4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2014, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * libcfs/libcfs/watchdog.c
34 * Author: Jacob Berkman <jacob@clusterfs.com>
37 #define DEBUG_SUBSYSTEM S_LNET
39 #include <linux/kthread.h>
40 #include <libcfs/libcfs.h>
41 #include "tracefile.h"
44 spinlock_t lcw_lock; /* check or change lcw_list */
45 int lcw_refcount; /* must hold lcw_pending_timers_lock */
46 struct timer_list lcw_timer; /* kernel timer */
47 struct list_head lcw_list; /* chain on pending list */
48 ktime_t lcw_last_touched;/* last touched stamp */
49 struct task_struct *lcw_task; /* owner task */
50 void (*lcw_callback)(pid_t, void *);
64 * The dispatcher will complete lcw_start_completion when it starts,
65 * and lcw_stop_completion when it exits.
66 * Wake lcw_event_waitq to signal timer callback dispatches.
68 static struct completion lcw_start_completion;
69 static struct completion lcw_stop_completion;
70 static wait_queue_head_t lcw_event_waitq;
73 * Set this and wake lcw_event_waitq to stop the dispatcher.
78 static unsigned long lcw_flags = 0;
81 * Number of outstanding watchdogs.
82 * When it hits 1, we start the dispatcher.
83 * When it hits 0, we stop the dispatcher.
85 static __u32 lcw_refcount = 0;
86 static DEFINE_MUTEX(lcw_refcount_mutex);
89 * List of timers that have fired that need their callbacks run by the
93 static DEFINE_SPINLOCK(lcw_pending_timers_lock);
94 static struct list_head lcw_pending_timers = LIST_HEAD_INIT(lcw_pending_timers);
96 /* Last time a watchdog expired */
97 static time64_t lcw_last_watchdog_time;
98 static int lcw_recent_watchdog_count;
101 lcw_dump(struct lc_watchdog *lcw)
105 if (lcw->lcw_task == NULL) {
106 LCONSOLE_WARN("Process %d was not found in the task "
107 "list; watchdog callback may be incomplete\n",
110 libcfs_debug_dumpstack(lcw->lcw_task);
117 static void lcw_cb(uintptr_t data)
119 struct lc_watchdog *lcw = (struct lc_watchdog *)data;
122 if (lcw->lcw_state != LC_WATCHDOG_ENABLED) {
127 lcw->lcw_state = LC_WATCHDOG_EXPIRED;
129 spin_lock_bh(&lcw->lcw_lock);
130 LASSERT(list_empty(&lcw->lcw_list));
132 spin_lock_bh(&lcw_pending_timers_lock);
133 lcw->lcw_refcount++; /* +1 for pending list */
134 list_add(&lcw->lcw_list, &lcw_pending_timers);
135 wake_up(&lcw_event_waitq);
137 spin_unlock_bh(&lcw_pending_timers_lock);
138 spin_unlock_bh(&lcw->lcw_lock);
142 static int is_watchdog_fired(void)
146 if (test_bit(LCW_FLAG_STOP, &lcw_flags))
149 spin_lock_bh(&lcw_pending_timers_lock);
150 rc = !list_empty(&lcw_pending_timers);
151 spin_unlock_bh(&lcw_pending_timers_lock);
155 static void lcw_dump_stack(struct lc_watchdog *lcw)
157 time64_t current_time = ktime_get_seconds();
158 struct timespec64 timediff;
161 timediff = ktime_to_timespec64(ktime_sub(ktime_get(),
162 lcw->lcw_last_touched));
164 /* LU-9235: Don't dump stack if the thread is just touched. */
165 if (timediff.tv_sec == 0)
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.
172 delta_time = current_time - lcw_last_watchdog_time;
173 if (delta_time < libcfs_watchdog_ratelimit &&
174 lcw_recent_watchdog_count > 3) {
175 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",
178 timediff.tv_nsec / (NSEC_PER_SEC / 100),
179 libcfs_watchdog_ratelimit);
181 if (delta_time < libcfs_watchdog_ratelimit) {
182 lcw_recent_watchdog_count++;
184 memcpy(&lcw_last_watchdog_time, ¤t_time,
185 sizeof(current_time));
186 lcw_recent_watchdog_count = 0;
189 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",
192 timediff.tv_nsec / (NSEC_PER_SEC / 100));
198 * Provided watchdog handlers
201 static void lc_watchdog_dumplog(pid_t pid, void *data)
203 libcfs_debug_dumplog_internal((void *)((uintptr_t)pid));
206 static int lcw_dispatch_main(void *data)
209 struct lc_watchdog *lcw;
210 struct list_head zombies = LIST_HEAD_INIT(zombies);
214 complete(&lcw_start_completion);
219 rc = wait_event_interruptible(lcw_event_waitq,
220 is_watchdog_fired());
221 CDEBUG(D_INFO, "Watchdog got woken up...\n");
222 if (test_bit(LCW_FLAG_STOP, &lcw_flags)) {
223 CDEBUG(D_INFO, "LCW_FLAG_STOP set, shutting down...\n");
225 spin_lock_bh(&lcw_pending_timers_lock);
226 rc = !list_empty(&lcw_pending_timers);
227 spin_unlock_bh(&lcw_pending_timers_lock);
229 CERROR("pending timers list was not empty at "
230 "time of watchdog dispatch shutdown\n");
235 spin_lock_bh(&lcw_pending_timers_lock);
236 while (!list_empty(&lcw_pending_timers)) {
239 lcw = list_entry(lcw_pending_timers.next,
240 struct lc_watchdog, lcw_list);
241 /* +1 ref for callback to make sure lwc wouldn't be
242 * deleted after releasing lcw_pending_timers_lock */
244 spin_unlock_bh(&lcw_pending_timers_lock);
247 spin_lock_bh(&lcw->lcw_lock);
248 spin_lock_bh(&lcw_pending_timers_lock);
250 if (list_empty(&lcw->lcw_list)) {
251 /* already removed from pending list */
252 lcw->lcw_refcount--; /* -1 ref for callback */
253 if (lcw->lcw_refcount == 0)
254 list_add(&lcw->lcw_list, &zombies);
255 spin_unlock_bh(&lcw->lcw_lock);
256 /* still hold lcw_pending_timers_lock */
260 list_del_init(&lcw->lcw_list);
261 lcw->lcw_refcount--; /* -1 ref for pending list */
263 spin_unlock_bh(&lcw_pending_timers_lock);
264 spin_unlock_bh(&lcw->lcw_lock);
266 CDEBUG(D_INFO, "found lcw for pid %d\n",
270 is_dumplog = lcw->lcw_callback == lc_watchdog_dumplog;
271 if (lcw->lcw_state != LC_WATCHDOG_DISABLED &&
272 (dumplog || !is_dumplog)) {
273 lcw->lcw_callback(lcw->lcw_pid, lcw->lcw_data);
274 if (dumplog && is_dumplog)
278 spin_lock_bh(&lcw_pending_timers_lock);
279 lcw->lcw_refcount--; /* -1 ref for callback */
280 if (lcw->lcw_refcount == 0)
281 list_add(&lcw->lcw_list, &zombies);
283 spin_unlock_bh(&lcw_pending_timers_lock);
285 while (!list_empty(&zombies)) {
286 lcw = list_entry(zombies.next,
287 struct lc_watchdog, lcw_list);
288 list_del_init(&lcw->lcw_list);
289 LIBCFS_FREE(lcw, sizeof(*lcw));
293 complete(&lcw_stop_completion);
298 static void lcw_dispatch_start(void)
300 struct task_struct *task;
303 LASSERT(lcw_refcount == 1);
305 init_completion(&lcw_stop_completion);
306 init_completion(&lcw_start_completion);
307 init_waitqueue_head(&lcw_event_waitq);
309 CDEBUG(D_INFO, "starting dispatch thread\n");
310 task = kthread_run(lcw_dispatch_main, NULL, "lc_watchdogd");
312 CERROR("error spawning watchdog dispatch thread: %ld\n",
317 wait_for_completion(&lcw_start_completion);
318 CDEBUG(D_INFO, "watchdog dispatcher initialization complete.\n");
323 static void lcw_dispatch_stop(void)
326 LASSERT(lcw_refcount == 0);
328 CDEBUG(D_INFO, "trying to stop watchdog dispatcher.\n");
330 set_bit(LCW_FLAG_STOP, &lcw_flags);
331 wake_up(&lcw_event_waitq);
333 wait_for_completion(&lcw_stop_completion);
335 CDEBUG(D_INFO, "watchdog dispatcher has shut down.\n");
340 struct lc_watchdog *lc_watchdog_add(int timeout,
341 void (*callback)(pid_t, void *),
344 struct lc_watchdog *lcw = NULL;
347 LIBCFS_ALLOC(lcw, sizeof(*lcw));
349 CDEBUG(D_INFO, "Could not allocate new lc_watchdog\n");
350 RETURN(ERR_PTR(-ENOMEM));
353 spin_lock_init(&lcw->lcw_lock);
354 lcw->lcw_refcount = 1; /* refcount for owner */
355 lcw->lcw_task = current;
356 lcw->lcw_pid = current_pid();
357 lcw->lcw_callback = (callback != NULL) ? callback : lc_watchdog_dumplog;
358 lcw->lcw_data = data;
359 lcw->lcw_state = LC_WATCHDOG_DISABLED;
361 INIT_LIST_HEAD(&lcw->lcw_list);
362 setup_timer(&lcw->lcw_timer, lcw_cb, (unsigned long)lcw);
364 mutex_lock(&lcw_refcount_mutex);
365 if (++lcw_refcount == 1)
366 lcw_dispatch_start();
367 mutex_unlock(&lcw_refcount_mutex);
369 /* Keep this working in case we enable them by default */
370 if (lcw->lcw_state == LC_WATCHDOG_ENABLED) {
371 lcw->lcw_last_touched = ktime_get();
372 mod_timer(&lcw->lcw_timer, cfs_time_seconds(timeout) +
378 EXPORT_SYMBOL(lc_watchdog_add);
380 static void lcw_update_time(struct lc_watchdog *lcw, const char *message)
382 ktime_t newtime = ktime_get();
384 if (lcw->lcw_state == LC_WATCHDOG_EXPIRED) {
385 ktime_t lapse = ktime_sub(newtime, lcw->lcw_last_touched);
386 struct timespec64 timediff;
388 timediff = ktime_to_timespec64(lapse);
389 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",
390 lcw->lcw_pid, message,
392 timediff.tv_nsec / (NSEC_PER_SEC / 100));
394 lcw->lcw_last_touched = newtime;
397 static void lc_watchdog_del_pending(struct lc_watchdog *lcw)
399 spin_lock_bh(&lcw->lcw_lock);
400 if (unlikely(!list_empty(&lcw->lcw_list))) {
401 spin_lock_bh(&lcw_pending_timers_lock);
402 list_del_init(&lcw->lcw_list);
403 lcw->lcw_refcount--; /* -1 ref for pending list */
404 spin_unlock_bh(&lcw_pending_timers_lock);
407 spin_unlock_bh(&lcw->lcw_lock);
410 void lc_watchdog_touch(struct lc_watchdog *lcw, int timeout)
413 LASSERT(lcw != NULL);
415 lc_watchdog_del_pending(lcw);
417 lcw_update_time(lcw, "resumed");
419 mod_timer(&lcw->lcw_timer, jiffies + cfs_time_seconds(timeout));
420 lcw->lcw_state = LC_WATCHDOG_ENABLED;
424 EXPORT_SYMBOL(lc_watchdog_touch);
426 void lc_watchdog_disable(struct lc_watchdog *lcw)
429 LASSERT(lcw != NULL);
431 lc_watchdog_del_pending(lcw);
433 lcw_update_time(lcw, "completed");
434 lcw->lcw_state = LC_WATCHDOG_DISABLED;
438 EXPORT_SYMBOL(lc_watchdog_disable);
440 void lc_watchdog_delete(struct lc_watchdog *lcw)
445 LASSERT(lcw != NULL);
447 del_timer(&lcw->lcw_timer);
449 lcw_update_time(lcw, "stopped");
451 spin_lock_bh(&lcw->lcw_lock);
452 spin_lock_bh(&lcw_pending_timers_lock);
453 if (unlikely(!list_empty(&lcw->lcw_list))) {
454 list_del_init(&lcw->lcw_list);
455 lcw->lcw_refcount--; /* -1 ref for pending list */
458 lcw->lcw_refcount--; /* -1 ref for owner */
459 dead = lcw->lcw_refcount == 0;
460 spin_unlock_bh(&lcw_pending_timers_lock);
461 spin_unlock_bh(&lcw->lcw_lock);
464 LIBCFS_FREE(lcw, sizeof(*lcw));
466 mutex_lock(&lcw_refcount_mutex);
467 if (--lcw_refcount == 0)
469 mutex_unlock(&lcw_refcount_mutex);
473 EXPORT_SYMBOL(lc_watchdog_delete);
475 #else /* !defined(WITH_WATCHDOG) */
477 struct lc_watchdog *lc_watchdog_add(int timeout,
478 void (*callback)(pid_t pid, void *),
481 static struct lc_watchdog watchdog;
484 EXPORT_SYMBOL(lc_watchdog_add);
486 void lc_watchdog_touch(struct lc_watchdog *lcw, int timeout)
489 EXPORT_SYMBOL(lc_watchdog_touch);
491 void lc_watchdog_disable(struct lc_watchdog *lcw)
494 EXPORT_SYMBOL(lc_watchdog_disable);
496 void lc_watchdog_delete(struct lc_watchdog *lcw)
499 EXPORT_SYMBOL(lc_watchdog_delete);