Whamcloud - gitweb
LU-9399 llite: register mountpoint before process llog
[fs/lustre-release.git] / 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, 2014, 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 struct lc_watchdog {
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 *);
51         void                    *lcw_data;
52
53         pid_t                   lcw_pid;
54
55         enum {
56                 LC_WATCHDOG_DISABLED,
57                 LC_WATCHDOG_ENABLED,
58                 LC_WATCHDOG_EXPIRED
59         } lcw_state;
60 };
61
62 #ifdef WITH_WATCHDOG
63 /*
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.
67  */
68 static struct completion lcw_start_completion;
69 static struct completion  lcw_stop_completion;
70 static wait_queue_head_t lcw_event_waitq;
71
72 /*
73  * Set this and wake lcw_event_waitq to stop the dispatcher.
74  */
75 enum {
76         LCW_FLAG_STOP = 0
77 };
78 static unsigned long lcw_flags = 0;
79
80 /*
81  * Number of outstanding watchdogs.
82  * When it hits 1, we start the dispatcher.
83  * When it hits 0, we stop the dispatcher.
84  */
85 static __u32         lcw_refcount = 0;
86 static DEFINE_MUTEX(lcw_refcount_mutex);
87
88 /*
89  * List of timers that have fired that need their callbacks run by the
90  * dispatcher.
91  */
92 /* BH lock! */
93 static DEFINE_SPINLOCK(lcw_pending_timers_lock);
94 static struct list_head lcw_pending_timers = LIST_HEAD_INIT(lcw_pending_timers);
95
96 /* Last time a watchdog expired */
97 static time64_t lcw_last_watchdog_time;
98 static int lcw_recent_watchdog_count;
99
100 static void
101 lcw_dump(struct lc_watchdog *lcw)
102 {
103         ENTRY;
104         rcu_read_lock();
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",
108                               (int)lcw->lcw_pid);
109         } else {
110                 libcfs_debug_dumpstack(lcw->lcw_task);
111         }
112
113         rcu_read_unlock();
114         EXIT;
115 }
116
117 static void lcw_cb(uintptr_t data)
118 {
119         struct lc_watchdog *lcw = (struct lc_watchdog *)data;
120         ENTRY;
121
122         if (lcw->lcw_state != LC_WATCHDOG_ENABLED) {
123                 EXIT;
124                 return;
125         }
126
127         lcw->lcw_state = LC_WATCHDOG_EXPIRED;
128
129         spin_lock_bh(&lcw->lcw_lock);
130         LASSERT(list_empty(&lcw->lcw_list));
131
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);
136
137         spin_unlock_bh(&lcw_pending_timers_lock);
138         spin_unlock_bh(&lcw->lcw_lock);
139         EXIT;
140 }
141
142 static int is_watchdog_fired(void)
143 {
144         int rc;
145
146         if (test_bit(LCW_FLAG_STOP, &lcw_flags))
147                 return 1;
148
149         spin_lock_bh(&lcw_pending_timers_lock);
150         rc = !list_empty(&lcw_pending_timers);
151         spin_unlock_bh(&lcw_pending_timers_lock);
152         return rc;
153 }
154
155 static void lcw_dump_stack(struct lc_watchdog *lcw)
156 {
157         time64_t current_time = ktime_get_seconds();
158         struct timespec64 timediff;
159         time64_t delta_time;
160
161         timediff = ktime_to_timespec64(ktime_sub(ktime_get(),
162                                        lcw->lcw_last_touched));
163         /*
164          * Check to see if we should throttle the watchdog timer to avoid
165          * too many dumps going to the console thus triggering an NMI.
166          */
167         delta_time = current_time - lcw_last_watchdog_time;
168         if (delta_time < libcfs_watchdog_ratelimit &&
169             lcw_recent_watchdog_count > 3) {
170                 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",
171                               (int)lcw->lcw_pid,
172                               timediff.tv_sec,
173                               timediff.tv_nsec / (NSEC_PER_SEC / 100),
174                               libcfs_watchdog_ratelimit);
175         } else {
176                 if (delta_time < libcfs_watchdog_ratelimit) {
177                         lcw_recent_watchdog_count++;
178                 } else {
179                         memcpy(&lcw_last_watchdog_time, &current_time,
180                                sizeof(current_time));
181                         lcw_recent_watchdog_count = 0;
182                 }
183
184                 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",
185                               (int)lcw->lcw_pid,
186                               timediff.tv_sec,
187                               timediff.tv_nsec / (NSEC_PER_SEC / 100));
188                 lcw_dump(lcw);
189         }
190 }
191
192 /*
193  * Provided watchdog handlers
194  */
195
196 static void lc_watchdog_dumplog(pid_t pid, void *data)
197 {
198         libcfs_debug_dumplog_internal((void *)((uintptr_t)pid));
199 }
200
201 static int lcw_dispatch_main(void *data)
202 {
203         int                 rc = 0;
204         struct lc_watchdog *lcw;
205         struct list_head zombies = LIST_HEAD_INIT(zombies);
206
207         ENTRY;
208
209         complete(&lcw_start_completion);
210
211         while (1) {
212                 int dumplog = 1;
213
214                 rc = wait_event_interruptible(lcw_event_waitq,
215                                               is_watchdog_fired());
216                 CDEBUG(D_INFO, "Watchdog got woken up...\n");
217                 if (test_bit(LCW_FLAG_STOP, &lcw_flags)) {
218                         CDEBUG(D_INFO, "LCW_FLAG_STOP set, shutting down...\n");
219
220                         spin_lock_bh(&lcw_pending_timers_lock);
221                         rc = !list_empty(&lcw_pending_timers);
222                         spin_unlock_bh(&lcw_pending_timers_lock);
223                         if (rc) {
224                                 CERROR("pending timers list was not empty at "
225                                        "time of watchdog dispatch shutdown\n");
226                         }
227                         break;
228                 }
229
230                 spin_lock_bh(&lcw_pending_timers_lock);
231                 while (!list_empty(&lcw_pending_timers)) {
232                         int is_dumplog;
233
234                         lcw = list_entry(lcw_pending_timers.next,
235                                          struct lc_watchdog, lcw_list);
236                         /* +1 ref for callback to make sure lwc wouldn't be
237                          * deleted after releasing lcw_pending_timers_lock */
238                         lcw->lcw_refcount++;
239                         spin_unlock_bh(&lcw_pending_timers_lock);
240
241                         /* lock ordering */
242                         spin_lock_bh(&lcw->lcw_lock);
243                         spin_lock_bh(&lcw_pending_timers_lock);
244
245                         if (list_empty(&lcw->lcw_list)) {
246                                 /* already removed from pending list */
247                                 lcw->lcw_refcount--; /* -1 ref for callback */
248                                 if (lcw->lcw_refcount == 0)
249                                         list_add(&lcw->lcw_list, &zombies);
250                                 spin_unlock_bh(&lcw->lcw_lock);
251                                 /* still hold lcw_pending_timers_lock */
252                                 continue;
253                         }
254
255                         list_del_init(&lcw->lcw_list);
256                         lcw->lcw_refcount--; /* -1 ref for pending list */
257
258                         spin_unlock_bh(&lcw_pending_timers_lock);
259                         spin_unlock_bh(&lcw->lcw_lock);
260
261                         CDEBUG(D_INFO, "found lcw for pid %d\n",
262                                lcw->lcw_pid);
263                         lcw_dump_stack(lcw);
264
265                         is_dumplog = lcw->lcw_callback == lc_watchdog_dumplog;
266                         if (lcw->lcw_state != LC_WATCHDOG_DISABLED &&
267                             (dumplog || !is_dumplog)) {
268                                 lcw->lcw_callback(lcw->lcw_pid, lcw->lcw_data);
269                                 if (dumplog && is_dumplog)
270                                         dumplog = 0;
271                         }
272
273                         spin_lock_bh(&lcw_pending_timers_lock);
274                         lcw->lcw_refcount--; /* -1 ref for callback */
275                         if (lcw->lcw_refcount == 0)
276                                 list_add(&lcw->lcw_list, &zombies);
277                 }
278                 spin_unlock_bh(&lcw_pending_timers_lock);
279
280                 while (!list_empty(&zombies)) {
281                         lcw = list_entry(zombies.next,
282                                              struct lc_watchdog, lcw_list);
283                         list_del_init(&lcw->lcw_list);
284                         LIBCFS_FREE(lcw, sizeof(*lcw));
285                 }
286         }
287
288         complete(&lcw_stop_completion);
289
290         RETURN(rc);
291 }
292
293 static void lcw_dispatch_start(void)
294 {
295         struct task_struct *task;
296
297         ENTRY;
298         LASSERT(lcw_refcount == 1);
299
300         init_completion(&lcw_stop_completion);
301         init_completion(&lcw_start_completion);
302         init_waitqueue_head(&lcw_event_waitq);
303
304         CDEBUG(D_INFO, "starting dispatch thread\n");
305         task = kthread_run(lcw_dispatch_main, NULL, "lc_watchdogd");
306         if (IS_ERR(task)) {
307                 CERROR("error spawning watchdog dispatch thread: %ld\n",
308                         PTR_ERR(task));
309                 EXIT;
310                 return;
311         }
312         wait_for_completion(&lcw_start_completion);
313         CDEBUG(D_INFO, "watchdog dispatcher initialization complete.\n");
314
315         EXIT;
316 }
317
318 static void lcw_dispatch_stop(void)
319 {
320         ENTRY;
321         LASSERT(lcw_refcount == 0);
322
323         CDEBUG(D_INFO, "trying to stop watchdog dispatcher.\n");
324
325         set_bit(LCW_FLAG_STOP, &lcw_flags);
326         wake_up(&lcw_event_waitq);
327
328         wait_for_completion(&lcw_stop_completion);
329
330         CDEBUG(D_INFO, "watchdog dispatcher has shut down.\n");
331
332         EXIT;
333 }
334
335 struct lc_watchdog *lc_watchdog_add(int timeout,
336                                     void (*callback)(pid_t, void *),
337                                     void *data)
338 {
339         struct lc_watchdog *lcw = NULL;
340         ENTRY;
341
342         LIBCFS_ALLOC(lcw, sizeof(*lcw));
343         if (lcw == NULL) {
344                 CDEBUG(D_INFO, "Could not allocate new lc_watchdog\n");
345                 RETURN(ERR_PTR(-ENOMEM));
346         }
347
348         spin_lock_init(&lcw->lcw_lock);
349         lcw->lcw_refcount = 1; /* refcount for owner */
350         lcw->lcw_task     = current;
351         lcw->lcw_pid      = current_pid();
352         lcw->lcw_callback = (callback != NULL) ? callback : lc_watchdog_dumplog;
353         lcw->lcw_data     = data;
354         lcw->lcw_state    = LC_WATCHDOG_DISABLED;
355
356         INIT_LIST_HEAD(&lcw->lcw_list);
357         setup_timer(&lcw->lcw_timer, lcw_cb, (unsigned long)lcw);
358
359         mutex_lock(&lcw_refcount_mutex);
360         if (++lcw_refcount == 1)
361                 lcw_dispatch_start();
362         mutex_unlock(&lcw_refcount_mutex);
363
364         /* Keep this working in case we enable them by default */
365         if (lcw->lcw_state == LC_WATCHDOG_ENABLED) {
366                 lcw->lcw_last_touched = ktime_get();
367                 mod_timer(&lcw->lcw_timer, cfs_time_seconds(timeout) +
368                           jiffies);
369         }
370
371         RETURN(lcw);
372 }
373 EXPORT_SYMBOL(lc_watchdog_add);
374
375 static void lcw_update_time(struct lc_watchdog *lcw, const char *message)
376 {
377         ktime_t newtime = ktime_get();
378
379         if (lcw->lcw_state == LC_WATCHDOG_EXPIRED) {
380                 ktime_t lapse = ktime_sub(newtime, lcw->lcw_last_touched);
381                 struct timespec64 timediff;
382
383                 timediff = ktime_to_timespec64(lapse);
384                 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",
385                               lcw->lcw_pid, message,
386                               timediff.tv_sec,
387                               timediff.tv_nsec / (NSEC_PER_SEC / 100));
388         }
389         lcw->lcw_last_touched = newtime;
390 }
391
392 static void lc_watchdog_del_pending(struct lc_watchdog *lcw)
393 {
394         spin_lock_bh(&lcw->lcw_lock);
395         if (unlikely(!list_empty(&lcw->lcw_list))) {
396                 spin_lock_bh(&lcw_pending_timers_lock);
397                 list_del_init(&lcw->lcw_list);
398                 lcw->lcw_refcount--; /* -1 ref for pending list */
399                 spin_unlock_bh(&lcw_pending_timers_lock);
400         }
401
402         spin_unlock_bh(&lcw->lcw_lock);
403 }
404
405 void lc_watchdog_touch(struct lc_watchdog *lcw, int timeout)
406 {
407         ENTRY;
408         LASSERT(lcw != NULL);
409
410         lc_watchdog_del_pending(lcw);
411
412         lcw_update_time(lcw, "resumed");
413
414         mod_timer(&lcw->lcw_timer, jiffies + cfs_time_seconds(timeout));
415         lcw->lcw_state = LC_WATCHDOG_ENABLED;
416
417         EXIT;
418 }
419 EXPORT_SYMBOL(lc_watchdog_touch);
420
421 void lc_watchdog_disable(struct lc_watchdog *lcw)
422 {
423         ENTRY;
424         LASSERT(lcw != NULL);
425
426         lc_watchdog_del_pending(lcw);
427
428         lcw_update_time(lcw, "completed");
429         lcw->lcw_state = LC_WATCHDOG_DISABLED;
430
431         EXIT;
432 }
433 EXPORT_SYMBOL(lc_watchdog_disable);
434
435 void lc_watchdog_delete(struct lc_watchdog *lcw)
436 {
437         int dead;
438
439         ENTRY;
440         LASSERT(lcw != NULL);
441
442         del_timer(&lcw->lcw_timer);
443
444         lcw_update_time(lcw, "stopped");
445
446         spin_lock_bh(&lcw->lcw_lock);
447         spin_lock_bh(&lcw_pending_timers_lock);
448         if (unlikely(!list_empty(&lcw->lcw_list))) {
449                 list_del_init(&lcw->lcw_list);
450                 lcw->lcw_refcount--; /* -1 ref for pending list */
451         }
452
453         lcw->lcw_refcount--; /* -1 ref for owner */
454         dead = lcw->lcw_refcount == 0;
455         spin_unlock_bh(&lcw_pending_timers_lock);
456         spin_unlock_bh(&lcw->lcw_lock);
457
458         if (dead)
459                 LIBCFS_FREE(lcw, sizeof(*lcw));
460
461         mutex_lock(&lcw_refcount_mutex);
462         if (--lcw_refcount == 0)
463                 lcw_dispatch_stop();
464         mutex_unlock(&lcw_refcount_mutex);
465
466         EXIT;
467 }
468 EXPORT_SYMBOL(lc_watchdog_delete);
469
470 #else   /* !defined(WITH_WATCHDOG) */
471
472 struct lc_watchdog *lc_watchdog_add(int timeout,
473                                     void (*callback)(pid_t pid, void *),
474                                     void *data)
475 {
476         static struct lc_watchdog      watchdog;
477         return &watchdog;
478 }
479 EXPORT_SYMBOL(lc_watchdog_add);
480
481 void lc_watchdog_touch(struct lc_watchdog *lcw, int timeout)
482 {
483 }
484 EXPORT_SYMBOL(lc_watchdog_touch);
485
486 void lc_watchdog_disable(struct lc_watchdog *lcw)
487 {
488 }
489 EXPORT_SYMBOL(lc_watchdog_disable);
490
491 void lc_watchdog_delete(struct lc_watchdog *lcw)
492 {
493 }
494 EXPORT_SYMBOL(lc_watchdog_delete);
495
496 #endif