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 <linux/igmp.h>
77 #include <linux/netfilter_ipv4.h>
78 #include <linux/netfilter_bridge.h>
79 #include <linux/mroute.h>
80 #include <linux/netlink.h>
81 #include <linux/tcp.h>
83 int sysctl_ip_default_ttl __read_mostly = IPDEFTTL;
84 EXPORT_SYMBOL(sysctl_ip_default_ttl);
86 static int ip_fragment(struct sock *sk, struct sk_buff *skb,
87 int (*output)(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 static int __ip_local_out_sk(struct sock *sk, struct sk_buff *skb)
99 struct iphdr *iph = ip_hdr(skb);
101 iph->tot_len = htons(skb->len);
103 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT, sk, skb, NULL,
104 skb_dst(skb)->dev, dst_output_sk);
107 int __ip_local_out(struct sk_buff *skb)
109 return __ip_local_out_sk(skb->sk, skb);
112 int ip_local_out_sk(struct sock *sk, struct sk_buff *skb)
116 err = __ip_local_out(skb);
117 if (likely(err == 1))
118 err = dst_output_sk(sk, skb);
122 EXPORT_SYMBOL_GPL(ip_local_out_sk);
124 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
126 int ttl = inet->uc_ttl;
129 ttl = ip4_dst_hoplimit(dst);
134 * Add an ip header to a skbuff and send it out.
137 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
138 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
140 struct inet_sock *inet = inet_sk(sk);
141 struct rtable *rt = skb_rtable(skb);
144 /* Build the IP header. */
145 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
146 skb_reset_network_header(skb);
150 iph->tos = inet->tos;
151 if (ip_dont_fragment(sk, &rt->dst))
152 iph->frag_off = htons(IP_DF);
155 iph->ttl = ip_select_ttl(inet, &rt->dst);
156 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
158 iph->protocol = sk->sk_protocol;
159 ip_select_ident(sock_net(sk), skb, sk);
161 if (opt && opt->opt.optlen) {
162 iph->ihl += opt->opt.optlen>>2;
163 ip_options_build(skb, &opt->opt, daddr, rt, 0);
166 skb->priority = sk->sk_priority;
167 skb->mark = sk->sk_mark;
170 return ip_local_out(skb);
172 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
174 static inline int ip_finish_output2(struct sock *sk, struct sk_buff *skb)
176 struct dst_entry *dst = skb_dst(skb);
177 struct rtable *rt = (struct rtable *)dst;
178 struct net_device *dev = dst->dev;
179 unsigned int hh_len = LL_RESERVED_SPACE(dev);
180 struct neighbour *neigh;
183 if (rt->rt_type == RTN_MULTICAST) {
184 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
185 } else if (rt->rt_type == RTN_BROADCAST)
186 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
188 /* Be paranoid, rather than too clever. */
189 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
190 struct sk_buff *skb2;
192 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
198 skb_set_owner_w(skb2, skb->sk);
204 nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
205 neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
206 if (unlikely(!neigh))
207 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
208 if (!IS_ERR(neigh)) {
209 int res = dst_neigh_output(dst, neigh, skb);
211 rcu_read_unlock_bh();
214 rcu_read_unlock_bh();
216 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
222 static int ip_finish_output_gso(struct sock *sk, struct sk_buff *skb)
224 netdev_features_t features;
225 struct sk_buff *segs;
228 /* common case: locally created skb or seglen is <= mtu */
229 if (((IPCB(skb)->flags & IPSKB_FORWARDED) == 0) ||
230 skb_gso_network_seglen(skb) <= ip_skb_dst_mtu(skb))
231 return ip_finish_output2(sk, skb);
233 /* Slowpath - GSO segment length is exceeding the dst MTU.
235 * This can happen in two cases:
236 * 1) TCP GRO packet, DF bit not set
237 * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
238 * from host network stack.
240 features = netif_skb_features(skb);
241 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
242 if (IS_ERR_OR_NULL(segs)) {
250 struct sk_buff *nskb = segs->next;
254 err = ip_fragment(sk, segs, ip_finish_output2);
264 static int ip_finish_output(struct sock *sk, struct sk_buff *skb)
266 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
267 /* Policy lookup after SNAT yielded a new policy */
268 if (skb_dst(skb)->xfrm) {
269 IPCB(skb)->flags |= IPSKB_REROUTED;
270 return dst_output_sk(sk, skb);
274 return ip_finish_output_gso(sk, skb);
276 if (skb->len > ip_skb_dst_mtu(skb))
277 return ip_fragment(sk, skb, ip_finish_output2);
279 return ip_finish_output2(sk, skb);
282 int ip_mc_output(struct sock *sk, struct sk_buff *skb)
284 struct rtable *rt = skb_rtable(skb);
285 struct net_device *dev = rt->dst.dev;
288 * If the indicated interface is up and running, send the packet.
290 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
293 skb->protocol = htons(ETH_P_IP);
296 * Multicasts are looped back for other local users
299 if (rt->rt_flags&RTCF_MULTICAST) {
301 #ifdef CONFIG_IP_MROUTE
302 /* Small optimization: do not loopback not local frames,
303 which returned after forwarding; they will be dropped
304 by ip_mr_input in any case.
305 Note, that local frames are looped back to be delivered
308 This check is duplicated in ip_mr_input at the moment.
311 ((rt->rt_flags & RTCF_LOCAL) ||
312 !(IPCB(skb)->flags & IPSKB_FORWARDED))
315 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
317 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
318 sk, newskb, NULL, newskb->dev,
322 /* Multicasts with ttl 0 must not go beyond the host */
324 if (ip_hdr(skb)->ttl == 0) {
330 if (rt->rt_flags&RTCF_BROADCAST) {
331 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
333 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, sk, newskb,
334 NULL, newskb->dev, dev_loopback_xmit);
337 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, sk, skb, NULL,
338 skb->dev, ip_finish_output,
339 !(IPCB(skb)->flags & IPSKB_REROUTED));
342 int ip_output(struct sock *sk, struct sk_buff *skb)
344 struct net_device *dev = skb_dst(skb)->dev;
346 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
349 skb->protocol = htons(ETH_P_IP);
351 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, sk, skb,
354 !(IPCB(skb)->flags & IPSKB_REROUTED));
358 * copy saddr and daddr, possibly using 64bit load/stores
360 * iph->saddr = fl4->saddr;
361 * iph->daddr = fl4->daddr;
363 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
365 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
366 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
367 memcpy(&iph->saddr, &fl4->saddr,
368 sizeof(fl4->saddr) + sizeof(fl4->daddr));
371 /* Note: skb->sk can be different from sk, in case of tunnels */
372 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
374 struct inet_sock *inet = inet_sk(sk);
375 struct ip_options_rcu *inet_opt;
381 /* Skip all of this if the packet is already routed,
382 * f.e. by something like SCTP.
385 inet_opt = rcu_dereference(inet->inet_opt);
387 rt = skb_rtable(skb);
391 /* Make sure we can route this packet. */
392 rt = (struct rtable *)__sk_dst_check(sk, 0);
396 /* Use correct destination address if we have options. */
397 daddr = inet->inet_daddr;
398 if (inet_opt && inet_opt->opt.srr)
399 daddr = inet_opt->opt.faddr;
401 /* If this fails, retransmit mechanism of transport layer will
402 * keep trying until route appears or the connection times
405 rt = ip_route_output_ports(sock_net(sk), fl4, sk,
406 daddr, inet->inet_saddr,
411 sk->sk_bound_dev_if);
414 sk_setup_caps(sk, &rt->dst);
416 skb_dst_set_noref(skb, &rt->dst);
419 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
422 /* OK, we know where to send it, allocate and build IP header. */
423 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
424 skb_reset_network_header(skb);
426 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
427 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
428 iph->frag_off = htons(IP_DF);
431 iph->ttl = ip_select_ttl(inet, &rt->dst);
432 iph->protocol = sk->sk_protocol;
433 ip_copy_addrs(iph, fl4);
435 /* Transport layer set skb->h.foo itself. */
437 if (inet_opt && inet_opt->opt.optlen) {
438 iph->ihl += inet_opt->opt.optlen >> 2;
439 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
442 ip_select_ident_segs(sock_net(sk), skb, sk,
443 skb_shinfo(skb)->gso_segs ?: 1);
445 /* TODO : should we use skb->sk here instead of sk ? */
446 skb->priority = sk->sk_priority;
447 skb->mark = sk->sk_mark;
449 res = ip_local_out(skb);
455 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
457 return -EHOSTUNREACH;
459 EXPORT_SYMBOL(ip_queue_xmit);
461 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
463 to->pkt_type = from->pkt_type;
464 to->priority = from->priority;
465 to->protocol = from->protocol;
467 skb_dst_copy(to, from);
469 to->mark = from->mark;
471 /* Copy the flags to each fragment. */
472 IPCB(to)->flags = IPCB(from)->flags;
474 #ifdef CONFIG_NET_SCHED
475 to->tc_index = from->tc_index;
478 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
479 to->ipvs_property = from->ipvs_property;
481 skb_copy_secmark(to, from);
484 static int ip_fragment(struct sock *sk, struct sk_buff *skb,
485 int (*output)(struct sock *, struct sk_buff *))
487 struct iphdr *iph = ip_hdr(skb);
488 unsigned int mtu = ip_skb_dst_mtu(skb);
490 if (unlikely(((iph->frag_off & htons(IP_DF)) && !skb->ignore_df) ||
491 (IPCB(skb)->frag_max_size &&
492 IPCB(skb)->frag_max_size > mtu))) {
493 struct rtable *rt = skb_rtable(skb);
494 struct net_device *dev = rt->dst.dev;
496 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
497 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
503 return ip_do_fragment(sk, skb, output);
507 * This IP datagram is too large to be sent in one piece. Break it up into
508 * smaller pieces (each of size equal to IP header plus
509 * a block of the data of the original IP data part) that will yet fit in a
510 * single device frame, and queue such a frame for sending.
513 int ip_do_fragment(struct sock *sk, struct sk_buff *skb,
514 int (*output)(struct sock *, struct sk_buff *))
518 struct net_device *dev;
519 struct sk_buff *skb2;
520 unsigned int mtu, hlen, left, len, ll_rs;
522 __be16 not_last_frag;
523 struct rtable *rt = skb_rtable(skb);
529 * Point into the IP datagram header.
534 mtu = ip_skb_dst_mtu(skb);
537 * Setup starting values.
541 mtu = mtu - hlen; /* Size of data space */
542 #if IS_ENABLED(CONFIG_BRIDGE_NETFILTER)
544 mtu -= nf_bridge_mtu_reduction(skb);
546 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
548 /* When frag_list is given, use it. First, check its validity:
549 * some transformers could create wrong frag_list or break existing
550 * one, it is not prohibited. In this case fall back to copying.
552 * LATER: this step can be merged to real generation of fragments,
553 * we can switch to copy when see the first bad fragment.
555 if (skb_has_frag_list(skb)) {
556 struct sk_buff *frag, *frag2;
557 int first_len = skb_pagelen(skb);
559 if (first_len - hlen > mtu ||
560 ((first_len - hlen) & 7) ||
561 ip_is_fragment(iph) ||
565 skb_walk_frags(skb, frag) {
566 /* Correct geometry. */
567 if (frag->len > mtu ||
568 ((frag->len & 7) && frag->next) ||
569 skb_headroom(frag) < hlen)
570 goto slow_path_clean;
572 /* Partially cloned skb? */
573 if (skb_shared(frag))
574 goto slow_path_clean;
579 frag->destructor = sock_wfree;
581 skb->truesize -= frag->truesize;
584 /* Everything is OK. Generate! */
588 frag = skb_shinfo(skb)->frag_list;
589 skb_frag_list_init(skb);
590 skb->data_len = first_len - skb_headlen(skb);
591 skb->len = first_len;
592 iph->tot_len = htons(first_len);
593 iph->frag_off = htons(IP_MF);
597 /* Prepare header of the next frame,
598 * before previous one went down. */
600 frag->ip_summed = CHECKSUM_NONE;
601 skb_reset_transport_header(frag);
602 __skb_push(frag, hlen);
603 skb_reset_network_header(frag);
604 memcpy(skb_network_header(frag), iph, hlen);
606 iph->tot_len = htons(frag->len);
607 ip_copy_metadata(frag, skb);
609 ip_options_fragment(frag);
610 offset += skb->len - hlen;
611 iph->frag_off = htons(offset>>3);
613 iph->frag_off |= htons(IP_MF);
614 /* Ready, complete checksum */
618 err = output(sk, skb);
621 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
631 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
640 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
644 skb_walk_frags(skb, frag2) {
648 frag2->destructor = NULL;
649 skb->truesize += frag2->truesize;
654 /* for offloaded checksums cleanup checksum before fragmentation */
655 if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
659 left = skb->len - hlen; /* Space per frame */
660 ptr = hlen; /* Where to start from */
662 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
665 * Fragment the datagram.
668 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
669 not_last_frag = iph->frag_off & htons(IP_MF);
672 * Keep copying data until we run out.
677 /* IF: it doesn't fit, use 'mtu' - the data space left */
680 /* IF: we are not sending up to and including the packet end
681 then align the next start on an eight byte boundary */
686 /* Allocate buffer */
687 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
694 * Set up data on packet
697 ip_copy_metadata(skb2, skb);
698 skb_reserve(skb2, ll_rs);
699 skb_put(skb2, len + hlen);
700 skb_reset_network_header(skb2);
701 skb2->transport_header = skb2->network_header + hlen;
704 * Charge the memory for the fragment to any owner
709 skb_set_owner_w(skb2, skb->sk);
712 * Copy the packet header into the new buffer.
715 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
718 * Copy a block of the IP datagram.
720 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
725 * Fill in the new header fields.
728 iph->frag_off = htons((offset >> 3));
730 /* ANK: dirty, but effective trick. Upgrade options only if
731 * the segment to be fragmented was THE FIRST (otherwise,
732 * options are already fixed) and make it ONCE
733 * on the initial skb, so that all the following fragments
734 * will inherit fixed options.
737 ip_options_fragment(skb);
740 * Added AC : If we are fragmenting a fragment that's not the
741 * last fragment then keep MF on each bit
743 if (left > 0 || not_last_frag)
744 iph->frag_off |= htons(IP_MF);
749 * Put this fragment into the sending queue.
751 iph->tot_len = htons(len + hlen);
755 err = output(sk, skb2);
759 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
762 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
767 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
770 EXPORT_SYMBOL(ip_do_fragment);
773 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
775 struct msghdr *msg = from;
777 if (skb->ip_summed == CHECKSUM_PARTIAL) {
778 if (copy_from_iter(to, len, &msg->msg_iter) != len)
782 if (csum_and_copy_from_iter(to, len, &csum, &msg->msg_iter) != len)
784 skb->csum = csum_block_add(skb->csum, csum, odd);
788 EXPORT_SYMBOL(ip_generic_getfrag);
791 csum_page(struct page *page, int offset, int copy)
796 csum = csum_partial(kaddr + offset, copy, 0);
801 static inline int ip_ufo_append_data(struct sock *sk,
802 struct sk_buff_head *queue,
803 int getfrag(void *from, char *to, int offset, int len,
804 int odd, struct sk_buff *skb),
805 void *from, int length, int hh_len, int fragheaderlen,
806 int transhdrlen, int maxfraglen, unsigned int flags)
811 /* There is support for UDP fragmentation offload by network
812 * device, so create one single skb packet containing complete
815 skb = skb_peek_tail(queue);
817 skb = sock_alloc_send_skb(sk,
818 hh_len + fragheaderlen + transhdrlen + 20,
819 (flags & MSG_DONTWAIT), &err);
824 /* reserve space for Hardware header */
825 skb_reserve(skb, hh_len);
827 /* create space for UDP/IP header */
828 skb_put(skb, fragheaderlen + transhdrlen);
830 /* initialize network header pointer */
831 skb_reset_network_header(skb);
833 /* initialize protocol header pointer */
834 skb->transport_header = skb->network_header + fragheaderlen;
838 __skb_queue_tail(queue, skb);
839 } else if (skb_is_gso(skb)) {
843 skb->ip_summed = CHECKSUM_PARTIAL;
844 /* specify the length of each IP datagram fragment */
845 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
846 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
849 return skb_append_datato_frags(sk, skb, getfrag, from,
850 (length - transhdrlen));
853 static int __ip_append_data(struct sock *sk,
855 struct sk_buff_head *queue,
856 struct inet_cork *cork,
857 struct page_frag *pfrag,
858 int getfrag(void *from, char *to, int offset,
859 int len, int odd, struct sk_buff *skb),
860 void *from, int length, int transhdrlen,
863 struct inet_sock *inet = inet_sk(sk);
866 struct ip_options *opt = cork->opt;
873 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
874 int csummode = CHECKSUM_NONE;
875 struct rtable *rt = (struct rtable *)cork->dst;
878 skb = skb_peek_tail(queue);
880 exthdrlen = !skb ? rt->dst.header_len : 0;
881 mtu = cork->fragsize;
882 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
883 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
884 tskey = sk->sk_tskey++;
886 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
888 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
889 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
890 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
892 if (cork->length + length > maxnonfragsize - fragheaderlen) {
893 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
894 mtu - (opt ? opt->optlen : 0));
899 * transhdrlen > 0 means that this is the first fragment and we wish
900 * it won't be fragmented in the future.
903 length + fragheaderlen <= mtu &&
904 rt->dst.dev->features & NETIF_F_V4_CSUM &&
906 csummode = CHECKSUM_PARTIAL;
908 cork->length += length;
909 if (((length > mtu) || (skb && skb_is_gso(skb))) &&
910 (sk->sk_protocol == IPPROTO_UDP) &&
911 (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len &&
912 (sk->sk_type == SOCK_DGRAM)) {
913 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
914 hh_len, fragheaderlen, transhdrlen,
921 /* So, what's going on in the loop below?
923 * We use calculated fragment length to generate chained skb,
924 * each of segments is IP fragment ready for sending to network after
925 * adding appropriate IP header.
932 /* Check if the remaining data fits into current packet. */
933 copy = mtu - skb->len;
935 copy = maxfraglen - skb->len;
938 unsigned int datalen;
939 unsigned int fraglen;
940 unsigned int fraggap;
941 unsigned int alloclen;
942 struct sk_buff *skb_prev;
946 fraggap = skb_prev->len - maxfraglen;
951 * If remaining data exceeds the mtu,
952 * we know we need more fragment(s).
954 datalen = length + fraggap;
955 if (datalen > mtu - fragheaderlen)
956 datalen = maxfraglen - fragheaderlen;
957 fraglen = datalen + fragheaderlen;
959 if ((flags & MSG_MORE) &&
960 !(rt->dst.dev->features&NETIF_F_SG))
965 alloclen += exthdrlen;
967 /* The last fragment gets additional space at tail.
968 * Note, with MSG_MORE we overallocate on fragments,
969 * because we have no idea what fragment will be
972 if (datalen == length + fraggap)
973 alloclen += rt->dst.trailer_len;
976 skb = sock_alloc_send_skb(sk,
977 alloclen + hh_len + 15,
978 (flags & MSG_DONTWAIT), &err);
981 if (atomic_read(&sk->sk_wmem_alloc) <=
983 skb = sock_wmalloc(sk,
984 alloclen + hh_len + 15, 1,
993 * Fill in the control structures
995 skb->ip_summed = csummode;
997 skb_reserve(skb, hh_len);
999 /* only the initial fragment is time stamped */
1000 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1002 skb_shinfo(skb)->tskey = tskey;
1006 * Find where to start putting bytes.
1008 data = skb_put(skb, fraglen + exthdrlen);
1009 skb_set_network_header(skb, exthdrlen);
1010 skb->transport_header = (skb->network_header +
1012 data += fragheaderlen + exthdrlen;
1015 skb->csum = skb_copy_and_csum_bits(
1016 skb_prev, maxfraglen,
1017 data + transhdrlen, fraggap, 0);
1018 skb_prev->csum = csum_sub(skb_prev->csum,
1021 pskb_trim_unique(skb_prev, maxfraglen);
1024 copy = datalen - transhdrlen - fraggap;
1025 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1032 length -= datalen - fraggap;
1035 csummode = CHECKSUM_NONE;
1038 * Put the packet on the pending queue.
1040 __skb_queue_tail(queue, skb);
1047 if (!(rt->dst.dev->features&NETIF_F_SG)) {
1051 if (getfrag(from, skb_put(skb, copy),
1052 offset, copy, off, skb) < 0) {
1053 __skb_trim(skb, off);
1058 int i = skb_shinfo(skb)->nr_frags;
1061 if (!sk_page_frag_refill(sk, pfrag))
1064 if (!skb_can_coalesce(skb, i, pfrag->page,
1067 if (i == MAX_SKB_FRAGS)
1070 __skb_fill_page_desc(skb, i, pfrag->page,
1072 skb_shinfo(skb)->nr_frags = ++i;
1073 get_page(pfrag->page);
1075 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1077 page_address(pfrag->page) + pfrag->offset,
1078 offset, copy, skb->len, skb) < 0)
1081 pfrag->offset += copy;
1082 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1084 skb->data_len += copy;
1085 skb->truesize += copy;
1086 atomic_add(copy, &sk->sk_wmem_alloc);
1097 cork->length -= length;
1098 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1102 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1103 struct ipcm_cookie *ipc, struct rtable **rtp)
1105 struct ip_options_rcu *opt;
1109 * setup for corking.
1114 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1116 if (unlikely(!cork->opt))
1119 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1120 cork->flags |= IPCORK_OPT;
1121 cork->addr = ipc->addr;
1127 * We steal reference to this route, caller should not release it
1130 cork->fragsize = ip_sk_use_pmtu(sk) ?
1131 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1132 cork->dst = &rt->dst;
1134 cork->ttl = ipc->ttl;
1135 cork->tos = ipc->tos;
1136 cork->priority = ipc->priority;
1137 cork->tx_flags = ipc->tx_flags;
1143 * ip_append_data() and ip_append_page() can make one large IP datagram
1144 * from many pieces of data. Each pieces will be holded on the socket
1145 * until ip_push_pending_frames() is called. Each piece can be a page
1148 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1149 * this interface potentially.
1151 * LATER: length must be adjusted by pad at tail, when it is required.
1153 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1154 int getfrag(void *from, char *to, int offset, int len,
1155 int odd, struct sk_buff *skb),
1156 void *from, int length, int transhdrlen,
1157 struct ipcm_cookie *ipc, struct rtable **rtp,
1160 struct inet_sock *inet = inet_sk(sk);
1163 if (flags&MSG_PROBE)
1166 if (skb_queue_empty(&sk->sk_write_queue)) {
1167 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1174 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1175 sk_page_frag(sk), getfrag,
1176 from, length, transhdrlen, flags);
1179 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1180 int offset, size_t size, int flags)
1182 struct inet_sock *inet = inet_sk(sk);
1183 struct sk_buff *skb;
1185 struct ip_options *opt = NULL;
1186 struct inet_cork *cork;
1191 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1196 if (flags&MSG_PROBE)
1199 if (skb_queue_empty(&sk->sk_write_queue))
1202 cork = &inet->cork.base;
1203 rt = (struct rtable *)cork->dst;
1204 if (cork->flags & IPCORK_OPT)
1207 if (!(rt->dst.dev->features&NETIF_F_SG))
1210 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1211 mtu = cork->fragsize;
1213 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1214 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1215 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1217 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1218 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1219 mtu - (opt ? opt->optlen : 0));
1223 skb = skb_peek_tail(&sk->sk_write_queue);
1227 cork->length += size;
1228 if ((size + skb->len > mtu) &&
1229 (sk->sk_protocol == IPPROTO_UDP) &&
1230 (rt->dst.dev->features & NETIF_F_UFO)) {
1231 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1232 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1236 if (skb_is_gso(skb)) {
1240 /* Check if the remaining data fits into current packet. */
1241 len = mtu - skb->len;
1243 len = maxfraglen - skb->len;
1246 struct sk_buff *skb_prev;
1250 fraggap = skb_prev->len - maxfraglen;
1252 alloclen = fragheaderlen + hh_len + fraggap + 15;
1253 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1254 if (unlikely(!skb)) {
1260 * Fill in the control structures
1262 skb->ip_summed = CHECKSUM_NONE;
1264 skb_reserve(skb, hh_len);
1267 * Find where to start putting bytes.
1269 skb_put(skb, fragheaderlen + fraggap);
1270 skb_reset_network_header(skb);
1271 skb->transport_header = (skb->network_header +
1274 skb->csum = skb_copy_and_csum_bits(skb_prev,
1276 skb_transport_header(skb),
1278 skb_prev->csum = csum_sub(skb_prev->csum,
1280 pskb_trim_unique(skb_prev, maxfraglen);
1284 * Put the packet on the pending queue.
1286 __skb_queue_tail(&sk->sk_write_queue, skb);
1293 if (skb_append_pagefrags(skb, page, offset, len)) {
1298 if (skb->ip_summed == CHECKSUM_NONE) {
1300 csum = csum_page(page, offset, len);
1301 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1305 skb->data_len += len;
1306 skb->truesize += len;
1307 atomic_add(len, &sk->sk_wmem_alloc);
1314 cork->length -= size;
1315 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1319 static void ip_cork_release(struct inet_cork *cork)
1321 cork->flags &= ~IPCORK_OPT;
1324 dst_release(cork->dst);
1329 * Combined all pending IP fragments on the socket as one IP datagram
1330 * and push them out.
1332 struct sk_buff *__ip_make_skb(struct sock *sk,
1334 struct sk_buff_head *queue,
1335 struct inet_cork *cork)
1337 struct sk_buff *skb, *tmp_skb;
1338 struct sk_buff **tail_skb;
1339 struct inet_sock *inet = inet_sk(sk);
1340 struct net *net = sock_net(sk);
1341 struct ip_options *opt = NULL;
1342 struct rtable *rt = (struct rtable *)cork->dst;
1347 skb = __skb_dequeue(queue);
1350 tail_skb = &(skb_shinfo(skb)->frag_list);
1352 /* move skb->data to ip header from ext header */
1353 if (skb->data < skb_network_header(skb))
1354 __skb_pull(skb, skb_network_offset(skb));
1355 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1356 __skb_pull(tmp_skb, skb_network_header_len(skb));
1357 *tail_skb = tmp_skb;
1358 tail_skb = &(tmp_skb->next);
1359 skb->len += tmp_skb->len;
1360 skb->data_len += tmp_skb->len;
1361 skb->truesize += tmp_skb->truesize;
1362 tmp_skb->destructor = NULL;
1366 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1367 * to fragment the frame generated here. No matter, what transforms
1368 * how transforms change size of the packet, it will come out.
1370 skb->ignore_df = ip_sk_ignore_df(sk);
1372 /* DF bit is set when we want to see DF on outgoing frames.
1373 * If ignore_df is set too, we still allow to fragment this frame
1375 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1376 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1377 (skb->len <= dst_mtu(&rt->dst) &&
1378 ip_dont_fragment(sk, &rt->dst)))
1381 if (cork->flags & IPCORK_OPT)
1386 else if (rt->rt_type == RTN_MULTICAST)
1389 ttl = ip_select_ttl(inet, &rt->dst);
1394 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1397 iph->protocol = sk->sk_protocol;
1398 ip_copy_addrs(iph, fl4);
1399 ip_select_ident(net, skb, sk);
1402 iph->ihl += opt->optlen>>2;
1403 ip_options_build(skb, opt, cork->addr, rt, 0);
1406 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1407 skb->mark = sk->sk_mark;
1409 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1413 skb_dst_set(skb, &rt->dst);
1415 if (iph->protocol == IPPROTO_ICMP)
1416 icmp_out_count(net, ((struct icmphdr *)
1417 skb_transport_header(skb))->type);
1419 ip_cork_release(cork);
1424 int ip_send_skb(struct net *net, struct sk_buff *skb)
1428 err = ip_local_out(skb);
1431 err = net_xmit_errno(err);
1433 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1439 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1441 struct sk_buff *skb;
1443 skb = ip_finish_skb(sk, fl4);
1447 /* Netfilter gets whole the not fragmented skb. */
1448 return ip_send_skb(sock_net(sk), skb);
1452 * Throw away all pending data on the socket.
1454 static void __ip_flush_pending_frames(struct sock *sk,
1455 struct sk_buff_head *queue,
1456 struct inet_cork *cork)
1458 struct sk_buff *skb;
1460 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1463 ip_cork_release(cork);
1466 void ip_flush_pending_frames(struct sock *sk)
1468 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1471 struct sk_buff *ip_make_skb(struct sock *sk,
1473 int getfrag(void *from, char *to, int offset,
1474 int len, int odd, struct sk_buff *skb),
1475 void *from, int length, int transhdrlen,
1476 struct ipcm_cookie *ipc, struct rtable **rtp,
1479 struct inet_cork cork;
1480 struct sk_buff_head queue;
1483 if (flags & MSG_PROBE)
1486 __skb_queue_head_init(&queue);
1491 err = ip_setup_cork(sk, &cork, ipc, rtp);
1493 return ERR_PTR(err);
1495 err = __ip_append_data(sk, fl4, &queue, &cork,
1496 ¤t->task_frag, getfrag,
1497 from, length, transhdrlen, flags);
1499 __ip_flush_pending_frames(sk, &queue, &cork);
1500 return ERR_PTR(err);
1503 return __ip_make_skb(sk, fl4, &queue, &cork);
1507 * Fetch data from kernel space and fill in checksum if needed.
1509 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1510 int len, int odd, struct sk_buff *skb)
1514 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1515 skb->csum = csum_block_add(skb->csum, csum, odd);
1520 * Generic function to send a packet as reply to another packet.
1521 * Used to send some TCP resets/acks so far.
1523 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1524 const struct ip_options *sopt,
1525 __be32 daddr, __be32 saddr,
1526 const struct ip_reply_arg *arg,
1529 struct ip_options_data replyopts;
1530 struct ipcm_cookie ipc;
1532 struct rtable *rt = skb_rtable(skb);
1533 struct net *net = sock_net(sk);
1534 struct sk_buff *nskb;
1537 if (__ip_options_echo(&replyopts.opt.opt, skb, sopt))
1546 if (replyopts.opt.opt.optlen) {
1547 ipc.opt = &replyopts.opt;
1549 if (replyopts.opt.opt.srr)
1550 daddr = replyopts.opt.opt.faddr;
1553 flowi4_init_output(&fl4, arg->bound_dev_if,
1554 IP4_REPLY_MARK(net, skb->mark),
1556 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1557 ip_reply_arg_flowi_flags(arg),
1559 tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1560 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1561 rt = ip_route_output_key(net, &fl4);
1565 inet_sk(sk)->tos = arg->tos;
1567 sk->sk_priority = skb->priority;
1568 sk->sk_protocol = ip_hdr(skb)->protocol;
1569 sk->sk_bound_dev_if = arg->bound_dev_if;
1570 sk->sk_sndbuf = sysctl_wmem_default;
1571 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1572 len, 0, &ipc, &rt, MSG_DONTWAIT);
1573 if (unlikely(err)) {
1574 ip_flush_pending_frames(sk);
1578 nskb = skb_peek(&sk->sk_write_queue);
1580 if (arg->csumoffset >= 0)
1581 *((__sum16 *)skb_transport_header(nskb) +
1582 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1584 nskb->ip_summed = CHECKSUM_NONE;
1585 skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1586 ip_push_pending_frames(sk, &fl4);
1592 void __init ip_init(void)
1597 #if defined(CONFIG_IP_MULTICAST)