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 if (ip_dont_fragment(sk, &rt->dst)) {
163 iph->frag_off = htons(IP_DF);
167 __ip_select_ident(net, iph, 1);
170 if (opt && opt->opt.optlen) {
171 iph->ihl += opt->opt.optlen>>2;
172 ip_options_build(skb, &opt->opt, daddr, rt, 0);
175 skb->priority = sk->sk_priority;
176 skb->mark = sk->sk_mark;
179 return ip_local_out(net, skb->sk, skb);
181 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
183 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
185 struct dst_entry *dst = skb_dst(skb);
186 struct rtable *rt = (struct rtable *)dst;
187 struct net_device *dev = dst->dev;
188 unsigned int hh_len = LL_RESERVED_SPACE(dev);
189 struct neighbour *neigh;
192 if (rt->rt_type == RTN_MULTICAST) {
193 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
194 } else if (rt->rt_type == RTN_BROADCAST)
195 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
197 /* Be paranoid, rather than too clever. */
198 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
199 struct sk_buff *skb2;
201 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
207 skb_set_owner_w(skb2, skb->sk);
212 if (lwtunnel_xmit_redirect(dst->lwtstate)) {
213 int res = lwtunnel_xmit(skb);
215 if (res < 0 || res == LWTUNNEL_XMIT_DONE)
220 nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
221 neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
222 if (unlikely(!neigh))
223 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
224 if (!IS_ERR(neigh)) {
225 int res = dst_neigh_output(dst, neigh, skb);
227 rcu_read_unlock_bh();
230 rcu_read_unlock_bh();
232 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
238 static int ip_finish_output_gso(struct net *net, struct sock *sk,
239 struct sk_buff *skb, unsigned int mtu)
241 netdev_features_t features;
242 struct sk_buff *segs;
245 /* common case: seglen is <= mtu
247 if (skb_gso_validate_mtu(skb, mtu))
248 return ip_finish_output2(net, sk, skb);
250 /* Slowpath - GSO segment length exceeds the egress MTU.
252 * This can happen in several cases:
253 * - Forwarding of a TCP GRO skb, when DF flag is not set.
254 * - Forwarding of an skb that arrived on a virtualization interface
255 * (virtio-net/vhost/tap) with TSO/GSO size set by other network
257 * - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
258 * interface with a smaller MTU.
259 * - Arriving GRO skb (or GSO skb in a virtualized environment) that is
260 * bridged to a NETIF_F_TSO tunnel stacked over an interface with an
263 features = netif_skb_features(skb);
264 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_SGO_CB_OFFSET);
265 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
266 if (IS_ERR_OR_NULL(segs)) {
274 struct sk_buff *nskb = segs->next;
278 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
288 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
292 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
293 /* Policy lookup after SNAT yielded a new policy */
294 if (skb_dst(skb)->xfrm) {
295 IPCB(skb)->flags |= IPSKB_REROUTED;
296 return dst_output(net, sk, skb);
299 mtu = ip_skb_dst_mtu(sk, skb);
301 return ip_finish_output_gso(net, sk, skb, mtu);
303 if (skb->len > mtu || IPCB(skb)->frag_max_size)
304 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
306 return ip_finish_output2(net, sk, skb);
309 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
311 struct rtable *rt = skb_rtable(skb);
312 struct net_device *dev = rt->dst.dev;
315 * If the indicated interface is up and running, send the packet.
317 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
320 skb->protocol = htons(ETH_P_IP);
323 * Multicasts are looped back for other local users
326 if (rt->rt_flags&RTCF_MULTICAST) {
328 #ifdef CONFIG_IP_MROUTE
329 /* Small optimization: do not loopback not local frames,
330 which returned after forwarding; they will be dropped
331 by ip_mr_input in any case.
332 Note, that local frames are looped back to be delivered
335 This check is duplicated in ip_mr_input at the moment.
338 ((rt->rt_flags & RTCF_LOCAL) ||
339 !(IPCB(skb)->flags & IPSKB_FORWARDED))
342 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
344 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
345 net, sk, newskb, NULL, newskb->dev,
349 /* Multicasts with ttl 0 must not go beyond the host */
351 if (ip_hdr(skb)->ttl == 0) {
357 if (rt->rt_flags&RTCF_BROADCAST) {
358 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
360 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
361 net, sk, newskb, NULL, newskb->dev,
365 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
366 net, sk, skb, NULL, skb->dev,
368 !(IPCB(skb)->flags & IPSKB_REROUTED));
371 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
373 struct net_device *dev = skb_dst(skb)->dev;
375 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
378 skb->protocol = htons(ETH_P_IP);
380 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
381 net, sk, skb, NULL, dev,
383 !(IPCB(skb)->flags & IPSKB_REROUTED));
387 * copy saddr and daddr, possibly using 64bit load/stores
389 * iph->saddr = fl4->saddr;
390 * iph->daddr = fl4->daddr;
392 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
394 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
395 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
397 iph->saddr = fl4->saddr;
398 iph->daddr = fl4->daddr;
401 /* Note: skb->sk can be different from sk, in case of tunnels */
402 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
404 struct inet_sock *inet = inet_sk(sk);
405 struct net *net = sock_net(sk);
406 struct ip_options_rcu *inet_opt;
412 /* Skip all of this if the packet is already routed,
413 * f.e. by something like SCTP.
416 inet_opt = rcu_dereference(inet->inet_opt);
418 rt = skb_rtable(skb);
422 /* Make sure we can route this packet. */
423 rt = (struct rtable *)__sk_dst_check(sk, 0);
427 /* Use correct destination address if we have options. */
428 daddr = inet->inet_daddr;
429 if (inet_opt && inet_opt->opt.srr)
430 daddr = inet_opt->opt.faddr;
432 /* If this fails, retransmit mechanism of transport layer will
433 * keep trying until route appears or the connection times
436 rt = ip_route_output_ports(net, fl4, sk,
437 daddr, inet->inet_saddr,
442 sk->sk_bound_dev_if);
445 sk_setup_caps(sk, &rt->dst);
447 skb_dst_set_noref(skb, &rt->dst);
450 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
453 /* OK, we know where to send it, allocate and build IP header. */
454 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
455 skb_reset_network_header(skb);
457 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
458 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
459 iph->frag_off = htons(IP_DF);
462 iph->ttl = ip_select_ttl(inet, &rt->dst);
463 iph->protocol = sk->sk_protocol;
464 ip_copy_addrs(iph, fl4);
466 /* Transport layer set skb->h.foo itself. */
468 if (inet_opt && inet_opt->opt.optlen) {
469 iph->ihl += inet_opt->opt.optlen >> 2;
470 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
473 ip_select_ident_segs(net, skb, sk,
474 skb_shinfo(skb)->gso_segs ?: 1);
476 /* TODO : should we use skb->sk here instead of sk ? */
477 skb->priority = sk->sk_priority;
478 skb->mark = sk->sk_mark;
480 res = ip_local_out(net, sk, skb);
486 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
488 return -EHOSTUNREACH;
490 EXPORT_SYMBOL(ip_queue_xmit);
492 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
494 to->pkt_type = from->pkt_type;
495 to->priority = from->priority;
496 to->protocol = from->protocol;
497 to->skb_iif = from->skb_iif;
499 skb_dst_copy(to, from);
501 to->mark = from->mark;
503 skb_copy_hash(to, from);
505 /* Copy the flags to each fragment. */
506 IPCB(to)->flags = IPCB(from)->flags;
508 #ifdef CONFIG_NET_SCHED
509 to->tc_index = from->tc_index;
512 #if IS_ENABLED(CONFIG_IP_VS)
513 to->ipvs_property = from->ipvs_property;
515 skb_copy_secmark(to, from);
518 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
520 int (*output)(struct net *, struct sock *, struct sk_buff *))
522 struct iphdr *iph = ip_hdr(skb);
524 if ((iph->frag_off & htons(IP_DF)) == 0)
525 return ip_do_fragment(net, sk, skb, output);
527 if (unlikely(!skb->ignore_df ||
528 (IPCB(skb)->frag_max_size &&
529 IPCB(skb)->frag_max_size > mtu))) {
530 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
531 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
537 return ip_do_fragment(net, sk, skb, output);
541 * This IP datagram is too large to be sent in one piece. Break it up into
542 * smaller pieces (each of size equal to IP header plus
543 * a block of the data of the original IP data part) that will yet fit in a
544 * single device frame, and queue such a frame for sending.
547 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
548 int (*output)(struct net *, struct sock *, struct sk_buff *))
552 struct sk_buff *skb2;
553 unsigned int mtu, hlen, left, len, ll_rs;
555 __be16 not_last_frag;
556 struct rtable *rt = skb_rtable(skb);
559 /* for offloaded checksums cleanup checksum before fragmentation */
560 if (skb->ip_summed == CHECKSUM_PARTIAL &&
561 (err = skb_checksum_help(skb)))
565 * Point into the IP datagram header.
570 mtu = ip_skb_dst_mtu(sk, skb);
571 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
572 mtu = IPCB(skb)->frag_max_size;
575 * Setup starting values.
579 mtu = mtu - hlen; /* Size of data space */
580 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
582 /* When frag_list is given, use it. First, check its validity:
583 * some transformers could create wrong frag_list or break existing
584 * one, it is not prohibited. In this case fall back to copying.
586 * LATER: this step can be merged to real generation of fragments,
587 * we can switch to copy when see the first bad fragment.
589 if (skb_has_frag_list(skb)) {
590 struct sk_buff *frag, *frag2;
591 int first_len = skb_pagelen(skb);
593 if (first_len - hlen > mtu ||
594 ((first_len - hlen) & 7) ||
595 ip_is_fragment(iph) ||
599 skb_walk_frags(skb, frag) {
600 /* Correct geometry. */
601 if (frag->len > mtu ||
602 ((frag->len & 7) && frag->next) ||
603 skb_headroom(frag) < hlen)
604 goto slow_path_clean;
606 /* Partially cloned skb? */
607 if (skb_shared(frag))
608 goto slow_path_clean;
613 frag->destructor = sock_wfree;
615 skb->truesize -= frag->truesize;
618 /* Everything is OK. Generate! */
622 frag = skb_shinfo(skb)->frag_list;
623 skb_frag_list_init(skb);
624 skb->data_len = first_len - skb_headlen(skb);
625 skb->len = first_len;
626 iph->tot_len = htons(first_len);
627 iph->frag_off = htons(IP_MF);
631 /* Prepare header of the next frame,
632 * before previous one went down. */
634 frag->ip_summed = CHECKSUM_NONE;
635 skb_reset_transport_header(frag);
636 __skb_push(frag, hlen);
637 skb_reset_network_header(frag);
638 memcpy(skb_network_header(frag), iph, hlen);
640 iph->tot_len = htons(frag->len);
641 ip_copy_metadata(frag, skb);
643 ip_options_fragment(frag);
644 offset += skb->len - hlen;
645 iph->frag_off = htons(offset>>3);
647 iph->frag_off |= htons(IP_MF);
648 /* Ready, complete checksum */
652 err = output(net, sk, skb);
655 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
665 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
674 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
678 skb_walk_frags(skb, frag2) {
682 frag2->destructor = NULL;
683 skb->truesize += frag2->truesize;
690 left = skb->len - hlen; /* Space per frame */
691 ptr = hlen; /* Where to start from */
693 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
696 * Fragment the datagram.
699 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
700 not_last_frag = iph->frag_off & htons(IP_MF);
703 * Keep copying data until we run out.
708 /* IF: it doesn't fit, use 'mtu' - the data space left */
711 /* IF: we are not sending up to and including the packet end
712 then align the next start on an eight byte boundary */
717 /* Allocate buffer */
718 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
725 * Set up data on packet
728 ip_copy_metadata(skb2, skb);
729 skb_reserve(skb2, ll_rs);
730 skb_put(skb2, len + hlen);
731 skb_reset_network_header(skb2);
732 skb2->transport_header = skb2->network_header + hlen;
735 * Charge the memory for the fragment to any owner
740 skb_set_owner_w(skb2, skb->sk);
743 * Copy the packet header into the new buffer.
746 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
749 * Copy a block of the IP datagram.
751 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
756 * Fill in the new header fields.
759 iph->frag_off = htons((offset >> 3));
761 if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
762 iph->frag_off |= htons(IP_DF);
764 /* ANK: dirty, but effective trick. Upgrade options only if
765 * the segment to be fragmented was THE FIRST (otherwise,
766 * options are already fixed) and make it ONCE
767 * on the initial skb, so that all the following fragments
768 * will inherit fixed options.
771 ip_options_fragment(skb);
774 * Added AC : If we are fragmenting a fragment that's not the
775 * last fragment then keep MF on each bit
777 if (left > 0 || not_last_frag)
778 iph->frag_off |= htons(IP_MF);
783 * Put this fragment into the sending queue.
785 iph->tot_len = htons(len + hlen);
789 err = output(net, sk, skb2);
793 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
796 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
801 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
804 EXPORT_SYMBOL(ip_do_fragment);
807 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
809 struct msghdr *msg = from;
811 if (skb->ip_summed == CHECKSUM_PARTIAL) {
812 if (copy_from_iter(to, len, &msg->msg_iter) != len)
816 if (csum_and_copy_from_iter(to, len, &csum, &msg->msg_iter) != len)
818 skb->csum = csum_block_add(skb->csum, csum, odd);
822 EXPORT_SYMBOL(ip_generic_getfrag);
825 csum_page(struct page *page, int offset, int copy)
830 csum = csum_partial(kaddr + offset, copy, 0);
835 static inline int ip_ufo_append_data(struct sock *sk,
836 struct sk_buff_head *queue,
837 int getfrag(void *from, char *to, int offset, int len,
838 int odd, struct sk_buff *skb),
839 void *from, int length, int hh_len, int fragheaderlen,
840 int transhdrlen, int maxfraglen, unsigned int flags)
845 /* There is support for UDP fragmentation offload by network
846 * device, so create one single skb packet containing complete
849 skb = skb_peek_tail(queue);
851 skb = sock_alloc_send_skb(sk,
852 hh_len + fragheaderlen + transhdrlen + 20,
853 (flags & MSG_DONTWAIT), &err);
858 /* reserve space for Hardware header */
859 skb_reserve(skb, hh_len);
861 /* create space for UDP/IP header */
862 skb_put(skb, fragheaderlen + transhdrlen);
864 /* initialize network header pointer */
865 skb_reset_network_header(skb);
867 /* initialize protocol header pointer */
868 skb->transport_header = skb->network_header + fragheaderlen;
872 __skb_queue_tail(queue, skb);
873 } else if (skb_is_gso(skb)) {
877 skb->ip_summed = CHECKSUM_PARTIAL;
878 /* specify the length of each IP datagram fragment */
879 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
880 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
883 return skb_append_datato_frags(sk, skb, getfrag, from,
884 (length - transhdrlen));
887 static int __ip_append_data(struct sock *sk,
889 struct sk_buff_head *queue,
890 struct inet_cork *cork,
891 struct page_frag *pfrag,
892 int getfrag(void *from, char *to, int offset,
893 int len, int odd, struct sk_buff *skb),
894 void *from, int length, int transhdrlen,
897 struct inet_sock *inet = inet_sk(sk);
900 struct ip_options *opt = cork->opt;
907 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
908 int csummode = CHECKSUM_NONE;
909 struct rtable *rt = (struct rtable *)cork->dst;
912 skb = skb_peek_tail(queue);
914 exthdrlen = !skb ? rt->dst.header_len : 0;
915 mtu = cork->fragsize;
916 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
917 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
918 tskey = sk->sk_tskey++;
920 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
922 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
923 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
924 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
926 if (cork->length + length > maxnonfragsize - fragheaderlen) {
927 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
928 mtu - (opt ? opt->optlen : 0));
933 * transhdrlen > 0 means that this is the first fragment and we wish
934 * it won't be fragmented in the future.
937 length + fragheaderlen <= mtu &&
938 rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
939 !(flags & MSG_MORE) &&
941 csummode = CHECKSUM_PARTIAL;
943 cork->length += length;
944 if ((skb && skb_is_gso(skb)) ||
946 (skb_queue_len(queue) <= 1) &&
947 (sk->sk_protocol == IPPROTO_UDP) &&
948 (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len &&
949 (sk->sk_type == SOCK_DGRAM) && !sk->sk_no_check_tx)) {
950 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
951 hh_len, fragheaderlen, transhdrlen,
958 /* So, what's going on in the loop below?
960 * We use calculated fragment length to generate chained skb,
961 * each of segments is IP fragment ready for sending to network after
962 * adding appropriate IP header.
969 /* Check if the remaining data fits into current packet. */
970 copy = mtu - skb->len;
972 copy = maxfraglen - skb->len;
975 unsigned int datalen;
976 unsigned int fraglen;
977 unsigned int fraggap;
978 unsigned int alloclen;
979 struct sk_buff *skb_prev;
983 fraggap = skb_prev->len - maxfraglen;
988 * If remaining data exceeds the mtu,
989 * we know we need more fragment(s).
991 datalen = length + fraggap;
992 if (datalen > mtu - fragheaderlen)
993 datalen = maxfraglen - fragheaderlen;
994 fraglen = datalen + fragheaderlen;
996 if ((flags & MSG_MORE) &&
997 !(rt->dst.dev->features&NETIF_F_SG))
1002 alloclen += exthdrlen;
1004 /* The last fragment gets additional space at tail.
1005 * Note, with MSG_MORE we overallocate on fragments,
1006 * because we have no idea what fragment will be
1009 if (datalen == length + fraggap)
1010 alloclen += rt->dst.trailer_len;
1013 skb = sock_alloc_send_skb(sk,
1014 alloclen + hh_len + 15,
1015 (flags & MSG_DONTWAIT), &err);
1018 if (atomic_read(&sk->sk_wmem_alloc) <=
1020 skb = sock_wmalloc(sk,
1021 alloclen + hh_len + 15, 1,
1030 * Fill in the control structures
1032 skb->ip_summed = csummode;
1034 skb_reserve(skb, hh_len);
1036 /* only the initial fragment is time stamped */
1037 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1039 skb_shinfo(skb)->tskey = tskey;
1043 * Find where to start putting bytes.
1045 data = skb_put(skb, fraglen + exthdrlen);
1046 skb_set_network_header(skb, exthdrlen);
1047 skb->transport_header = (skb->network_header +
1049 data += fragheaderlen + exthdrlen;
1052 skb->csum = skb_copy_and_csum_bits(
1053 skb_prev, maxfraglen,
1054 data + transhdrlen, fraggap, 0);
1055 skb_prev->csum = csum_sub(skb_prev->csum,
1058 pskb_trim_unique(skb_prev, maxfraglen);
1061 copy = datalen - transhdrlen - fraggap;
1062 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1069 length -= datalen - fraggap;
1072 csummode = CHECKSUM_NONE;
1075 * Put the packet on the pending queue.
1077 __skb_queue_tail(queue, skb);
1084 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1085 skb_tailroom(skb) >= copy) {
1089 if (getfrag(from, skb_put(skb, copy),
1090 offset, copy, off, skb) < 0) {
1091 __skb_trim(skb, off);
1096 int i = skb_shinfo(skb)->nr_frags;
1099 if (!sk_page_frag_refill(sk, pfrag))
1102 if (!skb_can_coalesce(skb, i, pfrag->page,
1105 if (i == MAX_SKB_FRAGS)
1108 __skb_fill_page_desc(skb, i, pfrag->page,
1110 skb_shinfo(skb)->nr_frags = ++i;
1111 get_page(pfrag->page);
1113 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1115 page_address(pfrag->page) + pfrag->offset,
1116 offset, copy, skb->len, skb) < 0)
1119 pfrag->offset += copy;
1120 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1122 skb->data_len += copy;
1123 skb->truesize += copy;
1124 atomic_add(copy, &sk->sk_wmem_alloc);
1135 cork->length -= length;
1136 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1140 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1141 struct ipcm_cookie *ipc, struct rtable **rtp)
1143 struct ip_options_rcu *opt;
1147 * setup for corking.
1152 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1154 if (unlikely(!cork->opt))
1157 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1158 cork->flags |= IPCORK_OPT;
1159 cork->addr = ipc->addr;
1165 cork->fragsize = ip_sk_use_pmtu(sk) ?
1166 dst_mtu(&rt->dst) : READ_ONCE(rt->dst.dev->mtu);
1168 if (!inetdev_valid_mtu(cork->fragsize))
1169 return -ENETUNREACH;
1171 cork->dst = &rt->dst;
1172 /* We stole this route, caller should not release it. */
1176 cork->ttl = ipc->ttl;
1177 cork->tos = ipc->tos;
1178 cork->priority = ipc->priority;
1179 cork->tx_flags = ipc->tx_flags;
1185 * ip_append_data() and ip_append_page() can make one large IP datagram
1186 * from many pieces of data. Each pieces will be holded on the socket
1187 * until ip_push_pending_frames() is called. Each piece can be a page
1190 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1191 * this interface potentially.
1193 * LATER: length must be adjusted by pad at tail, when it is required.
1195 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1196 int getfrag(void *from, char *to, int offset, int len,
1197 int odd, struct sk_buff *skb),
1198 void *from, int length, int transhdrlen,
1199 struct ipcm_cookie *ipc, struct rtable **rtp,
1202 struct inet_sock *inet = inet_sk(sk);
1205 if (flags&MSG_PROBE)
1208 if (skb_queue_empty(&sk->sk_write_queue)) {
1209 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1216 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1217 sk_page_frag(sk), getfrag,
1218 from, length, transhdrlen, flags);
1221 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1222 int offset, size_t size, int flags)
1224 struct inet_sock *inet = inet_sk(sk);
1225 struct sk_buff *skb;
1227 struct ip_options *opt = NULL;
1228 struct inet_cork *cork;
1233 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1238 if (flags&MSG_PROBE)
1241 if (skb_queue_empty(&sk->sk_write_queue))
1244 cork = &inet->cork.base;
1245 rt = (struct rtable *)cork->dst;
1246 if (cork->flags & IPCORK_OPT)
1249 if (!(rt->dst.dev->features&NETIF_F_SG))
1252 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1253 mtu = cork->fragsize;
1255 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1256 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1257 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1259 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1260 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1261 mtu - (opt ? opt->optlen : 0));
1265 skb = skb_peek_tail(&sk->sk_write_queue);
1269 if ((size + skb->len > mtu) &&
1270 (skb_queue_len(&sk->sk_write_queue) == 1) &&
1271 (sk->sk_protocol == IPPROTO_UDP) &&
1272 (rt->dst.dev->features & NETIF_F_UFO)) {
1273 if (skb->ip_summed != CHECKSUM_PARTIAL)
1276 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1277 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1279 cork->length += size;
1282 if (skb_is_gso(skb)) {
1286 /* Check if the remaining data fits into current packet. */
1287 len = mtu - skb->len;
1289 len = maxfraglen - skb->len;
1292 struct sk_buff *skb_prev;
1296 fraggap = skb_prev->len - maxfraglen;
1298 alloclen = fragheaderlen + hh_len + fraggap + 15;
1299 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1300 if (unlikely(!skb)) {
1306 * Fill in the control structures
1308 skb->ip_summed = CHECKSUM_NONE;
1310 skb_reserve(skb, hh_len);
1313 * Find where to start putting bytes.
1315 skb_put(skb, fragheaderlen + fraggap);
1316 skb_reset_network_header(skb);
1317 skb->transport_header = (skb->network_header +
1320 skb->csum = skb_copy_and_csum_bits(skb_prev,
1322 skb_transport_header(skb),
1324 skb_prev->csum = csum_sub(skb_prev->csum,
1326 pskb_trim_unique(skb_prev, maxfraglen);
1330 * Put the packet on the pending queue.
1332 __skb_queue_tail(&sk->sk_write_queue, skb);
1339 if (skb_append_pagefrags(skb, page, offset, len)) {
1344 if (skb->ip_summed == CHECKSUM_NONE) {
1346 csum = csum_page(page, offset, len);
1347 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1351 skb->data_len += len;
1352 skb->truesize += len;
1353 atomic_add(len, &sk->sk_wmem_alloc);
1360 cork->length -= size;
1361 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1365 static void ip_cork_release(struct inet_cork *cork)
1367 cork->flags &= ~IPCORK_OPT;
1370 dst_release(cork->dst);
1375 * Combined all pending IP fragments on the socket as one IP datagram
1376 * and push them out.
1378 struct sk_buff *__ip_make_skb(struct sock *sk,
1380 struct sk_buff_head *queue,
1381 struct inet_cork *cork)
1383 struct sk_buff *skb, *tmp_skb;
1384 struct sk_buff **tail_skb;
1385 struct inet_sock *inet = inet_sk(sk);
1386 struct net *net = sock_net(sk);
1387 struct ip_options *opt = NULL;
1388 struct rtable *rt = (struct rtable *)cork->dst;
1393 skb = __skb_dequeue(queue);
1396 tail_skb = &(skb_shinfo(skb)->frag_list);
1398 /* move skb->data to ip header from ext header */
1399 if (skb->data < skb_network_header(skb))
1400 __skb_pull(skb, skb_network_offset(skb));
1401 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1402 __skb_pull(tmp_skb, skb_network_header_len(skb));
1403 *tail_skb = tmp_skb;
1404 tail_skb = &(tmp_skb->next);
1405 skb->len += tmp_skb->len;
1406 skb->data_len += tmp_skb->len;
1407 skb->truesize += tmp_skb->truesize;
1408 tmp_skb->destructor = NULL;
1412 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1413 * to fragment the frame generated here. No matter, what transforms
1414 * how transforms change size of the packet, it will come out.
1416 skb->ignore_df = ip_sk_ignore_df(sk);
1418 /* DF bit is set when we want to see DF on outgoing frames.
1419 * If ignore_df is set too, we still allow to fragment this frame
1421 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1422 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1423 (skb->len <= dst_mtu(&rt->dst) &&
1424 ip_dont_fragment(sk, &rt->dst)))
1427 if (cork->flags & IPCORK_OPT)
1432 else if (rt->rt_type == RTN_MULTICAST)
1435 ttl = ip_select_ttl(inet, &rt->dst);
1440 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1443 iph->protocol = sk->sk_protocol;
1444 ip_copy_addrs(iph, fl4);
1445 ip_select_ident(net, skb, sk);
1448 iph->ihl += opt->optlen>>2;
1449 ip_options_build(skb, opt, cork->addr, rt, 0);
1452 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1453 skb->mark = sk->sk_mark;
1455 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1459 skb_dst_set(skb, &rt->dst);
1461 if (iph->protocol == IPPROTO_ICMP)
1462 icmp_out_count(net, ((struct icmphdr *)
1463 skb_transport_header(skb))->type);
1465 ip_cork_release(cork);
1470 int ip_send_skb(struct net *net, struct sk_buff *skb)
1474 err = ip_local_out(net, skb->sk, skb);
1477 err = net_xmit_errno(err);
1479 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1485 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1487 struct sk_buff *skb;
1489 skb = ip_finish_skb(sk, fl4);
1493 /* Netfilter gets whole the not fragmented skb. */
1494 return ip_send_skb(sock_net(sk), skb);
1498 * Throw away all pending data on the socket.
1500 static void __ip_flush_pending_frames(struct sock *sk,
1501 struct sk_buff_head *queue,
1502 struct inet_cork *cork)
1504 struct sk_buff *skb;
1506 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1509 ip_cork_release(cork);
1512 void ip_flush_pending_frames(struct sock *sk)
1514 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1517 struct sk_buff *ip_make_skb(struct sock *sk,
1519 int getfrag(void *from, char *to, int offset,
1520 int len, int odd, struct sk_buff *skb),
1521 void *from, int length, int transhdrlen,
1522 struct ipcm_cookie *ipc, struct rtable **rtp,
1525 struct inet_cork cork;
1526 struct sk_buff_head queue;
1529 if (flags & MSG_PROBE)
1532 __skb_queue_head_init(&queue);
1537 err = ip_setup_cork(sk, &cork, ipc, rtp);
1539 return ERR_PTR(err);
1541 err = __ip_append_data(sk, fl4, &queue, &cork,
1542 ¤t->task_frag, getfrag,
1543 from, length, transhdrlen, flags);
1545 __ip_flush_pending_frames(sk, &queue, &cork);
1546 return ERR_PTR(err);
1549 return __ip_make_skb(sk, fl4, &queue, &cork);
1553 * Fetch data from kernel space and fill in checksum if needed.
1555 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1556 int len, int odd, struct sk_buff *skb)
1560 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1561 skb->csum = csum_block_add(skb->csum, csum, odd);
1566 * Generic function to send a packet as reply to another packet.
1567 * Used to send some TCP resets/acks so far.
1569 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1570 const struct ip_options *sopt,
1571 __be32 daddr, __be32 saddr,
1572 const struct ip_reply_arg *arg,
1575 struct ip_options_data replyopts;
1576 struct ipcm_cookie ipc;
1578 struct rtable *rt = skb_rtable(skb);
1579 struct net *net = sock_net(sk);
1580 struct sk_buff *nskb;
1584 if (__ip_options_echo(&replyopts.opt.opt, skb, sopt))
1593 if (replyopts.opt.opt.optlen) {
1594 ipc.opt = &replyopts.opt;
1596 if (replyopts.opt.opt.srr)
1597 daddr = replyopts.opt.opt.faddr;
1600 oif = arg->bound_dev_if;
1601 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1604 flowi4_init_output(&fl4, oif,
1605 IP4_REPLY_MARK(net, skb->mark),
1607 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1608 ip_reply_arg_flowi_flags(arg),
1610 tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1611 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1612 rt = ip_route_output_key(net, &fl4);
1616 inet_sk(sk)->tos = arg->tos & ~INET_ECN_MASK;
1618 sk->sk_priority = skb->priority;
1619 sk->sk_protocol = ip_hdr(skb)->protocol;
1620 sk->sk_bound_dev_if = arg->bound_dev_if;
1621 sk->sk_sndbuf = sysctl_wmem_default;
1622 sk->sk_mark = fl4.flowi4_mark;
1623 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1624 len, 0, &ipc, &rt, MSG_DONTWAIT);
1625 if (unlikely(err)) {
1626 ip_flush_pending_frames(sk);
1630 nskb = skb_peek(&sk->sk_write_queue);
1632 if (arg->csumoffset >= 0)
1633 *((__sum16 *)skb_transport_header(nskb) +
1634 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1636 nskb->ip_summed = CHECKSUM_NONE;
1637 ip_push_pending_frames(sk, &fl4);
1643 void __init ip_init(void)
1648 #if defined(CONFIG_IP_MULTICAST)