]> git.proxmox.com Git - pve-ha-manager.git/blob - src/PVE/HA/Manager.pm
3aaecd6c89b395b27d2cfb40324b6d93e122ca64
[pve-ha-manager.git] / src / PVE / HA / Manager.pm
1 package PVE::HA::Manager;
2
3 use strict;
4 use warnings;
5 use Digest::MD5 qw(md5_base64);
6
7 use Data::Dumper;
8 use PVE::Tools;
9 use PVE::HA::NodeStatus;
10
11 my $fence_delay = 60;
12
13 sub new {
14 my ($this, $haenv) = @_;
15
16 my $class = ref($this) || $this;
17
18 my $ms = $haenv->read_manager_status();
19
20 $ms->{master_node} = $haenv->nodename();
21
22 my $ns = PVE::HA::NodeStatus->new($haenv, $ms->{node_status} || {});
23
24 # fixme: use separate class PVE::HA::ServiceStatus
25 my $ss = $ms->{service_status} || {};
26
27 my $self = bless {
28 haenv => $haenv,
29 ms => $ms, # master status
30 ns => $ns, # PVE::HA::NodeStatus
31 ss => $ss, # service status
32 }, $class;
33
34 return $self;
35 }
36
37 sub cleanup {
38 my ($self) = @_;
39
40 # todo: ?
41 }
42
43 sub flush_master_status {
44 my ($self) = @_;
45
46 my ($haenv, $ms, $ns, $ss) = ($self->{haenv}, $self->{ms}, $self->{ns}, $self->{ss});
47
48 $ms->{node_status} = $ns->{status};
49 $ms->{service_status} = $ss;
50 $ms->{timestamp} = $haenv->get_time();
51
52 $haenv->write_manager_status($ms);
53 }
54
55 sub select_service_node {
56 my ($groups, $online_node_usage, $service_conf, $current_node, $try_next) = @_;
57
58 my $group = { 'nodes' => { $service_conf->{node} => 1 } }; # default group
59
60 $group = $groups->{ids}->{$service_conf->{group}} if $service_conf->{group} &&
61 $groups->{ids}->{$service_conf->{group}};
62
63 my $pri_groups = {};
64 my $group_members = {};
65 foreach my $entry (keys %{$group->{nodes}}) {
66 my ($node, $pri) = ($entry, 0);
67 if ($entry =~ m/^(\S+):(\d+)$/) {
68 ($node, $pri) = ($1, $2);
69 }
70 next if !defined($online_node_usage->{$node}); # offline
71 $pri_groups->{$pri}->{$node} = 1;
72 $group_members->{$node} = $pri;
73 }
74
75
76 # add non-group members to unrestricted groups (priority -1)
77 if (!$group->{restricted}) {
78 my $pri = -1;
79 foreach my $node (keys %$online_node_usage) {
80 next if defined($group_members->{$node});
81 $pri_groups->{$pri}->{$node} = 1;
82 $group_members->{$node} = -1;
83 }
84 }
85
86
87 my @pri_list = sort {$b <=> $a} keys %$pri_groups;
88 return undef if !scalar(@pri_list);
89
90 if (!$try_next && $group->{nofailback} && defined($group_members->{$current_node})) {
91 return $current_node;
92 }
93
94 # select node from top priority node list
95
96 my $top_pri = $pri_list[0];
97
98 my @nodes = sort {
99 $online_node_usage->{$a} <=> $online_node_usage->{$b} || $a cmp $b
100 } keys %{$pri_groups->{$top_pri}};
101
102 my $found;
103 for (my $i = scalar(@nodes) - 1; $i >= 0; $i--) {
104 my $node = $nodes[$i];
105 if ($node eq $current_node) {
106 $found = $i;
107 last;
108 }
109 }
110
111 if ($try_next) {
112
113 if (defined($found) && ($found < (scalar(@nodes) - 1))) {
114 return $nodes[$found + 1];
115 } else {
116 return $nodes[0];
117 }
118
119 } else {
120
121 return $nodes[$found] if defined($found);
122
123 return $nodes[0];
124
125 }
126 }
127
128 my $uid_counter = 0;
129
130 sub compute_new_uuid {
131 my ($state) = @_;
132
133 $uid_counter++;
134 return md5_base64($state . $$ . time() . $uid_counter);
135 }
136
137 my $valid_service_states = {
138 stopped => 1,
139 request_stop => 1,
140 started => 1,
141 fence => 1,
142 migrate => 1,
143 relocate => 1,
144 freeze => 1,
145 error => 1,
146 };
147
148 sub recompute_online_node_usage {
149 my ($self) = @_;
150
151 my $online_node_usage = {};
152
153 my $online_nodes = $self->{ns}->list_online_nodes();
154
155 foreach my $node (@$online_nodes) {
156 $online_node_usage->{$node} = 0;
157 }
158
159 foreach my $sid (keys %{$self->{ss}}) {
160 my $sd = $self->{ss}->{$sid};
161 my $state = $sd->{state};
162 if (defined($online_node_usage->{$sd->{node}})) {
163 if (($state eq 'started') || ($state eq 'request_stop') ||
164 ($state eq 'fence') || ($state eq 'freeze') || ($state eq 'error')) {
165 $online_node_usage->{$sd->{node}}++;
166 } elsif (($state eq 'migrate') || ($state eq 'relocate')) {
167 $online_node_usage->{$sd->{target}}++;
168 } elsif ($state eq 'stopped') {
169 # do nothing
170 } else {
171 die "should not be reached";
172 }
173 }
174 }
175
176 $self->{online_node_usage} = $online_node_usage;
177 }
178
179 my $change_service_state = sub {
180 my ($self, $sid, $new_state, %params) = @_;
181
182 my ($haenv, $ss) = ($self->{haenv}, $self->{ss});
183
184 my $sd = $ss->{$sid} || die "no such service '$sid";
185
186 my $old_state = $sd->{state};
187 my $old_node = $sd->{node};
188
189 die "no state change" if $old_state eq $new_state; # just to be sure
190
191 die "invalid CRM service state '$new_state'\n" if !$valid_service_states->{$new_state};
192
193 foreach my $k (keys %$sd) { delete $sd->{$k}; };
194
195 $sd->{state} = $new_state;
196 $sd->{node} = $old_node;
197
198 my $text_state = '';
199 foreach my $k (sort keys %params) {
200 my $v = $params{$k};
201 $text_state .= ", " if $text_state;
202 $text_state .= "$k = $v";
203 $sd->{$k} = $v;
204 }
205
206 $self->recompute_online_node_usage();
207
208 $sd->{uid} = compute_new_uuid($new_state);
209
210
211 $text_state = " ($text_state)" if $text_state;
212 $haenv->log('info', "service '$sid': state changed from '${old_state}' to '${new_state}' $text_state");
213 };
214
215 # read LRM status for all nodes
216 sub read_lrm_status {
217 my ($self) = @_;
218
219 my $nodes = $self->{ns}->list_nodes();
220 my $haenv = $self->{haenv};
221
222 my $results = {};
223 my $modes = {};
224 foreach my $node (@$nodes) {
225 my $lrm_status = $haenv->read_lrm_status($node);
226 $modes->{$node} = $lrm_status->{mode} || 'active';
227 foreach my $uid (keys %{$lrm_status->{results}}) {
228 next if $results->{$uid}; # should not happen
229 $results->{$uid} = $lrm_status->{results}->{$uid};
230 }
231 }
232
233
234 return ($results, $modes);
235 }
236
237 # read new crm commands and save them into crm master status
238 sub update_crm_commands {
239 my ($self) = @_;
240
241 my ($haenv, $ms, $ns, $ss) = ($self->{haenv}, $self->{ms}, $self->{ns}, $self->{ss});
242
243 my $cmdlist = $haenv->read_crm_commands();
244
245 foreach my $cmd (split(/\n/, $cmdlist)) {
246 chomp $cmd;
247
248 if ($cmd =~ m/^(migrate|relocate)\s+(\S+)\s+(\S+)$/) {
249 my ($task, $sid, $node) = ($1, $2, $3);
250 if (my $sd = $ss->{$sid}) {
251 if (!$ns->node_is_online($node)) {
252 $haenv->log('err', "crm command error - node not online: $cmd");
253 } else {
254 if ($node eq $sd->{node}) {
255 $haenv->log('info', "ignore crm command - service already on target node: $cmd");
256 } else {
257 $haenv->log('info', "got crm command: $cmd");
258 $ss->{$sid}->{cmd} = [ $task, $node];
259 }
260 }
261 } else {
262 $haenv->log('err', "crm command error - no such service: $cmd");
263 }
264
265 } else {
266 $haenv->log('err', "unable to parse crm command: $cmd");
267 }
268 }
269
270 }
271
272 sub manage {
273 my ($self) = @_;
274
275 my ($haenv, $ms, $ns, $ss) = ($self->{haenv}, $self->{ms}, $self->{ns}, $self->{ss});
276
277 $ns->update($haenv->get_node_info());
278
279 if (!$ns->node_is_online($haenv->nodename())) {
280 $haenv->log('info', "master seems offline");
281 return;
282 }
283
284 my ($lrm_results, $lrm_modes) = $self->read_lrm_status();
285
286 my $sc = $haenv->read_service_config();
287
288 $self->{groups} = $haenv->read_group_config(); # update
289
290 # compute new service status
291
292 # add new service
293 foreach my $sid (sort keys %$sc) {
294 next if $ss->{$sid}; # already there
295 $haenv->log('info', "adding new service '$sid' on node '$sc->{$sid}->{node}'");
296 # assume we are running to avoid relocate running service at add
297 $ss->{$sid} = { state => 'started', node => $sc->{$sid}->{node},
298 uid => compute_new_uuid('started') };
299 }
300
301 # remove stale service from manager state
302 foreach my $sid (keys %$ss) {
303 next if $sc->{$sid};
304 $haenv->log('info', "removing stale service '$sid' (no config)");
305 delete $ss->{$sid};
306 }
307
308 $self->update_crm_commands();
309
310 for (;;) {
311 my $repeat = 0;
312
313 $self->recompute_online_node_usage();
314
315 foreach my $sid (keys %$ss) {
316 my $sd = $ss->{$sid};
317 my $cd = $sc->{$sid} || { state => 'disabled' };
318
319 my $lrm_res = $sd->{uid} ? $lrm_results->{$sd->{uid}} : undef;
320
321 my $last_state = $sd->{state};
322
323 if ($last_state eq 'stopped') {
324
325 $self->next_state_stopped($sid, $cd, $sd, $lrm_res);
326
327 } elsif ($last_state eq 'started') {
328
329 $self->next_state_started($sid, $cd, $sd, $lrm_res);
330
331 } elsif ($last_state eq 'migrate' || $last_state eq 'relocate') {
332
333 $self->next_state_migrate_relocate($sid, $cd, $sd, $lrm_res);
334
335 } elsif ($last_state eq 'fence') {
336
337 # do nothing here - wait until fenced
338
339 } elsif ($last_state eq 'request_stop') {
340
341 $self->next_state_request_stop($sid, $cd, $sd, $lrm_res);
342
343 } elsif ($last_state eq 'freeze') {
344
345 my $lrm_mode = $sd->{node} ? $lrm_modes->{$sd->{node}} : undef;
346 # unfreeze
347 &$change_service_state($self, $sid, 'started')
348 if $lrm_mode && $lrm_mode eq 'active';
349
350 } elsif ($last_state eq 'error') {
351
352 $self->next_state_error($sid, $cd, $sd, $lrm_res);
353
354 } else {
355
356 die "unknown service state '$last_state'";
357 }
358
359
360 my $lrm_mode = $sd->{node} ? $lrm_modes->{$sd->{node}} : undef;
361 $lrm_mode = 'unknown'if !$lrm_mode;
362 if (($sd->{state} eq 'started' || $sd->{state} eq 'stopped' ||
363 $sd->{state} eq 'request_stop') && ($lrm_mode ne 'active')) {
364 &$change_service_state($self, $sid, 'freeze');
365 }
366
367 $repeat = 1 if $sd->{state} ne $last_state;
368 }
369
370 # handle fencing
371 my $fenced_nodes = {};
372 foreach my $sid (keys %$ss) {
373 my $sd = $ss->{$sid};
374 next if $sd->{state} ne 'fence';
375
376 if (!defined($fenced_nodes->{$sd->{node}})) {
377 $fenced_nodes->{$sd->{node}} = $ns->fence_node($sd->{node}) || 0;
378 }
379
380 next if !$fenced_nodes->{$sd->{node}};
381
382 # node fence was successful - mark service as stopped
383 &$change_service_state($self, $sid, 'stopped');
384 }
385
386 last if !$repeat;
387 }
388
389 $self->flush_master_status();
390 }
391
392 # functions to compute next service states
393 # $cd: service configuration data (read only)
394 # $sd: service status data (read only)
395 #
396 # Note: use change_service_state() to alter state
397 #
398
399 sub next_state_request_stop {
400 my ($self, $sid, $cd, $sd, $lrm_res) = @_;
401
402 my $haenv = $self->{haenv};
403 my $ns = $self->{ns};
404
405 # check result from LRM daemon
406 if ($lrm_res) {
407 my $exit_code = $lrm_res->{exit_code};
408 if ($exit_code == 0) {
409 &$change_service_state($self, $sid, 'stopped');
410 return;
411 } else {
412 &$change_service_state($self, $sid, 'error'); # fixme: what state?
413 return;
414 }
415 }
416
417 if ($ns->node_is_offline_delayed($sd->{node}, $fence_delay)) {
418 &$change_service_state($self, $sid, 'fence');
419 return;
420 }
421 }
422
423 sub next_state_migrate_relocate {
424 my ($self, $sid, $cd, $sd, $lrm_res) = @_;
425
426 my $haenv = $self->{haenv};
427 my $ns = $self->{ns};
428
429 # check result from LRM daemon
430 if ($lrm_res) {
431 my $exit_code = $lrm_res->{exit_code};
432 if ($exit_code == 0) {
433 &$change_service_state($self, $sid, 'started', node => $sd->{target});
434 return;
435 } else {
436 $haenv->log('err', "service '$sid' - migration failed (exit code $exit_code)");
437 &$change_service_state($self, $sid, 'started', node => $sd->{node});
438 return;
439 }
440 }
441
442 if ($ns->node_is_offline_delayed($sd->{node}, $fence_delay)) {
443 &$change_service_state($self, $sid, 'fence');
444 return;
445 }
446 }
447
448
449 sub next_state_stopped {
450 my ($self, $sid, $cd, $sd, $lrm_res) = @_;
451
452 my $haenv = $self->{haenv};
453 my $ns = $self->{ns};
454
455 if ($sd->{node} ne $cd->{node}) {
456 # this can happen if we fence a node with active migrations
457 # hack: modify $sd (normally this should be considered read-only)
458 $haenv->log('info', "fixup service '$sid' location ($sd->{node} => $cd->{node}");
459 $sd->{node} = $cd->{node};
460 }
461
462 if ($sd->{cmd}) {
463 my ($cmd, $target) = @{$sd->{cmd}};
464 delete $sd->{cmd};
465
466 if ($cmd eq 'migrate' || $cmd eq 'relocate') {
467 if (!$ns->node_is_online($target)) {
468 $haenv->log('err', "ignore service '$sid' $cmd request - node '$target' not online");
469 } elsif ($sd->{node} eq $target) {
470 $haenv->log('info', "ignore service '$sid' $cmd request - service already on node '$target'");
471 } else {
472 eval {
473 $haenv->change_service_location($sid, $sd->{node}, $target);
474 $cd->{node} = $sd->{node} = $target; # fixme: $sd is read-only??!!
475 $haenv->log('info', "$cmd service '$sid' to node '$target' (stopped)");
476 };
477 if (my $err = $@) {
478 $haenv->log('err', "$cmd service '$sid' to node '$target' failed - $err");
479 }
480 }
481 } else {
482 $haenv->log('err', "unknown command '$cmd' for service '$sid'");
483 }
484 }
485
486 if ($cd->{state} eq 'disabled') {
487 # do nothing
488 return;
489 }
490
491 if ($cd->{state} eq 'enabled') {
492 if (my $node = select_service_node($self->{groups}, $self->{online_node_usage}, $cd, $sd->{node})) {
493 if ($node && ($sd->{node} ne $node)) {
494 eval {
495 $haenv->change_service_location($sid, $sd->{node}, $node);
496 $cd->{node} = $sd->{node} = $node; # fixme: $sd is read-only??!!
497 };
498 if (my $err = $@) {
499 $haenv->log('err', "move service '$sid' to node '$node' failed - $err");
500 } else {
501 &$change_service_state($self, $sid, 'started', node => $node);
502 }
503 } else {
504 &$change_service_state($self, $sid, 'started', node => $node);
505 }
506 } else {
507 # fixme: warn
508 }
509
510 return;
511 }
512
513 $haenv->log('err', "service '$sid' - unknown state '$cd->{state}' in service configuration");
514 }
515
516 sub next_state_started {
517 my ($self, $sid, $cd, $sd, $lrm_res) = @_;
518
519 my $haenv = $self->{haenv};
520 my $master_status = $self->{ms};
521 my $ns = $self->{ns};
522
523 if (!$ns->node_is_online($sd->{node})) {
524 if ($ns->node_is_offline_delayed($sd->{node}, $fence_delay)) {
525 &$change_service_state($self, $sid, 'fence');
526 }
527 return;
528 }
529
530 if ($cd->{state} eq 'disabled') {
531 &$change_service_state($self, $sid, 'request_stop');
532 return;
533 }
534
535 if ($cd->{state} eq 'enabled') {
536
537 if ($sd->{cmd}) {
538 my ($cmd, $target) = @{$sd->{cmd}};
539 delete $sd->{cmd};
540
541 if ($cmd eq 'migrate' || $cmd eq 'relocate') {
542 if (!$ns->node_is_online($target)) {
543 $haenv->log('err', "ignore service '$sid' $cmd request - node '$target' not online");
544 } elsif ($sd->{node} eq $target) {
545 $haenv->log('info', "ignore service '$sid' $cmd request - service already on node '$target'");
546 } else {
547 $haenv->log('info', "$cmd service '$sid' to node '$target' (running)");
548 &$change_service_state($self, $sid, $cmd, node => $sd->{node}, target => $target);
549 }
550 } else {
551 $haenv->log('err', "unknown command '$cmd' for service '$sid'");
552 }
553 } else {
554
555 my $try_next = 0;
556 if ($lrm_res) {
557 if ($lrm_res->{exit_code} == 1) {
558
559 my $try = $master_status->{relocate_trial}->{$sid} || 0;
560
561 if ($try < $cd->{max_relocate}) {
562
563 $try++;
564 $try_next = 1; # tell select_service_node to relocate
565
566 $haenv->log('warning', "starting service $sid on node".
567 " '$sd->{node}' failed, relocating service.");
568 $master_status->{relocate_trial}->{$sid} = $try;
569
570 } else {
571
572 $haenv->log('err', "recovery policy for service".
573 " $sid failed, entering error state!");
574 &$change_service_state($self, $sid, 'error');
575 return;
576
577 }
578 } elsif ($lrm_res->{exit_code} == 0) {
579 $master_status->{relocate_trial}->{$sid} = 0;
580 }
581 }
582
583 my $node = select_service_node($self->{groups}, $self->{online_node_usage},
584 $cd, $sd->{node}, $try_next);
585
586 if ($node && ($sd->{node} ne $node)) {
587 $haenv->log('info', "migrate service '$sid' to node '$node' (running)");
588 &$change_service_state($self, $sid, 'migrate', node => $sd->{node}, target => $node);
589 } else {
590 # do nothing
591 }
592 }
593
594 return;
595 }
596
597 $haenv->log('err', "service '$sid' - unknown state '$cd->{state}' in service configuration");
598 }
599
600 sub next_state_error {
601 my ($self, $sid, $cd, $sd, $lrm_res) = @_;
602
603 my $ns = $self->{ns};
604
605 if ($cd->{state} eq 'disabled') {
606 &$change_service_state($self, $sid, 'stopped');
607 return;
608 }
609
610 if ($ns->node_is_offline_delayed($sd->{node}, $fence_delay)) {
611 &$change_service_state($self, $sid, 'fence');
612 return;
613 }
614
615 }
616
617 1;