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 readability.
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>
77 #include <net/inetpeer.h>
78 #include <net/inet_ecn.h>
79 #include <net/lwtunnel.h>
80 #include <linux/bpf-cgroup.h>
81 #include <linux/igmp.h>
82 #include <linux/netfilter_ipv4.h>
83 #include <linux/netfilter_bridge.h>
84 #include <linux/netlink.h>
85 #include <linux/tcp.h>
88 ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
90 int (*output)(struct net *, struct sock *, struct sk_buff *));
92 /* Generate a checksum for an outgoing IP datagram. */
93 void ip_send_check(struct iphdr *iph)
96 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
98 EXPORT_SYMBOL(ip_send_check);
100 int __ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
102 struct iphdr *iph = ip_hdr(skb);
104 iph_set_totlen(iph, skb->len);
107 /* if egress device is enslaved to an L3 master device pass the
108 * skb to its handler for processing
110 skb = l3mdev_ip_out(sk, skb);
114 skb->protocol = htons(ETH_P_IP);
116 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
117 net, sk, skb, NULL, skb_dst(skb)->dev,
121 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
125 err = __ip_local_out(net, sk, skb);
126 if (likely(err == 1))
127 err = dst_output(net, sk, skb);
131 EXPORT_SYMBOL_GPL(ip_local_out);
133 static inline int ip_select_ttl(const struct inet_sock *inet,
134 const struct dst_entry *dst)
136 int ttl = READ_ONCE(inet->uc_ttl);
139 ttl = ip4_dst_hoplimit(dst);
144 * Add an ip header to a skbuff and send it out.
147 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
148 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt,
151 const struct inet_sock *inet = inet_sk(sk);
152 struct rtable *rt = skb_rtable(skb);
153 struct net *net = sock_net(sk);
156 /* Build the IP header. */
157 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
158 skb_reset_network_header(skb);
163 iph->ttl = ip_select_ttl(inet, &rt->dst);
164 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
166 iph->protocol = sk->sk_protocol;
167 /* Do not bother generating IPID for small packets (eg SYNACK) */
168 if (skb->len <= IPV4_MIN_MTU || ip_dont_fragment(sk, &rt->dst)) {
169 iph->frag_off = htons(IP_DF);
173 /* TCP packets here are SYNACK with fat IPv4/TCP options.
174 * Avoid using the hashed IP ident generator.
176 if (sk->sk_protocol == IPPROTO_TCP)
177 iph->id = (__force __be16)get_random_u16();
179 __ip_select_ident(net, iph, 1);
182 if (opt && opt->opt.optlen) {
183 iph->ihl += opt->opt.optlen>>2;
184 ip_options_build(skb, &opt->opt, daddr, rt);
187 skb->priority = READ_ONCE(sk->sk_priority);
189 skb->mark = READ_ONCE(sk->sk_mark);
192 return ip_local_out(net, skb->sk, skb);
194 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
196 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
198 struct dst_entry *dst = skb_dst(skb);
199 struct rtable *rt = (struct rtable *)dst;
200 struct net_device *dev = dst->dev;
201 unsigned int hh_len = LL_RESERVED_SPACE(dev);
202 struct neighbour *neigh;
203 bool is_v6gw = false;
205 if (rt->rt_type == RTN_MULTICAST) {
206 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
207 } else if (rt->rt_type == RTN_BROADCAST)
208 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
210 /* OUTOCTETS should be counted after fragment */
211 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
213 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
214 skb = skb_expand_head(skb, hh_len);
219 if (lwtunnel_xmit_redirect(dst->lwtstate)) {
220 int res = lwtunnel_xmit(skb);
222 if (res != LWTUNNEL_XMIT_CONTINUE)
227 neigh = ip_neigh_for_gw(rt, skb, &is_v6gw);
228 if (!IS_ERR(neigh)) {
231 sock_confirm_neigh(skb, neigh);
232 /* if crossing protocols, can not use the cached header */
233 res = neigh_output(neigh, skb, is_v6gw);
239 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
241 kfree_skb_reason(skb, SKB_DROP_REASON_NEIGH_CREATEFAIL);
242 return PTR_ERR(neigh);
245 static int ip_finish_output_gso(struct net *net, struct sock *sk,
246 struct sk_buff *skb, unsigned int mtu)
248 struct sk_buff *segs, *nskb;
249 netdev_features_t features;
252 /* common case: seglen is <= mtu
254 if (skb_gso_validate_network_len(skb, mtu))
255 return ip_finish_output2(net, sk, skb);
257 /* Slowpath - GSO segment length exceeds the egress MTU.
259 * This can happen in several cases:
260 * - Forwarding of a TCP GRO skb, when DF flag is not set.
261 * - Forwarding of an skb that arrived on a virtualization interface
262 * (virtio-net/vhost/tap) with TSO/GSO size set by other network
264 * - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
265 * interface with a smaller MTU.
266 * - Arriving GRO skb (or GSO skb in a virtualized environment) that is
267 * bridged to a NETIF_F_TSO tunnel stacked over an interface with an
270 features = netif_skb_features(skb);
271 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_GSO_CB_OFFSET);
272 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
273 if (IS_ERR_OR_NULL(segs)) {
280 skb_list_walk_safe(segs, segs, nskb) {
283 skb_mark_not_on_list(segs);
284 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
293 static int __ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
297 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
298 /* Policy lookup after SNAT yielded a new policy */
299 if (skb_dst(skb)->xfrm) {
300 IPCB(skb)->flags |= IPSKB_REROUTED;
301 return dst_output(net, sk, skb);
304 mtu = ip_skb_dst_mtu(sk, skb);
306 return ip_finish_output_gso(net, sk, skb, mtu);
308 if (skb->len > mtu || IPCB(skb)->frag_max_size)
309 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
311 return ip_finish_output2(net, sk, skb);
314 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
318 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
320 case NET_XMIT_SUCCESS:
321 return __ip_finish_output(net, sk, skb);
323 return __ip_finish_output(net, sk, skb) ? : ret;
325 kfree_skb_reason(skb, SKB_DROP_REASON_BPF_CGROUP_EGRESS);
330 static int ip_mc_finish_output(struct net *net, struct sock *sk,
333 struct rtable *new_rt;
337 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
342 case NET_XMIT_SUCCESS:
345 kfree_skb_reason(skb, SKB_DROP_REASON_BPF_CGROUP_EGRESS);
349 /* Reset rt_iif so that inet_iif() will return skb->skb_iif. Setting
350 * this to non-zero causes ipi_ifindex in in_pktinfo to be overwritten,
351 * see ipv4_pktinfo_prepare().
353 new_rt = rt_dst_clone(net->loopback_dev, skb_rtable(skb));
357 skb_dst_set(skb, &new_rt->dst);
360 err = dev_loopback_xmit(net, sk, skb);
361 return (do_cn && err) ? ret : err;
364 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
366 struct rtable *rt = skb_rtable(skb);
367 struct net_device *dev = rt->dst.dev;
370 * If the indicated interface is up and running, send the packet.
373 skb->protocol = htons(ETH_P_IP);
376 * Multicasts are looped back for other local users
379 if (rt->rt_flags&RTCF_MULTICAST) {
381 #ifdef CONFIG_IP_MROUTE
382 /* Small optimization: do not loopback not local frames,
383 which returned after forwarding; they will be dropped
384 by ip_mr_input in any case.
385 Note, that local frames are looped back to be delivered
388 This check is duplicated in ip_mr_input at the moment.
391 ((rt->rt_flags & RTCF_LOCAL) ||
392 !(IPCB(skb)->flags & IPSKB_FORWARDED))
395 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
397 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
398 net, sk, newskb, NULL, newskb->dev,
399 ip_mc_finish_output);
402 /* Multicasts with ttl 0 must not go beyond the host */
404 if (ip_hdr(skb)->ttl == 0) {
410 if (rt->rt_flags&RTCF_BROADCAST) {
411 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
413 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
414 net, sk, newskb, NULL, newskb->dev,
415 ip_mc_finish_output);
418 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
419 net, sk, skb, NULL, skb->dev,
421 !(IPCB(skb)->flags & IPSKB_REROUTED));
424 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
426 struct net_device *dev = skb_dst(skb)->dev, *indev = skb->dev;
429 skb->protocol = htons(ETH_P_IP);
431 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
432 net, sk, skb, indev, dev,
434 !(IPCB(skb)->flags & IPSKB_REROUTED));
436 EXPORT_SYMBOL(ip_output);
439 * copy saddr and daddr, possibly using 64bit load/stores
441 * iph->saddr = fl4->saddr;
442 * iph->daddr = fl4->daddr;
444 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
446 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
447 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
449 iph->saddr = fl4->saddr;
450 iph->daddr = fl4->daddr;
453 /* Note: skb->sk can be different from sk, in case of tunnels */
454 int __ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl,
457 struct inet_sock *inet = inet_sk(sk);
458 struct net *net = sock_net(sk);
459 struct ip_options_rcu *inet_opt;
465 /* Skip all of this if the packet is already routed,
466 * f.e. by something like SCTP.
469 inet_opt = rcu_dereference(inet->inet_opt);
471 rt = skb_rtable(skb);
475 /* Make sure we can route this packet. */
476 rt = (struct rtable *)__sk_dst_check(sk, 0);
480 /* Use correct destination address if we have options. */
481 daddr = inet->inet_daddr;
482 if (inet_opt && inet_opt->opt.srr)
483 daddr = inet_opt->opt.faddr;
485 /* If this fails, retransmit mechanism of transport layer will
486 * keep trying until route appears or the connection times
489 rt = ip_route_output_ports(net, fl4, sk,
490 daddr, inet->inet_saddr,
494 RT_CONN_FLAGS_TOS(sk, tos),
495 sk->sk_bound_dev_if);
498 sk_setup_caps(sk, &rt->dst);
500 skb_dst_set_noref(skb, &rt->dst);
503 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
506 /* OK, we know where to send it, allocate and build IP header. */
507 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
508 skb_reset_network_header(skb);
510 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (tos & 0xff));
511 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
512 iph->frag_off = htons(IP_DF);
515 iph->ttl = ip_select_ttl(inet, &rt->dst);
516 iph->protocol = sk->sk_protocol;
517 ip_copy_addrs(iph, fl4);
519 /* Transport layer set skb->h.foo itself. */
521 if (inet_opt && inet_opt->opt.optlen) {
522 iph->ihl += inet_opt->opt.optlen >> 2;
523 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt);
526 ip_select_ident_segs(net, skb, sk,
527 skb_shinfo(skb)->gso_segs ?: 1);
529 /* TODO : should we use skb->sk here instead of sk ? */
530 skb->priority = READ_ONCE(sk->sk_priority);
531 skb->mark = READ_ONCE(sk->sk_mark);
533 res = ip_local_out(net, sk, skb);
539 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
540 kfree_skb_reason(skb, SKB_DROP_REASON_IP_OUTNOROUTES);
541 return -EHOSTUNREACH;
543 EXPORT_SYMBOL(__ip_queue_xmit);
545 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
547 return __ip_queue_xmit(sk, skb, fl, inet_sk(sk)->tos);
549 EXPORT_SYMBOL(ip_queue_xmit);
551 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
553 to->pkt_type = from->pkt_type;
554 to->priority = from->priority;
555 to->protocol = from->protocol;
556 to->skb_iif = from->skb_iif;
558 skb_dst_copy(to, from);
560 to->mark = from->mark;
562 skb_copy_hash(to, from);
564 #ifdef CONFIG_NET_SCHED
565 to->tc_index = from->tc_index;
568 skb_ext_copy(to, from);
569 #if IS_ENABLED(CONFIG_IP_VS)
570 to->ipvs_property = from->ipvs_property;
572 skb_copy_secmark(to, from);
575 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
577 int (*output)(struct net *, struct sock *, struct sk_buff *))
579 struct iphdr *iph = ip_hdr(skb);
581 if ((iph->frag_off & htons(IP_DF)) == 0)
582 return ip_do_fragment(net, sk, skb, output);
584 if (unlikely(!skb->ignore_df ||
585 (IPCB(skb)->frag_max_size &&
586 IPCB(skb)->frag_max_size > mtu))) {
587 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
588 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
594 return ip_do_fragment(net, sk, skb, output);
597 void ip_fraglist_init(struct sk_buff *skb, struct iphdr *iph,
598 unsigned int hlen, struct ip_fraglist_iter *iter)
600 unsigned int first_len = skb_pagelen(skb);
602 iter->frag = skb_shinfo(skb)->frag_list;
603 skb_frag_list_init(skb);
609 skb->data_len = first_len - skb_headlen(skb);
610 skb->len = first_len;
611 iph->tot_len = htons(first_len);
612 iph->frag_off = htons(IP_MF);
615 EXPORT_SYMBOL(ip_fraglist_init);
617 void ip_fraglist_prepare(struct sk_buff *skb, struct ip_fraglist_iter *iter)
619 unsigned int hlen = iter->hlen;
620 struct iphdr *iph = iter->iph;
621 struct sk_buff *frag;
624 frag->ip_summed = CHECKSUM_NONE;
625 skb_reset_transport_header(frag);
626 __skb_push(frag, hlen);
627 skb_reset_network_header(frag);
628 memcpy(skb_network_header(frag), iph, hlen);
629 iter->iph = ip_hdr(frag);
631 iph->tot_len = htons(frag->len);
632 ip_copy_metadata(frag, skb);
633 iter->offset += skb->len - hlen;
634 iph->frag_off = htons(iter->offset >> 3);
636 iph->frag_off |= htons(IP_MF);
637 /* Ready, complete checksum */
640 EXPORT_SYMBOL(ip_fraglist_prepare);
642 void ip_frag_init(struct sk_buff *skb, unsigned int hlen,
643 unsigned int ll_rs, unsigned int mtu, bool DF,
644 struct ip_frag_state *state)
646 struct iphdr *iph = ip_hdr(skb);
650 state->ll_rs = ll_rs;
653 state->left = skb->len - hlen; /* Space per frame */
654 state->ptr = hlen; /* Where to start from */
656 state->offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
657 state->not_last_frag = iph->frag_off & htons(IP_MF);
659 EXPORT_SYMBOL(ip_frag_init);
661 static void ip_frag_ipcb(struct sk_buff *from, struct sk_buff *to,
664 /* Copy the flags to each fragment. */
665 IPCB(to)->flags = IPCB(from)->flags;
667 /* ANK: dirty, but effective trick. Upgrade options only if
668 * the segment to be fragmented was THE FIRST (otherwise,
669 * options are already fixed) and make it ONCE
670 * on the initial skb, so that all the following fragments
671 * will inherit fixed options.
674 ip_options_fragment(from);
677 struct sk_buff *ip_frag_next(struct sk_buff *skb, struct ip_frag_state *state)
679 unsigned int len = state->left;
680 struct sk_buff *skb2;
683 /* IF: it doesn't fit, use 'mtu' - the data space left */
684 if (len > state->mtu)
686 /* IF: we are not sending up to and including the packet end
687 then align the next start on an eight byte boundary */
688 if (len < state->left) {
692 /* Allocate buffer */
693 skb2 = alloc_skb(len + state->hlen + state->ll_rs, GFP_ATOMIC);
695 return ERR_PTR(-ENOMEM);
698 * Set up data on packet
701 ip_copy_metadata(skb2, skb);
702 skb_reserve(skb2, state->ll_rs);
703 skb_put(skb2, len + state->hlen);
704 skb_reset_network_header(skb2);
705 skb2->transport_header = skb2->network_header + state->hlen;
708 * Charge the memory for the fragment to any owner
713 skb_set_owner_w(skb2, skb->sk);
716 * Copy the packet header into the new buffer.
719 skb_copy_from_linear_data(skb, skb_network_header(skb2), state->hlen);
722 * Copy a block of the IP datagram.
724 if (skb_copy_bits(skb, state->ptr, skb_transport_header(skb2), len))
729 * Fill in the new header fields.
732 iph->frag_off = htons((state->offset >> 3));
734 iph->frag_off |= htons(IP_DF);
737 * Added AC : If we are fragmenting a fragment that's not the
738 * last fragment then keep MF on each bit
740 if (state->left > 0 || state->not_last_frag)
741 iph->frag_off |= htons(IP_MF);
743 state->offset += len;
745 iph->tot_len = htons(len + state->hlen);
751 EXPORT_SYMBOL(ip_frag_next);
754 * This IP datagram is too large to be sent in one piece. Break it up into
755 * smaller pieces (each of size equal to IP header plus
756 * a block of the data of the original IP data part) that will yet fit in a
757 * single device frame, and queue such a frame for sending.
760 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
761 int (*output)(struct net *, struct sock *, struct sk_buff *))
764 struct sk_buff *skb2;
765 bool mono_delivery_time = skb->mono_delivery_time;
766 struct rtable *rt = skb_rtable(skb);
767 unsigned int mtu, hlen, ll_rs;
768 struct ip_fraglist_iter iter;
769 ktime_t tstamp = skb->tstamp;
770 struct ip_frag_state state;
773 /* for offloaded checksums cleanup checksum before fragmentation */
774 if (skb->ip_summed == CHECKSUM_PARTIAL &&
775 (err = skb_checksum_help(skb)))
779 * Point into the IP datagram header.
784 mtu = ip_skb_dst_mtu(sk, skb);
785 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
786 mtu = IPCB(skb)->frag_max_size;
789 * Setup starting values.
793 mtu = mtu - hlen; /* Size of data space */
794 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
795 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
797 /* When frag_list is given, use it. First, check its validity:
798 * some transformers could create wrong frag_list or break existing
799 * one, it is not prohibited. In this case fall back to copying.
801 * LATER: this step can be merged to real generation of fragments,
802 * we can switch to copy when see the first bad fragment.
804 if (skb_has_frag_list(skb)) {
805 struct sk_buff *frag, *frag2;
806 unsigned int first_len = skb_pagelen(skb);
808 if (first_len - hlen > mtu ||
809 ((first_len - hlen) & 7) ||
810 ip_is_fragment(iph) ||
812 skb_headroom(skb) < ll_rs)
815 skb_walk_frags(skb, frag) {
816 /* Correct geometry. */
817 if (frag->len > mtu ||
818 ((frag->len & 7) && frag->next) ||
819 skb_headroom(frag) < hlen + ll_rs)
820 goto slow_path_clean;
822 /* Partially cloned skb? */
823 if (skb_shared(frag))
824 goto slow_path_clean;
829 frag->destructor = sock_wfree;
831 skb->truesize -= frag->truesize;
834 /* Everything is OK. Generate! */
835 ip_fraglist_init(skb, iph, hlen, &iter);
838 /* Prepare header of the next frame,
839 * before previous one went down. */
841 bool first_frag = (iter.offset == 0);
843 IPCB(iter.frag)->flags = IPCB(skb)->flags;
844 ip_fraglist_prepare(skb, &iter);
845 if (first_frag && IPCB(skb)->opt.optlen) {
846 /* ipcb->opt is not populated for frags
847 * coming from __ip_make_skb(),
848 * ip_options_fragment() needs optlen
850 IPCB(iter.frag)->opt.optlen =
851 IPCB(skb)->opt.optlen;
852 ip_options_fragment(iter.frag);
853 ip_send_check(iter.iph);
857 skb_set_delivery_time(skb, tstamp, mono_delivery_time);
858 err = output(net, sk, skb);
861 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
862 if (err || !iter.frag)
865 skb = ip_fraglist_next(&iter);
869 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
873 kfree_skb_list(iter.frag);
875 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
879 skb_walk_frags(skb, frag2) {
883 frag2->destructor = NULL;
884 skb->truesize += frag2->truesize;
890 * Fragment the datagram.
893 ip_frag_init(skb, hlen, ll_rs, mtu, IPCB(skb)->flags & IPSKB_FRAG_PMTU,
897 * Keep copying data until we run out.
900 while (state.left > 0) {
901 bool first_frag = (state.offset == 0);
903 skb2 = ip_frag_next(skb, &state);
908 ip_frag_ipcb(skb, skb2, first_frag);
911 * Put this fragment into the sending queue.
913 skb_set_delivery_time(skb2, tstamp, mono_delivery_time);
914 err = output(net, sk, skb2);
918 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
921 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
926 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
929 EXPORT_SYMBOL(ip_do_fragment);
932 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
934 struct msghdr *msg = from;
936 if (skb->ip_summed == CHECKSUM_PARTIAL) {
937 if (!copy_from_iter_full(to, len, &msg->msg_iter))
941 if (!csum_and_copy_from_iter_full(to, len, &csum, &msg->msg_iter))
943 skb->csum = csum_block_add(skb->csum, csum, odd);
947 EXPORT_SYMBOL(ip_generic_getfrag);
949 static int __ip_append_data(struct sock *sk,
951 struct sk_buff_head *queue,
952 struct inet_cork *cork,
953 struct page_frag *pfrag,
954 int getfrag(void *from, char *to, int offset,
955 int len, int odd, struct sk_buff *skb),
956 void *from, int length, int transhdrlen,
959 struct inet_sock *inet = inet_sk(sk);
960 struct ubuf_info *uarg = NULL;
962 struct ip_options *opt = cork->opt;
970 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
971 int csummode = CHECKSUM_NONE;
972 struct rtable *rt = (struct rtable *)cork->dst;
973 unsigned int wmem_alloc_delta = 0;
974 bool paged, extra_uref = false;
977 skb = skb_peek_tail(queue);
979 exthdrlen = !skb ? rt->dst.header_len : 0;
980 mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
981 paged = !!cork->gso_size;
983 if (cork->tx_flags & SKBTX_ANY_TSTAMP &&
984 READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_OPT_ID)
985 tskey = atomic_inc_return(&sk->sk_tskey) - 1;
987 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
989 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
990 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
991 maxnonfragsize = ip_sk_ignore_df(sk) ? IP_MAX_MTU : mtu;
993 if (cork->length + length > maxnonfragsize - fragheaderlen) {
994 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
995 mtu - (opt ? opt->optlen : 0));
1000 * transhdrlen > 0 means that this is the first fragment and we wish
1001 * it won't be fragmented in the future.
1004 length + fragheaderlen <= mtu &&
1005 rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
1006 (!(flags & MSG_MORE) || cork->gso_size) &&
1007 (!exthdrlen || (rt->dst.dev->features & NETIF_F_HW_ESP_TX_CSUM)))
1008 csummode = CHECKSUM_PARTIAL;
1010 if ((flags & MSG_ZEROCOPY) && length) {
1011 struct msghdr *msg = from;
1013 if (getfrag == ip_generic_getfrag && msg->msg_ubuf) {
1014 if (skb_zcopy(skb) && msg->msg_ubuf != skb_zcopy(skb))
1017 /* Leave uarg NULL if can't zerocopy, callers should
1018 * be able to handle it.
1020 if ((rt->dst.dev->features & NETIF_F_SG) &&
1021 csummode == CHECKSUM_PARTIAL) {
1024 uarg = msg->msg_ubuf;
1026 } else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1027 uarg = msg_zerocopy_realloc(sk, length, skb_zcopy(skb));
1030 extra_uref = !skb_zcopy(skb); /* only ref on new uarg */
1031 if (rt->dst.dev->features & NETIF_F_SG &&
1032 csummode == CHECKSUM_PARTIAL) {
1036 uarg_to_msgzc(uarg)->zerocopy = 0;
1037 skb_zcopy_set(skb, uarg, &extra_uref);
1040 } else if ((flags & MSG_SPLICE_PAGES) && length) {
1041 if (inet_test_bit(HDRINCL, sk))
1043 if (rt->dst.dev->features & NETIF_F_SG &&
1044 getfrag == ip_generic_getfrag)
1045 /* We need an empty buffer to attach stuff to */
1048 flags &= ~MSG_SPLICE_PAGES;
1051 cork->length += length;
1053 /* So, what's going on in the loop below?
1055 * We use calculated fragment length to generate chained skb,
1056 * each of segments is IP fragment ready for sending to network after
1057 * adding appropriate IP header.
1063 while (length > 0) {
1064 /* Check if the remaining data fits into current packet. */
1065 copy = mtu - skb->len;
1067 copy = maxfraglen - skb->len;
1070 unsigned int datalen;
1071 unsigned int fraglen;
1072 unsigned int fraggap;
1073 unsigned int alloclen, alloc_extra;
1074 unsigned int pagedlen;
1075 struct sk_buff *skb_prev;
1079 fraggap = skb_prev->len - maxfraglen;
1084 * If remaining data exceeds the mtu,
1085 * we know we need more fragment(s).
1087 datalen = length + fraggap;
1088 if (datalen > mtu - fragheaderlen)
1089 datalen = maxfraglen - fragheaderlen;
1090 fraglen = datalen + fragheaderlen;
1093 alloc_extra = hh_len + 15;
1094 alloc_extra += exthdrlen;
1096 /* The last fragment gets additional space at tail.
1097 * Note, with MSG_MORE we overallocate on fragments,
1098 * because we have no idea what fragment will be
1101 if (datalen == length + fraggap)
1102 alloc_extra += rt->dst.trailer_len;
1104 if ((flags & MSG_MORE) &&
1105 !(rt->dst.dev->features&NETIF_F_SG))
1108 (fraglen + alloc_extra < SKB_MAX_ALLOC ||
1109 !(rt->dst.dev->features & NETIF_F_SG)))
1112 alloclen = fragheaderlen + transhdrlen;
1113 pagedlen = datalen - transhdrlen;
1116 alloclen += alloc_extra;
1119 skb = sock_alloc_send_skb(sk, alloclen,
1120 (flags & MSG_DONTWAIT), &err);
1123 if (refcount_read(&sk->sk_wmem_alloc) + wmem_alloc_delta <=
1125 skb = alloc_skb(alloclen,
1134 * Fill in the control structures
1136 skb->ip_summed = csummode;
1138 skb_reserve(skb, hh_len);
1141 * Find where to start putting bytes.
1143 data = skb_put(skb, fraglen + exthdrlen - pagedlen);
1144 skb_set_network_header(skb, exthdrlen);
1145 skb->transport_header = (skb->network_header +
1147 data += fragheaderlen + exthdrlen;
1150 skb->csum = skb_copy_and_csum_bits(
1151 skb_prev, maxfraglen,
1152 data + transhdrlen, fraggap);
1153 skb_prev->csum = csum_sub(skb_prev->csum,
1156 pskb_trim_unique(skb_prev, maxfraglen);
1159 copy = datalen - transhdrlen - fraggap - pagedlen;
1160 /* [!] NOTE: copy will be negative if pagedlen>0
1161 * because then the equation reduces to -fraggap.
1163 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1167 } else if (flags & MSG_SPLICE_PAGES) {
1172 length -= copy + transhdrlen;
1175 csummode = CHECKSUM_NONE;
1177 /* only the initial fragment is time stamped */
1178 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1180 skb_shinfo(skb)->tskey = tskey;
1182 skb_zcopy_set(skb, uarg, &extra_uref);
1184 if ((flags & MSG_CONFIRM) && !skb_prev)
1185 skb_set_dst_pending_confirm(skb, 1);
1188 * Put the packet on the pending queue.
1190 if (!skb->destructor) {
1191 skb->destructor = sock_wfree;
1193 wmem_alloc_delta += skb->truesize;
1195 __skb_queue_tail(queue, skb);
1202 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1203 skb_tailroom(skb) >= copy) {
1207 if (getfrag(from, skb_put(skb, copy),
1208 offset, copy, off, skb) < 0) {
1209 __skb_trim(skb, off);
1213 } else if (flags & MSG_SPLICE_PAGES) {
1214 struct msghdr *msg = from;
1217 if (WARN_ON_ONCE(copy > msg->msg_iter.count))
1220 err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
1225 wmem_alloc_delta += copy;
1227 int i = skb_shinfo(skb)->nr_frags;
1230 if (!sk_page_frag_refill(sk, pfrag))
1233 skb_zcopy_downgrade_managed(skb);
1234 if (!skb_can_coalesce(skb, i, pfrag->page,
1237 if (i == MAX_SKB_FRAGS)
1240 __skb_fill_page_desc(skb, i, pfrag->page,
1242 skb_shinfo(skb)->nr_frags = ++i;
1243 get_page(pfrag->page);
1245 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1247 page_address(pfrag->page) + pfrag->offset,
1248 offset, copy, skb->len, skb) < 0)
1251 pfrag->offset += copy;
1252 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1253 skb_len_add(skb, copy);
1254 wmem_alloc_delta += copy;
1256 err = skb_zerocopy_iter_dgram(skb, from, copy);
1264 if (wmem_alloc_delta)
1265 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1271 net_zcopy_put_abort(uarg, extra_uref);
1272 cork->length -= length;
1273 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1274 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1278 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1279 struct ipcm_cookie *ipc, struct rtable **rtp)
1281 struct ip_options_rcu *opt;
1289 * setup for corking.
1294 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1296 if (unlikely(!cork->opt))
1299 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1300 cork->flags |= IPCORK_OPT;
1301 cork->addr = ipc->addr;
1304 cork->fragsize = ip_sk_use_pmtu(sk) ?
1305 dst_mtu(&rt->dst) : READ_ONCE(rt->dst.dev->mtu);
1307 if (!inetdev_valid_mtu(cork->fragsize))
1308 return -ENETUNREACH;
1310 cork->gso_size = ipc->gso_size;
1312 cork->dst = &rt->dst;
1313 /* We stole this route, caller should not release it. */
1317 cork->ttl = ipc->ttl;
1318 cork->tos = ipc->tos;
1319 cork->mark = ipc->sockc.mark;
1320 cork->priority = ipc->priority;
1321 cork->transmit_time = ipc->sockc.transmit_time;
1323 sock_tx_timestamp(sk, ipc->sockc.tsflags, &cork->tx_flags);
1329 * ip_append_data() can make one large IP datagram from many pieces of
1330 * data. Each piece will be held on the socket until
1331 * ip_push_pending_frames() is called. Each piece can be a page or
1334 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1335 * this interface potentially.
1337 * LATER: length must be adjusted by pad at tail, when it is required.
1339 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1340 int getfrag(void *from, char *to, int offset, int len,
1341 int odd, struct sk_buff *skb),
1342 void *from, int length, int transhdrlen,
1343 struct ipcm_cookie *ipc, struct rtable **rtp,
1346 struct inet_sock *inet = inet_sk(sk);
1349 if (flags&MSG_PROBE)
1352 if (skb_queue_empty(&sk->sk_write_queue)) {
1353 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1360 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1361 sk_page_frag(sk), getfrag,
1362 from, length, transhdrlen, flags);
1365 static void ip_cork_release(struct inet_cork *cork)
1367 cork->flags &= ~IPCORK_OPT;
1370 dst_release(cork->dst);
1375 * Combined all pending IP fragments on the socket as one IP datagram
1376 * and push them out.
1378 struct sk_buff *__ip_make_skb(struct sock *sk,
1380 struct sk_buff_head *queue,
1381 struct inet_cork *cork)
1383 struct sk_buff *skb, *tmp_skb;
1384 struct sk_buff **tail_skb;
1385 struct inet_sock *inet = inet_sk(sk);
1386 struct net *net = sock_net(sk);
1387 struct ip_options *opt = NULL;
1388 struct rtable *rt = (struct rtable *)cork->dst;
1393 skb = __skb_dequeue(queue);
1396 tail_skb = &(skb_shinfo(skb)->frag_list);
1398 /* move skb->data to ip header from ext header */
1399 if (skb->data < skb_network_header(skb))
1400 __skb_pull(skb, skb_network_offset(skb));
1401 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1402 __skb_pull(tmp_skb, skb_network_header_len(skb));
1403 *tail_skb = tmp_skb;
1404 tail_skb = &(tmp_skb->next);
1405 skb->len += tmp_skb->len;
1406 skb->data_len += tmp_skb->len;
1407 skb->truesize += tmp_skb->truesize;
1408 tmp_skb->destructor = NULL;
1412 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1413 * to fragment the frame generated here. No matter, what transforms
1414 * how transforms change size of the packet, it will come out.
1416 skb->ignore_df = ip_sk_ignore_df(sk);
1418 /* DF bit is set when we want to see DF on outgoing frames.
1419 * If ignore_df is set too, we still allow to fragment this frame
1421 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1422 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1423 (skb->len <= dst_mtu(&rt->dst) &&
1424 ip_dont_fragment(sk, &rt->dst)))
1427 if (cork->flags & IPCORK_OPT)
1432 else if (rt->rt_type == RTN_MULTICAST)
1435 ttl = ip_select_ttl(inet, &rt->dst);
1440 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1443 iph->protocol = sk->sk_protocol;
1444 ip_copy_addrs(iph, fl4);
1445 ip_select_ident(net, skb, sk);
1448 iph->ihl += opt->optlen >> 2;
1449 ip_options_build(skb, opt, cork->addr, rt);
1452 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1453 skb->mark = cork->mark;
1454 skb->tstamp = cork->transmit_time;
1456 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1460 skb_dst_set(skb, &rt->dst);
1462 if (iph->protocol == IPPROTO_ICMP) {
1465 /* For such sockets, transhdrlen is zero when do ip_append_data(),
1466 * so icmphdr does not in skb linear region and can not get icmp_type
1467 * by icmp_hdr(skb)->type.
1469 if (sk->sk_type == SOCK_RAW &&
1470 !inet_test_bit(HDRINCL, sk))
1471 icmp_type = fl4->fl4_icmp_type;
1473 icmp_type = icmp_hdr(skb)->type;
1474 icmp_out_count(net, icmp_type);
1477 ip_cork_release(cork);
1482 int ip_send_skb(struct net *net, struct sk_buff *skb)
1486 err = ip_local_out(net, skb->sk, skb);
1489 err = net_xmit_errno(err);
1491 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1497 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1499 struct sk_buff *skb;
1501 skb = ip_finish_skb(sk, fl4);
1505 /* Netfilter gets whole the not fragmented skb. */
1506 return ip_send_skb(sock_net(sk), skb);
1510 * Throw away all pending data on the socket.
1512 static void __ip_flush_pending_frames(struct sock *sk,
1513 struct sk_buff_head *queue,
1514 struct inet_cork *cork)
1516 struct sk_buff *skb;
1518 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1521 ip_cork_release(cork);
1524 void ip_flush_pending_frames(struct sock *sk)
1526 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1529 struct sk_buff *ip_make_skb(struct sock *sk,
1531 int getfrag(void *from, char *to, int offset,
1532 int len, int odd, struct sk_buff *skb),
1533 void *from, int length, int transhdrlen,
1534 struct ipcm_cookie *ipc, struct rtable **rtp,
1535 struct inet_cork *cork, unsigned int flags)
1537 struct sk_buff_head queue;
1540 if (flags & MSG_PROBE)
1543 __skb_queue_head_init(&queue);
1548 err = ip_setup_cork(sk, cork, ipc, rtp);
1550 return ERR_PTR(err);
1552 err = __ip_append_data(sk, fl4, &queue, cork,
1553 ¤t->task_frag, getfrag,
1554 from, length, transhdrlen, flags);
1556 __ip_flush_pending_frames(sk, &queue, cork);
1557 return ERR_PTR(err);
1560 return __ip_make_skb(sk, fl4, &queue, cork);
1564 * Fetch data from kernel space and fill in checksum if needed.
1566 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1567 int len, int odd, struct sk_buff *skb)
1571 csum = csum_partial_copy_nocheck(dptr+offset, to, len);
1572 skb->csum = csum_block_add(skb->csum, csum, odd);
1577 * Generic function to send a packet as reply to another packet.
1578 * Used to send some TCP resets/acks so far.
1580 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1581 const struct ip_options *sopt,
1582 __be32 daddr, __be32 saddr,
1583 const struct ip_reply_arg *arg,
1584 unsigned int len, u64 transmit_time, u32 txhash)
1586 struct ip_options_data replyopts;
1587 struct ipcm_cookie ipc;
1589 struct rtable *rt = skb_rtable(skb);
1590 struct net *net = sock_net(sk);
1591 struct sk_buff *nskb;
1595 if (__ip_options_echo(net, &replyopts.opt.opt, skb, sopt))
1600 ipc.sockc.transmit_time = transmit_time;
1602 if (replyopts.opt.opt.optlen) {
1603 ipc.opt = &replyopts.opt;
1605 if (replyopts.opt.opt.srr)
1606 daddr = replyopts.opt.opt.faddr;
1609 oif = arg->bound_dev_if;
1610 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1613 flowi4_init_output(&fl4, oif,
1614 IP4_REPLY_MARK(net, skb->mark) ?: sk->sk_mark,
1616 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1617 ip_reply_arg_flowi_flags(arg),
1619 tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1621 security_skb_classify_flow(skb, flowi4_to_flowi_common(&fl4));
1622 rt = ip_route_output_flow(net, &fl4, sk);
1626 inet_sk(sk)->tos = arg->tos & ~INET_ECN_MASK;
1628 sk->sk_protocol = ip_hdr(skb)->protocol;
1629 sk->sk_bound_dev_if = arg->bound_dev_if;
1630 sk->sk_sndbuf = READ_ONCE(sysctl_wmem_default);
1631 ipc.sockc.mark = fl4.flowi4_mark;
1632 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1633 len, 0, &ipc, &rt, MSG_DONTWAIT);
1634 if (unlikely(err)) {
1635 ip_flush_pending_frames(sk);
1639 nskb = skb_peek(&sk->sk_write_queue);
1641 if (arg->csumoffset >= 0)
1642 *((__sum16 *)skb_transport_header(nskb) +
1643 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1645 nskb->ip_summed = CHECKSUM_NONE;
1646 nskb->mono_delivery_time = !!transmit_time;
1648 skb_set_hash(nskb, txhash, PKT_HASH_TYPE_L4);
1649 ip_push_pending_frames(sk, &fl4);
1655 void __init ip_init(void)
1660 #if defined(CONFIG_IP_MULTICAST)