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;
30 char pathname[FILENAME_MAX + 1];
31 snprintfz(pathname, FILENAME_MAX, "%s/health", netdata_configured_varlib_dir);
32 if(mkdir(pathname, 0770) == -1 && errno != EEXIST)
33 fatal("Cannot create directory '%s'.", pathname);
36 // ----------------------------------------------------------------------------
37 // re-load health configuration
39 void health_reload_host(RRDHOST *host) {
40 char *path = health_config_dir();
42 // free all running alarms
45 while(host->templates)
46 rrdcalctemplate_free(host, host->templates);
49 rrdcalc_free(host, host->alarms);
53 // invalidate all previous entries in the alarm log
55 for(t = host->health_log.alarms ; t ; t = t->next) {
56 if(t->new_status != RRDCALC_STATUS_REMOVED)
57 t->flags |= HEALTH_ENTRY_FLAG_UPDATED;
61 // reset all thresholds to all charts
63 rrdset_foreach_read(st, host) {
69 // load the new alarms
71 health_readdir(host, path);
73 // link the loaded alarms to their charts
74 rrdset_foreach_write(st, host) {
75 rrdsetcalc_link_matching(st);
76 rrdcalctemplate_link_matching(st);
82 void health_reload(void) {
87 rrdhost_foreach_read(host)
88 health_reload_host(host);
93 // ----------------------------------------------------------------------------
94 // health main thread and friends
96 static inline int rrdcalc_value2status(calculated_number n) {
97 if(isnan(n) || isinf(n)) return RRDCALC_STATUS_UNDEFINED;
98 if(n) return RRDCALC_STATUS_RAISED;
99 return RRDCALC_STATUS_CLEAR;
102 #define ALARM_EXEC_COMMAND_LENGTH 8192
104 static inline void health_alarm_execute(RRDHOST *host, ALARM_ENTRY *ae) {
105 ae->flags |= HEALTH_ENTRY_FLAG_PROCESSED;
107 if(unlikely(ae->new_status < RRDCALC_STATUS_CLEAR)) {
108 // do not send notifications for internal statuses
109 debug(D_HEALTH, "Health not sending notification for alarm '%s.%s' status %s (internal statuses)", ae->chart, ae->name, rrdcalc_status2string(ae->new_status));
113 if(unlikely(ae->new_status <= RRDCALC_STATUS_CLEAR && (ae->flags & HEALTH_ENTRY_FLAG_NO_CLEAR_NOTIFICATION))) {
114 // do not send notifications for disabled statuses
115 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));
116 // mark it as run, so that we will send the same alarm if it happens again
120 // find the previous notification for the same alarm
121 // which we have run the exec script
122 // exception: alarms with HEALTH_ENTRY_FLAG_NO_CLEAR_NOTIFICATION set
123 if(likely(!(ae->flags & HEALTH_ENTRY_FLAG_NO_CLEAR_NOTIFICATION))) {
124 uint32_t id = ae->alarm_id;
126 for(t = ae->next; t ; t = t->next) {
127 if(t->alarm_id == id && t->flags & HEALTH_ENTRY_FLAG_EXEC_RUN)
132 // we have executed this alarm notification in the past
133 if(t && t->new_status == ae->new_status) {
134 // don't send the notification for the same status again
135 debug(D_HEALTH, "Health not sending again notification for alarm '%s.%s' status %s", ae->chart, ae->name
136 , rrdcalc_status2string(ae->new_status));
141 // we have not executed this alarm notification in the past
142 // so, don't send CLEAR notifications
143 if(unlikely(ae->new_status == RRDCALC_STATUS_CLEAR)) {
144 debug(D_HEALTH, "Health not sending notification for first initialization of alarm '%s.%s' status %s"
145 , ae->chart, ae->name, rrdcalc_status2string(ae->new_status));
151 static char command_to_run[ALARM_EXEC_COMMAND_LENGTH + 1];
154 const char *exec = (ae->exec) ? ae->exec : host->health_default_exec;
155 const char *recipient = (ae->recipient) ? ae->recipient : host->health_default_recipient;
157 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'",
164 (unsigned long)ae->when,
166 ae->chart?ae->chart:"NOCAHRT",
167 ae->family?ae->family:"NOFAMILY",
168 rrdcalc_status2string(ae->new_status),
169 rrdcalc_status2string(ae->old_status),
172 ae->source?ae->source:"UNKNOWN",
173 (uint32_t)ae->duration,
174 (uint32_t)ae->non_clear_duration,
175 ae->units?ae->units:"",
176 ae->info?ae->info:"",
177 ae->new_value_string,
181 ae->flags |= HEALTH_ENTRY_FLAG_EXEC_RUN;
182 ae->exec_run_timestamp = now_realtime_sec();
184 debug(D_HEALTH, "executing command '%s'", command_to_run);
185 FILE *fp = mypopen(command_to_run, &command_pid);
187 error("HEALTH: Cannot popen(\"%s\", \"r\").", command_to_run);
190 debug(D_HEALTH, "HEALTH reading from command");
191 char *s = fgets(command_to_run, FILENAME_MAX, fp);
193 ae->exec_code = mypclose(fp, command_pid);
194 debug(D_HEALTH, "done executing command - returned with code %d", ae->exec_code);
196 if(ae->exec_code != 0)
197 ae->flags |= HEALTH_ENTRY_FLAG_EXEC_FAILED;
200 health_alarm_log_save(host, ae);
204 static inline void health_process_notifications(RRDHOST *host, ALARM_ENTRY *ae) {
205 debug(D_HEALTH, "Health alarm '%s.%s' = %0.2Lf - changed status from %s to %s",
206 ae->chart?ae->chart:"NOCHART", ae->name,
208 rrdcalc_status2string(ae->old_status),
209 rrdcalc_status2string(ae->new_status)
212 health_alarm_execute(host, ae);
215 static inline void health_alarm_log_process(RRDHOST *host) {
216 static uint32_t stop_at_id = 0;
217 uint32_t first_waiting = (host->health_log.alarms)?host->health_log.alarms->unique_id:0;
218 time_t now = now_realtime_sec();
220 pthread_rwlock_rdlock(&host->health_log.alarm_log_rwlock);
223 for(ae = host->health_log.alarms; ae && ae->unique_id >= stop_at_id ; ae = ae->next) {
225 !(ae->flags & HEALTH_ENTRY_FLAG_PROCESSED) &&
226 !(ae->flags & HEALTH_ENTRY_FLAG_UPDATED)
229 if(unlikely(ae->unique_id < first_waiting))
230 first_waiting = ae->unique_id;
232 if(likely(now >= ae->delay_up_to_timestamp))
233 health_process_notifications(host, ae);
237 // remember this for the next iteration
238 stop_at_id = first_waiting;
240 pthread_rwlock_unlock(&host->health_log.alarm_log_rwlock);
242 if(host->health_log.count <= host->health_log.max)
245 // cleanup excess entries in the log
246 pthread_rwlock_wrlock(&host->health_log.alarm_log_rwlock);
248 ALARM_ENTRY *last = NULL;
249 unsigned int count = host->health_log.max * 2 / 3;
250 for(ae = host->health_log.alarms; ae && count ; count--, last = ae, ae = ae->next) ;
252 if(ae && last && last->next == ae)
258 debug(D_HEALTH, "Health removing alarm log entry with id: %u", ae->unique_id);
260 ALARM_ENTRY *t = ae->next;
262 health_alarm_log_free_one_nochecks_nounlink(ae);
265 host->health_log.count--;
268 pthread_rwlock_unlock(&host->health_log.alarm_log_rwlock);
271 static inline int rrdcalc_isrunnable(RRDCALC *rc, time_t now, time_t *next_run) {
272 if(unlikely(!rc->rrdset)) {
273 debug(D_HEALTH, "Health not running alarm '%s.%s'. It is not linked to a chart.", rc->chart?rc->chart:"NOCHART", rc->name);
277 if(unlikely(rc->next_update > now)) {
278 if (unlikely(*next_run > rc->next_update)) {
279 // update the next_run time of the main loop
280 // to run this alarm precisely the time required
281 *next_run = rc->next_update;
284 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));
288 if(unlikely(!rc->update_every)) {
289 debug(D_HEALTH, "Health not running alarm '%s.%s'. It does not have an update frequency", rc->chart?rc->chart:"NOCHART", rc->name);
293 if(unlikely(!rc->rrdset->last_collected_time.tv_sec || rc->rrdset->counter_done < 2)) {
294 debug(D_HEALTH, "Health not running alarm '%s.%s'. Chart is not fully collected yet.", rc->chart?rc->chart:"NOCHART", rc->name);
298 int update_every = rc->rrdset->update_every;
299 time_t first = rrdset_first_entry_t(rc->rrdset);
300 time_t last = rrdset_last_entry_t(rc->rrdset);
302 if(unlikely(now + update_every < first /* || now - update_every > last */)) {
304 , "Health not examining alarm '%s.%s' yet (wanted time is out of bounds - we need %lu but got %lu - %lu)."
305 , rc->chart ? rc->chart : "NOCHART", rc->name, (unsigned long) now, (unsigned long) first
306 , (unsigned long) last);
310 if(RRDCALC_HAS_DB_LOOKUP(rc)) {
311 time_t needed = now + rc->before + rc->after;
313 if(needed + update_every < first || needed - update_every > last) {
315 , "Health not examining alarm '%s.%s' yet (not enough data yet - we need %lu but got %lu - %lu)."
316 , rc->chart ? rc->chart : "NOCHART", rc->name, (unsigned long) needed, (unsigned long) first
317 , (unsigned long) last);
325 void *health_main(void *ptr) {
326 struct netdata_static_thread *static_thread = (struct netdata_static_thread *)ptr;
328 info("HEALTH thread created with task id %d", gettid());
330 if(pthread_setcanceltype(PTHREAD_CANCEL_DEFERRED, NULL) != 0)
331 error("Cannot set pthread cancel type to DEFERRED.");
333 if(pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, NULL) != 0)
334 error("Cannot set pthread cancel state to ENABLE.");
336 int min_run_every = (int)config_get_number("health", "run at least every seconds", 10);
337 if(min_run_every < 1) min_run_every = 1;
339 BUFFER *wb = buffer_create(100);
341 unsigned int loop = 0;
342 while(!netdata_exit) {
344 debug(D_HEALTH, "Health monitoring iteration no %u started", loop);
346 int oldstate, runnable = 0;
347 time_t now = now_realtime_sec();
348 time_t next_run = now + min_run_every;
351 if(unlikely(pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate) != 0))
352 error("Cannot set pthread cancel state to DISABLE.");
357 rrdhost_foreach_read(host) {
358 if(unlikely(!host->health_enabled)) continue;
360 rrdhost_rdlock(host);
362 // the first loop is to lookup values from the db
363 for(rc = host->alarms; rc; rc = rc->next) {
364 if(unlikely(!rrdcalc_isrunnable(rc, now, &next_run))) {
365 if(unlikely(rc->rrdcalc_flags & RRDCALC_FLAG_RUNNABLE))
366 rc->rrdcalc_flags &= ~RRDCALC_FLAG_RUNNABLE;
371 rc->old_value = rc->value;
372 rc->rrdcalc_flags |= RRDCALC_FLAG_RUNNABLE;
374 // 1. if there is database lookup, do it
375 // 2. if there is calculation expression, run it
377 if(unlikely(RRDCALC_HAS_DB_LOOKUP(rc))) {
378 /* time_t old_db_timestamp = rc->db_before; */
379 int value_is_null = 0;
381 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);
383 if(unlikely(ret != 200)) {
384 // database lookup failed
387 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);
389 if(unlikely(!(rc->rrdcalc_flags & RRDCALC_FLAG_DB_ERROR))) {
390 rc->rrdcalc_flags |= RRDCALC_FLAG_DB_ERROR;
391 error("Health on host '%s', alarm '%s.%s': database lookup returned error %d", host->hostname, rc->chart ? rc->chart : "NOCHART", rc->name, ret);
394 else if(unlikely(rc->rrdcalc_flags & RRDCALC_FLAG_DB_ERROR))
395 rc->rrdcalc_flags &= ~RRDCALC_FLAG_DB_ERROR;
397 /* - RRDCALC_FLAG_DB_STALE not currently used
398 if (unlikely(old_db_timestamp == rc->db_before)) {
401 debug(D_HEALTH, "Health on host '%s', alarm '%s.%s': database is stale", host->hostname, rc->chart?rc->chart:"NOCHART", rc->name);
403 if (unlikely(!(rc->rrdcalc_flags & RRDCALC_FLAG_DB_STALE))) {
404 rc->rrdcalc_flags |= RRDCALC_FLAG_DB_STALE;
405 error("Health on host '%s', alarm '%s.%s': database is stale", host->hostname, rc->chart?rc->chart:"NOCHART", rc->name);
408 else if (unlikely(rc->rrdcalc_flags & RRDCALC_FLAG_DB_STALE))
409 rc->rrdcalc_flags &= ~RRDCALC_FLAG_DB_STALE;
412 if(unlikely(value_is_null)) {
413 // collected value is null
417 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);
419 if(unlikely(!(rc->rrdcalc_flags & RRDCALC_FLAG_DB_NAN))) {
420 rc->rrdcalc_flags |= RRDCALC_FLAG_DB_NAN;
421 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);
424 else if(unlikely(rc->rrdcalc_flags & RRDCALC_FLAG_DB_NAN))
425 rc->rrdcalc_flags &= ~RRDCALC_FLAG_DB_NAN;
427 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);
430 if(unlikely(rc->calculation)) {
431 if(unlikely(!expression_evaluate(rc->calculation))) {
432 // calculation failed
436 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));
438 if(unlikely(!(rc->rrdcalc_flags & RRDCALC_FLAG_CALC_ERROR))) {
439 rc->rrdcalc_flags |= RRDCALC_FLAG_CALC_ERROR;
440 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));
444 if(unlikely(rc->rrdcalc_flags & RRDCALC_FLAG_CALC_ERROR))
445 rc->rrdcalc_flags &= ~RRDCALC_FLAG_CALC_ERROR;
447 debug(D_HEALTH, "Health on host '%s', alarm '%s.%s': expression '%s' gave value "
448 CALCULATED_NUMBER_FORMAT
449 ": %s (source: %s)", host->hostname, rc->chart ? rc->chart : "NOCHART", rc->name
450 , rc->calculation->parsed_as, rc->calculation->result,
451 buffer_tostring(rc->calculation->error_msg), rc->source
454 rc->value = rc->calculation->result;
458 rrdhost_unlock(host);
460 if(unlikely(runnable && !netdata_exit)) {
461 rrdhost_rdlock(host);
463 for(rc = host->alarms; rc; rc = rc->next) {
464 if(unlikely(!(rc->rrdcalc_flags & RRDCALC_FLAG_RUNNABLE)))
467 int warning_status = RRDCALC_STATUS_UNDEFINED;
468 int critical_status = RRDCALC_STATUS_UNDEFINED;
470 if(likely(rc->warning)) {
471 if(unlikely(!expression_evaluate(rc->warning))) {
472 // calculation failed
474 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));
476 if(unlikely(!(rc->rrdcalc_flags & RRDCALC_FLAG_WARN_ERROR))) {
477 rc->rrdcalc_flags |= RRDCALC_FLAG_WARN_ERROR;
478 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));
482 if(unlikely(rc->rrdcalc_flags & RRDCALC_FLAG_WARN_ERROR))
483 rc->rrdcalc_flags &= ~RRDCALC_FLAG_WARN_ERROR;
485 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);
487 warning_status = rrdcalc_value2status(rc->warning->result);
491 if(likely(rc->critical)) {
492 if(unlikely(!expression_evaluate(rc->critical))) {
493 // calculation failed
495 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));
497 if(unlikely(!(rc->rrdcalc_flags & RRDCALC_FLAG_CRIT_ERROR))) {
498 rc->rrdcalc_flags |= RRDCALC_FLAG_CRIT_ERROR;
499 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));
503 if(unlikely(rc->rrdcalc_flags & RRDCALC_FLAG_CRIT_ERROR))
504 rc->rrdcalc_flags &= ~RRDCALC_FLAG_CRIT_ERROR;
506 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);
508 critical_status = rrdcalc_value2status(rc->critical->result);
512 int status = RRDCALC_STATUS_UNDEFINED;
514 switch(warning_status) {
515 case RRDCALC_STATUS_CLEAR:
516 status = RRDCALC_STATUS_CLEAR;
519 case RRDCALC_STATUS_RAISED:
520 status = RRDCALC_STATUS_WARNING;
527 switch(critical_status) {
528 case RRDCALC_STATUS_CLEAR:
529 if(status == RRDCALC_STATUS_UNDEFINED)
530 status = RRDCALC_STATUS_CLEAR;
533 case RRDCALC_STATUS_RAISED:
534 status = RRDCALC_STATUS_CRITICAL;
541 if(status != rc->status) {
544 if(now > rc->delay_up_to_timestamp) {
545 rc->delay_up_current = rc->delay_up_duration;
546 rc->delay_down_current = rc->delay_down_duration;
548 rc->delay_up_to_timestamp = 0;
551 rc->delay_up_current = (int) (rc->delay_up_current * rc->delay_multiplier);
552 if(rc->delay_up_current > rc->delay_max_duration)
553 rc->delay_up_current = rc->delay_max_duration;
555 rc->delay_down_current = (int) (rc->delay_down_current * rc->delay_multiplier);
556 if(rc->delay_down_current > rc->delay_max_duration)
557 rc->delay_down_current = rc->delay_max_duration;
560 if(status > rc->status)
561 delay = rc->delay_up_current;
563 delay = rc->delay_down_current;
565 // COMMENTED: because we do need to send raising alarms
566 // if(now + delay < rc->delay_up_to_timestamp)
567 // delay = (int)(rc->delay_up_to_timestamp - now);
569 rc->delay_last = delay;
570 rc->delay_up_to_timestamp = now + delay;
572 host, rc->id, rc->next_event_id++, now, rc->name, rc->rrdset->id
573 , rc->rrdset->family, rc->exec, rc->recipient, now - rc->last_status_change
574 , rc->old_value, rc->value, rc->status, status, rc->source, rc->units, rc->info
575 , rc->delay_last, (rc->options & RRDCALC_FLAG_NO_CLEAR_NOTIFICATION)
576 ? HEALTH_ENTRY_FLAG_NO_CLEAR_NOTIFICATION : 0
578 rc->last_status_change = now;
582 rc->last_updated = now;
583 rc->next_update = now + rc->update_every;
585 if(next_run > rc->next_update)
586 next_run = rc->next_update;
589 rrdhost_unlock(host);
592 if(unlikely(netdata_exit))
595 // execute notifications
597 health_alarm_log_process(host);
599 if(unlikely(netdata_exit))
606 if(unlikely(pthread_setcancelstate(oldstate, NULL) != 0))
607 error("Cannot set pthread cancel state to RESTORE (%d).", oldstate);
609 if(unlikely(netdata_exit))
612 now = now_realtime_sec();
614 debug(D_HEALTH, "Health monitoring iteration no %u done. Next iteration in %d secs", loop, (int) (next_run - now));
615 sleep_usec(USEC_PER_SEC * (usec_t) (next_run - now));
618 debug(D_HEALTH, "Health monitoring iteration no %u done. Next iteration now", loop);
623 info("HEALTH thread exiting");
625 static_thread->enabled = 0;