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 <net/lwtunnel.h>
77 #include <linux/igmp.h>
78 #include <linux/netfilter_ipv4.h>
79 #include <linux/netfilter_bridge.h>
80 #include <linux/netlink.h>
81 #include <linux/tcp.h>
84 ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
86 int (*output)(struct net *, struct sock *, struct sk_buff *));
88 /* Generate a checksum for an outgoing IP datagram. */
89 void ip_send_check(struct iphdr *iph)
92 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
94 EXPORT_SYMBOL(ip_send_check);
96 int __ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
98 struct iphdr *iph = ip_hdr(skb);
100 iph->tot_len = htons(skb->len);
103 /* if egress device is enslaved to an L3 master device pass the
104 * skb to its handler for processing
106 skb = l3mdev_ip_out(sk, skb);
110 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
111 net, sk, skb, NULL, skb_dst(skb)->dev,
115 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
119 err = __ip_local_out(net, sk, skb);
120 if (likely(err == 1))
121 err = dst_output(net, sk, skb);
125 EXPORT_SYMBOL_GPL(ip_local_out);
127 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
129 int ttl = inet->uc_ttl;
132 ttl = ip4_dst_hoplimit(dst);
137 * Add an ip header to a skbuff and send it out.
140 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
141 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
143 struct inet_sock *inet = inet_sk(sk);
144 struct rtable *rt = skb_rtable(skb);
145 struct net *net = sock_net(sk);
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(net, 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(net, skb->sk, 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);
209 if (lwtunnel_xmit_redirect(dst->lwtstate)) {
210 int res = lwtunnel_xmit(skb);
212 if (res < 0 || res == LWTUNNEL_XMIT_DONE)
217 nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
218 neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
219 if (unlikely(!neigh))
220 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
221 if (!IS_ERR(neigh)) {
222 int res = dst_neigh_output(dst, neigh, skb);
224 rcu_read_unlock_bh();
227 rcu_read_unlock_bh();
229 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
235 static int ip_finish_output_gso(struct net *net, struct sock *sk,
236 struct sk_buff *skb, unsigned int mtu)
238 netdev_features_t features;
239 struct sk_buff *segs;
242 /* common case: fragmentation of segments is not allowed,
243 * or seglen is <= mtu
245 if (((IPCB(skb)->flags & IPSKB_FRAG_SEGS) == 0) ||
246 skb_gso_validate_mtu(skb, mtu))
247 return ip_finish_output2(net, sk, skb);
249 /* Slowpath - GSO segment length is exceeding the dst MTU.
251 * This can happen in two cases:
252 * 1) TCP GRO packet, DF bit not set
253 * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
254 * from host network stack.
256 features = netif_skb_features(skb);
257 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_SGO_CB_OFFSET);
258 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
259 if (IS_ERR_OR_NULL(segs)) {
267 struct sk_buff *nskb = segs->next;
271 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
281 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
285 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
286 /* Policy lookup after SNAT yielded a new policy */
287 if (skb_dst(skb)->xfrm) {
288 IPCB(skb)->flags |= IPSKB_REROUTED;
289 return dst_output(net, sk, skb);
292 mtu = ip_skb_dst_mtu(sk, skb);
294 return ip_finish_output_gso(net, sk, skb, mtu);
296 if (skb->len > mtu || (IPCB(skb)->flags & IPSKB_FRAG_PMTU))
297 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
299 return ip_finish_output2(net, sk, skb);
302 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
304 struct rtable *rt = skb_rtable(skb);
305 struct net_device *dev = rt->dst.dev;
308 * If the indicated interface is up and running, send the packet.
310 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
313 skb->protocol = htons(ETH_P_IP);
316 * Multicasts are looped back for other local users
319 if (rt->rt_flags&RTCF_MULTICAST) {
321 #ifdef CONFIG_IP_MROUTE
322 /* Small optimization: do not loopback not local frames,
323 which returned after forwarding; they will be dropped
324 by ip_mr_input in any case.
325 Note, that local frames are looped back to be delivered
328 This check is duplicated in ip_mr_input at the moment.
331 ((rt->rt_flags & RTCF_LOCAL) ||
332 !(IPCB(skb)->flags & IPSKB_FORWARDED))
335 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
337 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
338 net, sk, newskb, NULL, newskb->dev,
342 /* Multicasts with ttl 0 must not go beyond the host */
344 if (ip_hdr(skb)->ttl == 0) {
350 if (rt->rt_flags&RTCF_BROADCAST) {
351 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
353 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
354 net, sk, newskb, NULL, newskb->dev,
358 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
359 net, sk, skb, NULL, skb->dev,
361 !(IPCB(skb)->flags & IPSKB_REROUTED));
364 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
366 struct net_device *dev = skb_dst(skb)->dev;
368 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
371 skb->protocol = htons(ETH_P_IP);
373 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
374 net, sk, skb, NULL, dev,
376 !(IPCB(skb)->flags & IPSKB_REROUTED));
380 * copy saddr and daddr, possibly using 64bit load/stores
382 * iph->saddr = fl4->saddr;
383 * iph->daddr = fl4->daddr;
385 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
387 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
388 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
389 memcpy(&iph->saddr, &fl4->saddr,
390 sizeof(fl4->saddr) + sizeof(fl4->daddr));
393 /* Note: skb->sk can be different from sk, in case of tunnels */
394 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
396 struct inet_sock *inet = inet_sk(sk);
397 struct net *net = sock_net(sk);
398 struct ip_options_rcu *inet_opt;
404 /* Skip all of this if the packet is already routed,
405 * f.e. by something like SCTP.
408 inet_opt = rcu_dereference(inet->inet_opt);
410 rt = skb_rtable(skb);
414 /* Make sure we can route this packet. */
415 rt = (struct rtable *)__sk_dst_check(sk, 0);
419 /* Use correct destination address if we have options. */
420 daddr = inet->inet_daddr;
421 if (inet_opt && inet_opt->opt.srr)
422 daddr = inet_opt->opt.faddr;
424 /* If this fails, retransmit mechanism of transport layer will
425 * keep trying until route appears or the connection times
428 rt = ip_route_output_ports(net, fl4, sk,
429 daddr, inet->inet_saddr,
434 sk->sk_bound_dev_if);
437 sk_setup_caps(sk, &rt->dst);
439 skb_dst_set_noref(skb, &rt->dst);
442 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
445 /* OK, we know where to send it, allocate and build IP header. */
446 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
447 skb_reset_network_header(skb);
449 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
450 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
451 iph->frag_off = htons(IP_DF);
454 iph->ttl = ip_select_ttl(inet, &rt->dst);
455 iph->protocol = sk->sk_protocol;
456 ip_copy_addrs(iph, fl4);
458 /* Transport layer set skb->h.foo itself. */
460 if (inet_opt && inet_opt->opt.optlen) {
461 iph->ihl += inet_opt->opt.optlen >> 2;
462 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
465 ip_select_ident_segs(net, skb, sk,
466 skb_shinfo(skb)->gso_segs ?: 1);
468 /* TODO : should we use skb->sk here instead of sk ? */
469 skb->priority = sk->sk_priority;
470 skb->mark = sk->sk_mark;
472 res = ip_local_out(net, sk, skb);
478 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
480 return -EHOSTUNREACH;
482 EXPORT_SYMBOL(ip_queue_xmit);
484 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
486 to->pkt_type = from->pkt_type;
487 to->priority = from->priority;
488 to->protocol = from->protocol;
490 skb_dst_copy(to, from);
492 to->mark = from->mark;
494 /* Copy the flags to each fragment. */
495 IPCB(to)->flags = IPCB(from)->flags;
497 #ifdef CONFIG_NET_SCHED
498 to->tc_index = from->tc_index;
501 #if IS_ENABLED(CONFIG_IP_VS)
502 to->ipvs_property = from->ipvs_property;
504 skb_copy_secmark(to, from);
507 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
509 int (*output)(struct net *, struct sock *, struct sk_buff *))
511 struct iphdr *iph = ip_hdr(skb);
513 if ((iph->frag_off & htons(IP_DF)) == 0)
514 return ip_do_fragment(net, sk, skb, output);
516 if (unlikely(!skb->ignore_df ||
517 (IPCB(skb)->frag_max_size &&
518 IPCB(skb)->frag_max_size > mtu))) {
519 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
520 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
526 return ip_do_fragment(net, sk, skb, output);
530 * This IP datagram is too large to be sent in one piece. Break it up into
531 * smaller pieces (each of size equal to IP header plus
532 * a block of the data of the original IP data part) that will yet fit in a
533 * single device frame, and queue such a frame for sending.
536 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
537 int (*output)(struct net *, struct sock *, struct sk_buff *))
541 struct sk_buff *skb2;
542 unsigned int mtu, hlen, left, len, ll_rs;
544 __be16 not_last_frag;
545 struct rtable *rt = skb_rtable(skb);
548 /* for offloaded checksums cleanup checksum before fragmentation */
549 if (skb->ip_summed == CHECKSUM_PARTIAL &&
550 (err = skb_checksum_help(skb)))
554 * Point into the IP datagram header.
559 mtu = ip_skb_dst_mtu(sk, skb);
560 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
561 mtu = IPCB(skb)->frag_max_size;
564 * Setup starting values.
568 mtu = mtu - hlen; /* Size of data space */
569 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
571 /* When frag_list is given, use it. First, check its validity:
572 * some transformers could create wrong frag_list or break existing
573 * one, it is not prohibited. In this case fall back to copying.
575 * LATER: this step can be merged to real generation of fragments,
576 * we can switch to copy when see the first bad fragment.
578 if (skb_has_frag_list(skb)) {
579 struct sk_buff *frag, *frag2;
580 int first_len = skb_pagelen(skb);
582 if (first_len - hlen > mtu ||
583 ((first_len - hlen) & 7) ||
584 ip_is_fragment(iph) ||
588 skb_walk_frags(skb, frag) {
589 /* Correct geometry. */
590 if (frag->len > mtu ||
591 ((frag->len & 7) && frag->next) ||
592 skb_headroom(frag) < hlen)
593 goto slow_path_clean;
595 /* Partially cloned skb? */
596 if (skb_shared(frag))
597 goto slow_path_clean;
602 frag->destructor = sock_wfree;
604 skb->truesize -= frag->truesize;
607 /* Everything is OK. Generate! */
611 frag = skb_shinfo(skb)->frag_list;
612 skb_frag_list_init(skb);
613 skb->data_len = first_len - skb_headlen(skb);
614 skb->len = first_len;
615 iph->tot_len = htons(first_len);
616 iph->frag_off = htons(IP_MF);
620 /* Prepare header of the next frame,
621 * before previous one went down. */
623 frag->ip_summed = CHECKSUM_NONE;
624 skb_reset_transport_header(frag);
625 __skb_push(frag, hlen);
626 skb_reset_network_header(frag);
627 memcpy(skb_network_header(frag), iph, hlen);
629 iph->tot_len = htons(frag->len);
630 ip_copy_metadata(frag, skb);
632 ip_options_fragment(frag);
633 offset += skb->len - hlen;
634 iph->frag_off = htons(offset>>3);
636 iph->frag_off |= htons(IP_MF);
637 /* Ready, complete checksum */
641 err = output(net, sk, skb);
644 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
654 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
663 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
667 skb_walk_frags(skb, frag2) {
671 frag2->destructor = NULL;
672 skb->truesize += frag2->truesize;
679 left = skb->len - hlen; /* Space per frame */
680 ptr = hlen; /* Where to start from */
682 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
685 * Fragment the datagram.
688 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
689 not_last_frag = iph->frag_off & htons(IP_MF);
692 * Keep copying data until we run out.
697 /* IF: it doesn't fit, use 'mtu' - the data space left */
700 /* IF: we are not sending up to and including the packet end
701 then align the next start on an eight byte boundary */
706 /* Allocate buffer */
707 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
714 * Set up data on packet
717 ip_copy_metadata(skb2, skb);
718 skb_reserve(skb2, ll_rs);
719 skb_put(skb2, len + hlen);
720 skb_reset_network_header(skb2);
721 skb2->transport_header = skb2->network_header + hlen;
724 * Charge the memory for the fragment to any owner
729 skb_set_owner_w(skb2, skb->sk);
732 * Copy the packet header into the new buffer.
735 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
738 * Copy a block of the IP datagram.
740 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
745 * Fill in the new header fields.
748 iph->frag_off = htons((offset >> 3));
750 if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
751 iph->frag_off |= htons(IP_DF);
753 /* ANK: dirty, but effective trick. Upgrade options only if
754 * the segment to be fragmented was THE FIRST (otherwise,
755 * options are already fixed) and make it ONCE
756 * on the initial skb, so that all the following fragments
757 * will inherit fixed options.
760 ip_options_fragment(skb);
763 * Added AC : If we are fragmenting a fragment that's not the
764 * last fragment then keep MF on each bit
766 if (left > 0 || not_last_frag)
767 iph->frag_off |= htons(IP_MF);
772 * Put this fragment into the sending queue.
774 iph->tot_len = htons(len + hlen);
778 err = output(net, sk, skb2);
782 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
785 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
790 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
793 EXPORT_SYMBOL(ip_do_fragment);
796 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
798 struct msghdr *msg = from;
800 if (skb->ip_summed == CHECKSUM_PARTIAL) {
801 if (copy_from_iter(to, len, &msg->msg_iter) != len)
805 if (csum_and_copy_from_iter(to, len, &csum, &msg->msg_iter) != len)
807 skb->csum = csum_block_add(skb->csum, csum, odd);
811 EXPORT_SYMBOL(ip_generic_getfrag);
814 csum_page(struct page *page, int offset, int copy)
819 csum = csum_partial(kaddr + offset, copy, 0);
824 static inline int ip_ufo_append_data(struct sock *sk,
825 struct sk_buff_head *queue,
826 int getfrag(void *from, char *to, int offset, int len,
827 int odd, struct sk_buff *skb),
828 void *from, int length, int hh_len, int fragheaderlen,
829 int transhdrlen, int maxfraglen, unsigned int flags)
834 /* There is support for UDP fragmentation offload by network
835 * device, so create one single skb packet containing complete
838 skb = skb_peek_tail(queue);
840 skb = sock_alloc_send_skb(sk,
841 hh_len + fragheaderlen + transhdrlen + 20,
842 (flags & MSG_DONTWAIT), &err);
847 /* reserve space for Hardware header */
848 skb_reserve(skb, hh_len);
850 /* create space for UDP/IP header */
851 skb_put(skb, fragheaderlen + transhdrlen);
853 /* initialize network header pointer */
854 skb_reset_network_header(skb);
856 /* initialize protocol header pointer */
857 skb->transport_header = skb->network_header + fragheaderlen;
861 __skb_queue_tail(queue, skb);
862 } else if (skb_is_gso(skb)) {
866 skb->ip_summed = CHECKSUM_PARTIAL;
867 /* specify the length of each IP datagram fragment */
868 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
869 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
872 return skb_append_datato_frags(sk, skb, getfrag, from,
873 (length - transhdrlen));
876 static int __ip_append_data(struct sock *sk,
878 struct sk_buff_head *queue,
879 struct inet_cork *cork,
880 struct page_frag *pfrag,
881 int getfrag(void *from, char *to, int offset,
882 int len, int odd, struct sk_buff *skb),
883 void *from, int length, int transhdrlen,
886 struct inet_sock *inet = inet_sk(sk);
889 struct ip_options *opt = cork->opt;
896 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
897 int csummode = CHECKSUM_NONE;
898 struct rtable *rt = (struct rtable *)cork->dst;
901 skb = skb_peek_tail(queue);
903 exthdrlen = !skb ? rt->dst.header_len : 0;
904 mtu = cork->fragsize;
905 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
906 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
907 tskey = sk->sk_tskey++;
909 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
911 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
912 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
913 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
915 if (cork->length + length > maxnonfragsize - fragheaderlen) {
916 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
917 mtu - (opt ? opt->optlen : 0));
922 * transhdrlen > 0 means that this is the first fragment and we wish
923 * it won't be fragmented in the future.
926 length + fragheaderlen <= mtu &&
927 rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
928 !(flags & MSG_MORE) &&
930 csummode = CHECKSUM_PARTIAL;
932 cork->length += length;
933 if (((length > mtu) || (skb && skb_is_gso(skb))) &&
934 (sk->sk_protocol == IPPROTO_UDP) &&
935 (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len &&
936 (sk->sk_type == SOCK_DGRAM) && !sk->sk_no_check_tx) {
937 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
938 hh_len, fragheaderlen, transhdrlen,
945 /* So, what's going on in the loop below?
947 * We use calculated fragment length to generate chained skb,
948 * each of segments is IP fragment ready for sending to network after
949 * adding appropriate IP header.
956 /* Check if the remaining data fits into current packet. */
957 copy = mtu - skb->len;
959 copy = maxfraglen - skb->len;
962 unsigned int datalen;
963 unsigned int fraglen;
964 unsigned int fraggap;
965 unsigned int alloclen;
966 struct sk_buff *skb_prev;
970 fraggap = skb_prev->len - maxfraglen;
975 * If remaining data exceeds the mtu,
976 * we know we need more fragment(s).
978 datalen = length + fraggap;
979 if (datalen > mtu - fragheaderlen)
980 datalen = maxfraglen - fragheaderlen;
981 fraglen = datalen + fragheaderlen;
983 if ((flags & MSG_MORE) &&
984 !(rt->dst.dev->features&NETIF_F_SG))
989 alloclen += exthdrlen;
991 /* The last fragment gets additional space at tail.
992 * Note, with MSG_MORE we overallocate on fragments,
993 * because we have no idea what fragment will be
996 if (datalen == length + fraggap)
997 alloclen += rt->dst.trailer_len;
1000 skb = sock_alloc_send_skb(sk,
1001 alloclen + hh_len + 15,
1002 (flags & MSG_DONTWAIT), &err);
1005 if (atomic_read(&sk->sk_wmem_alloc) <=
1007 skb = sock_wmalloc(sk,
1008 alloclen + hh_len + 15, 1,
1017 * Fill in the control structures
1019 skb->ip_summed = csummode;
1021 skb_reserve(skb, hh_len);
1023 /* only the initial fragment is time stamped */
1024 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1026 skb_shinfo(skb)->tskey = tskey;
1030 * Find where to start putting bytes.
1032 data = skb_put(skb, fraglen + exthdrlen);
1033 skb_set_network_header(skb, exthdrlen);
1034 skb->transport_header = (skb->network_header +
1036 data += fragheaderlen + exthdrlen;
1039 skb->csum = skb_copy_and_csum_bits(
1040 skb_prev, maxfraglen,
1041 data + transhdrlen, fraggap, 0);
1042 skb_prev->csum = csum_sub(skb_prev->csum,
1045 pskb_trim_unique(skb_prev, maxfraglen);
1048 copy = datalen - transhdrlen - fraggap;
1049 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1056 length -= datalen - fraggap;
1059 csummode = CHECKSUM_NONE;
1062 * Put the packet on the pending queue.
1064 __skb_queue_tail(queue, skb);
1071 if (!(rt->dst.dev->features&NETIF_F_SG)) {
1075 if (getfrag(from, skb_put(skb, copy),
1076 offset, copy, off, skb) < 0) {
1077 __skb_trim(skb, off);
1082 int i = skb_shinfo(skb)->nr_frags;
1085 if (!sk_page_frag_refill(sk, pfrag))
1088 if (!skb_can_coalesce(skb, i, pfrag->page,
1091 if (i == MAX_SKB_FRAGS)
1094 __skb_fill_page_desc(skb, i, pfrag->page,
1096 skb_shinfo(skb)->nr_frags = ++i;
1097 get_page(pfrag->page);
1099 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1101 page_address(pfrag->page) + pfrag->offset,
1102 offset, copy, skb->len, skb) < 0)
1105 pfrag->offset += copy;
1106 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1108 skb->data_len += copy;
1109 skb->truesize += copy;
1110 atomic_add(copy, &sk->sk_wmem_alloc);
1121 cork->length -= length;
1122 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1126 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1127 struct ipcm_cookie *ipc, struct rtable **rtp)
1129 struct ip_options_rcu *opt;
1133 * setup for corking.
1138 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1140 if (unlikely(!cork->opt))
1143 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1144 cork->flags |= IPCORK_OPT;
1145 cork->addr = ipc->addr;
1151 * We steal reference to this route, caller should not release it
1154 cork->fragsize = ip_sk_use_pmtu(sk) ?
1155 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1156 cork->dst = &rt->dst;
1158 cork->ttl = ipc->ttl;
1159 cork->tos = ipc->tos;
1160 cork->priority = ipc->priority;
1161 cork->tx_flags = ipc->tx_flags;
1167 * ip_append_data() and ip_append_page() can make one large IP datagram
1168 * from many pieces of data. Each pieces will be holded on the socket
1169 * until ip_push_pending_frames() is called. Each piece can be a page
1172 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1173 * this interface potentially.
1175 * LATER: length must be adjusted by pad at tail, when it is required.
1177 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1178 int getfrag(void *from, char *to, int offset, int len,
1179 int odd, struct sk_buff *skb),
1180 void *from, int length, int transhdrlen,
1181 struct ipcm_cookie *ipc, struct rtable **rtp,
1184 struct inet_sock *inet = inet_sk(sk);
1187 if (flags&MSG_PROBE)
1190 if (skb_queue_empty(&sk->sk_write_queue)) {
1191 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1198 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1199 sk_page_frag(sk), getfrag,
1200 from, length, transhdrlen, flags);
1203 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1204 int offset, size_t size, int flags)
1206 struct inet_sock *inet = inet_sk(sk);
1207 struct sk_buff *skb;
1209 struct ip_options *opt = NULL;
1210 struct inet_cork *cork;
1215 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1220 if (flags&MSG_PROBE)
1223 if (skb_queue_empty(&sk->sk_write_queue))
1226 cork = &inet->cork.base;
1227 rt = (struct rtable *)cork->dst;
1228 if (cork->flags & IPCORK_OPT)
1231 if (!(rt->dst.dev->features&NETIF_F_SG))
1234 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1235 mtu = cork->fragsize;
1237 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1238 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1239 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1241 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1242 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1243 mtu - (opt ? opt->optlen : 0));
1247 skb = skb_peek_tail(&sk->sk_write_queue);
1251 if ((size + skb->len > mtu) &&
1252 (sk->sk_protocol == IPPROTO_UDP) &&
1253 (rt->dst.dev->features & NETIF_F_UFO)) {
1254 if (skb->ip_summed != CHECKSUM_PARTIAL)
1257 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1258 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1260 cork->length += size;
1263 if (skb_is_gso(skb)) {
1267 /* Check if the remaining data fits into current packet. */
1268 len = mtu - skb->len;
1270 len = maxfraglen - skb->len;
1273 struct sk_buff *skb_prev;
1277 fraggap = skb_prev->len - maxfraglen;
1279 alloclen = fragheaderlen + hh_len + fraggap + 15;
1280 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1281 if (unlikely(!skb)) {
1287 * Fill in the control structures
1289 skb->ip_summed = CHECKSUM_NONE;
1291 skb_reserve(skb, hh_len);
1294 * Find where to start putting bytes.
1296 skb_put(skb, fragheaderlen + fraggap);
1297 skb_reset_network_header(skb);
1298 skb->transport_header = (skb->network_header +
1301 skb->csum = skb_copy_and_csum_bits(skb_prev,
1303 skb_transport_header(skb),
1305 skb_prev->csum = csum_sub(skb_prev->csum,
1307 pskb_trim_unique(skb_prev, maxfraglen);
1311 * Put the packet on the pending queue.
1313 __skb_queue_tail(&sk->sk_write_queue, skb);
1320 if (skb_append_pagefrags(skb, page, offset, len)) {
1325 if (skb->ip_summed == CHECKSUM_NONE) {
1327 csum = csum_page(page, offset, len);
1328 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1332 skb->data_len += len;
1333 skb->truesize += len;
1334 atomic_add(len, &sk->sk_wmem_alloc);
1341 cork->length -= size;
1342 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1346 static void ip_cork_release(struct inet_cork *cork)
1348 cork->flags &= ~IPCORK_OPT;
1351 dst_release(cork->dst);
1356 * Combined all pending IP fragments on the socket as one IP datagram
1357 * and push them out.
1359 struct sk_buff *__ip_make_skb(struct sock *sk,
1361 struct sk_buff_head *queue,
1362 struct inet_cork *cork)
1364 struct sk_buff *skb, *tmp_skb;
1365 struct sk_buff **tail_skb;
1366 struct inet_sock *inet = inet_sk(sk);
1367 struct net *net = sock_net(sk);
1368 struct ip_options *opt = NULL;
1369 struct rtable *rt = (struct rtable *)cork->dst;
1374 skb = __skb_dequeue(queue);
1377 tail_skb = &(skb_shinfo(skb)->frag_list);
1379 /* move skb->data to ip header from ext header */
1380 if (skb->data < skb_network_header(skb))
1381 __skb_pull(skb, skb_network_offset(skb));
1382 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1383 __skb_pull(tmp_skb, skb_network_header_len(skb));
1384 *tail_skb = tmp_skb;
1385 tail_skb = &(tmp_skb->next);
1386 skb->len += tmp_skb->len;
1387 skb->data_len += tmp_skb->len;
1388 skb->truesize += tmp_skb->truesize;
1389 tmp_skb->destructor = NULL;
1393 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1394 * to fragment the frame generated here. No matter, what transforms
1395 * how transforms change size of the packet, it will come out.
1397 skb->ignore_df = ip_sk_ignore_df(sk);
1399 /* DF bit is set when we want to see DF on outgoing frames.
1400 * If ignore_df is set too, we still allow to fragment this frame
1402 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1403 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1404 (skb->len <= dst_mtu(&rt->dst) &&
1405 ip_dont_fragment(sk, &rt->dst)))
1408 if (cork->flags & IPCORK_OPT)
1413 else if (rt->rt_type == RTN_MULTICAST)
1416 ttl = ip_select_ttl(inet, &rt->dst);
1421 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1424 iph->protocol = sk->sk_protocol;
1425 ip_copy_addrs(iph, fl4);
1426 ip_select_ident(net, skb, sk);
1429 iph->ihl += opt->optlen>>2;
1430 ip_options_build(skb, opt, cork->addr, rt, 0);
1433 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1434 skb->mark = sk->sk_mark;
1436 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1440 skb_dst_set(skb, &rt->dst);
1442 if (iph->protocol == IPPROTO_ICMP)
1443 icmp_out_count(net, ((struct icmphdr *)
1444 skb_transport_header(skb))->type);
1446 ip_cork_release(cork);
1451 int ip_send_skb(struct net *net, struct sk_buff *skb)
1455 err = ip_local_out(net, skb->sk, skb);
1458 err = net_xmit_errno(err);
1460 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1466 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1468 struct sk_buff *skb;
1470 skb = ip_finish_skb(sk, fl4);
1474 /* Netfilter gets whole the not fragmented skb. */
1475 return ip_send_skb(sock_net(sk), skb);
1479 * Throw away all pending data on the socket.
1481 static void __ip_flush_pending_frames(struct sock *sk,
1482 struct sk_buff_head *queue,
1483 struct inet_cork *cork)
1485 struct sk_buff *skb;
1487 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1490 ip_cork_release(cork);
1493 void ip_flush_pending_frames(struct sock *sk)
1495 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1498 struct sk_buff *ip_make_skb(struct sock *sk,
1500 int getfrag(void *from, char *to, int offset,
1501 int len, int odd, struct sk_buff *skb),
1502 void *from, int length, int transhdrlen,
1503 struct ipcm_cookie *ipc, struct rtable **rtp,
1506 struct inet_cork cork;
1507 struct sk_buff_head queue;
1510 if (flags & MSG_PROBE)
1513 __skb_queue_head_init(&queue);
1518 err = ip_setup_cork(sk, &cork, ipc, rtp);
1520 return ERR_PTR(err);
1522 err = __ip_append_data(sk, fl4, &queue, &cork,
1523 ¤t->task_frag, getfrag,
1524 from, length, transhdrlen, flags);
1526 __ip_flush_pending_frames(sk, &queue, &cork);
1527 return ERR_PTR(err);
1530 return __ip_make_skb(sk, fl4, &queue, &cork);
1534 * Fetch data from kernel space and fill in checksum if needed.
1536 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1537 int len, int odd, struct sk_buff *skb)
1541 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1542 skb->csum = csum_block_add(skb->csum, csum, odd);
1547 * Generic function to send a packet as reply to another packet.
1548 * Used to send some TCP resets/acks so far.
1550 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1551 const struct ip_options *sopt,
1552 __be32 daddr, __be32 saddr,
1553 const struct ip_reply_arg *arg,
1556 struct ip_options_data replyopts;
1557 struct ipcm_cookie ipc;
1559 struct rtable *rt = skb_rtable(skb);
1560 struct net *net = sock_net(sk);
1561 struct sk_buff *nskb;
1565 if (__ip_options_echo(&replyopts.opt.opt, skb, sopt))
1574 if (replyopts.opt.opt.optlen) {
1575 ipc.opt = &replyopts.opt;
1577 if (replyopts.opt.opt.srr)
1578 daddr = replyopts.opt.opt.faddr;
1581 oif = arg->bound_dev_if;
1582 oif = oif ? : skb->skb_iif;
1584 flowi4_init_output(&fl4, oif,
1585 IP4_REPLY_MARK(net, skb->mark),
1587 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1588 ip_reply_arg_flowi_flags(arg),
1590 tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1591 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1592 rt = ip_route_output_key(net, &fl4);
1596 inet_sk(sk)->tos = arg->tos;
1598 sk->sk_priority = skb->priority;
1599 sk->sk_protocol = ip_hdr(skb)->protocol;
1600 sk->sk_bound_dev_if = arg->bound_dev_if;
1601 sk->sk_sndbuf = sysctl_wmem_default;
1602 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1603 len, 0, &ipc, &rt, MSG_DONTWAIT);
1604 if (unlikely(err)) {
1605 ip_flush_pending_frames(sk);
1609 nskb = skb_peek(&sk->sk_write_queue);
1611 if (arg->csumoffset >= 0)
1612 *((__sum16 *)skb_transport_header(nskb) +
1613 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1615 nskb->ip_summed = CHECKSUM_NONE;
1616 ip_push_pending_frames(sk, &fl4);
1622 void __init ip_init(void)
1627 #if defined(CONFIG_IP_MULTICAST)