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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * libcfs/libcfs/watchdog.c
  33  *
  34  * Author: Jacob Berkman <jacob@clusterfs.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LNET
  38
  39 #include <linux/kthread.h>
  40 #include <libcfs/libcfs.h>
  41 #include "tracefile.h"
  42
  43 #ifndef WITH_WATCHDOG
  44 #define WITH_WATCHDOG
  45 #endif
  46
  47 struct lc_watchdog {
  48         spinlock_t              lcw_lock;       /* check or change lcw_list */
  49         int                     lcw_refcount;   /* must hold lcw_pending_timers_lock */
  50         struct timer_list       lcw_timer;      /* kernel timer */
  51         struct list_head        lcw_list;       /* chain on pending list */
  52         ktime_t                 lcw_last_touched;/* last touched stamp */
  53         struct task_struct     *lcw_task;       /* owner task */
  54         void                    (*lcw_callback)(pid_t, void *);
  55         void                    *lcw_data;
  56
  57         pid_t                   lcw_pid;
  58
  59         enum {
  60                 LC_WATCHDOG_DISABLED,
  61                 LC_WATCHDOG_ENABLED,
  62                 LC_WATCHDOG_EXPIRED
  63         } lcw_state;
  64 };
  65
  66 #ifdef WITH_WATCHDOG
  67 /*
  68  * The dispatcher will complete lcw_start_completion when it starts,
  69  * and lcw_stop_completion when it exits.
  70  * Wake lcw_event_waitq to signal timer callback dispatches.
  71  */
  72 static struct completion lcw_start_completion;
  73 static struct completion  lcw_stop_completion;
  74 static wait_queue_head_t lcw_event_waitq;
  75
  76 /*
  77  * Set this and wake lcw_event_waitq to stop the dispatcher.
  78  */
  79 enum {
  80         LCW_FLAG_STOP = 0
  81 };
  82 static unsigned long lcw_flags = 0;
  83
  84 /*
  85  * Number of outstanding watchdogs.
  86  * When it hits 1, we start the dispatcher.
  87  * When it hits 0, we stop the dispatcher.
  88  */
  89 static __u32         lcw_refcount = 0;
  90 static DEFINE_MUTEX(lcw_refcount_mutex);
  91
  92 /*
  93  * List of timers that have fired that need their callbacks run by the
  94  * dispatcher.
  95  */
  96 /* BH lock! */
  97 static DEFINE_SPINLOCK(lcw_pending_timers_lock);
  98 static struct list_head lcw_pending_timers = LIST_HEAD_INIT(lcw_pending_timers);
  99
 100 /* Last time a watchdog expired */
 101 static time64_t lcw_last_watchdog_time;
 102 static int lcw_recent_watchdog_count;
 103
 104 static void
 105 lcw_dump(struct lc_watchdog *lcw)
 106 {
 107         ENTRY;
 108         rcu_read_lock();
 109        if (lcw->lcw_task == NULL) {
 110                 LCONSOLE_WARN("Process %d was not found in the task "
 111                               "list; watchdog callback may be incomplete\n",
 112                               (int)lcw->lcw_pid);
 113         } else {
 114                 libcfs_debug_dumpstack(lcw->lcw_task);
 115         }
 116
 117         rcu_read_unlock();
 118         EXIT;
 119 }
 120
 121 static void lcw_cb(cfs_timer_cb_arg_t data)
 122 {
 123         struct lc_watchdog *lcw = cfs_from_timer(lcw, data, lcw_timer);
 124         ENTRY;
 125
 126         if (lcw->lcw_state != LC_WATCHDOG_ENABLED) {
 127                 EXIT;
 128                 return;
 129         }
 130
 131         lcw->lcw_state = LC_WATCHDOG_EXPIRED;
 132
 133         spin_lock_bh(&lcw->lcw_lock);
 134         LASSERT(list_empty(&lcw->lcw_list));
 135
 136         spin_lock_bh(&lcw_pending_timers_lock);
 137         lcw->lcw_refcount++; /* +1 for pending list */
 138         list_add(&lcw->lcw_list, &lcw_pending_timers);
 139         wake_up(&lcw_event_waitq);
 140
 141         spin_unlock_bh(&lcw_pending_timers_lock);
 142         spin_unlock_bh(&lcw->lcw_lock);
 143         EXIT;
 144 }
 145
 146 static int is_watchdog_fired(void)
 147 {
 148         int rc;
 149
 150         if (test_bit(LCW_FLAG_STOP, &lcw_flags))
 151                 return 1;
 152
 153         spin_lock_bh(&lcw_pending_timers_lock);
 154         rc = !list_empty(&lcw_pending_timers);
 155         spin_unlock_bh(&lcw_pending_timers_lock);
 156         return rc;
 157 }
 158
 159 static void lcw_dump_stack(struct lc_watchdog *lcw)
 160 {
 161         time64_t current_time = ktime_get_seconds();
 162         struct timespec64 timediff;
 163         time64_t delta_time;
 164
 165         timediff = ktime_to_timespec64(ktime_sub(ktime_get(),
 166                                        lcw->lcw_last_touched));
 167
 168         /* LU-9235: Don't dump stack if the thread is just touched. */
 169         if (timediff.tv_sec == 0)
 170                 return;
 171
 172         /*
 173          * Check to see if we should throttle the watchdog timer to avoid
 174          * too many dumps going to the console thus triggering an NMI.
 175          */
 176         delta_time = current_time - lcw_last_watchdog_time;
 177         if (delta_time < libcfs_watchdog_ratelimit &&
 178             lcw_recent_watchdog_count > 3) {
 179                 LCONSOLE_WARN("Service thread pid %u was inactive for %lu.%.02lus. Watchdog stack traces are limited to 3 per %d seconds, skipping this one.\n",
 180                               (int)lcw->lcw_pid,
 181                               timediff.tv_sec,
 182                               timediff.tv_nsec / (NSEC_PER_SEC / 100),
 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 %lu.%.02lus. The thread might be hung, or it might only be slow and will resume later. Dumping the stack trace for debugging purposes:\n",
 194                               (int)lcw->lcw_pid,
 195                               timediff.tv_sec,
 196                               timediff.tv_nsec / (NSEC_PER_SEC / 100));
 197                 lcw_dump(lcw);
 198         }
 199 }
 200
 201 /*
 202  * Provided watchdog handlers
 203  */
 204
 205 static void lc_watchdog_dumplog(pid_t pid, void *data)
 206 {
 207         libcfs_debug_dumplog_internal((void *)((uintptr_t)pid));
 208 }
 209
 210 static int lcw_dispatch_main(void *data)
 211 {
 212         int                 rc = 0;
 213         struct lc_watchdog *lcw;
 214         struct list_head zombies = LIST_HEAD_INIT(zombies);
 215
 216         ENTRY;
 217
 218         complete(&lcw_start_completion);
 219
 220         while (1) {
 221                 int dumplog = 1;
 222
 223                 rc = wait_event_interruptible(lcw_event_waitq,
 224                                               is_watchdog_fired());
 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 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                         int is_dumplog;
 242
 243                         lcw = list_entry(lcw_pending_timers.next,
 244                                          struct lc_watchdog, lcw_list);
 245                         /* +1 ref for callback to make sure lwc wouldn't be
 246                          * deleted after releasing lcw_pending_timers_lock */
 247                         lcw->lcw_refcount++;
 248                         spin_unlock_bh(&lcw_pending_timers_lock);
 249
 250                         /* lock ordering */
 251                         spin_lock_bh(&lcw->lcw_lock);
 252                         spin_lock_bh(&lcw_pending_timers_lock);
 253
 254                         if (list_empty(&lcw->lcw_list)) {
 255                                 /* already removed from pending list */
 256                                 lcw->lcw_refcount--; /* -1 ref for callback */
 257                                 if (lcw->lcw_refcount == 0)
 258                                         list_add(&lcw->lcw_list, &zombies);
 259                                 spin_unlock_bh(&lcw->lcw_lock);
 260                                 /* still hold lcw_pending_timers_lock */
 261                                 continue;
 262                         }
 263
 264                         list_del_init(&lcw->lcw_list);
 265                         lcw->lcw_refcount--; /* -1 ref for pending list */
 266
 267                         spin_unlock_bh(&lcw_pending_timers_lock);
 268                         spin_unlock_bh(&lcw->lcw_lock);
 269
 270                         CDEBUG(D_INFO, "found lcw for pid %d\n",
 271                                lcw->lcw_pid);
 272                         lcw_dump_stack(lcw);
 273
 274                         is_dumplog = lcw->lcw_callback == lc_watchdog_dumplog;
 275                         if (lcw->lcw_state != LC_WATCHDOG_DISABLED &&
 276                             (dumplog || !is_dumplog)) {
 277                                 lcw->lcw_callback(lcw->lcw_pid, lcw->lcw_data);
 278                                 if (dumplog && is_dumplog)
 279                                         dumplog = 0;
 280                         }
 281
 282                         spin_lock_bh(&lcw_pending_timers_lock);
 283                         lcw->lcw_refcount--; /* -1 ref for callback */
 284                         if (lcw->lcw_refcount == 0)
 285                                 list_add(&lcw->lcw_list, &zombies);
 286                 }
 287                 spin_unlock_bh(&lcw_pending_timers_lock);
 288
 289                 while (!list_empty(&zombies)) {
 290                         lcw = list_entry(zombies.next,
 291                                              struct lc_watchdog, lcw_list);
 292                         list_del_init(&lcw->lcw_list);
 293                         LIBCFS_FREE(lcw, sizeof(*lcw));
 294                 }
 295         }
 296
 297         complete(&lcw_stop_completion);
 298
 299         RETURN(rc);
 300 }
 301
 302 static void lcw_dispatch_start(void)
 303 {
 304         struct task_struct *task;
 305
 306         ENTRY;
 307         LASSERT(lcw_refcount == 1);
 308
 309         init_completion(&lcw_stop_completion);
 310         init_completion(&lcw_start_completion);
 311         init_waitqueue_head(&lcw_event_waitq);
 312
 313         CDEBUG(D_INFO, "starting dispatch thread\n");
 314         task = kthread_run(lcw_dispatch_main, NULL, "lc_watchdogd");
 315         if (IS_ERR(task)) {
 316                 CERROR("error spawning watchdog dispatch thread: %ld\n",
 317                         PTR_ERR(task));
 318                 EXIT;
 319                 return;
 320         }
 321         wait_for_completion(&lcw_start_completion);
 322         CDEBUG(D_INFO, "watchdog dispatcher initialization complete.\n");
 323
 324         EXIT;
 325 }
 326
 327 static void lcw_dispatch_stop(void)
 328 {
 329         ENTRY;
 330         LASSERT(lcw_refcount == 0);
 331
 332         CDEBUG(D_INFO, "trying to stop watchdog dispatcher.\n");
 333
 334         set_bit(LCW_FLAG_STOP, &lcw_flags);
 335         wake_up(&lcw_event_waitq);
 336
 337         wait_for_completion(&lcw_stop_completion);
 338         clear_bit(LCW_FLAG_STOP, &lcw_flags);
 339
 340         CDEBUG(D_INFO, "watchdog dispatcher has shut down.\n");
 341
 342         EXIT;
 343 }
 344
 345 struct lc_watchdog *lc_watchdog_add(int timeout,
 346                                     void (*callback)(pid_t, void *),
 347                                     void *data)
 348 {
 349         struct lc_watchdog *lcw = NULL;
 350         ENTRY;
 351
 352         LIBCFS_ALLOC(lcw, sizeof(*lcw));
 353         if (lcw == NULL) {
 354                 CDEBUG(D_INFO, "Could not allocate new lc_watchdog\n");
 355                 RETURN(ERR_PTR(-ENOMEM));
 356         }
 357
 358         spin_lock_init(&lcw->lcw_lock);
 359         lcw->lcw_refcount = 1; /* refcount for owner */
 360         lcw->lcw_task     = current;
 361         lcw->lcw_pid      = current_pid();
 362         lcw->lcw_callback = (callback != NULL) ? callback : lc_watchdog_dumplog;
 363         lcw->lcw_data     = data;
 364         lcw->lcw_state    = LC_WATCHDOG_DISABLED;
 365
 366         INIT_LIST_HEAD(&lcw->lcw_list);
 367         cfs_timer_setup(&lcw->lcw_timer, lcw_cb, (unsigned long)lcw, 0);
 368
 369         mutex_lock(&lcw_refcount_mutex);
 370         if (++lcw_refcount == 1)
 371                 lcw_dispatch_start();
 372         mutex_unlock(&lcw_refcount_mutex);
 373
 374         /* Keep this working in case we enable them by default */
 375         if (lcw->lcw_state == LC_WATCHDOG_ENABLED) {
 376                 lcw->lcw_last_touched = ktime_get();
 377                 mod_timer(&lcw->lcw_timer, cfs_time_seconds(timeout) +
 378                           jiffies);
 379         }
 380
 381         RETURN(lcw);
 382 }
 383 EXPORT_SYMBOL(lc_watchdog_add);
 384
 385 static void lcw_update_time(struct lc_watchdog *lcw, const char *message)
 386 {
 387         ktime_t newtime = ktime_get();
 388
 389         if (lcw->lcw_state == LC_WATCHDOG_EXPIRED) {
 390                 ktime_t lapse = ktime_sub(newtime, lcw->lcw_last_touched);
 391                 struct timespec64 timediff;
 392
 393                 timediff = ktime_to_timespec64(lapse);
 394                 LCONSOLE_WARN("Service thread pid %u %s after %lu.%.02lus. This indicates the system was overloaded (too many service threads, or there were not enough hardware resources).\n",
 395                               lcw->lcw_pid, message,
 396                               timediff.tv_sec,
 397                               timediff.tv_nsec / (NSEC_PER_SEC / 100));
 398         }
 399         lcw->lcw_last_touched = newtime;
 400 }
 401
 402 static void lc_watchdog_del_pending(struct lc_watchdog *lcw)
 403 {
 404         spin_lock_bh(&lcw->lcw_lock);
 405         if (unlikely(!list_empty(&lcw->lcw_list))) {
 406                 spin_lock_bh(&lcw_pending_timers_lock);
 407                 list_del_init(&lcw->lcw_list);
 408                 lcw->lcw_refcount--; /* -1 ref for pending list */
 409                 spin_unlock_bh(&lcw_pending_timers_lock);
 410         }
 411
 412         spin_unlock_bh(&lcw->lcw_lock);
 413 }
 414
 415 void lc_watchdog_touch(struct lc_watchdog *lcw, int timeout)
 416 {
 417         ENTRY;
 418         LASSERT(lcw != NULL);
 419
 420         lc_watchdog_del_pending(lcw);
 421
 422         lcw_update_time(lcw, "resumed");
 423
 424         mod_timer(&lcw->lcw_timer, jiffies + cfs_time_seconds(timeout));
 425         lcw->lcw_state = LC_WATCHDOG_ENABLED;
 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         del_timer(&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(!list_empty(&lcw->lcw_list))) {
 459                 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 #else   /* !defined(WITH_WATCHDOG) */
 481
 482 struct lc_watchdog *lc_watchdog_add(int timeout,
 483                                     void (*callback)(pid_t pid, void *),
 484                                     void *data)
 485 {
 486         static struct lc_watchdog      watchdog;
 487         return &watchdog;
 488 }
 489 EXPORT_SYMBOL(lc_watchdog_add);
 490
 491 void lc_watchdog_touch(struct lc_watchdog *lcw, int timeout)
 492 {
 493 }
 494 EXPORT_SYMBOL(lc_watchdog_touch);
 495
 496 void lc_watchdog_disable(struct lc_watchdog *lcw)
 497 {
 498 }
 499 EXPORT_SYMBOL(lc_watchdog_disable);
 500
 501 void lc_watchdog_delete(struct lc_watchdog *lcw)
 502 {
 503 }
 504 EXPORT_SYMBOL(lc_watchdog_delete);
 505
 506 #endif