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)
295 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
301 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
302 /* Policy lookup after SNAT yielded a new policy */
303 if (skb_dst(skb)->xfrm) {
304 IPCB(skb)->flags |= IPSKB_REROUTED;
305 return dst_output(net, sk, skb);
308 mtu = ip_skb_dst_mtu(sk, skb);
310 return ip_finish_output_gso(net, sk, skb, mtu);
312 if (skb->len > mtu || (IPCB(skb)->flags & IPSKB_FRAG_PMTU))
313 return ip_fragment(net, sk, skb, mtu, ip_finish_output2);
315 return ip_finish_output2(net, sk, skb);
318 static int ip_mc_finish_output(struct net *net, struct sock *sk,
321 struct rtable *new_rt;
324 ret = BPF_CGROUP_RUN_PROG_INET_EGRESS(sk, skb);
330 /* Reset rt_iif so that inet_iif() will return skb->skb_iif. Setting
331 * this to non-zero causes ipi_ifindex in in_pktinfo to be overwritten,
332 * see ipv4_pktinfo_prepare().
334 new_rt = rt_dst_clone(net->loopback_dev, skb_rtable(skb));
338 skb_dst_set(skb, &new_rt->dst);
341 return dev_loopback_xmit(net, sk, skb);
344 int ip_mc_output(struct net *net, struct sock *sk, struct sk_buff *skb)
346 struct rtable *rt = skb_rtable(skb);
347 struct net_device *dev = rt->dst.dev;
350 * If the indicated interface is up and running, send the packet.
352 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
355 skb->protocol = htons(ETH_P_IP);
358 * Multicasts are looped back for other local users
361 if (rt->rt_flags&RTCF_MULTICAST) {
363 #ifdef CONFIG_IP_MROUTE
364 /* Small optimization: do not loopback not local frames,
365 which returned after forwarding; they will be dropped
366 by ip_mr_input in any case.
367 Note, that local frames are looped back to be delivered
370 This check is duplicated in ip_mr_input at the moment.
373 ((rt->rt_flags & RTCF_LOCAL) ||
374 !(IPCB(skb)->flags & IPSKB_FORWARDED))
377 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
379 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
380 net, sk, newskb, NULL, newskb->dev,
381 ip_mc_finish_output);
384 /* Multicasts with ttl 0 must not go beyond the host */
386 if (ip_hdr(skb)->ttl == 0) {
392 if (rt->rt_flags&RTCF_BROADCAST) {
393 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
395 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
396 net, sk, newskb, NULL, newskb->dev,
397 ip_mc_finish_output);
400 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
401 net, sk, skb, NULL, skb->dev,
403 !(IPCB(skb)->flags & IPSKB_REROUTED));
406 int ip_output(struct net *net, struct sock *sk, struct sk_buff *skb)
408 struct net_device *dev = skb_dst(skb)->dev;
410 IP_UPD_PO_STATS(net, IPSTATS_MIB_OUT, skb->len);
413 skb->protocol = htons(ETH_P_IP);
415 return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING,
416 net, sk, skb, NULL, dev,
418 !(IPCB(skb)->flags & IPSKB_REROUTED));
422 * copy saddr and daddr, possibly using 64bit load/stores
424 * iph->saddr = fl4->saddr;
425 * iph->daddr = fl4->daddr;
427 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
429 BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
430 offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
431 memcpy(&iph->saddr, &fl4->saddr,
432 sizeof(fl4->saddr) + sizeof(fl4->daddr));
435 /* Note: skb->sk can be different from sk, in case of tunnels */
436 int __ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl,
439 struct inet_sock *inet = inet_sk(sk);
440 struct net *net = sock_net(sk);
441 struct ip_options_rcu *inet_opt;
447 /* Skip all of this if the packet is already routed,
448 * f.e. by something like SCTP.
451 inet_opt = rcu_dereference(inet->inet_opt);
453 rt = skb_rtable(skb);
457 /* Make sure we can route this packet. */
458 rt = (struct rtable *)__sk_dst_check(sk, 0);
462 /* Use correct destination address if we have options. */
463 daddr = inet->inet_daddr;
464 if (inet_opt && inet_opt->opt.srr)
465 daddr = inet_opt->opt.faddr;
467 /* If this fails, retransmit mechanism of transport layer will
468 * keep trying until route appears or the connection times
471 rt = ip_route_output_ports(net, fl4, sk,
472 daddr, inet->inet_saddr,
476 RT_CONN_FLAGS_TOS(sk, tos),
477 sk->sk_bound_dev_if);
480 sk_setup_caps(sk, &rt->dst);
482 skb_dst_set_noref(skb, &rt->dst);
485 if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_gw_family)
488 /* OK, we know where to send it, allocate and build IP header. */
489 skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
490 skb_reset_network_header(skb);
492 *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (tos & 0xff));
493 if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
494 iph->frag_off = htons(IP_DF);
497 iph->ttl = ip_select_ttl(inet, &rt->dst);
498 iph->protocol = sk->sk_protocol;
499 ip_copy_addrs(iph, fl4);
501 /* Transport layer set skb->h.foo itself. */
503 if (inet_opt && inet_opt->opt.optlen) {
504 iph->ihl += inet_opt->opt.optlen >> 2;
505 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
508 ip_select_ident_segs(net, skb, sk,
509 skb_shinfo(skb)->gso_segs ?: 1);
511 /* TODO : should we use skb->sk here instead of sk ? */
512 skb->priority = sk->sk_priority;
513 skb->mark = sk->sk_mark;
515 res = ip_local_out(net, sk, skb);
521 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
523 return -EHOSTUNREACH;
525 EXPORT_SYMBOL(__ip_queue_xmit);
527 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
529 to->pkt_type = from->pkt_type;
530 to->priority = from->priority;
531 to->protocol = from->protocol;
532 to->skb_iif = from->skb_iif;
534 skb_dst_copy(to, from);
536 to->mark = from->mark;
538 skb_copy_hash(to, from);
540 /* Copy the flags to each fragment. */
541 IPCB(to)->flags = IPCB(from)->flags;
543 #ifdef CONFIG_NET_SCHED
544 to->tc_index = from->tc_index;
547 skb_ext_copy(to, from);
548 #if IS_ENABLED(CONFIG_IP_VS)
549 to->ipvs_property = from->ipvs_property;
551 skb_copy_secmark(to, from);
554 static int ip_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
556 int (*output)(struct net *, struct sock *, struct sk_buff *))
558 struct iphdr *iph = ip_hdr(skb);
560 if ((iph->frag_off & htons(IP_DF)) == 0)
561 return ip_do_fragment(net, sk, skb, output);
563 if (unlikely(!skb->ignore_df ||
564 (IPCB(skb)->frag_max_size &&
565 IPCB(skb)->frag_max_size > mtu))) {
566 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
567 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
573 return ip_do_fragment(net, sk, skb, output);
577 * This IP datagram is too large to be sent in one piece. Break it up into
578 * smaller pieces (each of size equal to IP header plus
579 * a block of the data of the original IP data part) that will yet fit in a
580 * single device frame, and queue such a frame for sending.
583 int ip_do_fragment(struct net *net, struct sock *sk, struct sk_buff *skb,
584 int (*output)(struct net *, struct sock *, struct sk_buff *))
588 struct sk_buff *skb2;
589 unsigned int mtu, hlen, left, len, ll_rs;
591 __be16 not_last_frag;
592 struct rtable *rt = skb_rtable(skb);
595 /* for offloaded checksums cleanup checksum before fragmentation */
596 if (skb->ip_summed == CHECKSUM_PARTIAL &&
597 (err = skb_checksum_help(skb)))
601 * Point into the IP datagram header.
606 mtu = ip_skb_dst_mtu(sk, skb);
607 if (IPCB(skb)->frag_max_size && IPCB(skb)->frag_max_size < mtu)
608 mtu = IPCB(skb)->frag_max_size;
611 * Setup starting values.
615 mtu = mtu - hlen; /* Size of data space */
616 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
617 ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
619 /* When frag_list is given, use it. First, check its validity:
620 * some transformers could create wrong frag_list or break existing
621 * one, it is not prohibited. In this case fall back to copying.
623 * LATER: this step can be merged to real generation of fragments,
624 * we can switch to copy when see the first bad fragment.
626 if (skb_has_frag_list(skb)) {
627 struct sk_buff *frag, *frag2;
628 unsigned int first_len = skb_pagelen(skb);
630 if (first_len - hlen > mtu ||
631 ((first_len - hlen) & 7) ||
632 ip_is_fragment(iph) ||
634 skb_headroom(skb) < ll_rs)
637 skb_walk_frags(skb, frag) {
638 /* Correct geometry. */
639 if (frag->len > mtu ||
640 ((frag->len & 7) && frag->next) ||
641 skb_headroom(frag) < hlen + ll_rs)
642 goto slow_path_clean;
644 /* Partially cloned skb? */
645 if (skb_shared(frag))
646 goto slow_path_clean;
651 frag->destructor = sock_wfree;
653 skb->truesize -= frag->truesize;
656 /* Everything is OK. Generate! */
660 frag = skb_shinfo(skb)->frag_list;
661 skb_frag_list_init(skb);
662 skb->data_len = first_len - skb_headlen(skb);
663 skb->len = first_len;
664 iph->tot_len = htons(first_len);
665 iph->frag_off = htons(IP_MF);
669 /* Prepare header of the next frame,
670 * before previous one went down. */
672 frag->ip_summed = CHECKSUM_NONE;
673 skb_reset_transport_header(frag);
674 __skb_push(frag, hlen);
675 skb_reset_network_header(frag);
676 memcpy(skb_network_header(frag), iph, hlen);
678 iph->tot_len = htons(frag->len);
679 ip_copy_metadata(frag, skb);
681 ip_options_fragment(frag);
682 offset += skb->len - hlen;
683 iph->frag_off = htons(offset>>3);
685 iph->frag_off |= htons(IP_MF);
686 /* Ready, complete checksum */
690 err = output(net, sk, skb);
693 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
699 skb_mark_not_on_list(skb);
703 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
707 kfree_skb_list(frag);
709 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
713 skb_walk_frags(skb, frag2) {
717 frag2->destructor = NULL;
718 skb->truesize += frag2->truesize;
725 left = skb->len - hlen; /* Space per frame */
726 ptr = hlen; /* Where to start from */
729 * Fragment the datagram.
732 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
733 not_last_frag = iph->frag_off & htons(IP_MF);
736 * Keep copying data until we run out.
741 /* IF: it doesn't fit, use 'mtu' - the data space left */
744 /* IF: we are not sending up to and including the packet end
745 then align the next start on an eight byte boundary */
750 /* Allocate buffer */
751 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
758 * Set up data on packet
761 ip_copy_metadata(skb2, skb);
762 skb_reserve(skb2, ll_rs);
763 skb_put(skb2, len + hlen);
764 skb_reset_network_header(skb2);
765 skb2->transport_header = skb2->network_header + hlen;
768 * Charge the memory for the fragment to any owner
773 skb_set_owner_w(skb2, skb->sk);
776 * Copy the packet header into the new buffer.
779 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
782 * Copy a block of the IP datagram.
784 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
789 * Fill in the new header fields.
792 iph->frag_off = htons((offset >> 3));
794 if (IPCB(skb)->flags & IPSKB_FRAG_PMTU)
795 iph->frag_off |= htons(IP_DF);
797 /* ANK: dirty, but effective trick. Upgrade options only if
798 * the segment to be fragmented was THE FIRST (otherwise,
799 * options are already fixed) and make it ONCE
800 * on the initial skb, so that all the following fragments
801 * will inherit fixed options.
804 ip_options_fragment(skb);
807 * Added AC : If we are fragmenting a fragment that's not the
808 * last fragment then keep MF on each bit
810 if (left > 0 || not_last_frag)
811 iph->frag_off |= htons(IP_MF);
816 * Put this fragment into the sending queue.
818 iph->tot_len = htons(len + hlen);
822 err = output(net, sk, skb2);
826 IP_INC_STATS(net, IPSTATS_MIB_FRAGCREATES);
829 IP_INC_STATS(net, IPSTATS_MIB_FRAGOKS);
834 IP_INC_STATS(net, IPSTATS_MIB_FRAGFAILS);
837 EXPORT_SYMBOL(ip_do_fragment);
840 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
842 struct msghdr *msg = from;
844 if (skb->ip_summed == CHECKSUM_PARTIAL) {
845 if (!copy_from_iter_full(to, len, &msg->msg_iter))
849 if (!csum_and_copy_from_iter_full(to, len, &csum, &msg->msg_iter))
851 skb->csum = csum_block_add(skb->csum, csum, odd);
855 EXPORT_SYMBOL(ip_generic_getfrag);
858 csum_page(struct page *page, int offset, int copy)
863 csum = csum_partial(kaddr + offset, copy, 0);
868 static int __ip_append_data(struct sock *sk,
870 struct sk_buff_head *queue,
871 struct inet_cork *cork,
872 struct page_frag *pfrag,
873 int getfrag(void *from, char *to, int offset,
874 int len, int odd, struct sk_buff *skb),
875 void *from, int length, int transhdrlen,
878 struct inet_sock *inet = inet_sk(sk);
879 struct ubuf_info *uarg = NULL;
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;
893 bool paged, extra_uref = false;
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 if (flags & MSG_ZEROCOPY && length && sock_flag(sk, SOCK_ZEROCOPY)) {
930 uarg = sock_zerocopy_realloc(sk, length, skb_zcopy(skb));
933 extra_uref = !skb_zcopy(skb); /* only ref on new uarg */
934 if (rt->dst.dev->features & NETIF_F_SG &&
935 csummode == CHECKSUM_PARTIAL) {
939 skb_zcopy_set(skb, uarg, &extra_uref);
943 cork->length += length;
945 /* So, what's going on in the loop below?
947 * We use calculated fragment length to generate chained skb,
948 * each of segments is IP fragment ready for sending to network after
949 * adding appropriate IP header.
956 /* Check if the remaining data fits into current packet. */
957 copy = mtu - skb->len;
959 copy = maxfraglen - skb->len;
962 unsigned int datalen;
963 unsigned int fraglen;
964 unsigned int fraggap;
965 unsigned int alloclen;
966 unsigned int pagedlen;
967 struct sk_buff *skb_prev;
971 fraggap = skb_prev->len - maxfraglen;
976 * If remaining data exceeds the mtu,
977 * we know we need more fragment(s).
979 datalen = length + fraggap;
980 if (datalen > mtu - fragheaderlen)
981 datalen = maxfraglen - fragheaderlen;
982 fraglen = datalen + fragheaderlen;
985 if ((flags & MSG_MORE) &&
986 !(rt->dst.dev->features&NETIF_F_SG))
991 alloclen = min_t(int, fraglen, MAX_HEADER);
992 pagedlen = fraglen - alloclen;
995 alloclen += exthdrlen;
997 /* The last fragment gets additional space at tail.
998 * Note, with MSG_MORE we overallocate on fragments,
999 * because we have no idea what fragment will be
1002 if (datalen == length + fraggap)
1003 alloclen += rt->dst.trailer_len;
1006 skb = sock_alloc_send_skb(sk,
1007 alloclen + hh_len + 15,
1008 (flags & MSG_DONTWAIT), &err);
1011 if (refcount_read(&sk->sk_wmem_alloc) + wmem_alloc_delta <=
1013 skb = alloc_skb(alloclen + hh_len + 15,
1022 * Fill in the control structures
1024 skb->ip_summed = csummode;
1026 skb_reserve(skb, hh_len);
1029 * Find where to start putting bytes.
1031 data = skb_put(skb, fraglen + exthdrlen - pagedlen);
1032 skb_set_network_header(skb, exthdrlen);
1033 skb->transport_header = (skb->network_header +
1035 data += fragheaderlen + exthdrlen;
1038 skb->csum = skb_copy_and_csum_bits(
1039 skb_prev, maxfraglen,
1040 data + transhdrlen, fraggap, 0);
1041 skb_prev->csum = csum_sub(skb_prev->csum,
1044 pskb_trim_unique(skb_prev, maxfraglen);
1047 copy = datalen - transhdrlen - fraggap - pagedlen;
1048 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1055 length -= copy + transhdrlen;
1058 csummode = CHECKSUM_NONE;
1060 /* only the initial fragment is time stamped */
1061 skb_shinfo(skb)->tx_flags = cork->tx_flags;
1063 skb_shinfo(skb)->tskey = tskey;
1065 skb_zcopy_set(skb, uarg, &extra_uref);
1067 if ((flags & MSG_CONFIRM) && !skb_prev)
1068 skb_set_dst_pending_confirm(skb, 1);
1071 * Put the packet on the pending queue.
1073 if (!skb->destructor) {
1074 skb->destructor = sock_wfree;
1076 wmem_alloc_delta += skb->truesize;
1078 __skb_queue_tail(queue, skb);
1085 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1086 skb_tailroom(skb) >= copy) {
1090 if (getfrag(from, skb_put(skb, copy),
1091 offset, copy, off, skb) < 0) {
1092 __skb_trim(skb, off);
1096 } else if (!uarg || !uarg->zerocopy) {
1097 int i = skb_shinfo(skb)->nr_frags;
1100 if (!sk_page_frag_refill(sk, pfrag))
1103 if (!skb_can_coalesce(skb, i, pfrag->page,
1106 if (i == MAX_SKB_FRAGS)
1109 __skb_fill_page_desc(skb, i, pfrag->page,
1111 skb_shinfo(skb)->nr_frags = ++i;
1112 get_page(pfrag->page);
1114 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1116 page_address(pfrag->page) + pfrag->offset,
1117 offset, copy, skb->len, skb) < 0)
1120 pfrag->offset += copy;
1121 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1123 skb->data_len += copy;
1124 skb->truesize += copy;
1125 wmem_alloc_delta += copy;
1127 err = skb_zerocopy_iter_dgram(skb, from, copy);
1135 if (wmem_alloc_delta)
1136 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1143 sock_zerocopy_put_abort(uarg, extra_uref);
1144 cork->length -= length;
1145 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1146 refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
1150 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1151 struct ipcm_cookie *ipc, struct rtable **rtp)
1153 struct ip_options_rcu *opt;
1161 * setup for corking.
1166 cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1168 if (unlikely(!cork->opt))
1171 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1172 cork->flags |= IPCORK_OPT;
1173 cork->addr = ipc->addr;
1177 * We steal reference to this route, caller should not release it
1180 cork->fragsize = ip_sk_use_pmtu(sk) ?
1181 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1183 cork->gso_size = ipc->gso_size;
1184 cork->dst = &rt->dst;
1186 cork->ttl = ipc->ttl;
1187 cork->tos = ipc->tos;
1188 cork->priority = ipc->priority;
1189 cork->transmit_time = ipc->sockc.transmit_time;
1191 sock_tx_timestamp(sk, ipc->sockc.tsflags, &cork->tx_flags);
1197 * ip_append_data() and ip_append_page() can make one large IP datagram
1198 * from many pieces of data. Each pieces will be holded on the socket
1199 * until ip_push_pending_frames() is called. Each piece can be a page
1202 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1203 * this interface potentially.
1205 * LATER: length must be adjusted by pad at tail, when it is required.
1207 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1208 int getfrag(void *from, char *to, int offset, int len,
1209 int odd, struct sk_buff *skb),
1210 void *from, int length, int transhdrlen,
1211 struct ipcm_cookie *ipc, struct rtable **rtp,
1214 struct inet_sock *inet = inet_sk(sk);
1217 if (flags&MSG_PROBE)
1220 if (skb_queue_empty(&sk->sk_write_queue)) {
1221 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1228 return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1229 sk_page_frag(sk), getfrag,
1230 from, length, transhdrlen, flags);
1233 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1234 int offset, size_t size, int flags)
1236 struct inet_sock *inet = inet_sk(sk);
1237 struct sk_buff *skb;
1239 struct ip_options *opt = NULL;
1240 struct inet_cork *cork;
1245 unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1250 if (flags&MSG_PROBE)
1253 if (skb_queue_empty(&sk->sk_write_queue))
1256 cork = &inet->cork.base;
1257 rt = (struct rtable *)cork->dst;
1258 if (cork->flags & IPCORK_OPT)
1261 if (!(rt->dst.dev->features&NETIF_F_SG))
1264 hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1265 mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
1267 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1268 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1269 maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1271 if (cork->length + size > maxnonfragsize - fragheaderlen) {
1272 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1273 mtu - (opt ? opt->optlen : 0));
1277 skb = skb_peek_tail(&sk->sk_write_queue);
1281 cork->length += size;
1284 /* Check if the remaining data fits into current packet. */
1285 len = mtu - skb->len;
1287 len = maxfraglen - skb->len;
1290 struct sk_buff *skb_prev;
1294 fraggap = skb_prev->len - maxfraglen;
1296 alloclen = fragheaderlen + hh_len + fraggap + 15;
1297 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1298 if (unlikely(!skb)) {
1304 * Fill in the control structures
1306 skb->ip_summed = CHECKSUM_NONE;
1308 skb_reserve(skb, hh_len);
1311 * Find where to start putting bytes.
1313 skb_put(skb, fragheaderlen + fraggap);
1314 skb_reset_network_header(skb);
1315 skb->transport_header = (skb->network_header +
1318 skb->csum = skb_copy_and_csum_bits(skb_prev,
1320 skb_transport_header(skb),
1322 skb_prev->csum = csum_sub(skb_prev->csum,
1324 pskb_trim_unique(skb_prev, maxfraglen);
1328 * Put the packet on the pending queue.
1330 __skb_queue_tail(&sk->sk_write_queue, skb);
1337 if (skb_append_pagefrags(skb, page, offset, len)) {
1342 if (skb->ip_summed == CHECKSUM_NONE) {
1344 csum = csum_page(page, offset, len);
1345 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1349 skb->data_len += len;
1350 skb->truesize += len;
1351 refcount_add(len, &sk->sk_wmem_alloc);
1358 cork->length -= size;
1359 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1363 static void ip_cork_release(struct inet_cork *cork)
1365 cork->flags &= ~IPCORK_OPT;
1368 dst_release(cork->dst);
1373 * Combined all pending IP fragments on the socket as one IP datagram
1374 * and push them out.
1376 struct sk_buff *__ip_make_skb(struct sock *sk,
1378 struct sk_buff_head *queue,
1379 struct inet_cork *cork)
1381 struct sk_buff *skb, *tmp_skb;
1382 struct sk_buff **tail_skb;
1383 struct inet_sock *inet = inet_sk(sk);
1384 struct net *net = sock_net(sk);
1385 struct ip_options *opt = NULL;
1386 struct rtable *rt = (struct rtable *)cork->dst;
1391 skb = __skb_dequeue(queue);
1394 tail_skb = &(skb_shinfo(skb)->frag_list);
1396 /* move skb->data to ip header from ext header */
1397 if (skb->data < skb_network_header(skb))
1398 __skb_pull(skb, skb_network_offset(skb));
1399 while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1400 __skb_pull(tmp_skb, skb_network_header_len(skb));
1401 *tail_skb = tmp_skb;
1402 tail_skb = &(tmp_skb->next);
1403 skb->len += tmp_skb->len;
1404 skb->data_len += tmp_skb->len;
1405 skb->truesize += tmp_skb->truesize;
1406 tmp_skb->destructor = NULL;
1410 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1411 * to fragment the frame generated here. No matter, what transforms
1412 * how transforms change size of the packet, it will come out.
1414 skb->ignore_df = ip_sk_ignore_df(sk);
1416 /* DF bit is set when we want to see DF on outgoing frames.
1417 * If ignore_df is set too, we still allow to fragment this frame
1419 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1420 inet->pmtudisc == IP_PMTUDISC_PROBE ||
1421 (skb->len <= dst_mtu(&rt->dst) &&
1422 ip_dont_fragment(sk, &rt->dst)))
1425 if (cork->flags & IPCORK_OPT)
1430 else if (rt->rt_type == RTN_MULTICAST)
1433 ttl = ip_select_ttl(inet, &rt->dst);
1438 iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1441 iph->protocol = sk->sk_protocol;
1442 ip_copy_addrs(iph, fl4);
1443 ip_select_ident(net, skb, sk);
1446 iph->ihl += opt->optlen>>2;
1447 ip_options_build(skb, opt, cork->addr, rt, 0);
1450 skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1451 skb->mark = sk->sk_mark;
1452 skb->tstamp = cork->transmit_time;
1454 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1458 skb_dst_set(skb, &rt->dst);
1460 if (iph->protocol == IPPROTO_ICMP)
1461 icmp_out_count(net, ((struct icmphdr *)
1462 skb_transport_header(skb))->type);
1464 ip_cork_release(cork);
1469 int ip_send_skb(struct net *net, struct sk_buff *skb)
1473 err = ip_local_out(net, skb->sk, skb);
1476 err = net_xmit_errno(err);
1478 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1484 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1486 struct sk_buff *skb;
1488 skb = ip_finish_skb(sk, fl4);
1492 /* Netfilter gets whole the not fragmented skb. */
1493 return ip_send_skb(sock_net(sk), skb);
1497 * Throw away all pending data on the socket.
1499 static void __ip_flush_pending_frames(struct sock *sk,
1500 struct sk_buff_head *queue,
1501 struct inet_cork *cork)
1503 struct sk_buff *skb;
1505 while ((skb = __skb_dequeue_tail(queue)) != NULL)
1508 ip_cork_release(cork);
1511 void ip_flush_pending_frames(struct sock *sk)
1513 __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1516 struct sk_buff *ip_make_skb(struct sock *sk,
1518 int getfrag(void *from, char *to, int offset,
1519 int len, int odd, struct sk_buff *skb),
1520 void *from, int length, int transhdrlen,
1521 struct ipcm_cookie *ipc, struct rtable **rtp,
1522 struct inet_cork *cork, unsigned int flags)
1524 struct sk_buff_head queue;
1527 if (flags & MSG_PROBE)
1530 __skb_queue_head_init(&queue);
1535 err = ip_setup_cork(sk, cork, ipc, rtp);
1537 return ERR_PTR(err);
1539 err = __ip_append_data(sk, fl4, &queue, cork,
1540 ¤t->task_frag, getfrag,
1541 from, length, transhdrlen, flags);
1543 __ip_flush_pending_frames(sk, &queue, cork);
1544 return ERR_PTR(err);
1547 return __ip_make_skb(sk, fl4, &queue, cork);
1551 * Fetch data from kernel space and fill in checksum if needed.
1553 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1554 int len, int odd, struct sk_buff *skb)
1558 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1559 skb->csum = csum_block_add(skb->csum, csum, odd);
1564 * Generic function to send a packet as reply to another packet.
1565 * Used to send some TCP resets/acks so far.
1567 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1568 const struct ip_options *sopt,
1569 __be32 daddr, __be32 saddr,
1570 const struct ip_reply_arg *arg,
1573 struct ip_options_data replyopts;
1574 struct ipcm_cookie ipc;
1576 struct rtable *rt = skb_rtable(skb);
1577 struct net *net = sock_net(sk);
1578 struct sk_buff *nskb;
1582 if (__ip_options_echo(net, &replyopts.opt.opt, skb, sopt))
1588 if (replyopts.opt.opt.optlen) {
1589 ipc.opt = &replyopts.opt;
1591 if (replyopts.opt.opt.srr)
1592 daddr = replyopts.opt.opt.faddr;
1595 oif = arg->bound_dev_if;
1596 if (!oif && netif_index_is_l3_master(net, skb->skb_iif))
1599 flowi4_init_output(&fl4, oif,
1600 IP4_REPLY_MARK(net, skb->mark) ?: sk->sk_mark,
1602 RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1603 ip_reply_arg_flowi_flags(arg),
1605 tcp_hdr(skb)->source, tcp_hdr(skb)->dest,
1607 security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1608 rt = ip_route_output_key(net, &fl4);
1612 inet_sk(sk)->tos = arg->tos;
1614 sk->sk_priority = skb->priority;
1615 sk->sk_protocol = ip_hdr(skb)->protocol;
1616 sk->sk_bound_dev_if = arg->bound_dev_if;
1617 sk->sk_sndbuf = sysctl_wmem_default;
1618 sk->sk_mark = fl4.flowi4_mark;
1619 err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1620 len, 0, &ipc, &rt, MSG_DONTWAIT);
1621 if (unlikely(err)) {
1622 ip_flush_pending_frames(sk);
1626 nskb = skb_peek(&sk->sk_write_queue);
1628 if (arg->csumoffset >= 0)
1629 *((__sum16 *)skb_transport_header(nskb) +
1630 arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1632 nskb->ip_summed = CHECKSUM_NONE;
1633 ip_push_pending_frames(sk, &fl4);
1639 void __init ip_init(void)
1644 #if defined(CONFIG_IP_MULTICAST)