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 <linux/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/bpf-cgroup.h>
79 #include <linux/igmp.h>
80 #include <linux/netfilter_ipv4.h>
81 #include <linux/netfilter_bridge.h>
82 #include <linux/netlink.h>
83 #include <linux/tcp.h>
86 ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
88 int (*output)(struct net *, struct sock *, struct sk_buff *));
90 /* Generate a checksum for an outgoing IP datagram. */
91 void ip_send_check(struct iphdr *iph)
94 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
96 EXPORT_SYMBOL(ip_send_check);
98 int __ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
100 struct iphdr *iph = ip_hdr(skb);
102 iph->tot_len = htons(skb->len);
105 /* if egress device is enslaved to an L3 master device pass the
106 * skb to its handler for processing
108 skb = l3mdev_ip_out(sk, skb);
112 skb->protocol = htons(ETH_P_IP);
114 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
115 net, sk, skb, NULL, skb_dst(skb)->dev,
119 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
123 err = __ip_local_out(net, sk, skb);
124 if (likely(err == 1))
125 err = dst_output(net, sk, skb);
129 EXPORT_SYMBOL_GPL(ip_local_out);
131 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
133 int ttl = inet->uc_ttl;
136 ttl = ip4_dst_hoplimit(dst);
141 * Add an ip header to a skbuff and send it out.
144 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
145 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
147 struct inet_sock *inet = inet_sk(sk);
148 struct rtable *rt = skb_rtable(skb);
149 struct net *net = sock_net(sk);
152 /* Build the IP header. */
153 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
154 skb_reset_network_header(skb);
158 iph->tos = inet->tos;
159 iph->ttl = ip_select_ttl(inet, &rt->dst);
160 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
162 iph->protocol = sk->sk_protocol;
163 if (ip_dont_fragment(sk, &rt->dst)) {
164 iph->frag_off = htons(IP_DF);
168 __ip_select_ident(net, iph, 1);
171 if (opt && opt->opt.optlen) {
172 iph->ihl += opt->opt.optlen>>2;
173 ip_options_build(skb, &opt->opt, daddr, rt, 0);
176 skb->priority = sk->sk_priority;
178 skb->mark = sk->sk_mark;
181 return ip_local_out(net, skb->sk, skb);
183 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
185 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
187 struct dst_entry *dst = skb_dst(skb);
188 struct rtable *rt = (struct rtable *)dst;
189 struct net_device *dev = dst->dev;
190 unsigned int hh_len = LL_RESERVED_SPACE(dev);
191 struct neighbour *neigh;
194 if (rt->rt_type == RTN_MULTICAST) {
195 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
196 } else if (rt->rt_type == RTN_BROADCAST)
197 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
199 /* Be paranoid, rather than too clever. */
200 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
201 struct sk_buff *skb2;
203 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
209 skb_set_owner_w(skb2, skb->sk);
214 if (lwtunnel_xmit_redirect(dst->lwtstate)) {
215 int res = lwtunnel_xmit(skb);
217 if (res < 0 || res == LWTUNNEL_XMIT_DONE)
222 nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
223 neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
224 if (unlikely(!neigh))
225 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
226 if (!IS_ERR(neigh)) {
229 sock_confirm_neigh(skb, neigh);
230 res = neigh_output(neigh, skb);
232 rcu_read_unlock_bh();
235 rcu_read_unlock_bh();
237 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
243 static int ip_finish_output_gso(struct net *net, struct sock *sk,
244 struct sk_buff *skb, unsigned int mtu)
246 netdev_features_t features;
247 struct sk_buff *segs;
250 /* common case: seglen is <= mtu
252 if (skb_gso_validate_mtu(skb, mtu))
253 return ip_finish_output2(net, sk, skb);
255 /* Slowpath - GSO segment length exceeds the egress MTU.
257 * This can happen in several cases:
258 * - Forwarding of a TCP GRO skb, when DF flag is not set.
259 * - Forwarding of an skb that arrived on a virtualization interface
260 * (virtio-net/vhost/tap) with TSO/GSO size set by other network
262 * - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
263 * interface with a smaller MTU.
264 * - Arriving GRO skb (or GSO skb in a virtualized environment) that is
265 * bridged to a NETIF_F_TSO tunnel stacked over an interface with an
268 features = netif_skb_features(skb);
269 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_SGO_CB_OFFSET);
270 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
271 if (IS_ERR_OR_NULL(segs)) {
279 struct sk_buff *nskb = segs->next;
283 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
293 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
298 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
304 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
305 /* Policy lookup after SNAT yielded a new policy */
306 if (skb_dst(skb)->xfrm) {
307 IPCB(skb)->flags |= IPSKB_REROUTED;
308 return dst_output(net, sk, skb);
311 mtu = ip_skb_dst_mtu(sk, skb);
313 return ip_finish_output_gso(net, sk, skb, mtu);
315 if (skb->len > mtu || IPCB(skb)->frag_max_size)
316 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
318 return ip_finish_output2(net, sk, skb);
321 static int ip_mc_finish_output(struct net *net, struct sock *sk,
326 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
332 return dev_loopback_xmit(net, sk, skb);
335 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
337 struct rtable *rt = skb_rtable(skb);
338 struct net_device *dev = rt->dst.dev;
341 * If the indicated interface is up and running, send the packet.
343 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
346 skb->protocol = htons(ETH_P_IP);
349 * Multicasts are looped back for other local users
352 if (rt->rt_flags&RTCF_MULTICAST) {
354 #ifdef CONFIG_IP_MROUTE
355 /* Small optimization: do not loopback not local frames,
356 which returned after forwarding; they will be dropped
357 by ip_mr_input in any case.
358 Note, that local frames are looped back to be delivered
361 This check is duplicated in ip_mr_input at the moment.
364 ((rt->rt_flags & RTCF_LOCAL) ||
365 !(IPCB(skb)->flags & IPSKB_FORWARDED))
368 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
370 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
371 net, sk, newskb, NULL, newskb->dev,
372 ip_mc_finish_output);
375 /* Multicasts with ttl 0 must not go beyond the host */
377 if (ip_hdr(skb)->ttl == 0) {
383 if (rt->rt_flags&RTCF_BROADCAST) {
384 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
386 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
387 net, sk, newskb, NULL, newskb->dev,
388 ip_mc_finish_output);
391 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
392 net, sk, skb, NULL, skb->dev,
394 !(IPCB(skb)->flags & IPSKB_REROUTED));
397 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
399 struct net_device *dev = skb_dst(skb)->dev;
401 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
404 skb->protocol = htons(ETH_P_IP);
406 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
407 net, sk, skb, NULL, dev,
409 !(IPCB(skb)->flags & IPSKB_REROUTED));
413 * copy saddr and daddr, possibly using 64bit load/stores
415 * iph->saddr = fl4->saddr;
416 * iph->daddr = fl4->daddr;
418 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
420 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
421 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
423 iph->saddr = fl4->saddr;
424 iph->daddr = fl4->daddr;
427 /* Note: skb->sk can be different from sk, in case of tunnels */
428 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
430 struct inet_sock *inet = inet_sk(sk);
431 struct net *net = sock_net(sk);
432 struct ip_options_rcu *inet_opt;
438 /* Skip all of this if the packet is already routed,
439 * f.e. by something like SCTP.
442 inet_opt = rcu_dereference(inet->inet_opt);
444 rt = skb_rtable(skb);
448 /* Make sure we can route this packet. */
449 rt = (struct rtable *)__sk_dst_check(sk, 0);
453 /* Use correct destination address if we have options. */
454 daddr = inet->inet_daddr;
455 if (inet_opt && inet_opt->opt.srr)
456 daddr = inet_opt->opt.faddr;
458 /* If this fails, retransmit mechanism of transport layer will
459 * keep trying until route appears or the connection times
462 rt = ip_route_output_ports(net, fl4, sk,
463 daddr, inet->inet_saddr,
468 sk->sk_bound_dev_if);
471 sk_setup_caps(sk, &rt->dst);
473 skb_dst_set_noref(skb, &rt->dst);
476 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
479 /* OK, we know where to send it, allocate and build IP header. */
480 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
481 skb_reset_network_header(skb);
483 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
484 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
485 iph->frag_off = htons(IP_DF);
488 iph->ttl = ip_select_ttl(inet, &rt->dst);
489 iph->protocol = sk->sk_protocol;
490 ip_copy_addrs(iph, fl4);
492 /* Transport layer set skb->h.foo itself. */
494 if (inet_opt && inet_opt->opt.optlen) {
495 iph->ihl += inet_opt->opt.optlen >> 2;
496 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
499 ip_select_ident_segs(net, skb, sk,
500 skb_shinfo(skb)->gso_segs ?: 1);
502 /* TODO : should we use skb->sk here instead of sk ? */
503 skb->priority = sk->sk_priority;
504 skb->mark = sk->sk_mark;
506 res = ip_local_out(net, sk, skb);
512 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
514 return -EHOSTUNREACH;
516 EXPORT_SYMBOL(ip_queue_xmit);
518 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
520 to->pkt_type = from->pkt_type;
521 to->priority = from->priority;
522 to->protocol = from->protocol;
523 to->skb_iif = from->skb_iif;
525 skb_dst_copy(to, from);
527 to->mark = from->mark;
529 skb_copy_hash(to, from);
531 /* Copy the flags to each fragment. */
532 IPCB(to)->flags = IPCB(from)->flags;
534 #ifdef CONFIG_NET_SCHED
535 to->tc_index = from->tc_index;
538 #if IS_ENABLED(CONFIG_IP_VS)
539 to->ipvs_property = from->ipvs_property;
541 skb_copy_secmark(to, from);
544 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
546 int (*output)(struct net *, struct sock *, struct sk_buff *))
548 struct iphdr *iph = ip_hdr(skb);
550 if ((iph->frag_off & htons(IP_DF)) == 0)
551 return ip_do_fragment(net, sk, skb, output);
553 if (unlikely(!skb->ignore_df ||
554 (IPCB(skb)->frag_max_size &&
555 IPCB(skb)->frag_max_size > mtu))) {
556 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
557 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
563 return ip_do_fragment(net, sk, skb, output);
567 * This IP datagram is too large to be sent in one piece. Break it up into
568 * smaller pieces (each of size equal to IP header plus
569 * a block of the data of the original IP data part) that will yet fit in a
570 * single device frame, and queue such a frame for sending.
573 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
574 int (*output)(struct net *, struct sock *, struct sk_buff *))
578 struct sk_buff *skb2;
579 unsigned int mtu, hlen, left, len, ll_rs;
581 __be16 not_last_frag;
582 struct rtable *rt = skb_rtable(skb);
585 /* for offloaded checksums cleanup checksum before fragmentation */
586 if (skb->ip_summed == CHECKSUM_PARTIAL &&
587 (err = skb_checksum_help(skb)))
591 * Point into the IP datagram header.
596 mtu = ip_skb_dst_mtu(sk, skb);
597 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
598 mtu = IPCB(skb)->frag_max_size;
601 * Setup starting values.
605 mtu = mtu - hlen; /* Size of data space */
606 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
607 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
609 /* When frag_list is given, use it. First, check its validity:
610 * some transformers could create wrong frag_list or break existing
611 * one, it is not prohibited. In this case fall back to copying.
613 * LATER: this step can be merged to real generation of fragments,
614 * we can switch to copy when see the first bad fragment.
616 if (skb_has_frag_list(skb)) {
617 struct sk_buff *frag, *frag2;
618 unsigned int first_len = skb_pagelen(skb);
620 if (first_len - hlen > mtu ||
621 ((first_len - hlen) & 7) ||
622 ip_is_fragment(iph) ||
624 skb_headroom(skb) < ll_rs)
627 skb_walk_frags(skb, frag) {
628 /* Correct geometry. */
629 if (frag->len > mtu ||
630 ((frag->len & 7) && frag->next) ||
631 skb_headroom(frag) < hlen + ll_rs)
632 goto slow_path_clean;
634 /* Partially cloned skb? */
635 if (skb_shared(frag))
636 goto slow_path_clean;
641 frag->destructor = sock_wfree;
643 skb->truesize -= frag->truesize;
646 /* Everything is OK. Generate! */
650 frag = skb_shinfo(skb)->frag_list;
651 skb_frag_list_init(skb);
652 skb->data_len = first_len - skb_headlen(skb);
653 skb->len = first_len;
654 iph->tot_len = htons(first_len);
655 iph->frag_off = htons(IP_MF);
659 /* Prepare header of the next frame,
660 * before previous one went down. */
662 frag->ip_summed = CHECKSUM_NONE;
663 skb_reset_transport_header(frag);
664 __skb_push(frag, hlen);
665 skb_reset_network_header(frag);
666 memcpy(skb_network_header(frag), iph, hlen);
668 iph->tot_len = htons(frag->len);
669 ip_copy_metadata(frag, skb);
671 ip_options_fragment(frag);
672 offset += skb->len - hlen;
673 iph->frag_off = htons(offset>>3);
675 iph->frag_off |= htons(IP_MF);
676 /* Ready, complete checksum */
680 err = output(net, sk, skb);
683 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
693 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
702 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
706 skb_walk_frags(skb, frag2) {
710 frag2->destructor = NULL;
711 skb->truesize += frag2->truesize;
718 left = skb->len - hlen; /* Space per frame */
719 ptr = hlen; /* Where to start from */
722 * Fragment the datagram.
725 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
726 not_last_frag = iph->frag_off & htons(IP_MF);
729 * Keep copying data until we run out.
734 /* IF: it doesn't fit, use 'mtu' - the data space left */
737 /* IF: we are not sending up to and including the packet end
738 then align the next start on an eight byte boundary */
743 /* Allocate buffer */
744 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
751 * Set up data on packet
754 ip_copy_metadata(skb2, skb);
755 skb_reserve(skb2, ll_rs);
756 skb_put(skb2, len + hlen);
757 skb_reset_network_header(skb2);
758 skb2->transport_header = skb2->network_header + hlen;
761 * Charge the memory for the fragment to any owner
766 skb_set_owner_w(skb2, skb->sk);
769 * Copy the packet header into the new buffer.
772 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
775 * Copy a block of the IP datagram.
777 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
782 * Fill in the new header fields.
785 iph->frag_off = htons((offset >> 3));
787 if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
788 iph->frag_off |= htons(IP_DF);
790 /* ANK: dirty, but effective trick. Upgrade options only if
791 * the segment to be fragmented was THE FIRST (otherwise,
792 * options are already fixed) and make it ONCE
793 * on the initial skb, so that all the following fragments
794 * will inherit fixed options.
797 ip_options_fragment(skb);
800 * Added AC : If we are fragmenting a fragment that's not the
801 * last fragment then keep MF on each bit
803 if (left > 0 || not_last_frag)
804 iph->frag_off |= htons(IP_MF);
809 * Put this fragment into the sending queue.
811 iph->tot_len = htons(len + hlen);
815 err = output(net, sk, skb2);
819 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
822 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
827 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
830 EXPORT_SYMBOL(ip_do_fragment);
833 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
835 struct msghdr *msg = from;
837 if (skb->ip_summed == CHECKSUM_PARTIAL) {
838 if (!copy_from_iter_full(to, len, &msg->msg_iter))
842 if (!csum_and_copy_from_iter_full(to, len, &csum, &msg->msg_iter))
844 skb->csum = csum_block_add(skb->csum, csum, odd);
848 EXPORT_SYMBOL(ip_generic_getfrag);
851 csum_page(struct page *page, int offset, int copy)
856 csum = csum_partial(kaddr + offset, copy, 0);
861 static int __ip_append_data(struct sock *sk,
863 struct sk_buff_head *queue,
864 struct inet_cork *cork,
865 struct page_frag *pfrag,
866 int getfrag(void *from, char *to, int offset,
867 int len, int odd, struct sk_buff *skb),
868 void *from, int length, int transhdrlen,
871 struct inet_sock *inet = inet_sk(sk);
874 struct ip_options *opt = cork->opt;
881 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
882 int csummode = CHECKSUM_NONE;
883 struct rtable *rt = (struct rtable *)cork->dst;
886 skb = skb_peek_tail(queue);
888 exthdrlen = !skb ? rt->dst.header_len : 0;
889 mtu = cork->fragsize;
890 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
891 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
892 tskey = sk->sk_tskey++;
894 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
896 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
897 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
898 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
900 if (cork->length + length > maxnonfragsize - fragheaderlen) {
901 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
902 mtu - (opt ? opt->optlen : 0));
907 * transhdrlen > 0 means that this is the first fragment and we wish
908 * it won't be fragmented in the future.
911 length + fragheaderlen <= mtu &&
912 rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
913 !(flags & MSG_MORE) &&
915 csummode = CHECKSUM_PARTIAL;
917 cork->length += length;
919 /* So, what's going on in the loop below?
921 * We use calculated fragment length to generate chained skb,
922 * each of segments is IP fragment ready for sending to network after
923 * adding appropriate IP header.
930 /* Check if the remaining data fits into current packet. */
931 copy = mtu - skb->len;
933 copy = maxfraglen - skb->len;
936 unsigned int datalen;
937 unsigned int fraglen;
938 unsigned int fraggap;
939 unsigned int alloclen;
940 struct sk_buff *skb_prev;
944 fraggap = skb_prev->len - maxfraglen;
949 * If remaining data exceeds the mtu,
950 * we know we need more fragment(s).
952 datalen = length + fraggap;
953 if (datalen > mtu - fragheaderlen)
954 datalen = maxfraglen - fragheaderlen;
955 fraglen = datalen + fragheaderlen;
957 if ((flags & MSG_MORE) &&
958 !(rt->dst.dev->features&NETIF_F_SG))
963 alloclen += exthdrlen;
965 /* The last fragment gets additional space at tail.
966 * Note, with MSG_MORE we overallocate on fragments,
967 * because we have no idea what fragment will be
970 if (datalen == length + fraggap)
971 alloclen += rt->dst.trailer_len;
974 skb = sock_alloc_send_skb(sk,
975 alloclen + hh_len + 15,
976 (flags & MSG_DONTWAIT), &err);
979 if (refcount_read(&sk->sk_wmem_alloc) <=
981 skb = sock_wmalloc(sk,
982 alloclen + hh_len + 15, 1,
991 * Fill in the control structures
993 skb->ip_summed = csummode;
995 skb_reserve(skb, hh_len);
997 /* only the initial fragment is time stamped */
998 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1000 skb_shinfo(skb)->tskey = tskey;
1004 * Find where to start putting bytes.
1006 data = skb_put(skb, fraglen + exthdrlen);
1007 skb_set_network_header(skb, exthdrlen);
1008 skb->transport_header = (skb->network_header +
1010 data += fragheaderlen + exthdrlen;
1013 skb->csum = skb_copy_and_csum_bits(
1014 skb_prev, maxfraglen,
1015 data + transhdrlen, fraggap, 0);
1016 skb_prev->csum = csum_sub(skb_prev->csum,
1019 pskb_trim_unique(skb_prev, maxfraglen);
1022 copy = datalen - transhdrlen - fraggap;
1023 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1030 length -= datalen - fraggap;
1033 csummode = CHECKSUM_NONE;
1035 if ((flags & MSG_CONFIRM) && !skb_prev)
1036 skb_set_dst_pending_confirm(skb, 1);
1039 * Put the packet on the pending queue.
1041 __skb_queue_tail(queue, skb);
1048 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1049 skb_tailroom(skb) >= copy) {
1053 if (getfrag(from, skb_put(skb, copy),
1054 offset, copy, off, skb) < 0) {
1055 __skb_trim(skb, off);
1060 int i = skb_shinfo(skb)->nr_frags;
1063 if (!sk_page_frag_refill(sk, pfrag))
1066 if (!skb_can_coalesce(skb, i, pfrag->page,
1069 if (i == MAX_SKB_FRAGS)
1072 __skb_fill_page_desc(skb, i, pfrag->page,
1074 skb_shinfo(skb)->nr_frags = ++i;
1075 get_page(pfrag->page);
1077 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1079 page_address(pfrag->page) + pfrag->offset,
1080 offset, copy, skb->len, skb) < 0)
1083 pfrag->offset += copy;
1084 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1086 skb->data_len += copy;
1087 skb->truesize += copy;
1088 refcount_add(copy, &sk->sk_wmem_alloc);
1099 cork->length -= length;
1100 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1104 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1105 struct ipcm_cookie *ipc, struct rtable **rtp)
1107 struct ip_options_rcu *opt;
1111 * setup for corking.
1116 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1118 if (unlikely(!cork->opt))
1121 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1122 cork->flags |= IPCORK_OPT;
1123 cork->addr = ipc->addr;
1129 cork->fragsize = ip_sk_use_pmtu(sk) ?
1130 dst_mtu(&rt->dst) : READ_ONCE(rt->dst.dev->mtu);
1132 if (!inetdev_valid_mtu(cork->fragsize))
1133 return -ENETUNREACH;
1135 cork->dst = &rt->dst;
1136 /* We stole this route, caller should not release it. */
1140 cork->ttl = ipc->ttl;
1141 cork->tos = ipc->tos;
1142 cork->priority = ipc->priority;
1143 cork->tx_flags = ipc->tx_flags;
1149 * ip_append_data() and ip_append_page() can make one large IP datagram
1150 * from many pieces of data. Each pieces will be holded on the socket
1151 * until ip_push_pending_frames() is called. Each piece can be a page
1154 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1155 * this interface potentially.
1157 * LATER: length must be adjusted by pad at tail, when it is required.
1159 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1160 int getfrag(void *from, char *to, int offset, int len,
1161 int odd, struct sk_buff *skb),
1162 void *from, int length, int transhdrlen,
1163 struct ipcm_cookie *ipc, struct rtable **rtp,
1166 struct inet_sock *inet = inet_sk(sk);
1169 if (flags&MSG_PROBE)
1172 if (skb_queue_empty(&sk->sk_write_queue)) {
1173 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1180 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1181 sk_page_frag(sk), getfrag,
1182 from, length, transhdrlen, flags);
1185 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1186 int offset, size_t size, int flags)
1188 struct inet_sock *inet = inet_sk(sk);
1189 struct sk_buff *skb;
1191 struct ip_options *opt = NULL;
1192 struct inet_cork *cork;
1197 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1202 if (flags&MSG_PROBE)
1205 if (skb_queue_empty(&sk->sk_write_queue))
1208 cork = &inet->cork.base;
1209 rt = (struct rtable *)cork->dst;
1210 if (cork->flags & IPCORK_OPT)
1213 if (!(rt->dst.dev->features&NETIF_F_SG))
1216 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1217 mtu = cork->fragsize;
1219 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1220 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1221 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1223 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1224 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1225 mtu - (opt ? opt->optlen : 0));
1229 skb = skb_peek_tail(&sk->sk_write_queue);
1233 cork->length += size;
1236 /* Check if the remaining data fits into current packet. */
1237 len = mtu - skb->len;
1239 len = maxfraglen - skb->len;
1242 struct sk_buff *skb_prev;
1246 fraggap = skb_prev->len - maxfraglen;
1248 alloclen = fragheaderlen + hh_len + fraggap + 15;
1249 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1250 if (unlikely(!skb)) {
1256 * Fill in the control structures
1258 skb->ip_summed = CHECKSUM_NONE;
1260 skb_reserve(skb, hh_len);
1263 * Find where to start putting bytes.
1265 skb_put(skb, fragheaderlen + fraggap);
1266 skb_reset_network_header(skb);
1267 skb->transport_header = (skb->network_header +
1270 skb->csum = skb_copy_and_csum_bits(skb_prev,
1272 skb_transport_header(skb),
1274 skb_prev->csum = csum_sub(skb_prev->csum,
1276 pskb_trim_unique(skb_prev, maxfraglen);
1280 * Put the packet on the pending queue.
1282 __skb_queue_tail(&sk->sk_write_queue, skb);
1289 if (skb_append_pagefrags(skb, page, offset, len)) {
1294 if (skb->ip_summed == CHECKSUM_NONE) {
1296 csum = csum_page(page, offset, len);
1297 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1301 skb->data_len += len;
1302 skb->truesize += len;
1303 refcount_add(len, &sk->sk_wmem_alloc);
1310 cork->length -= size;
1311 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1315 static void ip_cork_release(struct inet_cork *cork)
1317 cork->flags &= ~IPCORK_OPT;
1320 dst_release(cork->dst);
1325 * Combined all pending IP fragments on the socket as one IP datagram
1326 * and push them out.
1328 struct sk_buff *__ip_make_skb(struct sock *sk,
1330 struct sk_buff_head *queue,
1331 struct inet_cork *cork)
1333 struct sk_buff *skb, *tmp_skb;
1334 struct sk_buff **tail_skb;
1335 struct inet_sock *inet = inet_sk(sk);
1336 struct net *net = sock_net(sk);
1337 struct ip_options *opt = NULL;
1338 struct rtable *rt = (struct rtable *)cork->dst;
1343 skb = __skb_dequeue(queue);
1346 tail_skb = &(skb_shinfo(skb)->frag_list);
1348 /* move skb->data to ip header from ext header */
1349 if (skb->data < skb_network_header(skb))
1350 __skb_pull(skb, skb_network_offset(skb));
1351 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1352 __skb_pull(tmp_skb, skb_network_header_len(skb));
1353 *tail_skb = tmp_skb;
1354 tail_skb = &(tmp_skb->next);
1355 skb->len += tmp_skb->len;
1356 skb->data_len += tmp_skb->len;
1357 skb->truesize += tmp_skb->truesize;
1358 tmp_skb->destructor = NULL;
1362 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1363 * to fragment the frame generated here. No matter, what transforms
1364 * how transforms change size of the packet, it will come out.
1366 skb->ignore_df = ip_sk_ignore_df(sk);
1368 /* DF bit is set when we want to see DF on outgoing frames.
1369 * If ignore_df is set too, we still allow to fragment this frame
1371 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1372 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1373 (skb->len <= dst_mtu(&rt->dst) &&
1374 ip_dont_fragment(sk, &rt->dst)))
1377 if (cork->flags & IPCORK_OPT)
1382 else if (rt->rt_type == RTN_MULTICAST)
1385 ttl = ip_select_ttl(inet, &rt->dst);
1390 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1393 iph->protocol = sk->sk_protocol;
1394 ip_copy_addrs(iph, fl4);
1395 ip_select_ident(net, skb, sk);
1398 iph->ihl += opt->optlen>>2;
1399 ip_options_build(skb, opt, cork->addr, rt, 0);
1402 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1403 skb->mark = sk->sk_mark;
1405 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1409 skb_dst_set(skb, &rt->dst);
1411 if (iph->protocol == IPPROTO_ICMP)
1412 icmp_out_count(net, ((struct icmphdr *)
1413 skb_transport_header(skb))->type);
1415 ip_cork_release(cork);
1420 int ip_send_skb(struct net *net, struct sk_buff *skb)
1424 err = ip_local_out(net, skb->sk, skb);
1427 err = net_xmit_errno(err);
1429 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1435 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1437 struct sk_buff *skb;
1439 skb = ip_finish_skb(sk, fl4);
1443 /* Netfilter gets whole the not fragmented skb. */
1444 return ip_send_skb(sock_net(sk), skb);
1448 * Throw away all pending data on the socket.
1450 static void __ip_flush_pending_frames(struct sock *sk,
1451 struct sk_buff_head *queue,
1452 struct inet_cork *cork)
1454 struct sk_buff *skb;
1456 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1459 ip_cork_release(cork);
1462 void ip_flush_pending_frames(struct sock *sk)
1464 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1467 struct sk_buff *ip_make_skb(struct sock *sk,
1469 int getfrag(void *from, char *to, int offset,
1470 int len, int odd, struct sk_buff *skb),
1471 void *from, int length, int transhdrlen,
1472 struct ipcm_cookie *ipc, struct rtable **rtp,
1475 struct inet_cork cork;
1476 struct sk_buff_head queue;
1479 if (flags & MSG_PROBE)
1482 __skb_queue_head_init(&queue);
1487 err = ip_setup_cork(sk, &cork, ipc, rtp);
1489 return ERR_PTR(err);
1491 err = __ip_append_data(sk, fl4, &queue, &cork,
1492 ¤t->task_frag, getfrag,
1493 from, length, transhdrlen, flags);
1495 __ip_flush_pending_frames(sk, &queue, &cork);
1496 return ERR_PTR(err);
1499 return __ip_make_skb(sk, fl4, &queue, &cork);
1503 * Fetch data from kernel space and fill in checksum if needed.
1505 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1506 int len, int odd, struct sk_buff *skb)
1510 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1511 skb->csum = csum_block_add(skb->csum, csum, odd);
1516 * Generic function to send a packet as reply to another packet.
1517 * Used to send some TCP resets/acks so far.
1519 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1520 const struct ip_options *sopt,
1521 __be32 daddr, __be32 saddr,
1522 const struct ip_reply_arg *arg,
1525 struct ip_options_data replyopts;
1526 struct ipcm_cookie ipc;
1528 struct rtable *rt = skb_rtable(skb);
1529 struct net *net = sock_net(sk);
1530 struct sk_buff *nskb;
1534 if (__ip_options_echo(net, &replyopts.opt.opt, skb, sopt))
1543 if (replyopts.opt.opt.optlen) {
1544 ipc.opt = &replyopts.opt;
1546 if (replyopts.opt.opt.srr)
1547 daddr = replyopts.opt.opt.faddr;
1550 oif = arg->bound_dev_if;
1551 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1554 flowi4_init_output(&fl4, oif,
1555 IP4_REPLY_MARK(net, skb->mark),
1557 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1558 ip_reply_arg_flowi_flags(arg),
1560 tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1562 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1563 rt = ip_route_output_key(net, &fl4);
1567 inet_sk(sk)->tos = arg->tos & ~INET_ECN_MASK;
1569 sk->sk_priority = skb->priority;
1570 sk->sk_protocol = ip_hdr(skb)->protocol;
1571 sk->sk_bound_dev_if = arg->bound_dev_if;
1572 sk->sk_sndbuf = sysctl_wmem_default;
1573 sk->sk_mark = fl4.flowi4_mark;
1574 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1575 len, 0, &ipc, &rt, MSG_DONTWAIT);
1576 if (unlikely(err)) {
1577 ip_flush_pending_frames(sk);
1581 nskb = skb_peek(&sk->sk_write_queue);
1583 if (arg->csumoffset >= 0)
1584 *((__sum16 *)skb_transport_header(nskb) +
1585 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1587 nskb->ip_summed = CHECKSUM_NONE;
1588 ip_push_pending_frames(sk, &fl4);
1594 void __init ip_init(void)
1599 #if defined(CONFIG_IP_MULTICAST)