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 *skb)
229 struct sock *sk = skb->sk;
230 struct rtable *rt = (struct rtable*)skb->dst;
231 struct net_device *dev = rt->u.dst.dev;
234 * If the indicated interface is up and running, send the packet.
236 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
239 skb->protocol = htons(ETH_P_IP);
242 * Multicasts are looped back for other local users
245 if (rt->rt_flags&RTCF_MULTICAST) {
246 if ((!sk || inet_sk(sk)->mc_loop)
247 #ifdef CONFIG_IP_MROUTE
248 /* Small optimization: do not loopback not local frames,
249 which returned after forwarding; they will be dropped
250 by ip_mr_input in any case.
251 Note, that local frames are looped back to be delivered
254 This check is duplicated in ip_mr_input at the moment.
256 && ((rt->rt_flags&RTCF_LOCAL) || !(IPCB(skb)->flags&IPSKB_FORWARDED))
259 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
261 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
263 ip_dev_loopback_xmit);
266 /* Multicasts with ttl 0 must not go beyond the host */
268 if (skb->nh.iph->ttl == 0) {
274 if (rt->rt_flags&RTCF_BROADCAST) {
275 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
277 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
278 newskb->dev, ip_dev_loopback_xmit);
281 if (skb->len > dst_pmtu(&rt->u.dst))
282 return ip_fragment(skb, ip_finish_output);
284 return ip_finish_output(skb);
287 int ip_output(struct sk_buff *skb)
289 IP_INC_STATS(IPSTATS_MIB_OUTREQUESTS);
291 if (skb->len > dst_pmtu(skb->dst) && !skb_shinfo(skb)->tso_size)
292 return ip_fragment(skb, ip_finish_output);
294 return ip_finish_output(skb);
297 int ip_queue_xmit(struct sk_buff *skb, int ipfragok)
299 struct sock *sk = skb->sk;
300 struct inet_opt *inet = inet_sk(sk);
301 struct ip_options *opt = inet->opt;
305 /* Skip all of this if the packet is already routed,
306 * f.e. by something like SCTP.
308 rt = (struct rtable *) skb->dst;
312 /* Make sure we can route this packet. */
313 rt = (struct rtable *)__sk_dst_check(sk, 0);
317 /* Use correct destination address if we have options. */
323 struct flowi fl = { .oif = sk->sk_bound_dev_if,
326 .saddr = inet->saddr,
327 .tos = RT_CONN_FLAGS(sk) } },
328 .proto = sk->sk_protocol,
330 { .sport = inet->sport,
331 .dport = inet->dport } } };
333 /* If this fails, retransmit mechanism of transport layer will
334 * keep trying until route appears or the connection times
337 if (ip_route_output_flow(&rt, &fl, sk, 0))
340 __sk_dst_set(sk, &rt->u.dst);
341 tcp_v4_setup_caps(sk, &rt->u.dst);
343 skb->dst = dst_clone(&rt->u.dst);
346 if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
349 /* OK, we know where to send it, allocate and build IP header. */
350 iph = (struct iphdr *) skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
351 *((__u16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
352 iph->tot_len = htons(skb->len);
353 if (ip_dont_fragment(sk, &rt->u.dst) && !ipfragok)
354 iph->frag_off = htons(IP_DF);
357 iph->ttl = ip_select_ttl(inet, &rt->u.dst);
358 iph->protocol = sk->sk_protocol;
359 iph->saddr = rt->rt_src;
360 iph->daddr = rt->rt_dst;
362 /* Transport layer set skb->h.foo itself. */
364 if (opt && opt->optlen) {
365 iph->ihl += opt->optlen >> 2;
366 ip_options_build(skb, opt, inet->daddr, rt, 0);
369 ip_select_ident_more(iph, &rt->u.dst, sk, skb_shinfo(skb)->tso_segs);
371 /* Add an IP checksum. */
374 skb->priority = sk->sk_priority;
376 return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
380 IP_INC_STATS(IPSTATS_MIB_OUTNOROUTES);
382 return -EHOSTUNREACH;
386 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
388 to->pkt_type = from->pkt_type;
389 to->priority = from->priority;
390 to->protocol = from->protocol;
391 to->security = from->security;
392 to->dst = dst_clone(from->dst);
395 /* Copy the flags to each fragment. */
396 IPCB(to)->flags = IPCB(from)->flags;
398 #ifdef CONFIG_NET_SCHED
399 to->tc_index = from->tc_index;
401 #ifdef CONFIG_NETFILTER
402 to->nfmark = from->nfmark;
403 to->nfcache = from->nfcache;
404 /* Connection association is same as pre-frag packet */
405 nf_conntrack_put(to->nfct);
406 to->nfct = from->nfct;
407 nf_conntrack_get(to->nfct);
408 to->nfctinfo = from->nfctinfo;
409 #ifdef CONFIG_BRIDGE_NETFILTER
410 nf_bridge_put(to->nf_bridge);
411 to->nf_bridge = from->nf_bridge;
412 nf_bridge_get(to->nf_bridge);
414 #ifdef CONFIG_NETFILTER_DEBUG
415 to->nf_debug = from->nf_debug;
421 * This IP datagram is too large to be sent in one piece. Break it up into
422 * smaller pieces (each of size equal to IP header plus
423 * a block of the data of the original IP data part) that will yet fit in a
424 * single device frame, and queue such a frame for sending.
427 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*))
432 struct net_device *dev;
433 struct sk_buff *skb2;
434 unsigned int mtu, hlen, left, len, ll_rs;
437 struct rtable *rt = (struct rtable*)skb->dst;
443 * Point into the IP datagram header.
448 if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
449 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
450 htonl(dst_pmtu(&rt->u.dst)));
456 * Setup starting values.
460 mtu = dst_pmtu(&rt->u.dst) - hlen; /* Size of data space */
462 /* When frag_list is given, use it. First, check its validity:
463 * some transformers could create wrong frag_list or break existing
464 * one, it is not prohibited. In this case fall back to copying.
466 * LATER: this step can be merged to real generation of fragments,
467 * we can switch to copy when see the first bad fragment.
469 if (skb_shinfo(skb)->frag_list) {
470 struct sk_buff *frag;
471 int first_len = skb_pagelen(skb);
473 if (first_len - hlen > mtu ||
474 ((first_len - hlen) & 7) ||
475 (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
479 for (frag = skb_shinfo(skb)->frag_list; frag; frag = frag->next) {
480 /* Correct geometry. */
481 if (frag->len > mtu ||
482 ((frag->len & 7) && frag->next) ||
483 skb_headroom(frag) < hlen)
486 /* Partially cloned skb? */
487 if (skb_shared(frag))
491 /* Everything is OK. Generate! */
495 frag = skb_shinfo(skb)->frag_list;
496 skb_shinfo(skb)->frag_list = NULL;
497 skb->data_len = first_len - skb_headlen(skb);
498 skb->len = first_len;
499 iph->tot_len = htons(first_len);
500 iph->frag_off |= htons(IP_MF);
504 /* Prepare header of the next frame,
505 * before previous one went down. */
507 frag->h.raw = frag->data;
508 frag->nh.raw = __skb_push(frag, hlen);
509 memcpy(frag->nh.raw, iph, hlen);
511 iph->tot_len = htons(frag->len);
512 ip_copy_metadata(frag, skb);
514 ip_options_fragment(frag);
515 offset += skb->len - hlen;
516 iph->frag_off = htons(offset>>3);
517 if (frag->next != NULL)
518 iph->frag_off |= htons(IP_MF);
519 /* Ready, complete checksum */
534 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
543 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
548 left = skb->len - hlen; /* Space per frame */
549 ptr = raw + hlen; /* Where to start from */
551 #ifdef CONFIG_BRIDGE_NETFILTER
552 /* for bridged IP traffic encapsulated inside f.e. a vlan header,
553 * we need to make room for the encapsulating header */
554 ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, nf_bridge_pad(skb));
555 mtu -= nf_bridge_pad(skb);
557 ll_rs = LL_RESERVED_SPACE(rt->u.dst.dev);
560 * Fragment the datagram.
563 offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
564 not_last_frag = iph->frag_off & htons(IP_MF);
567 * Keep copying data until we run out.
572 /* IF: it doesn't fit, use 'mtu' - the data space left */
575 /* IF: we are not sending upto and including the packet end
576 then align the next start on an eight byte boundary */
584 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
585 NETDEBUG(printk(KERN_INFO "IP: frag: no memory for new fragment!\n"));
591 * Set up data on packet
594 ip_copy_metadata(skb2, skb);
595 skb_reserve(skb2, ll_rs);
596 skb_put(skb2, len + hlen);
597 skb2->nh.raw = skb2->data;
598 skb2->h.raw = skb2->data + hlen;
601 * Charge the memory for the fragment to any owner
606 skb_set_owner_w(skb2, skb->sk);
609 * Copy the packet header into the new buffer.
612 memcpy(skb2->nh.raw, skb->data, hlen);
615 * Copy a block of the IP datagram.
617 if (skb_copy_bits(skb, ptr, skb2->h.raw, len))
622 * Fill in the new header fields.
625 iph->frag_off = htons((offset >> 3));
627 /* ANK: dirty, but effective trick. Upgrade options only if
628 * the segment to be fragmented was THE FIRST (otherwise,
629 * options are already fixed) and make it ONCE
630 * on the initial skb, so that all the following fragments
631 * will inherit fixed options.
634 ip_options_fragment(skb);
637 * Added AC : If we are fragmenting a fragment that's not the
638 * last fragment then keep MF on each bit
640 if (left > 0 || not_last_frag)
641 iph->frag_off |= htons(IP_MF);
646 * Put this fragment into the sending queue.
649 IP_INC_STATS(IPSTATS_MIB_FRAGCREATES);
651 iph->tot_len = htons(len + hlen);
660 IP_INC_STATS(IPSTATS_MIB_FRAGOKS);
665 IP_INC_STATS(IPSTATS_MIB_FRAGFAILS);
670 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
672 struct iovec *iov = from;
674 if (skb->ip_summed == CHECKSUM_HW) {
675 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
678 unsigned int csum = 0;
679 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
681 skb->csum = csum_block_add(skb->csum, csum, odd);
686 static inline unsigned int
687 csum_page(struct page *page, int offset, int copy)
692 csum = csum_partial(kaddr + offset, copy, 0);
698 * ip_append_data() and ip_append_page() can make one large IP datagram
699 * from many pieces of data. Each pieces will be holded on the socket
700 * until ip_push_pending_frames() is called. Each piece can be a page
703 * Not only UDP, other transport protocols - e.g. raw sockets - can use
704 * this interface potentially.
706 * LATER: length must be adjusted by pad at tail, when it is required.
708 int ip_append_data(struct sock *sk,
709 int getfrag(void *from, char *to, int offset, int len,
710 int odd, struct sk_buff *skb),
711 void *from, int length, int transhdrlen,
712 struct ipcm_cookie *ipc, struct rtable *rt,
715 struct inet_opt *inet = inet_sk(sk);
718 struct ip_options *opt = NULL;
725 unsigned int maxfraglen, fragheaderlen;
726 int csummode = CHECKSUM_NONE;
731 if (skb_queue_empty(&sk->sk_write_queue)) {
737 if (inet->cork.opt == NULL) {
738 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
739 if (unlikely(inet->cork.opt == NULL))
742 memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
743 inet->cork.flags |= IPCORK_OPT;
744 inet->cork.addr = ipc->addr;
746 dst_hold(&rt->u.dst);
747 inet->cork.fragsize = mtu = dst_pmtu(&rt->u.dst);
749 inet->cork.length = 0;
750 sk->sk_sndmsg_page = NULL;
751 sk->sk_sndmsg_off = 0;
752 if ((exthdrlen = rt->u.dst.header_len) != 0) {
754 transhdrlen += exthdrlen;
758 if (inet->cork.flags & IPCORK_OPT)
759 opt = inet->cork.opt;
763 mtu = inet->cork.fragsize;
765 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
767 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
768 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
770 if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
771 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu-exthdrlen);
776 * transhdrlen > 0 means that this is the first fragment and we wish
777 * it won't be fragmented in the future.
780 length + fragheaderlen <= mtu &&
781 rt->u.dst.dev->features&(NETIF_F_IP_CSUM|NETIF_F_NO_CSUM|NETIF_F_HW_CSUM) &&
783 csummode = CHECKSUM_HW;
785 inet->cork.length += length;
787 /* So, what's going on in the loop below?
789 * We use calculated fragment length to generate chained skb,
790 * each of segments is IP fragment ready for sending to network after
791 * adding appropriate IP header.
794 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
798 /* Check if the remaining data fits into current packet. */
799 copy = mtu - skb->len;
801 copy = maxfraglen - skb->len;
804 unsigned int datalen;
805 unsigned int fraglen;
806 unsigned int fraggap;
807 unsigned int alloclen;
808 struct sk_buff *skb_prev;
812 fraggap = skb_prev->len - maxfraglen;
817 * If remaining data exceeds the mtu,
818 * we know we need more fragment(s).
820 datalen = length + fraggap;
821 if (datalen > mtu - fragheaderlen)
822 datalen = maxfraglen - fragheaderlen;
823 fraglen = datalen + fragheaderlen;
825 if ((flags & MSG_MORE) &&
826 !(rt->u.dst.dev->features&NETIF_F_SG))
829 alloclen = datalen + fragheaderlen;
831 /* The last fragment gets additional space at tail.
832 * Note, with MSG_MORE we overallocate on fragments,
833 * because we have no idea what fragment will be
836 if (datalen == length)
837 alloclen += rt->u.dst.trailer_len;
840 skb = sock_alloc_send_skb(sk,
841 alloclen + hh_len + 15,
842 (flags & MSG_DONTWAIT), &err);
845 if (atomic_read(&sk->sk_wmem_alloc) <=
847 skb = sock_wmalloc(sk,
848 alloclen + hh_len + 15, 1,
850 if (unlikely(skb == NULL))
857 * Fill in the control structures
859 skb->ip_summed = csummode;
861 skb_reserve(skb, hh_len);
864 * Find where to start putting bytes.
866 data = skb_put(skb, fraglen);
867 skb->nh.raw = data + exthdrlen;
868 data += fragheaderlen;
869 skb->h.raw = data + exthdrlen;
872 skb->csum = skb_copy_and_csum_bits(
873 skb_prev, maxfraglen,
874 data + transhdrlen, fraggap, 0);
875 skb_prev->csum = csum_sub(skb_prev->csum,
878 skb_trim(skb_prev, maxfraglen);
881 copy = datalen - transhdrlen - fraggap;
882 if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
889 length -= datalen - fraggap;
892 csummode = CHECKSUM_NONE;
895 * Put the packet on the pending queue.
897 __skb_queue_tail(&sk->sk_write_queue, skb);
904 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
908 if (getfrag(from, skb_put(skb, copy),
909 offset, copy, off, skb) < 0) {
910 __skb_trim(skb, off);
915 int i = skb_shinfo(skb)->nr_frags;
916 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
917 struct page *page = sk->sk_sndmsg_page;
918 int off = sk->sk_sndmsg_off;
921 if (page && (left = PAGE_SIZE - off) > 0) {
924 if (page != frag->page) {
925 if (i == MAX_SKB_FRAGS) {
930 skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
931 frag = &skb_shinfo(skb)->frags[i];
933 } else if (i < MAX_SKB_FRAGS) {
934 if (copy > PAGE_SIZE)
936 page = alloc_pages(sk->sk_allocation, 0);
941 sk->sk_sndmsg_page = page;
942 sk->sk_sndmsg_off = 0;
944 skb_fill_page_desc(skb, i, page, 0, 0);
945 frag = &skb_shinfo(skb)->frags[i];
946 skb->truesize += PAGE_SIZE;
947 atomic_add(PAGE_SIZE, &sk->sk_wmem_alloc);
952 if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
956 sk->sk_sndmsg_off += copy;
959 skb->data_len += copy;
968 inet->cork.length -= length;
969 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
973 ssize_t ip_append_page(struct sock *sk, struct page *page,
974 int offset, size_t size, int flags)
976 struct inet_opt *inet = inet_sk(sk);
979 struct ip_options *opt = NULL;
984 unsigned int maxfraglen, fragheaderlen, fraggap;
992 if (skb_queue_empty(&sk->sk_write_queue))
996 if (inet->cork.flags & IPCORK_OPT)
997 opt = inet->cork.opt;
999 if (!(rt->u.dst.dev->features&NETIF_F_SG))
1002 hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1003 mtu = inet->cork.fragsize;
1005 fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1006 maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1008 if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1009 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu);
1013 if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1016 inet->cork.length += size;
1021 /* Check if the remaining data fits into current packet. */
1022 len = mtu - skb->len;
1024 len = maxfraglen - skb->len;
1026 struct sk_buff *skb_prev;
1033 fraggap = skb_prev->len - maxfraglen;
1037 alloclen = fragheaderlen + hh_len + fraggap + 15;
1038 skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1039 if (unlikely(!skb)) {
1045 * Fill in the control structures
1047 skb->ip_summed = CHECKSUM_NONE;
1049 skb_reserve(skb, hh_len);
1052 * Find where to start putting bytes.
1054 data = skb_put(skb, fragheaderlen + fraggap);
1055 skb->nh.iph = iph = (struct iphdr *)data;
1056 data += fragheaderlen;
1060 skb->csum = skb_copy_and_csum_bits(
1061 skb_prev, maxfraglen,
1063 skb_prev->csum = csum_sub(skb_prev->csum,
1065 skb_trim(skb_prev, maxfraglen);
1069 * Put the packet on the pending queue.
1071 __skb_queue_tail(&sk->sk_write_queue, skb);
1075 i = skb_shinfo(skb)->nr_frags;
1078 if (skb_can_coalesce(skb, i, page, offset)) {
1079 skb_shinfo(skb)->frags[i-1].size += len;
1080 } else if (i < MAX_SKB_FRAGS) {
1082 skb_fill_page_desc(skb, i, page, offset, len);
1088 if (skb->ip_summed == CHECKSUM_NONE) {
1090 csum = csum_page(page, offset, len);
1091 skb->csum = csum_block_add(skb->csum, csum, skb->len);
1095 skb->data_len += len;
1102 inet->cork.length -= size;
1103 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1108 * Combined all pending IP fragments on the socket as one IP datagram
1109 * and push them out.
1111 int ip_push_pending_frames(struct sock *sk)
1113 struct sk_buff *skb, *tmp_skb;
1114 struct sk_buff **tail_skb;
1115 struct inet_opt *inet = inet_sk(sk);
1116 struct ip_options *opt = NULL;
1117 struct rtable *rt = inet->cork.rt;
1123 if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1125 tail_skb = &(skb_shinfo(skb)->frag_list);
1127 /* move skb->data to ip header from ext header */
1128 if (skb->data < skb->nh.raw)
1129 __skb_pull(skb, skb->nh.raw - skb->data);
1130 while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1131 __skb_pull(tmp_skb, skb->h.raw - skb->nh.raw);
1132 *tail_skb = tmp_skb;
1133 tail_skb = &(tmp_skb->next);
1134 skb->len += tmp_skb->len;
1135 skb->data_len += tmp_skb->len;
1136 skb->truesize += tmp_skb->truesize;
1137 __sock_put(tmp_skb->sk);
1138 tmp_skb->destructor = NULL;
1142 /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1143 * to fragment the frame generated here. No matter, what transforms
1144 * how transforms change size of the packet, it will come out.
1146 if (inet->pmtudisc != IP_PMTUDISC_DO)
1149 /* DF bit is set when we want to see DF on outgoing frames.
1150 * If local_df is set too, we still allow to fragment this frame
1152 if (inet->pmtudisc == IP_PMTUDISC_DO ||
1153 (!skb_shinfo(skb)->frag_list && ip_dont_fragment(sk, &rt->u.dst)))
1156 if (inet->cork.flags & IPCORK_OPT)
1157 opt = inet->cork.opt;
1159 if (rt->rt_type == RTN_MULTICAST)
1162 ttl = ip_select_ttl(inet, &rt->u.dst);
1164 iph = (struct iphdr *)skb->data;
1168 iph->ihl += opt->optlen>>2;
1169 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1171 iph->tos = inet->tos;
1172 iph->tot_len = htons(skb->len);
1175 __ip_select_ident(iph, &rt->u.dst, 0);
1177 iph->id = htons(inet->id++);
1180 iph->protocol = sk->sk_protocol;
1181 iph->saddr = rt->rt_src;
1182 iph->daddr = rt->rt_dst;
1185 skb->priority = sk->sk_priority;
1186 skb->dst = dst_clone(&rt->u.dst);
1188 /* Netfilter gets whole the not fragmented skb. */
1189 err = NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL,
1190 skb->dst->dev, dst_output);
1193 err = inet->recverr ? net_xmit_errno(err) : 0;
1199 inet->cork.flags &= ~IPCORK_OPT;
1200 if (inet->cork.opt) {
1201 kfree(inet->cork.opt);
1202 inet->cork.opt = NULL;
1204 if (inet->cork.rt) {
1205 ip_rt_put(inet->cork.rt);
1206 inet->cork.rt = NULL;
1211 IP_INC_STATS(IPSTATS_MIB_OUTDISCARDS);
1216 * Throw away all pending data on the socket.
1218 void ip_flush_pending_frames(struct sock *sk)
1220 struct inet_opt *inet = inet_sk(sk);
1221 struct sk_buff *skb;
1223 while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1226 inet->cork.flags &= ~IPCORK_OPT;
1227 if (inet->cork.opt) {
1228 kfree(inet->cork.opt);
1229 inet->cork.opt = NULL;
1231 if (inet->cork.rt) {
1232 ip_rt_put(inet->cork.rt);
1233 inet->cork.rt = NULL;
1239 * Fetch data from kernel space and fill in checksum if needed.
1241 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1242 int len, int odd, struct sk_buff *skb)
1246 csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1247 skb->csum = csum_block_add(skb->csum, csum, odd);
1252 * Generic function to send a packet as reply to another packet.
1253 * Used to send TCP resets so far. ICMP should use this function too.
1255 * Should run single threaded per socket because it uses the sock
1256 * structure to pass arguments.
1258 * LATER: switch from ip_build_xmit to ip_append_*
1260 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1263 struct inet_opt *inet = inet_sk(sk);
1265 struct ip_options opt;
1268 struct ipcm_cookie ipc;
1270 struct rtable *rt = (struct rtable*)skb->dst;
1272 if (ip_options_echo(&replyopts.opt, skb))
1275 daddr = ipc.addr = rt->rt_src;
1278 if (replyopts.opt.optlen) {
1279 ipc.opt = &replyopts.opt;
1282 daddr = replyopts.opt.faddr;
1286 struct flowi fl = { .nl_u = { .ip4_u =
1288 .saddr = rt->rt_spec_dst,
1289 .tos = RT_TOS(skb->nh.iph->tos) } },
1290 /* Not quite clean, but right. */
1292 { .sport = skb->h.th->dest,
1293 .dport = skb->h.th->source } },
1294 .proto = sk->sk_protocol };
1295 if (ip_route_output_key(&rt, &fl))
1299 /* And let IP do all the hard work.
1301 This chunk is not reenterable, hence spinlock.
1302 Note that it uses the fact, that this function is called
1303 with locally disabled BH and that sk cannot be already spinlocked.
1306 inet->tos = skb->nh.iph->tos;
1307 sk->sk_priority = skb->priority;
1308 sk->sk_protocol = skb->nh.iph->protocol;
1309 ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1310 &ipc, rt, MSG_DONTWAIT);
1311 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1312 if (arg->csumoffset >= 0)
1313 *((u16 *)skb->h.raw + arg->csumoffset) = csum_fold(csum_add(skb->csum, arg->csum));
1314 skb->ip_summed = CHECKSUM_NONE;
1315 ip_push_pending_frames(sk);
1324 * IP protocol layer initialiser
1327 static struct packet_type ip_packet_type = {
1328 .type = __constant_htons(ETH_P_IP),
1333 * IP registers the packet type and then calls the subprotocol initialisers
1336 void __init ip_init(void)
1338 dev_add_pack(&ip_packet_type);
1343 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1344 igmp_mc_proc_init();
1348 EXPORT_SYMBOL(ip_finish_output);
1349 EXPORT_SYMBOL(ip_fragment);
1350 EXPORT_SYMBOL(ip_generic_getfrag);
1351 EXPORT_SYMBOL(ip_queue_xmit);
1352 EXPORT_SYMBOL(ip_send_check);
1354 #ifdef CONFIG_SYSCTL
1355 EXPORT_SYMBOL(sysctl_ip_default_ttl);