4b04c3edd4a95518a878e294096424c053037912
[linux-2.6.git] / 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;
82 int sysctl_tcp_low_latency;
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 sock *sk, 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         u32 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((struct inet_timewait_sock *)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(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_HW) {
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_HW;
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[3];
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 = {
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         __u32 saddr = skb->nh.iph->saddr;
738         __u32 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         ireq = inet_rsk(req);
802         ireq->loc_addr = daddr;
803         ireq->rmt_addr = saddr;
804         ireq->opt = tcp_v4_save_options(sk, skb);
805         if (!want_cookie)
806                 TCP_ECN_create_request(req, skb->h.th);
807
808         if (want_cookie) {
809 #ifdef CONFIG_SYN_COOKIES
810                 syn_flood_warning(skb);
811 #endif
812                 isn = cookie_v4_init_sequence(sk, skb, &req->mss);
813         } else if (!isn) {
814                 struct inet_peer *peer = NULL;
815
816                 /* VJ's idea. We save last timestamp seen
817                  * from the destination in peer table, when entering
818                  * state TIME-WAIT, and check against it before
819                  * accepting new connection request.
820                  *
821                  * If "isn" is not zero, this request hit alive
822                  * timewait bucket, so that all the necessary checks
823                  * are made in the function processing timewait state.
824                  */
825                 if (tmp_opt.saw_tstamp &&
826                     tcp_death_row.sysctl_tw_recycle &&
827                     (dst = inet_csk_route_req(sk, req)) != NULL &&
828                     (peer = rt_get_peer((struct rtable *)dst)) != NULL &&
829                     peer->v4daddr == saddr) {
830                         if (xtime.tv_sec < peer->tcp_ts_stamp + TCP_PAWS_MSL &&
831                             (s32)(peer->tcp_ts - req->ts_recent) >
832                                                         TCP_PAWS_WINDOW) {
833                                 NET_INC_STATS_BH(LINUX_MIB_PAWSPASSIVEREJECTED);
834                                 dst_release(dst);
835                                 goto drop_and_free;
836                         }
837                 }
838                 /* Kill the following clause, if you dislike this way. */
839                 else if (!sysctl_tcp_syncookies &&
840                          (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
841                           (sysctl_max_syn_backlog >> 2)) &&
842                          (!peer || !peer->tcp_ts_stamp) &&
843                          (!dst || !dst_metric(dst, RTAX_RTT))) {
844                         /* Without syncookies last quarter of
845                          * backlog is filled with destinations,
846                          * proven to be alive.
847                          * It means that we continue to communicate
848                          * to destinations, already remembered
849                          * to the moment of synflood.
850                          */
851                         LIMIT_NETDEBUG(KERN_DEBUG "TCP: drop open "
852                                        "request from %u.%u.%u.%u/%u\n",
853                                        NIPQUAD(saddr),
854                                        ntohs(skb->h.th->source));
855                         dst_release(dst);
856                         goto drop_and_free;
857                 }
858
859                 isn = tcp_v4_init_sequence(sk, skb);
860         }
861         tcp_rsk(req)->snt_isn = isn;
862
863         if (tcp_v4_send_synack(sk, req, dst))
864                 goto drop_and_free;
865
866         if (want_cookie) {
867                 reqsk_free(req);
868         } else {
869                 inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
870         }
871         return 0;
872
873 drop_and_free:
874         reqsk_free(req);
875 drop:
876         return 0;
877 }
878
879
880 /*
881  * The three way handshake has completed - we got a valid synack -
882  * now create the new socket.
883  */
884 struct sock *tcp_v4_syn_recv_sock(struct sock *sk, struct sk_buff *skb,
885                                   struct request_sock *req,
886                                   struct dst_entry *dst)
887 {
888         struct inet_request_sock *ireq;
889         struct inet_sock *newinet;
890         struct tcp_sock *newtp;
891         struct sock *newsk;
892
893         if (sk_acceptq_is_full(sk))
894                 goto exit_overflow;
895
896         if (!dst && (dst = inet_csk_route_req(sk, req)) == NULL)
897                 goto exit;
898
899         newsk = tcp_create_openreq_child(sk, req, skb);
900         if (!newsk)
901                 goto exit;
902
903         newsk->sk_gso_type = SKB_GSO_TCPV4;
904         sk_setup_caps(newsk, dst);
905
906         newtp                 = tcp_sk(newsk);
907         newinet               = inet_sk(newsk);
908         ireq                  = inet_rsk(req);
909         newinet->daddr        = ireq->rmt_addr;
910         newinet->rcv_saddr    = ireq->loc_addr;
911         newinet->saddr        = ireq->loc_addr;
912         newinet->opt          = ireq->opt;
913         ireq->opt             = NULL;
914         newinet->mc_index     = inet_iif(skb);
915         newinet->mc_ttl       = skb->nh.iph->ttl;
916         inet_csk(newsk)->icsk_ext_hdr_len = 0;
917         if (newinet->opt)
918                 inet_csk(newsk)->icsk_ext_hdr_len = newinet->opt->optlen;
919         newinet->id = newtp->write_seq ^ jiffies;
920
921         tcp_mtup_init(newsk);
922         tcp_sync_mss(newsk, dst_mtu(dst));
923         newtp->advmss = dst_metric(dst, RTAX_ADVMSS);
924         tcp_initialize_rcv_mss(newsk);
925
926         __inet_hash(&tcp_hashinfo, newsk, 0);
927         __inet_inherit_port(&tcp_hashinfo, sk, newsk);
928
929         return newsk;
930
931 exit_overflow:
932         NET_INC_STATS_BH(LINUX_MIB_LISTENOVERFLOWS);
933 exit:
934         NET_INC_STATS_BH(LINUX_MIB_LISTENDROPS);
935         dst_release(dst);
936         return NULL;
937 }
938
939 static struct sock *tcp_v4_hnd_req(struct sock *sk, struct sk_buff *skb)
940 {
941         struct tcphdr *th = skb->h.th;
942         struct iphdr *iph = skb->nh.iph;
943         struct sock *nsk;
944         struct request_sock **prev;
945         /* Find possible connection requests. */
946         struct request_sock *req = inet_csk_search_req(sk, &prev, th->source,
947                                                        iph->saddr, iph->daddr);
948         if (req)
949                 return tcp_check_req(sk, skb, req, prev);
950
951         nsk = __inet_lookup_established(&tcp_hashinfo, skb->nh.iph->saddr,
952                                         th->source, skb->nh.iph->daddr,
953                                         ntohs(th->dest), inet_iif(skb));
954
955         if (nsk) {
956                 if (nsk->sk_state != TCP_TIME_WAIT) {
957                         bh_lock_sock(nsk);
958                         return nsk;
959                 }
960                 inet_twsk_put((struct inet_timewait_sock *)nsk);
961                 return NULL;
962         }
963
964 #ifdef CONFIG_SYN_COOKIES
965         if (!th->rst && !th->syn && th->ack)
966                 sk = cookie_v4_check(sk, skb, &(IPCB(skb)->opt));
967 #endif
968         return sk;
969 }
970
971 static int tcp_v4_checksum_init(struct sk_buff *skb)
972 {
973         if (skb->ip_summed == CHECKSUM_HW) {
974                 if (!tcp_v4_check(skb->h.th, skb->len, skb->nh.iph->saddr,
975                                   skb->nh.iph->daddr, skb->csum)) {
976                         skb->ip_summed = CHECKSUM_UNNECESSARY;
977                         return 0;
978                 }
979         }
980
981         skb->csum = csum_tcpudp_nofold(skb->nh.iph->saddr, skb->nh.iph->daddr,
982                                        skb->len, IPPROTO_TCP, 0);
983
984         if (skb->len <= 76) {
985                 return __skb_checksum_complete(skb);
986         }
987         return 0;
988 }
989
990
991 /* The socket must have it's spinlock held when we get
992  * here.
993  *
994  * We have a potential double-lock case here, so even when
995  * doing backlog processing we use the BH locking scheme.
996  * This is because we cannot sleep with the original spinlock
997  * held.
998  */
999 int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb)
1000 {
1001         if (sk->sk_state == TCP_ESTABLISHED) { /* Fast path */
1002                 TCP_CHECK_TIMER(sk);
1003                 if (tcp_rcv_established(sk, skb, skb->h.th, skb->len))
1004                         goto reset;
1005                 TCP_CHECK_TIMER(sk);
1006                 return 0;
1007         }
1008
1009         if (skb->len < (skb->h.th->doff << 2) || tcp_checksum_complete(skb))
1010                 goto csum_err;
1011
1012         if (sk->sk_state == TCP_LISTEN) {
1013                 struct sock *nsk = tcp_v4_hnd_req(sk, skb);
1014                 if (!nsk)
1015                         goto discard;
1016
1017                 if (nsk != sk) {
1018                         if (tcp_child_process(sk, nsk, skb))
1019                                 goto reset;
1020                         return 0;
1021                 }
1022         }
1023
1024         TCP_CHECK_TIMER(sk);
1025         if (tcp_rcv_state_process(sk, skb, skb->h.th, skb->len))
1026                 goto reset;
1027         TCP_CHECK_TIMER(sk);
1028         return 0;
1029
1030 reset:
1031         tcp_v4_send_reset(skb);
1032 discard:
1033         kfree_skb(skb);
1034         /* Be careful here. If this function gets more complicated and
1035          * gcc suffers from register pressure on the x86, sk (in %ebx)
1036          * might be destroyed here. This current version compiles correctly,
1037          * but you have been warned.
1038          */
1039         return 0;
1040
1041 csum_err:
1042         TCP_INC_STATS_BH(TCP_MIB_INERRS);
1043         goto discard;
1044 }
1045
1046 /*
1047  *      From tcp_input.c
1048  */
1049
1050 int tcp_v4_rcv(struct sk_buff *skb)
1051 {
1052         struct tcphdr *th;
1053         struct sock *sk;
1054         int ret;
1055
1056         if (skb->pkt_type != PACKET_HOST)
1057                 goto discard_it;
1058
1059         /* Count it even if it's bad */
1060         TCP_INC_STATS_BH(TCP_MIB_INSEGS);
1061
1062         if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
1063                 goto discard_it;
1064
1065         th = skb->h.th;
1066
1067         if (th->doff < sizeof(struct tcphdr) / 4)
1068                 goto bad_packet;
1069         if (!pskb_may_pull(skb, th->doff * 4))
1070                 goto discard_it;
1071
1072         /* An explanation is required here, I think.
1073          * Packet length and doff are validated by header prediction,
1074          * provided case of th->doff==0 is eliminated.
1075          * So, we defer the checks. */
1076         if ((skb->ip_summed != CHECKSUM_UNNECESSARY &&
1077              tcp_v4_checksum_init(skb)))
1078                 goto bad_packet;
1079
1080         th = skb->h.th;
1081         TCP_SKB_CB(skb)->seq = ntohl(th->seq);
1082         TCP_SKB_CB(skb)->end_seq = (TCP_SKB_CB(skb)->seq + th->syn + th->fin +
1083                                     skb->len - th->doff * 4);
1084         TCP_SKB_CB(skb)->ack_seq = ntohl(th->ack_seq);
1085         TCP_SKB_CB(skb)->when    = 0;
1086         TCP_SKB_CB(skb)->flags   = skb->nh.iph->tos;
1087         TCP_SKB_CB(skb)->sacked  = 0;
1088
1089         sk = __inet_lookup(&tcp_hashinfo, skb->nh.iph->saddr, th->source,
1090                            skb->nh.iph->daddr, ntohs(th->dest),
1091                            inet_iif(skb));
1092
1093         if (!sk)
1094                 goto no_tcp_socket;
1095
1096 process:
1097         if (sk->sk_state == TCP_TIME_WAIT)
1098                 goto do_time_wait;
1099
1100         if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb))
1101                 goto discard_and_relse;
1102         nf_reset(skb);
1103
1104         if (sk_filter(sk, skb, 0))
1105                 goto discard_and_relse;
1106
1107         skb->dev = NULL;
1108
1109         bh_lock_sock_nested(sk);
1110         ret = 0;
1111         if (!sock_owned_by_user(sk)) {
1112 #ifdef CONFIG_NET_DMA
1113                 struct tcp_sock *tp = tcp_sk(sk);
1114                 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1115                         tp->ucopy.dma_chan = get_softnet_dma();
1116                 if (tp->ucopy.dma_chan)
1117                         ret = tcp_v4_do_rcv(sk, skb);
1118                 else
1119 #endif
1120                 {
1121                         if (!tcp_prequeue(sk, skb))
1122                         ret = tcp_v4_do_rcv(sk, skb);
1123                 }
1124         } else
1125                 sk_add_backlog(sk, skb);
1126         bh_unlock_sock(sk);
1127
1128         sock_put(sk);
1129
1130         return ret;
1131
1132 no_tcp_socket:
1133         if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb))
1134                 goto discard_it;
1135
1136         if (skb->len < (th->doff << 2) || tcp_checksum_complete(skb)) {
1137 bad_packet:
1138                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
1139         } else {
1140                 tcp_v4_send_reset(skb);
1141         }
1142
1143 discard_it:
1144         /* Discard frame. */
1145         kfree_skb(skb);
1146         return 0;
1147
1148 discard_and_relse:
1149         sock_put(sk);
1150         goto discard_it;
1151
1152 do_time_wait:
1153         if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb)) {
1154                 inet_twsk_put((struct inet_timewait_sock *) sk);
1155                 goto discard_it;
1156         }
1157
1158         if (skb->len < (th->doff << 2) || tcp_checksum_complete(skb)) {
1159                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
1160                 inet_twsk_put((struct inet_timewait_sock *) sk);
1161                 goto discard_it;
1162         }
1163         switch (tcp_timewait_state_process((struct inet_timewait_sock *)sk,
1164                                            skb, th)) {
1165         case TCP_TW_SYN: {
1166                 struct sock *sk2 = inet_lookup_listener(&tcp_hashinfo,
1167                                                         skb->nh.iph->daddr,
1168                                                         ntohs(th->dest),
1169                                                         inet_iif(skb));
1170                 if (sk2) {
1171                         inet_twsk_deschedule((struct inet_timewait_sock *)sk,
1172                                              &tcp_death_row);
1173                         inet_twsk_put((struct inet_timewait_sock *)sk);
1174                         sk = sk2;
1175                         goto process;
1176                 }
1177                 /* Fall through to ACK */
1178         }
1179         case TCP_TW_ACK:
1180                 tcp_v4_timewait_ack(sk, skb);
1181                 break;
1182         case TCP_TW_RST:
1183                 goto no_tcp_socket;
1184         case TCP_TW_SUCCESS:;
1185         }
1186         goto discard_it;
1187 }
1188
1189 /* VJ's idea. Save last timestamp seen from this destination
1190  * and hold it at least for normal timewait interval to use for duplicate
1191  * segment detection in subsequent connections, before they enter synchronized
1192  * state.
1193  */
1194
1195 int tcp_v4_remember_stamp(struct sock *sk)
1196 {
1197         struct inet_sock *inet = inet_sk(sk);
1198         struct tcp_sock *tp = tcp_sk(sk);
1199         struct rtable *rt = (struct rtable *)__sk_dst_get(sk);
1200         struct inet_peer *peer = NULL;
1201         int release_it = 0;
1202
1203         if (!rt || rt->rt_dst != inet->daddr) {
1204                 peer = inet_getpeer(inet->daddr, 1);
1205                 release_it = 1;
1206         } else {
1207                 if (!rt->peer)
1208                         rt_bind_peer(rt, 1);
1209                 peer = rt->peer;
1210         }
1211
1212         if (peer) {
1213                 if ((s32)(peer->tcp_ts - tp->rx_opt.ts_recent) <= 0 ||
1214                     (peer->tcp_ts_stamp + TCP_PAWS_MSL < xtime.tv_sec &&
1215                      peer->tcp_ts_stamp <= tp->rx_opt.ts_recent_stamp)) {
1216                         peer->tcp_ts_stamp = tp->rx_opt.ts_recent_stamp;
1217                         peer->tcp_ts = tp->rx_opt.ts_recent;
1218                 }
1219                 if (release_it)
1220                         inet_putpeer(peer);
1221                 return 1;
1222         }
1223
1224         return 0;
1225 }
1226
1227 int tcp_v4_tw_remember_stamp(struct inet_timewait_sock *tw)
1228 {
1229         struct inet_peer *peer = inet_getpeer(tw->tw_daddr, 1);
1230
1231         if (peer) {
1232                 const struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
1233
1234                 if ((s32)(peer->tcp_ts - tcptw->tw_ts_recent) <= 0 ||
1235                     (peer->tcp_ts_stamp + TCP_PAWS_MSL < xtime.tv_sec &&
1236                      peer->tcp_ts_stamp <= tcptw->tw_ts_recent_stamp)) {
1237                         peer->tcp_ts_stamp = tcptw->tw_ts_recent_stamp;
1238                         peer->tcp_ts       = tcptw->tw_ts_recent;
1239                 }
1240                 inet_putpeer(peer);
1241                 return 1;
1242         }
1243
1244         return 0;
1245 }
1246
1247 struct inet_connection_sock_af_ops ipv4_specific = {
1248         .queue_xmit        = ip_queue_xmit,
1249         .send_check        = tcp_v4_send_check,
1250         .rebuild_header    = inet_sk_rebuild_header,
1251         .conn_request      = tcp_v4_conn_request,
1252         .syn_recv_sock     = tcp_v4_syn_recv_sock,
1253         .remember_stamp    = tcp_v4_remember_stamp,
1254         .net_header_len    = sizeof(struct iphdr),
1255         .setsockopt        = ip_setsockopt,
1256         .getsockopt        = ip_getsockopt,
1257         .addr2sockaddr     = inet_csk_addr2sockaddr,
1258         .sockaddr_len      = sizeof(struct sockaddr_in),
1259 #ifdef CONFIG_COMPAT
1260         .compat_setsockopt = compat_ip_setsockopt,
1261         .compat_getsockopt = compat_ip_getsockopt,
1262 #endif
1263 };
1264
1265 /* NOTE: A lot of things set to zero explicitly by call to
1266  *       sk_alloc() so need not be done here.
1267  */
1268 static int tcp_v4_init_sock(struct sock *sk)
1269 {
1270         struct inet_connection_sock *icsk = inet_csk(sk);
1271         struct tcp_sock *tp = tcp_sk(sk);
1272
1273         skb_queue_head_init(&tp->out_of_order_queue);
1274         tcp_init_xmit_timers(sk);
1275         tcp_prequeue_init(tp);
1276
1277         icsk->icsk_rto = TCP_TIMEOUT_INIT;
1278         tp->mdev = TCP_TIMEOUT_INIT;
1279
1280         /* So many TCP implementations out there (incorrectly) count the
1281          * initial SYN frame in their delayed-ACK and congestion control
1282          * algorithms that we must have the following bandaid to talk
1283          * efficiently to them.  -DaveM
1284          */
1285         tp->snd_cwnd = 2;
1286
1287         /* See draft-stevens-tcpca-spec-01 for discussion of the
1288          * initialization of these values.
1289          */
1290         tp->snd_ssthresh = 0x7fffffff;  /* Infinity */
1291         tp->snd_cwnd_clamp = ~0;
1292         tp->mss_cache = 536;
1293
1294         tp->reordering = sysctl_tcp_reordering;
1295         icsk->icsk_ca_ops = &tcp_init_congestion_ops;
1296
1297         sk->sk_state = TCP_CLOSE;
1298
1299         sk->sk_write_space = sk_stream_write_space;
1300         sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
1301
1302         icsk->icsk_af_ops = &ipv4_specific;
1303         icsk->icsk_sync_mss = tcp_sync_mss;
1304
1305         sk->sk_sndbuf = sysctl_tcp_wmem[1];
1306         sk->sk_rcvbuf = sysctl_tcp_rmem[1];
1307
1308         atomic_inc(&tcp_sockets_allocated);
1309
1310         return 0;
1311 }
1312
1313 int tcp_v4_destroy_sock(struct sock *sk)
1314 {
1315         struct tcp_sock *tp = tcp_sk(sk);
1316
1317         tcp_clear_xmit_timers(sk);
1318
1319         tcp_cleanup_congestion_control(sk);
1320
1321         /* Cleanup up the write buffer. */
1322         sk_stream_writequeue_purge(sk);
1323
1324         /* Cleans up our, hopefully empty, out_of_order_queue. */
1325         __skb_queue_purge(&tp->out_of_order_queue);
1326
1327 #ifdef CONFIG_NET_DMA
1328         /* Cleans up our sk_async_wait_queue */
1329         __skb_queue_purge(&sk->sk_async_wait_queue);
1330 #endif
1331
1332         /* Clean prequeue, it must be empty really */
1333         __skb_queue_purge(&tp->ucopy.prequeue);
1334
1335         /* Clean up a referenced TCP bind bucket. */
1336         if (inet_csk(sk)->icsk_bind_hash)
1337                 inet_put_port(&tcp_hashinfo, sk);
1338
1339         /*
1340          * If sendmsg cached page exists, toss it.
1341          */
1342         if (sk->sk_sndmsg_page) {
1343                 __free_page(sk->sk_sndmsg_page);
1344                 sk->sk_sndmsg_page = NULL;
1345         }
1346
1347         atomic_dec(&tcp_sockets_allocated);
1348
1349         return 0;
1350 }
1351
1352 EXPORT_SYMBOL(tcp_v4_destroy_sock);
1353
1354 #ifdef CONFIG_PROC_FS
1355 /* Proc filesystem TCP sock list dumping. */
1356
1357 static inline struct inet_timewait_sock *tw_head(struct hlist_head *head)
1358 {
1359         return hlist_empty(head) ? NULL :
1360                 list_entry(head->first, struct inet_timewait_sock, tw_node);
1361 }
1362
1363 static inline struct inet_timewait_sock *tw_next(struct inet_timewait_sock *tw)
1364 {
1365         return tw->tw_node.next ?
1366                 hlist_entry(tw->tw_node.next, typeof(*tw), tw_node) : NULL;
1367 }
1368
1369 static void *listening_get_next(struct seq_file *seq, void *cur)
1370 {
1371         struct inet_connection_sock *icsk;
1372         struct hlist_node *node;
1373         struct sock *sk = cur;
1374         struct tcp_iter_state* st = seq->private;
1375
1376         if (!sk) {
1377                 st->bucket = 0;
1378                 sk = sk_head(&tcp_hashinfo.listening_hash[0]);
1379                 goto get_sk;
1380         }
1381
1382         ++st->num;
1383
1384         if (st->state == TCP_SEQ_STATE_OPENREQ) {
1385                 struct request_sock *req = cur;
1386
1387                 icsk = inet_csk(st->syn_wait_sk);
1388                 req = req->dl_next;
1389                 while (1) {
1390                         while (req) {
1391                                 if (req->rsk_ops->family == st->family) {
1392                                         cur = req;
1393                                         goto out;
1394                                 }
1395                                 req = req->dl_next;
1396                         }
1397                         if (++st->sbucket >= TCP_SYNQ_HSIZE)
1398                                 break;
1399 get_req:
1400                         req = icsk->icsk_accept_queue.listen_opt->syn_table[st->sbucket];
1401                 }
1402                 sk        = sk_next(st->syn_wait_sk);
1403                 st->state = TCP_SEQ_STATE_LISTENING;
1404                 read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1405         } else {
1406                 icsk = inet_csk(sk);
1407                 read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1408                 if (reqsk_queue_len(&icsk->icsk_accept_queue))
1409                         goto start_req;
1410                 read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1411                 sk = sk_next(sk);
1412         }
1413 get_sk:
1414         sk_for_each_from(sk, node) {
1415                 if (sk->sk_family == st->family) {
1416                         cur = sk;
1417                         goto out;
1418                 }
1419                 icsk = inet_csk(sk);
1420                 read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1421                 if (reqsk_queue_len(&icsk->icsk_accept_queue)) {
1422 start_req:
1423                         st->uid         = sock_i_uid(sk);
1424                         st->syn_wait_sk = sk;
1425                         st->state       = TCP_SEQ_STATE_OPENREQ;
1426                         st->sbucket     = 0;
1427                         goto get_req;
1428                 }
1429                 read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1430         }
1431         if (++st->bucket < INET_LHTABLE_SIZE) {
1432                 sk = sk_head(&tcp_hashinfo.listening_hash[st->bucket]);
1433                 goto get_sk;
1434         }
1435         cur = NULL;
1436 out:
1437         return cur;
1438 }
1439
1440 static void *listening_get_idx(struct seq_file *seq, loff_t *pos)
1441 {
1442         void *rc = listening_get_next(seq, NULL);
1443
1444         while (rc && *pos) {
1445                 rc = listening_get_next(seq, rc);
1446                 --*pos;
1447         }
1448         return rc;
1449 }
1450
1451 static void *established_get_first(struct seq_file *seq)
1452 {
1453         struct tcp_iter_state* st = seq->private;
1454         void *rc = NULL;
1455
1456         for (st->bucket = 0; st->bucket < tcp_hashinfo.ehash_size; ++st->bucket) {
1457                 struct sock *sk;
1458                 struct hlist_node *node;
1459                 struct inet_timewait_sock *tw;
1460
1461                 /* We can reschedule _before_ having picked the target: */
1462                 cond_resched_softirq();
1463
1464                 read_lock(&tcp_hashinfo.ehash[st->bucket].lock);
1465                 sk_for_each(sk, node, &tcp_hashinfo.ehash[st->bucket].chain) {
1466                         if (sk->sk_family != st->family) {
1467                                 continue;
1468                         }
1469                         rc = sk;
1470                         goto out;
1471                 }
1472                 st->state = TCP_SEQ_STATE_TIME_WAIT;
1473                 inet_twsk_for_each(tw, node,
1474                                    &tcp_hashinfo.ehash[st->bucket + tcp_hashinfo.ehash_size].chain) {
1475                         if (tw->tw_family != st->family) {
1476                                 continue;
1477                         }
1478                         rc = tw;
1479                         goto out;
1480                 }
1481                 read_unlock(&tcp_hashinfo.ehash[st->bucket].lock);
1482                 st->state = TCP_SEQ_STATE_ESTABLISHED;
1483         }
1484 out:
1485         return rc;
1486 }
1487
1488 static void *established_get_next(struct seq_file *seq, void *cur)
1489 {
1490         struct sock *sk = cur;
1491         struct inet_timewait_sock *tw;
1492         struct hlist_node *node;
1493         struct tcp_iter_state* st = seq->private;
1494
1495         ++st->num;
1496
1497         if (st->state == TCP_SEQ_STATE_TIME_WAIT) {
1498                 tw = cur;
1499                 tw = tw_next(tw);
1500 get_tw:
1501                 while (tw && tw->tw_family != st->family) {
1502                         tw = tw_next(tw);
1503                 }
1504                 if (tw) {
1505                         cur = tw;
1506                         goto out;
1507                 }
1508                 read_unlock(&tcp_hashinfo.ehash[st->bucket].lock);
1509                 st->state = TCP_SEQ_STATE_ESTABLISHED;
1510
1511                 /* We can reschedule between buckets: */
1512                 cond_resched_softirq();
1513
1514                 if (++st->bucket < tcp_hashinfo.ehash_size) {
1515                         read_lock(&tcp_hashinfo.ehash[st->bucket].lock);
1516                         sk = sk_head(&tcp_hashinfo.ehash[st->bucket].chain);
1517                 } else {
1518                         cur = NULL;
1519                         goto out;
1520                 }
1521         } else
1522                 sk = sk_next(sk);
1523
1524         sk_for_each_from(sk, node) {
1525                 if (sk->sk_family == st->family)
1526                         goto found;
1527         }
1528
1529         st->state = TCP_SEQ_STATE_TIME_WAIT;
1530         tw = tw_head(&tcp_hashinfo.ehash[st->bucket + tcp_hashinfo.ehash_size].chain);
1531         goto get_tw;
1532 found:
1533         cur = sk;
1534 out:
1535         return cur;
1536 }
1537
1538 static void *established_get_idx(struct seq_file *seq, loff_t pos)
1539 {
1540         void *rc = established_get_first(seq);
1541
1542         while (rc && pos) {
1543                 rc = established_get_next(seq, rc);
1544                 --pos;
1545         }               
1546         return rc;
1547 }
1548
1549 static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
1550 {
1551         void *rc;
1552         struct tcp_iter_state* st = seq->private;
1553
1554         inet_listen_lock(&tcp_hashinfo);
1555         st->state = TCP_SEQ_STATE_LISTENING;
1556         rc        = listening_get_idx(seq, &pos);
1557
1558         if (!rc) {
1559                 inet_listen_unlock(&tcp_hashinfo);
1560                 local_bh_disable();
1561                 st->state = TCP_SEQ_STATE_ESTABLISHED;
1562                 rc        = established_get_idx(seq, pos);
1563         }
1564
1565         return rc;
1566 }
1567
1568 static void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
1569 {
1570         struct tcp_iter_state* st = seq->private;
1571         st->state = TCP_SEQ_STATE_LISTENING;
1572         st->num = 0;
1573         return *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
1574 }
1575
1576 static void *tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos)
1577 {
1578         void *rc = NULL;
1579         struct tcp_iter_state* st;
1580
1581         if (v == SEQ_START_TOKEN) {
1582                 rc = tcp_get_idx(seq, 0);
1583                 goto out;
1584         }
1585         st = seq->private;
1586
1587         switch (st->state) {
1588         case TCP_SEQ_STATE_OPENREQ:
1589         case TCP_SEQ_STATE_LISTENING:
1590                 rc = listening_get_next(seq, v);
1591                 if (!rc) {
1592                         inet_listen_unlock(&tcp_hashinfo);
1593                         local_bh_disable();
1594                         st->state = TCP_SEQ_STATE_ESTABLISHED;
1595                         rc        = established_get_first(seq);
1596                 }
1597                 break;
1598         case TCP_SEQ_STATE_ESTABLISHED:
1599         case TCP_SEQ_STATE_TIME_WAIT:
1600                 rc = established_get_next(seq, v);
1601                 break;
1602         }
1603 out:
1604         ++*pos;
1605         return rc;
1606 }
1607
1608 static void tcp_seq_stop(struct seq_file *seq, void *v)
1609 {
1610         struct tcp_iter_state* st = seq->private;
1611
1612         switch (st->state) {
1613         case TCP_SEQ_STATE_OPENREQ:
1614                 if (v) {
1615                         struct inet_connection_sock *icsk = inet_csk(st->syn_wait_sk);
1616                         read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1617                 }
1618         case TCP_SEQ_STATE_LISTENING:
1619                 if (v != SEQ_START_TOKEN)
1620                         inet_listen_unlock(&tcp_hashinfo);
1621                 break;
1622         case TCP_SEQ_STATE_TIME_WAIT:
1623         case TCP_SEQ_STATE_ESTABLISHED:
1624                 if (v)
1625                         read_unlock(&tcp_hashinfo.ehash[st->bucket].lock);
1626                 local_bh_enable();
1627                 break;
1628         }
1629 }
1630
1631 static int tcp_seq_open(struct inode *inode, struct file *file)
1632 {
1633         struct tcp_seq_afinfo *afinfo = PDE(inode)->data;
1634         struct seq_file *seq;
1635         struct tcp_iter_state *s;
1636         int rc;
1637
1638         if (unlikely(afinfo == NULL))
1639                 return -EINVAL;
1640
1641         s = kzalloc(sizeof(*s), GFP_KERNEL);
1642         if (!s)
1643                 return -ENOMEM;
1644         s->family               = afinfo->family;
1645         s->seq_ops.start        = tcp_seq_start;
1646         s->seq_ops.next         = tcp_seq_next;
1647         s->seq_ops.show         = afinfo->seq_show;
1648         s->seq_ops.stop         = tcp_seq_stop;
1649
1650         rc = seq_open(file, &s->seq_ops);
1651         if (rc)
1652                 goto out_kfree;
1653         seq          = file->private_data;
1654         seq->private = s;
1655 out:
1656         return rc;
1657 out_kfree:
1658         kfree(s);
1659         goto out;
1660 }
1661
1662 int tcp_proc_register(struct tcp_seq_afinfo *afinfo)
1663 {
1664         int rc = 0;
1665         struct proc_dir_entry *p;
1666
1667         if (!afinfo)
1668                 return -EINVAL;
1669         afinfo->seq_fops->owner         = afinfo->owner;
1670         afinfo->seq_fops->open          = tcp_seq_open;
1671         afinfo->seq_fops->read          = seq_read;
1672         afinfo->seq_fops->llseek        = seq_lseek;
1673         afinfo->seq_fops->release       = seq_release_private;
1674         
1675         p = proc_net_fops_create(afinfo->name, S_IRUGO, afinfo->seq_fops);
1676         if (p)
1677                 p->data = afinfo;
1678         else
1679                 rc = -ENOMEM;
1680         return rc;
1681 }
1682
1683 void tcp_proc_unregister(struct tcp_seq_afinfo *afinfo)
1684 {
1685         if (!afinfo)
1686                 return;
1687         proc_net_remove(afinfo->name);
1688         memset(afinfo->seq_fops, 0, sizeof(*afinfo->seq_fops)); 
1689 }
1690
1691 static void get_openreq4(struct sock *sk, struct request_sock *req,
1692                          char *tmpbuf, int i, int uid)
1693 {
1694         const struct inet_request_sock *ireq = inet_rsk(req);
1695         int ttd = req->expires - jiffies;
1696
1697         sprintf(tmpbuf, "%4d: %08X:%04X %08X:%04X"
1698                 " %02X %08X:%08X %02X:%08lX %08X %5d %8d %u %d %p",
1699                 i,
1700                 ireq->loc_addr,
1701                 ntohs(inet_sk(sk)->sport),
1702                 ireq->rmt_addr,
1703                 ntohs(ireq->rmt_port),
1704                 TCP_SYN_RECV,
1705                 0, 0, /* could print option size, but that is af dependent. */
1706                 1,    /* timers active (only the expire timer) */
1707                 jiffies_to_clock_t(ttd),
1708                 req->retrans,
1709                 uid,
1710                 0,  /* non standard timer */
1711                 0, /* open_requests have no inode */
1712                 atomic_read(&sk->sk_refcnt),
1713                 req);
1714 }
1715
1716 static void get_tcp4_sock(struct sock *sp, char *tmpbuf, int i)
1717 {
1718         int timer_active;
1719         unsigned long timer_expires;
1720         struct tcp_sock *tp = tcp_sk(sp);
1721         const struct inet_connection_sock *icsk = inet_csk(sp);
1722         struct inet_sock *inet = inet_sk(sp);
1723         unsigned int dest = inet->daddr;
1724         unsigned int src = inet->rcv_saddr;
1725         __u16 destp = ntohs(inet->dport);
1726         __u16 srcp = ntohs(inet->sport);
1727
1728         if (icsk->icsk_pending == ICSK_TIME_RETRANS) {
1729                 timer_active    = 1;
1730                 timer_expires   = icsk->icsk_timeout;
1731         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
1732                 timer_active    = 4;
1733                 timer_expires   = icsk->icsk_timeout;
1734         } else if (timer_pending(&sp->sk_timer)) {
1735                 timer_active    = 2;
1736                 timer_expires   = sp->sk_timer.expires;
1737         } else {
1738                 timer_active    = 0;
1739                 timer_expires = jiffies;
1740         }
1741
1742         sprintf(tmpbuf, "%4d: %08X:%04X %08X:%04X %02X %08X:%08X %02X:%08lX "
1743                         "%08X %5d %8d %lu %d %p %u %u %u %u %d",
1744                 i, src, srcp, dest, destp, sp->sk_state,
1745                 tp->write_seq - tp->snd_una,
1746                 (sp->sk_state == TCP_LISTEN) ? sp->sk_ack_backlog : (tp->rcv_nxt - tp->copied_seq),
1747                 timer_active,
1748                 jiffies_to_clock_t(timer_expires - jiffies),
1749                 icsk->icsk_retransmits,
1750                 sock_i_uid(sp),
1751                 icsk->icsk_probes_out,
1752                 sock_i_ino(sp),
1753                 atomic_read(&sp->sk_refcnt), sp,
1754                 icsk->icsk_rto,
1755                 icsk->icsk_ack.ato,
1756                 (icsk->icsk_ack.quick << 1) | icsk->icsk_ack.pingpong,
1757                 tp->snd_cwnd,
1758                 tp->snd_ssthresh >= 0xFFFF ? -1 : tp->snd_ssthresh);
1759 }
1760
1761 static void get_timewait4_sock(struct inet_timewait_sock *tw, char *tmpbuf, int i)
1762 {
1763         unsigned int dest, src;
1764         __u16 destp, srcp;
1765         int ttd = tw->tw_ttd - jiffies;
1766
1767         if (ttd < 0)
1768                 ttd = 0;
1769
1770         dest  = tw->tw_daddr;
1771         src   = tw->tw_rcv_saddr;
1772         destp = ntohs(tw->tw_dport);
1773         srcp  = ntohs(tw->tw_sport);
1774
1775         sprintf(tmpbuf, "%4d: %08X:%04X %08X:%04X"
1776                 " %02X %08X:%08X %02X:%08lX %08X %5d %8d %d %d %p",
1777                 i, src, srcp, dest, destp, tw->tw_substate, 0, 0,
1778                 3, jiffies_to_clock_t(ttd), 0, 0, 0, 0,
1779                 atomic_read(&tw->tw_refcnt), tw);
1780 }
1781
1782 #define TMPSZ 150
1783
1784 static int tcp4_seq_show(struct seq_file *seq, void *v)
1785 {
1786         struct tcp_iter_state* st;
1787         char tmpbuf[TMPSZ + 1];
1788
1789         if (v == SEQ_START_TOKEN) {
1790                 seq_printf(seq, "%-*s\n", TMPSZ - 1,
1791                            "  sl  local_address rem_address   st tx_queue "
1792                            "rx_queue tr tm->when retrnsmt   uid  timeout "
1793                            "inode");
1794                 goto out;
1795         }
1796         st = seq->private;
1797
1798         switch (st->state) {
1799         case TCP_SEQ_STATE_LISTENING:
1800         case TCP_SEQ_STATE_ESTABLISHED:
1801                 get_tcp4_sock(v, tmpbuf, st->num);
1802                 break;
1803         case TCP_SEQ_STATE_OPENREQ:
1804                 get_openreq4(st->syn_wait_sk, v, tmpbuf, st->num, st->uid);
1805                 break;
1806         case TCP_SEQ_STATE_TIME_WAIT:
1807                 get_timewait4_sock(v, tmpbuf, st->num);
1808                 break;
1809         }
1810         seq_printf(seq, "%-*s\n", TMPSZ - 1, tmpbuf);
1811 out:
1812         return 0;
1813 }
1814
1815 static struct file_operations tcp4_seq_fops;
1816 static struct tcp_seq_afinfo tcp4_seq_afinfo = {
1817         .owner          = THIS_MODULE,
1818         .name           = "tcp",
1819         .family         = AF_INET,
1820         .seq_show       = tcp4_seq_show,
1821         .seq_fops       = &tcp4_seq_fops,
1822 };
1823
1824 int __init tcp4_proc_init(void)
1825 {
1826         return tcp_proc_register(&tcp4_seq_afinfo);
1827 }
1828
1829 void tcp4_proc_exit(void)
1830 {
1831         tcp_proc_unregister(&tcp4_seq_afinfo);
1832 }
1833 #endif /* CONFIG_PROC_FS */
1834
1835 struct proto tcp_prot = {
1836         .name                   = "TCP",
1837         .owner                  = THIS_MODULE,
1838         .close                  = tcp_close,
1839         .connect                = tcp_v4_connect,
1840         .disconnect             = tcp_disconnect,
1841         .accept                 = inet_csk_accept,
1842         .ioctl                  = tcp_ioctl,
1843         .init                   = tcp_v4_init_sock,
1844         .destroy                = tcp_v4_destroy_sock,
1845         .shutdown               = tcp_shutdown,
1846         .setsockopt             = tcp_setsockopt,
1847         .getsockopt             = tcp_getsockopt,
1848         .sendmsg                = tcp_sendmsg,
1849         .recvmsg                = tcp_recvmsg,
1850         .backlog_rcv            = tcp_v4_do_rcv,
1851         .hash                   = tcp_v4_hash,
1852         .unhash                 = tcp_unhash,
1853         .get_port               = tcp_v4_get_port,
1854         .enter_memory_pressure  = tcp_enter_memory_pressure,
1855         .sockets_allocated      = &tcp_sockets_allocated,
1856         .orphan_count           = &tcp_orphan_count,
1857         .memory_allocated       = &tcp_memory_allocated,
1858         .memory_pressure        = &tcp_memory_pressure,
1859         .sysctl_mem             = sysctl_tcp_mem,
1860         .sysctl_wmem            = sysctl_tcp_wmem,
1861         .sysctl_rmem            = sysctl_tcp_rmem,
1862         .max_header             = MAX_TCP_HEADER,
1863         .obj_size               = sizeof(struct tcp_sock),
1864         .twsk_prot              = &tcp_timewait_sock_ops,
1865         .rsk_prot               = &tcp_request_sock_ops,
1866 #ifdef CONFIG_COMPAT
1867         .compat_setsockopt      = compat_tcp_setsockopt,
1868         .compat_getsockopt      = compat_tcp_getsockopt,
1869 #endif
1870 };
1871
1872 void __init tcp_v4_init(struct net_proto_family *ops)
1873 {
1874         if (inet_csk_ctl_sock_create(&tcp_socket, PF_INET, SOCK_RAW, IPPROTO_TCP) < 0)
1875                 panic("Failed to create the TCP control socket.\n");
1876 }
1877
1878 EXPORT_SYMBOL(ipv4_specific);
1879 EXPORT_SYMBOL(tcp_hashinfo);
1880 EXPORT_SYMBOL(tcp_prot);
1881 EXPORT_SYMBOL(tcp_unhash);
1882 EXPORT_SYMBOL(tcp_v4_conn_request);
1883 EXPORT_SYMBOL(tcp_v4_connect);
1884 EXPORT_SYMBOL(tcp_v4_do_rcv);
1885 EXPORT_SYMBOL(tcp_v4_remember_stamp);
1886 EXPORT_SYMBOL(tcp_v4_send_check);
1887 EXPORT_SYMBOL(tcp_v4_syn_recv_sock);
1888
1889 #ifdef CONFIG_PROC_FS
1890 EXPORT_SYMBOL(tcp_proc_register);
1891 EXPORT_SYMBOL(tcp_proc_unregister);
1892 #endif
1893 EXPORT_SYMBOL(sysctl_local_port_range);
1894 EXPORT_SYMBOL(sysctl_tcp_low_latency);
1895