[IPVS]: Fix sched registration race when checking for name collision.
[linux-2.6] / net / ipv4 / tcp_timer.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_timer.c,v 1.88 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/module.h>
24 #include <net/tcp.h>
25
26 int sysctl_tcp_syn_retries __read_mostly = TCP_SYN_RETRIES;
27 int sysctl_tcp_synack_retries __read_mostly = TCP_SYNACK_RETRIES;
28 int sysctl_tcp_keepalive_time __read_mostly = TCP_KEEPALIVE_TIME;
29 int sysctl_tcp_keepalive_probes __read_mostly = TCP_KEEPALIVE_PROBES;
30 int sysctl_tcp_keepalive_intvl __read_mostly = TCP_KEEPALIVE_INTVL;
31 int sysctl_tcp_retries1 __read_mostly = TCP_RETR1;
32 int sysctl_tcp_retries2 __read_mostly = TCP_RETR2;
33 int sysctl_tcp_orphan_retries __read_mostly;
34
35 static void tcp_write_timer(unsigned long);
36 static void tcp_delack_timer(unsigned long);
37 static void tcp_keepalive_timer (unsigned long data);
38
39 void tcp_init_xmit_timers(struct sock *sk)
40 {
41         inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
42                                   &tcp_keepalive_timer);
43 }
44
45 EXPORT_SYMBOL(tcp_init_xmit_timers);
46
47 static void tcp_write_err(struct sock *sk)
48 {
49         sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
50         sk->sk_error_report(sk);
51
52         tcp_done(sk);
53         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONTIMEOUT);
54 }
55
56 /* Do not allow orphaned sockets to eat all our resources.
57  * This is direct violation of TCP specs, but it is required
58  * to prevent DoS attacks. It is called when a retransmission timeout
59  * or zero probe timeout occurs on orphaned socket.
60  *
61  * Criteria is still not confirmed experimentally and may change.
62  * We kill the socket, if:
63  * 1. If number of orphaned sockets exceeds an administratively configured
64  *    limit.
65  * 2. If we have strong memory pressure.
66  */
67 static int tcp_out_of_resources(struct sock *sk, int do_reset)
68 {
69         struct tcp_sock *tp = tcp_sk(sk);
70         int orphans = atomic_read(&tcp_orphan_count);
71
72         /* If peer does not open window for long time, or did not transmit
73          * anything for long time, penalize it. */
74         if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
75                 orphans <<= 1;
76
77         /* If some dubious ICMP arrived, penalize even more. */
78         if (sk->sk_err_soft)
79                 orphans <<= 1;
80
81         if (tcp_too_many_orphans(sk, orphans)) {
82                 if (net_ratelimit())
83                         printk(KERN_INFO "Out of socket memory\n");
84
85                 /* Catch exceptional cases, when connection requires reset.
86                  *      1. Last segment was sent recently. */
87                 if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
88                     /*  2. Window is closed. */
89                     (!tp->snd_wnd && !tp->packets_out))
90                         do_reset = 1;
91                 if (do_reset)
92                         tcp_send_active_reset(sk, GFP_ATOMIC);
93                 tcp_done(sk);
94                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
95                 return 1;
96         }
97         return 0;
98 }
99
100 /* Calculate maximal number or retries on an orphaned socket. */
101 static int tcp_orphan_retries(struct sock *sk, int alive)
102 {
103         int retries = sysctl_tcp_orphan_retries; /* May be zero. */
104
105         /* We know from an ICMP that something is wrong. */
106         if (sk->sk_err_soft && !alive)
107                 retries = 0;
108
109         /* However, if socket sent something recently, select some safe
110          * number of retries. 8 corresponds to >100 seconds with minimal
111          * RTO of 200msec. */
112         if (retries == 0 && alive)
113                 retries = 8;
114         return retries;
115 }
116
117 /* A write timeout has occurred. Process the after effects. */
118 static int tcp_write_timeout(struct sock *sk)
119 {
120         struct inet_connection_sock *icsk = inet_csk(sk);
121         struct tcp_sock *tp = tcp_sk(sk);
122         int retry_until;
123         int mss;
124
125         if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
126                 if (icsk->icsk_retransmits)
127                         dst_negative_advice(&sk->sk_dst_cache);
128                 retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
129         } else {
130                 if (icsk->icsk_retransmits >= sysctl_tcp_retries1) {
131                         /* Black hole detection */
132                         if (sysctl_tcp_mtu_probing) {
133                                 if (!icsk->icsk_mtup.enabled) {
134                                         icsk->icsk_mtup.enabled = 1;
135                                         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
136                                 } else {
137                                         mss = min(sysctl_tcp_base_mss,
138                                                   tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)/2);
139                                         mss = max(mss, 68 - tp->tcp_header_len);
140                                         icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
141                                         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
142                                 }
143                         }
144
145                         dst_negative_advice(&sk->sk_dst_cache);
146                 }
147
148                 retry_until = sysctl_tcp_retries2;
149                 if (sock_flag(sk, SOCK_DEAD)) {
150                         const int alive = (icsk->icsk_rto < TCP_RTO_MAX);
151
152                         retry_until = tcp_orphan_retries(sk, alive);
153
154                         if (tcp_out_of_resources(sk, alive || icsk->icsk_retransmits < retry_until))
155                                 return 1;
156                 }
157         }
158
159         if (icsk->icsk_retransmits >= retry_until) {
160                 /* Has it gone just too far? */
161                 tcp_write_err(sk);
162                 return 1;
163         }
164         return 0;
165 }
166
167 static void tcp_delack_timer(unsigned long data)
168 {
169         struct sock *sk = (struct sock*)data;
170         struct tcp_sock *tp = tcp_sk(sk);
171         struct inet_connection_sock *icsk = inet_csk(sk);
172
173         bh_lock_sock(sk);
174         if (sock_owned_by_user(sk)) {
175                 /* Try again later. */
176                 icsk->icsk_ack.blocked = 1;
177                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOCKED);
178                 sk_reset_timer(sk, &icsk->icsk_delack_timer, jiffies + TCP_DELACK_MIN);
179                 goto out_unlock;
180         }
181
182         sk_stream_mem_reclaim(sk);
183
184         if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
185                 goto out;
186
187         if (time_after(icsk->icsk_ack.timeout, jiffies)) {
188                 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
189                 goto out;
190         }
191         icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
192
193         if (!skb_queue_empty(&tp->ucopy.prequeue)) {
194                 struct sk_buff *skb;
195
196                 NET_INC_STATS_BH(LINUX_MIB_TCPSCHEDULERFAILED);
197
198                 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
199                         sk->sk_backlog_rcv(sk, skb);
200
201                 tp->ucopy.memory = 0;
202         }
203
204         if (inet_csk_ack_scheduled(sk)) {
205                 if (!icsk->icsk_ack.pingpong) {
206                         /* Delayed ACK missed: inflate ATO. */
207                         icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
208                 } else {
209                         /* Delayed ACK missed: leave pingpong mode and
210                          * deflate ATO.
211                          */
212                         icsk->icsk_ack.pingpong = 0;
213                         icsk->icsk_ack.ato      = TCP_ATO_MIN;
214                 }
215                 tcp_send_ack(sk);
216                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKS);
217         }
218         TCP_CHECK_TIMER(sk);
219
220 out:
221         if (tcp_memory_pressure)
222                 sk_stream_mem_reclaim(sk);
223 out_unlock:
224         bh_unlock_sock(sk);
225         sock_put(sk);
226 }
227
228 static void tcp_probe_timer(struct sock *sk)
229 {
230         struct inet_connection_sock *icsk = inet_csk(sk);
231         struct tcp_sock *tp = tcp_sk(sk);
232         int max_probes;
233
234         if (tp->packets_out || !tcp_send_head(sk)) {
235                 icsk->icsk_probes_out = 0;
236                 return;
237         }
238
239         /* *WARNING* RFC 1122 forbids this
240          *
241          * It doesn't AFAIK, because we kill the retransmit timer -AK
242          *
243          * FIXME: We ought not to do it, Solaris 2.5 actually has fixing
244          * this behaviour in Solaris down as a bug fix. [AC]
245          *
246          * Let me to explain. icsk_probes_out is zeroed by incoming ACKs
247          * even if they advertise zero window. Hence, connection is killed only
248          * if we received no ACKs for normal connection timeout. It is not killed
249          * only because window stays zero for some time, window may be zero
250          * until armageddon and even later. We are in full accordance
251          * with RFCs, only probe timer combines both retransmission timeout
252          * and probe timeout in one bottle.                             --ANK
253          */
254         max_probes = sysctl_tcp_retries2;
255
256         if (sock_flag(sk, SOCK_DEAD)) {
257                 const int alive = ((icsk->icsk_rto << icsk->icsk_backoff) < TCP_RTO_MAX);
258
259                 max_probes = tcp_orphan_retries(sk, alive);
260
261                 if (tcp_out_of_resources(sk, alive || icsk->icsk_probes_out <= max_probes))
262                         return;
263         }
264
265         if (icsk->icsk_probes_out > max_probes) {
266                 tcp_write_err(sk);
267         } else {
268                 /* Only send another probe if we didn't close things up. */
269                 tcp_send_probe0(sk);
270         }
271 }
272
273 /*
274  *      The TCP retransmit timer.
275  */
276
277 static void tcp_retransmit_timer(struct sock *sk)
278 {
279         struct tcp_sock *tp = tcp_sk(sk);
280         struct inet_connection_sock *icsk = inet_csk(sk);
281
282         if (!tp->packets_out)
283                 goto out;
284
285         BUG_TRAP(!tcp_write_queue_empty(sk));
286
287         if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
288             !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
289                 /* Receiver dastardly shrinks window. Our retransmits
290                  * become zero probes, but we should not timeout this
291                  * connection. If the socket is an orphan, time it out,
292                  * we cannot allow such beasts to hang infinitely.
293                  */
294 #ifdef TCP_DEBUG
295                 if (1) {
296                         struct inet_sock *inet = inet_sk(sk);
297                         LIMIT_NETDEBUG(KERN_DEBUG "TCP: Treason uncloaked! Peer %u.%u.%u.%u:%u/%u shrinks window %u:%u. Repaired.\n",
298                                NIPQUAD(inet->daddr), ntohs(inet->dport),
299                                inet->num, tp->snd_una, tp->snd_nxt);
300                 }
301 #endif
302                 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
303                         tcp_write_err(sk);
304                         goto out;
305                 }
306                 tcp_enter_loss(sk, 0);
307                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
308                 __sk_dst_reset(sk);
309                 goto out_reset_timer;
310         }
311
312         if (tcp_write_timeout(sk))
313                 goto out;
314
315         if (icsk->icsk_retransmits == 0) {
316                 if (icsk->icsk_ca_state == TCP_CA_Disorder ||
317                     icsk->icsk_ca_state == TCP_CA_Recovery) {
318                         if (tcp_is_sack(tp)) {
319                                 if (icsk->icsk_ca_state == TCP_CA_Recovery)
320                                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERYFAIL);
321                                 else
322                                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKFAILURES);
323                         } else {
324                                 if (icsk->icsk_ca_state == TCP_CA_Recovery)
325                                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERYFAIL);
326                                 else
327                                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOFAILURES);
328                         }
329                 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
330                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSFAILURES);
331                 } else {
332                         NET_INC_STATS_BH(LINUX_MIB_TCPTIMEOUTS);
333                 }
334         }
335
336         if (tcp_use_frto(sk)) {
337                 tcp_enter_frto(sk);
338         } else {
339                 tcp_enter_loss(sk, 0);
340         }
341
342         if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
343                 /* Retransmission failed because of local congestion,
344                  * do not backoff.
345                  */
346                 if (!icsk->icsk_retransmits)
347                         icsk->icsk_retransmits = 1;
348                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
349                                           min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
350                                           TCP_RTO_MAX);
351                 goto out;
352         }
353
354         /* Increase the timeout each time we retransmit.  Note that
355          * we do not increase the rtt estimate.  rto is initialized
356          * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
357          * that doubling rto each time is the least we can get away with.
358          * In KA9Q, Karn uses this for the first few times, and then
359          * goes to quadratic.  netBSD doubles, but only goes up to *64,
360          * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
361          * defined in the protocol as the maximum possible RTT.  I guess
362          * we'll have to use something other than TCP to talk to the
363          * University of Mars.
364          *
365          * PAWS allows us longer timeouts and large windows, so once
366          * implemented ftp to mars will work nicely. We will have to fix
367          * the 120 second clamps though!
368          */
369         icsk->icsk_backoff++;
370         icsk->icsk_retransmits++;
371
372 out_reset_timer:
373         icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
374         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
375         if (icsk->icsk_retransmits > sysctl_tcp_retries1)
376                 __sk_dst_reset(sk);
377
378 out:;
379 }
380
381 static void tcp_write_timer(unsigned long data)
382 {
383         struct sock *sk = (struct sock*)data;
384         struct inet_connection_sock *icsk = inet_csk(sk);
385         int event;
386
387         bh_lock_sock(sk);
388         if (sock_owned_by_user(sk)) {
389                 /* Try again later */
390                 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + (HZ / 20));
391                 goto out_unlock;
392         }
393
394         if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
395                 goto out;
396
397         if (time_after(icsk->icsk_timeout, jiffies)) {
398                 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
399                 goto out;
400         }
401
402         event = icsk->icsk_pending;
403         icsk->icsk_pending = 0;
404
405         switch (event) {
406         case ICSK_TIME_RETRANS:
407                 tcp_retransmit_timer(sk);
408                 break;
409         case ICSK_TIME_PROBE0:
410                 tcp_probe_timer(sk);
411                 break;
412         }
413         TCP_CHECK_TIMER(sk);
414
415 out:
416         sk_stream_mem_reclaim(sk);
417 out_unlock:
418         bh_unlock_sock(sk);
419         sock_put(sk);
420 }
421
422 /*
423  *      Timer for listening sockets
424  */
425
426 static void tcp_synack_timer(struct sock *sk)
427 {
428         inet_csk_reqsk_queue_prune(sk, TCP_SYNQ_INTERVAL,
429                                    TCP_TIMEOUT_INIT, TCP_RTO_MAX);
430 }
431
432 void tcp_set_keepalive(struct sock *sk, int val)
433 {
434         if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
435                 return;
436
437         if (val && !sock_flag(sk, SOCK_KEEPOPEN))
438                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
439         else if (!val)
440                 inet_csk_delete_keepalive_timer(sk);
441 }
442
443
444 static void tcp_keepalive_timer (unsigned long data)
445 {
446         struct sock *sk = (struct sock *) data;
447         struct inet_connection_sock *icsk = inet_csk(sk);
448         struct tcp_sock *tp = tcp_sk(sk);
449         __u32 elapsed;
450
451         /* Only process if socket is not in use. */
452         bh_lock_sock(sk);
453         if (sock_owned_by_user(sk)) {
454                 /* Try again later. */
455                 inet_csk_reset_keepalive_timer (sk, HZ/20);
456                 goto out;
457         }
458
459         if (sk->sk_state == TCP_LISTEN) {
460                 tcp_synack_timer(sk);
461                 goto out;
462         }
463
464         if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
465                 if (tp->linger2 >= 0) {
466                         const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
467
468                         if (tmo > 0) {
469                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
470                                 goto out;
471                         }
472                 }
473                 tcp_send_active_reset(sk, GFP_ATOMIC);
474                 goto death;
475         }
476
477         if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
478                 goto out;
479
480         elapsed = keepalive_time_when(tp);
481
482         /* It is alive without keepalive 8) */
483         if (tp->packets_out || tcp_send_head(sk))
484                 goto resched;
485
486         elapsed = tcp_time_stamp - tp->rcv_tstamp;
487
488         if (elapsed >= keepalive_time_when(tp)) {
489                 if ((!tp->keepalive_probes && icsk->icsk_probes_out >= sysctl_tcp_keepalive_probes) ||
490                      (tp->keepalive_probes && icsk->icsk_probes_out >= tp->keepalive_probes)) {
491                         tcp_send_active_reset(sk, GFP_ATOMIC);
492                         tcp_write_err(sk);
493                         goto out;
494                 }
495                 if (tcp_write_wakeup(sk) <= 0) {
496                         icsk->icsk_probes_out++;
497                         elapsed = keepalive_intvl_when(tp);
498                 } else {
499                         /* If keepalive was lost due to local congestion,
500                          * try harder.
501                          */
502                         elapsed = TCP_RESOURCE_PROBE_INTERVAL;
503                 }
504         } else {
505                 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
506                 elapsed = keepalive_time_when(tp) - elapsed;
507         }
508
509         TCP_CHECK_TIMER(sk);
510         sk_stream_mem_reclaim(sk);
511
512 resched:
513         inet_csk_reset_keepalive_timer (sk, elapsed);
514         goto out;
515
516 death:
517         tcp_done(sk);
518
519 out:
520         bh_unlock_sock(sk);
521         sock_put(sk);
522 }