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 if (ip_dont_fragment(sk, &rt->dst)) {
166 iph->frag_off = htons(IP_DF);
170 __ip_select_ident(net, iph, 1);
173 if (opt && opt->opt.optlen) {
174 iph->ihl += opt->opt.optlen>>2;
175 ip_options_build(skb, &opt->opt, daddr, rt, 0);
178 skb->priority = sk->sk_priority;
180 skb->mark = sk->sk_mark;
183 return ip_local_out(net, skb->sk, skb);
185 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
187 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
189 struct dst_entry *dst = skb_dst(skb);
190 struct rtable *rt = (struct rtable *)dst;
191 struct net_device *dev = dst->dev;
192 unsigned int hh_len = LL_RESERVED_SPACE(dev);
193 struct neighbour *neigh;
194 bool is_v6gw = false;
196 if (rt->rt_type == RTN_MULTICAST) {
197 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
198 } else if (rt->rt_type == RTN_BROADCAST)
199 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
201 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
202 skb = skb_expand_head(skb, hh_len);
207 if (lwtunnel_xmit_redirect(dst->lwtstate)) {
208 int res = lwtunnel_xmit(skb);
210 if (res < 0 || res == LWTUNNEL_XMIT_DONE)
215 neigh = ip_neigh_for_gw(rt, skb, &is_v6gw);
216 if (!IS_ERR(neigh)) {
219 sock_confirm_neigh(skb, neigh);
220 /* if crossing protocols, can not use the cached header */
221 res = neigh_output(neigh, skb, is_v6gw);
222 rcu_read_unlock_bh();
225 rcu_read_unlock_bh();
227 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
233 static int ip_finish_output_gso(struct net *net, struct sock *sk,
234 struct sk_buff *skb, unsigned int mtu)
236 struct sk_buff *segs, *nskb;
237 netdev_features_t features;
240 /* common case: seglen is <= mtu
242 if (skb_gso_validate_network_len(skb, mtu))
243 return ip_finish_output2(net, sk, skb);
245 /* Slowpath - GSO segment length exceeds the egress MTU.
247 * This can happen in several cases:
248 * - Forwarding of a TCP GRO skb, when DF flag is not set.
249 * - Forwarding of an skb that arrived on a virtualization interface
250 * (virtio-net/vhost/tap) with TSO/GSO size set by other network
252 * - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
253 * interface with a smaller MTU.
254 * - Arriving GRO skb (or GSO skb in a virtualized environment) that is
255 * bridged to a NETIF_F_TSO tunnel stacked over an interface with an
258 features = netif_skb_features(skb);
259 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_GSO_CB_OFFSET);
260 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
261 if (IS_ERR_OR_NULL(segs)) {
268 skb_list_walk_safe(segs, segs, nskb) {
271 skb_mark_not_on_list(segs);
272 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
281 static int __ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
285 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
286 /* Policy lookup after SNAT yielded a new policy */
287 if (skb_dst(skb)->xfrm) {
288 IPCB(skb)->flags |= IPSKB_REROUTED;
289 return dst_output(net, sk, skb);
292 mtu = ip_skb_dst_mtu(sk, skb);
294 return ip_finish_output_gso(net, sk, skb, mtu);
296 if (skb->len > mtu || IPCB(skb)->frag_max_size)
297 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
299 return ip_finish_output2(net, sk, skb);
302 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
306 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
308 case NET_XMIT_SUCCESS:
309 return __ip_finish_output(net, sk, skb);
311 return __ip_finish_output(net, sk, skb) ? : ret;
318 static int ip_mc_finish_output(struct net *net, struct sock *sk,
321 struct rtable *new_rt;
325 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
330 case NET_XMIT_SUCCESS:
337 /* Reset rt_iif so that inet_iif() will return skb->skb_iif. Setting
338 * this to non-zero causes ipi_ifindex in in_pktinfo to be overwritten,
339 * see ipv4_pktinfo_prepare().
341 new_rt = rt_dst_clone(net->loopback_dev, skb_rtable(skb));
345 skb_dst_set(skb, &new_rt->dst);
348 err = dev_loopback_xmit(net, sk, skb);
349 return (do_cn && err) ? ret : err;
352 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
354 struct rtable *rt = skb_rtable(skb);
355 struct net_device *dev = rt->dst.dev;
358 * If the indicated interface is up and running, send the packet.
360 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
363 skb->protocol = htons(ETH_P_IP);
366 * Multicasts are looped back for other local users
369 if (rt->rt_flags&RTCF_MULTICAST) {
371 #ifdef CONFIG_IP_MROUTE
372 /* Small optimization: do not loopback not local frames,
373 which returned after forwarding; they will be dropped
374 by ip_mr_input in any case.
375 Note, that local frames are looped back to be delivered
378 This check is duplicated in ip_mr_input at the moment.
381 ((rt->rt_flags & RTCF_LOCAL) ||
382 !(IPCB(skb)->flags & IPSKB_FORWARDED))
385 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
387 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
388 net, sk, newskb, NULL, newskb->dev,
389 ip_mc_finish_output);
392 /* Multicasts with ttl 0 must not go beyond the host */
394 if (ip_hdr(skb)->ttl == 0) {
400 if (rt->rt_flags&RTCF_BROADCAST) {
401 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
403 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
404 net, sk, newskb, NULL, newskb->dev,
405 ip_mc_finish_output);
408 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
409 net, sk, skb, NULL, skb->dev,
411 !(IPCB(skb)->flags & IPSKB_REROUTED));
414 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
416 struct net_device *dev = skb_dst(skb)->dev, *indev = skb->dev;
418 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
421 skb->protocol = htons(ETH_P_IP);
423 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
424 net, sk, skb, indev, dev,
426 !(IPCB(skb)->flags & IPSKB_REROUTED));
428 EXPORT_SYMBOL(ip_output);
431 * copy saddr and daddr, possibly using 64bit load/stores
433 * iph->saddr = fl4->saddr;
434 * iph->daddr = fl4->daddr;
436 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
438 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
439 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
441 iph->saddr = fl4->saddr;
442 iph->daddr = fl4->daddr;
445 /* Note: skb->sk can be different from sk, in case of tunnels */
446 int __ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl,
449 struct inet_sock *inet = inet_sk(sk);
450 struct net *net = sock_net(sk);
451 struct ip_options_rcu *inet_opt;
457 /* Skip all of this if the packet is already routed,
458 * f.e. by something like SCTP.
461 inet_opt = rcu_dereference(inet->inet_opt);
463 rt = skb_rtable(skb);
467 /* Make sure we can route this packet. */
468 rt = (struct rtable *)__sk_dst_check(sk, 0);
472 /* Use correct destination address if we have options. */
473 daddr = inet->inet_daddr;
474 if (inet_opt && inet_opt->opt.srr)
475 daddr = inet_opt->opt.faddr;
477 /* If this fails, retransmit mechanism of transport layer will
478 * keep trying until route appears or the connection times
481 rt = ip_route_output_ports(net, fl4, sk,
482 daddr, inet->inet_saddr,
486 RT_CONN_FLAGS_TOS(sk, tos),
487 sk->sk_bound_dev_if);
490 sk_setup_caps(sk, &rt->dst);
492 skb_dst_set_noref(skb, &rt->dst);
495 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
498 /* OK, we know where to send it, allocate and build IP header. */
499 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
500 skb_reset_network_header(skb);
502 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (tos & 0xff));
503 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
504 iph->frag_off = htons(IP_DF);
507 iph->ttl = ip_select_ttl(inet, &rt->dst);
508 iph->protocol = sk->sk_protocol;
509 ip_copy_addrs(iph, fl4);
511 /* Transport layer set skb->h.foo itself. */
513 if (inet_opt && inet_opt->opt.optlen) {
514 iph->ihl += inet_opt->opt.optlen >> 2;
515 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
518 ip_select_ident_segs(net, skb, sk,
519 skb_shinfo(skb)->gso_segs ?: 1);
521 /* TODO : should we use skb->sk here instead of sk ? */
522 skb->priority = sk->sk_priority;
523 skb->mark = sk->sk_mark;
525 res = ip_local_out(net, sk, skb);
531 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
533 return -EHOSTUNREACH;
535 EXPORT_SYMBOL(__ip_queue_xmit);
537 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
539 return __ip_queue_xmit(sk, skb, fl, inet_sk(sk)->tos);
541 EXPORT_SYMBOL(ip_queue_xmit);
543 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
545 to->pkt_type = from->pkt_type;
546 to->priority = from->priority;
547 to->protocol = from->protocol;
548 to->skb_iif = from->skb_iif;
550 skb_dst_copy(to, from);
552 to->mark = from->mark;
554 skb_copy_hash(to, from);
556 #ifdef CONFIG_NET_SCHED
557 to->tc_index = from->tc_index;
560 skb_ext_copy(to, from);
561 #if IS_ENABLED(CONFIG_IP_VS)
562 to->ipvs_property = from->ipvs_property;
564 skb_copy_secmark(to, from);
567 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
569 int (*output)(struct net *, struct sock *, struct sk_buff *))
571 struct iphdr *iph = ip_hdr(skb);
573 if ((iph->frag_off & htons(IP_DF)) == 0)
574 return ip_do_fragment(net, sk, skb, output);
576 if (unlikely(!skb->ignore_df ||
577 (IPCB(skb)->frag_max_size &&
578 IPCB(skb)->frag_max_size > mtu))) {
579 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
580 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
586 return ip_do_fragment(net, sk, skb, output);
589 void ip_fraglist_init(struct sk_buff *skb, struct iphdr *iph,
590 unsigned int hlen, struct ip_fraglist_iter *iter)
592 unsigned int first_len = skb_pagelen(skb);
594 iter->frag = skb_shinfo(skb)->frag_list;
595 skb_frag_list_init(skb);
601 skb->data_len = first_len - skb_headlen(skb);
602 skb->len = first_len;
603 iph->tot_len = htons(first_len);
604 iph->frag_off = htons(IP_MF);
607 EXPORT_SYMBOL(ip_fraglist_init);
609 void ip_fraglist_prepare(struct sk_buff *skb, struct ip_fraglist_iter *iter)
611 unsigned int hlen = iter->hlen;
612 struct iphdr *iph = iter->iph;
613 struct sk_buff *frag;
616 frag->ip_summed = CHECKSUM_NONE;
617 skb_reset_transport_header(frag);
618 __skb_push(frag, hlen);
619 skb_reset_network_header(frag);
620 memcpy(skb_network_header(frag), iph, hlen);
621 iter->iph = ip_hdr(frag);
623 iph->tot_len = htons(frag->len);
624 ip_copy_metadata(frag, skb);
625 iter->offset += skb->len - hlen;
626 iph->frag_off = htons(iter->offset >> 3);
628 iph->frag_off |= htons(IP_MF);
629 /* Ready, complete checksum */
632 EXPORT_SYMBOL(ip_fraglist_prepare);
634 void ip_frag_init(struct sk_buff *skb, unsigned int hlen,
635 unsigned int ll_rs, unsigned int mtu, bool DF,
636 struct ip_frag_state *state)
638 struct iphdr *iph = ip_hdr(skb);
642 state->ll_rs = ll_rs;
645 state->left = skb->len - hlen; /* Space per frame */
646 state->ptr = hlen; /* Where to start from */
648 state->offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
649 state->not_last_frag = iph->frag_off & htons(IP_MF);
651 EXPORT_SYMBOL(ip_frag_init);
653 static void ip_frag_ipcb(struct sk_buff *from, struct sk_buff *to,
656 /* Copy the flags to each fragment. */
657 IPCB(to)->flags = IPCB(from)->flags;
659 /* ANK: dirty, but effective trick. Upgrade options only if
660 * the segment to be fragmented was THE FIRST (otherwise,
661 * options are already fixed) and make it ONCE
662 * on the initial skb, so that all the following fragments
663 * will inherit fixed options.
666 ip_options_fragment(from);
669 struct sk_buff *ip_frag_next(struct sk_buff *skb, struct ip_frag_state *state)
671 unsigned int len = state->left;
672 struct sk_buff *skb2;
675 /* IF: it doesn't fit, use 'mtu' - the data space left */
676 if (len > state->mtu)
678 /* IF: we are not sending up to and including the packet end
679 then align the next start on an eight byte boundary */
680 if (len < state->left) {
684 /* Allocate buffer */
685 skb2 = alloc_skb(len + state->hlen + state->ll_rs, GFP_ATOMIC);
687 return ERR_PTR(-ENOMEM);
690 * Set up data on packet
693 ip_copy_metadata(skb2, skb);
694 skb_reserve(skb2, state->ll_rs);
695 skb_put(skb2, len + state->hlen);
696 skb_reset_network_header(skb2);
697 skb2->transport_header = skb2->network_header + state->hlen;
700 * Charge the memory for the fragment to any owner
705 skb_set_owner_w(skb2, skb->sk);
708 * Copy the packet header into the new buffer.
711 skb_copy_from_linear_data(skb, skb_network_header(skb2), state->hlen);
714 * Copy a block of the IP datagram.
716 if (skb_copy_bits(skb, state->ptr, skb_transport_header(skb2), len))
721 * Fill in the new header fields.
724 iph->frag_off = htons((state->offset >> 3));
726 iph->frag_off |= htons(IP_DF);
729 * Added AC : If we are fragmenting a fragment that's not the
730 * last fragment then keep MF on each bit
732 if (state->left > 0 || state->not_last_frag)
733 iph->frag_off |= htons(IP_MF);
735 state->offset += len;
737 iph->tot_len = htons(len + state->hlen);
743 EXPORT_SYMBOL(ip_frag_next);
746 * This IP datagram is too large to be sent in one piece. Break it up into
747 * smaller pieces (each of size equal to IP header plus
748 * a block of the data of the original IP data part) that will yet fit in a
749 * single device frame, and queue such a frame for sending.
752 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
753 int (*output)(struct net *, struct sock *, struct sk_buff *))
756 struct sk_buff *skb2;
757 struct rtable *rt = skb_rtable(skb);
758 unsigned int mtu, hlen, ll_rs;
759 struct ip_fraglist_iter iter;
760 ktime_t tstamp = skb->tstamp;
761 struct ip_frag_state state;
764 /* for offloaded checksums cleanup checksum before fragmentation */
765 if (skb->ip_summed == CHECKSUM_PARTIAL &&
766 (err = skb_checksum_help(skb)))
770 * Point into the IP datagram header.
775 mtu = ip_skb_dst_mtu(sk, skb);
776 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
777 mtu = IPCB(skb)->frag_max_size;
780 * Setup starting values.
784 mtu = mtu - hlen; /* Size of data space */
785 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
786 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
788 /* When frag_list is given, use it. First, check its validity:
789 * some transformers could create wrong frag_list or break existing
790 * one, it is not prohibited. In this case fall back to copying.
792 * LATER: this step can be merged to real generation of fragments,
793 * we can switch to copy when see the first bad fragment.
795 if (skb_has_frag_list(skb)) {
796 struct sk_buff *frag, *frag2;
797 unsigned int first_len = skb_pagelen(skb);
799 if (first_len - hlen > mtu ||
800 ((first_len - hlen) & 7) ||
801 ip_is_fragment(iph) ||
803 skb_headroom(skb) < ll_rs)
806 skb_walk_frags(skb, frag) {
807 /* Correct geometry. */
808 if (frag->len > mtu ||
809 ((frag->len & 7) && frag->next) ||
810 skb_headroom(frag) < hlen + ll_rs)
811 goto slow_path_clean;
813 /* Partially cloned skb? */
814 if (skb_shared(frag))
815 goto slow_path_clean;
820 frag->destructor = sock_wfree;
822 skb->truesize -= frag->truesize;
825 /* Everything is OK. Generate! */
826 ip_fraglist_init(skb, iph, hlen, &iter);
829 ip_options_fragment(iter.frag);
832 /* Prepare header of the next frame,
833 * before previous one went down. */
835 IPCB(iter.frag)->flags = IPCB(skb)->flags;
836 ip_fraglist_prepare(skb, &iter);
839 skb->tstamp = tstamp;
840 err = output(net, sk, skb);
843 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
844 if (err || !iter.frag)
847 skb = ip_fraglist_next(&iter);
851 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
855 kfree_skb_list(iter.frag);
857 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
861 skb_walk_frags(skb, frag2) {
865 frag2->destructor = NULL;
866 skb->truesize += frag2->truesize;
872 * Fragment the datagram.
875 ip_frag_init(skb, hlen, ll_rs, mtu, IPCB(skb)->flags & IPSKB_FRAG_PMTU,
879 * Keep copying data until we run out.
882 while (state.left > 0) {
883 bool first_frag = (state.offset == 0);
885 skb2 = ip_frag_next(skb, &state);
890 ip_frag_ipcb(skb, skb2, first_frag);
893 * Put this fragment into the sending queue.
895 skb2->tstamp = tstamp;
896 err = output(net, sk, skb2);
900 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
903 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
908 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
911 EXPORT_SYMBOL(ip_do_fragment);
914 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
916 struct msghdr *msg = from;
918 if (skb->ip_summed == CHECKSUM_PARTIAL) {
919 if (!copy_from_iter_full(to, len, &msg->msg_iter))
923 if (!csum_and_copy_from_iter_full(to, len, &csum, &msg->msg_iter))
925 skb->csum = csum_block_add(skb->csum, csum, odd);
929 EXPORT_SYMBOL(ip_generic_getfrag);
932 csum_page(struct page *page, int offset, int copy)
937 csum = csum_partial(kaddr + offset, copy, 0);
942 static int __ip_append_data(struct sock *sk,
944 struct sk_buff_head *queue,
945 struct inet_cork *cork,
946 struct page_frag *pfrag,
947 int getfrag(void *from, char *to, int offset,
948 int len, int odd, struct sk_buff *skb),
949 void *from, int length, int transhdrlen,
952 struct inet_sock *inet = inet_sk(sk);
953 struct ubuf_info *uarg = NULL;
956 struct ip_options *opt = cork->opt;
963 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
964 int csummode = CHECKSUM_NONE;
965 struct rtable *rt = (struct rtable *)cork->dst;
966 unsigned int wmem_alloc_delta = 0;
967 bool paged, extra_uref = false;
970 skb = skb_peek_tail(queue);
972 exthdrlen = !skb ? rt->dst.header_len : 0;
973 mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
974 paged = !!cork->gso_size;
976 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
977 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
978 tskey = sk->sk_tskey++;
980 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
982 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
983 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
984 maxnonfragsize = ip_sk_ignore_df(sk) ? IP_MAX_MTU : mtu;
986 if (cork->length + length > maxnonfragsize - fragheaderlen) {
987 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
988 mtu - (opt ? opt->optlen : 0));
993 * transhdrlen > 0 means that this is the first fragment and we wish
994 * it won't be fragmented in the future.
997 length + fragheaderlen <= mtu &&
998 rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
999 (!(flags & MSG_MORE) || cork->gso_size) &&
1000 (!exthdrlen || (rt->dst.dev->features & NETIF_F_HW_ESP_TX_CSUM)))
1001 csummode = CHECKSUM_PARTIAL;
1003 if (flags & MSG_ZEROCOPY && length && sock_flag(sk, SOCK_ZEROCOPY)) {
1004 uarg = msg_zerocopy_realloc(sk, length, skb_zcopy(skb));
1007 extra_uref = !skb_zcopy(skb); /* only ref on new uarg */
1008 if (rt->dst.dev->features & NETIF_F_SG &&
1009 csummode == CHECKSUM_PARTIAL) {
1013 skb_zcopy_set(skb, uarg, &extra_uref);
1017 cork->length += length;
1019 /* So, what's going on in the loop below?
1021 * We use calculated fragment length to generate chained skb,
1022 * each of segments is IP fragment ready for sending to network after
1023 * adding appropriate IP header.
1029 while (length > 0) {
1030 /* Check if the remaining data fits into current packet. */
1031 copy = mtu - skb->len;
1033 copy = maxfraglen - skb->len;
1036 unsigned int datalen;
1037 unsigned int fraglen;
1038 unsigned int fraggap;
1039 unsigned int alloclen, alloc_extra;
1040 unsigned int pagedlen;
1041 struct sk_buff *skb_prev;
1045 fraggap = skb_prev->len - maxfraglen;
1050 * If remaining data exceeds the mtu,
1051 * we know we need more fragment(s).
1053 datalen = length + fraggap;
1054 if (datalen > mtu - fragheaderlen)
1055 datalen = maxfraglen - fragheaderlen;
1056 fraglen = datalen + fragheaderlen;
1059 alloc_extra = hh_len + 15;
1060 alloc_extra += exthdrlen;
1062 /* The last fragment gets additional space at tail.
1063 * Note, with MSG_MORE we overallocate on fragments,
1064 * because we have no idea what fragment will be
1067 if (datalen == length + fraggap)
1068 alloc_extra += rt->dst.trailer_len;
1070 if ((flags & MSG_MORE) &&
1071 !(rt->dst.dev->features&NETIF_F_SG))
1074 (fraglen + alloc_extra < SKB_MAX_ALLOC ||
1075 !(rt->dst.dev->features & NETIF_F_SG)))
1078 alloclen = min_t(int, fraglen, MAX_HEADER);
1079 pagedlen = fraglen - alloclen;
1082 alloclen += alloc_extra;
1085 skb = sock_alloc_send_skb(sk, alloclen,
1086 (flags & MSG_DONTWAIT), &err);
1089 if (refcount_read(&sk->sk_wmem_alloc) + wmem_alloc_delta <=
1091 skb = alloc_skb(alloclen,
1100 * Fill in the control structures
1102 skb->ip_summed = csummode;
1104 skb_reserve(skb, hh_len);
1107 * Find where to start putting bytes.
1109 data = skb_put(skb, fraglen + exthdrlen - pagedlen);
1110 skb_set_network_header(skb, exthdrlen);
1111 skb->transport_header = (skb->network_header +
1113 data += fragheaderlen + exthdrlen;
1116 skb->csum = skb_copy_and_csum_bits(
1117 skb_prev, maxfraglen,
1118 data + transhdrlen, fraggap);
1119 skb_prev->csum = csum_sub(skb_prev->csum,
1122 pskb_trim_unique(skb_prev, maxfraglen);
1125 copy = datalen - transhdrlen - fraggap - pagedlen;
1126 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1133 length -= copy + transhdrlen;
1136 csummode = CHECKSUM_NONE;
1138 /* only the initial fragment is time stamped */
1139 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1141 skb_shinfo(skb)->tskey = tskey;
1143 skb_zcopy_set(skb, uarg, &extra_uref);
1145 if ((flags & MSG_CONFIRM) && !skb_prev)
1146 skb_set_dst_pending_confirm(skb, 1);
1149 * Put the packet on the pending queue.
1151 if (!skb->destructor) {
1152 skb->destructor = sock_wfree;
1154 wmem_alloc_delta += skb->truesize;
1156 __skb_queue_tail(queue, skb);
1163 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1164 skb_tailroom(skb) >= copy) {
1168 if (getfrag(from, skb_put(skb, copy),
1169 offset, copy, off, skb) < 0) {
1170 __skb_trim(skb, off);
1174 } else if (!uarg || !uarg->zerocopy) {
1175 int i = skb_shinfo(skb)->nr_frags;
1178 if (!sk_page_frag_refill(sk, pfrag))
1181 if (!skb_can_coalesce(skb, i, pfrag->page,
1184 if (i == MAX_SKB_FRAGS)
1187 __skb_fill_page_desc(skb, i, pfrag->page,
1189 skb_shinfo(skb)->nr_frags = ++i;
1190 get_page(pfrag->page);
1192 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1194 page_address(pfrag->page) + pfrag->offset,
1195 offset, copy, skb->len, skb) < 0)
1198 pfrag->offset += copy;
1199 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1201 skb->data_len += copy;
1202 skb->truesize += copy;
1203 wmem_alloc_delta += copy;
1205 err = skb_zerocopy_iter_dgram(skb, from, copy);
1213 if (wmem_alloc_delta)
1214 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1220 net_zcopy_put_abort(uarg, extra_uref);
1221 cork->length -= length;
1222 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1223 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1227 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1228 struct ipcm_cookie *ipc, struct rtable **rtp)
1230 struct ip_options_rcu *opt;
1238 * setup for corking.
1243 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1245 if (unlikely(!cork->opt))
1248 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1249 cork->flags |= IPCORK_OPT;
1250 cork->addr = ipc->addr;
1253 cork->fragsize = ip_sk_use_pmtu(sk) ?
1254 dst_mtu(&rt->dst) : READ_ONCE(rt->dst.dev->mtu);
1256 if (!inetdev_valid_mtu(cork->fragsize))
1257 return -ENETUNREACH;
1259 cork->gso_size = ipc->gso_size;
1261 cork->dst = &rt->dst;
1262 /* We stole this route, caller should not release it. */
1266 cork->ttl = ipc->ttl;
1267 cork->tos = ipc->tos;
1268 cork->mark = ipc->sockc.mark;
1269 cork->priority = ipc->priority;
1270 cork->transmit_time = ipc->sockc.transmit_time;
1272 sock_tx_timestamp(sk, ipc->sockc.tsflags, &cork->tx_flags);
1278 * ip_append_data() and ip_append_page() can make one large IP datagram
1279 * from many pieces of data. Each pieces will be holded on the socket
1280 * until ip_push_pending_frames() is called. Each piece can be a page
1283 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1284 * this interface potentially.
1286 * LATER: length must be adjusted by pad at tail, when it is required.
1288 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1289 int getfrag(void *from, char *to, int offset, int len,
1290 int odd, struct sk_buff *skb),
1291 void *from, int length, int transhdrlen,
1292 struct ipcm_cookie *ipc, struct rtable **rtp,
1295 struct inet_sock *inet = inet_sk(sk);
1298 if (flags&MSG_PROBE)
1301 if (skb_queue_empty(&sk->sk_write_queue)) {
1302 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1309 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1310 sk_page_frag(sk), getfrag,
1311 from, length, transhdrlen, flags);
1314 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1315 int offset, size_t size, int flags)
1317 struct inet_sock *inet = inet_sk(sk);
1318 struct sk_buff *skb;
1320 struct ip_options *opt = NULL;
1321 struct inet_cork *cork;
1326 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1331 if (flags&MSG_PROBE)
1334 if (skb_queue_empty(&sk->sk_write_queue))
1337 cork = &inet->cork.base;
1338 rt = (struct rtable *)cork->dst;
1339 if (cork->flags & IPCORK_OPT)
1342 if (!(rt->dst.dev->features & NETIF_F_SG))
1345 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1346 mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
1348 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1349 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1350 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1352 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1353 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1354 mtu - (opt ? opt->optlen : 0));
1358 skb = skb_peek_tail(&sk->sk_write_queue);
1362 cork->length += size;
1365 /* Check if the remaining data fits into current packet. */
1366 len = mtu - skb->len;
1368 len = maxfraglen - skb->len;
1371 struct sk_buff *skb_prev;
1375 fraggap = skb_prev->len - maxfraglen;
1377 alloclen = fragheaderlen + hh_len + fraggap + 15;
1378 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1379 if (unlikely(!skb)) {
1385 * Fill in the control structures
1387 skb->ip_summed = CHECKSUM_NONE;
1389 skb_reserve(skb, hh_len);
1392 * Find where to start putting bytes.
1394 skb_put(skb, fragheaderlen + fraggap);
1395 skb_reset_network_header(skb);
1396 skb->transport_header = (skb->network_header +
1399 skb->csum = skb_copy_and_csum_bits(skb_prev,
1401 skb_transport_header(skb),
1403 skb_prev->csum = csum_sub(skb_prev->csum,
1405 pskb_trim_unique(skb_prev, maxfraglen);
1409 * Put the packet on the pending queue.
1411 __skb_queue_tail(&sk->sk_write_queue, skb);
1418 if (skb_append_pagefrags(skb, page, offset, len)) {
1423 if (skb->ip_summed == CHECKSUM_NONE) {
1425 csum = csum_page(page, offset, len);
1426 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1430 skb->data_len += len;
1431 skb->truesize += len;
1432 refcount_add(len, &sk->sk_wmem_alloc);
1439 cork->length -= size;
1440 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1444 static void ip_cork_release(struct inet_cork *cork)
1446 cork->flags &= ~IPCORK_OPT;
1449 dst_release(cork->dst);
1454 * Combined all pending IP fragments on the socket as one IP datagram
1455 * and push them out.
1457 struct sk_buff *__ip_make_skb(struct sock *sk,
1459 struct sk_buff_head *queue,
1460 struct inet_cork *cork)
1462 struct sk_buff *skb, *tmp_skb;
1463 struct sk_buff **tail_skb;
1464 struct inet_sock *inet = inet_sk(sk);
1465 struct net *net = sock_net(sk);
1466 struct ip_options *opt = NULL;
1467 struct rtable *rt = (struct rtable *)cork->dst;
1472 skb = __skb_dequeue(queue);
1475 tail_skb = &(skb_shinfo(skb)->frag_list);
1477 /* move skb->data to ip header from ext header */
1478 if (skb->data < skb_network_header(skb))
1479 __skb_pull(skb, skb_network_offset(skb));
1480 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1481 __skb_pull(tmp_skb, skb_network_header_len(skb));
1482 *tail_skb = tmp_skb;
1483 tail_skb = &(tmp_skb->next);
1484 skb->len += tmp_skb->len;
1485 skb->data_len += tmp_skb->len;
1486 skb->truesize += tmp_skb->truesize;
1487 tmp_skb->destructor = NULL;
1491 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1492 * to fragment the frame generated here. No matter, what transforms
1493 * how transforms change size of the packet, it will come out.
1495 skb->ignore_df = ip_sk_ignore_df(sk);
1497 /* DF bit is set when we want to see DF on outgoing frames.
1498 * If ignore_df is set too, we still allow to fragment this frame
1500 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1501 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1502 (skb->len <= dst_mtu(&rt->dst) &&
1503 ip_dont_fragment(sk, &rt->dst)))
1506 if (cork->flags & IPCORK_OPT)
1511 else if (rt->rt_type == RTN_MULTICAST)
1514 ttl = ip_select_ttl(inet, &rt->dst);
1519 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1522 iph->protocol = sk->sk_protocol;
1523 ip_copy_addrs(iph, fl4);
1524 ip_select_ident(net, skb, sk);
1527 iph->ihl += opt->optlen >> 2;
1528 ip_options_build(skb, opt, cork->addr, rt, 0);
1531 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1532 skb->mark = cork->mark;
1533 skb->tstamp = cork->transmit_time;
1535 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1539 skb_dst_set(skb, &rt->dst);
1541 if (iph->protocol == IPPROTO_ICMP)
1542 icmp_out_count(net, ((struct icmphdr *)
1543 skb_transport_header(skb))->type);
1545 ip_cork_release(cork);
1550 int ip_send_skb(struct net *net, struct sk_buff *skb)
1554 err = ip_local_out(net, skb->sk, skb);
1557 err = net_xmit_errno(err);
1559 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1565 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1567 struct sk_buff *skb;
1569 skb = ip_finish_skb(sk, fl4);
1573 /* Netfilter gets whole the not fragmented skb. */
1574 return ip_send_skb(sock_net(sk), skb);
1578 * Throw away all pending data on the socket.
1580 static void __ip_flush_pending_frames(struct sock *sk,
1581 struct sk_buff_head *queue,
1582 struct inet_cork *cork)
1584 struct sk_buff *skb;
1586 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1589 ip_cork_release(cork);
1592 void ip_flush_pending_frames(struct sock *sk)
1594 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1597 struct sk_buff *ip_make_skb(struct sock *sk,
1599 int getfrag(void *from, char *to, int offset,
1600 int len, int odd, struct sk_buff *skb),
1601 void *from, int length, int transhdrlen,
1602 struct ipcm_cookie *ipc, struct rtable **rtp,
1603 struct inet_cork *cork, unsigned int flags)
1605 struct sk_buff_head queue;
1608 if (flags & MSG_PROBE)
1611 __skb_queue_head_init(&queue);
1616 err = ip_setup_cork(sk, cork, ipc, rtp);
1618 return ERR_PTR(err);
1620 err = __ip_append_data(sk, fl4, &queue, cork,
1621 ¤t->task_frag, getfrag,
1622 from, length, transhdrlen, flags);
1624 __ip_flush_pending_frames(sk, &queue, cork);
1625 return ERR_PTR(err);
1628 return __ip_make_skb(sk, fl4, &queue, cork);
1632 * Fetch data from kernel space and fill in checksum if needed.
1634 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1635 int len, int odd, struct sk_buff *skb)
1639 csum = csum_partial_copy_nocheck(dptr+offset, to, len);
1640 skb->csum = csum_block_add(skb->csum, csum, odd);
1645 * Generic function to send a packet as reply to another packet.
1646 * Used to send some TCP resets/acks so far.
1648 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1649 const struct ip_options *sopt,
1650 __be32 daddr, __be32 saddr,
1651 const struct ip_reply_arg *arg,
1652 unsigned int len, u64 transmit_time)
1654 struct ip_options_data replyopts;
1655 struct ipcm_cookie ipc;
1657 struct rtable *rt = skb_rtable(skb);
1658 struct net *net = sock_net(sk);
1659 struct sk_buff *nskb;
1663 if (__ip_options_echo(net, &replyopts.opt.opt, skb, sopt))
1668 ipc.sockc.transmit_time = transmit_time;
1670 if (replyopts.opt.opt.optlen) {
1671 ipc.opt = &replyopts.opt;
1673 if (replyopts.opt.opt.srr)
1674 daddr = replyopts.opt.opt.faddr;
1677 oif = arg->bound_dev_if;
1678 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1681 flowi4_init_output(&fl4, oif,
1682 IP4_REPLY_MARK(net, skb->mark) ?: sk->sk_mark,
1684 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1685 ip_reply_arg_flowi_flags(arg),
1687 tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1689 security_skb_classify_flow(skb, flowi4_to_flowi_common(&fl4));
1690 rt = ip_route_output_key(net, &fl4);
1694 inet_sk(sk)->tos = arg->tos & ~INET_ECN_MASK;
1696 sk->sk_protocol = ip_hdr(skb)->protocol;
1697 sk->sk_bound_dev_if = arg->bound_dev_if;
1698 sk->sk_sndbuf = sysctl_wmem_default;
1699 ipc.sockc.mark = fl4.flowi4_mark;
1700 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1701 len, 0, &ipc, &rt, MSG_DONTWAIT);
1702 if (unlikely(err)) {
1703 ip_flush_pending_frames(sk);
1707 nskb = skb_peek(&sk->sk_write_queue);
1709 if (arg->csumoffset >= 0)
1710 *((__sum16 *)skb_transport_header(nskb) +
1711 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1713 nskb->ip_summed = CHECKSUM_NONE;
1714 ip_push_pending_frames(sk, &fl4);
1720 void __init ip_init(void)
1725 #if defined(CONFIG_IP_MULTICAST)