Merge commit 'v2.6.26-rc8' into x86/mce
[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/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 EXPORT_SYMBOL(sysctl_tcp_syncookies);
39
40 int sysctl_tcp_abort_on_overflow __read_mostly;
41
42 struct inet_timewait_death_row tcp_death_row = {
43         .sysctl_max_tw_buckets = NR_FILE * 2,
44         .period         = TCP_TIMEWAIT_LEN / INET_TWDR_TWKILL_SLOTS,
45         .death_lock     = __SPIN_LOCK_UNLOCKED(tcp_death_row.death_lock),
46         .hashinfo       = &tcp_hashinfo,
47         .tw_timer       = TIMER_INITIALIZER(inet_twdr_hangman, 0,
48                                             (unsigned long)&tcp_death_row),
49         .twkill_work    = __WORK_INITIALIZER(tcp_death_row.twkill_work,
50                                              inet_twdr_twkill_work),
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 = get_seconds();
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 = get_seconds();
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 inet_connection_sock *icsk = inet_csk(sk);
278         const struct tcp_sock *tp = tcp_sk(sk);
279         int recycle_ok = 0;
280
281         if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp)
282                 recycle_ok = icsk->icsk_af_ops->remember_stamp(sk);
283
284         if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets)
285                 tw = inet_twsk_alloc(sk, state);
286
287         if (tw != NULL) {
288                 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
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 inet6_timewait_sock *tw6;
302
303                         tw->tw_ipv6_offset = inet6_tw_offset(sk->sk_prot);
304                         tw6 = inet6_twsk((struct sock *)tw);
305                         ipv6_addr_copy(&tw6->tw_v6_daddr, &np->daddr);
306                         ipv6_addr_copy(&tw6->tw_v6_rcv_saddr, &np->rcv_saddr);
307                         tw->tw_ipv6only = np->ipv6only;
308                 }
309 #endif
310
311 #ifdef CONFIG_TCP_MD5SIG
312                 /*
313                  * The timewait bucket does not have the key DB from the
314                  * sock structure. We just make a quick copy of the
315                  * md5 key being used (if indeed we are using one)
316                  * so the timewait ack generating code has the key.
317                  */
318                 do {
319                         struct tcp_md5sig_key *key;
320                         memset(tcptw->tw_md5_key, 0, sizeof(tcptw->tw_md5_key));
321                         tcptw->tw_md5_keylen = 0;
322                         key = tp->af_specific->md5_lookup(sk, sk);
323                         if (key != NULL) {
324                                 memcpy(&tcptw->tw_md5_key, key->key, key->keylen);
325                                 tcptw->tw_md5_keylen = key->keylen;
326                                 if (tcp_alloc_md5sig_pool() == NULL)
327                                         BUG();
328                         }
329                 } while (0);
330 #endif
331
332                 /* Linkage updates. */
333                 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
334
335                 /* Get the TIME_WAIT timeout firing. */
336                 if (timeo < rto)
337                         timeo = rto;
338
339                 if (recycle_ok) {
340                         tw->tw_timeout = rto;
341                 } else {
342                         tw->tw_timeout = TCP_TIMEWAIT_LEN;
343                         if (state == TCP_TIME_WAIT)
344                                 timeo = TCP_TIMEWAIT_LEN;
345                 }
346
347                 inet_twsk_schedule(tw, &tcp_death_row, timeo,
348                                    TCP_TIMEWAIT_LEN);
349                 inet_twsk_put(tw);
350         } else {
351                 /* Sorry, if we're out of memory, just CLOSE this
352                  * socket up.  We've got bigger problems than
353                  * non-graceful socket closings.
354                  */
355                 LIMIT_NETDEBUG(KERN_INFO "TCP: time wait bucket table overflow\n");
356         }
357
358         tcp_update_metrics(sk);
359         tcp_done(sk);
360 }
361
362 void tcp_twsk_destructor(struct sock *sk)
363 {
364 #ifdef CONFIG_TCP_MD5SIG
365         struct tcp_timewait_sock *twsk = tcp_twsk(sk);
366         if (twsk->tw_md5_keylen)
367                 tcp_put_md5sig_pool();
368 #endif
369 }
370
371 EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
372
373 static inline void TCP_ECN_openreq_child(struct tcp_sock *tp,
374                                          struct request_sock *req)
375 {
376         tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
377 }
378
379 /* This is not only more efficient than what we used to do, it eliminates
380  * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
381  *
382  * Actually, we could lots of memory writes here. tp of listening
383  * socket contains all necessary default parameters.
384  */
385 struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb)
386 {
387         struct sock *newsk = inet_csk_clone(sk, req, GFP_ATOMIC);
388
389         if (newsk != NULL) {
390                 const struct inet_request_sock *ireq = inet_rsk(req);
391                 struct tcp_request_sock *treq = tcp_rsk(req);
392                 struct inet_connection_sock *newicsk = inet_csk(newsk);
393                 struct tcp_sock *newtp;
394
395                 /* Now setup tcp_sock */
396                 newtp = tcp_sk(newsk);
397                 newtp->pred_flags = 0;
398                 newtp->rcv_wup = newtp->copied_seq = newtp->rcv_nxt = treq->rcv_isn + 1;
399                 newtp->snd_sml = newtp->snd_una = newtp->snd_nxt = treq->snt_isn + 1;
400
401                 tcp_prequeue_init(newtp);
402
403                 tcp_init_wl(newtp, treq->snt_isn, treq->rcv_isn);
404
405                 newtp->srtt = 0;
406                 newtp->mdev = TCP_TIMEOUT_INIT;
407                 newicsk->icsk_rto = TCP_TIMEOUT_INIT;
408
409                 newtp->packets_out = 0;
410                 newtp->retrans_out = 0;
411                 newtp->sacked_out = 0;
412                 newtp->fackets_out = 0;
413                 newtp->snd_ssthresh = 0x7fffffff;
414
415                 /* So many TCP implementations out there (incorrectly) count the
416                  * initial SYN frame in their delayed-ACK and congestion control
417                  * algorithms that we must have the following bandaid to talk
418                  * efficiently to them.  -DaveM
419                  */
420                 newtp->snd_cwnd = 2;
421                 newtp->snd_cwnd_cnt = 0;
422                 newtp->bytes_acked = 0;
423
424                 newtp->frto_counter = 0;
425                 newtp->frto_highmark = 0;
426
427                 newicsk->icsk_ca_ops = &tcp_init_congestion_ops;
428
429                 tcp_set_ca_state(newsk, TCP_CA_Open);
430                 tcp_init_xmit_timers(newsk);
431                 skb_queue_head_init(&newtp->out_of_order_queue);
432                 newtp->write_seq = treq->snt_isn + 1;
433                 newtp->pushed_seq = newtp->write_seq;
434
435                 newtp->rx_opt.saw_tstamp = 0;
436
437                 newtp->rx_opt.dsack = 0;
438                 newtp->rx_opt.eff_sacks = 0;
439
440                 newtp->rx_opt.num_sacks = 0;
441                 newtp->urg_data = 0;
442
443                 if (sock_flag(newsk, SOCK_KEEPOPEN))
444                         inet_csk_reset_keepalive_timer(newsk,
445                                                        keepalive_time_when(newtp));
446
447                 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
448                 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
449                         if (sysctl_tcp_fack)
450                                 tcp_enable_fack(newtp);
451                 }
452                 newtp->window_clamp = req->window_clamp;
453                 newtp->rcv_ssthresh = req->rcv_wnd;
454                 newtp->rcv_wnd = req->rcv_wnd;
455                 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
456                 if (newtp->rx_opt.wscale_ok) {
457                         newtp->rx_opt.snd_wscale = ireq->snd_wscale;
458                         newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
459                 } else {
460                         newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
461                         newtp->window_clamp = min(newtp->window_clamp, 65535U);
462                 }
463                 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) <<
464                                   newtp->rx_opt.snd_wscale);
465                 newtp->max_window = newtp->snd_wnd;
466
467                 if (newtp->rx_opt.tstamp_ok) {
468                         newtp->rx_opt.ts_recent = req->ts_recent;
469                         newtp->rx_opt.ts_recent_stamp = get_seconds();
470                         newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
471                 } else {
472                         newtp->rx_opt.ts_recent_stamp = 0;
473                         newtp->tcp_header_len = sizeof(struct tcphdr);
474                 }
475 #ifdef CONFIG_TCP_MD5SIG
476                 newtp->md5sig_info = NULL;      /*XXX*/
477                 if (newtp->af_specific->md5_lookup(sk, newsk))
478                         newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
479 #endif
480                 if (skb->len >= TCP_MIN_RCVMSS+newtp->tcp_header_len)
481                         newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
482                 newtp->rx_opt.mss_clamp = req->mss;
483                 TCP_ECN_openreq_child(newtp, req);
484
485                 TCP_INC_STATS_BH(TCP_MIB_PASSIVEOPENS);
486         }
487         return newsk;
488 }
489
490 /*
491  *      Process an incoming packet for SYN_RECV sockets represented
492  *      as a request_sock.
493  */
494
495 struct sock *tcp_check_req(struct sock *sk,struct sk_buff *skb,
496                            struct request_sock *req,
497                            struct request_sock **prev)
498 {
499         const struct tcphdr *th = tcp_hdr(skb);
500         __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
501         int paws_reject = 0;
502         struct tcp_options_received tmp_opt;
503         struct sock *child;
504
505         tmp_opt.saw_tstamp = 0;
506         if (th->doff > (sizeof(struct tcphdr)>>2)) {
507                 tcp_parse_options(skb, &tmp_opt, 0);
508
509                 if (tmp_opt.saw_tstamp) {
510                         tmp_opt.ts_recent = req->ts_recent;
511                         /* We do not store true stamp, but it is not required,
512                          * it can be estimated (approximately)
513                          * from another data.
514                          */
515                         tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->retrans);
516                         paws_reject = tcp_paws_check(&tmp_opt, th->rst);
517                 }
518         }
519
520         /* Check for pure retransmitted SYN. */
521         if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
522             flg == TCP_FLAG_SYN &&
523             !paws_reject) {
524                 /*
525                  * RFC793 draws (Incorrectly! It was fixed in RFC1122)
526                  * this case on figure 6 and figure 8, but formal
527                  * protocol description says NOTHING.
528                  * To be more exact, it says that we should send ACK,
529                  * because this segment (at least, if it has no data)
530                  * is out of window.
531                  *
532                  *  CONCLUSION: RFC793 (even with RFC1122) DOES NOT
533                  *  describe SYN-RECV state. All the description
534                  *  is wrong, we cannot believe to it and should
535                  *  rely only on common sense and implementation
536                  *  experience.
537                  *
538                  * Enforce "SYN-ACK" according to figure 8, figure 6
539                  * of RFC793, fixed by RFC1122.
540                  */
541                 req->rsk_ops->rtx_syn_ack(sk, req);
542                 return NULL;
543         }
544
545         /* Further reproduces section "SEGMENT ARRIVES"
546            for state SYN-RECEIVED of RFC793.
547            It is broken, however, it does not work only
548            when SYNs are crossed.
549
550            You would think that SYN crossing is impossible here, since
551            we should have a SYN_SENT socket (from connect()) on our end,
552            but this is not true if the crossed SYNs were sent to both
553            ends by a malicious third party.  We must defend against this,
554            and to do that we first verify the ACK (as per RFC793, page
555            36) and reset if it is invalid.  Is this a true full defense?
556            To convince ourselves, let us consider a way in which the ACK
557            test can still pass in this 'malicious crossed SYNs' case.
558            Malicious sender sends identical SYNs (and thus identical sequence
559            numbers) to both A and B:
560
561                 A: gets SYN, seq=7
562                 B: gets SYN, seq=7
563
564            By our good fortune, both A and B select the same initial
565            send sequence number of seven :-)
566
567                 A: sends SYN|ACK, seq=7, ack_seq=8
568                 B: sends SYN|ACK, seq=7, ack_seq=8
569
570            So we are now A eating this SYN|ACK, ACK test passes.  So
571            does sequence test, SYN is truncated, and thus we consider
572            it a bare ACK.
573
574            If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
575            bare ACK.  Otherwise, we create an established connection.  Both
576            ends (listening sockets) accept the new incoming connection and try
577            to talk to each other. 8-)
578
579            Note: This case is both harmless, and rare.  Possibility is about the
580            same as us discovering intelligent life on another plant tomorrow.
581
582            But generally, we should (RFC lies!) to accept ACK
583            from SYNACK both here and in tcp_rcv_state_process().
584            tcp_rcv_state_process() does not, hence, we do not too.
585
586            Note that the case is absolutely generic:
587            we cannot optimize anything here without
588            violating protocol. All the checks must be made
589            before attempt to create socket.
590          */
591
592         /* RFC793 page 36: "If the connection is in any non-synchronized state ...
593          *                  and the incoming segment acknowledges something not yet
594          *                  sent (the segment carries an unacceptable ACK) ...
595          *                  a reset is sent."
596          *
597          * Invalid ACK: reset will be sent by listening socket
598          */
599         if ((flg & TCP_FLAG_ACK) &&
600             (TCP_SKB_CB(skb)->ack_seq != tcp_rsk(req)->snt_isn + 1))
601                 return sk;
602
603         /* Also, it would be not so bad idea to check rcv_tsecr, which
604          * is essentially ACK extension and too early or too late values
605          * should cause reset in unsynchronized states.
606          */
607
608         /* RFC793: "first check sequence number". */
609
610         if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
611                                           tcp_rsk(req)->rcv_isn + 1, tcp_rsk(req)->rcv_isn + 1 + req->rcv_wnd)) {
612                 /* Out of window: send ACK and drop. */
613                 if (!(flg & TCP_FLAG_RST))
614                         req->rsk_ops->send_ack(skb, req);
615                 if (paws_reject)
616                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
617                 return NULL;
618         }
619
620         /* In sequence, PAWS is OK. */
621
622         if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_isn + 1))
623                         req->ts_recent = tmp_opt.rcv_tsval;
624
625                 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
626                         /* Truncate SYN, it is out of window starting
627                            at tcp_rsk(req)->rcv_isn + 1. */
628                         flg &= ~TCP_FLAG_SYN;
629                 }
630
631                 /* RFC793: "second check the RST bit" and
632                  *         "fourth, check the SYN bit"
633                  */
634                 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
635                         TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
636                         goto embryonic_reset;
637                 }
638
639                 /* ACK sequence verified above, just make sure ACK is
640                  * set.  If ACK not set, just silently drop the packet.
641                  */
642                 if (!(flg & TCP_FLAG_ACK))
643                         return NULL;
644
645                 /* If TCP_DEFER_ACCEPT is set, drop bare ACK. */
646                 if (inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
647                     TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
648                         inet_rsk(req)->acked = 1;
649                         return NULL;
650                 }
651
652                 /* OK, ACK is valid, create big socket and
653                  * feed this segment to it. It will repeat all
654                  * the tests. THIS SEGMENT MUST MOVE SOCKET TO
655                  * ESTABLISHED STATE. If it will be dropped after
656                  * socket is created, wait for troubles.
657                  */
658                 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb,
659                                                                  req, NULL);
660                 if (child == NULL)
661                         goto listen_overflow;
662 #ifdef CONFIG_TCP_MD5SIG
663                 else {
664                         /* Copy over the MD5 key from the original socket */
665                         struct tcp_md5sig_key *key;
666                         struct tcp_sock *tp = tcp_sk(sk);
667                         key = tp->af_specific->md5_lookup(sk, child);
668                         if (key != NULL) {
669                                 /*
670                                  * We're using one, so create a matching key on the
671                                  * newsk structure. If we fail to get memory then we
672                                  * end up not copying the key across. Shucks.
673                                  */
674                                 char *newkey = kmemdup(key->key, key->keylen,
675                                                        GFP_ATOMIC);
676                                 if (newkey) {
677                                         if (!tcp_alloc_md5sig_pool())
678                                                 BUG();
679                                         tp->af_specific->md5_add(child, child,
680                                                                  newkey,
681                                                                  key->keylen);
682                                 }
683                         }
684                 }
685 #endif
686
687                 inet_csk_reqsk_queue_unlink(sk, req, prev);
688                 inet_csk_reqsk_queue_removed(sk, req);
689
690                 inet_csk_reqsk_queue_add(sk, req, child);
691                 return child;
692
693         listen_overflow:
694                 if (!sysctl_tcp_abort_on_overflow) {
695                         inet_rsk(req)->acked = 1;
696                         return NULL;
697                 }
698
699         embryonic_reset:
700                 NET_INC_STATS_BH(LINUX_MIB_EMBRYONICRSTS);
701                 if (!(flg & TCP_FLAG_RST))
702                         req->rsk_ops->send_reset(sk, skb);
703
704                 inet_csk_reqsk_queue_drop(sk, req, prev);
705                 return NULL;
706 }
707
708 /*
709  * Queue segment on the new socket if the new socket is active,
710  * otherwise we just shortcircuit this and continue with
711  * the new socket.
712  */
713
714 int tcp_child_process(struct sock *parent, struct sock *child,
715                       struct sk_buff *skb)
716 {
717         int ret = 0;
718         int state = child->sk_state;
719
720         if (!sock_owned_by_user(child)) {
721                 ret = tcp_rcv_state_process(child, skb, tcp_hdr(skb),
722                                             skb->len);
723                 /* Wakeup parent, send SIGIO */
724                 if (state == TCP_SYN_RECV && child->sk_state != state)
725                         parent->sk_data_ready(parent, 0);
726         } else {
727                 /* Alas, it is possible again, because we do lookup
728                  * in main socket hash table and lock on listening
729                  * socket does not protect us more.
730                  */
731                 sk_add_backlog(child, skb);
732         }
733
734         bh_unlock_sock(child);
735         sock_put(child);
736         return ret;
737 }
738
739 EXPORT_SYMBOL(tcp_check_req);
740 EXPORT_SYMBOL(tcp_child_process);
741 EXPORT_SYMBOL(tcp_create_openreq_child);
742 EXPORT_SYMBOL(tcp_timewait_state_process);