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_sk(struct sock *sk, struct sk_buff *skb)
101 struct net *net = dev_net(skb_dst(skb)->dev);
102 struct iphdr *iph = ip_hdr(skb);
104 iph->tot_len = htons(skb->len);
106 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
107 net, sk, skb, NULL, skb_dst(skb)->dev,
111 int __ip_local_out(struct sk_buff *skb)
113 return __ip_local_out_sk(skb->sk, skb);
116 int ip_local_out_sk(struct sock *sk, struct sk_buff *skb)
118 struct net *net = dev_net(skb_dst(skb)->dev);
121 err = __ip_local_out_sk(sk, skb);
122 if (likely(err == 1))
123 err = dst_output(net, sk, skb);
127 EXPORT_SYMBOL_GPL(ip_local_out_sk);
129 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
131 int ttl = inet->uc_ttl;
134 ttl = ip4_dst_hoplimit(dst);
139 * Add an ip header to a skbuff and send it out.
142 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
143 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
145 struct inet_sock *inet = inet_sk(sk);
146 struct rtable *rt = skb_rtable(skb);
149 /* Build the IP header. */
150 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
151 skb_reset_network_header(skb);
155 iph->tos = inet->tos;
156 iph->ttl = ip_select_ttl(inet, &rt->dst);
157 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
159 iph->protocol = sk->sk_protocol;
160 if (ip_dont_fragment(sk, &rt->dst)) {
161 iph->frag_off = htons(IP_DF);
165 __ip_select_ident(sock_net(sk), iph, 1);
168 if (opt && opt->opt.optlen) {
169 iph->ihl += opt->opt.optlen>>2;
170 ip_options_build(skb, &opt->opt, daddr, rt, 0);
173 skb->priority = sk->sk_priority;
174 skb->mark = sk->sk_mark;
177 return ip_local_out(skb);
179 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
181 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
183 struct dst_entry *dst = skb_dst(skb);
184 struct rtable *rt = (struct rtable *)dst;
185 struct net_device *dev = dst->dev;
186 unsigned int hh_len = LL_RESERVED_SPACE(dev);
187 struct neighbour *neigh;
190 if (rt->rt_type == RTN_MULTICAST) {
191 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
192 } else if (rt->rt_type == RTN_BROADCAST)
193 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
195 /* Be paranoid, rather than too clever. */
196 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
197 struct sk_buff *skb2;
199 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
205 skb_set_owner_w(skb2, skb->sk);
211 nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
212 neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
213 if (unlikely(!neigh))
214 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
215 if (!IS_ERR(neigh)) {
216 int res = dst_neigh_output(dst, neigh, skb);
218 rcu_read_unlock_bh();
221 rcu_read_unlock_bh();
223 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
229 static int ip_finish_output_gso(struct net *net, struct sock *sk,
230 struct sk_buff *skb, unsigned int mtu)
232 netdev_features_t features;
233 struct sk_buff *segs;
236 /* common case: locally created skb or seglen is <= mtu */
237 if (((IPCB(skb)->flags & IPSKB_FORWARDED) == 0) ||
238 skb_gso_network_seglen(skb) <= mtu)
239 return ip_finish_output2(net, sk, skb);
241 /* Slowpath - GSO segment length is exceeding the dst MTU.
243 * This can happen in two cases:
244 * 1) TCP GRO packet, DF bit not set
245 * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
246 * from host network stack.
248 features = netif_skb_features(skb);
249 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
250 if (IS_ERR_OR_NULL(segs)) {
258 struct sk_buff *nskb = segs->next;
262 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
272 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
276 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
277 /* Policy lookup after SNAT yielded a new policy */
278 if (skb_dst(skb)->xfrm) {
279 IPCB(skb)->flags |= IPSKB_REROUTED;
280 return dst_output(net, sk, skb);
283 mtu = ip_skb_dst_mtu(skb);
285 return ip_finish_output_gso(net, sk, skb, mtu);
287 if (skb->len > mtu || (IPCB(skb)->flags & IPSKB_FRAG_PMTU))
288 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
290 return ip_finish_output2(net, sk, skb);
293 int ip_mc_output(struct sock *sk, struct sk_buff *skb)
295 struct rtable *rt = skb_rtable(skb);
296 struct net_device *dev = rt->dst.dev;
297 struct net *net = dev_net(dev);
300 * If the indicated interface is up and running, send the packet.
302 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
305 skb->protocol = htons(ETH_P_IP);
308 * Multicasts are looped back for other local users
311 if (rt->rt_flags&RTCF_MULTICAST) {
313 #ifdef CONFIG_IP_MROUTE
314 /* Small optimization: do not loopback not local frames,
315 which returned after forwarding; they will be dropped
316 by ip_mr_input in any case.
317 Note, that local frames are looped back to be delivered
320 This check is duplicated in ip_mr_input at the moment.
323 ((rt->rt_flags & RTCF_LOCAL) ||
324 !(IPCB(skb)->flags & IPSKB_FORWARDED))
327 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
329 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
330 net, sk, newskb, NULL, newskb->dev,
334 /* Multicasts with ttl 0 must not go beyond the host */
336 if (ip_hdr(skb)->ttl == 0) {
342 if (rt->rt_flags&RTCF_BROADCAST) {
343 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
345 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
346 net, sk, newskb, NULL, newskb->dev,
350 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
351 net, sk, skb, NULL, skb->dev,
353 !(IPCB(skb)->flags & IPSKB_REROUTED));
356 int ip_output(struct sock *sk, struct sk_buff *skb)
358 struct net_device *dev = skb_dst(skb)->dev;
359 struct net *net = dev_net(dev);
361 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
364 skb->protocol = htons(ETH_P_IP);
366 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
367 net, sk, skb, NULL, dev,
369 !(IPCB(skb)->flags & IPSKB_REROUTED));
373 * copy saddr and daddr, possibly using 64bit load/stores
375 * iph->saddr = fl4->saddr;
376 * iph->daddr = fl4->daddr;
378 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
380 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
381 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
382 memcpy(&iph->saddr, &fl4->saddr,
383 sizeof(fl4->saddr) + sizeof(fl4->daddr));
386 /* Note: skb->sk can be different from sk, in case of tunnels */
387 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
389 struct inet_sock *inet = inet_sk(sk);
390 struct ip_options_rcu *inet_opt;
396 /* Skip all of this if the packet is already routed,
397 * f.e. by something like SCTP.
400 inet_opt = rcu_dereference(inet->inet_opt);
402 rt = skb_rtable(skb);
406 /* Make sure we can route this packet. */
407 rt = (struct rtable *)__sk_dst_check(sk, 0);
411 /* Use correct destination address if we have options. */
412 daddr = inet->inet_daddr;
413 if (inet_opt && inet_opt->opt.srr)
414 daddr = inet_opt->opt.faddr;
416 /* If this fails, retransmit mechanism of transport layer will
417 * keep trying until route appears or the connection times
420 rt = ip_route_output_ports(sock_net(sk), fl4, sk,
421 daddr, inet->inet_saddr,
426 sk->sk_bound_dev_if);
429 sk_setup_caps(sk, &rt->dst);
431 skb_dst_set_noref(skb, &rt->dst);
434 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
437 /* OK, we know where to send it, allocate and build IP header. */
438 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
439 skb_reset_network_header(skb);
441 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
442 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
443 iph->frag_off = htons(IP_DF);
446 iph->ttl = ip_select_ttl(inet, &rt->dst);
447 iph->protocol = sk->sk_protocol;
448 ip_copy_addrs(iph, fl4);
450 /* Transport layer set skb->h.foo itself. */
452 if (inet_opt && inet_opt->opt.optlen) {
453 iph->ihl += inet_opt->opt.optlen >> 2;
454 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
457 ip_select_ident_segs(sock_net(sk), skb, sk,
458 skb_shinfo(skb)->gso_segs ?: 1);
460 /* TODO : should we use skb->sk here instead of sk ? */
461 skb->priority = sk->sk_priority;
462 skb->mark = sk->sk_mark;
464 res = ip_local_out_sk(sk, skb);
470 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
472 return -EHOSTUNREACH;
474 EXPORT_SYMBOL(ip_queue_xmit);
476 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
478 to->pkt_type = from->pkt_type;
479 to->priority = from->priority;
480 to->protocol = from->protocol;
482 skb_dst_copy(to, from);
484 to->mark = from->mark;
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);
544 * Point into the IP datagram header.
549 mtu = ip_skb_dst_mtu(skb);
550 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
551 mtu = IPCB(skb)->frag_max_size;
554 * Setup starting values.
558 mtu = mtu - hlen; /* Size of data space */
559 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
561 /* When frag_list is given, use it. First, check its validity:
562 * some transformers could create wrong frag_list or break existing
563 * one, it is not prohibited. In this case fall back to copying.
565 * LATER: this step can be merged to real generation of fragments,
566 * we can switch to copy when see the first bad fragment.
568 if (skb_has_frag_list(skb)) {
569 struct sk_buff *frag, *frag2;
570 int first_len = skb_pagelen(skb);
572 if (first_len - hlen > mtu ||
573 ((first_len - hlen) & 7) ||
574 ip_is_fragment(iph) ||
578 skb_walk_frags(skb, frag) {
579 /* Correct geometry. */
580 if (frag->len > mtu ||
581 ((frag->len & 7) && frag->next) ||
582 skb_headroom(frag) < hlen)
583 goto slow_path_clean;
585 /* Partially cloned skb? */
586 if (skb_shared(frag))
587 goto slow_path_clean;
592 frag->destructor = sock_wfree;
594 skb->truesize -= frag->truesize;
597 /* Everything is OK. Generate! */
601 frag = skb_shinfo(skb)->frag_list;
602 skb_frag_list_init(skb);
603 skb->data_len = first_len - skb_headlen(skb);
604 skb->len = first_len;
605 iph->tot_len = htons(first_len);
606 iph->frag_off = htons(IP_MF);
610 /* Prepare header of the next frame,
611 * before previous one went down. */
613 frag->ip_summed = CHECKSUM_NONE;
614 skb_reset_transport_header(frag);
615 __skb_push(frag, hlen);
616 skb_reset_network_header(frag);
617 memcpy(skb_network_header(frag), iph, hlen);
619 iph->tot_len = htons(frag->len);
620 ip_copy_metadata(frag, skb);
622 ip_options_fragment(frag);
623 offset += skb->len - hlen;
624 iph->frag_off = htons(offset>>3);
626 iph->frag_off |= htons(IP_MF);
627 /* Ready, complete checksum */
631 err = output(net, sk, skb);
634 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
644 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
653 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
657 skb_walk_frags(skb, frag2) {
661 frag2->destructor = NULL;
662 skb->truesize += frag2->truesize;
667 /* for offloaded checksums cleanup checksum before fragmentation */
668 if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
672 left = skb->len - hlen; /* Space per frame */
673 ptr = hlen; /* Where to start from */
675 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
678 * Fragment the datagram.
681 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
682 not_last_frag = iph->frag_off & htons(IP_MF);
685 * Keep copying data until we run out.
690 /* IF: it doesn't fit, use 'mtu' - the data space left */
693 /* IF: we are not sending up to and including the packet end
694 then align the next start on an eight byte boundary */
699 /* Allocate buffer */
700 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
707 * Set up data on packet
710 ip_copy_metadata(skb2, skb);
711 skb_reserve(skb2, ll_rs);
712 skb_put(skb2, len + hlen);
713 skb_reset_network_header(skb2);
714 skb2->transport_header = skb2->network_header + hlen;
717 * Charge the memory for the fragment to any owner
722 skb_set_owner_w(skb2, skb->sk);
725 * Copy the packet header into the new buffer.
728 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
731 * Copy a block of the IP datagram.
733 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
738 * Fill in the new header fields.
741 iph->frag_off = htons((offset >> 3));
743 if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
744 iph->frag_off |= htons(IP_DF);
746 /* ANK: dirty, but effective trick. Upgrade options only if
747 * the segment to be fragmented was THE FIRST (otherwise,
748 * options are already fixed) and make it ONCE
749 * on the initial skb, so that all the following fragments
750 * will inherit fixed options.
753 ip_options_fragment(skb);
756 * Added AC : If we are fragmenting a fragment that's not the
757 * last fragment then keep MF on each bit
759 if (left > 0 || not_last_frag)
760 iph->frag_off |= htons(IP_MF);
765 * Put this fragment into the sending queue.
767 iph->tot_len = htons(len + hlen);
771 err = output(net, sk, skb2);
775 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
778 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
783 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
786 EXPORT_SYMBOL(ip_do_fragment);
789 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
791 struct msghdr *msg = from;
793 if (skb->ip_summed == CHECKSUM_PARTIAL) {
794 if (copy_from_iter(to, len, &msg->msg_iter) != len)
798 if (csum_and_copy_from_iter(to, len, &csum, &msg->msg_iter) != len)
800 skb->csum = csum_block_add(skb->csum, csum, odd);
804 EXPORT_SYMBOL(ip_generic_getfrag);
807 csum_page(struct page *page, int offset, int copy)
812 csum = csum_partial(kaddr + offset, copy, 0);
817 static inline int ip_ufo_append_data(struct sock *sk,
818 struct sk_buff_head *queue,
819 int getfrag(void *from, char *to, int offset, int len,
820 int odd, struct sk_buff *skb),
821 void *from, int length, int hh_len, int fragheaderlen,
822 int transhdrlen, int maxfraglen, unsigned int flags)
827 /* There is support for UDP fragmentation offload by network
828 * device, so create one single skb packet containing complete
831 skb = skb_peek_tail(queue);
833 skb = sock_alloc_send_skb(sk,
834 hh_len + fragheaderlen + transhdrlen + 20,
835 (flags & MSG_DONTWAIT), &err);
840 /* reserve space for Hardware header */
841 skb_reserve(skb, hh_len);
843 /* create space for UDP/IP header */
844 skb_put(skb, fragheaderlen + transhdrlen);
846 /* initialize network header pointer */
847 skb_reset_network_header(skb);
849 /* initialize protocol header pointer */
850 skb->transport_header = skb->network_header + fragheaderlen;
854 __skb_queue_tail(queue, skb);
855 } else if (skb_is_gso(skb)) {
859 skb->ip_summed = CHECKSUM_PARTIAL;
860 /* specify the length of each IP datagram fragment */
861 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
862 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
865 return skb_append_datato_frags(sk, skb, getfrag, from,
866 (length - transhdrlen));
869 static int __ip_append_data(struct sock *sk,
871 struct sk_buff_head *queue,
872 struct inet_cork *cork,
873 struct page_frag *pfrag,
874 int getfrag(void *from, char *to, int offset,
875 int len, int odd, struct sk_buff *skb),
876 void *from, int length, int transhdrlen,
879 struct inet_sock *inet = inet_sk(sk);
882 struct ip_options *opt = cork->opt;
889 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
890 int csummode = CHECKSUM_NONE;
891 struct rtable *rt = (struct rtable *)cork->dst;
894 skb = skb_peek_tail(queue);
896 exthdrlen = !skb ? rt->dst.header_len : 0;
897 mtu = cork->fragsize;
898 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
899 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
900 tskey = sk->sk_tskey++;
902 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
904 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
905 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
906 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
908 if (cork->length + length > maxnonfragsize - fragheaderlen) {
909 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
910 mtu - (opt ? opt->optlen : 0));
915 * transhdrlen > 0 means that this is the first fragment and we wish
916 * it won't be fragmented in the future.
919 length + fragheaderlen <= mtu &&
920 rt->dst.dev->features & NETIF_F_V4_CSUM &&
922 csummode = CHECKSUM_PARTIAL;
924 cork->length += length;
925 if (((length > mtu) || (skb && skb_is_gso(skb))) &&
926 (sk->sk_protocol == IPPROTO_UDP) &&
927 (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len &&
928 (sk->sk_type == SOCK_DGRAM)) {
929 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
930 hh_len, fragheaderlen, transhdrlen,
937 /* So, what's going on in the loop below?
939 * We use calculated fragment length to generate chained skb,
940 * each of segments is IP fragment ready for sending to network after
941 * adding appropriate IP header.
948 /* Check if the remaining data fits into current packet. */
949 copy = mtu - skb->len;
951 copy = maxfraglen - skb->len;
954 unsigned int datalen;
955 unsigned int fraglen;
956 unsigned int fraggap;
957 unsigned int alloclen;
958 struct sk_buff *skb_prev;
962 fraggap = skb_prev->len - maxfraglen;
967 * If remaining data exceeds the mtu,
968 * we know we need more fragment(s).
970 datalen = length + fraggap;
971 if (datalen > mtu - fragheaderlen)
972 datalen = maxfraglen - fragheaderlen;
973 fraglen = datalen + fragheaderlen;
975 if ((flags & MSG_MORE) &&
976 !(rt->dst.dev->features&NETIF_F_SG))
981 alloclen += exthdrlen;
983 /* The last fragment gets additional space at tail.
984 * Note, with MSG_MORE we overallocate on fragments,
985 * because we have no idea what fragment will be
988 if (datalen == length + fraggap)
989 alloclen += rt->dst.trailer_len;
992 skb = sock_alloc_send_skb(sk,
993 alloclen + hh_len + 15,
994 (flags & MSG_DONTWAIT), &err);
997 if (atomic_read(&sk->sk_wmem_alloc) <=
999 skb = sock_wmalloc(sk,
1000 alloclen + hh_len + 15, 1,
1009 * Fill in the control structures
1011 skb->ip_summed = csummode;
1013 skb_reserve(skb, hh_len);
1015 /* only the initial fragment is time stamped */
1016 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1018 skb_shinfo(skb)->tskey = tskey;
1022 * Find where to start putting bytes.
1024 data = skb_put(skb, fraglen + exthdrlen);
1025 skb_set_network_header(skb, exthdrlen);
1026 skb->transport_header = (skb->network_header +
1028 data += fragheaderlen + exthdrlen;
1031 skb->csum = skb_copy_and_csum_bits(
1032 skb_prev, maxfraglen,
1033 data + transhdrlen, fraggap, 0);
1034 skb_prev->csum = csum_sub(skb_prev->csum,
1037 pskb_trim_unique(skb_prev, maxfraglen);
1040 copy = datalen - transhdrlen - fraggap;
1041 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1048 length -= datalen - fraggap;
1051 csummode = CHECKSUM_NONE;
1054 * Put the packet on the pending queue.
1056 __skb_queue_tail(queue, skb);
1063 if (!(rt->dst.dev->features&NETIF_F_SG)) {
1067 if (getfrag(from, skb_put(skb, copy),
1068 offset, copy, off, skb) < 0) {
1069 __skb_trim(skb, off);
1074 int i = skb_shinfo(skb)->nr_frags;
1077 if (!sk_page_frag_refill(sk, pfrag))
1080 if (!skb_can_coalesce(skb, i, pfrag->page,
1083 if (i == MAX_SKB_FRAGS)
1086 __skb_fill_page_desc(skb, i, pfrag->page,
1088 skb_shinfo(skb)->nr_frags = ++i;
1089 get_page(pfrag->page);
1091 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1093 page_address(pfrag->page) + pfrag->offset,
1094 offset, copy, skb->len, skb) < 0)
1097 pfrag->offset += copy;
1098 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1100 skb->data_len += copy;
1101 skb->truesize += copy;
1102 atomic_add(copy, &sk->sk_wmem_alloc);
1113 cork->length -= length;
1114 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1118 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1119 struct ipcm_cookie *ipc, struct rtable **rtp)
1121 struct ip_options_rcu *opt;
1125 * setup for corking.
1130 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1132 if (unlikely(!cork->opt))
1135 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1136 cork->flags |= IPCORK_OPT;
1137 cork->addr = ipc->addr;
1143 * We steal reference to this route, caller should not release it
1146 cork->fragsize = ip_sk_use_pmtu(sk) ?
1147 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1148 cork->dst = &rt->dst;
1150 cork->ttl = ipc->ttl;
1151 cork->tos = ipc->tos;
1152 cork->priority = ipc->priority;
1153 cork->tx_flags = ipc->tx_flags;
1159 * ip_append_data() and ip_append_page() can make one large IP datagram
1160 * from many pieces of data. Each pieces will be holded on the socket
1161 * until ip_push_pending_frames() is called. Each piece can be a page
1164 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1165 * this interface potentially.
1167 * LATER: length must be adjusted by pad at tail, when it is required.
1169 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1170 int getfrag(void *from, char *to, int offset, int len,
1171 int odd, struct sk_buff *skb),
1172 void *from, int length, int transhdrlen,
1173 struct ipcm_cookie *ipc, struct rtable **rtp,
1176 struct inet_sock *inet = inet_sk(sk);
1179 if (flags&MSG_PROBE)
1182 if (skb_queue_empty(&sk->sk_write_queue)) {
1183 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1190 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1191 sk_page_frag(sk), getfrag,
1192 from, length, transhdrlen, flags);
1195 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1196 int offset, size_t size, int flags)
1198 struct inet_sock *inet = inet_sk(sk);
1199 struct sk_buff *skb;
1201 struct ip_options *opt = NULL;
1202 struct inet_cork *cork;
1207 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1212 if (flags&MSG_PROBE)
1215 if (skb_queue_empty(&sk->sk_write_queue))
1218 cork = &inet->cork.base;
1219 rt = (struct rtable *)cork->dst;
1220 if (cork->flags & IPCORK_OPT)
1223 if (!(rt->dst.dev->features&NETIF_F_SG))
1226 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1227 mtu = cork->fragsize;
1229 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1230 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1231 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1233 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1234 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1235 mtu - (opt ? opt->optlen : 0));
1239 skb = skb_peek_tail(&sk->sk_write_queue);
1243 cork->length += size;
1244 if ((size + skb->len > mtu) &&
1245 (sk->sk_protocol == IPPROTO_UDP) &&
1246 (rt->dst.dev->features & NETIF_F_UFO)) {
1247 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1248 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1252 if (skb_is_gso(skb)) {
1256 /* Check if the remaining data fits into current packet. */
1257 len = mtu - skb->len;
1259 len = maxfraglen - skb->len;
1262 struct sk_buff *skb_prev;
1266 fraggap = skb_prev->len - maxfraglen;
1268 alloclen = fragheaderlen + hh_len + fraggap + 15;
1269 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1270 if (unlikely(!skb)) {
1276 * Fill in the control structures
1278 skb->ip_summed = CHECKSUM_NONE;
1280 skb_reserve(skb, hh_len);
1283 * Find where to start putting bytes.
1285 skb_put(skb, fragheaderlen + fraggap);
1286 skb_reset_network_header(skb);
1287 skb->transport_header = (skb->network_header +
1290 skb->csum = skb_copy_and_csum_bits(skb_prev,
1292 skb_transport_header(skb),
1294 skb_prev->csum = csum_sub(skb_prev->csum,
1296 pskb_trim_unique(skb_prev, maxfraglen);
1300 * Put the packet on the pending queue.
1302 __skb_queue_tail(&sk->sk_write_queue, skb);
1309 if (skb_append_pagefrags(skb, page, offset, len)) {
1314 if (skb->ip_summed == CHECKSUM_NONE) {
1316 csum = csum_page(page, offset, len);
1317 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1321 skb->data_len += len;
1322 skb->truesize += len;
1323 atomic_add(len, &sk->sk_wmem_alloc);
1330 cork->length -= size;
1331 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1335 static void ip_cork_release(struct inet_cork *cork)
1337 cork->flags &= ~IPCORK_OPT;
1340 dst_release(cork->dst);
1345 * Combined all pending IP fragments on the socket as one IP datagram
1346 * and push them out.
1348 struct sk_buff *__ip_make_skb(struct sock *sk,
1350 struct sk_buff_head *queue,
1351 struct inet_cork *cork)
1353 struct sk_buff *skb, *tmp_skb;
1354 struct sk_buff **tail_skb;
1355 struct inet_sock *inet = inet_sk(sk);
1356 struct net *net = sock_net(sk);
1357 struct ip_options *opt = NULL;
1358 struct rtable *rt = (struct rtable *)cork->dst;
1363 skb = __skb_dequeue(queue);
1366 tail_skb = &(skb_shinfo(skb)->frag_list);
1368 /* move skb->data to ip header from ext header */
1369 if (skb->data < skb_network_header(skb))
1370 __skb_pull(skb, skb_network_offset(skb));
1371 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1372 __skb_pull(tmp_skb, skb_network_header_len(skb));
1373 *tail_skb = tmp_skb;
1374 tail_skb = &(tmp_skb->next);
1375 skb->len += tmp_skb->len;
1376 skb->data_len += tmp_skb->len;
1377 skb->truesize += tmp_skb->truesize;
1378 tmp_skb->destructor = NULL;
1382 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1383 * to fragment the frame generated here. No matter, what transforms
1384 * how transforms change size of the packet, it will come out.
1386 skb->ignore_df = ip_sk_ignore_df(sk);
1388 /* DF bit is set when we want to see DF on outgoing frames.
1389 * If ignore_df is set too, we still allow to fragment this frame
1391 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1392 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1393 (skb->len <= dst_mtu(&rt->dst) &&
1394 ip_dont_fragment(sk, &rt->dst)))
1397 if (cork->flags & IPCORK_OPT)
1402 else if (rt->rt_type == RTN_MULTICAST)
1405 ttl = ip_select_ttl(inet, &rt->dst);
1410 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1413 iph->protocol = sk->sk_protocol;
1414 ip_copy_addrs(iph, fl4);
1415 ip_select_ident(net, skb, sk);
1418 iph->ihl += opt->optlen>>2;
1419 ip_options_build(skb, opt, cork->addr, rt, 0);
1422 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1423 skb->mark = sk->sk_mark;
1425 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1429 skb_dst_set(skb, &rt->dst);
1431 if (iph->protocol == IPPROTO_ICMP)
1432 icmp_out_count(net, ((struct icmphdr *)
1433 skb_transport_header(skb))->type);
1435 ip_cork_release(cork);
1440 int ip_send_skb(struct net *net, struct sk_buff *skb)
1444 err = ip_local_out(skb);
1447 err = net_xmit_errno(err);
1449 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1455 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1457 struct sk_buff *skb;
1459 skb = ip_finish_skb(sk, fl4);
1463 /* Netfilter gets whole the not fragmented skb. */
1464 return ip_send_skb(sock_net(sk), skb);
1468 * Throw away all pending data on the socket.
1470 static void __ip_flush_pending_frames(struct sock *sk,
1471 struct sk_buff_head *queue,
1472 struct inet_cork *cork)
1474 struct sk_buff *skb;
1476 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1479 ip_cork_release(cork);
1482 void ip_flush_pending_frames(struct sock *sk)
1484 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1487 struct sk_buff *ip_make_skb(struct sock *sk,
1489 int getfrag(void *from, char *to, int offset,
1490 int len, int odd, struct sk_buff *skb),
1491 void *from, int length, int transhdrlen,
1492 struct ipcm_cookie *ipc, struct rtable **rtp,
1495 struct inet_cork cork;
1496 struct sk_buff_head queue;
1499 if (flags & MSG_PROBE)
1502 __skb_queue_head_init(&queue);
1507 err = ip_setup_cork(sk, &cork, ipc, rtp);
1509 return ERR_PTR(err);
1511 err = __ip_append_data(sk, fl4, &queue, &cork,
1512 ¤t->task_frag, getfrag,
1513 from, length, transhdrlen, flags);
1515 __ip_flush_pending_frames(sk, &queue, &cork);
1516 return ERR_PTR(err);
1519 return __ip_make_skb(sk, fl4, &queue, &cork);
1523 * Fetch data from kernel space and fill in checksum if needed.
1525 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1526 int len, int odd, struct sk_buff *skb)
1530 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1531 skb->csum = csum_block_add(skb->csum, csum, odd);
1536 * Generic function to send a packet as reply to another packet.
1537 * Used to send some TCP resets/acks so far.
1539 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1540 const struct ip_options *sopt,
1541 __be32 daddr, __be32 saddr,
1542 const struct ip_reply_arg *arg,
1545 struct ip_options_data replyopts;
1546 struct ipcm_cookie ipc;
1548 struct rtable *rt = skb_rtable(skb);
1549 struct net *net = sock_net(sk);
1550 struct sk_buff *nskb;
1554 if (__ip_options_echo(&replyopts.opt.opt, skb, sopt))
1563 if (replyopts.opt.opt.optlen) {
1564 ipc.opt = &replyopts.opt;
1566 if (replyopts.opt.opt.srr)
1567 daddr = replyopts.opt.opt.faddr;
1570 oif = arg->bound_dev_if;
1571 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1574 flowi4_init_output(&fl4, oif,
1575 IP4_REPLY_MARK(net, skb->mark),
1577 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1578 ip_reply_arg_flowi_flags(arg),
1580 tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1581 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1582 rt = ip_route_output_key(net, &fl4);
1586 inet_sk(sk)->tos = arg->tos;
1588 sk->sk_priority = skb->priority;
1589 sk->sk_protocol = ip_hdr(skb)->protocol;
1590 sk->sk_bound_dev_if = arg->bound_dev_if;
1591 sk->sk_sndbuf = sysctl_wmem_default;
1592 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1593 len, 0, &ipc, &rt, MSG_DONTWAIT);
1594 if (unlikely(err)) {
1595 ip_flush_pending_frames(sk);
1599 nskb = skb_peek(&sk->sk_write_queue);
1601 if (arg->csumoffset >= 0)
1602 *((__sum16 *)skb_transport_header(nskb) +
1603 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1605 nskb->ip_summed = CHECKSUM_NONE;
1606 skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1607 ip_push_pending_frames(sk, &fl4);
1613 void __init ip_init(void)
1618 #if defined(CONFIG_IP_MULTICAST)