]> git.proxmox.com Git - pve-ha-manager.git/blob - src/PVE/HA/LRM.pm
1894f3c7b4e89da110881d701b5597596df147e6
[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 mode => $self->{mode},
119 results => $self->{results},
120 timestamp => $haenv->get_time(),
121 };
122
123 eval { $haenv->write_lrm_status($lrm_status); };
124 if (my $err = $@) {
125 $haenv->log('err', "unable to write lrm status file - $err");
126 return 0;
127 }
128
129 return 1;
130 }
131
132 sub get_protected_ha_agent_lock {
133 my ($self) = @_;
134
135 my $haenv = $self->{haenv};
136
137 my $count = 0;
138 my $starttime = $haenv->get_time();
139
140 for (;;) {
141
142 if ($haenv->get_ha_agent_lock()) {
143 if ($self->{ha_agent_wd}) {
144 $haenv->watchdog_update($self->{ha_agent_wd});
145 } else {
146 my $wfh = $haenv->watchdog_open();
147 $self->{ha_agent_wd} = $wfh;
148 }
149 return 1;
150 }
151
152 last if ++$count > 5; # try max 5 time
153
154 my $delay = $haenv->get_time() - $starttime;
155 last if $delay > 5; # for max 5 seconds
156
157 $haenv->sleep(1);
158 }
159
160 return 0;
161 }
162
163 sub active_service_count {
164 my ($self) = @_;
165
166 my $haenv = $self->{haenv};
167
168 my $nodename = $haenv->nodename();
169
170 my $ss = $self->{service_status};
171
172 my $count = 0;
173
174 foreach my $sid (keys %$ss) {
175 my $sd = $ss->{$sid};
176 next if !$sd->{node};
177 next if $sd->{node} ne $nodename;
178 my $req_state = $sd->{state};
179 next if !defined($req_state);
180 next if $req_state eq 'stopped';
181 next if $req_state eq 'freeze';
182
183 $count++;
184 }
185
186 return $count;
187 }
188
189 my $wrote_lrm_status_at_startup = 0;
190
191 sub do_one_iteration {
192 my ($self) = @_;
193
194 my $haenv = $self->{haenv};
195
196 if (!$wrote_lrm_status_at_startup) {
197 if ($self->update_lrm_status()) {
198 $wrote_lrm_status_at_startup = 1;
199 } else {
200 # do nothing
201 $haenv->sleep(5);
202 return $self->{shutdown_request} ? 0 : 1;
203 }
204 }
205
206 my $status = $self->get_local_status();
207 my $state = $status->{state};
208
209 my $ms = $haenv->read_manager_status();
210 $self->{service_status} = $ms->{service_status} || {};
211
212 my $fence_request = PVE::HA::Tools::count_fenced_services($self->{service_status}, $haenv->nodename());
213
214 # do state changes first
215
216 my $ctime = $haenv->get_time();
217
218 if ($state eq 'wait_for_agent_lock') {
219
220 my $service_count = $self->active_service_count();
221
222 if (!$fence_request && $service_count && $haenv->quorate()) {
223 if ($self->get_protected_ha_agent_lock()) {
224 $self->set_local_status({ state => 'active' });
225 }
226 }
227
228 } elsif ($state eq 'lost_agent_lock') {
229
230 if (!$fence_request && $haenv->quorate()) {
231 if ($self->get_protected_ha_agent_lock()) {
232 $self->set_local_status({ state => 'active' });
233 }
234 }
235
236 } elsif ($state eq 'active') {
237
238 if ($fence_request) {
239 $haenv->log('err', "node need to be fenced - releasing agent_lock\n");
240 $self->set_local_status({ state => 'lost_agent_lock'});
241 } elsif (!$self->get_protected_ha_agent_lock()) {
242 $self->set_local_status({ state => 'lost_agent_lock'});
243 }
244 }
245
246 $status = $self->get_local_status();
247 $state = $status->{state};
248
249 # do work
250
251 if ($state eq 'wait_for_agent_lock') {
252
253 return 0 if $self->{shutdown_request};
254
255 $self->update_lrm_status();
256
257 $haenv->sleep(5);
258
259 } elsif ($state eq 'active') {
260
261 my $startime = $haenv->get_time();
262
263 my $max_time = 10;
264
265 my $shutdown = 0;
266
267 # do work (max_time seconds)
268 eval {
269 # fixme: set alert timer
270
271 if ($self->{shutdown_request}) {
272
273 if ($self->{mode} eq 'restart') {
274
275 my $service_count = $self->active_service_count();
276
277 if ($service_count == 0) {
278
279 if ($self->run_workers() == 0) {
280 if ($self->{ha_agent_wd}) {
281 $haenv->watchdog_close($self->{ha_agent_wd});
282 delete $self->{ha_agent_wd};
283 }
284
285 $shutdown = 1;
286
287 # restart with no or freezed services, release the lock
288 $haenv->release_ha_agent_lock();
289 }
290 }
291 } else {
292
293 if ($self->run_workers() == 0) {
294 if ($self->{shutdown_errors} == 0) {
295 if ($self->{ha_agent_wd}) {
296 $haenv->watchdog_close($self->{ha_agent_wd});
297 delete $self->{ha_agent_wd};
298 }
299
300 # shutdown with all services stopped thus release the lock
301 $haenv->release_ha_agent_lock();
302 }
303
304 $shutdown = 1;
305 }
306 }
307 } else {
308
309 $self->manage_resources();
310
311 }
312 };
313 if (my $err = $@) {
314 $haenv->log('err', "got unexpected error - $err");
315 }
316
317 $self->update_lrm_status();
318
319 return 0 if $shutdown;
320
321 $haenv->sleep_until($startime + $max_time);
322
323 } elsif ($state eq 'lost_agent_lock') {
324
325 # Note: watchdog is active an will triger soon!
326
327 # so we hope to get the lock back soon!
328
329 if ($self->{shutdown_request}) {
330
331 my $service_count = $self->active_service_count();
332
333 if ($service_count > 0) {
334 $haenv->log('err', "get shutdown request in state 'lost_agent_lock' - " .
335 "detected $service_count running services");
336
337 } else {
338
339 # all services are stopped, so we can close the watchdog
340
341 if ($self->{ha_agent_wd}) {
342 $haenv->watchdog_close($self->{ha_agent_wd});
343 delete $self->{ha_agent_wd};
344 }
345
346 return 0;
347 }
348 }
349
350 $haenv->sleep(5);
351
352 } else {
353
354 die "got unexpected status '$state'\n";
355
356 }
357
358 return 1;
359 }
360
361 sub run_workers {
362 my ($self) = @_;
363
364 my $haenv = $self->{haenv};
365
366 my $starttime = $haenv->get_time();
367
368 # start workers
369 my $max_workers = 4;
370
371 my $sc = $haenv->read_service_config();
372
373 while (($haenv->get_time() - $starttime) < 5) {
374 my $count = $self->check_active_workers();
375
376 foreach my $sid (keys %{$self->{workers}}) {
377 last if $count >= $max_workers;
378 my $w = $self->{workers}->{$sid};
379 my $cd = $sc->{$sid};
380 if (!$cd) {
381 $haenv->log('err', "missing resource configuration for '$sid'");
382 next;
383 }
384 if (!$w->{pid}) {
385 if ($haenv->can_fork()) {
386 my $pid = fork();
387 if (!defined($pid)) {
388 $haenv->log('err', "fork worker failed");
389 $count = 0; last; # abort, try later
390 } elsif ($pid == 0) {
391 $haenv->after_fork(); # cleanup
392
393 # do work
394 my $res = -1;
395 eval {
396 $res = $self->exec_resource_agent($sid, $cd, $w->{state}, $w->{target});
397 };
398 if (my $err = $@) {
399 $haenv->log('err', $err);
400 POSIX::_exit(-1);
401 }
402 POSIX::_exit($res);
403 } else {
404 $count++;
405 $w->{pid} = $pid;
406 }
407 } else {
408 my $res = -1;
409 eval {
410 $res = $self->exec_resource_agent($sid, $cd, $w->{state}, $w->{target});
411 $res = $res << 8 if $res > 0;
412 };
413 if (my $err = $@) {
414 $haenv->log('err', $err);
415 }
416 if (defined($w->{uid})) {
417 $self->resource_command_finished($sid, $w->{uid}, $res);
418 } else {
419 $self->stop_command_finished($sid, $res);
420 }
421 }
422 }
423 }
424
425 last if !$count;
426
427 $haenv->sleep(1);
428 }
429
430 return scalar(keys %{$self->{workers}});
431 }
432
433 sub manage_resources {
434 my ($self) = @_;
435
436 my $haenv = $self->{haenv};
437
438 my $nodename = $haenv->nodename();
439
440 my $ss = $self->{service_status};
441
442 foreach my $sid (keys %$ss) {
443 my $sd = $ss->{$sid};
444 next if !$sd->{node};
445 next if !$sd->{uid};
446 next if $sd->{node} ne $nodename;
447 my $req_state = $sd->{state};
448 next if !defined($req_state);
449 next if $req_state eq 'freeze';
450 $self->queue_resource_command($sid, $sd->{uid}, $req_state, $sd->{target});
451 }
452
453 return $self->run_workers();
454 }
455
456 sub queue_resource_command {
457 my ($self, $sid, $uid, $state, $target) = @_;
458
459 if (my $w = $self->{workers}->{$sid}) {
460 return if $w->{pid}; # already started
461 # else, delete and overwrite queue entry with new command
462 delete $self->{workers}->{$sid};
463 }
464
465 $self->{workers}->{$sid} = {
466 sid => $sid,
467 uid => $uid,
468 state => $state,
469 };
470
471 $self->{workers}->{$sid}->{target} = $target if $target;
472 }
473
474 sub check_active_workers {
475 my ($self) = @_;
476
477 # finish/count workers
478 my $count = 0;
479 foreach my $sid (keys %{$self->{workers}}) {
480 my $w = $self->{workers}->{$sid};
481 if (my $pid = $w->{pid}) {
482 # check status
483 my $waitpid = waitpid($pid, WNOHANG);
484 if (defined($waitpid) && ($waitpid == $pid)) {
485 if (defined($w->{uid})) {
486 $self->resource_command_finished($sid, $w->{uid}, $?);
487 } else {
488 $self->stop_command_finished($sid, $?);
489 }
490 } else {
491 $count++;
492 }
493 }
494 }
495
496 return $count;
497 }
498
499 sub stop_command_finished {
500 my ($self, $sid, $status) = @_;
501
502 my $haenv = $self->{haenv};
503
504 my $w = delete $self->{workers}->{$sid};
505 return if !$w; # should not happen
506
507 my $exit_code = -1;
508
509 if ($status == -1) {
510 $haenv->log('err', "resource agent $sid finished - failed to execute");
511 } elsif (my $sig = ($status & 127)) {
512 $haenv->log('err', "resource agent $sid finished - got signal $sig");
513 } else {
514 $exit_code = ($status >> 8);
515 }
516
517 if ($exit_code != 0) {
518 $self->{shutdown_errors}++;
519 }
520 }
521
522 sub resource_command_finished {
523 my ($self, $sid, $uid, $status) = @_;
524
525 my $haenv = $self->{haenv};
526
527 my $w = delete $self->{workers}->{$sid};
528 return if !$w; # should not happen
529
530 my $exit_code = -1;
531
532 if ($status == -1) {
533 $haenv->log('err', "resource agent $sid finished - failed to execute");
534 } elsif (my $sig = ($status & 127)) {
535 $haenv->log('err', "resource agent $sid finished - got signal $sig");
536 } else {
537 $exit_code = ($status >> 8);
538 }
539
540 $exit_code = $self->handle_service_exitcode($sid, $w->{state}, $exit_code);
541
542 $self->{results}->{$uid} = {
543 sid => $w->{sid},
544 state => $w->{state},
545 exit_code => $exit_code,
546 };
547
548 my $ss = $self->{service_status};
549
550 # compute hash of valid/existing uids
551 my $valid_uids = {};
552 foreach my $sid (keys %$ss) {
553 my $sd = $ss->{$sid};
554 next if !$sd->{uid};
555 $valid_uids->{$sd->{uid}} = 1;
556 }
557
558 my $results = {};
559 foreach my $id (keys %{$self->{results}}) {
560 next if !$valid_uids->{$id};
561 $results->{$id} = $self->{results}->{$id};
562 }
563 $self->{results} = $results;
564 }
565
566 # processes the exit code from a finished resource agent, so that the CRM knows
567 # if the LRM wants to retry an action based on the current recovery policies for
568 # the failed service, or the CRM itself must try to recover from the failure.
569 sub handle_service_exitcode {
570 my ($self, $sid, $cmd, $exit_code) = @_;
571
572 my $haenv = $self->{haenv};
573 my $tries = $self->{restart_tries};
574
575 my $sc = $haenv->read_service_config();
576 my $cd = $sc->{$sid};
577
578 if ($cmd eq 'started') {
579
580 if ($exit_code == SUCCESS) {
581
582 $tries->{$sid} = 0;
583
584 return $exit_code;
585
586 } elsif ($exit_code == ERROR) {
587
588 $tries->{$sid} = 0 if !defined($tries->{$sid});
589
590 $tries->{$sid}++;
591 if ($tries->{$sid} >= $cd->{max_restart}) {
592 $haenv->log('err', "unable to start service $sid on local node".
593 " after $tries->{$sid} retries");
594 $tries->{$sid} = 0;
595 return ERROR;
596 }
597
598 # tell CRM that we retry the start
599 return ETRY_AGAIN;
600 }
601 }
602
603 return $exit_code;
604
605 }
606
607 sub exec_resource_agent {
608 my ($self, $sid, $service_config, $cmd, @params) = @_;
609
610 # setup execution environment
611
612 $ENV{'PATH'} = '/sbin:/bin:/usr/sbin:/usr/bin';
613
614 my $haenv = $self->{haenv};
615
616 my $nodename = $haenv->nodename();
617
618 my (undef, $service_type, $service_name) = PVE::HA::Tools::parse_sid($sid);
619
620 my $plugin = PVE::HA::Resources->lookup($service_type);
621 if (!$plugin) {
622 $haenv->log('err', "service type '$service_type' not implemented");
623 return EUNKNOWN_SERVICE_TYPE;
624 }
625
626 if ($service_config->{node} ne $nodename) {
627 $haenv->log('err', "service '$sid' not on this node");
628 return EWRONG_NODE;
629 }
630
631 my $id = $service_name;
632
633 my $running = $plugin->check_running($haenv, $id);
634
635 if ($cmd eq 'started') {
636
637 return SUCCESS if $running;
638
639 $haenv->log("info", "starting service $sid");
640
641 $plugin->start($haenv, $id);
642
643 $running = $plugin->check_running($haenv, $id);
644
645 if ($running) {
646 $haenv->log("info", "service status $sid started");
647 return SUCCESS;
648 } else {
649 $haenv->log("warning", "unable to start service $sid");
650 return ERROR;
651 }
652
653 } elsif ($cmd eq 'request_stop' || $cmd eq 'stopped') {
654
655 return SUCCESS if !$running;
656
657 $haenv->log("info", "stopping service $sid");
658
659 $plugin->shutdown($haenv, $id);
660
661 $running = $plugin->check_running($haenv, $id);
662
663 if (!$running) {
664 $haenv->log("info", "service status $sid stopped");
665 return SUCCESS;
666 } else {
667 $haenv->log("info", "unable to stop stop service $sid (still running)");
668 return ERROR;
669 }
670
671 } elsif ($cmd eq 'migrate' || $cmd eq 'relocate') {
672
673 my $target = $params[0];
674 if (!defined($target)) {
675 die "$cmd '$sid' failed - missing target\n" if !defined($target);
676 return EINVALID_PARAMETER;
677 }
678
679 if ($service_config->{node} eq $target) {
680 # already there
681 return SUCCESS;
682 }
683
684 my $online = ($cmd eq 'migrate') ? 1 : 0;
685
686 my $res = $plugin->migrate($haenv, $id, $target, $online);
687
688 # something went wrong if service is still on this node
689 if (!$res) {
690 $haenv->log("err", "service $sid not moved (migration error)");
691 return ERROR;
692 }
693
694 return SUCCESS;
695
696 } elsif ($cmd eq 'error') {
697
698 if ($running) {
699 $haenv->log("err", "service $sid is in an error state while running");
700 } else {
701 $haenv->log("warning", "service $sid is not running and in an error state");
702 }
703 return SUCCESS; # error always succeeds
704
705 }
706
707 $haenv->log("err", "implement me (cmd '$cmd')");
708 return EUNKNOWN_COMMAND;
709 }
710
711
712 1;