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 IP_INC_STATS(net, IPSTATS_MIB_OUTREQUESTS);
106 iph_set_totlen(iph, skb->len);
109 /* if egress device is enslaved to an L3 master device pass the
110 * skb to its handler for processing
112 skb = l3mdev_ip_out(sk, skb);
116 skb->protocol = htons(ETH_P_IP);
118 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
119 net, sk, skb, NULL, skb_dst(skb)->dev,
123 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
127 err = __ip_local_out(net, sk, skb);
128 if (likely(err == 1))
129 err = dst_output(net, sk, skb);
133 EXPORT_SYMBOL_GPL(ip_local_out);
135 static inline int ip_select_ttl(const struct inet_sock *inet,
136 const struct dst_entry *dst)
138 int ttl = READ_ONCE(inet->uc_ttl);
141 ttl = ip4_dst_hoplimit(dst);
146 * Add an ip header to a skbuff and send it out.
149 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
150 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt,
153 const struct inet_sock *inet = inet_sk(sk);
154 struct rtable *rt = skb_rtable(skb);
155 struct net *net = sock_net(sk);
158 /* Build the IP header. */
159 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
160 skb_reset_network_header(skb);
165 iph->ttl = ip_select_ttl(inet, &rt->dst);
166 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
168 iph->protocol = sk->sk_protocol;
169 /* Do not bother generating IPID for small packets (eg SYNACK) */
170 if (skb->len <= IPV4_MIN_MTU || ip_dont_fragment(sk, &rt->dst)) {
171 iph->frag_off = htons(IP_DF);
175 /* TCP packets here are SYNACK with fat IPv4/TCP options.
176 * Avoid using the hashed IP ident generator.
178 if (sk->sk_protocol == IPPROTO_TCP)
179 iph->id = (__force __be16)get_random_u16();
181 __ip_select_ident(net, iph, 1);
184 if (opt && opt->opt.optlen) {
185 iph->ihl += opt->opt.optlen>>2;
186 ip_options_build(skb, &opt->opt, daddr, rt);
189 skb->priority = READ_ONCE(sk->sk_priority);
191 skb->mark = READ_ONCE(sk->sk_mark);
194 return ip_local_out(net, skb->sk, skb);
196 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
198 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
200 struct dst_entry *dst = skb_dst(skb);
201 struct rtable *rt = (struct rtable *)dst;
202 struct net_device *dev = dst->dev;
203 unsigned int hh_len = LL_RESERVED_SPACE(dev);
204 struct neighbour *neigh;
205 bool is_v6gw = false;
207 if (rt->rt_type == RTN_MULTICAST) {
208 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
209 } else if (rt->rt_type == RTN_BROADCAST)
210 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
212 /* OUTOCTETS should be counted after fragment */
213 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
215 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
216 skb = skb_expand_head(skb, hh_len);
221 if (lwtunnel_xmit_redirect(dst->lwtstate)) {
222 int res = lwtunnel_xmit(skb);
224 if (res != LWTUNNEL_XMIT_CONTINUE)
229 neigh = ip_neigh_for_gw(rt, skb, &is_v6gw);
230 if (!IS_ERR(neigh)) {
233 sock_confirm_neigh(skb, neigh);
234 /* if crossing protocols, can not use the cached header */
235 res = neigh_output(neigh, skb, is_v6gw);
241 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
243 kfree_skb_reason(skb, SKB_DROP_REASON_NEIGH_CREATEFAIL);
244 return PTR_ERR(neigh);
247 static int ip_finish_output_gso(struct net *net, struct sock *sk,
248 struct sk_buff *skb, unsigned int mtu)
250 struct sk_buff *segs, *nskb;
251 netdev_features_t features;
254 /* common case: seglen is <= mtu
256 if (skb_gso_validate_network_len(skb, mtu))
257 return ip_finish_output2(net, sk, skb);
259 /* Slowpath - GSO segment length exceeds the egress MTU.
261 * This can happen in several cases:
262 * - Forwarding of a TCP GRO skb, when DF flag is not set.
263 * - Forwarding of an skb that arrived on a virtualization interface
264 * (virtio-net/vhost/tap) with TSO/GSO size set by other network
266 * - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
267 * interface with a smaller MTU.
268 * - Arriving GRO skb (or GSO skb in a virtualized environment) that is
269 * bridged to a NETIF_F_TSO tunnel stacked over an interface with an
272 features = netif_skb_features(skb);
273 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_GSO_CB_OFFSET);
274 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
275 if (IS_ERR_OR_NULL(segs)) {
282 skb_list_walk_safe(segs, segs, nskb) {
285 skb_mark_not_on_list(segs);
286 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
295 static int __ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
299 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
300 /* Policy lookup after SNAT yielded a new policy */
301 if (skb_dst(skb)->xfrm) {
302 IPCB(skb)->flags |= IPSKB_REROUTED;
303 return dst_output(net, sk, skb);
306 mtu = ip_skb_dst_mtu(sk, skb);
308 return ip_finish_output_gso(net, sk, skb, mtu);
310 if (skb->len > mtu || IPCB(skb)->frag_max_size)
311 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
313 return ip_finish_output2(net, sk, skb);
316 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
320 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
322 case NET_XMIT_SUCCESS:
323 return __ip_finish_output(net, sk, skb);
325 return __ip_finish_output(net, sk, skb) ? : ret;
327 kfree_skb_reason(skb, SKB_DROP_REASON_BPF_CGROUP_EGRESS);
332 static int ip_mc_finish_output(struct net *net, struct sock *sk,
335 struct rtable *new_rt;
339 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
344 case NET_XMIT_SUCCESS:
347 kfree_skb_reason(skb, SKB_DROP_REASON_BPF_CGROUP_EGRESS);
351 /* Reset rt_iif so that inet_iif() will return skb->skb_iif. Setting
352 * this to non-zero causes ipi_ifindex in in_pktinfo to be overwritten,
353 * see ipv4_pktinfo_prepare().
355 new_rt = rt_dst_clone(net->loopback_dev, skb_rtable(skb));
359 skb_dst_set(skb, &new_rt->dst);
362 err = dev_loopback_xmit(net, sk, skb);
363 return (do_cn && err) ? ret : err;
366 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
368 struct rtable *rt = skb_rtable(skb);
369 struct net_device *dev = rt->dst.dev;
372 * If the indicated interface is up and running, send the packet.
375 skb->protocol = htons(ETH_P_IP);
378 * Multicasts are looped back for other local users
381 if (rt->rt_flags&RTCF_MULTICAST) {
383 #ifdef CONFIG_IP_MROUTE
384 /* Small optimization: do not loopback not local frames,
385 which returned after forwarding; they will be dropped
386 by ip_mr_input in any case.
387 Note, that local frames are looped back to be delivered
390 This check is duplicated in ip_mr_input at the moment.
393 ((rt->rt_flags & RTCF_LOCAL) ||
394 !(IPCB(skb)->flags & IPSKB_FORWARDED))
397 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
399 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
400 net, sk, newskb, NULL, newskb->dev,
401 ip_mc_finish_output);
404 /* Multicasts with ttl 0 must not go beyond the host */
406 if (ip_hdr(skb)->ttl == 0) {
412 if (rt->rt_flags&RTCF_BROADCAST) {
413 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
415 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
416 net, sk, newskb, NULL, newskb->dev,
417 ip_mc_finish_output);
420 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
421 net, sk, skb, NULL, skb->dev,
423 !(IPCB(skb)->flags & IPSKB_REROUTED));
426 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
428 struct net_device *dev = skb_dst(skb)->dev, *indev = skb->dev;
431 skb->protocol = htons(ETH_P_IP);
433 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
434 net, sk, skb, indev, dev,
436 !(IPCB(skb)->flags & IPSKB_REROUTED));
438 EXPORT_SYMBOL(ip_output);
441 * copy saddr and daddr, possibly using 64bit load/stores
443 * iph->saddr = fl4->saddr;
444 * iph->daddr = fl4->daddr;
446 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
448 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
449 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
451 iph->saddr = fl4->saddr;
452 iph->daddr = fl4->daddr;
455 /* Note: skb->sk can be different from sk, in case of tunnels */
456 int __ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl,
459 struct inet_sock *inet = inet_sk(sk);
460 struct net *net = sock_net(sk);
461 struct ip_options_rcu *inet_opt;
467 /* Skip all of this if the packet is already routed,
468 * f.e. by something like SCTP.
471 inet_opt = rcu_dereference(inet->inet_opt);
473 rt = skb_rtable(skb);
477 /* Make sure we can route this packet. */
478 rt = (struct rtable *)__sk_dst_check(sk, 0);
482 /* Use correct destination address if we have options. */
483 daddr = inet->inet_daddr;
484 if (inet_opt && inet_opt->opt.srr)
485 daddr = inet_opt->opt.faddr;
487 /* If this fails, retransmit mechanism of transport layer will
488 * keep trying until route appears or the connection times
491 rt = ip_route_output_ports(net, fl4, sk,
492 daddr, inet->inet_saddr,
496 RT_CONN_FLAGS_TOS(sk, tos),
497 sk->sk_bound_dev_if);
500 sk_setup_caps(sk, &rt->dst);
502 skb_dst_set_noref(skb, &rt->dst);
505 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
508 /* OK, we know where to send it, allocate and build IP header. */
509 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
510 skb_reset_network_header(skb);
512 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (tos & 0xff));
513 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
514 iph->frag_off = htons(IP_DF);
517 iph->ttl = ip_select_ttl(inet, &rt->dst);
518 iph->protocol = sk->sk_protocol;
519 ip_copy_addrs(iph, fl4);
521 /* Transport layer set skb->h.foo itself. */
523 if (inet_opt && inet_opt->opt.optlen) {
524 iph->ihl += inet_opt->opt.optlen >> 2;
525 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt);
528 ip_select_ident_segs(net, skb, sk,
529 skb_shinfo(skb)->gso_segs ?: 1);
531 /* TODO : should we use skb->sk here instead of sk ? */
532 skb->priority = READ_ONCE(sk->sk_priority);
533 skb->mark = READ_ONCE(sk->sk_mark);
535 res = ip_local_out(net, sk, skb);
541 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
542 kfree_skb_reason(skb, SKB_DROP_REASON_IP_OUTNOROUTES);
543 return -EHOSTUNREACH;
545 EXPORT_SYMBOL(__ip_queue_xmit);
547 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
549 return __ip_queue_xmit(sk, skb, fl, READ_ONCE(inet_sk(sk)->tos));
551 EXPORT_SYMBOL(ip_queue_xmit);
553 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
555 to->pkt_type = from->pkt_type;
556 to->priority = from->priority;
557 to->protocol = from->protocol;
558 to->skb_iif = from->skb_iif;
560 skb_dst_copy(to, from);
562 to->mark = from->mark;
564 skb_copy_hash(to, from);
566 #ifdef CONFIG_NET_SCHED
567 to->tc_index = from->tc_index;
570 skb_ext_copy(to, from);
571 #if IS_ENABLED(CONFIG_IP_VS)
572 to->ipvs_property = from->ipvs_property;
574 skb_copy_secmark(to, from);
577 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
579 int (*output)(struct net *, struct sock *, struct sk_buff *))
581 struct iphdr *iph = ip_hdr(skb);
583 if ((iph->frag_off & htons(IP_DF)) == 0)
584 return ip_do_fragment(net, sk, skb, output);
586 if (unlikely(!skb->ignore_df ||
587 (IPCB(skb)->frag_max_size &&
588 IPCB(skb)->frag_max_size > mtu))) {
589 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
590 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
596 return ip_do_fragment(net, sk, skb, output);
599 void ip_fraglist_init(struct sk_buff *skb, struct iphdr *iph,
600 unsigned int hlen, struct ip_fraglist_iter *iter)
602 unsigned int first_len = skb_pagelen(skb);
604 iter->frag = skb_shinfo(skb)->frag_list;
605 skb_frag_list_init(skb);
611 skb->data_len = first_len - skb_headlen(skb);
612 skb->len = first_len;
613 iph->tot_len = htons(first_len);
614 iph->frag_off = htons(IP_MF);
617 EXPORT_SYMBOL(ip_fraglist_init);
619 void ip_fraglist_prepare(struct sk_buff *skb, struct ip_fraglist_iter *iter)
621 unsigned int hlen = iter->hlen;
622 struct iphdr *iph = iter->iph;
623 struct sk_buff *frag;
626 frag->ip_summed = CHECKSUM_NONE;
627 skb_reset_transport_header(frag);
628 __skb_push(frag, hlen);
629 skb_reset_network_header(frag);
630 memcpy(skb_network_header(frag), iph, hlen);
631 iter->iph = ip_hdr(frag);
633 iph->tot_len = htons(frag->len);
634 ip_copy_metadata(frag, skb);
635 iter->offset += skb->len - hlen;
636 iph->frag_off = htons(iter->offset >> 3);
638 iph->frag_off |= htons(IP_MF);
639 /* Ready, complete checksum */
642 EXPORT_SYMBOL(ip_fraglist_prepare);
644 void ip_frag_init(struct sk_buff *skb, unsigned int hlen,
645 unsigned int ll_rs, unsigned int mtu, bool DF,
646 struct ip_frag_state *state)
648 struct iphdr *iph = ip_hdr(skb);
652 state->ll_rs = ll_rs;
655 state->left = skb->len - hlen; /* Space per frame */
656 state->ptr = hlen; /* Where to start from */
658 state->offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
659 state->not_last_frag = iph->frag_off & htons(IP_MF);
661 EXPORT_SYMBOL(ip_frag_init);
663 static void ip_frag_ipcb(struct sk_buff *from, struct sk_buff *to,
666 /* Copy the flags to each fragment. */
667 IPCB(to)->flags = IPCB(from)->flags;
669 /* ANK: dirty, but effective trick. Upgrade options only if
670 * the segment to be fragmented was THE FIRST (otherwise,
671 * options are already fixed) and make it ONCE
672 * on the initial skb, so that all the following fragments
673 * will inherit fixed options.
676 ip_options_fragment(from);
679 struct sk_buff *ip_frag_next(struct sk_buff *skb, struct ip_frag_state *state)
681 unsigned int len = state->left;
682 struct sk_buff *skb2;
685 /* IF: it doesn't fit, use 'mtu' - the data space left */
686 if (len > state->mtu)
688 /* IF: we are not sending up to and including the packet end
689 then align the next start on an eight byte boundary */
690 if (len < state->left) {
694 /* Allocate buffer */
695 skb2 = alloc_skb(len + state->hlen + state->ll_rs, GFP_ATOMIC);
697 return ERR_PTR(-ENOMEM);
700 * Set up data on packet
703 ip_copy_metadata(skb2, skb);
704 skb_reserve(skb2, state->ll_rs);
705 skb_put(skb2, len + state->hlen);
706 skb_reset_network_header(skb2);
707 skb2->transport_header = skb2->network_header + state->hlen;
710 * Charge the memory for the fragment to any owner
715 skb_set_owner_w(skb2, skb->sk);
718 * Copy the packet header into the new buffer.
721 skb_copy_from_linear_data(skb, skb_network_header(skb2), state->hlen);
724 * Copy a block of the IP datagram.
726 if (skb_copy_bits(skb, state->ptr, skb_transport_header(skb2), len))
731 * Fill in the new header fields.
734 iph->frag_off = htons((state->offset >> 3));
736 iph->frag_off |= htons(IP_DF);
739 * Added AC : If we are fragmenting a fragment that's not the
740 * last fragment then keep MF on each bit
742 if (state->left > 0 || state->not_last_frag)
743 iph->frag_off |= htons(IP_MF);
745 state->offset += len;
747 iph->tot_len = htons(len + state->hlen);
753 EXPORT_SYMBOL(ip_frag_next);
756 * This IP datagram is too large to be sent in one piece. Break it up into
757 * smaller pieces (each of size equal to IP header plus
758 * a block of the data of the original IP data part) that will yet fit in a
759 * single device frame, and queue such a frame for sending.
762 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
763 int (*output)(struct net *, struct sock *, struct sk_buff *))
766 struct sk_buff *skb2;
767 bool mono_delivery_time = skb->mono_delivery_time;
768 struct rtable *rt = skb_rtable(skb);
769 unsigned int mtu, hlen, ll_rs;
770 struct ip_fraglist_iter iter;
771 ktime_t tstamp = skb->tstamp;
772 struct ip_frag_state state;
775 /* for offloaded checksums cleanup checksum before fragmentation */
776 if (skb->ip_summed == CHECKSUM_PARTIAL &&
777 (err = skb_checksum_help(skb)))
781 * Point into the IP datagram header.
786 mtu = ip_skb_dst_mtu(sk, skb);
787 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
788 mtu = IPCB(skb)->frag_max_size;
791 * Setup starting values.
795 mtu = mtu - hlen; /* Size of data space */
796 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
797 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
799 /* When frag_list is given, use it. First, check its validity:
800 * some transformers could create wrong frag_list or break existing
801 * one, it is not prohibited. In this case fall back to copying.
803 * LATER: this step can be merged to real generation of fragments,
804 * we can switch to copy when see the first bad fragment.
806 if (skb_has_frag_list(skb)) {
807 struct sk_buff *frag, *frag2;
808 unsigned int first_len = skb_pagelen(skb);
810 if (first_len - hlen > mtu ||
811 ((first_len - hlen) & 7) ||
812 ip_is_fragment(iph) ||
814 skb_headroom(skb) < ll_rs)
817 skb_walk_frags(skb, frag) {
818 /* Correct geometry. */
819 if (frag->len > mtu ||
820 ((frag->len & 7) && frag->next) ||
821 skb_headroom(frag) < hlen + ll_rs)
822 goto slow_path_clean;
824 /* Partially cloned skb? */
825 if (skb_shared(frag))
826 goto slow_path_clean;
831 frag->destructor = sock_wfree;
833 skb->truesize -= frag->truesize;
836 /* Everything is OK. Generate! */
837 ip_fraglist_init(skb, iph, hlen, &iter);
840 /* Prepare header of the next frame,
841 * before previous one went down. */
843 bool first_frag = (iter.offset == 0);
845 IPCB(iter.frag)->flags = IPCB(skb)->flags;
846 ip_fraglist_prepare(skb, &iter);
847 if (first_frag && IPCB(skb)->opt.optlen) {
848 /* ipcb->opt is not populated for frags
849 * coming from __ip_make_skb(),
850 * ip_options_fragment() needs optlen
852 IPCB(iter.frag)->opt.optlen =
853 IPCB(skb)->opt.optlen;
854 ip_options_fragment(iter.frag);
855 ip_send_check(iter.iph);
859 skb_set_delivery_time(skb, tstamp, mono_delivery_time);
860 err = output(net, sk, skb);
863 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
864 if (err || !iter.frag)
867 skb = ip_fraglist_next(&iter);
871 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
875 kfree_skb_list(iter.frag);
877 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
881 skb_walk_frags(skb, frag2) {
885 frag2->destructor = NULL;
886 skb->truesize += frag2->truesize;
892 * Fragment the datagram.
895 ip_frag_init(skb, hlen, ll_rs, mtu, IPCB(skb)->flags & IPSKB_FRAG_PMTU,
899 * Keep copying data until we run out.
902 while (state.left > 0) {
903 bool first_frag = (state.offset == 0);
905 skb2 = ip_frag_next(skb, &state);
910 ip_frag_ipcb(skb, skb2, first_frag);
913 * Put this fragment into the sending queue.
915 skb_set_delivery_time(skb2, tstamp, mono_delivery_time);
916 err = output(net, sk, skb2);
920 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
923 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
928 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
931 EXPORT_SYMBOL(ip_do_fragment);
934 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
936 struct msghdr *msg = from;
938 if (skb->ip_summed == CHECKSUM_PARTIAL) {
939 if (!copy_from_iter_full(to, len, &msg->msg_iter))
943 if (!csum_and_copy_from_iter_full(to, len, &csum, &msg->msg_iter))
945 skb->csum = csum_block_add(skb->csum, csum, odd);
949 EXPORT_SYMBOL(ip_generic_getfrag);
951 static int __ip_append_data(struct sock *sk,
953 struct sk_buff_head *queue,
954 struct inet_cork *cork,
955 struct page_frag *pfrag,
956 int getfrag(void *from, char *to, int offset,
957 int len, int odd, struct sk_buff *skb),
958 void *from, int length, int transhdrlen,
961 struct inet_sock *inet = inet_sk(sk);
962 struct ubuf_info *uarg = NULL;
964 struct ip_options *opt = cork->opt;
972 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
973 int csummode = CHECKSUM_NONE;
974 struct rtable *rt = (struct rtable *)cork->dst;
975 unsigned int wmem_alloc_delta = 0;
976 bool paged, extra_uref = false;
979 skb = skb_peek_tail(queue);
981 exthdrlen = !skb ? rt->dst.header_len : 0;
982 mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
983 paged = !!cork->gso_size;
985 if (cork->tx_flags & SKBTX_ANY_TSTAMP &&
986 READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_OPT_ID)
987 tskey = atomic_inc_return(&sk->sk_tskey) - 1;
989 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
991 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
992 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
993 maxnonfragsize = ip_sk_ignore_df(sk) ? IP_MAX_MTU : mtu;
995 if (cork->length + length > maxnonfragsize - fragheaderlen) {
996 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
997 mtu - (opt ? opt->optlen : 0));
1002 * transhdrlen > 0 means that this is the first fragment and we wish
1003 * it won't be fragmented in the future.
1006 length + fragheaderlen <= mtu &&
1007 rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
1008 (!(flags & MSG_MORE) || cork->gso_size) &&
1009 (!exthdrlen || (rt->dst.dev->features & NETIF_F_HW_ESP_TX_CSUM)))
1010 csummode = CHECKSUM_PARTIAL;
1012 if ((flags & MSG_ZEROCOPY) && length) {
1013 struct msghdr *msg = from;
1015 if (getfrag == ip_generic_getfrag && msg->msg_ubuf) {
1016 if (skb_zcopy(skb) && msg->msg_ubuf != skb_zcopy(skb))
1019 /* Leave uarg NULL if can't zerocopy, callers should
1020 * be able to handle it.
1022 if ((rt->dst.dev->features & NETIF_F_SG) &&
1023 csummode == CHECKSUM_PARTIAL) {
1026 uarg = msg->msg_ubuf;
1028 } else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1029 uarg = msg_zerocopy_realloc(sk, length, skb_zcopy(skb));
1032 extra_uref = !skb_zcopy(skb); /* only ref on new uarg */
1033 if (rt->dst.dev->features & NETIF_F_SG &&
1034 csummode == CHECKSUM_PARTIAL) {
1038 uarg_to_msgzc(uarg)->zerocopy = 0;
1039 skb_zcopy_set(skb, uarg, &extra_uref);
1042 } else if ((flags & MSG_SPLICE_PAGES) && length) {
1043 if (inet_test_bit(HDRINCL, sk))
1045 if (rt->dst.dev->features & NETIF_F_SG &&
1046 getfrag == ip_generic_getfrag)
1047 /* We need an empty buffer to attach stuff to */
1050 flags &= ~MSG_SPLICE_PAGES;
1053 cork->length += length;
1055 /* So, what's going on in the loop below?
1057 * We use calculated fragment length to generate chained skb,
1058 * each of segments is IP fragment ready for sending to network after
1059 * adding appropriate IP header.
1065 while (length > 0) {
1066 /* Check if the remaining data fits into current packet. */
1067 copy = mtu - skb->len;
1069 copy = maxfraglen - skb->len;
1072 unsigned int datalen;
1073 unsigned int fraglen;
1074 unsigned int fraggap;
1075 unsigned int alloclen, alloc_extra;
1076 unsigned int pagedlen;
1077 struct sk_buff *skb_prev;
1081 fraggap = skb_prev->len - maxfraglen;
1086 * If remaining data exceeds the mtu,
1087 * we know we need more fragment(s).
1089 datalen = length + fraggap;
1090 if (datalen > mtu - fragheaderlen)
1091 datalen = maxfraglen - fragheaderlen;
1092 fraglen = datalen + fragheaderlen;
1095 alloc_extra = hh_len + 15;
1096 alloc_extra += exthdrlen;
1098 /* The last fragment gets additional space at tail.
1099 * Note, with MSG_MORE we overallocate on fragments,
1100 * because we have no idea what fragment will be
1103 if (datalen == length + fraggap)
1104 alloc_extra += rt->dst.trailer_len;
1106 if ((flags & MSG_MORE) &&
1107 !(rt->dst.dev->features&NETIF_F_SG))
1110 (fraglen + alloc_extra < SKB_MAX_ALLOC ||
1111 !(rt->dst.dev->features & NETIF_F_SG)))
1114 alloclen = fragheaderlen + transhdrlen;
1115 pagedlen = datalen - transhdrlen;
1118 alloclen += alloc_extra;
1121 skb = sock_alloc_send_skb(sk, alloclen,
1122 (flags & MSG_DONTWAIT), &err);
1125 if (refcount_read(&sk->sk_wmem_alloc) + wmem_alloc_delta <=
1127 skb = alloc_skb(alloclen,
1136 * Fill in the control structures
1138 skb->ip_summed = csummode;
1140 skb_reserve(skb, hh_len);
1143 * Find where to start putting bytes.
1145 data = skb_put(skb, fraglen + exthdrlen - pagedlen);
1146 skb_set_network_header(skb, exthdrlen);
1147 skb->transport_header = (skb->network_header +
1149 data += fragheaderlen + exthdrlen;
1152 skb->csum = skb_copy_and_csum_bits(
1153 skb_prev, maxfraglen,
1154 data + transhdrlen, fraggap);
1155 skb_prev->csum = csum_sub(skb_prev->csum,
1158 pskb_trim_unique(skb_prev, maxfraglen);
1161 copy = datalen - transhdrlen - fraggap - pagedlen;
1162 /* [!] NOTE: copy will be negative if pagedlen>0
1163 * because then the equation reduces to -fraggap.
1165 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1169 } else if (flags & MSG_SPLICE_PAGES) {
1174 length -= copy + transhdrlen;
1177 csummode = CHECKSUM_NONE;
1179 /* only the initial fragment is time stamped */
1180 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1182 skb_shinfo(skb)->tskey = tskey;
1184 skb_zcopy_set(skb, uarg, &extra_uref);
1186 if ((flags & MSG_CONFIRM) && !skb_prev)
1187 skb_set_dst_pending_confirm(skb, 1);
1190 * Put the packet on the pending queue.
1192 if (!skb->destructor) {
1193 skb->destructor = sock_wfree;
1195 wmem_alloc_delta += skb->truesize;
1197 __skb_queue_tail(queue, skb);
1204 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1205 skb_tailroom(skb) >= copy) {
1209 if (getfrag(from, skb_put(skb, copy),
1210 offset, copy, off, skb) < 0) {
1211 __skb_trim(skb, off);
1215 } else if (flags & MSG_SPLICE_PAGES) {
1216 struct msghdr *msg = from;
1219 if (WARN_ON_ONCE(copy > msg->msg_iter.count))
1222 err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
1227 wmem_alloc_delta += copy;
1229 int i = skb_shinfo(skb)->nr_frags;
1232 if (!sk_page_frag_refill(sk, pfrag))
1235 skb_zcopy_downgrade_managed(skb);
1236 if (!skb_can_coalesce(skb, i, pfrag->page,
1239 if (i == MAX_SKB_FRAGS)
1242 __skb_fill_page_desc(skb, i, pfrag->page,
1244 skb_shinfo(skb)->nr_frags = ++i;
1245 get_page(pfrag->page);
1247 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1249 page_address(pfrag->page) + pfrag->offset,
1250 offset, copy, skb->len, skb) < 0)
1253 pfrag->offset += copy;
1254 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1255 skb_len_add(skb, copy);
1256 wmem_alloc_delta += copy;
1258 err = skb_zerocopy_iter_dgram(skb, from, copy);
1266 if (wmem_alloc_delta)
1267 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1273 net_zcopy_put_abort(uarg, extra_uref);
1274 cork->length -= length;
1275 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1276 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1280 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1281 struct ipcm_cookie *ipc, struct rtable **rtp)
1283 struct ip_options_rcu *opt;
1291 * setup for corking.
1296 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1298 if (unlikely(!cork->opt))
1301 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1302 cork->flags |= IPCORK_OPT;
1303 cork->addr = ipc->addr;
1306 cork->fragsize = ip_sk_use_pmtu(sk) ?
1307 dst_mtu(&rt->dst) : READ_ONCE(rt->dst.dev->mtu);
1309 if (!inetdev_valid_mtu(cork->fragsize))
1310 return -ENETUNREACH;
1312 cork->gso_size = ipc->gso_size;
1314 cork->dst = &rt->dst;
1315 /* We stole this route, caller should not release it. */
1319 cork->ttl = ipc->ttl;
1320 cork->tos = ipc->tos;
1321 cork->mark = ipc->sockc.mark;
1322 cork->priority = ipc->priority;
1323 cork->transmit_time = ipc->sockc.transmit_time;
1325 sock_tx_timestamp(sk, ipc->sockc.tsflags, &cork->tx_flags);
1331 * ip_append_data() can make one large IP datagram from many pieces of
1332 * data. Each piece will be held on the socket until
1333 * ip_push_pending_frames() is called. Each piece can be a page or
1336 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1337 * this interface potentially.
1339 * LATER: length must be adjusted by pad at tail, when it is required.
1341 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1342 int getfrag(void *from, char *to, int offset, int len,
1343 int odd, struct sk_buff *skb),
1344 void *from, int length, int transhdrlen,
1345 struct ipcm_cookie *ipc, struct rtable **rtp,
1348 struct inet_sock *inet = inet_sk(sk);
1351 if (flags&MSG_PROBE)
1354 if (skb_queue_empty(&sk->sk_write_queue)) {
1355 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1362 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1363 sk_page_frag(sk), getfrag,
1364 from, length, transhdrlen, flags);
1367 static void ip_cork_release(struct inet_cork *cork)
1369 cork->flags &= ~IPCORK_OPT;
1372 dst_release(cork->dst);
1377 * Combined all pending IP fragments on the socket as one IP datagram
1378 * and push them out.
1380 struct sk_buff *__ip_make_skb(struct sock *sk,
1382 struct sk_buff_head *queue,
1383 struct inet_cork *cork)
1385 struct sk_buff *skb, *tmp_skb;
1386 struct sk_buff **tail_skb;
1387 struct inet_sock *inet = inet_sk(sk);
1388 struct net *net = sock_net(sk);
1389 struct ip_options *opt = NULL;
1390 struct rtable *rt = (struct rtable *)cork->dst;
1395 skb = __skb_dequeue(queue);
1398 tail_skb = &(skb_shinfo(skb)->frag_list);
1400 /* move skb->data to ip header from ext header */
1401 if (skb->data < skb_network_header(skb))
1402 __skb_pull(skb, skb_network_offset(skb));
1403 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1404 __skb_pull(tmp_skb, skb_network_header_len(skb));
1405 *tail_skb = tmp_skb;
1406 tail_skb = &(tmp_skb->next);
1407 skb->len += tmp_skb->len;
1408 skb->data_len += tmp_skb->len;
1409 skb->truesize += tmp_skb->truesize;
1410 tmp_skb->destructor = NULL;
1414 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1415 * to fragment the frame generated here. No matter, what transforms
1416 * how transforms change size of the packet, it will come out.
1418 skb->ignore_df = ip_sk_ignore_df(sk);
1420 /* DF bit is set when we want to see DF on outgoing frames.
1421 * If ignore_df is set too, we still allow to fragment this frame
1423 pmtudisc = READ_ONCE(inet->pmtudisc);
1424 if (pmtudisc == IP_PMTUDISC_DO ||
1425 pmtudisc == IP_PMTUDISC_PROBE ||
1426 (skb->len <= dst_mtu(&rt->dst) &&
1427 ip_dont_fragment(sk, &rt->dst)))
1430 if (cork->flags & IPCORK_OPT)
1435 else if (rt->rt_type == RTN_MULTICAST)
1436 ttl = READ_ONCE(inet->mc_ttl);
1438 ttl = ip_select_ttl(inet, &rt->dst);
1443 iph->tos = (cork->tos != -1) ? cork->tos : READ_ONCE(inet->tos);
1446 iph->protocol = sk->sk_protocol;
1447 ip_copy_addrs(iph, fl4);
1448 ip_select_ident(net, skb, sk);
1451 iph->ihl += opt->optlen >> 2;
1452 ip_options_build(skb, opt, cork->addr, rt);
1455 skb->priority = (cork->tos != -1) ? cork->priority: READ_ONCE(sk->sk_priority);
1456 skb->mark = cork->mark;
1457 skb->tstamp = cork->transmit_time;
1459 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1463 skb_dst_set(skb, &rt->dst);
1465 if (iph->protocol == IPPROTO_ICMP) {
1468 /* For such sockets, transhdrlen is zero when do ip_append_data(),
1469 * so icmphdr does not in skb linear region and can not get icmp_type
1470 * by icmp_hdr(skb)->type.
1472 if (sk->sk_type == SOCK_RAW &&
1473 !inet_test_bit(HDRINCL, sk))
1474 icmp_type = fl4->fl4_icmp_type;
1476 icmp_type = icmp_hdr(skb)->type;
1477 icmp_out_count(net, icmp_type);
1480 ip_cork_release(cork);
1485 int ip_send_skb(struct net *net, struct sk_buff *skb)
1489 err = ip_local_out(net, skb->sk, skb);
1492 err = net_xmit_errno(err);
1494 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1500 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1502 struct sk_buff *skb;
1504 skb = ip_finish_skb(sk, fl4);
1508 /* Netfilter gets whole the not fragmented skb. */
1509 return ip_send_skb(sock_net(sk), skb);
1513 * Throw away all pending data on the socket.
1515 static void __ip_flush_pending_frames(struct sock *sk,
1516 struct sk_buff_head *queue,
1517 struct inet_cork *cork)
1519 struct sk_buff *skb;
1521 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1524 ip_cork_release(cork);
1527 void ip_flush_pending_frames(struct sock *sk)
1529 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1532 struct sk_buff *ip_make_skb(struct sock *sk,
1534 int getfrag(void *from, char *to, int offset,
1535 int len, int odd, struct sk_buff *skb),
1536 void *from, int length, int transhdrlen,
1537 struct ipcm_cookie *ipc, struct rtable **rtp,
1538 struct inet_cork *cork, unsigned int flags)
1540 struct sk_buff_head queue;
1543 if (flags & MSG_PROBE)
1546 __skb_queue_head_init(&queue);
1551 err = ip_setup_cork(sk, cork, ipc, rtp);
1553 return ERR_PTR(err);
1555 err = __ip_append_data(sk, fl4, &queue, cork,
1556 ¤t->task_frag, getfrag,
1557 from, length, transhdrlen, flags);
1559 __ip_flush_pending_frames(sk, &queue, cork);
1560 return ERR_PTR(err);
1563 return __ip_make_skb(sk, fl4, &queue, cork);
1567 * Fetch data from kernel space and fill in checksum if needed.
1569 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1570 int len, int odd, struct sk_buff *skb)
1574 csum = csum_partial_copy_nocheck(dptr+offset, to, len);
1575 skb->csum = csum_block_add(skb->csum, csum, odd);
1580 * Generic function to send a packet as reply to another packet.
1581 * Used to send some TCP resets/acks so far.
1583 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1584 const struct ip_options *sopt,
1585 __be32 daddr, __be32 saddr,
1586 const struct ip_reply_arg *arg,
1587 unsigned int len, u64 transmit_time, u32 txhash)
1589 struct ip_options_data replyopts;
1590 struct ipcm_cookie ipc;
1592 struct rtable *rt = skb_rtable(skb);
1593 struct net *net = sock_net(sk);
1594 struct sk_buff *nskb;
1598 if (__ip_options_echo(net, &replyopts.opt.opt, skb, sopt))
1603 ipc.sockc.transmit_time = transmit_time;
1605 if (replyopts.opt.opt.optlen) {
1606 ipc.opt = &replyopts.opt;
1608 if (replyopts.opt.opt.srr)
1609 daddr = replyopts.opt.opt.faddr;
1612 oif = arg->bound_dev_if;
1613 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1616 flowi4_init_output(&fl4, oif,
1617 IP4_REPLY_MARK(net, skb->mark) ?: sk->sk_mark,
1619 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1620 ip_reply_arg_flowi_flags(arg),
1622 tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1624 security_skb_classify_flow(skb, flowi4_to_flowi_common(&fl4));
1625 rt = ip_route_output_flow(net, &fl4, sk);
1629 inet_sk(sk)->tos = arg->tos & ~INET_ECN_MASK;
1631 sk->sk_protocol = ip_hdr(skb)->protocol;
1632 sk->sk_bound_dev_if = arg->bound_dev_if;
1633 sk->sk_sndbuf = READ_ONCE(sysctl_wmem_default);
1634 ipc.sockc.mark = fl4.flowi4_mark;
1635 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1636 len, 0, &ipc, &rt, MSG_DONTWAIT);
1637 if (unlikely(err)) {
1638 ip_flush_pending_frames(sk);
1642 nskb = skb_peek(&sk->sk_write_queue);
1644 if (arg->csumoffset >= 0)
1645 *((__sum16 *)skb_transport_header(nskb) +
1646 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1648 nskb->ip_summed = CHECKSUM_NONE;
1649 nskb->mono_delivery_time = !!transmit_time;
1651 skb_set_hash(nskb, txhash, PKT_HASH_TYPE_L4);
1652 ip_push_pending_frames(sk, &fl4);
1658 void __init ip_init(void)
1663 #if defined(CONFIG_IP_MULTICAST)