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/lwtunnel.h>
77 #include <linux/igmp.h>
78 #include <linux/netfilter_ipv4.h>
79 #include <linux/netfilter_bridge.h>
80 #include <linux/netlink.h>
81 #include <linux/tcp.h>
84 ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
86 int (*output)(struct net *, struct sock *, struct sk_buff *));
88 /* Generate a checksum for an outgoing IP datagram. */
89 void ip_send_check(struct iphdr *iph)
92 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
94 EXPORT_SYMBOL(ip_send_check);
96 int __ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
98 struct iphdr *iph = ip_hdr(skb);
100 iph->tot_len = htons(skb->len);
103 /* if egress device is enslaved to an L3 master device pass the
104 * skb to its handler for processing
106 skb = l3mdev_ip_out(sk, skb);
110 skb->protocol = htons(ETH_P_IP);
112 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
113 net, sk, skb, NULL, skb_dst(skb)->dev,
117 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
121 err = __ip_local_out(net, sk, skb);
122 if (likely(err == 1))
123 err = dst_output(net, sk, skb);
127 EXPORT_SYMBOL_GPL(ip_local_out);
129 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
131 int ttl = inet->uc_ttl;
134 ttl = ip4_dst_hoplimit(dst);
139 * Add an ip header to a skbuff and send it out.
142 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
143 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
145 struct inet_sock *inet = inet_sk(sk);
146 struct rtable *rt = skb_rtable(skb);
147 struct net *net = sock_net(sk);
150 /* Build the IP header. */
151 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
152 skb_reset_network_header(skb);
156 iph->tos = inet->tos;
157 iph->ttl = ip_select_ttl(inet, &rt->dst);
158 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
160 iph->protocol = sk->sk_protocol;
161 if (ip_dont_fragment(sk, &rt->dst)) {
162 iph->frag_off = htons(IP_DF);
166 __ip_select_ident(net, iph, 1);
169 if (opt && opt->opt.optlen) {
170 iph->ihl += opt->opt.optlen>>2;
171 ip_options_build(skb, &opt->opt, daddr, rt, 0);
174 skb->priority = sk->sk_priority;
175 skb->mark = sk->sk_mark;
178 return ip_local_out(net, skb->sk, skb);
180 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
182 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
184 struct dst_entry *dst = skb_dst(skb);
185 struct rtable *rt = (struct rtable *)dst;
186 struct net_device *dev = dst->dev;
187 unsigned int hh_len = LL_RESERVED_SPACE(dev);
188 struct neighbour *neigh;
191 if (rt->rt_type == RTN_MULTICAST) {
192 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
193 } else if (rt->rt_type == RTN_BROADCAST)
194 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
196 /* Be paranoid, rather than too clever. */
197 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
198 struct sk_buff *skb2;
200 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
206 skb_set_owner_w(skb2, skb->sk);
211 if (lwtunnel_xmit_redirect(dst->lwtstate)) {
212 int res = lwtunnel_xmit(skb);
214 if (res < 0 || res == LWTUNNEL_XMIT_DONE)
219 nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
220 neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
221 if (unlikely(!neigh))
222 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
223 if (!IS_ERR(neigh)) {
224 int res = dst_neigh_output(dst, neigh, skb);
226 rcu_read_unlock_bh();
229 rcu_read_unlock_bh();
231 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
237 static int ip_finish_output_gso(struct net *net, struct sock *sk,
238 struct sk_buff *skb, unsigned int mtu)
240 netdev_features_t features;
241 struct sk_buff *segs;
244 /* common case: seglen is <= mtu
246 if (skb_gso_validate_mtu(skb, mtu))
247 return ip_finish_output2(net, sk, skb);
249 /* Slowpath - GSO segment length exceeds the egress MTU.
251 * This can happen in several cases:
252 * - Forwarding of a TCP GRO skb, when DF flag is not set.
253 * - Forwarding of an skb that arrived on a virtualization interface
254 * (virtio-net/vhost/tap) with TSO/GSO size set by other network
256 * - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
257 * interface with a smaller MTU.
258 * - Arriving GRO skb (or GSO skb in a virtualized environment) that is
259 * bridged to a NETIF_F_TSO tunnel stacked over an interface with an
262 features = netif_skb_features(skb);
263 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_SGO_CB_OFFSET);
264 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
265 if (IS_ERR_OR_NULL(segs)) {
273 struct sk_buff *nskb = segs->next;
277 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
287 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
291 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
292 /* Policy lookup after SNAT yielded a new policy */
293 if (skb_dst(skb)->xfrm) {
294 IPCB(skb)->flags |= IPSKB_REROUTED;
295 return dst_output(net, sk, skb);
298 mtu = ip_skb_dst_mtu(sk, skb);
300 return ip_finish_output_gso(net, sk, skb, mtu);
302 if (skb->len > mtu || (IPCB(skb)->flags & IPSKB_FRAG_PMTU))
303 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
305 return ip_finish_output2(net, sk, skb);
308 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
310 struct rtable *rt = skb_rtable(skb);
311 struct net_device *dev = rt->dst.dev;
314 * If the indicated interface is up and running, send the packet.
316 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
319 skb->protocol = htons(ETH_P_IP);
322 * Multicasts are looped back for other local users
325 if (rt->rt_flags&RTCF_MULTICAST) {
327 #ifdef CONFIG_IP_MROUTE
328 /* Small optimization: do not loopback not local frames,
329 which returned after forwarding; they will be dropped
330 by ip_mr_input in any case.
331 Note, that local frames are looped back to be delivered
334 This check is duplicated in ip_mr_input at the moment.
337 ((rt->rt_flags & RTCF_LOCAL) ||
338 !(IPCB(skb)->flags & IPSKB_FORWARDED))
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 /* Multicasts with ttl 0 must not go beyond the host */
350 if (ip_hdr(skb)->ttl == 0) {
356 if (rt->rt_flags&RTCF_BROADCAST) {
357 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
359 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
360 net, sk, newskb, NULL, newskb->dev,
364 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
365 net, sk, skb, NULL, skb->dev,
367 !(IPCB(skb)->flags & IPSKB_REROUTED));
370 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
372 struct net_device *dev = skb_dst(skb)->dev;
374 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
377 skb->protocol = htons(ETH_P_IP);
379 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
380 net, sk, skb, NULL, dev,
382 !(IPCB(skb)->flags & IPSKB_REROUTED));
386 * copy saddr and daddr, possibly using 64bit load/stores
388 * iph->saddr = fl4->saddr;
389 * iph->daddr = fl4->daddr;
391 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
393 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
394 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
395 memcpy(&iph->saddr, &fl4->saddr,
396 sizeof(fl4->saddr) + sizeof(fl4->daddr));
399 /* Note: skb->sk can be different from sk, in case of tunnels */
400 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
402 struct inet_sock *inet = inet_sk(sk);
403 struct net *net = sock_net(sk);
404 struct ip_options_rcu *inet_opt;
410 /* Skip all of this if the packet is already routed,
411 * f.e. by something like SCTP.
414 inet_opt = rcu_dereference(inet->inet_opt);
416 rt = skb_rtable(skb);
420 /* Make sure we can route this packet. */
421 rt = (struct rtable *)__sk_dst_check(sk, 0);
425 /* Use correct destination address if we have options. */
426 daddr = inet->inet_daddr;
427 if (inet_opt && inet_opt->opt.srr)
428 daddr = inet_opt->opt.faddr;
430 /* If this fails, retransmit mechanism of transport layer will
431 * keep trying until route appears or the connection times
434 rt = ip_route_output_ports(net, fl4, sk,
435 daddr, inet->inet_saddr,
440 sk->sk_bound_dev_if);
443 sk_setup_caps(sk, &rt->dst);
445 skb_dst_set_noref(skb, &rt->dst);
448 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
451 /* OK, we know where to send it, allocate and build IP header. */
452 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
453 skb_reset_network_header(skb);
455 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
456 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
457 iph->frag_off = htons(IP_DF);
460 iph->ttl = ip_select_ttl(inet, &rt->dst);
461 iph->protocol = sk->sk_protocol;
462 ip_copy_addrs(iph, fl4);
464 /* Transport layer set skb->h.foo itself. */
466 if (inet_opt && inet_opt->opt.optlen) {
467 iph->ihl += inet_opt->opt.optlen >> 2;
468 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
471 ip_select_ident_segs(net, skb, sk,
472 skb_shinfo(skb)->gso_segs ?: 1);
474 /* TODO : should we use skb->sk here instead of sk ? */
475 skb->priority = sk->sk_priority;
476 skb->mark = sk->sk_mark;
478 res = ip_local_out(net, sk, skb);
484 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
486 return -EHOSTUNREACH;
488 EXPORT_SYMBOL(ip_queue_xmit);
490 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
492 to->pkt_type = from->pkt_type;
493 to->priority = from->priority;
494 to->protocol = from->protocol;
496 skb_dst_copy(to, from);
498 to->mark = from->mark;
500 /* Copy the flags to each fragment. */
501 IPCB(to)->flags = IPCB(from)->flags;
503 #ifdef CONFIG_NET_SCHED
504 to->tc_index = from->tc_index;
507 #if IS_ENABLED(CONFIG_IP_VS)
508 to->ipvs_property = from->ipvs_property;
510 skb_copy_secmark(to, from);
513 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
515 int (*output)(struct net *, struct sock *, struct sk_buff *))
517 struct iphdr *iph = ip_hdr(skb);
519 if ((iph->frag_off & htons(IP_DF)) == 0)
520 return ip_do_fragment(net, sk, skb, output);
522 if (unlikely(!skb->ignore_df ||
523 (IPCB(skb)->frag_max_size &&
524 IPCB(skb)->frag_max_size > mtu))) {
525 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
526 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
532 return ip_do_fragment(net, sk, skb, output);
536 * This IP datagram is too large to be sent in one piece. Break it up into
537 * smaller pieces (each of size equal to IP header plus
538 * a block of the data of the original IP data part) that will yet fit in a
539 * single device frame, and queue such a frame for sending.
542 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
543 int (*output)(struct net *, struct sock *, struct sk_buff *))
547 struct sk_buff *skb2;
548 unsigned int mtu, hlen, left, len, ll_rs;
550 __be16 not_last_frag;
551 struct rtable *rt = skb_rtable(skb);
554 /* for offloaded checksums cleanup checksum before fragmentation */
555 if (skb->ip_summed == CHECKSUM_PARTIAL &&
556 (err = skb_checksum_help(skb)))
560 * Point into the IP datagram header.
565 mtu = ip_skb_dst_mtu(sk, skb);
566 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
567 mtu = IPCB(skb)->frag_max_size;
570 * Setup starting values.
574 mtu = mtu - hlen; /* Size of data space */
575 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
577 /* When frag_list is given, use it. First, check its validity:
578 * some transformers could create wrong frag_list or break existing
579 * one, it is not prohibited. In this case fall back to copying.
581 * LATER: this step can be merged to real generation of fragments,
582 * we can switch to copy when see the first bad fragment.
584 if (skb_has_frag_list(skb)) {
585 struct sk_buff *frag, *frag2;
586 int first_len = skb_pagelen(skb);
588 if (first_len - hlen > mtu ||
589 ((first_len - hlen) & 7) ||
590 ip_is_fragment(iph) ||
594 skb_walk_frags(skb, frag) {
595 /* Correct geometry. */
596 if (frag->len > mtu ||
597 ((frag->len & 7) && frag->next) ||
598 skb_headroom(frag) < hlen)
599 goto slow_path_clean;
601 /* Partially cloned skb? */
602 if (skb_shared(frag))
603 goto slow_path_clean;
608 frag->destructor = sock_wfree;
610 skb->truesize -= frag->truesize;
613 /* Everything is OK. Generate! */
617 frag = skb_shinfo(skb)->frag_list;
618 skb_frag_list_init(skb);
619 skb->data_len = first_len - skb_headlen(skb);
620 skb->len = first_len;
621 iph->tot_len = htons(first_len);
622 iph->frag_off = htons(IP_MF);
626 /* Prepare header of the next frame,
627 * before previous one went down. */
629 frag->ip_summed = CHECKSUM_NONE;
630 skb_reset_transport_header(frag);
631 __skb_push(frag, hlen);
632 skb_reset_network_header(frag);
633 memcpy(skb_network_header(frag), iph, hlen);
635 iph->tot_len = htons(frag->len);
636 ip_copy_metadata(frag, skb);
638 ip_options_fragment(frag);
639 offset += skb->len - hlen;
640 iph->frag_off = htons(offset>>3);
642 iph->frag_off |= htons(IP_MF);
643 /* Ready, complete checksum */
647 err = output(net, sk, skb);
650 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
660 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
669 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
673 skb_walk_frags(skb, frag2) {
677 frag2->destructor = NULL;
678 skb->truesize += frag2->truesize;
685 left = skb->len - hlen; /* Space per frame */
686 ptr = hlen; /* Where to start from */
688 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
691 * Fragment the datagram.
694 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
695 not_last_frag = iph->frag_off & htons(IP_MF);
698 * Keep copying data until we run out.
703 /* IF: it doesn't fit, use 'mtu' - the data space left */
706 /* IF: we are not sending up to and including the packet end
707 then align the next start on an eight byte boundary */
712 /* Allocate buffer */
713 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
720 * Set up data on packet
723 ip_copy_metadata(skb2, skb);
724 skb_reserve(skb2, ll_rs);
725 skb_put(skb2, len + hlen);
726 skb_reset_network_header(skb2);
727 skb2->transport_header = skb2->network_header + hlen;
730 * Charge the memory for the fragment to any owner
735 skb_set_owner_w(skb2, skb->sk);
738 * Copy the packet header into the new buffer.
741 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
744 * Copy a block of the IP datagram.
746 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
751 * Fill in the new header fields.
754 iph->frag_off = htons((offset >> 3));
756 if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
757 iph->frag_off |= htons(IP_DF);
759 /* ANK: dirty, but effective trick. Upgrade options only if
760 * the segment to be fragmented was THE FIRST (otherwise,
761 * options are already fixed) and make it ONCE
762 * on the initial skb, so that all the following fragments
763 * will inherit fixed options.
766 ip_options_fragment(skb);
769 * Added AC : If we are fragmenting a fragment that's not the
770 * last fragment then keep MF on each bit
772 if (left > 0 || not_last_frag)
773 iph->frag_off |= htons(IP_MF);
778 * Put this fragment into the sending queue.
780 iph->tot_len = htons(len + hlen);
784 err = output(net, sk, skb2);
788 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
791 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
796 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
799 EXPORT_SYMBOL(ip_do_fragment);
802 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
804 struct msghdr *msg = from;
806 if (skb->ip_summed == CHECKSUM_PARTIAL) {
807 if (copy_from_iter(to, len, &msg->msg_iter) != len)
811 if (csum_and_copy_from_iter(to, len, &csum, &msg->msg_iter) != len)
813 skb->csum = csum_block_add(skb->csum, csum, odd);
817 EXPORT_SYMBOL(ip_generic_getfrag);
820 csum_page(struct page *page, int offset, int copy)
825 csum = csum_partial(kaddr + offset, copy, 0);
830 static inline int ip_ufo_append_data(struct sock *sk,
831 struct sk_buff_head *queue,
832 int getfrag(void *from, char *to, int offset, int len,
833 int odd, struct sk_buff *skb),
834 void *from, int length, int hh_len, int fragheaderlen,
835 int transhdrlen, int maxfraglen, unsigned int flags)
840 /* There is support for UDP fragmentation offload by network
841 * device, so create one single skb packet containing complete
844 skb = skb_peek_tail(queue);
846 skb = sock_alloc_send_skb(sk,
847 hh_len + fragheaderlen + transhdrlen + 20,
848 (flags & MSG_DONTWAIT), &err);
853 /* reserve space for Hardware header */
854 skb_reserve(skb, hh_len);
856 /* create space for UDP/IP header */
857 skb_put(skb, fragheaderlen + transhdrlen);
859 /* initialize network header pointer */
860 skb_reset_network_header(skb);
862 /* initialize protocol header pointer */
863 skb->transport_header = skb->network_header + fragheaderlen;
867 __skb_queue_tail(queue, skb);
868 } else if (skb_is_gso(skb)) {
872 skb->ip_summed = CHECKSUM_PARTIAL;
873 /* specify the length of each IP datagram fragment */
874 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
875 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
878 return skb_append_datato_frags(sk, skb, getfrag, from,
879 (length - transhdrlen));
882 static int __ip_append_data(struct sock *sk,
884 struct sk_buff_head *queue,
885 struct inet_cork *cork,
886 struct page_frag *pfrag,
887 int getfrag(void *from, char *to, int offset,
888 int len, int odd, struct sk_buff *skb),
889 void *from, int length, int transhdrlen,
892 struct inet_sock *inet = inet_sk(sk);
895 struct ip_options *opt = cork->opt;
902 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
903 int csummode = CHECKSUM_NONE;
904 struct rtable *rt = (struct rtable *)cork->dst;
907 skb = skb_peek_tail(queue);
909 exthdrlen = !skb ? rt->dst.header_len : 0;
910 mtu = cork->fragsize;
911 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
912 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
913 tskey = sk->sk_tskey++;
915 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
917 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
918 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
919 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
921 if (cork->length + length > maxnonfragsize - fragheaderlen) {
922 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
923 mtu - (opt ? opt->optlen : 0));
928 * transhdrlen > 0 means that this is the first fragment and we wish
929 * it won't be fragmented in the future.
932 length + fragheaderlen <= mtu &&
933 rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
934 !(flags & MSG_MORE) &&
936 csummode = CHECKSUM_PARTIAL;
938 cork->length += length;
939 if (((length > mtu) || (skb && skb_is_gso(skb))) &&
940 (sk->sk_protocol == IPPROTO_UDP) &&
941 (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len &&
942 (sk->sk_type == SOCK_DGRAM) && !sk->sk_no_check_tx) {
943 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
944 hh_len, fragheaderlen, transhdrlen,
951 /* So, what's going on in the loop below?
953 * We use calculated fragment length to generate chained skb,
954 * each of segments is IP fragment ready for sending to network after
955 * adding appropriate IP header.
962 /* Check if the remaining data fits into current packet. */
963 copy = mtu - skb->len;
965 copy = maxfraglen - skb->len;
968 unsigned int datalen;
969 unsigned int fraglen;
970 unsigned int fraggap;
971 unsigned int alloclen;
972 struct sk_buff *skb_prev;
976 fraggap = skb_prev->len - maxfraglen;
981 * If remaining data exceeds the mtu,
982 * we know we need more fragment(s).
984 datalen = length + fraggap;
985 if (datalen > mtu - fragheaderlen)
986 datalen = maxfraglen - fragheaderlen;
987 fraglen = datalen + fragheaderlen;
989 if ((flags & MSG_MORE) &&
990 !(rt->dst.dev->features&NETIF_F_SG))
995 alloclen += exthdrlen;
997 /* The last fragment gets additional space at tail.
998 * Note, with MSG_MORE we overallocate on fragments,
999 * because we have no idea what fragment will be
1002 if (datalen == length + fraggap)
1003 alloclen += rt->dst.trailer_len;
1006 skb = sock_alloc_send_skb(sk,
1007 alloclen + hh_len + 15,
1008 (flags & MSG_DONTWAIT), &err);
1011 if (atomic_read(&sk->sk_wmem_alloc) <=
1013 skb = sock_wmalloc(sk,
1014 alloclen + hh_len + 15, 1,
1023 * Fill in the control structures
1025 skb->ip_summed = csummode;
1027 skb_reserve(skb, hh_len);
1029 /* only the initial fragment is time stamped */
1030 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1032 skb_shinfo(skb)->tskey = tskey;
1036 * Find where to start putting bytes.
1038 data = skb_put(skb, fraglen + exthdrlen);
1039 skb_set_network_header(skb, exthdrlen);
1040 skb->transport_header = (skb->network_header +
1042 data += fragheaderlen + exthdrlen;
1045 skb->csum = skb_copy_and_csum_bits(
1046 skb_prev, maxfraglen,
1047 data + transhdrlen, fraggap, 0);
1048 skb_prev->csum = csum_sub(skb_prev->csum,
1051 pskb_trim_unique(skb_prev, maxfraglen);
1054 copy = datalen - transhdrlen - fraggap;
1055 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1062 length -= datalen - fraggap;
1065 csummode = CHECKSUM_NONE;
1068 * Put the packet on the pending queue.
1070 __skb_queue_tail(queue, skb);
1077 if (!(rt->dst.dev->features&NETIF_F_SG)) {
1081 if (getfrag(from, skb_put(skb, copy),
1082 offset, copy, off, skb) < 0) {
1083 __skb_trim(skb, off);
1088 int i = skb_shinfo(skb)->nr_frags;
1091 if (!sk_page_frag_refill(sk, pfrag))
1094 if (!skb_can_coalesce(skb, i, pfrag->page,
1097 if (i == MAX_SKB_FRAGS)
1100 __skb_fill_page_desc(skb, i, pfrag->page,
1102 skb_shinfo(skb)->nr_frags = ++i;
1103 get_page(pfrag->page);
1105 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1107 page_address(pfrag->page) + pfrag->offset,
1108 offset, copy, skb->len, skb) < 0)
1111 pfrag->offset += copy;
1112 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1114 skb->data_len += copy;
1115 skb->truesize += copy;
1116 atomic_add(copy, &sk->sk_wmem_alloc);
1127 cork->length -= length;
1128 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1132 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1133 struct ipcm_cookie *ipc, struct rtable **rtp)
1135 struct ip_options_rcu *opt;
1139 * setup for corking.
1144 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1146 if (unlikely(!cork->opt))
1149 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1150 cork->flags |= IPCORK_OPT;
1151 cork->addr = ipc->addr;
1157 * We steal reference to this route, caller should not release it
1160 cork->fragsize = ip_sk_use_pmtu(sk) ?
1161 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1162 cork->dst = &rt->dst;
1164 cork->ttl = ipc->ttl;
1165 cork->tos = ipc->tos;
1166 cork->priority = ipc->priority;
1167 cork->tx_flags = ipc->tx_flags;
1173 * ip_append_data() and ip_append_page() can make one large IP datagram
1174 * from many pieces of data. Each pieces will be holded on the socket
1175 * until ip_push_pending_frames() is called. Each piece can be a page
1178 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1179 * this interface potentially.
1181 * LATER: length must be adjusted by pad at tail, when it is required.
1183 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1184 int getfrag(void *from, char *to, int offset, int len,
1185 int odd, struct sk_buff *skb),
1186 void *from, int length, int transhdrlen,
1187 struct ipcm_cookie *ipc, struct rtable **rtp,
1190 struct inet_sock *inet = inet_sk(sk);
1193 if (flags&MSG_PROBE)
1196 if (skb_queue_empty(&sk->sk_write_queue)) {
1197 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1204 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1205 sk_page_frag(sk), getfrag,
1206 from, length, transhdrlen, flags);
1209 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1210 int offset, size_t size, int flags)
1212 struct inet_sock *inet = inet_sk(sk);
1213 struct sk_buff *skb;
1215 struct ip_options *opt = NULL;
1216 struct inet_cork *cork;
1221 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1226 if (flags&MSG_PROBE)
1229 if (skb_queue_empty(&sk->sk_write_queue))
1232 cork = &inet->cork.base;
1233 rt = (struct rtable *)cork->dst;
1234 if (cork->flags & IPCORK_OPT)
1237 if (!(rt->dst.dev->features&NETIF_F_SG))
1240 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1241 mtu = cork->fragsize;
1243 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1244 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1245 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1247 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1248 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1249 mtu - (opt ? opt->optlen : 0));
1253 skb = skb_peek_tail(&sk->sk_write_queue);
1257 if ((size + skb->len > mtu) &&
1258 (sk->sk_protocol == IPPROTO_UDP) &&
1259 (rt->dst.dev->features & NETIF_F_UFO)) {
1260 if (skb->ip_summed != CHECKSUM_PARTIAL)
1263 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1264 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1266 cork->length += size;
1269 if (skb_is_gso(skb)) {
1273 /* Check if the remaining data fits into current packet. */
1274 len = mtu - skb->len;
1276 len = maxfraglen - skb->len;
1279 struct sk_buff *skb_prev;
1283 fraggap = skb_prev->len - maxfraglen;
1285 alloclen = fragheaderlen + hh_len + fraggap + 15;
1286 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1287 if (unlikely(!skb)) {
1293 * Fill in the control structures
1295 skb->ip_summed = CHECKSUM_NONE;
1297 skb_reserve(skb, hh_len);
1300 * Find where to start putting bytes.
1302 skb_put(skb, fragheaderlen + fraggap);
1303 skb_reset_network_header(skb);
1304 skb->transport_header = (skb->network_header +
1307 skb->csum = skb_copy_and_csum_bits(skb_prev,
1309 skb_transport_header(skb),
1311 skb_prev->csum = csum_sub(skb_prev->csum,
1313 pskb_trim_unique(skb_prev, maxfraglen);
1317 * Put the packet on the pending queue.
1319 __skb_queue_tail(&sk->sk_write_queue, skb);
1326 if (skb_append_pagefrags(skb, page, offset, len)) {
1331 if (skb->ip_summed == CHECKSUM_NONE) {
1333 csum = csum_page(page, offset, len);
1334 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1338 skb->data_len += len;
1339 skb->truesize += len;
1340 atomic_add(len, &sk->sk_wmem_alloc);
1347 cork->length -= size;
1348 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1352 static void ip_cork_release(struct inet_cork *cork)
1354 cork->flags &= ~IPCORK_OPT;
1357 dst_release(cork->dst);
1362 * Combined all pending IP fragments on the socket as one IP datagram
1363 * and push them out.
1365 struct sk_buff *__ip_make_skb(struct sock *sk,
1367 struct sk_buff_head *queue,
1368 struct inet_cork *cork)
1370 struct sk_buff *skb, *tmp_skb;
1371 struct sk_buff **tail_skb;
1372 struct inet_sock *inet = inet_sk(sk);
1373 struct net *net = sock_net(sk);
1374 struct ip_options *opt = NULL;
1375 struct rtable *rt = (struct rtable *)cork->dst;
1380 skb = __skb_dequeue(queue);
1383 tail_skb = &(skb_shinfo(skb)->frag_list);
1385 /* move skb->data to ip header from ext header */
1386 if (skb->data < skb_network_header(skb))
1387 __skb_pull(skb, skb_network_offset(skb));
1388 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1389 __skb_pull(tmp_skb, skb_network_header_len(skb));
1390 *tail_skb = tmp_skb;
1391 tail_skb = &(tmp_skb->next);
1392 skb->len += tmp_skb->len;
1393 skb->data_len += tmp_skb->len;
1394 skb->truesize += tmp_skb->truesize;
1395 tmp_skb->destructor = NULL;
1399 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1400 * to fragment the frame generated here. No matter, what transforms
1401 * how transforms change size of the packet, it will come out.
1403 skb->ignore_df = ip_sk_ignore_df(sk);
1405 /* DF bit is set when we want to see DF on outgoing frames.
1406 * If ignore_df is set too, we still allow to fragment this frame
1408 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1409 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1410 (skb->len <= dst_mtu(&rt->dst) &&
1411 ip_dont_fragment(sk, &rt->dst)))
1414 if (cork->flags & IPCORK_OPT)
1419 else if (rt->rt_type == RTN_MULTICAST)
1422 ttl = ip_select_ttl(inet, &rt->dst);
1427 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1430 iph->protocol = sk->sk_protocol;
1431 ip_copy_addrs(iph, fl4);
1432 ip_select_ident(net, skb, sk);
1435 iph->ihl += opt->optlen>>2;
1436 ip_options_build(skb, opt, cork->addr, rt, 0);
1439 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1440 skb->mark = sk->sk_mark;
1442 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1446 skb_dst_set(skb, &rt->dst);
1448 if (iph->protocol == IPPROTO_ICMP)
1449 icmp_out_count(net, ((struct icmphdr *)
1450 skb_transport_header(skb))->type);
1452 ip_cork_release(cork);
1457 int ip_send_skb(struct net *net, struct sk_buff *skb)
1461 err = ip_local_out(net, skb->sk, skb);
1464 err = net_xmit_errno(err);
1466 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1472 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1474 struct sk_buff *skb;
1476 skb = ip_finish_skb(sk, fl4);
1480 /* Netfilter gets whole the not fragmented skb. */
1481 return ip_send_skb(sock_net(sk), skb);
1485 * Throw away all pending data on the socket.
1487 static void __ip_flush_pending_frames(struct sock *sk,
1488 struct sk_buff_head *queue,
1489 struct inet_cork *cork)
1491 struct sk_buff *skb;
1493 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1496 ip_cork_release(cork);
1499 void ip_flush_pending_frames(struct sock *sk)
1501 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1504 struct sk_buff *ip_make_skb(struct sock *sk,
1506 int getfrag(void *from, char *to, int offset,
1507 int len, int odd, struct sk_buff *skb),
1508 void *from, int length, int transhdrlen,
1509 struct ipcm_cookie *ipc, struct rtable **rtp,
1512 struct inet_cork cork;
1513 struct sk_buff_head queue;
1516 if (flags & MSG_PROBE)
1519 __skb_queue_head_init(&queue);
1524 err = ip_setup_cork(sk, &cork, ipc, rtp);
1526 return ERR_PTR(err);
1528 err = __ip_append_data(sk, fl4, &queue, &cork,
1529 ¤t->task_frag, getfrag,
1530 from, length, transhdrlen, flags);
1532 __ip_flush_pending_frames(sk, &queue, &cork);
1533 return ERR_PTR(err);
1536 return __ip_make_skb(sk, fl4, &queue, &cork);
1540 * Fetch data from kernel space and fill in checksum if needed.
1542 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1543 int len, int odd, struct sk_buff *skb)
1547 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1548 skb->csum = csum_block_add(skb->csum, csum, odd);
1553 * Generic function to send a packet as reply to another packet.
1554 * Used to send some TCP resets/acks so far.
1556 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1557 const struct ip_options *sopt,
1558 __be32 daddr, __be32 saddr,
1559 const struct ip_reply_arg *arg,
1562 struct ip_options_data replyopts;
1563 struct ipcm_cookie ipc;
1565 struct rtable *rt = skb_rtable(skb);
1566 struct net *net = sock_net(sk);
1567 struct sk_buff *nskb;
1571 if (__ip_options_echo(&replyopts.opt.opt, skb, sopt))
1580 if (replyopts.opt.opt.optlen) {
1581 ipc.opt = &replyopts.opt;
1583 if (replyopts.opt.opt.srr)
1584 daddr = replyopts.opt.opt.faddr;
1587 oif = arg->bound_dev_if;
1588 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1591 flowi4_init_output(&fl4, oif,
1592 IP4_REPLY_MARK(net, skb->mark),
1594 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1595 ip_reply_arg_flowi_flags(arg),
1597 tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1598 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1599 rt = ip_route_output_key(net, &fl4);
1603 inet_sk(sk)->tos = arg->tos;
1605 sk->sk_priority = skb->priority;
1606 sk->sk_protocol = ip_hdr(skb)->protocol;
1607 sk->sk_bound_dev_if = arg->bound_dev_if;
1608 sk->sk_sndbuf = sysctl_wmem_default;
1609 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1610 len, 0, &ipc, &rt, MSG_DONTWAIT);
1611 if (unlikely(err)) {
1612 ip_flush_pending_frames(sk);
1616 nskb = skb_peek(&sk->sk_write_queue);
1618 if (arg->csumoffset >= 0)
1619 *((__sum16 *)skb_transport_header(nskb) +
1620 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1622 nskb->ip_summed = CHECKSUM_NONE;
1623 ip_push_pending_frames(sk, &fl4);
1629 void __init ip_init(void)
1634 #if defined(CONFIG_IP_MULTICAST)