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 readibility.
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/lwtunnel.h>
78 #include <linux/bpf-cgroup.h>
79 #include <linux/igmp.h>
80 #include <linux/netfilter_ipv4.h>
81 #include <linux/netfilter_bridge.h>
82 #include <linux/netlink.h>
83 #include <linux/tcp.h>
86 ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
88 int (*output)(struct net *, struct sock *, struct sk_buff *));
90 /* Generate a checksum for an outgoing IP datagram. */
91 void ip_send_check(struct iphdr *iph)
94 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
96 EXPORT_SYMBOL(ip_send_check);
98 int __ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
100 struct iphdr *iph = ip_hdr(skb);
102 iph->tot_len = htons(skb->len);
105 /* if egress device is enslaved to an L3 master device pass the
106 * skb to its handler for processing
108 skb = l3mdev_ip_out(sk, skb);
112 skb->protocol = htons(ETH_P_IP);
114 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
115 net, sk, skb, NULL, skb_dst(skb)->dev,
119 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
123 err = __ip_local_out(net, sk, skb);
124 if (likely(err == 1))
125 err = dst_output(net, sk, skb);
129 EXPORT_SYMBOL_GPL(ip_local_out);
131 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
133 int ttl = inet->uc_ttl;
136 ttl = ip4_dst_hoplimit(dst);
141 * Add an ip header to a skbuff and send it out.
144 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
145 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
147 struct inet_sock *inet = inet_sk(sk);
148 struct rtable *rt = skb_rtable(skb);
149 struct net *net = sock_net(sk);
152 /* Build the IP header. */
153 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
154 skb_reset_network_header(skb);
158 iph->tos = inet->tos;
159 iph->ttl = ip_select_ttl(inet, &rt->dst);
160 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
162 iph->protocol = sk->sk_protocol;
163 if (ip_dont_fragment(sk, &rt->dst)) {
164 iph->frag_off = htons(IP_DF);
168 __ip_select_ident(net, iph, 1);
171 if (opt && opt->opt.optlen) {
172 iph->ihl += opt->opt.optlen>>2;
173 ip_options_build(skb, &opt->opt, daddr, rt, 0);
176 skb->priority = sk->sk_priority;
178 skb->mark = sk->sk_mark;
181 return ip_local_out(net, skb->sk, skb);
183 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
185 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
187 struct dst_entry *dst = skb_dst(skb);
188 struct rtable *rt = (struct rtable *)dst;
189 struct net_device *dev = dst->dev;
190 unsigned int hh_len = LL_RESERVED_SPACE(dev);
191 struct neighbour *neigh;
192 bool is_v6gw = false;
194 if (rt->rt_type == RTN_MULTICAST) {
195 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
196 } else if (rt->rt_type == RTN_BROADCAST)
197 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
199 /* Be paranoid, rather than too clever. */
200 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
201 struct sk_buff *skb2;
203 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
209 skb_set_owner_w(skb2, skb->sk);
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 netdev_features_t features;
244 struct sk_buff *segs;
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_SGO_CB_OFFSET);
267 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
268 if (IS_ERR_OR_NULL(segs)) {
276 struct sk_buff *nskb = segs->next;
279 skb_mark_not_on_list(segs);
280 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)->flags & IPSKB_FRAG_PMTU))
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;
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:
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;
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, NULL, dev,
435 !(IPCB(skb)->flags & IPSKB_REROUTED));
439 * copy saddr and daddr, possibly using 64bit load/stores
441 * iph->saddr = fl4->saddr;
442 * iph->daddr = fl4->daddr;
444 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
446 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
447 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
448 memcpy(&iph->saddr, &fl4->saddr,
449 sizeof(fl4->saddr) + sizeof(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 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
546 to->pkt_type = from->pkt_type;
547 to->priority = from->priority;
548 to->protocol = from->protocol;
549 to->skb_iif = from->skb_iif;
551 skb_dst_copy(to, from);
553 to->mark = from->mark;
555 skb_copy_hash(to, from);
557 #ifdef CONFIG_NET_SCHED
558 to->tc_index = from->tc_index;
561 skb_ext_copy(to, from);
562 #if IS_ENABLED(CONFIG_IP_VS)
563 to->ipvs_property = from->ipvs_property;
565 skb_copy_secmark(to, from);
568 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
570 int (*output)(struct net *, struct sock *, struct sk_buff *))
572 struct iphdr *iph = ip_hdr(skb);
574 if ((iph->frag_off & htons(IP_DF)) == 0)
575 return ip_do_fragment(net, sk, skb, output);
577 if (unlikely(!skb->ignore_df ||
578 (IPCB(skb)->frag_max_size &&
579 IPCB(skb)->frag_max_size > mtu))) {
580 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
581 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
587 return ip_do_fragment(net, sk, skb, output);
590 void ip_fraglist_init(struct sk_buff *skb, struct iphdr *iph,
591 unsigned int hlen, struct ip_fraglist_iter *iter)
593 unsigned int first_len = skb_pagelen(skb);
595 iter->frag = skb_shinfo(skb)->frag_list;
596 skb_frag_list_init(skb);
602 skb->data_len = first_len - skb_headlen(skb);
603 skb->len = first_len;
604 iph->tot_len = htons(first_len);
605 iph->frag_off = htons(IP_MF);
608 EXPORT_SYMBOL(ip_fraglist_init);
610 static void ip_fraglist_ipcb_prepare(struct sk_buff *skb,
611 struct ip_fraglist_iter *iter)
613 struct sk_buff *to = iter->frag;
615 /* Copy the flags to each fragment. */
616 IPCB(to)->flags = IPCB(skb)->flags;
618 if (iter->offset == 0)
619 ip_options_fragment(to);
622 void ip_fraglist_prepare(struct sk_buff *skb, struct ip_fraglist_iter *iter)
624 unsigned int hlen = iter->hlen;
625 struct iphdr *iph = iter->iph;
626 struct sk_buff *frag;
629 frag->ip_summed = CHECKSUM_NONE;
630 skb_reset_transport_header(frag);
631 __skb_push(frag, hlen);
632 skb_reset_network_header(frag);
633 memcpy(skb_network_header(frag), iph, hlen);
634 iter->iph = ip_hdr(frag);
636 iph->tot_len = htons(frag->len);
637 ip_copy_metadata(frag, skb);
638 iter->offset += skb->len - hlen;
639 iph->frag_off = htons(iter->offset >> 3);
641 iph->frag_off |= htons(IP_MF);
642 /* Ready, complete checksum */
645 EXPORT_SYMBOL(ip_fraglist_prepare);
647 void ip_frag_init(struct sk_buff *skb, unsigned int hlen,
648 unsigned int ll_rs, unsigned int mtu,
649 struct ip_frag_state *state)
651 struct iphdr *iph = ip_hdr(skb);
654 state->ll_rs = ll_rs;
657 state->left = skb->len - hlen; /* Space per frame */
658 state->ptr = hlen; /* Where to start from */
660 state->offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
661 state->not_last_frag = iph->frag_off & htons(IP_MF);
663 EXPORT_SYMBOL(ip_frag_init);
665 static void ip_frag_ipcb(struct sk_buff *from, struct sk_buff *to,
666 bool first_frag, struct ip_frag_state *state)
668 /* Copy the flags to each fragment. */
669 IPCB(to)->flags = IPCB(from)->flags;
671 if (IPCB(from)->flags & IPSKB_FRAG_PMTU)
672 state->iph->frag_off |= htons(IP_DF);
674 /* ANK: dirty, but effective trick. Upgrade options only if
675 * the segment to be fragmented was THE FIRST (otherwise,
676 * options are already fixed) and make it ONCE
677 * on the initial skb, so that all the following fragments
678 * will inherit fixed options.
681 ip_options_fragment(from);
684 struct sk_buff *ip_frag_next(struct sk_buff *skb, struct ip_frag_state *state)
686 unsigned int len = state->left;
687 struct sk_buff *skb2;
691 /* IF: it doesn't fit, use 'mtu' - the data space left */
692 if (len > state->mtu)
694 /* IF: we are not sending up to and including the packet end
695 then align the next start on an eight byte boundary */
696 if (len < state->left) {
700 /* Allocate buffer */
701 skb2 = alloc_skb(len + state->hlen + state->ll_rs, GFP_ATOMIC);
703 return ERR_PTR(-ENOMEM);
706 * Set up data on packet
709 ip_copy_metadata(skb2, skb);
710 skb_reserve(skb2, state->ll_rs);
711 skb_put(skb2, len + state->hlen);
712 skb_reset_network_header(skb2);
713 skb2->transport_header = skb2->network_header + state->hlen;
716 * Charge the memory for the fragment to any owner
721 skb_set_owner_w(skb2, skb->sk);
724 * Copy the packet header into the new buffer.
727 skb_copy_from_linear_data(skb, skb_network_header(skb2), state->hlen);
730 * Copy a block of the IP datagram.
732 if (skb_copy_bits(skb, state->ptr, skb_transport_header(skb2), len))
737 * Fill in the new header fields.
740 iph->frag_off = htons((state->offset >> 3));
743 * Added AC : If we are fragmenting a fragment that's not the
744 * last fragment then keep MF on each bit
746 if (state->left > 0 || state->not_last_frag)
747 iph->frag_off |= htons(IP_MF);
749 state->offset += len;
751 iph->tot_len = htons(len + state->hlen);
757 EXPORT_SYMBOL(ip_frag_next);
760 * This IP datagram is too large to be sent in one piece. Break it up into
761 * smaller pieces (each of size equal to IP header plus
762 * a block of the data of the original IP data part) that will yet fit in a
763 * single device frame, and queue such a frame for sending.
766 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
767 int (*output)(struct net *, struct sock *, struct sk_buff *))
770 struct sk_buff *skb2;
771 struct rtable *rt = skb_rtable(skb);
772 unsigned int mtu, hlen, ll_rs;
773 struct ip_fraglist_iter iter;
774 ktime_t tstamp = skb->tstamp;
775 struct ip_frag_state state;
778 /* for offloaded checksums cleanup checksum before fragmentation */
779 if (skb->ip_summed == CHECKSUM_PARTIAL &&
780 (err = skb_checksum_help(skb)))
784 * Point into the IP datagram header.
789 mtu = ip_skb_dst_mtu(sk, skb);
790 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
791 mtu = IPCB(skb)->frag_max_size;
794 * Setup starting values.
798 mtu = mtu - hlen; /* Size of data space */
799 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
800 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
802 /* When frag_list is given, use it. First, check its validity:
803 * some transformers could create wrong frag_list or break existing
804 * one, it is not prohibited. In this case fall back to copying.
806 * LATER: this step can be merged to real generation of fragments,
807 * we can switch to copy when see the first bad fragment.
809 if (skb_has_frag_list(skb)) {
810 struct sk_buff *frag, *frag2;
811 unsigned int first_len = skb_pagelen(skb);
813 if (first_len - hlen > mtu ||
814 ((first_len - hlen) & 7) ||
815 ip_is_fragment(iph) ||
817 skb_headroom(skb) < ll_rs)
820 skb_walk_frags(skb, frag) {
821 /* Correct geometry. */
822 if (frag->len > mtu ||
823 ((frag->len & 7) && frag->next) ||
824 skb_headroom(frag) < hlen + ll_rs)
825 goto slow_path_clean;
827 /* Partially cloned skb? */
828 if (skb_shared(frag))
829 goto slow_path_clean;
834 frag->destructor = sock_wfree;
836 skb->truesize -= frag->truesize;
839 /* Everything is OK. Generate! */
840 ip_fraglist_init(skb, iph, hlen, &iter);
843 /* Prepare header of the next frame,
844 * before previous one went down. */
846 ip_fraglist_ipcb_prepare(skb, &iter);
847 ip_fraglist_prepare(skb, &iter);
850 skb->tstamp = tstamp;
851 err = output(net, sk, skb);
854 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
855 if (err || !iter.frag)
858 skb = ip_fraglist_next(&iter);
862 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
866 kfree_skb_list(iter.frag);
868 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
872 skb_walk_frags(skb, frag2) {
876 frag2->destructor = NULL;
877 skb->truesize += frag2->truesize;
883 * Fragment the datagram.
886 ip_frag_init(skb, hlen, ll_rs, mtu, &state);
889 * Keep copying data until we run out.
892 while (state.left > 0) {
893 bool first_frag = (state.offset == 0);
895 skb2 = ip_frag_next(skb, &state);
900 ip_frag_ipcb(skb, skb2, first_frag, &state);
903 * Put this fragment into the sending queue.
905 skb2->tstamp = tstamp;
906 err = output(net, sk, skb2);
910 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
913 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
918 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
921 EXPORT_SYMBOL(ip_do_fragment);
924 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
926 struct msghdr *msg = from;
928 if (skb->ip_summed == CHECKSUM_PARTIAL) {
929 if (!copy_from_iter_full(to, len, &msg->msg_iter))
933 if (!csum_and_copy_from_iter_full(to, len, &csum, &msg->msg_iter))
935 skb->csum = csum_block_add(skb->csum, csum, odd);
939 EXPORT_SYMBOL(ip_generic_getfrag);
942 csum_page(struct page *page, int offset, int copy)
947 csum = csum_partial(kaddr + offset, copy, 0);
952 static int __ip_append_data(struct sock *sk,
954 struct sk_buff_head *queue,
955 struct inet_cork *cork,
956 struct page_frag *pfrag,
957 int getfrag(void *from, char *to, int offset,
958 int len, int odd, struct sk_buff *skb),
959 void *from, int length, int transhdrlen,
962 struct inet_sock *inet = inet_sk(sk);
963 struct ubuf_info *uarg = NULL;
966 struct ip_options *opt = cork->opt;
973 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
974 int csummode = CHECKSUM_NONE;
975 struct rtable *rt = (struct rtable *)cork->dst;
976 unsigned int wmem_alloc_delta = 0;
977 bool paged, extra_uref = false;
980 skb = skb_peek_tail(queue);
982 exthdrlen = !skb ? rt->dst.header_len : 0;
983 mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
984 paged = !!cork->gso_size;
986 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
987 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
988 tskey = sk->sk_tskey++;
990 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
992 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
993 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
994 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
996 if (cork->length + length > maxnonfragsize - fragheaderlen) {
997 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
998 mtu - (opt ? opt->optlen : 0));
1003 * transhdrlen > 0 means that this is the first fragment and we wish
1004 * it won't be fragmented in the future.
1007 length + fragheaderlen <= mtu &&
1008 rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
1009 (!(flags & MSG_MORE) || cork->gso_size) &&
1010 (!exthdrlen || (rt->dst.dev->features & NETIF_F_HW_ESP_TX_CSUM)))
1011 csummode = CHECKSUM_PARTIAL;
1013 if (flags & MSG_ZEROCOPY && length && sock_flag(sk, SOCK_ZEROCOPY)) {
1014 uarg = sock_zerocopy_realloc(sk, length, skb_zcopy(skb));
1017 extra_uref = !skb_zcopy(skb); /* only ref on new uarg */
1018 if (rt->dst.dev->features & NETIF_F_SG &&
1019 csummode == CHECKSUM_PARTIAL) {
1023 skb_zcopy_set(skb, uarg, &extra_uref);
1027 cork->length += length;
1029 /* So, what's going on in the loop below?
1031 * We use calculated fragment length to generate chained skb,
1032 * each of segments is IP fragment ready for sending to network after
1033 * adding appropriate IP header.
1039 while (length > 0) {
1040 /* Check if the remaining data fits into current packet. */
1041 copy = mtu - skb->len;
1043 copy = maxfraglen - skb->len;
1046 unsigned int datalen;
1047 unsigned int fraglen;
1048 unsigned int fraggap;
1049 unsigned int alloclen;
1050 unsigned int pagedlen;
1051 struct sk_buff *skb_prev;
1055 fraggap = skb_prev->len - maxfraglen;
1060 * If remaining data exceeds the mtu,
1061 * we know we need more fragment(s).
1063 datalen = length + fraggap;
1064 if (datalen > mtu - fragheaderlen)
1065 datalen = maxfraglen - fragheaderlen;
1066 fraglen = datalen + fragheaderlen;
1069 if ((flags & MSG_MORE) &&
1070 !(rt->dst.dev->features&NETIF_F_SG))
1075 alloclen = min_t(int, fraglen, MAX_HEADER);
1076 pagedlen = fraglen - alloclen;
1079 alloclen += exthdrlen;
1081 /* The last fragment gets additional space at tail.
1082 * Note, with MSG_MORE we overallocate on fragments,
1083 * because we have no idea what fragment will be
1086 if (datalen == length + fraggap)
1087 alloclen += rt->dst.trailer_len;
1090 skb = sock_alloc_send_skb(sk,
1091 alloclen + hh_len + 15,
1092 (flags & MSG_DONTWAIT), &err);
1095 if (refcount_read(&sk->sk_wmem_alloc) + wmem_alloc_delta <=
1097 skb = alloc_skb(alloclen + hh_len + 15,
1106 * Fill in the control structures
1108 skb->ip_summed = csummode;
1110 skb_reserve(skb, hh_len);
1113 * Find where to start putting bytes.
1115 data = skb_put(skb, fraglen + exthdrlen - pagedlen);
1116 skb_set_network_header(skb, exthdrlen);
1117 skb->transport_header = (skb->network_header +
1119 data += fragheaderlen + exthdrlen;
1122 skb->csum = skb_copy_and_csum_bits(
1123 skb_prev, maxfraglen,
1124 data + transhdrlen, fraggap, 0);
1125 skb_prev->csum = csum_sub(skb_prev->csum,
1128 pskb_trim_unique(skb_prev, maxfraglen);
1131 copy = datalen - transhdrlen - fraggap - pagedlen;
1132 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1139 length -= copy + transhdrlen;
1142 csummode = CHECKSUM_NONE;
1144 /* only the initial fragment is time stamped */
1145 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1147 skb_shinfo(skb)->tskey = tskey;
1149 skb_zcopy_set(skb, uarg, &extra_uref);
1151 if ((flags & MSG_CONFIRM) && !skb_prev)
1152 skb_set_dst_pending_confirm(skb, 1);
1155 * Put the packet on the pending queue.
1157 if (!skb->destructor) {
1158 skb->destructor = sock_wfree;
1160 wmem_alloc_delta += skb->truesize;
1162 __skb_queue_tail(queue, skb);
1169 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1170 skb_tailroom(skb) >= copy) {
1174 if (getfrag(from, skb_put(skb, copy),
1175 offset, copy, off, skb) < 0) {
1176 __skb_trim(skb, off);
1180 } else if (!uarg || !uarg->zerocopy) {
1181 int i = skb_shinfo(skb)->nr_frags;
1184 if (!sk_page_frag_refill(sk, pfrag))
1187 if (!skb_can_coalesce(skb, i, pfrag->page,
1190 if (i == MAX_SKB_FRAGS)
1193 __skb_fill_page_desc(skb, i, pfrag->page,
1195 skb_shinfo(skb)->nr_frags = ++i;
1196 get_page(pfrag->page);
1198 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1200 page_address(pfrag->page) + pfrag->offset,
1201 offset, copy, skb->len, skb) < 0)
1204 pfrag->offset += copy;
1205 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1207 skb->data_len += copy;
1208 skb->truesize += copy;
1209 wmem_alloc_delta += copy;
1211 err = skb_zerocopy_iter_dgram(skb, from, copy);
1219 if (wmem_alloc_delta)
1220 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1227 sock_zerocopy_put_abort(uarg, extra_uref);
1228 cork->length -= length;
1229 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1230 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1234 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1235 struct ipcm_cookie *ipc, struct rtable **rtp)
1237 struct ip_options_rcu *opt;
1245 * setup for corking.
1250 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1252 if (unlikely(!cork->opt))
1255 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1256 cork->flags |= IPCORK_OPT;
1257 cork->addr = ipc->addr;
1261 * We steal reference to this route, caller should not release it
1264 cork->fragsize = ip_sk_use_pmtu(sk) ?
1265 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1267 cork->gso_size = ipc->gso_size;
1268 cork->dst = &rt->dst;
1270 cork->ttl = ipc->ttl;
1271 cork->tos = ipc->tos;
1272 cork->mark = ipc->sockc.mark;
1273 cork->priority = ipc->priority;
1274 cork->transmit_time = ipc->sockc.transmit_time;
1276 sock_tx_timestamp(sk, ipc->sockc.tsflags, &cork->tx_flags);
1282 * ip_append_data() and ip_append_page() can make one large IP datagram
1283 * from many pieces of data. Each pieces will be holded on the socket
1284 * until ip_push_pending_frames() is called. Each piece can be a page
1287 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1288 * this interface potentially.
1290 * LATER: length must be adjusted by pad at tail, when it is required.
1292 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1293 int getfrag(void *from, char *to, int offset, int len,
1294 int odd, struct sk_buff *skb),
1295 void *from, int length, int transhdrlen,
1296 struct ipcm_cookie *ipc, struct rtable **rtp,
1299 struct inet_sock *inet = inet_sk(sk);
1302 if (flags&MSG_PROBE)
1305 if (skb_queue_empty(&sk->sk_write_queue)) {
1306 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1313 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1314 sk_page_frag(sk), getfrag,
1315 from, length, transhdrlen, flags);
1318 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1319 int offset, size_t size, int flags)
1321 struct inet_sock *inet = inet_sk(sk);
1322 struct sk_buff *skb;
1324 struct ip_options *opt = NULL;
1325 struct inet_cork *cork;
1330 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1335 if (flags&MSG_PROBE)
1338 if (skb_queue_empty(&sk->sk_write_queue))
1341 cork = &inet->cork.base;
1342 rt = (struct rtable *)cork->dst;
1343 if (cork->flags & IPCORK_OPT)
1346 if (!(rt->dst.dev->features&NETIF_F_SG))
1349 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1350 mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
1352 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1353 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1354 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1356 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1357 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1358 mtu - (opt ? opt->optlen : 0));
1362 skb = skb_peek_tail(&sk->sk_write_queue);
1366 cork->length += size;
1369 /* Check if the remaining data fits into current packet. */
1370 len = mtu - skb->len;
1372 len = maxfraglen - skb->len;
1375 struct sk_buff *skb_prev;
1379 fraggap = skb_prev->len - maxfraglen;
1381 alloclen = fragheaderlen + hh_len + fraggap + 15;
1382 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1383 if (unlikely(!skb)) {
1389 * Fill in the control structures
1391 skb->ip_summed = CHECKSUM_NONE;
1393 skb_reserve(skb, hh_len);
1396 * Find where to start putting bytes.
1398 skb_put(skb, fragheaderlen + fraggap);
1399 skb_reset_network_header(skb);
1400 skb->transport_header = (skb->network_header +
1403 skb->csum = skb_copy_and_csum_bits(skb_prev,
1405 skb_transport_header(skb),
1407 skb_prev->csum = csum_sub(skb_prev->csum,
1409 pskb_trim_unique(skb_prev, maxfraglen);
1413 * Put the packet on the pending queue.
1415 __skb_queue_tail(&sk->sk_write_queue, skb);
1422 if (skb_append_pagefrags(skb, page, offset, len)) {
1427 if (skb->ip_summed == CHECKSUM_NONE) {
1429 csum = csum_page(page, offset, len);
1430 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1434 skb->data_len += len;
1435 skb->truesize += len;
1436 refcount_add(len, &sk->sk_wmem_alloc);
1443 cork->length -= size;
1444 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1448 static void ip_cork_release(struct inet_cork *cork)
1450 cork->flags &= ~IPCORK_OPT;
1453 dst_release(cork->dst);
1458 * Combined all pending IP fragments on the socket as one IP datagram
1459 * and push them out.
1461 struct sk_buff *__ip_make_skb(struct sock *sk,
1463 struct sk_buff_head *queue,
1464 struct inet_cork *cork)
1466 struct sk_buff *skb, *tmp_skb;
1467 struct sk_buff **tail_skb;
1468 struct inet_sock *inet = inet_sk(sk);
1469 struct net *net = sock_net(sk);
1470 struct ip_options *opt = NULL;
1471 struct rtable *rt = (struct rtable *)cork->dst;
1476 skb = __skb_dequeue(queue);
1479 tail_skb = &(skb_shinfo(skb)->frag_list);
1481 /* move skb->data to ip header from ext header */
1482 if (skb->data < skb_network_header(skb))
1483 __skb_pull(skb, skb_network_offset(skb));
1484 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1485 __skb_pull(tmp_skb, skb_network_header_len(skb));
1486 *tail_skb = tmp_skb;
1487 tail_skb = &(tmp_skb->next);
1488 skb->len += tmp_skb->len;
1489 skb->data_len += tmp_skb->len;
1490 skb->truesize += tmp_skb->truesize;
1491 tmp_skb->destructor = NULL;
1495 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1496 * to fragment the frame generated here. No matter, what transforms
1497 * how transforms change size of the packet, it will come out.
1499 skb->ignore_df = ip_sk_ignore_df(sk);
1501 /* DF bit is set when we want to see DF on outgoing frames.
1502 * If ignore_df is set too, we still allow to fragment this frame
1504 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1505 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1506 (skb->len <= dst_mtu(&rt->dst) &&
1507 ip_dont_fragment(sk, &rt->dst)))
1510 if (cork->flags & IPCORK_OPT)
1515 else if (rt->rt_type == RTN_MULTICAST)
1518 ttl = ip_select_ttl(inet, &rt->dst);
1523 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1526 iph->protocol = sk->sk_protocol;
1527 ip_copy_addrs(iph, fl4);
1528 ip_select_ident(net, skb, sk);
1531 iph->ihl += opt->optlen>>2;
1532 ip_options_build(skb, opt, cork->addr, rt, 0);
1535 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1536 skb->mark = cork->mark;
1537 skb->tstamp = cork->transmit_time;
1539 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1543 skb_dst_set(skb, &rt->dst);
1545 if (iph->protocol == IPPROTO_ICMP)
1546 icmp_out_count(net, ((struct icmphdr *)
1547 skb_transport_header(skb))->type);
1549 ip_cork_release(cork);
1554 int ip_send_skb(struct net *net, struct sk_buff *skb)
1558 err = ip_local_out(net, skb->sk, skb);
1561 err = net_xmit_errno(err);
1563 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1569 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1571 struct sk_buff *skb;
1573 skb = ip_finish_skb(sk, fl4);
1577 /* Netfilter gets whole the not fragmented skb. */
1578 return ip_send_skb(sock_net(sk), skb);
1582 * Throw away all pending data on the socket.
1584 static void __ip_flush_pending_frames(struct sock *sk,
1585 struct sk_buff_head *queue,
1586 struct inet_cork *cork)
1588 struct sk_buff *skb;
1590 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1593 ip_cork_release(cork);
1596 void ip_flush_pending_frames(struct sock *sk)
1598 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1601 struct sk_buff *ip_make_skb(struct sock *sk,
1603 int getfrag(void *from, char *to, int offset,
1604 int len, int odd, struct sk_buff *skb),
1605 void *from, int length, int transhdrlen,
1606 struct ipcm_cookie *ipc, struct rtable **rtp,
1607 struct inet_cork *cork, unsigned int flags)
1609 struct sk_buff_head queue;
1612 if (flags & MSG_PROBE)
1615 __skb_queue_head_init(&queue);
1620 err = ip_setup_cork(sk, cork, ipc, rtp);
1622 return ERR_PTR(err);
1624 err = __ip_append_data(sk, fl4, &queue, cork,
1625 ¤t->task_frag, getfrag,
1626 from, length, transhdrlen, flags);
1628 __ip_flush_pending_frames(sk, &queue, cork);
1629 return ERR_PTR(err);
1632 return __ip_make_skb(sk, fl4, &queue, cork);
1636 * Fetch data from kernel space and fill in checksum if needed.
1638 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1639 int len, int odd, struct sk_buff *skb)
1643 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1644 skb->csum = csum_block_add(skb->csum, csum, odd);
1649 * Generic function to send a packet as reply to another packet.
1650 * Used to send some TCP resets/acks so far.
1652 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1653 const struct ip_options *sopt,
1654 __be32 daddr, __be32 saddr,
1655 const struct ip_reply_arg *arg,
1656 unsigned int len, u64 transmit_time)
1658 struct ip_options_data replyopts;
1659 struct ipcm_cookie ipc;
1661 struct rtable *rt = skb_rtable(skb);
1662 struct net *net = sock_net(sk);
1663 struct sk_buff *nskb;
1667 if (__ip_options_echo(net, &replyopts.opt.opt, skb, sopt))
1672 ipc.sockc.transmit_time = transmit_time;
1674 if (replyopts.opt.opt.optlen) {
1675 ipc.opt = &replyopts.opt;
1677 if (replyopts.opt.opt.srr)
1678 daddr = replyopts.opt.opt.faddr;
1681 oif = arg->bound_dev_if;
1682 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1685 flowi4_init_output(&fl4, oif,
1686 IP4_REPLY_MARK(net, skb->mark) ?: sk->sk_mark,
1688 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1689 ip_reply_arg_flowi_flags(arg),
1691 tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1693 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1694 rt = ip_route_output_key(net, &fl4);
1698 inet_sk(sk)->tos = arg->tos;
1700 sk->sk_protocol = ip_hdr(skb)->protocol;
1701 sk->sk_bound_dev_if = arg->bound_dev_if;
1702 sk->sk_sndbuf = sysctl_wmem_default;
1703 sk->sk_mark = fl4.flowi4_mark;
1704 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1705 len, 0, &ipc, &rt, MSG_DONTWAIT);
1706 if (unlikely(err)) {
1707 ip_flush_pending_frames(sk);
1711 nskb = skb_peek(&sk->sk_write_queue);
1713 if (arg->csumoffset >= 0)
1714 *((__sum16 *)skb_transport_header(nskb) +
1715 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1717 nskb->ip_summed = CHECKSUM_NONE;
1718 ip_push_pending_frames(sk, &fl4);
1724 void __init ip_init(void)
1729 #if defined(CONFIG_IP_MULTICAST)