1df6cd46066bb05b40e504d1d979019773ffb5d4
[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/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_tw_recycle;
39 int sysctl_tcp_max_tw_buckets = NR_FILE*2;
40
41 int sysctl_tcp_syncookies = SYNC_INIT; 
42 int sysctl_tcp_abort_on_overflow;
43
44 static void tcp_tw_schedule(struct tcp_tw_bucket *tw, int timeo);
45
46 static __inline__ int tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
47 {
48         if (seq == s_win)
49                 return 1;
50         if (after(end_seq, s_win) && before(seq, e_win))
51                 return 1;
52         return (seq == e_win && seq == end_seq);
53 }
54
55 /* New-style handling of TIME_WAIT sockets. */
56
57 int tcp_tw_count;
58
59
60 /* Must be called with locally disabled BHs. */
61 static void tcp_timewait_kill(struct tcp_tw_bucket *tw)
62 {
63         struct inet_ehash_bucket *ehead;
64         struct inet_bind_hashbucket *bhead;
65         struct inet_bind_bucket *tb;
66
67         /* Unlink from established hashes. */
68         ehead = &tcp_ehash[tw->tw_hashent];
69         write_lock(&ehead->lock);
70         if (hlist_unhashed(&tw->tw_node)) {
71                 write_unlock(&ehead->lock);
72                 return;
73         }
74         __hlist_del(&tw->tw_node);
75         sk_node_init(&tw->tw_node);
76         write_unlock(&ehead->lock);
77
78         /* Disassociate with bind bucket. */
79         bhead = &tcp_bhash[inet_bhashfn(tw->tw_num, tcp_bhash_size)];
80         spin_lock(&bhead->lock);
81         tb = tw->tw_tb;
82         __hlist_del(&tw->tw_bind_node);
83         tw->tw_tb = NULL;
84         inet_bind_bucket_destroy(tcp_bucket_cachep, tb);
85         spin_unlock(&bhead->lock);
86
87 #ifdef SOCK_REFCNT_DEBUG
88         if (atomic_read(&tw->tw_refcnt) != 1) {
89                 printk(KERN_DEBUG "tw_bucket %p refcnt=%d\n", tw,
90                        atomic_read(&tw->tw_refcnt));
91         }
92 #endif
93         tcp_tw_put(tw);
94 }
95
96 /* 
97  * * Main purpose of TIME-WAIT state is to close connection gracefully,
98  *   when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
99  *   (and, probably, tail of data) and one or more our ACKs are lost.
100  * * What is TIME-WAIT timeout? It is associated with maximal packet
101  *   lifetime in the internet, which results in wrong conclusion, that
102  *   it is set to catch "old duplicate segments" wandering out of their path.
103  *   It is not quite correct. This timeout is calculated so that it exceeds
104  *   maximal retransmission timeout enough to allow to lose one (or more)
105  *   segments sent by peer and our ACKs. This time may be calculated from RTO.
106  * * When TIME-WAIT socket receives RST, it means that another end
107  *   finally closed and we are allowed to kill TIME-WAIT too.
108  * * Second purpose of TIME-WAIT is catching old duplicate segments.
109  *   Well, certainly it is pure paranoia, but if we load TIME-WAIT
110  *   with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
111  * * If we invented some more clever way to catch duplicates
112  *   (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
113  *
114  * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
115  * When you compare it to RFCs, please, read section SEGMENT ARRIVES
116  * from the very beginning.
117  *
118  * NOTE. With recycling (and later with fin-wait-2) TW bucket
119  * is _not_ stateless. It means, that strictly speaking we must
120  * spinlock it. I do not want! Well, probability of misbehaviour
121  * is ridiculously low and, seems, we could use some mb() tricks
122  * to avoid misread sequence numbers, states etc.  --ANK
123  */
124 enum tcp_tw_status
125 tcp_timewait_state_process(struct tcp_tw_bucket *tw, struct sk_buff *skb,
126                            struct tcphdr *th, unsigned len)
127 {
128         struct tcp_options_received tmp_opt;
129         int paws_reject = 0;
130
131         tmp_opt.saw_tstamp = 0;
132         if (th->doff > (sizeof(struct tcphdr) >> 2) && tw->tw_ts_recent_stamp) {
133                 tcp_parse_options(skb, &tmp_opt, 0);
134
135                 if (tmp_opt.saw_tstamp) {
136                         tmp_opt.ts_recent          = tw->tw_ts_recent;
137                         tmp_opt.ts_recent_stamp = tw->tw_ts_recent_stamp;
138                         paws_reject = tcp_paws_check(&tmp_opt, th->rst);
139                 }
140         }
141
142         if (tw->tw_substate == TCP_FIN_WAIT2) {
143                 /* Just repeat all the checks of tcp_rcv_state_process() */
144
145                 /* Out of window, send ACK */
146                 if (paws_reject ||
147                     !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
148                                    tw->tw_rcv_nxt,
149                                    tw->tw_rcv_nxt + tw->tw_rcv_wnd))
150                         return TCP_TW_ACK;
151
152                 if (th->rst)
153                         goto kill;
154
155                 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tw->tw_rcv_nxt))
156                         goto kill_with_rst;
157
158                 /* Dup ACK? */
159                 if (!after(TCP_SKB_CB(skb)->end_seq, tw->tw_rcv_nxt) ||
160                     TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
161                         tcp_tw_put(tw);
162                         return TCP_TW_SUCCESS;
163                 }
164
165                 /* New data or FIN. If new data arrive after half-duplex close,
166                  * reset.
167                  */
168                 if (!th->fin ||
169                     TCP_SKB_CB(skb)->end_seq != tw->tw_rcv_nxt + 1) {
170 kill_with_rst:
171                         tcp_tw_deschedule(tw);
172                         tcp_tw_put(tw);
173                         return TCP_TW_RST;
174                 }
175
176                 /* FIN arrived, enter true time-wait state. */
177                 tw->tw_substate = TCP_TIME_WAIT;
178                 tw->tw_rcv_nxt  = TCP_SKB_CB(skb)->end_seq;
179                 if (tmp_opt.saw_tstamp) {
180                         tw->tw_ts_recent_stamp  = xtime.tv_sec;
181                         tw->tw_ts_recent        = tmp_opt.rcv_tsval;
182                 }
183
184                 /* I am shamed, but failed to make it more elegant.
185                  * Yes, it is direct reference to IP, which is impossible
186                  * to generalize to IPv6. Taking into account that IPv6
187                  * do not undertsnad recycling in any case, it not
188                  * a big problem in practice. --ANK */
189                 if (tw->tw_family == AF_INET &&
190                     sysctl_tcp_tw_recycle && tw->tw_ts_recent_stamp &&
191                     tcp_v4_tw_remember_stamp(tw))
192                         tcp_tw_schedule(tw, tw->tw_timeout);
193                 else
194                         tcp_tw_schedule(tw, TCP_TIMEWAIT_LEN);
195                 return TCP_TW_ACK;
196         }
197
198         /*
199          *      Now real TIME-WAIT state.
200          *
201          *      RFC 1122:
202          *      "When a connection is [...] on TIME-WAIT state [...]
203          *      [a TCP] MAY accept a new SYN from the remote TCP to
204          *      reopen the connection directly, if it:
205          *      
206          *      (1)  assigns its initial sequence number for the new
207          *      connection to be larger than the largest sequence
208          *      number it used on the previous connection incarnation,
209          *      and
210          *
211          *      (2)  returns to TIME-WAIT state if the SYN turns out 
212          *      to be an old duplicate".
213          */
214
215         if (!paws_reject &&
216             (TCP_SKB_CB(skb)->seq == tw->tw_rcv_nxt &&
217              (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
218                 /* In window segment, it may be only reset or bare ack. */
219
220                 if (th->rst) {
221                         /* This is TIME_WAIT assasination, in two flavors.
222                          * Oh well... nobody has a sufficient solution to this
223                          * protocol bug yet.
224                          */
225                         if (sysctl_tcp_rfc1337 == 0) {
226 kill:
227                                 tcp_tw_deschedule(tw);
228                                 tcp_tw_put(tw);
229                                 return TCP_TW_SUCCESS;
230                         }
231                 }
232                 tcp_tw_schedule(tw, TCP_TIMEWAIT_LEN);
233
234                 if (tmp_opt.saw_tstamp) {
235                         tw->tw_ts_recent        = tmp_opt.rcv_tsval;
236                         tw->tw_ts_recent_stamp  = xtime.tv_sec;
237                 }
238
239                 tcp_tw_put(tw);
240                 return TCP_TW_SUCCESS;
241         }
242
243         /* Out of window segment.
244
245            All the segments are ACKed immediately.
246
247            The only exception is new SYN. We accept it, if it is
248            not old duplicate and we are not in danger to be killed
249            by delayed old duplicates. RFC check is that it has
250            newer sequence number works at rates <40Mbit/sec.
251            However, if paws works, it is reliable AND even more,
252            we even may relax silly seq space cutoff.
253
254            RED-PEN: we violate main RFC requirement, if this SYN will appear
255            old duplicate (i.e. we receive RST in reply to SYN-ACK),
256            we must return socket to time-wait state. It is not good,
257            but not fatal yet.
258          */
259
260         if (th->syn && !th->rst && !th->ack && !paws_reject &&
261             (after(TCP_SKB_CB(skb)->seq, tw->tw_rcv_nxt) ||
262              (tmp_opt.saw_tstamp && (s32)(tw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
263                 u32 isn = tw->tw_snd_nxt + 65535 + 2;
264                 if (isn == 0)
265                         isn++;
266                 TCP_SKB_CB(skb)->when = isn;
267                 return TCP_TW_SYN;
268         }
269
270         if (paws_reject)
271                 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
272
273         if(!th->rst) {
274                 /* In this case we must reset the TIMEWAIT timer.
275                  *
276                  * If it is ACKless SYN it may be both old duplicate
277                  * and new good SYN with random sequence number <rcv_nxt.
278                  * Do not reschedule in the last case.
279                  */
280                 if (paws_reject || th->ack)
281                         tcp_tw_schedule(tw, TCP_TIMEWAIT_LEN);
282
283                 /* Send ACK. Note, we do not put the bucket,
284                  * it will be released by caller.
285                  */
286                 return TCP_TW_ACK;
287         }
288         tcp_tw_put(tw);
289         return TCP_TW_SUCCESS;
290 }
291
292 /* Enter the time wait state.  This is called with locally disabled BH.
293  * Essentially we whip up a timewait bucket, copy the
294  * relevant info into it from the SK, and mess with hash chains
295  * and list linkage.
296  */
297 static void __tcp_tw_hashdance(struct sock *sk, struct tcp_tw_bucket *tw)
298 {
299         struct inet_ehash_bucket *ehead = &tcp_ehash[sk->sk_hashent];
300         struct inet_bind_hashbucket *bhead;
301
302         /* Step 1: Put TW into bind hash. Original socket stays there too.
303            Note, that any socket with inet_sk(sk)->num != 0 MUST be bound in
304            binding cache, even if it is closed.
305          */
306         bhead = &tcp_bhash[inet_bhashfn(inet_sk(sk)->num, tcp_bhash_size)];
307         spin_lock(&bhead->lock);
308         tw->tw_tb = tcp_sk(sk)->bind_hash;
309         BUG_TRAP(tcp_sk(sk)->bind_hash);
310         tw_add_bind_node(tw, &tw->tw_tb->owners);
311         spin_unlock(&bhead->lock);
312
313         write_lock(&ehead->lock);
314
315         /* Step 2: Remove SK from established hash. */
316         if (__sk_del_node_init(sk))
317                 sock_prot_dec_use(sk->sk_prot);
318
319         /* Step 3: Hash TW into TIMEWAIT half of established hash table. */
320         tw_add_node(tw, &(ehead + tcp_ehash_size)->chain);
321         atomic_inc(&tw->tw_refcnt);
322
323         write_unlock(&ehead->lock);
324 }
325
326 /* 
327  * Move a socket to time-wait or dead fin-wait-2 state.
328  */ 
329 void tcp_time_wait(struct sock *sk, int state, int timeo)
330 {
331         struct tcp_tw_bucket *tw = NULL;
332         struct tcp_sock *tp = tcp_sk(sk);
333         int recycle_ok = 0;
334
335         if (sysctl_tcp_tw_recycle && tp->rx_opt.ts_recent_stamp)
336                 recycle_ok = tp->af_specific->remember_stamp(sk);
337
338         if (tcp_tw_count < sysctl_tcp_max_tw_buckets)
339                 tw = kmem_cache_alloc(tcp_timewait_cachep, SLAB_ATOMIC);
340
341         if(tw != NULL) {
342                 struct inet_sock *inet = inet_sk(sk);
343                 int rto = (tp->rto<<2) - (tp->rto>>1);
344
345                 /* Give us an identity. */
346                 tw->tw_daddr            = inet->daddr;
347                 tw->tw_rcv_saddr        = inet->rcv_saddr;
348                 tw->tw_bound_dev_if     = sk->sk_bound_dev_if;
349                 tw->tw_num              = inet->num;
350                 tw->tw_state            = TCP_TIME_WAIT;
351                 tw->tw_substate         = state;
352                 tw->tw_sport            = inet->sport;
353                 tw->tw_dport            = inet->dport;
354                 tw->tw_family           = sk->sk_family;
355                 tw->tw_reuse            = sk->sk_reuse;
356                 tw->tw_rcv_wscale       = tp->rx_opt.rcv_wscale;
357                 atomic_set(&tw->tw_refcnt, 1);
358
359                 tw->tw_hashent          = sk->sk_hashent;
360                 tw->tw_rcv_nxt          = tp->rcv_nxt;
361                 tw->tw_snd_nxt          = tp->snd_nxt;
362                 tw->tw_rcv_wnd          = tcp_receive_window(tp);
363                 tw->tw_ts_recent        = tp->rx_opt.ts_recent;
364                 tw->tw_ts_recent_stamp  = tp->rx_opt.ts_recent_stamp;
365                 tw_dead_node_init(tw);
366
367 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
368                 if (tw->tw_family == PF_INET6) {
369                         struct ipv6_pinfo *np = inet6_sk(sk);
370
371                         ipv6_addr_copy(&tw->tw_v6_daddr, &np->daddr);
372                         ipv6_addr_copy(&tw->tw_v6_rcv_saddr, &np->rcv_saddr);
373                         tw->tw_v6_ipv6only = np->ipv6only;
374                 } else {
375                         memset(&tw->tw_v6_daddr, 0, sizeof(tw->tw_v6_daddr));
376                         memset(&tw->tw_v6_rcv_saddr, 0, sizeof(tw->tw_v6_rcv_saddr));
377                         tw->tw_v6_ipv6only = 0;
378                 }
379 #endif
380                 /* Linkage updates. */
381                 __tcp_tw_hashdance(sk, tw);
382
383                 /* Get the TIME_WAIT timeout firing. */
384                 if (timeo < rto)
385                         timeo = rto;
386
387                 if (recycle_ok) {
388                         tw->tw_timeout = rto;
389                 } else {
390                         tw->tw_timeout = TCP_TIMEWAIT_LEN;
391                         if (state == TCP_TIME_WAIT)
392                                 timeo = TCP_TIMEWAIT_LEN;
393                 }
394
395                 tcp_tw_schedule(tw, timeo);
396                 tcp_tw_put(tw);
397         } else {
398                 /* Sorry, if we're out of memory, just CLOSE this
399                  * socket up.  We've got bigger problems than
400                  * non-graceful socket closings.
401                  */
402                 if (net_ratelimit())
403                         printk(KERN_INFO "TCP: time wait bucket table overflow\n");
404         }
405
406         tcp_update_metrics(sk);
407         tcp_done(sk);
408 }
409
410 /* Kill off TIME_WAIT sockets once their lifetime has expired. */
411 static int tcp_tw_death_row_slot;
412
413 static void tcp_twkill(unsigned long);
414
415 /* TIME_WAIT reaping mechanism. */
416 #define TCP_TWKILL_SLOTS        8       /* Please keep this a power of 2. */
417 #define TCP_TWKILL_PERIOD       (TCP_TIMEWAIT_LEN/TCP_TWKILL_SLOTS)
418
419 #define TCP_TWKILL_QUOTA        100
420
421 static struct hlist_head tcp_tw_death_row[TCP_TWKILL_SLOTS];
422 static DEFINE_SPINLOCK(tw_death_lock);
423 static struct timer_list tcp_tw_timer = TIMER_INITIALIZER(tcp_twkill, 0, 0);
424 static void twkill_work(void *);
425 static DECLARE_WORK(tcp_twkill_work, twkill_work, NULL);
426 static u32 twkill_thread_slots;
427
428 /* Returns non-zero if quota exceeded.  */
429 static int tcp_do_twkill_work(int slot, unsigned int quota)
430 {
431         struct tcp_tw_bucket *tw;
432         struct hlist_node *node;
433         unsigned int killed;
434         int ret;
435
436         /* NOTE: compare this to previous version where lock
437          * was released after detaching chain. It was racy,
438          * because tw buckets are scheduled in not serialized context
439          * in 2.3 (with netfilter), and with softnet it is common, because
440          * soft irqs are not sequenced.
441          */
442         killed = 0;
443         ret = 0;
444 rescan:
445         tw_for_each_inmate(tw, node, &tcp_tw_death_row[slot]) {
446                 __tw_del_dead_node(tw);
447                 spin_unlock(&tw_death_lock);
448                 tcp_timewait_kill(tw);
449                 tcp_tw_put(tw);
450                 killed++;
451                 spin_lock(&tw_death_lock);
452                 if (killed > quota) {
453                         ret = 1;
454                         break;
455                 }
456
457                 /* While we dropped tw_death_lock, another cpu may have
458                  * killed off the next TW bucket in the list, therefore
459                  * do a fresh re-read of the hlist head node with the
460                  * lock reacquired.  We still use the hlist traversal
461                  * macro in order to get the prefetches.
462                  */
463                 goto rescan;
464         }
465
466         tcp_tw_count -= killed;
467         NET_ADD_STATS_BH(LINUX_MIB_TIMEWAITED, killed);
468
469         return ret;
470 }
471
472 static void tcp_twkill(unsigned long dummy)
473 {
474         int need_timer, ret;
475
476         spin_lock(&tw_death_lock);
477
478         if (tcp_tw_count == 0)
479                 goto out;
480
481         need_timer = 0;
482         ret = tcp_do_twkill_work(tcp_tw_death_row_slot, TCP_TWKILL_QUOTA);
483         if (ret) {
484                 twkill_thread_slots |= (1 << tcp_tw_death_row_slot);
485                 mb();
486                 schedule_work(&tcp_twkill_work);
487                 need_timer = 1;
488         } else {
489                 /* We purged the entire slot, anything left?  */
490                 if (tcp_tw_count)
491                         need_timer = 1;
492         }
493         tcp_tw_death_row_slot =
494                 ((tcp_tw_death_row_slot + 1) & (TCP_TWKILL_SLOTS - 1));
495         if (need_timer)
496                 mod_timer(&tcp_tw_timer, jiffies + TCP_TWKILL_PERIOD);
497 out:
498         spin_unlock(&tw_death_lock);
499 }
500
501 extern void twkill_slots_invalid(void);
502
503 static void twkill_work(void *dummy)
504 {
505         int i;
506
507         if ((TCP_TWKILL_SLOTS - 1) > (sizeof(twkill_thread_slots) * 8))
508                 twkill_slots_invalid();
509
510         while (twkill_thread_slots) {
511                 spin_lock_bh(&tw_death_lock);
512                 for (i = 0; i < TCP_TWKILL_SLOTS; i++) {
513                         if (!(twkill_thread_slots & (1 << i)))
514                                 continue;
515
516                         while (tcp_do_twkill_work(i, TCP_TWKILL_QUOTA) != 0) {
517                                 if (need_resched()) {
518                                         spin_unlock_bh(&tw_death_lock);
519                                         schedule();
520                                         spin_lock_bh(&tw_death_lock);
521                                 }
522                         }
523
524                         twkill_thread_slots &= ~(1 << i);
525                 }
526                 spin_unlock_bh(&tw_death_lock);
527         }
528 }
529
530 /* These are always called from BH context.  See callers in
531  * tcp_input.c to verify this.
532  */
533
534 /* This is for handling early-kills of TIME_WAIT sockets. */
535 void tcp_tw_deschedule(struct tcp_tw_bucket *tw)
536 {
537         spin_lock(&tw_death_lock);
538         if (tw_del_dead_node(tw)) {
539                 tcp_tw_put(tw);
540                 if (--tcp_tw_count == 0)
541                         del_timer(&tcp_tw_timer);
542         }
543         spin_unlock(&tw_death_lock);
544         tcp_timewait_kill(tw);
545 }
546
547 /* Short-time timewait calendar */
548
549 static int tcp_twcal_hand = -1;
550 static int tcp_twcal_jiffie;
551 static void tcp_twcal_tick(unsigned long);
552 static struct timer_list tcp_twcal_timer =
553                 TIMER_INITIALIZER(tcp_twcal_tick, 0, 0);
554 static struct hlist_head tcp_twcal_row[TCP_TW_RECYCLE_SLOTS];
555
556 static void tcp_tw_schedule(struct tcp_tw_bucket *tw, int timeo)
557 {
558         struct hlist_head *list;
559         int slot;
560
561         /* timeout := RTO * 3.5
562          *
563          * 3.5 = 1+2+0.5 to wait for two retransmits.
564          *
565          * RATIONALE: if FIN arrived and we entered TIME-WAIT state,
566          * our ACK acking that FIN can be lost. If N subsequent retransmitted
567          * FINs (or previous seqments) are lost (probability of such event
568          * is p^(N+1), where p is probability to lose single packet and
569          * time to detect the loss is about RTO*(2^N - 1) with exponential
570          * backoff). Normal timewait length is calculated so, that we
571          * waited at least for one retransmitted FIN (maximal RTO is 120sec).
572          * [ BTW Linux. following BSD, violates this requirement waiting
573          *   only for 60sec, we should wait at least for 240 secs.
574          *   Well, 240 consumes too much of resources 8)
575          * ]
576          * This interval is not reduced to catch old duplicate and
577          * responces to our wandering segments living for two MSLs.
578          * However, if we use PAWS to detect
579          * old duplicates, we can reduce the interval to bounds required
580          * by RTO, rather than MSL. So, if peer understands PAWS, we
581          * kill tw bucket after 3.5*RTO (it is important that this number
582          * is greater than TS tick!) and detect old duplicates with help
583          * of PAWS.
584          */
585         slot = (timeo + (1<<TCP_TW_RECYCLE_TICK) - 1) >> TCP_TW_RECYCLE_TICK;
586
587         spin_lock(&tw_death_lock);
588
589         /* Unlink it, if it was scheduled */
590         if (tw_del_dead_node(tw))
591                 tcp_tw_count--;
592         else
593                 atomic_inc(&tw->tw_refcnt);
594
595         if (slot >= TCP_TW_RECYCLE_SLOTS) {
596                 /* Schedule to slow timer */
597                 if (timeo >= TCP_TIMEWAIT_LEN) {
598                         slot = TCP_TWKILL_SLOTS-1;
599                 } else {
600                         slot = (timeo + TCP_TWKILL_PERIOD-1) / TCP_TWKILL_PERIOD;
601                         if (slot >= TCP_TWKILL_SLOTS)
602                                 slot = TCP_TWKILL_SLOTS-1;
603                 }
604                 tw->tw_ttd = jiffies + timeo;
605                 slot = (tcp_tw_death_row_slot + slot) & (TCP_TWKILL_SLOTS - 1);
606                 list = &tcp_tw_death_row[slot];
607         } else {
608                 tw->tw_ttd = jiffies + (slot << TCP_TW_RECYCLE_TICK);
609
610                 if (tcp_twcal_hand < 0) {
611                         tcp_twcal_hand = 0;
612                         tcp_twcal_jiffie = jiffies;
613                         tcp_twcal_timer.expires = tcp_twcal_jiffie + (slot<<TCP_TW_RECYCLE_TICK);
614                         add_timer(&tcp_twcal_timer);
615                 } else {
616                         if (time_after(tcp_twcal_timer.expires, jiffies + (slot<<TCP_TW_RECYCLE_TICK)))
617                                 mod_timer(&tcp_twcal_timer, jiffies + (slot<<TCP_TW_RECYCLE_TICK));
618                         slot = (tcp_twcal_hand + slot)&(TCP_TW_RECYCLE_SLOTS-1);
619                 }
620                 list = &tcp_twcal_row[slot];
621         }
622
623         hlist_add_head(&tw->tw_death_node, list);
624
625         if (tcp_tw_count++ == 0)
626                 mod_timer(&tcp_tw_timer, jiffies+TCP_TWKILL_PERIOD);
627         spin_unlock(&tw_death_lock);
628 }
629
630 void tcp_twcal_tick(unsigned long dummy)
631 {
632         int n, slot;
633         unsigned long j;
634         unsigned long now = jiffies;
635         int killed = 0;
636         int adv = 0;
637
638         spin_lock(&tw_death_lock);
639         if (tcp_twcal_hand < 0)
640                 goto out;
641
642         slot = tcp_twcal_hand;
643         j = tcp_twcal_jiffie;
644
645         for (n=0; n<TCP_TW_RECYCLE_SLOTS; n++) {
646                 if (time_before_eq(j, now)) {
647                         struct hlist_node *node, *safe;
648                         struct tcp_tw_bucket *tw;
649
650                         tw_for_each_inmate_safe(tw, node, safe,
651                                            &tcp_twcal_row[slot]) {
652                                 __tw_del_dead_node(tw);
653                                 tcp_timewait_kill(tw);
654                                 tcp_tw_put(tw);
655                                 killed++;
656                         }
657                 } else {
658                         if (!adv) {
659                                 adv = 1;
660                                 tcp_twcal_jiffie = j;
661                                 tcp_twcal_hand = slot;
662                         }
663
664                         if (!hlist_empty(&tcp_twcal_row[slot])) {
665                                 mod_timer(&tcp_twcal_timer, j);
666                                 goto out;
667                         }
668                 }
669                 j += (1<<TCP_TW_RECYCLE_TICK);
670                 slot = (slot+1)&(TCP_TW_RECYCLE_SLOTS-1);
671         }
672         tcp_twcal_hand = -1;
673
674 out:
675         if ((tcp_tw_count -= killed) == 0)
676                 del_timer(&tcp_tw_timer);
677         NET_ADD_STATS_BH(LINUX_MIB_TIMEWAITKILLED, killed);
678         spin_unlock(&tw_death_lock);
679 }
680
681 /* This is not only more efficient than what we used to do, it eliminates
682  * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
683  *
684  * Actually, we could lots of memory writes here. tp of listening
685  * socket contains all necessary default parameters.
686  */
687 struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb)
688 {
689         /* allocate the newsk from the same slab of the master sock,
690          * if not, at sk_free time we'll try to free it from the wrong
691          * slabcache (i.e. is it TCPv4 or v6?), this is handled thru sk->sk_prot -acme */
692         struct sock *newsk = sk_alloc(PF_INET, GFP_ATOMIC, sk->sk_prot, 0);
693
694         if(newsk != NULL) {
695                 struct inet_request_sock *ireq = inet_rsk(req);
696                 struct tcp_request_sock *treq = tcp_rsk(req);
697                 struct tcp_sock *newtp;
698                 struct sk_filter *filter;
699
700                 memcpy(newsk, sk, sizeof(struct tcp_sock));
701                 newsk->sk_state = TCP_SYN_RECV;
702
703                 /* SANITY */
704                 sk_node_init(&newsk->sk_node);
705                 tcp_sk(newsk)->bind_hash = NULL;
706
707                 /* Clone the TCP header template */
708                 inet_sk(newsk)->dport = ireq->rmt_port;
709
710                 sock_lock_init(newsk);
711                 bh_lock_sock(newsk);
712
713                 rwlock_init(&newsk->sk_dst_lock);
714                 newsk->sk_dst_cache = NULL;
715                 atomic_set(&newsk->sk_rmem_alloc, 0);
716                 skb_queue_head_init(&newsk->sk_receive_queue);
717                 atomic_set(&newsk->sk_wmem_alloc, 0);
718                 skb_queue_head_init(&newsk->sk_write_queue);
719                 atomic_set(&newsk->sk_omem_alloc, 0);
720                 newsk->sk_wmem_queued = 0;
721                 newsk->sk_forward_alloc = 0;
722
723                 sock_reset_flag(newsk, SOCK_DONE);
724                 newsk->sk_userlocks = sk->sk_userlocks & ~SOCK_BINDPORT_LOCK;
725                 newsk->sk_backlog.head = newsk->sk_backlog.tail = NULL;
726                 newsk->sk_send_head = NULL;
727                 rwlock_init(&newsk->sk_callback_lock);
728                 skb_queue_head_init(&newsk->sk_error_queue);
729                 newsk->sk_write_space = sk_stream_write_space;
730
731                 if ((filter = newsk->sk_filter) != NULL)
732                         sk_filter_charge(newsk, filter);
733
734                 if (unlikely(xfrm_sk_clone_policy(newsk))) {
735                         /* It is still raw copy of parent, so invalidate
736                          * destructor and make plain sk_free() */
737                         newsk->sk_destruct = NULL;
738                         sk_free(newsk);
739                         return NULL;
740                 }
741
742                 /* Now setup tcp_sock */
743                 newtp = tcp_sk(newsk);
744                 newtp->pred_flags = 0;
745                 newtp->rcv_nxt = treq->rcv_isn + 1;
746                 newtp->snd_nxt = treq->snt_isn + 1;
747                 newtp->snd_una = treq->snt_isn + 1;
748                 newtp->snd_sml = treq->snt_isn + 1;
749
750                 tcp_prequeue_init(newtp);
751
752                 tcp_init_wl(newtp, treq->snt_isn, treq->rcv_isn);
753
754                 newtp->retransmits = 0;
755                 newtp->backoff = 0;
756                 newtp->srtt = 0;
757                 newtp->mdev = TCP_TIMEOUT_INIT;
758                 newtp->rto = TCP_TIMEOUT_INIT;
759
760                 newtp->packets_out = 0;
761                 newtp->left_out = 0;
762                 newtp->retrans_out = 0;
763                 newtp->sacked_out = 0;
764                 newtp->fackets_out = 0;
765                 newtp->snd_ssthresh = 0x7fffffff;
766
767                 /* So many TCP implementations out there (incorrectly) count the
768                  * initial SYN frame in their delayed-ACK and congestion control
769                  * algorithms that we must have the following bandaid to talk
770                  * efficiently to them.  -DaveM
771                  */
772                 newtp->snd_cwnd = 2;
773                 newtp->snd_cwnd_cnt = 0;
774
775                 newtp->frto_counter = 0;
776                 newtp->frto_highmark = 0;
777
778                 newtp->ca_ops = &tcp_reno;
779
780                 tcp_set_ca_state(newtp, TCP_CA_Open);
781                 tcp_init_xmit_timers(newsk);
782                 skb_queue_head_init(&newtp->out_of_order_queue);
783                 newtp->rcv_wup = treq->rcv_isn + 1;
784                 newtp->write_seq = treq->snt_isn + 1;
785                 newtp->pushed_seq = newtp->write_seq;
786                 newtp->copied_seq = treq->rcv_isn + 1;
787
788                 newtp->rx_opt.saw_tstamp = 0;
789
790                 newtp->rx_opt.dsack = 0;
791                 newtp->rx_opt.eff_sacks = 0;
792
793                 newtp->probes_out = 0;
794                 newtp->rx_opt.num_sacks = 0;
795                 newtp->urg_data = 0;
796                 /* Deinitialize accept_queue to trap illegal accesses. */
797                 memset(&newtp->accept_queue, 0, sizeof(newtp->accept_queue));
798
799                 /* Back to base struct sock members. */
800                 newsk->sk_err = 0;
801                 newsk->sk_priority = 0;
802                 atomic_set(&newsk->sk_refcnt, 2);
803
804                 /*
805                  * Increment the counter in the same struct proto as the master
806                  * sock (sk_refcnt_debug_inc uses newsk->sk_prot->socks, that
807                  * is the same as sk->sk_prot->socks, as this field was copied
808                  * with memcpy), same rationale as the first comment in this
809                  * function.
810                  *
811                  * This _changes_ the previous behaviour, where
812                  * tcp_create_openreq_child always was incrementing the
813                  * equivalent to tcp_prot->socks (inet_sock_nr), so this have
814                  * to be taken into account in all callers. -acme
815                  */
816                 sk_refcnt_debug_inc(newsk);
817
818                 atomic_inc(&tcp_sockets_allocated);
819
820                 if (sock_flag(newsk, SOCK_KEEPOPEN))
821                         tcp_reset_keepalive_timer(newsk,
822                                                   keepalive_time_when(newtp));
823                 newsk->sk_socket = NULL;
824                 newsk->sk_sleep = NULL;
825
826                 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
827                 if((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
828                         if (sysctl_tcp_fack)
829                                 newtp->rx_opt.sack_ok |= 2;
830                 }
831                 newtp->window_clamp = req->window_clamp;
832                 newtp->rcv_ssthresh = req->rcv_wnd;
833                 newtp->rcv_wnd = req->rcv_wnd;
834                 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
835                 if (newtp->rx_opt.wscale_ok) {
836                         newtp->rx_opt.snd_wscale = ireq->snd_wscale;
837                         newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
838                 } else {
839                         newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
840                         newtp->window_clamp = min(newtp->window_clamp, 65535U);
841                 }
842                 newtp->snd_wnd = ntohs(skb->h.th->window) << newtp->rx_opt.snd_wscale;
843                 newtp->max_window = newtp->snd_wnd;
844
845                 if (newtp->rx_opt.tstamp_ok) {
846                         newtp->rx_opt.ts_recent = req->ts_recent;
847                         newtp->rx_opt.ts_recent_stamp = xtime.tv_sec;
848                         newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
849                 } else {
850                         newtp->rx_opt.ts_recent_stamp = 0;
851                         newtp->tcp_header_len = sizeof(struct tcphdr);
852                 }
853                 if (skb->len >= TCP_MIN_RCVMSS+newtp->tcp_header_len)
854                         newtp->ack.last_seg_size = skb->len-newtp->tcp_header_len;
855                 newtp->rx_opt.mss_clamp = req->mss;
856                 TCP_ECN_openreq_child(newtp, req);
857                 if (newtp->ecn_flags&TCP_ECN_OK)
858                         sock_set_flag(newsk, SOCK_NO_LARGESEND);
859
860                 TCP_INC_STATS_BH(TCP_MIB_PASSIVEOPENS);
861         }
862         return newsk;
863 }
864
865 /* 
866  *      Process an incoming packet for SYN_RECV sockets represented
867  *      as a request_sock.
868  */
869
870 struct sock *tcp_check_req(struct sock *sk,struct sk_buff *skb,
871                            struct request_sock *req,
872                            struct request_sock **prev)
873 {
874         struct tcphdr *th = skb->h.th;
875         struct tcp_sock *tp = tcp_sk(sk);
876         u32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
877         int paws_reject = 0;
878         struct tcp_options_received tmp_opt;
879         struct sock *child;
880
881         tmp_opt.saw_tstamp = 0;
882         if (th->doff > (sizeof(struct tcphdr)>>2)) {
883                 tcp_parse_options(skb, &tmp_opt, 0);
884
885                 if (tmp_opt.saw_tstamp) {
886                         tmp_opt.ts_recent = req->ts_recent;
887                         /* We do not store true stamp, but it is not required,
888                          * it can be estimated (approximately)
889                          * from another data.
890                          */
891                         tmp_opt.ts_recent_stamp = xtime.tv_sec - ((TCP_TIMEOUT_INIT/HZ)<<req->retrans);
892                         paws_reject = tcp_paws_check(&tmp_opt, th->rst);
893                 }
894         }
895
896         /* Check for pure retransmitted SYN. */
897         if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
898             flg == TCP_FLAG_SYN &&
899             !paws_reject) {
900                 /*
901                  * RFC793 draws (Incorrectly! It was fixed in RFC1122)
902                  * this case on figure 6 and figure 8, but formal
903                  * protocol description says NOTHING.
904                  * To be more exact, it says that we should send ACK,
905                  * because this segment (at least, if it has no data)
906                  * is out of window.
907                  *
908                  *  CONCLUSION: RFC793 (even with RFC1122) DOES NOT
909                  *  describe SYN-RECV state. All the description
910                  *  is wrong, we cannot believe to it and should
911                  *  rely only on common sense and implementation
912                  *  experience.
913                  *
914                  * Enforce "SYN-ACK" according to figure 8, figure 6
915                  * of RFC793, fixed by RFC1122.
916                  */
917                 req->rsk_ops->rtx_syn_ack(sk, req, NULL);
918                 return NULL;
919         }
920
921         /* Further reproduces section "SEGMENT ARRIVES"
922            for state SYN-RECEIVED of RFC793.
923            It is broken, however, it does not work only
924            when SYNs are crossed.
925
926            You would think that SYN crossing is impossible here, since
927            we should have a SYN_SENT socket (from connect()) on our end,
928            but this is not true if the crossed SYNs were sent to both
929            ends by a malicious third party.  We must defend against this,
930            and to do that we first verify the ACK (as per RFC793, page
931            36) and reset if it is invalid.  Is this a true full defense?
932            To convince ourselves, let us consider a way in which the ACK
933            test can still pass in this 'malicious crossed SYNs' case.
934            Malicious sender sends identical SYNs (and thus identical sequence
935            numbers) to both A and B:
936
937                 A: gets SYN, seq=7
938                 B: gets SYN, seq=7
939
940            By our good fortune, both A and B select the same initial
941            send sequence number of seven :-)
942
943                 A: sends SYN|ACK, seq=7, ack_seq=8
944                 B: sends SYN|ACK, seq=7, ack_seq=8
945
946            So we are now A eating this SYN|ACK, ACK test passes.  So
947            does sequence test, SYN is truncated, and thus we consider
948            it a bare ACK.
949
950            If tp->defer_accept, we silently drop this bare ACK.  Otherwise,
951            we create an established connection.  Both ends (listening sockets)
952            accept the new incoming connection and try to talk to each other. 8-)
953
954            Note: This case is both harmless, and rare.  Possibility is about the
955            same as us discovering intelligent life on another plant tomorrow.
956
957            But generally, we should (RFC lies!) to accept ACK
958            from SYNACK both here and in tcp_rcv_state_process().
959            tcp_rcv_state_process() does not, hence, we do not too.
960
961            Note that the case is absolutely generic:
962            we cannot optimize anything here without
963            violating protocol. All the checks must be made
964            before attempt to create socket.
965          */
966
967         /* RFC793 page 36: "If the connection is in any non-synchronized state ...
968          *                  and the incoming segment acknowledges something not yet
969          *                  sent (the segment carries an unaccaptable ACK) ...
970          *                  a reset is sent."
971          *
972          * Invalid ACK: reset will be sent by listening socket
973          */
974         if ((flg & TCP_FLAG_ACK) &&
975             (TCP_SKB_CB(skb)->ack_seq != tcp_rsk(req)->snt_isn + 1))
976                 return sk;
977
978         /* Also, it would be not so bad idea to check rcv_tsecr, which
979          * is essentially ACK extension and too early or too late values
980          * should cause reset in unsynchronized states.
981          */
982
983         /* RFC793: "first check sequence number". */
984
985         if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
986                                           tcp_rsk(req)->rcv_isn + 1, tcp_rsk(req)->rcv_isn + 1 + req->rcv_wnd)) {
987                 /* Out of window: send ACK and drop. */
988                 if (!(flg & TCP_FLAG_RST))
989                         req->rsk_ops->send_ack(skb, req);
990                 if (paws_reject)
991                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
992                 return NULL;
993         }
994
995         /* In sequence, PAWS is OK. */
996
997         if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_isn + 1))
998                         req->ts_recent = tmp_opt.rcv_tsval;
999
1000                 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
1001                         /* Truncate SYN, it is out of window starting
1002                            at tcp_rsk(req)->rcv_isn + 1. */
1003                         flg &= ~TCP_FLAG_SYN;
1004                 }
1005
1006                 /* RFC793: "second check the RST bit" and
1007                  *         "fourth, check the SYN bit"
1008                  */
1009                 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN))
1010                         goto embryonic_reset;
1011
1012                 /* ACK sequence verified above, just make sure ACK is
1013                  * set.  If ACK not set, just silently drop the packet.
1014                  */
1015                 if (!(flg & TCP_FLAG_ACK))
1016                         return NULL;
1017
1018                 /* If TCP_DEFER_ACCEPT is set, drop bare ACK. */
1019                 if (tp->defer_accept && TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
1020                         inet_rsk(req)->acked = 1;
1021                         return NULL;
1022                 }
1023
1024                 /* OK, ACK is valid, create big socket and
1025                  * feed this segment to it. It will repeat all
1026                  * the tests. THIS SEGMENT MUST MOVE SOCKET TO
1027                  * ESTABLISHED STATE. If it will be dropped after
1028                  * socket is created, wait for troubles.
1029                  */
1030                 child = tp->af_specific->syn_recv_sock(sk, skb, req, NULL);
1031                 if (child == NULL)
1032                         goto listen_overflow;
1033
1034                 tcp_synq_unlink(tp, req, prev);
1035                 tcp_synq_removed(sk, req);
1036
1037                 tcp_acceptq_queue(sk, req, child);
1038                 return child;
1039
1040         listen_overflow:
1041                 if (!sysctl_tcp_abort_on_overflow) {
1042                         inet_rsk(req)->acked = 1;
1043                         return NULL;
1044                 }
1045
1046         embryonic_reset:
1047                 NET_INC_STATS_BH(LINUX_MIB_EMBRYONICRSTS);
1048                 if (!(flg & TCP_FLAG_RST))
1049                         req->rsk_ops->send_reset(skb);
1050
1051                 tcp_synq_drop(sk, req, prev);
1052                 return NULL;
1053 }
1054
1055 /*
1056  * Queue segment on the new socket if the new socket is active,
1057  * otherwise we just shortcircuit this and continue with
1058  * the new socket.
1059  */
1060
1061 int tcp_child_process(struct sock *parent, struct sock *child,
1062                       struct sk_buff *skb)
1063 {
1064         int ret = 0;
1065         int state = child->sk_state;
1066
1067         if (!sock_owned_by_user(child)) {
1068                 ret = tcp_rcv_state_process(child, skb, skb->h.th, skb->len);
1069
1070                 /* Wakeup parent, send SIGIO */
1071                 if (state == TCP_SYN_RECV && child->sk_state != state)
1072                         parent->sk_data_ready(parent, 0);
1073         } else {
1074                 /* Alas, it is possible again, because we do lookup
1075                  * in main socket hash table and lock on listening
1076                  * socket does not protect us more.
1077                  */
1078                 sk_add_backlog(child, skb);
1079         }
1080
1081         bh_unlock_sock(child);
1082         sock_put(child);
1083         return ret;
1084 }
1085
1086 EXPORT_SYMBOL(tcp_check_req);
1087 EXPORT_SYMBOL(tcp_child_process);
1088 EXPORT_SYMBOL(tcp_create_openreq_child);
1089 EXPORT_SYMBOL(tcp_timewait_state_process);
1090 EXPORT_SYMBOL(tcp_tw_deschedule);