Merge git://git.skbuff.net/gitroot/yoshfuji/linux-2.6.14+advapi-fix/
[linux-2.6] / net / ipv4 / tcp_minisocks.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_minisocks.c,v 1.15 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *              Florian La Roche, <flla@stud.uni-sb.de>
15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *              Jorge Cwik, <jorge@laser.satlink.net>
21  */
22
23 #include <linux/config.h>
24 #include <linux/mm.h>
25 #include <linux/module.h>
26 #include <linux/sysctl.h>
27 #include <linux/workqueue.h>
28 #include <net/tcp.h>
29 #include <net/inet_common.h>
30 #include <net/xfrm.h>
31
32 #ifdef CONFIG_SYSCTL
33 #define SYNC_INIT 0 /* let the user enable it */
34 #else
35 #define SYNC_INIT 1
36 #endif
37
38 int sysctl_tcp_syncookies = SYNC_INIT; 
39 int sysctl_tcp_abort_on_overflow;
40
41 struct inet_timewait_death_row tcp_death_row = {
42         .sysctl_max_tw_buckets = NR_FILE * 2,
43         .period         = TCP_TIMEWAIT_LEN / INET_TWDR_TWKILL_SLOTS,
44         .death_lock     = SPIN_LOCK_UNLOCKED,
45         .hashinfo       = &tcp_hashinfo,
46         .tw_timer       = TIMER_INITIALIZER(inet_twdr_hangman, 0,
47                                             (unsigned long)&tcp_death_row),
48         .twkill_work    = __WORK_INITIALIZER(tcp_death_row.twkill_work,
49                                              inet_twdr_twkill_work,
50                                              &tcp_death_row),
51 /* Short-time timewait calendar */
52
53         .twcal_hand     = -1,
54         .twcal_timer    = TIMER_INITIALIZER(inet_twdr_twcal_tick, 0,
55                                             (unsigned long)&tcp_death_row),
56 };
57
58 EXPORT_SYMBOL_GPL(tcp_death_row);
59
60 static __inline__ int tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
61 {
62         if (seq == s_win)
63                 return 1;
64         if (after(end_seq, s_win) && before(seq, e_win))
65                 return 1;
66         return (seq == e_win && seq == end_seq);
67 }
68
69 /* 
70  * * Main purpose of TIME-WAIT state is to close connection gracefully,
71  *   when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
72  *   (and, probably, tail of data) and one or more our ACKs are lost.
73  * * What is TIME-WAIT timeout? It is associated with maximal packet
74  *   lifetime in the internet, which results in wrong conclusion, that
75  *   it is set to catch "old duplicate segments" wandering out of their path.
76  *   It is not quite correct. This timeout is calculated so that it exceeds
77  *   maximal retransmission timeout enough to allow to lose one (or more)
78  *   segments sent by peer and our ACKs. This time may be calculated from RTO.
79  * * When TIME-WAIT socket receives RST, it means that another end
80  *   finally closed and we are allowed to kill TIME-WAIT too.
81  * * Second purpose of TIME-WAIT is catching old duplicate segments.
82  *   Well, certainly it is pure paranoia, but if we load TIME-WAIT
83  *   with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
84  * * If we invented some more clever way to catch duplicates
85  *   (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
86  *
87  * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
88  * When you compare it to RFCs, please, read section SEGMENT ARRIVES
89  * from the very beginning.
90  *
91  * NOTE. With recycling (and later with fin-wait-2) TW bucket
92  * is _not_ stateless. It means, that strictly speaking we must
93  * spinlock it. I do not want! Well, probability of misbehaviour
94  * is ridiculously low and, seems, we could use some mb() tricks
95  * to avoid misread sequence numbers, states etc.  --ANK
96  */
97 enum tcp_tw_status
98 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
99                            const struct tcphdr *th)
100 {
101         struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
102         struct tcp_options_received tmp_opt;
103         int paws_reject = 0;
104
105         tmp_opt.saw_tstamp = 0;
106         if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
107                 tcp_parse_options(skb, &tmp_opt, 0);
108
109                 if (tmp_opt.saw_tstamp) {
110                         tmp_opt.ts_recent       = tcptw->tw_ts_recent;
111                         tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
112                         paws_reject = tcp_paws_check(&tmp_opt, th->rst);
113                 }
114         }
115
116         if (tw->tw_substate == TCP_FIN_WAIT2) {
117                 /* Just repeat all the checks of tcp_rcv_state_process() */
118
119                 /* Out of window, send ACK */
120                 if (paws_reject ||
121                     !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
122                                    tcptw->tw_rcv_nxt,
123                                    tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
124                         return TCP_TW_ACK;
125
126                 if (th->rst)
127                         goto kill;
128
129                 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
130                         goto kill_with_rst;
131
132                 /* Dup ACK? */
133                 if (!after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
134                     TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
135                         inet_twsk_put(tw);
136                         return TCP_TW_SUCCESS;
137                 }
138
139                 /* New data or FIN. If new data arrive after half-duplex close,
140                  * reset.
141                  */
142                 if (!th->fin ||
143                     TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) {
144 kill_with_rst:
145                         inet_twsk_deschedule(tw, &tcp_death_row);
146                         inet_twsk_put(tw);
147                         return TCP_TW_RST;
148                 }
149
150                 /* FIN arrived, enter true time-wait state. */
151                 tw->tw_substate   = TCP_TIME_WAIT;
152                 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
153                 if (tmp_opt.saw_tstamp) {
154                         tcptw->tw_ts_recent_stamp = xtime.tv_sec;
155                         tcptw->tw_ts_recent       = tmp_opt.rcv_tsval;
156                 }
157
158                 /* I am shamed, but failed to make it more elegant.
159                  * Yes, it is direct reference to IP, which is impossible
160                  * to generalize to IPv6. Taking into account that IPv6
161                  * do not understand recycling in any case, it not
162                  * a big problem in practice. --ANK */
163                 if (tw->tw_family == AF_INET &&
164                     tcp_death_row.sysctl_tw_recycle && tcptw->tw_ts_recent_stamp &&
165                     tcp_v4_tw_remember_stamp(tw))
166                         inet_twsk_schedule(tw, &tcp_death_row, tw->tw_timeout,
167                                            TCP_TIMEWAIT_LEN);
168                 else
169                         inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
170                                            TCP_TIMEWAIT_LEN);
171                 return TCP_TW_ACK;
172         }
173
174         /*
175          *      Now real TIME-WAIT state.
176          *
177          *      RFC 1122:
178          *      "When a connection is [...] on TIME-WAIT state [...]
179          *      [a TCP] MAY accept a new SYN from the remote TCP to
180          *      reopen the connection directly, if it:
181          *      
182          *      (1)  assigns its initial sequence number for the new
183          *      connection to be larger than the largest sequence
184          *      number it used on the previous connection incarnation,
185          *      and
186          *
187          *      (2)  returns to TIME-WAIT state if the SYN turns out 
188          *      to be an old duplicate".
189          */
190
191         if (!paws_reject &&
192             (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
193              (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
194                 /* In window segment, it may be only reset or bare ack. */
195
196                 if (th->rst) {
197                         /* This is TIME_WAIT assassination, in two flavors.
198                          * Oh well... nobody has a sufficient solution to this
199                          * protocol bug yet.
200                          */
201                         if (sysctl_tcp_rfc1337 == 0) {
202 kill:
203                                 inet_twsk_deschedule(tw, &tcp_death_row);
204                                 inet_twsk_put(tw);
205                                 return TCP_TW_SUCCESS;
206                         }
207                 }
208                 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
209                                    TCP_TIMEWAIT_LEN);
210
211                 if (tmp_opt.saw_tstamp) {
212                         tcptw->tw_ts_recent       = tmp_opt.rcv_tsval;
213                         tcptw->tw_ts_recent_stamp = xtime.tv_sec;
214                 }
215
216                 inet_twsk_put(tw);
217                 return TCP_TW_SUCCESS;
218         }
219
220         /* Out of window segment.
221
222            All the segments are ACKed immediately.
223
224            The only exception is new SYN. We accept it, if it is
225            not old duplicate and we are not in danger to be killed
226            by delayed old duplicates. RFC check is that it has
227            newer sequence number works at rates <40Mbit/sec.
228            However, if paws works, it is reliable AND even more,
229            we even may relax silly seq space cutoff.
230
231            RED-PEN: we violate main RFC requirement, if this SYN will appear
232            old duplicate (i.e. we receive RST in reply to SYN-ACK),
233            we must return socket to time-wait state. It is not good,
234            but not fatal yet.
235          */
236
237         if (th->syn && !th->rst && !th->ack && !paws_reject &&
238             (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
239              (tmp_opt.saw_tstamp &&
240               (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
241                 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
242                 if (isn == 0)
243                         isn++;
244                 TCP_SKB_CB(skb)->when = isn;
245                 return TCP_TW_SYN;
246         }
247
248         if (paws_reject)
249                 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
250
251         if(!th->rst) {
252                 /* In this case we must reset the TIMEWAIT timer.
253                  *
254                  * If it is ACKless SYN it may be both old duplicate
255                  * and new good SYN with random sequence number <rcv_nxt.
256                  * Do not reschedule in the last case.
257                  */
258                 if (paws_reject || th->ack)
259                         inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
260                                            TCP_TIMEWAIT_LEN);
261
262                 /* Send ACK. Note, we do not put the bucket,
263                  * it will be released by caller.
264                  */
265                 return TCP_TW_ACK;
266         }
267         inet_twsk_put(tw);
268         return TCP_TW_SUCCESS;
269 }
270
271 /* 
272  * Move a socket to time-wait or dead fin-wait-2 state.
273  */ 
274 void tcp_time_wait(struct sock *sk, int state, int timeo)
275 {
276         struct inet_timewait_sock *tw = NULL;
277         const struct tcp_sock *tp = tcp_sk(sk);
278         int recycle_ok = 0;
279
280         if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp)
281                 recycle_ok = tp->af_specific->remember_stamp(sk);
282
283         if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets)
284                 tw = inet_twsk_alloc(sk, state);
285
286         if (tw != NULL) {
287                 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
288                 const struct inet_connection_sock *icsk = inet_csk(sk);
289                 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
290
291                 tw->tw_rcv_wscale       = tp->rx_opt.rcv_wscale;
292                 tcptw->tw_rcv_nxt       = tp->rcv_nxt;
293                 tcptw->tw_snd_nxt       = tp->snd_nxt;
294                 tcptw->tw_rcv_wnd       = tcp_receive_window(tp);
295                 tcptw->tw_ts_recent     = tp->rx_opt.ts_recent;
296                 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
297
298 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
299                 if (tw->tw_family == PF_INET6) {
300                         struct ipv6_pinfo *np = inet6_sk(sk);
301                         struct tcp6_timewait_sock *tcp6tw = tcp6_twsk((struct sock *)tw);
302
303                         ipv6_addr_copy(&tcp6tw->tw_v6_daddr, &np->daddr);
304                         ipv6_addr_copy(&tcp6tw->tw_v6_rcv_saddr, &np->rcv_saddr);
305                         tw->tw_ipv6only = np->ipv6only;
306                 }
307 #endif
308                 /* Linkage updates. */
309                 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
310
311                 /* Get the TIME_WAIT timeout firing. */
312                 if (timeo < rto)
313                         timeo = rto;
314
315                 if (recycle_ok) {
316                         tw->tw_timeout = rto;
317                 } else {
318                         tw->tw_timeout = TCP_TIMEWAIT_LEN;
319                         if (state == TCP_TIME_WAIT)
320                                 timeo = TCP_TIMEWAIT_LEN;
321                 }
322
323                 inet_twsk_schedule(tw, &tcp_death_row, timeo,
324                                    TCP_TIMEWAIT_LEN);
325                 inet_twsk_put(tw);
326         } else {
327                 /* Sorry, if we're out of memory, just CLOSE this
328                  * socket up.  We've got bigger problems than
329                  * non-graceful socket closings.
330                  */
331                 if (net_ratelimit())
332                         printk(KERN_INFO "TCP: time wait bucket table overflow\n");
333         }
334
335         tcp_update_metrics(sk);
336         tcp_done(sk);
337 }
338
339 /* This is not only more efficient than what we used to do, it eliminates
340  * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
341  *
342  * Actually, we could lots of memory writes here. tp of listening
343  * socket contains all necessary default parameters.
344  */
345 struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb)
346 {
347         struct sock *newsk = inet_csk_clone(sk, req, GFP_ATOMIC);
348
349         if (newsk != NULL) {
350                 const struct inet_request_sock *ireq = inet_rsk(req);
351                 struct tcp_request_sock *treq = tcp_rsk(req);
352                 struct inet_connection_sock *newicsk = inet_csk(sk);
353                 struct tcp_sock *newtp;
354
355                 /* Now setup tcp_sock */
356                 newtp = tcp_sk(newsk);
357                 newtp->pred_flags = 0;
358                 newtp->rcv_nxt = treq->rcv_isn + 1;
359                 newtp->snd_nxt = newtp->snd_una = newtp->snd_sml = treq->snt_isn + 1;
360
361                 tcp_prequeue_init(newtp);
362
363                 tcp_init_wl(newtp, treq->snt_isn, treq->rcv_isn);
364
365                 newtp->srtt = 0;
366                 newtp->mdev = TCP_TIMEOUT_INIT;
367                 newicsk->icsk_rto = TCP_TIMEOUT_INIT;
368
369                 newtp->packets_out = 0;
370                 newtp->left_out = 0;
371                 newtp->retrans_out = 0;
372                 newtp->sacked_out = 0;
373                 newtp->fackets_out = 0;
374                 newtp->snd_ssthresh = 0x7fffffff;
375
376                 /* So many TCP implementations out there (incorrectly) count the
377                  * initial SYN frame in their delayed-ACK and congestion control
378                  * algorithms that we must have the following bandaid to talk
379                  * efficiently to them.  -DaveM
380                  */
381                 newtp->snd_cwnd = 2;
382                 newtp->snd_cwnd_cnt = 0;
383                 newtp->bytes_acked = 0;
384
385                 newtp->frto_counter = 0;
386                 newtp->frto_highmark = 0;
387
388                 newicsk->icsk_ca_ops = &tcp_init_congestion_ops;
389
390                 tcp_set_ca_state(newsk, TCP_CA_Open);
391                 tcp_init_xmit_timers(newsk);
392                 skb_queue_head_init(&newtp->out_of_order_queue);
393                 newtp->rcv_wup = treq->rcv_isn + 1;
394                 newtp->write_seq = treq->snt_isn + 1;
395                 newtp->pushed_seq = newtp->write_seq;
396                 newtp->copied_seq = treq->rcv_isn + 1;
397
398                 newtp->rx_opt.saw_tstamp = 0;
399
400                 newtp->rx_opt.dsack = 0;
401                 newtp->rx_opt.eff_sacks = 0;
402
403                 newtp->rx_opt.num_sacks = 0;
404                 newtp->urg_data = 0;
405
406                 if (sock_flag(newsk, SOCK_KEEPOPEN))
407                         inet_csk_reset_keepalive_timer(newsk,
408                                                        keepalive_time_when(newtp));
409
410                 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
411                 if((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
412                         if (sysctl_tcp_fack)
413                                 newtp->rx_opt.sack_ok |= 2;
414                 }
415                 newtp->window_clamp = req->window_clamp;
416                 newtp->rcv_ssthresh = req->rcv_wnd;
417                 newtp->rcv_wnd = req->rcv_wnd;
418                 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
419                 if (newtp->rx_opt.wscale_ok) {
420                         newtp->rx_opt.snd_wscale = ireq->snd_wscale;
421                         newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
422                 } else {
423                         newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
424                         newtp->window_clamp = min(newtp->window_clamp, 65535U);
425                 }
426                 newtp->snd_wnd = ntohs(skb->h.th->window) << newtp->rx_opt.snd_wscale;
427                 newtp->max_window = newtp->snd_wnd;
428
429                 if (newtp->rx_opt.tstamp_ok) {
430                         newtp->rx_opt.ts_recent = req->ts_recent;
431                         newtp->rx_opt.ts_recent_stamp = xtime.tv_sec;
432                         newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
433                 } else {
434                         newtp->rx_opt.ts_recent_stamp = 0;
435                         newtp->tcp_header_len = sizeof(struct tcphdr);
436                 }
437                 if (skb->len >= TCP_MIN_RCVMSS+newtp->tcp_header_len)
438                         newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
439                 newtp->rx_opt.mss_clamp = req->mss;
440                 TCP_ECN_openreq_child(newtp, req);
441                 if (newtp->ecn_flags&TCP_ECN_OK)
442                         sock_set_flag(newsk, SOCK_NO_LARGESEND);
443
444                 TCP_INC_STATS_BH(TCP_MIB_PASSIVEOPENS);
445         }
446         return newsk;
447 }
448
449 /* 
450  *      Process an incoming packet for SYN_RECV sockets represented
451  *      as a request_sock.
452  */
453
454 struct sock *tcp_check_req(struct sock *sk,struct sk_buff *skb,
455                            struct request_sock *req,
456                            struct request_sock **prev)
457 {
458         struct tcphdr *th = skb->h.th;
459         struct tcp_sock *tp = tcp_sk(sk);
460         u32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
461         int paws_reject = 0;
462         struct tcp_options_received tmp_opt;
463         struct sock *child;
464
465         tmp_opt.saw_tstamp = 0;
466         if (th->doff > (sizeof(struct tcphdr)>>2)) {
467                 tcp_parse_options(skb, &tmp_opt, 0);
468
469                 if (tmp_opt.saw_tstamp) {
470                         tmp_opt.ts_recent = req->ts_recent;
471                         /* We do not store true stamp, but it is not required,
472                          * it can be estimated (approximately)
473                          * from another data.
474                          */
475                         tmp_opt.ts_recent_stamp = xtime.tv_sec - ((TCP_TIMEOUT_INIT/HZ)<<req->retrans);
476                         paws_reject = tcp_paws_check(&tmp_opt, th->rst);
477                 }
478         }
479
480         /* Check for pure retransmitted SYN. */
481         if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
482             flg == TCP_FLAG_SYN &&
483             !paws_reject) {
484                 /*
485                  * RFC793 draws (Incorrectly! It was fixed in RFC1122)
486                  * this case on figure 6 and figure 8, but formal
487                  * protocol description says NOTHING.
488                  * To be more exact, it says that we should send ACK,
489                  * because this segment (at least, if it has no data)
490                  * is out of window.
491                  *
492                  *  CONCLUSION: RFC793 (even with RFC1122) DOES NOT
493                  *  describe SYN-RECV state. All the description
494                  *  is wrong, we cannot believe to it and should
495                  *  rely only on common sense and implementation
496                  *  experience.
497                  *
498                  * Enforce "SYN-ACK" according to figure 8, figure 6
499                  * of RFC793, fixed by RFC1122.
500                  */
501                 req->rsk_ops->rtx_syn_ack(sk, req, NULL);
502                 return NULL;
503         }
504
505         /* Further reproduces section "SEGMENT ARRIVES"
506            for state SYN-RECEIVED of RFC793.
507            It is broken, however, it does not work only
508            when SYNs are crossed.
509
510            You would think that SYN crossing is impossible here, since
511            we should have a SYN_SENT socket (from connect()) on our end,
512            but this is not true if the crossed SYNs were sent to both
513            ends by a malicious third party.  We must defend against this,
514            and to do that we first verify the ACK (as per RFC793, page
515            36) and reset if it is invalid.  Is this a true full defense?
516            To convince ourselves, let us consider a way in which the ACK
517            test can still pass in this 'malicious crossed SYNs' case.
518            Malicious sender sends identical SYNs (and thus identical sequence
519            numbers) to both A and B:
520
521                 A: gets SYN, seq=7
522                 B: gets SYN, seq=7
523
524            By our good fortune, both A and B select the same initial
525            send sequence number of seven :-)
526
527                 A: sends SYN|ACK, seq=7, ack_seq=8
528                 B: sends SYN|ACK, seq=7, ack_seq=8
529
530            So we are now A eating this SYN|ACK, ACK test passes.  So
531            does sequence test, SYN is truncated, and thus we consider
532            it a bare ACK.
533
534            If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
535            bare ACK.  Otherwise, we create an established connection.  Both
536            ends (listening sockets) accept the new incoming connection and try
537            to talk to each other. 8-)
538
539            Note: This case is both harmless, and rare.  Possibility is about the
540            same as us discovering intelligent life on another plant tomorrow.
541
542            But generally, we should (RFC lies!) to accept ACK
543            from SYNACK both here and in tcp_rcv_state_process().
544            tcp_rcv_state_process() does not, hence, we do not too.
545
546            Note that the case is absolutely generic:
547            we cannot optimize anything here without
548            violating protocol. All the checks must be made
549            before attempt to create socket.
550          */
551
552         /* RFC793 page 36: "If the connection is in any non-synchronized state ...
553          *                  and the incoming segment acknowledges something not yet
554          *                  sent (the segment carries an unacceptable ACK) ...
555          *                  a reset is sent."
556          *
557          * Invalid ACK: reset will be sent by listening socket
558          */
559         if ((flg & TCP_FLAG_ACK) &&
560             (TCP_SKB_CB(skb)->ack_seq != tcp_rsk(req)->snt_isn + 1))
561                 return sk;
562
563         /* Also, it would be not so bad idea to check rcv_tsecr, which
564          * is essentially ACK extension and too early or too late values
565          * should cause reset in unsynchronized states.
566          */
567
568         /* RFC793: "first check sequence number". */
569
570         if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
571                                           tcp_rsk(req)->rcv_isn + 1, tcp_rsk(req)->rcv_isn + 1 + req->rcv_wnd)) {
572                 /* Out of window: send ACK and drop. */
573                 if (!(flg & TCP_FLAG_RST))
574                         req->rsk_ops->send_ack(skb, req);
575                 if (paws_reject)
576                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
577                 return NULL;
578         }
579
580         /* In sequence, PAWS is OK. */
581
582         if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_isn + 1))
583                         req->ts_recent = tmp_opt.rcv_tsval;
584
585                 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
586                         /* Truncate SYN, it is out of window starting
587                            at tcp_rsk(req)->rcv_isn + 1. */
588                         flg &= ~TCP_FLAG_SYN;
589                 }
590
591                 /* RFC793: "second check the RST bit" and
592                  *         "fourth, check the SYN bit"
593                  */
594                 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN))
595                         goto embryonic_reset;
596
597                 /* ACK sequence verified above, just make sure ACK is
598                  * set.  If ACK not set, just silently drop the packet.
599                  */
600                 if (!(flg & TCP_FLAG_ACK))
601                         return NULL;
602
603                 /* If TCP_DEFER_ACCEPT is set, drop bare ACK. */
604                 if (inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
605                     TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
606                         inet_rsk(req)->acked = 1;
607                         return NULL;
608                 }
609
610                 /* OK, ACK is valid, create big socket and
611                  * feed this segment to it. It will repeat all
612                  * the tests. THIS SEGMENT MUST MOVE SOCKET TO
613                  * ESTABLISHED STATE. If it will be dropped after
614                  * socket is created, wait for troubles.
615                  */
616                 child = tp->af_specific->syn_recv_sock(sk, skb, req, NULL);
617                 if (child == NULL)
618                         goto listen_overflow;
619
620                 inet_csk_reqsk_queue_unlink(sk, req, prev);
621                 inet_csk_reqsk_queue_removed(sk, req);
622
623                 inet_csk_reqsk_queue_add(sk, req, child);
624                 return child;
625
626         listen_overflow:
627                 if (!sysctl_tcp_abort_on_overflow) {
628                         inet_rsk(req)->acked = 1;
629                         return NULL;
630                 }
631
632         embryonic_reset:
633                 NET_INC_STATS_BH(LINUX_MIB_EMBRYONICRSTS);
634                 if (!(flg & TCP_FLAG_RST))
635                         req->rsk_ops->send_reset(skb);
636
637                 inet_csk_reqsk_queue_drop(sk, req, prev);
638                 return NULL;
639 }
640
641 /*
642  * Queue segment on the new socket if the new socket is active,
643  * otherwise we just shortcircuit this and continue with
644  * the new socket.
645  */
646
647 int tcp_child_process(struct sock *parent, struct sock *child,
648                       struct sk_buff *skb)
649 {
650         int ret = 0;
651         int state = child->sk_state;
652
653         if (!sock_owned_by_user(child)) {
654                 ret = tcp_rcv_state_process(child, skb, skb->h.th, skb->len);
655
656                 /* Wakeup parent, send SIGIO */
657                 if (state == TCP_SYN_RECV && child->sk_state != state)
658                         parent->sk_data_ready(parent, 0);
659         } else {
660                 /* Alas, it is possible again, because we do lookup
661                  * in main socket hash table and lock on listening
662                  * socket does not protect us more.
663                  */
664                 sk_add_backlog(child, skb);
665         }
666
667         bh_unlock_sock(child);
668         sock_put(child);
669         return ret;
670 }
671
672 EXPORT_SYMBOL(tcp_check_req);
673 EXPORT_SYMBOL(tcp_child_process);
674 EXPORT_SYMBOL(tcp_create_openreq_child);
675 EXPORT_SYMBOL(tcp_timewait_state_process);