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/inet_ecn.h>
77 #include <net/lwtunnel.h>
78 #include <linux/igmp.h>
79 #include <linux/netfilter_ipv4.h>
80 #include <linux/netfilter_bridge.h>
81 #include <linux/netlink.h>
82 #include <linux/tcp.h>
85 ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
87 int (*output)(struct net *, struct sock *, struct sk_buff *));
89 /* Generate a checksum for an outgoing IP datagram. */
90 void ip_send_check(struct iphdr *iph)
93 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
95 EXPORT_SYMBOL(ip_send_check);
97 int __ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
99 struct iphdr *iph = ip_hdr(skb);
101 iph->tot_len = htons(skb->len);
104 /* if egress device is enslaved to an L3 master device pass the
105 * skb to its handler for processing
107 skb = l3mdev_ip_out(sk, skb);
111 skb->protocol = htons(ETH_P_IP);
113 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
114 net, sk, skb, NULL, skb_dst(skb)->dev,
118 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
122 err = __ip_local_out(net, sk, skb);
123 if (likely(err == 1))
124 err = dst_output(net, sk, skb);
128 EXPORT_SYMBOL_GPL(ip_local_out);
130 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
132 int ttl = inet->uc_ttl;
135 ttl = ip4_dst_hoplimit(dst);
140 * Add an ip header to a skbuff and send it out.
143 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
144 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
146 struct inet_sock *inet = inet_sk(sk);
147 struct rtable *rt = skb_rtable(skb);
148 struct net *net = sock_net(sk);
151 /* Build the IP header. */
152 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
153 skb_reset_network_header(skb);
157 iph->tos = inet->tos;
158 iph->ttl = ip_select_ttl(inet, &rt->dst);
159 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
161 iph->protocol = sk->sk_protocol;
162 /* Do not bother generating IPID for small packets (eg SYNACK) */
163 if (skb->len <= IPV4_MIN_MTU || ip_dont_fragment(sk, &rt->dst)) {
164 iph->frag_off = htons(IP_DF);
168 /* TCP packets here are SYNACK with fat IPv4/TCP options.
169 * Avoid using the hashed IP ident generator.
171 if (sk->sk_protocol == IPPROTO_TCP)
172 iph->id = (__force __be16)prandom_u32();
174 __ip_select_ident(net, iph, 1);
177 if (opt && opt->opt.optlen) {
178 iph->ihl += opt->opt.optlen>>2;
179 ip_options_build(skb, &opt->opt, daddr, rt, 0);
182 skb->priority = sk->sk_priority;
183 skb->mark = sk->sk_mark;
186 return ip_local_out(net, skb->sk, skb);
188 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
190 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
192 struct dst_entry *dst = skb_dst(skb);
193 struct rtable *rt = (struct rtable *)dst;
194 struct net_device *dev = dst->dev;
195 unsigned int hh_len = LL_RESERVED_SPACE(dev);
196 struct neighbour *neigh;
199 if (rt->rt_type == RTN_MULTICAST) {
200 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
201 } else if (rt->rt_type == RTN_BROADCAST)
202 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
204 /* Be paranoid, rather than too clever. */
205 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
206 struct sk_buff *skb2;
208 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
214 skb_set_owner_w(skb2, skb->sk);
219 if (lwtunnel_xmit_redirect(dst->lwtstate)) {
220 int res = lwtunnel_xmit(skb);
222 if (res < 0 || res == LWTUNNEL_XMIT_DONE)
227 nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
228 neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
229 if (unlikely(!neigh))
230 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
231 if (!IS_ERR(neigh)) {
232 int res = dst_neigh_output(dst, neigh, skb);
234 rcu_read_unlock_bh();
237 rcu_read_unlock_bh();
239 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
245 static int ip_finish_output_gso(struct net *net, struct sock *sk,
246 struct sk_buff *skb, unsigned int mtu)
248 netdev_features_t features;
249 struct sk_buff *segs;
252 /* common case: seglen is <= mtu
254 if (skb_gso_validate_mtu(skb, mtu))
255 return ip_finish_output2(net, sk, skb);
257 /* Slowpath - GSO segment length exceeds the egress MTU.
259 * This can happen in several cases:
260 * - Forwarding of a TCP GRO skb, when DF flag is not set.
261 * - Forwarding of an skb that arrived on a virtualization interface
262 * (virtio-net/vhost/tap) with TSO/GSO size set by other network
264 * - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
265 * interface with a smaller MTU.
266 * - Arriving GRO skb (or GSO skb in a virtualized environment) that is
267 * bridged to a NETIF_F_TSO tunnel stacked over an interface with an
270 features = netif_skb_features(skb);
271 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_SGO_CB_OFFSET);
272 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
273 if (IS_ERR_OR_NULL(segs)) {
281 struct sk_buff *nskb = segs->next;
285 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
295 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
299 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
300 /* Policy lookup after SNAT yielded a new policy */
301 if (skb_dst(skb)->xfrm) {
302 IPCB(skb)->flags |= IPSKB_REROUTED;
303 return dst_output(net, sk, skb);
306 mtu = ip_skb_dst_mtu(sk, skb);
308 return ip_finish_output_gso(net, sk, skb, mtu);
310 if (skb->len > mtu || IPCB(skb)->frag_max_size)
311 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
313 return ip_finish_output2(net, sk, skb);
316 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
318 struct rtable *rt = skb_rtable(skb);
319 struct net_device *dev = rt->dst.dev;
322 * If the indicated interface is up and running, send the packet.
324 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
327 skb->protocol = htons(ETH_P_IP);
330 * Multicasts are looped back for other local users
333 if (rt->rt_flags&RTCF_MULTICAST) {
335 #ifdef CONFIG_IP_MROUTE
336 /* Small optimization: do not loopback not local frames,
337 which returned after forwarding; they will be dropped
338 by ip_mr_input in any case.
339 Note, that local frames are looped back to be delivered
342 This check is duplicated in ip_mr_input at the moment.
345 ((rt->rt_flags & RTCF_LOCAL) ||
346 !(IPCB(skb)->flags & IPSKB_FORWARDED))
349 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
351 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
352 net, sk, newskb, NULL, newskb->dev,
356 /* Multicasts with ttl 0 must not go beyond the host */
358 if (ip_hdr(skb)->ttl == 0) {
364 if (rt->rt_flags&RTCF_BROADCAST) {
365 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
367 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
368 net, sk, newskb, NULL, newskb->dev,
372 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
373 net, sk, skb, NULL, skb->dev,
375 !(IPCB(skb)->flags & IPSKB_REROUTED));
378 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
380 struct net_device *dev = skb_dst(skb)->dev;
382 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
385 skb->protocol = htons(ETH_P_IP);
387 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
388 net, sk, skb, NULL, dev,
390 !(IPCB(skb)->flags & IPSKB_REROUTED));
394 * copy saddr and daddr, possibly using 64bit load/stores
396 * iph->saddr = fl4->saddr;
397 * iph->daddr = fl4->daddr;
399 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
401 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
402 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
404 iph->saddr = fl4->saddr;
405 iph->daddr = fl4->daddr;
408 /* Note: skb->sk can be different from sk, in case of tunnels */
409 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
411 struct inet_sock *inet = inet_sk(sk);
412 struct net *net = sock_net(sk);
413 struct ip_options_rcu *inet_opt;
419 /* Skip all of this if the packet is already routed,
420 * f.e. by something like SCTP.
423 inet_opt = rcu_dereference(inet->inet_opt);
425 rt = skb_rtable(skb);
429 /* Make sure we can route this packet. */
430 rt = (struct rtable *)__sk_dst_check(sk, 0);
434 /* Use correct destination address if we have options. */
435 daddr = inet->inet_daddr;
436 if (inet_opt && inet_opt->opt.srr)
437 daddr = inet_opt->opt.faddr;
439 /* If this fails, retransmit mechanism of transport layer will
440 * keep trying until route appears or the connection times
443 rt = ip_route_output_ports(net, fl4, sk,
444 daddr, inet->inet_saddr,
449 sk->sk_bound_dev_if);
452 sk_setup_caps(sk, &rt->dst);
454 skb_dst_set_noref(skb, &rt->dst);
457 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
460 /* OK, we know where to send it, allocate and build IP header. */
461 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
462 skb_reset_network_header(skb);
464 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
465 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
466 iph->frag_off = htons(IP_DF);
469 iph->ttl = ip_select_ttl(inet, &rt->dst);
470 iph->protocol = sk->sk_protocol;
471 ip_copy_addrs(iph, fl4);
473 /* Transport layer set skb->h.foo itself. */
475 if (inet_opt && inet_opt->opt.optlen) {
476 iph->ihl += inet_opt->opt.optlen >> 2;
477 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
480 ip_select_ident_segs(net, skb, sk,
481 skb_shinfo(skb)->gso_segs ?: 1);
483 /* TODO : should we use skb->sk here instead of sk ? */
484 skb->priority = sk->sk_priority;
485 skb->mark = sk->sk_mark;
487 res = ip_local_out(net, sk, skb);
493 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
495 return -EHOSTUNREACH;
497 EXPORT_SYMBOL(ip_queue_xmit);
499 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
501 to->pkt_type = from->pkt_type;
502 to->priority = from->priority;
503 to->protocol = from->protocol;
504 to->skb_iif = from->skb_iif;
506 skb_dst_copy(to, from);
508 to->mark = from->mark;
510 skb_copy_hash(to, from);
512 /* Copy the flags to each fragment. */
513 IPCB(to)->flags = IPCB(from)->flags;
515 #ifdef CONFIG_NET_SCHED
516 to->tc_index = from->tc_index;
519 #if IS_ENABLED(CONFIG_IP_VS)
520 to->ipvs_property = from->ipvs_property;
522 skb_copy_secmark(to, from);
525 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
527 int (*output)(struct net *, struct sock *, struct sk_buff *))
529 struct iphdr *iph = ip_hdr(skb);
531 if ((iph->frag_off & htons(IP_DF)) == 0)
532 return ip_do_fragment(net, sk, skb, output);
534 if (unlikely(!skb->ignore_df ||
535 (IPCB(skb)->frag_max_size &&
536 IPCB(skb)->frag_max_size > mtu))) {
537 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
538 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
544 return ip_do_fragment(net, sk, skb, output);
548 * This IP datagram is too large to be sent in one piece. Break it up into
549 * smaller pieces (each of size equal to IP header plus
550 * a block of the data of the original IP data part) that will yet fit in a
551 * single device frame, and queue such a frame for sending.
554 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
555 int (*output)(struct net *, struct sock *, struct sk_buff *))
559 struct sk_buff *skb2;
560 unsigned int mtu, hlen, left, len, ll_rs;
562 __be16 not_last_frag;
563 struct rtable *rt = skb_rtable(skb);
566 /* for offloaded checksums cleanup checksum before fragmentation */
567 if (skb->ip_summed == CHECKSUM_PARTIAL &&
568 (err = skb_checksum_help(skb)))
572 * Point into the IP datagram header.
577 mtu = ip_skb_dst_mtu(sk, skb);
578 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
579 mtu = IPCB(skb)->frag_max_size;
582 * Setup starting values.
586 mtu = mtu - hlen; /* Size of data space */
587 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
589 /* When frag_list is given, use it. First, check its validity:
590 * some transformers could create wrong frag_list or break existing
591 * one, it is not prohibited. In this case fall back to copying.
593 * LATER: this step can be merged to real generation of fragments,
594 * we can switch to copy when see the first bad fragment.
596 if (skb_has_frag_list(skb)) {
597 struct sk_buff *frag, *frag2;
598 int first_len = skb_pagelen(skb);
600 if (first_len - hlen > mtu ||
601 ((first_len - hlen) & 7) ||
602 ip_is_fragment(iph) ||
606 skb_walk_frags(skb, frag) {
607 /* Correct geometry. */
608 if (frag->len > mtu ||
609 ((frag->len & 7) && frag->next) ||
610 skb_headroom(frag) < hlen)
611 goto slow_path_clean;
613 /* Partially cloned skb? */
614 if (skb_shared(frag))
615 goto slow_path_clean;
620 frag->destructor = sock_wfree;
622 skb->truesize -= frag->truesize;
625 /* Everything is OK. Generate! */
629 frag = skb_shinfo(skb)->frag_list;
630 skb_frag_list_init(skb);
631 skb->data_len = first_len - skb_headlen(skb);
632 skb->len = first_len;
633 iph->tot_len = htons(first_len);
634 iph->frag_off = htons(IP_MF);
638 /* Prepare header of the next frame,
639 * before previous one went down. */
641 frag->ip_summed = CHECKSUM_NONE;
642 skb_reset_transport_header(frag);
643 __skb_push(frag, hlen);
644 skb_reset_network_header(frag);
645 memcpy(skb_network_header(frag), iph, hlen);
647 iph->tot_len = htons(frag->len);
648 ip_copy_metadata(frag, skb);
650 ip_options_fragment(frag);
651 offset += skb->len - hlen;
652 iph->frag_off = htons(offset>>3);
654 iph->frag_off |= htons(IP_MF);
655 /* Ready, complete checksum */
659 err = output(net, sk, skb);
662 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
672 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
681 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
685 skb_walk_frags(skb, frag2) {
689 frag2->destructor = NULL;
690 skb->truesize += frag2->truesize;
697 left = skb->len - hlen; /* Space per frame */
698 ptr = hlen; /* Where to start from */
700 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
703 * Fragment the datagram.
706 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
707 not_last_frag = iph->frag_off & htons(IP_MF);
710 * Keep copying data until we run out.
715 /* IF: it doesn't fit, use 'mtu' - the data space left */
718 /* IF: we are not sending up to and including the packet end
719 then align the next start on an eight byte boundary */
724 /* Allocate buffer */
725 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
732 * Set up data on packet
735 ip_copy_metadata(skb2, skb);
736 skb_reserve(skb2, ll_rs);
737 skb_put(skb2, len + hlen);
738 skb_reset_network_header(skb2);
739 skb2->transport_header = skb2->network_header + hlen;
742 * Charge the memory for the fragment to any owner
747 skb_set_owner_w(skb2, skb->sk);
750 * Copy the packet header into the new buffer.
753 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
756 * Copy a block of the IP datagram.
758 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
763 * Fill in the new header fields.
766 iph->frag_off = htons((offset >> 3));
768 if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
769 iph->frag_off |= htons(IP_DF);
771 /* ANK: dirty, but effective trick. Upgrade options only if
772 * the segment to be fragmented was THE FIRST (otherwise,
773 * options are already fixed) and make it ONCE
774 * on the initial skb, so that all the following fragments
775 * will inherit fixed options.
778 ip_options_fragment(skb);
781 * Added AC : If we are fragmenting a fragment that's not the
782 * last fragment then keep MF on each bit
784 if (left > 0 || not_last_frag)
785 iph->frag_off |= htons(IP_MF);
790 * Put this fragment into the sending queue.
792 iph->tot_len = htons(len + hlen);
796 err = output(net, sk, skb2);
800 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
803 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
808 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
811 EXPORT_SYMBOL(ip_do_fragment);
814 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
816 struct msghdr *msg = from;
818 if (skb->ip_summed == CHECKSUM_PARTIAL) {
819 if (copy_from_iter(to, len, &msg->msg_iter) != len)
823 if (csum_and_copy_from_iter(to, len, &csum, &msg->msg_iter) != len)
825 skb->csum = csum_block_add(skb->csum, csum, odd);
829 EXPORT_SYMBOL(ip_generic_getfrag);
832 csum_page(struct page *page, int offset, int copy)
837 csum = csum_partial(kaddr + offset, copy, 0);
842 static inline int ip_ufo_append_data(struct sock *sk,
843 struct sk_buff_head *queue,
844 int getfrag(void *from, char *to, int offset, int len,
845 int odd, struct sk_buff *skb),
846 void *from, int length, int hh_len, int fragheaderlen,
847 int transhdrlen, int maxfraglen, unsigned int flags)
852 /* There is support for UDP fragmentation offload by network
853 * device, so create one single skb packet containing complete
856 skb = skb_peek_tail(queue);
858 skb = sock_alloc_send_skb(sk,
859 hh_len + fragheaderlen + transhdrlen + 20,
860 (flags & MSG_DONTWAIT), &err);
865 /* reserve space for Hardware header */
866 skb_reserve(skb, hh_len);
868 /* create space for UDP/IP header */
869 skb_put(skb, fragheaderlen + transhdrlen);
871 /* initialize network header pointer */
872 skb_reset_network_header(skb);
874 /* initialize protocol header pointer */
875 skb->transport_header = skb->network_header + fragheaderlen;
879 __skb_queue_tail(queue, skb);
880 } else if (skb_is_gso(skb)) {
884 skb->ip_summed = CHECKSUM_PARTIAL;
885 /* specify the length of each IP datagram fragment */
886 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
887 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
890 return skb_append_datato_frags(sk, skb, getfrag, from,
891 (length - transhdrlen));
894 static int __ip_append_data(struct sock *sk,
896 struct sk_buff_head *queue,
897 struct inet_cork *cork,
898 struct page_frag *pfrag,
899 int getfrag(void *from, char *to, int offset,
900 int len, int odd, struct sk_buff *skb),
901 void *from, int length, int transhdrlen,
904 struct inet_sock *inet = inet_sk(sk);
907 struct ip_options *opt = cork->opt;
914 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
915 int csummode = CHECKSUM_NONE;
916 struct rtable *rt = (struct rtable *)cork->dst;
919 skb = skb_peek_tail(queue);
921 exthdrlen = !skb ? rt->dst.header_len : 0;
922 mtu = cork->fragsize;
923 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
924 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
925 tskey = sk->sk_tskey++;
927 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
929 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
930 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
931 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
933 if (cork->length + length > maxnonfragsize - fragheaderlen) {
934 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
935 mtu - (opt ? opt->optlen : 0));
940 * transhdrlen > 0 means that this is the first fragment and we wish
941 * it won't be fragmented in the future.
944 length + fragheaderlen <= mtu &&
945 rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
946 !(flags & MSG_MORE) &&
948 csummode = CHECKSUM_PARTIAL;
950 cork->length += length;
951 if ((skb && skb_is_gso(skb)) ||
953 (skb_queue_len(queue) <= 1) &&
954 (sk->sk_protocol == IPPROTO_UDP) &&
955 (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len &&
956 (sk->sk_type == SOCK_DGRAM) && !sk->sk_no_check_tx)) {
957 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
958 hh_len, fragheaderlen, transhdrlen,
965 /* So, what's going on in the loop below?
967 * We use calculated fragment length to generate chained skb,
968 * each of segments is IP fragment ready for sending to network after
969 * adding appropriate IP header.
976 /* Check if the remaining data fits into current packet. */
977 copy = mtu - skb->len;
979 copy = maxfraglen - skb->len;
982 unsigned int datalen;
983 unsigned int fraglen;
984 unsigned int fraggap;
985 unsigned int alloclen;
986 struct sk_buff *skb_prev;
990 fraggap = skb_prev->len - maxfraglen;
995 * If remaining data exceeds the mtu,
996 * we know we need more fragment(s).
998 datalen = length + fraggap;
999 if (datalen > mtu - fragheaderlen)
1000 datalen = maxfraglen - fragheaderlen;
1001 fraglen = datalen + fragheaderlen;
1003 if ((flags & MSG_MORE) &&
1004 !(rt->dst.dev->features&NETIF_F_SG))
1009 alloclen += exthdrlen;
1011 /* The last fragment gets additional space at tail.
1012 * Note, with MSG_MORE we overallocate on fragments,
1013 * because we have no idea what fragment will be
1016 if (datalen == length + fraggap)
1017 alloclen += rt->dst.trailer_len;
1020 skb = sock_alloc_send_skb(sk,
1021 alloclen + hh_len + 15,
1022 (flags & MSG_DONTWAIT), &err);
1025 if (atomic_read(&sk->sk_wmem_alloc) <=
1027 skb = sock_wmalloc(sk,
1028 alloclen + hh_len + 15, 1,
1037 * Fill in the control structures
1039 skb->ip_summed = csummode;
1041 skb_reserve(skb, hh_len);
1043 /* only the initial fragment is time stamped */
1044 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1046 skb_shinfo(skb)->tskey = tskey;
1050 * Find where to start putting bytes.
1052 data = skb_put(skb, fraglen + exthdrlen);
1053 skb_set_network_header(skb, exthdrlen);
1054 skb->transport_header = (skb->network_header +
1056 data += fragheaderlen + exthdrlen;
1059 skb->csum = skb_copy_and_csum_bits(
1060 skb_prev, maxfraglen,
1061 data + transhdrlen, fraggap, 0);
1062 skb_prev->csum = csum_sub(skb_prev->csum,
1065 pskb_trim_unique(skb_prev, maxfraglen);
1068 copy = datalen - transhdrlen - fraggap;
1069 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1076 length -= datalen - fraggap;
1079 csummode = CHECKSUM_NONE;
1082 * Put the packet on the pending queue.
1084 __skb_queue_tail(queue, skb);
1091 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1092 skb_tailroom(skb) >= copy) {
1096 if (getfrag(from, skb_put(skb, copy),
1097 offset, copy, off, skb) < 0) {
1098 __skb_trim(skb, off);
1103 int i = skb_shinfo(skb)->nr_frags;
1106 if (!sk_page_frag_refill(sk, pfrag))
1109 if (!skb_can_coalesce(skb, i, pfrag->page,
1112 if (i == MAX_SKB_FRAGS)
1115 __skb_fill_page_desc(skb, i, pfrag->page,
1117 skb_shinfo(skb)->nr_frags = ++i;
1118 get_page(pfrag->page);
1120 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1122 page_address(pfrag->page) + pfrag->offset,
1123 offset, copy, skb->len, skb) < 0)
1126 pfrag->offset += copy;
1127 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1129 skb->data_len += copy;
1130 skb->truesize += copy;
1131 atomic_add(copy, &sk->sk_wmem_alloc);
1142 cork->length -= length;
1143 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1147 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1148 struct ipcm_cookie *ipc, struct rtable **rtp)
1150 struct ip_options_rcu *opt;
1154 * setup for corking.
1159 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1161 if (unlikely(!cork->opt))
1164 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1165 cork->flags |= IPCORK_OPT;
1166 cork->addr = ipc->addr;
1172 cork->fragsize = ip_sk_use_pmtu(sk) ?
1173 dst_mtu(&rt->dst) : READ_ONCE(rt->dst.dev->mtu);
1175 if (!inetdev_valid_mtu(cork->fragsize))
1176 return -ENETUNREACH;
1178 cork->dst = &rt->dst;
1179 /* We stole this route, caller should not release it. */
1183 cork->ttl = ipc->ttl;
1184 cork->tos = ipc->tos;
1185 cork->priority = ipc->priority;
1186 cork->tx_flags = ipc->tx_flags;
1192 * ip_append_data() and ip_append_page() can make one large IP datagram
1193 * from many pieces of data. Each pieces will be holded on the socket
1194 * until ip_push_pending_frames() is called. Each piece can be a page
1197 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1198 * this interface potentially.
1200 * LATER: length must be adjusted by pad at tail, when it is required.
1202 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1203 int getfrag(void *from, char *to, int offset, int len,
1204 int odd, struct sk_buff *skb),
1205 void *from, int length, int transhdrlen,
1206 struct ipcm_cookie *ipc, struct rtable **rtp,
1209 struct inet_sock *inet = inet_sk(sk);
1212 if (flags&MSG_PROBE)
1215 if (skb_queue_empty(&sk->sk_write_queue)) {
1216 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1223 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1224 sk_page_frag(sk), getfrag,
1225 from, length, transhdrlen, flags);
1228 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1229 int offset, size_t size, int flags)
1231 struct inet_sock *inet = inet_sk(sk);
1232 struct sk_buff *skb;
1234 struct ip_options *opt = NULL;
1235 struct inet_cork *cork;
1240 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1245 if (flags&MSG_PROBE)
1248 if (skb_queue_empty(&sk->sk_write_queue))
1251 cork = &inet->cork.base;
1252 rt = (struct rtable *)cork->dst;
1253 if (cork->flags & IPCORK_OPT)
1256 if (!(rt->dst.dev->features&NETIF_F_SG))
1259 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1260 mtu = cork->fragsize;
1262 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1263 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1264 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1266 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1267 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1268 mtu - (opt ? opt->optlen : 0));
1272 skb = skb_peek_tail(&sk->sk_write_queue);
1276 if ((size + skb->len > mtu) &&
1277 (skb_queue_len(&sk->sk_write_queue) == 1) &&
1278 (sk->sk_protocol == IPPROTO_UDP) &&
1279 (rt->dst.dev->features & NETIF_F_UFO)) {
1280 if (skb->ip_summed != CHECKSUM_PARTIAL)
1283 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1284 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1286 cork->length += size;
1289 if (skb_is_gso(skb)) {
1293 /* Check if the remaining data fits into current packet. */
1294 len = mtu - skb->len;
1296 len = maxfraglen - skb->len;
1299 struct sk_buff *skb_prev;
1303 fraggap = skb_prev->len - maxfraglen;
1305 alloclen = fragheaderlen + hh_len + fraggap + 15;
1306 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1307 if (unlikely(!skb)) {
1313 * Fill in the control structures
1315 skb->ip_summed = CHECKSUM_NONE;
1317 skb_reserve(skb, hh_len);
1320 * Find where to start putting bytes.
1322 skb_put(skb, fragheaderlen + fraggap);
1323 skb_reset_network_header(skb);
1324 skb->transport_header = (skb->network_header +
1327 skb->csum = skb_copy_and_csum_bits(skb_prev,
1329 skb_transport_header(skb),
1331 skb_prev->csum = csum_sub(skb_prev->csum,
1333 pskb_trim_unique(skb_prev, maxfraglen);
1337 * Put the packet on the pending queue.
1339 __skb_queue_tail(&sk->sk_write_queue, skb);
1346 if (skb_append_pagefrags(skb, page, offset, len)) {
1351 if (skb->ip_summed == CHECKSUM_NONE) {
1353 csum = csum_page(page, offset, len);
1354 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1358 skb->data_len += len;
1359 skb->truesize += len;
1360 atomic_add(len, &sk->sk_wmem_alloc);
1367 cork->length -= size;
1368 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1372 static void ip_cork_release(struct inet_cork *cork)
1374 cork->flags &= ~IPCORK_OPT;
1377 dst_release(cork->dst);
1382 * Combined all pending IP fragments on the socket as one IP datagram
1383 * and push them out.
1385 struct sk_buff *__ip_make_skb(struct sock *sk,
1387 struct sk_buff_head *queue,
1388 struct inet_cork *cork)
1390 struct sk_buff *skb, *tmp_skb;
1391 struct sk_buff **tail_skb;
1392 struct inet_sock *inet = inet_sk(sk);
1393 struct net *net = sock_net(sk);
1394 struct ip_options *opt = NULL;
1395 struct rtable *rt = (struct rtable *)cork->dst;
1400 skb = __skb_dequeue(queue);
1403 tail_skb = &(skb_shinfo(skb)->frag_list);
1405 /* move skb->data to ip header from ext header */
1406 if (skb->data < skb_network_header(skb))
1407 __skb_pull(skb, skb_network_offset(skb));
1408 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1409 __skb_pull(tmp_skb, skb_network_header_len(skb));
1410 *tail_skb = tmp_skb;
1411 tail_skb = &(tmp_skb->next);
1412 skb->len += tmp_skb->len;
1413 skb->data_len += tmp_skb->len;
1414 skb->truesize += tmp_skb->truesize;
1415 tmp_skb->destructor = NULL;
1419 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1420 * to fragment the frame generated here. No matter, what transforms
1421 * how transforms change size of the packet, it will come out.
1423 skb->ignore_df = ip_sk_ignore_df(sk);
1425 /* DF bit is set when we want to see DF on outgoing frames.
1426 * If ignore_df is set too, we still allow to fragment this frame
1428 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1429 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1430 (skb->len <= dst_mtu(&rt->dst) &&
1431 ip_dont_fragment(sk, &rt->dst)))
1434 if (cork->flags & IPCORK_OPT)
1439 else if (rt->rt_type == RTN_MULTICAST)
1442 ttl = ip_select_ttl(inet, &rt->dst);
1447 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1450 iph->protocol = sk->sk_protocol;
1451 ip_copy_addrs(iph, fl4);
1452 ip_select_ident(net, skb, sk);
1455 iph->ihl += opt->optlen>>2;
1456 ip_options_build(skb, opt, cork->addr, rt, 0);
1459 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1460 skb->mark = sk->sk_mark;
1462 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1466 skb_dst_set(skb, &rt->dst);
1468 if (iph->protocol == IPPROTO_ICMP)
1469 icmp_out_count(net, ((struct icmphdr *)
1470 skb_transport_header(skb))->type);
1472 ip_cork_release(cork);
1477 int ip_send_skb(struct net *net, struct sk_buff *skb)
1481 err = ip_local_out(net, skb->sk, skb);
1484 err = net_xmit_errno(err);
1486 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1492 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1494 struct sk_buff *skb;
1496 skb = ip_finish_skb(sk, fl4);
1500 /* Netfilter gets whole the not fragmented skb. */
1501 return ip_send_skb(sock_net(sk), skb);
1505 * Throw away all pending data on the socket.
1507 static void __ip_flush_pending_frames(struct sock *sk,
1508 struct sk_buff_head *queue,
1509 struct inet_cork *cork)
1511 struct sk_buff *skb;
1513 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1516 ip_cork_release(cork);
1519 void ip_flush_pending_frames(struct sock *sk)
1521 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1524 struct sk_buff *ip_make_skb(struct sock *sk,
1526 int getfrag(void *from, char *to, int offset,
1527 int len, int odd, struct sk_buff *skb),
1528 void *from, int length, int transhdrlen,
1529 struct ipcm_cookie *ipc, struct rtable **rtp,
1532 struct inet_cork cork;
1533 struct sk_buff_head queue;
1536 if (flags & MSG_PROBE)
1539 __skb_queue_head_init(&queue);
1544 err = ip_setup_cork(sk, &cork, ipc, rtp);
1546 return ERR_PTR(err);
1548 err = __ip_append_data(sk, fl4, &queue, &cork,
1549 ¤t->task_frag, getfrag,
1550 from, length, transhdrlen, flags);
1552 __ip_flush_pending_frames(sk, &queue, &cork);
1553 return ERR_PTR(err);
1556 return __ip_make_skb(sk, fl4, &queue, &cork);
1560 * Fetch data from kernel space and fill in checksum if needed.
1562 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1563 int len, int odd, struct sk_buff *skb)
1567 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1568 skb->csum = csum_block_add(skb->csum, csum, odd);
1573 * Generic function to send a packet as reply to another packet.
1574 * Used to send some TCP resets/acks so far.
1576 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1577 const struct ip_options *sopt,
1578 __be32 daddr, __be32 saddr,
1579 const struct ip_reply_arg *arg,
1582 struct ip_options_data replyopts;
1583 struct ipcm_cookie ipc;
1585 struct rtable *rt = skb_rtable(skb);
1586 struct net *net = sock_net(sk);
1587 struct sk_buff *nskb;
1591 if (__ip_options_echo(&replyopts.opt.opt, skb, sopt))
1600 if (replyopts.opt.opt.optlen) {
1601 ipc.opt = &replyopts.opt;
1603 if (replyopts.opt.opt.srr)
1604 daddr = replyopts.opt.opt.faddr;
1607 oif = arg->bound_dev_if;
1608 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1611 flowi4_init_output(&fl4, oif,
1612 IP4_REPLY_MARK(net, skb->mark),
1614 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1615 ip_reply_arg_flowi_flags(arg),
1617 tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1618 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1619 rt = ip_route_output_key(net, &fl4);
1623 inet_sk(sk)->tos = arg->tos & ~INET_ECN_MASK;
1625 sk->sk_priority = skb->priority;
1626 sk->sk_protocol = ip_hdr(skb)->protocol;
1627 sk->sk_bound_dev_if = arg->bound_dev_if;
1628 sk->sk_sndbuf = sysctl_wmem_default;
1629 sk->sk_mark = fl4.flowi4_mark;
1630 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1631 len, 0, &ipc, &rt, MSG_DONTWAIT);
1632 if (unlikely(err)) {
1633 ip_flush_pending_frames(sk);
1637 nskb = skb_peek(&sk->sk_write_queue);
1639 if (arg->csumoffset >= 0)
1640 *((__sum16 *)skb_transport_header(nskb) +
1641 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1643 nskb->ip_summed = CHECKSUM_NONE;
1644 ip_push_pending_frames(sk, &fl4);
1650 void __init ip_init(void)
1655 #if defined(CONFIG_IP_MULTICAST)