lnet/libcfs/watchdog.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2004 Cluster File Systems, Inc.
   5  *   Author: Jacob Berkman <jacob@clusterfs.com>
   6  *
   7  *   This file is part of Lustre, http://www.lustre.org.
   8  *
   9  *   Lustre is free software; you can redistribute it and/or
  10  *   modify it under the terms of version 2 of the GNU General Public
  11  *   License as published by the Free Software Foundation.
  12  *
  13  *   Lustre is distributed in the hope that it will be useful,
  14  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  *   GNU General Public License for more details.
  17  *
  18  *   You should have received a copy of the GNU General Public License
  19  *   along with Lustre; if not, write to the Free Software
  20  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  */
  22
  23 #define DEBUG_SUBSYSTEM S_LNET
  24
  25 #include <libcfs/kp30.h>
  26 #include <libcfs/libcfs.h>
  27 #include "tracefile.h"
  28
  29 struct lc_watchdog {
  30         cfs_timer_t       lcw_timer; /* kernel timer */
  31         struct list_head  lcw_list;
  32         struct timeval    lcw_last_touched;
  33         cfs_task_t       *lcw_task;
  34
  35         void            (*lcw_callback)(pid_t, void *);
  36         void             *lcw_data;
  37
  38         pid_t             lcw_pid;
  39         cfs_duration_t    lcw_time; /* time until watchdog fires, jiffies */
  40
  41         enum {
  42                 LC_WATCHDOG_DISABLED,
  43                 LC_WATCHDOG_ENABLED,
  44                 LC_WATCHDOG_EXPIRED
  45         } lcw_state;
  46 };
  47
  48 #ifdef WITH_WATCHDOG
  49 /*
  50  * The dispatcher will complete lcw_start_completion when it starts,
  51  * and lcw_stop_completion when it exits.
  52  * Wake lcw_event_waitq to signal timer callback dispatches.
  53  */
  54 static struct completion lcw_start_completion;
  55 static struct completion lcw_stop_completion;
  56 static wait_queue_head_t lcw_event_waitq;
  57
  58 /*
  59  * Set this and wake lcw_event_waitq to stop the dispatcher.
  60  */
  61 enum {
  62         LCW_FLAG_STOP = 0
  63 };
  64 static unsigned long lcw_flags = 0;
  65
  66 /*
  67  * Number of outstanding watchdogs.
  68  * When it hits 1, we start the dispatcher.
  69  * When it hits 0, we stop the distpatcher.
  70  */
  71 static __u32         lcw_refcount = 0;
  72 static DECLARE_MUTEX(lcw_refcount_sem);
  73
  74 /*
  75  * List of timers that have fired that need their callbacks run by the
  76  * dispatcher.
  77  */
  78 static spinlock_t lcw_pending_timers_lock = SPIN_LOCK_UNLOCKED; /* BH lock! */
  79 static struct list_head lcw_pending_timers = \
  80         LIST_HEAD_INIT(lcw_pending_timers);
  81
  82 #ifdef HAVE_TASKLIST_LOCK
  83 static void
  84 lcw_dump(struct lc_watchdog *lcw)
  85 {
  86         cfs_task_t *tsk;
  87         ENTRY;
  88
  89         read_lock(&tasklist_lock);
  90         tsk = find_task_by_pid(lcw->lcw_pid);
  91
  92         if (tsk == NULL) {
  93                 CWARN("Process %d was not found in the task list; "
  94                       "watchdog callback may be incomplete\n", (int)lcw->lcw_pid);
  95         } else if (tsk != lcw->lcw_task) {
  96                 CWARN("The current process %d did not set the watchdog; "
  97                       "watchdog callback may be incomplete\n", (int)lcw->lcw_pid);
  98         } else {
  99                 libcfs_debug_dumpstack(tsk);
 100         }
 101
 102         read_unlock(&tasklist_lock);
 103         EXIT;
 104 }
 105 #else
 106 static void
 107 lcw_dump(struct lc_watchdog *lcw)
 108 {
 109         CERROR("unable to dump stack because of missing export\n");
 110 }
 111 #endif
 112
 113 static void lcw_cb(unsigned long data)
 114 {
 115         struct lc_watchdog *lcw = (struct lc_watchdog *)data;
 116
 117         ENTRY;
 118
 119         if (lcw->lcw_state != LC_WATCHDOG_ENABLED) {
 120                 EXIT;
 121                 return;
 122         }
 123
 124         lcw->lcw_state = LC_WATCHDOG_EXPIRED;
 125
 126         /* NB this warning should appear on the console, but may not get into
 127          * the logs since we're running in a softirq handler */
 128
 129         CWARN("Watchdog triggered for pid %d: it was inactive for %lds\n",
 130               (int)lcw->lcw_pid, cfs_duration_sec(lcw->lcw_time));
 131         lcw_dump(lcw);
 132
 133         spin_lock_bh(&lcw_pending_timers_lock);
 134
 135         if (list_empty(&lcw->lcw_list)) {
 136                 list_add(&lcw->lcw_list, &lcw_pending_timers);
 137                 wake_up(&lcw_event_waitq);
 138         }
 139
 140         spin_unlock_bh(&lcw_pending_timers_lock);
 141
 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 = !list_empty(&lcw_pending_timers);
 154         spin_unlock_bh(&lcw_pending_timers_lock);
 155         return rc;
 156 }
 157
 158 static int lcw_dispatch_main(void *data)
 159 {
 160         int                 rc = 0;
 161         unsigned long       flags;
 162         struct lc_watchdog *lcw;
 163
 164         ENTRY;
 165
 166         cfs_daemonize("lc_watchdogd");
 167
 168         SIGNAL_MASK_LOCK(current, flags);
 169         sigfillset(&current->blocked);
 170         RECALC_SIGPENDING;
 171         SIGNAL_MASK_UNLOCK(current, flags);
 172
 173         complete(&lcw_start_completion);
 174
 175         while (1) {
 176                 wait_event_interruptible(lcw_event_waitq, is_watchdog_fired());
 177                 CDEBUG(D_INFO, "Watchdog got woken up...\n");
 178                 if (test_bit(LCW_FLAG_STOP, &lcw_flags)) {
 179                         CDEBUG(D_INFO, "LCW_FLAG_STOP was set, shutting down...\n");
 180
 181                         spin_lock_bh(&lcw_pending_timers_lock);
 182                         rc = !list_empty(&lcw_pending_timers);
 183                         spin_unlock_bh(&lcw_pending_timers_lock);
 184                         if (rc) {
 185                                 CERROR("pending timers list was not empty at "
 186                                        "time of watchdog dispatch shutdown\n");
 187                         }
 188                         break;
 189                 }
 190
 191                 spin_lock_bh(&lcw_pending_timers_lock);
 192                 while (!list_empty(&lcw_pending_timers)) {
 193
 194                         lcw = list_entry(lcw_pending_timers.next,
 195                                          struct lc_watchdog,
 196                                          lcw_list);
 197                         list_del_init(&lcw->lcw_list);
 198                         spin_unlock_bh(&lcw_pending_timers_lock);
 199
 200                         CDEBUG(D_INFO, "found lcw for pid %d: inactive for "
 201                                "%lds\n", (int)lcw->lcw_pid,
 202                                cfs_duration_sec(lcw->lcw_time));
 203
 204                         if (lcw->lcw_state != LC_WATCHDOG_DISABLED)
 205                                 lcw->lcw_callback(lcw->lcw_pid, lcw->lcw_data);
 206
 207                         spin_lock_bh(&lcw_pending_timers_lock);
 208                 }
 209                 spin_unlock_bh(&lcw_pending_timers_lock);
 210         }
 211
 212         complete(&lcw_stop_completion);
 213
 214         RETURN(rc);
 215 }
 216
 217 static void lcw_dispatch_start(void)
 218 {
 219         int rc;
 220
 221         ENTRY;
 222         LASSERT(lcw_refcount == 1);
 223
 224         init_completion(&lcw_stop_completion);
 225         init_completion(&lcw_start_completion);
 226         init_waitqueue_head(&lcw_event_waitq);
 227
 228         CDEBUG(D_INFO, "starting dispatch thread\n");
 229         rc = kernel_thread(lcw_dispatch_main, NULL, 0);
 230         if (rc < 0) {
 231                 CERROR("error spawning watchdog dispatch thread: %d\n", rc);
 232                 EXIT;
 233                 return;
 234         }
 235         wait_for_completion(&lcw_start_completion);
 236         CDEBUG(D_INFO, "watchdog dispatcher initialization complete.\n");
 237
 238         EXIT;
 239 }
 240
 241 static void lcw_dispatch_stop(void)
 242 {
 243         ENTRY;
 244         LASSERT(lcw_refcount == 0);
 245
 246         CDEBUG(D_INFO, "trying to stop watchdog dispatcher.\n");
 247
 248         set_bit(LCW_FLAG_STOP, &lcw_flags);
 249         wake_up(&lcw_event_waitq);
 250
 251         wait_for_completion(&lcw_stop_completion);
 252
 253         CDEBUG(D_INFO, "watchdog dispatcher has shut down.\n");
 254
 255         EXIT;
 256 }
 257
 258 struct lc_watchdog *lc_watchdog_add(int timeout_ms,
 259                                     void (*callback)(pid_t, void *),
 260                                     void *data)
 261 {
 262         struct lc_watchdog *lcw = NULL;
 263         ENTRY;
 264
 265         LIBCFS_ALLOC(lcw, sizeof(*lcw));
 266         if (lcw == NULL) {
 267                 CDEBUG(D_INFO, "Could not allocate new lc_watchdog\n");
 268                 RETURN(ERR_PTR(-ENOMEM));
 269         }
 270
 271         lcw->lcw_task     = cfs_current();
 272         lcw->lcw_pid      = cfs_curproc_pid();
 273         lcw->lcw_time     = cfs_time_seconds(timeout_ms) / 1000;
 274         lcw->lcw_callback = (callback != NULL) ? callback : lc_watchdog_dumplog;
 275         lcw->lcw_data     = data;
 276         lcw->lcw_state    = LC_WATCHDOG_DISABLED;
 277
 278         INIT_LIST_HEAD(&lcw->lcw_list);
 279
 280         lcw->lcw_timer.function = lcw_cb;
 281         lcw->lcw_timer.data = (unsigned long)lcw;
 282         lcw->lcw_timer.expires = jiffies + lcw->lcw_time;
 283         init_timer(&lcw->lcw_timer);
 284
 285         down(&lcw_refcount_sem);
 286         if (++lcw_refcount == 1)
 287                 lcw_dispatch_start();
 288         up(&lcw_refcount_sem);
 289
 290         /* Keep this working in case we enable them by default */
 291         if (lcw->lcw_state == LC_WATCHDOG_ENABLED) {
 292                 do_gettimeofday(&lcw->lcw_last_touched);
 293                 add_timer(&lcw->lcw_timer);
 294         }
 295
 296         RETURN(lcw);
 297 }
 298 EXPORT_SYMBOL(lc_watchdog_add);
 299
 300 static void lcw_update_time(struct lc_watchdog *lcw, const char *message)
 301 {
 302         struct timeval newtime;
 303         struct timeval timediff;
 304
 305         do_gettimeofday(&newtime);
 306         if (lcw->lcw_state == LC_WATCHDOG_EXPIRED) {
 307                 cfs_timeval_sub(&newtime, &lcw->lcw_last_touched, &timediff);
 308                 CWARN("Expired watchdog for pid %d %s after %lu.%.4lus\n",
 309                       lcw->lcw_pid,
 310                       message,
 311                       timediff.tv_sec,
 312                       timediff.tv_usec / 100);
 313         }
 314         lcw->lcw_last_touched = newtime;
 315 }
 316
 317 void lc_watchdog_touch_ms(struct lc_watchdog *lcw, int timeout_ms)
 318 {
 319         ENTRY;
 320         LASSERT(lcw != NULL);
 321
 322         spin_lock_bh(&lcw_pending_timers_lock);
 323         list_del_init(&lcw->lcw_list);
 324         spin_unlock_bh(&lcw_pending_timers_lock);
 325
 326         lcw_update_time(lcw, "touched");
 327         lcw->lcw_state = LC_WATCHDOG_ENABLED;
 328
 329         mod_timer(&lcw->lcw_timer, jiffies +
 330                   cfs_time_seconds(timeout_ms) / 1000);
 331
 332         EXIT;
 333 }
 334 EXPORT_SYMBOL(lc_watchdog_touch_ms);
 335
 336 /* deprecated - use above instead */
 337 void lc_watchdog_touch(struct lc_watchdog *lcw)
 338 {
 339         lc_watchdog_touch_ms(lcw, cfs_duration_sec(lcw->lcw_time) * 1000);
 340 }
 341 EXPORT_SYMBOL(lc_watchdog_touch);
 342
 343 void lc_watchdog_disable(struct lc_watchdog *lcw)
 344 {
 345         ENTRY;
 346         LASSERT(lcw != NULL);
 347
 348         spin_lock_bh(&lcw_pending_timers_lock);
 349         if (!list_empty(&lcw->lcw_list))
 350                 list_del_init(&lcw->lcw_list);
 351         spin_unlock_bh(&lcw_pending_timers_lock);
 352
 353         lcw_update_time(lcw, "disabled");
 354         lcw->lcw_state = LC_WATCHDOG_DISABLED;
 355
 356         EXIT;
 357 }
 358 EXPORT_SYMBOL(lc_watchdog_disable);
 359
 360 void lc_watchdog_delete(struct lc_watchdog *lcw)
 361 {
 362         ENTRY;
 363         LASSERT(lcw != NULL);
 364
 365         del_timer(&lcw->lcw_timer);
 366
 367         lcw_update_time(lcw, "deleted");
 368
 369         spin_lock_bh(&lcw_pending_timers_lock);
 370         if (!list_empty(&lcw->lcw_list))
 371                 list_del_init(&lcw->lcw_list);
 372         spin_unlock_bh(&lcw_pending_timers_lock);
 373
 374         down(&lcw_refcount_sem);
 375         if (--lcw_refcount == 0)
 376                 lcw_dispatch_stop();
 377         up(&lcw_refcount_sem);
 378
 379         LIBCFS_FREE(lcw, sizeof(*lcw));
 380
 381         EXIT;
 382 }
 383 EXPORT_SYMBOL(lc_watchdog_delete);
 384
 385 /*
 386  * Provided watchdog handlers
 387  */
 388
 389 void lc_watchdog_dumplog(pid_t pid, void *data)
 390 {
 391         libcfs_debug_dumplog_internal((void *)((unsigned long)pid));
 392 }
 393 EXPORT_SYMBOL(lc_watchdog_dumplog);
 394
 395 #else   /* !defined(WITH_WATCHDOG) */
 396
 397 struct lc_watchdog *lc_watchdog_add(int timeout_ms,
 398                                     void (*callback)(pid_t pid, void *),
 399                                     void *data)
 400 {
 401         static struct lc_watchdog      watchdog;
 402         return &watchdog;
 403 }
 404 EXPORT_SYMBOL(lc_watchdog_add);
 405
 406 void lc_watchdog_touch_ms(struct lc_watchdog *lcw, int timeout_ms)
 407 {
 408 }
 409 EXPORT_SYMBOL(lc_watchdog_touch_ms);
 410
 411 void lc_watchdog_touch(struct lc_watchdog *lcw)
 412 {
 413 }
 414 EXPORT_SYMBOL(lc_watchdog_touch);
 415
 416 void lc_watchdog_disable(struct lc_watchdog *lcw)
 417 {
 418 }
 419 EXPORT_SYMBOL(lc_watchdog_disable);
 420
 421 void lc_watchdog_delete(struct lc_watchdog *lcw)
 422 {
 423 }
 424 EXPORT_SYMBOL(lc_watchdog_delete);
 425
 426 #endif
 427