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 $
10 * Authors: Ross Biro, <bir7@leland.Stanford.Edu>
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>
74 #include <linux/skbuff.h>
79 #include <net/checksum.h>
80 #include <net/inetpeer.h>
81 #include <net/checksum.h>
82 #include <linux/igmp.h>
83 #include <linux/netfilter_ipv4.h>
84 #include <linux/netfilter_bridge.h>
85 #include <linux/mroute.h>
86 #include <linux/netlink.h>
89 * Shall we try to damage output packets if routing dev changes?
92 int sysctl_ip_dynaddr;
93 int sysctl_ip_default_ttl = IPDEFTTL;
95 /* Generate a checksum for an outgoing IP datagram. */
96 __inline__ void ip_send_check(struct iphdr *iph)
99 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
102 /* dev_loopback_xmit for use with netfilter. */
103 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
105 newskb->mac.raw = newskb->data;
106 __skb_pull(newskb, newskb->nh.raw - newskb->data);
107 newskb->pkt_type = PACKET_LOOPBACK;
108 newskb->ip_summed = CHECKSUM_UNNECESSARY;
109 BUG_TRAP(newskb->dst);
111 #ifdef CONFIG_NETFILTER_DEBUG
112 nf_debug_ip_loopback_xmit(newskb);
118 static inline int ip_select_ttl(struct inet_opt *inet, struct dst_entry *dst)
120 int ttl = inet->uc_ttl;
123 ttl = dst_metric(dst, RTAX_HOPLIMIT);
128 * Add an ip header to a skbuff and send it out.
131 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
132 u32 saddr, u32 daddr, struct ip_options *opt)
134 struct inet_opt *inet = inet_sk(sk);
135 struct rtable *rt = (struct rtable *)skb->dst;
138 /* Build the IP header. */
140 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr) + opt->optlen);
142 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr));
146 iph->tos = inet->tos;
147 if (ip_dont_fragment(sk, &rt->u.dst))
148 iph->frag_off = htons(IP_DF);
151 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
152 iph->daddr = rt->rt_dst;
153 iph->saddr = rt->rt_src;
154 iph->protocol = sk->sk_protocol;
155 iph->tot_len = htons(skb->len);
156 ip_select_ident(iph, &rt->u.dst, sk);
159 if (opt && opt->optlen) {
160 iph->ihl += opt->optlen>>2;
161 ip_options_build(skb, opt, daddr, rt, 0);
165 skb->priority = sk->sk_priority;
168 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
172 static inline int ip_finish_output2(struct sk_buff *skb)
174 struct dst_entry *dst = skb->dst;
175 struct hh_cache *hh = dst->hh;
176 struct net_device *dev = dst->dev;
177 int hh_len = LL_RESERVED_SPACE(dev);
179 /* Be paranoid, rather than too clever. */
180 if (unlikely(skb_headroom(skb) < hh_len && dev->hard_header)) {
181 struct sk_buff *skb2;
183 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
189 skb_set_owner_w(skb2, skb->sk);
194 #ifdef CONFIG_NETFILTER_DEBUG
195 nf_debug_ip_finish_output2(skb);
196 #endif /*CONFIG_NETFILTER_DEBUG*/
201 read_lock_bh(&hh->hh_lock);
202 hh_alen = HH_DATA_ALIGN(hh->hh_len);
203 memcpy(skb->data - hh_alen, hh->hh_data, hh_alen);
204 read_unlock_bh(&hh->hh_lock);
205 skb_push(skb, hh->hh_len);
206 return hh->hh_output(skb);
207 } else if (dst->neighbour)
208 return dst->neighbour->output(skb);
211 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
216 int ip_finish_output(struct sk_buff *skb)
218 struct net_device *dev = skb->dst->dev;
221 skb->protocol = htons(ETH_P_IP);
223 return NF_HOOK(PF_INET, NF_IP_POST_ROUTING, skb, NULL, dev,
227 int ip_mc_output(struct sk_buff **pskb)
229 struct sk_buff *skb = *pskb;
230 struct sock *sk = skb->sk;
231 struct rtable *rt = (struct rtable*)skb->dst;
232 struct net_device *dev = rt->u.dst.dev;
235 * If the indicated interface is up and running, send the packet.
237 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
240 skb->protocol = htons(ETH_P_IP);
243 * Multicasts are looped back for other local users
246 if (rt->rt_flags&RTCF_MULTICAST) {
247 if ((!sk || inet_sk(sk)->mc_loop)
248 #ifdef CONFIG_IP_MROUTE
249 /* Small optimization: do not loopback not local frames,
250 which returned after forwarding; they will be dropped
251 by ip_mr_input in any case.
252 Note, that local frames are looped back to be delivered
255 This check is duplicated in ip_mr_input at the moment.
257 && ((rt->rt_flags&RTCF_LOCAL) || !(IPCB(skb)->flags&IPSKB_FORWARDED))
260 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
262 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
264 ip_dev_loopback_xmit);
267 /* Multicasts with ttl 0 must not go beyond the host */
269 if (skb->nh.iph->ttl == 0) {
275 if (rt->rt_flags&RTCF_BROADCAST) {
276 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
278 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
279 newskb->dev, ip_dev_loopback_xmit);
282 if (skb->len > dst_pmtu(&rt->u.dst))
283 return ip_fragment(skb, ip_finish_output);
285 return ip_finish_output(skb);
288 int ip_output(struct sk_buff **pskb)
290 struct sk_buff *skb = *pskb;
292 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
294 if (skb->len > dst_pmtu(skb->dst) && !skb_shinfo(skb)->tso_size)
295 return ip_fragment(skb, ip_finish_output);
297 return ip_finish_output(skb);
300 int ip_queue_xmit(struct sk_buff *skb, int ipfragok)
302 struct sock *sk = skb->sk;
303 struct inet_opt *inet = inet_sk(sk);
304 struct ip_options *opt = inet->opt;
308 /* Skip all of this if the packet is already routed,
309 * f.e. by something like SCTP.
311 rt = (struct rtable *) skb->dst;
315 /* Make sure we can route this packet. */
316 rt = (struct rtable *)__sk_dst_check(sk, 0);
320 /* Use correct destination address if we have options. */
326 struct flowi fl = { .oif = sk->sk_bound_dev_if,
329 .saddr = inet->saddr,
330 .tos = RT_CONN_FLAGS(sk) } },
331 .proto = sk->sk_protocol,
333 { .sport = inet->sport,
334 .dport = inet->dport } } };
336 /* If this fails, retransmit mechanism of transport layer will
337 * keep trying until route appears or the connection times
340 if (ip_route_output_flow(&rt, &fl, sk, 0))
343 __sk_dst_set(sk, &rt->u.dst);
344 tcp_v4_setup_caps(sk, &rt->u.dst);
346 skb->dst = dst_clone(&rt->u.dst);
349 if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
352 /* OK, we know where to send it, allocate and build IP header. */
353 iph = (struct iphdr *) skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
354 *((__u16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
355 iph->tot_len = htons(skb->len);
356 if (ip_dont_fragment(sk, &rt->u.dst) && !ipfragok)
357 iph->frag_off = htons(IP_DF);
360 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
361 iph->protocol = sk->sk_protocol;
362 iph->saddr = rt->rt_src;
363 iph->daddr = rt->rt_dst;
365 /* Transport layer set skb->h.foo itself. */
367 if (opt && opt->optlen) {
368 iph->ihl += opt->optlen >> 2;
369 ip_options_build(skb, opt, inet->daddr, rt, 0);
372 ip_select_ident_more(iph, &rt->u.dst, sk, skb_shinfo(skb)->tso_segs);
374 /* Add an IP checksum. */
377 skb->priority = sk->sk_priority;
379 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
383 IP_INC_STATS(IPSTATS_MIB_OUTNOROUTES);
385 return -EHOSTUNREACH;
389 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
391 to->pkt_type = from->pkt_type;
392 to->priority = from->priority;
393 to->protocol = from->protocol;
394 to->security = from->security;
395 to->dst = dst_clone(from->dst);
398 /* Copy the flags to each fragment. */
399 IPCB(to)->flags = IPCB(from)->flags;
401 #ifdef CONFIG_NET_SCHED
402 to->tc_index = from->tc_index;
404 #ifdef CONFIG_NETFILTER
405 to->nfmark = from->nfmark;
406 to->nfcache = from->nfcache;
407 /* Connection association is same as pre-frag packet */
408 nf_conntrack_put(to->nfct);
409 to->nfct = from->nfct;
410 nf_conntrack_get(to->nfct);
411 to->nfctinfo = from->nfctinfo;
412 #ifdef CONFIG_BRIDGE_NETFILTER
413 nf_bridge_put(to->nf_bridge);
414 to->nf_bridge = from->nf_bridge;
415 nf_bridge_get(to->nf_bridge);
417 #ifdef CONFIG_NETFILTER_DEBUG
418 to->nf_debug = from->nf_debug;
424 * This IP datagram is too large to be sent in one piece. Break it up into
425 * smaller pieces (each of size equal to IP header plus
426 * a block of the data of the original IP data part) that will yet fit in a
427 * single device frame, and queue such a frame for sending.
430 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*))
435 struct net_device *dev;
436 struct sk_buff *skb2;
437 unsigned int mtu, hlen, left, len, ll_rs;
440 struct rtable *rt = (struct rtable*)skb->dst;
446 * Point into the IP datagram header.
451 if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
452 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
453 htonl(dst_pmtu(&rt->u.dst)));
459 * Setup starting values.
463 mtu = dst_pmtu(&rt->u.dst) - hlen; /* Size of data space */
465 /* When frag_list is given, use it. First, check its validity:
466 * some transformers could create wrong frag_list or break existing
467 * one, it is not prohibited. In this case fall back to copying.
469 * LATER: this step can be merged to real generation of fragments,
470 * we can switch to copy when see the first bad fragment.
472 if (skb_shinfo(skb)->frag_list) {
473 struct sk_buff *frag;
474 int first_len = skb_pagelen(skb);
476 if (first_len - hlen > mtu ||
477 ((first_len - hlen) & 7) ||
478 (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
482 for (frag = skb_shinfo(skb)->frag_list; frag; frag = frag->next) {
483 /* Correct geometry. */
484 if (frag->len > mtu ||
485 ((frag->len & 7) && frag->next) ||
486 skb_headroom(frag) < hlen)
489 /* Partially cloned skb? */
490 if (skb_shared(frag))
494 /* Everything is OK. Generate! */
498 frag = skb_shinfo(skb)->frag_list;
499 skb_shinfo(skb)->frag_list = NULL;
500 skb->data_len = first_len - skb_headlen(skb);
501 skb->len = first_len;
502 iph->tot_len = htons(first_len);
503 iph->frag_off |= htons(IP_MF);
507 /* Prepare header of the next frame,
508 * before previous one went down. */
510 frag->h.raw = frag->data;
511 frag->nh.raw = __skb_push(frag, hlen);
512 memcpy(frag->nh.raw, iph, hlen);
514 iph->tot_len = htons(frag->len);
515 ip_copy_metadata(frag, skb);
517 ip_options_fragment(frag);
518 offset += skb->len - hlen;
519 iph->frag_off = htons(offset>>3);
520 if (frag->next != NULL)
521 iph->frag_off |= htons(IP_MF);
522 /* Ready, complete checksum */
537 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
546 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
551 left = skb->len - hlen; /* Space per frame */
552 ptr = raw + hlen; /* Where to start from */
554 #ifdef CONFIG_BRIDGE_NETFILTER
555 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
556 * we need to make room for the encapsulating header */
557 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, nf_bridge_pad(skb));
558 mtu -= nf_bridge_pad(skb);
560 ll_rs = LL_RESERVED_SPACE(rt->u.dst.dev);
563 * Fragment the datagram.
566 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
567 not_last_frag = iph->frag_off & htons(IP_MF);
570 * Keep copying data until we run out.
575 /* IF: it doesn't fit, use 'mtu' - the data space left */
578 /* IF: we are not sending upto and including the packet end
579 then align the next start on an eight byte boundary */
587 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
588 NETDEBUG(printk(KERN_INFO "IP: frag: no memory for new fragment!\n"));
594 * Set up data on packet
597 ip_copy_metadata(skb2, skb);
598 skb_reserve(skb2, ll_rs);
599 skb_put(skb2, len + hlen);
600 skb2->nh.raw = skb2->data;
601 skb2->h.raw = skb2->data + hlen;
604 * Charge the memory for the fragment to any owner
609 skb_set_owner_w(skb2, skb->sk);
612 * Copy the packet header into the new buffer.
615 memcpy(skb2->nh.raw, skb->data, hlen);
618 * Copy a block of the IP datagram.
620 if (skb_copy_bits(skb, ptr, skb2->h.raw, len))
625 * Fill in the new header fields.
628 iph->frag_off = htons((offset >> 3));
630 /* ANK: dirty, but effective trick. Upgrade options only if
631 * the segment to be fragmented was THE FIRST (otherwise,
632 * options are already fixed) and make it ONCE
633 * on the initial skb, so that all the following fragments
634 * will inherit fixed options.
637 ip_options_fragment(skb);
640 * Added AC : If we are fragmenting a fragment that's not the
641 * last fragment then keep MF on each bit
643 if (left > 0 || not_last_frag)
644 iph->frag_off |= htons(IP_MF);
649 * Put this fragment into the sending queue.
652 IP_INC_STATS(IPSTATS_MIB_FRAGCREATES);
654 iph->tot_len = htons(len + hlen);
663 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
668 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
673 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
675 struct iovec *iov = from;
677 if (skb->ip_summed == CHECKSUM_HW) {
678 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
681 unsigned int csum = 0;
682 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
684 skb->csum = csum_block_add(skb->csum, csum, odd);
689 static inline unsigned int
690 csum_page(struct page *page, int offset, int copy)
695 csum = csum_partial(kaddr + offset, copy, 0);
701 * ip_append_data() and ip_append_page() can make one large IP datagram
702 * from many pieces of data. Each pieces will be holded on the socket
703 * until ip_push_pending_frames() is called. Each piece can be a page
706 * Not only UDP, other transport protocols - e.g. raw sockets - can use
707 * this interface potentially.
709 * LATER: length must be adjusted by pad at tail, when it is required.
711 int ip_append_data(struct sock *sk,
712 int getfrag(void *from, char *to, int offset, int len,
713 int odd, struct sk_buff *skb),
714 void *from, int length, int transhdrlen,
715 struct ipcm_cookie *ipc, struct rtable *rt,
718 struct inet_opt *inet = inet_sk(sk);
721 struct ip_options *opt = NULL;
728 unsigned int maxfraglen, fragheaderlen;
729 int csummode = CHECKSUM_NONE;
734 if (skb_queue_empty(&sk->sk_write_queue)) {
740 if (inet->cork.opt == NULL) {
741 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
742 if (unlikely(inet->cork.opt == NULL))
745 memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
746 inet->cork.flags |= IPCORK_OPT;
747 inet->cork.addr = ipc->addr;
749 dst_hold(&rt->u.dst);
750 inet->cork.fragsize = mtu = dst_pmtu(&rt->u.dst);
752 inet->cork.length = 0;
753 sk->sk_sndmsg_page = NULL;
754 sk->sk_sndmsg_off = 0;
755 if ((exthdrlen = rt->u.dst.header_len) != 0) {
757 transhdrlen += exthdrlen;
761 if (inet->cork.flags & IPCORK_OPT)
762 opt = inet->cork.opt;
766 mtu = inet->cork.fragsize;
768 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
770 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
771 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
773 if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
774 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu-exthdrlen);
779 * transhdrlen > 0 means that this is the first fragment and we wish
780 * it won't be fragmented in the future.
783 length + fragheaderlen <= mtu &&
784 rt->u.dst.dev->features&(NETIF_F_IP_CSUM|NETIF_F_NO_CSUM|NETIF_F_HW_CSUM) &&
786 csummode = CHECKSUM_HW;
788 inet->cork.length += length;
790 /* So, what's going on in the loop below?
792 * We use calculated fragment length to generate chained skb,
793 * each of segments is IP fragment ready for sending to network after
794 * adding appropriate IP header.
797 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
801 /* Check if the remaining data fits into current packet. */
802 copy = mtu - skb->len;
804 copy = maxfraglen - skb->len;
807 unsigned int datalen;
808 unsigned int fraglen;
809 unsigned int fraggap;
810 unsigned int alloclen;
811 struct sk_buff *skb_prev;
815 fraggap = skb_prev->len - maxfraglen;
820 * If remaining data exceeds the mtu,
821 * we know we need more fragment(s).
823 datalen = length + fraggap;
824 if (datalen > mtu - fragheaderlen)
825 datalen = maxfraglen - fragheaderlen;
826 fraglen = datalen + fragheaderlen;
828 if ((flags & MSG_MORE) &&
829 !(rt->u.dst.dev->features&NETIF_F_SG))
832 alloclen = datalen + fragheaderlen;
834 /* The last fragment gets additional space at tail.
835 * Note, with MSG_MORE we overallocate on fragments,
836 * because we have no idea what fragment will be
839 if (datalen == length)
840 alloclen += rt->u.dst.trailer_len;
843 skb = sock_alloc_send_skb(sk,
844 alloclen + hh_len + 15,
845 (flags & MSG_DONTWAIT), &err);
848 if (atomic_read(&sk->sk_wmem_alloc) <=
850 skb = sock_wmalloc(sk,
851 alloclen + hh_len + 15, 1,
853 if (unlikely(skb == NULL))
860 * Fill in the control structures
862 skb->ip_summed = csummode;
864 skb_reserve(skb, hh_len);
867 * Find where to start putting bytes.
869 data = skb_put(skb, fraglen);
870 skb->nh.raw = data + exthdrlen;
871 data += fragheaderlen;
872 skb->h.raw = data + exthdrlen;
875 skb->csum = skb_copy_and_csum_bits(
876 skb_prev, maxfraglen,
877 data + transhdrlen, fraggap, 0);
878 skb_prev->csum = csum_sub(skb_prev->csum,
881 skb_trim(skb_prev, maxfraglen);
884 copy = datalen - transhdrlen - fraggap;
885 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
892 length -= datalen - fraggap;
895 csummode = CHECKSUM_NONE;
898 * Put the packet on the pending queue.
900 __skb_queue_tail(&sk->sk_write_queue, skb);
907 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
911 if (getfrag(from, skb_put(skb, copy),
912 offset, copy, off, skb) < 0) {
913 __skb_trim(skb, off);
918 int i = skb_shinfo(skb)->nr_frags;
919 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
920 struct page *page = sk->sk_sndmsg_page;
921 int off = sk->sk_sndmsg_off;
924 if (page && (left = PAGE_SIZE - off) > 0) {
927 if (page != frag->page) {
928 if (i == MAX_SKB_FRAGS) {
933 skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
934 frag = &skb_shinfo(skb)->frags[i];
936 } else if (i < MAX_SKB_FRAGS) {
937 if (copy > PAGE_SIZE)
939 page = alloc_pages(sk->sk_allocation, 0);
944 sk->sk_sndmsg_page = page;
945 sk->sk_sndmsg_off = 0;
947 skb_fill_page_desc(skb, i, page, 0, 0);
948 frag = &skb_shinfo(skb)->frags[i];
949 skb->truesize += PAGE_SIZE;
950 atomic_add(PAGE_SIZE, &sk->sk_wmem_alloc);
955 if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
959 sk->sk_sndmsg_off += copy;
962 skb->data_len += copy;
971 inet->cork.length -= length;
972 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
976 ssize_t ip_append_page(struct sock *sk, struct page *page,
977 int offset, size_t size, int flags)
979 struct inet_opt *inet = inet_sk(sk);
982 struct ip_options *opt = NULL;
987 unsigned int maxfraglen, fragheaderlen, fraggap;
995 if (skb_queue_empty(&sk->sk_write_queue))
999 if (inet->cork.flags & IPCORK_OPT)
1000 opt = inet->cork.opt;
1002 if (!(rt->u.dst.dev->features&NETIF_F_SG))
1005 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1006 mtu = inet->cork.fragsize;
1008 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1009 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1011 if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1012 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu);
1016 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1019 inet->cork.length += size;
1024 /* Check if the remaining data fits into current packet. */
1025 len = mtu - skb->len;
1027 len = maxfraglen - skb->len;
1029 struct sk_buff *skb_prev;
1036 fraggap = skb_prev->len - maxfraglen;
1040 alloclen = fragheaderlen + hh_len + fraggap + 15;
1041 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1042 if (unlikely(!skb)) {
1048 * Fill in the control structures
1050 skb->ip_summed = CHECKSUM_NONE;
1052 skb_reserve(skb, hh_len);
1055 * Find where to start putting bytes.
1057 data = skb_put(skb, fragheaderlen + fraggap);
1058 skb->nh.iph = iph = (struct iphdr *)data;
1059 data += fragheaderlen;
1063 skb->csum = skb_copy_and_csum_bits(
1064 skb_prev, maxfraglen,
1066 skb_prev->csum = csum_sub(skb_prev->csum,
1068 skb_trim(skb_prev, maxfraglen);
1072 * Put the packet on the pending queue.
1074 __skb_queue_tail(&sk->sk_write_queue, skb);
1078 i = skb_shinfo(skb)->nr_frags;
1081 if (skb_can_coalesce(skb, i, page, offset)) {
1082 skb_shinfo(skb)->frags[i-1].size += len;
1083 } else if (i < MAX_SKB_FRAGS) {
1085 skb_fill_page_desc(skb, i, page, offset, len);
1091 if (skb->ip_summed == CHECKSUM_NONE) {
1093 csum = csum_page(page, offset, len);
1094 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1098 skb->data_len += len;
1105 inet->cork.length -= size;
1106 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1111 * Combined all pending IP fragments on the socket as one IP datagram
1112 * and push them out.
1114 int ip_push_pending_frames(struct sock *sk)
1116 struct sk_buff *skb, *tmp_skb;
1117 struct sk_buff **tail_skb;
1118 struct inet_opt *inet = inet_sk(sk);
1119 struct ip_options *opt = NULL;
1120 struct rtable *rt = inet->cork.rt;
1126 if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1128 tail_skb = &(skb_shinfo(skb)->frag_list);
1130 /* move skb->data to ip header from ext header */
1131 if (skb->data < skb->nh.raw)
1132 __skb_pull(skb, skb->nh.raw - skb->data);
1133 while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1134 __skb_pull(tmp_skb, skb->h.raw - skb->nh.raw);
1135 *tail_skb = tmp_skb;
1136 tail_skb = &(tmp_skb->next);
1137 skb->len += tmp_skb->len;
1138 skb->data_len += tmp_skb->len;
1139 skb->truesize += tmp_skb->truesize;
1140 __sock_put(tmp_skb->sk);
1141 tmp_skb->destructor = NULL;
1145 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1146 * to fragment the frame generated here. No matter, what transforms
1147 * how transforms change size of the packet, it will come out.
1149 if (inet->pmtudisc != IP_PMTUDISC_DO)
1152 /* DF bit is set when we want to see DF on outgoing frames.
1153 * If local_df is set too, we still allow to fragment this frame
1155 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1156 (!skb_shinfo(skb)->frag_list && ip_dont_fragment(sk, &rt->u.dst)))
1159 if (inet->cork.flags & IPCORK_OPT)
1160 opt = inet->cork.opt;
1162 if (rt->rt_type == RTN_MULTICAST)
1165 ttl = ip_select_ttl(inet, &rt->u.dst);
1167 iph = (struct iphdr *)skb->data;
1171 iph->ihl += opt->optlen>>2;
1172 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1174 iph->tos = inet->tos;
1175 iph->tot_len = htons(skb->len);
1178 __ip_select_ident(iph, &rt->u.dst, 0);
1180 iph->id = htons(inet->id++);
1183 iph->protocol = sk->sk_protocol;
1184 iph->saddr = rt->rt_src;
1185 iph->daddr = rt->rt_dst;
1188 skb->priority = sk->sk_priority;
1189 skb->dst = dst_clone(&rt->u.dst);
1191 /* Netfilter gets whole the not fragmented skb. */
1192 err = NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL,
1193 skb->dst->dev, dst_output);
1196 err = inet->recverr ? net_xmit_errno(err) : 0;
1202 inet->cork.flags &= ~IPCORK_OPT;
1203 if (inet->cork.opt) {
1204 kfree(inet->cork.opt);
1205 inet->cork.opt = NULL;
1207 if (inet->cork.rt) {
1208 ip_rt_put(inet->cork.rt);
1209 inet->cork.rt = NULL;
1214 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1219 * Throw away all pending data on the socket.
1221 void ip_flush_pending_frames(struct sock *sk)
1223 struct inet_opt *inet = inet_sk(sk);
1224 struct sk_buff *skb;
1226 while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1229 inet->cork.flags &= ~IPCORK_OPT;
1230 if (inet->cork.opt) {
1231 kfree(inet->cork.opt);
1232 inet->cork.opt = NULL;
1234 if (inet->cork.rt) {
1235 ip_rt_put(inet->cork.rt);
1236 inet->cork.rt = NULL;
1242 * Fetch data from kernel space and fill in checksum if needed.
1244 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1245 int len, int odd, struct sk_buff *skb)
1249 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1250 skb->csum = csum_block_add(skb->csum, csum, odd);
1255 * Generic function to send a packet as reply to another packet.
1256 * Used to send TCP resets so far. ICMP should use this function too.
1258 * Should run single threaded per socket because it uses the sock
1259 * structure to pass arguments.
1261 * LATER: switch from ip_build_xmit to ip_append_*
1263 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1266 struct inet_opt *inet = inet_sk(sk);
1268 struct ip_options opt;
1271 struct ipcm_cookie ipc;
1273 struct rtable *rt = (struct rtable*)skb->dst;
1275 if (ip_options_echo(&replyopts.opt, skb))
1278 daddr = ipc.addr = rt->rt_src;
1281 if (replyopts.opt.optlen) {
1282 ipc.opt = &replyopts.opt;
1285 daddr = replyopts.opt.faddr;
1289 struct flowi fl = { .nl_u = { .ip4_u =
1291 .saddr = rt->rt_spec_dst,
1292 .tos = RT_TOS(skb->nh.iph->tos) } },
1293 /* Not quite clean, but right. */
1295 { .sport = skb->h.th->dest,
1296 .dport = skb->h.th->source } },
1297 .proto = sk->sk_protocol };
1298 if (ip_route_output_key(&rt, &fl))
1302 /* And let IP do all the hard work.
1304 This chunk is not reenterable, hence spinlock.
1305 Note that it uses the fact, that this function is called
1306 with locally disabled BH and that sk cannot be already spinlocked.
1309 inet->tos = skb->nh.iph->tos;
1310 sk->sk_priority = skb->priority;
1311 sk->sk_protocol = skb->nh.iph->protocol;
1312 ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1313 &ipc, rt, MSG_DONTWAIT);
1314 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1315 if (arg->csumoffset >= 0)
1316 *((u16 *)skb->h.raw + arg->csumoffset) = csum_fold(csum_add(skb->csum, arg->csum));
1317 skb->ip_summed = CHECKSUM_NONE;
1318 ip_push_pending_frames(sk);
1327 * IP protocol layer initialiser
1330 static struct packet_type ip_packet_type = {
1331 .type = __constant_htons(ETH_P_IP),
1336 * IP registers the packet type and then calls the subprotocol initialisers
1339 void __init ip_init(void)
1341 dev_add_pack(&ip_packet_type);
1346 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1347 igmp_mc_proc_init();
1351 EXPORT_SYMBOL(ip_finish_output);
1352 EXPORT_SYMBOL(ip_fragment);
1353 EXPORT_SYMBOL(ip_generic_getfrag);
1354 EXPORT_SYMBOL(ip_queue_xmit);
1355 EXPORT_SYMBOL(ip_send_check);
1357 #ifdef CONFIG_SYSCTL
1358 EXPORT_SYMBOL(sysctl_ip_default_ttl);