[NETLINK]: Do precise netlink message allocations where possible
[linux-2.6] / net / ipv4 / tcp_ipv4.c
1 /*
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.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Version:     $Id: tcp_ipv4.c,v 1.240 2002/02/01 22:01:04 davem Exp $
9  *
10  *              IPv4 specific functions
11  *
12  *
13  *              code split from:
14  *              linux/ipv4/tcp.c
15  *              linux/ipv4/tcp_input.c
16  *              linux/ipv4/tcp_output.c
17  *
18  *              See tcp.c for author information
19  *
20  *      This program is free software; you can redistribute it and/or
21  *      modify it under the terms of the GNU General Public License
22  *      as published by the Free Software Foundation; either version
23  *      2 of the License, or (at your option) any later version.
24  */
25
26 /*
27  * Changes:
28  *              David S. Miller :       New socket lookup architecture.
29  *                                      This code is dedicated to John Dyson.
30  *              David S. Miller :       Change semantics of established hash,
31  *                                      half is devoted to TIME_WAIT sockets
32  *                                      and the rest go in the other half.
33  *              Andi Kleen :            Add support for syncookies and fixed
34  *                                      some bugs: ip options weren't passed to
35  *                                      the TCP layer, missed a check for an
36  *                                      ACK bit.
37  *              Andi Kleen :            Implemented fast path mtu discovery.
38  *                                      Fixed many serious bugs in the
39  *                                      request_sock handling and moved
40  *                                      most of it into the af independent code.
41  *                                      Added tail drop and some other bugfixes.
42  *                                      Added new listen semantics.
43  *              Mike McLagan    :       Routing by source
44  *      Juan Jose Ciarlante:            ip_dynaddr bits
45  *              Andi Kleen:             various fixes.
46  *      Vitaly E. Lavrov        :       Transparent proxy revived after year
47  *                                      coma.
48  *      Andi Kleen              :       Fix new listen.
49  *      Andi Kleen              :       Fix accept error reporting.
50  *      YOSHIFUJI Hideaki @USAGI and:   Support IPV6_V6ONLY socket option, which
51  *      Alexey Kuznetsov                allow both IPv4 and IPv6 sockets to bind
52  *                                      a single port at the same time.
53  */
54
55
56 #include <linux/types.h>
57 #include <linux/fcntl.h>
58 #include <linux/module.h>
59 #include <linux/random.h>
60 #include <linux/cache.h>
61 #include <linux/jhash.h>
62 #include <linux/init.h>
63 #include <linux/times.h>
64
65 #include <net/icmp.h>
66 #include <net/inet_hashtables.h>
67 #include <net/tcp.h>
68 #include <net/transp_v6.h>
69 #include <net/ipv6.h>
70 #include <net/inet_common.h>
71 #include <net/timewait_sock.h>
72 #include <net/xfrm.h>
73 #include <net/netdma.h>
74
75 #include <linux/inet.h>
76 #include <linux/ipv6.h>
77 #include <linux/stddef.h>
78 #include <linux/proc_fs.h>
79 #include <linux/seq_file.h>
80
81 int sysctl_tcp_tw_reuse __read_mostly;
82 int sysctl_tcp_low_latency __read_mostly;
83
84 /* Check TCP sequence numbers in ICMP packets. */
85 #define ICMP_MIN_LENGTH 8
86
87 /* Socket used for sending RSTs */
88 static struct socket *tcp_socket;
89
90 void tcp_v4_send_check(struct sock *sk, int len, struct sk_buff *skb);
91
92 struct inet_hashinfo __cacheline_aligned tcp_hashinfo = {
93         .lhash_lock     = __RW_LOCK_UNLOCKED(tcp_hashinfo.lhash_lock),
94         .lhash_users    = ATOMIC_INIT(0),
95         .lhash_wait     = __WAIT_QUEUE_HEAD_INITIALIZER(tcp_hashinfo.lhash_wait),
96 };
97
98 static int tcp_v4_get_port(struct sock *sk, unsigned short snum)
99 {
100         return inet_csk_get_port(&tcp_hashinfo, sk, snum,
101                                  inet_csk_bind_conflict);
102 }
103
104 static void tcp_v4_hash(struct sock *sk)
105 {
106         inet_hash(&tcp_hashinfo, sk);
107 }
108
109 void tcp_unhash(struct sock *sk)
110 {
111         inet_unhash(&tcp_hashinfo, sk);
112 }
113
114 static inline __u32 tcp_v4_init_sequence(struct sk_buff *skb)
115 {
116         return secure_tcp_sequence_number(skb->nh.iph->daddr,
117                                           skb->nh.iph->saddr,
118                                           skb->h.th->dest,
119                                           skb->h.th->source);
120 }
121
122 int tcp_twsk_unique(struct sock *sk, struct sock *sktw, void *twp)
123 {
124         const struct tcp_timewait_sock *tcptw = tcp_twsk(sktw);
125         struct tcp_sock *tp = tcp_sk(sk);
126
127         /* With PAWS, it is safe from the viewpoint
128            of data integrity. Even without PAWS it is safe provided sequence
129            spaces do not overlap i.e. at data rates <= 80Mbit/sec.
130
131            Actually, the idea is close to VJ's one, only timestamp cache is
132            held not per host, but per port pair and TW bucket is used as state
133            holder.
134
135            If TW bucket has been already destroyed we fall back to VJ's scheme
136            and use initial timestamp retrieved from peer table.
137          */
138         if (tcptw->tw_ts_recent_stamp &&
139             (twp == NULL || (sysctl_tcp_tw_reuse &&
140                              xtime.tv_sec - tcptw->tw_ts_recent_stamp > 1))) {
141                 tp->write_seq = tcptw->tw_snd_nxt + 65535 + 2;
142                 if (tp->write_seq == 0)
143                         tp->write_seq = 1;
144                 tp->rx_opt.ts_recent       = tcptw->tw_ts_recent;
145                 tp->rx_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
146                 sock_hold(sktw);
147                 return 1;
148         }
149
150         return 0;
151 }
152
153 EXPORT_SYMBOL_GPL(tcp_twsk_unique);
154
155 /* This will initiate an outgoing connection. */
156 int tcp_v4_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
157 {
158         struct inet_sock *inet = inet_sk(sk);
159         struct tcp_sock *tp = tcp_sk(sk);
160         struct sockaddr_in *usin = (struct sockaddr_in *)uaddr;
161         struct rtable *rt;
162         __be32 daddr, nexthop;
163         int tmp;
164         int err;
165
166         if (addr_len < sizeof(struct sockaddr_in))
167                 return -EINVAL;
168
169         if (usin->sin_family != AF_INET)
170                 return -EAFNOSUPPORT;
171
172         nexthop = daddr = usin->sin_addr.s_addr;
173         if (inet->opt && inet->opt->srr) {
174                 if (!daddr)
175                         return -EINVAL;
176                 nexthop = inet->opt->faddr;
177         }
178
179         tmp = ip_route_connect(&rt, nexthop, inet->saddr,
180                                RT_CONN_FLAGS(sk), sk->sk_bound_dev_if,
181                                IPPROTO_TCP,
182                                inet->sport, usin->sin_port, sk);
183         if (tmp < 0)
184                 return tmp;
185
186         if (rt->rt_flags & (RTCF_MULTICAST | RTCF_BROADCAST)) {
187                 ip_rt_put(rt);
188                 return -ENETUNREACH;
189         }
190
191         if (!inet->opt || !inet->opt->srr)
192                 daddr = rt->rt_dst;
193
194         if (!inet->saddr)
195                 inet->saddr = rt->rt_src;
196         inet->rcv_saddr = inet->saddr;
197
198         if (tp->rx_opt.ts_recent_stamp && inet->daddr != daddr) {
199                 /* Reset inherited state */
200                 tp->rx_opt.ts_recent       = 0;
201                 tp->rx_opt.ts_recent_stamp = 0;
202                 tp->write_seq              = 0;
203         }
204
205         if (tcp_death_row.sysctl_tw_recycle &&
206             !tp->rx_opt.ts_recent_stamp && rt->rt_dst == daddr) {
207                 struct inet_peer *peer = rt_get_peer(rt);
208
209                 /* VJ's idea. We save last timestamp seen from
210                  * the destination in peer table, when entering state TIME-WAIT
211                  * and initialize rx_opt.ts_recent from it, when trying new connection.
212                  */
213
214                 if (peer && peer->tcp_ts_stamp + TCP_PAWS_MSL >= xtime.tv_sec) {
215                         tp->rx_opt.ts_recent_stamp = peer->tcp_ts_stamp;
216                         tp->rx_opt.ts_recent = peer->tcp_ts;
217                 }
218         }
219
220         inet->dport = usin->sin_port;
221         inet->daddr = daddr;
222
223         inet_csk(sk)->icsk_ext_hdr_len = 0;
224         if (inet->opt)
225                 inet_csk(sk)->icsk_ext_hdr_len = inet->opt->optlen;
226
227         tp->rx_opt.mss_clamp = 536;
228
229         /* Socket identity is still unknown (sport may be zero).
230          * However we set state to SYN-SENT and not releasing socket
231          * lock select source port, enter ourselves into the hash tables and
232          * complete initialization after this.
233          */
234         tcp_set_state(sk, TCP_SYN_SENT);
235         err = inet_hash_connect(&tcp_death_row, sk);
236         if (err)
237                 goto failure;
238
239         err = ip_route_newports(&rt, IPPROTO_TCP, inet->sport, inet->dport, sk);
240         if (err)
241                 goto failure;
242
243         /* OK, now commit destination to socket.  */
244         sk->sk_gso_type = SKB_GSO_TCPV4;
245         sk_setup_caps(sk, &rt->u.dst);
246
247         if (!tp->write_seq)
248                 tp->write_seq = secure_tcp_sequence_number(inet->saddr,
249                                                            inet->daddr,
250                                                            inet->sport,
251                                                            usin->sin_port);
252
253         inet->id = tp->write_seq ^ jiffies;
254
255         err = tcp_connect(sk);
256         rt = NULL;
257         if (err)
258                 goto failure;
259
260         return 0;
261
262 failure:
263         /* This unhashes the socket and releases the local port, if necessary. */
264         tcp_set_state(sk, TCP_CLOSE);
265         ip_rt_put(rt);
266         sk->sk_route_caps = 0;
267         inet->dport = 0;
268         return err;
269 }
270
271 /*
272  * This routine does path mtu discovery as defined in RFC1191.
273  */
274 static void do_pmtu_discovery(struct sock *sk, struct iphdr *iph, u32 mtu)
275 {
276         struct dst_entry *dst;
277         struct inet_sock *inet = inet_sk(sk);
278
279         /* We are not interested in TCP_LISTEN and open_requests (SYN-ACKs
280          * send out by Linux are always <576bytes so they should go through
281          * unfragmented).
282          */
283         if (sk->sk_state == TCP_LISTEN)
284                 return;
285
286         /* We don't check in the destentry if pmtu discovery is forbidden
287          * on this route. We just assume that no packet_to_big packets
288          * are send back when pmtu discovery is not active.
289          * There is a small race when the user changes this flag in the
290          * route, but I think that's acceptable.
291          */
292         if ((dst = __sk_dst_check(sk, 0)) == NULL)
293                 return;
294
295         dst->ops->update_pmtu(dst, mtu);
296
297         /* Something is about to be wrong... Remember soft error
298          * for the case, if this connection will not able to recover.
299          */
300         if (mtu < dst_mtu(dst) && ip_dont_fragment(sk, dst))
301                 sk->sk_err_soft = EMSGSIZE;
302
303         mtu = dst_mtu(dst);
304
305         if (inet->pmtudisc != IP_PMTUDISC_DONT &&
306             inet_csk(sk)->icsk_pmtu_cookie > mtu) {
307                 tcp_sync_mss(sk, mtu);
308
309                 /* Resend the TCP packet because it's
310                  * clear that the old packet has been
311                  * dropped. This is the new "fast" path mtu
312                  * discovery.
313                  */
314                 tcp_simple_retransmit(sk);
315         } /* else let the usual retransmit timer handle it */
316 }
317
318 /*
319  * This routine is called by the ICMP module when it gets some
320  * sort of error condition.  If err < 0 then the socket should
321  * be closed and the error returned to the user.  If err > 0
322  * it's just the icmp type << 8 | icmp code.  After adjustment
323  * header points to the first 8 bytes of the tcp header.  We need
324  * to find the appropriate port.
325  *
326  * The locking strategy used here is very "optimistic". When
327  * someone else accesses the socket the ICMP is just dropped
328  * and for some paths there is no check at all.
329  * A more general error queue to queue errors for later handling
330  * is probably better.
331  *
332  */
333
334 void tcp_v4_err(struct sk_buff *skb, u32 info)
335 {
336         struct iphdr *iph = (struct iphdr *)skb->data;
337         struct tcphdr *th = (struct tcphdr *)(skb->data + (iph->ihl << 2));
338         struct tcp_sock *tp;
339         struct inet_sock *inet;
340         int type = skb->h.icmph->type;
341         int code = skb->h.icmph->code;
342         struct sock *sk;
343         __u32 seq;
344         int err;
345
346         if (skb->len < (iph->ihl << 2) + 8) {
347                 ICMP_INC_STATS_BH(ICMP_MIB_INERRORS);
348                 return;
349         }
350
351         sk = inet_lookup(&tcp_hashinfo, iph->daddr, th->dest, iph->saddr,
352                          th->source, inet_iif(skb));
353         if (!sk) {
354                 ICMP_INC_STATS_BH(ICMP_MIB_INERRORS);
355                 return;
356         }
357         if (sk->sk_state == TCP_TIME_WAIT) {
358                 inet_twsk_put(inet_twsk(sk));
359                 return;
360         }
361
362         bh_lock_sock(sk);
363         /* If too many ICMPs get dropped on busy
364          * servers this needs to be solved differently.
365          */
366         if (sock_owned_by_user(sk))
367                 NET_INC_STATS_BH(LINUX_MIB_LOCKDROPPEDICMPS);
368
369         if (sk->sk_state == TCP_CLOSE)
370                 goto out;
371
372         tp = tcp_sk(sk);
373         seq = ntohl(th->seq);
374         if (sk->sk_state != TCP_LISTEN &&
375             !between(seq, tp->snd_una, tp->snd_nxt)) {
376                 NET_INC_STATS_BH(LINUX_MIB_OUTOFWINDOWICMPS);
377                 goto out;
378         }
379
380         switch (type) {
381         case ICMP_SOURCE_QUENCH:
382                 /* Just silently ignore these. */
383                 goto out;
384         case ICMP_PARAMETERPROB:
385                 err = EPROTO;
386                 break;
387         case ICMP_DEST_UNREACH:
388                 if (code > NR_ICMP_UNREACH)
389                         goto out;
390
391                 if (code == ICMP_FRAG_NEEDED) { /* PMTU discovery (RFC1191) */
392                         if (!sock_owned_by_user(sk))
393                                 do_pmtu_discovery(sk, iph, info);
394                         goto out;
395                 }
396
397                 err = icmp_err_convert[code].errno;
398                 break;
399         case ICMP_TIME_EXCEEDED:
400                 err = EHOSTUNREACH;
401                 break;
402         default:
403                 goto out;
404         }
405
406         switch (sk->sk_state) {
407                 struct request_sock *req, **prev;
408         case TCP_LISTEN:
409                 if (sock_owned_by_user(sk))
410                         goto out;
411
412                 req = inet_csk_search_req(sk, &prev, th->dest,
413                                           iph->daddr, iph->saddr);
414                 if (!req)
415                         goto out;
416
417                 /* ICMPs are not backlogged, hence we cannot get
418                    an established socket here.
419                  */
420                 BUG_TRAP(!req->sk);
421
422                 if (seq != tcp_rsk(req)->snt_isn) {
423                         NET_INC_STATS_BH(LINUX_MIB_OUTOFWINDOWICMPS);
424                         goto out;
425                 }
426
427                 /*
428                  * Still in SYN_RECV, just remove it silently.
429                  * There is no good way to pass the error to the newly
430                  * created socket, and POSIX does not want network
431                  * errors returned from accept().
432                  */
433                 inet_csk_reqsk_queue_drop(sk, req, prev);
434                 goto out;
435
436         case TCP_SYN_SENT:
437         case TCP_SYN_RECV:  /* Cannot happen.
438                                It can f.e. if SYNs crossed.
439                              */
440                 if (!sock_owned_by_user(sk)) {
441                         sk->sk_err = err;
442
443                         sk->sk_error_report(sk);
444
445                         tcp_done(sk);
446                 } else {
447                         sk->sk_err_soft = err;
448                 }
449                 goto out;
450         }
451
452         /* If we've already connected we will keep trying
453          * until we time out, or the user gives up.
454          *
455          * rfc1122 4.2.3.9 allows to consider as hard errors
456          * only PROTO_UNREACH and PORT_UNREACH (well, FRAG_FAILED too,
457          * but it is obsoleted by pmtu discovery).
458          *
459          * Note, that in modern internet, where routing is unreliable
460          * and in each dark corner broken firewalls sit, sending random
461          * errors ordered by their masters even this two messages finally lose
462          * their original sense (even Linux sends invalid PORT_UNREACHs)
463          *
464          * Now we are in compliance with RFCs.
465          *                                                      --ANK (980905)
466          */
467
468         inet = inet_sk(sk);
469         if (!sock_owned_by_user(sk) && inet->recverr) {
470                 sk->sk_err = err;
471                 sk->sk_error_report(sk);
472         } else  { /* Only an error on timeout */
473                 sk->sk_err_soft = err;
474         }
475
476 out:
477         bh_unlock_sock(sk);
478         sock_put(sk);
479 }
480
481 /* This routine computes an IPv4 TCP checksum. */
482 void tcp_v4_send_check(struct sock *sk, int len, struct sk_buff *skb)
483 {
484         struct inet_sock *inet = inet_sk(sk);
485         struct tcphdr *th = skb->h.th;
486
487         if (skb->ip_summed == CHECKSUM_PARTIAL) {
488                 th->check = ~tcp_v4_check(th, len, inet->saddr, inet->daddr, 0);
489                 skb->csum = offsetof(struct tcphdr, check);
490         } else {
491                 th->check = tcp_v4_check(th, len, inet->saddr, inet->daddr,
492                                          csum_partial((char *)th,
493                                                       th->doff << 2,
494                                                       skb->csum));
495         }
496 }
497
498 int tcp_v4_gso_send_check(struct sk_buff *skb)
499 {
500         struct iphdr *iph;
501         struct tcphdr *th;
502
503         if (!pskb_may_pull(skb, sizeof(*th)))
504                 return -EINVAL;
505
506         iph = skb->nh.iph;
507         th = skb->h.th;
508
509         th->check = 0;
510         th->check = ~tcp_v4_check(th, skb->len, iph->saddr, iph->daddr, 0);
511         skb->csum = offsetof(struct tcphdr, check);
512         skb->ip_summed = CHECKSUM_PARTIAL;
513         return 0;
514 }
515
516 /*
517  *      This routine will send an RST to the other tcp.
518  *
519  *      Someone asks: why I NEVER use socket parameters (TOS, TTL etc.)
520  *                    for reset.
521  *      Answer: if a packet caused RST, it is not for a socket
522  *              existing in our system, if it is matched to a socket,
523  *              it is just duplicate segment or bug in other side's TCP.
524  *              So that we build reply only basing on parameters
525  *              arrived with segment.
526  *      Exception: precedence violation. We do not implement it in any case.
527  */
528
529 static void tcp_v4_send_reset(struct sk_buff *skb)
530 {
531         struct tcphdr *th = skb->h.th;
532         struct tcphdr rth;
533         struct ip_reply_arg arg;
534
535         /* Never send a reset in response to a reset. */
536         if (th->rst)
537                 return;
538
539         if (((struct rtable *)skb->dst)->rt_type != RTN_LOCAL)
540                 return;
541
542         /* Swap the send and the receive. */
543         memset(&rth, 0, sizeof(struct tcphdr));
544         rth.dest   = th->source;
545         rth.source = th->dest;
546         rth.doff   = sizeof(struct tcphdr) / 4;
547         rth.rst    = 1;
548
549         if (th->ack) {
550                 rth.seq = th->ack_seq;
551         } else {
552                 rth.ack = 1;
553                 rth.ack_seq = htonl(ntohl(th->seq) + th->syn + th->fin +
554                                     skb->len - (th->doff << 2));
555         }
556
557         memset(&arg, 0, sizeof arg);
558         arg.iov[0].iov_base = (unsigned char *)&rth;
559         arg.iov[0].iov_len  = sizeof rth;
560         arg.csum = csum_tcpudp_nofold(skb->nh.iph->daddr,
561                                       skb->nh.iph->saddr, /*XXX*/
562                                       sizeof(struct tcphdr), IPPROTO_TCP, 0);
563         arg.csumoffset = offsetof(struct tcphdr, check) / 2;
564
565         ip_send_reply(tcp_socket->sk, skb, &arg, sizeof rth);
566
567         TCP_INC_STATS_BH(TCP_MIB_OUTSEGS);
568         TCP_INC_STATS_BH(TCP_MIB_OUTRSTS);
569 }
570
571 /* The code following below sending ACKs in SYN-RECV and TIME-WAIT states
572    outside socket context is ugly, certainly. What can I do?
573  */
574
575 static void tcp_v4_send_ack(struct sk_buff *skb, u32 seq, u32 ack,
576                             u32 win, u32 ts)
577 {
578         struct tcphdr *th = skb->h.th;
579         struct {
580                 struct tcphdr th;
581                 u32 tsopt[TCPOLEN_TSTAMP_ALIGNED >> 2];
582         } rep;
583         struct ip_reply_arg arg;
584
585         memset(&rep.th, 0, sizeof(struct tcphdr));
586         memset(&arg, 0, sizeof arg);
587
588         arg.iov[0].iov_base = (unsigned char *)&rep;
589         arg.iov[0].iov_len  = sizeof(rep.th);
590         if (ts) {
591                 rep.tsopt[0] = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
592                                      (TCPOPT_TIMESTAMP << 8) |
593                                      TCPOLEN_TIMESTAMP);
594                 rep.tsopt[1] = htonl(tcp_time_stamp);
595                 rep.tsopt[2] = htonl(ts);
596                 arg.iov[0].iov_len = sizeof(rep);
597         }
598
599         /* Swap the send and the receive. */
600         rep.th.dest    = th->source;
601         rep.th.source  = th->dest;
602         rep.th.doff    = arg.iov[0].iov_len / 4;
603         rep.th.seq     = htonl(seq);
604         rep.th.ack_seq = htonl(ack);
605         rep.th.ack     = 1;
606         rep.th.window  = htons(win);
607
608         arg.csum = csum_tcpudp_nofold(skb->nh.iph->daddr,
609                                       skb->nh.iph->saddr, /*XXX*/
610                                       arg.iov[0].iov_len, IPPROTO_TCP, 0);
611         arg.csumoffset = offsetof(struct tcphdr, check) / 2;
612
613         ip_send_reply(tcp_socket->sk, skb, &arg, arg.iov[0].iov_len);
614
615         TCP_INC_STATS_BH(TCP_MIB_OUTSEGS);
616 }
617
618 static void tcp_v4_timewait_ack(struct sock *sk, struct sk_buff *skb)
619 {
620         struct inet_timewait_sock *tw = inet_twsk(sk);
621         const struct tcp_timewait_sock *tcptw = tcp_twsk(sk);
622
623         tcp_v4_send_ack(skb, tcptw->tw_snd_nxt, tcptw->tw_rcv_nxt,
624                         tcptw->tw_rcv_wnd >> tw->tw_rcv_wscale, tcptw->tw_ts_recent);
625
626         inet_twsk_put(tw);
627 }
628
629 static void tcp_v4_reqsk_send_ack(struct sk_buff *skb, struct request_sock *req)
630 {
631         tcp_v4_send_ack(skb, tcp_rsk(req)->snt_isn + 1, tcp_rsk(req)->rcv_isn + 1, req->rcv_wnd,
632                         req->ts_recent);
633 }
634
635 /*
636  *      Send a SYN-ACK after having received an ACK.
637  *      This still operates on a request_sock only, not on a big
638  *      socket.
639  */
640 static int tcp_v4_send_synack(struct sock *sk, struct request_sock *req,
641                               struct dst_entry *dst)
642 {
643         const struct inet_request_sock *ireq = inet_rsk(req);
644         int err = -1;
645         struct sk_buff * skb;
646
647         /* First, grab a route. */
648         if (!dst && (dst = inet_csk_route_req(sk, req)) == NULL)
649                 goto out;
650
651         skb = tcp_make_synack(sk, dst, req);
652
653         if (skb) {
654                 struct tcphdr *th = skb->h.th;
655
656                 th->check = tcp_v4_check(th, skb->len,
657                                          ireq->loc_addr,
658                                          ireq->rmt_addr,
659                                          csum_partial((char *)th, skb->len,
660                                                       skb->csum));
661
662                 err = ip_build_and_send_pkt(skb, sk, ireq->loc_addr,
663                                             ireq->rmt_addr,
664                                             ireq->opt);
665                 if (err == NET_XMIT_CN)
666                         err = 0;
667         }
668
669 out:
670         dst_release(dst);
671         return err;
672 }
673
674 /*
675  *      IPv4 request_sock destructor.
676  */
677 static void tcp_v4_reqsk_destructor(struct request_sock *req)
678 {
679         kfree(inet_rsk(req)->opt);
680 }
681
682 #ifdef CONFIG_SYN_COOKIES
683 static void syn_flood_warning(struct sk_buff *skb)
684 {
685         static unsigned long warntime;
686
687         if (time_after(jiffies, (warntime + HZ * 60))) {
688                 warntime = jiffies;
689                 printk(KERN_INFO
690                        "possible SYN flooding on port %d. Sending cookies.\n",
691                        ntohs(skb->h.th->dest));
692         }
693 }
694 #endif
695
696 /*
697  * Save and compile IPv4 options into the request_sock if needed.
698  */
699 static struct ip_options *tcp_v4_save_options(struct sock *sk,
700                                               struct sk_buff *skb)
701 {
702         struct ip_options *opt = &(IPCB(skb)->opt);
703         struct ip_options *dopt = NULL;
704
705         if (opt && opt->optlen) {
706                 int opt_size = optlength(opt);
707                 dopt = kmalloc(opt_size, GFP_ATOMIC);
708                 if (dopt) {
709                         if (ip_options_echo(dopt, skb)) {
710                                 kfree(dopt);
711                                 dopt = NULL;
712                         }
713                 }
714         }
715         return dopt;
716 }
717
718 struct request_sock_ops tcp_request_sock_ops __read_mostly = {
719         .family         =       PF_INET,
720         .obj_size       =       sizeof(struct tcp_request_sock),
721         .rtx_syn_ack    =       tcp_v4_send_synack,
722         .send_ack       =       tcp_v4_reqsk_send_ack,
723         .destructor     =       tcp_v4_reqsk_destructor,
724         .send_reset     =       tcp_v4_send_reset,
725 };
726
727 static struct timewait_sock_ops tcp_timewait_sock_ops = {
728         .twsk_obj_size  = sizeof(struct tcp_timewait_sock),
729         .twsk_unique    = tcp_twsk_unique,
730 };
731
732 int tcp_v4_conn_request(struct sock *sk, struct sk_buff *skb)
733 {
734         struct inet_request_sock *ireq;
735         struct tcp_options_received tmp_opt;
736         struct request_sock *req;
737         __be32 saddr = skb->nh.iph->saddr;
738         __be32 daddr = skb->nh.iph->daddr;
739         __u32 isn = TCP_SKB_CB(skb)->when;
740         struct dst_entry *dst = NULL;
741 #ifdef CONFIG_SYN_COOKIES
742         int want_cookie = 0;
743 #else
744 #define want_cookie 0 /* Argh, why doesn't gcc optimize this :( */
745 #endif
746
747         /* Never answer to SYNs send to broadcast or multicast */
748         if (((struct rtable *)skb->dst)->rt_flags &
749             (RTCF_BROADCAST | RTCF_MULTICAST))
750                 goto drop;
751
752         /* TW buckets are converted to open requests without
753          * limitations, they conserve resources and peer is
754          * evidently real one.
755          */
756         if (inet_csk_reqsk_queue_is_full(sk) && !isn) {
757 #ifdef CONFIG_SYN_COOKIES
758                 if (sysctl_tcp_syncookies) {
759                         want_cookie = 1;
760                 } else
761 #endif
762                 goto drop;
763         }
764
765         /* Accept backlog is full. If we have already queued enough
766          * of warm entries in syn queue, drop request. It is better than
767          * clogging syn queue with openreqs with exponentially increasing
768          * timeout.
769          */
770         if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1)
771                 goto drop;
772
773         req = reqsk_alloc(&tcp_request_sock_ops);
774         if (!req)
775                 goto drop;
776
777         tcp_clear_options(&tmp_opt);
778         tmp_opt.mss_clamp = 536;
779         tmp_opt.user_mss  = tcp_sk(sk)->rx_opt.user_mss;
780
781         tcp_parse_options(skb, &tmp_opt, 0);
782
783         if (want_cookie) {
784                 tcp_clear_options(&tmp_opt);
785                 tmp_opt.saw_tstamp = 0;
786         }
787
788         if (tmp_opt.saw_tstamp && !tmp_opt.rcv_tsval) {
789                 /* Some OSes (unknown ones, but I see them on web server, which
790                  * contains information interesting only for windows'
791                  * users) do not send their stamp in SYN. It is easy case.
792                  * We simply do not advertise TS support.
793                  */
794                 tmp_opt.saw_tstamp = 0;
795                 tmp_opt.tstamp_ok  = 0;
796         }
797         tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
798
799         tcp_openreq_init(req, &tmp_opt, skb);
800
801         if (security_inet_conn_request(sk, skb, req))
802                 goto drop_and_free;
803
804         ireq = inet_rsk(req);
805         ireq->loc_addr = daddr;
806         ireq->rmt_addr = saddr;
807         ireq->opt = tcp_v4_save_options(sk, skb);
808         if (!want_cookie)
809                 TCP_ECN_create_request(req, skb->h.th);
810
811         if (want_cookie) {
812 #ifdef CONFIG_SYN_COOKIES
813                 syn_flood_warning(skb);
814 #endif
815                 isn = cookie_v4_init_sequence(sk, skb, &req->mss);
816         } else if (!isn) {
817                 struct inet_peer *peer = NULL;
818
819                 /* VJ's idea. We save last timestamp seen
820                  * from the destination in peer table, when entering
821                  * state TIME-WAIT, and check against it before
822                  * accepting new connection request.
823                  *
824                  * If "isn" is not zero, this request hit alive
825                  * timewait bucket, so that all the necessary checks
826                  * are made in the function processing timewait state.
827                  */
828                 if (tmp_opt.saw_tstamp &&
829                     tcp_death_row.sysctl_tw_recycle &&
830                     (dst = inet_csk_route_req(sk, req)) != NULL &&
831                     (peer = rt_get_peer((struct rtable *)dst)) != NULL &&
832                     peer->v4daddr == saddr) {
833                         if (xtime.tv_sec < peer->tcp_ts_stamp + TCP_PAWS_MSL &&
834                             (s32)(peer->tcp_ts - req->ts_recent) >
835                                                         TCP_PAWS_WINDOW) {
836                                 NET_INC_STATS_BH(LINUX_MIB_PAWSPASSIVEREJECTED);
837                                 dst_release(dst);
838                                 goto drop_and_free;
839                         }
840                 }
841                 /* Kill the following clause, if you dislike this way. */
842                 else if (!sysctl_tcp_syncookies &&
843                          (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
844                           (sysctl_max_syn_backlog >> 2)) &&
845                          (!peer || !peer->tcp_ts_stamp) &&
846                          (!dst || !dst_metric(dst, RTAX_RTT))) {
847                         /* Without syncookies last quarter of
848                          * backlog is filled with destinations,
849                          * proven to be alive.
850                          * It means that we continue to communicate
851                          * to destinations, already remembered
852                          * to the moment of synflood.
853                          */
854                         LIMIT_NETDEBUG(KERN_DEBUG "TCP: drop open "
855                                        "request from %u.%u.%u.%u/%u\n",
856                                        NIPQUAD(saddr),
857                                        ntohs(skb->h.th->source));
858                         dst_release(dst);
859                         goto drop_and_free;
860                 }
861
862                 isn = tcp_v4_init_sequence(skb);
863         }
864         tcp_rsk(req)->snt_isn = isn;
865
866         if (tcp_v4_send_synack(sk, req, dst))
867                 goto drop_and_free;
868
869         if (want_cookie) {
870                 reqsk_free(req);
871         } else {
872                 inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
873         }
874         return 0;
875
876 drop_and_free:
877         reqsk_free(req);
878 drop:
879         return 0;
880 }
881
882
883 /*
884  * The three way handshake has completed - we got a valid synack -
885  * now create the new socket.
886  */
887 struct sock *tcp_v4_syn_recv_sock(struct sock *sk, struct sk_buff *skb,
888                                   struct request_sock *req,
889                                   struct dst_entry *dst)
890 {
891         struct inet_request_sock *ireq;
892         struct inet_sock *newinet;
893         struct tcp_sock *newtp;
894         struct sock *newsk;
895
896         if (sk_acceptq_is_full(sk))
897                 goto exit_overflow;
898
899         if (!dst && (dst = inet_csk_route_req(sk, req)) == NULL)
900                 goto exit;
901
902         newsk = tcp_create_openreq_child(sk, req, skb);
903         if (!newsk)
904                 goto exit;
905
906         newsk->sk_gso_type = SKB_GSO_TCPV4;
907         sk_setup_caps(newsk, dst);
908
909         newtp                 = tcp_sk(newsk);
910         newinet               = inet_sk(newsk);
911         ireq                  = inet_rsk(req);
912         newinet->daddr        = ireq->rmt_addr;
913         newinet->rcv_saddr    = ireq->loc_addr;
914         newinet->saddr        = ireq->loc_addr;
915         newinet->opt          = ireq->opt;
916         ireq->opt             = NULL;
917         newinet->mc_index     = inet_iif(skb);
918         newinet->mc_ttl       = skb->nh.iph->ttl;
919         inet_csk(newsk)->icsk_ext_hdr_len = 0;
920         if (newinet->opt)
921                 inet_csk(newsk)->icsk_ext_hdr_len = newinet->opt->optlen;
922         newinet->id = newtp->write_seq ^ jiffies;
923
924         tcp_mtup_init(newsk);
925         tcp_sync_mss(newsk, dst_mtu(dst));
926         newtp->advmss = dst_metric(dst, RTAX_ADVMSS);
927         tcp_initialize_rcv_mss(newsk);
928
929         __inet_hash(&tcp_hashinfo, newsk, 0);
930         __inet_inherit_port(&tcp_hashinfo, sk, newsk);
931
932         return newsk;
933
934 exit_overflow:
935         NET_INC_STATS_BH(LINUX_MIB_LISTENOVERFLOWS);
936 exit:
937         NET_INC_STATS_BH(LINUX_MIB_LISTENDROPS);
938         dst_release(dst);
939         return NULL;
940 }
941
942 static struct sock *tcp_v4_hnd_req(struct sock *sk, struct sk_buff *skb)
943 {
944         struct tcphdr *th = skb->h.th;
945         struct iphdr *iph = skb->nh.iph;
946         struct sock *nsk;
947         struct request_sock **prev;
948         /* Find possible connection requests. */
949         struct request_sock *req = inet_csk_search_req(sk, &prev, th->source,
950                                                        iph->saddr, iph->daddr);
951         if (req)
952                 return tcp_check_req(sk, skb, req, prev);
953
954         nsk = inet_lookup_established(&tcp_hashinfo, skb->nh.iph->saddr,
955                                       th->source, skb->nh.iph->daddr,
956                                       th->dest, inet_iif(skb));
957
958         if (nsk) {
959                 if (nsk->sk_state != TCP_TIME_WAIT) {
960                         bh_lock_sock(nsk);
961                         return nsk;
962                 }
963                 inet_twsk_put(inet_twsk(nsk));
964                 return NULL;
965         }
966
967 #ifdef CONFIG_SYN_COOKIES
968         if (!th->rst && !th->syn && th->ack)
969                 sk = cookie_v4_check(sk, skb, &(IPCB(skb)->opt));
970 #endif
971         return sk;
972 }
973
974 static int tcp_v4_checksum_init(struct sk_buff *skb)
975 {
976         if (skb->ip_summed == CHECKSUM_COMPLETE) {
977                 if (!tcp_v4_check(skb->h.th, skb->len, skb->nh.iph->saddr,
978                                   skb->nh.iph->daddr, skb->csum)) {
979                         skb->ip_summed = CHECKSUM_UNNECESSARY;
980                         return 0;
981                 }
982         }
983
984         skb->csum = csum_tcpudp_nofold(skb->nh.iph->saddr, skb->nh.iph->daddr,
985                                        skb->len, IPPROTO_TCP, 0);
986
987         if (skb->len <= 76) {
988                 return __skb_checksum_complete(skb);
989         }
990         return 0;
991 }
992
993
994 /* The socket must have it's spinlock held when we get
995  * here.
996  *
997  * We have a potential double-lock case here, so even when
998  * doing backlog processing we use the BH locking scheme.
999  * This is because we cannot sleep with the original spinlock
1000  * held.
1001  */
1002 int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb)
1003 {
1004         if (sk->sk_state == TCP_ESTABLISHED) { /* Fast path */
1005                 TCP_CHECK_TIMER(sk);
1006                 if (tcp_rcv_established(sk, skb, skb->h.th, skb->len))
1007                         goto reset;
1008                 TCP_CHECK_TIMER(sk);
1009                 return 0;
1010         }
1011
1012         if (skb->len < (skb->h.th->doff << 2) || tcp_checksum_complete(skb))
1013                 goto csum_err;
1014
1015         if (sk->sk_state == TCP_LISTEN) {
1016                 struct sock *nsk = tcp_v4_hnd_req(sk, skb);
1017                 if (!nsk)
1018                         goto discard;
1019
1020                 if (nsk != sk) {
1021                         if (tcp_child_process(sk, nsk, skb))
1022                                 goto reset;
1023                         return 0;
1024                 }
1025         }
1026
1027         TCP_CHECK_TIMER(sk);
1028         if (tcp_rcv_state_process(sk, skb, skb->h.th, skb->len))
1029                 goto reset;
1030         TCP_CHECK_TIMER(sk);
1031         return 0;
1032
1033 reset:
1034         tcp_v4_send_reset(skb);
1035 discard:
1036         kfree_skb(skb);
1037         /* Be careful here. If this function gets more complicated and
1038          * gcc suffers from register pressure on the x86, sk (in %ebx)
1039          * might be destroyed here. This current version compiles correctly,
1040          * but you have been warned.
1041          */
1042         return 0;
1043
1044 csum_err:
1045         TCP_INC_STATS_BH(TCP_MIB_INERRS);
1046         goto discard;
1047 }
1048
1049 /*
1050  *      From tcp_input.c
1051  */
1052
1053 int tcp_v4_rcv(struct sk_buff *skb)
1054 {
1055         struct tcphdr *th;
1056         struct sock *sk;
1057         int ret;
1058
1059         if (skb->pkt_type != PACKET_HOST)
1060                 goto discard_it;
1061
1062         /* Count it even if it's bad */
1063         TCP_INC_STATS_BH(TCP_MIB_INSEGS);
1064
1065         if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
1066                 goto discard_it;
1067
1068         th = skb->h.th;
1069
1070         if (th->doff < sizeof(struct tcphdr) / 4)
1071                 goto bad_packet;
1072         if (!pskb_may_pull(skb, th->doff * 4))
1073                 goto discard_it;
1074
1075         /* An explanation is required here, I think.
1076          * Packet length and doff are validated by header prediction,
1077          * provided case of th->doff==0 is eliminated.
1078          * So, we defer the checks. */
1079         if ((skb->ip_summed != CHECKSUM_UNNECESSARY &&
1080              tcp_v4_checksum_init(skb)))
1081                 goto bad_packet;
1082
1083         th = skb->h.th;
1084         TCP_SKB_CB(skb)->seq = ntohl(th->seq);
1085         TCP_SKB_CB(skb)->end_seq = (TCP_SKB_CB(skb)->seq + th->syn + th->fin +
1086                                     skb->len - th->doff * 4);
1087         TCP_SKB_CB(skb)->ack_seq = ntohl(th->ack_seq);
1088         TCP_SKB_CB(skb)->when    = 0;
1089         TCP_SKB_CB(skb)->flags   = skb->nh.iph->tos;
1090         TCP_SKB_CB(skb)->sacked  = 0;
1091
1092         sk = __inet_lookup(&tcp_hashinfo, skb->nh.iph->saddr, th->source,
1093                            skb->nh.iph->daddr, th->dest,
1094                            inet_iif(skb));
1095
1096         if (!sk)
1097                 goto no_tcp_socket;
1098
1099 process:
1100         if (sk->sk_state == TCP_TIME_WAIT)
1101                 goto do_time_wait;
1102
1103         if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb))
1104                 goto discard_and_relse;
1105         nf_reset(skb);
1106
1107         if (sk_filter(sk, skb))
1108                 goto discard_and_relse;
1109
1110         skb->dev = NULL;
1111
1112         bh_lock_sock_nested(sk);
1113         ret = 0;
1114         if (!sock_owned_by_user(sk)) {
1115 #ifdef CONFIG_NET_DMA
1116                 struct tcp_sock *tp = tcp_sk(sk);
1117                 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1118                         tp->ucopy.dma_chan = get_softnet_dma();
1119                 if (tp->ucopy.dma_chan)
1120                         ret = tcp_v4_do_rcv(sk, skb);
1121                 else
1122 #endif
1123                 {
1124                         if (!tcp_prequeue(sk, skb))
1125                         ret = tcp_v4_do_rcv(sk, skb);
1126                 }
1127         } else
1128                 sk_add_backlog(sk, skb);
1129         bh_unlock_sock(sk);
1130
1131         sock_put(sk);
1132
1133         return ret;
1134
1135 no_tcp_socket:
1136         if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb))
1137                 goto discard_it;
1138
1139         if (skb->len < (th->doff << 2) || tcp_checksum_complete(skb)) {
1140 bad_packet:
1141                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
1142         } else {
1143                 tcp_v4_send_reset(skb);
1144         }
1145
1146 discard_it:
1147         /* Discard frame. */
1148         kfree_skb(skb);
1149         return 0;
1150
1151 discard_and_relse:
1152         sock_put(sk);
1153         goto discard_it;
1154
1155 do_time_wait:
1156         if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb)) {
1157                 inet_twsk_put(inet_twsk(sk));
1158                 goto discard_it;
1159         }
1160
1161         if (skb->len < (th->doff << 2) || tcp_checksum_complete(skb)) {
1162                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
1163                 inet_twsk_put(inet_twsk(sk));
1164                 goto discard_it;
1165         }
1166         switch (tcp_timewait_state_process(inet_twsk(sk), skb, th)) {
1167         case TCP_TW_SYN: {
1168                 struct sock *sk2 = inet_lookup_listener(&tcp_hashinfo,
1169                                                         skb->nh.iph->daddr,
1170                                                         th->dest,
1171                                                         inet_iif(skb));
1172                 if (sk2) {
1173                         inet_twsk_deschedule(inet_twsk(sk), &tcp_death_row);
1174                         inet_twsk_put(inet_twsk(sk));
1175                         sk = sk2;
1176                         goto process;
1177                 }
1178                 /* Fall through to ACK */
1179         }
1180         case TCP_TW_ACK:
1181                 tcp_v4_timewait_ack(sk, skb);
1182                 break;
1183         case TCP_TW_RST:
1184                 goto no_tcp_socket;
1185         case TCP_TW_SUCCESS:;
1186         }
1187         goto discard_it;
1188 }
1189
1190 /* VJ's idea. Save last timestamp seen from this destination
1191  * and hold it at least for normal timewait interval to use for duplicate
1192  * segment detection in subsequent connections, before they enter synchronized
1193  * state.
1194  */
1195
1196 int tcp_v4_remember_stamp(struct sock *sk)
1197 {
1198         struct inet_sock *inet = inet_sk(sk);
1199         struct tcp_sock *tp = tcp_sk(sk);
1200         struct rtable *rt = (struct rtable *)__sk_dst_get(sk);
1201         struct inet_peer *peer = NULL;
1202         int release_it = 0;
1203
1204         if (!rt || rt->rt_dst != inet->daddr) {
1205                 peer = inet_getpeer(inet->daddr, 1);
1206                 release_it = 1;
1207         } else {
1208                 if (!rt->peer)
1209                         rt_bind_peer(rt, 1);
1210                 peer = rt->peer;
1211         }
1212
1213         if (peer) {
1214                 if ((s32)(peer->tcp_ts - tp->rx_opt.ts_recent) <= 0 ||
1215                     (peer->tcp_ts_stamp + TCP_PAWS_MSL < xtime.tv_sec &&
1216                      peer->tcp_ts_stamp <= tp->rx_opt.ts_recent_stamp)) {
1217                         peer->tcp_ts_stamp = tp->rx_opt.ts_recent_stamp;
1218                         peer->tcp_ts = tp->rx_opt.ts_recent;
1219                 }
1220                 if (release_it)
1221                         inet_putpeer(peer);
1222                 return 1;
1223         }
1224
1225         return 0;
1226 }
1227
1228 int tcp_v4_tw_remember_stamp(struct inet_timewait_sock *tw)
1229 {
1230         struct inet_peer *peer = inet_getpeer(tw->tw_daddr, 1);
1231
1232         if (peer) {
1233                 const struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
1234
1235                 if ((s32)(peer->tcp_ts - tcptw->tw_ts_recent) <= 0 ||
1236                     (peer->tcp_ts_stamp + TCP_PAWS_MSL < xtime.tv_sec &&
1237                      peer->tcp_ts_stamp <= tcptw->tw_ts_recent_stamp)) {
1238                         peer->tcp_ts_stamp = tcptw->tw_ts_recent_stamp;
1239                         peer->tcp_ts       = tcptw->tw_ts_recent;
1240                 }
1241                 inet_putpeer(peer);
1242                 return 1;
1243         }
1244
1245         return 0;
1246 }
1247
1248 struct inet_connection_sock_af_ops ipv4_specific = {
1249         .queue_xmit        = ip_queue_xmit,
1250         .send_check        = tcp_v4_send_check,
1251         .rebuild_header    = inet_sk_rebuild_header,
1252         .conn_request      = tcp_v4_conn_request,
1253         .syn_recv_sock     = tcp_v4_syn_recv_sock,
1254         .remember_stamp    = tcp_v4_remember_stamp,
1255         .net_header_len    = sizeof(struct iphdr),
1256         .setsockopt        = ip_setsockopt,
1257         .getsockopt        = ip_getsockopt,
1258         .addr2sockaddr     = inet_csk_addr2sockaddr,
1259         .sockaddr_len      = sizeof(struct sockaddr_in),
1260 #ifdef CONFIG_COMPAT
1261         .compat_setsockopt = compat_ip_setsockopt,
1262         .compat_getsockopt = compat_ip_getsockopt,
1263 #endif
1264 };
1265
1266 /* NOTE: A lot of things set to zero explicitly by call to
1267  *       sk_alloc() so need not be done here.
1268  */
1269 static int tcp_v4_init_sock(struct sock *sk)
1270 {
1271         struct inet_connection_sock *icsk = inet_csk(sk);
1272         struct tcp_sock *tp = tcp_sk(sk);
1273
1274         skb_queue_head_init(&tp->out_of_order_queue);
1275         tcp_init_xmit_timers(sk);
1276         tcp_prequeue_init(tp);
1277
1278         icsk->icsk_rto = TCP_TIMEOUT_INIT;
1279         tp->mdev = TCP_TIMEOUT_INIT;
1280
1281         /* So many TCP implementations out there (incorrectly) count the
1282          * initial SYN frame in their delayed-ACK and congestion control
1283          * algorithms that we must have the following bandaid to talk
1284          * efficiently to them.  -DaveM
1285          */
1286         tp->snd_cwnd = 2;
1287
1288         /* See draft-stevens-tcpca-spec-01 for discussion of the
1289          * initialization of these values.
1290          */
1291         tp->snd_ssthresh = 0x7fffffff;  /* Infinity */
1292         tp->snd_cwnd_clamp = ~0;
1293         tp->mss_cache = 536;
1294
1295         tp->reordering = sysctl_tcp_reordering;
1296         icsk->icsk_ca_ops = &tcp_init_congestion_ops;
1297
1298         sk->sk_state = TCP_CLOSE;
1299
1300         sk->sk_write_space = sk_stream_write_space;
1301         sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
1302
1303         icsk->icsk_af_ops = &ipv4_specific;
1304         icsk->icsk_sync_mss = tcp_sync_mss;
1305
1306         sk->sk_sndbuf = sysctl_tcp_wmem[1];
1307         sk->sk_rcvbuf = sysctl_tcp_rmem[1];
1308
1309         atomic_inc(&tcp_sockets_allocated);
1310
1311         return 0;
1312 }
1313
1314 int tcp_v4_destroy_sock(struct sock *sk)
1315 {
1316         struct tcp_sock *tp = tcp_sk(sk);
1317
1318         tcp_clear_xmit_timers(sk);
1319
1320         tcp_cleanup_congestion_control(sk);
1321
1322         /* Cleanup up the write buffer. */
1323         sk_stream_writequeue_purge(sk);
1324
1325         /* Cleans up our, hopefully empty, out_of_order_queue. */
1326         __skb_queue_purge(&tp->out_of_order_queue);
1327
1328 #ifdef CONFIG_NET_DMA
1329         /* Cleans up our sk_async_wait_queue */
1330         __skb_queue_purge(&sk->sk_async_wait_queue);
1331 #endif
1332
1333         /* Clean prequeue, it must be empty really */
1334         __skb_queue_purge(&tp->ucopy.prequeue);
1335
1336         /* Clean up a referenced TCP bind bucket. */
1337         if (inet_csk(sk)->icsk_bind_hash)
1338                 inet_put_port(&tcp_hashinfo, sk);
1339
1340         /*
1341          * If sendmsg cached page exists, toss it.
1342          */
1343         if (sk->sk_sndmsg_page) {
1344                 __free_page(sk->sk_sndmsg_page);
1345                 sk->sk_sndmsg_page = NULL;
1346         }
1347
1348         atomic_dec(&tcp_sockets_allocated);
1349
1350         return 0;
1351 }
1352
1353 EXPORT_SYMBOL(tcp_v4_destroy_sock);
1354
1355 #ifdef CONFIG_PROC_FS
1356 /* Proc filesystem TCP sock list dumping. */
1357
1358 static inline struct inet_timewait_sock *tw_head(struct hlist_head *head)
1359 {
1360         return hlist_empty(head) ? NULL :
1361                 list_entry(head->first, struct inet_timewait_sock, tw_node);
1362 }
1363
1364 static inline struct inet_timewait_sock *tw_next(struct inet_timewait_sock *tw)
1365 {
1366         return tw->tw_node.next ?
1367                 hlist_entry(tw->tw_node.next, typeof(*tw), tw_node) : NULL;
1368 }
1369
1370 static void *listening_get_next(struct seq_file *seq, void *cur)
1371 {
1372         struct inet_connection_sock *icsk;
1373         struct hlist_node *node;
1374         struct sock *sk = cur;
1375         struct tcp_iter_state* st = seq->private;
1376
1377         if (!sk) {
1378                 st->bucket = 0;
1379                 sk = sk_head(&tcp_hashinfo.listening_hash[0]);
1380                 goto get_sk;
1381         }
1382
1383         ++st->num;
1384
1385         if (st->state == TCP_SEQ_STATE_OPENREQ) {
1386                 struct request_sock *req = cur;
1387
1388                 icsk = inet_csk(st->syn_wait_sk);
1389                 req = req->dl_next;
1390                 while (1) {
1391                         while (req) {
1392                                 if (req->rsk_ops->family == st->family) {
1393                                         cur = req;
1394                                         goto out;
1395                                 }
1396                                 req = req->dl_next;
1397                         }
1398                         if (++st->sbucket >= icsk->icsk_accept_queue.listen_opt->nr_table_entries)
1399                                 break;
1400 get_req:
1401                         req = icsk->icsk_accept_queue.listen_opt->syn_table[st->sbucket];
1402                 }
1403                 sk        = sk_next(st->syn_wait_sk);
1404                 st->state = TCP_SEQ_STATE_LISTENING;
1405                 read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1406         } else {
1407                 icsk = inet_csk(sk);
1408                 read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1409                 if (reqsk_queue_len(&icsk->icsk_accept_queue))
1410                         goto start_req;
1411                 read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1412                 sk = sk_next(sk);
1413         }
1414 get_sk:
1415         sk_for_each_from(sk, node) {
1416                 if (sk->sk_family == st->family) {
1417                         cur = sk;
1418                         goto out;
1419                 }
1420                 icsk = inet_csk(sk);
1421                 read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1422                 if (reqsk_queue_len(&icsk->icsk_accept_queue)) {
1423 start_req:
1424                         st->uid         = sock_i_uid(sk);
1425                         st->syn_wait_sk = sk;
1426                         st->state       = TCP_SEQ_STATE_OPENREQ;
1427                         st->sbucket     = 0;
1428                         goto get_req;
1429                 }
1430                 read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1431         }
1432         if (++st->bucket < INET_LHTABLE_SIZE) {
1433                 sk = sk_head(&tcp_hashinfo.listening_hash[st->bucket]);
1434                 goto get_sk;
1435         }
1436         cur = NULL;
1437 out:
1438         return cur;
1439 }
1440
1441 static void *listening_get_idx(struct seq_file *seq, loff_t *pos)
1442 {
1443         void *rc = listening_get_next(seq, NULL);
1444
1445         while (rc && *pos) {
1446                 rc = listening_get_next(seq, rc);
1447                 --*pos;
1448         }
1449         return rc;
1450 }
1451
1452 static void *established_get_first(struct seq_file *seq)
1453 {
1454         struct tcp_iter_state* st = seq->private;
1455         void *rc = NULL;
1456
1457         for (st->bucket = 0; st->bucket < tcp_hashinfo.ehash_size; ++st->bucket) {
1458                 struct sock *sk;
1459                 struct hlist_node *node;
1460                 struct inet_timewait_sock *tw;
1461
1462                 /* We can reschedule _before_ having picked the target: */
1463                 cond_resched_softirq();
1464
1465                 read_lock(&tcp_hashinfo.ehash[st->bucket].lock);
1466                 sk_for_each(sk, node, &tcp_hashinfo.ehash[st->bucket].chain) {
1467                         if (sk->sk_family != st->family) {
1468                                 continue;
1469                         }
1470                         rc = sk;
1471                         goto out;
1472                 }
1473                 st->state = TCP_SEQ_STATE_TIME_WAIT;
1474                 inet_twsk_for_each(tw, node,
1475                                    &tcp_hashinfo.ehash[st->bucket + tcp_hashinfo.ehash_size].chain) {
1476                         if (tw->tw_family != st->family) {
1477                                 continue;
1478                         }
1479                         rc = tw;
1480                         goto out;
1481                 }
1482                 read_unlock(&tcp_hashinfo.ehash[st->bucket].lock);
1483                 st->state = TCP_SEQ_STATE_ESTABLISHED;
1484         }
1485 out:
1486         return rc;
1487 }
1488
1489 static void *established_get_next(struct seq_file *seq, void *cur)
1490 {
1491         struct sock *sk = cur;
1492         struct inet_timewait_sock *tw;
1493         struct hlist_node *node;
1494         struct tcp_iter_state* st = seq->private;
1495
1496         ++st->num;
1497
1498         if (st->state == TCP_SEQ_STATE_TIME_WAIT) {
1499                 tw = cur;
1500                 tw = tw_next(tw);
1501 get_tw:
1502                 while (tw && tw->tw_family != st->family) {
1503                         tw = tw_next(tw);
1504                 }
1505                 if (tw) {
1506                         cur = tw;
1507                         goto out;
1508                 }
1509                 read_unlock(&tcp_hashinfo.ehash[st->bucket].lock);
1510                 st->state = TCP_SEQ_STATE_ESTABLISHED;
1511
1512                 /* We can reschedule between buckets: */
1513                 cond_resched_softirq();
1514
1515                 if (++st->bucket < tcp_hashinfo.ehash_size) {
1516                         read_lock(&tcp_hashinfo.ehash[st->bucket].lock);
1517                         sk = sk_head(&tcp_hashinfo.ehash[st->bucket].chain);
1518                 } else {
1519                         cur = NULL;
1520                         goto out;
1521                 }
1522         } else
1523                 sk = sk_next(sk);
1524
1525         sk_for_each_from(sk, node) {
1526                 if (sk->sk_family == st->family)
1527                         goto found;
1528         }
1529
1530         st->state = TCP_SEQ_STATE_TIME_WAIT;
1531         tw = tw_head(&tcp_hashinfo.ehash[st->bucket + tcp_hashinfo.ehash_size].chain);
1532         goto get_tw;
1533 found:
1534         cur = sk;
1535 out:
1536         return cur;
1537 }
1538
1539 static void *established_get_idx(struct seq_file *seq, loff_t pos)
1540 {
1541         void *rc = established_get_first(seq);
1542
1543         while (rc && pos) {
1544                 rc = established_get_next(seq, rc);
1545                 --pos;
1546         }               
1547         return rc;
1548 }
1549
1550 static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
1551 {
1552         void *rc;
1553         struct tcp_iter_state* st = seq->private;
1554
1555         inet_listen_lock(&tcp_hashinfo);
1556         st->state = TCP_SEQ_STATE_LISTENING;
1557         rc        = listening_get_idx(seq, &pos);
1558
1559         if (!rc) {
1560                 inet_listen_unlock(&tcp_hashinfo);
1561                 local_bh_disable();
1562                 st->state = TCP_SEQ_STATE_ESTABLISHED;
1563                 rc        = established_get_idx(seq, pos);
1564         }
1565
1566         return rc;
1567 }
1568
1569 static void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
1570 {
1571         struct tcp_iter_state* st = seq->private;
1572         st->state = TCP_SEQ_STATE_LISTENING;
1573         st->num = 0;
1574         return *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
1575 }
1576
1577 static void *tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos)
1578 {
1579         void *rc = NULL;
1580         struct tcp_iter_state* st;
1581
1582         if (v == SEQ_START_TOKEN) {
1583                 rc = tcp_get_idx(seq, 0);
1584                 goto out;
1585         }
1586         st = seq->private;
1587
1588         switch (st->state) {
1589         case TCP_SEQ_STATE_OPENREQ:
1590         case TCP_SEQ_STATE_LISTENING:
1591                 rc = listening_get_next(seq, v);
1592                 if (!rc) {
1593                         inet_listen_unlock(&tcp_hashinfo);
1594                         local_bh_disable();
1595                         st->state = TCP_SEQ_STATE_ESTABLISHED;
1596                         rc        = established_get_first(seq);
1597                 }
1598                 break;
1599         case TCP_SEQ_STATE_ESTABLISHED:
1600         case TCP_SEQ_STATE_TIME_WAIT:
1601                 rc = established_get_next(seq, v);
1602                 break;
1603         }
1604 out:
1605         ++*pos;
1606         return rc;
1607 }
1608
1609 static void tcp_seq_stop(struct seq_file *seq, void *v)
1610 {
1611         struct tcp_iter_state* st = seq->private;
1612
1613         switch (st->state) {
1614         case TCP_SEQ_STATE_OPENREQ:
1615                 if (v) {
1616                         struct inet_connection_sock *icsk = inet_csk(st->syn_wait_sk);
1617                         read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1618                 }
1619         case TCP_SEQ_STATE_LISTENING:
1620                 if (v != SEQ_START_TOKEN)
1621                         inet_listen_unlock(&tcp_hashinfo);
1622                 break;
1623         case TCP_SEQ_STATE_TIME_WAIT:
1624         case TCP_SEQ_STATE_ESTABLISHED:
1625                 if (v)
1626                         read_unlock(&tcp_hashinfo.ehash[st->bucket].lock);
1627                 local_bh_enable();
1628                 break;
1629         }
1630 }
1631
1632 static int tcp_seq_open(struct inode *inode, struct file *file)
1633 {
1634         struct tcp_seq_afinfo *afinfo = PDE(inode)->data;
1635         struct seq_file *seq;
1636         struct tcp_iter_state *s;
1637         int rc;
1638
1639         if (unlikely(afinfo == NULL))
1640                 return -EINVAL;
1641
1642         s = kzalloc(sizeof(*s), GFP_KERNEL);
1643         if (!s)
1644                 return -ENOMEM;
1645         s->family               = afinfo->family;
1646         s->seq_ops.start        = tcp_seq_start;
1647         s->seq_ops.next         = tcp_seq_next;
1648         s->seq_ops.show         = afinfo->seq_show;
1649         s->seq_ops.stop         = tcp_seq_stop;
1650
1651         rc = seq_open(file, &s->seq_ops);
1652         if (rc)
1653                 goto out_kfree;
1654         seq          = file->private_data;
1655         seq->private = s;
1656 out:
1657         return rc;
1658 out_kfree:
1659         kfree(s);
1660         goto out;
1661 }
1662
1663 int tcp_proc_register(struct tcp_seq_afinfo *afinfo)
1664 {
1665         int rc = 0;
1666         struct proc_dir_entry *p;
1667
1668         if (!afinfo)
1669                 return -EINVAL;
1670         afinfo->seq_fops->owner         = afinfo->owner;
1671         afinfo->seq_fops->open          = tcp_seq_open;
1672         afinfo->seq_fops->read          = seq_read;
1673         afinfo->seq_fops->llseek        = seq_lseek;
1674         afinfo->seq_fops->release       = seq_release_private;
1675         
1676         p = proc_net_fops_create(afinfo->name, S_IRUGO, afinfo->seq_fops);
1677         if (p)
1678                 p->data = afinfo;
1679         else
1680                 rc = -ENOMEM;
1681         return rc;
1682 }
1683
1684 void tcp_proc_unregister(struct tcp_seq_afinfo *afinfo)
1685 {
1686         if (!afinfo)
1687                 return;
1688         proc_net_remove(afinfo->name);
1689         memset(afinfo->seq_fops, 0, sizeof(*afinfo->seq_fops)); 
1690 }
1691
1692 static void get_openreq4(struct sock *sk, struct request_sock *req,
1693                          char *tmpbuf, int i, int uid)
1694 {
1695         const struct inet_request_sock *ireq = inet_rsk(req);
1696         int ttd = req->expires - jiffies;
1697
1698         sprintf(tmpbuf, "%4d: %08X:%04X %08X:%04X"
1699                 " %02X %08X:%08X %02X:%08lX %08X %5d %8d %u %d %p",
1700                 i,
1701                 ireq->loc_addr,
1702                 ntohs(inet_sk(sk)->sport),
1703                 ireq->rmt_addr,
1704                 ntohs(ireq->rmt_port),
1705                 TCP_SYN_RECV,
1706                 0, 0, /* could print option size, but that is af dependent. */
1707                 1,    /* timers active (only the expire timer) */
1708                 jiffies_to_clock_t(ttd),
1709                 req->retrans,
1710                 uid,
1711                 0,  /* non standard timer */
1712                 0, /* open_requests have no inode */
1713                 atomic_read(&sk->sk_refcnt),
1714                 req);
1715 }
1716
1717 static void get_tcp4_sock(struct sock *sp, char *tmpbuf, int i)
1718 {
1719         int timer_active;
1720         unsigned long timer_expires;
1721         struct tcp_sock *tp = tcp_sk(sp);
1722         const struct inet_connection_sock *icsk = inet_csk(sp);
1723         struct inet_sock *inet = inet_sk(sp);
1724         unsigned int dest = inet->daddr;
1725         unsigned int src = inet->rcv_saddr;
1726         __u16 destp = ntohs(inet->dport);
1727         __u16 srcp = ntohs(inet->sport);
1728
1729         if (icsk->icsk_pending == ICSK_TIME_RETRANS) {
1730                 timer_active    = 1;
1731                 timer_expires   = icsk->icsk_timeout;
1732         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
1733                 timer_active    = 4;
1734                 timer_expires   = icsk->icsk_timeout;
1735         } else if (timer_pending(&sp->sk_timer)) {
1736                 timer_active    = 2;
1737                 timer_expires   = sp->sk_timer.expires;
1738         } else {
1739                 timer_active    = 0;
1740                 timer_expires = jiffies;
1741         }
1742
1743         sprintf(tmpbuf, "%4d: %08X:%04X %08X:%04X %02X %08X:%08X %02X:%08lX "
1744                         "%08X %5d %8d %lu %d %p %u %u %u %u %d",
1745                 i, src, srcp, dest, destp, sp->sk_state,
1746                 tp->write_seq - tp->snd_una,
1747                 (sp->sk_state == TCP_LISTEN) ? sp->sk_ack_backlog : (tp->rcv_nxt - tp->copied_seq),
1748                 timer_active,
1749                 jiffies_to_clock_t(timer_expires - jiffies),
1750                 icsk->icsk_retransmits,
1751                 sock_i_uid(sp),
1752                 icsk->icsk_probes_out,
1753                 sock_i_ino(sp),
1754                 atomic_read(&sp->sk_refcnt), sp,
1755                 icsk->icsk_rto,
1756                 icsk->icsk_ack.ato,
1757                 (icsk->icsk_ack.quick << 1) | icsk->icsk_ack.pingpong,
1758                 tp->snd_cwnd,
1759                 tp->snd_ssthresh >= 0xFFFF ? -1 : tp->snd_ssthresh);
1760 }
1761
1762 static void get_timewait4_sock(struct inet_timewait_sock *tw, char *tmpbuf, int i)
1763 {
1764         __be32 dest, src;
1765         __u16 destp, srcp;
1766         int ttd = tw->tw_ttd - jiffies;
1767
1768         if (ttd < 0)
1769                 ttd = 0;
1770
1771         dest  = tw->tw_daddr;
1772         src   = tw->tw_rcv_saddr;
1773         destp = ntohs(tw->tw_dport);
1774         srcp  = ntohs(tw->tw_sport);
1775
1776         sprintf(tmpbuf, "%4d: %08X:%04X %08X:%04X"
1777                 " %02X %08X:%08X %02X:%08lX %08X %5d %8d %d %d %p",
1778                 i, src, srcp, dest, destp, tw->tw_substate, 0, 0,
1779                 3, jiffies_to_clock_t(ttd), 0, 0, 0, 0,
1780                 atomic_read(&tw->tw_refcnt), tw);
1781 }
1782
1783 #define TMPSZ 150
1784
1785 static int tcp4_seq_show(struct seq_file *seq, void *v)
1786 {
1787         struct tcp_iter_state* st;
1788         char tmpbuf[TMPSZ + 1];
1789
1790         if (v == SEQ_START_TOKEN) {
1791                 seq_printf(seq, "%-*s\n", TMPSZ - 1,
1792                            "  sl  local_address rem_address   st tx_queue "
1793                            "rx_queue tr tm->when retrnsmt   uid  timeout "
1794                            "inode");
1795                 goto out;
1796         }
1797         st = seq->private;
1798
1799         switch (st->state) {
1800         case TCP_SEQ_STATE_LISTENING:
1801         case TCP_SEQ_STATE_ESTABLISHED:
1802                 get_tcp4_sock(v, tmpbuf, st->num);
1803                 break;
1804         case TCP_SEQ_STATE_OPENREQ:
1805                 get_openreq4(st->syn_wait_sk, v, tmpbuf, st->num, st->uid);
1806                 break;
1807         case TCP_SEQ_STATE_TIME_WAIT:
1808                 get_timewait4_sock(v, tmpbuf, st->num);
1809                 break;
1810         }
1811         seq_printf(seq, "%-*s\n", TMPSZ - 1, tmpbuf);
1812 out:
1813         return 0;
1814 }
1815
1816 static struct file_operations tcp4_seq_fops;
1817 static struct tcp_seq_afinfo tcp4_seq_afinfo = {
1818         .owner          = THIS_MODULE,
1819         .name           = "tcp",
1820         .family         = AF_INET,
1821         .seq_show       = tcp4_seq_show,
1822         .seq_fops       = &tcp4_seq_fops,
1823 };
1824
1825 int __init tcp4_proc_init(void)
1826 {
1827         return tcp_proc_register(&tcp4_seq_afinfo);
1828 }
1829
1830 void tcp4_proc_exit(void)
1831 {
1832         tcp_proc_unregister(&tcp4_seq_afinfo);
1833 }
1834 #endif /* CONFIG_PROC_FS */
1835
1836 struct proto tcp_prot = {
1837         .name                   = "TCP",
1838         .owner                  = THIS_MODULE,
1839         .close                  = tcp_close,
1840         .connect                = tcp_v4_connect,
1841         .disconnect             = tcp_disconnect,
1842         .accept                 = inet_csk_accept,
1843         .ioctl                  = tcp_ioctl,
1844         .init                   = tcp_v4_init_sock,
1845         .destroy                = tcp_v4_destroy_sock,
1846         .shutdown               = tcp_shutdown,
1847         .setsockopt             = tcp_setsockopt,
1848         .getsockopt             = tcp_getsockopt,
1849         .sendmsg                = tcp_sendmsg,
1850         .recvmsg                = tcp_recvmsg,
1851         .backlog_rcv            = tcp_v4_do_rcv,
1852         .hash                   = tcp_v4_hash,
1853         .unhash                 = tcp_unhash,
1854         .get_port               = tcp_v4_get_port,
1855         .enter_memory_pressure  = tcp_enter_memory_pressure,
1856         .sockets_allocated      = &tcp_sockets_allocated,
1857         .orphan_count           = &tcp_orphan_count,
1858         .memory_allocated       = &tcp_memory_allocated,
1859         .memory_pressure        = &tcp_memory_pressure,
1860         .sysctl_mem             = sysctl_tcp_mem,
1861         .sysctl_wmem            = sysctl_tcp_wmem,
1862         .sysctl_rmem            = sysctl_tcp_rmem,
1863         .max_header             = MAX_TCP_HEADER,
1864         .obj_size               = sizeof(struct tcp_sock),
1865         .twsk_prot              = &tcp_timewait_sock_ops,
1866         .rsk_prot               = &tcp_request_sock_ops,
1867 #ifdef CONFIG_COMPAT
1868         .compat_setsockopt      = compat_tcp_setsockopt,
1869         .compat_getsockopt      = compat_tcp_getsockopt,
1870 #endif
1871 };
1872
1873 void __init tcp_v4_init(struct net_proto_family *ops)
1874 {
1875         if (inet_csk_ctl_sock_create(&tcp_socket, PF_INET, SOCK_RAW, IPPROTO_TCP) < 0)
1876                 panic("Failed to create the TCP control socket.\n");
1877 }
1878
1879 EXPORT_SYMBOL(ipv4_specific);
1880 EXPORT_SYMBOL(tcp_hashinfo);
1881 EXPORT_SYMBOL(tcp_prot);
1882 EXPORT_SYMBOL(tcp_unhash);
1883 EXPORT_SYMBOL(tcp_v4_conn_request);
1884 EXPORT_SYMBOL(tcp_v4_connect);
1885 EXPORT_SYMBOL(tcp_v4_do_rcv);
1886 EXPORT_SYMBOL(tcp_v4_remember_stamp);
1887 EXPORT_SYMBOL(tcp_v4_send_check);
1888 EXPORT_SYMBOL(tcp_v4_syn_recv_sock);
1889
1890 #ifdef CONFIG_PROC_FS
1891 EXPORT_SYMBOL(tcp_proc_register);
1892 EXPORT_SYMBOL(tcp_proc_unregister);
1893 #endif
1894 EXPORT_SYMBOL(sysctl_local_port_range);
1895 EXPORT_SYMBOL(sysctl_tcp_low_latency);
1896