1 // SPDX-License-Identifier: GPL-2.0-only
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * The Internet Protocol (IP) output module.
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Donald Becker, <becker@super.org>
12 * Alan Cox, <Alan.Cox@linux.org>
14 * Stefan Becker, <stefanb@yello.ping.de>
15 * Jorge Cwik, <jorge@laser.satlink.net>
16 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
17 * Hirokazu Takahashi, <taka@valinux.co.jp>
19 * See ip_input.c for original log
22 * Alan Cox : Missing nonblock feature in ip_build_xmit.
23 * Mike Kilburn : htons() missing in ip_build_xmit.
24 * Bradford Johnson: Fix faulty handling of some frames when
26 * Alexander Demenshin: Missing sk/skb free in ip_queue_xmit
27 * (in case if packet not accepted by
28 * output firewall rules)
29 * Mike McLagan : Routing by source
30 * Alexey Kuznetsov: use new route cache
31 * Andi Kleen: Fix broken PMTU recovery and remove
32 * some redundant tests.
33 * Vitaly E. Lavrov : Transparent proxy revived after year coma.
34 * Andi Kleen : Replace ip_reply with ip_send_reply.
35 * Andi Kleen : Split fast and slow ip_build_xmit path
36 * for decreased register pressure on x86
37 * and more readability.
38 * Marc Boucher : When call_out_firewall returns FW_QUEUE,
39 * silently drop skb instead of failing with -EPERM.
40 * Detlev Wengorz : Copy protocol for fragments.
41 * Hirokazu Takahashi: HW checksumming for outgoing UDP
43 * Hirokazu Takahashi: sendfile() on UDP works now.
46 #include <linux/uaccess.h>
47 #include <linux/module.h>
48 #include <linux/types.h>
49 #include <linux/kernel.h>
51 #include <linux/string.h>
52 #include <linux/errno.h>
53 #include <linux/highmem.h>
54 #include <linux/slab.h>
56 #include <linux/socket.h>
57 #include <linux/sockios.h>
59 #include <linux/inet.h>
60 #include <linux/netdevice.h>
61 #include <linux/etherdevice.h>
62 #include <linux/proc_fs.h>
63 #include <linux/stat.h>
64 #include <linux/init.h>
68 #include <net/protocol.h>
69 #include <net/route.h>
71 #include <linux/skbuff.h>
75 #include <net/checksum.h>
77 #include <net/inetpeer.h>
78 #include <net/inet_ecn.h>
79 #include <net/lwtunnel.h>
80 #include <net/inet_dscp.h>
81 #include <linux/bpf-cgroup.h>
82 #include <linux/igmp.h>
83 #include <linux/netfilter_ipv4.h>
84 #include <linux/netfilter_bridge.h>
85 #include <linux/netlink.h>
86 #include <linux/tcp.h>
89 ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
91 int (*output)(struct net *, struct sock *, struct sk_buff *));
93 /* Generate a checksum for an outgoing IP datagram. */
94 void ip_send_check(struct iphdr *iph)
97 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
99 EXPORT_SYMBOL(ip_send_check);
101 int __ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
103 struct iphdr *iph = ip_hdr(skb);
105 IP_INC_STATS(net, IPSTATS_MIB_OUTREQUESTS);
107 iph_set_totlen(iph, skb->len);
110 /* if egress device is enslaved to an L3 master device pass the
111 * skb to its handler for processing
113 skb = l3mdev_ip_out(sk, skb);
117 skb->protocol = htons(ETH_P_IP);
119 return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT,
120 net, sk, skb, NULL, skb_dst(skb)->dev,
124 int ip_local_out(struct net *net, struct sock *sk, struct sk_buff *skb)
128 err = __ip_local_out(net, sk, skb);
129 if (likely(err == 1))
130 err = dst_output(net, sk, skb);
134 EXPORT_SYMBOL_GPL(ip_local_out);
136 static inline int ip_select_ttl(const struct inet_sock *inet,
137 const struct dst_entry *dst)
139 int ttl = READ_ONCE(inet->uc_ttl);
142 ttl = ip4_dst_hoplimit(dst);
147 * Add an ip header to a skbuff and send it out.
150 int ip_build_and_send_pkt(struct sk_buff *skb, const struct sock *sk,
151 __be32 saddr, __be32 daddr, struct ip_options_rcu *opt,
154 const struct inet_sock *inet = inet_sk(sk);
155 struct rtable *rt = skb_rtable(skb);
156 struct net *net = sock_net(sk);
159 /* Build the IP header. */
160 skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
161 skb_reset_network_header(skb);
166 iph->ttl = ip_select_ttl(inet, &rt->dst);
167 iph->daddr = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
169 iph->protocol = sk->sk_protocol;
170 /* Do not bother generating IPID for small packets (eg SYNACK) */
171 if (skb->len <= IPV4_MIN_MTU || ip_dont_fragment(sk, &rt->dst)) {
172 iph->frag_off = htons(IP_DF);
176 /* TCP packets here are SYNACK with fat IPv4/TCP options.
177 * Avoid using the hashed IP ident generator.
179 if (sk->sk_protocol == IPPROTO_TCP)
180 iph->id = (__force __be16)get_random_u16();
182 __ip_select_ident(net, iph, 1);
185 if (opt && opt->opt.optlen) {
186 iph->ihl += opt->opt.optlen>>2;
187 ip_options_build(skb, &opt->opt, daddr, rt);
190 skb->priority = READ_ONCE(sk->sk_priority);
192 skb->mark = READ_ONCE(sk->sk_mark);
195 return ip_local_out(net, skb->sk, skb);
197 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
199 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
201 struct dst_entry *dst = skb_dst(skb);
202 struct rtable *rt = dst_rtable(dst);
203 struct net_device *dev = dst->dev;
204 unsigned int hh_len = LL_RESERVED_SPACE(dev);
205 struct neighbour *neigh;
206 bool is_v6gw = false;
208 if (rt->rt_type == RTN_MULTICAST) {
209 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
210 } else if (rt->rt_type == RTN_BROADCAST)
211 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
213 /* OUTOCTETS should be counted after fragment */
214 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
216 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
217 skb = skb_expand_head(skb, hh_len);
222 if (lwtunnel_xmit_redirect(dst->lwtstate)) {
223 int res = lwtunnel_xmit(skb);
225 if (res != LWTUNNEL_XMIT_CONTINUE)
230 neigh = ip_neigh_for_gw(rt, skb, &is_v6gw);
231 if (!IS_ERR(neigh)) {
234 sock_confirm_neigh(skb, neigh);
235 /* if crossing protocols, can not use the cached header */
236 res = neigh_output(neigh, skb, is_v6gw);
242 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
244 kfree_skb_reason(skb, SKB_DROP_REASON_NEIGH_CREATEFAIL);
245 return PTR_ERR(neigh);
248 static int ip_finish_output_gso(struct net *net, struct sock *sk,
249 struct sk_buff *skb, unsigned int mtu)
251 struct sk_buff *segs, *nskb;
252 netdev_features_t features;
255 /* common case: seglen is <= mtu
257 if (skb_gso_validate_network_len(skb, mtu))
258 return ip_finish_output2(net, sk, skb);
260 /* Slowpath - GSO segment length exceeds the egress MTU.
262 * This can happen in several cases:
263 * - Forwarding of a TCP GRO skb, when DF flag is not set.
264 * - Forwarding of an skb that arrived on a virtualization interface
265 * (virtio-net/vhost/tap) with TSO/GSO size set by other network
267 * - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
268 * interface with a smaller MTU.
269 * - Arriving GRO skb (or GSO skb in a virtualized environment) that is
270 * bridged to a NETIF_F_TSO tunnel stacked over an interface with an
273 features = netif_skb_features(skb);
274 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_GSO_CB_OFFSET);
275 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
276 if (IS_ERR_OR_NULL(segs)) {
283 skb_list_walk_safe(segs, segs, nskb) {
286 skb_mark_not_on_list(segs);
287 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
296 static int __ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
300 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
301 /* Policy lookup after SNAT yielded a new policy */
302 if (skb_dst(skb)->xfrm) {
303 IPCB(skb)->flags |= IPSKB_REROUTED;
304 return dst_output(net, sk, skb);
307 mtu = ip_skb_dst_mtu(sk, skb);
309 return ip_finish_output_gso(net, sk, skb, mtu);
311 if (skb->len > mtu || IPCB(skb)->frag_max_size)
312 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
314 return ip_finish_output2(net, sk, skb);
317 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
321 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
323 case NET_XMIT_SUCCESS:
324 return __ip_finish_output(net, sk, skb);
326 return __ip_finish_output(net, sk, skb) ? : ret;
328 kfree_skb_reason(skb, SKB_DROP_REASON_BPF_CGROUP_EGRESS);
333 static int ip_mc_finish_output(struct net *net, struct sock *sk,
336 struct rtable *new_rt;
340 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
345 case NET_XMIT_SUCCESS:
348 kfree_skb_reason(skb, SKB_DROP_REASON_BPF_CGROUP_EGRESS);
352 /* Reset rt_iif so that inet_iif() will return skb->skb_iif. Setting
353 * this to non-zero causes ipi_ifindex in in_pktinfo to be overwritten,
354 * see ipv4_pktinfo_prepare().
356 new_rt = rt_dst_clone(net->loopback_dev, skb_rtable(skb));
360 skb_dst_set(skb, &new_rt->dst);
363 err = dev_loopback_xmit(net, sk, skb);
364 return (do_cn && err) ? ret : err;
367 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
369 struct rtable *rt = skb_rtable(skb);
370 struct net_device *dev = rt->dst.dev;
373 * If the indicated interface is up and running, send the packet.
376 skb->protocol = htons(ETH_P_IP);
379 * Multicasts are looped back for other local users
382 if (rt->rt_flags&RTCF_MULTICAST) {
384 #ifdef CONFIG_IP_MROUTE
385 /* Small optimization: do not loopback not local frames,
386 which returned after forwarding; they will be dropped
387 by ip_mr_input in any case.
388 Note, that local frames are looped back to be delivered
391 This check is duplicated in ip_mr_input at the moment.
394 ((rt->rt_flags & RTCF_LOCAL) ||
395 !(IPCB(skb)->flags & IPSKB_FORWARDED))
398 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
400 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
401 net, sk, newskb, NULL, newskb->dev,
402 ip_mc_finish_output);
405 /* Multicasts with ttl 0 must not go beyond the host */
407 if (ip_hdr(skb)->ttl == 0) {
413 if (rt->rt_flags&RTCF_BROADCAST) {
414 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
416 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
417 net, sk, newskb, NULL, newskb->dev,
418 ip_mc_finish_output);
421 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
422 net, sk, skb, NULL, skb->dev,
424 !(IPCB(skb)->flags & IPSKB_REROUTED));
427 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
429 struct net_device *dev = skb_dst(skb)->dev, *indev = skb->dev;
432 skb->protocol = htons(ETH_P_IP);
434 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
435 net, sk, skb, indev, dev,
437 !(IPCB(skb)->flags & IPSKB_REROUTED));
439 EXPORT_SYMBOL(ip_output);
442 * copy saddr and daddr, possibly using 64bit load/stores
444 * iph->saddr = fl4->saddr;
445 * iph->daddr = fl4->daddr;
447 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
449 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
450 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
452 iph->saddr = fl4->saddr;
453 iph->daddr = fl4->daddr;
456 /* Note: skb->sk can be different from sk, in case of tunnels */
457 int __ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl,
460 struct inet_sock *inet = inet_sk(sk);
461 struct net *net = sock_net(sk);
462 struct ip_options_rcu *inet_opt;
468 /* Skip all of this if the packet is already routed,
469 * f.e. by something like SCTP.
472 inet_opt = rcu_dereference(inet->inet_opt);
474 rt = skb_rtable(skb);
478 /* Make sure we can route this packet. */
479 rt = dst_rtable(__sk_dst_check(sk, 0));
483 /* Use correct destination address if we have options. */
484 daddr = inet->inet_daddr;
485 if (inet_opt && inet_opt->opt.srr)
486 daddr = inet_opt->opt.faddr;
488 /* If this fails, retransmit mechanism of transport layer will
489 * keep trying until route appears or the connection times
492 rt = ip_route_output_ports(net, fl4, sk,
493 daddr, inet->inet_saddr,
497 tos & INET_DSCP_MASK,
498 sk->sk_bound_dev_if);
501 sk_setup_caps(sk, &rt->dst);
503 skb_dst_set_noref(skb, &rt->dst);
506 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
509 /* OK, we know where to send it, allocate and build IP header. */
510 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
511 skb_reset_network_header(skb);
513 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (tos & 0xff));
514 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
515 iph->frag_off = htons(IP_DF);
518 iph->ttl = ip_select_ttl(inet, &rt->dst);
519 iph->protocol = sk->sk_protocol;
520 ip_copy_addrs(iph, fl4);
522 /* Transport layer set skb->h.foo itself. */
524 if (inet_opt && inet_opt->opt.optlen) {
525 iph->ihl += inet_opt->opt.optlen >> 2;
526 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt);
529 ip_select_ident_segs(net, skb, sk,
530 skb_shinfo(skb)->gso_segs ?: 1);
532 /* TODO : should we use skb->sk here instead of sk ? */
533 skb->priority = READ_ONCE(sk->sk_priority);
534 skb->mark = READ_ONCE(sk->sk_mark);
536 res = ip_local_out(net, sk, skb);
542 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
543 kfree_skb_reason(skb, SKB_DROP_REASON_IP_OUTNOROUTES);
544 return -EHOSTUNREACH;
546 EXPORT_SYMBOL(__ip_queue_xmit);
548 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
550 return __ip_queue_xmit(sk, skb, fl, READ_ONCE(inet_sk(sk)->tos));
552 EXPORT_SYMBOL(ip_queue_xmit);
554 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
556 to->pkt_type = from->pkt_type;
557 to->priority = from->priority;
558 to->protocol = from->protocol;
559 to->skb_iif = from->skb_iif;
561 skb_dst_copy(to, from);
563 to->mark = from->mark;
565 skb_copy_hash(to, from);
567 #ifdef CONFIG_NET_SCHED
568 to->tc_index = from->tc_index;
571 skb_ext_copy(to, from);
572 #if IS_ENABLED(CONFIG_IP_VS)
573 to->ipvs_property = from->ipvs_property;
575 skb_copy_secmark(to, from);
578 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
580 int (*output)(struct net *, struct sock *, struct sk_buff *))
582 struct iphdr *iph = ip_hdr(skb);
584 if ((iph->frag_off & htons(IP_DF)) == 0)
585 return ip_do_fragment(net, sk, skb, output);
587 if (unlikely(!skb->ignore_df ||
588 (IPCB(skb)->frag_max_size &&
589 IPCB(skb)->frag_max_size > mtu))) {
590 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
591 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
597 return ip_do_fragment(net, sk, skb, output);
600 void ip_fraglist_init(struct sk_buff *skb, struct iphdr *iph,
601 unsigned int hlen, struct ip_fraglist_iter *iter)
603 unsigned int first_len = skb_pagelen(skb);
605 iter->frag = skb_shinfo(skb)->frag_list;
606 skb_frag_list_init(skb);
612 skb->data_len = first_len - skb_headlen(skb);
613 skb->len = first_len;
614 iph->tot_len = htons(first_len);
615 iph->frag_off = htons(IP_MF);
618 EXPORT_SYMBOL(ip_fraglist_init);
620 void ip_fraglist_prepare(struct sk_buff *skb, struct ip_fraglist_iter *iter)
622 unsigned int hlen = iter->hlen;
623 struct iphdr *iph = iter->iph;
624 struct sk_buff *frag;
627 frag->ip_summed = CHECKSUM_NONE;
628 skb_reset_transport_header(frag);
629 __skb_push(frag, hlen);
630 skb_reset_network_header(frag);
631 memcpy(skb_network_header(frag), iph, hlen);
632 iter->iph = ip_hdr(frag);
634 iph->tot_len = htons(frag->len);
635 ip_copy_metadata(frag, skb);
636 iter->offset += skb->len - hlen;
637 iph->frag_off = htons(iter->offset >> 3);
639 iph->frag_off |= htons(IP_MF);
640 /* Ready, complete checksum */
643 EXPORT_SYMBOL(ip_fraglist_prepare);
645 void ip_frag_init(struct sk_buff *skb, unsigned int hlen,
646 unsigned int ll_rs, unsigned int mtu, bool DF,
647 struct ip_frag_state *state)
649 struct iphdr *iph = ip_hdr(skb);
653 state->ll_rs = ll_rs;
656 state->left = skb->len - hlen; /* Space per frame */
657 state->ptr = hlen; /* Where to start from */
659 state->offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
660 state->not_last_frag = iph->frag_off & htons(IP_MF);
662 EXPORT_SYMBOL(ip_frag_init);
664 static void ip_frag_ipcb(struct sk_buff *from, struct sk_buff *to,
667 /* Copy the flags to each fragment. */
668 IPCB(to)->flags = IPCB(from)->flags;
670 /* ANK: dirty, but effective trick. Upgrade options only if
671 * the segment to be fragmented was THE FIRST (otherwise,
672 * options are already fixed) and make it ONCE
673 * on the initial skb, so that all the following fragments
674 * will inherit fixed options.
677 ip_options_fragment(from);
680 struct sk_buff *ip_frag_next(struct sk_buff *skb, struct ip_frag_state *state)
682 unsigned int len = state->left;
683 struct sk_buff *skb2;
686 /* IF: it doesn't fit, use 'mtu' - the data space left */
687 if (len > state->mtu)
689 /* IF: we are not sending up to and including the packet end
690 then align the next start on an eight byte boundary */
691 if (len < state->left) {
695 /* Allocate buffer */
696 skb2 = alloc_skb(len + state->hlen + state->ll_rs, GFP_ATOMIC);
698 return ERR_PTR(-ENOMEM);
701 * Set up data on packet
704 ip_copy_metadata(skb2, skb);
705 skb_reserve(skb2, state->ll_rs);
706 skb_put(skb2, len + state->hlen);
707 skb_reset_network_header(skb2);
708 skb2->transport_header = skb2->network_header + state->hlen;
711 * Charge the memory for the fragment to any owner
716 skb_set_owner_w(skb2, skb->sk);
719 * Copy the packet header into the new buffer.
722 skb_copy_from_linear_data(skb, skb_network_header(skb2), state->hlen);
725 * Copy a block of the IP datagram.
727 if (skb_copy_bits(skb, state->ptr, skb_transport_header(skb2), len))
732 * Fill in the new header fields.
735 iph->frag_off = htons((state->offset >> 3));
737 iph->frag_off |= htons(IP_DF);
740 * Added AC : If we are fragmenting a fragment that's not the
741 * last fragment then keep MF on each bit
743 if (state->left > 0 || state->not_last_frag)
744 iph->frag_off |= htons(IP_MF);
746 state->offset += len;
748 iph->tot_len = htons(len + state->hlen);
754 EXPORT_SYMBOL(ip_frag_next);
757 * This IP datagram is too large to be sent in one piece. Break it up into
758 * smaller pieces (each of size equal to IP header plus
759 * a block of the data of the original IP data part) that will yet fit in a
760 * single device frame, and queue such a frame for sending.
763 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
764 int (*output)(struct net *, struct sock *, struct sk_buff *))
767 struct sk_buff *skb2;
768 u8 tstamp_type = skb->tstamp_type;
769 struct rtable *rt = skb_rtable(skb);
770 unsigned int mtu, hlen, ll_rs;
771 struct ip_fraglist_iter iter;
772 ktime_t tstamp = skb->tstamp;
773 struct ip_frag_state state;
776 /* for offloaded checksums cleanup checksum before fragmentation */
777 if (skb->ip_summed == CHECKSUM_PARTIAL &&
778 (err = skb_checksum_help(skb)))
782 * Point into the IP datagram header.
787 mtu = ip_skb_dst_mtu(sk, skb);
788 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
789 mtu = IPCB(skb)->frag_max_size;
792 * Setup starting values.
796 mtu = mtu - hlen; /* Size of data space */
797 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
798 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
800 /* When frag_list is given, use it. First, check its validity:
801 * some transformers could create wrong frag_list or break existing
802 * one, it is not prohibited. In this case fall back to copying.
804 * LATER: this step can be merged to real generation of fragments,
805 * we can switch to copy when see the first bad fragment.
807 if (skb_has_frag_list(skb)) {
808 struct sk_buff *frag, *frag2;
809 unsigned int first_len = skb_pagelen(skb);
811 if (first_len - hlen > mtu ||
812 ((first_len - hlen) & 7) ||
813 ip_is_fragment(iph) ||
815 skb_headroom(skb) < ll_rs)
818 skb_walk_frags(skb, frag) {
819 /* Correct geometry. */
820 if (frag->len > mtu ||
821 ((frag->len & 7) && frag->next) ||
822 skb_headroom(frag) < hlen + ll_rs)
823 goto slow_path_clean;
825 /* Partially cloned skb? */
826 if (skb_shared(frag))
827 goto slow_path_clean;
832 frag->destructor = sock_wfree;
834 skb->truesize -= frag->truesize;
837 /* Everything is OK. Generate! */
838 ip_fraglist_init(skb, iph, hlen, &iter);
841 /* Prepare header of the next frame,
842 * before previous one went down. */
844 bool first_frag = (iter.offset == 0);
846 IPCB(iter.frag)->flags = IPCB(skb)->flags;
847 ip_fraglist_prepare(skb, &iter);
848 if (first_frag && IPCB(skb)->opt.optlen) {
849 /* ipcb->opt is not populated for frags
850 * coming from __ip_make_skb(),
851 * ip_options_fragment() needs optlen
853 IPCB(iter.frag)->opt.optlen =
854 IPCB(skb)->opt.optlen;
855 ip_options_fragment(iter.frag);
856 ip_send_check(iter.iph);
860 skb_set_delivery_time(skb, tstamp, tstamp_type);
861 err = output(net, sk, skb);
864 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
865 if (err || !iter.frag)
868 skb = ip_fraglist_next(&iter);
872 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
876 kfree_skb_list(iter.frag);
878 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
882 skb_walk_frags(skb, frag2) {
886 frag2->destructor = NULL;
887 skb->truesize += frag2->truesize;
893 * Fragment the datagram.
896 ip_frag_init(skb, hlen, ll_rs, mtu, IPCB(skb)->flags & IPSKB_FRAG_PMTU,
900 * Keep copying data until we run out.
903 while (state.left > 0) {
904 bool first_frag = (state.offset == 0);
906 skb2 = ip_frag_next(skb, &state);
911 ip_frag_ipcb(skb, skb2, first_frag);
914 * Put this fragment into the sending queue.
916 skb_set_delivery_time(skb2, tstamp, tstamp_type);
917 err = output(net, sk, skb2);
921 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
924 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
929 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
932 EXPORT_SYMBOL(ip_do_fragment);
935 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
937 struct msghdr *msg = from;
939 if (skb->ip_summed == CHECKSUM_PARTIAL) {
940 if (!copy_from_iter_full(to, len, &msg->msg_iter))
944 if (!csum_and_copy_from_iter_full(to, len, &csum, &msg->msg_iter))
946 skb->csum = csum_block_add(skb->csum, csum, odd);
950 EXPORT_SYMBOL(ip_generic_getfrag);
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;
965 struct ip_options *opt = cork->opt;
973 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
974 int csummode = CHECKSUM_NONE;
975 struct rtable *rt = dst_rtable(cork->dst);
976 bool paged, hold_tskey, extra_uref = false;
977 unsigned int wmem_alloc_delta = 0;
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 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
988 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
989 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
990 maxnonfragsize = ip_sk_ignore_df(sk) ? IP_MAX_MTU : mtu;
992 if (cork->length + length > maxnonfragsize - fragheaderlen) {
993 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
994 mtu - (opt ? opt->optlen : 0));
999 * transhdrlen > 0 means that this is the first fragment and we wish
1000 * it won't be fragmented in the future.
1003 length + fragheaderlen <= mtu &&
1004 rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
1005 (!(flags & MSG_MORE) || cork->gso_size) &&
1006 (!exthdrlen || (rt->dst.dev->features & NETIF_F_HW_ESP_TX_CSUM)))
1007 csummode = CHECKSUM_PARTIAL;
1009 if ((flags & MSG_ZEROCOPY) && length) {
1010 struct msghdr *msg = from;
1012 if (getfrag == ip_generic_getfrag && msg->msg_ubuf) {
1013 if (skb_zcopy(skb) && msg->msg_ubuf != skb_zcopy(skb))
1016 /* Leave uarg NULL if can't zerocopy, callers should
1017 * be able to handle it.
1019 if ((rt->dst.dev->features & NETIF_F_SG) &&
1020 csummode == CHECKSUM_PARTIAL) {
1023 uarg = msg->msg_ubuf;
1025 } else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1026 uarg = msg_zerocopy_realloc(sk, length, skb_zcopy(skb));
1029 extra_uref = !skb_zcopy(skb); /* only ref on new uarg */
1030 if (rt->dst.dev->features & NETIF_F_SG &&
1031 csummode == CHECKSUM_PARTIAL) {
1035 uarg_to_msgzc(uarg)->zerocopy = 0;
1036 skb_zcopy_set(skb, uarg, &extra_uref);
1039 } else if ((flags & MSG_SPLICE_PAGES) && length) {
1040 if (inet_test_bit(HDRINCL, sk))
1042 if (rt->dst.dev->features & NETIF_F_SG &&
1043 getfrag == ip_generic_getfrag)
1044 /* We need an empty buffer to attach stuff to */
1047 flags &= ~MSG_SPLICE_PAGES;
1050 cork->length += length;
1052 hold_tskey = cork->tx_flags & SKBTX_ANY_TSTAMP &&
1053 READ_ONCE(sk->sk_tsflags) & SOF_TIMESTAMPING_OPT_ID;
1055 tskey = atomic_inc_return(&sk->sk_tskey) - 1;
1057 /* So, what's going on in the loop below?
1059 * We use calculated fragment length to generate chained skb,
1060 * each of segments is IP fragment ready for sending to network after
1061 * adding appropriate IP header.
1067 while (length > 0) {
1068 /* Check if the remaining data fits into current packet. */
1069 copy = mtu - skb->len;
1071 copy = maxfraglen - skb->len;
1074 unsigned int datalen;
1075 unsigned int fraglen;
1076 unsigned int fraggap;
1077 unsigned int alloclen, alloc_extra;
1078 unsigned int pagedlen;
1079 struct sk_buff *skb_prev;
1083 fraggap = skb_prev->len - maxfraglen;
1088 * If remaining data exceeds the mtu,
1089 * we know we need more fragment(s).
1091 datalen = length + fraggap;
1092 if (datalen > mtu - fragheaderlen)
1093 datalen = maxfraglen - fragheaderlen;
1094 fraglen = datalen + fragheaderlen;
1097 alloc_extra = hh_len + 15;
1098 alloc_extra += exthdrlen;
1100 /* The last fragment gets additional space at tail.
1101 * Note, with MSG_MORE we overallocate on fragments,
1102 * because we have no idea what fragment will be
1105 if (datalen == length + fraggap)
1106 alloc_extra += rt->dst.trailer_len;
1108 if ((flags & MSG_MORE) &&
1109 !(rt->dst.dev->features&NETIF_F_SG))
1112 (fraglen + alloc_extra < SKB_MAX_ALLOC ||
1113 !(rt->dst.dev->features & NETIF_F_SG)))
1116 alloclen = fragheaderlen + transhdrlen;
1117 pagedlen = datalen - transhdrlen;
1120 alloclen += alloc_extra;
1123 skb = sock_alloc_send_skb(sk, alloclen,
1124 (flags & MSG_DONTWAIT), &err);
1127 if (refcount_read(&sk->sk_wmem_alloc) + wmem_alloc_delta <=
1129 skb = alloc_skb(alloclen,
1138 * Fill in the control structures
1140 skb->ip_summed = csummode;
1142 skb_reserve(skb, hh_len);
1145 * Find where to start putting bytes.
1147 data = skb_put(skb, fraglen + exthdrlen - pagedlen);
1148 skb_set_network_header(skb, exthdrlen);
1149 skb->transport_header = (skb->network_header +
1151 data += fragheaderlen + exthdrlen;
1154 skb->csum = skb_copy_and_csum_bits(
1155 skb_prev, maxfraglen,
1156 data + transhdrlen, fraggap);
1157 skb_prev->csum = csum_sub(skb_prev->csum,
1160 pskb_trim_unique(skb_prev, maxfraglen);
1163 copy = datalen - transhdrlen - fraggap - pagedlen;
1164 /* [!] NOTE: copy will be negative if pagedlen>0
1165 * because then the equation reduces to -fraggap.
1167 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1171 } else if (flags & MSG_SPLICE_PAGES) {
1176 length -= copy + transhdrlen;
1179 csummode = CHECKSUM_NONE;
1181 /* only the initial fragment is time stamped */
1182 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1184 skb_shinfo(skb)->tskey = tskey;
1186 skb_zcopy_set(skb, uarg, &extra_uref);
1188 if ((flags & MSG_CONFIRM) && !skb_prev)
1189 skb_set_dst_pending_confirm(skb, 1);
1192 * Put the packet on the pending queue.
1194 if (!skb->destructor) {
1195 skb->destructor = sock_wfree;
1197 wmem_alloc_delta += skb->truesize;
1199 __skb_queue_tail(queue, skb);
1206 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1207 skb_tailroom(skb) >= copy) {
1211 if (getfrag(from, skb_put(skb, copy),
1212 offset, copy, off, skb) < 0) {
1213 __skb_trim(skb, off);
1217 } else if (flags & MSG_SPLICE_PAGES) {
1218 struct msghdr *msg = from;
1221 if (WARN_ON_ONCE(copy > msg->msg_iter.count))
1224 err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
1229 wmem_alloc_delta += copy;
1231 int i = skb_shinfo(skb)->nr_frags;
1234 if (!sk_page_frag_refill(sk, pfrag))
1237 skb_zcopy_downgrade_managed(skb);
1238 if (!skb_can_coalesce(skb, i, pfrag->page,
1241 if (i == MAX_SKB_FRAGS)
1244 __skb_fill_page_desc(skb, i, pfrag->page,
1246 skb_shinfo(skb)->nr_frags = ++i;
1247 get_page(pfrag->page);
1249 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1251 page_address(pfrag->page) + pfrag->offset,
1252 offset, copy, skb->len, skb) < 0)
1255 pfrag->offset += copy;
1256 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1257 skb_len_add(skb, copy);
1258 wmem_alloc_delta += copy;
1260 err = skb_zerocopy_iter_dgram(skb, from, copy);
1268 if (wmem_alloc_delta)
1269 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1275 net_zcopy_put_abort(uarg, extra_uref);
1276 cork->length -= length;
1277 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1278 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1280 atomic_dec(&sk->sk_tskey);
1284 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1285 struct ipcm_cookie *ipc, struct rtable **rtp)
1287 struct ip_options_rcu *opt;
1294 cork->fragsize = ip_sk_use_pmtu(sk) ?
1295 dst_mtu(&rt->dst) : READ_ONCE(rt->dst.dev->mtu);
1297 if (!inetdev_valid_mtu(cork->fragsize))
1298 return -ENETUNREACH;
1301 * setup for corking.
1306 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1308 if (unlikely(!cork->opt))
1311 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1312 cork->flags |= IPCORK_OPT;
1313 cork->addr = ipc->addr;
1316 cork->gso_size = ipc->gso_size;
1318 cork->dst = &rt->dst;
1319 /* We stole this route, caller should not release it. */
1323 cork->ttl = ipc->ttl;
1324 cork->tos = ipc->tos;
1325 cork->mark = ipc->sockc.mark;
1326 cork->priority = ipc->priority;
1327 cork->transmit_time = ipc->sockc.transmit_time;
1329 sock_tx_timestamp(sk, ipc->sockc.tsflags, &cork->tx_flags);
1335 * ip_append_data() can make one large IP datagram from many pieces of
1336 * data. Each piece will be held on the socket until
1337 * ip_push_pending_frames() is called. Each piece can be a page or
1340 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1341 * this interface potentially.
1343 * LATER: length must be adjusted by pad at tail, when it is required.
1345 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1346 int getfrag(void *from, char *to, int offset, int len,
1347 int odd, struct sk_buff *skb),
1348 void *from, int length, int transhdrlen,
1349 struct ipcm_cookie *ipc, struct rtable **rtp,
1352 struct inet_sock *inet = inet_sk(sk);
1355 if (flags&MSG_PROBE)
1358 if (skb_queue_empty(&sk->sk_write_queue)) {
1359 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1366 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1367 sk_page_frag(sk), getfrag,
1368 from, length, transhdrlen, flags);
1371 static void ip_cork_release(struct inet_cork *cork)
1373 cork->flags &= ~IPCORK_OPT;
1376 dst_release(cork->dst);
1381 * Combined all pending IP fragments on the socket as one IP datagram
1382 * and push them out.
1384 struct sk_buff *__ip_make_skb(struct sock *sk,
1386 struct sk_buff_head *queue,
1387 struct inet_cork *cork)
1389 struct sk_buff *skb, *tmp_skb;
1390 struct sk_buff **tail_skb;
1391 struct inet_sock *inet = inet_sk(sk);
1392 struct net *net = sock_net(sk);
1393 struct ip_options *opt = NULL;
1394 struct rtable *rt = dst_rtable(cork->dst);
1399 skb = __skb_dequeue(queue);
1402 tail_skb = &(skb_shinfo(skb)->frag_list);
1404 /* move skb->data to ip header from ext header */
1405 if (skb->data < skb_network_header(skb))
1406 __skb_pull(skb, skb_network_offset(skb));
1407 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1408 __skb_pull(tmp_skb, skb_network_header_len(skb));
1409 *tail_skb = tmp_skb;
1410 tail_skb = &(tmp_skb->next);
1411 skb->len += tmp_skb->len;
1412 skb->data_len += tmp_skb->len;
1413 skb->truesize += tmp_skb->truesize;
1414 tmp_skb->destructor = NULL;
1418 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1419 * to fragment the frame generated here. No matter, what transforms
1420 * how transforms change size of the packet, it will come out.
1422 skb->ignore_df = ip_sk_ignore_df(sk);
1424 /* DF bit is set when we want to see DF on outgoing frames.
1425 * If ignore_df is set too, we still allow to fragment this frame
1427 pmtudisc = READ_ONCE(inet->pmtudisc);
1428 if (pmtudisc == IP_PMTUDISC_DO ||
1429 pmtudisc == IP_PMTUDISC_PROBE ||
1430 (skb->len <= dst_mtu(&rt->dst) &&
1431 ip_dont_fragment(sk, &rt->dst)))
1434 if (cork->flags & IPCORK_OPT)
1439 else if (rt->rt_type == RTN_MULTICAST)
1440 ttl = READ_ONCE(inet->mc_ttl);
1442 ttl = ip_select_ttl(inet, &rt->dst);
1447 iph->tos = (cork->tos != -1) ? cork->tos : READ_ONCE(inet->tos);
1450 iph->protocol = sk->sk_protocol;
1451 ip_copy_addrs(iph, fl4);
1452 ip_select_ident(net, skb, sk);
1455 iph->ihl += opt->optlen >> 2;
1456 ip_options_build(skb, opt, cork->addr, rt);
1459 skb->priority = (cork->tos != -1) ? cork->priority: READ_ONCE(sk->sk_priority);
1460 skb->mark = cork->mark;
1462 skb_set_delivery_time(skb, cork->transmit_time, SKB_CLOCK_MONOTONIC);
1464 skb_set_delivery_type_by_clockid(skb, cork->transmit_time, sk->sk_clockid);
1466 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1470 skb_dst_set(skb, &rt->dst);
1472 if (iph->protocol == IPPROTO_ICMP) {
1475 /* For such sockets, transhdrlen is zero when do ip_append_data(),
1476 * so icmphdr does not in skb linear region and can not get icmp_type
1477 * by icmp_hdr(skb)->type.
1479 if (sk->sk_type == SOCK_RAW &&
1480 !(fl4->flowi4_flags & FLOWI_FLAG_KNOWN_NH))
1481 icmp_type = fl4->fl4_icmp_type;
1483 icmp_type = icmp_hdr(skb)->type;
1484 icmp_out_count(net, icmp_type);
1487 ip_cork_release(cork);
1492 int ip_send_skb(struct net *net, struct sk_buff *skb)
1496 err = ip_local_out(net, skb->sk, skb);
1499 err = net_xmit_errno(err);
1501 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1507 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1509 struct sk_buff *skb;
1511 skb = ip_finish_skb(sk, fl4);
1515 /* Netfilter gets whole the not fragmented skb. */
1516 return ip_send_skb(sock_net(sk), skb);
1520 * Throw away all pending data on the socket.
1522 static void __ip_flush_pending_frames(struct sock *sk,
1523 struct sk_buff_head *queue,
1524 struct inet_cork *cork)
1526 struct sk_buff *skb;
1528 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1531 ip_cork_release(cork);
1534 void ip_flush_pending_frames(struct sock *sk)
1536 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1539 struct sk_buff *ip_make_skb(struct sock *sk,
1541 int getfrag(void *from, char *to, int offset,
1542 int len, int odd, struct sk_buff *skb),
1543 void *from, int length, int transhdrlen,
1544 struct ipcm_cookie *ipc, struct rtable **rtp,
1545 struct inet_cork *cork, unsigned int flags)
1547 struct sk_buff_head queue;
1550 if (flags & MSG_PROBE)
1553 __skb_queue_head_init(&queue);
1558 err = ip_setup_cork(sk, cork, ipc, rtp);
1560 return ERR_PTR(err);
1562 err = __ip_append_data(sk, fl4, &queue, cork,
1563 ¤t->task_frag, getfrag,
1564 from, length, transhdrlen, flags);
1566 __ip_flush_pending_frames(sk, &queue, cork);
1567 return ERR_PTR(err);
1570 return __ip_make_skb(sk, fl4, &queue, cork);
1574 * Fetch data from kernel space and fill in checksum if needed.
1576 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1577 int len, int odd, struct sk_buff *skb)
1581 csum = csum_partial_copy_nocheck(dptr+offset, to, len);
1582 skb->csum = csum_block_add(skb->csum, csum, odd);
1587 * Generic function to send a packet as reply to another packet.
1588 * Used to send some TCP resets/acks so far.
1590 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1591 const struct ip_options *sopt,
1592 __be32 daddr, __be32 saddr,
1593 const struct ip_reply_arg *arg,
1594 unsigned int len, u64 transmit_time, u32 txhash)
1596 struct ip_options_data replyopts;
1597 struct ipcm_cookie ipc;
1599 struct rtable *rt = skb_rtable(skb);
1600 struct net *net = sock_net(sk);
1601 struct sk_buff *nskb;
1605 if (__ip_options_echo(net, &replyopts.opt.opt, skb, sopt))
1610 ipc.sockc.transmit_time = transmit_time;
1612 if (replyopts.opt.opt.optlen) {
1613 ipc.opt = &replyopts.opt;
1615 if (replyopts.opt.opt.srr)
1616 daddr = replyopts.opt.opt.faddr;
1619 oif = arg->bound_dev_if;
1620 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1623 flowi4_init_output(&fl4, oif,
1624 IP4_REPLY_MARK(net, skb->mark) ?: sk->sk_mark,
1625 arg->tos & INET_DSCP_MASK,
1626 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1627 ip_reply_arg_flowi_flags(arg),
1629 tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1631 security_skb_classify_flow(skb, flowi4_to_flowi_common(&fl4));
1632 rt = ip_route_output_flow(net, &fl4, sk);
1636 inet_sk(sk)->tos = arg->tos & ~INET_ECN_MASK;
1638 sk->sk_protocol = ip_hdr(skb)->protocol;
1639 sk->sk_bound_dev_if = arg->bound_dev_if;
1640 sk->sk_sndbuf = READ_ONCE(sysctl_wmem_default);
1641 ipc.sockc.mark = fl4.flowi4_mark;
1642 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1643 len, 0, &ipc, &rt, MSG_DONTWAIT);
1644 if (unlikely(err)) {
1645 ip_flush_pending_frames(sk);
1649 nskb = skb_peek(&sk->sk_write_queue);
1651 if (arg->csumoffset >= 0)
1652 *((__sum16 *)skb_transport_header(nskb) +
1653 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1655 nskb->ip_summed = CHECKSUM_NONE;
1657 nskb->tstamp_type = SKB_CLOCK_MONOTONIC;
1659 skb_set_hash(nskb, txhash, PKT_HASH_TYPE_L4);
1660 ip_push_pending_frames(sk, &fl4);
1666 void __init ip_init(void)
1671 #if defined(CONFIG_IP_MULTICAST)