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 /* Generate a checksum for an outgoing IP datagram. */
87 void ip_send_check(struct iphdr *iph)
90 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
92 EXPORT_SYMBOL(ip_send_check);
94 int __ip_local_out(struct sk_buff *skb)
96 struct iphdr *iph = ip_hdr(skb);
98 iph->tot_len = htons(skb->len);
100 skb->protocol = htons(ETH_P_IP);
102 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT, skb, NULL,
103 skb_dst(skb)->dev, dst_output);
106 int ip_local_out_sk(struct sock *sk, struct sk_buff *skb)
110 err = __ip_local_out(skb);
111 if (likely(err == 1))
112 err = dst_output_sk(sk, skb);
116 EXPORT_SYMBOL_GPL(ip_local_out_sk);
118 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
120 int ttl = inet->uc_ttl;
123 ttl = ip4_dst_hoplimit(dst);
128 * Add an ip header to a skbuff and send it out.
131 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
132 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
134 struct inet_sock *inet = inet_sk(sk);
135 struct rtable *rt = skb_rtable(skb);
138 /* Build the IP header. */
139 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
140 skb_reset_network_header(skb);
144 iph->tos = inet->tos;
145 if (ip_dont_fragment(sk, &rt->dst))
146 iph->frag_off = htons(IP_DF);
149 iph->ttl = ip_select_ttl(inet, &rt->dst);
150 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
152 iph->protocol = sk->sk_protocol;
153 ip_select_ident(sock_net(sk), skb, sk);
155 if (opt && opt->opt.optlen) {
156 iph->ihl += opt->opt.optlen>>2;
157 ip_options_build(skb, &opt->opt, daddr, rt, 0);
160 skb->priority = sk->sk_priority;
162 skb->mark = sk->sk_mark;
165 return ip_local_out(skb);
167 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
169 static inline int ip_finish_output2(struct sk_buff *skb)
171 struct dst_entry *dst = skb_dst(skb);
172 struct rtable *rt = (struct rtable *)dst;
173 struct net_device *dev = dst->dev;
174 unsigned int hh_len = LL_RESERVED_SPACE(dev);
175 struct neighbour *neigh;
178 if (rt->rt_type == RTN_MULTICAST) {
179 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
180 } else if (rt->rt_type == RTN_BROADCAST)
181 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
183 /* Be paranoid, rather than too clever. */
184 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
185 struct sk_buff *skb2;
187 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
193 skb_set_owner_w(skb2, skb->sk);
199 nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
200 neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
201 if (unlikely(!neigh))
202 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
203 if (!IS_ERR(neigh)) {
204 int res = dst_neigh_output(dst, neigh, skb);
206 rcu_read_unlock_bh();
209 rcu_read_unlock_bh();
211 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
217 static int ip_finish_output_gso(struct sk_buff *skb)
219 netdev_features_t features;
220 struct sk_buff *segs;
223 /* common case: seglen is <= mtu */
224 if (skb_gso_network_seglen(skb) <= ip_skb_dst_mtu(skb))
225 return ip_finish_output2(skb);
227 /* Slowpath - GSO segment length is exceeding the dst MTU.
229 * This can happen in two cases:
230 * 1) TCP GRO packet, DF bit not set
231 * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
232 * from host network stack.
234 features = netif_skb_features(skb);
235 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
236 if (IS_ERR_OR_NULL(segs)) {
244 struct sk_buff *nskb = segs->next;
248 err = ip_fragment(segs, ip_finish_output2);
258 static int ip_finish_output(struct sk_buff *skb)
260 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
261 /* Policy lookup after SNAT yielded a new policy */
262 if (skb_dst(skb)->xfrm != NULL) {
263 IPCB(skb)->flags |= IPSKB_REROUTED;
264 return dst_output(skb);
268 return ip_finish_output_gso(skb);
270 if (skb->len > ip_skb_dst_mtu(skb))
271 return ip_fragment(skb, ip_finish_output2);
273 return ip_finish_output2(skb);
276 int ip_mc_output(struct sock *sk, struct sk_buff *skb)
278 struct rtable *rt = skb_rtable(skb);
279 struct net_device *dev = rt->dst.dev;
282 * If the indicated interface is up and running, send the packet.
284 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
287 skb->protocol = htons(ETH_P_IP);
290 * Multicasts are looped back for other local users
293 if (rt->rt_flags&RTCF_MULTICAST) {
295 #ifdef CONFIG_IP_MROUTE
296 /* Small optimization: do not loopback not local frames,
297 which returned after forwarding; they will be dropped
298 by ip_mr_input in any case.
299 Note, that local frames are looped back to be delivered
302 This check is duplicated in ip_mr_input at the moment.
305 ((rt->rt_flags & RTCF_LOCAL) ||
306 !(IPCB(skb)->flags & IPSKB_FORWARDED))
309 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
311 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
312 newskb, NULL, newskb->dev,
316 /* Multicasts with ttl 0 must not go beyond the host */
318 if (ip_hdr(skb)->ttl == 0) {
324 if (rt->rt_flags&RTCF_BROADCAST) {
325 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
327 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
328 NULL, newskb->dev, dev_loopback_xmit);
331 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
332 skb->dev, ip_finish_output,
333 !(IPCB(skb)->flags & IPSKB_REROUTED));
336 int ip_output(struct sock *sk, struct sk_buff *skb)
338 struct net_device *dev = skb_dst(skb)->dev;
340 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
343 skb->protocol = htons(ETH_P_IP);
345 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
347 !(IPCB(skb)->flags & IPSKB_REROUTED));
351 * copy saddr and daddr, possibly using 64bit load/stores
353 * iph->saddr = fl4->saddr;
354 * iph->daddr = fl4->daddr;
356 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
358 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
359 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
360 memcpy(&iph->saddr, &fl4->saddr,
361 sizeof(fl4->saddr) + sizeof(fl4->daddr));
364 /* Note: skb->sk can be different from sk, in case of tunnels */
365 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
367 struct inet_sock *inet = inet_sk(sk);
368 struct ip_options_rcu *inet_opt;
374 /* Skip all of this if the packet is already routed,
375 * f.e. by something like SCTP.
378 inet_opt = rcu_dereference(inet->inet_opt);
380 rt = skb_rtable(skb);
384 /* Make sure we can route this packet. */
385 rt = (struct rtable *)__sk_dst_check(sk, 0);
389 /* Use correct destination address if we have options. */
390 daddr = inet->inet_daddr;
391 if (inet_opt && inet_opt->opt.srr)
392 daddr = inet_opt->opt.faddr;
394 /* If this fails, retransmit mechanism of transport layer will
395 * keep trying until route appears or the connection times
398 rt = ip_route_output_ports(sock_net(sk), fl4, sk,
399 daddr, inet->inet_saddr,
404 sk->sk_bound_dev_if);
407 sk_setup_caps(sk, &rt->dst);
409 skb_dst_set_noref(skb, &rt->dst);
412 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
415 /* OK, we know where to send it, allocate and build IP header. */
416 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
417 skb_reset_network_header(skb);
419 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
420 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
421 iph->frag_off = htons(IP_DF);
424 iph->ttl = ip_select_ttl(inet, &rt->dst);
425 iph->protocol = sk->sk_protocol;
426 ip_copy_addrs(iph, fl4);
428 /* Transport layer set skb->h.foo itself. */
430 if (inet_opt && inet_opt->opt.optlen) {
431 iph->ihl += inet_opt->opt.optlen >> 2;
432 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
435 ip_select_ident_segs(sock_net(sk), skb, sk,
436 skb_shinfo(skb)->gso_segs ?: 1);
438 /* TODO : should we use skb->sk here instead of sk ? */
439 skb->priority = sk->sk_priority;
440 skb->mark = sk->sk_mark;
442 res = ip_local_out_sk(sk, skb);
448 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
450 return -EHOSTUNREACH;
452 EXPORT_SYMBOL(ip_queue_xmit);
455 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
457 to->pkt_type = from->pkt_type;
458 to->priority = from->priority;
459 to->protocol = from->protocol;
461 skb_dst_copy(to, from);
463 to->mark = from->mark;
465 /* Copy the flags to each fragment. */
466 IPCB(to)->flags = IPCB(from)->flags;
468 #ifdef CONFIG_NET_SCHED
469 to->tc_index = from->tc_index;
472 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
473 to->ipvs_property = from->ipvs_property;
475 skb_copy_secmark(to, from);
479 * This IP datagram is too large to be sent in one piece. Break it up into
480 * smaller pieces (each of size equal to IP header plus
481 * a block of the data of the original IP data part) that will yet fit in a
482 * single device frame, and queue such a frame for sending.
485 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
489 struct net_device *dev;
490 struct sk_buff *skb2;
491 unsigned int mtu, hlen, left, len, ll_rs;
493 __be16 not_last_frag;
494 struct rtable *rt = skb_rtable(skb);
500 * Point into the IP datagram header.
505 mtu = ip_skb_dst_mtu(skb);
506 if (unlikely(((iph->frag_off & htons(IP_DF)) && !skb->ignore_df) ||
507 (IPCB(skb)->frag_max_size &&
508 IPCB(skb)->frag_max_size > mtu))) {
509 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
510 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
517 * Setup starting values.
521 mtu = mtu - hlen; /* Size of data space */
522 #ifdef CONFIG_BRIDGE_NETFILTER
524 mtu -= nf_bridge_mtu_reduction(skb);
526 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
528 /* When frag_list is given, use it. First, check its validity:
529 * some transformers could create wrong frag_list or break existing
530 * one, it is not prohibited. In this case fall back to copying.
532 * LATER: this step can be merged to real generation of fragments,
533 * we can switch to copy when see the first bad fragment.
535 if (skb_has_frag_list(skb)) {
536 struct sk_buff *frag, *frag2;
537 int first_len = skb_pagelen(skb);
539 if (first_len - hlen > mtu ||
540 ((first_len - hlen) & 7) ||
541 ip_is_fragment(iph) ||
545 skb_walk_frags(skb, frag) {
546 /* Correct geometry. */
547 if (frag->len > mtu ||
548 ((frag->len & 7) && frag->next) ||
549 skb_headroom(frag) < hlen)
550 goto slow_path_clean;
552 /* Partially cloned skb? */
553 if (skb_shared(frag))
554 goto slow_path_clean;
559 frag->destructor = sock_wfree;
561 skb->truesize -= frag->truesize;
564 /* Everything is OK. Generate! */
568 frag = skb_shinfo(skb)->frag_list;
569 skb_frag_list_init(skb);
570 skb->data_len = first_len - skb_headlen(skb);
571 skb->len = first_len;
572 iph->tot_len = htons(first_len);
573 iph->frag_off = htons(IP_MF);
577 /* Prepare header of the next frame,
578 * before previous one went down. */
580 frag->ip_summed = CHECKSUM_NONE;
581 skb_reset_transport_header(frag);
582 __skb_push(frag, hlen);
583 skb_reset_network_header(frag);
584 memcpy(skb_network_header(frag), iph, hlen);
586 iph->tot_len = htons(frag->len);
587 ip_copy_metadata(frag, skb);
589 ip_options_fragment(frag);
590 offset += skb->len - hlen;
591 iph->frag_off = htons(offset>>3);
592 if (frag->next != NULL)
593 iph->frag_off |= htons(IP_MF);
594 /* Ready, complete checksum */
601 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
611 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
620 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
624 skb_walk_frags(skb, frag2) {
628 frag2->destructor = NULL;
629 skb->truesize += frag2->truesize;
634 /* for offloaded checksums cleanup checksum before fragmentation */
635 if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
639 left = skb->len - hlen; /* Space per frame */
640 ptr = hlen; /* Where to start from */
642 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
643 * we need to make room for the encapsulating header
645 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
648 * Fragment the datagram.
651 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
652 not_last_frag = iph->frag_off & htons(IP_MF);
655 * Keep copying data until we run out.
660 /* IF: it doesn't fit, use 'mtu' - the data space left */
663 /* IF: we are not sending up to and including the packet end
664 then align the next start on an eight byte boundary */
672 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
673 NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
679 * Set up data on packet
682 ip_copy_metadata(skb2, skb);
683 skb_reserve(skb2, ll_rs);
684 skb_put(skb2, len + hlen);
685 skb_reset_network_header(skb2);
686 skb2->transport_header = skb2->network_header + hlen;
689 * Charge the memory for the fragment to any owner
694 skb_set_owner_w(skb2, skb->sk);
697 * Copy the packet header into the new buffer.
700 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
703 * Copy a block of the IP datagram.
705 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
710 * Fill in the new header fields.
713 iph->frag_off = htons((offset >> 3));
715 /* ANK: dirty, but effective trick. Upgrade options only if
716 * the segment to be fragmented was THE FIRST (otherwise,
717 * options are already fixed) and make it ONCE
718 * on the initial skb, so that all the following fragments
719 * will inherit fixed options.
722 ip_options_fragment(skb);
725 * Added AC : If we are fragmenting a fragment that's not the
726 * last fragment then keep MF on each bit
728 if (left > 0 || not_last_frag)
729 iph->frag_off |= htons(IP_MF);
734 * Put this fragment into the sending queue.
736 iph->tot_len = htons(len + hlen);
744 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
747 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
752 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
755 EXPORT_SYMBOL(ip_fragment);
758 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
760 struct iovec *iov = from;
762 if (skb->ip_summed == CHECKSUM_PARTIAL) {
763 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
767 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
769 skb->csum = csum_block_add(skb->csum, csum, odd);
773 EXPORT_SYMBOL(ip_generic_getfrag);
776 csum_page(struct page *page, int offset, int copy)
781 csum = csum_partial(kaddr + offset, copy, 0);
786 static inline int ip_ufo_append_data(struct sock *sk,
787 struct sk_buff_head *queue,
788 int getfrag(void *from, char *to, int offset, int len,
789 int odd, struct sk_buff *skb),
790 void *from, int length, int hh_len, int fragheaderlen,
791 int transhdrlen, int maxfraglen, unsigned int flags)
796 /* There is support for UDP fragmentation offload by network
797 * device, so create one single skb packet containing complete
800 if ((skb = skb_peek_tail(queue)) == NULL) {
801 skb = sock_alloc_send_skb(sk,
802 hh_len + fragheaderlen + transhdrlen + 20,
803 (flags & MSG_DONTWAIT), &err);
808 /* reserve space for Hardware header */
809 skb_reserve(skb, hh_len);
811 /* create space for UDP/IP header */
812 skb_put(skb, fragheaderlen + transhdrlen);
814 /* initialize network header pointer */
815 skb_reset_network_header(skb);
817 /* initialize protocol header pointer */
818 skb->transport_header = skb->network_header + fragheaderlen;
823 __skb_queue_tail(queue, skb);
824 } else if (skb_is_gso(skb)) {
828 skb->ip_summed = CHECKSUM_PARTIAL;
829 /* specify the length of each IP datagram fragment */
830 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
831 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
834 return skb_append_datato_frags(sk, skb, getfrag, from,
835 (length - transhdrlen));
838 static int __ip_append_data(struct sock *sk,
840 struct sk_buff_head *queue,
841 struct inet_cork *cork,
842 struct page_frag *pfrag,
843 int getfrag(void *from, char *to, int offset,
844 int len, int odd, struct sk_buff *skb),
845 void *from, int length, int transhdrlen,
848 struct inet_sock *inet = inet_sk(sk);
851 struct ip_options *opt = cork->opt;
858 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
859 int csummode = CHECKSUM_NONE;
860 struct rtable *rt = (struct rtable *)cork->dst;
862 skb = skb_peek_tail(queue);
864 exthdrlen = !skb ? rt->dst.header_len : 0;
865 mtu = cork->fragsize;
867 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
869 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
870 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
871 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
873 if (cork->length + length > maxnonfragsize - fragheaderlen) {
874 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
875 mtu - (opt ? opt->optlen : 0));
880 * transhdrlen > 0 means that this is the first fragment and we wish
881 * it won't be fragmented in the future.
884 length + fragheaderlen <= mtu &&
885 rt->dst.dev->features & NETIF_F_V4_CSUM &&
887 csummode = CHECKSUM_PARTIAL;
889 cork->length += length;
890 if ((skb && skb_is_gso(skb)) ||
891 (((length + (skb ? skb->len : fragheaderlen)) > mtu) &&
892 (skb_queue_len(queue) <= 1) &&
893 (sk->sk_protocol == IPPROTO_UDP) &&
894 (rt->dst.dev->features & NETIF_F_UFO) && !dst_xfrm(&rt->dst) &&
895 (sk->sk_type == SOCK_DGRAM))) {
896 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
897 hh_len, fragheaderlen, transhdrlen,
904 /* So, what's going on in the loop below?
906 * We use calculated fragment length to generate chained skb,
907 * each of segments is IP fragment ready for sending to network after
908 * adding appropriate IP header.
915 /* Check if the remaining data fits into current packet. */
916 copy = mtu - skb->len;
918 copy = maxfraglen - skb->len;
921 unsigned int datalen;
922 unsigned int fraglen;
923 unsigned int fraggap;
924 unsigned int alloclen;
925 struct sk_buff *skb_prev;
929 fraggap = skb_prev->len - maxfraglen;
934 * If remaining data exceeds the mtu,
935 * we know we need more fragment(s).
937 datalen = length + fraggap;
938 if (datalen > mtu - fragheaderlen)
939 datalen = maxfraglen - fragheaderlen;
940 fraglen = datalen + fragheaderlen;
942 if ((flags & MSG_MORE) &&
943 !(rt->dst.dev->features&NETIF_F_SG))
948 alloclen += exthdrlen;
950 /* The last fragment gets additional space at tail.
951 * Note, with MSG_MORE we overallocate on fragments,
952 * because we have no idea what fragment will be
955 if (datalen == length + fraggap)
956 alloclen += rt->dst.trailer_len;
959 skb = sock_alloc_send_skb(sk,
960 alloclen + hh_len + 15,
961 (flags & MSG_DONTWAIT), &err);
964 if (atomic_read(&sk->sk_wmem_alloc) <=
966 skb = sock_wmalloc(sk,
967 alloclen + hh_len + 15, 1,
969 if (unlikely(skb == NULL))
972 /* only the initial fragment is
980 * Fill in the control structures
982 skb->ip_summed = csummode;
984 skb_reserve(skb, hh_len);
985 skb_shinfo(skb)->tx_flags = cork->tx_flags;
988 * Find where to start putting bytes.
990 data = skb_put(skb, fraglen + exthdrlen);
991 skb_set_network_header(skb, exthdrlen);
992 skb->transport_header = (skb->network_header +
994 data += fragheaderlen + exthdrlen;
997 skb->csum = skb_copy_and_csum_bits(
998 skb_prev, maxfraglen,
999 data + transhdrlen, fraggap, 0);
1000 skb_prev->csum = csum_sub(skb_prev->csum,
1003 pskb_trim_unique(skb_prev, maxfraglen);
1006 copy = datalen - transhdrlen - fraggap;
1007 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1014 length -= datalen - fraggap;
1017 csummode = CHECKSUM_NONE;
1020 * Put the packet on the pending queue.
1022 __skb_queue_tail(queue, skb);
1029 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1030 skb_tailroom(skb) >= copy) {
1034 if (getfrag(from, skb_put(skb, copy),
1035 offset, copy, off, skb) < 0) {
1036 __skb_trim(skb, off);
1041 int i = skb_shinfo(skb)->nr_frags;
1044 if (!sk_page_frag_refill(sk, pfrag))
1047 if (!skb_can_coalesce(skb, i, pfrag->page,
1050 if (i == MAX_SKB_FRAGS)
1053 __skb_fill_page_desc(skb, i, pfrag->page,
1055 skb_shinfo(skb)->nr_frags = ++i;
1056 get_page(pfrag->page);
1058 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1060 page_address(pfrag->page) + pfrag->offset,
1061 offset, copy, skb->len, skb) < 0)
1064 pfrag->offset += copy;
1065 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1067 skb->data_len += copy;
1068 skb->truesize += copy;
1069 atomic_add(copy, &sk->sk_wmem_alloc);
1080 cork->length -= length;
1081 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1085 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1086 struct ipcm_cookie *ipc, struct rtable **rtp)
1088 struct ip_options_rcu *opt;
1092 * setup for corking.
1096 if (cork->opt == NULL) {
1097 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1099 if (unlikely(cork->opt == NULL))
1102 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1103 cork->flags |= IPCORK_OPT;
1104 cork->addr = ipc->addr;
1110 cork->fragsize = ip_sk_use_pmtu(sk) ?
1111 dst_mtu(&rt->dst) : ACCESS_ONCE(rt->dst.dev->mtu);
1113 if (!inetdev_valid_mtu(cork->fragsize))
1114 return -ENETUNREACH;
1116 cork->dst = &rt->dst;
1117 /* We stole this route, caller should not release it. */
1121 cork->ttl = ipc->ttl;
1122 cork->tos = ipc->tos;
1123 cork->priority = ipc->priority;
1124 cork->tx_flags = ipc->tx_flags;
1130 * ip_append_data() and ip_append_page() can make one large IP datagram
1131 * from many pieces of data. Each pieces will be holded on the socket
1132 * until ip_push_pending_frames() is called. Each piece can be a page
1135 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1136 * this interface potentially.
1138 * LATER: length must be adjusted by pad at tail, when it is required.
1140 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1141 int getfrag(void *from, char *to, int offset, int len,
1142 int odd, struct sk_buff *skb),
1143 void *from, int length, int transhdrlen,
1144 struct ipcm_cookie *ipc, struct rtable **rtp,
1147 struct inet_sock *inet = inet_sk(sk);
1150 if (flags&MSG_PROBE)
1153 if (skb_queue_empty(&sk->sk_write_queue)) {
1154 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1161 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1162 sk_page_frag(sk), getfrag,
1163 from, length, transhdrlen, flags);
1166 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1167 int offset, size_t size, int flags)
1169 struct inet_sock *inet = inet_sk(sk);
1170 struct sk_buff *skb;
1172 struct ip_options *opt = NULL;
1173 struct inet_cork *cork;
1178 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1183 if (flags&MSG_PROBE)
1186 if (skb_queue_empty(&sk->sk_write_queue))
1189 cork = &inet->cork.base;
1190 rt = (struct rtable *)cork->dst;
1191 if (cork->flags & IPCORK_OPT)
1194 if (!(rt->dst.dev->features&NETIF_F_SG))
1197 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1198 mtu = cork->fragsize;
1200 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1201 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1202 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1204 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1205 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1206 mtu - (opt ? opt->optlen : 0));
1210 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1213 cork->length += size;
1214 if ((size + skb->len > mtu) &&
1215 (skb_queue_len(&sk->sk_write_queue) == 1) &&
1216 (sk->sk_protocol == IPPROTO_UDP) &&
1217 (rt->dst.dev->features & NETIF_F_UFO)) {
1218 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1219 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1226 if (skb_is_gso(skb))
1230 /* Check if the remaining data fits into current packet. */
1231 len = mtu - skb->len;
1233 len = maxfraglen - skb->len;
1236 struct sk_buff *skb_prev;
1240 fraggap = skb_prev->len - maxfraglen;
1242 alloclen = fragheaderlen + hh_len + fraggap + 15;
1243 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1244 if (unlikely(!skb)) {
1250 * Fill in the control structures
1252 skb->ip_summed = CHECKSUM_NONE;
1254 skb_reserve(skb, hh_len);
1257 * Find where to start putting bytes.
1259 skb_put(skb, fragheaderlen + fraggap);
1260 skb_reset_network_header(skb);
1261 skb->transport_header = (skb->network_header +
1264 skb->csum = skb_copy_and_csum_bits(skb_prev,
1266 skb_transport_header(skb),
1268 skb_prev->csum = csum_sub(skb_prev->csum,
1270 pskb_trim_unique(skb_prev, maxfraglen);
1274 * Put the packet on the pending queue.
1276 __skb_queue_tail(&sk->sk_write_queue, skb);
1280 i = skb_shinfo(skb)->nr_frags;
1283 if (skb_can_coalesce(skb, i, page, offset)) {
1284 skb_frag_size_add(&skb_shinfo(skb)->frags[i-1], len);
1285 } else if (i < MAX_SKB_FRAGS) {
1287 skb_fill_page_desc(skb, i, page, offset, len);
1293 if (skb->ip_summed == CHECKSUM_NONE) {
1295 csum = csum_page(page, offset, len);
1296 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1300 skb->data_len += len;
1301 skb->truesize += len;
1302 atomic_add(len, &sk->sk_wmem_alloc);
1309 cork->length -= size;
1310 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1314 static void ip_cork_release(struct inet_cork *cork)
1316 cork->flags &= ~IPCORK_OPT;
1319 dst_release(cork->dst);
1324 * Combined all pending IP fragments on the socket as one IP datagram
1325 * and push them out.
1327 struct sk_buff *__ip_make_skb(struct sock *sk,
1329 struct sk_buff_head *queue,
1330 struct inet_cork *cork)
1332 struct sk_buff *skb, *tmp_skb;
1333 struct sk_buff **tail_skb;
1334 struct inet_sock *inet = inet_sk(sk);
1335 struct net *net = sock_net(sk);
1336 struct ip_options *opt = NULL;
1337 struct rtable *rt = (struct rtable *)cork->dst;
1342 if ((skb = __skb_dequeue(queue)) == NULL)
1344 tail_skb = &(skb_shinfo(skb)->frag_list);
1346 /* move skb->data to ip header from ext header */
1347 if (skb->data < skb_network_header(skb))
1348 __skb_pull(skb, skb_network_offset(skb));
1349 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1350 __skb_pull(tmp_skb, skb_network_header_len(skb));
1351 *tail_skb = tmp_skb;
1352 tail_skb = &(tmp_skb->next);
1353 skb->len += tmp_skb->len;
1354 skb->data_len += tmp_skb->len;
1355 skb->truesize += tmp_skb->truesize;
1356 tmp_skb->destructor = NULL;
1360 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1361 * to fragment the frame generated here. No matter, what transforms
1362 * how transforms change size of the packet, it will come out.
1364 skb->ignore_df = ip_sk_ignore_df(sk);
1366 /* DF bit is set when we want to see DF on outgoing frames.
1367 * If ignore_df is set too, we still allow to fragment this frame
1369 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1370 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1371 (skb->len <= dst_mtu(&rt->dst) &&
1372 ip_dont_fragment(sk, &rt->dst)))
1375 if (cork->flags & IPCORK_OPT)
1380 else if (rt->rt_type == RTN_MULTICAST)
1383 ttl = ip_select_ttl(inet, &rt->dst);
1388 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1391 iph->protocol = sk->sk_protocol;
1392 ip_copy_addrs(iph, fl4);
1393 ip_select_ident(net, skb, sk);
1396 iph->ihl += opt->optlen>>2;
1397 ip_options_build(skb, opt, cork->addr, rt, 0);
1400 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1401 skb->mark = sk->sk_mark;
1403 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1407 skb_dst_set(skb, &rt->dst);
1409 if (iph->protocol == IPPROTO_ICMP)
1410 icmp_out_count(net, ((struct icmphdr *)
1411 skb_transport_header(skb))->type);
1413 ip_cork_release(cork);
1418 int ip_send_skb(struct net *net, struct sk_buff *skb)
1422 err = ip_local_out(skb);
1425 err = net_xmit_errno(err);
1427 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1433 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1435 struct sk_buff *skb;
1437 skb = ip_finish_skb(sk, fl4);
1441 /* Netfilter gets whole the not fragmented skb. */
1442 return ip_send_skb(sock_net(sk), skb);
1446 * Throw away all pending data on the socket.
1448 static void __ip_flush_pending_frames(struct sock *sk,
1449 struct sk_buff_head *queue,
1450 struct inet_cork *cork)
1452 struct sk_buff *skb;
1454 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1457 ip_cork_release(cork);
1460 void ip_flush_pending_frames(struct sock *sk)
1462 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1465 struct sk_buff *ip_make_skb(struct sock *sk,
1467 int getfrag(void *from, char *to, int offset,
1468 int len, int odd, struct sk_buff *skb),
1469 void *from, int length, int transhdrlen,
1470 struct ipcm_cookie *ipc, struct rtable **rtp,
1473 struct inet_cork cork;
1474 struct sk_buff_head queue;
1477 if (flags & MSG_PROBE)
1480 __skb_queue_head_init(&queue);
1485 err = ip_setup_cork(sk, &cork, ipc, rtp);
1487 return ERR_PTR(err);
1489 err = __ip_append_data(sk, fl4, &queue, &cork,
1490 ¤t->task_frag, getfrag,
1491 from, length, transhdrlen, flags);
1493 __ip_flush_pending_frames(sk, &queue, &cork);
1494 return ERR_PTR(err);
1497 return __ip_make_skb(sk, fl4, &queue, &cork);
1501 * Fetch data from kernel space and fill in checksum if needed.
1503 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1504 int len, int odd, struct sk_buff *skb)
1508 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1509 skb->csum = csum_block_add(skb->csum, csum, odd);
1514 * Generic function to send a packet as reply to another packet.
1515 * Used to send some TCP resets/acks so far.
1517 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb, __be32 daddr,
1518 __be32 saddr, const struct ip_reply_arg *arg,
1521 struct ip_options_data replyopts;
1522 struct ipcm_cookie ipc;
1524 struct rtable *rt = skb_rtable(skb);
1525 struct net *net = sock_net(sk);
1526 struct sk_buff *nskb;
1529 if (ip_options_echo(&replyopts.opt.opt, skb))
1538 if (replyopts.opt.opt.optlen) {
1539 ipc.opt = &replyopts.opt;
1541 if (replyopts.opt.opt.srr)
1542 daddr = replyopts.opt.opt.faddr;
1545 flowi4_init_output(&fl4, arg->bound_dev_if,
1546 IP4_REPLY_MARK(net, skb->mark),
1548 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1549 ip_reply_arg_flowi_flags(arg),
1551 tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1552 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1553 rt = ip_route_output_key(net, &fl4);
1557 inet_sk(sk)->tos = arg->tos;
1559 sk->sk_priority = skb->priority;
1560 sk->sk_protocol = ip_hdr(skb)->protocol;
1561 sk->sk_bound_dev_if = arg->bound_dev_if;
1562 sk->sk_sndbuf = sysctl_wmem_default;
1563 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1564 len, 0, &ipc, &rt, MSG_DONTWAIT);
1565 if (unlikely(err)) {
1566 ip_flush_pending_frames(sk);
1570 nskb = skb_peek(&sk->sk_write_queue);
1572 if (arg->csumoffset >= 0)
1573 *((__sum16 *)skb_transport_header(nskb) +
1574 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1576 nskb->ip_summed = CHECKSUM_NONE;
1577 skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1578 ip_push_pending_frames(sk, &fl4);
1584 void __init ip_init(void)
1589 #if defined(CONFIG_IP_MULTICAST)