lnet/libcfs/watchdog.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/libcfs/watchdog.c
  37  *
  38  * Author: Jacob Berkman <jacob@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_LNET
  42
  43 #include <libcfs/kp30.h>
  44 #include <libcfs/libcfs.h>
  45 #include "tracefile.h"
  46
  47 struct lc_watchdog {
  48         cfs_timer_t       lcw_timer; /* kernel timer */
  49         struct list_head  lcw_list;
  50         cfs_time_t        lcw_last_touched;
  51         cfs_task_t       *lcw_task;
  52
  53         void            (*lcw_callback)(pid_t, void *);
  54         void             *lcw_data;
  55
  56         pid_t             lcw_pid;
  57
  58         enum {
  59                 LC_WATCHDOG_DISABLED,
  60                 LC_WATCHDOG_ENABLED,
  61                 LC_WATCHDOG_EXPIRED
  62         } lcw_state;
  63 };
  64
  65 #ifdef WITH_WATCHDOG
  66 /*
  67  * The dispatcher will complete lcw_start_completion when it starts,
  68  * and lcw_stop_completion when it exits.
  69  * Wake lcw_event_waitq to signal timer callback dispatches.
  70  */
  71 static struct completion lcw_start_completion;
  72 static struct completion lcw_stop_completion;
  73 static wait_queue_head_t lcw_event_waitq;
  74
  75 /*
  76  * Set this and wake lcw_event_waitq to stop the dispatcher.
  77  */
  78 enum {
  79         LCW_FLAG_STOP = 0
  80 };
  81 static unsigned long lcw_flags = 0;
  82
  83 /*
  84  * Number of outstanding watchdogs.
  85  * When it hits 1, we start the dispatcher.
  86  * When it hits 0, we stop the distpatcher.
  87  */
  88 static __u32         lcw_refcount = 0;
  89 static DECLARE_MUTEX(lcw_refcount_sem);
  90
  91 /*
  92  * List of timers that have fired that need their callbacks run by the
  93  * dispatcher.
  94  */
  95 static spinlock_t lcw_pending_timers_lock = SPIN_LOCK_UNLOCKED; /* BH lock! */
  96 static struct list_head lcw_pending_timers = \
  97         LIST_HEAD_INIT(lcw_pending_timers);
  98
  99 /* Last time a watchdog expired */
 100 static cfs_time_t lcw_last_watchdog_time;
 101 static int lcw_recent_watchdog_count;
 102 static spinlock_t lcw_last_watchdog_lock = SPIN_LOCK_UNLOCKED;
 103
 104 static void
 105 lcw_dump(struct lc_watchdog *lcw)
 106 {
 107         ENTRY;
 108
 109 #if defined(HAVE_TASKLIST_LOCK)
 110         read_lock(&tasklist_lock);
 111 #elif defined(HAVE_TASK_RCU)
 112         rcu_read_lock();
 113 #else
 114         CERROR("unable to dump stack because of missing export\n");
 115         RETURN_EXIT;
 116 #endif
 117         if (lcw->lcw_task == NULL) {
 118                 CWARN("Process %d was not found in the task list; "
 119                       "watchdog callback may be incomplete\n",
 120                       (int)lcw->lcw_pid);
 121         } else {
 122                 libcfs_debug_dumpstack(lcw->lcw_task);
 123         }
 124
 125 #if defined(HAVE_TASKLIST_LOCK)
 126         read_unlock(&tasklist_lock);
 127 #elif defined(HAVE_TASK_RCU)
 128         rcu_read_unlock();
 129 #endif
 130         EXIT;
 131 }
 132
 133 static void lcw_cb(unsigned long data)
 134 {
 135         struct lc_watchdog *lcw = (struct lc_watchdog *)data;
 136         cfs_time_t current_time;
 137         cfs_duration_t delta_time;
 138         struct timeval timediff;
 139
 140         ENTRY;
 141
 142         if (lcw->lcw_state != LC_WATCHDOG_ENABLED) {
 143                 EXIT;
 144                 return;
 145         }
 146
 147         lcw->lcw_state = LC_WATCHDOG_EXPIRED;
 148         current_time = cfs_time_current();
 149
 150         delta_time = cfs_time_sub(current_time, lcw->lcw_last_touched);
 151         cfs_duration_usec(delta_time, &timediff);
 152
 153         /* Check to see if we should throttle the watchdog timer to avoid
 154          * too many dumps going to the console thus triggering an NMI.
 155          * Normally we would not hold the spin lock over the CWARN but in
 156          * this case we hold it to ensure non ratelimited lcw_dumps are not
 157          * interleaved on the console making them hard to read. */
 158         spin_lock_bh(&lcw_last_watchdog_lock);
 159         delta_time = cfs_duration_sec(cfs_time_sub(current_time,
 160                                                    lcw_last_watchdog_time));
 161
 162         if (delta_time < libcfs_watchdog_ratelimit &&
 163             lcw_recent_watchdog_count > 3) {
 164                 CWARN("Refusing to fire watchdog for pid %d: it was inactive "
 165                       "for %lu.%.02lus. Rate limiting 1 per %d seconds.\n",
 166                       (int)lcw->lcw_pid, timediff.tv_sec,
 167                       timediff.tv_usec / 10000, libcfs_watchdog_ratelimit);
 168         } else {
 169                 if (delta_time < libcfs_watchdog_ratelimit) {
 170                         lcw_recent_watchdog_count++;
 171                 } else {
 172                         memcpy(&lcw_last_watchdog_time, &current_time,
 173                                sizeof(current_time));
 174                         lcw_recent_watchdog_count = 0;
 175                 }
 176
 177                 /* This warning should appear on the console, but may not get
 178                  * into the logs since we're running in a softirq handler */
 179                 CWARN("Watchdog triggered for pid %d: it was inactive for "
 180                       "%lu.%.02lus\n", (int)lcw->lcw_pid, timediff.tv_sec,
 181                       timediff.tv_usec / 10000);
 182                 lcw_dump(lcw);
 183         }
 184
 185         spin_unlock_bh(&lcw_last_watchdog_lock);
 186         spin_lock_bh(&lcw_pending_timers_lock);
 187
 188         if (list_empty(&lcw->lcw_list)) {
 189                 list_add(&lcw->lcw_list, &lcw_pending_timers);
 190                 wake_up(&lcw_event_waitq);
 191         }
 192
 193         spin_unlock_bh(&lcw_pending_timers_lock);
 194
 195         EXIT;
 196 }
 197
 198 static int is_watchdog_fired(void)
 199 {
 200         int rc;
 201
 202         if (test_bit(LCW_FLAG_STOP, &lcw_flags))
 203                 return 1;
 204
 205         spin_lock_bh(&lcw_pending_timers_lock);
 206         rc = !list_empty(&lcw_pending_timers);
 207         spin_unlock_bh(&lcw_pending_timers_lock);
 208         return rc;
 209 }
 210
 211 static int lcw_dispatch_main(void *data)
 212 {
 213         int                 rc = 0;
 214         unsigned long       flags;
 215         struct lc_watchdog *lcw;
 216
 217         ENTRY;
 218
 219         cfs_daemonize("lc_watchdogd");
 220
 221         SIGNAL_MASK_LOCK(current, flags);
 222         sigfillset(&current->blocked);
 223         RECALC_SIGPENDING;
 224         SIGNAL_MASK_UNLOCK(current, flags);
 225
 226         complete(&lcw_start_completion);
 227
 228         while (1) {
 229                 wait_event_interruptible(lcw_event_waitq, is_watchdog_fired());
 230                 CDEBUG(D_INFO, "Watchdog got woken up...\n");
 231                 if (test_bit(LCW_FLAG_STOP, &lcw_flags)) {
 232                         CDEBUG(D_INFO, "LCW_FLAG_STOP was set, shutting down...\n");
 233
 234                         spin_lock_bh(&lcw_pending_timers_lock);
 235                         rc = !list_empty(&lcw_pending_timers);
 236                         spin_unlock_bh(&lcw_pending_timers_lock);
 237                         if (rc) {
 238                                 CERROR("pending timers list was not empty at "
 239                                        "time of watchdog dispatch shutdown\n");
 240                         }
 241                         break;
 242                 }
 243
 244                 spin_lock_bh(&lcw_pending_timers_lock);
 245                 while (!list_empty(&lcw_pending_timers)) {
 246
 247                         lcw = list_entry(lcw_pending_timers.next,
 248                                          struct lc_watchdog,
 249                                          lcw_list);
 250                         list_del_init(&lcw->lcw_list);
 251                         spin_unlock_bh(&lcw_pending_timers_lock);
 252
 253                         CDEBUG(D_INFO, "found lcw for pid %d\n",
 254                                (int)lcw->lcw_pid);
 255
 256                         if (lcw->lcw_state != LC_WATCHDOG_DISABLED)
 257                                 lcw->lcw_callback(lcw->lcw_pid, lcw->lcw_data);
 258
 259                         spin_lock_bh(&lcw_pending_timers_lock);
 260                 }
 261                 spin_unlock_bh(&lcw_pending_timers_lock);
 262         }
 263
 264         complete(&lcw_stop_completion);
 265
 266         RETURN(rc);
 267 }
 268
 269 static void lcw_dispatch_start(void)
 270 {
 271         int rc;
 272
 273         ENTRY;
 274         LASSERT(lcw_refcount == 1);
 275
 276         init_completion(&lcw_stop_completion);
 277         init_completion(&lcw_start_completion);
 278         init_waitqueue_head(&lcw_event_waitq);
 279
 280         CDEBUG(D_INFO, "starting dispatch thread\n");
 281         rc = kernel_thread(lcw_dispatch_main, NULL, 0);
 282         if (rc < 0) {
 283                 CERROR("error spawning watchdog dispatch thread: %d\n", rc);
 284                 EXIT;
 285                 return;
 286         }
 287         wait_for_completion(&lcw_start_completion);
 288         CDEBUG(D_INFO, "watchdog dispatcher initialization complete.\n");
 289
 290         EXIT;
 291 }
 292
 293 static void lcw_dispatch_stop(void)
 294 {
 295         ENTRY;
 296         LASSERT(lcw_refcount == 0);
 297
 298         CDEBUG(D_INFO, "trying to stop watchdog dispatcher.\n");
 299
 300         set_bit(LCW_FLAG_STOP, &lcw_flags);
 301         wake_up(&lcw_event_waitq);
 302
 303         wait_for_completion(&lcw_stop_completion);
 304
 305         CDEBUG(D_INFO, "watchdog dispatcher has shut down.\n");
 306
 307         EXIT;
 308 }
 309
 310 struct lc_watchdog *lc_watchdog_add(int timeout,
 311                                     void (*callback)(pid_t, void *),
 312                                     void *data)
 313 {
 314         struct lc_watchdog *lcw = NULL;
 315         ENTRY;
 316
 317         LIBCFS_ALLOC(lcw, sizeof(*lcw));
 318         if (lcw == NULL) {
 319                 CDEBUG(D_INFO, "Could not allocate new lc_watchdog\n");
 320                 RETURN(ERR_PTR(-ENOMEM));
 321         }
 322
 323         lcw->lcw_task     = cfs_current();
 324         lcw->lcw_pid      = cfs_curproc_pid();
 325         lcw->lcw_callback = (callback != NULL) ? callback : lc_watchdog_dumplog;
 326         lcw->lcw_data     = data;
 327         lcw->lcw_state    = LC_WATCHDOG_DISABLED;
 328
 329         INIT_LIST_HEAD(&lcw->lcw_list);
 330
 331         lcw->lcw_timer.function = lcw_cb;
 332         lcw->lcw_timer.data = (unsigned long)lcw;
 333         lcw->lcw_timer.expires = jiffies + cfs_time_seconds(timeout);
 334         init_timer(&lcw->lcw_timer);
 335
 336         down(&lcw_refcount_sem);
 337         if (++lcw_refcount == 1)
 338                 lcw_dispatch_start();
 339         up(&lcw_refcount_sem);
 340
 341         /* Keep this working in case we enable them by default */
 342         if (lcw->lcw_state == LC_WATCHDOG_ENABLED) {
 343                 lcw->lcw_last_touched = cfs_time_current();
 344                 add_timer(&lcw->lcw_timer);
 345         }
 346
 347         RETURN(lcw);
 348 }
 349 EXPORT_SYMBOL(lc_watchdog_add);
 350
 351 static void lcw_update_time(struct lc_watchdog *lcw, const char *message)
 352 {
 353         cfs_time_t newtime = cfs_time_current();;
 354
 355         if (lcw->lcw_state == LC_WATCHDOG_EXPIRED) {
 356                 struct timeval timediff;
 357                 cfs_time_t delta_time = cfs_time_sub(newtime,
 358                                                      lcw->lcw_last_touched);
 359                 cfs_duration_usec(delta_time, &timediff);
 360
 361                 CWARN("Expired watchdog for pid %d %s after %lu.%.02lus\n",
 362                       lcw->lcw_pid, message, timediff.tv_sec,
 363                       timediff.tv_usec / 10000);
 364         }
 365         lcw->lcw_last_touched = newtime;
 366 }
 367
 368 void lc_watchdog_touch(struct lc_watchdog *lcw, int timeout)
 369 {
 370         ENTRY;
 371         LASSERT(lcw != NULL);
 372
 373         spin_lock_bh(&lcw_pending_timers_lock);
 374         list_del_init(&lcw->lcw_list);
 375         spin_unlock_bh(&lcw_pending_timers_lock);
 376
 377         lcw_update_time(lcw, "touched");
 378         lcw->lcw_state = LC_WATCHDOG_ENABLED;
 379
 380         mod_timer(&lcw->lcw_timer, jiffies + cfs_time_seconds(timeout));
 381
 382         EXIT;
 383 }
 384 EXPORT_SYMBOL(lc_watchdog_touch);
 385
 386 void lc_watchdog_disable(struct lc_watchdog *lcw)
 387 {
 388         ENTRY;
 389         LASSERT(lcw != NULL);
 390
 391         spin_lock_bh(&lcw_pending_timers_lock);
 392         if (!list_empty(&lcw->lcw_list))
 393                 list_del_init(&lcw->lcw_list);
 394         spin_unlock_bh(&lcw_pending_timers_lock);
 395
 396         lcw_update_time(lcw, "disabled");
 397         lcw->lcw_state = LC_WATCHDOG_DISABLED;
 398
 399         EXIT;
 400 }
 401 EXPORT_SYMBOL(lc_watchdog_disable);
 402
 403 void lc_watchdog_delete(struct lc_watchdog *lcw)
 404 {
 405         ENTRY;
 406         LASSERT(lcw != NULL);
 407
 408         del_timer(&lcw->lcw_timer);
 409
 410         lcw_update_time(lcw, "deleted");
 411
 412         spin_lock_bh(&lcw_pending_timers_lock);
 413         if (!list_empty(&lcw->lcw_list))
 414                 list_del_init(&lcw->lcw_list);
 415         spin_unlock_bh(&lcw_pending_timers_lock);
 416
 417         down(&lcw_refcount_sem);
 418         if (--lcw_refcount == 0)
 419                 lcw_dispatch_stop();
 420         up(&lcw_refcount_sem);
 421
 422         LIBCFS_FREE(lcw, sizeof(*lcw));
 423
 424         EXIT;
 425 }
 426 EXPORT_SYMBOL(lc_watchdog_delete);
 427
 428 /*
 429  * Provided watchdog handlers
 430  */
 431
 432 void lc_watchdog_dumplog(pid_t pid, void *data)
 433 {
 434         libcfs_debug_dumplog_internal((void *)((unsigned long)pid));
 435 }
 436 EXPORT_SYMBOL(lc_watchdog_dumplog);
 437
 438 #else   /* !defined(WITH_WATCHDOG) */
 439
 440 struct lc_watchdog *lc_watchdog_add(int timeout,
 441                                     void (*callback)(pid_t pid, void *),
 442                                     void *data)
 443 {
 444         static struct lc_watchdog      watchdog;
 445         return &watchdog;
 446 }
 447 EXPORT_SYMBOL(lc_watchdog_add);
 448
 449 void lc_watchdog_touch(struct lc_watchdog *lcw, int timeout)
 450 {
 451 }
 452 EXPORT_SYMBOL(lc_watchdog_touch);
 453
 454 void lc_watchdog_disable(struct lc_watchdog *lcw)
 455 {
 456 }
 457 EXPORT_SYMBOL(lc_watchdog_disable);
 458
 459 void lc_watchdog_delete(struct lc_watchdog *lcw)
 460 {
 461 }
 462 EXPORT_SYMBOL(lc_watchdog_delete);
 463
 464 #endif