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>
76 #include <net/inetpeer.h>
77 #include <net/inet_ecn.h>
78 #include <net/lwtunnel.h>
79 #include <linux/bpf-cgroup.h>
80 #include <linux/igmp.h>
81 #include <linux/netfilter_ipv4.h>
82 #include <linux/netfilter_bridge.h>
83 #include <linux/netlink.h>
84 #include <linux/tcp.h>
87 ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
89 int (*output)(struct net *, struct sock *, struct sk_buff *));
91 /* Generate a checksum for an outgoing IP datagram. */
92 void ip_send_check(struct iphdr *iph)
95 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
97 EXPORT_SYMBOL(ip_send_check);
99 int __ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
101 struct iphdr *iph = ip_hdr(skb);
103 iph->tot_len = htons(skb->len);
106 /* if egress device is enslaved to an L3 master device pass the
107 * skb to its handler for processing
109 skb = l3mdev_ip_out(sk, skb);
113 skb->protocol = htons(ETH_P_IP);
115 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
116 net, sk, skb, NULL, skb_dst(skb)->dev,
120 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
124 err = __ip_local_out(net, sk, skb);
125 if (likely(err == 1))
126 err = dst_output(net, sk, skb);
130 EXPORT_SYMBOL_GPL(ip_local_out);
132 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
134 int ttl = inet->uc_ttl;
137 ttl = ip4_dst_hoplimit(dst);
142 * Add an ip header to a skbuff and send it out.
145 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
146 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt,
149 struct inet_sock *inet = inet_sk(sk);
150 struct rtable *rt = skb_rtable(skb);
151 struct net *net = sock_net(sk);
154 /* Build the IP header. */
155 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
156 skb_reset_network_header(skb);
161 iph->ttl = ip_select_ttl(inet, &rt->dst);
162 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
164 iph->protocol = sk->sk_protocol;
165 /* Do not bother generating IPID for small packets (eg SYNACK) */
166 if (skb->len <= IPV4_MIN_MTU || ip_dont_fragment(sk, &rt->dst)) {
167 iph->frag_off = htons(IP_DF);
171 /* TCP packets here are SYNACK with fat IPv4/TCP options.
172 * Avoid using the hashed IP ident generator.
174 if (sk->sk_protocol == IPPROTO_TCP)
175 iph->id = (__force __be16)prandom_u32();
177 __ip_select_ident(net, iph, 1);
180 if (opt && opt->opt.optlen) {
181 iph->ihl += opt->opt.optlen>>2;
182 ip_options_build(skb, &opt->opt, daddr, rt, 0);
185 skb->priority = sk->sk_priority;
187 skb->mark = sk->sk_mark;
190 return ip_local_out(net, skb->sk, skb);
192 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
194 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
196 struct dst_entry *dst = skb_dst(skb);
197 struct rtable *rt = (struct rtable *)dst;
198 struct net_device *dev = dst->dev;
199 unsigned int hh_len = LL_RESERVED_SPACE(dev);
200 struct neighbour *neigh;
201 bool is_v6gw = false;
203 if (rt->rt_type == RTN_MULTICAST) {
204 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
205 } else if (rt->rt_type == RTN_BROADCAST)
206 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
208 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
209 skb = skb_expand_head(skb, hh_len);
214 if (lwtunnel_xmit_redirect(dst->lwtstate)) {
215 int res = lwtunnel_xmit(skb);
217 if (res < 0 || res == LWTUNNEL_XMIT_DONE)
222 neigh = ip_neigh_for_gw(rt, skb, &is_v6gw);
223 if (!IS_ERR(neigh)) {
226 sock_confirm_neigh(skb, neigh);
227 /* if crossing protocols, can not use the cached header */
228 res = neigh_output(neigh, skb, is_v6gw);
229 rcu_read_unlock_bh();
232 rcu_read_unlock_bh();
234 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
240 static int ip_finish_output_gso(struct net *net, struct sock *sk,
241 struct sk_buff *skb, unsigned int mtu)
243 struct sk_buff *segs, *nskb;
244 netdev_features_t features;
247 /* common case: seglen is <= mtu
249 if (skb_gso_validate_network_len(skb, mtu))
250 return ip_finish_output2(net, sk, skb);
252 /* Slowpath - GSO segment length exceeds the egress MTU.
254 * This can happen in several cases:
255 * - Forwarding of a TCP GRO skb, when DF flag is not set.
256 * - Forwarding of an skb that arrived on a virtualization interface
257 * (virtio-net/vhost/tap) with TSO/GSO size set by other network
259 * - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
260 * interface with a smaller MTU.
261 * - Arriving GRO skb (or GSO skb in a virtualized environment) that is
262 * bridged to a NETIF_F_TSO tunnel stacked over an interface with an
265 features = netif_skb_features(skb);
266 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_GSO_CB_OFFSET);
267 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
268 if (IS_ERR_OR_NULL(segs)) {
275 skb_list_walk_safe(segs, segs, nskb) {
278 skb_mark_not_on_list(segs);
279 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
288 static int __ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
292 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
293 /* Policy lookup after SNAT yielded a new policy */
294 if (skb_dst(skb)->xfrm) {
295 IPCB(skb)->flags |= IPSKB_REROUTED;
296 return dst_output(net, sk, skb);
299 mtu = ip_skb_dst_mtu(sk, skb);
301 return ip_finish_output_gso(net, sk, skb, mtu);
303 if (skb->len > mtu || IPCB(skb)->frag_max_size)
304 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
306 return ip_finish_output2(net, sk, skb);
309 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
313 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
315 case NET_XMIT_SUCCESS:
316 return __ip_finish_output(net, sk, skb);
318 return __ip_finish_output(net, sk, skb) ? : ret;
325 static int ip_mc_finish_output(struct net *net, struct sock *sk,
328 struct rtable *new_rt;
332 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
337 case NET_XMIT_SUCCESS:
344 /* Reset rt_iif so that inet_iif() will return skb->skb_iif. Setting
345 * this to non-zero causes ipi_ifindex in in_pktinfo to be overwritten,
346 * see ipv4_pktinfo_prepare().
348 new_rt = rt_dst_clone(net->loopback_dev, skb_rtable(skb));
352 skb_dst_set(skb, &new_rt->dst);
355 err = dev_loopback_xmit(net, sk, skb);
356 return (do_cn && err) ? ret : err;
359 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
361 struct rtable *rt = skb_rtable(skb);
362 struct net_device *dev = rt->dst.dev;
365 * If the indicated interface is up and running, send the packet.
367 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
370 skb->protocol = htons(ETH_P_IP);
373 * Multicasts are looped back for other local users
376 if (rt->rt_flags&RTCF_MULTICAST) {
378 #ifdef CONFIG_IP_MROUTE
379 /* Small optimization: do not loopback not local frames,
380 which returned after forwarding; they will be dropped
381 by ip_mr_input in any case.
382 Note, that local frames are looped back to be delivered
385 This check is duplicated in ip_mr_input at the moment.
388 ((rt->rt_flags & RTCF_LOCAL) ||
389 !(IPCB(skb)->flags & IPSKB_FORWARDED))
392 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
394 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
395 net, sk, newskb, NULL, newskb->dev,
396 ip_mc_finish_output);
399 /* Multicasts with ttl 0 must not go beyond the host */
401 if (ip_hdr(skb)->ttl == 0) {
407 if (rt->rt_flags&RTCF_BROADCAST) {
408 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
410 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
411 net, sk, newskb, NULL, newskb->dev,
412 ip_mc_finish_output);
415 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
416 net, sk, skb, NULL, skb->dev,
418 !(IPCB(skb)->flags & IPSKB_REROUTED));
421 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
423 struct net_device *dev = skb_dst(skb)->dev, *indev = skb->dev;
425 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
428 skb->protocol = htons(ETH_P_IP);
430 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
431 net, sk, skb, indev, dev,
433 !(IPCB(skb)->flags & IPSKB_REROUTED));
435 EXPORT_SYMBOL(ip_output);
438 * copy saddr and daddr, possibly using 64bit load/stores
440 * iph->saddr = fl4->saddr;
441 * iph->daddr = fl4->daddr;
443 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
445 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
446 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
448 iph->saddr = fl4->saddr;
449 iph->daddr = fl4->daddr;
452 /* Note: skb->sk can be different from sk, in case of tunnels */
453 int __ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl,
456 struct inet_sock *inet = inet_sk(sk);
457 struct net *net = sock_net(sk);
458 struct ip_options_rcu *inet_opt;
464 /* Skip all of this if the packet is already routed,
465 * f.e. by something like SCTP.
468 inet_opt = rcu_dereference(inet->inet_opt);
470 rt = skb_rtable(skb);
474 /* Make sure we can route this packet. */
475 rt = (struct rtable *)__sk_dst_check(sk, 0);
479 /* Use correct destination address if we have options. */
480 daddr = inet->inet_daddr;
481 if (inet_opt && inet_opt->opt.srr)
482 daddr = inet_opt->opt.faddr;
484 /* If this fails, retransmit mechanism of transport layer will
485 * keep trying until route appears or the connection times
488 rt = ip_route_output_ports(net, fl4, sk,
489 daddr, inet->inet_saddr,
493 RT_CONN_FLAGS_TOS(sk, tos),
494 sk->sk_bound_dev_if);
497 sk_setup_caps(sk, &rt->dst);
499 skb_dst_set_noref(skb, &rt->dst);
502 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
505 /* OK, we know where to send it, allocate and build IP header. */
506 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
507 skb_reset_network_header(skb);
509 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (tos & 0xff));
510 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
511 iph->frag_off = htons(IP_DF);
514 iph->ttl = ip_select_ttl(inet, &rt->dst);
515 iph->protocol = sk->sk_protocol;
516 ip_copy_addrs(iph, fl4);
518 /* Transport layer set skb->h.foo itself. */
520 if (inet_opt && inet_opt->opt.optlen) {
521 iph->ihl += inet_opt->opt.optlen >> 2;
522 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
525 ip_select_ident_segs(net, skb, sk,
526 skb_shinfo(skb)->gso_segs ?: 1);
528 /* TODO : should we use skb->sk here instead of sk ? */
529 skb->priority = sk->sk_priority;
530 skb->mark = sk->sk_mark;
532 res = ip_local_out(net, sk, skb);
538 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
540 return -EHOSTUNREACH;
542 EXPORT_SYMBOL(__ip_queue_xmit);
544 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
546 return __ip_queue_xmit(sk, skb, fl, inet_sk(sk)->tos);
548 EXPORT_SYMBOL(ip_queue_xmit);
550 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
552 to->pkt_type = from->pkt_type;
553 to->priority = from->priority;
554 to->protocol = from->protocol;
555 to->skb_iif = from->skb_iif;
557 skb_dst_copy(to, from);
559 to->mark = from->mark;
561 skb_copy_hash(to, from);
563 #ifdef CONFIG_NET_SCHED
564 to->tc_index = from->tc_index;
567 skb_ext_copy(to, from);
568 #if IS_ENABLED(CONFIG_IP_VS)
569 to->ipvs_property = from->ipvs_property;
571 skb_copy_secmark(to, from);
574 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
576 int (*output)(struct net *, struct sock *, struct sk_buff *))
578 struct iphdr *iph = ip_hdr(skb);
580 if ((iph->frag_off & htons(IP_DF)) == 0)
581 return ip_do_fragment(net, sk, skb, output);
583 if (unlikely(!skb->ignore_df ||
584 (IPCB(skb)->frag_max_size &&
585 IPCB(skb)->frag_max_size > mtu))) {
586 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
587 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
593 return ip_do_fragment(net, sk, skb, output);
596 void ip_fraglist_init(struct sk_buff *skb, struct iphdr *iph,
597 unsigned int hlen, struct ip_fraglist_iter *iter)
599 unsigned int first_len = skb_pagelen(skb);
601 iter->frag = skb_shinfo(skb)->frag_list;
602 skb_frag_list_init(skb);
608 skb->data_len = first_len - skb_headlen(skb);
609 skb->len = first_len;
610 iph->tot_len = htons(first_len);
611 iph->frag_off = htons(IP_MF);
614 EXPORT_SYMBOL(ip_fraglist_init);
616 void ip_fraglist_prepare(struct sk_buff *skb, struct ip_fraglist_iter *iter)
618 unsigned int hlen = iter->hlen;
619 struct iphdr *iph = iter->iph;
620 struct sk_buff *frag;
623 frag->ip_summed = CHECKSUM_NONE;
624 skb_reset_transport_header(frag);
625 __skb_push(frag, hlen);
626 skb_reset_network_header(frag);
627 memcpy(skb_network_header(frag), iph, hlen);
628 iter->iph = ip_hdr(frag);
630 iph->tot_len = htons(frag->len);
631 ip_copy_metadata(frag, skb);
632 iter->offset += skb->len - hlen;
633 iph->frag_off = htons(iter->offset >> 3);
635 iph->frag_off |= htons(IP_MF);
636 /* Ready, complete checksum */
639 EXPORT_SYMBOL(ip_fraglist_prepare);
641 void ip_frag_init(struct sk_buff *skb, unsigned int hlen,
642 unsigned int ll_rs, unsigned int mtu, bool DF,
643 struct ip_frag_state *state)
645 struct iphdr *iph = ip_hdr(skb);
649 state->ll_rs = ll_rs;
652 state->left = skb->len - hlen; /* Space per frame */
653 state->ptr = hlen; /* Where to start from */
655 state->offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
656 state->not_last_frag = iph->frag_off & htons(IP_MF);
658 EXPORT_SYMBOL(ip_frag_init);
660 static void ip_frag_ipcb(struct sk_buff *from, struct sk_buff *to,
663 /* Copy the flags to each fragment. */
664 IPCB(to)->flags = IPCB(from)->flags;
666 /* ANK: dirty, but effective trick. Upgrade options only if
667 * the segment to be fragmented was THE FIRST (otherwise,
668 * options are already fixed) and make it ONCE
669 * on the initial skb, so that all the following fragments
670 * will inherit fixed options.
673 ip_options_fragment(from);
676 struct sk_buff *ip_frag_next(struct sk_buff *skb, struct ip_frag_state *state)
678 unsigned int len = state->left;
679 struct sk_buff *skb2;
682 /* IF: it doesn't fit, use 'mtu' - the data space left */
683 if (len > state->mtu)
685 /* IF: we are not sending up to and including the packet end
686 then align the next start on an eight byte boundary */
687 if (len < state->left) {
691 /* Allocate buffer */
692 skb2 = alloc_skb(len + state->hlen + state->ll_rs, GFP_ATOMIC);
694 return ERR_PTR(-ENOMEM);
697 * Set up data on packet
700 ip_copy_metadata(skb2, skb);
701 skb_reserve(skb2, state->ll_rs);
702 skb_put(skb2, len + state->hlen);
703 skb_reset_network_header(skb2);
704 skb2->transport_header = skb2->network_header + state->hlen;
707 * Charge the memory for the fragment to any owner
712 skb_set_owner_w(skb2, skb->sk);
715 * Copy the packet header into the new buffer.
718 skb_copy_from_linear_data(skb, skb_network_header(skb2), state->hlen);
721 * Copy a block of the IP datagram.
723 if (skb_copy_bits(skb, state->ptr, skb_transport_header(skb2), len))
728 * Fill in the new header fields.
731 iph->frag_off = htons((state->offset >> 3));
733 iph->frag_off |= htons(IP_DF);
736 * Added AC : If we are fragmenting a fragment that's not the
737 * last fragment then keep MF on each bit
739 if (state->left > 0 || state->not_last_frag)
740 iph->frag_off |= htons(IP_MF);
742 state->offset += len;
744 iph->tot_len = htons(len + state->hlen);
750 EXPORT_SYMBOL(ip_frag_next);
753 * This IP datagram is too large to be sent in one piece. Break it up into
754 * smaller pieces (each of size equal to IP header plus
755 * a block of the data of the original IP data part) that will yet fit in a
756 * single device frame, and queue such a frame for sending.
759 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
760 int (*output)(struct net *, struct sock *, struct sk_buff *))
763 struct sk_buff *skb2;
764 struct rtable *rt = skb_rtable(skb);
765 unsigned int mtu, hlen, ll_rs;
766 struct ip_fraglist_iter iter;
767 ktime_t tstamp = skb->tstamp;
768 struct ip_frag_state state;
771 /* for offloaded checksums cleanup checksum before fragmentation */
772 if (skb->ip_summed == CHECKSUM_PARTIAL &&
773 (err = skb_checksum_help(skb)))
777 * Point into the IP datagram header.
782 mtu = ip_skb_dst_mtu(sk, skb);
783 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
784 mtu = IPCB(skb)->frag_max_size;
787 * Setup starting values.
791 mtu = mtu - hlen; /* Size of data space */
792 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
793 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
795 /* When frag_list is given, use it. First, check its validity:
796 * some transformers could create wrong frag_list or break existing
797 * one, it is not prohibited. In this case fall back to copying.
799 * LATER: this step can be merged to real generation of fragments,
800 * we can switch to copy when see the first bad fragment.
802 if (skb_has_frag_list(skb)) {
803 struct sk_buff *frag, *frag2;
804 unsigned int first_len = skb_pagelen(skb);
806 if (first_len - hlen > mtu ||
807 ((first_len - hlen) & 7) ||
808 ip_is_fragment(iph) ||
810 skb_headroom(skb) < ll_rs)
813 skb_walk_frags(skb, frag) {
814 /* Correct geometry. */
815 if (frag->len > mtu ||
816 ((frag->len & 7) && frag->next) ||
817 skb_headroom(frag) < hlen + ll_rs)
818 goto slow_path_clean;
820 /* Partially cloned skb? */
821 if (skb_shared(frag))
822 goto slow_path_clean;
827 frag->destructor = sock_wfree;
829 skb->truesize -= frag->truesize;
832 /* Everything is OK. Generate! */
833 ip_fraglist_init(skb, iph, hlen, &iter);
836 /* Prepare header of the next frame,
837 * before previous one went down. */
839 bool first_frag = (iter.offset == 0);
841 IPCB(iter.frag)->flags = IPCB(skb)->flags;
842 ip_fraglist_prepare(skb, &iter);
843 if (first_frag && IPCB(skb)->opt.optlen) {
844 /* ipcb->opt is not populated for frags
845 * coming from __ip_make_skb(),
846 * ip_options_fragment() needs optlen
848 IPCB(iter.frag)->opt.optlen =
849 IPCB(skb)->opt.optlen;
850 ip_options_fragment(iter.frag);
851 ip_send_check(iter.iph);
855 skb->tstamp = tstamp;
856 err = output(net, sk, skb);
859 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
860 if (err || !iter.frag)
863 skb = ip_fraglist_next(&iter);
867 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
871 kfree_skb_list(iter.frag);
873 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
877 skb_walk_frags(skb, frag2) {
881 frag2->destructor = NULL;
882 skb->truesize += frag2->truesize;
888 * Fragment the datagram.
891 ip_frag_init(skb, hlen, ll_rs, mtu, IPCB(skb)->flags & IPSKB_FRAG_PMTU,
895 * Keep copying data until we run out.
898 while (state.left > 0) {
899 bool first_frag = (state.offset == 0);
901 skb2 = ip_frag_next(skb, &state);
906 ip_frag_ipcb(skb, skb2, first_frag);
909 * Put this fragment into the sending queue.
911 skb2->tstamp = tstamp;
912 err = output(net, sk, skb2);
916 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
919 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
924 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
927 EXPORT_SYMBOL(ip_do_fragment);
930 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
932 struct msghdr *msg = from;
934 if (skb->ip_summed == CHECKSUM_PARTIAL) {
935 if (!copy_from_iter_full(to, len, &msg->msg_iter))
939 if (!csum_and_copy_from_iter_full(to, len, &csum, &msg->msg_iter))
941 skb->csum = csum_block_add(skb->csum, csum, odd);
945 EXPORT_SYMBOL(ip_generic_getfrag);
948 csum_page(struct page *page, int offset, int copy)
953 csum = csum_partial(kaddr + offset, copy, 0);
958 static int __ip_append_data(struct sock *sk,
960 struct sk_buff_head *queue,
961 struct inet_cork *cork,
962 struct page_frag *pfrag,
963 int getfrag(void *from, char *to, int offset,
964 int len, int odd, struct sk_buff *skb),
965 void *from, int length, int transhdrlen,
968 struct inet_sock *inet = inet_sk(sk);
969 struct ubuf_info *uarg = NULL;
972 struct ip_options *opt = cork->opt;
979 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
980 int csummode = CHECKSUM_NONE;
981 struct rtable *rt = (struct rtable *)cork->dst;
982 unsigned int wmem_alloc_delta = 0;
983 bool paged, extra_uref = false;
986 skb = skb_peek_tail(queue);
988 exthdrlen = !skb ? rt->dst.header_len : 0;
989 mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
990 paged = !!cork->gso_size;
992 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
993 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
994 tskey = sk->sk_tskey++;
996 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
998 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
999 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1000 maxnonfragsize = ip_sk_ignore_df(sk) ? IP_MAX_MTU : mtu;
1002 if (cork->length + length > maxnonfragsize - fragheaderlen) {
1003 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1004 mtu - (opt ? opt->optlen : 0));
1009 * transhdrlen > 0 means that this is the first fragment and we wish
1010 * it won't be fragmented in the future.
1013 length + fragheaderlen <= mtu &&
1014 rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
1015 (!(flags & MSG_MORE) || cork->gso_size) &&
1016 (!exthdrlen || (rt->dst.dev->features & NETIF_F_HW_ESP_TX_CSUM)))
1017 csummode = CHECKSUM_PARTIAL;
1019 if (flags & MSG_ZEROCOPY && length && sock_flag(sk, SOCK_ZEROCOPY)) {
1020 uarg = msg_zerocopy_realloc(sk, length, skb_zcopy(skb));
1023 extra_uref = !skb_zcopy(skb); /* only ref on new uarg */
1024 if (rt->dst.dev->features & NETIF_F_SG &&
1025 csummode == CHECKSUM_PARTIAL) {
1029 skb_zcopy_set(skb, uarg, &extra_uref);
1033 cork->length += length;
1035 /* So, what's going on in the loop below?
1037 * We use calculated fragment length to generate chained skb,
1038 * each of segments is IP fragment ready for sending to network after
1039 * adding appropriate IP header.
1045 while (length > 0) {
1046 /* Check if the remaining data fits into current packet. */
1047 copy = mtu - skb->len;
1049 copy = maxfraglen - skb->len;
1052 unsigned int datalen;
1053 unsigned int fraglen;
1054 unsigned int fraggap;
1055 unsigned int alloclen, alloc_extra;
1056 unsigned int pagedlen;
1057 struct sk_buff *skb_prev;
1061 fraggap = skb_prev->len - maxfraglen;
1066 * If remaining data exceeds the mtu,
1067 * we know we need more fragment(s).
1069 datalen = length + fraggap;
1070 if (datalen > mtu - fragheaderlen)
1071 datalen = maxfraglen - fragheaderlen;
1072 fraglen = datalen + fragheaderlen;
1075 alloc_extra = hh_len + 15;
1076 alloc_extra += exthdrlen;
1078 /* The last fragment gets additional space at tail.
1079 * Note, with MSG_MORE we overallocate on fragments,
1080 * because we have no idea what fragment will be
1083 if (datalen == length + fraggap)
1084 alloc_extra += rt->dst.trailer_len;
1086 if ((flags & MSG_MORE) &&
1087 !(rt->dst.dev->features&NETIF_F_SG))
1090 (fraglen + alloc_extra < SKB_MAX_ALLOC ||
1091 !(rt->dst.dev->features & NETIF_F_SG)))
1094 alloclen = min_t(int, fraglen, MAX_HEADER);
1095 pagedlen = fraglen - alloclen;
1098 alloclen += alloc_extra;
1101 skb = sock_alloc_send_skb(sk, alloclen,
1102 (flags & MSG_DONTWAIT), &err);
1105 if (refcount_read(&sk->sk_wmem_alloc) + wmem_alloc_delta <=
1107 skb = alloc_skb(alloclen,
1116 * Fill in the control structures
1118 skb->ip_summed = csummode;
1120 skb_reserve(skb, hh_len);
1123 * Find where to start putting bytes.
1125 data = skb_put(skb, fraglen + exthdrlen - pagedlen);
1126 skb_set_network_header(skb, exthdrlen);
1127 skb->transport_header = (skb->network_header +
1129 data += fragheaderlen + exthdrlen;
1132 skb->csum = skb_copy_and_csum_bits(
1133 skb_prev, maxfraglen,
1134 data + transhdrlen, fraggap);
1135 skb_prev->csum = csum_sub(skb_prev->csum,
1138 pskb_trim_unique(skb_prev, maxfraglen);
1141 copy = datalen - transhdrlen - fraggap - pagedlen;
1142 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1149 length -= copy + transhdrlen;
1152 csummode = CHECKSUM_NONE;
1154 /* only the initial fragment is time stamped */
1155 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1157 skb_shinfo(skb)->tskey = tskey;
1159 skb_zcopy_set(skb, uarg, &extra_uref);
1161 if ((flags & MSG_CONFIRM) && !skb_prev)
1162 skb_set_dst_pending_confirm(skb, 1);
1165 * Put the packet on the pending queue.
1167 if (!skb->destructor) {
1168 skb->destructor = sock_wfree;
1170 wmem_alloc_delta += skb->truesize;
1172 __skb_queue_tail(queue, skb);
1179 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1180 skb_tailroom(skb) >= copy) {
1184 if (getfrag(from, skb_put(skb, copy),
1185 offset, copy, off, skb) < 0) {
1186 __skb_trim(skb, off);
1190 } else if (!uarg || !uarg->zerocopy) {
1191 int i = skb_shinfo(skb)->nr_frags;
1194 if (!sk_page_frag_refill(sk, pfrag))
1197 if (!skb_can_coalesce(skb, i, pfrag->page,
1200 if (i == MAX_SKB_FRAGS)
1203 __skb_fill_page_desc(skb, i, pfrag->page,
1205 skb_shinfo(skb)->nr_frags = ++i;
1206 get_page(pfrag->page);
1208 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1210 page_address(pfrag->page) + pfrag->offset,
1211 offset, copy, skb->len, skb) < 0)
1214 pfrag->offset += copy;
1215 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1217 skb->data_len += copy;
1218 skb->truesize += copy;
1219 wmem_alloc_delta += copy;
1221 err = skb_zerocopy_iter_dgram(skb, from, copy);
1229 if (wmem_alloc_delta)
1230 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1236 net_zcopy_put_abort(uarg, extra_uref);
1237 cork->length -= length;
1238 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1239 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1243 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1244 struct ipcm_cookie *ipc, struct rtable **rtp)
1246 struct ip_options_rcu *opt;
1254 * setup for corking.
1259 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1261 if (unlikely(!cork->opt))
1264 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1265 cork->flags |= IPCORK_OPT;
1266 cork->addr = ipc->addr;
1269 cork->fragsize = ip_sk_use_pmtu(sk) ?
1270 dst_mtu(&rt->dst) : READ_ONCE(rt->dst.dev->mtu);
1272 if (!inetdev_valid_mtu(cork->fragsize))
1273 return -ENETUNREACH;
1275 cork->gso_size = ipc->gso_size;
1277 cork->dst = &rt->dst;
1278 /* We stole this route, caller should not release it. */
1282 cork->ttl = ipc->ttl;
1283 cork->tos = ipc->tos;
1284 cork->mark = ipc->sockc.mark;
1285 cork->priority = ipc->priority;
1286 cork->transmit_time = ipc->sockc.transmit_time;
1288 sock_tx_timestamp(sk, ipc->sockc.tsflags, &cork->tx_flags);
1294 * ip_append_data() and ip_append_page() can make one large IP datagram
1295 * from many pieces of data. Each pieces will be holded on the socket
1296 * until ip_push_pending_frames() is called. Each piece can be a page
1299 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1300 * this interface potentially.
1302 * LATER: length must be adjusted by pad at tail, when it is required.
1304 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1305 int getfrag(void *from, char *to, int offset, int len,
1306 int odd, struct sk_buff *skb),
1307 void *from, int length, int transhdrlen,
1308 struct ipcm_cookie *ipc, struct rtable **rtp,
1311 struct inet_sock *inet = inet_sk(sk);
1314 if (flags&MSG_PROBE)
1317 if (skb_queue_empty(&sk->sk_write_queue)) {
1318 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1325 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1326 sk_page_frag(sk), getfrag,
1327 from, length, transhdrlen, flags);
1330 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1331 int offset, size_t size, int flags)
1333 struct inet_sock *inet = inet_sk(sk);
1334 struct sk_buff *skb;
1336 struct ip_options *opt = NULL;
1337 struct inet_cork *cork;
1342 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1347 if (flags&MSG_PROBE)
1350 if (skb_queue_empty(&sk->sk_write_queue))
1353 cork = &inet->cork.base;
1354 rt = (struct rtable *)cork->dst;
1355 if (cork->flags & IPCORK_OPT)
1358 if (!(rt->dst.dev->features & NETIF_F_SG))
1361 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1362 mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
1364 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1365 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1366 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1368 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1369 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1370 mtu - (opt ? opt->optlen : 0));
1374 skb = skb_peek_tail(&sk->sk_write_queue);
1378 cork->length += size;
1381 /* Check if the remaining data fits into current packet. */
1382 len = mtu - skb->len;
1384 len = maxfraglen - skb->len;
1387 struct sk_buff *skb_prev;
1391 fraggap = skb_prev->len - maxfraglen;
1393 alloclen = fragheaderlen + hh_len + fraggap + 15;
1394 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1395 if (unlikely(!skb)) {
1401 * Fill in the control structures
1403 skb->ip_summed = CHECKSUM_NONE;
1405 skb_reserve(skb, hh_len);
1408 * Find where to start putting bytes.
1410 skb_put(skb, fragheaderlen + fraggap);
1411 skb_reset_network_header(skb);
1412 skb->transport_header = (skb->network_header +
1415 skb->csum = skb_copy_and_csum_bits(skb_prev,
1417 skb_transport_header(skb),
1419 skb_prev->csum = csum_sub(skb_prev->csum,
1421 pskb_trim_unique(skb_prev, maxfraglen);
1425 * Put the packet on the pending queue.
1427 __skb_queue_tail(&sk->sk_write_queue, skb);
1434 if (skb_append_pagefrags(skb, page, offset, len)) {
1439 if (skb->ip_summed == CHECKSUM_NONE) {
1441 csum = csum_page(page, offset, len);
1442 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1446 skb->data_len += len;
1447 skb->truesize += len;
1448 refcount_add(len, &sk->sk_wmem_alloc);
1455 cork->length -= size;
1456 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1460 static void ip_cork_release(struct inet_cork *cork)
1462 cork->flags &= ~IPCORK_OPT;
1465 dst_release(cork->dst);
1470 * Combined all pending IP fragments on the socket as one IP datagram
1471 * and push them out.
1473 struct sk_buff *__ip_make_skb(struct sock *sk,
1475 struct sk_buff_head *queue,
1476 struct inet_cork *cork)
1478 struct sk_buff *skb, *tmp_skb;
1479 struct sk_buff **tail_skb;
1480 struct inet_sock *inet = inet_sk(sk);
1481 struct net *net = sock_net(sk);
1482 struct ip_options *opt = NULL;
1483 struct rtable *rt = (struct rtable *)cork->dst;
1488 skb = __skb_dequeue(queue);
1491 tail_skb = &(skb_shinfo(skb)->frag_list);
1493 /* move skb->data to ip header from ext header */
1494 if (skb->data < skb_network_header(skb))
1495 __skb_pull(skb, skb_network_offset(skb));
1496 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1497 __skb_pull(tmp_skb, skb_network_header_len(skb));
1498 *tail_skb = tmp_skb;
1499 tail_skb = &(tmp_skb->next);
1500 skb->len += tmp_skb->len;
1501 skb->data_len += tmp_skb->len;
1502 skb->truesize += tmp_skb->truesize;
1503 tmp_skb->destructor = NULL;
1507 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1508 * to fragment the frame generated here. No matter, what transforms
1509 * how transforms change size of the packet, it will come out.
1511 skb->ignore_df = ip_sk_ignore_df(sk);
1513 /* DF bit is set when we want to see DF on outgoing frames.
1514 * If ignore_df is set too, we still allow to fragment this frame
1516 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1517 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1518 (skb->len <= dst_mtu(&rt->dst) &&
1519 ip_dont_fragment(sk, &rt->dst)))
1522 if (cork->flags & IPCORK_OPT)
1527 else if (rt->rt_type == RTN_MULTICAST)
1530 ttl = ip_select_ttl(inet, &rt->dst);
1535 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1538 iph->protocol = sk->sk_protocol;
1539 ip_copy_addrs(iph, fl4);
1540 ip_select_ident(net, skb, sk);
1543 iph->ihl += opt->optlen >> 2;
1544 ip_options_build(skb, opt, cork->addr, rt, 0);
1547 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1548 skb->mark = cork->mark;
1549 skb->tstamp = cork->transmit_time;
1551 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1555 skb_dst_set(skb, &rt->dst);
1557 if (iph->protocol == IPPROTO_ICMP)
1558 icmp_out_count(net, ((struct icmphdr *)
1559 skb_transport_header(skb))->type);
1561 ip_cork_release(cork);
1566 int ip_send_skb(struct net *net, struct sk_buff *skb)
1570 err = ip_local_out(net, skb->sk, skb);
1573 err = net_xmit_errno(err);
1575 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1581 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1583 struct sk_buff *skb;
1585 skb = ip_finish_skb(sk, fl4);
1589 /* Netfilter gets whole the not fragmented skb. */
1590 return ip_send_skb(sock_net(sk), skb);
1594 * Throw away all pending data on the socket.
1596 static void __ip_flush_pending_frames(struct sock *sk,
1597 struct sk_buff_head *queue,
1598 struct inet_cork *cork)
1600 struct sk_buff *skb;
1602 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1605 ip_cork_release(cork);
1608 void ip_flush_pending_frames(struct sock *sk)
1610 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1613 struct sk_buff *ip_make_skb(struct sock *sk,
1615 int getfrag(void *from, char *to, int offset,
1616 int len, int odd, struct sk_buff *skb),
1617 void *from, int length, int transhdrlen,
1618 struct ipcm_cookie *ipc, struct rtable **rtp,
1619 struct inet_cork *cork, unsigned int flags)
1621 struct sk_buff_head queue;
1624 if (flags & MSG_PROBE)
1627 __skb_queue_head_init(&queue);
1632 err = ip_setup_cork(sk, cork, ipc, rtp);
1634 return ERR_PTR(err);
1636 err = __ip_append_data(sk, fl4, &queue, cork,
1637 ¤t->task_frag, getfrag,
1638 from, length, transhdrlen, flags);
1640 __ip_flush_pending_frames(sk, &queue, cork);
1641 return ERR_PTR(err);
1644 return __ip_make_skb(sk, fl4, &queue, cork);
1648 * Fetch data from kernel space and fill in checksum if needed.
1650 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1651 int len, int odd, struct sk_buff *skb)
1655 csum = csum_partial_copy_nocheck(dptr+offset, to, len);
1656 skb->csum = csum_block_add(skb->csum, csum, odd);
1661 * Generic function to send a packet as reply to another packet.
1662 * Used to send some TCP resets/acks so far.
1664 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1665 const struct ip_options *sopt,
1666 __be32 daddr, __be32 saddr,
1667 const struct ip_reply_arg *arg,
1668 unsigned int len, u64 transmit_time)
1670 struct ip_options_data replyopts;
1671 struct ipcm_cookie ipc;
1673 struct rtable *rt = skb_rtable(skb);
1674 struct net *net = sock_net(sk);
1675 struct sk_buff *nskb;
1679 if (__ip_options_echo(net, &replyopts.opt.opt, skb, sopt))
1684 ipc.sockc.transmit_time = transmit_time;
1686 if (replyopts.opt.opt.optlen) {
1687 ipc.opt = &replyopts.opt;
1689 if (replyopts.opt.opt.srr)
1690 daddr = replyopts.opt.opt.faddr;
1693 oif = arg->bound_dev_if;
1694 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1697 flowi4_init_output(&fl4, oif,
1698 IP4_REPLY_MARK(net, skb->mark) ?: sk->sk_mark,
1700 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1701 ip_reply_arg_flowi_flags(arg),
1703 tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1705 security_skb_classify_flow(skb, flowi4_to_flowi_common(&fl4));
1706 rt = ip_route_output_key(net, &fl4);
1710 inet_sk(sk)->tos = arg->tos & ~INET_ECN_MASK;
1712 sk->sk_protocol = ip_hdr(skb)->protocol;
1713 sk->sk_bound_dev_if = arg->bound_dev_if;
1714 sk->sk_sndbuf = sysctl_wmem_default;
1715 ipc.sockc.mark = fl4.flowi4_mark;
1716 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1717 len, 0, &ipc, &rt, MSG_DONTWAIT);
1718 if (unlikely(err)) {
1719 ip_flush_pending_frames(sk);
1723 nskb = skb_peek(&sk->sk_write_queue);
1725 if (arg->csumoffset >= 0)
1726 *((__sum16 *)skb_transport_header(nskb) +
1727 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1729 nskb->ip_summed = CHECKSUM_NONE;
1730 ip_push_pending_frames(sk, &fl4);
1736 void __init ip_init(void)
1741 #if defined(CONFIG_IP_MULTICAST)