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 <linux/igmp.h>
78 #include <linux/netfilter_ipv4.h>
79 #include <linux/netfilter_bridge.h>
80 #include <linux/mroute.h>
81 #include <linux/netlink.h>
82 #include <linux/tcp.h>
84 int sysctl_ip_default_ttl __read_mostly = IPDEFTTL;
85 EXPORT_SYMBOL(sysctl_ip_default_ttl);
88 ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
90 int (*output)(struct net *, struct sock *, struct sk_buff *));
92 /* Generate a checksum for an outgoing IP datagram. */
93 void ip_send_check(struct iphdr *iph)
96 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
98 EXPORT_SYMBOL(ip_send_check);
100 int __ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
102 struct iphdr *iph = ip_hdr(skb);
104 iph->tot_len = htons(skb->len);
107 skb->protocol = htons(ETH_P_IP);
109 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
110 net, sk, skb, NULL, skb_dst(skb)->dev,
114 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
118 err = __ip_local_out(net, sk, skb);
119 if (likely(err == 1))
120 err = dst_output(net, sk, skb);
124 EXPORT_SYMBOL_GPL(ip_local_out);
126 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
128 int ttl = inet->uc_ttl;
131 ttl = ip4_dst_hoplimit(dst);
136 * Add an ip header to a skbuff and send it out.
139 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
140 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
142 struct inet_sock *inet = inet_sk(sk);
143 struct rtable *rt = skb_rtable(skb);
144 struct net *net = sock_net(sk);
147 /* Build the IP header. */
148 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
149 skb_reset_network_header(skb);
153 iph->tos = inet->tos;
154 iph->ttl = ip_select_ttl(inet, &rt->dst);
155 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
157 iph->protocol = sk->sk_protocol;
158 if (ip_dont_fragment(sk, &rt->dst)) {
159 iph->frag_off = htons(IP_DF);
163 __ip_select_ident(net, iph, 1);
166 if (opt && opt->opt.optlen) {
167 iph->ihl += opt->opt.optlen>>2;
168 ip_options_build(skb, &opt->opt, daddr, rt, 0);
171 skb->priority = sk->sk_priority;
172 skb->mark = sk->sk_mark;
175 return ip_local_out(net, skb->sk, skb);
177 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
179 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
181 struct dst_entry *dst = skb_dst(skb);
182 struct rtable *rt = (struct rtable *)dst;
183 struct net_device *dev = dst->dev;
184 unsigned int hh_len = LL_RESERVED_SPACE(dev);
185 struct neighbour *neigh;
188 if (rt->rt_type == RTN_MULTICAST) {
189 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
190 } else if (rt->rt_type == RTN_BROADCAST)
191 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
193 /* Be paranoid, rather than too clever. */
194 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
195 struct sk_buff *skb2;
197 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
203 skb_set_owner_w(skb2, skb->sk);
209 nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
210 neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
211 if (unlikely(!neigh))
212 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
213 if (!IS_ERR(neigh)) {
214 int res = dst_neigh_output(dst, neigh, skb);
216 rcu_read_unlock_bh();
219 rcu_read_unlock_bh();
221 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
227 static int ip_finish_output_gso(struct net *net, struct sock *sk,
228 struct sk_buff *skb, unsigned int mtu)
230 netdev_features_t features;
231 struct sk_buff *segs;
234 /* common case: locally created skb or seglen is <= mtu */
235 if (((IPCB(skb)->flags & IPSKB_FORWARDED) == 0) ||
236 skb_gso_network_seglen(skb) <= mtu)
237 return ip_finish_output2(net, sk, skb);
239 /* Slowpath - GSO segment length is exceeding the dst MTU.
241 * This can happen in two cases:
242 * 1) TCP GRO packet, DF bit not set
243 * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
244 * from host network stack.
246 features = netif_skb_features(skb);
247 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_SGO_CB_OFFSET);
248 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
249 if (IS_ERR_OR_NULL(segs)) {
257 struct sk_buff *nskb = segs->next;
261 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
271 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
275 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
276 /* Policy lookup after SNAT yielded a new policy */
277 if (skb_dst(skb)->xfrm) {
278 IPCB(skb)->flags |= IPSKB_REROUTED;
279 return dst_output(net, sk, skb);
282 mtu = ip_skb_dst_mtu(skb);
284 return ip_finish_output_gso(net, sk, skb, mtu);
286 if (skb->len > mtu || IPCB(skb)->frag_max_size)
287 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
289 return ip_finish_output2(net, sk, skb);
292 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
294 struct rtable *rt = skb_rtable(skb);
295 struct net_device *dev = rt->dst.dev;
298 * If the indicated interface is up and running, send the packet.
300 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
303 skb->protocol = htons(ETH_P_IP);
306 * Multicasts are looped back for other local users
309 if (rt->rt_flags&RTCF_MULTICAST) {
311 #ifdef CONFIG_IP_MROUTE
312 /* Small optimization: do not loopback not local frames,
313 which returned after forwarding; they will be dropped
314 by ip_mr_input in any case.
315 Note, that local frames are looped back to be delivered
318 This check is duplicated in ip_mr_input at the moment.
321 ((rt->rt_flags & RTCF_LOCAL) ||
322 !(IPCB(skb)->flags & IPSKB_FORWARDED))
325 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
327 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
328 net, sk, newskb, NULL, newskb->dev,
332 /* Multicasts with ttl 0 must not go beyond the host */
334 if (ip_hdr(skb)->ttl == 0) {
340 if (rt->rt_flags&RTCF_BROADCAST) {
341 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
343 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
344 net, sk, newskb, NULL, newskb->dev,
348 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
349 net, sk, skb, NULL, skb->dev,
351 !(IPCB(skb)->flags & IPSKB_REROUTED));
354 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
356 struct net_device *dev = skb_dst(skb)->dev;
358 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
361 skb->protocol = htons(ETH_P_IP);
363 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
364 net, sk, skb, NULL, dev,
366 !(IPCB(skb)->flags & IPSKB_REROUTED));
370 * copy saddr and daddr, possibly using 64bit load/stores
372 * iph->saddr = fl4->saddr;
373 * iph->daddr = fl4->daddr;
375 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
377 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
378 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
380 iph->saddr = fl4->saddr;
381 iph->daddr = fl4->daddr;
384 /* Note: skb->sk can be different from sk, in case of tunnels */
385 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
387 struct inet_sock *inet = inet_sk(sk);
388 struct net *net = sock_net(sk);
389 struct ip_options_rcu *inet_opt;
395 /* Skip all of this if the packet is already routed,
396 * f.e. by something like SCTP.
399 inet_opt = rcu_dereference(inet->inet_opt);
401 rt = skb_rtable(skb);
405 /* Make sure we can route this packet. */
406 rt = (struct rtable *)__sk_dst_check(sk, 0);
410 /* Use correct destination address if we have options. */
411 daddr = inet->inet_daddr;
412 if (inet_opt && inet_opt->opt.srr)
413 daddr = inet_opt->opt.faddr;
415 /* If this fails, retransmit mechanism of transport layer will
416 * keep trying until route appears or the connection times
419 rt = ip_route_output_ports(net, fl4, sk,
420 daddr, inet->inet_saddr,
425 sk->sk_bound_dev_if);
428 sk_setup_caps(sk, &rt->dst);
430 skb_dst_set_noref(skb, &rt->dst);
433 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
436 /* OK, we know where to send it, allocate and build IP header. */
437 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
438 skb_reset_network_header(skb);
440 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
441 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
442 iph->frag_off = htons(IP_DF);
445 iph->ttl = ip_select_ttl(inet, &rt->dst);
446 iph->protocol = sk->sk_protocol;
447 ip_copy_addrs(iph, fl4);
449 /* Transport layer set skb->h.foo itself. */
451 if (inet_opt && inet_opt->opt.optlen) {
452 iph->ihl += inet_opt->opt.optlen >> 2;
453 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
456 ip_select_ident_segs(net, skb, sk,
457 skb_shinfo(skb)->gso_segs ?: 1);
459 /* TODO : should we use skb->sk here instead of sk ? */
460 skb->priority = sk->sk_priority;
461 skb->mark = sk->sk_mark;
463 res = ip_local_out(net, sk, skb);
469 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
471 return -EHOSTUNREACH;
473 EXPORT_SYMBOL(ip_queue_xmit);
475 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
477 to->pkt_type = from->pkt_type;
478 to->priority = from->priority;
479 to->protocol = from->protocol;
480 to->skb_iif = from->skb_iif;
482 skb_dst_copy(to, from);
484 to->mark = from->mark;
486 skb_copy_hash(to, from);
488 /* Copy the flags to each fragment. */
489 IPCB(to)->flags = IPCB(from)->flags;
491 #ifdef CONFIG_NET_SCHED
492 to->tc_index = from->tc_index;
495 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
496 to->ipvs_property = from->ipvs_property;
498 skb_copy_secmark(to, from);
501 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
503 int (*output)(struct net *, struct sock *, struct sk_buff *))
505 struct iphdr *iph = ip_hdr(skb);
507 if ((iph->frag_off & htons(IP_DF)) == 0)
508 return ip_do_fragment(net, sk, skb, output);
510 if (unlikely(!skb->ignore_df ||
511 (IPCB(skb)->frag_max_size &&
512 IPCB(skb)->frag_max_size > mtu))) {
513 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
514 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
520 return ip_do_fragment(net, sk, skb, output);
524 * This IP datagram is too large to be sent in one piece. Break it up into
525 * smaller pieces (each of size equal to IP header plus
526 * a block of the data of the original IP data part) that will yet fit in a
527 * single device frame, and queue such a frame for sending.
530 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
531 int (*output)(struct net *, struct sock *, struct sk_buff *))
535 struct net_device *dev;
536 struct sk_buff *skb2;
537 unsigned int mtu, hlen, left, len, ll_rs;
539 __be16 not_last_frag;
540 struct rtable *rt = skb_rtable(skb);
545 /* for offloaded checksums cleanup checksum before fragmentation */
546 if (skb->ip_summed == CHECKSUM_PARTIAL &&
547 (err = skb_checksum_help(skb)))
551 * Point into the IP datagram header.
556 mtu = ip_skb_dst_mtu(skb);
557 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
558 mtu = IPCB(skb)->frag_max_size;
561 * Setup starting values.
565 mtu = mtu - hlen; /* Size of data space */
566 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
568 /* When frag_list is given, use it. First, check its validity:
569 * some transformers could create wrong frag_list or break existing
570 * one, it is not prohibited. In this case fall back to copying.
572 * LATER: this step can be merged to real generation of fragments,
573 * we can switch to copy when see the first bad fragment.
575 if (skb_has_frag_list(skb)) {
576 struct sk_buff *frag, *frag2;
577 int first_len = skb_pagelen(skb);
579 if (first_len - hlen > mtu ||
580 ((first_len - hlen) & 7) ||
581 ip_is_fragment(iph) ||
585 skb_walk_frags(skb, frag) {
586 /* Correct geometry. */
587 if (frag->len > mtu ||
588 ((frag->len & 7) && frag->next) ||
589 skb_headroom(frag) < hlen)
590 goto slow_path_clean;
592 /* Partially cloned skb? */
593 if (skb_shared(frag))
594 goto slow_path_clean;
599 frag->destructor = sock_wfree;
601 skb->truesize -= frag->truesize;
604 /* Everything is OK. Generate! */
608 frag = skb_shinfo(skb)->frag_list;
609 skb_frag_list_init(skb);
610 skb->data_len = first_len - skb_headlen(skb);
611 skb->len = first_len;
612 iph->tot_len = htons(first_len);
613 iph->frag_off = htons(IP_MF);
617 /* Prepare header of the next frame,
618 * before previous one went down. */
620 frag->ip_summed = CHECKSUM_NONE;
621 skb_reset_transport_header(frag);
622 __skb_push(frag, hlen);
623 skb_reset_network_header(frag);
624 memcpy(skb_network_header(frag), iph, hlen);
626 iph->tot_len = htons(frag->len);
627 ip_copy_metadata(frag, skb);
629 ip_options_fragment(frag);
630 offset += skb->len - hlen;
631 iph->frag_off = htons(offset>>3);
633 iph->frag_off |= htons(IP_MF);
634 /* Ready, complete checksum */
638 err = output(net, sk, skb);
641 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
651 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
660 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
664 skb_walk_frags(skb, frag2) {
668 frag2->destructor = NULL;
669 skb->truesize += frag2->truesize;
676 left = skb->len - hlen; /* Space per frame */
677 ptr = hlen; /* Where to start from */
679 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
682 * Fragment the datagram.
685 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
686 not_last_frag = iph->frag_off & htons(IP_MF);
689 * Keep copying data until we run out.
694 /* IF: it doesn't fit, use 'mtu' - the data space left */
697 /* IF: we are not sending up to and including the packet end
698 then align the next start on an eight byte boundary */
703 /* Allocate buffer */
704 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
711 * Set up data on packet
714 ip_copy_metadata(skb2, skb);
715 skb_reserve(skb2, ll_rs);
716 skb_put(skb2, len + hlen);
717 skb_reset_network_header(skb2);
718 skb2->transport_header = skb2->network_header + hlen;
721 * Charge the memory for the fragment to any owner
726 skb_set_owner_w(skb2, skb->sk);
729 * Copy the packet header into the new buffer.
732 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
735 * Copy a block of the IP datagram.
737 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
742 * Fill in the new header fields.
745 iph->frag_off = htons((offset >> 3));
747 if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
748 iph->frag_off |= htons(IP_DF);
750 /* ANK: dirty, but effective trick. Upgrade options only if
751 * the segment to be fragmented was THE FIRST (otherwise,
752 * options are already fixed) and make it ONCE
753 * on the initial skb, so that all the following fragments
754 * will inherit fixed options.
757 ip_options_fragment(skb);
760 * Added AC : If we are fragmenting a fragment that's not the
761 * last fragment then keep MF on each bit
763 if (left > 0 || not_last_frag)
764 iph->frag_off |= htons(IP_MF);
769 * Put this fragment into the sending queue.
771 iph->tot_len = htons(len + hlen);
775 err = output(net, sk, skb2);
779 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
782 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
787 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
790 EXPORT_SYMBOL(ip_do_fragment);
793 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
795 struct msghdr *msg = from;
797 if (skb->ip_summed == CHECKSUM_PARTIAL) {
798 if (copy_from_iter(to, len, &msg->msg_iter) != len)
802 if (csum_and_copy_from_iter(to, len, &csum, &msg->msg_iter) != len)
804 skb->csum = csum_block_add(skb->csum, csum, odd);
808 EXPORT_SYMBOL(ip_generic_getfrag);
811 csum_page(struct page *page, int offset, int copy)
816 csum = csum_partial(kaddr + offset, copy, 0);
821 static inline int ip_ufo_append_data(struct sock *sk,
822 struct sk_buff_head *queue,
823 int getfrag(void *from, char *to, int offset, int len,
824 int odd, struct sk_buff *skb),
825 void *from, int length, int hh_len, int fragheaderlen,
826 int transhdrlen, int maxfraglen, unsigned int flags)
831 /* There is support for UDP fragmentation offload by network
832 * device, so create one single skb packet containing complete
835 skb = skb_peek_tail(queue);
837 skb = sock_alloc_send_skb(sk,
838 hh_len + fragheaderlen + transhdrlen + 20,
839 (flags & MSG_DONTWAIT), &err);
844 /* reserve space for Hardware header */
845 skb_reserve(skb, hh_len);
847 /* create space for UDP/IP header */
848 skb_put(skb, fragheaderlen + transhdrlen);
850 /* initialize network header pointer */
851 skb_reset_network_header(skb);
853 /* initialize protocol header pointer */
854 skb->transport_header = skb->network_header + fragheaderlen;
858 __skb_queue_tail(queue, skb);
859 } else if (skb_is_gso(skb)) {
863 skb->ip_summed = CHECKSUM_PARTIAL;
864 /* specify the length of each IP datagram fragment */
865 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
866 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
869 return skb_append_datato_frags(sk, skb, getfrag, from,
870 (length - transhdrlen));
873 static int __ip_append_data(struct sock *sk,
875 struct sk_buff_head *queue,
876 struct inet_cork *cork,
877 struct page_frag *pfrag,
878 int getfrag(void *from, char *to, int offset,
879 int len, int odd, struct sk_buff *skb),
880 void *from, int length, int transhdrlen,
883 struct inet_sock *inet = inet_sk(sk);
886 struct ip_options *opt = cork->opt;
893 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
894 int csummode = CHECKSUM_NONE;
895 struct rtable *rt = (struct rtable *)cork->dst;
898 skb = skb_peek_tail(queue);
900 exthdrlen = !skb ? rt->dst.header_len : 0;
901 mtu = cork->fragsize;
902 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
903 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
904 tskey = sk->sk_tskey++;
906 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
908 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
909 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
910 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
912 if (cork->length + length > maxnonfragsize - fragheaderlen) {
913 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
914 mtu - (opt ? opt->optlen : 0));
919 * transhdrlen > 0 means that this is the first fragment and we wish
920 * it won't be fragmented in the future.
923 length + fragheaderlen <= mtu &&
924 rt->dst.dev->features & NETIF_F_V4_CSUM &&
925 !(flags & MSG_MORE) &&
927 csummode = CHECKSUM_PARTIAL;
929 cork->length += length;
930 if ((skb && skb_is_gso(skb)) ||
931 (((length + (skb ? skb->len : fragheaderlen)) > mtu) &&
932 (skb_queue_len(queue) <= 1) &&
933 (sk->sk_protocol == IPPROTO_UDP) &&
934 (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len &&
935 (sk->sk_type == SOCK_DGRAM) && !sk->sk_no_check_tx)) {
936 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
937 hh_len, fragheaderlen, transhdrlen,
944 /* So, what's going on in the loop below?
946 * We use calculated fragment length to generate chained skb,
947 * each of segments is IP fragment ready for sending to network after
948 * adding appropriate IP header.
955 /* Check if the remaining data fits into current packet. */
956 copy = mtu - skb->len;
958 copy = maxfraglen - skb->len;
961 unsigned int datalen;
962 unsigned int fraglen;
963 unsigned int fraggap;
964 unsigned int alloclen;
965 struct sk_buff *skb_prev;
969 fraggap = skb_prev->len - maxfraglen;
974 * If remaining data exceeds the mtu,
975 * we know we need more fragment(s).
977 datalen = length + fraggap;
978 if (datalen > mtu - fragheaderlen)
979 datalen = maxfraglen - fragheaderlen;
980 fraglen = datalen + fragheaderlen;
982 if ((flags & MSG_MORE) &&
983 !(rt->dst.dev->features&NETIF_F_SG))
988 alloclen += exthdrlen;
990 /* The last fragment gets additional space at tail.
991 * Note, with MSG_MORE we overallocate on fragments,
992 * because we have no idea what fragment will be
995 if (datalen == length + fraggap)
996 alloclen += rt->dst.trailer_len;
999 skb = sock_alloc_send_skb(sk,
1000 alloclen + hh_len + 15,
1001 (flags & MSG_DONTWAIT), &err);
1004 if (atomic_read(&sk->sk_wmem_alloc) <=
1006 skb = sock_wmalloc(sk,
1007 alloclen + hh_len + 15, 1,
1016 * Fill in the control structures
1018 skb->ip_summed = csummode;
1020 skb_reserve(skb, hh_len);
1022 /* only the initial fragment is time stamped */
1023 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1025 skb_shinfo(skb)->tskey = tskey;
1029 * Find where to start putting bytes.
1031 data = skb_put(skb, fraglen + exthdrlen);
1032 skb_set_network_header(skb, exthdrlen);
1033 skb->transport_header = (skb->network_header +
1035 data += fragheaderlen + exthdrlen;
1038 skb->csum = skb_copy_and_csum_bits(
1039 skb_prev, maxfraglen,
1040 data + transhdrlen, fraggap, 0);
1041 skb_prev->csum = csum_sub(skb_prev->csum,
1044 pskb_trim_unique(skb_prev, maxfraglen);
1047 copy = datalen - transhdrlen - fraggap;
1048 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1055 length -= datalen - fraggap;
1058 csummode = CHECKSUM_NONE;
1061 * Put the packet on the pending queue.
1063 __skb_queue_tail(queue, skb);
1070 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1071 skb_tailroom(skb) >= copy) {
1075 if (getfrag(from, skb_put(skb, copy),
1076 offset, copy, off, skb) < 0) {
1077 __skb_trim(skb, off);
1082 int i = skb_shinfo(skb)->nr_frags;
1085 if (!sk_page_frag_refill(sk, pfrag))
1088 if (!skb_can_coalesce(skb, i, pfrag->page,
1091 if (i == MAX_SKB_FRAGS)
1094 __skb_fill_page_desc(skb, i, pfrag->page,
1096 skb_shinfo(skb)->nr_frags = ++i;
1097 get_page(pfrag->page);
1099 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1101 page_address(pfrag->page) + pfrag->offset,
1102 offset, copy, skb->len, skb) < 0)
1105 pfrag->offset += copy;
1106 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1108 skb->data_len += copy;
1109 skb->truesize += copy;
1110 atomic_add(copy, &sk->sk_wmem_alloc);
1121 cork->length -= length;
1122 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1126 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1127 struct ipcm_cookie *ipc, struct rtable **rtp)
1129 struct ip_options_rcu *opt;
1133 * setup for corking.
1138 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1140 if (unlikely(!cork->opt))
1143 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1144 cork->flags |= IPCORK_OPT;
1145 cork->addr = ipc->addr;
1151 cork->fragsize = ip_sk_use_pmtu(sk) ?
1152 dst_mtu(&rt->dst) : READ_ONCE(rt->dst.dev->mtu);
1154 if (!inetdev_valid_mtu(cork->fragsize))
1155 return -ENETUNREACH;
1157 cork->dst = &rt->dst;
1158 /* We stole this route, caller should not release it. */
1162 cork->ttl = ipc->ttl;
1163 cork->tos = ipc->tos;
1164 cork->priority = ipc->priority;
1165 cork->tx_flags = ipc->tx_flags;
1171 * ip_append_data() and ip_append_page() can make one large IP datagram
1172 * from many pieces of data. Each pieces will be holded on the socket
1173 * until ip_push_pending_frames() is called. Each piece can be a page
1176 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1177 * this interface potentially.
1179 * LATER: length must be adjusted by pad at tail, when it is required.
1181 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1182 int getfrag(void *from, char *to, int offset, int len,
1183 int odd, struct sk_buff *skb),
1184 void *from, int length, int transhdrlen,
1185 struct ipcm_cookie *ipc, struct rtable **rtp,
1188 struct inet_sock *inet = inet_sk(sk);
1191 if (flags&MSG_PROBE)
1194 if (skb_queue_empty(&sk->sk_write_queue)) {
1195 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1202 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1203 sk_page_frag(sk), getfrag,
1204 from, length, transhdrlen, flags);
1207 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1208 int offset, size_t size, int flags)
1210 struct inet_sock *inet = inet_sk(sk);
1211 struct sk_buff *skb;
1213 struct ip_options *opt = NULL;
1214 struct inet_cork *cork;
1219 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1224 if (flags&MSG_PROBE)
1227 if (skb_queue_empty(&sk->sk_write_queue))
1230 cork = &inet->cork.base;
1231 rt = (struct rtable *)cork->dst;
1232 if (cork->flags & IPCORK_OPT)
1235 if (!(rt->dst.dev->features&NETIF_F_SG))
1238 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1239 mtu = cork->fragsize;
1241 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1242 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1243 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1245 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1246 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1247 mtu - (opt ? opt->optlen : 0));
1251 skb = skb_peek_tail(&sk->sk_write_queue);
1255 if ((size + skb->len > mtu) &&
1256 (skb_queue_len(&sk->sk_write_queue) == 1) &&
1257 (sk->sk_protocol == IPPROTO_UDP) &&
1258 (rt->dst.dev->features & NETIF_F_UFO)) {
1259 if (skb->ip_summed != CHECKSUM_PARTIAL)
1262 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1263 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1265 cork->length += size;
1268 if (skb_is_gso(skb)) {
1272 /* Check if the remaining data fits into current packet. */
1273 len = mtu - skb->len;
1275 len = maxfraglen - skb->len;
1278 struct sk_buff *skb_prev;
1282 fraggap = skb_prev->len - maxfraglen;
1284 alloclen = fragheaderlen + hh_len + fraggap + 15;
1285 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1286 if (unlikely(!skb)) {
1292 * Fill in the control structures
1294 skb->ip_summed = CHECKSUM_NONE;
1296 skb_reserve(skb, hh_len);
1299 * Find where to start putting bytes.
1301 skb_put(skb, fragheaderlen + fraggap);
1302 skb_reset_network_header(skb);
1303 skb->transport_header = (skb->network_header +
1306 skb->csum = skb_copy_and_csum_bits(skb_prev,
1308 skb_transport_header(skb),
1310 skb_prev->csum = csum_sub(skb_prev->csum,
1312 pskb_trim_unique(skb_prev, maxfraglen);
1316 * Put the packet on the pending queue.
1318 __skb_queue_tail(&sk->sk_write_queue, skb);
1325 if (skb_append_pagefrags(skb, page, offset, len)) {
1330 if (skb->ip_summed == CHECKSUM_NONE) {
1332 csum = csum_page(page, offset, len);
1333 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1337 skb->data_len += len;
1338 skb->truesize += len;
1339 atomic_add(len, &sk->sk_wmem_alloc);
1346 cork->length -= size;
1347 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1351 static void ip_cork_release(struct inet_cork *cork)
1353 cork->flags &= ~IPCORK_OPT;
1356 dst_release(cork->dst);
1361 * Combined all pending IP fragments on the socket as one IP datagram
1362 * and push them out.
1364 struct sk_buff *__ip_make_skb(struct sock *sk,
1366 struct sk_buff_head *queue,
1367 struct inet_cork *cork)
1369 struct sk_buff *skb, *tmp_skb;
1370 struct sk_buff **tail_skb;
1371 struct inet_sock *inet = inet_sk(sk);
1372 struct net *net = sock_net(sk);
1373 struct ip_options *opt = NULL;
1374 struct rtable *rt = (struct rtable *)cork->dst;
1379 skb = __skb_dequeue(queue);
1382 tail_skb = &(skb_shinfo(skb)->frag_list);
1384 /* move skb->data to ip header from ext header */
1385 if (skb->data < skb_network_header(skb))
1386 __skb_pull(skb, skb_network_offset(skb));
1387 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1388 __skb_pull(tmp_skb, skb_network_header_len(skb));
1389 *tail_skb = tmp_skb;
1390 tail_skb = &(tmp_skb->next);
1391 skb->len += tmp_skb->len;
1392 skb->data_len += tmp_skb->len;
1393 skb->truesize += tmp_skb->truesize;
1394 tmp_skb->destructor = NULL;
1398 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1399 * to fragment the frame generated here. No matter, what transforms
1400 * how transforms change size of the packet, it will come out.
1402 skb->ignore_df = ip_sk_ignore_df(sk);
1404 /* DF bit is set when we want to see DF on outgoing frames.
1405 * If ignore_df is set too, we still allow to fragment this frame
1407 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1408 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1409 (skb->len <= dst_mtu(&rt->dst) &&
1410 ip_dont_fragment(sk, &rt->dst)))
1413 if (cork->flags & IPCORK_OPT)
1418 else if (rt->rt_type == RTN_MULTICAST)
1421 ttl = ip_select_ttl(inet, &rt->dst);
1426 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1429 iph->protocol = sk->sk_protocol;
1430 ip_copy_addrs(iph, fl4);
1431 ip_select_ident(net, skb, sk);
1434 iph->ihl += opt->optlen>>2;
1435 ip_options_build(skb, opt, cork->addr, rt, 0);
1438 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1439 skb->mark = sk->sk_mark;
1441 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1445 skb_dst_set(skb, &rt->dst);
1447 if (iph->protocol == IPPROTO_ICMP)
1448 icmp_out_count(net, ((struct icmphdr *)
1449 skb_transport_header(skb))->type);
1451 ip_cork_release(cork);
1456 int ip_send_skb(struct net *net, struct sk_buff *skb)
1460 err = ip_local_out(net, skb->sk, skb);
1463 err = net_xmit_errno(err);
1465 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1471 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1473 struct sk_buff *skb;
1475 skb = ip_finish_skb(sk, fl4);
1479 /* Netfilter gets whole the not fragmented skb. */
1480 return ip_send_skb(sock_net(sk), skb);
1484 * Throw away all pending data on the socket.
1486 static void __ip_flush_pending_frames(struct sock *sk,
1487 struct sk_buff_head *queue,
1488 struct inet_cork *cork)
1490 struct sk_buff *skb;
1492 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1495 ip_cork_release(cork);
1498 void ip_flush_pending_frames(struct sock *sk)
1500 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1503 struct sk_buff *ip_make_skb(struct sock *sk,
1505 int getfrag(void *from, char *to, int offset,
1506 int len, int odd, struct sk_buff *skb),
1507 void *from, int length, int transhdrlen,
1508 struct ipcm_cookie *ipc, struct rtable **rtp,
1511 struct inet_cork cork;
1512 struct sk_buff_head queue;
1515 if (flags & MSG_PROBE)
1518 __skb_queue_head_init(&queue);
1523 err = ip_setup_cork(sk, &cork, ipc, rtp);
1525 return ERR_PTR(err);
1527 err = __ip_append_data(sk, fl4, &queue, &cork,
1528 ¤t->task_frag, getfrag,
1529 from, length, transhdrlen, flags);
1531 __ip_flush_pending_frames(sk, &queue, &cork);
1532 return ERR_PTR(err);
1535 return __ip_make_skb(sk, fl4, &queue, &cork);
1539 * Fetch data from kernel space and fill in checksum if needed.
1541 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1542 int len, int odd, struct sk_buff *skb)
1546 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1547 skb->csum = csum_block_add(skb->csum, csum, odd);
1552 * Generic function to send a packet as reply to another packet.
1553 * Used to send some TCP resets/acks so far.
1555 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1556 const struct ip_options *sopt,
1557 __be32 daddr, __be32 saddr,
1558 const struct ip_reply_arg *arg,
1561 struct ip_options_data replyopts;
1562 struct ipcm_cookie ipc;
1564 struct rtable *rt = skb_rtable(skb);
1565 struct net *net = sock_net(sk);
1566 struct sk_buff *nskb;
1570 if (__ip_options_echo(&replyopts.opt.opt, skb, sopt))
1579 if (replyopts.opt.opt.optlen) {
1580 ipc.opt = &replyopts.opt;
1582 if (replyopts.opt.opt.srr)
1583 daddr = replyopts.opt.opt.faddr;
1586 oif = arg->bound_dev_if;
1587 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1590 flowi4_init_output(&fl4, oif,
1591 IP4_REPLY_MARK(net, skb->mark),
1593 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1594 ip_reply_arg_flowi_flags(arg),
1596 tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1597 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1598 rt = ip_route_output_key(net, &fl4);
1602 inet_sk(sk)->tos = arg->tos & ~INET_ECN_MASK;
1604 sk->sk_priority = skb->priority;
1605 sk->sk_protocol = ip_hdr(skb)->protocol;
1606 sk->sk_bound_dev_if = arg->bound_dev_if;
1607 sk->sk_sndbuf = sysctl_wmem_default;
1608 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1609 len, 0, &ipc, &rt, MSG_DONTWAIT);
1610 if (unlikely(err)) {
1611 ip_flush_pending_frames(sk);
1615 nskb = skb_peek(&sk->sk_write_queue);
1617 if (arg->csumoffset >= 0)
1618 *((__sum16 *)skb_transport_header(nskb) +
1619 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1621 nskb->ip_summed = CHECKSUM_NONE;
1622 ip_push_pending_frames(sk, &fl4);
1628 void __init ip_init(void)
1633 #if defined(CONFIG_IP_MULTICAST)