1 #define NETDATA_HEALTH_INTERNALS
4 int default_health_enabled = 1;
6 // ----------------------------------------------------------------------------
7 // health initialization
9 inline char *health_config_dir(void) {
10 char buffer[FILENAME_MAX + 1];
11 snprintfz(buffer, FILENAME_MAX, "%s/health.d", netdata_configured_config_dir);
12 return config_get("health", "health configuration directory", buffer);
15 void health_init(void) {
16 debug(D_HEALTH, "Health configuration initializing");
18 if(!rrdpush_exclusive) {
19 if(!(default_health_enabled = config_get_boolean("health", "enabled", 1))) {
20 debug(D_HEALTH, "Health is disabled.");
25 info("Health is disabled - setup alarms at the central netdata.");
26 config_set_boolean("health", "enabled", 0);
27 default_health_enabled = 0;
31 // ----------------------------------------------------------------------------
32 // re-load health configuration
34 void health_reload_host(RRDHOST *host) {
35 char *path = health_config_dir();
37 // free all running alarms
40 while(host->templates)
41 rrdcalctemplate_free(host, host->templates);
44 rrdcalc_free(host, host->alarms);
48 // invalidate all previous entries in the alarm log
50 for(t = host->health_log.alarms ; t ; t = t->next) {
51 if(t->new_status != RRDCALC_STATUS_REMOVED)
52 t->flags |= HEALTH_ENTRY_FLAG_UPDATED;
56 // reset all thresholds to all charts
58 rrdset_foreach_read(st, host) {
64 // load the new alarms
66 health_readdir(host, path);
68 // link the loaded alarms to their charts
69 rrdset_foreach_write(st, host) {
70 rrdsetcalc_link_matching(st);
71 rrdcalctemplate_link_matching(st);
77 void health_reload(void) {
82 rrdhost_foreach_read(host)
83 health_reload_host(host);
88 // ----------------------------------------------------------------------------
89 // health main thread and friends
91 static inline int rrdcalc_value2status(calculated_number n) {
92 if(isnan(n) || isinf(n)) return RRDCALC_STATUS_UNDEFINED;
93 if(n) return RRDCALC_STATUS_RAISED;
94 return RRDCALC_STATUS_CLEAR;
97 #define ALARM_EXEC_COMMAND_LENGTH 8192
99 static inline void health_alarm_execute(RRDHOST *host, ALARM_ENTRY *ae) {
100 ae->flags |= HEALTH_ENTRY_FLAG_PROCESSED;
102 if(unlikely(ae->new_status < RRDCALC_STATUS_CLEAR)) {
103 // do not send notifications for internal statuses
104 debug(D_HEALTH, "Health not sending notification for alarm '%s.%s' status %s (internal statuses)", ae->chart, ae->name, rrdcalc_status2string(ae->new_status));
108 if(unlikely(ae->new_status <= RRDCALC_STATUS_CLEAR && (ae->flags & HEALTH_ENTRY_FLAG_NO_CLEAR_NOTIFICATION))) {
109 // do not send notifications for disabled statuses
110 debug(D_HEALTH, "Health not sending notification for alarm '%s.%s' status %s (it has no-clear-notification enabled)", ae->chart, ae->name, rrdcalc_status2string(ae->new_status));
111 // mark it as run, so that we will send the same alarm if it happens again
115 // find the previous notification for the same alarm
116 // which we have run the exec script
117 // exception: alarms with HEALTH_ENTRY_FLAG_NO_CLEAR_NOTIFICATION set
118 if(likely(!(ae->flags & HEALTH_ENTRY_FLAG_NO_CLEAR_NOTIFICATION))) {
119 uint32_t id = ae->alarm_id;
121 for(t = ae->next; t ; t = t->next) {
122 if(t->alarm_id == id && t->flags & HEALTH_ENTRY_FLAG_EXEC_RUN)
127 // we have executed this alarm notification in the past
128 if(t && t->new_status == ae->new_status) {
129 // don't send the notification for the same status again
130 debug(D_HEALTH, "Health not sending again notification for alarm '%s.%s' status %s", ae->chart, ae->name
131 , rrdcalc_status2string(ae->new_status));
136 // we have not executed this alarm notification in the past
137 // so, don't send CLEAR notifications
138 if(unlikely(ae->new_status == RRDCALC_STATUS_CLEAR)) {
139 debug(D_HEALTH, "Health not sending notification for first initialization of alarm '%s.%s' status %s"
140 , ae->chart, ae->name, rrdcalc_status2string(ae->new_status));
146 static char command_to_run[ALARM_EXEC_COMMAND_LENGTH + 1];
149 const char *exec = (ae->exec) ? ae->exec : host->health_default_exec;
150 const char *recipient = (ae->recipient) ? ae->recipient : host->health_default_recipient;
152 snprintfz(command_to_run, ALARM_EXEC_COMMAND_LENGTH, "exec %s '%s' '%s' '%u' '%u' '%u' '%lu' '%s' '%s' '%s' '%s' '%s' '%0.0Lf' '%0.0Lf' '%s' '%u' '%u' '%s' '%s' '%s' '%s'",
159 (unsigned long)ae->when,
161 ae->chart?ae->chart:"NOCAHRT",
162 ae->family?ae->family:"NOFAMILY",
163 rrdcalc_status2string(ae->new_status),
164 rrdcalc_status2string(ae->old_status),
167 ae->source?ae->source:"UNKNOWN",
168 (uint32_t)ae->duration,
169 (uint32_t)ae->non_clear_duration,
170 ae->units?ae->units:"",
171 ae->info?ae->info:"",
172 ae->new_value_string,
176 ae->flags |= HEALTH_ENTRY_FLAG_EXEC_RUN;
177 ae->exec_run_timestamp = now_realtime_sec();
179 debug(D_HEALTH, "executing command '%s'", command_to_run);
180 FILE *fp = mypopen(command_to_run, &command_pid);
182 error("HEALTH: Cannot popen(\"%s\", \"r\").", command_to_run);
185 debug(D_HEALTH, "HEALTH reading from command");
186 char *s = fgets(command_to_run, FILENAME_MAX, fp);
188 ae->exec_code = mypclose(fp, command_pid);
189 debug(D_HEALTH, "done executing command - returned with code %d", ae->exec_code);
191 if(ae->exec_code != 0)
192 ae->flags |= HEALTH_ENTRY_FLAG_EXEC_FAILED;
195 health_alarm_log_save(host, ae);
199 static inline void health_process_notifications(RRDHOST *host, ALARM_ENTRY *ae) {
200 debug(D_HEALTH, "Health alarm '%s.%s' = %0.2Lf - changed status from %s to %s",
201 ae->chart?ae->chart:"NOCHART", ae->name,
203 rrdcalc_status2string(ae->old_status),
204 rrdcalc_status2string(ae->new_status)
207 health_alarm_execute(host, ae);
210 static inline void health_alarm_log_process(RRDHOST *host) {
211 static uint32_t stop_at_id = 0;
212 uint32_t first_waiting = (host->health_log.alarms)?host->health_log.alarms->unique_id:0;
213 time_t now = now_realtime_sec();
215 pthread_rwlock_rdlock(&host->health_log.alarm_log_rwlock);
218 for(ae = host->health_log.alarms; ae && ae->unique_id >= stop_at_id ; ae = ae->next) {
220 !(ae->flags & HEALTH_ENTRY_FLAG_PROCESSED) &&
221 !(ae->flags & HEALTH_ENTRY_FLAG_UPDATED)
224 if(unlikely(ae->unique_id < first_waiting))
225 first_waiting = ae->unique_id;
227 if(likely(now >= ae->delay_up_to_timestamp))
228 health_process_notifications(host, ae);
232 // remember this for the next iteration
233 stop_at_id = first_waiting;
235 pthread_rwlock_unlock(&host->health_log.alarm_log_rwlock);
237 if(host->health_log.count <= host->health_log.max)
240 // cleanup excess entries in the log
241 pthread_rwlock_wrlock(&host->health_log.alarm_log_rwlock);
243 ALARM_ENTRY *last = NULL;
244 unsigned int count = host->health_log.max * 2 / 3;
245 for(ae = host->health_log.alarms; ae && count ; count--, last = ae, ae = ae->next) ;
247 if(ae && last && last->next == ae)
253 debug(D_HEALTH, "Health removing alarm log entry with id: %u", ae->unique_id);
255 ALARM_ENTRY *t = ae->next;
257 health_alarm_log_free_one_nochecks_nounlink(ae);
260 host->health_log.count--;
263 pthread_rwlock_unlock(&host->health_log.alarm_log_rwlock);
266 static inline int rrdcalc_isrunnable(RRDCALC *rc, time_t now, time_t *next_run) {
267 if(unlikely(!rc->rrdset)) {
268 debug(D_HEALTH, "Health not running alarm '%s.%s'. It is not linked to a chart.", rc->chart?rc->chart:"NOCHART", rc->name);
272 if(unlikely(rc->next_update > now)) {
273 if (unlikely(*next_run > rc->next_update)) {
274 // update the next_run time of the main loop
275 // to run this alarm precisely the time required
276 *next_run = rc->next_update;
279 debug(D_HEALTH, "Health not examining alarm '%s.%s' yet (will do in %d secs).", rc->chart?rc->chart:"NOCHART", rc->name, (int) (rc->next_update - now));
283 if(unlikely(!rc->update_every)) {
284 debug(D_HEALTH, "Health not running alarm '%s.%s'. It does not have an update frequency", rc->chart?rc->chart:"NOCHART", rc->name);
288 if(unlikely(!rc->rrdset->last_collected_time.tv_sec || rc->rrdset->counter_done < 2)) {
289 debug(D_HEALTH, "Health not running alarm '%s.%s'. Chart is not fully collected yet.", rc->chart?rc->chart:"NOCHART", rc->name);
293 int update_every = rc->rrdset->update_every;
294 time_t first = rrdset_first_entry_t(rc->rrdset);
295 time_t last = rrdset_last_entry_t(rc->rrdset);
297 if(unlikely(now + update_every < first /* || now - update_every > last */)) {
299 , "Health not examining alarm '%s.%s' yet (wanted time is out of bounds - we need %lu but got %lu - %lu)."
300 , rc->chart ? rc->chart : "NOCHART", rc->name, (unsigned long) now, (unsigned long) first
301 , (unsigned long) last);
305 if(RRDCALC_HAS_DB_LOOKUP(rc)) {
306 time_t needed = now + rc->before + rc->after;
308 if(needed + update_every < first || needed - update_every > last) {
310 , "Health not examining alarm '%s.%s' yet (not enough data yet - we need %lu but got %lu - %lu)."
311 , rc->chart ? rc->chart : "NOCHART", rc->name, (unsigned long) needed, (unsigned long) first
312 , (unsigned long) last);
320 void *health_main(void *ptr) {
321 struct netdata_static_thread *static_thread = (struct netdata_static_thread *)ptr;
323 info("HEALTH thread created with task id %d", gettid());
325 if(pthread_setcanceltype(PTHREAD_CANCEL_DEFERRED, NULL) != 0)
326 error("Cannot set pthread cancel type to DEFERRED.");
328 if(pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, NULL) != 0)
329 error("Cannot set pthread cancel state to ENABLE.");
331 int min_run_every = (int)config_get_number("health", "run at least every seconds", 10);
332 if(min_run_every < 1) min_run_every = 1;
334 BUFFER *wb = buffer_create(100);
336 time_t now = now_realtime_sec();
337 time_t now_boottime = now_boottime_sec();
338 time_t last_now = now;
339 time_t last_now_boottime = now_boottime;
340 time_t hibernation_delay = config_get_number("health", "postpone alarms during hibernation for seconds", 60);
342 unsigned int loop = 0;
343 while(!netdata_exit) {
345 debug(D_HEALTH, "Health monitoring iteration no %u started", loop);
347 int oldstate, runnable = 0, apply_hibernation_delay = 0;
348 time_t next_run = now + min_run_every;
351 // detect if boottime and realtime have twice the difference
352 // in which case we assume the system was just waken from hibernation
353 if(unlikely(now - last_now > 2 * (now_boottime - last_now_boottime)))
354 apply_hibernation_delay = 1;
357 last_now_boottime = now_boottime;
359 if(unlikely(pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate) != 0))
360 error("Cannot set pthread cancel state to DISABLE.");
365 rrdhost_foreach_read(host) {
366 if(unlikely(apply_hibernation_delay)) {
368 info("Postponing alarm checks for %ld seconds, on host '%s', due to boottime discrepancy (realtime dt: %ld, boottime dt: %ld)."
371 , (long)(now - last_now)
372 , (long)(now_boottime - last_now_boottime)
375 host->health_delay_up_to = now + hibernation_delay;
378 if(unlikely(!host->health_enabled || now < host->health_delay_up_to))
381 rrdhost_rdlock(host);
383 // the first loop is to lookup values from the db
384 for(rc = host->alarms; rc; rc = rc->next) {
385 if(unlikely(!rrdcalc_isrunnable(rc, now, &next_run))) {
386 if(unlikely(rc->rrdcalc_flags & RRDCALC_FLAG_RUNNABLE))
387 rc->rrdcalc_flags &= ~RRDCALC_FLAG_RUNNABLE;
392 rc->old_value = rc->value;
393 rc->rrdcalc_flags |= RRDCALC_FLAG_RUNNABLE;
395 // ------------------------------------------------------------
396 // if there is database lookup, do it
398 if(unlikely(RRDCALC_HAS_DB_LOOKUP(rc))) {
399 /* time_t old_db_timestamp = rc->db_before; */
400 int value_is_null = 0;
402 int ret = rrd2value(rc->rrdset
416 if(unlikely(ret != 200)) {
417 // database lookup failed
419 rc->rrdcalc_flags |= RRDCALC_FLAG_DB_ERROR;
422 , "Health on host '%s', alarm '%s.%s': database lookup returned error %d"
424 , rc->chart ? rc->chart : "NOCHART"
430 rc->rrdcalc_flags &= ~RRDCALC_FLAG_DB_ERROR;
432 /* - RRDCALC_FLAG_DB_STALE not currently used
433 if (unlikely(old_db_timestamp == rc->db_before)) {
436 debug(D_HEALTH, "Health on host '%s', alarm '%s.%s': database is stale", host->hostname, rc->chart?rc->chart:"NOCHART", rc->name);
438 if (unlikely(!(rc->rrdcalc_flags & RRDCALC_FLAG_DB_STALE))) {
439 rc->rrdcalc_flags |= RRDCALC_FLAG_DB_STALE;
440 error("Health on host '%s', alarm '%s.%s': database is stale", host->hostname, rc->chart?rc->chart:"NOCHART", rc->name);
443 else if (unlikely(rc->rrdcalc_flags & RRDCALC_FLAG_DB_STALE))
444 rc->rrdcalc_flags &= ~RRDCALC_FLAG_DB_STALE;
447 if(unlikely(value_is_null)) {
448 // collected value is null
450 rc->rrdcalc_flags |= RRDCALC_FLAG_DB_NAN;
453 , "Health on host '%s', alarm '%s.%s': database lookup returned empty value (possibly value is not collected yet)"
455 , rc->chart ? rc->chart : "NOCHART"
460 rc->rrdcalc_flags &= ~RRDCALC_FLAG_DB_NAN;
463 , "Health on host '%s', alarm '%s.%s': database lookup gave value " CALCULATED_NUMBER_FORMAT
465 , rc->chart ? rc->chart : "NOCHART"
471 // ------------------------------------------------------------
472 // if there is calculation expression, run it
474 if(unlikely(rc->calculation)) {
475 if(unlikely(!expression_evaluate(rc->calculation))) {
476 // calculation failed
478 rc->rrdcalc_flags |= RRDCALC_FLAG_CALC_ERROR;
481 , "Health on host '%s', alarm '%s.%s': expression '%s' failed: %s"
483 , rc->chart ? rc->chart : "NOCHART"
485 , rc->calculation->parsed_as
486 , buffer_tostring(rc->calculation->error_msg)
490 rc->rrdcalc_flags &= ~RRDCALC_FLAG_CALC_ERROR;
492 debug(D_HEALTH, "Health on host '%s', alarm '%s.%s': expression '%s' gave value " CALCULATED_NUMBER_FORMAT ": %s (source: %s)"
494 , rc->chart ? rc->chart : "NOCHART"
496 , rc->calculation->parsed_as
497 , rc->calculation->result
498 , buffer_tostring(rc->calculation->error_msg)
502 rc->value = rc->calculation->result;
506 rrdhost_unlock(host);
508 if(unlikely(runnable && !netdata_exit)) {
509 rrdhost_rdlock(host);
511 for(rc = host->alarms; rc; rc = rc->next) {
512 if(unlikely(!(rc->rrdcalc_flags & RRDCALC_FLAG_RUNNABLE)))
515 int warning_status = RRDCALC_STATUS_UNDEFINED;
516 int critical_status = RRDCALC_STATUS_UNDEFINED;
518 // --------------------------------------------------------
519 // check the warning expression
521 if(likely(rc->warning)) {
522 if(unlikely(!expression_evaluate(rc->warning))) {
523 // calculation failed
524 rc->rrdcalc_flags |= RRDCALC_FLAG_WARN_ERROR;
527 , "Health on host '%s', alarm '%s.%s': warning expression failed with error: %s"
529 , rc->chart ? rc->chart : "NOCHART"
531 , buffer_tostring(rc->warning->error_msg)
535 rc->rrdcalc_flags &= ~RRDCALC_FLAG_WARN_ERROR;
537 , "Health on host '%s', alarm '%s.%s': warning expression gave value " CALCULATED_NUMBER_FORMAT ": %s (source: %s)"
539 , rc->chart ? rc->chart : "NOCHART"
541 , rc->warning->result
542 , buffer_tostring(rc->warning->error_msg)
545 warning_status = rrdcalc_value2status(rc->warning->result);
549 // --------------------------------------------------------
550 // check the critical expression
552 if(likely(rc->critical)) {
553 if(unlikely(!expression_evaluate(rc->critical))) {
554 // calculation failed
555 rc->rrdcalc_flags |= RRDCALC_FLAG_CRIT_ERROR;
558 , "Health on host '%s', alarm '%s.%s': critical expression failed with error: %s"
560 , rc->chart ? rc->chart : "NOCHART"
562 , buffer_tostring(rc->critical->error_msg)
566 rc->rrdcalc_flags &= ~RRDCALC_FLAG_CRIT_ERROR;
568 , "Health on host '%s', alarm '%s.%s': critical expression gave value " CALCULATED_NUMBER_FORMAT ": %s (source: %s)"
570 , rc->chart ? rc->chart : "NOCHART"
572 , rc->critical->result
573 , buffer_tostring(rc->critical->error_msg)
576 critical_status = rrdcalc_value2status(rc->critical->result);
580 // --------------------------------------------------------
581 // decide the final alarm status
583 int status = RRDCALC_STATUS_UNDEFINED;
585 switch(warning_status) {
586 case RRDCALC_STATUS_CLEAR:
587 status = RRDCALC_STATUS_CLEAR;
590 case RRDCALC_STATUS_RAISED:
591 status = RRDCALC_STATUS_WARNING;
598 switch(critical_status) {
599 case RRDCALC_STATUS_CLEAR:
600 if(status == RRDCALC_STATUS_UNDEFINED)
601 status = RRDCALC_STATUS_CLEAR;
604 case RRDCALC_STATUS_RAISED:
605 status = RRDCALC_STATUS_CRITICAL;
612 // --------------------------------------------------------
613 // check if the new status and the old differ
615 if(status != rc->status) {
618 // apply trigger hysteresis
620 if(now > rc->delay_up_to_timestamp) {
621 rc->delay_up_current = rc->delay_up_duration;
622 rc->delay_down_current = rc->delay_down_duration;
624 rc->delay_up_to_timestamp = 0;
627 rc->delay_up_current = (int) (rc->delay_up_current * rc->delay_multiplier);
628 if(rc->delay_up_current > rc->delay_max_duration)
629 rc->delay_up_current = rc->delay_max_duration;
631 rc->delay_down_current = (int) (rc->delay_down_current * rc->delay_multiplier);
632 if(rc->delay_down_current > rc->delay_max_duration)
633 rc->delay_down_current = rc->delay_max_duration;
636 if(status > rc->status)
637 delay = rc->delay_up_current;
639 delay = rc->delay_down_current;
641 // COMMENTED: because we do need to send raising alarms
642 // if(now + delay < rc->delay_up_to_timestamp)
643 // delay = (int)(rc->delay_up_to_timestamp - now);
645 rc->delay_last = delay;
646 rc->delay_up_to_timestamp = now + delay;
648 // add the alarm into the log
653 , rc->next_event_id++
660 , now - rc->last_status_change
669 , (rc->options & RRDCALC_FLAG_NO_CLEAR_NOTIFICATION) ? HEALTH_ENTRY_FLAG_NO_CLEAR_NOTIFICATION : 0
672 rc->last_status_change = now;
676 rc->last_updated = now;
677 rc->next_update = now + rc->update_every;
679 if(next_run > rc->next_update)
680 next_run = rc->next_update;
683 rrdhost_unlock(host);
686 if(unlikely(netdata_exit))
689 // execute notifications
691 health_alarm_log_process(host);
693 if(unlikely(netdata_exit))
696 } /* rrdhost_foreach */
700 if(unlikely(pthread_setcancelstate(oldstate, NULL) != 0))
701 error("Cannot set pthread cancel state to RESTORE (%d).", oldstate);
703 if(unlikely(netdata_exit))
706 now = now_realtime_sec();
708 debug(D_HEALTH, "Health monitoring iteration no %u done. Next iteration in %d secs", loop, (int) (next_run - now));
709 sleep_usec(USEC_PER_SEC * (usec_t) (next_run - now));
710 now = now_realtime_sec();
713 debug(D_HEALTH, "Health monitoring iteration no %u done. Next iteration now", loop);
715 now_boottime = now_boottime_sec();
721 info("HEALTH thread exiting");
723 static_thread->enabled = 0;