ac4ce48a6599ddec8ca5d5e3e9f84ece9a11d569
[linux-2.6.git] / 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/mm.h>
24 #include <linux/module.h>
25 #include <linux/sysctl.h>
26 #include <linux/workqueue.h>
27 #include <net/tcp.h>
28 #include <net/inet_common.h>
29 #include <net/xfrm.h>
30
31 #ifdef CONFIG_SYSCTL
32 #define SYNC_INIT 0 /* let the user enable it */
33 #else
34 #define SYNC_INIT 1
35 #endif
36
37 int sysctl_tcp_syncookies __read_mostly = SYNC_INIT;
38 int sysctl_tcp_abort_on_overflow __read_mostly;
39
40 struct inet_timewait_death_row tcp_death_row = {
41         .sysctl_max_tw_buckets = NR_FILE * 2,
42         .period         = TCP_TIMEWAIT_LEN / INET_TWDR_TWKILL_SLOTS,
43         .death_lock     = __SPIN_LOCK_UNLOCKED(tcp_death_row.death_lock),
44         .hashinfo       = &tcp_hashinfo,
45         .tw_timer       = TIMER_INITIALIZER(inet_twdr_hangman, 0,
46                                             (unsigned long)&tcp_death_row),
47         .twkill_work    = __WORK_INITIALIZER(tcp_death_row.twkill_work,
48                                              inet_twdr_twkill_work),
49 /* Short-time timewait calendar */
50
51         .twcal_hand     = -1,
52         .twcal_timer    = TIMER_INITIALIZER(inet_twdr_twcal_tick, 0,
53                                             (unsigned long)&tcp_death_row),
54 };
55
56 EXPORT_SYMBOL_GPL(tcp_death_row);
57
58 static __inline__ int tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
59 {
60         if (seq == s_win)
61                 return 1;
62         if (after(end_seq, s_win) && before(seq, e_win))
63                 return 1;
64         return (seq == e_win && seq == end_seq);
65 }
66
67 /*
68  * * Main purpose of TIME-WAIT state is to close connection gracefully,
69  *   when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
70  *   (and, probably, tail of data) and one or more our ACKs are lost.
71  * * What is TIME-WAIT timeout? It is associated with maximal packet
72  *   lifetime in the internet, which results in wrong conclusion, that
73  *   it is set to catch "old duplicate segments" wandering out of their path.
74  *   It is not quite correct. This timeout is calculated so that it exceeds
75  *   maximal retransmission timeout enough to allow to lose one (or more)
76  *   segments sent by peer and our ACKs. This time may be calculated from RTO.
77  * * When TIME-WAIT socket receives RST, it means that another end
78  *   finally closed and we are allowed to kill TIME-WAIT too.
79  * * Second purpose of TIME-WAIT is catching old duplicate segments.
80  *   Well, certainly it is pure paranoia, but if we load TIME-WAIT
81  *   with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
82  * * If we invented some more clever way to catch duplicates
83  *   (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
84  *
85  * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
86  * When you compare it to RFCs, please, read section SEGMENT ARRIVES
87  * from the very beginning.
88  *
89  * NOTE. With recycling (and later with fin-wait-2) TW bucket
90  * is _not_ stateless. It means, that strictly speaking we must
91  * spinlock it. I do not want! Well, probability of misbehaviour
92  * is ridiculously low and, seems, we could use some mb() tricks
93  * to avoid misread sequence numbers, states etc.  --ANK
94  */
95 enum tcp_tw_status
96 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
97                            const struct tcphdr *th)
98 {
99         struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
100         struct tcp_options_received tmp_opt;
101         int paws_reject = 0;
102
103         tmp_opt.saw_tstamp = 0;
104         if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
105                 tcp_parse_options(skb, &tmp_opt, 0);
106
107                 if (tmp_opt.saw_tstamp) {
108                         tmp_opt.ts_recent       = tcptw->tw_ts_recent;
109                         tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
110                         paws_reject = tcp_paws_check(&tmp_opt, th->rst);
111                 }
112         }
113
114         if (tw->tw_substate == TCP_FIN_WAIT2) {
115                 /* Just repeat all the checks of tcp_rcv_state_process() */
116
117                 /* Out of window, send ACK */
118                 if (paws_reject ||
119                     !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
120                                    tcptw->tw_rcv_nxt,
121                                    tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
122                         return TCP_TW_ACK;
123
124                 if (th->rst)
125                         goto kill;
126
127                 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
128                         goto kill_with_rst;
129
130                 /* Dup ACK? */
131                 if (!after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
132                     TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
133                         inet_twsk_put(tw);
134                         return TCP_TW_SUCCESS;
135                 }
136
137                 /* New data or FIN. If new data arrive after half-duplex close,
138                  * reset.
139                  */
140                 if (!th->fin ||
141                     TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) {
142 kill_with_rst:
143                         inet_twsk_deschedule(tw, &tcp_death_row);
144                         inet_twsk_put(tw);
145                         return TCP_TW_RST;
146                 }
147
148                 /* FIN arrived, enter true time-wait state. */
149                 tw->tw_substate   = TCP_TIME_WAIT;
150                 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
151                 if (tmp_opt.saw_tstamp) {
152                         tcptw->tw_ts_recent_stamp = get_seconds();
153                         tcptw->tw_ts_recent       = tmp_opt.rcv_tsval;
154                 }
155
156                 /* I am shamed, but failed to make it more elegant.
157                  * Yes, it is direct reference to IP, which is impossible
158                  * to generalize to IPv6. Taking into account that IPv6
159                  * do not understand recycling in any case, it not
160                  * a big problem in practice. --ANK */
161                 if (tw->tw_family == AF_INET &&
162                     tcp_death_row.sysctl_tw_recycle && tcptw->tw_ts_recent_stamp &&
163                     tcp_v4_tw_remember_stamp(tw))
164                         inet_twsk_schedule(tw, &tcp_death_row, tw->tw_timeout,
165                                            TCP_TIMEWAIT_LEN);
166                 else
167                         inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
168                                            TCP_TIMEWAIT_LEN);
169                 return TCP_TW_ACK;
170         }
171
172         /*
173          *      Now real TIME-WAIT state.
174          *
175          *      RFC 1122:
176          *      "When a connection is [...] on TIME-WAIT state [...]
177          *      [a TCP] MAY accept a new SYN from the remote TCP to
178          *      reopen the connection directly, if it:
179          *
180          *      (1)  assigns its initial sequence number for the new
181          *      connection to be larger than the largest sequence
182          *      number it used on the previous connection incarnation,
183          *      and
184          *
185          *      (2)  returns to TIME-WAIT state if the SYN turns out
186          *      to be an old duplicate".
187          */
188
189         if (!paws_reject &&
190             (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
191              (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
192                 /* In window segment, it may be only reset or bare ack. */
193
194                 if (th->rst) {
195                         /* This is TIME_WAIT assassination, in two flavors.
196                          * Oh well... nobody has a sufficient solution to this
197                          * protocol bug yet.
198                          */
199                         if (sysctl_tcp_rfc1337 == 0) {
200 kill:
201                                 inet_twsk_deschedule(tw, &tcp_death_row);
202                                 inet_twsk_put(tw);
203                                 return TCP_TW_SUCCESS;
204                         }
205                 }
206                 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
207                                    TCP_TIMEWAIT_LEN);
208
209                 if (tmp_opt.saw_tstamp) {
210                         tcptw->tw_ts_recent       = tmp_opt.rcv_tsval;
211                         tcptw->tw_ts_recent_stamp = get_seconds();
212                 }
213
214                 inet_twsk_put(tw);
215                 return TCP_TW_SUCCESS;
216         }
217
218         /* Out of window segment.
219
220            All the segments are ACKed immediately.
221
222            The only exception is new SYN. We accept it, if it is
223            not old duplicate and we are not in danger to be killed
224            by delayed old duplicates. RFC check is that it has
225            newer sequence number works at rates <40Mbit/sec.
226            However, if paws works, it is reliable AND even more,
227            we even may relax silly seq space cutoff.
228
229            RED-PEN: we violate main RFC requirement, if this SYN will appear
230            old duplicate (i.e. we receive RST in reply to SYN-ACK),
231            we must return socket to time-wait state. It is not good,
232            but not fatal yet.
233          */
234
235         if (th->syn && !th->rst && !th->ack && !paws_reject &&
236             (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
237              (tmp_opt.saw_tstamp &&
238               (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
239                 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
240                 if (isn == 0)
241                         isn++;
242                 TCP_SKB_CB(skb)->when = isn;
243                 return TCP_TW_SYN;
244         }
245
246         if (paws_reject)
247                 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
248
249         if(!th->rst) {
250                 /* In this case we must reset the TIMEWAIT timer.
251                  *
252                  * If it is ACKless SYN it may be both old duplicate
253                  * and new good SYN with random sequence number <rcv_nxt.
254                  * Do not reschedule in the last case.
255                  */
256                 if (paws_reject || th->ack)
257                         inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
258                                            TCP_TIMEWAIT_LEN);
259
260                 /* Send ACK. Note, we do not put the bucket,
261                  * it will be released by caller.
262                  */
263                 return TCP_TW_ACK;
264         }
265         inet_twsk_put(tw);
266         return TCP_TW_SUCCESS;
267 }
268
269 /*
270  * Move a socket to time-wait or dead fin-wait-2 state.
271  */
272 void tcp_time_wait(struct sock *sk, int state, int timeo)
273 {
274         struct inet_timewait_sock *tw = NULL;
275         const struct inet_connection_sock *icsk = inet_csk(sk);
276         const struct tcp_sock *tp = tcp_sk(sk);
277         int recycle_ok = 0;
278
279         if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp)
280                 recycle_ok = icsk->icsk_af_ops->remember_stamp(sk);
281
282         if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets)
283                 tw = inet_twsk_alloc(sk, state);
284
285         if (tw != NULL) {
286                 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
287                 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
288
289                 tw->tw_rcv_wscale       = tp->rx_opt.rcv_wscale;
290                 tcptw->tw_rcv_nxt       = tp->rcv_nxt;
291                 tcptw->tw_snd_nxt       = tp->snd_nxt;
292                 tcptw->tw_rcv_wnd       = tcp_receive_window(tp);
293                 tcptw->tw_ts_recent     = tp->rx_opt.ts_recent;
294                 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
295
296 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
297                 if (tw->tw_family == PF_INET6) {
298                         struct ipv6_pinfo *np = inet6_sk(sk);
299                         struct inet6_timewait_sock *tw6;
300
301                         tw->tw_ipv6_offset = inet6_tw_offset(sk->sk_prot);
302                         tw6 = inet6_twsk((struct sock *)tw);
303                         ipv6_addr_copy(&tw6->tw_v6_daddr, &np->daddr);
304                         ipv6_addr_copy(&tw6->tw_v6_rcv_saddr, &np->rcv_saddr);
305                         tw->tw_ipv6only = np->ipv6only;
306                 }
307 #endif
308
309 #ifdef CONFIG_TCP_MD5SIG
310                 /*
311                  * The timewait bucket does not have the key DB from the
312                  * sock structure. We just make a quick copy of the
313                  * md5 key being used (if indeed we are using one)
314                  * so the timewait ack generating code has the key.
315                  */
316                 do {
317                         struct tcp_md5sig_key *key;
318                         memset(tcptw->tw_md5_key, 0, sizeof(tcptw->tw_md5_key));
319                         tcptw->tw_md5_keylen = 0;
320                         key = tp->af_specific->md5_lookup(sk, sk);
321                         if (key != NULL) {
322                                 memcpy(&tcptw->tw_md5_key, key->key, key->keylen);
323                                 tcptw->tw_md5_keylen = key->keylen;
324                                 if (tcp_alloc_md5sig_pool() == NULL)
325                                         BUG();
326                         }
327                 } while(0);
328 #endif
329
330                 /* Linkage updates. */
331                 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
332
333                 /* Get the TIME_WAIT timeout firing. */
334                 if (timeo < rto)
335                         timeo = rto;
336
337                 if (recycle_ok) {
338                         tw->tw_timeout = rto;
339                 } else {
340                         tw->tw_timeout = TCP_TIMEWAIT_LEN;
341                         if (state == TCP_TIME_WAIT)
342                                 timeo = TCP_TIMEWAIT_LEN;
343                 }
344
345                 inet_twsk_schedule(tw, &tcp_death_row, timeo,
346                                    TCP_TIMEWAIT_LEN);
347                 inet_twsk_put(tw);
348         } else {
349                 /* Sorry, if we're out of memory, just CLOSE this
350                  * socket up.  We've got bigger problems than
351                  * non-graceful socket closings.
352                  */
353                 LIMIT_NETDEBUG(KERN_INFO "TCP: time wait bucket table overflow\n");
354         }
355
356         tcp_update_metrics(sk);
357         tcp_done(sk);
358 }
359
360 void tcp_twsk_destructor(struct sock *sk)
361 {
362 #ifdef CONFIG_TCP_MD5SIG
363         struct tcp_timewait_sock *twsk = tcp_twsk(sk);
364         if (twsk->tw_md5_keylen)
365                 tcp_put_md5sig_pool();
366 #endif
367 }
368
369 EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
370
371 /* This is not only more efficient than what we used to do, it eliminates
372  * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
373  *
374  * Actually, we could lots of memory writes here. tp of listening
375  * socket contains all necessary default parameters.
376  */
377 struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb)
378 {
379         struct sock *newsk = inet_csk_clone(sk, req, GFP_ATOMIC);
380
381         if (newsk != NULL) {
382                 const struct inet_request_sock *ireq = inet_rsk(req);
383                 struct tcp_request_sock *treq = tcp_rsk(req);
384                 struct inet_connection_sock *newicsk = inet_csk(newsk);
385                 struct tcp_sock *newtp;
386
387                 /* Now setup tcp_sock */
388                 newtp = tcp_sk(newsk);
389                 newtp->pred_flags = 0;
390                 newtp->rcv_wup = newtp->copied_seq = newtp->rcv_nxt = treq->rcv_isn + 1;
391                 newtp->snd_sml = newtp->snd_una = newtp->snd_nxt = treq->snt_isn + 1;
392
393                 tcp_prequeue_init(newtp);
394
395                 tcp_init_wl(newtp, treq->snt_isn, treq->rcv_isn);
396
397                 newtp->srtt = 0;
398                 newtp->mdev = TCP_TIMEOUT_INIT;
399                 newicsk->icsk_rto = TCP_TIMEOUT_INIT;
400
401                 newtp->packets_out = 0;
402                 newtp->left_out = 0;
403                 newtp->retrans_out = 0;
404                 newtp->sacked_out = 0;
405                 newtp->fackets_out = 0;
406                 newtp->snd_ssthresh = 0x7fffffff;
407
408                 /* So many TCP implementations out there (incorrectly) count the
409                  * initial SYN frame in their delayed-ACK and congestion control
410                  * algorithms that we must have the following bandaid to talk
411                  * efficiently to them.  -DaveM
412                  */
413                 newtp->snd_cwnd = 2;
414                 newtp->snd_cwnd_cnt = 0;
415                 newtp->bytes_acked = 0;
416
417                 newtp->frto_counter = 0;
418                 newtp->frto_highmark = 0;
419
420                 newicsk->icsk_ca_ops = &tcp_init_congestion_ops;
421
422                 tcp_set_ca_state(newsk, TCP_CA_Open);
423                 tcp_init_xmit_timers(newsk);
424                 skb_queue_head_init(&newtp->out_of_order_queue);
425                 newtp->write_seq = treq->snt_isn + 1;
426                 newtp->pushed_seq = newtp->write_seq;
427
428                 newtp->rx_opt.saw_tstamp = 0;
429
430                 newtp->rx_opt.dsack = 0;
431                 newtp->rx_opt.eff_sacks = 0;
432
433                 newtp->rx_opt.num_sacks = 0;
434                 newtp->urg_data = 0;
435
436                 if (sock_flag(newsk, SOCK_KEEPOPEN))
437                         inet_csk_reset_keepalive_timer(newsk,
438                                                        keepalive_time_when(newtp));
439
440                 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
441                 if((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
442                         if (sysctl_tcp_fack)
443                                 newtp->rx_opt.sack_ok |= 2;
444                 }
445                 newtp->window_clamp = req->window_clamp;
446                 newtp->rcv_ssthresh = req->rcv_wnd;
447                 newtp->rcv_wnd = req->rcv_wnd;
448                 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
449                 if (newtp->rx_opt.wscale_ok) {
450                         newtp->rx_opt.snd_wscale = ireq->snd_wscale;
451                         newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
452                 } else {
453                         newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
454                         newtp->window_clamp = min(newtp->window_clamp, 65535U);
455                 }
456                 newtp->snd_wnd = ntohs(skb->h.th->window) << newtp->rx_opt.snd_wscale;
457                 newtp->max_window = newtp->snd_wnd;
458
459                 if (newtp->rx_opt.tstamp_ok) {
460                         newtp->rx_opt.ts_recent = req->ts_recent;
461                         newtp->rx_opt.ts_recent_stamp = get_seconds();
462                         newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
463                 } else {
464                         newtp->rx_opt.ts_recent_stamp = 0;
465                         newtp->tcp_header_len = sizeof(struct tcphdr);
466                 }
467 #ifdef CONFIG_TCP_MD5SIG
468                 newtp->md5sig_info = NULL;      /*XXX*/
469                 if (newtp->af_specific->md5_lookup(sk, newsk))
470                         newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
471 #endif
472                 if (skb->len >= TCP_MIN_RCVMSS+newtp->tcp_header_len)
473                         newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
474                 newtp->rx_opt.mss_clamp = req->mss;
475                 TCP_ECN_openreq_child(newtp, req);
476
477                 TCP_INC_STATS_BH(TCP_MIB_PASSIVEOPENS);
478         }
479         return newsk;
480 }
481
482 /*
483  *      Process an incoming packet for SYN_RECV sockets represented
484  *      as a request_sock.
485  */
486
487 struct sock *tcp_check_req(struct sock *sk,struct sk_buff *skb,
488                            struct request_sock *req,
489                            struct request_sock **prev)
490 {
491         struct tcphdr *th = skb->h.th;
492         __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
493         int paws_reject = 0;
494         struct tcp_options_received tmp_opt;
495         struct sock *child;
496
497         tmp_opt.saw_tstamp = 0;
498         if (th->doff > (sizeof(struct tcphdr)>>2)) {
499                 tcp_parse_options(skb, &tmp_opt, 0);
500
501                 if (tmp_opt.saw_tstamp) {
502                         tmp_opt.ts_recent = req->ts_recent;
503                         /* We do not store true stamp, but it is not required,
504                          * it can be estimated (approximately)
505                          * from another data.
506                          */
507                         tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->retrans);
508                         paws_reject = tcp_paws_check(&tmp_opt, th->rst);
509                 }
510         }
511
512         /* Check for pure retransmitted SYN. */
513         if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
514             flg == TCP_FLAG_SYN &&
515             !paws_reject) {
516                 /*
517                  * RFC793 draws (Incorrectly! It was fixed in RFC1122)
518                  * this case on figure 6 and figure 8, but formal
519                  * protocol description says NOTHING.
520                  * To be more exact, it says that we should send ACK,
521                  * because this segment (at least, if it has no data)
522                  * is out of window.
523                  *
524                  *  CONCLUSION: RFC793 (even with RFC1122) DOES NOT
525                  *  describe SYN-RECV state. All the description
526                  *  is wrong, we cannot believe to it and should
527                  *  rely only on common sense and implementation
528                  *  experience.
529                  *
530                  * Enforce "SYN-ACK" according to figure 8, figure 6
531                  * of RFC793, fixed by RFC1122.
532                  */
533                 req->rsk_ops->rtx_syn_ack(sk, req, NULL);
534                 return NULL;
535         }
536
537         /* Further reproduces section "SEGMENT ARRIVES"
538            for state SYN-RECEIVED of RFC793.
539            It is broken, however, it does not work only
540            when SYNs are crossed.
541
542            You would think that SYN crossing is impossible here, since
543            we should have a SYN_SENT socket (from connect()) on our end,
544            but this is not true if the crossed SYNs were sent to both
545            ends by a malicious third party.  We must defend against this,
546            and to do that we first verify the ACK (as per RFC793, page
547            36) and reset if it is invalid.  Is this a true full defense?
548            To convince ourselves, let us consider a way in which the ACK
549            test can still pass in this 'malicious crossed SYNs' case.
550            Malicious sender sends identical SYNs (and thus identical sequence
551            numbers) to both A and B:
552
553                 A: gets SYN, seq=7
554                 B: gets SYN, seq=7
555
556            By our good fortune, both A and B select the same initial
557            send sequence number of seven :-)
558
559                 A: sends SYN|ACK, seq=7, ack_seq=8
560                 B: sends SYN|ACK, seq=7, ack_seq=8
561
562            So we are now A eating this SYN|ACK, ACK test passes.  So
563            does sequence test, SYN is truncated, and thus we consider
564            it a bare ACK.
565
566            If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
567            bare ACK.  Otherwise, we create an established connection.  Both
568            ends (listening sockets) accept the new incoming connection and try
569            to talk to each other. 8-)
570
571            Note: This case is both harmless, and rare.  Possibility is about the
572            same as us discovering intelligent life on another plant tomorrow.
573
574            But generally, we should (RFC lies!) to accept ACK
575            from SYNACK both here and in tcp_rcv_state_process().
576            tcp_rcv_state_process() does not, hence, we do not too.
577
578            Note that the case is absolutely generic:
579            we cannot optimize anything here without
580            violating protocol. All the checks must be made
581            before attempt to create socket.
582          */
583
584         /* RFC793 page 36: "If the connection is in any non-synchronized state ...
585          *                  and the incoming segment acknowledges something not yet
586          *                  sent (the segment carries an unacceptable ACK) ...
587          *                  a reset is sent."
588          *
589          * Invalid ACK: reset will be sent by listening socket
590          */
591         if ((flg & TCP_FLAG_ACK) &&
592             (TCP_SKB_CB(skb)->ack_seq != tcp_rsk(req)->snt_isn + 1))
593                 return sk;
594
595         /* Also, it would be not so bad idea to check rcv_tsecr, which
596          * is essentially ACK extension and too early or too late values
597          * should cause reset in unsynchronized states.
598          */
599
600         /* RFC793: "first check sequence number". */
601
602         if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
603                                           tcp_rsk(req)->rcv_isn + 1, tcp_rsk(req)->rcv_isn + 1 + req->rcv_wnd)) {
604                 /* Out of window: send ACK and drop. */
605                 if (!(flg & TCP_FLAG_RST))
606                         req->rsk_ops->send_ack(skb, req);
607                 if (paws_reject)
608                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
609                 return NULL;
610         }
611
612         /* In sequence, PAWS is OK. */
613
614         if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_isn + 1))
615                         req->ts_recent = tmp_opt.rcv_tsval;
616
617                 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
618                         /* Truncate SYN, it is out of window starting
619                            at tcp_rsk(req)->rcv_isn + 1. */
620                         flg &= ~TCP_FLAG_SYN;
621                 }
622
623                 /* RFC793: "second check the RST bit" and
624                  *         "fourth, check the SYN bit"
625                  */
626                 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
627                         TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
628                         goto embryonic_reset;
629                 }
630
631                 /* ACK sequence verified above, just make sure ACK is
632                  * set.  If ACK not set, just silently drop the packet.
633                  */
634                 if (!(flg & TCP_FLAG_ACK))
635                         return NULL;
636
637                 /* If TCP_DEFER_ACCEPT is set, drop bare ACK. */
638                 if (inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
639                     TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
640                         inet_rsk(req)->acked = 1;
641                         return NULL;
642                 }
643
644                 /* OK, ACK is valid, create big socket and
645                  * feed this segment to it. It will repeat all
646                  * the tests. THIS SEGMENT MUST MOVE SOCKET TO
647                  * ESTABLISHED STATE. If it will be dropped after
648                  * socket is created, wait for troubles.
649                  */
650                 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb,
651                                                                  req, NULL);
652                 if (child == NULL)
653                         goto listen_overflow;
654 #ifdef CONFIG_TCP_MD5SIG
655                 else {
656                         /* Copy over the MD5 key from the original socket */
657                         struct tcp_md5sig_key *key;
658                         struct tcp_sock *tp = tcp_sk(sk);
659                         key = tp->af_specific->md5_lookup(sk, child);
660                         if (key != NULL) {
661                                 /*
662                                  * We're using one, so create a matching key on the
663                                  * newsk structure. If we fail to get memory then we
664                                  * end up not copying the key across. Shucks.
665                                  */
666                                 char *newkey = kmemdup(key->key, key->keylen,
667                                                        GFP_ATOMIC);
668                                 if (newkey) {
669                                         if (!tcp_alloc_md5sig_pool())
670                                                 BUG();
671                                         tp->af_specific->md5_add(child, child,
672                                                                  newkey,
673                                                                  key->keylen);
674                                 }
675                         }
676                 }
677 #endif
678
679                 inet_csk_reqsk_queue_unlink(sk, req, prev);
680                 inet_csk_reqsk_queue_removed(sk, req);
681
682                 inet_csk_reqsk_queue_add(sk, req, child);
683                 return child;
684
685         listen_overflow:
686                 if (!sysctl_tcp_abort_on_overflow) {
687                         inet_rsk(req)->acked = 1;
688                         return NULL;
689                 }
690
691         embryonic_reset:
692                 NET_INC_STATS_BH(LINUX_MIB_EMBRYONICRSTS);
693                 if (!(flg & TCP_FLAG_RST))
694                         req->rsk_ops->send_reset(sk, skb);
695
696                 inet_csk_reqsk_queue_drop(sk, req, prev);
697                 return NULL;
698 }
699
700 /*
701  * Queue segment on the new socket if the new socket is active,
702  * otherwise we just shortcircuit this and continue with
703  * the new socket.
704  */
705
706 int tcp_child_process(struct sock *parent, struct sock *child,
707                       struct sk_buff *skb)
708 {
709         int ret = 0;
710         int state = child->sk_state;
711
712         if (!sock_owned_by_user(child)) {
713                 ret = tcp_rcv_state_process(child, skb, skb->h.th, skb->len);
714
715                 /* Wakeup parent, send SIGIO */
716                 if (state == TCP_SYN_RECV && child->sk_state != state)
717                         parent->sk_data_ready(parent, 0);
718         } else {
719                 /* Alas, it is possible again, because we do lookup
720                  * in main socket hash table and lock on listening
721                  * socket does not protect us more.
722                  */
723                 sk_add_backlog(child, skb);
724         }
725
726         bh_unlock_sock(child);
727         sock_put(child);
728         return ret;
729 }
730
731 EXPORT_SYMBOL(tcp_check_req);
732 EXPORT_SYMBOL(tcp_child_process);
733 EXPORT_SYMBOL(tcp_create_openreq_child);
734 EXPORT_SYMBOL(tcp_timewait_state_process);