1 // SPDX-License-Identifier: GPL-2.0-only
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * The Internet Protocol (IP) output module.
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Donald Becker, <becker@super.org>
12 * Alan Cox, <Alan.Cox@linux.org>
14 * Stefan Becker, <stefanb@yello.ping.de>
15 * Jorge Cwik, <jorge@laser.satlink.net>
16 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
17 * Hirokazu Takahashi, <taka@valinux.co.jp>
19 * See ip_input.c for original log
22 * Alan Cox : Missing nonblock feature in ip_build_xmit.
23 * Mike Kilburn : htons() missing in ip_build_xmit.
24 * Bradford Johnson: Fix faulty handling of some frames when
26 * Alexander Demenshin: Missing sk/skb free in ip_queue_xmit
27 * (in case if packet not accepted by
28 * output firewall rules)
29 * Mike McLagan : Routing by source
30 * Alexey Kuznetsov: use new route cache
31 * Andi Kleen: Fix broken PMTU recovery and remove
32 * some redundant tests.
33 * Vitaly E. Lavrov : Transparent proxy revived after year coma.
34 * Andi Kleen : Replace ip_reply with ip_send_reply.
35 * Andi Kleen : Split fast and slow ip_build_xmit path
36 * for decreased register pressure on x86
37 * and more readibility.
38 * Marc Boucher : When call_out_firewall returns FW_QUEUE,
39 * silently drop skb instead of failing with -EPERM.
40 * Detlev Wengorz : Copy protocol for fragments.
41 * Hirokazu Takahashi: HW checksumming for outgoing UDP
43 * Hirokazu Takahashi: sendfile() on UDP works now.
46 #include <linux/uaccess.h>
47 #include <linux/module.h>
48 #include <linux/types.h>
49 #include <linux/kernel.h>
51 #include <linux/string.h>
52 #include <linux/errno.h>
53 #include <linux/highmem.h>
54 #include <linux/slab.h>
56 #include <linux/socket.h>
57 #include <linux/sockios.h>
59 #include <linux/inet.h>
60 #include <linux/netdevice.h>
61 #include <linux/etherdevice.h>
62 #include <linux/proc_fs.h>
63 #include <linux/stat.h>
64 #include <linux/init.h>
68 #include <net/protocol.h>
69 #include <net/route.h>
71 #include <linux/skbuff.h>
75 #include <net/checksum.h>
76 #include <net/inetpeer.h>
77 #include <net/lwtunnel.h>
78 #include <linux/bpf-cgroup.h>
79 #include <linux/igmp.h>
80 #include <linux/netfilter_ipv4.h>
81 #include <linux/netfilter_bridge.h>
82 #include <linux/netlink.h>
83 #include <linux/tcp.h>
86 ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
88 int (*output)(struct net *, struct sock *, struct sk_buff *));
90 /* Generate a checksum for an outgoing IP datagram. */
91 void ip_send_check(struct iphdr *iph)
94 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
96 EXPORT_SYMBOL(ip_send_check);
98 int __ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
100 struct iphdr *iph = ip_hdr(skb);
102 iph->tot_len = htons(skb->len);
105 /* if egress device is enslaved to an L3 master device pass the
106 * skb to its handler for processing
108 skb = l3mdev_ip_out(sk, skb);
112 skb->protocol = htons(ETH_P_IP);
114 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
115 net, sk, skb, NULL, skb_dst(skb)->dev,
119 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
123 err = __ip_local_out(net, sk, skb);
124 if (likely(err == 1))
125 err = dst_output(net, sk, skb);
129 EXPORT_SYMBOL_GPL(ip_local_out);
131 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
133 int ttl = inet->uc_ttl;
136 ttl = ip4_dst_hoplimit(dst);
141 * Add an ip header to a skbuff and send it out.
144 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
145 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
147 struct inet_sock *inet = inet_sk(sk);
148 struct rtable *rt = skb_rtable(skb);
149 struct net *net = sock_net(sk);
152 /* Build the IP header. */
153 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
154 skb_reset_network_header(skb);
158 iph->tos = inet->tos;
159 iph->ttl = ip_select_ttl(inet, &rt->dst);
160 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
162 iph->protocol = sk->sk_protocol;
163 if (ip_dont_fragment(sk, &rt->dst)) {
164 iph->frag_off = htons(IP_DF);
168 __ip_select_ident(net, iph, 1);
171 if (opt && opt->opt.optlen) {
172 iph->ihl += opt->opt.optlen>>2;
173 ip_options_build(skb, &opt->opt, daddr, rt, 0);
176 skb->priority = sk->sk_priority;
178 skb->mark = sk->sk_mark;
181 return ip_local_out(net, skb->sk, skb);
183 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
185 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
187 struct dst_entry *dst = skb_dst(skb);
188 struct rtable *rt = (struct rtable *)dst;
189 struct net_device *dev = dst->dev;
190 unsigned int hh_len = LL_RESERVED_SPACE(dev);
191 struct neighbour *neigh;
192 bool is_v6gw = false;
194 if (rt->rt_type == RTN_MULTICAST) {
195 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
196 } else if (rt->rt_type == RTN_BROADCAST)
197 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
199 /* Be paranoid, rather than too clever. */
200 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
201 struct sk_buff *skb2;
203 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
209 skb_set_owner_w(skb2, skb->sk);
214 if (lwtunnel_xmit_redirect(dst->lwtstate)) {
215 int res = lwtunnel_xmit(skb);
217 if (res < 0 || res == LWTUNNEL_XMIT_DONE)
222 neigh = ip_neigh_for_gw(rt, skb, &is_v6gw);
223 if (!IS_ERR(neigh)) {
226 sock_confirm_neigh(skb, neigh);
227 /* if crossing protocols, can not use the cached header */
228 res = neigh_output(neigh, skb, is_v6gw);
229 rcu_read_unlock_bh();
232 rcu_read_unlock_bh();
234 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
240 static int ip_finish_output_gso(struct net *net, struct sock *sk,
241 struct sk_buff *skb, unsigned int mtu)
243 struct sk_buff *segs, *nskb;
244 netdev_features_t features;
247 /* common case: seglen is <= mtu
249 if (skb_gso_validate_network_len(skb, mtu))
250 return ip_finish_output2(net, sk, skb);
252 /* Slowpath - GSO segment length exceeds the egress MTU.
254 * This can happen in several cases:
255 * - Forwarding of a TCP GRO skb, when DF flag is not set.
256 * - Forwarding of an skb that arrived on a virtualization interface
257 * (virtio-net/vhost/tap) with TSO/GSO size set by other network
259 * - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
260 * interface with a smaller MTU.
261 * - Arriving GRO skb (or GSO skb in a virtualized environment) that is
262 * bridged to a NETIF_F_TSO tunnel stacked over an interface with an
265 features = netif_skb_features(skb);
266 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_GSO_CB_OFFSET);
267 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
268 if (IS_ERR_OR_NULL(segs)) {
275 skb_list_walk_safe(segs, segs, nskb) {
278 skb_mark_not_on_list(segs);
279 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
288 static int __ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
292 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
293 /* Policy lookup after SNAT yielded a new policy */
294 if (skb_dst(skb)->xfrm) {
295 IPCB(skb)->flags |= IPSKB_REROUTED;
296 return dst_output(net, sk, skb);
299 mtu = ip_skb_dst_mtu(sk, skb);
301 return ip_finish_output_gso(net, sk, skb, mtu);
303 if (skb->len > mtu || (IPCB(skb)->flags & IPSKB_FRAG_PMTU))
304 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
306 return ip_finish_output2(net, sk, skb);
309 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
313 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
315 case NET_XMIT_SUCCESS:
316 return __ip_finish_output(net, sk, skb);
318 return __ip_finish_output(net, sk, skb) ? : ret;
325 static int ip_mc_finish_output(struct net *net, struct sock *sk,
328 struct rtable *new_rt;
332 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
337 case NET_XMIT_SUCCESS:
344 /* Reset rt_iif so that inet_iif() will return skb->skb_iif. Setting
345 * this to non-zero causes ipi_ifindex in in_pktinfo to be overwritten,
346 * see ipv4_pktinfo_prepare().
348 new_rt = rt_dst_clone(net->loopback_dev, skb_rtable(skb));
352 skb_dst_set(skb, &new_rt->dst);
355 err = dev_loopback_xmit(net, sk, skb);
356 return (do_cn && err) ? ret : err;
359 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
361 struct rtable *rt = skb_rtable(skb);
362 struct net_device *dev = rt->dst.dev;
365 * If the indicated interface is up and running, send the packet.
367 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
370 skb->protocol = htons(ETH_P_IP);
373 * Multicasts are looped back for other local users
376 if (rt->rt_flags&RTCF_MULTICAST) {
378 #ifdef CONFIG_IP_MROUTE
379 /* Small optimization: do not loopback not local frames,
380 which returned after forwarding; they will be dropped
381 by ip_mr_input in any case.
382 Note, that local frames are looped back to be delivered
385 This check is duplicated in ip_mr_input at the moment.
388 ((rt->rt_flags & RTCF_LOCAL) ||
389 !(IPCB(skb)->flags & IPSKB_FORWARDED))
392 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
394 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
395 net, sk, newskb, NULL, newskb->dev,
396 ip_mc_finish_output);
399 /* Multicasts with ttl 0 must not go beyond the host */
401 if (ip_hdr(skb)->ttl == 0) {
407 if (rt->rt_flags&RTCF_BROADCAST) {
408 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
410 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
411 net, sk, newskb, NULL, newskb->dev,
412 ip_mc_finish_output);
415 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
416 net, sk, skb, NULL, skb->dev,
418 !(IPCB(skb)->flags & IPSKB_REROUTED));
421 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
423 struct net_device *dev = skb_dst(skb)->dev, *indev = skb->dev;
425 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
428 skb->protocol = htons(ETH_P_IP);
430 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
431 net, sk, skb, indev, dev,
433 !(IPCB(skb)->flags & IPSKB_REROUTED));
437 * copy saddr and daddr, possibly using 64bit load/stores
439 * iph->saddr = fl4->saddr;
440 * iph->daddr = fl4->daddr;
442 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
444 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
445 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
446 memcpy(&iph->saddr, &fl4->saddr,
447 sizeof(fl4->saddr) + sizeof(fl4->daddr));
450 /* Note: skb->sk can be different from sk, in case of tunnels */
451 int __ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl,
454 struct inet_sock *inet = inet_sk(sk);
455 struct net *net = sock_net(sk);
456 struct ip_options_rcu *inet_opt;
462 /* Skip all of this if the packet is already routed,
463 * f.e. by something like SCTP.
466 inet_opt = rcu_dereference(inet->inet_opt);
468 rt = skb_rtable(skb);
472 /* Make sure we can route this packet. */
473 rt = (struct rtable *)__sk_dst_check(sk, 0);
477 /* Use correct destination address if we have options. */
478 daddr = inet->inet_daddr;
479 if (inet_opt && inet_opt->opt.srr)
480 daddr = inet_opt->opt.faddr;
482 /* If this fails, retransmit mechanism of transport layer will
483 * keep trying until route appears or the connection times
486 rt = ip_route_output_ports(net, fl4, sk,
487 daddr, inet->inet_saddr,
491 RT_CONN_FLAGS_TOS(sk, tos),
492 sk->sk_bound_dev_if);
495 sk_setup_caps(sk, &rt->dst);
497 skb_dst_set_noref(skb, &rt->dst);
500 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
503 /* OK, we know where to send it, allocate and build IP header. */
504 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
505 skb_reset_network_header(skb);
507 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (tos & 0xff));
508 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
509 iph->frag_off = htons(IP_DF);
512 iph->ttl = ip_select_ttl(inet, &rt->dst);
513 iph->protocol = sk->sk_protocol;
514 ip_copy_addrs(iph, fl4);
516 /* Transport layer set skb->h.foo itself. */
518 if (inet_opt && inet_opt->opt.optlen) {
519 iph->ihl += inet_opt->opt.optlen >> 2;
520 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
523 ip_select_ident_segs(net, skb, sk,
524 skb_shinfo(skb)->gso_segs ?: 1);
526 /* TODO : should we use skb->sk here instead of sk ? */
527 skb->priority = sk->sk_priority;
528 skb->mark = sk->sk_mark;
530 res = ip_local_out(net, sk, skb);
536 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
538 return -EHOSTUNREACH;
540 EXPORT_SYMBOL(__ip_queue_xmit);
542 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
544 return __ip_queue_xmit(sk, skb, fl, inet_sk(sk)->tos);
546 EXPORT_SYMBOL(ip_queue_xmit);
548 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
550 to->pkt_type = from->pkt_type;
551 to->priority = from->priority;
552 to->protocol = from->protocol;
553 to->skb_iif = from->skb_iif;
555 skb_dst_copy(to, from);
557 to->mark = from->mark;
559 skb_copy_hash(to, from);
561 #ifdef CONFIG_NET_SCHED
562 to->tc_index = from->tc_index;
565 skb_ext_copy(to, from);
566 #if IS_ENABLED(CONFIG_IP_VS)
567 to->ipvs_property = from->ipvs_property;
569 skb_copy_secmark(to, from);
572 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
574 int (*output)(struct net *, struct sock *, struct sk_buff *))
576 struct iphdr *iph = ip_hdr(skb);
578 if ((iph->frag_off & htons(IP_DF)) == 0)
579 return ip_do_fragment(net, sk, skb, output);
581 if (unlikely(!skb->ignore_df ||
582 (IPCB(skb)->frag_max_size &&
583 IPCB(skb)->frag_max_size > mtu))) {
584 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
585 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
591 return ip_do_fragment(net, sk, skb, output);
594 void ip_fraglist_init(struct sk_buff *skb, struct iphdr *iph,
595 unsigned int hlen, struct ip_fraglist_iter *iter)
597 unsigned int first_len = skb_pagelen(skb);
599 iter->frag = skb_shinfo(skb)->frag_list;
600 skb_frag_list_init(skb);
606 skb->data_len = first_len - skb_headlen(skb);
607 skb->len = first_len;
608 iph->tot_len = htons(first_len);
609 iph->frag_off = htons(IP_MF);
612 EXPORT_SYMBOL(ip_fraglist_init);
614 static void ip_fraglist_ipcb_prepare(struct sk_buff *skb,
615 struct ip_fraglist_iter *iter)
617 struct sk_buff *to = iter->frag;
619 /* Copy the flags to each fragment. */
620 IPCB(to)->flags = IPCB(skb)->flags;
622 if (iter->offset == 0)
623 ip_options_fragment(to);
626 void ip_fraglist_prepare(struct sk_buff *skb, struct ip_fraglist_iter *iter)
628 unsigned int hlen = iter->hlen;
629 struct iphdr *iph = iter->iph;
630 struct sk_buff *frag;
633 frag->ip_summed = CHECKSUM_NONE;
634 skb_reset_transport_header(frag);
635 __skb_push(frag, hlen);
636 skb_reset_network_header(frag);
637 memcpy(skb_network_header(frag), iph, hlen);
638 iter->iph = ip_hdr(frag);
640 iph->tot_len = htons(frag->len);
641 ip_copy_metadata(frag, skb);
642 iter->offset += skb->len - hlen;
643 iph->frag_off = htons(iter->offset >> 3);
645 iph->frag_off |= htons(IP_MF);
646 /* Ready, complete checksum */
649 EXPORT_SYMBOL(ip_fraglist_prepare);
651 void ip_frag_init(struct sk_buff *skb, unsigned int hlen,
652 unsigned int ll_rs, unsigned int mtu, bool DF,
653 struct ip_frag_state *state)
655 struct iphdr *iph = ip_hdr(skb);
659 state->ll_rs = ll_rs;
662 state->left = skb->len - hlen; /* Space per frame */
663 state->ptr = hlen; /* Where to start from */
665 state->offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
666 state->not_last_frag = iph->frag_off & htons(IP_MF);
668 EXPORT_SYMBOL(ip_frag_init);
670 static void ip_frag_ipcb(struct sk_buff *from, struct sk_buff *to,
671 bool first_frag, struct ip_frag_state *state)
673 /* Copy the flags to each fragment. */
674 IPCB(to)->flags = IPCB(from)->flags;
676 /* ANK: dirty, but effective trick. Upgrade options only if
677 * the segment to be fragmented was THE FIRST (otherwise,
678 * options are already fixed) and make it ONCE
679 * on the initial skb, so that all the following fragments
680 * will inherit fixed options.
683 ip_options_fragment(from);
686 struct sk_buff *ip_frag_next(struct sk_buff *skb, struct ip_frag_state *state)
688 unsigned int len = state->left;
689 struct sk_buff *skb2;
693 /* IF: it doesn't fit, use 'mtu' - the data space left */
694 if (len > state->mtu)
696 /* IF: we are not sending up to and including the packet end
697 then align the next start on an eight byte boundary */
698 if (len < state->left) {
702 /* Allocate buffer */
703 skb2 = alloc_skb(len + state->hlen + state->ll_rs, GFP_ATOMIC);
705 return ERR_PTR(-ENOMEM);
708 * Set up data on packet
711 ip_copy_metadata(skb2, skb);
712 skb_reserve(skb2, state->ll_rs);
713 skb_put(skb2, len + state->hlen);
714 skb_reset_network_header(skb2);
715 skb2->transport_header = skb2->network_header + state->hlen;
718 * Charge the memory for the fragment to any owner
723 skb_set_owner_w(skb2, skb->sk);
726 * Copy the packet header into the new buffer.
729 skb_copy_from_linear_data(skb, skb_network_header(skb2), state->hlen);
732 * Copy a block of the IP datagram.
734 if (skb_copy_bits(skb, state->ptr, skb_transport_header(skb2), len))
739 * Fill in the new header fields.
742 iph->frag_off = htons((state->offset >> 3));
744 iph->frag_off |= htons(IP_DF);
747 * Added AC : If we are fragmenting a fragment that's not the
748 * last fragment then keep MF on each bit
750 if (state->left > 0 || state->not_last_frag)
751 iph->frag_off |= htons(IP_MF);
753 state->offset += len;
755 iph->tot_len = htons(len + state->hlen);
761 EXPORT_SYMBOL(ip_frag_next);
764 * This IP datagram is too large to be sent in one piece. Break it up into
765 * smaller pieces (each of size equal to IP header plus
766 * a block of the data of the original IP data part) that will yet fit in a
767 * single device frame, and queue such a frame for sending.
770 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
771 int (*output)(struct net *, struct sock *, struct sk_buff *))
774 struct sk_buff *skb2;
775 struct rtable *rt = skb_rtable(skb);
776 unsigned int mtu, hlen, ll_rs;
777 struct ip_fraglist_iter iter;
778 ktime_t tstamp = skb->tstamp;
779 struct ip_frag_state state;
782 /* for offloaded checksums cleanup checksum before fragmentation */
783 if (skb->ip_summed == CHECKSUM_PARTIAL &&
784 (err = skb_checksum_help(skb)))
788 * Point into the IP datagram header.
793 mtu = ip_skb_dst_mtu(sk, skb);
794 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
795 mtu = IPCB(skb)->frag_max_size;
798 * Setup starting values.
802 mtu = mtu - hlen; /* Size of data space */
803 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
804 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
806 /* When frag_list is given, use it. First, check its validity:
807 * some transformers could create wrong frag_list or break existing
808 * one, it is not prohibited. In this case fall back to copying.
810 * LATER: this step can be merged to real generation of fragments,
811 * we can switch to copy when see the first bad fragment.
813 if (skb_has_frag_list(skb)) {
814 struct sk_buff *frag, *frag2;
815 unsigned int first_len = skb_pagelen(skb);
817 if (first_len - hlen > mtu ||
818 ((first_len - hlen) & 7) ||
819 ip_is_fragment(iph) ||
821 skb_headroom(skb) < ll_rs)
824 skb_walk_frags(skb, frag) {
825 /* Correct geometry. */
826 if (frag->len > mtu ||
827 ((frag->len & 7) && frag->next) ||
828 skb_headroom(frag) < hlen + ll_rs)
829 goto slow_path_clean;
831 /* Partially cloned skb? */
832 if (skb_shared(frag))
833 goto slow_path_clean;
838 frag->destructor = sock_wfree;
840 skb->truesize -= frag->truesize;
843 /* Everything is OK. Generate! */
844 ip_fraglist_init(skb, iph, hlen, &iter);
847 /* Prepare header of the next frame,
848 * before previous one went down. */
850 ip_fraglist_ipcb_prepare(skb, &iter);
851 ip_fraglist_prepare(skb, &iter);
854 skb->tstamp = tstamp;
855 err = output(net, sk, skb);
858 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
859 if (err || !iter.frag)
862 skb = ip_fraglist_next(&iter);
866 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
870 kfree_skb_list(iter.frag);
872 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
876 skb_walk_frags(skb, frag2) {
880 frag2->destructor = NULL;
881 skb->truesize += frag2->truesize;
887 * Fragment the datagram.
890 ip_frag_init(skb, hlen, ll_rs, mtu, IPCB(skb)->flags & IPSKB_FRAG_PMTU,
894 * Keep copying data until we run out.
897 while (state.left > 0) {
898 bool first_frag = (state.offset == 0);
900 skb2 = ip_frag_next(skb, &state);
905 ip_frag_ipcb(skb, skb2, first_frag, &state);
908 * Put this fragment into the sending queue.
910 skb2->tstamp = tstamp;
911 err = output(net, sk, skb2);
915 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
918 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
923 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
926 EXPORT_SYMBOL(ip_do_fragment);
929 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
931 struct msghdr *msg = from;
933 if (skb->ip_summed == CHECKSUM_PARTIAL) {
934 if (!copy_from_iter_full(to, len, &msg->msg_iter))
938 if (!csum_and_copy_from_iter_full(to, len, &csum, &msg->msg_iter))
940 skb->csum = csum_block_add(skb->csum, csum, odd);
944 EXPORT_SYMBOL(ip_generic_getfrag);
947 csum_page(struct page *page, int offset, int copy)
952 csum = csum_partial(kaddr + offset, copy, 0);
957 static int __ip_append_data(struct sock *sk,
959 struct sk_buff_head *queue,
960 struct inet_cork *cork,
961 struct page_frag *pfrag,
962 int getfrag(void *from, char *to, int offset,
963 int len, int odd, struct sk_buff *skb),
964 void *from, int length, int transhdrlen,
967 struct inet_sock *inet = inet_sk(sk);
968 struct ubuf_info *uarg = NULL;
971 struct ip_options *opt = cork->opt;
978 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
979 int csummode = CHECKSUM_NONE;
980 struct rtable *rt = (struct rtable *)cork->dst;
981 unsigned int wmem_alloc_delta = 0;
982 bool paged, extra_uref = false;
985 skb = skb_peek_tail(queue);
987 exthdrlen = !skb ? rt->dst.header_len : 0;
988 mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
989 paged = !!cork->gso_size;
991 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
992 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
993 tskey = sk->sk_tskey++;
995 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
997 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
998 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
999 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1001 if (cork->length + length > maxnonfragsize - fragheaderlen) {
1002 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1003 mtu - (opt ? opt->optlen : 0));
1008 * transhdrlen > 0 means that this is the first fragment and we wish
1009 * it won't be fragmented in the future.
1012 length + fragheaderlen <= mtu &&
1013 rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
1014 (!(flags & MSG_MORE) || cork->gso_size) &&
1015 (!exthdrlen || (rt->dst.dev->features & NETIF_F_HW_ESP_TX_CSUM)))
1016 csummode = CHECKSUM_PARTIAL;
1018 if (flags & MSG_ZEROCOPY && length && sock_flag(sk, SOCK_ZEROCOPY)) {
1019 uarg = sock_zerocopy_realloc(sk, length, skb_zcopy(skb));
1022 extra_uref = !skb_zcopy(skb); /* only ref on new uarg */
1023 if (rt->dst.dev->features & NETIF_F_SG &&
1024 csummode == CHECKSUM_PARTIAL) {
1028 skb_zcopy_set(skb, uarg, &extra_uref);
1032 cork->length += length;
1034 /* So, what's going on in the loop below?
1036 * We use calculated fragment length to generate chained skb,
1037 * each of segments is IP fragment ready for sending to network after
1038 * adding appropriate IP header.
1044 while (length > 0) {
1045 /* Check if the remaining data fits into current packet. */
1046 copy = mtu - skb->len;
1048 copy = maxfraglen - skb->len;
1051 unsigned int datalen;
1052 unsigned int fraglen;
1053 unsigned int fraggap;
1054 unsigned int alloclen;
1055 unsigned int pagedlen;
1056 struct sk_buff *skb_prev;
1060 fraggap = skb_prev->len - maxfraglen;
1065 * If remaining data exceeds the mtu,
1066 * we know we need more fragment(s).
1068 datalen = length + fraggap;
1069 if (datalen > mtu - fragheaderlen)
1070 datalen = maxfraglen - fragheaderlen;
1071 fraglen = datalen + fragheaderlen;
1074 if ((flags & MSG_MORE) &&
1075 !(rt->dst.dev->features&NETIF_F_SG))
1080 alloclen = min_t(int, fraglen, MAX_HEADER);
1081 pagedlen = fraglen - alloclen;
1084 alloclen += exthdrlen;
1086 /* The last fragment gets additional space at tail.
1087 * Note, with MSG_MORE we overallocate on fragments,
1088 * because we have no idea what fragment will be
1091 if (datalen == length + fraggap)
1092 alloclen += rt->dst.trailer_len;
1095 skb = sock_alloc_send_skb(sk,
1096 alloclen + hh_len + 15,
1097 (flags & MSG_DONTWAIT), &err);
1100 if (refcount_read(&sk->sk_wmem_alloc) + wmem_alloc_delta <=
1102 skb = alloc_skb(alloclen + hh_len + 15,
1111 * Fill in the control structures
1113 skb->ip_summed = csummode;
1115 skb_reserve(skb, hh_len);
1118 * Find where to start putting bytes.
1120 data = skb_put(skb, fraglen + exthdrlen - pagedlen);
1121 skb_set_network_header(skb, exthdrlen);
1122 skb->transport_header = (skb->network_header +
1124 data += fragheaderlen + exthdrlen;
1127 skb->csum = skb_copy_and_csum_bits(
1128 skb_prev, maxfraglen,
1129 data + transhdrlen, fraggap, 0);
1130 skb_prev->csum = csum_sub(skb_prev->csum,
1133 pskb_trim_unique(skb_prev, maxfraglen);
1136 copy = datalen - transhdrlen - fraggap - pagedlen;
1137 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1144 length -= copy + transhdrlen;
1147 csummode = CHECKSUM_NONE;
1149 /* only the initial fragment is time stamped */
1150 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1152 skb_shinfo(skb)->tskey = tskey;
1154 skb_zcopy_set(skb, uarg, &extra_uref);
1156 if ((flags & MSG_CONFIRM) && !skb_prev)
1157 skb_set_dst_pending_confirm(skb, 1);
1160 * Put the packet on the pending queue.
1162 if (!skb->destructor) {
1163 skb->destructor = sock_wfree;
1165 wmem_alloc_delta += skb->truesize;
1167 __skb_queue_tail(queue, skb);
1174 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1175 skb_tailroom(skb) >= copy) {
1179 if (getfrag(from, skb_put(skb, copy),
1180 offset, copy, off, skb) < 0) {
1181 __skb_trim(skb, off);
1185 } else if (!uarg || !uarg->zerocopy) {
1186 int i = skb_shinfo(skb)->nr_frags;
1189 if (!sk_page_frag_refill(sk, pfrag))
1192 if (!skb_can_coalesce(skb, i, pfrag->page,
1195 if (i == MAX_SKB_FRAGS)
1198 __skb_fill_page_desc(skb, i, pfrag->page,
1200 skb_shinfo(skb)->nr_frags = ++i;
1201 get_page(pfrag->page);
1203 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1205 page_address(pfrag->page) + pfrag->offset,
1206 offset, copy, skb->len, skb) < 0)
1209 pfrag->offset += copy;
1210 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1212 skb->data_len += copy;
1213 skb->truesize += copy;
1214 wmem_alloc_delta += copy;
1216 err = skb_zerocopy_iter_dgram(skb, from, copy);
1224 if (wmem_alloc_delta)
1225 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1232 sock_zerocopy_put_abort(uarg, extra_uref);
1233 cork->length -= length;
1234 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1235 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1239 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1240 struct ipcm_cookie *ipc, struct rtable **rtp)
1242 struct ip_options_rcu *opt;
1250 * setup for corking.
1255 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1257 if (unlikely(!cork->opt))
1260 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1261 cork->flags |= IPCORK_OPT;
1262 cork->addr = ipc->addr;
1265 cork->fragsize = ip_sk_use_pmtu(sk) ?
1266 dst_mtu(&rt->dst) : READ_ONCE(rt->dst.dev->mtu);
1268 if (!inetdev_valid_mtu(cork->fragsize))
1269 return -ENETUNREACH;
1271 cork->gso_size = ipc->gso_size;
1273 cork->dst = &rt->dst;
1274 /* We stole this route, caller should not release it. */
1278 cork->ttl = ipc->ttl;
1279 cork->tos = ipc->tos;
1280 cork->mark = ipc->sockc.mark;
1281 cork->priority = ipc->priority;
1282 cork->transmit_time = ipc->sockc.transmit_time;
1284 sock_tx_timestamp(sk, ipc->sockc.tsflags, &cork->tx_flags);
1290 * ip_append_data() and ip_append_page() can make one large IP datagram
1291 * from many pieces of data. Each pieces will be holded on the socket
1292 * until ip_push_pending_frames() is called. Each piece can be a page
1295 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1296 * this interface potentially.
1298 * LATER: length must be adjusted by pad at tail, when it is required.
1300 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1301 int getfrag(void *from, char *to, int offset, int len,
1302 int odd, struct sk_buff *skb),
1303 void *from, int length, int transhdrlen,
1304 struct ipcm_cookie *ipc, struct rtable **rtp,
1307 struct inet_sock *inet = inet_sk(sk);
1310 if (flags&MSG_PROBE)
1313 if (skb_queue_empty(&sk->sk_write_queue)) {
1314 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1321 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1322 sk_page_frag(sk), getfrag,
1323 from, length, transhdrlen, flags);
1326 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1327 int offset, size_t size, int flags)
1329 struct inet_sock *inet = inet_sk(sk);
1330 struct sk_buff *skb;
1332 struct ip_options *opt = NULL;
1333 struct inet_cork *cork;
1338 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1343 if (flags&MSG_PROBE)
1346 if (skb_queue_empty(&sk->sk_write_queue))
1349 cork = &inet->cork.base;
1350 rt = (struct rtable *)cork->dst;
1351 if (cork->flags & IPCORK_OPT)
1354 if (!(rt->dst.dev->features&NETIF_F_SG))
1357 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1358 mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
1360 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1361 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1362 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1364 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1365 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1366 mtu - (opt ? opt->optlen : 0));
1370 skb = skb_peek_tail(&sk->sk_write_queue);
1374 cork->length += size;
1377 /* Check if the remaining data fits into current packet. */
1378 len = mtu - skb->len;
1380 len = maxfraglen - skb->len;
1383 struct sk_buff *skb_prev;
1387 fraggap = skb_prev->len - maxfraglen;
1389 alloclen = fragheaderlen + hh_len + fraggap + 15;
1390 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1391 if (unlikely(!skb)) {
1397 * Fill in the control structures
1399 skb->ip_summed = CHECKSUM_NONE;
1401 skb_reserve(skb, hh_len);
1404 * Find where to start putting bytes.
1406 skb_put(skb, fragheaderlen + fraggap);
1407 skb_reset_network_header(skb);
1408 skb->transport_header = (skb->network_header +
1411 skb->csum = skb_copy_and_csum_bits(skb_prev,
1413 skb_transport_header(skb),
1415 skb_prev->csum = csum_sub(skb_prev->csum,
1417 pskb_trim_unique(skb_prev, maxfraglen);
1421 * Put the packet on the pending queue.
1423 __skb_queue_tail(&sk->sk_write_queue, skb);
1430 if (skb_append_pagefrags(skb, page, offset, len)) {
1435 if (skb->ip_summed == CHECKSUM_NONE) {
1437 csum = csum_page(page, offset, len);
1438 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1442 skb->data_len += len;
1443 skb->truesize += len;
1444 refcount_add(len, &sk->sk_wmem_alloc);
1451 cork->length -= size;
1452 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1456 static void ip_cork_release(struct inet_cork *cork)
1458 cork->flags &= ~IPCORK_OPT;
1461 dst_release(cork->dst);
1466 * Combined all pending IP fragments on the socket as one IP datagram
1467 * and push them out.
1469 struct sk_buff *__ip_make_skb(struct sock *sk,
1471 struct sk_buff_head *queue,
1472 struct inet_cork *cork)
1474 struct sk_buff *skb, *tmp_skb;
1475 struct sk_buff **tail_skb;
1476 struct inet_sock *inet = inet_sk(sk);
1477 struct net *net = sock_net(sk);
1478 struct ip_options *opt = NULL;
1479 struct rtable *rt = (struct rtable *)cork->dst;
1484 skb = __skb_dequeue(queue);
1487 tail_skb = &(skb_shinfo(skb)->frag_list);
1489 /* move skb->data to ip header from ext header */
1490 if (skb->data < skb_network_header(skb))
1491 __skb_pull(skb, skb_network_offset(skb));
1492 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1493 __skb_pull(tmp_skb, skb_network_header_len(skb));
1494 *tail_skb = tmp_skb;
1495 tail_skb = &(tmp_skb->next);
1496 skb->len += tmp_skb->len;
1497 skb->data_len += tmp_skb->len;
1498 skb->truesize += tmp_skb->truesize;
1499 tmp_skb->destructor = NULL;
1503 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1504 * to fragment the frame generated here. No matter, what transforms
1505 * how transforms change size of the packet, it will come out.
1507 skb->ignore_df = ip_sk_ignore_df(sk);
1509 /* DF bit is set when we want to see DF on outgoing frames.
1510 * If ignore_df is set too, we still allow to fragment this frame
1512 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1513 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1514 (skb->len <= dst_mtu(&rt->dst) &&
1515 ip_dont_fragment(sk, &rt->dst)))
1518 if (cork->flags & IPCORK_OPT)
1523 else if (rt->rt_type == RTN_MULTICAST)
1526 ttl = ip_select_ttl(inet, &rt->dst);
1531 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1534 iph->protocol = sk->sk_protocol;
1535 ip_copy_addrs(iph, fl4);
1536 ip_select_ident(net, skb, sk);
1539 iph->ihl += opt->optlen>>2;
1540 ip_options_build(skb, opt, cork->addr, rt, 0);
1543 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1544 skb->mark = cork->mark;
1545 skb->tstamp = cork->transmit_time;
1547 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1551 skb_dst_set(skb, &rt->dst);
1553 if (iph->protocol == IPPROTO_ICMP)
1554 icmp_out_count(net, ((struct icmphdr *)
1555 skb_transport_header(skb))->type);
1557 ip_cork_release(cork);
1562 int ip_send_skb(struct net *net, struct sk_buff *skb)
1566 err = ip_local_out(net, skb->sk, skb);
1569 err = net_xmit_errno(err);
1571 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1577 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1579 struct sk_buff *skb;
1581 skb = ip_finish_skb(sk, fl4);
1585 /* Netfilter gets whole the not fragmented skb. */
1586 return ip_send_skb(sock_net(sk), skb);
1590 * Throw away all pending data on the socket.
1592 static void __ip_flush_pending_frames(struct sock *sk,
1593 struct sk_buff_head *queue,
1594 struct inet_cork *cork)
1596 struct sk_buff *skb;
1598 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1601 ip_cork_release(cork);
1604 void ip_flush_pending_frames(struct sock *sk)
1606 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1609 struct sk_buff *ip_make_skb(struct sock *sk,
1611 int getfrag(void *from, char *to, int offset,
1612 int len, int odd, struct sk_buff *skb),
1613 void *from, int length, int transhdrlen,
1614 struct ipcm_cookie *ipc, struct rtable **rtp,
1615 struct inet_cork *cork, unsigned int flags)
1617 struct sk_buff_head queue;
1620 if (flags & MSG_PROBE)
1623 __skb_queue_head_init(&queue);
1628 err = ip_setup_cork(sk, cork, ipc, rtp);
1630 return ERR_PTR(err);
1632 err = __ip_append_data(sk, fl4, &queue, cork,
1633 ¤t->task_frag, getfrag,
1634 from, length, transhdrlen, flags);
1636 __ip_flush_pending_frames(sk, &queue, cork);
1637 return ERR_PTR(err);
1640 return __ip_make_skb(sk, fl4, &queue, cork);
1644 * Fetch data from kernel space and fill in checksum if needed.
1646 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1647 int len, int odd, struct sk_buff *skb)
1651 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1652 skb->csum = csum_block_add(skb->csum, csum, odd);
1657 * Generic function to send a packet as reply to another packet.
1658 * Used to send some TCP resets/acks so far.
1660 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1661 const struct ip_options *sopt,
1662 __be32 daddr, __be32 saddr,
1663 const struct ip_reply_arg *arg,
1664 unsigned int len, u64 transmit_time)
1666 struct ip_options_data replyopts;
1667 struct ipcm_cookie ipc;
1669 struct rtable *rt = skb_rtable(skb);
1670 struct net *net = sock_net(sk);
1671 struct sk_buff *nskb;
1675 if (__ip_options_echo(net, &replyopts.opt.opt, skb, sopt))
1680 ipc.sockc.transmit_time = transmit_time;
1682 if (replyopts.opt.opt.optlen) {
1683 ipc.opt = &replyopts.opt;
1685 if (replyopts.opt.opt.srr)
1686 daddr = replyopts.opt.opt.faddr;
1689 oif = arg->bound_dev_if;
1690 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1693 flowi4_init_output(&fl4, oif,
1694 IP4_REPLY_MARK(net, skb->mark) ?: sk->sk_mark,
1696 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1697 ip_reply_arg_flowi_flags(arg),
1699 tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1701 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1702 rt = ip_route_output_key(net, &fl4);
1706 inet_sk(sk)->tos = arg->tos;
1708 sk->sk_protocol = ip_hdr(skb)->protocol;
1709 sk->sk_bound_dev_if = arg->bound_dev_if;
1710 sk->sk_sndbuf = sysctl_wmem_default;
1711 ipc.sockc.mark = fl4.flowi4_mark;
1712 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1713 len, 0, &ipc, &rt, MSG_DONTWAIT);
1714 if (unlikely(err)) {
1715 ip_flush_pending_frames(sk);
1719 nskb = skb_peek(&sk->sk_write_queue);
1721 if (arg->csumoffset >= 0)
1722 *((__sum16 *)skb_transport_header(nskb) +
1723 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1725 nskb->ip_summed = CHECKSUM_NONE;
1726 ip_push_pending_frames(sk, &fl4);
1732 void __init ip_init(void)
1737 #if defined(CONFIG_IP_MULTICAST)