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>
72 #include <linux/skbuff.h>
76 #include <net/checksum.h>
77 #include <net/inetpeer.h>
78 #include <net/checksum.h>
79 #include <linux/igmp.h>
80 #include <linux/netfilter_ipv4.h>
81 #include <linux/netfilter_bridge.h>
82 #include <linux/mroute.h>
83 #include <linux/netlink.h>
84 #include <linux/tcp.h>
86 int sysctl_ip_default_ttl = IPDEFTTL;
88 static int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*));
90 /* Generate a checksum for an outgoing IP datagram. */
91 __inline__ void ip_send_check(struct iphdr *iph)
94 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
97 /* dev_loopback_xmit for use with netfilter. */
98 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
100 newskb->mac.raw = newskb->data;
101 __skb_pull(newskb, newskb->nh.raw - newskb->data);
102 newskb->pkt_type = PACKET_LOOPBACK;
103 newskb->ip_summed = CHECKSUM_UNNECESSARY;
104 BUG_TRAP(newskb->dst);
109 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
111 int ttl = inet->uc_ttl;
114 ttl = dst_metric(dst, RTAX_HOPLIMIT);
119 * Add an ip header to a skbuff and send it out.
122 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
123 u32 saddr, u32 daddr, struct ip_options *opt)
125 struct inet_sock *inet = inet_sk(sk);
126 struct rtable *rt = (struct rtable *)skb->dst;
129 /* Build the IP header. */
131 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr) + opt->optlen);
133 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr));
137 iph->tos = inet->tos;
138 if (ip_dont_fragment(sk, &rt->u.dst))
139 iph->frag_off = htons(IP_DF);
142 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
143 iph->daddr = rt->rt_dst;
144 iph->saddr = rt->rt_src;
145 iph->protocol = sk->sk_protocol;
146 iph->tot_len = htons(skb->len);
147 ip_select_ident(iph, &rt->u.dst, sk);
150 if (opt && opt->optlen) {
151 iph->ihl += opt->optlen>>2;
152 ip_options_build(skb, opt, daddr, rt, 0);
156 skb->priority = sk->sk_priority;
159 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
163 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
165 static inline int ip_finish_output2(struct sk_buff *skb)
167 struct dst_entry *dst = skb->dst;
168 struct hh_cache *hh = dst->hh;
169 struct net_device *dev = dst->dev;
170 int hh_len = LL_RESERVED_SPACE(dev);
172 /* Be paranoid, rather than too clever. */
173 if (unlikely(skb_headroom(skb) < hh_len && dev->hard_header)) {
174 struct sk_buff *skb2;
176 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
182 skb_set_owner_w(skb2, skb->sk);
190 read_lock_bh(&hh->hh_lock);
191 hh_alen = HH_DATA_ALIGN(hh->hh_len);
192 memcpy(skb->data - hh_alen, hh->hh_data, hh_alen);
193 read_unlock_bh(&hh->hh_lock);
194 skb_push(skb, hh->hh_len);
195 return hh->hh_output(skb);
196 } else if (dst->neighbour)
197 return dst->neighbour->output(skb);
200 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
205 static inline int ip_finish_output(struct sk_buff *skb)
207 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
208 /* Policy lookup after SNAT yielded a new policy */
209 if (skb->dst->xfrm != NULL)
210 return xfrm4_output_finish(skb);
212 if (skb->len > dst_mtu(skb->dst) &&
213 !(skb_shinfo(skb)->ufo_size || skb_shinfo(skb)->tso_size))
214 return ip_fragment(skb, ip_finish_output2);
216 return ip_finish_output2(skb);
219 int ip_mc_output(struct sk_buff *skb)
221 struct sock *sk = skb->sk;
222 struct rtable *rt = (struct rtable*)skb->dst;
223 struct net_device *dev = rt->u.dst.dev;
226 * If the indicated interface is up and running, send the packet.
228 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
231 skb->protocol = htons(ETH_P_IP);
234 * Multicasts are looped back for other local users
237 if (rt->rt_flags&RTCF_MULTICAST) {
238 if ((!sk || inet_sk(sk)->mc_loop)
239 #ifdef CONFIG_IP_MROUTE
240 /* Small optimization: do not loopback not local frames,
241 which returned after forwarding; they will be dropped
242 by ip_mr_input in any case.
243 Note, that local frames are looped back to be delivered
246 This check is duplicated in ip_mr_input at the moment.
248 && ((rt->rt_flags&RTCF_LOCAL) || !(IPCB(skb)->flags&IPSKB_FORWARDED))
251 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
253 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
255 ip_dev_loopback_xmit);
258 /* Multicasts with ttl 0 must not go beyond the host */
260 if (skb->nh.iph->ttl == 0) {
266 if (rt->rt_flags&RTCF_BROADCAST) {
267 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
269 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
270 newskb->dev, ip_dev_loopback_xmit);
273 return NF_HOOK(PF_INET, NF_IP_POST_ROUTING, skb, NULL, skb->dev,
277 int ip_output(struct sk_buff *skb)
279 struct net_device *dev = skb->dst->dev;
281 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
284 skb->protocol = htons(ETH_P_IP);
286 return NF_HOOK(PF_INET, NF_IP_POST_ROUTING, skb, NULL, dev,
290 int ip_queue_xmit(struct sk_buff *skb, int ipfragok)
292 struct sock *sk = skb->sk;
293 struct inet_sock *inet = inet_sk(sk);
294 struct ip_options *opt = inet->opt;
298 /* Skip all of this if the packet is already routed,
299 * f.e. by something like SCTP.
301 rt = (struct rtable *) skb->dst;
305 /* Make sure we can route this packet. */
306 rt = (struct rtable *)__sk_dst_check(sk, 0);
310 /* Use correct destination address if we have options. */
316 struct flowi fl = { .oif = sk->sk_bound_dev_if,
319 .saddr = inet->saddr,
320 .tos = RT_CONN_FLAGS(sk) } },
321 .proto = sk->sk_protocol,
323 { .sport = inet->sport,
324 .dport = inet->dport } } };
326 /* If this fails, retransmit mechanism of transport layer will
327 * keep trying until route appears or the connection times
330 if (ip_route_output_flow(&rt, &fl, sk, 0))
333 sk_setup_caps(sk, &rt->u.dst);
335 skb->dst = dst_clone(&rt->u.dst);
338 if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
341 /* OK, we know where to send it, allocate and build IP header. */
342 iph = (struct iphdr *) skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
343 *((__u16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
344 iph->tot_len = htons(skb->len);
345 if (ip_dont_fragment(sk, &rt->u.dst) && !ipfragok)
346 iph->frag_off = htons(IP_DF);
349 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
350 iph->protocol = sk->sk_protocol;
351 iph->saddr = rt->rt_src;
352 iph->daddr = rt->rt_dst;
354 /* Transport layer set skb->h.foo itself. */
356 if (opt && opt->optlen) {
357 iph->ihl += opt->optlen >> 2;
358 ip_options_build(skb, opt, inet->daddr, rt, 0);
361 ip_select_ident_more(iph, &rt->u.dst, sk,
362 (skb_shinfo(skb)->tso_segs ?: 1) - 1);
364 /* Add an IP checksum. */
367 skb->priority = sk->sk_priority;
369 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
373 IP_INC_STATS(IPSTATS_MIB_OUTNOROUTES);
375 return -EHOSTUNREACH;
379 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
381 to->pkt_type = from->pkt_type;
382 to->priority = from->priority;
383 to->protocol = from->protocol;
384 dst_release(to->dst);
385 to->dst = dst_clone(from->dst);
388 /* Copy the flags to each fragment. */
389 IPCB(to)->flags = IPCB(from)->flags;
391 #ifdef CONFIG_NET_SCHED
392 to->tc_index = from->tc_index;
394 #ifdef CONFIG_NETFILTER
395 to->nfmark = from->nfmark;
396 /* Connection association is same as pre-frag packet */
397 nf_conntrack_put(to->nfct);
398 to->nfct = from->nfct;
399 nf_conntrack_get(to->nfct);
400 to->nfctinfo = from->nfctinfo;
401 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
402 to->ipvs_property = from->ipvs_property;
404 #ifdef CONFIG_BRIDGE_NETFILTER
405 nf_bridge_put(to->nf_bridge);
406 to->nf_bridge = from->nf_bridge;
407 nf_bridge_get(to->nf_bridge);
413 * This IP datagram is too large to be sent in one piece. Break it up into
414 * smaller pieces (each of size equal to IP header plus
415 * a block of the data of the original IP data part) that will yet fit in a
416 * single device frame, and queue such a frame for sending.
419 static int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*))
424 struct net_device *dev;
425 struct sk_buff *skb2;
426 unsigned int mtu, hlen, left, len, ll_rs;
428 __be16 not_last_frag;
429 struct rtable *rt = (struct rtable*)skb->dst;
435 * Point into the IP datagram header.
440 if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
441 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
442 htonl(dst_mtu(&rt->u.dst)));
448 * Setup starting values.
452 mtu = dst_mtu(&rt->u.dst) - hlen; /* Size of data space */
453 IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
455 /* When frag_list is given, use it. First, check its validity:
456 * some transformers could create wrong frag_list or break existing
457 * one, it is not prohibited. In this case fall back to copying.
459 * LATER: this step can be merged to real generation of fragments,
460 * we can switch to copy when see the first bad fragment.
462 if (skb_shinfo(skb)->frag_list) {
463 struct sk_buff *frag;
464 int first_len = skb_pagelen(skb);
466 if (first_len - hlen > mtu ||
467 ((first_len - hlen) & 7) ||
468 (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
472 for (frag = skb_shinfo(skb)->frag_list; frag; frag = frag->next) {
473 /* Correct geometry. */
474 if (frag->len > mtu ||
475 ((frag->len & 7) && frag->next) ||
476 skb_headroom(frag) < hlen)
479 /* Partially cloned skb? */
480 if (skb_shared(frag))
487 frag->destructor = sock_wfree;
488 skb->truesize -= frag->truesize;
492 /* Everything is OK. Generate! */
496 frag = skb_shinfo(skb)->frag_list;
497 skb_shinfo(skb)->frag_list = NULL;
498 skb->data_len = first_len - skb_headlen(skb);
499 skb->len = first_len;
500 iph->tot_len = htons(first_len);
501 iph->frag_off = htons(IP_MF);
505 /* Prepare header of the next frame,
506 * before previous one went down. */
508 frag->ip_summed = CHECKSUM_NONE;
509 frag->h.raw = frag->data;
510 frag->nh.raw = __skb_push(frag, hlen);
511 memcpy(frag->nh.raw, iph, hlen);
513 iph->tot_len = htons(frag->len);
514 ip_copy_metadata(frag, skb);
516 ip_options_fragment(frag);
517 offset += skb->len - hlen;
518 iph->frag_off = htons(offset>>3);
519 if (frag->next != NULL)
520 iph->frag_off |= htons(IP_MF);
521 /* Ready, complete checksum */
536 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
545 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
550 left = skb->len - hlen; /* Space per frame */
551 ptr = raw + hlen; /* Where to start from */
553 #ifdef CONFIG_BRIDGE_NETFILTER
554 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
555 * we need to make room for the encapsulating header */
556 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, nf_bridge_pad(skb));
557 mtu -= nf_bridge_pad(skb);
559 ll_rs = LL_RESERVED_SPACE(rt->u.dst.dev);
562 * Fragment the datagram.
565 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
566 not_last_frag = iph->frag_off & htons(IP_MF);
569 * Keep copying data until we run out.
574 /* IF: it doesn't fit, use 'mtu' - the data space left */
577 /* IF: we are not sending upto and including the packet end
578 then align the next start on an eight byte boundary */
586 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
587 NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
593 * Set up data on packet
596 ip_copy_metadata(skb2, skb);
597 skb_reserve(skb2, ll_rs);
598 skb_put(skb2, len + hlen);
599 skb2->nh.raw = skb2->data;
600 skb2->h.raw = skb2->data + hlen;
603 * Charge the memory for the fragment to any owner
608 skb_set_owner_w(skb2, skb->sk);
611 * Copy the packet header into the new buffer.
614 memcpy(skb2->nh.raw, skb->data, hlen);
617 * Copy a block of the IP datagram.
619 if (skb_copy_bits(skb, ptr, skb2->h.raw, len))
624 * Fill in the new header fields.
627 iph->frag_off = htons((offset >> 3));
629 /* ANK: dirty, but effective trick. Upgrade options only if
630 * the segment to be fragmented was THE FIRST (otherwise,
631 * options are already fixed) and make it ONCE
632 * on the initial skb, so that all the following fragments
633 * will inherit fixed options.
636 ip_options_fragment(skb);
639 * Added AC : If we are fragmenting a fragment that's not the
640 * last fragment then keep MF on each bit
642 if (left > 0 || not_last_frag)
643 iph->frag_off |= htons(IP_MF);
648 * Put this fragment into the sending queue.
651 IP_INC_STATS(IPSTATS_MIB_FRAGCREATES);
653 iph->tot_len = htons(len + hlen);
662 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
667 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
672 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
674 struct iovec *iov = from;
676 if (skb->ip_summed == CHECKSUM_HW) {
677 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
680 unsigned int csum = 0;
681 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
683 skb->csum = csum_block_add(skb->csum, csum, odd);
688 static inline unsigned int
689 csum_page(struct page *page, int offset, int copy)
694 csum = csum_partial(kaddr + offset, copy, 0);
699 static inline int ip_ufo_append_data(struct sock *sk,
700 int getfrag(void *from, char *to, int offset, int len,
701 int odd, struct sk_buff *skb),
702 void *from, int length, int hh_len, int fragheaderlen,
703 int transhdrlen, int mtu,unsigned int flags)
708 /* There is support for UDP fragmentation offload by network
709 * device, so create one single skb packet containing complete
712 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL) {
713 skb = sock_alloc_send_skb(sk,
714 hh_len + fragheaderlen + transhdrlen + 20,
715 (flags & MSG_DONTWAIT), &err);
720 /* reserve space for Hardware header */
721 skb_reserve(skb, hh_len);
723 /* create space for UDP/IP header */
724 skb_put(skb,fragheaderlen + transhdrlen);
726 /* initialize network header pointer */
727 skb->nh.raw = skb->data;
729 /* initialize protocol header pointer */
730 skb->h.raw = skb->data + fragheaderlen;
732 skb->ip_summed = CHECKSUM_HW;
734 sk->sk_sndmsg_off = 0;
737 err = skb_append_datato_frags(sk,skb, getfrag, from,
738 (length - transhdrlen));
740 /* specify the length of each IP datagram fragment*/
741 skb_shinfo(skb)->ufo_size = (mtu - fragheaderlen);
742 __skb_queue_tail(&sk->sk_write_queue, skb);
746 /* There is not enough support do UFO ,
747 * so follow normal path
754 * ip_append_data() and ip_append_page() can make one large IP datagram
755 * from many pieces of data. Each pieces will be holded on the socket
756 * until ip_push_pending_frames() is called. Each piece can be a page
759 * Not only UDP, other transport protocols - e.g. raw sockets - can use
760 * this interface potentially.
762 * LATER: length must be adjusted by pad at tail, when it is required.
764 int ip_append_data(struct sock *sk,
765 int getfrag(void *from, char *to, int offset, int len,
766 int odd, struct sk_buff *skb),
767 void *from, int length, int transhdrlen,
768 struct ipcm_cookie *ipc, struct rtable *rt,
771 struct inet_sock *inet = inet_sk(sk);
774 struct ip_options *opt = NULL;
781 unsigned int maxfraglen, fragheaderlen;
782 int csummode = CHECKSUM_NONE;
787 if (skb_queue_empty(&sk->sk_write_queue)) {
793 if (inet->cork.opt == NULL) {
794 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
795 if (unlikely(inet->cork.opt == NULL))
798 memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
799 inet->cork.flags |= IPCORK_OPT;
800 inet->cork.addr = ipc->addr;
802 dst_hold(&rt->u.dst);
803 inet->cork.fragsize = mtu = dst_mtu(rt->u.dst.path);
805 inet->cork.length = 0;
806 sk->sk_sndmsg_page = NULL;
807 sk->sk_sndmsg_off = 0;
808 if ((exthdrlen = rt->u.dst.header_len) != 0) {
810 transhdrlen += exthdrlen;
814 if (inet->cork.flags & IPCORK_OPT)
815 opt = inet->cork.opt;
819 mtu = inet->cork.fragsize;
821 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
823 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
824 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
826 if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
827 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu-exthdrlen);
832 * transhdrlen > 0 means that this is the first fragment and we wish
833 * it won't be fragmented in the future.
836 length + fragheaderlen <= mtu &&
837 rt->u.dst.dev->features&(NETIF_F_IP_CSUM|NETIF_F_NO_CSUM|NETIF_F_HW_CSUM) &&
839 csummode = CHECKSUM_HW;
841 inet->cork.length += length;
842 if (((length > mtu) && (sk->sk_protocol == IPPROTO_UDP)) &&
843 (rt->u.dst.dev->features & NETIF_F_UFO)) {
845 if(ip_ufo_append_data(sk, getfrag, from, length, hh_len,
846 fragheaderlen, transhdrlen, mtu, flags))
852 /* So, what's going on in the loop below?
854 * We use calculated fragment length to generate chained skb,
855 * each of segments is IP fragment ready for sending to network after
856 * adding appropriate IP header.
859 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
863 /* Check if the remaining data fits into current packet. */
864 copy = mtu - skb->len;
866 copy = maxfraglen - skb->len;
869 unsigned int datalen;
870 unsigned int fraglen;
871 unsigned int fraggap;
872 unsigned int alloclen;
873 struct sk_buff *skb_prev;
877 fraggap = skb_prev->len - maxfraglen;
882 * If remaining data exceeds the mtu,
883 * we know we need more fragment(s).
885 datalen = length + fraggap;
886 if (datalen > mtu - fragheaderlen)
887 datalen = maxfraglen - fragheaderlen;
888 fraglen = datalen + fragheaderlen;
890 if ((flags & MSG_MORE) &&
891 !(rt->u.dst.dev->features&NETIF_F_SG))
894 alloclen = datalen + fragheaderlen;
896 /* The last fragment gets additional space at tail.
897 * Note, with MSG_MORE we overallocate on fragments,
898 * because we have no idea what fragment will be
901 if (datalen == length)
902 alloclen += rt->u.dst.trailer_len;
905 skb = sock_alloc_send_skb(sk,
906 alloclen + hh_len + 15,
907 (flags & MSG_DONTWAIT), &err);
910 if (atomic_read(&sk->sk_wmem_alloc) <=
912 skb = sock_wmalloc(sk,
913 alloclen + hh_len + 15, 1,
915 if (unlikely(skb == NULL))
922 * Fill in the control structures
924 skb->ip_summed = csummode;
926 skb_reserve(skb, hh_len);
929 * Find where to start putting bytes.
931 data = skb_put(skb, fraglen);
932 skb->nh.raw = data + exthdrlen;
933 data += fragheaderlen;
934 skb->h.raw = data + exthdrlen;
937 skb->csum = skb_copy_and_csum_bits(
938 skb_prev, maxfraglen,
939 data + transhdrlen, fraggap, 0);
940 skb_prev->csum = csum_sub(skb_prev->csum,
943 skb_trim(skb_prev, maxfraglen);
946 copy = datalen - transhdrlen - fraggap;
947 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
954 length -= datalen - fraggap;
957 csummode = CHECKSUM_NONE;
960 * Put the packet on the pending queue.
962 __skb_queue_tail(&sk->sk_write_queue, skb);
969 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
973 if (getfrag(from, skb_put(skb, copy),
974 offset, copy, off, skb) < 0) {
975 __skb_trim(skb, off);
980 int i = skb_shinfo(skb)->nr_frags;
981 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
982 struct page *page = sk->sk_sndmsg_page;
983 int off = sk->sk_sndmsg_off;
986 if (page && (left = PAGE_SIZE - off) > 0) {
989 if (page != frag->page) {
990 if (i == MAX_SKB_FRAGS) {
995 skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
996 frag = &skb_shinfo(skb)->frags[i];
998 } else if (i < MAX_SKB_FRAGS) {
999 if (copy > PAGE_SIZE)
1001 page = alloc_pages(sk->sk_allocation, 0);
1006 sk->sk_sndmsg_page = page;
1007 sk->sk_sndmsg_off = 0;
1009 skb_fill_page_desc(skb, i, page, 0, 0);
1010 frag = &skb_shinfo(skb)->frags[i];
1011 skb->truesize += PAGE_SIZE;
1012 atomic_add(PAGE_SIZE, &sk->sk_wmem_alloc);
1017 if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1021 sk->sk_sndmsg_off += copy;
1024 skb->data_len += copy;
1033 inet->cork.length -= length;
1034 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1038 ssize_t ip_append_page(struct sock *sk, struct page *page,
1039 int offset, size_t size, int flags)
1041 struct inet_sock *inet = inet_sk(sk);
1042 struct sk_buff *skb;
1044 struct ip_options *opt = NULL;
1049 unsigned int maxfraglen, fragheaderlen, fraggap;
1054 if (flags&MSG_PROBE)
1057 if (skb_queue_empty(&sk->sk_write_queue))
1061 if (inet->cork.flags & IPCORK_OPT)
1062 opt = inet->cork.opt;
1064 if (!(rt->u.dst.dev->features&NETIF_F_SG))
1067 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1068 mtu = inet->cork.fragsize;
1070 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1071 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1073 if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1074 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu);
1078 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1081 inet->cork.length += size;
1082 if ((sk->sk_protocol == IPPROTO_UDP) &&
1083 (rt->u.dst.dev->features & NETIF_F_UFO))
1084 skb_shinfo(skb)->ufo_size = (mtu - fragheaderlen);
1090 if (skb_shinfo(skb)->ufo_size)
1094 /* Check if the remaining data fits into current packet. */
1095 len = mtu - skb->len;
1097 len = maxfraglen - skb->len;
1100 struct sk_buff *skb_prev;
1106 fraggap = skb_prev->len - maxfraglen;
1108 alloclen = fragheaderlen + hh_len + fraggap + 15;
1109 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1110 if (unlikely(!skb)) {
1116 * Fill in the control structures
1118 skb->ip_summed = CHECKSUM_NONE;
1120 skb_reserve(skb, hh_len);
1123 * Find where to start putting bytes.
1125 data = skb_put(skb, fragheaderlen + fraggap);
1126 skb->nh.iph = iph = (struct iphdr *)data;
1127 data += fragheaderlen;
1131 skb->csum = skb_copy_and_csum_bits(
1132 skb_prev, maxfraglen,
1134 skb_prev->csum = csum_sub(skb_prev->csum,
1136 skb_trim(skb_prev, maxfraglen);
1140 * Put the packet on the pending queue.
1142 __skb_queue_tail(&sk->sk_write_queue, skb);
1146 i = skb_shinfo(skb)->nr_frags;
1149 if (skb_can_coalesce(skb, i, page, offset)) {
1150 skb_shinfo(skb)->frags[i-1].size += len;
1151 } else if (i < MAX_SKB_FRAGS) {
1153 skb_fill_page_desc(skb, i, page, offset, len);
1159 if (skb->ip_summed == CHECKSUM_NONE) {
1161 csum = csum_page(page, offset, len);
1162 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1166 skb->data_len += len;
1173 inet->cork.length -= size;
1174 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1179 * Combined all pending IP fragments on the socket as one IP datagram
1180 * and push them out.
1182 int ip_push_pending_frames(struct sock *sk)
1184 struct sk_buff *skb, *tmp_skb;
1185 struct sk_buff **tail_skb;
1186 struct inet_sock *inet = inet_sk(sk);
1187 struct ip_options *opt = NULL;
1188 struct rtable *rt = inet->cork.rt;
1194 if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1196 tail_skb = &(skb_shinfo(skb)->frag_list);
1198 /* move skb->data to ip header from ext header */
1199 if (skb->data < skb->nh.raw)
1200 __skb_pull(skb, skb->nh.raw - skb->data);
1201 while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1202 __skb_pull(tmp_skb, skb->h.raw - skb->nh.raw);
1203 *tail_skb = tmp_skb;
1204 tail_skb = &(tmp_skb->next);
1205 skb->len += tmp_skb->len;
1206 skb->data_len += tmp_skb->len;
1207 skb->truesize += tmp_skb->truesize;
1208 __sock_put(tmp_skb->sk);
1209 tmp_skb->destructor = NULL;
1213 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1214 * to fragment the frame generated here. No matter, what transforms
1215 * how transforms change size of the packet, it will come out.
1217 if (inet->pmtudisc != IP_PMTUDISC_DO)
1220 /* DF bit is set when we want to see DF on outgoing frames.
1221 * If local_df is set too, we still allow to fragment this frame
1223 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1224 (skb->len <= dst_mtu(&rt->u.dst) &&
1225 ip_dont_fragment(sk, &rt->u.dst)))
1228 if (inet->cork.flags & IPCORK_OPT)
1229 opt = inet->cork.opt;
1231 if (rt->rt_type == RTN_MULTICAST)
1234 ttl = ip_select_ttl(inet, &rt->u.dst);
1236 iph = (struct iphdr *)skb->data;
1240 iph->ihl += opt->optlen>>2;
1241 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1243 iph->tos = inet->tos;
1244 iph->tot_len = htons(skb->len);
1247 __ip_select_ident(iph, &rt->u.dst, 0);
1249 iph->id = htons(inet->id++);
1252 iph->protocol = sk->sk_protocol;
1253 iph->saddr = rt->rt_src;
1254 iph->daddr = rt->rt_dst;
1257 skb->priority = sk->sk_priority;
1258 skb->dst = dst_clone(&rt->u.dst);
1260 /* Netfilter gets whole the not fragmented skb. */
1261 err = NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL,
1262 skb->dst->dev, dst_output);
1265 err = inet->recverr ? net_xmit_errno(err) : 0;
1271 inet->cork.flags &= ~IPCORK_OPT;
1272 kfree(inet->cork.opt);
1273 inet->cork.opt = NULL;
1274 if (inet->cork.rt) {
1275 ip_rt_put(inet->cork.rt);
1276 inet->cork.rt = NULL;
1281 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1286 * Throw away all pending data on the socket.
1288 void ip_flush_pending_frames(struct sock *sk)
1290 struct inet_sock *inet = inet_sk(sk);
1291 struct sk_buff *skb;
1293 while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1296 inet->cork.flags &= ~IPCORK_OPT;
1297 kfree(inet->cork.opt);
1298 inet->cork.opt = NULL;
1299 if (inet->cork.rt) {
1300 ip_rt_put(inet->cork.rt);
1301 inet->cork.rt = NULL;
1307 * Fetch data from kernel space and fill in checksum if needed.
1309 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1310 int len, int odd, struct sk_buff *skb)
1314 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1315 skb->csum = csum_block_add(skb->csum, csum, odd);
1320 * Generic function to send a packet as reply to another packet.
1321 * Used to send TCP resets so far. ICMP should use this function too.
1323 * Should run single threaded per socket because it uses the sock
1324 * structure to pass arguments.
1326 * LATER: switch from ip_build_xmit to ip_append_*
1328 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1331 struct inet_sock *inet = inet_sk(sk);
1333 struct ip_options opt;
1336 struct ipcm_cookie ipc;
1338 struct rtable *rt = (struct rtable*)skb->dst;
1340 if (ip_options_echo(&replyopts.opt, skb))
1343 daddr = ipc.addr = rt->rt_src;
1346 if (replyopts.opt.optlen) {
1347 ipc.opt = &replyopts.opt;
1350 daddr = replyopts.opt.faddr;
1354 struct flowi fl = { .nl_u = { .ip4_u =
1356 .saddr = rt->rt_spec_dst,
1357 .tos = RT_TOS(skb->nh.iph->tos) } },
1358 /* Not quite clean, but right. */
1360 { .sport = skb->h.th->dest,
1361 .dport = skb->h.th->source } },
1362 .proto = sk->sk_protocol };
1363 if (ip_route_output_key(&rt, &fl))
1367 /* And let IP do all the hard work.
1369 This chunk is not reenterable, hence spinlock.
1370 Note that it uses the fact, that this function is called
1371 with locally disabled BH and that sk cannot be already spinlocked.
1374 inet->tos = skb->nh.iph->tos;
1375 sk->sk_priority = skb->priority;
1376 sk->sk_protocol = skb->nh.iph->protocol;
1377 ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1378 &ipc, rt, MSG_DONTWAIT);
1379 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1380 if (arg->csumoffset >= 0)
1381 *((u16 *)skb->h.raw + arg->csumoffset) = csum_fold(csum_add(skb->csum, arg->csum));
1382 skb->ip_summed = CHECKSUM_NONE;
1383 ip_push_pending_frames(sk);
1391 void __init ip_init(void)
1396 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1397 igmp_mc_proc_init();
1401 EXPORT_SYMBOL(ip_generic_getfrag);
1402 EXPORT_SYMBOL(ip_queue_xmit);
1403 EXPORT_SYMBOL(ip_send_check);