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 = 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 = sk->sk_priority;
189 skb->mark = 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 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
211 skb = skb_expand_head(skb, hh_len);
216 if (lwtunnel_xmit_redirect(dst->lwtstate)) {
217 int res = lwtunnel_xmit(skb);
219 if (res < 0 || res == LWTUNNEL_XMIT_DONE)
224 neigh = ip_neigh_for_gw(rt, skb, &is_v6gw);
225 if (!IS_ERR(neigh)) {
228 sock_confirm_neigh(skb, neigh);
229 /* if crossing protocols, can not use the cached header */
230 res = neigh_output(neigh, skb, is_v6gw);
236 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
238 kfree_skb_reason(skb, SKB_DROP_REASON_NEIGH_CREATEFAIL);
242 static int ip_finish_output_gso(struct net *net, struct sock *sk,
243 struct sk_buff *skb, unsigned int mtu)
245 struct sk_buff *segs, *nskb;
246 netdev_features_t features;
249 /* common case: seglen is <= mtu
251 if (skb_gso_validate_network_len(skb, mtu))
252 return ip_finish_output2(net, sk, skb);
254 /* Slowpath - GSO segment length exceeds the egress MTU.
256 * This can happen in several cases:
257 * - Forwarding of a TCP GRO skb, when DF flag is not set.
258 * - Forwarding of an skb that arrived on a virtualization interface
259 * (virtio-net/vhost/tap) with TSO/GSO size set by other network
261 * - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
262 * interface with a smaller MTU.
263 * - Arriving GRO skb (or GSO skb in a virtualized environment) that is
264 * bridged to a NETIF_F_TSO tunnel stacked over an interface with an
267 features = netif_skb_features(skb);
268 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_GSO_CB_OFFSET);
269 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
270 if (IS_ERR_OR_NULL(segs)) {
277 skb_list_walk_safe(segs, segs, nskb) {
280 skb_mark_not_on_list(segs);
281 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
290 static int __ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
294 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
295 /* Policy lookup after SNAT yielded a new policy */
296 if (skb_dst(skb)->xfrm) {
297 IPCB(skb)->flags |= IPSKB_REROUTED;
298 return dst_output(net, sk, skb);
301 mtu = ip_skb_dst_mtu(sk, skb);
303 return ip_finish_output_gso(net, sk, skb, mtu);
305 if (skb->len > mtu || IPCB(skb)->frag_max_size)
306 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
308 return ip_finish_output2(net, sk, skb);
311 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
315 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
317 case NET_XMIT_SUCCESS:
318 return __ip_finish_output(net, sk, skb);
320 return __ip_finish_output(net, sk, skb) ? : ret;
322 kfree_skb_reason(skb, SKB_DROP_REASON_BPF_CGROUP_EGRESS);
327 static int ip_mc_finish_output(struct net *net, struct sock *sk,
330 struct rtable *new_rt;
334 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
339 case NET_XMIT_SUCCESS:
342 kfree_skb_reason(skb, SKB_DROP_REASON_BPF_CGROUP_EGRESS);
346 /* Reset rt_iif so that inet_iif() will return skb->skb_iif. Setting
347 * this to non-zero causes ipi_ifindex in in_pktinfo to be overwritten,
348 * see ipv4_pktinfo_prepare().
350 new_rt = rt_dst_clone(net->loopback_dev, skb_rtable(skb));
354 skb_dst_set(skb, &new_rt->dst);
357 err = dev_loopback_xmit(net, sk, skb);
358 return (do_cn && err) ? ret : err;
361 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
363 struct rtable *rt = skb_rtable(skb);
364 struct net_device *dev = rt->dst.dev;
367 * If the indicated interface is up and running, send the packet.
369 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
372 skb->protocol = htons(ETH_P_IP);
375 * Multicasts are looped back for other local users
378 if (rt->rt_flags&RTCF_MULTICAST) {
380 #ifdef CONFIG_IP_MROUTE
381 /* Small optimization: do not loopback not local frames,
382 which returned after forwarding; they will be dropped
383 by ip_mr_input in any case.
384 Note, that local frames are looped back to be delivered
387 This check is duplicated in ip_mr_input at the moment.
390 ((rt->rt_flags & RTCF_LOCAL) ||
391 !(IPCB(skb)->flags & IPSKB_FORWARDED))
394 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
396 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
397 net, sk, newskb, NULL, newskb->dev,
398 ip_mc_finish_output);
401 /* Multicasts with ttl 0 must not go beyond the host */
403 if (ip_hdr(skb)->ttl == 0) {
409 if (rt->rt_flags&RTCF_BROADCAST) {
410 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
412 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
413 net, sk, newskb, NULL, newskb->dev,
414 ip_mc_finish_output);
417 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
418 net, sk, skb, NULL, skb->dev,
420 !(IPCB(skb)->flags & IPSKB_REROUTED));
423 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
425 struct net_device *dev = skb_dst(skb)->dev, *indev = skb->dev;
427 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
430 skb->protocol = htons(ETH_P_IP);
432 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
433 net, sk, skb, indev, dev,
435 !(IPCB(skb)->flags & IPSKB_REROUTED));
437 EXPORT_SYMBOL(ip_output);
440 * copy saddr and daddr, possibly using 64bit load/stores
442 * iph->saddr = fl4->saddr;
443 * iph->daddr = fl4->daddr;
445 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
447 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
448 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
450 iph->saddr = fl4->saddr;
451 iph->daddr = fl4->daddr;
454 /* Note: skb->sk can be different from sk, in case of tunnels */
455 int __ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl,
458 struct inet_sock *inet = inet_sk(sk);
459 struct net *net = sock_net(sk);
460 struct ip_options_rcu *inet_opt;
466 /* Skip all of this if the packet is already routed,
467 * f.e. by something like SCTP.
470 inet_opt = rcu_dereference(inet->inet_opt);
472 rt = skb_rtable(skb);
476 /* Make sure we can route this packet. */
477 rt = (struct rtable *)__sk_dst_check(sk, 0);
481 /* Use correct destination address if we have options. */
482 daddr = inet->inet_daddr;
483 if (inet_opt && inet_opt->opt.srr)
484 daddr = inet_opt->opt.faddr;
486 /* If this fails, retransmit mechanism of transport layer will
487 * keep trying until route appears or the connection times
490 rt = ip_route_output_ports(net, fl4, sk,
491 daddr, inet->inet_saddr,
495 RT_CONN_FLAGS_TOS(sk, tos),
496 sk->sk_bound_dev_if);
499 sk_setup_caps(sk, &rt->dst);
501 skb_dst_set_noref(skb, &rt->dst);
504 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
507 /* OK, we know where to send it, allocate and build IP header. */
508 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
509 skb_reset_network_header(skb);
511 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (tos & 0xff));
512 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
513 iph->frag_off = htons(IP_DF);
516 iph->ttl = ip_select_ttl(inet, &rt->dst);
517 iph->protocol = sk->sk_protocol;
518 ip_copy_addrs(iph, fl4);
520 /* Transport layer set skb->h.foo itself. */
522 if (inet_opt && inet_opt->opt.optlen) {
523 iph->ihl += inet_opt->opt.optlen >> 2;
524 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt);
527 ip_select_ident_segs(net, skb, sk,
528 skb_shinfo(skb)->gso_segs ?: 1);
530 /* TODO : should we use skb->sk here instead of sk ? */
531 skb->priority = sk->sk_priority;
532 skb->mark = sk->sk_mark;
534 res = ip_local_out(net, sk, skb);
540 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
541 kfree_skb_reason(skb, SKB_DROP_REASON_IP_OUTNOROUTES);
542 return -EHOSTUNREACH;
544 EXPORT_SYMBOL(__ip_queue_xmit);
546 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
548 return __ip_queue_xmit(sk, skb, fl, inet_sk(sk)->tos);
550 EXPORT_SYMBOL(ip_queue_xmit);
552 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
554 to->pkt_type = from->pkt_type;
555 to->priority = from->priority;
556 to->protocol = from->protocol;
557 to->skb_iif = from->skb_iif;
559 skb_dst_copy(to, from);
561 to->mark = from->mark;
563 skb_copy_hash(to, from);
565 #ifdef CONFIG_NET_SCHED
566 to->tc_index = from->tc_index;
569 skb_ext_copy(to, from);
570 #if IS_ENABLED(CONFIG_IP_VS)
571 to->ipvs_property = from->ipvs_property;
573 skb_copy_secmark(to, from);
576 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
578 int (*output)(struct net *, struct sock *, struct sk_buff *))
580 struct iphdr *iph = ip_hdr(skb);
582 if ((iph->frag_off & htons(IP_DF)) == 0)
583 return ip_do_fragment(net, sk, skb, output);
585 if (unlikely(!skb->ignore_df ||
586 (IPCB(skb)->frag_max_size &&
587 IPCB(skb)->frag_max_size > mtu))) {
588 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
589 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
595 return ip_do_fragment(net, sk, skb, output);
598 void ip_fraglist_init(struct sk_buff *skb, struct iphdr *iph,
599 unsigned int hlen, struct ip_fraglist_iter *iter)
601 unsigned int first_len = skb_pagelen(skb);
603 iter->frag = skb_shinfo(skb)->frag_list;
604 skb_frag_list_init(skb);
610 skb->data_len = first_len - skb_headlen(skb);
611 skb->len = first_len;
612 iph->tot_len = htons(first_len);
613 iph->frag_off = htons(IP_MF);
616 EXPORT_SYMBOL(ip_fraglist_init);
618 void ip_fraglist_prepare(struct sk_buff *skb, struct ip_fraglist_iter *iter)
620 unsigned int hlen = iter->hlen;
621 struct iphdr *iph = iter->iph;
622 struct sk_buff *frag;
625 frag->ip_summed = CHECKSUM_NONE;
626 skb_reset_transport_header(frag);
627 __skb_push(frag, hlen);
628 skb_reset_network_header(frag);
629 memcpy(skb_network_header(frag), iph, hlen);
630 iter->iph = ip_hdr(frag);
632 iph->tot_len = htons(frag->len);
633 ip_copy_metadata(frag, skb);
634 iter->offset += skb->len - hlen;
635 iph->frag_off = htons(iter->offset >> 3);
637 iph->frag_off |= htons(IP_MF);
638 /* Ready, complete checksum */
641 EXPORT_SYMBOL(ip_fraglist_prepare);
643 void ip_frag_init(struct sk_buff *skb, unsigned int hlen,
644 unsigned int ll_rs, unsigned int mtu, bool DF,
645 struct ip_frag_state *state)
647 struct iphdr *iph = ip_hdr(skb);
651 state->ll_rs = ll_rs;
654 state->left = skb->len - hlen; /* Space per frame */
655 state->ptr = hlen; /* Where to start from */
657 state->offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
658 state->not_last_frag = iph->frag_off & htons(IP_MF);
660 EXPORT_SYMBOL(ip_frag_init);
662 static void ip_frag_ipcb(struct sk_buff *from, struct sk_buff *to,
665 /* Copy the flags to each fragment. */
666 IPCB(to)->flags = IPCB(from)->flags;
668 /* ANK: dirty, but effective trick. Upgrade options only if
669 * the segment to be fragmented was THE FIRST (otherwise,
670 * options are already fixed) and make it ONCE
671 * on the initial skb, so that all the following fragments
672 * will inherit fixed options.
675 ip_options_fragment(from);
678 struct sk_buff *ip_frag_next(struct sk_buff *skb, struct ip_frag_state *state)
680 unsigned int len = state->left;
681 struct sk_buff *skb2;
684 /* IF: it doesn't fit, use 'mtu' - the data space left */
685 if (len > state->mtu)
687 /* IF: we are not sending up to and including the packet end
688 then align the next start on an eight byte boundary */
689 if (len < state->left) {
693 /* Allocate buffer */
694 skb2 = alloc_skb(len + state->hlen + state->ll_rs, GFP_ATOMIC);
696 return ERR_PTR(-ENOMEM);
699 * Set up data on packet
702 ip_copy_metadata(skb2, skb);
703 skb_reserve(skb2, state->ll_rs);
704 skb_put(skb2, len + state->hlen);
705 skb_reset_network_header(skb2);
706 skb2->transport_header = skb2->network_header + state->hlen;
709 * Charge the memory for the fragment to any owner
714 skb_set_owner_w(skb2, skb->sk);
717 * Copy the packet header into the new buffer.
720 skb_copy_from_linear_data(skb, skb_network_header(skb2), state->hlen);
723 * Copy a block of the IP datagram.
725 if (skb_copy_bits(skb, state->ptr, skb_transport_header(skb2), len))
730 * Fill in the new header fields.
733 iph->frag_off = htons((state->offset >> 3));
735 iph->frag_off |= htons(IP_DF);
738 * Added AC : If we are fragmenting a fragment that's not the
739 * last fragment then keep MF on each bit
741 if (state->left > 0 || state->not_last_frag)
742 iph->frag_off |= htons(IP_MF);
744 state->offset += len;
746 iph->tot_len = htons(len + state->hlen);
752 EXPORT_SYMBOL(ip_frag_next);
755 * This IP datagram is too large to be sent in one piece. Break it up into
756 * smaller pieces (each of size equal to IP header plus
757 * a block of the data of the original IP data part) that will yet fit in a
758 * single device frame, and queue such a frame for sending.
761 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
762 int (*output)(struct net *, struct sock *, struct sk_buff *))
765 struct sk_buff *skb2;
766 bool mono_delivery_time = skb->mono_delivery_time;
767 struct rtable *rt = skb_rtable(skb);
768 unsigned int mtu, hlen, ll_rs;
769 struct ip_fraglist_iter iter;
770 ktime_t tstamp = skb->tstamp;
771 struct ip_frag_state state;
774 /* for offloaded checksums cleanup checksum before fragmentation */
775 if (skb->ip_summed == CHECKSUM_PARTIAL &&
776 (err = skb_checksum_help(skb)))
780 * Point into the IP datagram header.
785 mtu = ip_skb_dst_mtu(sk, skb);
786 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
787 mtu = IPCB(skb)->frag_max_size;
790 * Setup starting values.
794 mtu = mtu - hlen; /* Size of data space */
795 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
796 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
798 /* When frag_list is given, use it. First, check its validity:
799 * some transformers could create wrong frag_list or break existing
800 * one, it is not prohibited. In this case fall back to copying.
802 * LATER: this step can be merged to real generation of fragments,
803 * we can switch to copy when see the first bad fragment.
805 if (skb_has_frag_list(skb)) {
806 struct sk_buff *frag, *frag2;
807 unsigned int first_len = skb_pagelen(skb);
809 if (first_len - hlen > mtu ||
810 ((first_len - hlen) & 7) ||
811 ip_is_fragment(iph) ||
813 skb_headroom(skb) < ll_rs)
816 skb_walk_frags(skb, frag) {
817 /* Correct geometry. */
818 if (frag->len > mtu ||
819 ((frag->len & 7) && frag->next) ||
820 skb_headroom(frag) < hlen + ll_rs)
821 goto slow_path_clean;
823 /* Partially cloned skb? */
824 if (skb_shared(frag))
825 goto slow_path_clean;
830 frag->destructor = sock_wfree;
832 skb->truesize -= frag->truesize;
835 /* Everything is OK. Generate! */
836 ip_fraglist_init(skb, iph, hlen, &iter);
839 /* Prepare header of the next frame,
840 * before previous one went down. */
842 bool first_frag = (iter.offset == 0);
844 IPCB(iter.frag)->flags = IPCB(skb)->flags;
845 ip_fraglist_prepare(skb, &iter);
846 if (first_frag && IPCB(skb)->opt.optlen) {
847 /* ipcb->opt is not populated for frags
848 * coming from __ip_make_skb(),
849 * ip_options_fragment() needs optlen
851 IPCB(iter.frag)->opt.optlen =
852 IPCB(skb)->opt.optlen;
853 ip_options_fragment(iter.frag);
854 ip_send_check(iter.iph);
858 skb_set_delivery_time(skb, tstamp, mono_delivery_time);
859 err = output(net, sk, skb);
862 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
863 if (err || !iter.frag)
866 skb = ip_fraglist_next(&iter);
870 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
874 kfree_skb_list(iter.frag);
876 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
880 skb_walk_frags(skb, frag2) {
884 frag2->destructor = NULL;
885 skb->truesize += frag2->truesize;
891 * Fragment the datagram.
894 ip_frag_init(skb, hlen, ll_rs, mtu, IPCB(skb)->flags & IPSKB_FRAG_PMTU,
898 * Keep copying data until we run out.
901 while (state.left > 0) {
902 bool first_frag = (state.offset == 0);
904 skb2 = ip_frag_next(skb, &state);
909 ip_frag_ipcb(skb, skb2, first_frag);
912 * Put this fragment into the sending queue.
914 skb_set_delivery_time(skb2, tstamp, mono_delivery_time);
915 err = output(net, sk, skb2);
919 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
922 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
927 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
930 EXPORT_SYMBOL(ip_do_fragment);
933 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
935 struct msghdr *msg = from;
937 if (skb->ip_summed == CHECKSUM_PARTIAL) {
938 if (!copy_from_iter_full(to, len, &msg->msg_iter))
942 if (!csum_and_copy_from_iter_full(to, len, &csum, &msg->msg_iter))
944 skb->csum = csum_block_add(skb->csum, csum, odd);
948 EXPORT_SYMBOL(ip_generic_getfrag);
950 static int __ip_append_data(struct sock *sk,
952 struct sk_buff_head *queue,
953 struct inet_cork *cork,
954 struct page_frag *pfrag,
955 int getfrag(void *from, char *to, int offset,
956 int len, int odd, struct sk_buff *skb),
957 void *from, int length, int transhdrlen,
960 struct inet_sock *inet = inet_sk(sk);
961 struct ubuf_info *uarg = NULL;
963 struct ip_options *opt = cork->opt;
971 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
972 int csummode = CHECKSUM_NONE;
973 struct rtable *rt = (struct rtable *)cork->dst;
974 unsigned int wmem_alloc_delta = 0;
975 bool paged, extra_uref = false;
978 skb = skb_peek_tail(queue);
980 exthdrlen = !skb ? rt->dst.header_len : 0;
981 mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
982 paged = !!cork->gso_size;
984 if (cork->tx_flags & SKBTX_ANY_TSTAMP &&
985 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
986 tskey = atomic_inc_return(&sk->sk_tskey) - 1;
988 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
990 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
991 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
992 maxnonfragsize = ip_sk_ignore_df(sk) ? IP_MAX_MTU : mtu;
994 if (cork->length + length > maxnonfragsize - fragheaderlen) {
995 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
996 mtu - (opt ? opt->optlen : 0));
1001 * transhdrlen > 0 means that this is the first fragment and we wish
1002 * it won't be fragmented in the future.
1005 length + fragheaderlen <= mtu &&
1006 rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
1007 (!(flags & MSG_MORE) || cork->gso_size) &&
1008 (!exthdrlen || (rt->dst.dev->features & NETIF_F_HW_ESP_TX_CSUM)))
1009 csummode = CHECKSUM_PARTIAL;
1011 if ((flags & MSG_ZEROCOPY) && length) {
1012 struct msghdr *msg = from;
1014 if (getfrag == ip_generic_getfrag && msg->msg_ubuf) {
1015 if (skb_zcopy(skb) && msg->msg_ubuf != skb_zcopy(skb))
1018 /* Leave uarg NULL if can't zerocopy, callers should
1019 * be able to handle it.
1021 if ((rt->dst.dev->features & NETIF_F_SG) &&
1022 csummode == CHECKSUM_PARTIAL) {
1025 uarg = msg->msg_ubuf;
1027 } else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1028 uarg = msg_zerocopy_realloc(sk, length, skb_zcopy(skb));
1031 extra_uref = !skb_zcopy(skb); /* only ref on new uarg */
1032 if (rt->dst.dev->features & NETIF_F_SG &&
1033 csummode == CHECKSUM_PARTIAL) {
1037 uarg_to_msgzc(uarg)->zerocopy = 0;
1038 skb_zcopy_set(skb, uarg, &extra_uref);
1041 } else if ((flags & MSG_SPLICE_PAGES) && length) {
1044 if (rt->dst.dev->features & NETIF_F_SG &&
1045 getfrag == ip_generic_getfrag)
1046 /* We need an empty buffer to attach stuff to */
1049 flags &= ~MSG_SPLICE_PAGES;
1052 cork->length += length;
1054 /* So, what's going on in the loop below?
1056 * We use calculated fragment length to generate chained skb,
1057 * each of segments is IP fragment ready for sending to network after
1058 * adding appropriate IP header.
1064 while (length > 0) {
1065 /* Check if the remaining data fits into current packet. */
1066 copy = mtu - skb->len;
1068 copy = maxfraglen - skb->len;
1071 unsigned int datalen;
1072 unsigned int fraglen;
1073 unsigned int fraggap;
1074 unsigned int alloclen, alloc_extra;
1075 unsigned int pagedlen;
1076 struct sk_buff *skb_prev;
1080 fraggap = skb_prev->len - maxfraglen;
1085 * If remaining data exceeds the mtu,
1086 * we know we need more fragment(s).
1088 datalen = length + fraggap;
1089 if (datalen > mtu - fragheaderlen)
1090 datalen = maxfraglen - fragheaderlen;
1091 fraglen = datalen + fragheaderlen;
1094 alloc_extra = hh_len + 15;
1095 alloc_extra += exthdrlen;
1097 /* The last fragment gets additional space at tail.
1098 * Note, with MSG_MORE we overallocate on fragments,
1099 * because we have no idea what fragment will be
1102 if (datalen == length + fraggap)
1103 alloc_extra += rt->dst.trailer_len;
1105 if ((flags & MSG_MORE) &&
1106 !(rt->dst.dev->features&NETIF_F_SG))
1109 (fraglen + alloc_extra < SKB_MAX_ALLOC ||
1110 !(rt->dst.dev->features & NETIF_F_SG)))
1113 alloclen = fragheaderlen + transhdrlen;
1114 pagedlen = datalen - transhdrlen;
1117 alloclen += alloc_extra;
1120 skb = sock_alloc_send_skb(sk, alloclen,
1121 (flags & MSG_DONTWAIT), &err);
1124 if (refcount_read(&sk->sk_wmem_alloc) + wmem_alloc_delta <=
1126 skb = alloc_skb(alloclen,
1135 * Fill in the control structures
1137 skb->ip_summed = csummode;
1139 skb_reserve(skb, hh_len);
1142 * Find where to start putting bytes.
1144 data = skb_put(skb, fraglen + exthdrlen - pagedlen);
1145 skb_set_network_header(skb, exthdrlen);
1146 skb->transport_header = (skb->network_header +
1148 data += fragheaderlen + exthdrlen;
1151 skb->csum = skb_copy_and_csum_bits(
1152 skb_prev, maxfraglen,
1153 data + transhdrlen, fraggap);
1154 skb_prev->csum = csum_sub(skb_prev->csum,
1157 pskb_trim_unique(skb_prev, maxfraglen);
1160 copy = datalen - transhdrlen - fraggap - pagedlen;
1161 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1168 length -= copy + transhdrlen;
1171 csummode = CHECKSUM_NONE;
1173 /* only the initial fragment is time stamped */
1174 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1176 skb_shinfo(skb)->tskey = tskey;
1178 skb_zcopy_set(skb, uarg, &extra_uref);
1180 if ((flags & MSG_CONFIRM) && !skb_prev)
1181 skb_set_dst_pending_confirm(skb, 1);
1184 * Put the packet on the pending queue.
1186 if (!skb->destructor) {
1187 skb->destructor = sock_wfree;
1189 wmem_alloc_delta += skb->truesize;
1191 __skb_queue_tail(queue, skb);
1198 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1199 skb_tailroom(skb) >= copy) {
1203 if (getfrag(from, skb_put(skb, copy),
1204 offset, copy, off, skb) < 0) {
1205 __skb_trim(skb, off);
1209 } else if (flags & MSG_SPLICE_PAGES) {
1210 struct msghdr *msg = from;
1212 err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
1217 wmem_alloc_delta += copy;
1219 int i = skb_shinfo(skb)->nr_frags;
1222 if (!sk_page_frag_refill(sk, pfrag))
1225 skb_zcopy_downgrade_managed(skb);
1226 if (!skb_can_coalesce(skb, i, pfrag->page,
1229 if (i == MAX_SKB_FRAGS)
1232 __skb_fill_page_desc(skb, i, pfrag->page,
1234 skb_shinfo(skb)->nr_frags = ++i;
1235 get_page(pfrag->page);
1237 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1239 page_address(pfrag->page) + pfrag->offset,
1240 offset, copy, skb->len, skb) < 0)
1243 pfrag->offset += copy;
1244 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1245 skb_len_add(skb, copy);
1246 wmem_alloc_delta += copy;
1248 err = skb_zerocopy_iter_dgram(skb, from, copy);
1256 if (wmem_alloc_delta)
1257 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1263 net_zcopy_put_abort(uarg, extra_uref);
1264 cork->length -= length;
1265 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1266 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1270 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1271 struct ipcm_cookie *ipc, struct rtable **rtp)
1273 struct ip_options_rcu *opt;
1281 * setup for corking.
1286 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1288 if (unlikely(!cork->opt))
1291 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1292 cork->flags |= IPCORK_OPT;
1293 cork->addr = ipc->addr;
1296 cork->fragsize = ip_sk_use_pmtu(sk) ?
1297 dst_mtu(&rt->dst) : READ_ONCE(rt->dst.dev->mtu);
1299 if (!inetdev_valid_mtu(cork->fragsize))
1300 return -ENETUNREACH;
1302 cork->gso_size = ipc->gso_size;
1304 cork->dst = &rt->dst;
1305 /* We stole this route, caller should not release it. */
1309 cork->ttl = ipc->ttl;
1310 cork->tos = ipc->tos;
1311 cork->mark = ipc->sockc.mark;
1312 cork->priority = ipc->priority;
1313 cork->transmit_time = ipc->sockc.transmit_time;
1315 sock_tx_timestamp(sk, ipc->sockc.tsflags, &cork->tx_flags);
1321 * ip_append_data() can make one large IP datagram from many pieces of
1322 * data. Each piece will be held on the socket until
1323 * ip_push_pending_frames() is called. Each piece can be a page or
1326 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1327 * this interface potentially.
1329 * LATER: length must be adjusted by pad at tail, when it is required.
1331 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1332 int getfrag(void *from, char *to, int offset, int len,
1333 int odd, struct sk_buff *skb),
1334 void *from, int length, int transhdrlen,
1335 struct ipcm_cookie *ipc, struct rtable **rtp,
1338 struct inet_sock *inet = inet_sk(sk);
1341 if (flags&MSG_PROBE)
1344 if (skb_queue_empty(&sk->sk_write_queue)) {
1345 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1352 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1353 sk_page_frag(sk), getfrag,
1354 from, length, transhdrlen, flags);
1357 static void ip_cork_release(struct inet_cork *cork)
1359 cork->flags &= ~IPCORK_OPT;
1362 dst_release(cork->dst);
1367 * Combined all pending IP fragments on the socket as one IP datagram
1368 * and push them out.
1370 struct sk_buff *__ip_make_skb(struct sock *sk,
1372 struct sk_buff_head *queue,
1373 struct inet_cork *cork)
1375 struct sk_buff *skb, *tmp_skb;
1376 struct sk_buff **tail_skb;
1377 struct inet_sock *inet = inet_sk(sk);
1378 struct net *net = sock_net(sk);
1379 struct ip_options *opt = NULL;
1380 struct rtable *rt = (struct rtable *)cork->dst;
1385 skb = __skb_dequeue(queue);
1388 tail_skb = &(skb_shinfo(skb)->frag_list);
1390 /* move skb->data to ip header from ext header */
1391 if (skb->data < skb_network_header(skb))
1392 __skb_pull(skb, skb_network_offset(skb));
1393 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1394 __skb_pull(tmp_skb, skb_network_header_len(skb));
1395 *tail_skb = tmp_skb;
1396 tail_skb = &(tmp_skb->next);
1397 skb->len += tmp_skb->len;
1398 skb->data_len += tmp_skb->len;
1399 skb->truesize += tmp_skb->truesize;
1400 tmp_skb->destructor = NULL;
1404 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1405 * to fragment the frame generated here. No matter, what transforms
1406 * how transforms change size of the packet, it will come out.
1408 skb->ignore_df = ip_sk_ignore_df(sk);
1410 /* DF bit is set when we want to see DF on outgoing frames.
1411 * If ignore_df is set too, we still allow to fragment this frame
1413 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1414 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1415 (skb->len <= dst_mtu(&rt->dst) &&
1416 ip_dont_fragment(sk, &rt->dst)))
1419 if (cork->flags & IPCORK_OPT)
1424 else if (rt->rt_type == RTN_MULTICAST)
1427 ttl = ip_select_ttl(inet, &rt->dst);
1432 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1435 iph->protocol = sk->sk_protocol;
1436 ip_copy_addrs(iph, fl4);
1437 ip_select_ident(net, skb, sk);
1440 iph->ihl += opt->optlen >> 2;
1441 ip_options_build(skb, opt, cork->addr, rt);
1444 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1445 skb->mark = cork->mark;
1446 skb->tstamp = cork->transmit_time;
1448 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1452 skb_dst_set(skb, &rt->dst);
1454 if (iph->protocol == IPPROTO_ICMP) {
1457 /* For such sockets, transhdrlen is zero when do ip_append_data(),
1458 * so icmphdr does not in skb linear region and can not get icmp_type
1459 * by icmp_hdr(skb)->type.
1461 if (sk->sk_type == SOCK_RAW && !inet_sk(sk)->hdrincl)
1462 icmp_type = fl4->fl4_icmp_type;
1464 icmp_type = icmp_hdr(skb)->type;
1465 icmp_out_count(net, icmp_type);
1468 ip_cork_release(cork);
1473 int ip_send_skb(struct net *net, struct sk_buff *skb)
1477 err = ip_local_out(net, skb->sk, skb);
1480 err = net_xmit_errno(err);
1482 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1488 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1490 struct sk_buff *skb;
1492 skb = ip_finish_skb(sk, fl4);
1496 /* Netfilter gets whole the not fragmented skb. */
1497 return ip_send_skb(sock_net(sk), skb);
1501 * Throw away all pending data on the socket.
1503 static void __ip_flush_pending_frames(struct sock *sk,
1504 struct sk_buff_head *queue,
1505 struct inet_cork *cork)
1507 struct sk_buff *skb;
1509 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1512 ip_cork_release(cork);
1515 void ip_flush_pending_frames(struct sock *sk)
1517 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1520 struct sk_buff *ip_make_skb(struct sock *sk,
1522 int getfrag(void *from, char *to, int offset,
1523 int len, int odd, struct sk_buff *skb),
1524 void *from, int length, int transhdrlen,
1525 struct ipcm_cookie *ipc, struct rtable **rtp,
1526 struct inet_cork *cork, unsigned int flags)
1528 struct sk_buff_head queue;
1531 if (flags & MSG_PROBE)
1534 __skb_queue_head_init(&queue);
1539 err = ip_setup_cork(sk, cork, ipc, rtp);
1541 return ERR_PTR(err);
1543 err = __ip_append_data(sk, fl4, &queue, cork,
1544 ¤t->task_frag, getfrag,
1545 from, length, transhdrlen, flags);
1547 __ip_flush_pending_frames(sk, &queue, cork);
1548 return ERR_PTR(err);
1551 return __ip_make_skb(sk, fl4, &queue, cork);
1555 * Fetch data from kernel space and fill in checksum if needed.
1557 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1558 int len, int odd, struct sk_buff *skb)
1562 csum = csum_partial_copy_nocheck(dptr+offset, to, len);
1563 skb->csum = csum_block_add(skb->csum, csum, odd);
1568 * Generic function to send a packet as reply to another packet.
1569 * Used to send some TCP resets/acks so far.
1571 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1572 const struct ip_options *sopt,
1573 __be32 daddr, __be32 saddr,
1574 const struct ip_reply_arg *arg,
1575 unsigned int len, u64 transmit_time, u32 txhash)
1577 struct ip_options_data replyopts;
1578 struct ipcm_cookie ipc;
1580 struct rtable *rt = skb_rtable(skb);
1581 struct net *net = sock_net(sk);
1582 struct sk_buff *nskb;
1586 if (__ip_options_echo(net, &replyopts.opt.opt, skb, sopt))
1591 ipc.sockc.transmit_time = transmit_time;
1593 if (replyopts.opt.opt.optlen) {
1594 ipc.opt = &replyopts.opt;
1596 if (replyopts.opt.opt.srr)
1597 daddr = replyopts.opt.opt.faddr;
1600 oif = arg->bound_dev_if;
1601 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1604 flowi4_init_output(&fl4, oif,
1605 IP4_REPLY_MARK(net, skb->mark) ?: sk->sk_mark,
1607 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1608 ip_reply_arg_flowi_flags(arg),
1610 tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1612 security_skb_classify_flow(skb, flowi4_to_flowi_common(&fl4));
1613 rt = ip_route_output_flow(net, &fl4, sk);
1617 inet_sk(sk)->tos = arg->tos & ~INET_ECN_MASK;
1619 sk->sk_protocol = ip_hdr(skb)->protocol;
1620 sk->sk_bound_dev_if = arg->bound_dev_if;
1621 sk->sk_sndbuf = READ_ONCE(sysctl_wmem_default);
1622 ipc.sockc.mark = fl4.flowi4_mark;
1623 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1624 len, 0, &ipc, &rt, MSG_DONTWAIT);
1625 if (unlikely(err)) {
1626 ip_flush_pending_frames(sk);
1630 nskb = skb_peek(&sk->sk_write_queue);
1632 if (arg->csumoffset >= 0)
1633 *((__sum16 *)skb_transport_header(nskb) +
1634 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1636 nskb->ip_summed = CHECKSUM_NONE;
1637 nskb->mono_delivery_time = !!transmit_time;
1639 skb_set_hash(nskb, txhash, PKT_HASH_TYPE_L4);
1640 ip_push_pending_frames(sk, &fl4);
1646 void __init ip_init(void)
1651 #if defined(CONFIG_IP_MULTICAST)