)->flags & IPSKB_REROUTED));
353 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
355 struct net_device *dev = skb_dst(skb)->dev;
357 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
360 skb->protocol = htons(ETH_P_IP);
362 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
363 net, sk, skb, NULL, dev,
365 !(IPCB(skb)->flags & IPSKB_REROUTED));
369 * copy saddr and daddr, possibly using 64bit load/stores
371 * iph->saddr = fl4->saddr;
372 * iph->daddr = fl4->daddr;
374 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
376 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
377 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
378 memcpy(&iph->saddr, &fl4->saddr,
379 sizeof(fl4->saddr) + sizeof(fl4->daddr));
382 /* Note: skb->sk can be different from sk, in case of tunnels */
383 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
385 struct inet_sock *inet = inet_sk(sk);
386 struct net *net = sock_net(sk);
387 struct ip_options_rcu *inet_opt;
393 /* Skip all of this if the packet is already routed,
394 * f.e. by something like SCTP.
397 inet_opt = rcu_dereference(inet->inet_opt);
399 rt = skb_rtable(skb);
403 /* Make sure we can route this packet. */
404 rt = (struct rtable *)__sk_dst_check(sk, 0);
408 /* Use correct destination address if we have options. */
409 daddr = inet->inet_daddr;
410 if (inet_opt && inet_opt->opt.srr)
411 daddr = inet_opt->opt.faddr;
413 /* If this fails, retransmit mechanism of transport layer will
414 * keep trying until route appears or the connection times
417 rt = ip_route_output_ports(net, fl4, sk,
418 daddr, inet->inet_saddr,
423 sk->sk_bound_dev_if);
426 sk_setup_caps(sk, &rt->dst);
428 skb_dst_set_noref(skb, &rt->dst);
431 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
434 /* OK, we know where to send it, allocate and build IP header. */
435 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
436 skb_reset_network_header(skb);
438 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
439 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
440 iph->frag_off = htons(IP_DF);
443 iph->ttl = ip_select_ttl(inet, &rt->dst);
444 iph->protocol = sk->sk_protocol;
445 ip_copy_addrs(iph, fl4);
447 /* Transport layer set skb->h.foo itself. */
449 if (inet_opt && inet_opt->opt.optlen) {
450 iph->ihl += inet_opt->opt.optlen >> 2;
451 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
454 ip_select_ident_segs(net, skb, sk,
455 skb_shinfo(skb)->gso_segs ?: 1);
457 /* TODO : should we use skb->sk here instead of sk ? */
458 skb->priority = sk->sk_priority;
459 skb->mark = sk->sk_mark;
461 res = ip_local_out(net, sk, skb);
467 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
469 return -EHOSTUNREACH;
471 EXPORT_SYMBOL(ip_queue_xmit);
473 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
475 to->pkt_type = from->pkt_type;
476 to->priority = from->priority;
477 to->protocol = from->protocol;
478 to->skb_iif = from->skb_iif;
480 skb_dst_copy(to, from);
482 to->mark = from->mark;
484 skb_copy_hash(to, from);
486 /* Copy the flags to each fragment. */
487 IPCB(to)->flags = IPCB(from)->flags;
489 #ifdef CONFIG_NET_SCHED
490 to->tc_index = from->tc_index;
493 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
494 to->ipvs_property = from->ipvs_property;
496 skb_copy_secmark(to, from);
499 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
501 int (*output)(struct net *, struct sock *, struct sk_buff *))
503 struct iphdr *iph = ip_hdr(skb);
505 if ((iph->frag_off & htons(IP_DF)) == 0)
506 return ip_do_fragment(net, sk, skb, output);
508 if (unlikely(!skb->ignore_df ||
509 (IPCB(skb)->frag_max_size &&
510 IPCB(skb)->frag_max_size > mtu))) {
511 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
512 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
518 return ip_do_fragment(net, sk, skb, output);
522 * This IP datagram is too large to be sent in one piece. Break it up into
523 * smaller pieces (each of size equal to IP header plus
524 * a block of the data of the original IP data part) that will yet fit in a
525 * single device frame, and queue such a frame for sending.
528 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
529 int (*output)(struct net *, struct sock *, struct sk_buff *))
533 struct net_device *dev;
534 struct sk_buff *skb2;
535 unsigned int mtu, hlen, left, len, ll_rs;
537 __be16 not_last_frag;
538 struct rtable *rt = skb_rtable(skb);
543 /* for offloaded checksums cleanup checksum before fragmentation */
544 if (skb->ip_summed == CHECKSUM_PARTIAL &&
545 (err = skb_checksum_help(skb)))
549 * Point into the IP datagram header.
554 mtu = ip_skb_dst_mtu(skb);
555 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
556 mtu = IPCB(skb)->frag_max_size;
559 * Setup starting values.
563 mtu = mtu - hlen; /* Size of data space */
564 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
566 /* When frag_list is given, use it. First, check its validity:
567 * some transformers could create wrong frag_list or break existing
568 * one, it is not prohibited. In this case fall back to copying.
570 * LATER: this step can be merged to real generation of fragments,
571 * we can switch to copy when see the first bad fragment.
573 if (skb_has_frag_list(skb)) {
574 struct sk_buff *frag, *frag2;
575 int first_len = skb_pagelen(skb);
577 if (first_len - hlen > mtu ||
578 ((first_len - hlen) & 7) ||
579 ip_is_fragment(iph) ||
583 skb_walk_frags(skb, frag) {
584 /* Correct geometry. */
585 if (frag->len > mtu ||
586 ((frag->len & 7) && frag->next) ||
587 skb_headroom(frag) < hlen)
588 goto slow_path_clean;
590 /* Partially cloned skb? */
591 if (skb_shared(frag))
592 goto slow_path_clean;
597 frag->destructor = sock_wfree;
599 skb->truesize -= frag->truesize;
602 /* Everything is OK. Generate! */
606 frag = skb_shinfo(skb)->frag_list;
607 skb_frag_list_init(skb);
608 skb->data_len = first_len - skb_headlen(skb);
609 skb->len = first_len;
610 iph->tot_len = htons(first_len);
611 iph->frag_off = htons(IP_MF);
615 /* Prepare header of the next frame,
616 * before previous one went down. */
618 frag->ip_summed = CHECKSUM_NONE;
619 skb_reset_transport_header(frag);
620 __skb_push(frag, hlen);
621 skb_reset_network_header(frag);
622 memcpy(skb_network_header(frag), iph, hlen);
624 iph->tot_len = htons(frag->len);
625 ip_copy_metadata(frag, skb);
627 ip_options_fragment(frag);
628 offset += skb->len - hlen;
629 iph->frag_off = htons(offset>>3);
631 iph->frag_off |= htons(IP_MF);
632 /* Ready, complete checksum */
636 err = output(net, sk, skb);
639 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
649 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
658 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
662 skb_walk_frags(skb, frag2) {
666 frag2->destructor = NULL;
667 skb->truesize += frag2->truesize;
674 left = skb->len - hlen; /* Space per frame */
675 ptr = hlen; /* Where to start from */
677 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
680 * Fragment the datagram.
683 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
684 not_last_frag = iph->frag_off & htons(IP_MF);
687 * Keep copying data until we run out.
692 /* IF: it doesn't fit, use 'mtu' - the data space left */
695 /* IF: we are not sending up to and including the packet end
696 then align the next start on an eight byte boundary */
701 /* Allocate buffer */
702 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
709 * Set up data on packet
712 ip_copy_metadata(skb2, skb);
713 skb_reserve(skb2, ll_rs);
714 skb_put(skb2, len + hlen);
715 skb_reset_network_header(skb2);
716 skb2->transport_header = skb2->network_header + hlen;
719 * Charge the memory for the fragment to any owner
724 skb_set_owner_w(skb2, skb->sk);
727 * Copy the packet header into the new buffer.
730 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
733 * Copy a block of the IP datagram.
735 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
740 * Fill in the new header fields.
743 iph->frag_off = htons((offset >> 3));
745 if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
746 iph->frag_off |= htons(IP_DF);
748 /* ANK: dirty, but effective trick. Upgrade options only if
749 * the segment to be fragmented was THE FIRST (otherwise,
750 * options are already fixed) and make it ONCE
751 * on the initial skb, so that all the following fragments
752 * will inherit fixed options.
755 ip_options_fragment(skb);
758 * Added AC : If we are fragmenting a fragment that's not the
759 * last fragment then keep MF on each bit
761 if (left > 0 || not_last_frag)
762 iph->frag_off |= htons(IP_MF);
767 * Put this fragment into the sending queue.
769 iph->tot_len = htons(len + hlen);
773 err = output(net, sk, skb2);
777 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
780 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
785 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
788 EXPORT_SYMBOL(ip_do_fragment);
791 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
793 struct msghdr *msg = from;
795 if (skb->ip_summed == CHECKSUM_PARTIAL) {
796 if (copy_from_iter(to, len, &msg->msg_iter) != len)
800 if (csum_and_copy_from_iter(to, len, &csum, &msg->msg_iter) != len)
802 skb->csum = csum_block_add(skb->csum, csum, odd);
806 EXPORT_SYMBOL(ip_generic_getfrag);
809 csum_page(struct page *page, int offset, int copy)
814 csum = csum_partial(kaddr + offset, copy, 0);
819 static inline int ip_ufo_append_data(struct sock *sk,
820 struct sk_buff_head *queue,
821 int getfrag(void *from, char *to, int offset, int len,
822 int odd, struct sk_buff *skb),
823 void *from, int length, int hh_len, int fragheaderlen,
824 int transhdrlen, int maxfraglen, unsigned int flags)
829 /* There is support for UDP fragmentation offload by network
830 * device, so create one single skb packet containing complete
833 skb = skb_peek_tail(queue);
835 skb = sock_alloc_send_skb(sk,
836 hh_len + fragheaderlen + transhdrlen + 20,
837 (flags & MSG_DONTWAIT), &err);
842 /* reserve space for Hardware header */
843 skb_reserve(skb, hh_len);
845 /* create space for UDP/IP header */
846 skb_put(skb, fragheaderlen + transhdrlen);
848 /* initialize network header pointer */
849 skb_reset_network_header(skb);
851 /* initialize protocol header pointer */
852 skb->transport_header = skb->network_header + fragheaderlen;
856 __skb_queue_tail(queue, skb);
857 } else if (skb_is_gso(skb)) {
861 skb->ip_summed = CHECKSUM_PARTIAL;
862 /* specify the length of each IP datagram fragment */
863 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
864 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
867 return skb_append_datato_frags(sk, skb, getfrag, from,
868 (length - transhdrlen));
871 static int __ip_append_data(struct sock *sk,
873 struct sk_buff_head *queue,
874 struct inet_cork *cork,
875 struct page_frag *pfrag,
876 int getfrag(void *from, char *to, int offset,
877 int len, int odd, struct sk_buff *skb),
878 void *from, int length, int transhdrlen,
881 struct inet_sock *inet = inet_sk(sk);
884 struct ip_options *opt = cork->opt;
891 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
892 int csummode = CHECKSUM_NONE;
893 struct rtable *rt = (struct rtable *)cork->dst;
896 skb = skb_peek_tail(queue);
898 exthdrlen = !skb ? rt->dst.header_len : 0;
899 mtu = cork->fragsize;
900 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
901 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
902 tskey = sk->sk_tskey++;
904 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
906 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
907 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
908 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
910 if (cork->length + length > maxnonfragsize - fragheaderlen) {
911 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
912 mtu - (opt ? opt->optlen : 0));
917 * transhdrlen > 0 means that this is the first fragment and we wish
918 * it won't be fragmented in the future.
921 length + fragheaderlen <= mtu &&
922 rt->dst.dev->features & NETIF_F_V4_CSUM &&
923 !(flags & MSG_MORE) &&
925 csummode = CHECKSUM_PARTIAL;
927 cork->length += length;
928 if ((skb && skb_is_gso(skb)) ||
929 (((length + (skb ? skb->len : fragheaderlen)) > mtu) &&
930 (skb_queue_len(queue) <= 1) &&
931 (sk->sk_protocol == IPPROTO_UDP) &&
932 (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len &&
933 (sk->sk_type == SOCK_DGRAM) && !sk->sk_no_check_tx)) {
934 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
935 hh_len, fragheaderlen, transhdrlen,
942 /* So, what's going on in the loop below?
944 * We use calculated fragment length to generate chained skb,
945 * each of segments is IP fragment ready for sending to network after
946 * adding appropriate IP header.
953 /* Check if the remaining data fits into current packet. */
954 copy = mtu - skb->len;
956 copy = maxfraglen - skb->len;
959 unsigned int datalen;
960 unsigned int fraglen;
961 unsigned int fraggap;
962 unsigned int alloclen;
963 struct sk_buff *skb_prev;
967 fraggap = skb_prev->len - maxfraglen;
972 * If remaining data exceeds the mtu,
973 * we know we need more fragment(s).
975 datalen = length + fraggap;
976 if (datalen > mtu - fragheaderlen)
977 datalen = maxfraglen - fragheaderlen;
978 fraglen = datalen + fragheaderlen;
980 if ((flags & MSG_MORE) &&
981 !(rt->dst.dev->features&NETIF_F_SG))
986 alloclen += exthdrlen;
988 /* The last fragment gets additional space at tail.
989 * Note, with MSG_MORE we overallocate on fragments,
990 * because we have no idea what fragment will be
993 if (datalen == length + fraggap)
994 alloclen += rt->dst.trailer_len;
997 skb = sock_alloc_send_skb(sk,
998 alloclen + hh_len + 15,
999 (flags & MSG_DONTWAIT), &err);
1002 if (atomic_read(&sk->sk_wmem_alloc) <=
1004 skb = sock_wmalloc(sk,
1005 alloclen + hh_len + 15, 1,
1014 * Fill in the control structures
1016 skb->ip_summed = csummode;
1018 skb_reserve(skb, hh_len);
1020 /* only the initial fragment is time stamped */
1021 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1023 skb_shinfo(skb)->tskey = tskey;
1027 * Find where to start putting bytes.
1029 data = skb_put(skb, fraglen + exthdrlen);
1030 skb_set_network_header(skb, exthdrlen);
1031 skb->transport_header = (skb->network_header +
1033 data += fragheaderlen + exthdrlen;
1036 skb->csum = skb_copy_and_csum_bits(
1037 skb_prev, maxfraglen,
1038 data + transhdrlen, fraggap, 0);
1039 skb_prev->csum = csum_sub(skb_prev->csum,
1042 pskb_trim_unique(skb_prev, maxfraglen);
1045 copy = datalen - transhdrlen - fraggap;
1046 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1053 length -= datalen - fraggap;
1056 csummode = CHECKSUM_NONE;
1059 * Put the packet on the pending queue.
1061 __skb_queue_tail(queue, skb);
1068 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1069 skb_tailroom(skb) >= copy) {
1073 if (getfrag(from, skb_put(skb, copy),
1074 offset, copy, off, skb) < 0) {
1075 __skb_trim(skb, off);
1080 int i = skb_shinfo(skb)->nr_frags;
1083 if (!sk_page_frag_refill(sk, pfrag))
1086 if (!skb_can_coalesce(skb, i, pfrag->page,
1089 if (i == MAX_SKB_FRAGS)
1092 __skb_fill_page_desc(skb, i, pfrag->page,
1094 skb_shinfo(skb)->nr_frags = ++i;
1095 get_page(pfrag->page);
1097 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1099 page_address(pfrag->page) + pfrag->offset,
1100 offset, copy, skb->len, skb) < 0)
1103 pfrag->offset += copy;
1104 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1106 skb->data_len += copy;
1107 skb->truesize += copy;
1108 atomic_add(copy, &sk->sk_wmem_alloc);
1119 cork->length -= length;
1120 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1124 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1125 struct ipcm_cookie *ipc, struct rtable **rtp)
1127 struct ip_options_rcu *opt;
1131 * setup for corking.
1136 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1138 if (unlikely(!cork->opt))
1141 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1142 cork->flags |= IPCORK_OPT;
1143 cork->addr = ipc->addr;
1149 * We steal reference to this route, caller should not release it
1152 cork->fragsize = ip_sk_use_pmtu(sk) ?
1153 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1154 cork->dst = &rt->dst;
1156 cork->ttl = ipc->ttl;
1157 cork->tos = ipc->tos;
1158 cork->priority = ipc->priority;
1159 cork->tx_flags = ipc->tx_flags;
1165 * ip_append_data() and ip_append_page() can make one large IP datagram
1166 * from many pieces of data. Each pieces will be holded on the socket
1167 * until ip_push_pending_frames() is called. Each piece can be a page
1170 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1171 * this interface potentially.
1173 * LATER: length must be adjusted by pad at tail, when it is required.
1175 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1176 int getfrag(void *from, char *to, int offset, int len,
1177 int odd, struct sk_buff *skb),
1178 void *from, int length, int transhdrlen,
1179 struct ipcm_cookie *ipc, struct rtable **rtp,
1182 struct inet_sock *inet = inet_sk(sk);
1185 if (flags&MSG_PROBE)
1188 if (skb_queue_empty(&sk->sk_write_queue)) {
1189 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1196 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1197 sk_page_frag(sk), getfrag,
1198 from, length, transhdrlen, flags);
1201 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1202 int offset, size_t size, int flags)
1204 struct inet_sock *inet = inet_sk(sk);
1205 struct sk_buff *skb;
1207 struct ip_options *opt = NULL;
1208 struct inet_cork *cork;
1213 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1218 if (flags&MSG_PROBE)
1221 if (skb_queue_empty(&sk->sk_write_queue))
1224 cork = &inet->cork.base;
1225 rt = (struct rtable *)cork->dst;
1226 if (cork->flags & IPCORK_OPT)
1229 if (!(rt->dst.dev->features&NETIF_F_SG))
1232 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1233 mtu = cork->fragsize;
1235 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1236 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1237 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1239 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1240 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1241 mtu - (opt ? opt->optlen : 0));
1245 skb = skb_peek_tail(&sk->sk_write_queue);
1249 if ((size + skb->len > mtu) &&
1250 (skb_queue_len(&sk->sk_write_queue) == 1) &&
1251 (sk->sk_protocol == IPPROTO_UDP) &&
1252 (rt->dst.dev->features & NETIF_F_UFO)) {
1253 if (skb->ip_summed != CHECKSUM_PARTIAL)
1256 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1257 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1259 cork->length += size;
1262 if (skb_is_gso(skb)) {
1266 /* Check if the remaining data fits into current packet. */
1267 len = mtu - skb->len;
1269 len = maxfraglen - skb->len;
1272 struct sk_buff *skb_prev;
1276 fraggap = skb_prev->len - maxfraglen;
1278 alloclen = fragheaderlen + hh_len + fraggap + 15;
1279 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1280 if (unlikely(!skb)) {
1286 * Fill in the control structures
1288 skb->ip_summed = CHECKSUM_NONE;
1290 skb_reserve(skb, hh_len);
1293 * Find where to start putting bytes.
1295 skb_put(skb, fragheaderlen + fraggap);
1296 skb_reset_network_header(skb);
1297 skb->transport_header = (skb->network_header +
1300 skb->csum = skb_copy_and_csum_bits(skb_prev,
1302 skb_transport_header(skb),
1304 skb_prev->csum = csum_sub(skb_prev->csum,
1306 pskb_trim_unique(skb_prev, maxfraglen);
1310 * Put the packet on the pending queue.
1312 __skb_queue_tail(&sk->sk_write_queue, skb);
1319 if (skb_append_pagefrags(skb, page, offset, len)) {
1324 if (skb->ip_summed == CHECKSUM_NONE) {
1326 csum = csum_page(page, offset, len);
1327 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1331 skb->data_len += len;
1332 skb->truesize += len;
1333 atomic_add(len, &sk->sk_wmem_alloc);
1340 cork->length -= size;
1341 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1345 static void ip_cork_release(struct inet_cork *cork)
1347 cork->flags &= ~IPCORK_OPT;
1350 dst_release(cork->dst);
1355 * Combined all pending IP fragments on the socket as one IP datagram
1356 * and push them out.
1358 struct sk_buff *__ip_make_skb(struct sock *sk,
1360 struct sk_buff_head *queue,
1361 struct inet_cork *cork)
1363 struct sk_buff *skb, *tmp_skb;
1364 struct sk_buff **tail_skb;
1365 struct inet_sock *inet = inet_sk(sk);
1366 struct net *net = sock_net(sk);
1367 struct ip_options *opt = NULL;
1368 struct rtable *rt = (struct rtable *)cork->dst;
1373 skb = __skb_dequeue(queue);
1376 tail_skb = &(skb_shinfo(skb)->frag_list);
1378 /* move skb->data to ip header from ext header */
1379 if (skb->data < skb_network_header(skb))
1380 __skb_pull(skb, skb_network_offset(skb));
1381 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1382 __skb_pull(tmp_skb, skb_network_header_len(skb));
1383 *tail_skb = tmp_skb;
1384 tail_skb = &(tmp_skb->next);
1385 skb->len += tmp_skb->len;
1386 skb->data_len += tmp_skb->len;
1387 skb->truesize += tmp_skb->truesize;
1388 tmp_skb->destructor = NULL;
1392 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1393 * to fragment the frame generated here. No matter, what transforms
1394 * how transforms change size of the packet, it will come out.
1396 skb->ignore_df = ip_sk_ignore_df(sk);
1398 /* DF bit is set when we want to see DF on outgoing frames.
1399 * If ignore_df is set too, we still allow to fragment this frame
1401 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1402 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1403 (skb->len <= dst_mtu(&rt->dst) &&
1404 ip_dont_fragment(sk, &rt->dst)))
1407 if (cork->flags & IPCORK_OPT)
1412 else if (rt->rt_type == RTN_MULTICAST)
1415 ttl = ip_select_ttl(inet, &rt->dst);
1420 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1423 iph->protocol = sk->sk_protocol;
1424 ip_copy_addrs(iph, fl4);
1425 ip_select_ident(net, skb, sk);
1428 iph->ihl += opt->optlen>>2;
1429 ip_options_build(skb, opt, cork->addr, rt, 0);
1432 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1433 skb->mark = sk->sk_mark;
1435 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1439 skb_dst_set(skb, &rt->dst);
1441 if (iph->protocol == IPPROTO_ICMP)
1442 icmp_out_count(net, ((struct icmphdr *)
1443 skb_transport_header(skb))->type);
1445 ip_cork_release(cork);
1450 int ip_send_skb(struct net *net, struct sk_buff *skb)
1454 err = ip_local_out(net, skb->sk, skb);
1457 err = net_xmit_errno(err);
1459 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1465 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1467 struct sk_buff *skb;
1469 skb = ip_finish_skb(sk, fl4);
1473 /* Netfilter gets whole the not fragmented skb. */
1474 return ip_send_skb(sock_net(sk), skb);
1478 * Throw away all pending data on the socket.
1480 static void __ip_flush_pending_frames(struct sock *sk,
1481 struct sk_buff_head *queue,
1482 struct inet_cork *cork)
1484 struct sk_buff *skb;
1486 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1489 ip_cork_release(cork);
1492 void ip_flush_pending_frames(struct sock *sk)
1494 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1497 struct sk_buff *ip_make_skb(struct sock *sk,
1499 int getfrag(void *from, char *to, int offset,
1500 int len, int odd, struct sk_buff *skb),
1501 void *from, int length, int transhdrlen,
1502 struct ipcm_cookie *ipc, struct rtable **rtp,
1505 struct inet_cork cork;
1506 struct sk_buff_head queue;
1509 if (flags & MSG_PROBE)
1512 __skb_queue_head_init(&queue);
1517 err = ip_setup_cork(sk, &cork, ipc, rtp);
1519 return ERR_PTR(err);
1521 err = __ip_append_data(sk, fl4, &queue, &cork,
1522 ¤t->task_frag, getfrag,
1523 from, length, transhdrlen, flags);
1525 __ip_flush_pending_frames(sk, &queue, &cork);
1526 return ERR_PTR(err);
1529 return __ip_make_skb(sk, fl4, &queue, &cork);
1533 * Fetch data from kernel space and fill in checksum if needed.
1535 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1536 int len, int odd, struct sk_buff *skb)
1540 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1541 skb->csum = csum_block_add(skb->csum, csum, odd);
1546 * Generic function to send a packet as reply to another packet.
1547 * Used to send some TCP resets/acks so far.
1549 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1550 const struct ip_options *sopt,
1551 __be32 daddr, __be32 saddr,
1552 const struct ip_reply_arg *arg,
1555 struct ip_options_data replyopts;
1556 struct ipcm_cookie ipc;
1558 struct rtable *rt = skb_rtable(skb);
1559 struct net *net = sock_net(sk);
1560 struct sk_buff *nskb;
1564 if (__ip_options_echo(&replyopts.opt.opt, skb, sopt))
1573 if (replyopts.opt.opt.optlen) {
1574 ipc.opt = &replyopts.opt;
1576 if (replyopts.opt.opt.srr)
1577 daddr = replyopts.opt.opt.faddr;
1580 oif = arg->bound_dev_if;
1581 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1584 flowi4_init_output(&fl4, oif,
1585 IP4_REPLY_MARK(net, skb->mark),
1587 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1588 ip_reply_arg_flowi_flags(arg),
1590 tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1592 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1593 rt = ip_route_output_key(net, &fl4);
1597 inet_sk(sk)->tos = arg->tos;
1599 sk->sk_priority = skb->priority;
1600 sk->sk_protocol = ip_hdr(skb)->protocol;
1601 sk->sk_bound_dev_if = arg->bound_dev_if;
1602 sk->sk_sndbuf = sysctl_wmem_default;
1603 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1604 len, 0, &ipc, &rt, MSG_DONTWAIT);
1605 if (unlikely(err)) {
1606 ip_flush_pending_frames(sk);
1610 nskb = skb_peek(&sk->sk_write_queue);
1612 if (arg->csumoffset >= 0)
1613 *((__sum16 *)skb_transport_header(nskb) +
1614 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1616 nskb->ip_summed = CHECKSUM_NONE;
1617 ip_push_pending_frames(sk, &fl4);
1623 void __init ip_init(void)
1628 #if defined(CONFIG_IP_MULTICAST)