libcfs/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  * libcfs/libcfs/watchdog.c
  37  *
  38  * Author: Jacob Berkman <jacob@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_LNET
  42
  43 #include <libcfs/libcfs.h>
  44 #include "tracefile.h"
  45
  46 struct lc_watchdog {
  47         cfs_timer_t       lcw_timer; /* kernel timer */
  48         struct list_head  lcw_list;
  49         struct timeval    lcw_last_touched;
  50         cfs_task_t       *lcw_task;
  51
  52         void            (*lcw_callback)(pid_t, void *);
  53         void             *lcw_data;
  54
  55         pid_t             lcw_pid;
  56         cfs_duration_t    lcw_time; /* time until watchdog fires, jiffies */
  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 cfs_waitq_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         CFS_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 #ifdef HAVE_TASKLIST_LOCK
 105 static void
 106 lcw_dump(struct lc_watchdog *lcw)
 107 {
 108         cfs_task_t *tsk;
 109         ENTRY;
 110
 111         read_lock(&tasklist_lock);
 112         tsk = find_task_by_pid(lcw->lcw_pid);
 113
 114         if (tsk == NULL) {
 115                 CWARN("Process " LPPID " was not found in the task list; "
 116                       "watchdog callback may be incomplete\n", lcw->lcw_pid);
 117         } else if (tsk != lcw->lcw_task) {
 118                 CWARN("The current process " LPPID " did not set the watchdog; "
 119                       "watchdog callback may be incomplete\n", lcw->lcw_pid);
 120         } else {
 121                 libcfs_debug_dumpstack(tsk);
 122         }
 123
 124         read_unlock(&tasklist_lock);
 125         EXIT;
 126 }
 127 #else
 128 static void
 129 lcw_dump(struct lc_watchdog *lcw)
 130 {
 131         CERROR("unable to dump stack because of missing export\n");
 132 }
 133 #endif
 134
 135 static void lcw_cb(ulong_ptr_t data)
 136 {
 137         struct lc_watchdog *lcw = (struct lc_watchdog *)data;
 138         cfs_time_t current_time;
 139         cfs_duration_t delta_time;
 140
 141         ENTRY;
 142
 143         if (lcw->lcw_state != LC_WATCHDOG_ENABLED) {
 144                 EXIT;
 145                 return;
 146         }
 147
 148         lcw->lcw_state = LC_WATCHDOG_EXPIRED;
 149         current_time = cfs_time_current();
 150
 151         /* Check to see if we should throttle the watchdog timer to avoid
 152          * too many dumps going to the console thus triggering an NMI.
 153          * Normally we would not hold the spin lock over the CWARN but in
 154          * this case we hold it to ensure non ratelimited lcw_dumps are not
 155          * interleaved on the console making them hard to read. */
 156         spin_lock_bh(&lcw_last_watchdog_lock);
 157         delta_time = cfs_duration_sec(current_time - lcw_last_watchdog_time);
 158
 159         if (delta_time < libcfs_watchdog_ratelimit && lcw_recent_watchdog_count > 3) {
 160                 CWARN("Refusing to fire watchdog for pid %d: it was inactive "
 161                       "for %ldms. Rate limiting 1 per %d seconds.\n",
 162                       (int)lcw->lcw_pid,cfs_duration_sec(lcw->lcw_time) * 1000,
 163                       libcfs_watchdog_ratelimit);
 164         } else {
 165                 if (delta_time < libcfs_watchdog_ratelimit) {
 166                         lcw_recent_watchdog_count++;
 167                 } else {
 168                         memcpy(&lcw_last_watchdog_time, &current_time,
 169                                sizeof(current_time));
 170                         lcw_recent_watchdog_count = 0;
 171                 }
 172
 173                 /* This warning should appear on the console, but may not get
 174                  * into the logs since we're running in a softirq handler */
 175                 CWARN("Watchdog triggered for pid %d: it was inactive for %lds\n",
 176                       (int)lcw->lcw_pid, cfs_duration_sec(lcw->lcw_time));
 177                 lcw_dump(lcw);
 178         }
 179
 180         spin_unlock_bh(&lcw_last_watchdog_lock);
 181         spin_lock_bh(&lcw_pending_timers_lock);
 182
 183         if (list_empty(&lcw->lcw_list)) {
 184                 list_add(&lcw->lcw_list, &lcw_pending_timers);
 185                 cfs_waitq_signal(&lcw_event_waitq);
 186         }
 187
 188         spin_unlock_bh(&lcw_pending_timers_lock);
 189
 190         EXIT;
 191 }
 192
 193 static int is_watchdog_fired(void)
 194 {
 195         int rc;
 196
 197         if (test_bit(LCW_FLAG_STOP, &lcw_flags))
 198                 return 1;
 199
 200         spin_lock_bh(&lcw_pending_timers_lock);
 201         rc = !list_empty(&lcw_pending_timers);
 202         spin_unlock_bh(&lcw_pending_timers_lock);
 203         return rc;
 204 }
 205
 206 static int lcw_dispatch_main(void *data)
 207 {
 208         int                 rc = 0;
 209         unsigned long       flags;
 210         struct lc_watchdog *lcw;
 211
 212         ENTRY;
 213
 214         cfs_daemonize("lc_watchdogd");
 215
 216         SIGNAL_MASK_LOCK(current, flags);
 217         sigfillset(&current->blocked);
 218         RECALC_SIGPENDING;
 219         SIGNAL_MASK_UNLOCK(current, flags);
 220
 221         complete(&lcw_start_completion);
 222
 223         while (1) {
 224                 cfs_wait_event_interruptible(lcw_event_waitq, is_watchdog_fired(), rc);
 225                 CDEBUG(D_INFO, "Watchdog got woken up...\n");
 226                 if (test_bit(LCW_FLAG_STOP, &lcw_flags)) {
 227                         CDEBUG(D_INFO, "LCW_FLAG_STOP was set, shutting down...\n");
 228
 229                         spin_lock_bh(&lcw_pending_timers_lock);
 230                         rc = !list_empty(&lcw_pending_timers);
 231                         spin_unlock_bh(&lcw_pending_timers_lock);
 232                         if (rc) {
 233                                 CERROR("pending timers list was not empty at "
 234                                        "time of watchdog dispatch shutdown\n");
 235                         }
 236                         break;
 237                 }
 238
 239                 spin_lock_bh(&lcw_pending_timers_lock);
 240                 while (!list_empty(&lcw_pending_timers)) {
 241
 242                         lcw = list_entry(lcw_pending_timers.next,
 243                                          struct lc_watchdog,
 244                                          lcw_list);
 245                         list_del_init(&lcw->lcw_list);
 246                         spin_unlock_bh(&lcw_pending_timers_lock);
 247
 248                         CDEBUG(D_INFO, "found lcw for pid " LPPID ": inactive for "
 249                                "%lds\n", lcw->lcw_pid, cfs_duration_sec(lcw->lcw_time));
 250
 251                         if (lcw->lcw_state != LC_WATCHDOG_DISABLED)
 252                                 lcw->lcw_callback(lcw->lcw_pid, lcw->lcw_data);
 253
 254                         spin_lock_bh(&lcw_pending_timers_lock);
 255                 }
 256                 spin_unlock_bh(&lcw_pending_timers_lock);
 257         }
 258
 259         complete(&lcw_stop_completion);
 260
 261         RETURN(rc);
 262 }
 263
 264 static void lcw_dispatch_start(void)
 265 {
 266         int rc;
 267
 268         ENTRY;
 269         LASSERT(lcw_refcount == 1);
 270
 271         init_completion(&lcw_stop_completion);
 272         init_completion(&lcw_start_completion);
 273         cfs_waitq_init(&lcw_event_waitq);
 274
 275         CDEBUG(D_INFO, "starting dispatch thread\n");
 276         rc = kernel_thread(lcw_dispatch_main, NULL, 0);
 277         if (rc < 0) {
 278                 CERROR("error spawning watchdog dispatch thread: %d\n", rc);
 279                 EXIT;
 280                 return;
 281         }
 282         wait_for_completion(&lcw_start_completion);
 283         CDEBUG(D_INFO, "watchdog dispatcher initialization complete.\n");
 284
 285         EXIT;
 286 }
 287
 288 static void lcw_dispatch_stop(void)
 289 {
 290         ENTRY;
 291         LASSERT(lcw_refcount == 0);
 292
 293         CDEBUG(D_INFO, "trying to stop watchdog dispatcher.\n");
 294
 295         set_bit(LCW_FLAG_STOP, &lcw_flags);
 296         cfs_waitq_signal(&lcw_event_waitq);
 297
 298         wait_for_completion(&lcw_stop_completion);
 299
 300         CDEBUG(D_INFO, "watchdog dispatcher has shut down.\n");
 301
 302         EXIT;
 303 }
 304
 305 struct lc_watchdog *lc_watchdog_add(int timeout_ms,
 306                                     void (*callback)(pid_t, void *),
 307                                     void *data)
 308 {
 309         struct lc_watchdog *lcw = NULL;
 310         ENTRY;
 311
 312         LIBCFS_ALLOC(lcw, sizeof(*lcw));
 313         if (lcw == NULL) {
 314                 CDEBUG(D_INFO, "Could not allocate new lc_watchdog\n");
 315                 RETURN(ERR_PTR(-ENOMEM));
 316         }
 317
 318         lcw->lcw_task     = cfs_current();
 319         lcw->lcw_pid      = cfs_curproc_pid();
 320         lcw->lcw_time     = cfs_time_seconds(timeout_ms) / 1000;
 321         lcw->lcw_callback = (callback != NULL) ? callback : lc_watchdog_dumplog;
 322         lcw->lcw_data     = data;
 323         lcw->lcw_state    = LC_WATCHDOG_DISABLED;
 324
 325         CFS_INIT_LIST_HEAD(&lcw->lcw_list);
 326         cfs_timer_init(&lcw->lcw_timer, lcw_cb, lcw);
 327
 328         down(&lcw_refcount_sem);
 329         if (++lcw_refcount == 1)
 330                 lcw_dispatch_start();
 331         up(&lcw_refcount_sem);
 332
 333         /* Keep this working in case we enable them by default */
 334         if (lcw->lcw_state == LC_WATCHDOG_ENABLED) {
 335                 do_gettimeofday(&lcw->lcw_last_touched);
 336                 cfs_timer_arm(&lcw->lcw_timer, lcw->lcw_time +
 337                               cfs_time_current());
 338         }
 339
 340         RETURN(lcw);
 341 }
 342 EXPORT_SYMBOL(lc_watchdog_add);
 343
 344 static void lcw_update_time(struct lc_watchdog *lcw, const char *message)
 345 {
 346         struct timeval newtime;
 347         struct timeval timediff;
 348
 349         do_gettimeofday(&newtime);
 350         if (lcw->lcw_state == LC_WATCHDOG_EXPIRED) {
 351                 cfs_timeval_sub(&newtime, &lcw->lcw_last_touched, &timediff);
 352                 CWARN("Expired watchdog for pid " LPPID " %s after %lu.%.4lus\n",
 353                       lcw->lcw_pid,
 354                       message,
 355                       timediff.tv_sec,
 356                       timediff.tv_usec / 100);
 357         }
 358         lcw->lcw_last_touched = newtime;
 359 }
 360
 361 void lc_watchdog_touch_ms(struct lc_watchdog *lcw, int timeout_ms)
 362 {
 363         ENTRY;
 364         LASSERT(lcw != NULL);
 365
 366         spin_lock_bh(&lcw_pending_timers_lock);
 367         list_del_init(&lcw->lcw_list);
 368         spin_unlock_bh(&lcw_pending_timers_lock);
 369
 370         lcw_update_time(lcw, "touched");
 371         lcw->lcw_state = LC_WATCHDOG_ENABLED;
 372
 373         cfs_timer_arm(&lcw->lcw_timer, cfs_time_current() +
 374                       cfs_time_seconds(timeout_ms) / 1000);
 375
 376         EXIT;
 377 }
 378 EXPORT_SYMBOL(lc_watchdog_touch_ms);
 379
 380 /* deprecated - use above instead */
 381 void lc_watchdog_touch(struct lc_watchdog *lcw)
 382 {
 383         lc_watchdog_touch_ms(lcw, (int)cfs_duration_sec(lcw->lcw_time) * 1000);
 384 }
 385 EXPORT_SYMBOL(lc_watchdog_touch);
 386
 387 void lc_watchdog_disable(struct lc_watchdog *lcw)
 388 {
 389         ENTRY;
 390         LASSERT(lcw != NULL);
 391
 392         spin_lock_bh(&lcw_pending_timers_lock);
 393         if (!list_empty(&lcw->lcw_list))
 394                 list_del_init(&lcw->lcw_list);
 395         spin_unlock_bh(&lcw_pending_timers_lock);
 396
 397         lcw_update_time(lcw, "disabled");
 398         lcw->lcw_state = LC_WATCHDOG_DISABLED;
 399
 400         EXIT;
 401 }
 402 EXPORT_SYMBOL(lc_watchdog_disable);
 403
 404 void lc_watchdog_delete(struct lc_watchdog *lcw)
 405 {
 406         ENTRY;
 407         LASSERT(lcw != NULL);
 408
 409         cfs_timer_disarm(&lcw->lcw_timer);
 410
 411         lcw_update_time(lcw, "deleted");
 412
 413         spin_lock_bh(&lcw_pending_timers_lock);
 414         if (!list_empty(&lcw->lcw_list))
 415                 list_del_init(&lcw->lcw_list);
 416         spin_unlock_bh(&lcw_pending_timers_lock);
 417
 418         down(&lcw_refcount_sem);
 419         if (--lcw_refcount == 0)
 420                 lcw_dispatch_stop();
 421         up(&lcw_refcount_sem);
 422
 423         LIBCFS_FREE(lcw, sizeof(*lcw));
 424
 425         EXIT;
 426 }
 427 EXPORT_SYMBOL(lc_watchdog_delete);
 428
 429 /*
 430  * Provided watchdog handlers
 431  */
 432
 433 void lc_watchdog_dumplog(pid_t pid, void *data)
 434 {
 435         libcfs_debug_dumplog_internal((void *)((long_ptr_t)pid));
 436 }
 437 EXPORT_SYMBOL(lc_watchdog_dumplog);
 438
 439 #else   /* !defined(WITH_WATCHDOG) */
 440
 441 struct lc_watchdog *lc_watchdog_add(int timeout_ms,
 442                                     void (*callback)(pid_t pid, void *),
 443                                     void *data)
 444 {
 445         static struct lc_watchdog      watchdog;
 446         return &watchdog;
 447 }
 448 EXPORT_SYMBOL(lc_watchdog_add);
 449
 450 void lc_watchdog_touch_ms(struct lc_watchdog *lcw, int timeout_ms)
 451 {
 452 }
 453 EXPORT_SYMBOL(lc_watchdog_touch_ms);
 454
 455 void lc_watchdog_touch(struct lc_watchdog *lcw)
 456 {
 457 }
 458 EXPORT_SYMBOL(lc_watchdog_touch);
 459
 460 void lc_watchdog_disable(struct lc_watchdog *lcw)
 461 {
 462 }
 463 EXPORT_SYMBOL(lc_watchdog_disable);
 464
 465 void lc_watchdog_delete(struct lc_watchdog *lcw)
 466 {
 467 }
 468 EXPORT_SYMBOL(lc_watchdog_delete);
 469
 470 #endif