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 static 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)
120 err = __ip_local_out(skb);
121 if (likely(err == 1))
122 err = dst_output(sk, skb);
126 EXPORT_SYMBOL_GPL(ip_local_out_sk);
128 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
130 int ttl = inet->uc_ttl;
133 ttl = ip4_dst_hoplimit(dst);
138 * Add an ip header to a skbuff and send it out.
141 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
142 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
144 struct inet_sock *inet = inet_sk(sk);
145 struct rtable *rt = skb_rtable(skb);
148 /* Build the IP header. */
149 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
150 skb_reset_network_header(skb);
154 iph->tos = inet->tos;
155 iph->ttl = ip_select_ttl(inet, &rt->dst);
156 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
158 iph->protocol = sk->sk_protocol;
159 if (ip_dont_fragment(sk, &rt->dst)) {
160 iph->frag_off = htons(IP_DF);
164 __ip_select_ident(sock_net(sk), iph, 1);
167 if (opt && opt->opt.optlen) {
168 iph->ihl += opt->opt.optlen>>2;
169 ip_options_build(skb, &opt->opt, daddr, rt, 0);
172 skb->priority = sk->sk_priority;
173 skb->mark = sk->sk_mark;
176 return ip_local_out(skb);
178 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
180 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
182 struct dst_entry *dst = skb_dst(skb);
183 struct rtable *rt = (struct rtable *)dst;
184 struct net_device *dev = dst->dev;
185 unsigned int hh_len = LL_RESERVED_SPACE(dev);
186 struct neighbour *neigh;
189 if (rt->rt_type == RTN_MULTICAST) {
190 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
191 } else if (rt->rt_type == RTN_BROADCAST)
192 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
194 /* Be paranoid, rather than too clever. */
195 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
196 struct sk_buff *skb2;
198 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
204 skb_set_owner_w(skb2, skb->sk);
210 nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
211 neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
212 if (unlikely(!neigh))
213 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
214 if (!IS_ERR(neigh)) {
215 int res = dst_neigh_output(dst, neigh, skb);
217 rcu_read_unlock_bh();
220 rcu_read_unlock_bh();
222 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
228 static int ip_finish_output_gso(struct net *net, struct sock *sk,
229 struct sk_buff *skb, unsigned int mtu)
231 netdev_features_t features;
232 struct sk_buff *segs;
235 /* common case: locally created skb or seglen is <= mtu */
236 if (((IPCB(skb)->flags & IPSKB_FORWARDED) == 0) ||
237 skb_gso_network_seglen(skb) <= mtu)
238 return ip_finish_output2(net, sk, skb);
240 /* Slowpath - GSO segment length is exceeding the dst MTU.
242 * This can happen in two cases:
243 * 1) TCP GRO packet, DF bit not set
244 * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
245 * from host network stack.
247 features = netif_skb_features(skb);
248 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
249 if (IS_ERR_OR_NULL(segs)) {
257 struct sk_buff *nskb = segs->next;
261 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
271 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
275 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
276 /* Policy lookup after SNAT yielded a new policy */
277 if (skb_dst(skb)->xfrm) {
278 IPCB(skb)->flags |= IPSKB_REROUTED;
279 return dst_output(sk, skb);
282 mtu = ip_skb_dst_mtu(skb);
284 return ip_finish_output_gso(net, sk, skb, mtu);
286 if (skb->len > mtu || (IPCB(skb)->flags & IPSKB_FRAG_PMTU))
287 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
289 return ip_finish_output2(net, sk, skb);
292 int ip_mc_output(struct sock *sk, struct sk_buff *skb)
294 struct rtable *rt = skb_rtable(skb);
295 struct net_device *dev = rt->dst.dev;
296 struct net *net = dev_net(dev);
299 * If the indicated interface is up and running, send the packet.
301 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
304 skb->protocol = htons(ETH_P_IP);
307 * Multicasts are looped back for other local users
310 if (rt->rt_flags&RTCF_MULTICAST) {
312 #ifdef CONFIG_IP_MROUTE
313 /* Small optimization: do not loopback not local frames,
314 which returned after forwarding; they will be dropped
315 by ip_mr_input in any case.
316 Note, that local frames are looped back to be delivered
319 This check is duplicated in ip_mr_input at the moment.
322 ((rt->rt_flags & RTCF_LOCAL) ||
323 !(IPCB(skb)->flags & IPSKB_FORWARDED))
326 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
328 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
329 net, sk, newskb, NULL, newskb->dev,
333 /* Multicasts with ttl 0 must not go beyond the host */
335 if (ip_hdr(skb)->ttl == 0) {
341 if (rt->rt_flags&RTCF_BROADCAST) {
342 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
344 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
345 net, sk, newskb, NULL, newskb->dev,
349 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
350 net, sk, skb, NULL, skb->dev,
352 !(IPCB(skb)->flags & IPSKB_REROUTED));
355 int ip_output(struct sock *sk, struct sk_buff *skb)
357 struct net_device *dev = skb_dst(skb)->dev;
358 struct net *net = dev_net(dev);
360 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
363 skb->protocol = htons(ETH_P_IP);
365 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
366 net, sk, skb, NULL, dev,
368 !(IPCB(skb)->flags & IPSKB_REROUTED));
372 * copy saddr and daddr, possibly using 64bit load/stores
374 * iph->saddr = fl4->saddr;
375 * iph->daddr = fl4->daddr;
377 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
379 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
380 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
381 memcpy(&iph->saddr, &fl4->saddr,
382 sizeof(fl4->saddr) + sizeof(fl4->daddr));
385 /* Note: skb->sk can be different from sk, in case of tunnels */
386 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
388 struct inet_sock *inet = inet_sk(sk);
389 struct ip_options_rcu *inet_opt;
395 /* Skip all of this if the packet is already routed,
396 * f.e. by something like SCTP.
399 inet_opt = rcu_dereference(inet->inet_opt);
401 rt = skb_rtable(skb);
405 /* Make sure we can route this packet. */
406 rt = (struct rtable *)__sk_dst_check(sk, 0);
410 /* Use correct destination address if we have options. */
411 daddr = inet->inet_daddr;
412 if (inet_opt && inet_opt->opt.srr)
413 daddr = inet_opt->opt.faddr;
415 /* If this fails, retransmit mechanism of transport layer will
416 * keep trying until route appears or the connection times
419 rt = ip_route_output_ports(sock_net(sk), fl4, sk,
420 daddr, inet->inet_saddr,
425 sk->sk_bound_dev_if);
428 sk_setup_caps(sk, &rt->dst);
430 skb_dst_set_noref(skb, &rt->dst);
433 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
436 /* OK, we know where to send it, allocate and build IP header. */
437 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
438 skb_reset_network_header(skb);
440 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
441 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
442 iph->frag_off = htons(IP_DF);
445 iph->ttl = ip_select_ttl(inet, &rt->dst);
446 iph->protocol = sk->sk_protocol;
447 ip_copy_addrs(iph, fl4);
449 /* Transport layer set skb->h.foo itself. */
451 if (inet_opt && inet_opt->opt.optlen) {
452 iph->ihl += inet_opt->opt.optlen >> 2;
453 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
456 ip_select_ident_segs(sock_net(sk), skb, sk,
457 skb_shinfo(skb)->gso_segs ?: 1);
459 /* TODO : should we use skb->sk here instead of sk ? */
460 skb->priority = sk->sk_priority;
461 skb->mark = sk->sk_mark;
463 res = ip_local_out(skb);
469 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
471 return -EHOSTUNREACH;
473 EXPORT_SYMBOL(ip_queue_xmit);
475 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
477 to->pkt_type = from->pkt_type;
478 to->priority = from->priority;
479 to->protocol = from->protocol;
481 skb_dst_copy(to, from);
483 to->mark = from->mark;
485 /* Copy the flags to each fragment. */
486 IPCB(to)->flags = IPCB(from)->flags;
488 #ifdef CONFIG_NET_SCHED
489 to->tc_index = from->tc_index;
492 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
493 to->ipvs_property = from->ipvs_property;
495 skb_copy_secmark(to, from);
498 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
500 int (*output)(struct net *, struct sock *, struct sk_buff *))
502 struct iphdr *iph = ip_hdr(skb);
504 if ((iph->frag_off & htons(IP_DF)) == 0)
505 return ip_do_fragment(net, sk, skb, output);
507 if (unlikely(!skb->ignore_df ||
508 (IPCB(skb)->frag_max_size &&
509 IPCB(skb)->frag_max_size > mtu))) {
510 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
511 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
517 return ip_do_fragment(net, sk, skb, output);
521 * This IP datagram is too large to be sent in one piece. Break it up into
522 * smaller pieces (each of size equal to IP header plus
523 * a block of the data of the original IP data part) that will yet fit in a
524 * single device frame, and queue such a frame for sending.
527 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
528 int (*output)(struct net *, struct sock *, struct sk_buff *))
532 struct net_device *dev;
533 struct sk_buff *skb2;
534 unsigned int mtu, hlen, left, len, ll_rs;
536 __be16 not_last_frag;
537 struct rtable *rt = skb_rtable(skb);
543 * Point into the IP datagram header.
548 mtu = ip_skb_dst_mtu(skb);
549 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
550 mtu = IPCB(skb)->frag_max_size;
553 * Setup starting values.
557 mtu = mtu - hlen; /* Size of data space */
558 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
560 /* When frag_list is given, use it. First, check its validity:
561 * some transformers could create wrong frag_list or break existing
562 * one, it is not prohibited. In this case fall back to copying.
564 * LATER: this step can be merged to real generation of fragments,
565 * we can switch to copy when see the first bad fragment.
567 if (skb_has_frag_list(skb)) {
568 struct sk_buff *frag, *frag2;
569 int first_len = skb_pagelen(skb);
571 if (first_len - hlen > mtu ||
572 ((first_len - hlen) & 7) ||
573 ip_is_fragment(iph) ||
577 skb_walk_frags(skb, frag) {
578 /* Correct geometry. */
579 if (frag->len > mtu ||
580 ((frag->len & 7) && frag->next) ||
581 skb_headroom(frag) < hlen)
582 goto slow_path_clean;
584 /* Partially cloned skb? */
585 if (skb_shared(frag))
586 goto slow_path_clean;
591 frag->destructor = sock_wfree;
593 skb->truesize -= frag->truesize;
596 /* Everything is OK. Generate! */
600 frag = skb_shinfo(skb)->frag_list;
601 skb_frag_list_init(skb);
602 skb->data_len = first_len - skb_headlen(skb);
603 skb->len = first_len;
604 iph->tot_len = htons(first_len);
605 iph->frag_off = htons(IP_MF);
609 /* Prepare header of the next frame,
610 * before previous one went down. */
612 frag->ip_summed = CHECKSUM_NONE;
613 skb_reset_transport_header(frag);
614 __skb_push(frag, hlen);
615 skb_reset_network_header(frag);
616 memcpy(skb_network_header(frag), iph, hlen);
618 iph->tot_len = htons(frag->len);
619 ip_copy_metadata(frag, skb);
621 ip_options_fragment(frag);
622 offset += skb->len - hlen;
623 iph->frag_off = htons(offset>>3);
625 iph->frag_off |= htons(IP_MF);
626 /* Ready, complete checksum */
630 err = output(net, sk, skb);
633 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
643 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
652 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
656 skb_walk_frags(skb, frag2) {
660 frag2->destructor = NULL;
661 skb->truesize += frag2->truesize;
666 /* for offloaded checksums cleanup checksum before fragmentation */
667 if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
671 left = skb->len - hlen; /* Space per frame */
672 ptr = hlen; /* Where to start from */
674 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
677 * Fragment the datagram.
680 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
681 not_last_frag = iph->frag_off & htons(IP_MF);
684 * Keep copying data until we run out.
689 /* IF: it doesn't fit, use 'mtu' - the data space left */
692 /* IF: we are not sending up to and including the packet end
693 then align the next start on an eight byte boundary */
698 /* Allocate buffer */
699 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
706 * Set up data on packet
709 ip_copy_metadata(skb2, skb);
710 skb_reserve(skb2, ll_rs);
711 skb_put(skb2, len + hlen);
712 skb_reset_network_header(skb2);
713 skb2->transport_header = skb2->network_header + hlen;
716 * Charge the memory for the fragment to any owner
721 skb_set_owner_w(skb2, skb->sk);
724 * Copy the packet header into the new buffer.
727 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
730 * Copy a block of the IP datagram.
732 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
737 * Fill in the new header fields.
740 iph->frag_off = htons((offset >> 3));
742 if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
743 iph->frag_off |= htons(IP_DF);
745 /* ANK: dirty, but effective trick. Upgrade options only if
746 * the segment to be fragmented was THE FIRST (otherwise,
747 * options are already fixed) and make it ONCE
748 * on the initial skb, so that all the following fragments
749 * will inherit fixed options.
752 ip_options_fragment(skb);
755 * Added AC : If we are fragmenting a fragment that's not the
756 * last fragment then keep MF on each bit
758 if (left > 0 || not_last_frag)
759 iph->frag_off |= htons(IP_MF);
764 * Put this fragment into the sending queue.
766 iph->tot_len = htons(len + hlen);
770 err = output(net, sk, skb2);
774 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
777 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
782 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
785 EXPORT_SYMBOL(ip_do_fragment);
788 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
790 struct msghdr *msg = from;
792 if (skb->ip_summed == CHECKSUM_PARTIAL) {
793 if (copy_from_iter(to, len, &msg->msg_iter) != len)
797 if (csum_and_copy_from_iter(to, len, &csum, &msg->msg_iter) != len)
799 skb->csum = csum_block_add(skb->csum, csum, odd);
803 EXPORT_SYMBOL(ip_generic_getfrag);
806 csum_page(struct page *page, int offset, int copy)
811 csum = csum_partial(kaddr + offset, copy, 0);
816 static inline int ip_ufo_append_data(struct sock *sk,
817 struct sk_buff_head *queue,
818 int getfrag(void *from, char *to, int offset, int len,
819 int odd, struct sk_buff *skb),
820 void *from, int length, int hh_len, int fragheaderlen,
821 int transhdrlen, int maxfraglen, unsigned int flags)
826 /* There is support for UDP fragmentation offload by network
827 * device, so create one single skb packet containing complete
830 skb = skb_peek_tail(queue);
832 skb = sock_alloc_send_skb(sk,
833 hh_len + fragheaderlen + transhdrlen + 20,
834 (flags & MSG_DONTWAIT), &err);
839 /* reserve space for Hardware header */
840 skb_reserve(skb, hh_len);
842 /* create space for UDP/IP header */
843 skb_put(skb, fragheaderlen + transhdrlen);
845 /* initialize network header pointer */
846 skb_reset_network_header(skb);
848 /* initialize protocol header pointer */
849 skb->transport_header = skb->network_header + fragheaderlen;
853 __skb_queue_tail(queue, skb);
854 } else if (skb_is_gso(skb)) {
858 skb->ip_summed = CHECKSUM_PARTIAL;
859 /* specify the length of each IP datagram fragment */
860 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
861 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
864 return skb_append_datato_frags(sk, skb, getfrag, from,
865 (length - transhdrlen));
868 static int __ip_append_data(struct sock *sk,
870 struct sk_buff_head *queue,
871 struct inet_cork *cork,
872 struct page_frag *pfrag,
873 int getfrag(void *from, char *to, int offset,
874 int len, int odd, struct sk_buff *skb),
875 void *from, int length, int transhdrlen,
878 struct inet_sock *inet = inet_sk(sk);
881 struct ip_options *opt = cork->opt;
888 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
889 int csummode = CHECKSUM_NONE;
890 struct rtable *rt = (struct rtable *)cork->dst;
893 skb = skb_peek_tail(queue);
895 exthdrlen = !skb ? rt->dst.header_len : 0;
896 mtu = cork->fragsize;
897 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
898 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
899 tskey = sk->sk_tskey++;
901 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
903 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
904 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
905 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
907 if (cork->length + length > maxnonfragsize - fragheaderlen) {
908 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
909 mtu - (opt ? opt->optlen : 0));
914 * transhdrlen > 0 means that this is the first fragment and we wish
915 * it won't be fragmented in the future.
918 length + fragheaderlen <= mtu &&
919 rt->dst.dev->features & NETIF_F_V4_CSUM &&
921 csummode = CHECKSUM_PARTIAL;
923 cork->length += length;
924 if (((length > mtu) || (skb && skb_is_gso(skb))) &&
925 (sk->sk_protocol == IPPROTO_UDP) &&
926 (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len &&
927 (sk->sk_type == SOCK_DGRAM)) {
928 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
929 hh_len, fragheaderlen, transhdrlen,
936 /* So, what's going on in the loop below?
938 * We use calculated fragment length to generate chained skb,
939 * each of segments is IP fragment ready for sending to network after
940 * adding appropriate IP header.
947 /* Check if the remaining data fits into current packet. */
948 copy = mtu - skb->len;
950 copy = maxfraglen - skb->len;
953 unsigned int datalen;
954 unsigned int fraglen;
955 unsigned int fraggap;
956 unsigned int alloclen;
957 struct sk_buff *skb_prev;
961 fraggap = skb_prev->len - maxfraglen;
966 * If remaining data exceeds the mtu,
967 * we know we need more fragment(s).
969 datalen = length + fraggap;
970 if (datalen > mtu - fragheaderlen)
971 datalen = maxfraglen - fragheaderlen;
972 fraglen = datalen + fragheaderlen;
974 if ((flags & MSG_MORE) &&
975 !(rt->dst.dev->features&NETIF_F_SG))
980 alloclen += exthdrlen;
982 /* The last fragment gets additional space at tail.
983 * Note, with MSG_MORE we overallocate on fragments,
984 * because we have no idea what fragment will be
987 if (datalen == length + fraggap)
988 alloclen += rt->dst.trailer_len;
991 skb = sock_alloc_send_skb(sk,
992 alloclen + hh_len + 15,
993 (flags & MSG_DONTWAIT), &err);
996 if (atomic_read(&sk->sk_wmem_alloc) <=
998 skb = sock_wmalloc(sk,
999 alloclen + hh_len + 15, 1,
1008 * Fill in the control structures
1010 skb->ip_summed = csummode;
1012 skb_reserve(skb, hh_len);
1014 /* only the initial fragment is time stamped */
1015 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1017 skb_shinfo(skb)->tskey = tskey;
1021 * Find where to start putting bytes.
1023 data = skb_put(skb, fraglen + exthdrlen);
1024 skb_set_network_header(skb, exthdrlen);
1025 skb->transport_header = (skb->network_header +
1027 data += fragheaderlen + exthdrlen;
1030 skb->csum = skb_copy_and_csum_bits(
1031 skb_prev, maxfraglen,
1032 data + transhdrlen, fraggap, 0);
1033 skb_prev->csum = csum_sub(skb_prev->csum,
1036 pskb_trim_unique(skb_prev, maxfraglen);
1039 copy = datalen - transhdrlen - fraggap;
1040 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1047 length -= datalen - fraggap;
1050 csummode = CHECKSUM_NONE;
1053 * Put the packet on the pending queue.
1055 __skb_queue_tail(queue, skb);
1062 if (!(rt->dst.dev->features&NETIF_F_SG)) {
1066 if (getfrag(from, skb_put(skb, copy),
1067 offset, copy, off, skb) < 0) {
1068 __skb_trim(skb, off);
1073 int i = skb_shinfo(skb)->nr_frags;
1076 if (!sk_page_frag_refill(sk, pfrag))
1079 if (!skb_can_coalesce(skb, i, pfrag->page,
1082 if (i == MAX_SKB_FRAGS)
1085 __skb_fill_page_desc(skb, i, pfrag->page,
1087 skb_shinfo(skb)->nr_frags = ++i;
1088 get_page(pfrag->page);
1090 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1092 page_address(pfrag->page) + pfrag->offset,
1093 offset, copy, skb->len, skb) < 0)
1096 pfrag->offset += copy;
1097 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1099 skb->data_len += copy;
1100 skb->truesize += copy;
1101 atomic_add(copy, &sk->sk_wmem_alloc);
1112 cork->length -= length;
1113 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1117 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1118 struct ipcm_cookie *ipc, struct rtable **rtp)
1120 struct ip_options_rcu *opt;
1124 * setup for corking.
1129 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1131 if (unlikely(!cork->opt))
1134 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1135 cork->flags |= IPCORK_OPT;
1136 cork->addr = ipc->addr;
1142 * We steal reference to this route, caller should not release it
1145 cork->fragsize = ip_sk_use_pmtu(sk) ?
1146 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1147 cork->dst = &rt->dst;
1149 cork->ttl = ipc->ttl;
1150 cork->tos = ipc->tos;
1151 cork->priority = ipc->priority;
1152 cork->tx_flags = ipc->tx_flags;
1158 * ip_append_data() and ip_append_page() can make one large IP datagram
1159 * from many pieces of data. Each pieces will be holded on the socket
1160 * until ip_push_pending_frames() is called. Each piece can be a page
1163 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1164 * this interface potentially.
1166 * LATER: length must be adjusted by pad at tail, when it is required.
1168 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1169 int getfrag(void *from, char *to, int offset, int len,
1170 int odd, struct sk_buff *skb),
1171 void *from, int length, int transhdrlen,
1172 struct ipcm_cookie *ipc, struct rtable **rtp,
1175 struct inet_sock *inet = inet_sk(sk);
1178 if (flags&MSG_PROBE)
1181 if (skb_queue_empty(&sk->sk_write_queue)) {
1182 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1189 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1190 sk_page_frag(sk), getfrag,
1191 from, length, transhdrlen, flags);
1194 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1195 int offset, size_t size, int flags)
1197 struct inet_sock *inet = inet_sk(sk);
1198 struct sk_buff *skb;
1200 struct ip_options *opt = NULL;
1201 struct inet_cork *cork;
1206 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1211 if (flags&MSG_PROBE)
1214 if (skb_queue_empty(&sk->sk_write_queue))
1217 cork = &inet->cork.base;
1218 rt = (struct rtable *)cork->dst;
1219 if (cork->flags & IPCORK_OPT)
1222 if (!(rt->dst.dev->features&NETIF_F_SG))
1225 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1226 mtu = cork->fragsize;
1228 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1229 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1230 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1232 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1233 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1234 mtu - (opt ? opt->optlen : 0));
1238 skb = skb_peek_tail(&sk->sk_write_queue);
1242 cork->length += size;
1243 if ((size + skb->len > mtu) &&
1244 (sk->sk_protocol == IPPROTO_UDP) &&
1245 (rt->dst.dev->features & NETIF_F_UFO)) {
1246 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1247 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1251 if (skb_is_gso(skb)) {
1255 /* Check if the remaining data fits into current packet. */
1256 len = mtu - skb->len;
1258 len = maxfraglen - skb->len;
1261 struct sk_buff *skb_prev;
1265 fraggap = skb_prev->len - maxfraglen;
1267 alloclen = fragheaderlen + hh_len + fraggap + 15;
1268 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1269 if (unlikely(!skb)) {
1275 * Fill in the control structures
1277 skb->ip_summed = CHECKSUM_NONE;
1279 skb_reserve(skb, hh_len);
1282 * Find where to start putting bytes.
1284 skb_put(skb, fragheaderlen + fraggap);
1285 skb_reset_network_header(skb);
1286 skb->transport_header = (skb->network_header +
1289 skb->csum = skb_copy_and_csum_bits(skb_prev,
1291 skb_transport_header(skb),
1293 skb_prev->csum = csum_sub(skb_prev->csum,
1295 pskb_trim_unique(skb_prev, maxfraglen);
1299 * Put the packet on the pending queue.
1301 __skb_queue_tail(&sk->sk_write_queue, skb);
1308 if (skb_append_pagefrags(skb, page, offset, len)) {
1313 if (skb->ip_summed == CHECKSUM_NONE) {
1315 csum = csum_page(page, offset, len);
1316 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1320 skb->data_len += len;
1321 skb->truesize += len;
1322 atomic_add(len, &sk->sk_wmem_alloc);
1329 cork->length -= size;
1330 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1334 static void ip_cork_release(struct inet_cork *cork)
1336 cork->flags &= ~IPCORK_OPT;
1339 dst_release(cork->dst);
1344 * Combined all pending IP fragments on the socket as one IP datagram
1345 * and push them out.
1347 struct sk_buff *__ip_make_skb(struct sock *sk,
1349 struct sk_buff_head *queue,
1350 struct inet_cork *cork)
1352 struct sk_buff *skb, *tmp_skb;
1353 struct sk_buff **tail_skb;
1354 struct inet_sock *inet = inet_sk(sk);
1355 struct net *net = sock_net(sk);
1356 struct ip_options *opt = NULL;
1357 struct rtable *rt = (struct rtable *)cork->dst;
1362 skb = __skb_dequeue(queue);
1365 tail_skb = &(skb_shinfo(skb)->frag_list);
1367 /* move skb->data to ip header from ext header */
1368 if (skb->data < skb_network_header(skb))
1369 __skb_pull(skb, skb_network_offset(skb));
1370 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1371 __skb_pull(tmp_skb, skb_network_header_len(skb));
1372 *tail_skb = tmp_skb;
1373 tail_skb = &(tmp_skb->next);
1374 skb->len += tmp_skb->len;
1375 skb->data_len += tmp_skb->len;
1376 skb->truesize += tmp_skb->truesize;
1377 tmp_skb->destructor = NULL;
1381 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1382 * to fragment the frame generated here. No matter, what transforms
1383 * how transforms change size of the packet, it will come out.
1385 skb->ignore_df = ip_sk_ignore_df(sk);
1387 /* DF bit is set when we want to see DF on outgoing frames.
1388 * If ignore_df is set too, we still allow to fragment this frame
1390 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1391 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1392 (skb->len <= dst_mtu(&rt->dst) &&
1393 ip_dont_fragment(sk, &rt->dst)))
1396 if (cork->flags & IPCORK_OPT)
1401 else if (rt->rt_type == RTN_MULTICAST)
1404 ttl = ip_select_ttl(inet, &rt->dst);
1409 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1412 iph->protocol = sk->sk_protocol;
1413 ip_copy_addrs(iph, fl4);
1414 ip_select_ident(net, skb, sk);
1417 iph->ihl += opt->optlen>>2;
1418 ip_options_build(skb, opt, cork->addr, rt, 0);
1421 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1422 skb->mark = sk->sk_mark;
1424 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1428 skb_dst_set(skb, &rt->dst);
1430 if (iph->protocol == IPPROTO_ICMP)
1431 icmp_out_count(net, ((struct icmphdr *)
1432 skb_transport_header(skb))->type);
1434 ip_cork_release(cork);
1439 int ip_send_skb(struct net *net, struct sk_buff *skb)
1443 err = ip_local_out(skb);
1446 err = net_xmit_errno(err);
1448 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1454 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1456 struct sk_buff *skb;
1458 skb = ip_finish_skb(sk, fl4);
1462 /* Netfilter gets whole the not fragmented skb. */
1463 return ip_send_skb(sock_net(sk), skb);
1467 * Throw away all pending data on the socket.
1469 static void __ip_flush_pending_frames(struct sock *sk,
1470 struct sk_buff_head *queue,
1471 struct inet_cork *cork)
1473 struct sk_buff *skb;
1475 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1478 ip_cork_release(cork);
1481 void ip_flush_pending_frames(struct sock *sk)
1483 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1486 struct sk_buff *ip_make_skb(struct sock *sk,
1488 int getfrag(void *from, char *to, int offset,
1489 int len, int odd, struct sk_buff *skb),
1490 void *from, int length, int transhdrlen,
1491 struct ipcm_cookie *ipc, struct rtable **rtp,
1494 struct inet_cork cork;
1495 struct sk_buff_head queue;
1498 if (flags & MSG_PROBE)
1501 __skb_queue_head_init(&queue);
1506 err = ip_setup_cork(sk, &cork, ipc, rtp);
1508 return ERR_PTR(err);
1510 err = __ip_append_data(sk, fl4, &queue, &cork,
1511 ¤t->task_frag, getfrag,
1512 from, length, transhdrlen, flags);
1514 __ip_flush_pending_frames(sk, &queue, &cork);
1515 return ERR_PTR(err);
1518 return __ip_make_skb(sk, fl4, &queue, &cork);
1522 * Fetch data from kernel space and fill in checksum if needed.
1524 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1525 int len, int odd, struct sk_buff *skb)
1529 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1530 skb->csum = csum_block_add(skb->csum, csum, odd);
1535 * Generic function to send a packet as reply to another packet.
1536 * Used to send some TCP resets/acks so far.
1538 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1539 const struct ip_options *sopt,
1540 __be32 daddr, __be32 saddr,
1541 const struct ip_reply_arg *arg,
1544 struct ip_options_data replyopts;
1545 struct ipcm_cookie ipc;
1547 struct rtable *rt = skb_rtable(skb);
1548 struct net *net = sock_net(sk);
1549 struct sk_buff *nskb;
1553 if (__ip_options_echo(&replyopts.opt.opt, skb, sopt))
1562 if (replyopts.opt.opt.optlen) {
1563 ipc.opt = &replyopts.opt;
1565 if (replyopts.opt.opt.srr)
1566 daddr = replyopts.opt.opt.faddr;
1569 oif = arg->bound_dev_if;
1570 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1573 flowi4_init_output(&fl4, oif,
1574 IP4_REPLY_MARK(net, skb->mark),
1576 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1577 ip_reply_arg_flowi_flags(arg),
1579 tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1580 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1581 rt = ip_route_output_key(net, &fl4);
1585 inet_sk(sk)->tos = arg->tos;
1587 sk->sk_priority = skb->priority;
1588 sk->sk_protocol = ip_hdr(skb)->protocol;
1589 sk->sk_bound_dev_if = arg->bound_dev_if;
1590 sk->sk_sndbuf = sysctl_wmem_default;
1591 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1592 len, 0, &ipc, &rt, MSG_DONTWAIT);
1593 if (unlikely(err)) {
1594 ip_flush_pending_frames(sk);
1598 nskb = skb_peek(&sk->sk_write_queue);
1600 if (arg->csumoffset >= 0)
1601 *((__sum16 *)skb_transport_header(nskb) +
1602 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1604 nskb->ip_summed = CHECKSUM_NONE;
1605 skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1606 ip_push_pending_frames(sk, &fl4);
1612 void __init ip_init(void)
1617 #if defined(CONFIG_IP_MULTICAST)