2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * The Internet Protocol (IP) output module.
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Donald Becker, <becker@super.org>
11 * Alan Cox, <Alan.Cox@linux.org>
13 * Stefan Becker, <stefanb@yello.ping.de>
14 * Jorge Cwik, <jorge@laser.satlink.net>
15 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
16 * Hirokazu Takahashi, <taka@valinux.co.jp>
18 * See ip_input.c for original log
21 * Alan Cox : Missing nonblock feature in ip_build_xmit.
22 * Mike Kilburn : htons() missing in ip_build_xmit.
23 * Bradford Johnson: Fix faulty handling of some frames when
25 * Alexander Demenshin: Missing sk/skb free in ip_queue_xmit
26 * (in case if packet not accepted by
27 * output firewall rules)
28 * Mike McLagan : Routing by source
29 * Alexey Kuznetsov: use new route cache
30 * Andi Kleen: Fix broken PMTU recovery and remove
31 * some redundant tests.
32 * Vitaly E. Lavrov : Transparent proxy revived after year coma.
33 * Andi Kleen : Replace ip_reply with ip_send_reply.
34 * Andi Kleen : Split fast and slow ip_build_xmit path
35 * for decreased register pressure on x86
36 * and more readibility.
37 * Marc Boucher : When call_out_firewall returns FW_QUEUE,
38 * silently drop skb instead of failing with -EPERM.
39 * Detlev Wengorz : Copy protocol for fragments.
40 * Hirokazu Takahashi: HW checksumming for outgoing UDP
42 * Hirokazu Takahashi: sendfile() on UDP works now.
45 #include <linux/uaccess.h>
46 #include <linux/module.h>
47 #include <linux/types.h>
48 #include <linux/kernel.h>
50 #include <linux/string.h>
51 #include <linux/errno.h>
52 #include <linux/highmem.h>
53 #include <linux/slab.h>
55 #include <linux/socket.h>
56 #include <linux/sockios.h>
58 #include <linux/inet.h>
59 #include <linux/netdevice.h>
60 #include <linux/etherdevice.h>
61 #include <linux/proc_fs.h>
62 #include <linux/stat.h>
63 #include <linux/init.h>
67 #include <net/protocol.h>
68 #include <net/route.h>
70 #include <linux/skbuff.h>
74 #include <net/checksum.h>
75 #include <net/inetpeer.h>
76 #include <net/lwtunnel.h>
77 #include <linux/bpf-cgroup.h>
78 #include <linux/igmp.h>
79 #include <linux/netfilter_ipv4.h>
80 #include <linux/netfilter_bridge.h>
81 #include <linux/netlink.h>
82 #include <linux/tcp.h>
85 ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
87 int (*output)(struct net *, struct sock *, struct sk_buff *));
89 /* Generate a checksum for an outgoing IP datagram. */
90 void ip_send_check(struct iphdr *iph)
93 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
95 EXPORT_SYMBOL(ip_send_check);
97 int __ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
99 struct iphdr *iph = ip_hdr(skb);
101 iph->tot_len = htons(skb->len);
104 /* if egress device is enslaved to an L3 master device pass the
105 * skb to its handler for processing
107 skb = l3mdev_ip_out(sk, skb);
111 skb->protocol = htons(ETH_P_IP);
113 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
114 net, sk, skb, NULL, skb_dst(skb)->dev,
118 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
122 err = __ip_local_out(net, sk, skb);
123 if (likely(err == 1))
124 err = dst_output(net, sk, skb);
128 EXPORT_SYMBOL_GPL(ip_local_out);
130 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
132 int ttl = inet->uc_ttl;
135 ttl = ip4_dst_hoplimit(dst);
140 * Add an ip header to a skbuff and send it out.
143 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
144 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
146 struct inet_sock *inet = inet_sk(sk);
147 struct rtable *rt = skb_rtable(skb);
148 struct net *net = sock_net(sk);
151 /* Build the IP header. */
152 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
153 skb_reset_network_header(skb);
157 iph->tos = inet->tos;
158 iph->ttl = ip_select_ttl(inet, &rt->dst);
159 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
161 iph->protocol = sk->sk_protocol;
162 if (ip_dont_fragment(sk, &rt->dst)) {
163 iph->frag_off = htons(IP_DF);
167 __ip_select_ident(net, iph, 1);
170 if (opt && opt->opt.optlen) {
171 iph->ihl += opt->opt.optlen>>2;
172 ip_options_build(skb, &opt->opt, daddr, rt, 0);
175 skb->priority = sk->sk_priority;
177 skb->mark = sk->sk_mark;
180 return ip_local_out(net, skb->sk, skb);
182 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
184 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
186 struct dst_entry *dst = skb_dst(skb);
187 struct rtable *rt = (struct rtable *)dst;
188 struct net_device *dev = dst->dev;
189 unsigned int hh_len = LL_RESERVED_SPACE(dev);
190 struct neighbour *neigh;
193 if (rt->rt_type == RTN_MULTICAST) {
194 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
195 } else if (rt->rt_type == RTN_BROADCAST)
196 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
198 /* Be paranoid, rather than too clever. */
199 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
200 struct sk_buff *skb2;
202 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
208 skb_set_owner_w(skb2, skb->sk);
213 if (lwtunnel_xmit_redirect(dst->lwtstate)) {
214 int res = lwtunnel_xmit(skb);
216 if (res < 0 || res == LWTUNNEL_XMIT_DONE)
221 nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
222 neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
223 if (unlikely(!neigh))
224 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
225 if (!IS_ERR(neigh)) {
228 sock_confirm_neigh(skb, neigh);
229 res = neigh_output(neigh, skb);
231 rcu_read_unlock_bh();
234 rcu_read_unlock_bh();
236 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
242 static int ip_finish_output_gso(struct net *net, struct sock *sk,
243 struct sk_buff *skb, unsigned int mtu)
245 netdev_features_t features;
246 struct sk_buff *segs;
249 /* common case: seglen is <= mtu
251 if (skb_gso_validate_network_len(skb, mtu))
252 return ip_finish_output2(net, sk, skb);
254 /* Slowpath - GSO segment length exceeds the egress MTU.
256 * This can happen in several cases:
257 * - Forwarding of a TCP GRO skb, when DF flag is not set.
258 * - Forwarding of an skb that arrived on a virtualization interface
259 * (virtio-net/vhost/tap) with TSO/GSO size set by other network
261 * - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
262 * interface with a smaller MTU.
263 * - Arriving GRO skb (or GSO skb in a virtualized environment) that is
264 * bridged to a NETIF_F_TSO tunnel stacked over an interface with an
267 features = netif_skb_features(skb);
268 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_SGO_CB_OFFSET);
269 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
270 if (IS_ERR_OR_NULL(segs)) {
278 struct sk_buff *nskb = segs->next;
282 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
292 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
297 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
303 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
304 /* Policy lookup after SNAT yielded a new policy */
305 if (skb_dst(skb)->xfrm) {
306 IPCB(skb)->flags |= IPSKB_REROUTED;
307 return dst_output(net, sk, skb);
310 mtu = ip_skb_dst_mtu(sk, skb);
312 return ip_finish_output_gso(net, sk, skb, mtu);
314 if (skb->len > mtu || (IPCB(skb)->flags & IPSKB_FRAG_PMTU))
315 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
317 return ip_finish_output2(net, sk, skb);
320 static int ip_mc_finish_output(struct net *net, struct sock *sk,
325 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
331 return dev_loopback_xmit(net, sk, skb);
334 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
336 struct rtable *rt = skb_rtable(skb);
337 struct net_device *dev = rt->dst.dev;
340 * If the indicated interface is up and running, send the packet.
342 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
345 skb->protocol = htons(ETH_P_IP);
348 * Multicasts are looped back for other local users
351 if (rt->rt_flags&RTCF_MULTICAST) {
353 #ifdef CONFIG_IP_MROUTE
354 /* Small optimization: do not loopback not local frames,
355 which returned after forwarding; they will be dropped
356 by ip_mr_input in any case.
357 Note, that local frames are looped back to be delivered
360 This check is duplicated in ip_mr_input at the moment.
363 ((rt->rt_flags & RTCF_LOCAL) ||
364 !(IPCB(skb)->flags & IPSKB_FORWARDED))
367 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
369 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
370 net, sk, newskb, NULL, newskb->dev,
371 ip_mc_finish_output);
374 /* Multicasts with ttl 0 must not go beyond the host */
376 if (ip_hdr(skb)->ttl == 0) {
382 if (rt->rt_flags&RTCF_BROADCAST) {
383 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
385 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
386 net, sk, newskb, NULL, newskb->dev,
387 ip_mc_finish_output);
390 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
391 net, sk, skb, NULL, skb->dev,
393 !(IPCB(skb)->flags & IPSKB_REROUTED));
396 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
398 struct net_device *dev = skb_dst(skb)->dev;
400 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
403 skb->protocol = htons(ETH_P_IP);
405 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
406 net, sk, skb, NULL, dev,
408 !(IPCB(skb)->flags & IPSKB_REROUTED));
412 * copy saddr and daddr, possibly using 64bit load/stores
414 * iph->saddr = fl4->saddr;
415 * iph->daddr = fl4->daddr;
417 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
419 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
420 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
421 memcpy(&iph->saddr, &fl4->saddr,
422 sizeof(fl4->saddr) + sizeof(fl4->daddr));
425 /* Note: skb->sk can be different from sk, in case of tunnels */
426 int __ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl,
429 struct inet_sock *inet = inet_sk(sk);
430 struct net *net = sock_net(sk);
431 struct ip_options_rcu *inet_opt;
437 /* Skip all of this if the packet is already routed,
438 * f.e. by something like SCTP.
441 inet_opt = rcu_dereference(inet->inet_opt);
443 rt = skb_rtable(skb);
447 /* Make sure we can route this packet. */
448 rt = (struct rtable *)__sk_dst_check(sk, 0);
452 /* Use correct destination address if we have options. */
453 daddr = inet->inet_daddr;
454 if (inet_opt && inet_opt->opt.srr)
455 daddr = inet_opt->opt.faddr;
457 /* If this fails, retransmit mechanism of transport layer will
458 * keep trying until route appears or the connection times
461 rt = ip_route_output_ports(net, fl4, sk,
462 daddr, inet->inet_saddr,
466 RT_CONN_FLAGS_TOS(sk, tos),
467 sk->sk_bound_dev_if);
470 sk_setup_caps(sk, &rt->dst);
472 skb_dst_set_noref(skb, &rt->dst);
475 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
478 /* OK, we know where to send it, allocate and build IP header. */
479 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
480 skb_reset_network_header(skb);
482 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (tos & 0xff));
483 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
484 iph->frag_off = htons(IP_DF);
487 iph->ttl = ip_select_ttl(inet, &rt->dst);
488 iph->protocol = sk->sk_protocol;
489 ip_copy_addrs(iph, fl4);
491 /* Transport layer set skb->h.foo itself. */
493 if (inet_opt && inet_opt->opt.optlen) {
494 iph->ihl += inet_opt->opt.optlen >> 2;
495 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
498 ip_select_ident_segs(net, skb, sk,
499 skb_shinfo(skb)->gso_segs ?: 1);
501 /* TODO : should we use skb->sk here instead of sk ? */
502 skb->priority = sk->sk_priority;
503 skb->mark = sk->sk_mark;
505 res = ip_local_out(net, sk, skb);
511 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
513 return -EHOSTUNREACH;
515 EXPORT_SYMBOL(__ip_queue_xmit);
517 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
519 to->pkt_type = from->pkt_type;
520 to->priority = from->priority;
521 to->protocol = from->protocol;
523 skb_dst_copy(to, from);
525 to->mark = from->mark;
527 /* Copy the flags to each fragment. */
528 IPCB(to)->flags = IPCB(from)->flags;
530 #ifdef CONFIG_NET_SCHED
531 to->tc_index = from->tc_index;
534 #if IS_ENABLED(CONFIG_IP_VS)
535 to->ipvs_property = from->ipvs_property;
537 skb_copy_secmark(to, from);
540 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
542 int (*output)(struct net *, struct sock *, struct sk_buff *))
544 struct iphdr *iph = ip_hdr(skb);
546 if ((iph->frag_off & htons(IP_DF)) == 0)
547 return ip_do_fragment(net, sk, skb, output);
549 if (unlikely(!skb->ignore_df ||
550 (IPCB(skb)->frag_max_size &&
551 IPCB(skb)->frag_max_size > mtu))) {
552 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
553 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
559 return ip_do_fragment(net, sk, skb, output);
563 * This IP datagram is too large to be sent in one piece. Break it up into
564 * smaller pieces (each of size equal to IP header plus
565 * a block of the data of the original IP data part) that will yet fit in a
566 * single device frame, and queue such a frame for sending.
569 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
570 int (*output)(struct net *, struct sock *, struct sk_buff *))
574 struct sk_buff *skb2;
575 unsigned int mtu, hlen, left, len, ll_rs;
577 __be16 not_last_frag;
578 struct rtable *rt = skb_rtable(skb);
581 /* for offloaded checksums cleanup checksum before fragmentation */
582 if (skb->ip_summed == CHECKSUM_PARTIAL &&
583 (err = skb_checksum_help(skb)))
587 * Point into the IP datagram header.
592 mtu = ip_skb_dst_mtu(sk, skb);
593 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
594 mtu = IPCB(skb)->frag_max_size;
597 * Setup starting values.
601 mtu = mtu - hlen; /* Size of data space */
602 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
603 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
605 /* When frag_list is given, use it. First, check its validity:
606 * some transformers could create wrong frag_list or break existing
607 * one, it is not prohibited. In this case fall back to copying.
609 * LATER: this step can be merged to real generation of fragments,
610 * we can switch to copy when see the first bad fragment.
612 if (skb_has_frag_list(skb)) {
613 struct sk_buff *frag, *frag2;
614 unsigned int first_len = skb_pagelen(skb);
616 if (first_len - hlen > mtu ||
617 ((first_len - hlen) & 7) ||
618 ip_is_fragment(iph) ||
620 skb_headroom(skb) < ll_rs)
623 skb_walk_frags(skb, frag) {
624 /* Correct geometry. */
625 if (frag->len > mtu ||
626 ((frag->len & 7) && frag->next) ||
627 skb_headroom(frag) < hlen + ll_rs)
628 goto slow_path_clean;
630 /* Partially cloned skb? */
631 if (skb_shared(frag))
632 goto slow_path_clean;
637 frag->destructor = sock_wfree;
639 skb->truesize -= frag->truesize;
642 /* Everything is OK. Generate! */
646 frag = skb_shinfo(skb)->frag_list;
647 skb_frag_list_init(skb);
648 skb->data_len = first_len - skb_headlen(skb);
649 skb->len = first_len;
650 iph->tot_len = htons(first_len);
651 iph->frag_off = htons(IP_MF);
655 /* Prepare header of the next frame,
656 * before previous one went down. */
658 frag->ip_summed = CHECKSUM_NONE;
659 skb_reset_transport_header(frag);
660 __skb_push(frag, hlen);
661 skb_reset_network_header(frag);
662 memcpy(skb_network_header(frag), iph, hlen);
664 iph->tot_len = htons(frag->len);
665 ip_copy_metadata(frag, skb);
667 ip_options_fragment(frag);
668 offset += skb->len - hlen;
669 iph->frag_off = htons(offset>>3);
671 iph->frag_off |= htons(IP_MF);
672 /* Ready, complete checksum */
676 err = output(net, sk, skb);
679 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
689 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
698 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
702 skb_walk_frags(skb, frag2) {
706 frag2->destructor = NULL;
707 skb->truesize += frag2->truesize;
714 left = skb->len - hlen; /* Space per frame */
715 ptr = hlen; /* Where to start from */
718 * Fragment the datagram.
721 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
722 not_last_frag = iph->frag_off & htons(IP_MF);
725 * Keep copying data until we run out.
730 /* IF: it doesn't fit, use 'mtu' - the data space left */
733 /* IF: we are not sending up to and including the packet end
734 then align the next start on an eight byte boundary */
739 /* Allocate buffer */
740 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
747 * Set up data on packet
750 ip_copy_metadata(skb2, skb);
751 skb_reserve(skb2, ll_rs);
752 skb_put(skb2, len + hlen);
753 skb_reset_network_header(skb2);
754 skb2->transport_header = skb2->network_header + hlen;
757 * Charge the memory for the fragment to any owner
762 skb_set_owner_w(skb2, skb->sk);
765 * Copy the packet header into the new buffer.
768 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
771 * Copy a block of the IP datagram.
773 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
778 * Fill in the new header fields.
781 iph->frag_off = htons((offset >> 3));
783 if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
784 iph->frag_off |= htons(IP_DF);
786 /* ANK: dirty, but effective trick. Upgrade options only if
787 * the segment to be fragmented was THE FIRST (otherwise,
788 * options are already fixed) and make it ONCE
789 * on the initial skb, so that all the following fragments
790 * will inherit fixed options.
793 ip_options_fragment(skb);
796 * Added AC : If we are fragmenting a fragment that's not the
797 * last fragment then keep MF on each bit
799 if (left > 0 || not_last_frag)
800 iph->frag_off |= htons(IP_MF);
805 * Put this fragment into the sending queue.
807 iph->tot_len = htons(len + hlen);
811 err = output(net, sk, skb2);
815 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
818 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
823 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
826 EXPORT_SYMBOL(ip_do_fragment);
829 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
831 struct msghdr *msg = from;
833 if (skb->ip_summed == CHECKSUM_PARTIAL) {
834 if (!copy_from_iter_full(to, len, &msg->msg_iter))
838 if (!csum_and_copy_from_iter_full(to, len, &csum, &msg->msg_iter))
840 skb->csum = csum_block_add(skb->csum, csum, odd);
844 EXPORT_SYMBOL(ip_generic_getfrag);
847 csum_page(struct page *page, int offset, int copy)
852 csum = csum_partial(kaddr + offset, copy, 0);
857 static int __ip_append_data(struct sock *sk,
859 struct sk_buff_head *queue,
860 struct inet_cork *cork,
861 struct page_frag *pfrag,
862 int getfrag(void *from, char *to, int offset,
863 int len, int odd, struct sk_buff *skb),
864 void *from, int length, int transhdrlen,
867 struct inet_sock *inet = inet_sk(sk);
870 struct ip_options *opt = cork->opt;
877 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
878 int csummode = CHECKSUM_NONE;
879 struct rtable *rt = (struct rtable *)cork->dst;
880 unsigned int wmem_alloc_delta = 0;
884 skb = skb_peek_tail(queue);
886 exthdrlen = !skb ? rt->dst.header_len : 0;
887 mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
888 paged = !!cork->gso_size;
890 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
891 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
892 tskey = sk->sk_tskey++;
894 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
896 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
897 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
898 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
900 if (cork->length + length > maxnonfragsize - fragheaderlen) {
901 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
902 mtu - (opt ? opt->optlen : 0));
907 * transhdrlen > 0 means that this is the first fragment and we wish
908 * it won't be fragmented in the future.
911 length + fragheaderlen <= mtu &&
912 rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
913 (!(flags & MSG_MORE) || cork->gso_size) &&
914 (!exthdrlen || (rt->dst.dev->features & NETIF_F_HW_ESP_TX_CSUM)))
915 csummode = CHECKSUM_PARTIAL;
917 cork->length += length;
919 /* So, what's going on in the loop below?
921 * We use calculated fragment length to generate chained skb,
922 * each of segments is IP fragment ready for sending to network after
923 * adding appropriate IP header.
930 /* Check if the remaining data fits into current packet. */
931 copy = mtu - skb->len;
933 copy = maxfraglen - skb->len;
936 unsigned int datalen;
937 unsigned int fraglen;
938 unsigned int fraggap;
939 unsigned int alloclen;
940 unsigned int pagedlen = 0;
941 struct sk_buff *skb_prev;
945 fraggap = skb_prev->len - maxfraglen;
950 * If remaining data exceeds the mtu,
951 * we know we need more fragment(s).
953 datalen = length + fraggap;
954 if (datalen > mtu - fragheaderlen)
955 datalen = maxfraglen - fragheaderlen;
956 fraglen = datalen + fragheaderlen;
958 if ((flags & MSG_MORE) &&
959 !(rt->dst.dev->features&NETIF_F_SG))
964 alloclen = min_t(int, fraglen, MAX_HEADER);
965 pagedlen = fraglen - alloclen;
968 alloclen += exthdrlen;
970 /* The last fragment gets additional space at tail.
971 * Note, with MSG_MORE we overallocate on fragments,
972 * because we have no idea what fragment will be
975 if (datalen == length + fraggap)
976 alloclen += rt->dst.trailer_len;
979 skb = sock_alloc_send_skb(sk,
980 alloclen + hh_len + 15,
981 (flags & MSG_DONTWAIT), &err);
984 if (refcount_read(&sk->sk_wmem_alloc) + wmem_alloc_delta <=
986 skb = alloc_skb(alloclen + hh_len + 15,
995 * Fill in the control structures
997 skb->ip_summed = csummode;
999 skb_reserve(skb, hh_len);
1001 /* only the initial fragment is time stamped */
1002 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1004 skb_shinfo(skb)->tskey = tskey;
1008 * Find where to start putting bytes.
1010 data = skb_put(skb, fraglen + exthdrlen - pagedlen);
1011 skb_set_network_header(skb, exthdrlen);
1012 skb->transport_header = (skb->network_header +
1014 data += fragheaderlen + exthdrlen;
1017 skb->csum = skb_copy_and_csum_bits(
1018 skb_prev, maxfraglen,
1019 data + transhdrlen, fraggap, 0);
1020 skb_prev->csum = csum_sub(skb_prev->csum,
1023 pskb_trim_unique(skb_prev, maxfraglen);
1026 copy = datalen - transhdrlen - fraggap - pagedlen;
1027 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1034 length -= copy + transhdrlen;
1037 csummode = CHECKSUM_NONE;
1039 if ((flags & MSG_CONFIRM) && !skb_prev)
1040 skb_set_dst_pending_confirm(skb, 1);
1043 * Put the packet on the pending queue.
1045 if (!skb->destructor) {
1046 skb->destructor = sock_wfree;
1048 wmem_alloc_delta += skb->truesize;
1050 __skb_queue_tail(queue, skb);
1057 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1058 skb_tailroom(skb) >= copy) {
1062 if (getfrag(from, skb_put(skb, copy),
1063 offset, copy, off, skb) < 0) {
1064 __skb_trim(skb, off);
1069 int i = skb_shinfo(skb)->nr_frags;
1072 if (!sk_page_frag_refill(sk, pfrag))
1075 if (!skb_can_coalesce(skb, i, pfrag->page,
1078 if (i == MAX_SKB_FRAGS)
1081 __skb_fill_page_desc(skb, i, pfrag->page,
1083 skb_shinfo(skb)->nr_frags = ++i;
1084 get_page(pfrag->page);
1086 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1088 page_address(pfrag->page) + pfrag->offset,
1089 offset, copy, skb->len, skb) < 0)
1092 pfrag->offset += copy;
1093 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1095 skb->data_len += copy;
1096 skb->truesize += copy;
1097 wmem_alloc_delta += copy;
1103 if (wmem_alloc_delta)
1104 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1110 cork->length -= length;
1111 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1112 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1116 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1117 struct ipcm_cookie *ipc, struct rtable **rtp)
1119 struct ip_options_rcu *opt;
1127 * setup for corking.
1132 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1134 if (unlikely(!cork->opt))
1137 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1138 cork->flags |= IPCORK_OPT;
1139 cork->addr = ipc->addr;
1143 * We steal reference to this route, caller should not release it
1146 cork->fragsize = ip_sk_use_pmtu(sk) ?
1147 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1149 cork->gso_size = sk->sk_type == SOCK_DGRAM &&
1150 sk->sk_protocol == IPPROTO_UDP ? ipc->gso_size : 0;
1151 cork->dst = &rt->dst;
1153 cork->ttl = ipc->ttl;
1154 cork->tos = ipc->tos;
1155 cork->priority = ipc->priority;
1156 cork->tx_flags = ipc->tx_flags;
1157 cork->transmit_time = ipc->sockc.transmit_time;
1163 * ip_append_data() and ip_append_page() can make one large IP datagram
1164 * from many pieces of data. Each pieces will be holded on the socket
1165 * until ip_push_pending_frames() is called. Each piece can be a page
1168 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1169 * this interface potentially.
1171 * LATER: length must be adjusted by pad at tail, when it is required.
1173 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1174 int getfrag(void *from, char *to, int offset, int len,
1175 int odd, struct sk_buff *skb),
1176 void *from, int length, int transhdrlen,
1177 struct ipcm_cookie *ipc, struct rtable **rtp,
1180 struct inet_sock *inet = inet_sk(sk);
1183 if (flags&MSG_PROBE)
1186 if (skb_queue_empty(&sk->sk_write_queue)) {
1187 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1194 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1195 sk_page_frag(sk), getfrag,
1196 from, length, transhdrlen, flags);
1199 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1200 int offset, size_t size, int flags)
1202 struct inet_sock *inet = inet_sk(sk);
1203 struct sk_buff *skb;
1205 struct ip_options *opt = NULL;
1206 struct inet_cork *cork;
1211 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1216 if (flags&MSG_PROBE)
1219 if (skb_queue_empty(&sk->sk_write_queue))
1222 cork = &inet->cork.base;
1223 rt = (struct rtable *)cork->dst;
1224 if (cork->flags & IPCORK_OPT)
1227 if (!(rt->dst.dev->features&NETIF_F_SG))
1230 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1231 mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
1233 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1234 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1235 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1237 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1238 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1239 mtu - (opt ? opt->optlen : 0));
1243 skb = skb_peek_tail(&sk->sk_write_queue);
1247 cork->length += size;
1250 /* Check if the remaining data fits into current packet. */
1251 len = mtu - skb->len;
1253 len = maxfraglen - skb->len;
1256 struct sk_buff *skb_prev;
1260 fraggap = skb_prev->len - maxfraglen;
1262 alloclen = fragheaderlen + hh_len + fraggap + 15;
1263 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1264 if (unlikely(!skb)) {
1270 * Fill in the control structures
1272 skb->ip_summed = CHECKSUM_NONE;
1274 skb_reserve(skb, hh_len);
1277 * Find where to start putting bytes.
1279 skb_put(skb, fragheaderlen + fraggap);
1280 skb_reset_network_header(skb);
1281 skb->transport_header = (skb->network_header +
1284 skb->csum = skb_copy_and_csum_bits(skb_prev,
1286 skb_transport_header(skb),
1288 skb_prev->csum = csum_sub(skb_prev->csum,
1290 pskb_trim_unique(skb_prev, maxfraglen);
1294 * Put the packet on the pending queue.
1296 __skb_queue_tail(&sk->sk_write_queue, skb);
1303 if (skb_append_pagefrags(skb, page, offset, len)) {
1308 if (skb->ip_summed == CHECKSUM_NONE) {
1310 csum = csum_page(page, offset, len);
1311 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1315 skb->data_len += len;
1316 skb->truesize += len;
1317 refcount_add(len, &sk->sk_wmem_alloc);
1324 cork->length -= size;
1325 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1329 static void ip_cork_release(struct inet_cork *cork)
1331 cork->flags &= ~IPCORK_OPT;
1334 dst_release(cork->dst);
1339 * Combined all pending IP fragments on the socket as one IP datagram
1340 * and push them out.
1342 struct sk_buff *__ip_make_skb(struct sock *sk,
1344 struct sk_buff_head *queue,
1345 struct inet_cork *cork)
1347 struct sk_buff *skb, *tmp_skb;
1348 struct sk_buff **tail_skb;
1349 struct inet_sock *inet = inet_sk(sk);
1350 struct net *net = sock_net(sk);
1351 struct ip_options *opt = NULL;
1352 struct rtable *rt = (struct rtable *)cork->dst;
1357 skb = __skb_dequeue(queue);
1360 tail_skb = &(skb_shinfo(skb)->frag_list);
1362 /* move skb->data to ip header from ext header */
1363 if (skb->data < skb_network_header(skb))
1364 __skb_pull(skb, skb_network_offset(skb));
1365 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1366 __skb_pull(tmp_skb, skb_network_header_len(skb));
1367 *tail_skb = tmp_skb;
1368 tail_skb = &(tmp_skb->next);
1369 skb->len += tmp_skb->len;
1370 skb->data_len += tmp_skb->len;
1371 skb->truesize += tmp_skb->truesize;
1372 tmp_skb->destructor = NULL;
1376 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1377 * to fragment the frame generated here. No matter, what transforms
1378 * how transforms change size of the packet, it will come out.
1380 skb->ignore_df = ip_sk_ignore_df(sk);
1382 /* DF bit is set when we want to see DF on outgoing frames.
1383 * If ignore_df is set too, we still allow to fragment this frame
1385 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1386 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1387 (skb->len <= dst_mtu(&rt->dst) &&
1388 ip_dont_fragment(sk, &rt->dst)))
1391 if (cork->flags & IPCORK_OPT)
1396 else if (rt->rt_type == RTN_MULTICAST)
1399 ttl = ip_select_ttl(inet, &rt->dst);
1404 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1407 iph->protocol = sk->sk_protocol;
1408 ip_copy_addrs(iph, fl4);
1409 ip_select_ident(net, skb, sk);
1412 iph->ihl += opt->optlen>>2;
1413 ip_options_build(skb, opt, cork->addr, rt, 0);
1416 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1417 skb->mark = sk->sk_mark;
1418 skb->tstamp = cork->transmit_time;
1420 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1424 skb_dst_set(skb, &rt->dst);
1426 if (iph->protocol == IPPROTO_ICMP)
1427 icmp_out_count(net, ((struct icmphdr *)
1428 skb_transport_header(skb))->type);
1430 ip_cork_release(cork);
1435 int ip_send_skb(struct net *net, struct sk_buff *skb)
1439 err = ip_local_out(net, skb->sk, skb);
1442 err = net_xmit_errno(err);
1444 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1450 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1452 struct sk_buff *skb;
1454 skb = ip_finish_skb(sk, fl4);
1458 /* Netfilter gets whole the not fragmented skb. */
1459 return ip_send_skb(sock_net(sk), skb);
1463 * Throw away all pending data on the socket.
1465 static void __ip_flush_pending_frames(struct sock *sk,
1466 struct sk_buff_head *queue,
1467 struct inet_cork *cork)
1469 struct sk_buff *skb;
1471 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1474 ip_cork_release(cork);
1477 void ip_flush_pending_frames(struct sock *sk)
1479 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1482 struct sk_buff *ip_make_skb(struct sock *sk,
1484 int getfrag(void *from, char *to, int offset,
1485 int len, int odd, struct sk_buff *skb),
1486 void *from, int length, int transhdrlen,
1487 struct ipcm_cookie *ipc, struct rtable **rtp,
1488 struct inet_cork *cork, unsigned int flags)
1490 struct sk_buff_head queue;
1493 if (flags & MSG_PROBE)
1496 __skb_queue_head_init(&queue);
1501 err = ip_setup_cork(sk, cork, ipc, rtp);
1503 return ERR_PTR(err);
1505 err = __ip_append_data(sk, fl4, &queue, cork,
1506 ¤t->task_frag, getfrag,
1507 from, length, transhdrlen, flags);
1509 __ip_flush_pending_frames(sk, &queue, cork);
1510 return ERR_PTR(err);
1513 return __ip_make_skb(sk, fl4, &queue, cork);
1517 * Fetch data from kernel space and fill in checksum if needed.
1519 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1520 int len, int odd, struct sk_buff *skb)
1524 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1525 skb->csum = csum_block_add(skb->csum, csum, odd);
1530 * Generic function to send a packet as reply to another packet.
1531 * Used to send some TCP resets/acks so far.
1533 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1534 const struct ip_options *sopt,
1535 __be32 daddr, __be32 saddr,
1536 const struct ip_reply_arg *arg,
1539 struct ip_options_data replyopts;
1540 struct ipcm_cookie ipc;
1542 struct rtable *rt = skb_rtable(skb);
1543 struct net *net = sock_net(sk);
1544 struct sk_buff *nskb;
1548 if (__ip_options_echo(net, &replyopts.opt.opt, skb, sopt))
1556 ipc.sockc.transmit_time = 0;
1558 if (replyopts.opt.opt.optlen) {
1559 ipc.opt = &replyopts.opt;
1561 if (replyopts.opt.opt.srr)
1562 daddr = replyopts.opt.opt.faddr;
1565 oif = arg->bound_dev_if;
1566 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1569 flowi4_init_output(&fl4, oif,
1570 IP4_REPLY_MARK(net, skb->mark) ?: sk->sk_mark,
1572 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1573 ip_reply_arg_flowi_flags(arg),
1575 tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1577 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1578 rt = ip_route_output_key(net, &fl4);
1582 inet_sk(sk)->tos = arg->tos;
1584 sk->sk_priority = skb->priority;
1585 sk->sk_protocol = ip_hdr(skb)->protocol;
1586 sk->sk_bound_dev_if = arg->bound_dev_if;
1587 sk->sk_sndbuf = sysctl_wmem_default;
1588 sk->sk_mark = fl4.flowi4_mark;
1589 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1590 len, 0, &ipc, &rt, MSG_DONTWAIT);
1591 if (unlikely(err)) {
1592 ip_flush_pending_frames(sk);
1596 nskb = skb_peek(&sk->sk_write_queue);
1598 if (arg->csumoffset >= 0)
1599 *((__sum16 *)skb_transport_header(nskb) +
1600 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1602 nskb->ip_summed = CHECKSUM_NONE;
1603 ip_push_pending_frames(sk, &fl4);
1609 void __init ip_init(void)
1614 #if defined(CONFIG_IP_MULTICAST)