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(!central_netdata_to_push_data) {
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 unsigned int loop = 0;
337 while(!netdata_exit) {
339 debug(D_HEALTH, "Health monitoring iteration no %u started", loop);
341 int oldstate, runnable = 0;
342 time_t now = now_realtime_sec();
343 time_t next_run = now + min_run_every;
346 if(unlikely(pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate) != 0))
347 error("Cannot set pthread cancel state to DISABLE.");
352 rrdhost_foreach_read(host) {
353 if(unlikely(!host->health_enabled)) continue;
355 rrdhost_rdlock(host);
357 // the first loop is to lookup values from the db
358 for(rc = host->alarms; rc; rc = rc->next) {
359 if(unlikely(!rrdcalc_isrunnable(rc, now, &next_run))) {
360 if(unlikely(rc->rrdcalc_flags & RRDCALC_FLAG_RUNNABLE))
361 rc->rrdcalc_flags &= ~RRDCALC_FLAG_RUNNABLE;
366 rc->old_value = rc->value;
367 rc->rrdcalc_flags |= RRDCALC_FLAG_RUNNABLE;
369 // 1. if there is database lookup, do it
370 // 2. if there is calculation expression, run it
372 if(unlikely(RRDCALC_HAS_DB_LOOKUP(rc))) {
373 /* time_t old_db_timestamp = rc->db_before; */
374 int value_is_null = 0;
376 int ret = rrd2value(rc->rrdset, wb, &rc->value, rc->dimensions, 1, rc->after, rc->before, rc->group, rc->options, &rc->db_after, &rc->db_before, &value_is_null);
378 if(unlikely(ret != 200)) {
379 // database lookup failed
382 debug(D_HEALTH, "Health on host '%s', alarm '%s.%s': database lookup returned error %d", host->hostname, rc->chart ? rc->chart : "NOCHART", rc->name, ret);
384 if(unlikely(!(rc->rrdcalc_flags & RRDCALC_FLAG_DB_ERROR))) {
385 rc->rrdcalc_flags |= RRDCALC_FLAG_DB_ERROR;
386 error("Health on host '%s', alarm '%s.%s': database lookup returned error %d", host->hostname, rc->chart ? rc->chart : "NOCHART", rc->name, ret);
389 else if(unlikely(rc->rrdcalc_flags & RRDCALC_FLAG_DB_ERROR))
390 rc->rrdcalc_flags &= ~RRDCALC_FLAG_DB_ERROR;
392 /* - RRDCALC_FLAG_DB_STALE not currently used
393 if (unlikely(old_db_timestamp == rc->db_before)) {
396 debug(D_HEALTH, "Health on host '%s', alarm '%s.%s': database is stale", host->hostname, rc->chart?rc->chart:"NOCHART", rc->name);
398 if (unlikely(!(rc->rrdcalc_flags & RRDCALC_FLAG_DB_STALE))) {
399 rc->rrdcalc_flags |= RRDCALC_FLAG_DB_STALE;
400 error("Health on host '%s', alarm '%s.%s': database is stale", host->hostname, rc->chart?rc->chart:"NOCHART", rc->name);
403 else if (unlikely(rc->rrdcalc_flags & RRDCALC_FLAG_DB_STALE))
404 rc->rrdcalc_flags &= ~RRDCALC_FLAG_DB_STALE;
407 if(unlikely(value_is_null)) {
408 // collected value is null
412 debug(D_HEALTH, "Health on host '%s', alarm '%s.%s': database lookup returned empty value (possibly value is not collected yet)", host->hostname, rc->chart ? rc->chart : "NOCHART", rc->name);
414 if(unlikely(!(rc->rrdcalc_flags & RRDCALC_FLAG_DB_NAN))) {
415 rc->rrdcalc_flags |= RRDCALC_FLAG_DB_NAN;
416 error("Health on host '%s', alarm '%s.%s': database lookup returned empty value (possibly value is not collected yet)", host->hostname, rc->chart ? rc->chart : "NOCHART", rc->name);
419 else if(unlikely(rc->rrdcalc_flags & RRDCALC_FLAG_DB_NAN))
420 rc->rrdcalc_flags &= ~RRDCALC_FLAG_DB_NAN;
422 debug(D_HEALTH, "Health on host '%s', alarm '%s.%s': database lookup gave value " CALCULATED_NUMBER_FORMAT, host->hostname, rc->chart ? rc->chart : "NOCHART", rc->name, rc->value);
425 if(unlikely(rc->calculation)) {
426 if(unlikely(!expression_evaluate(rc->calculation))) {
427 // calculation failed
431 debug(D_HEALTH, "Health on host '%s', alarm '%s.%s': expression '%s' failed: %s", host->hostname, rc->chart ? rc->chart : "NOCHART", rc->name, rc->calculation->parsed_as, buffer_tostring(rc->calculation->error_msg));
433 if(unlikely(!(rc->rrdcalc_flags & RRDCALC_FLAG_CALC_ERROR))) {
434 rc->rrdcalc_flags |= RRDCALC_FLAG_CALC_ERROR;
435 error("Health on host '%s', alarm '%s.%s': expression '%s' failed: %s", rc->chart ? rc->chart : "NOCHART", host->hostname, rc->name, rc->calculation->parsed_as, buffer_tostring(rc->calculation->error_msg));
439 if(unlikely(rc->rrdcalc_flags & RRDCALC_FLAG_CALC_ERROR))
440 rc->rrdcalc_flags &= ~RRDCALC_FLAG_CALC_ERROR;
442 debug(D_HEALTH, "Health on host '%s', alarm '%s.%s': expression '%s' gave value "
443 CALCULATED_NUMBER_FORMAT
444 ": %s (source: %s)", host->hostname, rc->chart ? rc->chart : "NOCHART", rc->name
445 , rc->calculation->parsed_as, rc->calculation->result,
446 buffer_tostring(rc->calculation->error_msg), rc->source
449 rc->value = rc->calculation->result;
453 rrdhost_unlock(host);
455 if(unlikely(runnable && !netdata_exit)) {
456 rrdhost_rdlock(host);
458 for(rc = host->alarms; rc; rc = rc->next) {
459 if(unlikely(!(rc->rrdcalc_flags & RRDCALC_FLAG_RUNNABLE)))
462 int warning_status = RRDCALC_STATUS_UNDEFINED;
463 int critical_status = RRDCALC_STATUS_UNDEFINED;
465 if(likely(rc->warning)) {
466 if(unlikely(!expression_evaluate(rc->warning))) {
467 // calculation failed
469 debug(D_HEALTH, "Health on host '%s', alarm '%s.%s': warning expression failed with error: %s", host->hostname, rc->chart ? rc->chart : "NOCHART", rc->name, buffer_tostring(rc->warning->error_msg));
471 if(unlikely(!(rc->rrdcalc_flags & RRDCALC_FLAG_WARN_ERROR))) {
472 rc->rrdcalc_flags |= RRDCALC_FLAG_WARN_ERROR;
473 error("Health on host '%s', alarm '%s.%s': warning expression failed with error: %s", host->hostname, rc->chart ? rc->chart : "NOCHART", rc->name, buffer_tostring(rc->warning->error_msg));
477 if(unlikely(rc->rrdcalc_flags & RRDCALC_FLAG_WARN_ERROR))
478 rc->rrdcalc_flags &= ~RRDCALC_FLAG_WARN_ERROR;
480 debug(D_HEALTH, "Health on host '%s', alarm '%s.%s': warning expression gave value " CALCULATED_NUMBER_FORMAT ": %s (source: %s)", host->hostname, rc->chart ? rc->chart : "NOCHART", rc->name, rc->warning->result, buffer_tostring(rc->warning->error_msg), rc->source);
482 warning_status = rrdcalc_value2status(rc->warning->result);
486 if(likely(rc->critical)) {
487 if(unlikely(!expression_evaluate(rc->critical))) {
488 // calculation failed
490 debug(D_HEALTH, "Health on host '%s', alarm '%s.%s': critical expression failed with error: %s", host->hostname, rc->chart ? rc->chart : "NOCHART", rc->name, buffer_tostring(rc->critical->error_msg));
492 if(unlikely(!(rc->rrdcalc_flags & RRDCALC_FLAG_CRIT_ERROR))) {
493 rc->rrdcalc_flags |= RRDCALC_FLAG_CRIT_ERROR;
494 error("Health on host '%s', alarm '%s.%s': critical expression failed with error: %s", host->hostname, rc->chart ? rc->chart : "NOCHART", rc->name, buffer_tostring(rc->critical->error_msg));
498 if(unlikely(rc->rrdcalc_flags & RRDCALC_FLAG_CRIT_ERROR))
499 rc->rrdcalc_flags &= ~RRDCALC_FLAG_CRIT_ERROR;
501 debug(D_HEALTH, "Health on host '%s', alarm '%s.%s': critical expression gave value " CALCULATED_NUMBER_FORMAT ": %s (source: %s)", host->hostname, rc->chart ? rc->chart : "NOCHART", rc->name , rc->critical->result, buffer_tostring(rc->critical->error_msg), rc->source);
503 critical_status = rrdcalc_value2status(rc->critical->result);
507 int status = RRDCALC_STATUS_UNDEFINED;
509 switch(warning_status) {
510 case RRDCALC_STATUS_CLEAR:
511 status = RRDCALC_STATUS_CLEAR;
514 case RRDCALC_STATUS_RAISED:
515 status = RRDCALC_STATUS_WARNING;
522 switch(critical_status) {
523 case RRDCALC_STATUS_CLEAR:
524 if(status == RRDCALC_STATUS_UNDEFINED)
525 status = RRDCALC_STATUS_CLEAR;
528 case RRDCALC_STATUS_RAISED:
529 status = RRDCALC_STATUS_CRITICAL;
536 if(status != rc->status) {
539 if(now > rc->delay_up_to_timestamp) {
540 rc->delay_up_current = rc->delay_up_duration;
541 rc->delay_down_current = rc->delay_down_duration;
543 rc->delay_up_to_timestamp = 0;
546 rc->delay_up_current = (int) (rc->delay_up_current * rc->delay_multiplier);
547 if(rc->delay_up_current > rc->delay_max_duration)
548 rc->delay_up_current = rc->delay_max_duration;
550 rc->delay_down_current = (int) (rc->delay_down_current * rc->delay_multiplier);
551 if(rc->delay_down_current > rc->delay_max_duration)
552 rc->delay_down_current = rc->delay_max_duration;
555 if(status > rc->status)
556 delay = rc->delay_up_current;
558 delay = rc->delay_down_current;
560 // COMMENTED: because we do need to send raising alarms
561 // if(now + delay < rc->delay_up_to_timestamp)
562 // delay = (int)(rc->delay_up_to_timestamp - now);
564 rc->delay_last = delay;
565 rc->delay_up_to_timestamp = now + delay;
567 host, rc->id, rc->next_event_id++, now, rc->name, rc->rrdset->id
568 , rc->rrdset->family, rc->exec, rc->recipient, now - rc->last_status_change
569 , rc->old_value, rc->value, rc->status, status, rc->source, rc->units, rc->info
570 , rc->delay_last, (rc->options & RRDCALC_FLAG_NO_CLEAR_NOTIFICATION)
571 ? HEALTH_ENTRY_FLAG_NO_CLEAR_NOTIFICATION : 0
573 rc->last_status_change = now;
577 rc->last_updated = now;
578 rc->next_update = now + rc->update_every;
580 if(next_run > rc->next_update)
581 next_run = rc->next_update;
584 rrdhost_unlock(host);
587 if(unlikely(netdata_exit))
590 // execute notifications
592 health_alarm_log_process(host);
594 if(unlikely(netdata_exit))
601 if(unlikely(pthread_setcancelstate(oldstate, NULL) != 0))
602 error("Cannot set pthread cancel state to RESTORE (%d).", oldstate);
604 if(unlikely(netdata_exit))
607 now = now_realtime_sec();
609 debug(D_HEALTH, "Health monitoring iteration no %u done. Next iteration in %d secs", loop, (int) (next_run - now));
610 sleep_usec(USEC_PER_SEC * (usec_t) (next_run - now));
613 debug(D_HEALTH, "Health monitoring iteration no %u done. Next iteration now", loop);
618 info("HEALTH thread exiting");
620 static_thread->enabled = 0;