Merge master.kernel.org:/pub/scm/linux/kernel/git/jejb/scsi-for-linus-2.6
[linux-2.6] / net / ipv4 / inet_connection_sock.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  *              Support for INET connection oriented protocols.
7  *
8  * Authors:     See the TCP sources
9  *
10  *              This program is free software; you can redistribute it and/or
11  *              modify it under the terms of the GNU General Public License
12  *              as published by the Free Software Foundation; either version
13  *              2 of the License, or(at your option) any later version.
14  */
15
16 #include <linux/config.h>
17 #include <linux/module.h>
18 #include <linux/jhash.h>
19
20 #include <net/inet_connection_sock.h>
21 #include <net/inet_hashtables.h>
22 #include <net/inet_timewait_sock.h>
23 #include <net/ip.h>
24 #include <net/route.h>
25 #include <net/tcp_states.h>
26 #include <net/xfrm.h>
27
28 #ifdef INET_CSK_DEBUG
29 const char inet_csk_timer_bug_msg[] = "inet_csk BUG: unknown timer value\n";
30 EXPORT_SYMBOL(inet_csk_timer_bug_msg);
31 #endif
32
33 /*
34  * This array holds the first and last local port number.
35  * For high-usage systems, use sysctl to change this to
36  * 32768-61000
37  */
38 int sysctl_local_port_range[2] = { 1024, 4999 };
39
40 static inline int inet_csk_bind_conflict(struct sock *sk, struct inet_bind_bucket *tb)
41 {
42         const u32 sk_rcv_saddr = inet_rcv_saddr(sk);
43         struct sock *sk2;
44         struct hlist_node *node;
45         int reuse = sk->sk_reuse;
46
47         sk_for_each_bound(sk2, node, &tb->owners) {
48                 if (sk != sk2 &&
49                     !inet_v6_ipv6only(sk2) &&
50                     (!sk->sk_bound_dev_if ||
51                      !sk2->sk_bound_dev_if ||
52                      sk->sk_bound_dev_if == sk2->sk_bound_dev_if)) {
53                         if (!reuse || !sk2->sk_reuse ||
54                             sk2->sk_state == TCP_LISTEN) {
55                                 const u32 sk2_rcv_saddr = inet_rcv_saddr(sk2);
56                                 if (!sk2_rcv_saddr || !sk_rcv_saddr ||
57                                     sk2_rcv_saddr == sk_rcv_saddr)
58                                         break;
59                         }
60                 }
61         }
62         return node != NULL;
63 }
64
65 /* Obtain a reference to a local port for the given sock,
66  * if snum is zero it means select any available local port.
67  */
68 int inet_csk_get_port(struct inet_hashinfo *hashinfo,
69                       struct sock *sk, unsigned short snum)
70 {
71         struct inet_bind_hashbucket *head;
72         struct hlist_node *node;
73         struct inet_bind_bucket *tb;
74         int ret;
75
76         local_bh_disable();
77         if (!snum) {
78                 int low = sysctl_local_port_range[0];
79                 int high = sysctl_local_port_range[1];
80                 int remaining = (high - low) + 1;
81                 int rover;
82
83                 spin_lock(&hashinfo->portalloc_lock);
84                 if (hashinfo->port_rover < low)
85                         rover = low;
86                 else
87                         rover = hashinfo->port_rover;
88                 do {
89                         rover++;
90                         if (rover > high)
91                                 rover = low;
92                         head = &hashinfo->bhash[inet_bhashfn(rover, hashinfo->bhash_size)];
93                         spin_lock(&head->lock);
94                         inet_bind_bucket_for_each(tb, node, &head->chain)
95                                 if (tb->port == rover)
96                                         goto next;
97                         break;
98                 next:
99                         spin_unlock(&head->lock);
100                 } while (--remaining > 0);
101                 hashinfo->port_rover = rover;
102                 spin_unlock(&hashinfo->portalloc_lock);
103
104                 /* Exhausted local port range during search?  It is not
105                  * possible for us to be holding one of the bind hash
106                  * locks if this test triggers, because if 'remaining'
107                  * drops to zero, we broke out of the do/while loop at
108                  * the top level, not from the 'break;' statement.
109                  */
110                 ret = 1;
111                 if (remaining <= 0)
112                         goto fail;
113
114                 /* OK, here is the one we will use.  HEAD is
115                  * non-NULL and we hold it's mutex.
116                  */
117                 snum = rover;
118         } else {
119                 head = &hashinfo->bhash[inet_bhashfn(snum, hashinfo->bhash_size)];
120                 spin_lock(&head->lock);
121                 inet_bind_bucket_for_each(tb, node, &head->chain)
122                         if (tb->port == snum)
123                                 goto tb_found;
124         }
125         tb = NULL;
126         goto tb_not_found;
127 tb_found:
128         if (!hlist_empty(&tb->owners)) {
129                 if (sk->sk_reuse > 1)
130                         goto success;
131                 if (tb->fastreuse > 0 &&
132                     sk->sk_reuse && sk->sk_state != TCP_LISTEN) {
133                         goto success;
134                 } else {
135                         ret = 1;
136                         if (inet_csk_bind_conflict(sk, tb))
137                                 goto fail_unlock;
138                 }
139         }
140 tb_not_found:
141         ret = 1;
142         if (!tb && (tb = inet_bind_bucket_create(hashinfo->bind_bucket_cachep, head, snum)) == NULL)
143                 goto fail_unlock;
144         if (hlist_empty(&tb->owners)) {
145                 if (sk->sk_reuse && sk->sk_state != TCP_LISTEN)
146                         tb->fastreuse = 1;
147                 else
148                         tb->fastreuse = 0;
149         } else if (tb->fastreuse &&
150                    (!sk->sk_reuse || sk->sk_state == TCP_LISTEN))
151                 tb->fastreuse = 0;
152 success:
153         if (!inet_csk(sk)->icsk_bind_hash)
154                 inet_bind_hash(sk, tb, snum);
155         BUG_TRAP(inet_csk(sk)->icsk_bind_hash == tb);
156         ret = 0;
157
158 fail_unlock:
159         spin_unlock(&head->lock);
160 fail:
161         local_bh_enable();
162         return ret;
163 }
164
165 EXPORT_SYMBOL_GPL(inet_csk_get_port);
166
167 /*
168  * Wait for an incoming connection, avoid race conditions. This must be called
169  * with the socket locked.
170  */
171 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
172 {
173         struct inet_connection_sock *icsk = inet_csk(sk);
174         DEFINE_WAIT(wait);
175         int err;
176
177         /*
178          * True wake-one mechanism for incoming connections: only
179          * one process gets woken up, not the 'whole herd'.
180          * Since we do not 'race & poll' for established sockets
181          * anymore, the common case will execute the loop only once.
182          *
183          * Subtle issue: "add_wait_queue_exclusive()" will be added
184          * after any current non-exclusive waiters, and we know that
185          * it will always _stay_ after any new non-exclusive waiters
186          * because all non-exclusive waiters are added at the
187          * beginning of the wait-queue. As such, it's ok to "drop"
188          * our exclusiveness temporarily when we get woken up without
189          * having to remove and re-insert us on the wait queue.
190          */
191         for (;;) {
192                 prepare_to_wait_exclusive(sk->sk_sleep, &wait,
193                                           TASK_INTERRUPTIBLE);
194                 release_sock(sk);
195                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
196                         timeo = schedule_timeout(timeo);
197                 lock_sock(sk);
198                 err = 0;
199                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
200                         break;
201                 err = -EINVAL;
202                 if (sk->sk_state != TCP_LISTEN)
203                         break;
204                 err = sock_intr_errno(timeo);
205                 if (signal_pending(current))
206                         break;
207                 err = -EAGAIN;
208                 if (!timeo)
209                         break;
210         }
211         finish_wait(sk->sk_sleep, &wait);
212         return err;
213 }
214
215 /*
216  * This will accept the next outstanding connection.
217  */
218 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err)
219 {
220         struct inet_connection_sock *icsk = inet_csk(sk);
221         struct sock *newsk;
222         int error;
223
224         lock_sock(sk);
225
226         /* We need to make sure that this socket is listening,
227          * and that it has something pending.
228          */
229         error = -EINVAL;
230         if (sk->sk_state != TCP_LISTEN)
231                 goto out_err;
232
233         /* Find already established connection */
234         if (reqsk_queue_empty(&icsk->icsk_accept_queue)) {
235                 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
236
237                 /* If this is a non blocking socket don't sleep */
238                 error = -EAGAIN;
239                 if (!timeo)
240                         goto out_err;
241
242                 error = inet_csk_wait_for_connect(sk, timeo);
243                 if (error)
244                         goto out_err;
245         }
246
247         newsk = reqsk_queue_get_child(&icsk->icsk_accept_queue, sk);
248         BUG_TRAP(newsk->sk_state != TCP_SYN_RECV);
249 out:
250         release_sock(sk);
251         return newsk;
252 out_err:
253         newsk = NULL;
254         *err = error;
255         goto out;
256 }
257
258 EXPORT_SYMBOL(inet_csk_accept);
259
260 /*
261  * Using different timers for retransmit, delayed acks and probes
262  * We may wish use just one timer maintaining a list of expire jiffies 
263  * to optimize.
264  */
265 void inet_csk_init_xmit_timers(struct sock *sk,
266                                void (*retransmit_handler)(unsigned long),
267                                void (*delack_handler)(unsigned long),
268                                void (*keepalive_handler)(unsigned long))
269 {
270         struct inet_connection_sock *icsk = inet_csk(sk);
271
272         init_timer(&icsk->icsk_retransmit_timer);
273         init_timer(&icsk->icsk_delack_timer);
274         init_timer(&sk->sk_timer);
275
276         icsk->icsk_retransmit_timer.function = retransmit_handler;
277         icsk->icsk_delack_timer.function     = delack_handler;
278         sk->sk_timer.function                = keepalive_handler;
279
280         icsk->icsk_retransmit_timer.data = 
281                 icsk->icsk_delack_timer.data =
282                         sk->sk_timer.data  = (unsigned long)sk;
283
284         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
285 }
286
287 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
288
289 void inet_csk_clear_xmit_timers(struct sock *sk)
290 {
291         struct inet_connection_sock *icsk = inet_csk(sk);
292
293         icsk->icsk_pending = icsk->icsk_ack.pending = icsk->icsk_ack.blocked = 0;
294
295         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
296         sk_stop_timer(sk, &icsk->icsk_delack_timer);
297         sk_stop_timer(sk, &sk->sk_timer);
298 }
299
300 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
301
302 void inet_csk_delete_keepalive_timer(struct sock *sk)
303 {
304         sk_stop_timer(sk, &sk->sk_timer);
305 }
306
307 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
308
309 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
310 {
311         sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
312 }
313
314 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
315
316 struct dst_entry* inet_csk_route_req(struct sock *sk,
317                                      const struct request_sock *req)
318 {
319         struct rtable *rt;
320         const struct inet_request_sock *ireq = inet_rsk(req);
321         struct ip_options *opt = inet_rsk(req)->opt;
322         struct flowi fl = { .oif = sk->sk_bound_dev_if,
323                             .nl_u = { .ip4_u =
324                                       { .daddr = ((opt && opt->srr) ?
325                                                   opt->faddr :
326                                                   ireq->rmt_addr),
327                                         .saddr = ireq->loc_addr,
328                                         .tos = RT_CONN_FLAGS(sk) } },
329                             .proto = sk->sk_protocol,
330                             .uli_u = { .ports =
331                                        { .sport = inet_sk(sk)->sport,
332                                          .dport = ireq->rmt_port } } };
333
334         if (ip_route_output_flow(&rt, &fl, sk, 0)) {
335                 IP_INC_STATS_BH(IPSTATS_MIB_OUTNOROUTES);
336                 return NULL;
337         }
338         if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway) {
339                 ip_rt_put(rt);
340                 IP_INC_STATS_BH(IPSTATS_MIB_OUTNOROUTES);
341                 return NULL;
342         }
343         return &rt->u.dst;
344 }
345
346 EXPORT_SYMBOL_GPL(inet_csk_route_req);
347
348 static inline u32 inet_synq_hash(const u32 raddr, const u16 rport,
349                                  const u32 rnd, const u16 synq_hsize)
350 {
351         return jhash_2words(raddr, (u32)rport, rnd) & (synq_hsize - 1);
352 }
353
354 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
355 #define AF_INET_FAMILY(fam) ((fam) == AF_INET)
356 #else
357 #define AF_INET_FAMILY(fam) 1
358 #endif
359
360 struct request_sock *inet_csk_search_req(const struct sock *sk,
361                                          struct request_sock ***prevp,
362                                          const __u16 rport, const __u32 raddr,
363                                          const __u32 laddr)
364 {
365         const struct inet_connection_sock *icsk = inet_csk(sk);
366         struct listen_sock *lopt = icsk->icsk_accept_queue.listen_opt;
367         struct request_sock *req, **prev;
368
369         for (prev = &lopt->syn_table[inet_synq_hash(raddr, rport, lopt->hash_rnd,
370                                                     lopt->nr_table_entries)];
371              (req = *prev) != NULL;
372              prev = &req->dl_next) {
373                 const struct inet_request_sock *ireq = inet_rsk(req);
374
375                 if (ireq->rmt_port == rport &&
376                     ireq->rmt_addr == raddr &&
377                     ireq->loc_addr == laddr &&
378                     AF_INET_FAMILY(req->rsk_ops->family)) {
379                         BUG_TRAP(!req->sk);
380                         *prevp = prev;
381                         break;
382                 }
383         }
384
385         return req;
386 }
387
388 EXPORT_SYMBOL_GPL(inet_csk_search_req);
389
390 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
391                                    const unsigned timeout)
392 {
393         struct inet_connection_sock *icsk = inet_csk(sk);
394         struct listen_sock *lopt = icsk->icsk_accept_queue.listen_opt;
395         const u32 h = inet_synq_hash(inet_rsk(req)->rmt_addr, inet_rsk(req)->rmt_port,
396                                      lopt->hash_rnd, lopt->nr_table_entries);
397
398         reqsk_queue_hash_req(&icsk->icsk_accept_queue, h, req, timeout);
399         inet_csk_reqsk_queue_added(sk, timeout);
400 }
401
402 /* Only thing we need from tcp.h */
403 extern int sysctl_tcp_synack_retries;
404
405 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
406
407 void inet_csk_reqsk_queue_prune(struct sock *parent,
408                                 const unsigned long interval,
409                                 const unsigned long timeout,
410                                 const unsigned long max_rto)
411 {
412         struct inet_connection_sock *icsk = inet_csk(parent);
413         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
414         struct listen_sock *lopt = queue->listen_opt;
415         int max_retries = icsk->icsk_syn_retries ? : sysctl_tcp_synack_retries;
416         int thresh = max_retries;
417         unsigned long now = jiffies;
418         struct request_sock **reqp, *req;
419         int i, budget;
420
421         if (lopt == NULL || lopt->qlen == 0)
422                 return;
423
424         /* Normally all the openreqs are young and become mature
425          * (i.e. converted to established socket) for first timeout.
426          * If synack was not acknowledged for 3 seconds, it means
427          * one of the following things: synack was lost, ack was lost,
428          * rtt is high or nobody planned to ack (i.e. synflood).
429          * When server is a bit loaded, queue is populated with old
430          * open requests, reducing effective size of queue.
431          * When server is well loaded, queue size reduces to zero
432          * after several minutes of work. It is not synflood,
433          * it is normal operation. The solution is pruning
434          * too old entries overriding normal timeout, when
435          * situation becomes dangerous.
436          *
437          * Essentially, we reserve half of room for young
438          * embrions; and abort old ones without pity, if old
439          * ones are about to clog our table.
440          */
441         if (lopt->qlen>>(lopt->max_qlen_log-1)) {
442                 int young = (lopt->qlen_young<<1);
443
444                 while (thresh > 2) {
445                         if (lopt->qlen < young)
446                                 break;
447                         thresh--;
448                         young <<= 1;
449                 }
450         }
451
452         if (queue->rskq_defer_accept)
453                 max_retries = queue->rskq_defer_accept;
454
455         budget = 2 * (lopt->nr_table_entries / (timeout / interval));
456         i = lopt->clock_hand;
457
458         do {
459                 reqp=&lopt->syn_table[i];
460                 while ((req = *reqp) != NULL) {
461                         if (time_after_eq(now, req->expires)) {
462                                 if ((req->retrans < thresh ||
463                                      (inet_rsk(req)->acked && req->retrans < max_retries))
464                                     && !req->rsk_ops->rtx_syn_ack(parent, req, NULL)) {
465                                         unsigned long timeo;
466
467                                         if (req->retrans++ == 0)
468                                                 lopt->qlen_young--;
469                                         timeo = min((timeout << req->retrans), max_rto);
470                                         req->expires = now + timeo;
471                                         reqp = &req->dl_next;
472                                         continue;
473                                 }
474
475                                 /* Drop this request */
476                                 inet_csk_reqsk_queue_unlink(parent, req, reqp);
477                                 reqsk_queue_removed(queue, req);
478                                 reqsk_free(req);
479                                 continue;
480                         }
481                         reqp = &req->dl_next;
482                 }
483
484                 i = (i + 1) & (lopt->nr_table_entries - 1);
485
486         } while (--budget > 0);
487
488         lopt->clock_hand = i;
489
490         if (lopt->qlen)
491                 inet_csk_reset_keepalive_timer(parent, interval);
492 }
493
494 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_prune);
495
496 struct sock *inet_csk_clone(struct sock *sk, const struct request_sock *req,
497                             const unsigned int __nocast priority)
498 {
499         struct sock *newsk = sk_clone(sk, priority);
500
501         if (newsk != NULL) {
502                 struct inet_connection_sock *newicsk = inet_csk(newsk);
503
504                 newsk->sk_state = TCP_SYN_RECV;
505                 newicsk->icsk_bind_hash = NULL;
506
507                 inet_sk(newsk)->dport = inet_rsk(req)->rmt_port;
508                 newsk->sk_write_space = sk_stream_write_space;
509
510                 newicsk->icsk_retransmits = 0;
511                 newicsk->icsk_backoff     = 0;
512                 newicsk->icsk_probes_out  = 0;
513
514                 /* Deinitialize accept_queue to trap illegal accesses. */
515                 memset(&newicsk->icsk_accept_queue, 0, sizeof(newicsk->icsk_accept_queue));
516         }
517         return newsk;
518 }
519
520 EXPORT_SYMBOL_GPL(inet_csk_clone);
521
522 /*
523  * At this point, there should be no process reference to this
524  * socket, and thus no user references at all.  Therefore we
525  * can assume the socket waitqueue is inactive and nobody will
526  * try to jump onto it.
527  */
528 void inet_csk_destroy_sock(struct sock *sk)
529 {
530         BUG_TRAP(sk->sk_state == TCP_CLOSE);
531         BUG_TRAP(sock_flag(sk, SOCK_DEAD));
532
533         /* It cannot be in hash table! */
534         BUG_TRAP(sk_unhashed(sk));
535
536         /* If it has not 0 inet_sk(sk)->num, it must be bound */
537         BUG_TRAP(!inet_sk(sk)->num || inet_csk(sk)->icsk_bind_hash);
538
539         sk->sk_prot->destroy(sk);
540
541         sk_stream_kill_queues(sk);
542
543         xfrm_sk_free_policy(sk);
544
545         sk_refcnt_debug_release(sk);
546
547         atomic_dec(sk->sk_prot->orphan_count);
548         sock_put(sk);
549 }
550
551 EXPORT_SYMBOL(inet_csk_destroy_sock);
552
553 int inet_csk_listen_start(struct sock *sk, const int nr_table_entries)
554 {
555         struct inet_sock *inet = inet_sk(sk);
556         struct inet_connection_sock *icsk = inet_csk(sk);
557         int rc = reqsk_queue_alloc(&icsk->icsk_accept_queue, nr_table_entries);
558
559         if (rc != 0)
560                 return rc;
561
562         sk->sk_max_ack_backlog = 0;
563         sk->sk_ack_backlog = 0;
564         inet_csk_delack_init(sk);
565
566         /* There is race window here: we announce ourselves listening,
567          * but this transition is still not validated by get_port().
568          * It is OK, because this socket enters to hash table only
569          * after validation is complete.
570          */
571         sk->sk_state = TCP_LISTEN;
572         if (!sk->sk_prot->get_port(sk, inet->num)) {
573                 inet->sport = htons(inet->num);
574
575                 sk_dst_reset(sk);
576                 sk->sk_prot->hash(sk);
577
578                 return 0;
579         }
580
581         sk->sk_state = TCP_CLOSE;
582         __reqsk_queue_destroy(&icsk->icsk_accept_queue);
583         return -EADDRINUSE;
584 }
585
586 EXPORT_SYMBOL_GPL(inet_csk_listen_start);
587
588 /*
589  *      This routine closes sockets which have been at least partially
590  *      opened, but not yet accepted.
591  */
592 void inet_csk_listen_stop(struct sock *sk)
593 {
594         struct inet_connection_sock *icsk = inet_csk(sk);
595         struct request_sock *acc_req;
596         struct request_sock *req;
597
598         inet_csk_delete_keepalive_timer(sk);
599
600         /* make all the listen_opt local to us */
601         acc_req = reqsk_queue_yank_acceptq(&icsk->icsk_accept_queue);
602
603         /* Following specs, it would be better either to send FIN
604          * (and enter FIN-WAIT-1, it is normal close)
605          * or to send active reset (abort).
606          * Certainly, it is pretty dangerous while synflood, but it is
607          * bad justification for our negligence 8)
608          * To be honest, we are not able to make either
609          * of the variants now.                 --ANK
610          */
611         reqsk_queue_destroy(&icsk->icsk_accept_queue);
612
613         while ((req = acc_req) != NULL) {
614                 struct sock *child = req->sk;
615
616                 acc_req = req->dl_next;
617
618                 local_bh_disable();
619                 bh_lock_sock(child);
620                 BUG_TRAP(!sock_owned_by_user(child));
621                 sock_hold(child);
622
623                 sk->sk_prot->disconnect(child, O_NONBLOCK);
624
625                 sock_orphan(child);
626
627                 atomic_inc(sk->sk_prot->orphan_count);
628
629                 inet_csk_destroy_sock(child);
630
631                 bh_unlock_sock(child);
632                 local_bh_enable();
633                 sock_put(child);
634
635                 sk_acceptq_removed(sk);
636                 __reqsk_free(req);
637         }
638         BUG_TRAP(!sk->sk_ack_backlog);
639 }
640
641 EXPORT_SYMBOL_GPL(inet_csk_listen_stop);