[NET] CORE: Introducing new memory accounting interface.
[linux-2.6.git] / 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 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
118 {
119         /* Black hole detection */
120         if (sysctl_tcp_mtu_probing) {
121                 if (!icsk->icsk_mtup.enabled) {
122                         icsk->icsk_mtup.enabled = 1;
123                         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
124                 } else {
125                         struct tcp_sock *tp = tcp_sk(sk);
126                         int mss;
127
128                         mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
129                         mss = min(sysctl_tcp_base_mss, mss);
130                         mss = max(mss, 68 - tp->tcp_header_len);
131                         icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
132                         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
133                 }
134         }
135 }
136
137 /* A write timeout has occurred. Process the after effects. */
138 static int tcp_write_timeout(struct sock *sk)
139 {
140         struct inet_connection_sock *icsk = inet_csk(sk);
141         int retry_until;
142
143         if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
144                 if (icsk->icsk_retransmits)
145                         dst_negative_advice(&sk->sk_dst_cache);
146                 retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
147         } else {
148                 if (icsk->icsk_retransmits >= sysctl_tcp_retries1) {
149                         /* Black hole detection */
150                         tcp_mtu_probing(icsk, sk);
151
152                         dst_negative_advice(&sk->sk_dst_cache);
153                 }
154
155                 retry_until = sysctl_tcp_retries2;
156                 if (sock_flag(sk, SOCK_DEAD)) {
157                         const int alive = (icsk->icsk_rto < TCP_RTO_MAX);
158
159                         retry_until = tcp_orphan_retries(sk, alive);
160
161                         if (tcp_out_of_resources(sk, alive || icsk->icsk_retransmits < retry_until))
162                                 return 1;
163                 }
164         }
165
166         if (icsk->icsk_retransmits >= retry_until) {
167                 /* Has it gone just too far? */
168                 tcp_write_err(sk);
169                 return 1;
170         }
171         return 0;
172 }
173
174 static void tcp_delack_timer(unsigned long data)
175 {
176         struct sock *sk = (struct sock*)data;
177         struct tcp_sock *tp = tcp_sk(sk);
178         struct inet_connection_sock *icsk = inet_csk(sk);
179
180         bh_lock_sock(sk);
181         if (sock_owned_by_user(sk)) {
182                 /* Try again later. */
183                 icsk->icsk_ack.blocked = 1;
184                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOCKED);
185                 sk_reset_timer(sk, &icsk->icsk_delack_timer, jiffies + TCP_DELACK_MIN);
186                 goto out_unlock;
187         }
188
189         sk_mem_reclaim(sk);
190
191         if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
192                 goto out;
193
194         if (time_after(icsk->icsk_ack.timeout, jiffies)) {
195                 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
196                 goto out;
197         }
198         icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
199
200         if (!skb_queue_empty(&tp->ucopy.prequeue)) {
201                 struct sk_buff *skb;
202
203                 NET_INC_STATS_BH(LINUX_MIB_TCPSCHEDULERFAILED);
204
205                 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
206                         sk->sk_backlog_rcv(sk, skb);
207
208                 tp->ucopy.memory = 0;
209         }
210
211         if (inet_csk_ack_scheduled(sk)) {
212                 if (!icsk->icsk_ack.pingpong) {
213                         /* Delayed ACK missed: inflate ATO. */
214                         icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
215                 } else {
216                         /* Delayed ACK missed: leave pingpong mode and
217                          * deflate ATO.
218                          */
219                         icsk->icsk_ack.pingpong = 0;
220                         icsk->icsk_ack.ato      = TCP_ATO_MIN;
221                 }
222                 tcp_send_ack(sk);
223                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKS);
224         }
225         TCP_CHECK_TIMER(sk);
226
227 out:
228         if (tcp_memory_pressure)
229                 sk_mem_reclaim(sk);
230 out_unlock:
231         bh_unlock_sock(sk);
232         sock_put(sk);
233 }
234
235 static void tcp_probe_timer(struct sock *sk)
236 {
237         struct inet_connection_sock *icsk = inet_csk(sk);
238         struct tcp_sock *tp = tcp_sk(sk);
239         int max_probes;
240
241         if (tp->packets_out || !tcp_send_head(sk)) {
242                 icsk->icsk_probes_out = 0;
243                 return;
244         }
245
246         /* *WARNING* RFC 1122 forbids this
247          *
248          * It doesn't AFAIK, because we kill the retransmit timer -AK
249          *
250          * FIXME: We ought not to do it, Solaris 2.5 actually has fixing
251          * this behaviour in Solaris down as a bug fix. [AC]
252          *
253          * Let me to explain. icsk_probes_out is zeroed by incoming ACKs
254          * even if they advertise zero window. Hence, connection is killed only
255          * if we received no ACKs for normal connection timeout. It is not killed
256          * only because window stays zero for some time, window may be zero
257          * until armageddon and even later. We are in full accordance
258          * with RFCs, only probe timer combines both retransmission timeout
259          * and probe timeout in one bottle.                             --ANK
260          */
261         max_probes = sysctl_tcp_retries2;
262
263         if (sock_flag(sk, SOCK_DEAD)) {
264                 const int alive = ((icsk->icsk_rto << icsk->icsk_backoff) < TCP_RTO_MAX);
265
266                 max_probes = tcp_orphan_retries(sk, alive);
267
268                 if (tcp_out_of_resources(sk, alive || icsk->icsk_probes_out <= max_probes))
269                         return;
270         }
271
272         if (icsk->icsk_probes_out > max_probes) {
273                 tcp_write_err(sk);
274         } else {
275                 /* Only send another probe if we didn't close things up. */
276                 tcp_send_probe0(sk);
277         }
278 }
279
280 /*
281  *      The TCP retransmit timer.
282  */
283
284 static void tcp_retransmit_timer(struct sock *sk)
285 {
286         struct tcp_sock *tp = tcp_sk(sk);
287         struct inet_connection_sock *icsk = inet_csk(sk);
288
289         if (!tp->packets_out)
290                 goto out;
291
292         BUG_TRAP(!tcp_write_queue_empty(sk));
293
294         if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
295             !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
296                 /* Receiver dastardly shrinks window. Our retransmits
297                  * become zero probes, but we should not timeout this
298                  * connection. If the socket is an orphan, time it out,
299                  * we cannot allow such beasts to hang infinitely.
300                  */
301 #ifdef TCP_DEBUG
302                 if (1) {
303                         struct inet_sock *inet = inet_sk(sk);
304                         LIMIT_NETDEBUG(KERN_DEBUG "TCP: Treason uncloaked! Peer %u.%u.%u.%u:%u/%u shrinks window %u:%u. Repaired.\n",
305                                NIPQUAD(inet->daddr), ntohs(inet->dport),
306                                inet->num, tp->snd_una, tp->snd_nxt);
307                 }
308 #endif
309                 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
310                         tcp_write_err(sk);
311                         goto out;
312                 }
313                 tcp_enter_loss(sk, 0);
314                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
315                 __sk_dst_reset(sk);
316                 goto out_reset_timer;
317         }
318
319         if (tcp_write_timeout(sk))
320                 goto out;
321
322         if (icsk->icsk_retransmits == 0) {
323                 if (icsk->icsk_ca_state == TCP_CA_Disorder ||
324                     icsk->icsk_ca_state == TCP_CA_Recovery) {
325                         if (tcp_is_sack(tp)) {
326                                 if (icsk->icsk_ca_state == TCP_CA_Recovery)
327                                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERYFAIL);
328                                 else
329                                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKFAILURES);
330                         } else {
331                                 if (icsk->icsk_ca_state == TCP_CA_Recovery)
332                                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERYFAIL);
333                                 else
334                                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOFAILURES);
335                         }
336                 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
337                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSFAILURES);
338                 } else {
339                         NET_INC_STATS_BH(LINUX_MIB_TCPTIMEOUTS);
340                 }
341         }
342
343         if (tcp_use_frto(sk)) {
344                 tcp_enter_frto(sk);
345         } else {
346                 tcp_enter_loss(sk, 0);
347         }
348
349         if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
350                 /* Retransmission failed because of local congestion,
351                  * do not backoff.
352                  */
353                 if (!icsk->icsk_retransmits)
354                         icsk->icsk_retransmits = 1;
355                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
356                                           min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
357                                           TCP_RTO_MAX);
358                 goto out;
359         }
360
361         /* Increase the timeout each time we retransmit.  Note that
362          * we do not increase the rtt estimate.  rto is initialized
363          * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
364          * that doubling rto each time is the least we can get away with.
365          * In KA9Q, Karn uses this for the first few times, and then
366          * goes to quadratic.  netBSD doubles, but only goes up to *64,
367          * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
368          * defined in the protocol as the maximum possible RTT.  I guess
369          * we'll have to use something other than TCP to talk to the
370          * University of Mars.
371          *
372          * PAWS allows us longer timeouts and large windows, so once
373          * implemented ftp to mars will work nicely. We will have to fix
374          * the 120 second clamps though!
375          */
376         icsk->icsk_backoff++;
377         icsk->icsk_retransmits++;
378
379 out_reset_timer:
380         icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
381         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
382         if (icsk->icsk_retransmits > sysctl_tcp_retries1)
383                 __sk_dst_reset(sk);
384
385 out:;
386 }
387
388 static void tcp_write_timer(unsigned long data)
389 {
390         struct sock *sk = (struct sock*)data;
391         struct inet_connection_sock *icsk = inet_csk(sk);
392         int event;
393
394         bh_lock_sock(sk);
395         if (sock_owned_by_user(sk)) {
396                 /* Try again later */
397                 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + (HZ / 20));
398                 goto out_unlock;
399         }
400
401         if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
402                 goto out;
403
404         if (time_after(icsk->icsk_timeout, jiffies)) {
405                 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
406                 goto out;
407         }
408
409         event = icsk->icsk_pending;
410         icsk->icsk_pending = 0;
411
412         switch (event) {
413         case ICSK_TIME_RETRANS:
414                 tcp_retransmit_timer(sk);
415                 break;
416         case ICSK_TIME_PROBE0:
417                 tcp_probe_timer(sk);
418                 break;
419         }
420         TCP_CHECK_TIMER(sk);
421
422 out:
423         sk_mem_reclaim(sk);
424 out_unlock:
425         bh_unlock_sock(sk);
426         sock_put(sk);
427 }
428
429 /*
430  *      Timer for listening sockets
431  */
432
433 static void tcp_synack_timer(struct sock *sk)
434 {
435         inet_csk_reqsk_queue_prune(sk, TCP_SYNQ_INTERVAL,
436                                    TCP_TIMEOUT_INIT, TCP_RTO_MAX);
437 }
438
439 void tcp_set_keepalive(struct sock *sk, int val)
440 {
441         if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
442                 return;
443
444         if (val && !sock_flag(sk, SOCK_KEEPOPEN))
445                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
446         else if (!val)
447                 inet_csk_delete_keepalive_timer(sk);
448 }
449
450
451 static void tcp_keepalive_timer (unsigned long data)
452 {
453         struct sock *sk = (struct sock *) data;
454         struct inet_connection_sock *icsk = inet_csk(sk);
455         struct tcp_sock *tp = tcp_sk(sk);
456         __u32 elapsed;
457
458         /* Only process if socket is not in use. */
459         bh_lock_sock(sk);
460         if (sock_owned_by_user(sk)) {
461                 /* Try again later. */
462                 inet_csk_reset_keepalive_timer (sk, HZ/20);
463                 goto out;
464         }
465
466         if (sk->sk_state == TCP_LISTEN) {
467                 tcp_synack_timer(sk);
468                 goto out;
469         }
470
471         if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
472                 if (tp->linger2 >= 0) {
473                         const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
474
475                         if (tmo > 0) {
476                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
477                                 goto out;
478                         }
479                 }
480                 tcp_send_active_reset(sk, GFP_ATOMIC);
481                 goto death;
482         }
483
484         if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
485                 goto out;
486
487         elapsed = keepalive_time_when(tp);
488
489         /* It is alive without keepalive 8) */
490         if (tp->packets_out || tcp_send_head(sk))
491                 goto resched;
492
493         elapsed = tcp_time_stamp - tp->rcv_tstamp;
494
495         if (elapsed >= keepalive_time_when(tp)) {
496                 if ((!tp->keepalive_probes && icsk->icsk_probes_out >= sysctl_tcp_keepalive_probes) ||
497                      (tp->keepalive_probes && icsk->icsk_probes_out >= tp->keepalive_probes)) {
498                         tcp_send_active_reset(sk, GFP_ATOMIC);
499                         tcp_write_err(sk);
500                         goto out;
501                 }
502                 if (tcp_write_wakeup(sk) <= 0) {
503                         icsk->icsk_probes_out++;
504                         elapsed = keepalive_intvl_when(tp);
505                 } else {
506                         /* If keepalive was lost due to local congestion,
507                          * try harder.
508                          */
509                         elapsed = TCP_RESOURCE_PROBE_INTERVAL;
510                 }
511         } else {
512                 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
513                 elapsed = keepalive_time_when(tp) - elapsed;
514         }
515
516         TCP_CHECK_TIMER(sk);
517         sk_mem_reclaim(sk);
518
519 resched:
520         inet_csk_reset_keepalive_timer (sk, elapsed);
521         goto out;
522
523 death:
524         tcp_done(sk);
525
526 out:
527         bh_unlock_sock(sk);
528         sock_put(sk);
529 }