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 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT, skb, NULL,
101 skb_dst(skb)->dev, dst_output);
104 int ip_local_out_sk(struct sock *sk, struct sk_buff *skb)
108 err = __ip_local_out(skb);
109 if (likely(err == 1))
110 err = dst_output_sk(sk, skb);
114 EXPORT_SYMBOL_GPL(ip_local_out_sk);
116 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
118 int ttl = inet->uc_ttl;
121 ttl = ip4_dst_hoplimit(dst);
126 * Add an ip header to a skbuff and send it out.
129 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
130 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
132 struct inet_sock *inet = inet_sk(sk);
133 struct rtable *rt = skb_rtable(skb);
136 /* Build the IP header. */
137 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
138 skb_reset_network_header(skb);
142 iph->tos = inet->tos;
143 if (ip_dont_fragment(sk, &rt->dst))
144 iph->frag_off = htons(IP_DF);
147 iph->ttl = ip_select_ttl(inet, &rt->dst);
148 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
150 iph->protocol = sk->sk_protocol;
151 ip_select_ident(skb, sk);
153 if (opt && opt->opt.optlen) {
154 iph->ihl += opt->opt.optlen>>2;
155 ip_options_build(skb, &opt->opt, daddr, rt, 0);
158 skb->priority = sk->sk_priority;
159 skb->mark = sk->sk_mark;
162 return ip_local_out(skb);
164 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
166 static inline int ip_finish_output2(struct sk_buff *skb)
168 struct dst_entry *dst = skb_dst(skb);
169 struct rtable *rt = (struct rtable *)dst;
170 struct net_device *dev = dst->dev;
171 unsigned int hh_len = LL_RESERVED_SPACE(dev);
172 struct neighbour *neigh;
175 if (rt->rt_type == RTN_MULTICAST) {
176 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
177 } else if (rt->rt_type == RTN_BROADCAST)
178 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
180 /* Be paranoid, rather than too clever. */
181 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
182 struct sk_buff *skb2;
184 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
190 skb_set_owner_w(skb2, skb->sk);
196 nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
197 neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
198 if (unlikely(!neigh))
199 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
200 if (!IS_ERR(neigh)) {
201 int res = dst_neigh_output(dst, neigh, skb);
203 rcu_read_unlock_bh();
206 rcu_read_unlock_bh();
208 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
214 static int ip_finish_output_gso(struct sk_buff *skb)
216 netdev_features_t features;
217 struct sk_buff *segs;
220 /* common case: locally created skb or seglen is <= mtu */
221 if (((IPCB(skb)->flags & IPSKB_FORWARDED) == 0) ||
222 skb_gso_network_seglen(skb) <= ip_skb_dst_mtu(skb))
223 return ip_finish_output2(skb);
225 /* Slowpath - GSO segment length is exceeding the dst MTU.
227 * This can happen in two cases:
228 * 1) TCP GRO packet, DF bit not set
229 * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
230 * from host network stack.
232 features = netif_skb_features(skb);
233 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
242 struct sk_buff *nskb = segs->next;
246 err = ip_fragment(segs, ip_finish_output2);
256 static int ip_finish_output(struct sk_buff *skb)
258 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
259 /* Policy lookup after SNAT yielded a new policy */
260 if (skb_dst(skb)->xfrm != NULL) {
261 IPCB(skb)->flags |= IPSKB_REROUTED;
262 return dst_output(skb);
266 return ip_finish_output_gso(skb);
268 if (skb->len > ip_skb_dst_mtu(skb))
269 return ip_fragment(skb, ip_finish_output2);
271 return ip_finish_output2(skb);
274 int ip_mc_output(struct sock *sk, struct sk_buff *skb)
276 struct rtable *rt = skb_rtable(skb);
277 struct net_device *dev = rt->dst.dev;
280 * If the indicated interface is up and running, send the packet.
282 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
285 skb->protocol = htons(ETH_P_IP);
288 * Multicasts are looped back for other local users
291 if (rt->rt_flags&RTCF_MULTICAST) {
293 #ifdef CONFIG_IP_MROUTE
294 /* Small optimization: do not loopback not local frames,
295 which returned after forwarding; they will be dropped
296 by ip_mr_input in any case.
297 Note, that local frames are looped back to be delivered
300 This check is duplicated in ip_mr_input at the moment.
303 ((rt->rt_flags & RTCF_LOCAL) ||
304 !(IPCB(skb)->flags & IPSKB_FORWARDED))
307 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
309 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
310 newskb, NULL, newskb->dev,
314 /* Multicasts with ttl 0 must not go beyond the host */
316 if (ip_hdr(skb)->ttl == 0) {
322 if (rt->rt_flags&RTCF_BROADCAST) {
323 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
325 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
326 NULL, newskb->dev, dev_loopback_xmit);
329 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
330 skb->dev, ip_finish_output,
331 !(IPCB(skb)->flags & IPSKB_REROUTED));
334 int ip_output(struct sock *sk, struct sk_buff *skb)
336 struct net_device *dev = skb_dst(skb)->dev;
338 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
341 skb->protocol = htons(ETH_P_IP);
343 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
345 !(IPCB(skb)->flags & IPSKB_REROUTED));
349 * copy saddr and daddr, possibly using 64bit load/stores
351 * iph->saddr = fl4->saddr;
352 * iph->daddr = fl4->daddr;
354 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
356 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
357 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
358 memcpy(&iph->saddr, &fl4->saddr,
359 sizeof(fl4->saddr) + sizeof(fl4->daddr));
362 /* Note: skb->sk can be different from sk, in case of tunnels */
363 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
365 struct inet_sock *inet = inet_sk(sk);
366 struct ip_options_rcu *inet_opt;
372 /* Skip all of this if the packet is already routed,
373 * f.e. by something like SCTP.
376 inet_opt = rcu_dereference(inet->inet_opt);
378 rt = skb_rtable(skb);
382 /* Make sure we can route this packet. */
383 rt = (struct rtable *)__sk_dst_check(sk, 0);
387 /* Use correct destination address if we have options. */
388 daddr = inet->inet_daddr;
389 if (inet_opt && inet_opt->opt.srr)
390 daddr = inet_opt->opt.faddr;
392 /* If this fails, retransmit mechanism of transport layer will
393 * keep trying until route appears or the connection times
396 rt = ip_route_output_ports(sock_net(sk), fl4, sk,
397 daddr, inet->inet_saddr,
402 sk->sk_bound_dev_if);
405 sk_setup_caps(sk, &rt->dst);
407 skb_dst_set_noref(skb, &rt->dst);
410 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
413 /* OK, we know where to send it, allocate and build IP header. */
414 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
415 skb_reset_network_header(skb);
417 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
418 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
419 iph->frag_off = htons(IP_DF);
422 iph->ttl = ip_select_ttl(inet, &rt->dst);
423 iph->protocol = sk->sk_protocol;
424 ip_copy_addrs(iph, fl4);
426 /* Transport layer set skb->h.foo itself. */
428 if (inet_opt && inet_opt->opt.optlen) {
429 iph->ihl += inet_opt->opt.optlen >> 2;
430 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
433 ip_select_ident_segs(skb, sk, skb_shinfo(skb)->gso_segs ?: 1);
435 /* TODO : should we use skb->sk here instead of sk ? */
436 skb->priority = sk->sk_priority;
437 skb->mark = sk->sk_mark;
439 res = ip_local_out(skb);
445 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
447 return -EHOSTUNREACH;
449 EXPORT_SYMBOL(ip_queue_xmit);
452 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
454 to->pkt_type = from->pkt_type;
455 to->priority = from->priority;
456 to->protocol = from->protocol;
458 skb_dst_copy(to, from);
460 to->mark = from->mark;
462 /* Copy the flags to each fragment. */
463 IPCB(to)->flags = IPCB(from)->flags;
465 #ifdef CONFIG_NET_SCHED
466 to->tc_index = from->tc_index;
469 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
470 to->ipvs_property = from->ipvs_property;
472 skb_copy_secmark(to, from);
476 * This IP datagram is too large to be sent in one piece. Break it up into
477 * smaller pieces (each of size equal to IP header plus
478 * a block of the data of the original IP data part) that will yet fit in a
479 * single device frame, and queue such a frame for sending.
482 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
486 struct net_device *dev;
487 struct sk_buff *skb2;
488 unsigned int mtu, hlen, left, len, ll_rs;
490 __be16 not_last_frag;
491 struct rtable *rt = skb_rtable(skb);
497 * Point into the IP datagram header.
502 mtu = ip_skb_dst_mtu(skb);
503 if (unlikely(((iph->frag_off & htons(IP_DF)) && !skb->ignore_df) ||
504 (IPCB(skb)->frag_max_size &&
505 IPCB(skb)->frag_max_size > mtu))) {
506 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
507 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
514 * Setup starting values.
518 mtu = mtu - hlen; /* Size of data space */
519 #ifdef CONFIG_BRIDGE_NETFILTER
521 mtu -= nf_bridge_mtu_reduction(skb);
523 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
525 /* When frag_list is given, use it. First, check its validity:
526 * some transformers could create wrong frag_list or break existing
527 * one, it is not prohibited. In this case fall back to copying.
529 * LATER: this step can be merged to real generation of fragments,
530 * we can switch to copy when see the first bad fragment.
532 if (skb_has_frag_list(skb)) {
533 struct sk_buff *frag, *frag2;
534 int first_len = skb_pagelen(skb);
536 if (first_len - hlen > mtu ||
537 ((first_len - hlen) & 7) ||
538 ip_is_fragment(iph) ||
542 skb_walk_frags(skb, frag) {
543 /* Correct geometry. */
544 if (frag->len > mtu ||
545 ((frag->len & 7) && frag->next) ||
546 skb_headroom(frag) < hlen)
547 goto slow_path_clean;
549 /* Partially cloned skb? */
550 if (skb_shared(frag))
551 goto slow_path_clean;
556 frag->destructor = sock_wfree;
558 skb->truesize -= frag->truesize;
561 /* Everything is OK. Generate! */
565 frag = skb_shinfo(skb)->frag_list;
566 skb_frag_list_init(skb);
567 skb->data_len = first_len - skb_headlen(skb);
568 skb->len = first_len;
569 iph->tot_len = htons(first_len);
570 iph->frag_off = htons(IP_MF);
574 /* Prepare header of the next frame,
575 * before previous one went down. */
577 frag->ip_summed = CHECKSUM_NONE;
578 skb_reset_transport_header(frag);
579 __skb_push(frag, hlen);
580 skb_reset_network_header(frag);
581 memcpy(skb_network_header(frag), iph, hlen);
583 iph->tot_len = htons(frag->len);
584 ip_copy_metadata(frag, skb);
586 ip_options_fragment(frag);
587 offset += skb->len - hlen;
588 iph->frag_off = htons(offset>>3);
589 if (frag->next != NULL)
590 iph->frag_off |= htons(IP_MF);
591 /* Ready, complete checksum */
598 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
608 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
617 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
621 skb_walk_frags(skb, frag2) {
625 frag2->destructor = NULL;
626 skb->truesize += frag2->truesize;
631 /* for offloaded checksums cleanup checksum before fragmentation */
632 if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
636 left = skb->len - hlen; /* Space per frame */
637 ptr = hlen; /* Where to start from */
639 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
640 * we need to make room for the encapsulating header
642 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
645 * Fragment the datagram.
648 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
649 not_last_frag = iph->frag_off & htons(IP_MF);
652 * Keep copying data until we run out.
657 /* IF: it doesn't fit, use 'mtu' - the data space left */
660 /* IF: we are not sending up to and including the packet end
661 then align the next start on an eight byte boundary */
669 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
670 NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
676 * Set up data on packet
679 ip_copy_metadata(skb2, skb);
680 skb_reserve(skb2, ll_rs);
681 skb_put(skb2, len + hlen);
682 skb_reset_network_header(skb2);
683 skb2->transport_header = skb2->network_header + hlen;
686 * Charge the memory for the fragment to any owner
691 skb_set_owner_w(skb2, skb->sk);
694 * Copy the packet header into the new buffer.
697 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
700 * Copy a block of the IP datagram.
702 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
707 * Fill in the new header fields.
710 iph->frag_off = htons((offset >> 3));
712 /* ANK: dirty, but effective trick. Upgrade options only if
713 * the segment to be fragmented was THE FIRST (otherwise,
714 * options are already fixed) and make it ONCE
715 * on the initial skb, so that all the following fragments
716 * will inherit fixed options.
719 ip_options_fragment(skb);
722 * Added AC : If we are fragmenting a fragment that's not the
723 * last fragment then keep MF on each bit
725 if (left > 0 || not_last_frag)
726 iph->frag_off |= htons(IP_MF);
731 * Put this fragment into the sending queue.
733 iph->tot_len = htons(len + hlen);
741 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
744 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
749 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
752 EXPORT_SYMBOL(ip_fragment);
755 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
757 struct iovec *iov = from;
759 if (skb->ip_summed == CHECKSUM_PARTIAL) {
760 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
764 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
766 skb->csum = csum_block_add(skb->csum, csum, odd);
770 EXPORT_SYMBOL(ip_generic_getfrag);
773 csum_page(struct page *page, int offset, int copy)
778 csum = csum_partial(kaddr + offset, copy, 0);
783 static inline int ip_ufo_append_data(struct sock *sk,
784 struct sk_buff_head *queue,
785 int getfrag(void *from, char *to, int offset, int len,
786 int odd, struct sk_buff *skb),
787 void *from, int length, int hh_len, int fragheaderlen,
788 int transhdrlen, int maxfraglen, unsigned int flags)
793 /* There is support for UDP fragmentation offload by network
794 * device, so create one single skb packet containing complete
797 if ((skb = skb_peek_tail(queue)) == NULL) {
798 skb = sock_alloc_send_skb(sk,
799 hh_len + fragheaderlen + transhdrlen + 20,
800 (flags & MSG_DONTWAIT), &err);
805 /* reserve space for Hardware header */
806 skb_reserve(skb, hh_len);
808 /* create space for UDP/IP header */
809 skb_put(skb, fragheaderlen + transhdrlen);
811 /* initialize network header pointer */
812 skb_reset_network_header(skb);
814 /* initialize protocol header pointer */
815 skb->transport_header = skb->network_header + fragheaderlen;
820 __skb_queue_tail(queue, skb);
821 } else if (skb_is_gso(skb)) {
825 skb->ip_summed = CHECKSUM_PARTIAL;
826 /* specify the length of each IP datagram fragment */
827 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
828 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
831 return skb_append_datato_frags(sk, skb, getfrag, from,
832 (length - transhdrlen));
835 static int __ip_append_data(struct sock *sk,
837 struct sk_buff_head *queue,
838 struct inet_cork *cork,
839 struct page_frag *pfrag,
840 int getfrag(void *from, char *to, int offset,
841 int len, int odd, struct sk_buff *skb),
842 void *from, int length, int transhdrlen,
845 struct inet_sock *inet = inet_sk(sk);
848 struct ip_options *opt = cork->opt;
855 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
856 int csummode = CHECKSUM_NONE;
857 struct rtable *rt = (struct rtable *)cork->dst;
859 skb = skb_peek_tail(queue);
861 exthdrlen = !skb ? rt->dst.header_len : 0;
862 mtu = cork->fragsize;
864 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
866 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
867 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
868 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
870 if (cork->length + length > maxnonfragsize - fragheaderlen) {
871 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
872 mtu - (opt ? opt->optlen : 0));
877 * transhdrlen > 0 means that this is the first fragment and we wish
878 * it won't be fragmented in the future.
881 length + fragheaderlen <= mtu &&
882 rt->dst.dev->features & NETIF_F_V4_CSUM &&
884 csummode = CHECKSUM_PARTIAL;
886 cork->length += length;
887 if (((length > mtu) || (skb && skb_is_gso(skb))) &&
888 (sk->sk_protocol == IPPROTO_UDP) &&
889 (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len) {
890 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
891 hh_len, fragheaderlen, transhdrlen,
898 /* So, what's going on in the loop below?
900 * We use calculated fragment length to generate chained skb,
901 * each of segments is IP fragment ready for sending to network after
902 * adding appropriate IP header.
909 /* Check if the remaining data fits into current packet. */
910 copy = mtu - skb->len;
912 copy = maxfraglen - skb->len;
915 unsigned int datalen;
916 unsigned int fraglen;
917 unsigned int fraggap;
918 unsigned int alloclen;
919 struct sk_buff *skb_prev;
923 fraggap = skb_prev->len - maxfraglen;
928 * If remaining data exceeds the mtu,
929 * we know we need more fragment(s).
931 datalen = length + fraggap;
932 if (datalen > mtu - fragheaderlen)
933 datalen = maxfraglen - fragheaderlen;
934 fraglen = datalen + fragheaderlen;
936 if ((flags & MSG_MORE) &&
937 !(rt->dst.dev->features&NETIF_F_SG))
942 alloclen += exthdrlen;
944 /* The last fragment gets additional space at tail.
945 * Note, with MSG_MORE we overallocate on fragments,
946 * because we have no idea what fragment will be
949 if (datalen == length + fraggap)
950 alloclen += rt->dst.trailer_len;
953 skb = sock_alloc_send_skb(sk,
954 alloclen + hh_len + 15,
955 (flags & MSG_DONTWAIT), &err);
958 if (atomic_read(&sk->sk_wmem_alloc) <=
960 skb = sock_wmalloc(sk,
961 alloclen + hh_len + 15, 1,
963 if (unlikely(skb == NULL))
966 /* only the initial fragment is
974 * Fill in the control structures
976 skb->ip_summed = csummode;
978 skb_reserve(skb, hh_len);
979 skb_shinfo(skb)->tx_flags = cork->tx_flags;
982 * Find where to start putting bytes.
984 data = skb_put(skb, fraglen + exthdrlen);
985 skb_set_network_header(skb, exthdrlen);
986 skb->transport_header = (skb->network_header +
988 data += fragheaderlen + exthdrlen;
991 skb->csum = skb_copy_and_csum_bits(
992 skb_prev, maxfraglen,
993 data + transhdrlen, fraggap, 0);
994 skb_prev->csum = csum_sub(skb_prev->csum,
997 pskb_trim_unique(skb_prev, maxfraglen);
1000 copy = datalen - transhdrlen - fraggap;
1001 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1008 length -= datalen - fraggap;
1011 csummode = CHECKSUM_NONE;
1014 * Put the packet on the pending queue.
1016 __skb_queue_tail(queue, skb);
1023 if (!(rt->dst.dev->features&NETIF_F_SG)) {
1027 if (getfrag(from, skb_put(skb, copy),
1028 offset, copy, off, skb) < 0) {
1029 __skb_trim(skb, off);
1034 int i = skb_shinfo(skb)->nr_frags;
1037 if (!sk_page_frag_refill(sk, pfrag))
1040 if (!skb_can_coalesce(skb, i, pfrag->page,
1043 if (i == MAX_SKB_FRAGS)
1046 __skb_fill_page_desc(skb, i, pfrag->page,
1048 skb_shinfo(skb)->nr_frags = ++i;
1049 get_page(pfrag->page);
1051 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1053 page_address(pfrag->page) + pfrag->offset,
1054 offset, copy, skb->len, skb) < 0)
1057 pfrag->offset += copy;
1058 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1060 skb->data_len += copy;
1061 skb->truesize += copy;
1062 atomic_add(copy, &sk->sk_wmem_alloc);
1073 cork->length -= length;
1074 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1078 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1079 struct ipcm_cookie *ipc, struct rtable **rtp)
1081 struct ip_options_rcu *opt;
1085 * setup for corking.
1089 if (cork->opt == NULL) {
1090 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1092 if (unlikely(cork->opt == NULL))
1095 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1096 cork->flags |= IPCORK_OPT;
1097 cork->addr = ipc->addr;
1103 * We steal reference to this route, caller should not release it
1106 cork->fragsize = ip_sk_use_pmtu(sk) ?
1107 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1108 cork->dst = &rt->dst;
1110 cork->ttl = ipc->ttl;
1111 cork->tos = ipc->tos;
1112 cork->priority = ipc->priority;
1113 cork->tx_flags = ipc->tx_flags;
1119 * ip_append_data() and ip_append_page() can make one large IP datagram
1120 * from many pieces of data. Each pieces will be holded on the socket
1121 * until ip_push_pending_frames() is called. Each piece can be a page
1124 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1125 * this interface potentially.
1127 * LATER: length must be adjusted by pad at tail, when it is required.
1129 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1130 int getfrag(void *from, char *to, int offset, int len,
1131 int odd, struct sk_buff *skb),
1132 void *from, int length, int transhdrlen,
1133 struct ipcm_cookie *ipc, struct rtable **rtp,
1136 struct inet_sock *inet = inet_sk(sk);
1139 if (flags&MSG_PROBE)
1142 if (skb_queue_empty(&sk->sk_write_queue)) {
1143 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1150 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1151 sk_page_frag(sk), getfrag,
1152 from, length, transhdrlen, flags);
1155 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1156 int offset, size_t size, int flags)
1158 struct inet_sock *inet = inet_sk(sk);
1159 struct sk_buff *skb;
1161 struct ip_options *opt = NULL;
1162 struct inet_cork *cork;
1167 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1172 if (flags&MSG_PROBE)
1175 if (skb_queue_empty(&sk->sk_write_queue))
1178 cork = &inet->cork.base;
1179 rt = (struct rtable *)cork->dst;
1180 if (cork->flags & IPCORK_OPT)
1183 if (!(rt->dst.dev->features&NETIF_F_SG))
1186 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1187 mtu = cork->fragsize;
1189 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1190 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1191 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1193 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1194 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1195 mtu - (opt ? opt->optlen : 0));
1199 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1202 cork->length += size;
1203 if ((size + skb->len > mtu) &&
1204 (sk->sk_protocol == IPPROTO_UDP) &&
1205 (rt->dst.dev->features & NETIF_F_UFO)) {
1206 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1207 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1214 if (skb_is_gso(skb))
1218 /* Check if the remaining data fits into current packet. */
1219 len = mtu - skb->len;
1221 len = maxfraglen - skb->len;
1224 struct sk_buff *skb_prev;
1228 fraggap = skb_prev->len - maxfraglen;
1230 alloclen = fragheaderlen + hh_len + fraggap + 15;
1231 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1232 if (unlikely(!skb)) {
1238 * Fill in the control structures
1240 skb->ip_summed = CHECKSUM_NONE;
1242 skb_reserve(skb, hh_len);
1245 * Find where to start putting bytes.
1247 skb_put(skb, fragheaderlen + fraggap);
1248 skb_reset_network_header(skb);
1249 skb->transport_header = (skb->network_header +
1252 skb->csum = skb_copy_and_csum_bits(skb_prev,
1254 skb_transport_header(skb),
1256 skb_prev->csum = csum_sub(skb_prev->csum,
1258 pskb_trim_unique(skb_prev, maxfraglen);
1262 * Put the packet on the pending queue.
1264 __skb_queue_tail(&sk->sk_write_queue, skb);
1268 i = skb_shinfo(skb)->nr_frags;
1271 if (skb_can_coalesce(skb, i, page, offset)) {
1272 skb_frag_size_add(&skb_shinfo(skb)->frags[i-1], len);
1273 } else if (i < MAX_SKB_FRAGS) {
1275 skb_fill_page_desc(skb, i, page, offset, len);
1281 if (skb->ip_summed == CHECKSUM_NONE) {
1283 csum = csum_page(page, offset, len);
1284 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1288 skb->data_len += len;
1289 skb->truesize += len;
1290 atomic_add(len, &sk->sk_wmem_alloc);
1297 cork->length -= size;
1298 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1302 static void ip_cork_release(struct inet_cork *cork)
1304 cork->flags &= ~IPCORK_OPT;
1307 dst_release(cork->dst);
1312 * Combined all pending IP fragments on the socket as one IP datagram
1313 * and push them out.
1315 struct sk_buff *__ip_make_skb(struct sock *sk,
1317 struct sk_buff_head *queue,
1318 struct inet_cork *cork)
1320 struct sk_buff *skb, *tmp_skb;
1321 struct sk_buff **tail_skb;
1322 struct inet_sock *inet = inet_sk(sk);
1323 struct net *net = sock_net(sk);
1324 struct ip_options *opt = NULL;
1325 struct rtable *rt = (struct rtable *)cork->dst;
1330 if ((skb = __skb_dequeue(queue)) == NULL)
1332 tail_skb = &(skb_shinfo(skb)->frag_list);
1334 /* move skb->data to ip header from ext header */
1335 if (skb->data < skb_network_header(skb))
1336 __skb_pull(skb, skb_network_offset(skb));
1337 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1338 __skb_pull(tmp_skb, skb_network_header_len(skb));
1339 *tail_skb = tmp_skb;
1340 tail_skb = &(tmp_skb->next);
1341 skb->len += tmp_skb->len;
1342 skb->data_len += tmp_skb->len;
1343 skb->truesize += tmp_skb->truesize;
1344 tmp_skb->destructor = NULL;
1348 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1349 * to fragment the frame generated here. No matter, what transforms
1350 * how transforms change size of the packet, it will come out.
1352 skb->ignore_df = ip_sk_ignore_df(sk);
1354 /* DF bit is set when we want to see DF on outgoing frames.
1355 * If ignore_df is set too, we still allow to fragment this frame
1357 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1358 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1359 (skb->len <= dst_mtu(&rt->dst) &&
1360 ip_dont_fragment(sk, &rt->dst)))
1363 if (cork->flags & IPCORK_OPT)
1368 else if (rt->rt_type == RTN_MULTICAST)
1371 ttl = ip_select_ttl(inet, &rt->dst);
1376 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1379 iph->protocol = sk->sk_protocol;
1380 ip_copy_addrs(iph, fl4);
1381 ip_select_ident(skb, sk);
1384 iph->ihl += opt->optlen>>2;
1385 ip_options_build(skb, opt, cork->addr, rt, 0);
1388 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1389 skb->mark = sk->sk_mark;
1391 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1395 skb_dst_set(skb, &rt->dst);
1397 if (iph->protocol == IPPROTO_ICMP)
1398 icmp_out_count(net, ((struct icmphdr *)
1399 skb_transport_header(skb))->type);
1401 ip_cork_release(cork);
1406 int ip_send_skb(struct net *net, struct sk_buff *skb)
1410 err = ip_local_out(skb);
1413 err = net_xmit_errno(err);
1415 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1421 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1423 struct sk_buff *skb;
1425 skb = ip_finish_skb(sk, fl4);
1429 /* Netfilter gets whole the not fragmented skb. */
1430 return ip_send_skb(sock_net(sk), skb);
1434 * Throw away all pending data on the socket.
1436 static void __ip_flush_pending_frames(struct sock *sk,
1437 struct sk_buff_head *queue,
1438 struct inet_cork *cork)
1440 struct sk_buff *skb;
1442 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1445 ip_cork_release(cork);
1448 void ip_flush_pending_frames(struct sock *sk)
1450 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1453 struct sk_buff *ip_make_skb(struct sock *sk,
1455 int getfrag(void *from, char *to, int offset,
1456 int len, int odd, struct sk_buff *skb),
1457 void *from, int length, int transhdrlen,
1458 struct ipcm_cookie *ipc, struct rtable **rtp,
1461 struct inet_cork cork;
1462 struct sk_buff_head queue;
1465 if (flags & MSG_PROBE)
1468 __skb_queue_head_init(&queue);
1473 err = ip_setup_cork(sk, &cork, ipc, rtp);
1475 return ERR_PTR(err);
1477 err = __ip_append_data(sk, fl4, &queue, &cork,
1478 ¤t->task_frag, getfrag,
1479 from, length, transhdrlen, flags);
1481 __ip_flush_pending_frames(sk, &queue, &cork);
1482 return ERR_PTR(err);
1485 return __ip_make_skb(sk, fl4, &queue, &cork);
1489 * Fetch data from kernel space and fill in checksum if needed.
1491 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1492 int len, int odd, struct sk_buff *skb)
1496 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1497 skb->csum = csum_block_add(skb->csum, csum, odd);
1502 * Generic function to send a packet as reply to another packet.
1503 * Used to send some TCP resets/acks so far.
1505 * Use a fake percpu inet socket to avoid false sharing and contention.
1507 static DEFINE_PER_CPU(struct inet_sock, unicast_sock) = {
1510 .skc_refcnt = ATOMIC_INIT(1),
1512 .sk_wmem_alloc = ATOMIC_INIT(1),
1513 .sk_allocation = GFP_ATOMIC,
1514 .sk_flags = (1UL << SOCK_USE_WRITE_QUEUE),
1516 .pmtudisc = IP_PMTUDISC_WANT,
1520 void ip_send_unicast_reply(struct net *net, struct sk_buff *skb, __be32 daddr,
1521 __be32 saddr, const struct ip_reply_arg *arg,
1524 struct ip_options_data replyopts;
1525 struct ipcm_cookie ipc;
1527 struct rtable *rt = skb_rtable(skb);
1528 struct sk_buff *nskb;
1530 struct inet_sock *inet;
1532 if (ip_options_echo(&replyopts.opt.opt, skb))
1541 if (replyopts.opt.opt.optlen) {
1542 ipc.opt = &replyopts.opt;
1544 if (replyopts.opt.opt.srr)
1545 daddr = replyopts.opt.opt.faddr;
1548 flowi4_init_output(&fl4, arg->bound_dev_if,
1549 IP4_REPLY_MARK(net, skb->mark),
1551 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1552 ip_reply_arg_flowi_flags(arg),
1554 tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1555 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1556 rt = ip_route_output_key(net, &fl4);
1560 inet = &get_cpu_var(unicast_sock);
1562 inet->tos = arg->tos;
1564 sk->sk_priority = skb->priority;
1565 sk->sk_protocol = ip_hdr(skb)->protocol;
1566 sk->sk_bound_dev_if = arg->bound_dev_if;
1567 sock_net_set(sk, net);
1568 __skb_queue_head_init(&sk->sk_write_queue);
1569 sk->sk_sndbuf = sysctl_wmem_default;
1570 ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1571 &ipc, &rt, MSG_DONTWAIT);
1572 nskb = skb_peek(&sk->sk_write_queue);
1574 if (arg->csumoffset >= 0)
1575 *((__sum16 *)skb_transport_header(nskb) +
1576 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1578 nskb->ip_summed = CHECKSUM_NONE;
1580 skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1581 ip_push_pending_frames(sk, &fl4);
1584 put_cpu_var(unicast_sock);
1589 void __init ip_init(void)
1594 #if defined(CONFIG_IP_MULTICAST)