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);
86 static int ip_fragment(struct sock *sk, struct sk_buff *skb,
88 int (*output)(struct sock *, struct sk_buff *));
90 /* Generate a checksum for an outgoing IP datagram. */
91 void ip_send_check(struct iphdr *iph)
94 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
96 EXPORT_SYMBOL(ip_send_check);
98 static int __ip_local_out_sk(struct sock *sk, struct sk_buff *skb)
100 struct iphdr *iph = ip_hdr(skb);
102 iph->tot_len = htons(skb->len);
104 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT, sk, skb, NULL,
105 skb_dst(skb)->dev, dst_output_sk);
108 int __ip_local_out(struct sk_buff *skb)
110 return __ip_local_out_sk(skb->sk, skb);
113 int ip_local_out_sk(struct sock *sk, struct sk_buff *skb)
117 err = __ip_local_out(skb);
118 if (likely(err == 1))
119 err = dst_output_sk(sk, skb);
123 EXPORT_SYMBOL_GPL(ip_local_out_sk);
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, 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);
145 /* Build the IP header. */
146 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
147 skb_reset_network_header(skb);
151 iph->tos = inet->tos;
152 if (ip_dont_fragment(sk, &rt->dst))
153 iph->frag_off = htons(IP_DF);
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 ip_select_ident(sock_net(sk), skb, sk);
162 if (opt && opt->opt.optlen) {
163 iph->ihl += opt->opt.optlen>>2;
164 ip_options_build(skb, &opt->opt, daddr, rt, 0);
167 skb->priority = sk->sk_priority;
168 skb->mark = sk->sk_mark;
171 return ip_local_out(skb);
173 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
175 static inline int ip_finish_output2(struct sock *sk, struct sk_buff *skb)
177 struct dst_entry *dst = skb_dst(skb);
178 struct rtable *rt = (struct rtable *)dst;
179 struct net_device *dev = dst->dev;
180 unsigned int hh_len = LL_RESERVED_SPACE(dev);
181 struct neighbour *neigh;
184 if (rt->rt_type == RTN_MULTICAST) {
185 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
186 } else if (rt->rt_type == RTN_BROADCAST)
187 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
189 /* Be paranoid, rather than too clever. */
190 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
191 struct sk_buff *skb2;
193 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
199 skb_set_owner_w(skb2, skb->sk);
205 nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
206 neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
207 if (unlikely(!neigh))
208 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
209 if (!IS_ERR(neigh)) {
210 int res = dst_neigh_output(dst, neigh, skb);
212 rcu_read_unlock_bh();
215 rcu_read_unlock_bh();
217 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
223 static int ip_finish_output_gso(struct sock *sk, struct sk_buff *skb,
226 netdev_features_t features;
227 struct sk_buff *segs;
230 /* common case: locally created skb or seglen is <= mtu */
231 if (((IPCB(skb)->flags & IPSKB_FORWARDED) == 0) ||
232 skb_gso_network_seglen(skb) <= mtu)
233 return ip_finish_output2(sk, skb);
235 /* Slowpath - GSO segment length is exceeding the dst MTU.
237 * This can happen in two cases:
238 * 1) TCP GRO packet, DF bit not set
239 * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
240 * from host network stack.
242 features = netif_skb_features(skb);
243 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
244 if (IS_ERR_OR_NULL(segs)) {
252 struct sk_buff *nskb = segs->next;
256 err = ip_fragment(sk, segs, mtu, ip_finish_output2);
266 static int ip_finish_output(struct sock *sk, struct sk_buff *skb)
270 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
271 /* Policy lookup after SNAT yielded a new policy */
272 if (skb_dst(skb)->xfrm) {
273 IPCB(skb)->flags |= IPSKB_REROUTED;
274 return dst_output_sk(sk, skb);
277 mtu = ip_skb_dst_mtu(skb);
279 return ip_finish_output_gso(sk, skb, mtu);
281 if (skb->len > mtu || (IPCB(skb)->flags & IPSKB_FRAG_PMTU))
282 return ip_fragment(sk, skb, mtu, ip_finish_output2);
284 return ip_finish_output2(sk, skb);
287 int ip_mc_output(struct sock *sk, struct sk_buff *skb)
289 struct rtable *rt = skb_rtable(skb);
290 struct net_device *dev = rt->dst.dev;
293 * If the indicated interface is up and running, send the packet.
295 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
298 skb->protocol = htons(ETH_P_IP);
301 * Multicasts are looped back for other local users
304 if (rt->rt_flags&RTCF_MULTICAST) {
306 #ifdef CONFIG_IP_MROUTE
307 /* Small optimization: do not loopback not local frames,
308 which returned after forwarding; they will be dropped
309 by ip_mr_input in any case.
310 Note, that local frames are looped back to be delivered
313 This check is duplicated in ip_mr_input at the moment.
316 ((rt->rt_flags & RTCF_LOCAL) ||
317 !(IPCB(skb)->flags & IPSKB_FORWARDED))
320 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
322 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
323 sk, newskb, NULL, newskb->dev,
327 /* Multicasts with ttl 0 must not go beyond the host */
329 if (ip_hdr(skb)->ttl == 0) {
335 if (rt->rt_flags&RTCF_BROADCAST) {
336 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
338 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, sk, newskb,
339 NULL, newskb->dev, dev_loopback_xmit);
342 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, sk, skb, NULL,
343 skb->dev, ip_finish_output,
344 !(IPCB(skb)->flags & IPSKB_REROUTED));
347 int ip_output(struct sock *sk, struct sk_buff *skb)
349 struct net_device *dev = skb_dst(skb)->dev;
351 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
354 skb->protocol = htons(ETH_P_IP);
356 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, sk, skb,
359 !(IPCB(skb)->flags & IPSKB_REROUTED));
363 * copy saddr and daddr, possibly using 64bit load/stores
365 * iph->saddr = fl4->saddr;
366 * iph->daddr = fl4->daddr;
368 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
370 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
371 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
372 memcpy(&iph->saddr, &fl4->saddr,
373 sizeof(fl4->saddr) + sizeof(fl4->daddr));
376 /* Note: skb->sk can be different from sk, in case of tunnels */
377 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
379 struct inet_sock *inet = inet_sk(sk);
380 struct ip_options_rcu *inet_opt;
386 /* Skip all of this if the packet is already routed,
387 * f.e. by something like SCTP.
390 inet_opt = rcu_dereference(inet->inet_opt);
392 rt = skb_rtable(skb);
396 /* Make sure we can route this packet. */
397 rt = (struct rtable *)__sk_dst_check(sk, 0);
401 /* Use correct destination address if we have options. */
402 daddr = inet->inet_daddr;
403 if (inet_opt && inet_opt->opt.srr)
404 daddr = inet_opt->opt.faddr;
406 /* If this fails, retransmit mechanism of transport layer will
407 * keep trying until route appears or the connection times
410 rt = ip_route_output_ports(sock_net(sk), fl4, sk,
411 daddr, inet->inet_saddr,
416 sk->sk_bound_dev_if);
419 sk_setup_caps(sk, &rt->dst);
421 skb_dst_set_noref(skb, &rt->dst);
424 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
427 /* OK, we know where to send it, allocate and build IP header. */
428 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
429 skb_reset_network_header(skb);
431 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
432 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
433 iph->frag_off = htons(IP_DF);
436 iph->ttl = ip_select_ttl(inet, &rt->dst);
437 iph->protocol = sk->sk_protocol;
438 ip_copy_addrs(iph, fl4);
440 /* Transport layer set skb->h.foo itself. */
442 if (inet_opt && inet_opt->opt.optlen) {
443 iph->ihl += inet_opt->opt.optlen >> 2;
444 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
447 ip_select_ident_segs(sock_net(sk), skb, sk,
448 skb_shinfo(skb)->gso_segs ?: 1);
450 /* TODO : should we use skb->sk here instead of sk ? */
451 skb->priority = sk->sk_priority;
452 skb->mark = sk->sk_mark;
454 res = ip_local_out(skb);
460 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
462 return -EHOSTUNREACH;
464 EXPORT_SYMBOL(ip_queue_xmit);
466 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
468 to->pkt_type = from->pkt_type;
469 to->priority = from->priority;
470 to->protocol = from->protocol;
472 skb_dst_copy(to, from);
474 to->mark = from->mark;
476 /* Copy the flags to each fragment. */
477 IPCB(to)->flags = IPCB(from)->flags;
479 #ifdef CONFIG_NET_SCHED
480 to->tc_index = from->tc_index;
483 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
484 to->ipvs_property = from->ipvs_property;
486 skb_copy_secmark(to, from);
489 static int ip_fragment(struct sock *sk, struct sk_buff *skb,
491 int (*output)(struct sock *, struct sk_buff *))
493 struct iphdr *iph = ip_hdr(skb);
495 if ((iph->frag_off & htons(IP_DF)) == 0)
496 return ip_do_fragment(sk, skb, output);
498 if (unlikely(!skb->ignore_df ||
499 (IPCB(skb)->frag_max_size &&
500 IPCB(skb)->frag_max_size > mtu))) {
501 struct rtable *rt = skb_rtable(skb);
502 struct net_device *dev = rt->dst.dev;
504 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
505 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
511 return ip_do_fragment(sk, skb, output);
515 * This IP datagram is too large to be sent in one piece. Break it up into
516 * smaller pieces (each of size equal to IP header plus
517 * a block of the data of the original IP data part) that will yet fit in a
518 * single device frame, and queue such a frame for sending.
521 int ip_do_fragment(struct sock *sk, struct sk_buff *skb,
522 int (*output)(struct sock *, struct sk_buff *))
526 struct net_device *dev;
527 struct sk_buff *skb2;
528 unsigned int mtu, hlen, left, len, ll_rs;
530 __be16 not_last_frag;
531 struct rtable *rt = skb_rtable(skb);
537 * Point into the IP datagram header.
542 mtu = ip_skb_dst_mtu(skb);
543 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
544 mtu = IPCB(skb)->frag_max_size;
547 * Setup starting values.
551 mtu = mtu - hlen; /* Size of data space */
552 #if IS_ENABLED(CONFIG_BRIDGE_NETFILTER)
554 mtu -= nf_bridge_mtu_reduction(skb);
556 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
558 /* When frag_list is given, use it. First, check its validity:
559 * some transformers could create wrong frag_list or break existing
560 * one, it is not prohibited. In this case fall back to copying.
562 * LATER: this step can be merged to real generation of fragments,
563 * we can switch to copy when see the first bad fragment.
565 if (skb_has_frag_list(skb)) {
566 struct sk_buff *frag, *frag2;
567 int first_len = skb_pagelen(skb);
569 if (first_len - hlen > mtu ||
570 ((first_len - hlen) & 7) ||
571 ip_is_fragment(iph) ||
575 skb_walk_frags(skb, frag) {
576 /* Correct geometry. */
577 if (frag->len > mtu ||
578 ((frag->len & 7) && frag->next) ||
579 skb_headroom(frag) < hlen)
580 goto slow_path_clean;
582 /* Partially cloned skb? */
583 if (skb_shared(frag))
584 goto slow_path_clean;
589 frag->destructor = sock_wfree;
591 skb->truesize -= frag->truesize;
594 /* Everything is OK. Generate! */
598 frag = skb_shinfo(skb)->frag_list;
599 skb_frag_list_init(skb);
600 skb->data_len = first_len - skb_headlen(skb);
601 skb->len = first_len;
602 iph->tot_len = htons(first_len);
603 iph->frag_off = htons(IP_MF);
607 /* Prepare header of the next frame,
608 * before previous one went down. */
610 frag->ip_summed = CHECKSUM_NONE;
611 skb_reset_transport_header(frag);
612 __skb_push(frag, hlen);
613 skb_reset_network_header(frag);
614 memcpy(skb_network_header(frag), iph, hlen);
616 iph->tot_len = htons(frag->len);
617 ip_copy_metadata(frag, skb);
619 ip_options_fragment(frag);
620 offset += skb->len - hlen;
621 iph->frag_off = htons(offset>>3);
623 iph->frag_off |= htons(IP_MF);
624 /* Ready, complete checksum */
628 err = output(sk, skb);
631 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
641 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
650 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
654 skb_walk_frags(skb, frag2) {
658 frag2->destructor = NULL;
659 skb->truesize += frag2->truesize;
664 /* for offloaded checksums cleanup checksum before fragmentation */
665 if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
669 left = skb->len - hlen; /* Space per frame */
670 ptr = hlen; /* Where to start from */
672 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
675 * Fragment the datagram.
678 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
679 not_last_frag = iph->frag_off & htons(IP_MF);
682 * Keep copying data until we run out.
687 /* IF: it doesn't fit, use 'mtu' - the data space left */
690 /* IF: we are not sending up to and including the packet end
691 then align the next start on an eight byte boundary */
696 /* Allocate buffer */
697 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
704 * Set up data on packet
707 ip_copy_metadata(skb2, skb);
708 skb_reserve(skb2, ll_rs);
709 skb_put(skb2, len + hlen);
710 skb_reset_network_header(skb2);
711 skb2->transport_header = skb2->network_header + hlen;
714 * Charge the memory for the fragment to any owner
719 skb_set_owner_w(skb2, skb->sk);
722 * Copy the packet header into the new buffer.
725 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
728 * Copy a block of the IP datagram.
730 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
735 * Fill in the new header fields.
738 iph->frag_off = htons((offset >> 3));
740 if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
741 iph->frag_off |= htons(IP_DF);
743 /* ANK: dirty, but effective trick. Upgrade options only if
744 * the segment to be fragmented was THE FIRST (otherwise,
745 * options are already fixed) and make it ONCE
746 * on the initial skb, so that all the following fragments
747 * will inherit fixed options.
750 ip_options_fragment(skb);
753 * Added AC : If we are fragmenting a fragment that's not the
754 * last fragment then keep MF on each bit
756 if (left > 0 || not_last_frag)
757 iph->frag_off |= htons(IP_MF);
762 * Put this fragment into the sending queue.
764 iph->tot_len = htons(len + hlen);
768 err = output(sk, skb2);
772 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
775 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
780 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
783 EXPORT_SYMBOL(ip_do_fragment);
786 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
788 struct msghdr *msg = from;
790 if (skb->ip_summed == CHECKSUM_PARTIAL) {
791 if (copy_from_iter(to, len, &msg->msg_iter) != len)
795 if (csum_and_copy_from_iter(to, len, &csum, &msg->msg_iter) != len)
797 skb->csum = csum_block_add(skb->csum, csum, odd);
801 EXPORT_SYMBOL(ip_generic_getfrag);
804 csum_page(struct page *page, int offset, int copy)
809 csum = csum_partial(kaddr + offset, copy, 0);
814 static inline int ip_ufo_append_data(struct sock *sk,
815 struct sk_buff_head *queue,
816 int getfrag(void *from, char *to, int offset, int len,
817 int odd, struct sk_buff *skb),
818 void *from, int length, int hh_len, int fragheaderlen,
819 int transhdrlen, int maxfraglen, unsigned int flags)
824 /* There is support for UDP fragmentation offload by network
825 * device, so create one single skb packet containing complete
828 skb = skb_peek_tail(queue);
830 skb = sock_alloc_send_skb(sk,
831 hh_len + fragheaderlen + transhdrlen + 20,
832 (flags & MSG_DONTWAIT), &err);
837 /* reserve space for Hardware header */
838 skb_reserve(skb, hh_len);
840 /* create space for UDP/IP header */
841 skb_put(skb, fragheaderlen + transhdrlen);
843 /* initialize network header pointer */
844 skb_reset_network_header(skb);
846 /* initialize protocol header pointer */
847 skb->transport_header = skb->network_header + fragheaderlen;
851 __skb_queue_tail(queue, skb);
852 } else if (skb_is_gso(skb)) {
856 skb->ip_summed = CHECKSUM_PARTIAL;
857 /* specify the length of each IP datagram fragment */
858 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
859 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
862 return skb_append_datato_frags(sk, skb, getfrag, from,
863 (length - transhdrlen));
866 static int __ip_append_data(struct sock *sk,
868 struct sk_buff_head *queue,
869 struct inet_cork *cork,
870 struct page_frag *pfrag,
871 int getfrag(void *from, char *to, int offset,
872 int len, int odd, struct sk_buff *skb),
873 void *from, int length, int transhdrlen,
876 struct inet_sock *inet = inet_sk(sk);
879 struct ip_options *opt = cork->opt;
886 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
887 int csummode = CHECKSUM_NONE;
888 struct rtable *rt = (struct rtable *)cork->dst;
891 skb = skb_peek_tail(queue);
893 exthdrlen = !skb ? rt->dst.header_len : 0;
894 mtu = cork->fragsize;
895 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
896 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
897 tskey = sk->sk_tskey++;
899 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
901 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
902 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
903 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
905 if (cork->length + length > maxnonfragsize - fragheaderlen) {
906 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
907 mtu - (opt ? opt->optlen : 0));
912 * transhdrlen > 0 means that this is the first fragment and we wish
913 * it won't be fragmented in the future.
916 length + fragheaderlen <= mtu &&
917 rt->dst.dev->features & NETIF_F_V4_CSUM &&
919 csummode = CHECKSUM_PARTIAL;
921 cork->length += length;
922 if (((length > mtu) || (skb && skb_is_gso(skb))) &&
923 (sk->sk_protocol == IPPROTO_UDP) &&
924 (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len &&
925 (sk->sk_type == SOCK_DGRAM)) {
926 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
927 hh_len, fragheaderlen, transhdrlen,
934 /* So, what's going on in the loop below?
936 * We use calculated fragment length to generate chained skb,
937 * each of segments is IP fragment ready for sending to network after
938 * adding appropriate IP header.
945 /* Check if the remaining data fits into current packet. */
946 copy = mtu - skb->len;
948 copy = maxfraglen - skb->len;
951 unsigned int datalen;
952 unsigned int fraglen;
953 unsigned int fraggap;
954 unsigned int alloclen;
955 struct sk_buff *skb_prev;
959 fraggap = skb_prev->len - maxfraglen;
964 * If remaining data exceeds the mtu,
965 * we know we need more fragment(s).
967 datalen = length + fraggap;
968 if (datalen > mtu - fragheaderlen)
969 datalen = maxfraglen - fragheaderlen;
970 fraglen = datalen + fragheaderlen;
972 if ((flags & MSG_MORE) &&
973 !(rt->dst.dev->features&NETIF_F_SG))
978 alloclen += exthdrlen;
980 /* The last fragment gets additional space at tail.
981 * Note, with MSG_MORE we overallocate on fragments,
982 * because we have no idea what fragment will be
985 if (datalen == length + fraggap)
986 alloclen += rt->dst.trailer_len;
989 skb = sock_alloc_send_skb(sk,
990 alloclen + hh_len + 15,
991 (flags & MSG_DONTWAIT), &err);
994 if (atomic_read(&sk->sk_wmem_alloc) <=
996 skb = sock_wmalloc(sk,
997 alloclen + hh_len + 15, 1,
1006 * Fill in the control structures
1008 skb->ip_summed = csummode;
1010 skb_reserve(skb, hh_len);
1012 /* only the initial fragment is time stamped */
1013 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1015 skb_shinfo(skb)->tskey = tskey;
1019 * Find where to start putting bytes.
1021 data = skb_put(skb, fraglen + exthdrlen);
1022 skb_set_network_header(skb, exthdrlen);
1023 skb->transport_header = (skb->network_header +
1025 data += fragheaderlen + exthdrlen;
1028 skb->csum = skb_copy_and_csum_bits(
1029 skb_prev, maxfraglen,
1030 data + transhdrlen, fraggap, 0);
1031 skb_prev->csum = csum_sub(skb_prev->csum,
1034 pskb_trim_unique(skb_prev, maxfraglen);
1037 copy = datalen - transhdrlen - fraggap;
1038 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1045 length -= datalen - fraggap;
1048 csummode = CHECKSUM_NONE;
1051 * Put the packet on the pending queue.
1053 __skb_queue_tail(queue, skb);
1060 if (!(rt->dst.dev->features&NETIF_F_SG)) {
1064 if (getfrag(from, skb_put(skb, copy),
1065 offset, copy, off, skb) < 0) {
1066 __skb_trim(skb, off);
1071 int i = skb_shinfo(skb)->nr_frags;
1074 if (!sk_page_frag_refill(sk, pfrag))
1077 if (!skb_can_coalesce(skb, i, pfrag->page,
1080 if (i == MAX_SKB_FRAGS)
1083 __skb_fill_page_desc(skb, i, pfrag->page,
1085 skb_shinfo(skb)->nr_frags = ++i;
1086 get_page(pfrag->page);
1088 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1090 page_address(pfrag->page) + pfrag->offset,
1091 offset, copy, skb->len, skb) < 0)
1094 pfrag->offset += copy;
1095 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1097 skb->data_len += copy;
1098 skb->truesize += copy;
1099 atomic_add(copy, &sk->sk_wmem_alloc);
1110 cork->length -= length;
1111 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1115 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1116 struct ipcm_cookie *ipc, struct rtable **rtp)
1118 struct ip_options_rcu *opt;
1122 * setup for corking.
1127 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1129 if (unlikely(!cork->opt))
1132 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1133 cork->flags |= IPCORK_OPT;
1134 cork->addr = ipc->addr;
1140 * We steal reference to this route, caller should not release it
1143 cork->fragsize = ip_sk_use_pmtu(sk) ?
1144 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1145 cork->dst = &rt->dst;
1147 cork->ttl = ipc->ttl;
1148 cork->tos = ipc->tos;
1149 cork->priority = ipc->priority;
1150 cork->tx_flags = ipc->tx_flags;
1156 * ip_append_data() and ip_append_page() can make one large IP datagram
1157 * from many pieces of data. Each pieces will be holded on the socket
1158 * until ip_push_pending_frames() is called. Each piece can be a page
1161 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1162 * this interface potentially.
1164 * LATER: length must be adjusted by pad at tail, when it is required.
1166 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1167 int getfrag(void *from, char *to, int offset, int len,
1168 int odd, struct sk_buff *skb),
1169 void *from, int length, int transhdrlen,
1170 struct ipcm_cookie *ipc, struct rtable **rtp,
1173 struct inet_sock *inet = inet_sk(sk);
1176 if (flags&MSG_PROBE)
1179 if (skb_queue_empty(&sk->sk_write_queue)) {
1180 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1187 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1188 sk_page_frag(sk), getfrag,
1189 from, length, transhdrlen, flags);
1192 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1193 int offset, size_t size, int flags)
1195 struct inet_sock *inet = inet_sk(sk);
1196 struct sk_buff *skb;
1198 struct ip_options *opt = NULL;
1199 struct inet_cork *cork;
1204 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1209 if (flags&MSG_PROBE)
1212 if (skb_queue_empty(&sk->sk_write_queue))
1215 cork = &inet->cork.base;
1216 rt = (struct rtable *)cork->dst;
1217 if (cork->flags & IPCORK_OPT)
1220 if (!(rt->dst.dev->features&NETIF_F_SG))
1223 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1224 mtu = cork->fragsize;
1226 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1227 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1228 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1230 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1231 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1232 mtu - (opt ? opt->optlen : 0));
1236 skb = skb_peek_tail(&sk->sk_write_queue);
1240 cork->length += size;
1241 if ((size + skb->len > mtu) &&
1242 (sk->sk_protocol == IPPROTO_UDP) &&
1243 (rt->dst.dev->features & NETIF_F_UFO)) {
1244 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1245 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1249 if (skb_is_gso(skb)) {
1253 /* Check if the remaining data fits into current packet. */
1254 len = mtu - skb->len;
1256 len = maxfraglen - skb->len;
1259 struct sk_buff *skb_prev;
1263 fraggap = skb_prev->len - maxfraglen;
1265 alloclen = fragheaderlen + hh_len + fraggap + 15;
1266 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1267 if (unlikely(!skb)) {
1273 * Fill in the control structures
1275 skb->ip_summed = CHECKSUM_NONE;
1277 skb_reserve(skb, hh_len);
1280 * Find where to start putting bytes.
1282 skb_put(skb, fragheaderlen + fraggap);
1283 skb_reset_network_header(skb);
1284 skb->transport_header = (skb->network_header +
1287 skb->csum = skb_copy_and_csum_bits(skb_prev,
1289 skb_transport_header(skb),
1291 skb_prev->csum = csum_sub(skb_prev->csum,
1293 pskb_trim_unique(skb_prev, maxfraglen);
1297 * Put the packet on the pending queue.
1299 __skb_queue_tail(&sk->sk_write_queue, skb);
1306 if (skb_append_pagefrags(skb, page, offset, len)) {
1311 if (skb->ip_summed == CHECKSUM_NONE) {
1313 csum = csum_page(page, offset, len);
1314 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1318 skb->data_len += len;
1319 skb->truesize += len;
1320 atomic_add(len, &sk->sk_wmem_alloc);
1327 cork->length -= size;
1328 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1332 static void ip_cork_release(struct inet_cork *cork)
1334 cork->flags &= ~IPCORK_OPT;
1337 dst_release(cork->dst);
1342 * Combined all pending IP fragments on the socket as one IP datagram
1343 * and push them out.
1345 struct sk_buff *__ip_make_skb(struct sock *sk,
1347 struct sk_buff_head *queue,
1348 struct inet_cork *cork)
1350 struct sk_buff *skb, *tmp_skb;
1351 struct sk_buff **tail_skb;
1352 struct inet_sock *inet = inet_sk(sk);
1353 struct net *net = sock_net(sk);
1354 struct ip_options *opt = NULL;
1355 struct rtable *rt = (struct rtable *)cork->dst;
1360 skb = __skb_dequeue(queue);
1363 tail_skb = &(skb_shinfo(skb)->frag_list);
1365 /* move skb->data to ip header from ext header */
1366 if (skb->data < skb_network_header(skb))
1367 __skb_pull(skb, skb_network_offset(skb));
1368 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1369 __skb_pull(tmp_skb, skb_network_header_len(skb));
1370 *tail_skb = tmp_skb;
1371 tail_skb = &(tmp_skb->next);
1372 skb->len += tmp_skb->len;
1373 skb->data_len += tmp_skb->len;
1374 skb->truesize += tmp_skb->truesize;
1375 tmp_skb->destructor = NULL;
1379 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1380 * to fragment the frame generated here. No matter, what transforms
1381 * how transforms change size of the packet, it will come out.
1383 skb->ignore_df = ip_sk_ignore_df(sk);
1385 /* DF bit is set when we want to see DF on outgoing frames.
1386 * If ignore_df is set too, we still allow to fragment this frame
1388 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1389 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1390 (skb->len <= dst_mtu(&rt->dst) &&
1391 ip_dont_fragment(sk, &rt->dst)))
1394 if (cork->flags & IPCORK_OPT)
1399 else if (rt->rt_type == RTN_MULTICAST)
1402 ttl = ip_select_ttl(inet, &rt->dst);
1407 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1410 iph->protocol = sk->sk_protocol;
1411 ip_copy_addrs(iph, fl4);
1412 ip_select_ident(net, skb, sk);
1415 iph->ihl += opt->optlen>>2;
1416 ip_options_build(skb, opt, cork->addr, rt, 0);
1419 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1420 skb->mark = sk->sk_mark;
1422 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1426 skb_dst_set(skb, &rt->dst);
1428 if (iph->protocol == IPPROTO_ICMP)
1429 icmp_out_count(net, ((struct icmphdr *)
1430 skb_transport_header(skb))->type);
1432 ip_cork_release(cork);
1437 int ip_send_skb(struct net *net, struct sk_buff *skb)
1441 err = ip_local_out(skb);
1444 err = net_xmit_errno(err);
1446 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1452 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1454 struct sk_buff *skb;
1456 skb = ip_finish_skb(sk, fl4);
1460 /* Netfilter gets whole the not fragmented skb. */
1461 return ip_send_skb(sock_net(sk), skb);
1465 * Throw away all pending data on the socket.
1467 static void __ip_flush_pending_frames(struct sock *sk,
1468 struct sk_buff_head *queue,
1469 struct inet_cork *cork)
1471 struct sk_buff *skb;
1473 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1476 ip_cork_release(cork);
1479 void ip_flush_pending_frames(struct sock *sk)
1481 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1484 struct sk_buff *ip_make_skb(struct sock *sk,
1486 int getfrag(void *from, char *to, int offset,
1487 int len, int odd, struct sk_buff *skb),
1488 void *from, int length, int transhdrlen,
1489 struct ipcm_cookie *ipc, struct rtable **rtp,
1492 struct inet_cork cork;
1493 struct sk_buff_head queue;
1496 if (flags & MSG_PROBE)
1499 __skb_queue_head_init(&queue);
1504 err = ip_setup_cork(sk, &cork, ipc, rtp);
1506 return ERR_PTR(err);
1508 err = __ip_append_data(sk, fl4, &queue, &cork,
1509 ¤t->task_frag, getfrag,
1510 from, length, transhdrlen, flags);
1512 __ip_flush_pending_frames(sk, &queue, &cork);
1513 return ERR_PTR(err);
1516 return __ip_make_skb(sk, fl4, &queue, &cork);
1520 * Fetch data from kernel space and fill in checksum if needed.
1522 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1523 int len, int odd, struct sk_buff *skb)
1527 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1528 skb->csum = csum_block_add(skb->csum, csum, odd);
1533 * Generic function to send a packet as reply to another packet.
1534 * Used to send some TCP resets/acks so far.
1536 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1537 const struct ip_options *sopt,
1538 __be32 daddr, __be32 saddr,
1539 const struct ip_reply_arg *arg,
1542 struct ip_options_data replyopts;
1543 struct ipcm_cookie ipc;
1545 struct rtable *rt = skb_rtable(skb);
1546 struct net *net = sock_net(sk);
1547 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 flowi4_init_output(&fl4, arg->bound_dev_if,
1567 IP4_REPLY_MARK(net, skb->mark),
1569 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1570 ip_reply_arg_flowi_flags(arg),
1572 tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1573 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1574 rt = ip_route_output_key(net, &fl4);
1578 inet_sk(sk)->tos = arg->tos;
1580 sk->sk_priority = skb->priority;
1581 sk->sk_protocol = ip_hdr(skb)->protocol;
1582 sk->sk_bound_dev_if = arg->bound_dev_if;
1583 sk->sk_sndbuf = sysctl_wmem_default;
1584 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1585 len, 0, &ipc, &rt, MSG_DONTWAIT);
1586 if (unlikely(err)) {
1587 ip_flush_pending_frames(sk);
1591 nskb = skb_peek(&sk->sk_write_queue);
1593 if (arg->csumoffset >= 0)
1594 *((__sum16 *)skb_transport_header(nskb) +
1595 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1597 nskb->ip_summed = CHECKSUM_NONE;
1598 skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1599 ip_push_pending_frames(sk, &fl4);
1605 void __init ip_init(void)
1610 #if defined(CONFIG_IP_MULTICAST)