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);
105 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
106 net, sk, skb, NULL, skb_dst(skb)->dev,
110 int ip_local_out(struct sock *sk, struct sk_buff *skb)
112 struct net *net = dev_net(skb_dst(skb)->dev);
115 err = __ip_local_out(net, sk, skb);
116 if (likely(err == 1))
117 err = dst_output(net, sk, skb);
121 EXPORT_SYMBOL_GPL(ip_local_out);
123 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
125 int ttl = inet->uc_ttl;
128 ttl = ip4_dst_hoplimit(dst);
133 * Add an ip header to a skbuff and send it out.
136 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
137 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
139 struct inet_sock *inet = inet_sk(sk);
140 struct rtable *rt = skb_rtable(skb);
141 struct net *net = sock_net(sk);
144 /* Build the IP header. */
145 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
146 skb_reset_network_header(skb);
150 iph->tos = inet->tos;
151 iph->ttl = ip_select_ttl(inet, &rt->dst);
152 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
154 iph->protocol = sk->sk_protocol;
155 if (ip_dont_fragment(sk, &rt->dst)) {
156 iph->frag_off = htons(IP_DF);
160 __ip_select_ident(net, iph, 1);
163 if (opt && opt->opt.optlen) {
164 iph->ihl += opt->opt.optlen>>2;
165 ip_options_build(skb, &opt->opt, daddr, rt, 0);
168 skb->priority = sk->sk_priority;
169 skb->mark = sk->sk_mark;
172 return ip_local_out(skb->sk, skb);
174 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
176 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
178 struct dst_entry *dst = skb_dst(skb);
179 struct rtable *rt = (struct rtable *)dst;
180 struct net_device *dev = dst->dev;
181 unsigned int hh_len = LL_RESERVED_SPACE(dev);
182 struct neighbour *neigh;
185 if (rt->rt_type == RTN_MULTICAST) {
186 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
187 } else if (rt->rt_type == RTN_BROADCAST)
188 IP_UPD_PO_STATS(net, 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 nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
207 neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
208 if (unlikely(!neigh))
209 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
210 if (!IS_ERR(neigh)) {
211 int res = dst_neigh_output(dst, neigh, skb);
213 rcu_read_unlock_bh();
216 rcu_read_unlock_bh();
218 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
224 static int ip_finish_output_gso(struct net *net, struct sock *sk,
225 struct sk_buff *skb, unsigned int mtu)
227 netdev_features_t features;
228 struct sk_buff *segs;
231 /* common case: locally created skb or seglen is <= mtu */
232 if (((IPCB(skb)->flags & IPSKB_FORWARDED) == 0) ||
233 skb_gso_network_seglen(skb) <= mtu)
234 return ip_finish_output2(net, sk, skb);
236 /* Slowpath - GSO segment length is exceeding the dst MTU.
238 * This can happen in two cases:
239 * 1) TCP GRO packet, DF bit not set
240 * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
241 * from host network stack.
243 features = netif_skb_features(skb);
244 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
245 if (IS_ERR_OR_NULL(segs)) {
253 struct sk_buff *nskb = segs->next;
257 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
267 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
271 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
272 /* Policy lookup after SNAT yielded a new policy */
273 if (skb_dst(skb)->xfrm) {
274 IPCB(skb)->flags |= IPSKB_REROUTED;
275 return dst_output(net, sk, skb);
278 mtu = ip_skb_dst_mtu(skb);
280 return ip_finish_output_gso(net, sk, skb, mtu);
282 if (skb->len > mtu || (IPCB(skb)->flags & IPSKB_FRAG_PMTU))
283 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
285 return ip_finish_output2(net, sk, skb);
288 int ip_mc_output(struct sock *sk, struct sk_buff *skb)
290 struct rtable *rt = skb_rtable(skb);
291 struct net_device *dev = rt->dst.dev;
292 struct net *net = dev_net(dev);
295 * If the indicated interface is up and running, send the packet.
297 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
300 skb->protocol = htons(ETH_P_IP);
303 * Multicasts are looped back for other local users
306 if (rt->rt_flags&RTCF_MULTICAST) {
308 #ifdef CONFIG_IP_MROUTE
309 /* Small optimization: do not loopback not local frames,
310 which returned after forwarding; they will be dropped
311 by ip_mr_input in any case.
312 Note, that local frames are looped back to be delivered
315 This check is duplicated in ip_mr_input at the moment.
318 ((rt->rt_flags & RTCF_LOCAL) ||
319 !(IPCB(skb)->flags & IPSKB_FORWARDED))
322 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
324 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
325 net, sk, newskb, NULL, newskb->dev,
329 /* Multicasts with ttl 0 must not go beyond the host */
331 if (ip_hdr(skb)->ttl == 0) {
337 if (rt->rt_flags&RTCF_BROADCAST) {
338 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
340 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
341 net, sk, newskb, NULL, newskb->dev,
345 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
346 net, sk, skb, NULL, skb->dev,
348 !(IPCB(skb)->flags & IPSKB_REROUTED));
351 int ip_output(struct sock *sk, struct sk_buff *skb)
353 struct net_device *dev = skb_dst(skb)->dev;
354 struct net *net = dev_net(dev);
356 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
359 skb->protocol = htons(ETH_P_IP);
361 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
362 net, sk, skb, NULL, dev,
364 !(IPCB(skb)->flags & IPSKB_REROUTED));
368 * copy saddr and daddr, possibly using 64bit load/stores
370 * iph->saddr = fl4->saddr;
371 * iph->daddr = fl4->daddr;
373 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
375 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
376 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
377 memcpy(&iph->saddr, &fl4->saddr,
378 sizeof(fl4->saddr) + sizeof(fl4->daddr));
381 /* Note: skb->sk can be different from sk, in case of tunnels */
382 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
384 struct inet_sock *inet = inet_sk(sk);
385 struct net *net = sock_net(sk);
386 struct ip_options_rcu *inet_opt;
392 /* Skip all of this if the packet is already routed,
393 * f.e. by something like SCTP.
396 inet_opt = rcu_dereference(inet->inet_opt);
398 rt = skb_rtable(skb);
402 /* Make sure we can route this packet. */
403 rt = (struct rtable *)__sk_dst_check(sk, 0);
407 /* Use correct destination address if we have options. */
408 daddr = inet->inet_daddr;
409 if (inet_opt && inet_opt->opt.srr)
410 daddr = inet_opt->opt.faddr;
412 /* If this fails, retransmit mechanism of transport layer will
413 * keep trying until route appears or the connection times
416 rt = ip_route_output_ports(net, fl4, sk,
417 daddr, inet->inet_saddr,
422 sk->sk_bound_dev_if);
425 sk_setup_caps(sk, &rt->dst);
427 skb_dst_set_noref(skb, &rt->dst);
430 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
433 /* OK, we know where to send it, allocate and build IP header. */
434 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
435 skb_reset_network_header(skb);
437 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
438 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
439 iph->frag_off = htons(IP_DF);
442 iph->ttl = ip_select_ttl(inet, &rt->dst);
443 iph->protocol = sk->sk_protocol;
444 ip_copy_addrs(iph, fl4);
446 /* Transport layer set skb->h.foo itself. */
448 if (inet_opt && inet_opt->opt.optlen) {
449 iph->ihl += inet_opt->opt.optlen >> 2;
450 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
453 ip_select_ident_segs(net, skb, sk,
454 skb_shinfo(skb)->gso_segs ?: 1);
456 /* TODO : should we use skb->sk here instead of sk ? */
457 skb->priority = sk->sk_priority;
458 skb->mark = sk->sk_mark;
460 res = ip_local_out(sk, skb);
466 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
468 return -EHOSTUNREACH;
470 EXPORT_SYMBOL(ip_queue_xmit);
472 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
474 to->pkt_type = from->pkt_type;
475 to->priority = from->priority;
476 to->protocol = from->protocol;
478 skb_dst_copy(to, from);
480 to->mark = from->mark;
482 /* Copy the flags to each fragment. */
483 IPCB(to)->flags = IPCB(from)->flags;
485 #ifdef CONFIG_NET_SCHED
486 to->tc_index = from->tc_index;
489 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
490 to->ipvs_property = from->ipvs_property;
492 skb_copy_secmark(to, from);
495 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
497 int (*output)(struct net *, struct sock *, struct sk_buff *))
499 struct iphdr *iph = ip_hdr(skb);
501 if ((iph->frag_off & htons(IP_DF)) == 0)
502 return ip_do_fragment(net, sk, skb, output);
504 if (unlikely(!skb->ignore_df ||
505 (IPCB(skb)->frag_max_size &&
506 IPCB(skb)->frag_max_size > mtu))) {
507 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
508 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
514 return ip_do_fragment(net, sk, skb, output);
518 * This IP datagram is too large to be sent in one piece. Break it up into
519 * smaller pieces (each of size equal to IP header plus
520 * a block of the data of the original IP data part) that will yet fit in a
521 * single device frame, and queue such a frame for sending.
524 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
525 int (*output)(struct net *, struct sock *, struct sk_buff *))
529 struct net_device *dev;
530 struct sk_buff *skb2;
531 unsigned int mtu, hlen, left, len, ll_rs;
533 __be16 not_last_frag;
534 struct rtable *rt = skb_rtable(skb);
540 * Point into the IP datagram header.
545 mtu = ip_skb_dst_mtu(skb);
546 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
547 mtu = IPCB(skb)->frag_max_size;
550 * Setup starting values.
554 mtu = mtu - hlen; /* Size of data space */
555 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
557 /* When frag_list is given, use it. First, check its validity:
558 * some transformers could create wrong frag_list or break existing
559 * one, it is not prohibited. In this case fall back to copying.
561 * LATER: this step can be merged to real generation of fragments,
562 * we can switch to copy when see the first bad fragment.
564 if (skb_has_frag_list(skb)) {
565 struct sk_buff *frag, *frag2;
566 int first_len = skb_pagelen(skb);
568 if (first_len - hlen > mtu ||
569 ((first_len - hlen) & 7) ||
570 ip_is_fragment(iph) ||
574 skb_walk_frags(skb, frag) {
575 /* Correct geometry. */
576 if (frag->len > mtu ||
577 ((frag->len & 7) && frag->next) ||
578 skb_headroom(frag) < hlen)
579 goto slow_path_clean;
581 /* Partially cloned skb? */
582 if (skb_shared(frag))
583 goto slow_path_clean;
588 frag->destructor = sock_wfree;
590 skb->truesize -= frag->truesize;
593 /* Everything is OK. Generate! */
597 frag = skb_shinfo(skb)->frag_list;
598 skb_frag_list_init(skb);
599 skb->data_len = first_len - skb_headlen(skb);
600 skb->len = first_len;
601 iph->tot_len = htons(first_len);
602 iph->frag_off = htons(IP_MF);
606 /* Prepare header of the next frame,
607 * before previous one went down. */
609 frag->ip_summed = CHECKSUM_NONE;
610 skb_reset_transport_header(frag);
611 __skb_push(frag, hlen);
612 skb_reset_network_header(frag);
613 memcpy(skb_network_header(frag), iph, hlen);
615 iph->tot_len = htons(frag->len);
616 ip_copy_metadata(frag, skb);
618 ip_options_fragment(frag);
619 offset += skb->len - hlen;
620 iph->frag_off = htons(offset>>3);
622 iph->frag_off |= htons(IP_MF);
623 /* Ready, complete checksum */
627 err = output(net, sk, skb);
630 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
640 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
649 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
653 skb_walk_frags(skb, frag2) {
657 frag2->destructor = NULL;
658 skb->truesize += frag2->truesize;
663 /* for offloaded checksums cleanup checksum before fragmentation */
664 if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
668 left = skb->len - hlen; /* Space per frame */
669 ptr = hlen; /* Where to start from */
671 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
674 * Fragment the datagram.
677 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
678 not_last_frag = iph->frag_off & htons(IP_MF);
681 * Keep copying data until we run out.
686 /* IF: it doesn't fit, use 'mtu' - the data space left */
689 /* IF: we are not sending up to and including the packet end
690 then align the next start on an eight byte boundary */
695 /* Allocate buffer */
696 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
703 * Set up data on packet
706 ip_copy_metadata(skb2, skb);
707 skb_reserve(skb2, ll_rs);
708 skb_put(skb2, len + hlen);
709 skb_reset_network_header(skb2);
710 skb2->transport_header = skb2->network_header + hlen;
713 * Charge the memory for the fragment to any owner
718 skb_set_owner_w(skb2, skb->sk);
721 * Copy the packet header into the new buffer.
724 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
727 * Copy a block of the IP datagram.
729 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
734 * Fill in the new header fields.
737 iph->frag_off = htons((offset >> 3));
739 if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
740 iph->frag_off |= htons(IP_DF);
742 /* ANK: dirty, but effective trick. Upgrade options only if
743 * the segment to be fragmented was THE FIRST (otherwise,
744 * options are already fixed) and make it ONCE
745 * on the initial skb, so that all the following fragments
746 * will inherit fixed options.
749 ip_options_fragment(skb);
752 * Added AC : If we are fragmenting a fragment that's not the
753 * last fragment then keep MF on each bit
755 if (left > 0 || not_last_frag)
756 iph->frag_off |= htons(IP_MF);
761 * Put this fragment into the sending queue.
763 iph->tot_len = htons(len + hlen);
767 err = output(net, sk, skb2);
771 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
774 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
779 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
782 EXPORT_SYMBOL(ip_do_fragment);
785 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
787 struct msghdr *msg = from;
789 if (skb->ip_summed == CHECKSUM_PARTIAL) {
790 if (copy_from_iter(to, len, &msg->msg_iter) != len)
794 if (csum_and_copy_from_iter(to, len, &csum, &msg->msg_iter) != len)
796 skb->csum = csum_block_add(skb->csum, csum, odd);
800 EXPORT_SYMBOL(ip_generic_getfrag);
803 csum_page(struct page *page, int offset, int copy)
808 csum = csum_partial(kaddr + offset, copy, 0);
813 static inline int ip_ufo_append_data(struct sock *sk,
814 struct sk_buff_head *queue,
815 int getfrag(void *from, char *to, int offset, int len,
816 int odd, struct sk_buff *skb),
817 void *from, int length, int hh_len, int fragheaderlen,
818 int transhdrlen, int maxfraglen, unsigned int flags)
823 /* There is support for UDP fragmentation offload by network
824 * device, so create one single skb packet containing complete
827 skb = skb_peek_tail(queue);
829 skb = sock_alloc_send_skb(sk,
830 hh_len + fragheaderlen + transhdrlen + 20,
831 (flags & MSG_DONTWAIT), &err);
836 /* reserve space for Hardware header */
837 skb_reserve(skb, hh_len);
839 /* create space for UDP/IP header */
840 skb_put(skb, fragheaderlen + transhdrlen);
842 /* initialize network header pointer */
843 skb_reset_network_header(skb);
845 /* initialize protocol header pointer */
846 skb->transport_header = skb->network_header + fragheaderlen;
850 __skb_queue_tail(queue, skb);
851 } else if (skb_is_gso(skb)) {
855 skb->ip_summed = CHECKSUM_PARTIAL;
856 /* specify the length of each IP datagram fragment */
857 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
858 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
861 return skb_append_datato_frags(sk, skb, getfrag, from,
862 (length - transhdrlen));
865 static int __ip_append_data(struct sock *sk,
867 struct sk_buff_head *queue,
868 struct inet_cork *cork,
869 struct page_frag *pfrag,
870 int getfrag(void *from, char *to, int offset,
871 int len, int odd, struct sk_buff *skb),
872 void *from, int length, int transhdrlen,
875 struct inet_sock *inet = inet_sk(sk);
878 struct ip_options *opt = cork->opt;
885 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
886 int csummode = CHECKSUM_NONE;
887 struct rtable *rt = (struct rtable *)cork->dst;
890 skb = skb_peek_tail(queue);
892 exthdrlen = !skb ? rt->dst.header_len : 0;
893 mtu = cork->fragsize;
894 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
895 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
896 tskey = sk->sk_tskey++;
898 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
900 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
901 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
902 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
904 if (cork->length + length > maxnonfragsize - fragheaderlen) {
905 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
906 mtu - (opt ? opt->optlen : 0));
911 * transhdrlen > 0 means that this is the first fragment and we wish
912 * it won't be fragmented in the future.
915 length + fragheaderlen <= mtu &&
916 rt->dst.dev->features & NETIF_F_V4_CSUM &&
918 csummode = CHECKSUM_PARTIAL;
920 cork->length += length;
921 if (((length > mtu) || (skb && skb_is_gso(skb))) &&
922 (sk->sk_protocol == IPPROTO_UDP) &&
923 (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len &&
924 (sk->sk_type == SOCK_DGRAM)) {
925 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
926 hh_len, fragheaderlen, transhdrlen,
933 /* So, what's going on in the loop below?
935 * We use calculated fragment length to generate chained skb,
936 * each of segments is IP fragment ready for sending to network after
937 * adding appropriate IP header.
944 /* Check if the remaining data fits into current packet. */
945 copy = mtu - skb->len;
947 copy = maxfraglen - skb->len;
950 unsigned int datalen;
951 unsigned int fraglen;
952 unsigned int fraggap;
953 unsigned int alloclen;
954 struct sk_buff *skb_prev;
958 fraggap = skb_prev->len - maxfraglen;
963 * If remaining data exceeds the mtu,
964 * we know we need more fragment(s).
966 datalen = length + fraggap;
967 if (datalen > mtu - fragheaderlen)
968 datalen = maxfraglen - fragheaderlen;
969 fraglen = datalen + fragheaderlen;
971 if ((flags & MSG_MORE) &&
972 !(rt->dst.dev->features&NETIF_F_SG))
977 alloclen += exthdrlen;
979 /* The last fragment gets additional space at tail.
980 * Note, with MSG_MORE we overallocate on fragments,
981 * because we have no idea what fragment will be
984 if (datalen == length + fraggap)
985 alloclen += rt->dst.trailer_len;
988 skb = sock_alloc_send_skb(sk,
989 alloclen + hh_len + 15,
990 (flags & MSG_DONTWAIT), &err);
993 if (atomic_read(&sk->sk_wmem_alloc) <=
995 skb = sock_wmalloc(sk,
996 alloclen + hh_len + 15, 1,
1005 * Fill in the control structures
1007 skb->ip_summed = csummode;
1009 skb_reserve(skb, hh_len);
1011 /* only the initial fragment is time stamped */
1012 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1014 skb_shinfo(skb)->tskey = tskey;
1018 * Find where to start putting bytes.
1020 data = skb_put(skb, fraglen + exthdrlen);
1021 skb_set_network_header(skb, exthdrlen);
1022 skb->transport_header = (skb->network_header +
1024 data += fragheaderlen + exthdrlen;
1027 skb->csum = skb_copy_and_csum_bits(
1028 skb_prev, maxfraglen,
1029 data + transhdrlen, fraggap, 0);
1030 skb_prev->csum = csum_sub(skb_prev->csum,
1033 pskb_trim_unique(skb_prev, maxfraglen);
1036 copy = datalen - transhdrlen - fraggap;
1037 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1044 length -= datalen - fraggap;
1047 csummode = CHECKSUM_NONE;
1050 * Put the packet on the pending queue.
1052 __skb_queue_tail(queue, skb);
1059 if (!(rt->dst.dev->features&NETIF_F_SG)) {
1063 if (getfrag(from, skb_put(skb, copy),
1064 offset, copy, off, skb) < 0) {
1065 __skb_trim(skb, off);
1070 int i = skb_shinfo(skb)->nr_frags;
1073 if (!sk_page_frag_refill(sk, pfrag))
1076 if (!skb_can_coalesce(skb, i, pfrag->page,
1079 if (i == MAX_SKB_FRAGS)
1082 __skb_fill_page_desc(skb, i, pfrag->page,
1084 skb_shinfo(skb)->nr_frags = ++i;
1085 get_page(pfrag->page);
1087 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1089 page_address(pfrag->page) + pfrag->offset,
1090 offset, copy, skb->len, skb) < 0)
1093 pfrag->offset += copy;
1094 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1096 skb->data_len += copy;
1097 skb->truesize += copy;
1098 atomic_add(copy, &sk->sk_wmem_alloc);
1109 cork->length -= length;
1110 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1114 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1115 struct ipcm_cookie *ipc, struct rtable **rtp)
1117 struct ip_options_rcu *opt;
1121 * setup for corking.
1126 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1128 if (unlikely(!cork->opt))
1131 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1132 cork->flags |= IPCORK_OPT;
1133 cork->addr = ipc->addr;
1139 * We steal reference to this route, caller should not release it
1142 cork->fragsize = ip_sk_use_pmtu(sk) ?
1143 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1144 cork->dst = &rt->dst;
1146 cork->ttl = ipc->ttl;
1147 cork->tos = ipc->tos;
1148 cork->priority = ipc->priority;
1149 cork->tx_flags = ipc->tx_flags;
1155 * ip_append_data() and ip_append_page() can make one large IP datagram
1156 * from many pieces of data. Each pieces will be holded on the socket
1157 * until ip_push_pending_frames() is called. Each piece can be a page
1160 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1161 * this interface potentially.
1163 * LATER: length must be adjusted by pad at tail, when it is required.
1165 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1166 int getfrag(void *from, char *to, int offset, int len,
1167 int odd, struct sk_buff *skb),
1168 void *from, int length, int transhdrlen,
1169 struct ipcm_cookie *ipc, struct rtable **rtp,
1172 struct inet_sock *inet = inet_sk(sk);
1175 if (flags&MSG_PROBE)
1178 if (skb_queue_empty(&sk->sk_write_queue)) {
1179 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1186 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1187 sk_page_frag(sk), getfrag,
1188 from, length, transhdrlen, flags);
1191 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1192 int offset, size_t size, int flags)
1194 struct inet_sock *inet = inet_sk(sk);
1195 struct sk_buff *skb;
1197 struct ip_options *opt = NULL;
1198 struct inet_cork *cork;
1203 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1208 if (flags&MSG_PROBE)
1211 if (skb_queue_empty(&sk->sk_write_queue))
1214 cork = &inet->cork.base;
1215 rt = (struct rtable *)cork->dst;
1216 if (cork->flags & IPCORK_OPT)
1219 if (!(rt->dst.dev->features&NETIF_F_SG))
1222 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1223 mtu = cork->fragsize;
1225 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1226 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1227 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1229 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1230 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1231 mtu - (opt ? opt->optlen : 0));
1235 skb = skb_peek_tail(&sk->sk_write_queue);
1239 cork->length += size;
1240 if ((size + skb->len > mtu) &&
1241 (sk->sk_protocol == IPPROTO_UDP) &&
1242 (rt->dst.dev->features & NETIF_F_UFO)) {
1243 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1244 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1248 if (skb_is_gso(skb)) {
1252 /* Check if the remaining data fits into current packet. */
1253 len = mtu - skb->len;
1255 len = maxfraglen - skb->len;
1258 struct sk_buff *skb_prev;
1262 fraggap = skb_prev->len - maxfraglen;
1264 alloclen = fragheaderlen + hh_len + fraggap + 15;
1265 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1266 if (unlikely(!skb)) {
1272 * Fill in the control structures
1274 skb->ip_summed = CHECKSUM_NONE;
1276 skb_reserve(skb, hh_len);
1279 * Find where to start putting bytes.
1281 skb_put(skb, fragheaderlen + fraggap);
1282 skb_reset_network_header(skb);
1283 skb->transport_header = (skb->network_header +
1286 skb->csum = skb_copy_and_csum_bits(skb_prev,
1288 skb_transport_header(skb),
1290 skb_prev->csum = csum_sub(skb_prev->csum,
1292 pskb_trim_unique(skb_prev, maxfraglen);
1296 * Put the packet on the pending queue.
1298 __skb_queue_tail(&sk->sk_write_queue, skb);
1305 if (skb_append_pagefrags(skb, page, offset, len)) {
1310 if (skb->ip_summed == CHECKSUM_NONE) {
1312 csum = csum_page(page, offset, len);
1313 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1317 skb->data_len += len;
1318 skb->truesize += len;
1319 atomic_add(len, &sk->sk_wmem_alloc);
1326 cork->length -= size;
1327 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1331 static void ip_cork_release(struct inet_cork *cork)
1333 cork->flags &= ~IPCORK_OPT;
1336 dst_release(cork->dst);
1341 * Combined all pending IP fragments on the socket as one IP datagram
1342 * and push them out.
1344 struct sk_buff *__ip_make_skb(struct sock *sk,
1346 struct sk_buff_head *queue,
1347 struct inet_cork *cork)
1349 struct sk_buff *skb, *tmp_skb;
1350 struct sk_buff **tail_skb;
1351 struct inet_sock *inet = inet_sk(sk);
1352 struct net *net = sock_net(sk);
1353 struct ip_options *opt = NULL;
1354 struct rtable *rt = (struct rtable *)cork->dst;
1359 skb = __skb_dequeue(queue);
1362 tail_skb = &(skb_shinfo(skb)->frag_list);
1364 /* move skb->data to ip header from ext header */
1365 if (skb->data < skb_network_header(skb))
1366 __skb_pull(skb, skb_network_offset(skb));
1367 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1368 __skb_pull(tmp_skb, skb_network_header_len(skb));
1369 *tail_skb = tmp_skb;
1370 tail_skb = &(tmp_skb->next);
1371 skb->len += tmp_skb->len;
1372 skb->data_len += tmp_skb->len;
1373 skb->truesize += tmp_skb->truesize;
1374 tmp_skb->destructor = NULL;
1378 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1379 * to fragment the frame generated here. No matter, what transforms
1380 * how transforms change size of the packet, it will come out.
1382 skb->ignore_df = ip_sk_ignore_df(sk);
1384 /* DF bit is set when we want to see DF on outgoing frames.
1385 * If ignore_df is set too, we still allow to fragment this frame
1387 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1388 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1389 (skb->len <= dst_mtu(&rt->dst) &&
1390 ip_dont_fragment(sk, &rt->dst)))
1393 if (cork->flags & IPCORK_OPT)
1398 else if (rt->rt_type == RTN_MULTICAST)
1401 ttl = ip_select_ttl(inet, &rt->dst);
1406 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1409 iph->protocol = sk->sk_protocol;
1410 ip_copy_addrs(iph, fl4);
1411 ip_select_ident(net, skb, sk);
1414 iph->ihl += opt->optlen>>2;
1415 ip_options_build(skb, opt, cork->addr, rt, 0);
1418 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1419 skb->mark = sk->sk_mark;
1421 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1425 skb_dst_set(skb, &rt->dst);
1427 if (iph->protocol == IPPROTO_ICMP)
1428 icmp_out_count(net, ((struct icmphdr *)
1429 skb_transport_header(skb))->type);
1431 ip_cork_release(cork);
1436 int ip_send_skb(struct net *net, struct sk_buff *skb)
1440 err = ip_local_out(skb->sk, skb);
1443 err = net_xmit_errno(err);
1445 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1451 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1453 struct sk_buff *skb;
1455 skb = ip_finish_skb(sk, fl4);
1459 /* Netfilter gets whole the not fragmented skb. */
1460 return ip_send_skb(sock_net(sk), skb);
1464 * Throw away all pending data on the socket.
1466 static void __ip_flush_pending_frames(struct sock *sk,
1467 struct sk_buff_head *queue,
1468 struct inet_cork *cork)
1470 struct sk_buff *skb;
1472 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1475 ip_cork_release(cork);
1478 void ip_flush_pending_frames(struct sock *sk)
1480 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1483 struct sk_buff *ip_make_skb(struct sock *sk,
1485 int getfrag(void *from, char *to, int offset,
1486 int len, int odd, struct sk_buff *skb),
1487 void *from, int length, int transhdrlen,
1488 struct ipcm_cookie *ipc, struct rtable **rtp,
1491 struct inet_cork cork;
1492 struct sk_buff_head queue;
1495 if (flags & MSG_PROBE)
1498 __skb_queue_head_init(&queue);
1503 err = ip_setup_cork(sk, &cork, ipc, rtp);
1505 return ERR_PTR(err);
1507 err = __ip_append_data(sk, fl4, &queue, &cork,
1508 ¤t->task_frag, getfrag,
1509 from, length, transhdrlen, flags);
1511 __ip_flush_pending_frames(sk, &queue, &cork);
1512 return ERR_PTR(err);
1515 return __ip_make_skb(sk, fl4, &queue, &cork);
1519 * Fetch data from kernel space and fill in checksum if needed.
1521 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1522 int len, int odd, struct sk_buff *skb)
1526 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1527 skb->csum = csum_block_add(skb->csum, csum, odd);
1532 * Generic function to send a packet as reply to another packet.
1533 * Used to send some TCP resets/acks so far.
1535 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1536 const struct ip_options *sopt,
1537 __be32 daddr, __be32 saddr,
1538 const struct ip_reply_arg *arg,
1541 struct ip_options_data replyopts;
1542 struct ipcm_cookie ipc;
1544 struct rtable *rt = skb_rtable(skb);
1545 struct net *net = sock_net(sk);
1546 struct sk_buff *nskb;
1550 if (__ip_options_echo(&replyopts.opt.opt, skb, sopt))
1559 if (replyopts.opt.opt.optlen) {
1560 ipc.opt = &replyopts.opt;
1562 if (replyopts.opt.opt.srr)
1563 daddr = replyopts.opt.opt.faddr;
1566 oif = arg->bound_dev_if;
1567 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1570 flowi4_init_output(&fl4, oif,
1571 IP4_REPLY_MARK(net, skb->mark),
1573 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1574 ip_reply_arg_flowi_flags(arg),
1576 tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1577 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1578 rt = ip_route_output_key(net, &fl4);
1582 inet_sk(sk)->tos = arg->tos;
1584 sk->sk_priority = skb->priority;
1585 sk->sk_protocol = ip_hdr(skb)->protocol;
1586 sk->sk_bound_dev_if = arg->bound_dev_if;
1587 sk->sk_sndbuf = sysctl_wmem_default;
1588 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1589 len, 0, &ipc, &rt, MSG_DONTWAIT);
1590 if (unlikely(err)) {
1591 ip_flush_pending_frames(sk);
1595 nskb = skb_peek(&sk->sk_write_queue);
1597 if (arg->csumoffset >= 0)
1598 *((__sum16 *)skb_transport_header(nskb) +
1599 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1601 nskb->ip_summed = CHECKSUM_NONE;
1602 skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1603 ip_push_pending_frames(sk, &fl4);
1609 void __init ip_init(void)
1614 #if defined(CONFIG_IP_MULTICAST)