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.
8 * Version: $Id: ip_output.c,v 1.100 2002/02/01 22:01:03 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Donald Becker, <becker@super.org>
13 * Alan Cox, <Alan.Cox@linux.org>
15 * Stefan Becker, <stefanb@yello.ping.de>
16 * Jorge Cwik, <jorge@laser.satlink.net>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Hirokazu Takahashi, <taka@valinux.co.jp>
20 * See ip_input.c for original log
23 * Alan Cox : Missing nonblock feature in ip_build_xmit.
24 * Mike Kilburn : htons() missing in ip_build_xmit.
25 * Bradford Johnson: Fix faulty handling of some frames when
27 * Alexander Demenshin: Missing sk/skb free in ip_queue_xmit
28 * (in case if packet not accepted by
29 * output firewall rules)
30 * Mike McLagan : Routing by source
31 * Alexey Kuznetsov: use new route cache
32 * Andi Kleen: Fix broken PMTU recovery and remove
33 * some redundant tests.
34 * Vitaly E. Lavrov : Transparent proxy revived after year coma.
35 * Andi Kleen : Replace ip_reply with ip_send_reply.
36 * Andi Kleen : Split fast and slow ip_build_xmit path
37 * for decreased register pressure on x86
38 * and more readibility.
39 * Marc Boucher : When call_out_firewall returns FW_QUEUE,
40 * silently drop skb instead of failing with -EPERM.
41 * Detlev Wengorz : Copy protocol for fragments.
42 * Hirokazu Takahashi: HW checksumming for outgoing UDP
44 * Hirokazu Takahashi: sendfile() on UDP works now.
47 #include <asm/uaccess.h>
48 #include <asm/system.h>
49 #include <linux/module.h>
50 #include <linux/types.h>
51 #include <linux/kernel.h>
52 #include <linux/sched.h>
54 #include <linux/string.h>
55 #include <linux/errno.h>
56 #include <linux/config.h>
58 #include <linux/socket.h>
59 #include <linux/sockios.h>
61 #include <linux/inet.h>
62 #include <linux/netdevice.h>
63 #include <linux/etherdevice.h>
64 #include <linux/proc_fs.h>
65 #include <linux/stat.h>
66 #include <linux/init.h>
70 #include <net/protocol.h>
71 #include <net/route.h>
73 #include <linux/skbuff.h>
77 #include <net/checksum.h>
78 #include <net/inetpeer.h>
79 #include <net/checksum.h>
80 #include <linux/igmp.h>
81 #include <linux/netfilter_ipv4.h>
82 #include <linux/netfilter_bridge.h>
83 #include <linux/mroute.h>
84 #include <linux/netlink.h>
85 #include <linux/tcp.h>
87 int sysctl_ip_default_ttl = IPDEFTTL;
89 /* Generate a checksum for an outgoing IP datagram. */
90 __inline__ void ip_send_check(struct iphdr *iph)
93 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
96 /* dev_loopback_xmit for use with netfilter. */
97 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
99 newskb->mac.raw = newskb->data;
100 __skb_pull(newskb, newskb->nh.raw - newskb->data);
101 newskb->pkt_type = PACKET_LOOPBACK;
102 newskb->ip_summed = CHECKSUM_UNNECESSARY;
103 BUG_TRAP(newskb->dst);
108 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
110 int ttl = inet->uc_ttl;
113 ttl = dst_metric(dst, RTAX_HOPLIMIT);
118 * Add an ip header to a skbuff and send it out.
121 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
122 u32 saddr, u32 daddr, struct ip_options *opt)
124 struct inet_sock *inet = inet_sk(sk);
125 struct rtable *rt = (struct rtable *)skb->dst;
128 /* Build the IP header. */
130 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr) + opt->optlen);
132 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr));
136 iph->tos = inet->tos;
137 if (ip_dont_fragment(sk, &rt->u.dst))
138 iph->frag_off = htons(IP_DF);
141 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
142 iph->daddr = rt->rt_dst;
143 iph->saddr = rt->rt_src;
144 iph->protocol = sk->sk_protocol;
145 iph->tot_len = htons(skb->len);
146 ip_select_ident(iph, &rt->u.dst, sk);
149 if (opt && opt->optlen) {
150 iph->ihl += opt->optlen>>2;
151 ip_options_build(skb, opt, daddr, rt, 0);
155 skb->priority = sk->sk_priority;
158 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
162 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
164 static inline int ip_finish_output2(struct sk_buff *skb)
166 struct dst_entry *dst = skb->dst;
167 struct hh_cache *hh = dst->hh;
168 struct net_device *dev = dst->dev;
169 int hh_len = LL_RESERVED_SPACE(dev);
171 /* Be paranoid, rather than too clever. */
172 if (unlikely(skb_headroom(skb) < hh_len && dev->hard_header)) {
173 struct sk_buff *skb2;
175 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
181 skb_set_owner_w(skb2, skb->sk);
189 read_lock_bh(&hh->hh_lock);
190 hh_alen = HH_DATA_ALIGN(hh->hh_len);
191 memcpy(skb->data - hh_alen, hh->hh_data, hh_alen);
192 read_unlock_bh(&hh->hh_lock);
193 skb_push(skb, hh->hh_len);
194 return hh->hh_output(skb);
195 } else if (dst->neighbour)
196 return dst->neighbour->output(skb);
199 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
204 static inline int ip_finish_output(struct sk_buff *skb)
206 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
207 /* Policy lookup after SNAT yielded a new policy */
208 if (skb->dst->xfrm != NULL) {
209 IPCB(skb)->flags |= IPSKB_REROUTED;
210 return dst_output(skb);
213 if (skb->len > dst_mtu(skb->dst) &&
214 !(skb_shinfo(skb)->ufo_size || skb_shinfo(skb)->tso_size))
215 return ip_fragment(skb, ip_finish_output2);
217 return ip_finish_output2(skb);
220 int ip_mc_output(struct sk_buff *skb)
222 struct sock *sk = skb->sk;
223 struct rtable *rt = (struct rtable*)skb->dst;
224 struct net_device *dev = rt->u.dst.dev;
227 * If the indicated interface is up and running, send the packet.
229 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
232 skb->protocol = htons(ETH_P_IP);
235 * Multicasts are looped back for other local users
238 if (rt->rt_flags&RTCF_MULTICAST) {
239 if ((!sk || inet_sk(sk)->mc_loop)
240 #ifdef CONFIG_IP_MROUTE
241 /* Small optimization: do not loopback not local frames,
242 which returned after forwarding; they will be dropped
243 by ip_mr_input in any case.
244 Note, that local frames are looped back to be delivered
247 This check is duplicated in ip_mr_input at the moment.
249 && ((rt->rt_flags&RTCF_LOCAL) || !(IPCB(skb)->flags&IPSKB_FORWARDED))
252 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
254 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
256 ip_dev_loopback_xmit);
259 /* Multicasts with ttl 0 must not go beyond the host */
261 if (skb->nh.iph->ttl == 0) {
267 if (rt->rt_flags&RTCF_BROADCAST) {
268 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
270 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
271 newskb->dev, ip_dev_loopback_xmit);
274 return NF_HOOK_COND(PF_INET, NF_IP_POST_ROUTING, skb, NULL, skb->dev,
276 !(IPCB(skb)->flags & IPSKB_REROUTED));
279 int ip_output(struct sk_buff *skb)
281 struct net_device *dev = skb->dst->dev;
283 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
286 skb->protocol = htons(ETH_P_IP);
288 return NF_HOOK_COND(PF_INET, NF_IP_POST_ROUTING, skb, NULL, dev,
290 !(IPCB(skb)->flags & IPSKB_REROUTED));
293 int ip_queue_xmit(struct sk_buff *skb, int ipfragok)
295 struct sock *sk = skb->sk;
296 struct inet_sock *inet = inet_sk(sk);
297 struct ip_options *opt = inet->opt;
301 /* Skip all of this if the packet is already routed,
302 * f.e. by something like SCTP.
304 rt = (struct rtable *) skb->dst;
308 /* Make sure we can route this packet. */
309 rt = (struct rtable *)__sk_dst_check(sk, 0);
313 /* Use correct destination address if we have options. */
319 struct flowi fl = { .oif = sk->sk_bound_dev_if,
322 .saddr = inet->saddr,
323 .tos = RT_CONN_FLAGS(sk) } },
324 .proto = sk->sk_protocol,
326 { .sport = inet->sport,
327 .dport = inet->dport } } };
329 /* If this fails, retransmit mechanism of transport layer will
330 * keep trying until route appears or the connection times
333 if (ip_route_output_flow(&rt, &fl, sk, 0))
336 sk_setup_caps(sk, &rt->u.dst);
338 skb->dst = dst_clone(&rt->u.dst);
341 if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
344 /* OK, we know where to send it, allocate and build IP header. */
345 iph = (struct iphdr *) skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
346 *((__u16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
347 iph->tot_len = htons(skb->len);
348 if (ip_dont_fragment(sk, &rt->u.dst) && !ipfragok)
349 iph->frag_off = htons(IP_DF);
352 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
353 iph->protocol = sk->sk_protocol;
354 iph->saddr = rt->rt_src;
355 iph->daddr = rt->rt_dst;
357 /* Transport layer set skb->h.foo itself. */
359 if (opt && opt->optlen) {
360 iph->ihl += opt->optlen >> 2;
361 ip_options_build(skb, opt, inet->daddr, rt, 0);
364 ip_select_ident_more(iph, &rt->u.dst, sk,
365 (skb_shinfo(skb)->tso_segs ?: 1) - 1);
367 /* Add an IP checksum. */
370 skb->priority = sk->sk_priority;
372 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
376 IP_INC_STATS(IPSTATS_MIB_OUTNOROUTES);
378 return -EHOSTUNREACH;
382 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
384 to->pkt_type = from->pkt_type;
385 to->priority = from->priority;
386 to->protocol = from->protocol;
387 dst_release(to->dst);
388 to->dst = dst_clone(from->dst);
391 /* Copy the flags to each fragment. */
392 IPCB(to)->flags = IPCB(from)->flags;
394 #ifdef CONFIG_NET_SCHED
395 to->tc_index = from->tc_index;
397 #ifdef CONFIG_NETFILTER
398 to->nfmark = from->nfmark;
399 /* Connection association is same as pre-frag packet */
400 nf_conntrack_put(to->nfct);
401 to->nfct = from->nfct;
402 nf_conntrack_get(to->nfct);
403 to->nfctinfo = from->nfctinfo;
404 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
405 to->ipvs_property = from->ipvs_property;
407 #ifdef CONFIG_BRIDGE_NETFILTER
408 nf_bridge_put(to->nf_bridge);
409 to->nf_bridge = from->nf_bridge;
410 nf_bridge_get(to->nf_bridge);
413 skb_copy_secmark(to, from);
417 * This IP datagram is too large to be sent in one piece. Break it up into
418 * smaller pieces (each of size equal to IP header plus
419 * a block of the data of the original IP data part) that will yet fit in a
420 * single device frame, and queue such a frame for sending.
423 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*))
428 struct net_device *dev;
429 struct sk_buff *skb2;
430 unsigned int mtu, hlen, left, len, ll_rs;
432 __be16 not_last_frag;
433 struct rtable *rt = (struct rtable*)skb->dst;
439 * Point into the IP datagram header.
444 if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
445 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
446 htonl(dst_mtu(&rt->u.dst)));
452 * Setup starting values.
456 mtu = dst_mtu(&rt->u.dst) - hlen; /* Size of data space */
457 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
459 /* When frag_list is given, use it. First, check its validity:
460 * some transformers could create wrong frag_list or break existing
461 * one, it is not prohibited. In this case fall back to copying.
463 * LATER: this step can be merged to real generation of fragments,
464 * we can switch to copy when see the first bad fragment.
466 if (skb_shinfo(skb)->frag_list) {
467 struct sk_buff *frag;
468 int first_len = skb_pagelen(skb);
470 if (first_len - hlen > mtu ||
471 ((first_len - hlen) & 7) ||
472 (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
476 for (frag = skb_shinfo(skb)->frag_list; frag; frag = frag->next) {
477 /* Correct geometry. */
478 if (frag->len > mtu ||
479 ((frag->len & 7) && frag->next) ||
480 skb_headroom(frag) < hlen)
483 /* Partially cloned skb? */
484 if (skb_shared(frag))
491 frag->destructor = sock_wfree;
492 skb->truesize -= frag->truesize;
496 /* Everything is OK. Generate! */
500 frag = skb_shinfo(skb)->frag_list;
501 skb_shinfo(skb)->frag_list = NULL;
502 skb->data_len = first_len - skb_headlen(skb);
503 skb->len = first_len;
504 iph->tot_len = htons(first_len);
505 iph->frag_off = htons(IP_MF);
509 /* Prepare header of the next frame,
510 * before previous one went down. */
512 frag->ip_summed = CHECKSUM_NONE;
513 frag->h.raw = frag->data;
514 frag->nh.raw = __skb_push(frag, hlen);
515 memcpy(frag->nh.raw, iph, hlen);
517 iph->tot_len = htons(frag->len);
518 ip_copy_metadata(frag, skb);
520 ip_options_fragment(frag);
521 offset += skb->len - hlen;
522 iph->frag_off = htons(offset>>3);
523 if (frag->next != NULL)
524 iph->frag_off |= htons(IP_MF);
525 /* Ready, complete checksum */
540 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
549 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
554 left = skb->len - hlen; /* Space per frame */
555 ptr = raw + hlen; /* Where to start from */
557 #ifdef CONFIG_BRIDGE_NETFILTER
558 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
559 * we need to make room for the encapsulating header */
560 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, nf_bridge_pad(skb));
561 mtu -= nf_bridge_pad(skb);
563 ll_rs = LL_RESERVED_SPACE(rt->u.dst.dev);
566 * Fragment the datagram.
569 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
570 not_last_frag = iph->frag_off & htons(IP_MF);
573 * Keep copying data until we run out.
578 /* IF: it doesn't fit, use 'mtu' - the data space left */
581 /* IF: we are not sending upto and including the packet end
582 then align the next start on an eight byte boundary */
590 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
591 NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
597 * Set up data on packet
600 ip_copy_metadata(skb2, skb);
601 skb_reserve(skb2, ll_rs);
602 skb_put(skb2, len + hlen);
603 skb2->nh.raw = skb2->data;
604 skb2->h.raw = skb2->data + hlen;
607 * Charge the memory for the fragment to any owner
612 skb_set_owner_w(skb2, skb->sk);
615 * Copy the packet header into the new buffer.
618 memcpy(skb2->nh.raw, skb->data, hlen);
621 * Copy a block of the IP datagram.
623 if (skb_copy_bits(skb, ptr, skb2->h.raw, len))
628 * Fill in the new header fields.
631 iph->frag_off = htons((offset >> 3));
633 /* ANK: dirty, but effective trick. Upgrade options only if
634 * the segment to be fragmented was THE FIRST (otherwise,
635 * options are already fixed) and make it ONCE
636 * on the initial skb, so that all the following fragments
637 * will inherit fixed options.
640 ip_options_fragment(skb);
643 * Added AC : If we are fragmenting a fragment that's not the
644 * last fragment then keep MF on each bit
646 if (left > 0 || not_last_frag)
647 iph->frag_off |= htons(IP_MF);
652 * Put this fragment into the sending queue.
655 IP_INC_STATS(IPSTATS_MIB_FRAGCREATES);
657 iph->tot_len = htons(len + hlen);
666 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
671 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
675 EXPORT_SYMBOL(ip_fragment);
678 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
680 struct iovec *iov = from;
682 if (skb->ip_summed == CHECKSUM_HW) {
683 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
686 unsigned int csum = 0;
687 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
689 skb->csum = csum_block_add(skb->csum, csum, odd);
694 static inline unsigned int
695 csum_page(struct page *page, int offset, int copy)
700 csum = csum_partial(kaddr + offset, copy, 0);
705 static inline int ip_ufo_append_data(struct sock *sk,
706 int getfrag(void *from, char *to, int offset, int len,
707 int odd, struct sk_buff *skb),
708 void *from, int length, int hh_len, int fragheaderlen,
709 int transhdrlen, int mtu,unsigned int flags)
714 /* There is support for UDP fragmentation offload by network
715 * device, so create one single skb packet containing complete
718 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL) {
719 skb = sock_alloc_send_skb(sk,
720 hh_len + fragheaderlen + transhdrlen + 20,
721 (flags & MSG_DONTWAIT), &err);
726 /* reserve space for Hardware header */
727 skb_reserve(skb, hh_len);
729 /* create space for UDP/IP header */
730 skb_put(skb,fragheaderlen + transhdrlen);
732 /* initialize network header pointer */
733 skb->nh.raw = skb->data;
735 /* initialize protocol header pointer */
736 skb->h.raw = skb->data + fragheaderlen;
738 skb->ip_summed = CHECKSUM_HW;
740 sk->sk_sndmsg_off = 0;
743 err = skb_append_datato_frags(sk,skb, getfrag, from,
744 (length - transhdrlen));
746 /* specify the length of each IP datagram fragment*/
747 skb_shinfo(skb)->ufo_size = (mtu - fragheaderlen);
748 __skb_queue_tail(&sk->sk_write_queue, skb);
752 /* There is not enough support do UFO ,
753 * so follow normal path
760 * ip_append_data() and ip_append_page() can make one large IP datagram
761 * from many pieces of data. Each pieces will be holded on the socket
762 * until ip_push_pending_frames() is called. Each piece can be a page
765 * Not only UDP, other transport protocols - e.g. raw sockets - can use
766 * this interface potentially.
768 * LATER: length must be adjusted by pad at tail, when it is required.
770 int ip_append_data(struct sock *sk,
771 int getfrag(void *from, char *to, int offset, int len,
772 int odd, struct sk_buff *skb),
773 void *from, int length, int transhdrlen,
774 struct ipcm_cookie *ipc, struct rtable *rt,
777 struct inet_sock *inet = inet_sk(sk);
780 struct ip_options *opt = NULL;
787 unsigned int maxfraglen, fragheaderlen;
788 int csummode = CHECKSUM_NONE;
793 if (skb_queue_empty(&sk->sk_write_queue)) {
799 if (inet->cork.opt == NULL) {
800 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
801 if (unlikely(inet->cork.opt == NULL))
804 memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
805 inet->cork.flags |= IPCORK_OPT;
806 inet->cork.addr = ipc->addr;
808 dst_hold(&rt->u.dst);
809 inet->cork.fragsize = mtu = dst_mtu(rt->u.dst.path);
811 inet->cork.length = 0;
812 sk->sk_sndmsg_page = NULL;
813 sk->sk_sndmsg_off = 0;
814 if ((exthdrlen = rt->u.dst.header_len) != 0) {
816 transhdrlen += exthdrlen;
820 if (inet->cork.flags & IPCORK_OPT)
821 opt = inet->cork.opt;
825 mtu = inet->cork.fragsize;
827 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
829 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
830 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
832 if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
833 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu-exthdrlen);
838 * transhdrlen > 0 means that this is the first fragment and we wish
839 * it won't be fragmented in the future.
842 length + fragheaderlen <= mtu &&
843 rt->u.dst.dev->features & NETIF_F_ALL_CSUM &&
845 csummode = CHECKSUM_HW;
847 inet->cork.length += length;
848 if (((length > mtu) && (sk->sk_protocol == IPPROTO_UDP)) &&
849 (rt->u.dst.dev->features & NETIF_F_UFO)) {
851 err = ip_ufo_append_data(sk, getfrag, from, length, hh_len,
852 fragheaderlen, transhdrlen, mtu,
859 /* So, what's going on in the loop below?
861 * We use calculated fragment length to generate chained skb,
862 * each of segments is IP fragment ready for sending to network after
863 * adding appropriate IP header.
866 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
870 /* Check if the remaining data fits into current packet. */
871 copy = mtu - skb->len;
873 copy = maxfraglen - skb->len;
876 unsigned int datalen;
877 unsigned int fraglen;
878 unsigned int fraggap;
879 unsigned int alloclen;
880 struct sk_buff *skb_prev;
884 fraggap = skb_prev->len - maxfraglen;
889 * If remaining data exceeds the mtu,
890 * we know we need more fragment(s).
892 datalen = length + fraggap;
893 if (datalen > mtu - fragheaderlen)
894 datalen = maxfraglen - fragheaderlen;
895 fraglen = datalen + fragheaderlen;
897 if ((flags & MSG_MORE) &&
898 !(rt->u.dst.dev->features&NETIF_F_SG))
901 alloclen = datalen + fragheaderlen;
903 /* The last fragment gets additional space at tail.
904 * Note, with MSG_MORE we overallocate on fragments,
905 * because we have no idea what fragment will be
908 if (datalen == length + fraggap)
909 alloclen += rt->u.dst.trailer_len;
912 skb = sock_alloc_send_skb(sk,
913 alloclen + hh_len + 15,
914 (flags & MSG_DONTWAIT), &err);
917 if (atomic_read(&sk->sk_wmem_alloc) <=
919 skb = sock_wmalloc(sk,
920 alloclen + hh_len + 15, 1,
922 if (unlikely(skb == NULL))
929 * Fill in the control structures
931 skb->ip_summed = csummode;
933 skb_reserve(skb, hh_len);
936 * Find where to start putting bytes.
938 data = skb_put(skb, fraglen);
939 skb->nh.raw = data + exthdrlen;
940 data += fragheaderlen;
941 skb->h.raw = data + exthdrlen;
944 skb->csum = skb_copy_and_csum_bits(
945 skb_prev, maxfraglen,
946 data + transhdrlen, fraggap, 0);
947 skb_prev->csum = csum_sub(skb_prev->csum,
950 skb_trim(skb_prev, maxfraglen);
953 copy = datalen - transhdrlen - fraggap;
954 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
961 length -= datalen - fraggap;
964 csummode = CHECKSUM_NONE;
967 * Put the packet on the pending queue.
969 __skb_queue_tail(&sk->sk_write_queue, skb);
976 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
980 if (getfrag(from, skb_put(skb, copy),
981 offset, copy, off, skb) < 0) {
982 __skb_trim(skb, off);
987 int i = skb_shinfo(skb)->nr_frags;
988 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
989 struct page *page = sk->sk_sndmsg_page;
990 int off = sk->sk_sndmsg_off;
993 if (page && (left = PAGE_SIZE - off) > 0) {
996 if (page != frag->page) {
997 if (i == MAX_SKB_FRAGS) {
1002 skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
1003 frag = &skb_shinfo(skb)->frags[i];
1005 } else if (i < MAX_SKB_FRAGS) {
1006 if (copy > PAGE_SIZE)
1008 page = alloc_pages(sk->sk_allocation, 0);
1013 sk->sk_sndmsg_page = page;
1014 sk->sk_sndmsg_off = 0;
1016 skb_fill_page_desc(skb, i, page, 0, 0);
1017 frag = &skb_shinfo(skb)->frags[i];
1018 skb->truesize += PAGE_SIZE;
1019 atomic_add(PAGE_SIZE, &sk->sk_wmem_alloc);
1024 if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1028 sk->sk_sndmsg_off += copy;
1031 skb->data_len += copy;
1040 inet->cork.length -= length;
1041 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1045 ssize_t ip_append_page(struct sock *sk, struct page *page,
1046 int offset, size_t size, int flags)
1048 struct inet_sock *inet = inet_sk(sk);
1049 struct sk_buff *skb;
1051 struct ip_options *opt = NULL;
1056 unsigned int maxfraglen, fragheaderlen, fraggap;
1061 if (flags&MSG_PROBE)
1064 if (skb_queue_empty(&sk->sk_write_queue))
1068 if (inet->cork.flags & IPCORK_OPT)
1069 opt = inet->cork.opt;
1071 if (!(rt->u.dst.dev->features&NETIF_F_SG))
1074 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1075 mtu = inet->cork.fragsize;
1077 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1078 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1080 if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1081 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu);
1085 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1088 inet->cork.length += size;
1089 if ((sk->sk_protocol == IPPROTO_UDP) &&
1090 (rt->u.dst.dev->features & NETIF_F_UFO))
1091 skb_shinfo(skb)->ufo_size = (mtu - fragheaderlen);
1097 if (skb_shinfo(skb)->ufo_size)
1101 /* Check if the remaining data fits into current packet. */
1102 len = mtu - skb->len;
1104 len = maxfraglen - skb->len;
1107 struct sk_buff *skb_prev;
1113 fraggap = skb_prev->len - maxfraglen;
1115 alloclen = fragheaderlen + hh_len + fraggap + 15;
1116 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1117 if (unlikely(!skb)) {
1123 * Fill in the control structures
1125 skb->ip_summed = CHECKSUM_NONE;
1127 skb_reserve(skb, hh_len);
1130 * Find where to start putting bytes.
1132 data = skb_put(skb, fragheaderlen + fraggap);
1133 skb->nh.iph = iph = (struct iphdr *)data;
1134 data += fragheaderlen;
1138 skb->csum = skb_copy_and_csum_bits(
1139 skb_prev, maxfraglen,
1141 skb_prev->csum = csum_sub(skb_prev->csum,
1143 skb_trim(skb_prev, maxfraglen);
1147 * Put the packet on the pending queue.
1149 __skb_queue_tail(&sk->sk_write_queue, skb);
1153 i = skb_shinfo(skb)->nr_frags;
1156 if (skb_can_coalesce(skb, i, page, offset)) {
1157 skb_shinfo(skb)->frags[i-1].size += len;
1158 } else if (i < MAX_SKB_FRAGS) {
1160 skb_fill_page_desc(skb, i, page, offset, len);
1166 if (skb->ip_summed == CHECKSUM_NONE) {
1168 csum = csum_page(page, offset, len);
1169 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1173 skb->data_len += len;
1180 inet->cork.length -= size;
1181 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1186 * Combined all pending IP fragments on the socket as one IP datagram
1187 * and push them out.
1189 int ip_push_pending_frames(struct sock *sk)
1191 struct sk_buff *skb, *tmp_skb;
1192 struct sk_buff **tail_skb;
1193 struct inet_sock *inet = inet_sk(sk);
1194 struct ip_options *opt = NULL;
1195 struct rtable *rt = inet->cork.rt;
1201 if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1203 tail_skb = &(skb_shinfo(skb)->frag_list);
1205 /* move skb->data to ip header from ext header */
1206 if (skb->data < skb->nh.raw)
1207 __skb_pull(skb, skb->nh.raw - skb->data);
1208 while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1209 __skb_pull(tmp_skb, skb->h.raw - skb->nh.raw);
1210 *tail_skb = tmp_skb;
1211 tail_skb = &(tmp_skb->next);
1212 skb->len += tmp_skb->len;
1213 skb->data_len += tmp_skb->len;
1214 skb->truesize += tmp_skb->truesize;
1215 __sock_put(tmp_skb->sk);
1216 tmp_skb->destructor = NULL;
1220 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1221 * to fragment the frame generated here. No matter, what transforms
1222 * how transforms change size of the packet, it will come out.
1224 if (inet->pmtudisc != IP_PMTUDISC_DO)
1227 /* DF bit is set when we want to see DF on outgoing frames.
1228 * If local_df is set too, we still allow to fragment this frame
1230 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1231 (skb->len <= dst_mtu(&rt->u.dst) &&
1232 ip_dont_fragment(sk, &rt->u.dst)))
1235 if (inet->cork.flags & IPCORK_OPT)
1236 opt = inet->cork.opt;
1238 if (rt->rt_type == RTN_MULTICAST)
1241 ttl = ip_select_ttl(inet, &rt->u.dst);
1243 iph = (struct iphdr *)skb->data;
1247 iph->ihl += opt->optlen>>2;
1248 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1250 iph->tos = inet->tos;
1251 iph->tot_len = htons(skb->len);
1253 ip_select_ident(iph, &rt->u.dst, sk);
1255 iph->protocol = sk->sk_protocol;
1256 iph->saddr = rt->rt_src;
1257 iph->daddr = rt->rt_dst;
1260 skb->priority = sk->sk_priority;
1261 skb->dst = dst_clone(&rt->u.dst);
1263 /* Netfilter gets whole the not fragmented skb. */
1264 err = NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL,
1265 skb->dst->dev, dst_output);
1268 err = inet->recverr ? net_xmit_errno(err) : 0;
1274 inet->cork.flags &= ~IPCORK_OPT;
1275 kfree(inet->cork.opt);
1276 inet->cork.opt = NULL;
1277 if (inet->cork.rt) {
1278 ip_rt_put(inet->cork.rt);
1279 inet->cork.rt = NULL;
1284 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1289 * Throw away all pending data on the socket.
1291 void ip_flush_pending_frames(struct sock *sk)
1293 struct inet_sock *inet = inet_sk(sk);
1294 struct sk_buff *skb;
1296 while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1299 inet->cork.flags &= ~IPCORK_OPT;
1300 kfree(inet->cork.opt);
1301 inet->cork.opt = NULL;
1302 if (inet->cork.rt) {
1303 ip_rt_put(inet->cork.rt);
1304 inet->cork.rt = NULL;
1310 * Fetch data from kernel space and fill in checksum if needed.
1312 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1313 int len, int odd, struct sk_buff *skb)
1317 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1318 skb->csum = csum_block_add(skb->csum, csum, odd);
1323 * Generic function to send a packet as reply to another packet.
1324 * Used to send TCP resets so far. ICMP should use this function too.
1326 * Should run single threaded per socket because it uses the sock
1327 * structure to pass arguments.
1329 * LATER: switch from ip_build_xmit to ip_append_*
1331 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1334 struct inet_sock *inet = inet_sk(sk);
1336 struct ip_options opt;
1339 struct ipcm_cookie ipc;
1341 struct rtable *rt = (struct rtable*)skb->dst;
1343 if (ip_options_echo(&replyopts.opt, skb))
1346 daddr = ipc.addr = rt->rt_src;
1349 if (replyopts.opt.optlen) {
1350 ipc.opt = &replyopts.opt;
1353 daddr = replyopts.opt.faddr;
1357 struct flowi fl = { .nl_u = { .ip4_u =
1359 .saddr = rt->rt_spec_dst,
1360 .tos = RT_TOS(skb->nh.iph->tos) } },
1361 /* Not quite clean, but right. */
1363 { .sport = skb->h.th->dest,
1364 .dport = skb->h.th->source } },
1365 .proto = sk->sk_protocol };
1366 if (ip_route_output_key(&rt, &fl))
1370 /* And let IP do all the hard work.
1372 This chunk is not reenterable, hence spinlock.
1373 Note that it uses the fact, that this function is called
1374 with locally disabled BH and that sk cannot be already spinlocked.
1377 inet->tos = skb->nh.iph->tos;
1378 sk->sk_priority = skb->priority;
1379 sk->sk_protocol = skb->nh.iph->protocol;
1380 ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1381 &ipc, rt, MSG_DONTWAIT);
1382 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1383 if (arg->csumoffset >= 0)
1384 *((u16 *)skb->h.raw + arg->csumoffset) = csum_fold(csum_add(skb->csum, arg->csum));
1385 skb->ip_summed = CHECKSUM_NONE;
1386 ip_push_pending_frames(sk);
1394 void __init ip_init(void)
1399 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1400 igmp_mc_proc_init();
1404 EXPORT_SYMBOL(ip_generic_getfrag);
1405 EXPORT_SYMBOL(ip_queue_xmit);
1406 EXPORT_SYMBOL(ip_send_check);