libcfs/libcfs/watchdog.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, Intel Corporation.
  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         spinlock_t  lcw_lock;     /* check or change lcw_list */
  48         int             lcw_refcount; /* must hold lcw_pending_timers_lock */
  49         struct timer_list     lcw_timer;    /* kernel timer */
  50         cfs_list_t      lcw_list;     /* chain on pending list */
  51         cfs_time_t      lcw_last_touched; /* last touched stamp */
  52         struct task_struct     *lcw_task;     /* owner task */
  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 dispatcher.
  87  */
  88 static __u32         lcw_refcount = 0;
  89 static DEFINE_MUTEX(lcw_refcount_mutex);
  90
  91 /*
  92  * List of timers that have fired that need their callbacks run by the
  93  * dispatcher.
  94  */
  95 /* BH lock! */
  96 static DEFINE_SPINLOCK(lcw_pending_timers_lock);
  97 static cfs_list_t lcw_pending_timers = 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
 103 static void
 104 lcw_dump(struct lc_watchdog *lcw)
 105 {
 106         ENTRY;
 107         rcu_read_lock();
 108        if (lcw->lcw_task == NULL) {
 109                 LCONSOLE_WARN("Process " LPPID " was not found in the task "
 110                               "list; watchdog callback may be incomplete\n",
 111                               (int)lcw->lcw_pid);
 112         } else {
 113                 libcfs_debug_dumpstack(lcw->lcw_task);
 114         }
 115
 116         rcu_read_unlock();
 117         EXIT;
 118 }
 119
 120 static void lcw_cb(ulong_ptr_t data)
 121 {
 122         struct lc_watchdog *lcw = (struct lc_watchdog *)data;
 123         ENTRY;
 124
 125         if (lcw->lcw_state != LC_WATCHDOG_ENABLED) {
 126                 EXIT;
 127                 return;
 128         }
 129
 130         lcw->lcw_state = LC_WATCHDOG_EXPIRED;
 131
 132         spin_lock_bh(&lcw->lcw_lock);
 133         LASSERT(cfs_list_empty(&lcw->lcw_list));
 134
 135         spin_lock_bh(&lcw_pending_timers_lock);
 136         lcw->lcw_refcount++; /* +1 for pending list */
 137         cfs_list_add(&lcw->lcw_list, &lcw_pending_timers);
 138         wake_up(&lcw_event_waitq);
 139
 140         spin_unlock_bh(&lcw_pending_timers_lock);
 141         spin_unlock_bh(&lcw->lcw_lock);
 142         EXIT;
 143 }
 144
 145 static int is_watchdog_fired(void)
 146 {
 147         int rc;
 148
 149         if (test_bit(LCW_FLAG_STOP, &lcw_flags))
 150                 return 1;
 151
 152         spin_lock_bh(&lcw_pending_timers_lock);
 153         rc = !cfs_list_empty(&lcw_pending_timers);
 154         spin_unlock_bh(&lcw_pending_timers_lock);
 155         return rc;
 156 }
 157
 158 static void lcw_dump_stack(struct lc_watchdog *lcw)
 159 {
 160         cfs_time_t      current_time;
 161         cfs_duration_t  delta_time;
 162         struct timeval  timediff;
 163
 164         current_time = cfs_time_current();
 165         delta_time = cfs_time_sub(current_time, lcw->lcw_last_touched);
 166         cfs_duration_usec(delta_time, &timediff);
 167
 168         /*
 169          * Check to see if we should throttle the watchdog timer to avoid
 170          * too many dumps going to the console thus triggering an NMI.
 171          */
 172         delta_time = cfs_duration_sec(cfs_time_sub(current_time,
 173                                                    lcw_last_watchdog_time));
 174
 175         if (delta_time < libcfs_watchdog_ratelimit &&
 176             lcw_recent_watchdog_count > 3) {
 177                 LCONSOLE_WARN("Service thread pid %u was inactive for "
 178                               "%lu.%.02lus. Watchdog stack traces are limited "
 179                               "to 3 per %d seconds, skipping this one.\n",
 180                               (int)lcw->lcw_pid,
 181                               timediff.tv_sec,
 182                               timediff.tv_usec / 10000,
 183                               libcfs_watchdog_ratelimit);
 184         } else {
 185                 if (delta_time < libcfs_watchdog_ratelimit) {
 186                         lcw_recent_watchdog_count++;
 187                 } else {
 188                         memcpy(&lcw_last_watchdog_time, &current_time,
 189                                sizeof(current_time));
 190                         lcw_recent_watchdog_count = 0;
 191                 }
 192
 193                 LCONSOLE_WARN("Service thread pid %u was inactive for "
 194                               "%lu.%.02lus. The thread might be hung, or it "
 195                               "might only be slow and will resume later. "
 196                               "Dumping the stack trace for debugging purposes:"
 197                               "\n",
 198                               (int)lcw->lcw_pid,
 199                               timediff.tv_sec,
 200                               timediff.tv_usec / 10000);
 201                 lcw_dump(lcw);
 202         }
 203 }
 204
 205 static int lcw_dispatch_main(void *data)
 206 {
 207         int                 rc = 0;
 208         struct lc_watchdog *lcw;
 209         CFS_LIST_HEAD      (zombies);
 210
 211         ENTRY;
 212
 213         complete(&lcw_start_completion);
 214
 215         while (1) {
 216                 int dumplog = 1;
 217
 218                 rc = wait_event_interruptible(lcw_event_waitq,
 219                                               is_watchdog_fired());
 220                 CDEBUG(D_INFO, "Watchdog got woken up...\n");
 221                 if (test_bit(LCW_FLAG_STOP, &lcw_flags)) {
 222                         CDEBUG(D_INFO, "LCW_FLAG_STOP set, shutting down...\n");
 223
 224                         spin_lock_bh(&lcw_pending_timers_lock);
 225                         rc = !cfs_list_empty(&lcw_pending_timers);
 226                         spin_unlock_bh(&lcw_pending_timers_lock);
 227                         if (rc) {
 228                                 CERROR("pending timers list was not empty at "
 229                                        "time of watchdog dispatch shutdown\n");
 230                         }
 231                         break;
 232                 }
 233
 234                 spin_lock_bh(&lcw_pending_timers_lock);
 235                 while (!cfs_list_empty(&lcw_pending_timers)) {
 236                         int is_dumplog;
 237
 238                         lcw = cfs_list_entry(lcw_pending_timers.next,
 239                                              struct lc_watchdog, lcw_list);
 240                         /* +1 ref for callback to make sure lwc wouldn't be
 241                          * deleted after releasing lcw_pending_timers_lock */
 242                         lcw->lcw_refcount++;
 243                         spin_unlock_bh(&lcw_pending_timers_lock);
 244
 245                         /* lock ordering */
 246                         spin_lock_bh(&lcw->lcw_lock);
 247                         spin_lock_bh(&lcw_pending_timers_lock);
 248
 249                         if (cfs_list_empty(&lcw->lcw_list)) {
 250                                 /* already removed from pending list */
 251                                 lcw->lcw_refcount--; /* -1 ref for callback */
 252                                 if (lcw->lcw_refcount == 0)
 253                                         cfs_list_add(&lcw->lcw_list, &zombies);
 254                                 spin_unlock_bh(&lcw->lcw_lock);
 255                                 /* still hold lcw_pending_timers_lock */
 256                                 continue;
 257                         }
 258
 259                         cfs_list_del_init(&lcw->lcw_list);
 260                         lcw->lcw_refcount--; /* -1 ref for pending list */
 261
 262                         spin_unlock_bh(&lcw_pending_timers_lock);
 263                         spin_unlock_bh(&lcw->lcw_lock);
 264
 265                         CDEBUG(D_INFO, "found lcw for pid " LPPID "\n",
 266                                lcw->lcw_pid);
 267                         lcw_dump_stack(lcw);
 268
 269                         is_dumplog = lcw->lcw_callback == lc_watchdog_dumplog;
 270                         if (lcw->lcw_state != LC_WATCHDOG_DISABLED &&
 271                             (dumplog || !is_dumplog)) {
 272                                 lcw->lcw_callback(lcw->lcw_pid, lcw->lcw_data);
 273                                 if (dumplog && is_dumplog)
 274                                         dumplog = 0;
 275                         }
 276
 277                         spin_lock_bh(&lcw_pending_timers_lock);
 278                         lcw->lcw_refcount--; /* -1 ref for callback */
 279                         if (lcw->lcw_refcount == 0)
 280                                 cfs_list_add(&lcw->lcw_list, &zombies);
 281                 }
 282                 spin_unlock_bh(&lcw_pending_timers_lock);
 283
 284                 while (!cfs_list_empty(&zombies)) {
 285                         lcw = cfs_list_entry(lcw_pending_timers.next,
 286                                          struct lc_watchdog, lcw_list);
 287                         cfs_list_del(&lcw->lcw_list);
 288                         LIBCFS_FREE(lcw, sizeof(*lcw));
 289                 }
 290         }
 291
 292         complete(&lcw_stop_completion);
 293
 294         RETURN(rc);
 295 }
 296
 297 static void lcw_dispatch_start(void)
 298 {
 299         struct task_struct *task;
 300
 301         ENTRY;
 302         LASSERT(lcw_refcount == 1);
 303
 304         init_completion(&lcw_stop_completion);
 305         init_completion(&lcw_start_completion);
 306         init_waitqueue_head(&lcw_event_waitq);
 307
 308         CDEBUG(D_INFO, "starting dispatch thread\n");
 309         task = kthread_run(lcw_dispatch_main, NULL, "lc_watchdogd");
 310         if (IS_ERR(task)) {
 311                 CERROR("error spawning watchdog dispatch thread: %ld\n",
 312                         PTR_ERR(task));
 313                 EXIT;
 314                 return;
 315         }
 316         wait_for_completion(&lcw_start_completion);
 317         CDEBUG(D_INFO, "watchdog dispatcher initialization complete.\n");
 318
 319         EXIT;
 320 }
 321
 322 static void lcw_dispatch_stop(void)
 323 {
 324         ENTRY;
 325         LASSERT(lcw_refcount == 0);
 326
 327         CDEBUG(D_INFO, "trying to stop watchdog dispatcher.\n");
 328
 329         set_bit(LCW_FLAG_STOP, &lcw_flags);
 330         wake_up(&lcw_event_waitq);
 331
 332         wait_for_completion(&lcw_stop_completion);
 333
 334         CDEBUG(D_INFO, "watchdog dispatcher has shut down.\n");
 335
 336         EXIT;
 337 }
 338
 339 struct lc_watchdog *lc_watchdog_add(int timeout,
 340                                     void (*callback)(pid_t, void *),
 341                                     void *data)
 342 {
 343         struct lc_watchdog *lcw = NULL;
 344         ENTRY;
 345
 346         LIBCFS_ALLOC(lcw, sizeof(*lcw));
 347         if (lcw == NULL) {
 348                 CDEBUG(D_INFO, "Could not allocate new lc_watchdog\n");
 349                 RETURN(ERR_PTR(-ENOMEM));
 350         }
 351
 352         spin_lock_init(&lcw->lcw_lock);
 353         lcw->lcw_refcount = 1; /* refcount for owner */
 354         lcw->lcw_task     = current;
 355         lcw->lcw_pid      = current_pid();
 356         lcw->lcw_callback = (callback != NULL) ? callback : lc_watchdog_dumplog;
 357         lcw->lcw_data     = data;
 358         lcw->lcw_state    = LC_WATCHDOG_DISABLED;
 359
 360         CFS_INIT_LIST_HEAD(&lcw->lcw_list);
 361         cfs_timer_init(&lcw->lcw_timer, lcw_cb, lcw);
 362
 363         mutex_lock(&lcw_refcount_mutex);
 364         if (++lcw_refcount == 1)
 365                 lcw_dispatch_start();
 366         mutex_unlock(&lcw_refcount_mutex);
 367
 368         /* Keep this working in case we enable them by default */
 369         if (lcw->lcw_state == LC_WATCHDOG_ENABLED) {
 370                 lcw->lcw_last_touched = cfs_time_current();
 371                 cfs_timer_arm(&lcw->lcw_timer, cfs_time_seconds(timeout) +
 372                               cfs_time_current());
 373         }
 374
 375         RETURN(lcw);
 376 }
 377 EXPORT_SYMBOL(lc_watchdog_add);
 378
 379 static void lcw_update_time(struct lc_watchdog *lcw, const char *message)
 380 {
 381         cfs_time_t newtime = cfs_time_current();;
 382
 383         if (lcw->lcw_state == LC_WATCHDOG_EXPIRED) {
 384                 struct timeval timediff;
 385                 cfs_time_t delta_time = cfs_time_sub(newtime,
 386                                                      lcw->lcw_last_touched);
 387                 cfs_duration_usec(delta_time, &timediff);
 388
 389                 LCONSOLE_WARN("Service thread pid %u %s after %lu.%.02lus. "
 390                               "This indicates the system was overloaded (too "
 391                               "many service threads, or there were not enough "
 392                               "hardware resources).\n",
 393                               lcw->lcw_pid,
 394                               message,
 395                               timediff.tv_sec,
 396                               timediff.tv_usec / 10000);
 397         }
 398         lcw->lcw_last_touched = newtime;
 399 }
 400
 401 static void lc_watchdog_del_pending(struct lc_watchdog *lcw)
 402 {
 403         spin_lock_bh(&lcw->lcw_lock);
 404         if (unlikely(!cfs_list_empty(&lcw->lcw_list))) {
 405                 spin_lock_bh(&lcw_pending_timers_lock);
 406                 cfs_list_del_init(&lcw->lcw_list);
 407                 lcw->lcw_refcount--; /* -1 ref for pending list */
 408                 spin_unlock_bh(&lcw_pending_timers_lock);
 409         }
 410
 411         spin_unlock_bh(&lcw->lcw_lock);
 412 }
 413
 414 void lc_watchdog_touch(struct lc_watchdog *lcw, int timeout)
 415 {
 416         ENTRY;
 417         LASSERT(lcw != NULL);
 418
 419         lc_watchdog_del_pending(lcw);
 420
 421         lcw_update_time(lcw, "resumed");
 422         lcw->lcw_state = LC_WATCHDOG_ENABLED;
 423
 424         cfs_timer_arm(&lcw->lcw_timer, cfs_time_current() +
 425                       cfs_time_seconds(timeout));
 426
 427         EXIT;
 428 }
 429 EXPORT_SYMBOL(lc_watchdog_touch);
 430
 431 void lc_watchdog_disable(struct lc_watchdog *lcw)
 432 {
 433         ENTRY;
 434         LASSERT(lcw != NULL);
 435
 436         lc_watchdog_del_pending(lcw);
 437
 438         lcw_update_time(lcw, "completed");
 439         lcw->lcw_state = LC_WATCHDOG_DISABLED;
 440
 441         EXIT;
 442 }
 443 EXPORT_SYMBOL(lc_watchdog_disable);
 444
 445 void lc_watchdog_delete(struct lc_watchdog *lcw)
 446 {
 447         int dead;
 448
 449         ENTRY;
 450         LASSERT(lcw != NULL);
 451
 452         cfs_timer_disarm(&lcw->lcw_timer);
 453
 454         lcw_update_time(lcw, "stopped");
 455
 456         spin_lock_bh(&lcw->lcw_lock);
 457         spin_lock_bh(&lcw_pending_timers_lock);
 458         if (unlikely(!cfs_list_empty(&lcw->lcw_list))) {
 459                 cfs_list_del_init(&lcw->lcw_list);
 460                 lcw->lcw_refcount--; /* -1 ref for pending list */
 461         }
 462
 463         lcw->lcw_refcount--; /* -1 ref for owner */
 464         dead = lcw->lcw_refcount == 0;
 465         spin_unlock_bh(&lcw_pending_timers_lock);
 466         spin_unlock_bh(&lcw->lcw_lock);
 467
 468         if (dead)
 469                 LIBCFS_FREE(lcw, sizeof(*lcw));
 470
 471         mutex_lock(&lcw_refcount_mutex);
 472         if (--lcw_refcount == 0)
 473                 lcw_dispatch_stop();
 474         mutex_unlock(&lcw_refcount_mutex);
 475
 476         EXIT;
 477 }
 478 EXPORT_SYMBOL(lc_watchdog_delete);
 479
 480 /*
 481  * Provided watchdog handlers
 482  */
 483
 484 void lc_watchdog_dumplog(pid_t pid, void *data)
 485 {
 486         libcfs_debug_dumplog_internal((void *)((long_ptr_t)pid));
 487 }
 488 EXPORT_SYMBOL(lc_watchdog_dumplog);
 489
 490 #else   /* !defined(WITH_WATCHDOG) */
 491
 492 struct lc_watchdog *lc_watchdog_add(int timeout,
 493                                     void (*callback)(pid_t pid, void *),
 494                                     void *data)
 495 {
 496         static struct lc_watchdog      watchdog;
 497         return &watchdog;
 498 }
 499 EXPORT_SYMBOL(lc_watchdog_add);
 500
 501 void lc_watchdog_touch(struct lc_watchdog *lcw, int timeout)
 502 {
 503 }
 504 EXPORT_SYMBOL(lc_watchdog_touch);
 505
 506 void lc_watchdog_disable(struct lc_watchdog *lcw)
 507 {
 508 }
 509 EXPORT_SYMBOL(lc_watchdog_disable);
 510
 511 void lc_watchdog_delete(struct lc_watchdog *lcw)
 512 {
 513 }
 514 EXPORT_SYMBOL(lc_watchdog_delete);
 515
 516 #endif