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: seglen is <= mtu
244 if (skb_gso_validate_mtu(skb, mtu))
245 return ip_finish_output2(net, sk, skb);
247 /* Slowpath - GSO segment length exceeds the egress MTU.
249 * This can happen in several cases:
250 * - Forwarding of a TCP GRO skb, when DF flag is not set.
251 * - Forwarding of an skb that arrived on a virtualization interface
252 * (virtio-net/vhost/tap) with TSO/GSO size set by other network
254 * - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
255 * interface with a smaller MTU.
256 * - Arriving GRO skb (or GSO skb in a virtualized environment) that is
257 * bridged to a NETIF_F_TSO tunnel stacked over an interface with an
260 features = netif_skb_features(skb);
261 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_SGO_CB_OFFSET);
262 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
263 if (IS_ERR_OR_NULL(segs)) {
271 struct sk_buff *nskb = segs->next;
275 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
285 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
289 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
290 /* Policy lookup after SNAT yielded a new policy */
291 if (skb_dst(skb)->xfrm) {
292 IPCB(skb)->flags |= IPSKB_REROUTED;
293 return dst_output(net, sk, skb);
296 mtu = ip_skb_dst_mtu(sk, skb);
298 return ip_finish_output_gso(net, sk, skb, mtu);
300 if (skb->len > mtu || (IPCB(skb)->flags & IPSKB_FRAG_PMTU))
301 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
303 return ip_finish_output2(net, sk, skb);
306 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
308 struct rtable *rt = skb_rtable(skb);
309 struct net_device *dev = rt->dst.dev;
312 * If the indicated interface is up and running, send the packet.
314 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
317 skb->protocol = htons(ETH_P_IP);
320 * Multicasts are looped back for other local users
323 if (rt->rt_flags&RTCF_MULTICAST) {
325 #ifdef CONFIG_IP_MROUTE
326 /* Small optimization: do not loopback not local frames,
327 which returned after forwarding; they will be dropped
328 by ip_mr_input in any case.
329 Note, that local frames are looped back to be delivered
332 This check is duplicated in ip_mr_input at the moment.
335 ((rt->rt_flags & RTCF_LOCAL) ||
336 !(IPCB(skb)->flags & IPSKB_FORWARDED))
339 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
341 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
342 net, sk, newskb, NULL, newskb->dev,
346 /* Multicasts with ttl 0 must not go beyond the host */
348 if (ip_hdr(skb)->ttl == 0) {
354 if (rt->rt_flags&RTCF_BROADCAST) {
355 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
357 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
358 net, sk, newskb, NULL, newskb->dev,
362 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
363 net, sk, skb, NULL, skb->dev,
365 !(IPCB(skb)->flags & IPSKB_REROUTED));
368 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
370 struct net_device *dev = skb_dst(skb)->dev;
372 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
375 skb->protocol = htons(ETH_P_IP);
377 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
378 net, sk, skb, NULL, dev,
380 !(IPCB(skb)->flags & IPSKB_REROUTED));
384 * copy saddr and daddr, possibly using 64bit load/stores
386 * iph->saddr = fl4->saddr;
387 * iph->daddr = fl4->daddr;
389 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
391 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
392 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
393 memcpy(&iph->saddr, &fl4->saddr,
394 sizeof(fl4->saddr) + sizeof(fl4->daddr));
397 /* Note: skb->sk can be different from sk, in case of tunnels */
398 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
400 struct inet_sock *inet = inet_sk(sk);
401 struct net *net = sock_net(sk);
402 struct ip_options_rcu *inet_opt;
408 /* Skip all of this if the packet is already routed,
409 * f.e. by something like SCTP.
412 inet_opt = rcu_dereference(inet->inet_opt);
414 rt = skb_rtable(skb);
418 /* Make sure we can route this packet. */
419 rt = (struct rtable *)__sk_dst_check(sk, 0);
423 /* Use correct destination address if we have options. */
424 daddr = inet->inet_daddr;
425 if (inet_opt && inet_opt->opt.srr)
426 daddr = inet_opt->opt.faddr;
428 /* If this fails, retransmit mechanism of transport layer will
429 * keep trying until route appears or the connection times
432 rt = ip_route_output_ports(net, fl4, sk,
433 daddr, inet->inet_saddr,
438 sk->sk_bound_dev_if);
441 sk_setup_caps(sk, &rt->dst);
443 skb_dst_set_noref(skb, &rt->dst);
446 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
449 /* OK, we know where to send it, allocate and build IP header. */
450 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
451 skb_reset_network_header(skb);
453 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
454 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
455 iph->frag_off = htons(IP_DF);
458 iph->ttl = ip_select_ttl(inet, &rt->dst);
459 iph->protocol = sk->sk_protocol;
460 ip_copy_addrs(iph, fl4);
462 /* Transport layer set skb->h.foo itself. */
464 if (inet_opt && inet_opt->opt.optlen) {
465 iph->ihl += inet_opt->opt.optlen >> 2;
466 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
469 ip_select_ident_segs(net, skb, sk,
470 skb_shinfo(skb)->gso_segs ?: 1);
472 /* TODO : should we use skb->sk here instead of sk ? */
473 skb->priority = sk->sk_priority;
474 skb->mark = sk->sk_mark;
476 res = ip_local_out(net, sk, skb);
482 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
484 return -EHOSTUNREACH;
486 EXPORT_SYMBOL(ip_queue_xmit);
488 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
490 to->pkt_type = from->pkt_type;
491 to->priority = from->priority;
492 to->protocol = from->protocol;
494 skb_dst_copy(to, from);
496 to->mark = from->mark;
498 /* Copy the flags to each fragment. */
499 IPCB(to)->flags = IPCB(from)->flags;
501 #ifdef CONFIG_NET_SCHED
502 to->tc_index = from->tc_index;
505 #if IS_ENABLED(CONFIG_IP_VS)
506 to->ipvs_property = from->ipvs_property;
508 skb_copy_secmark(to, from);
511 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
513 int (*output)(struct net *, struct sock *, struct sk_buff *))
515 struct iphdr *iph = ip_hdr(skb);
517 if ((iph->frag_off & htons(IP_DF)) == 0)
518 return ip_do_fragment(net, sk, skb, output);
520 if (unlikely(!skb->ignore_df ||
521 (IPCB(skb)->frag_max_size &&
522 IPCB(skb)->frag_max_size > mtu))) {
523 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
524 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
530 return ip_do_fragment(net, sk, skb, output);
534 * This IP datagram is too large to be sent in one piece. Break it up into
535 * smaller pieces (each of size equal to IP header plus
536 * a block of the data of the original IP data part) that will yet fit in a
537 * single device frame, and queue such a frame for sending.
540 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
541 int (*output)(struct net *, struct sock *, struct sk_buff *))
545 struct sk_buff *skb2;
546 unsigned int mtu, hlen, left, len, ll_rs;
548 __be16 not_last_frag;
549 struct rtable *rt = skb_rtable(skb);
552 /* for offloaded checksums cleanup checksum before fragmentation */
553 if (skb->ip_summed == CHECKSUM_PARTIAL &&
554 (err = skb_checksum_help(skb)))
558 * Point into the IP datagram header.
563 mtu = ip_skb_dst_mtu(sk, skb);
564 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
565 mtu = IPCB(skb)->frag_max_size;
568 * Setup starting values.
572 mtu = mtu - hlen; /* Size of data space */
573 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
575 /* When frag_list is given, use it. First, check its validity:
576 * some transformers could create wrong frag_list or break existing
577 * one, it is not prohibited. In this case fall back to copying.
579 * LATER: this step can be merged to real generation of fragments,
580 * we can switch to copy when see the first bad fragment.
582 if (skb_has_frag_list(skb)) {
583 struct sk_buff *frag, *frag2;
584 int first_len = skb_pagelen(skb);
586 if (first_len - hlen > mtu ||
587 ((first_len - hlen) & 7) ||
588 ip_is_fragment(iph) ||
592 skb_walk_frags(skb, frag) {
593 /* Correct geometry. */
594 if (frag->len > mtu ||
595 ((frag->len & 7) && frag->next) ||
596 skb_headroom(frag) < hlen)
597 goto slow_path_clean;
599 /* Partially cloned skb? */
600 if (skb_shared(frag))
601 goto slow_path_clean;
606 frag->destructor = sock_wfree;
608 skb->truesize -= frag->truesize;
611 /* Everything is OK. Generate! */
615 frag = skb_shinfo(skb)->frag_list;
616 skb_frag_list_init(skb);
617 skb->data_len = first_len - skb_headlen(skb);
618 skb->len = first_len;
619 iph->tot_len = htons(first_len);
620 iph->frag_off = htons(IP_MF);
624 /* Prepare header of the next frame,
625 * before previous one went down. */
627 frag->ip_summed = CHECKSUM_NONE;
628 skb_reset_transport_header(frag);
629 __skb_push(frag, hlen);
630 skb_reset_network_header(frag);
631 memcpy(skb_network_header(frag), iph, hlen);
633 iph->tot_len = htons(frag->len);
634 ip_copy_metadata(frag, skb);
636 ip_options_fragment(frag);
637 offset += skb->len - hlen;
638 iph->frag_off = htons(offset>>3);
640 iph->frag_off |= htons(IP_MF);
641 /* Ready, complete checksum */
645 err = output(net, sk, skb);
648 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
658 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
667 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
671 skb_walk_frags(skb, frag2) {
675 frag2->destructor = NULL;
676 skb->truesize += frag2->truesize;
683 left = skb->len - hlen; /* Space per frame */
684 ptr = hlen; /* Where to start from */
686 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
689 * Fragment the datagram.
692 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
693 not_last_frag = iph->frag_off & htons(IP_MF);
696 * Keep copying data until we run out.
701 /* IF: it doesn't fit, use 'mtu' - the data space left */
704 /* IF: we are not sending up to and including the packet end
705 then align the next start on an eight byte boundary */
710 /* Allocate buffer */
711 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
718 * Set up data on packet
721 ip_copy_metadata(skb2, skb);
722 skb_reserve(skb2, ll_rs);
723 skb_put(skb2, len + hlen);
724 skb_reset_network_header(skb2);
725 skb2->transport_header = skb2->network_header + hlen;
728 * Charge the memory for the fragment to any owner
733 skb_set_owner_w(skb2, skb->sk);
736 * Copy the packet header into the new buffer.
739 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
742 * Copy a block of the IP datagram.
744 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
749 * Fill in the new header fields.
752 iph->frag_off = htons((offset >> 3));
754 if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
755 iph->frag_off |= htons(IP_DF);
757 /* ANK: dirty, but effective trick. Upgrade options only if
758 * the segment to be fragmented was THE FIRST (otherwise,
759 * options are already fixed) and make it ONCE
760 * on the initial skb, so that all the following fragments
761 * will inherit fixed options.
764 ip_options_fragment(skb);
767 * Added AC : If we are fragmenting a fragment that's not the
768 * last fragment then keep MF on each bit
770 if (left > 0 || not_last_frag)
771 iph->frag_off |= htons(IP_MF);
776 * Put this fragment into the sending queue.
778 iph->tot_len = htons(len + hlen);
782 err = output(net, sk, skb2);
786 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
789 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
794 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
797 EXPORT_SYMBOL(ip_do_fragment);
800 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
802 struct msghdr *msg = from;
804 if (skb->ip_summed == CHECKSUM_PARTIAL) {
805 if (copy_from_iter(to, len, &msg->msg_iter) != len)
809 if (csum_and_copy_from_iter(to, len, &csum, &msg->msg_iter) != len)
811 skb->csum = csum_block_add(skb->csum, csum, odd);
815 EXPORT_SYMBOL(ip_generic_getfrag);
818 csum_page(struct page *page, int offset, int copy)
823 csum = csum_partial(kaddr + offset, copy, 0);
828 static inline int ip_ufo_append_data(struct sock *sk,
829 struct sk_buff_head *queue,
830 int getfrag(void *from, char *to, int offset, int len,
831 int odd, struct sk_buff *skb),
832 void *from, int length, int hh_len, int fragheaderlen,
833 int transhdrlen, int maxfraglen, unsigned int flags)
838 /* There is support for UDP fragmentation offload by network
839 * device, so create one single skb packet containing complete
842 skb = skb_peek_tail(queue);
844 skb = sock_alloc_send_skb(sk,
845 hh_len + fragheaderlen + transhdrlen + 20,
846 (flags & MSG_DONTWAIT), &err);
851 /* reserve space for Hardware header */
852 skb_reserve(skb, hh_len);
854 /* create space for UDP/IP header */
855 skb_put(skb, fragheaderlen + transhdrlen);
857 /* initialize network header pointer */
858 skb_reset_network_header(skb);
860 /* initialize protocol header pointer */
861 skb->transport_header = skb->network_header + fragheaderlen;
865 __skb_queue_tail(queue, skb);
866 } else if (skb_is_gso(skb)) {
870 skb->ip_summed = CHECKSUM_PARTIAL;
871 /* specify the length of each IP datagram fragment */
872 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
873 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
876 return skb_append_datato_frags(sk, skb, getfrag, from,
877 (length - transhdrlen));
880 static int __ip_append_data(struct sock *sk,
882 struct sk_buff_head *queue,
883 struct inet_cork *cork,
884 struct page_frag *pfrag,
885 int getfrag(void *from, char *to, int offset,
886 int len, int odd, struct sk_buff *skb),
887 void *from, int length, int transhdrlen,
890 struct inet_sock *inet = inet_sk(sk);
893 struct ip_options *opt = cork->opt;
900 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
901 int csummode = CHECKSUM_NONE;
902 struct rtable *rt = (struct rtable *)cork->dst;
905 skb = skb_peek_tail(queue);
907 exthdrlen = !skb ? rt->dst.header_len : 0;
908 mtu = cork->fragsize;
909 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
910 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
911 tskey = sk->sk_tskey++;
913 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
915 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
916 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
917 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
919 if (cork->length + length > maxnonfragsize - fragheaderlen) {
920 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
921 mtu - (opt ? opt->optlen : 0));
926 * transhdrlen > 0 means that this is the first fragment and we wish
927 * it won't be fragmented in the future.
930 length + fragheaderlen <= mtu &&
931 rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
932 !(flags & MSG_MORE) &&
934 csummode = CHECKSUM_PARTIAL;
936 cork->length += length;
937 if (((length > mtu) || (skb && skb_is_gso(skb))) &&
938 (sk->sk_protocol == IPPROTO_UDP) &&
939 (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len &&
940 (sk->sk_type == SOCK_DGRAM) && !sk->sk_no_check_tx) {
941 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
942 hh_len, fragheaderlen, transhdrlen,
949 /* So, what's going on in the loop below?
951 * We use calculated fragment length to generate chained skb,
952 * each of segments is IP fragment ready for sending to network after
953 * adding appropriate IP header.
960 /* Check if the remaining data fits into current packet. */
961 copy = mtu - skb->len;
963 copy = maxfraglen - skb->len;
966 unsigned int datalen;
967 unsigned int fraglen;
968 unsigned int fraggap;
969 unsigned int alloclen;
970 struct sk_buff *skb_prev;
974 fraggap = skb_prev->len - maxfraglen;
979 * If remaining data exceeds the mtu,
980 * we know we need more fragment(s).
982 datalen = length + fraggap;
983 if (datalen > mtu - fragheaderlen)
984 datalen = maxfraglen - fragheaderlen;
985 fraglen = datalen + fragheaderlen;
987 if ((flags & MSG_MORE) &&
988 !(rt->dst.dev->features&NETIF_F_SG))
993 alloclen += exthdrlen;
995 /* The last fragment gets additional space at tail.
996 * Note, with MSG_MORE we overallocate on fragments,
997 * because we have no idea what fragment will be
1000 if (datalen == length + fraggap)
1001 alloclen += rt->dst.trailer_len;
1004 skb = sock_alloc_send_skb(sk,
1005 alloclen + hh_len + 15,
1006 (flags & MSG_DONTWAIT), &err);
1009 if (atomic_read(&sk->sk_wmem_alloc) <=
1011 skb = sock_wmalloc(sk,
1012 alloclen + hh_len + 15, 1,
1021 * Fill in the control structures
1023 skb->ip_summed = csummode;
1025 skb_reserve(skb, hh_len);
1027 /* only the initial fragment is time stamped */
1028 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1030 skb_shinfo(skb)->tskey = tskey;
1034 * Find where to start putting bytes.
1036 data = skb_put(skb, fraglen + exthdrlen);
1037 skb_set_network_header(skb, exthdrlen);
1038 skb->transport_header = (skb->network_header +
1040 data += fragheaderlen + exthdrlen;
1043 skb->csum = skb_copy_and_csum_bits(
1044 skb_prev, maxfraglen,
1045 data + transhdrlen, fraggap, 0);
1046 skb_prev->csum = csum_sub(skb_prev->csum,
1049 pskb_trim_unique(skb_prev, maxfraglen);
1052 copy = datalen - transhdrlen - fraggap;
1053 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1060 length -= datalen - fraggap;
1063 csummode = CHECKSUM_NONE;
1066 * Put the packet on the pending queue.
1068 __skb_queue_tail(queue, skb);
1075 if (!(rt->dst.dev->features&NETIF_F_SG)) {
1079 if (getfrag(from, skb_put(skb, copy),
1080 offset, copy, off, skb) < 0) {
1081 __skb_trim(skb, off);
1086 int i = skb_shinfo(skb)->nr_frags;
1089 if (!sk_page_frag_refill(sk, pfrag))
1092 if (!skb_can_coalesce(skb, i, pfrag->page,
1095 if (i == MAX_SKB_FRAGS)
1098 __skb_fill_page_desc(skb, i, pfrag->page,
1100 skb_shinfo(skb)->nr_frags = ++i;
1101 get_page(pfrag->page);
1103 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1105 page_address(pfrag->page) + pfrag->offset,
1106 offset, copy, skb->len, skb) < 0)
1109 pfrag->offset += copy;
1110 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1112 skb->data_len += copy;
1113 skb->truesize += copy;
1114 atomic_add(copy, &sk->sk_wmem_alloc);
1125 cork->length -= length;
1126 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1130 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1131 struct ipcm_cookie *ipc, struct rtable **rtp)
1133 struct ip_options_rcu *opt;
1137 * setup for corking.
1142 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1144 if (unlikely(!cork->opt))
1147 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1148 cork->flags |= IPCORK_OPT;
1149 cork->addr = ipc->addr;
1155 * We steal reference to this route, caller should not release it
1158 cork->fragsize = ip_sk_use_pmtu(sk) ?
1159 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1160 cork->dst = &rt->dst;
1162 cork->ttl = ipc->ttl;
1163 cork->tos = ipc->tos;
1164 cork->priority = ipc->priority;
1165 cork->tx_flags = ipc->tx_flags;
1171 * ip_append_data() and ip_append_page() can make one large IP datagram
1172 * from many pieces of data. Each pieces will be holded on the socket
1173 * until ip_push_pending_frames() is called. Each piece can be a page
1176 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1177 * this interface potentially.
1179 * LATER: length must be adjusted by pad at tail, when it is required.
1181 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1182 int getfrag(void *from, char *to, int offset, int len,
1183 int odd, struct sk_buff *skb),
1184 void *from, int length, int transhdrlen,
1185 struct ipcm_cookie *ipc, struct rtable **rtp,
1188 struct inet_sock *inet = inet_sk(sk);
1191 if (flags&MSG_PROBE)
1194 if (skb_queue_empty(&sk->sk_write_queue)) {
1195 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1202 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1203 sk_page_frag(sk), getfrag,
1204 from, length, transhdrlen, flags);
1207 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1208 int offset, size_t size, int flags)
1210 struct inet_sock *inet = inet_sk(sk);
1211 struct sk_buff *skb;
1213 struct ip_options *opt = NULL;
1214 struct inet_cork *cork;
1219 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1224 if (flags&MSG_PROBE)
1227 if (skb_queue_empty(&sk->sk_write_queue))
1230 cork = &inet->cork.base;
1231 rt = (struct rtable *)cork->dst;
1232 if (cork->flags & IPCORK_OPT)
1235 if (!(rt->dst.dev->features&NETIF_F_SG))
1238 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1239 mtu = cork->fragsize;
1241 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1242 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1243 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1245 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1246 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1247 mtu - (opt ? opt->optlen : 0));
1251 skb = skb_peek_tail(&sk->sk_write_queue);
1255 if ((size + skb->len > mtu) &&
1256 (sk->sk_protocol == IPPROTO_UDP) &&
1257 (rt->dst.dev->features & NETIF_F_UFO)) {
1258 if (skb->ip_summed != CHECKSUM_PARTIAL)
1261 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1262 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1264 cork->length += size;
1267 if (skb_is_gso(skb)) {
1271 /* Check if the remaining data fits into current packet. */
1272 len = mtu - skb->len;
1274 len = maxfraglen - skb->len;
1277 struct sk_buff *skb_prev;
1281 fraggap = skb_prev->len - maxfraglen;
1283 alloclen = fragheaderlen + hh_len + fraggap + 15;
1284 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1285 if (unlikely(!skb)) {
1291 * Fill in the control structures
1293 skb->ip_summed = CHECKSUM_NONE;
1295 skb_reserve(skb, hh_len);
1298 * Find where to start putting bytes.
1300 skb_put(skb, fragheaderlen + fraggap);
1301 skb_reset_network_header(skb);
1302 skb->transport_header = (skb->network_header +
1305 skb->csum = skb_copy_and_csum_bits(skb_prev,
1307 skb_transport_header(skb),
1309 skb_prev->csum = csum_sub(skb_prev->csum,
1311 pskb_trim_unique(skb_prev, maxfraglen);
1315 * Put the packet on the pending queue.
1317 __skb_queue_tail(&sk->sk_write_queue, skb);
1324 if (skb_append_pagefrags(skb, page, offset, len)) {
1329 if (skb->ip_summed == CHECKSUM_NONE) {
1331 csum = csum_page(page, offset, len);
1332 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1336 skb->data_len += len;
1337 skb->truesize += len;
1338 atomic_add(len, &sk->sk_wmem_alloc);
1345 cork->length -= size;
1346 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1350 static void ip_cork_release(struct inet_cork *cork)
1352 cork->flags &= ~IPCORK_OPT;
1355 dst_release(cork->dst);
1360 * Combined all pending IP fragments on the socket as one IP datagram
1361 * and push them out.
1363 struct sk_buff *__ip_make_skb(struct sock *sk,
1365 struct sk_buff_head *queue,
1366 struct inet_cork *cork)
1368 struct sk_buff *skb, *tmp_skb;
1369 struct sk_buff **tail_skb;
1370 struct inet_sock *inet = inet_sk(sk);
1371 struct net *net = sock_net(sk);
1372 struct ip_options *opt = NULL;
1373 struct rtable *rt = (struct rtable *)cork->dst;
1378 skb = __skb_dequeue(queue);
1381 tail_skb = &(skb_shinfo(skb)->frag_list);
1383 /* move skb->data to ip header from ext header */
1384 if (skb->data < skb_network_header(skb))
1385 __skb_pull(skb, skb_network_offset(skb));
1386 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1387 __skb_pull(tmp_skb, skb_network_header_len(skb));
1388 *tail_skb = tmp_skb;
1389 tail_skb = &(tmp_skb->next);
1390 skb->len += tmp_skb->len;
1391 skb->data_len += tmp_skb->len;
1392 skb->truesize += tmp_skb->truesize;
1393 tmp_skb->destructor = NULL;
1397 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1398 * to fragment the frame generated here. No matter, what transforms
1399 * how transforms change size of the packet, it will come out.
1401 skb->ignore_df = ip_sk_ignore_df(sk);
1403 /* DF bit is set when we want to see DF on outgoing frames.
1404 * If ignore_df is set too, we still allow to fragment this frame
1406 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1407 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1408 (skb->len <= dst_mtu(&rt->dst) &&
1409 ip_dont_fragment(sk, &rt->dst)))
1412 if (cork->flags & IPCORK_OPT)
1417 else if (rt->rt_type == RTN_MULTICAST)
1420 ttl = ip_select_ttl(inet, &rt->dst);
1425 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1428 iph->protocol = sk->sk_protocol;
1429 ip_copy_addrs(iph, fl4);
1430 ip_select_ident(net, skb, sk);
1433 iph->ihl += opt->optlen>>2;
1434 ip_options_build(skb, opt, cork->addr, rt, 0);
1437 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1438 skb->mark = sk->sk_mark;
1440 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1444 skb_dst_set(skb, &rt->dst);
1446 if (iph->protocol == IPPROTO_ICMP)
1447 icmp_out_count(net, ((struct icmphdr *)
1448 skb_transport_header(skb))->type);
1450 ip_cork_release(cork);
1455 int ip_send_skb(struct net *net, struct sk_buff *skb)
1459 err = ip_local_out(net, skb->sk, skb);
1462 err = net_xmit_errno(err);
1464 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1470 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1472 struct sk_buff *skb;
1474 skb = ip_finish_skb(sk, fl4);
1478 /* Netfilter gets whole the not fragmented skb. */
1479 return ip_send_skb(sock_net(sk), skb);
1483 * Throw away all pending data on the socket.
1485 static void __ip_flush_pending_frames(struct sock *sk,
1486 struct sk_buff_head *queue,
1487 struct inet_cork *cork)
1489 struct sk_buff *skb;
1491 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1494 ip_cork_release(cork);
1497 void ip_flush_pending_frames(struct sock *sk)
1499 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1502 struct sk_buff *ip_make_skb(struct sock *sk,
1504 int getfrag(void *from, char *to, int offset,
1505 int len, int odd, struct sk_buff *skb),
1506 void *from, int length, int transhdrlen,
1507 struct ipcm_cookie *ipc, struct rtable **rtp,
1510 struct inet_cork cork;
1511 struct sk_buff_head queue;
1514 if (flags & MSG_PROBE)
1517 __skb_queue_head_init(&queue);
1522 err = ip_setup_cork(sk, &cork, ipc, rtp);
1524 return ERR_PTR(err);
1526 err = __ip_append_data(sk, fl4, &queue, &cork,
1527 ¤t->task_frag, getfrag,
1528 from, length, transhdrlen, flags);
1530 __ip_flush_pending_frames(sk, &queue, &cork);
1531 return ERR_PTR(err);
1534 return __ip_make_skb(sk, fl4, &queue, &cork);
1538 * Fetch data from kernel space and fill in checksum if needed.
1540 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1541 int len, int odd, struct sk_buff *skb)
1545 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1546 skb->csum = csum_block_add(skb->csum, csum, odd);
1551 * Generic function to send a packet as reply to another packet.
1552 * Used to send some TCP resets/acks so far.
1554 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1555 const struct ip_options *sopt,
1556 __be32 daddr, __be32 saddr,
1557 const struct ip_reply_arg *arg,
1560 struct ip_options_data replyopts;
1561 struct ipcm_cookie ipc;
1563 struct rtable *rt = skb_rtable(skb);
1564 struct net *net = sock_net(sk);
1565 struct sk_buff *nskb;
1569 if (__ip_options_echo(&replyopts.opt.opt, skb, sopt))
1578 if (replyopts.opt.opt.optlen) {
1579 ipc.opt = &replyopts.opt;
1581 if (replyopts.opt.opt.srr)
1582 daddr = replyopts.opt.opt.faddr;
1585 oif = arg->bound_dev_if;
1586 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1589 flowi4_init_output(&fl4, oif,
1590 IP4_REPLY_MARK(net, skb->mark),
1592 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1593 ip_reply_arg_flowi_flags(arg),
1595 tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1597 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1598 rt = ip_route_output_key(net, &fl4);
1602 inet_sk(sk)->tos = arg->tos;
1604 sk->sk_priority = skb->priority;
1605 sk->sk_protocol = ip_hdr(skb)->protocol;
1606 sk->sk_bound_dev_if = arg->bound_dev_if;
1607 sk->sk_sndbuf = sysctl_wmem_default;
1608 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1609 len, 0, &ipc, &rt, MSG_DONTWAIT);
1610 if (unlikely(err)) {
1611 ip_flush_pending_frames(sk);
1615 nskb = skb_peek(&sk->sk_write_queue);
1617 if (arg->csumoffset >= 0)
1618 *((__sum16 *)skb_transport_header(nskb) +
1619 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1621 nskb->ip_summed = CHECKSUM_NONE;
1622 ip_push_pending_frames(sk, &fl4);
1628 void __init ip_init(void)
1633 #if defined(CONFIG_IP_MULTICAST)