2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * The Internet Protocol (IP) output module.
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Donald Becker, <becker@super.org>
11 * Alan Cox, <Alan.Cox@linux.org>
13 * Stefan Becker, <stefanb@yello.ping.de>
14 * Jorge Cwik, <jorge@laser.satlink.net>
15 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
16 * Hirokazu Takahashi, <taka@valinux.co.jp>
18 * See ip_input.c for original log
21 * Alan Cox : Missing nonblock feature in ip_build_xmit.
22 * Mike Kilburn : htons() missing in ip_build_xmit.
23 * Bradford Johnson: Fix faulty handling of some frames when
25 * Alexander Demenshin: Missing sk/skb free in ip_queue_xmit
26 * (in case if packet not accepted by
27 * output firewall rules)
28 * Mike McLagan : Routing by source
29 * Alexey Kuznetsov: use new route cache
30 * Andi Kleen: Fix broken PMTU recovery and remove
31 * some redundant tests.
32 * Vitaly E. Lavrov : Transparent proxy revived after year coma.
33 * Andi Kleen : Replace ip_reply with ip_send_reply.
34 * Andi Kleen : Split fast and slow ip_build_xmit path
35 * for decreased register pressure on x86
36 * and more readibility.
37 * Marc Boucher : When call_out_firewall returns FW_QUEUE,
38 * silently drop skb instead of failing with -EPERM.
39 * Detlev Wengorz : Copy protocol for fragments.
40 * Hirokazu Takahashi: HW checksumming for outgoing UDP
42 * Hirokazu Takahashi: sendfile() on UDP works now.
45 #include <linux/uaccess.h>
46 #include <linux/module.h>
47 #include <linux/types.h>
48 #include <linux/kernel.h>
50 #include <linux/string.h>
51 #include <linux/errno.h>
52 #include <linux/highmem.h>
53 #include <linux/slab.h>
55 #include <linux/socket.h>
56 #include <linux/sockios.h>
58 #include <linux/inet.h>
59 #include <linux/netdevice.h>
60 #include <linux/etherdevice.h>
61 #include <linux/proc_fs.h>
62 #include <linux/stat.h>
63 #include <linux/init.h>
67 #include <net/protocol.h>
68 #include <net/route.h>
70 #include <linux/skbuff.h>
74 #include <net/checksum.h>
75 #include <net/inetpeer.h>
76 #include <net/inet_ecn.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 /* Do not bother generating IPID for small packets (eg SYNACK) */
164 if (skb->len <= IPV4_MIN_MTU || ip_dont_fragment(sk, &rt->dst)) {
165 iph->frag_off = htons(IP_DF);
169 /* TCP packets here are SYNACK with fat IPv4/TCP options.
170 * Avoid using the hashed IP ident generator.
172 if (sk->sk_protocol == IPPROTO_TCP)
173 iph->id = (__force __be16)prandom_u32();
175 __ip_select_ident(net, iph, 1);
178 if (opt && opt->opt.optlen) {
179 iph->ihl += opt->opt.optlen>>2;
180 ip_options_build(skb, &opt->opt, daddr, rt, 0);
183 skb->priority = sk->sk_priority;
185 skb->mark = sk->sk_mark;
188 return ip_local_out(net, skb->sk, skb);
190 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
192 static int ip_finish_output2(struct net *net, struct sock *sk, struct sk_buff *skb)
194 struct dst_entry *dst = skb_dst(skb);
195 struct rtable *rt = (struct rtable *)dst;
196 struct net_device *dev = dst->dev;
197 unsigned int hh_len = LL_RESERVED_SPACE(dev);
198 struct neighbour *neigh;
201 if (rt->rt_type == RTN_MULTICAST) {
202 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTMCAST, skb->len);
203 } else if (rt->rt_type == RTN_BROADCAST)
204 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUTBCAST, skb->len);
206 /* Be paranoid, rather than too clever. */
207 if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
208 struct sk_buff *skb2;
210 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
216 skb_set_owner_w(skb2, skb->sk);
221 if (lwtunnel_xmit_redirect(dst->lwtstate)) {
222 int res = lwtunnel_xmit(skb);
224 if (res < 0 || res == LWTUNNEL_XMIT_DONE)
229 nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
230 neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
231 if (unlikely(!neigh))
232 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
233 if (!IS_ERR(neigh)) {
236 sock_confirm_neigh(skb, neigh);
237 res = neigh_output(neigh, skb);
239 rcu_read_unlock_bh();
242 rcu_read_unlock_bh();
244 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
250 static int ip_finish_output_gso(struct net *net, struct sock *sk,
251 struct sk_buff *skb, unsigned int mtu)
253 netdev_features_t features;
254 struct sk_buff *segs;
257 /* common case: seglen is <= mtu
259 if (skb_gso_validate_network_len(skb, mtu))
260 return ip_finish_output2(net, sk, skb);
262 /* Slowpath - GSO segment length exceeds the egress MTU.
264 * This can happen in several cases:
265 * - Forwarding of a TCP GRO skb, when DF flag is not set.
266 * - Forwarding of an skb that arrived on a virtualization interface
267 * (virtio-net/vhost/tap) with TSO/GSO size set by other network
269 * - Local GSO skb transmitted on an NETIF_F_TSO tunnel stacked over an
270 * interface with a smaller MTU.
271 * - Arriving GRO skb (or GSO skb in a virtualized environment) that is
272 * bridged to a NETIF_F_TSO tunnel stacked over an interface with an
275 features = netif_skb_features(skb);
276 BUILD_BUG_ON(sizeof(*IPCB(skb)) > SKB_SGO_CB_OFFSET);
277 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
278 if (IS_ERR_OR_NULL(segs)) {
286 struct sk_buff *nskb = segs->next;
290 err = ip_fragment(net, sk, segs, mtu, ip_finish_output2);
300 static int ip_finish_output(struct net *net, struct sock *sk, struct sk_buff *skb)
305 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
311 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
312 /* Policy lookup after SNAT yielded a new policy */
313 if (skb_dst(skb)->xfrm) {
314 IPCB(skb)->flags |= IPSKB_REROUTED;
315 return dst_output(net, sk, skb);
318 mtu = ip_skb_dst_mtu(sk, skb);
320 return ip_finish_output_gso(net, sk, skb, mtu);
322 if (skb->len > mtu || IPCB(skb)->frag_max_size)
323 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
325 return ip_finish_output2(net, sk, skb);
328 static int ip_mc_finish_output(struct net *net, struct sock *sk,
333 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
339 return dev_loopback_xmit(net, sk, skb);
342 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
344 struct rtable *rt = skb_rtable(skb);
345 struct net_device *dev = rt->dst.dev;
348 * If the indicated interface is up and running, send the packet.
350 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
353 skb->protocol = htons(ETH_P_IP);
356 * Multicasts are looped back for other local users
359 if (rt->rt_flags&RTCF_MULTICAST) {
361 #ifdef CONFIG_IP_MROUTE
362 /* Small optimization: do not loopback not local frames,
363 which returned after forwarding; they will be dropped
364 by ip_mr_input in any case.
365 Note, that local frames are looped back to be delivered
368 This check is duplicated in ip_mr_input at the moment.
371 ((rt->rt_flags & RTCF_LOCAL) ||
372 !(IPCB(skb)->flags & IPSKB_FORWARDED))
375 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
377 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
378 net, sk, newskb, NULL, newskb->dev,
379 ip_mc_finish_output);
382 /* Multicasts with ttl 0 must not go beyond the host */
384 if (ip_hdr(skb)->ttl == 0) {
390 if (rt->rt_flags&RTCF_BROADCAST) {
391 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
393 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
394 net, sk, newskb, NULL, newskb->dev,
395 ip_mc_finish_output);
398 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
399 net, sk, skb, NULL, skb->dev,
401 !(IPCB(skb)->flags & IPSKB_REROUTED));
404 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
406 struct net_device *dev = skb_dst(skb)->dev;
408 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
411 skb->protocol = htons(ETH_P_IP);
413 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
414 net, sk, skb, NULL, dev,
416 !(IPCB(skb)->flags & IPSKB_REROUTED));
420 * copy saddr and daddr, possibly using 64bit load/stores
422 * iph->saddr = fl4->saddr;
423 * iph->daddr = fl4->daddr;
425 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
427 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
428 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
430 iph->saddr = fl4->saddr;
431 iph->daddr = fl4->daddr;
434 /* Note: skb->sk can be different from sk, in case of tunnels */
435 int __ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl,
438 struct inet_sock *inet = inet_sk(sk);
439 struct net *net = sock_net(sk);
440 struct ip_options_rcu *inet_opt;
446 /* Skip all of this if the packet is already routed,
447 * f.e. by something like SCTP.
450 inet_opt = rcu_dereference(inet->inet_opt);
452 rt = skb_rtable(skb);
456 /* Make sure we can route this packet. */
457 rt = (struct rtable *)__sk_dst_check(sk, 0);
461 /* Use correct destination address if we have options. */
462 daddr = inet->inet_daddr;
463 if (inet_opt && inet_opt->opt.srr)
464 daddr = inet_opt->opt.faddr;
466 /* If this fails, retransmit mechanism of transport layer will
467 * keep trying until route appears or the connection times
470 rt = ip_route_output_ports(net, fl4, sk,
471 daddr, inet->inet_saddr,
475 RT_CONN_FLAGS_TOS(sk, tos),
476 sk->sk_bound_dev_if);
479 sk_setup_caps(sk, &rt->dst);
481 skb_dst_set_noref(skb, &rt->dst);
484 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
487 /* OK, we know where to send it, allocate and build IP header. */
488 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
489 skb_reset_network_header(skb);
491 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (tos & 0xff));
492 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
493 iph->frag_off = htons(IP_DF);
496 iph->ttl = ip_select_ttl(inet, &rt->dst);
497 iph->protocol = sk->sk_protocol;
498 ip_copy_addrs(iph, fl4);
500 /* Transport layer set skb->h.foo itself. */
502 if (inet_opt && inet_opt->opt.optlen) {
503 iph->ihl += inet_opt->opt.optlen >> 2;
504 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
507 ip_select_ident_segs(net, skb, sk,
508 skb_shinfo(skb)->gso_segs ?: 1);
510 /* TODO : should we use skb->sk here instead of sk ? */
511 skb->priority = sk->sk_priority;
512 skb->mark = sk->sk_mark;
514 res = ip_local_out(net, sk, skb);
520 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
522 return -EHOSTUNREACH;
524 EXPORT_SYMBOL(__ip_queue_xmit);
526 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
528 to->pkt_type = from->pkt_type;
529 to->priority = from->priority;
530 to->protocol = from->protocol;
531 to->skb_iif = from->skb_iif;
533 skb_dst_copy(to, from);
535 to->mark = from->mark;
537 skb_copy_hash(to, from);
539 /* Copy the flags to each fragment. */
540 IPCB(to)->flags = IPCB(from)->flags;
542 #ifdef CONFIG_NET_SCHED
543 to->tc_index = from->tc_index;
546 #if IS_ENABLED(CONFIG_IP_VS)
547 to->ipvs_property = from->ipvs_property;
549 skb_copy_secmark(to, from);
552 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
554 int (*output)(struct net *, struct sock *, struct sk_buff *))
556 struct iphdr *iph = ip_hdr(skb);
558 if ((iph->frag_off & htons(IP_DF)) == 0)
559 return ip_do_fragment(net, sk, skb, output);
561 if (unlikely(!skb->ignore_df ||
562 (IPCB(skb)->frag_max_size &&
563 IPCB(skb)->frag_max_size > mtu))) {
564 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
565 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
571 return ip_do_fragment(net, sk, skb, output);
575 * This IP datagram is too large to be sent in one piece. Break it up into
576 * smaller pieces (each of size equal to IP header plus
577 * a block of the data of the original IP data part) that will yet fit in a
578 * single device frame, and queue such a frame for sending.
581 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
582 int (*output)(struct net *, struct sock *, struct sk_buff *))
586 struct sk_buff *skb2;
587 unsigned int mtu, hlen, left, len, ll_rs;
589 __be16 not_last_frag;
590 struct rtable *rt = skb_rtable(skb);
593 /* for offloaded checksums cleanup checksum before fragmentation */
594 if (skb->ip_summed == CHECKSUM_PARTIAL &&
595 (err = skb_checksum_help(skb)))
599 * Point into the IP datagram header.
604 mtu = ip_skb_dst_mtu(sk, skb);
605 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
606 mtu = IPCB(skb)->frag_max_size;
609 * Setup starting values.
613 mtu = mtu - hlen; /* Size of data space */
614 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
615 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
617 /* When frag_list is given, use it. First, check its validity:
618 * some transformers could create wrong frag_list or break existing
619 * one, it is not prohibited. In this case fall back to copying.
621 * LATER: this step can be merged to real generation of fragments,
622 * we can switch to copy when see the first bad fragment.
624 if (skb_has_frag_list(skb)) {
625 struct sk_buff *frag, *frag2;
626 unsigned int first_len = skb_pagelen(skb);
628 if (first_len - hlen > mtu ||
629 ((first_len - hlen) & 7) ||
630 ip_is_fragment(iph) ||
632 skb_headroom(skb) < ll_rs)
635 skb_walk_frags(skb, frag) {
636 /* Correct geometry. */
637 if (frag->len > mtu ||
638 ((frag->len & 7) && frag->next) ||
639 skb_headroom(frag) < hlen + ll_rs)
640 goto slow_path_clean;
642 /* Partially cloned skb? */
643 if (skb_shared(frag))
644 goto slow_path_clean;
649 frag->destructor = sock_wfree;
651 skb->truesize -= frag->truesize;
654 /* Everything is OK. Generate! */
658 frag = skb_shinfo(skb)->frag_list;
659 skb_frag_list_init(skb);
660 skb->data_len = first_len - skb_headlen(skb);
661 skb->len = first_len;
662 iph->tot_len = htons(first_len);
663 iph->frag_off = htons(IP_MF);
667 /* Prepare header of the next frame,
668 * before previous one went down. */
670 frag->ip_summed = CHECKSUM_NONE;
671 skb_reset_transport_header(frag);
672 __skb_push(frag, hlen);
673 skb_reset_network_header(frag);
674 memcpy(skb_network_header(frag), iph, hlen);
676 iph->tot_len = htons(frag->len);
677 ip_copy_metadata(frag, skb);
679 ip_options_fragment(frag);
680 offset += skb->len - hlen;
681 iph->frag_off = htons(offset>>3);
683 iph->frag_off |= htons(IP_MF);
684 /* Ready, complete checksum */
688 err = output(net, sk, skb);
691 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
701 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
710 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
714 skb_walk_frags(skb, frag2) {
718 frag2->destructor = NULL;
719 skb->truesize += frag2->truesize;
726 left = skb->len - hlen; /* Space per frame */
727 ptr = hlen; /* Where to start from */
730 * Fragment the datagram.
733 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
734 not_last_frag = iph->frag_off & htons(IP_MF);
737 * Keep copying data until we run out.
742 /* IF: it doesn't fit, use 'mtu' - the data space left */
745 /* IF: we are not sending up to and including the packet end
746 then align the next start on an eight byte boundary */
751 /* Allocate buffer */
752 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
759 * Set up data on packet
762 ip_copy_metadata(skb2, skb);
763 skb_reserve(skb2, ll_rs);
764 skb_put(skb2, len + hlen);
765 skb_reset_network_header(skb2);
766 skb2->transport_header = skb2->network_header + hlen;
769 * Charge the memory for the fragment to any owner
774 skb_set_owner_w(skb2, skb->sk);
777 * Copy the packet header into the new buffer.
780 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
783 * Copy a block of the IP datagram.
785 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
790 * Fill in the new header fields.
793 iph->frag_off = htons((offset >> 3));
795 if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
796 iph->frag_off |= htons(IP_DF);
798 /* ANK: dirty, but effective trick. Upgrade options only if
799 * the segment to be fragmented was THE FIRST (otherwise,
800 * options are already fixed) and make it ONCE
801 * on the initial skb, so that all the following fragments
802 * will inherit fixed options.
805 ip_options_fragment(skb);
808 * Added AC : If we are fragmenting a fragment that's not the
809 * last fragment then keep MF on each bit
811 if (left > 0 || not_last_frag)
812 iph->frag_off |= htons(IP_MF);
817 * Put this fragment into the sending queue.
819 iph->tot_len = htons(len + hlen);
823 err = output(net, sk, skb2);
827 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
830 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
835 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
838 EXPORT_SYMBOL(ip_do_fragment);
841 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
843 struct msghdr *msg = from;
845 if (skb->ip_summed == CHECKSUM_PARTIAL) {
846 if (!copy_from_iter_full(to, len, &msg->msg_iter))
850 if (!csum_and_copy_from_iter_full(to, len, &csum, &msg->msg_iter))
852 skb->csum = csum_block_add(skb->csum, csum, odd);
856 EXPORT_SYMBOL(ip_generic_getfrag);
859 csum_page(struct page *page, int offset, int copy)
864 csum = csum_partial(kaddr + offset, copy, 0);
869 static int __ip_append_data(struct sock *sk,
871 struct sk_buff_head *queue,
872 struct inet_cork *cork,
873 struct page_frag *pfrag,
874 int getfrag(void *from, char *to, int offset,
875 int len, int odd, struct sk_buff *skb),
876 void *from, int length, int transhdrlen,
879 struct inet_sock *inet = inet_sk(sk);
882 struct ip_options *opt = cork->opt;
889 unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
890 int csummode = CHECKSUM_NONE;
891 struct rtable *rt = (struct rtable *)cork->dst;
892 unsigned int wmem_alloc_delta = 0;
896 skb = skb_peek_tail(queue);
898 exthdrlen = !skb ? rt->dst.header_len : 0;
899 mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
900 paged = !!cork->gso_size;
902 if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
903 sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
904 tskey = sk->sk_tskey++;
906 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
908 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
909 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
910 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
912 if (cork->length + length > maxnonfragsize - fragheaderlen) {
913 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
914 mtu - (opt ? opt->optlen : 0));
919 * transhdrlen > 0 means that this is the first fragment and we wish
920 * it won't be fragmented in the future.
923 length + fragheaderlen <= mtu &&
924 rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
925 (!(flags & MSG_MORE) || cork->gso_size) &&
926 (!exthdrlen || (rt->dst.dev->features & NETIF_F_HW_ESP_TX_CSUM)))
927 csummode = CHECKSUM_PARTIAL;
929 cork->length += length;
931 /* So, what's going on in the loop below?
933 * We use calculated fragment length to generate chained skb,
934 * each of segments is IP fragment ready for sending to network after
935 * adding appropriate IP header.
942 /* Check if the remaining data fits into current packet. */
943 copy = mtu - skb->len;
945 copy = maxfraglen - skb->len;
948 unsigned int datalen;
949 unsigned int fraglen;
950 unsigned int fraggap;
951 unsigned int alloclen, alloc_extra;
952 unsigned int pagedlen;
953 struct sk_buff *skb_prev;
957 fraggap = skb_prev->len - maxfraglen;
962 * If remaining data exceeds the mtu,
963 * we know we need more fragment(s).
965 datalen = length + fraggap;
966 if (datalen > mtu - fragheaderlen)
967 datalen = maxfraglen - fragheaderlen;
968 fraglen = datalen + fragheaderlen;
971 alloc_extra = hh_len + 15;
972 alloc_extra += exthdrlen;
974 /* The last fragment gets additional space at tail.
975 * Note, with MSG_MORE we overallocate on fragments,
976 * because we have no idea what fragment will be
979 if (datalen == length + fraggap)
980 alloc_extra += rt->dst.trailer_len;
982 if ((flags & MSG_MORE) &&
983 !(rt->dst.dev->features&NETIF_F_SG))
986 (fraglen + alloc_extra < SKB_MAX_ALLOC ||
987 !(rt->dst.dev->features & NETIF_F_SG)))
990 alloclen = min_t(int, fraglen, MAX_HEADER);
991 pagedlen = fraglen - alloclen;
994 alloclen += alloc_extra;
997 skb = sock_alloc_send_skb(sk, alloclen,
998 (flags & MSG_DONTWAIT), &err);
1001 if (refcount_read(&sk->sk_wmem_alloc) + wmem_alloc_delta <=
1003 skb = alloc_skb(alloclen,
1012 * Fill in the control structures
1014 skb->ip_summed = csummode;
1016 skb_reserve(skb, hh_len);
1018 /* only the initial fragment is time stamped */
1019 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1021 skb_shinfo(skb)->tskey = tskey;
1025 * Find where to start putting bytes.
1027 data = skb_put(skb, fraglen + exthdrlen - pagedlen);
1028 skb_set_network_header(skb, exthdrlen);
1029 skb->transport_header = (skb->network_header +
1031 data += fragheaderlen + exthdrlen;
1034 skb->csum = skb_copy_and_csum_bits(
1035 skb_prev, maxfraglen,
1036 data + transhdrlen, fraggap, 0);
1037 skb_prev->csum = csum_sub(skb_prev->csum,
1040 pskb_trim_unique(skb_prev, maxfraglen);
1043 copy = datalen - transhdrlen - fraggap - pagedlen;
1044 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1051 length -= copy + transhdrlen;
1054 csummode = CHECKSUM_NONE;
1056 if ((flags & MSG_CONFIRM) && !skb_prev)
1057 skb_set_dst_pending_confirm(skb, 1);
1060 * Put the packet on the pending queue.
1062 if (!skb->destructor) {
1063 skb->destructor = sock_wfree;
1065 wmem_alloc_delta += skb->truesize;
1067 __skb_queue_tail(queue, skb);
1074 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1075 skb_tailroom(skb) >= copy) {
1079 if (getfrag(from, skb_put(skb, copy),
1080 offset, copy, off, skb) < 0) {
1081 __skb_trim(skb, off);
1086 int i = skb_shinfo(skb)->nr_frags;
1089 if (!sk_page_frag_refill(sk, pfrag))
1092 if (!skb_can_coalesce(skb, i, pfrag->page,
1095 if (i == MAX_SKB_FRAGS)
1098 __skb_fill_page_desc(skb, i, pfrag->page,
1100 skb_shinfo(skb)->nr_frags = ++i;
1101 get_page(pfrag->page);
1103 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1105 page_address(pfrag->page) + pfrag->offset,
1106 offset, copy, skb->len, skb) < 0)
1109 pfrag->offset += copy;
1110 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1112 skb->data_len += copy;
1113 skb->truesize += copy;
1114 wmem_alloc_delta += copy;
1120 if (wmem_alloc_delta)
1121 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1127 cork->length -= length;
1128 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1129 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1133 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1134 struct ipcm_cookie *ipc, struct rtable **rtp)
1136 struct ip_options_rcu *opt;
1144 * setup for corking.
1149 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1151 if (unlikely(!cork->opt))
1154 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1155 cork->flags |= IPCORK_OPT;
1156 cork->addr = ipc->addr;
1159 cork->fragsize = ip_sk_use_pmtu(sk) ?
1160 dst_mtu(&rt->dst) : READ_ONCE(rt->dst.dev->mtu);
1162 if (!inetdev_valid_mtu(cork->fragsize))
1163 return -ENETUNREACH;
1165 cork->gso_size = ipc->gso_size;
1167 cork->dst = &rt->dst;
1168 /* We stole this route, caller should not release it. */
1172 cork->ttl = ipc->ttl;
1173 cork->tos = ipc->tos;
1174 cork->priority = ipc->priority;
1175 cork->transmit_time = ipc->sockc.transmit_time;
1177 sock_tx_timestamp(sk, ipc->sockc.tsflags, &cork->tx_flags);
1183 * ip_append_data() and ip_append_page() can make one large IP datagram
1184 * from many pieces of data. Each pieces will be holded on the socket
1185 * until ip_push_pending_frames() is called. Each piece can be a page
1188 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1189 * this interface potentially.
1191 * LATER: length must be adjusted by pad at tail, when it is required.
1193 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1194 int getfrag(void *from, char *to, int offset, int len,
1195 int odd, struct sk_buff *skb),
1196 void *from, int length, int transhdrlen,
1197 struct ipcm_cookie *ipc, struct rtable **rtp,
1200 struct inet_sock *inet = inet_sk(sk);
1203 if (flags&MSG_PROBE)
1206 if (skb_queue_empty(&sk->sk_write_queue)) {
1207 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1214 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1215 sk_page_frag(sk), getfrag,
1216 from, length, transhdrlen, flags);
1219 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1220 int offset, size_t size, int flags)
1222 struct inet_sock *inet = inet_sk(sk);
1223 struct sk_buff *skb;
1225 struct ip_options *opt = NULL;
1226 struct inet_cork *cork;
1231 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1236 if (flags&MSG_PROBE)
1239 if (skb_queue_empty(&sk->sk_write_queue))
1242 cork = &inet->cork.base;
1243 rt = (struct rtable *)cork->dst;
1244 if (cork->flags & IPCORK_OPT)
1247 if (!(rt->dst.dev->features&NETIF_F_SG))
1250 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1251 mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
1253 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1254 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1255 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1257 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1258 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1259 mtu - (opt ? opt->optlen : 0));
1263 skb = skb_peek_tail(&sk->sk_write_queue);
1267 cork->length += size;
1270 /* Check if the remaining data fits into current packet. */
1271 len = mtu - skb->len;
1273 len = maxfraglen - skb->len;
1276 struct sk_buff *skb_prev;
1280 fraggap = skb_prev->len - maxfraglen;
1282 alloclen = fragheaderlen + hh_len + fraggap + 15;
1283 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1284 if (unlikely(!skb)) {
1290 * Fill in the control structures
1292 skb->ip_summed = CHECKSUM_NONE;
1294 skb_reserve(skb, hh_len);
1297 * Find where to start putting bytes.
1299 skb_put(skb, fragheaderlen + fraggap);
1300 skb_reset_network_header(skb);
1301 skb->transport_header = (skb->network_header +
1304 skb->csum = skb_copy_and_csum_bits(skb_prev,
1306 skb_transport_header(skb),
1308 skb_prev->csum = csum_sub(skb_prev->csum,
1310 pskb_trim_unique(skb_prev, maxfraglen);
1314 * Put the packet on the pending queue.
1316 __skb_queue_tail(&sk->sk_write_queue, skb);
1323 if (skb_append_pagefrags(skb, page, offset, len)) {
1328 if (skb->ip_summed == CHECKSUM_NONE) {
1330 csum = csum_page(page, offset, len);
1331 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1335 skb->data_len += len;
1336 skb->truesize += len;
1337 refcount_add(len, &sk->sk_wmem_alloc);
1344 cork->length -= size;
1345 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1349 static void ip_cork_release(struct inet_cork *cork)
1351 cork->flags &= ~IPCORK_OPT;
1354 dst_release(cork->dst);
1359 * Combined all pending IP fragments on the socket as one IP datagram
1360 * and push them out.
1362 struct sk_buff *__ip_make_skb(struct sock *sk,
1364 struct sk_buff_head *queue,
1365 struct inet_cork *cork)
1367 struct sk_buff *skb, *tmp_skb;
1368 struct sk_buff **tail_skb;
1369 struct inet_sock *inet = inet_sk(sk);
1370 struct net *net = sock_net(sk);
1371 struct ip_options *opt = NULL;
1372 struct rtable *rt = (struct rtable *)cork->dst;
1377 skb = __skb_dequeue(queue);
1380 tail_skb = &(skb_shinfo(skb)->frag_list);
1382 /* move skb->data to ip header from ext header */
1383 if (skb->data < skb_network_header(skb))
1384 __skb_pull(skb, skb_network_offset(skb));
1385 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1386 __skb_pull(tmp_skb, skb_network_header_len(skb));
1387 *tail_skb = tmp_skb;
1388 tail_skb = &(tmp_skb->next);
1389 skb->len += tmp_skb->len;
1390 skb->data_len += tmp_skb->len;
1391 skb->truesize += tmp_skb->truesize;
1392 tmp_skb->destructor = NULL;
1396 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1397 * to fragment the frame generated here. No matter, what transforms
1398 * how transforms change size of the packet, it will come out.
1400 skb->ignore_df = ip_sk_ignore_df(sk);
1402 /* DF bit is set when we want to see DF on outgoing frames.
1403 * If ignore_df is set too, we still allow to fragment this frame
1405 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1406 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1407 (skb->len <= dst_mtu(&rt->dst) &&
1408 ip_dont_fragment(sk, &rt->dst)))
1411 if (cork->flags & IPCORK_OPT)
1416 else if (rt->rt_type == RTN_MULTICAST)
1419 ttl = ip_select_ttl(inet, &rt->dst);
1424 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1427 iph->protocol = sk->sk_protocol;
1428 ip_copy_addrs(iph, fl4);
1429 ip_select_ident(net, skb, sk);
1432 iph->ihl += opt->optlen>>2;
1433 ip_options_build(skb, opt, cork->addr, rt, 0);
1436 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1437 skb->mark = sk->sk_mark;
1438 skb->tstamp = cork->transmit_time;
1440 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1444 skb_dst_set(skb, &rt->dst);
1446 if (iph->protocol == IPPROTO_ICMP) {
1449 /* For such sockets, transhdrlen is zero when do ip_append_data(),
1450 * so icmphdr does not in skb linear region and can not get icmp_type
1451 * by icmp_hdr(skb)->type.
1453 if (sk->sk_type == SOCK_RAW && !inet_sk(sk)->hdrincl)
1454 icmp_type = fl4->fl4_icmp_type;
1456 icmp_type = icmp_hdr(skb)->type;
1457 icmp_out_count(net, icmp_type);
1460 ip_cork_release(cork);
1465 int ip_send_skb(struct net *net, struct sk_buff *skb)
1469 err = ip_local_out(net, skb->sk, skb);
1472 err = net_xmit_errno(err);
1474 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1480 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1482 struct sk_buff *skb;
1484 skb = ip_finish_skb(sk, fl4);
1488 /* Netfilter gets whole the not fragmented skb. */
1489 return ip_send_skb(sock_net(sk), skb);
1493 * Throw away all pending data on the socket.
1495 static void __ip_flush_pending_frames(struct sock *sk,
1496 struct sk_buff_head *queue,
1497 struct inet_cork *cork)
1499 struct sk_buff *skb;
1501 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1504 ip_cork_release(cork);
1507 void ip_flush_pending_frames(struct sock *sk)
1509 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1512 struct sk_buff *ip_make_skb(struct sock *sk,
1514 int getfrag(void *from, char *to, int offset,
1515 int len, int odd, struct sk_buff *skb),
1516 void *from, int length, int transhdrlen,
1517 struct ipcm_cookie *ipc, struct rtable **rtp,
1518 struct inet_cork *cork, unsigned int flags)
1520 struct sk_buff_head queue;
1523 if (flags & MSG_PROBE)
1526 __skb_queue_head_init(&queue);
1531 err = ip_setup_cork(sk, cork, ipc, rtp);
1533 return ERR_PTR(err);
1535 err = __ip_append_data(sk, fl4, &queue, cork,
1536 ¤t->task_frag, getfrag,
1537 from, length, transhdrlen, flags);
1539 __ip_flush_pending_frames(sk, &queue, cork);
1540 return ERR_PTR(err);
1543 return __ip_make_skb(sk, fl4, &queue, cork);
1547 * Fetch data from kernel space and fill in checksum if needed.
1549 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1550 int len, int odd, struct sk_buff *skb)
1554 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1555 skb->csum = csum_block_add(skb->csum, csum, odd);
1560 * Generic function to send a packet as reply to another packet.
1561 * Used to send some TCP resets/acks so far.
1563 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1564 const struct ip_options *sopt,
1565 __be32 daddr, __be32 saddr,
1566 const struct ip_reply_arg *arg,
1569 struct ip_options_data replyopts;
1570 struct ipcm_cookie ipc;
1572 struct rtable *rt = skb_rtable(skb);
1573 struct net *net = sock_net(sk);
1574 struct sk_buff *nskb;
1578 if (__ip_options_echo(net, &replyopts.opt.opt, skb, sopt))
1584 if (replyopts.opt.opt.optlen) {
1585 ipc.opt = &replyopts.opt;
1587 if (replyopts.opt.opt.srr)
1588 daddr = replyopts.opt.opt.faddr;
1591 oif = arg->bound_dev_if;
1592 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1595 flowi4_init_output(&fl4, oif,
1596 IP4_REPLY_MARK(net, skb->mark) ?: sk->sk_mark,
1598 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1599 ip_reply_arg_flowi_flags(arg),
1601 tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1603 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1604 rt = ip_route_output_key(net, &fl4);
1608 inet_sk(sk)->tos = arg->tos & ~INET_ECN_MASK;
1610 sk->sk_priority = skb->priority;
1611 sk->sk_protocol = ip_hdr(skb)->protocol;
1612 sk->sk_bound_dev_if = arg->bound_dev_if;
1613 sk->sk_sndbuf = sysctl_wmem_default;
1614 sk->sk_mark = fl4.flowi4_mark;
1615 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1616 len, 0, &ipc, &rt, MSG_DONTWAIT);
1617 if (unlikely(err)) {
1618 ip_flush_pending_frames(sk);
1622 nskb = skb_peek(&sk->sk_write_queue);
1624 if (arg->csumoffset >= 0)
1625 *((__sum16 *)skb_transport_header(nskb) +
1626 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1628 nskb->ip_summed = CHECKSUM_NONE;
1629 ip_push_pending_frames(sk, &fl4);
1635 void __init ip_init(void)
1640 #if defined(CONFIG_IP_MULTICAST)