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 <asm/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 <linux/igmp.h>
77 #include <linux/netfilter_ipv4.h>
78 #include <linux/netfilter_bridge.h>
79 #include <linux/mroute.h>
80 #include <linux/netlink.h>
81 #include <linux/tcp.h>
83 int sysctl_ip_default_ttl __read_mostly
= IPDEFTTL
;
84 EXPORT_SYMBOL(sysctl_ip_default_ttl
);
86 static int ip_fragment(struct sock
*sk
, struct sk_buff
*skb
,
88 int (*output
)(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 static int __ip_local_out_sk(struct sock
*sk
, struct sk_buff
*skb
)
100 struct net
*net
= dev_net(skb_dst(skb
)->dev
);
101 struct iphdr
*iph
= ip_hdr(skb
);
103 iph
->tot_len
= htons(skb
->len
);
105 return nf_hook(NFPROTO_IPV4
, NF_INET_LOCAL_OUT
,
106 net
, sk
, skb
, NULL
, skb_dst(skb
)->dev
,
110 int __ip_local_out(struct sk_buff
*skb
)
112 return __ip_local_out_sk(skb
->sk
, skb
);
115 int ip_local_out_sk(struct sock
*sk
, struct sk_buff
*skb
)
119 err
= __ip_local_out(skb
);
120 if (likely(err
== 1))
121 err
= dst_output(sk
, skb
);
125 EXPORT_SYMBOL_GPL(ip_local_out_sk
);
127 static inline int ip_select_ttl(struct inet_sock
*inet
, struct dst_entry
*dst
)
129 int ttl
= inet
->uc_ttl
;
132 ttl
= ip4_dst_hoplimit(dst
);
137 * Add an ip header to a skbuff and send it out.
140 int ip_build_and_send_pkt(struct sk_buff
*skb
, const struct sock
*sk
,
141 __be32 saddr
, __be32 daddr
, struct ip_options_rcu
*opt
)
143 struct inet_sock
*inet
= inet_sk(sk
);
144 struct rtable
*rt
= skb_rtable(skb
);
147 /* Build the IP header. */
148 skb_push(skb
, sizeof(struct iphdr
) + (opt
? opt
->opt
.optlen
: 0));
149 skb_reset_network_header(skb
);
153 iph
->tos
= inet
->tos
;
154 iph
->ttl
= ip_select_ttl(inet
, &rt
->dst
);
155 iph
->daddr
= (opt
&& opt
->opt
.srr
? opt
->opt
.faddr
: daddr
);
157 iph
->protocol
= sk
->sk_protocol
;
158 if (ip_dont_fragment(sk
, &rt
->dst
)) {
159 iph
->frag_off
= htons(IP_DF
);
163 __ip_select_ident(sock_net(sk
), iph
, 1);
166 if (opt
&& opt
->opt
.optlen
) {
167 iph
->ihl
+= opt
->opt
.optlen
>>2;
168 ip_options_build(skb
, &opt
->opt
, daddr
, rt
, 0);
171 skb
->priority
= sk
->sk_priority
;
172 skb
->mark
= sk
->sk_mark
;
175 return ip_local_out(skb
);
177 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt
);
179 static int ip_finish_output2(struct sock
*sk
, struct sk_buff
*skb
)
181 struct dst_entry
*dst
= skb_dst(skb
);
182 struct rtable
*rt
= (struct rtable
*)dst
;
183 struct net_device
*dev
= dst
->dev
;
184 struct net
*net
= dev_net(dev
);
185 unsigned int hh_len
= LL_RESERVED_SPACE(dev
);
186 struct neighbour
*neigh
;
189 if (rt
->rt_type
== RTN_MULTICAST
) {
190 IP_UPD_PO_STATS(net
, IPSTATS_MIB_OUTMCAST
, skb
->len
);
191 } else if (rt
->rt_type
== RTN_BROADCAST
)
192 IP_UPD_PO_STATS(net
, IPSTATS_MIB_OUTBCAST
, skb
->len
);
194 /* Be paranoid, rather than too clever. */
195 if (unlikely(skb_headroom(skb
) < hh_len
&& dev
->header_ops
)) {
196 struct sk_buff
*skb2
;
198 skb2
= skb_realloc_headroom(skb
, LL_RESERVED_SPACE(dev
));
204 skb_set_owner_w(skb2
, skb
->sk
);
210 nexthop
= (__force u32
) rt_nexthop(rt
, ip_hdr(skb
)->daddr
);
211 neigh
= __ipv4_neigh_lookup_noref(dev
, nexthop
);
212 if (unlikely(!neigh
))
213 neigh
= __neigh_create(&arp_tbl
, &nexthop
, dev
, false);
214 if (!IS_ERR(neigh
)) {
215 int res
= dst_neigh_output(dst
, neigh
, skb
);
217 rcu_read_unlock_bh();
220 rcu_read_unlock_bh();
222 net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
228 static int ip_finish_output_gso(struct sock
*sk
, struct sk_buff
*skb
,
231 netdev_features_t features
;
232 struct sk_buff
*segs
;
235 /* common case: locally created skb or seglen is <= mtu */
236 if (((IPCB(skb
)->flags
& IPSKB_FORWARDED
) == 0) ||
237 skb_gso_network_seglen(skb
) <= mtu
)
238 return ip_finish_output2(sk
, skb
);
240 /* Slowpath - GSO segment length is exceeding the dst MTU.
242 * This can happen in two cases:
243 * 1) TCP GRO packet, DF bit not set
244 * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
245 * from host network stack.
247 features
= netif_skb_features(skb
);
248 segs
= skb_gso_segment(skb
, features
& ~NETIF_F_GSO_MASK
);
249 if (IS_ERR_OR_NULL(segs
)) {
257 struct sk_buff
*nskb
= segs
->next
;
261 err
= ip_fragment(sk
, segs
, mtu
, ip_finish_output2
);
271 static int ip_finish_output(struct net
*net
, struct sock
*sk
, struct sk_buff
*skb
)
275 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
276 /* Policy lookup after SNAT yielded a new policy */
277 if (skb_dst(skb
)->xfrm
) {
278 IPCB(skb
)->flags
|= IPSKB_REROUTED
;
279 return dst_output(sk
, skb
);
282 mtu
= ip_skb_dst_mtu(skb
);
284 return ip_finish_output_gso(sk
, skb
, mtu
);
286 if (skb
->len
> mtu
|| (IPCB(skb
)->flags
& IPSKB_FRAG_PMTU
))
287 return ip_fragment(sk
, skb
, mtu
, ip_finish_output2
);
289 return ip_finish_output2(sk
, skb
);
292 int ip_mc_output(struct sock
*sk
, struct sk_buff
*skb
)
294 struct rtable
*rt
= skb_rtable(skb
);
295 struct net_device
*dev
= rt
->dst
.dev
;
296 struct net
*net
= dev_net(dev
);
299 * If the indicated interface is up and running, send the packet.
301 IP_UPD_PO_STATS(net
, IPSTATS_MIB_OUT
, skb
->len
);
304 skb
->protocol
= htons(ETH_P_IP
);
307 * Multicasts are looped back for other local users
310 if (rt
->rt_flags
&RTCF_MULTICAST
) {
312 #ifdef CONFIG_IP_MROUTE
313 /* Small optimization: do not loopback not local frames,
314 which returned after forwarding; they will be dropped
315 by ip_mr_input in any case.
316 Note, that local frames are looped back to be delivered
319 This check is duplicated in ip_mr_input at the moment.
322 ((rt
->rt_flags
& RTCF_LOCAL
) ||
323 !(IPCB(skb
)->flags
& IPSKB_FORWARDED
))
326 struct sk_buff
*newskb
= skb_clone(skb
, GFP_ATOMIC
);
328 NF_HOOK(NFPROTO_IPV4
, NF_INET_POST_ROUTING
,
329 net
, sk
, newskb
, NULL
, newskb
->dev
,
333 /* Multicasts with ttl 0 must not go beyond the host */
335 if (ip_hdr(skb
)->ttl
== 0) {
341 if (rt
->rt_flags
&RTCF_BROADCAST
) {
342 struct sk_buff
*newskb
= skb_clone(skb
, GFP_ATOMIC
);
344 NF_HOOK(NFPROTO_IPV4
, NF_INET_POST_ROUTING
,
345 net
, sk
, newskb
, NULL
, newskb
->dev
,
349 return NF_HOOK_COND(NFPROTO_IPV4
, NF_INET_POST_ROUTING
,
350 net
, sk
, skb
, NULL
, skb
->dev
,
352 !(IPCB(skb
)->flags
& IPSKB_REROUTED
));
355 int ip_output(struct sock
*sk
, struct sk_buff
*skb
)
357 struct net_device
*dev
= skb_dst(skb
)->dev
;
358 struct net
*net
= dev_net(dev
);
360 IP_UPD_PO_STATS(net
, IPSTATS_MIB_OUT
, skb
->len
);
363 skb
->protocol
= htons(ETH_P_IP
);
365 return NF_HOOK_COND(NFPROTO_IPV4
, NF_INET_POST_ROUTING
,
366 net
, sk
, skb
, NULL
, dev
,
368 !(IPCB(skb
)->flags
& IPSKB_REROUTED
));
372 * copy saddr and daddr, possibly using 64bit load/stores
374 * iph->saddr = fl4->saddr;
375 * iph->daddr = fl4->daddr;
377 static void ip_copy_addrs(struct iphdr
*iph
, const struct flowi4
*fl4
)
379 BUILD_BUG_ON(offsetof(typeof(*fl4
), daddr
) !=
380 offsetof(typeof(*fl4
), saddr
) + sizeof(fl4
->saddr
));
381 memcpy(&iph
->saddr
, &fl4
->saddr
,
382 sizeof(fl4
->saddr
) + sizeof(fl4
->daddr
));
385 /* Note: skb->sk can be different from sk, in case of tunnels */
386 int ip_queue_xmit(struct sock
*sk
, struct sk_buff
*skb
, struct flowi
*fl
)
388 struct inet_sock
*inet
= inet_sk(sk
);
389 struct ip_options_rcu
*inet_opt
;
395 /* Skip all of this if the packet is already routed,
396 * f.e. by something like SCTP.
399 inet_opt
= rcu_dereference(inet
->inet_opt
);
401 rt
= skb_rtable(skb
);
405 /* Make sure we can route this packet. */
406 rt
= (struct rtable
*)__sk_dst_check(sk
, 0);
410 /* Use correct destination address if we have options. */
411 daddr
= inet
->inet_daddr
;
412 if (inet_opt
&& inet_opt
->opt
.srr
)
413 daddr
= inet_opt
->opt
.faddr
;
415 /* If this fails, retransmit mechanism of transport layer will
416 * keep trying until route appears or the connection times
419 rt
= ip_route_output_ports(sock_net(sk
), fl4
, sk
,
420 daddr
, inet
->inet_saddr
,
425 sk
->sk_bound_dev_if
);
428 sk_setup_caps(sk
, &rt
->dst
);
430 skb_dst_set_noref(skb
, &rt
->dst
);
433 if (inet_opt
&& inet_opt
->opt
.is_strictroute
&& rt
->rt_uses_gateway
)
436 /* OK, we know where to send it, allocate and build IP header. */
437 skb_push(skb
, sizeof(struct iphdr
) + (inet_opt
? inet_opt
->opt
.optlen
: 0));
438 skb_reset_network_header(skb
);
440 *((__be16
*)iph
) = htons((4 << 12) | (5 << 8) | (inet
->tos
& 0xff));
441 if (ip_dont_fragment(sk
, &rt
->dst
) && !skb
->ignore_df
)
442 iph
->frag_off
= htons(IP_DF
);
445 iph
->ttl
= ip_select_ttl(inet
, &rt
->dst
);
446 iph
->protocol
= sk
->sk_protocol
;
447 ip_copy_addrs(iph
, fl4
);
449 /* Transport layer set skb->h.foo itself. */
451 if (inet_opt
&& inet_opt
->opt
.optlen
) {
452 iph
->ihl
+= inet_opt
->opt
.optlen
>> 2;
453 ip_options_build(skb
, &inet_opt
->opt
, inet
->inet_daddr
, rt
, 0);
456 ip_select_ident_segs(sock_net(sk
), skb
, sk
,
457 skb_shinfo(skb
)->gso_segs
?: 1);
459 /* TODO : should we use skb->sk here instead of sk ? */
460 skb
->priority
= sk
->sk_priority
;
461 skb
->mark
= sk
->sk_mark
;
463 res
= ip_local_out(skb
);
469 IP_INC_STATS(sock_net(sk
), IPSTATS_MIB_OUTNOROUTES
);
471 return -EHOSTUNREACH
;
473 EXPORT_SYMBOL(ip_queue_xmit
);
475 static void ip_copy_metadata(struct sk_buff
*to
, struct sk_buff
*from
)
477 to
->pkt_type
= from
->pkt_type
;
478 to
->priority
= from
->priority
;
479 to
->protocol
= from
->protocol
;
481 skb_dst_copy(to
, from
);
483 to
->mark
= from
->mark
;
485 /* Copy the flags to each fragment. */
486 IPCB(to
)->flags
= IPCB(from
)->flags
;
488 #ifdef CONFIG_NET_SCHED
489 to
->tc_index
= from
->tc_index
;
492 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
493 to
->ipvs_property
= from
->ipvs_property
;
495 skb_copy_secmark(to
, from
);
498 static int ip_fragment(struct sock
*sk
, struct sk_buff
*skb
,
500 int (*output
)(struct sock
*, struct sk_buff
*))
502 struct iphdr
*iph
= ip_hdr(skb
);
504 if ((iph
->frag_off
& htons(IP_DF
)) == 0)
505 return ip_do_fragment(sk
, skb
, output
);
507 if (unlikely(!skb
->ignore_df
||
508 (IPCB(skb
)->frag_max_size
&&
509 IPCB(skb
)->frag_max_size
> mtu
))) {
510 struct net
*net
= dev_net(skb_rtable(skb
)->dst
.dev
);
512 IP_INC_STATS(net
, IPSTATS_MIB_FRAGFAILS
);
513 icmp_send(skb
, ICMP_DEST_UNREACH
, ICMP_FRAG_NEEDED
,
519 return ip_do_fragment(sk
, skb
, output
);
523 * This IP datagram is too large to be sent in one piece. Break it up into
524 * smaller pieces (each of size equal to IP header plus
525 * a block of the data of the original IP data part) that will yet fit in a
526 * single device frame, and queue such a frame for sending.
529 int ip_do_fragment(struct sock
*sk
, struct sk_buff
*skb
,
530 int (*output
)(struct sock
*, struct sk_buff
*))
534 struct net_device
*dev
;
535 struct sk_buff
*skb2
;
536 unsigned int mtu
, hlen
, left
, len
, ll_rs
;
538 __be16 not_last_frag
;
539 struct rtable
*rt
= skb_rtable(skb
);
547 * Point into the IP datagram header.
552 mtu
= ip_skb_dst_mtu(skb
);
553 if (IPCB(skb
)->frag_max_size
&& IPCB(skb
)->frag_max_size
< mtu
)
554 mtu
= IPCB(skb
)->frag_max_size
;
557 * Setup starting values.
561 mtu
= mtu
- hlen
; /* Size of data space */
562 IPCB(skb
)->flags
|= IPSKB_FRAG_COMPLETE
;
564 /* When frag_list is given, use it. First, check its validity:
565 * some transformers could create wrong frag_list or break existing
566 * one, it is not prohibited. In this case fall back to copying.
568 * LATER: this step can be merged to real generation of fragments,
569 * we can switch to copy when see the first bad fragment.
571 if (skb_has_frag_list(skb
)) {
572 struct sk_buff
*frag
, *frag2
;
573 int first_len
= skb_pagelen(skb
);
575 if (first_len
- hlen
> mtu
||
576 ((first_len
- hlen
) & 7) ||
577 ip_is_fragment(iph
) ||
581 skb_walk_frags(skb
, frag
) {
582 /* Correct geometry. */
583 if (frag
->len
> mtu
||
584 ((frag
->len
& 7) && frag
->next
) ||
585 skb_headroom(frag
) < hlen
)
586 goto slow_path_clean
;
588 /* Partially cloned skb? */
589 if (skb_shared(frag
))
590 goto slow_path_clean
;
595 frag
->destructor
= sock_wfree
;
597 skb
->truesize
-= frag
->truesize
;
600 /* Everything is OK. Generate! */
604 frag
= skb_shinfo(skb
)->frag_list
;
605 skb_frag_list_init(skb
);
606 skb
->data_len
= first_len
- skb_headlen(skb
);
607 skb
->len
= first_len
;
608 iph
->tot_len
= htons(first_len
);
609 iph
->frag_off
= htons(IP_MF
);
613 /* Prepare header of the next frame,
614 * before previous one went down. */
616 frag
->ip_summed
= CHECKSUM_NONE
;
617 skb_reset_transport_header(frag
);
618 __skb_push(frag
, hlen
);
619 skb_reset_network_header(frag
);
620 memcpy(skb_network_header(frag
), iph
, hlen
);
622 iph
->tot_len
= htons(frag
->len
);
623 ip_copy_metadata(frag
, skb
);
625 ip_options_fragment(frag
);
626 offset
+= skb
->len
- hlen
;
627 iph
->frag_off
= htons(offset
>>3);
629 iph
->frag_off
|= htons(IP_MF
);
630 /* Ready, complete checksum */
634 err
= output(sk
, skb
);
637 IP_INC_STATS(net
, IPSTATS_MIB_FRAGCREATES
);
647 IP_INC_STATS(net
, IPSTATS_MIB_FRAGOKS
);
656 IP_INC_STATS(net
, IPSTATS_MIB_FRAGFAILS
);
660 skb_walk_frags(skb
, frag2
) {
664 frag2
->destructor
= NULL
;
665 skb
->truesize
+= frag2
->truesize
;
670 /* for offloaded checksums cleanup checksum before fragmentation */
671 if ((skb
->ip_summed
== CHECKSUM_PARTIAL
) && skb_checksum_help(skb
))
675 left
= skb
->len
- hlen
; /* Space per frame */
676 ptr
= hlen
; /* Where to start from */
678 ll_rs
= LL_RESERVED_SPACE(rt
->dst
.dev
);
681 * Fragment the datagram.
684 offset
= (ntohs(iph
->frag_off
) & IP_OFFSET
) << 3;
685 not_last_frag
= iph
->frag_off
& htons(IP_MF
);
688 * Keep copying data until we run out.
693 /* IF: it doesn't fit, use 'mtu' - the data space left */
696 /* IF: we are not sending up to and including the packet end
697 then align the next start on an eight byte boundary */
702 /* Allocate buffer */
703 skb2
= alloc_skb(len
+ hlen
+ ll_rs
, GFP_ATOMIC
);
710 * Set up data on packet
713 ip_copy_metadata(skb2
, skb
);
714 skb_reserve(skb2
, ll_rs
);
715 skb_put(skb2
, len
+ hlen
);
716 skb_reset_network_header(skb2
);
717 skb2
->transport_header
= skb2
->network_header
+ hlen
;
720 * Charge the memory for the fragment to any owner
725 skb_set_owner_w(skb2
, skb
->sk
);
728 * Copy the packet header into the new buffer.
731 skb_copy_from_linear_data(skb
, skb_network_header(skb2
), hlen
);
734 * Copy a block of the IP datagram.
736 if (skb_copy_bits(skb
, ptr
, skb_transport_header(skb2
), len
))
741 * Fill in the new header fields.
744 iph
->frag_off
= htons((offset
>> 3));
746 if (IPCB(skb
)->flags
& IPSKB_FRAG_PMTU
)
747 iph
->frag_off
|= htons(IP_DF
);
749 /* ANK: dirty, but effective trick. Upgrade options only if
750 * the segment to be fragmented was THE FIRST (otherwise,
751 * options are already fixed) and make it ONCE
752 * on the initial skb, so that all the following fragments
753 * will inherit fixed options.
756 ip_options_fragment(skb
);
759 * Added AC : If we are fragmenting a fragment that's not the
760 * last fragment then keep MF on each bit
762 if (left
> 0 || not_last_frag
)
763 iph
->frag_off
|= htons(IP_MF
);
768 * Put this fragment into the sending queue.
770 iph
->tot_len
= htons(len
+ hlen
);
774 err
= output(sk
, skb2
);
778 IP_INC_STATS(net
, IPSTATS_MIB_FRAGCREATES
);
781 IP_INC_STATS(net
, IPSTATS_MIB_FRAGOKS
);
786 IP_INC_STATS(net
, IPSTATS_MIB_FRAGFAILS
);
789 EXPORT_SYMBOL(ip_do_fragment
);
792 ip_generic_getfrag(void *from
, char *to
, int offset
, int len
, int odd
, struct sk_buff
*skb
)
794 struct msghdr
*msg
= from
;
796 if (skb
->ip_summed
== CHECKSUM_PARTIAL
) {
797 if (copy_from_iter(to
, len
, &msg
->msg_iter
) != len
)
801 if (csum_and_copy_from_iter(to
, len
, &csum
, &msg
->msg_iter
) != len
)
803 skb
->csum
= csum_block_add(skb
->csum
, csum
, odd
);
807 EXPORT_SYMBOL(ip_generic_getfrag
);
810 csum_page(struct page
*page
, int offset
, int copy
)
815 csum
= csum_partial(kaddr
+ offset
, copy
, 0);
820 static inline int ip_ufo_append_data(struct sock
*sk
,
821 struct sk_buff_head
*queue
,
822 int getfrag(void *from
, char *to
, int offset
, int len
,
823 int odd
, struct sk_buff
*skb
),
824 void *from
, int length
, int hh_len
, int fragheaderlen
,
825 int transhdrlen
, int maxfraglen
, unsigned int flags
)
830 /* There is support for UDP fragmentation offload by network
831 * device, so create one single skb packet containing complete
834 skb
= skb_peek_tail(queue
);
836 skb
= sock_alloc_send_skb(sk
,
837 hh_len
+ fragheaderlen
+ transhdrlen
+ 20,
838 (flags
& MSG_DONTWAIT
), &err
);
843 /* reserve space for Hardware header */
844 skb_reserve(skb
, hh_len
);
846 /* create space for UDP/IP header */
847 skb_put(skb
, fragheaderlen
+ transhdrlen
);
849 /* initialize network header pointer */
850 skb_reset_network_header(skb
);
852 /* initialize protocol header pointer */
853 skb
->transport_header
= skb
->network_header
+ fragheaderlen
;
857 __skb_queue_tail(queue
, skb
);
858 } else if (skb_is_gso(skb
)) {
862 skb
->ip_summed
= CHECKSUM_PARTIAL
;
863 /* specify the length of each IP datagram fragment */
864 skb_shinfo(skb
)->gso_size
= maxfraglen
- fragheaderlen
;
865 skb_shinfo(skb
)->gso_type
= SKB_GSO_UDP
;
868 return skb_append_datato_frags(sk
, skb
, getfrag
, from
,
869 (length
- transhdrlen
));
872 static int __ip_append_data(struct sock
*sk
,
874 struct sk_buff_head
*queue
,
875 struct inet_cork
*cork
,
876 struct page_frag
*pfrag
,
877 int getfrag(void *from
, char *to
, int offset
,
878 int len
, int odd
, struct sk_buff
*skb
),
879 void *from
, int length
, int transhdrlen
,
882 struct inet_sock
*inet
= inet_sk(sk
);
885 struct ip_options
*opt
= cork
->opt
;
892 unsigned int maxfraglen
, fragheaderlen
, maxnonfragsize
;
893 int csummode
= CHECKSUM_NONE
;
894 struct rtable
*rt
= (struct rtable
*)cork
->dst
;
897 skb
= skb_peek_tail(queue
);
899 exthdrlen
= !skb
? rt
->dst
.header_len
: 0;
900 mtu
= cork
->fragsize
;
901 if (cork
->tx_flags
& SKBTX_ANY_SW_TSTAMP
&&
902 sk
->sk_tsflags
& SOF_TIMESTAMPING_OPT_ID
)
903 tskey
= sk
->sk_tskey
++;
905 hh_len
= LL_RESERVED_SPACE(rt
->dst
.dev
);
907 fragheaderlen
= sizeof(struct iphdr
) + (opt
? opt
->optlen
: 0);
908 maxfraglen
= ((mtu
- fragheaderlen
) & ~7) + fragheaderlen
;
909 maxnonfragsize
= ip_sk_ignore_df(sk
) ? 0xFFFF : mtu
;
911 if (cork
->length
+ length
> maxnonfragsize
- fragheaderlen
) {
912 ip_local_error(sk
, EMSGSIZE
, fl4
->daddr
, inet
->inet_dport
,
913 mtu
- (opt
? opt
->optlen
: 0));
918 * transhdrlen > 0 means that this is the first fragment and we wish
919 * it won't be fragmented in the future.
922 length
+ fragheaderlen
<= mtu
&&
923 rt
->dst
.dev
->features
& NETIF_F_V4_CSUM
&&
925 csummode
= CHECKSUM_PARTIAL
;
927 cork
->length
+= length
;
928 if (((length
> mtu
) || (skb
&& skb_is_gso(skb
))) &&
929 (sk
->sk_protocol
== IPPROTO_UDP
) &&
930 (rt
->dst
.dev
->features
& NETIF_F_UFO
) && !rt
->dst
.header_len
&&
931 (sk
->sk_type
== SOCK_DGRAM
)) {
932 err
= ip_ufo_append_data(sk
, queue
, getfrag
, from
, length
,
933 hh_len
, fragheaderlen
, transhdrlen
,
940 /* So, what's going on in the loop below?
942 * We use calculated fragment length to generate chained skb,
943 * each of segments is IP fragment ready for sending to network after
944 * adding appropriate IP header.
951 /* Check if the remaining data fits into current packet. */
952 copy
= mtu
- skb
->len
;
954 copy
= maxfraglen
- skb
->len
;
957 unsigned int datalen
;
958 unsigned int fraglen
;
959 unsigned int fraggap
;
960 unsigned int alloclen
;
961 struct sk_buff
*skb_prev
;
965 fraggap
= skb_prev
->len
- maxfraglen
;
970 * If remaining data exceeds the mtu,
971 * we know we need more fragment(s).
973 datalen
= length
+ fraggap
;
974 if (datalen
> mtu
- fragheaderlen
)
975 datalen
= maxfraglen
- fragheaderlen
;
976 fraglen
= datalen
+ fragheaderlen
;
978 if ((flags
& MSG_MORE
) &&
979 !(rt
->dst
.dev
->features
&NETIF_F_SG
))
984 alloclen
+= exthdrlen
;
986 /* The last fragment gets additional space at tail.
987 * Note, with MSG_MORE we overallocate on fragments,
988 * because we have no idea what fragment will be
991 if (datalen
== length
+ fraggap
)
992 alloclen
+= rt
->dst
.trailer_len
;
995 skb
= sock_alloc_send_skb(sk
,
996 alloclen
+ hh_len
+ 15,
997 (flags
& MSG_DONTWAIT
), &err
);
1000 if (atomic_read(&sk
->sk_wmem_alloc
) <=
1002 skb
= sock_wmalloc(sk
,
1003 alloclen
+ hh_len
+ 15, 1,
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
);
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
;
1044 if (copy
> 0 && getfrag(from
, data
+ transhdrlen
, offset
, copy
, fraggap
, skb
) < 0) {
1051 length
-= datalen
- fraggap
;
1054 csummode
= CHECKSUM_NONE
;
1057 * Put the packet on the pending queue.
1059 __skb_queue_tail(queue
, skb
);
1066 if (!(rt
->dst
.dev
->features
&NETIF_F_SG
)) {
1070 if (getfrag(from
, skb_put(skb
, copy
),
1071 offset
, copy
, off
, skb
) < 0) {
1072 __skb_trim(skb
, off
);
1077 int i
= skb_shinfo(skb
)->nr_frags
;
1080 if (!sk_page_frag_refill(sk
, pfrag
))
1083 if (!skb_can_coalesce(skb
, i
, pfrag
->page
,
1086 if (i
== MAX_SKB_FRAGS
)
1089 __skb_fill_page_desc(skb
, i
, pfrag
->page
,
1091 skb_shinfo(skb
)->nr_frags
= ++i
;
1092 get_page(pfrag
->page
);
1094 copy
= min_t(int, copy
, pfrag
->size
- pfrag
->offset
);
1096 page_address(pfrag
->page
) + pfrag
->offset
,
1097 offset
, copy
, skb
->len
, skb
) < 0)
1100 pfrag
->offset
+= copy
;
1101 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
1103 skb
->data_len
+= copy
;
1104 skb
->truesize
+= copy
;
1105 atomic_add(copy
, &sk
->sk_wmem_alloc
);
1116 cork
->length
-= length
;
1117 IP_INC_STATS(sock_net(sk
), IPSTATS_MIB_OUTDISCARDS
);
1121 static int ip_setup_cork(struct sock
*sk
, struct inet_cork
*cork
,
1122 struct ipcm_cookie
*ipc
, struct rtable
**rtp
)
1124 struct ip_options_rcu
*opt
;
1128 * setup for corking.
1133 cork
->opt
= kmalloc(sizeof(struct ip_options
) + 40,
1135 if (unlikely(!cork
->opt
))
1138 memcpy(cork
->opt
, &opt
->opt
, sizeof(struct ip_options
) + opt
->opt
.optlen
);
1139 cork
->flags
|= IPCORK_OPT
;
1140 cork
->addr
= ipc
->addr
;
1146 * We steal reference to this route, caller should not release it
1149 cork
->fragsize
= ip_sk_use_pmtu(sk
) ?
1150 dst_mtu(&rt
->dst
) : rt
->dst
.dev
->mtu
;
1151 cork
->dst
= &rt
->dst
;
1153 cork
->ttl
= ipc
->ttl
;
1154 cork
->tos
= ipc
->tos
;
1155 cork
->priority
= ipc
->priority
;
1156 cork
->tx_flags
= ipc
->tx_flags
;
1162 * ip_append_data() and ip_append_page() can make one large IP datagram
1163 * from many pieces of data. Each pieces will be holded on the socket
1164 * until ip_push_pending_frames() is called. Each piece can be a page
1167 * Not only UDP, other transport protocols - e.g. raw sockets - can use
1168 * this interface potentially.
1170 * LATER: length must be adjusted by pad at tail, when it is required.
1172 int ip_append_data(struct sock
*sk
, struct flowi4
*fl4
,
1173 int getfrag(void *from
, char *to
, int offset
, int len
,
1174 int odd
, struct sk_buff
*skb
),
1175 void *from
, int length
, int transhdrlen
,
1176 struct ipcm_cookie
*ipc
, struct rtable
**rtp
,
1179 struct inet_sock
*inet
= inet_sk(sk
);
1182 if (flags
&MSG_PROBE
)
1185 if (skb_queue_empty(&sk
->sk_write_queue
)) {
1186 err
= ip_setup_cork(sk
, &inet
->cork
.base
, ipc
, rtp
);
1193 return __ip_append_data(sk
, fl4
, &sk
->sk_write_queue
, &inet
->cork
.base
,
1194 sk_page_frag(sk
), getfrag
,
1195 from
, length
, transhdrlen
, flags
);
1198 ssize_t
ip_append_page(struct sock
*sk
, struct flowi4
*fl4
, struct page
*page
,
1199 int offset
, size_t size
, int flags
)
1201 struct inet_sock
*inet
= inet_sk(sk
);
1202 struct sk_buff
*skb
;
1204 struct ip_options
*opt
= NULL
;
1205 struct inet_cork
*cork
;
1210 unsigned int maxfraglen
, fragheaderlen
, fraggap
, maxnonfragsize
;
1215 if (flags
&MSG_PROBE
)
1218 if (skb_queue_empty(&sk
->sk_write_queue
))
1221 cork
= &inet
->cork
.base
;
1222 rt
= (struct rtable
*)cork
->dst
;
1223 if (cork
->flags
& IPCORK_OPT
)
1226 if (!(rt
->dst
.dev
->features
&NETIF_F_SG
))
1229 hh_len
= LL_RESERVED_SPACE(rt
->dst
.dev
);
1230 mtu
= cork
->fragsize
;
1232 fragheaderlen
= sizeof(struct iphdr
) + (opt
? opt
->optlen
: 0);
1233 maxfraglen
= ((mtu
- fragheaderlen
) & ~7) + fragheaderlen
;
1234 maxnonfragsize
= ip_sk_ignore_df(sk
) ? 0xFFFF : mtu
;
1236 if (cork
->length
+ size
> maxnonfragsize
- fragheaderlen
) {
1237 ip_local_error(sk
, EMSGSIZE
, fl4
->daddr
, inet
->inet_dport
,
1238 mtu
- (opt
? opt
->optlen
: 0));
1242 skb
= skb_peek_tail(&sk
->sk_write_queue
);
1246 cork
->length
+= size
;
1247 if ((size
+ skb
->len
> mtu
) &&
1248 (sk
->sk_protocol
== IPPROTO_UDP
) &&
1249 (rt
->dst
.dev
->features
& NETIF_F_UFO
)) {
1250 skb_shinfo(skb
)->gso_size
= mtu
- fragheaderlen
;
1251 skb_shinfo(skb
)->gso_type
= SKB_GSO_UDP
;
1255 if (skb_is_gso(skb
)) {
1259 /* Check if the remaining data fits into current packet. */
1260 len
= mtu
- skb
->len
;
1262 len
= maxfraglen
- skb
->len
;
1265 struct sk_buff
*skb_prev
;
1269 fraggap
= skb_prev
->len
- maxfraglen
;
1271 alloclen
= fragheaderlen
+ hh_len
+ fraggap
+ 15;
1272 skb
= sock_wmalloc(sk
, alloclen
, 1, sk
->sk_allocation
);
1273 if (unlikely(!skb
)) {
1279 * Fill in the control structures
1281 skb
->ip_summed
= CHECKSUM_NONE
;
1283 skb_reserve(skb
, hh_len
);
1286 * Find where to start putting bytes.
1288 skb_put(skb
, fragheaderlen
+ fraggap
);
1289 skb_reset_network_header(skb
);
1290 skb
->transport_header
= (skb
->network_header
+
1293 skb
->csum
= skb_copy_and_csum_bits(skb_prev
,
1295 skb_transport_header(skb
),
1297 skb_prev
->csum
= csum_sub(skb_prev
->csum
,
1299 pskb_trim_unique(skb_prev
, maxfraglen
);
1303 * Put the packet on the pending queue.
1305 __skb_queue_tail(&sk
->sk_write_queue
, skb
);
1312 if (skb_append_pagefrags(skb
, page
, offset
, len
)) {
1317 if (skb
->ip_summed
== CHECKSUM_NONE
) {
1319 csum
= csum_page(page
, offset
, len
);
1320 skb
->csum
= csum_block_add(skb
->csum
, csum
, skb
->len
);
1324 skb
->data_len
+= len
;
1325 skb
->truesize
+= len
;
1326 atomic_add(len
, &sk
->sk_wmem_alloc
);
1333 cork
->length
-= size
;
1334 IP_INC_STATS(sock_net(sk
), IPSTATS_MIB_OUTDISCARDS
);
1338 static void ip_cork_release(struct inet_cork
*cork
)
1340 cork
->flags
&= ~IPCORK_OPT
;
1343 dst_release(cork
->dst
);
1348 * Combined all pending IP fragments on the socket as one IP datagram
1349 * and push them out.
1351 struct sk_buff
*__ip_make_skb(struct sock
*sk
,
1353 struct sk_buff_head
*queue
,
1354 struct inet_cork
*cork
)
1356 struct sk_buff
*skb
, *tmp_skb
;
1357 struct sk_buff
**tail_skb
;
1358 struct inet_sock
*inet
= inet_sk(sk
);
1359 struct net
*net
= sock_net(sk
);
1360 struct ip_options
*opt
= NULL
;
1361 struct rtable
*rt
= (struct rtable
*)cork
->dst
;
1366 skb
= __skb_dequeue(queue
);
1369 tail_skb
= &(skb_shinfo(skb
)->frag_list
);
1371 /* move skb->data to ip header from ext header */
1372 if (skb
->data
< skb_network_header(skb
))
1373 __skb_pull(skb
, skb_network_offset(skb
));
1374 while ((tmp_skb
= __skb_dequeue(queue
)) != NULL
) {
1375 __skb_pull(tmp_skb
, skb_network_header_len(skb
));
1376 *tail_skb
= tmp_skb
;
1377 tail_skb
= &(tmp_skb
->next
);
1378 skb
->len
+= tmp_skb
->len
;
1379 skb
->data_len
+= tmp_skb
->len
;
1380 skb
->truesize
+= tmp_skb
->truesize
;
1381 tmp_skb
->destructor
= NULL
;
1385 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1386 * to fragment the frame generated here. No matter, what transforms
1387 * how transforms change size of the packet, it will come out.
1389 skb
->ignore_df
= ip_sk_ignore_df(sk
);
1391 /* DF bit is set when we want to see DF on outgoing frames.
1392 * If ignore_df is set too, we still allow to fragment this frame
1394 if (inet
->pmtudisc
== IP_PMTUDISC_DO
||
1395 inet
->pmtudisc
== IP_PMTUDISC_PROBE
||
1396 (skb
->len
<= dst_mtu(&rt
->dst
) &&
1397 ip_dont_fragment(sk
, &rt
->dst
)))
1400 if (cork
->flags
& IPCORK_OPT
)
1405 else if (rt
->rt_type
== RTN_MULTICAST
)
1408 ttl
= ip_select_ttl(inet
, &rt
->dst
);
1413 iph
->tos
= (cork
->tos
!= -1) ? cork
->tos
: inet
->tos
;
1416 iph
->protocol
= sk
->sk_protocol
;
1417 ip_copy_addrs(iph
, fl4
);
1418 ip_select_ident(net
, skb
, sk
);
1421 iph
->ihl
+= opt
->optlen
>>2;
1422 ip_options_build(skb
, opt
, cork
->addr
, rt
, 0);
1425 skb
->priority
= (cork
->tos
!= -1) ? cork
->priority
: sk
->sk_priority
;
1426 skb
->mark
= sk
->sk_mark
;
1428 * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1432 skb_dst_set(skb
, &rt
->dst
);
1434 if (iph
->protocol
== IPPROTO_ICMP
)
1435 icmp_out_count(net
, ((struct icmphdr
*)
1436 skb_transport_header(skb
))->type
);
1438 ip_cork_release(cork
);
1443 int ip_send_skb(struct net
*net
, struct sk_buff
*skb
)
1447 err
= ip_local_out(skb
);
1450 err
= net_xmit_errno(err
);
1452 IP_INC_STATS(net
, IPSTATS_MIB_OUTDISCARDS
);
1458 int ip_push_pending_frames(struct sock
*sk
, struct flowi4
*fl4
)
1460 struct sk_buff
*skb
;
1462 skb
= ip_finish_skb(sk
, fl4
);
1466 /* Netfilter gets whole the not fragmented skb. */
1467 return ip_send_skb(sock_net(sk
), skb
);
1471 * Throw away all pending data on the socket.
1473 static void __ip_flush_pending_frames(struct sock
*sk
,
1474 struct sk_buff_head
*queue
,
1475 struct inet_cork
*cork
)
1477 struct sk_buff
*skb
;
1479 while ((skb
= __skb_dequeue_tail(queue
)) != NULL
)
1482 ip_cork_release(cork
);
1485 void ip_flush_pending_frames(struct sock
*sk
)
1487 __ip_flush_pending_frames(sk
, &sk
->sk_write_queue
, &inet_sk(sk
)->cork
.base
);
1490 struct sk_buff
*ip_make_skb(struct sock
*sk
,
1492 int getfrag(void *from
, char *to
, int offset
,
1493 int len
, int odd
, struct sk_buff
*skb
),
1494 void *from
, int length
, int transhdrlen
,
1495 struct ipcm_cookie
*ipc
, struct rtable
**rtp
,
1498 struct inet_cork cork
;
1499 struct sk_buff_head queue
;
1502 if (flags
& MSG_PROBE
)
1505 __skb_queue_head_init(&queue
);
1510 err
= ip_setup_cork(sk
, &cork
, ipc
, rtp
);
1512 return ERR_PTR(err
);
1514 err
= __ip_append_data(sk
, fl4
, &queue
, &cork
,
1515 ¤t
->task_frag
, getfrag
,
1516 from
, length
, transhdrlen
, flags
);
1518 __ip_flush_pending_frames(sk
, &queue
, &cork
);
1519 return ERR_PTR(err
);
1522 return __ip_make_skb(sk
, fl4
, &queue
, &cork
);
1526 * Fetch data from kernel space and fill in checksum if needed.
1528 static int ip_reply_glue_bits(void *dptr
, char *to
, int offset
,
1529 int len
, int odd
, struct sk_buff
*skb
)
1533 csum
= csum_partial_copy_nocheck(dptr
+offset
, to
, len
, 0);
1534 skb
->csum
= csum_block_add(skb
->csum
, csum
, odd
);
1539 * Generic function to send a packet as reply to another packet.
1540 * Used to send some TCP resets/acks so far.
1542 void ip_send_unicast_reply(struct sock
*sk
, struct sk_buff
*skb
,
1543 const struct ip_options
*sopt
,
1544 __be32 daddr
, __be32 saddr
,
1545 const struct ip_reply_arg
*arg
,
1548 struct ip_options_data replyopts
;
1549 struct ipcm_cookie ipc
;
1551 struct rtable
*rt
= skb_rtable(skb
);
1552 struct net
*net
= sock_net(sk
);
1553 struct sk_buff
*nskb
;
1557 if (__ip_options_echo(&replyopts
.opt
.opt
, skb
, sopt
))
1566 if (replyopts
.opt
.opt
.optlen
) {
1567 ipc
.opt
= &replyopts
.opt
;
1569 if (replyopts
.opt
.opt
.srr
)
1570 daddr
= replyopts
.opt
.opt
.faddr
;
1573 oif
= arg
->bound_dev_if
;
1574 if (!oif
&& netif_index_is_vrf(net
, skb
->skb_iif
))
1577 flowi4_init_output(&fl4
, oif
,
1578 IP4_REPLY_MARK(net
, skb
->mark
),
1580 RT_SCOPE_UNIVERSE
, ip_hdr(skb
)->protocol
,
1581 ip_reply_arg_flowi_flags(arg
),
1583 tcp_hdr(skb
)->source
, tcp_hdr(skb
)->dest
);
1584 security_skb_classify_flow(skb
, flowi4_to_flowi(&fl4
));
1585 rt
= ip_route_output_key(net
, &fl4
);
1589 inet_sk(sk
)->tos
= arg
->tos
;
1591 sk
->sk_priority
= skb
->priority
;
1592 sk
->sk_protocol
= ip_hdr(skb
)->protocol
;
1593 sk
->sk_bound_dev_if
= arg
->bound_dev_if
;
1594 sk
->sk_sndbuf
= sysctl_wmem_default
;
1595 err
= ip_append_data(sk
, &fl4
, ip_reply_glue_bits
, arg
->iov
->iov_base
,
1596 len
, 0, &ipc
, &rt
, MSG_DONTWAIT
);
1597 if (unlikely(err
)) {
1598 ip_flush_pending_frames(sk
);
1602 nskb
= skb_peek(&sk
->sk_write_queue
);
1604 if (arg
->csumoffset
>= 0)
1605 *((__sum16
*)skb_transport_header(nskb
) +
1606 arg
->csumoffset
) = csum_fold(csum_add(nskb
->csum
,
1608 nskb
->ip_summed
= CHECKSUM_NONE
;
1609 skb_set_queue_mapping(nskb
, skb_get_queue_mapping(skb
));
1610 ip_push_pending_frames(sk
, &fl4
);
1616 void __init
ip_init(void)
1621 #if defined(CONFIG_IP_MULTICAST)