]> git.proxmox.com Git - pve-ha-manager.git/blob - src/PVE/HA/LRM.pm
improve verbosity of API status call
[pve-ha-manager.git] / src / PVE / HA / LRM.pm
1 package PVE::HA::LRM;
2
3 # Local Resource Manager
4
5 use strict;
6 use warnings;
7 use Data::Dumper;
8 use POSIX qw(:sys_wait_h);
9
10 use PVE::SafeSyslog;
11 use PVE::Tools;
12 use PVE::HA::Tools ':exit_codes';
13 use PVE::HA::Resources;
14
15 # Server can have several states:
16
17 my $valid_states = {
18 wait_for_agent_lock => "waiting for agent lock",
19 active => "got agent_lock",
20 lost_agent_lock => "lost agent_lock",
21 };
22
23 sub new {
24 my ($this, $haenv) = @_;
25
26 my $class = ref($this) || $this;
27
28 my $self = bless {
29 haenv => $haenv,
30 status => { state => 'startup' },
31 workers => {},
32 results => {},
33 restart_tries => {},
34 shutdown_request => 0,
35 shutdown_errors => 0,
36 # mode can be: active, reboot, shutdown, restart
37 mode => 'active',
38 }, $class;
39
40 $self->set_local_status({ state => 'wait_for_agent_lock' });
41
42 return $self;
43 }
44
45 sub shutdown_request {
46 my ($self) = @_;
47
48 return if $self->{shutdown_request}; # already in shutdown mode
49
50 my $haenv = $self->{haenv};
51
52 my $nodename = $haenv->nodename();
53
54 my $shutdown = $haenv->is_node_shutdown();
55
56 if ($shutdown) {
57 $haenv->log('info', "shutdown LRM, stop all services");
58 $self->{mode} = 'shutdown';
59
60 # queue stop jobs for all services
61
62 my $ss = $self->{service_status};
63
64 foreach my $sid (keys %$ss) {
65 my $sd = $ss->{$sid};
66 next if !$sd->{node};
67 next if $sd->{node} ne $nodename;
68 # Note: use undef uid to mark shutdown/stop jobs
69 $self->queue_resource_command($sid, undef, 'request_stop');
70 }
71
72 } else {
73 $haenv->log('info', "restart LRM, freeze all services");
74 $self->{mode} = 'restart';
75 }
76
77 $self->{shutdown_request} = 1;
78
79 eval { $self->update_lrm_status(); };
80 if (my $err = $@) {
81 $self->log('err', "unable to update lrm status file - $err");
82 }
83 }
84
85 sub get_local_status {
86 my ($self) = @_;
87
88 return $self->{status};
89 }
90
91 sub set_local_status {
92 my ($self, $new) = @_;
93
94 die "invalid state '$new->{state}'" if !$valid_states->{$new->{state}};
95
96 my $haenv = $self->{haenv};
97
98 my $old = $self->{status};
99
100 # important: only update if if really changed
101 return if $old->{state} eq $new->{state};
102
103 $haenv->log('info', "status change $old->{state} => $new->{state}");
104
105 $new->{state_change_time} = $haenv->get_time();
106
107 $self->{status} = $new;
108 }
109
110 sub update_lrm_status {
111 my ($self) = @_;
112
113 my $haenv = $self->{haenv};
114
115 return 0 if !$haenv->quorate();
116
117 my $lrm_status = {
118 state => $self->{status}->{state},
119 mode => $self->{mode},
120 results => $self->{results},
121 timestamp => $haenv->get_time(),
122 };
123
124 eval { $haenv->write_lrm_status($lrm_status); };
125 if (my $err = $@) {
126 $haenv->log('err', "unable to write lrm status file - $err");
127 return 0;
128 }
129
130 return 1;
131 }
132
133 sub get_protected_ha_agent_lock {
134 my ($self) = @_;
135
136 my $haenv = $self->{haenv};
137
138 my $count = 0;
139 my $starttime = $haenv->get_time();
140
141 for (;;) {
142
143 if ($haenv->get_ha_agent_lock()) {
144 if ($self->{ha_agent_wd}) {
145 $haenv->watchdog_update($self->{ha_agent_wd});
146 } else {
147 my $wfh = $haenv->watchdog_open();
148 $self->{ha_agent_wd} = $wfh;
149 }
150 return 1;
151 }
152
153 last if ++$count > 5; # try max 5 time
154
155 my $delay = $haenv->get_time() - $starttime;
156 last if $delay > 5; # for max 5 seconds
157
158 $haenv->sleep(1);
159 }
160
161 return 0;
162 }
163
164 sub active_service_count {
165 my ($self) = @_;
166
167 my $haenv = $self->{haenv};
168
169 my $nodename = $haenv->nodename();
170
171 my $ss = $self->{service_status};
172
173 my $count = 0;
174
175 foreach my $sid (keys %$ss) {
176 my $sd = $ss->{$sid};
177 next if !$sd->{node};
178 next if $sd->{node} ne $nodename;
179 my $req_state = $sd->{state};
180 next if !defined($req_state);
181 next if $req_state eq 'stopped';
182 next if $req_state eq 'freeze';
183
184 $count++;
185 }
186
187 return $count;
188 }
189
190 my $wrote_lrm_status_at_startup = 0;
191
192 sub do_one_iteration {
193 my ($self) = @_;
194
195 my $haenv = $self->{haenv};
196
197 if (!$wrote_lrm_status_at_startup) {
198 if ($self->update_lrm_status()) {
199 $wrote_lrm_status_at_startup = 1;
200 } else {
201 # do nothing
202 $haenv->sleep(5);
203 return $self->{shutdown_request} ? 0 : 1;
204 }
205 }
206
207 my $status = $self->get_local_status();
208 my $state = $status->{state};
209
210 my $ms = $haenv->read_manager_status();
211 $self->{service_status} = $ms->{service_status} || {};
212
213 my $fence_request = PVE::HA::Tools::count_fenced_services($self->{service_status}, $haenv->nodename());
214
215 # do state changes first
216
217 my $ctime = $haenv->get_time();
218
219 if ($state eq 'wait_for_agent_lock') {
220
221 my $service_count = $self->active_service_count();
222
223 if (!$fence_request && $service_count && $haenv->quorate()) {
224 if ($self->get_protected_ha_agent_lock()) {
225 $self->set_local_status({ state => 'active' });
226 }
227 }
228
229 } elsif ($state eq 'lost_agent_lock') {
230
231 if (!$fence_request && $haenv->quorate()) {
232 if ($self->get_protected_ha_agent_lock()) {
233 $self->set_local_status({ state => 'active' });
234 }
235 }
236
237 } elsif ($state eq 'active') {
238
239 if ($fence_request) {
240 $haenv->log('err', "node need to be fenced - releasing agent_lock\n");
241 $self->set_local_status({ state => 'lost_agent_lock'});
242 } elsif (!$self->get_protected_ha_agent_lock()) {
243 $self->set_local_status({ state => 'lost_agent_lock'});
244 }
245 }
246
247 $status = $self->get_local_status();
248 $state = $status->{state};
249
250 # do work
251
252 if ($state eq 'wait_for_agent_lock') {
253
254 return 0 if $self->{shutdown_request};
255
256 $self->update_lrm_status();
257
258 $haenv->sleep(5);
259
260 } elsif ($state eq 'active') {
261
262 my $startime = $haenv->get_time();
263
264 my $max_time = 10;
265
266 my $shutdown = 0;
267
268 # do work (max_time seconds)
269 eval {
270 # fixme: set alert timer
271
272 if ($self->{shutdown_request}) {
273
274 if ($self->{mode} eq 'restart') {
275
276 my $service_count = $self->active_service_count();
277
278 if ($service_count == 0) {
279
280 if ($self->run_workers() == 0) {
281 if ($self->{ha_agent_wd}) {
282 $haenv->watchdog_close($self->{ha_agent_wd});
283 delete $self->{ha_agent_wd};
284 }
285
286 $shutdown = 1;
287
288 # restart with no or freezed services, release the lock
289 $haenv->release_ha_agent_lock();
290 }
291 }
292 } else {
293
294 if ($self->run_workers() == 0) {
295 if ($self->{shutdown_errors} == 0) {
296 if ($self->{ha_agent_wd}) {
297 $haenv->watchdog_close($self->{ha_agent_wd});
298 delete $self->{ha_agent_wd};
299 }
300
301 # shutdown with all services stopped thus release the lock
302 $haenv->release_ha_agent_lock();
303 }
304
305 $shutdown = 1;
306 }
307 }
308 } else {
309
310 $self->manage_resources();
311
312 }
313 };
314 if (my $err = $@) {
315 $haenv->log('err', "got unexpected error - $err");
316 }
317
318 $self->update_lrm_status();
319
320 return 0 if $shutdown;
321
322 $haenv->sleep_until($startime + $max_time);
323
324 } elsif ($state eq 'lost_agent_lock') {
325
326 # Note: watchdog is active an will triger soon!
327
328 # so we hope to get the lock back soon!
329
330 if ($self->{shutdown_request}) {
331
332 my $service_count = $self->active_service_count();
333
334 if ($service_count > 0) {
335 $haenv->log('err', "get shutdown request in state 'lost_agent_lock' - " .
336 "detected $service_count running services");
337
338 } else {
339
340 # all services are stopped, so we can close the watchdog
341
342 if ($self->{ha_agent_wd}) {
343 $haenv->watchdog_close($self->{ha_agent_wd});
344 delete $self->{ha_agent_wd};
345 }
346
347 return 0;
348 }
349 }
350
351 $haenv->sleep(5);
352
353 } else {
354
355 die "got unexpected status '$state'\n";
356
357 }
358
359 return 1;
360 }
361
362 sub run_workers {
363 my ($self) = @_;
364
365 my $haenv = $self->{haenv};
366
367 my $starttime = $haenv->get_time();
368
369 # start workers
370 my $max_workers = 4;
371
372 my $sc = $haenv->read_service_config();
373
374 while (($haenv->get_time() - $starttime) < 5) {
375 my $count = $self->check_active_workers();
376
377 foreach my $sid (keys %{$self->{workers}}) {
378 last if $count >= $max_workers;
379 my $w = $self->{workers}->{$sid};
380 if (!$w->{pid}) {
381 if ($haenv->can_fork()) {
382 my $pid = fork();
383 if (!defined($pid)) {
384 $haenv->log('err', "fork worker failed");
385 $count = 0; last; # abort, try later
386 } elsif ($pid == 0) {
387 $haenv->after_fork(); # cleanup
388
389 # do work
390 my $res = -1;
391 eval {
392 $res = $self->exec_resource_agent($sid, $sc->{$sid}, $w->{state}, $w->{target});
393 };
394 if (my $err = $@) {
395 $haenv->log('err', $err);
396 POSIX::_exit(-1);
397 }
398 POSIX::_exit($res);
399 } else {
400 $count++;
401 $w->{pid} = $pid;
402 }
403 } else {
404 my $res = -1;
405 eval {
406 $res = $self->exec_resource_agent($sid, $sc->{$sid}, $w->{state}, $w->{target});
407 $res = $res << 8 if $res > 0;
408 };
409 if (my $err = $@) {
410 $haenv->log('err', $err);
411 }
412 if (defined($w->{uid})) {
413 $self->resource_command_finished($sid, $w->{uid}, $res);
414 } else {
415 $self->stop_command_finished($sid, $res);
416 }
417 }
418 }
419 }
420
421 last if !$count;
422
423 $haenv->sleep(1);
424 }
425
426 return scalar(keys %{$self->{workers}});
427 }
428
429 sub manage_resources {
430 my ($self) = @_;
431
432 my $haenv = $self->{haenv};
433
434 my $nodename = $haenv->nodename();
435
436 my $ss = $self->{service_status};
437
438 foreach my $sid (keys %$ss) {
439 my $sd = $ss->{$sid};
440 next if !$sd->{node};
441 next if !$sd->{uid};
442 next if $sd->{node} ne $nodename;
443 my $req_state = $sd->{state};
444 next if !defined($req_state);
445 next if $req_state eq 'freeze';
446 $self->queue_resource_command($sid, $sd->{uid}, $req_state, $sd->{target});
447 }
448
449 return $self->run_workers();
450 }
451
452 sub queue_resource_command {
453 my ($self, $sid, $uid, $state, $target) = @_;
454
455 if (my $w = $self->{workers}->{$sid}) {
456 return if $w->{pid}; # already started
457 # else, delete and overwrite queue entry with new command
458 delete $self->{workers}->{$sid};
459 }
460
461 $self->{workers}->{$sid} = {
462 sid => $sid,
463 uid => $uid,
464 state => $state,
465 };
466
467 $self->{workers}->{$sid}->{target} = $target if $target;
468 }
469
470 sub check_active_workers {
471 my ($self) = @_;
472
473 # finish/count workers
474 my $count = 0;
475 foreach my $sid (keys %{$self->{workers}}) {
476 my $w = $self->{workers}->{$sid};
477 if (my $pid = $w->{pid}) {
478 # check status
479 my $waitpid = waitpid($pid, WNOHANG);
480 if (defined($waitpid) && ($waitpid == $pid)) {
481 if (defined($w->{uid})) {
482 $self->resource_command_finished($sid, $w->{uid}, $?);
483 } else {
484 $self->stop_command_finished($sid, $?);
485 }
486 } else {
487 $count++;
488 }
489 }
490 }
491
492 return $count;
493 }
494
495 sub stop_command_finished {
496 my ($self, $sid, $status) = @_;
497
498 my $haenv = $self->{haenv};
499
500 my $w = delete $self->{workers}->{$sid};
501 return if !$w; # should not happen
502
503 my $exit_code = -1;
504
505 if ($status == -1) {
506 $haenv->log('err', "resource agent $sid finished - failed to execute");
507 } elsif (my $sig = ($status & 127)) {
508 $haenv->log('err', "resource agent $sid finished - got signal $sig");
509 } else {
510 $exit_code = ($status >> 8);
511 }
512
513 if ($exit_code != 0) {
514 $self->{shutdown_errors}++;
515 }
516 }
517
518 sub resource_command_finished {
519 my ($self, $sid, $uid, $status) = @_;
520
521 my $haenv = $self->{haenv};
522
523 my $w = delete $self->{workers}->{$sid};
524 return if !$w; # should not happen
525
526 my $exit_code = -1;
527
528 if ($status == -1) {
529 $haenv->log('err', "resource agent $sid finished - failed to execute");
530 } elsif (my $sig = ($status & 127)) {
531 $haenv->log('err', "resource agent $sid finished - got signal $sig");
532 } else {
533 $exit_code = ($status >> 8);
534 }
535
536 $exit_code = $self->handle_service_exitcode($sid, $w->{state}, $exit_code);
537
538 $self->{results}->{$uid} = {
539 sid => $w->{sid},
540 state => $w->{state},
541 exit_code => $exit_code,
542 };
543
544 my $ss = $self->{service_status};
545
546 # compute hash of valid/existing uids
547 my $valid_uids = {};
548 foreach my $sid (keys %$ss) {
549 my $sd = $ss->{$sid};
550 next if !$sd->{uid};
551 $valid_uids->{$sd->{uid}} = 1;
552 }
553
554 my $results = {};
555 foreach my $id (keys %{$self->{results}}) {
556 next if !$valid_uids->{$id};
557 $results->{$id} = $self->{results}->{$id};
558 }
559 $self->{results} = $results;
560 }
561
562 # processes the exit code from a finished resource agent, so that the CRM knows
563 # if the LRM wants to retry an action based on the current recovery policies for
564 # the failed service, or the CRM itself must try to recover from the failure.
565 sub handle_service_exitcode {
566 my ($self, $sid, $cmd, $exit_code) = @_;
567
568 my $haenv = $self->{haenv};
569 my $tries = $self->{restart_tries};
570
571 my $sc = $haenv->read_service_config();
572
573 my $max_restart = 0;
574
575 if (my $cd = $sc->{$sid}) {
576 $max_restart = $cd->{max_restart};
577 }
578
579 if ($cmd eq 'started') {
580
581 if ($exit_code == SUCCESS) {
582
583 $tries->{$sid} = 0;
584
585 return $exit_code;
586
587 } elsif ($exit_code == ERROR) {
588
589 $tries->{$sid} = 0 if !defined($tries->{$sid});
590
591 $tries->{$sid}++;
592 if ($tries->{$sid} >= $max_restart) {
593 $haenv->log('err', "unable to start service $sid on local node".
594 " after $tries->{$sid} retries");
595 $tries->{$sid} = 0;
596 return ERROR;
597 }
598
599 # tell CRM that we retry the start
600 return ETRY_AGAIN;
601 }
602 }
603
604 return $exit_code;
605
606 }
607
608 sub exec_resource_agent {
609 my ($self, $sid, $service_config, $cmd, @params) = @_;
610
611 # setup execution environment
612
613 $ENV{'PATH'} = '/sbin:/bin:/usr/sbin:/usr/bin';
614
615 my $haenv = $self->{haenv};
616
617 my $nodename = $haenv->nodename();
618
619 my (undef, $service_type, $service_name) = PVE::HA::Tools::parse_sid($sid);
620
621 my $plugin = PVE::HA::Resources->lookup($service_type);
622 if (!$plugin) {
623 $haenv->log('err', "service type '$service_type' not implemented");
624 return EUNKNOWN_SERVICE_TYPE;
625 }
626
627 if (!$service_config) {
628 $haenv->log('err', "missing resource configuration for '$sid'");
629 return EUNKNOWN_SERVICE;
630 }
631
632 if ($service_config->{node} ne $nodename) {
633 $haenv->log('err', "service '$sid' not on this node");
634 return EWRONG_NODE;
635 }
636
637 my $id = $service_name;
638
639 my $running = $plugin->check_running($haenv, $id);
640
641 if ($cmd eq 'started') {
642
643 return SUCCESS if $running;
644
645 $haenv->log("info", "starting service $sid");
646
647 $plugin->start($haenv, $id);
648
649 $running = $plugin->check_running($haenv, $id);
650
651 if ($running) {
652 $haenv->log("info", "service status $sid started");
653 return SUCCESS;
654 } else {
655 $haenv->log("warning", "unable to start service $sid");
656 return ERROR;
657 }
658
659 } elsif ($cmd eq 'request_stop' || $cmd eq 'stopped') {
660
661 return SUCCESS if !$running;
662
663 $haenv->log("info", "stopping service $sid");
664
665 $plugin->shutdown($haenv, $id);
666
667 $running = $plugin->check_running($haenv, $id);
668
669 if (!$running) {
670 $haenv->log("info", "service status $sid stopped");
671 return SUCCESS;
672 } else {
673 $haenv->log("info", "unable to stop stop service $sid (still running)");
674 return ERROR;
675 }
676
677 } elsif ($cmd eq 'migrate' || $cmd eq 'relocate') {
678
679 my $target = $params[0];
680 if (!defined($target)) {
681 die "$cmd '$sid' failed - missing target\n" if !defined($target);
682 return EINVALID_PARAMETER;
683 }
684
685 if ($service_config->{node} eq $target) {
686 # already there
687 return SUCCESS;
688 }
689
690 my $online = ($cmd eq 'migrate') ? 1 : 0;
691
692 my $res = $plugin->migrate($haenv, $id, $target, $online);
693
694 # something went wrong if service is still on this node
695 if (!$res) {
696 $haenv->log("err", "service $sid not moved (migration error)");
697 return ERROR;
698 }
699
700 return SUCCESS;
701
702 } elsif ($cmd eq 'error') {
703
704 if ($running) {
705 $haenv->log("err", "service $sid is in an error state while running");
706 } else {
707 $haenv->log("warning", "service $sid is not running and in an error state");
708 }
709 return SUCCESS; # error always succeeds
710
711 }
712
713 $haenv->log("err", "implement me (cmd '$cmd')");
714 return EUNKNOWN_COMMAND;
715 }
716
717
718 1;