]> git.proxmox.com Git - pve-ha-manager.git/blob - src/PVE/HA/LRM.pm
update lrm status on each iteration
[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;
13
14 # Server can have several states:
15
16 my $valid_states = {
17 wait_for_agent_lock => "waiting for agent lock",
18 active => "got agent_lock",
19 lost_agent_lock => "lost agent_lock",
20 };
21
22 sub new {
23 my ($this, $haenv) = @_;
24
25 my $class = ref($this) || $this;
26
27 my $self = bless {
28 haenv => $haenv,
29 status => { state => 'startup' },
30 workers => {},
31 results => {},
32 shutdown_request => 0,
33 # mode can be: active, reboot, shutdown, restart
34 mode => 'active',
35 }, $class;
36
37 $self->set_local_status({ state => 'wait_for_agent_lock' });
38
39 return $self;
40 }
41
42 sub shutdown_request {
43 my ($self) = @_;
44
45 $self->{shutdown_request} = 1;
46
47 $self->{mode} = 'restart'; # fixme: detect shutdown/reboot
48
49 eval { $self->update_lrm_status(); };
50 if (my $err = $@) {
51 $self->log('err', "unable to update lrm status file - $err");
52 }
53 }
54
55 sub get_local_status {
56 my ($self) = @_;
57
58 return $self->{status};
59 }
60
61 sub set_local_status {
62 my ($self, $new) = @_;
63
64 die "invalid state '$new->{state}'" if !$valid_states->{$new->{state}};
65
66 my $haenv = $self->{haenv};
67
68 my $old = $self->{status};
69
70 # important: only update if if really changed
71 return if $old->{state} eq $new->{state};
72
73 $haenv->log('info', "status change $old->{state} => $new->{state}");
74
75 $new->{state_change_time} = $haenv->get_time();
76
77 $self->{status} = $new;
78 }
79
80 sub update_lrm_status {
81 my ($self) = @_;
82
83 my $haenv = $self->{haenv};
84
85 return 0 if !$haenv->quorate();
86
87 my $lrm_status = {
88 mode => $self->{mode},
89 results => $self->{results},
90 timestamp => $haenv->get_time(),
91 };
92
93 eval { $haenv->write_lrm_status($lrm_status); };
94 if (my $err = $@) {
95 $haenv->log('err', "unable to write lrm status file - $err");
96 return 0;
97 }
98
99 return 1;
100 }
101
102 sub get_protected_ha_agent_lock {
103 my ($self) = @_;
104
105 my $haenv = $self->{haenv};
106
107 my $count = 0;
108 my $starttime = $haenv->get_time();
109
110 for (;;) {
111
112 if ($haenv->get_ha_agent_lock()) {
113 if ($self->{ha_agent_wd}) {
114 $haenv->watchdog_update($self->{ha_agent_wd});
115 } else {
116 my $wfh = $haenv->watchdog_open();
117 $self->{ha_agent_wd} = $wfh;
118 }
119 return 1;
120 }
121
122 last if ++$count > 5; # try max 5 time
123
124 my $delay = $haenv->get_time() - $starttime;
125 last if $delay > 5; # for max 5 seconds
126
127 $haenv->sleep(1);
128 }
129
130 return 0;
131 }
132
133 sub active_service_count {
134 my ($self) = @_;
135
136 my $haenv = $self->{haenv};
137
138 my $nodename = $haenv->nodename();
139
140 my $ss = $self->{service_status};
141
142 my $count = 0;
143
144 foreach my $sid (keys %$ss) {
145 my $sd = $ss->{$sid};
146 next if !$sd->{node};
147 next if $sd->{node} ne $nodename;
148 my $req_state = $sd->{state};
149 next if !defined($req_state);
150 next if $req_state eq 'stopped';
151 next if $req_state eq 'freeze';
152
153 $count++;
154 }
155
156 return $count;
157 }
158
159 my $wrote_lrm_status_at_startup = 0;
160
161 sub do_one_iteration {
162 my ($self) = @_;
163
164 my $haenv = $self->{haenv};
165
166 if (!$wrote_lrm_status_at_startup) {
167 if ($self->update_lrm_status()) {
168 $wrote_lrm_status_at_startup = 1;
169 } else {
170 # do nothing
171 $haenv->sleep(5);
172 return $self->{shutdown_request} ? 0 : 1;
173 }
174 }
175
176 my $status = $self->get_local_status();
177 my $state = $status->{state};
178
179 my $ms = $haenv->read_manager_status();
180 $self->{service_status} = $ms->{service_status} || {};
181
182 my $fence_request = PVE::HA::Tools::count_fenced_services($self->{service_status}, $haenv->nodename());
183
184 # do state changes first
185
186 my $ctime = $haenv->get_time();
187
188 if ($state eq 'wait_for_agent_lock') {
189
190 my $service_count = $self->active_service_count();
191
192 if (!$fence_request && $service_count && $haenv->quorate()) {
193 if ($self->get_protected_ha_agent_lock()) {
194 $self->set_local_status({ state => 'active' });
195 }
196 }
197
198 } elsif ($state eq 'lost_agent_lock') {
199
200 if (!$fence_request && $haenv->quorate()) {
201 if ($self->get_protected_ha_agent_lock()) {
202 $self->set_local_status({ state => 'active' });
203 }
204 }
205
206 } elsif ($state eq 'active') {
207
208 if ($fence_request) {
209 $haenv->log('err', "node need to be fenced - releasing agent_lock\n");
210 $self->set_local_status({ state => 'lost_agent_lock'});
211 } elsif (!$self->get_protected_ha_agent_lock()) {
212 $self->set_local_status({ state => 'lost_agent_lock'});
213 }
214 }
215
216 $status = $self->get_local_status();
217 $state = $status->{state};
218
219 # do work
220
221 if ($state eq 'wait_for_agent_lock') {
222
223 return 0 if $self->{shutdown_request};
224
225 $self->update_lrm_status();
226
227 $haenv->sleep(5);
228
229 } elsif ($state eq 'active') {
230
231 my $startime = $haenv->get_time();
232
233 my $max_time = 10;
234
235 my $shutdown = 0;
236
237 # do work (max_time seconds)
238 eval {
239 # fixme: set alert timer
240
241 if ($self->{shutdown_request}) {
242
243 # fixme: request service stop or relocate ?
244
245 my $service_count = $self->active_service_count();
246
247 if ($service_count == 0) {
248
249 if ($self->{ha_agent_wd}) {
250 $haenv->watchdog_close($self->{ha_agent_wd});
251 delete $self->{ha_agent_wd};
252 }
253
254 $shutdown = 1;
255 }
256 } else {
257
258 $self->manage_resources();
259
260 }
261 };
262 if (my $err = $@) {
263 $haenv->log('err', "got unexpected error - $err");
264 }
265
266 $self->update_lrm_status();
267
268 return 0 if $shutdown;
269
270 $haenv->sleep_until($startime + $max_time);
271
272 } elsif ($state eq 'lost_agent_lock') {
273
274 # Note: watchdog is active an will triger soon!
275
276 # so we hope to get the lock back soon!
277
278 if ($self->{shutdown_request}) {
279
280 my $service_count = $self->active_service_count();
281
282 if ($service_count > 0) {
283 $haenv->log('err', "get shutdown request in state 'lost_agent_lock' - " .
284 "detected $service_count running services");
285
286 } else {
287
288 # all services are stopped, so we can close the watchdog
289
290 if ($self->{ha_agent_wd}) {
291 $haenv->watchdog_close($self->{ha_agent_wd});
292 delete $self->{ha_agent_wd};
293 }
294
295 return 0;
296 }
297 }
298
299 $haenv->sleep(5);
300
301 } else {
302
303 die "got unexpected status '$state'\n";
304
305 }
306
307 return 1;
308 }
309
310 sub manage_resources {
311 my ($self) = @_;
312
313 my $haenv = $self->{haenv};
314
315 my $nodename = $haenv->nodename();
316
317 my $ss = $self->{service_status};
318
319 foreach my $sid (keys %$ss) {
320 my $sd = $ss->{$sid};
321 next if !$sd->{node};
322 next if !$sd->{uid};
323 next if $sd->{node} ne $nodename;
324 my $req_state = $sd->{state};
325 next if !defined($req_state);
326 next if $req_state eq 'freeze';
327 eval {
328 $self->queue_resource_command($sid, $sd->{uid}, $req_state, $sd->{target});
329 };
330 if (my $err = $@) {
331 $haenv->log('err', "unable to run resource agent for '$sid' - $err"); # fixme
332 }
333 }
334
335 my $starttime = $haenv->get_time();
336
337 # start workers
338 my $max_workers = 4;
339
340 my $sc = $haenv->read_service_config();
341
342 while (($haenv->get_time() - $starttime) < 5) {
343 my $count = $self->check_active_workers();
344
345 foreach my $sid (keys %{$self->{workers}}) {
346 last if $count >= $max_workers;
347 my $w = $self->{workers}->{$sid};
348 my $cd = $sc->{$sid};
349 if (!$cd) {
350 $haenv->log('err', "missing resource configuration for '$sid'");
351 next;
352 }
353 if (!$w->{pid}) {
354 if ($haenv->can_fork()) {
355 my $pid = fork();
356 if (!defined($pid)) {
357 $haenv->log('err', "fork worker failed");
358 $count = 0; last; # abort, try later
359 } elsif ($pid == 0) {
360 # do work
361 my $res = -1;
362 eval {
363 $res = $haenv->exec_resource_agent($sid, $cd, $w->{state}, $w->{target});
364 };
365 if (my $err = $@) {
366 $haenv->log('err', $err);
367 POSIX::_exit(-1);
368 }
369 POSIX::_exit($res);
370 } else {
371 $count++;
372 $w->{pid} = $pid;
373 }
374 } else {
375 my $res = -1;
376 eval {
377 $res = $haenv->exec_resource_agent($sid, $cd, $w->{state}, $w->{target});
378 };
379 if (my $err = $@) {
380 $haenv->log('err', $err);
381 }
382 $self->resource_command_finished($sid, $w->{uid}, $res);
383 }
384 }
385 }
386
387 last if !$count;
388
389 $haenv->sleep(1);
390 }
391 }
392
393 # fixme: use a queue an limit number of parallel workers?
394 sub queue_resource_command {
395 my ($self, $sid, $uid, $state, $target) = @_;
396
397 if (my $w = $self->{workers}->{$sid}) {
398 return if $w->{pid}; # already started
399 # else, delete and overwrite queue entry with new command
400 delete $self->{workers}->{$sid};
401 }
402
403 $self->{workers}->{$sid} = {
404 sid => $sid,
405 uid => $uid,
406 state => $state,
407 };
408
409 $self->{workers}->{$sid}->{target} = $target if $target;
410 }
411
412 sub check_active_workers {
413 my ($self) = @_;
414
415 # finish/count workers
416 my $count = 0;
417 foreach my $sid (keys %{$self->{workers}}) {
418 my $w = $self->{workers}->{$sid};
419 if (my $pid = $w->{pid}) {
420 # check status
421 my $waitpid = waitpid($pid, WNOHANG);
422 if (defined($waitpid) && ($waitpid == $pid)) {
423 $self->resource_command_finished($sid, $w->{uid}, $?);
424 } else {
425 $count++;
426 }
427 }
428 }
429
430 return $count;
431 }
432
433 sub resource_command_finished {
434 my ($self, $sid, $uid, $status) = @_;
435
436 my $haenv = $self->{haenv};
437
438 my $w = delete $self->{workers}->{$sid};
439 return if !$w; # should not happen
440
441 my $exit_code = -1;
442
443 if ($status == -1) {
444 $haenv->log('err', "resource agent $sid finished - failed to execute");
445 } elsif (my $sig = ($status & 127)) {
446 $haenv->log('err', "resource agent $sid finished - got signal $sig");
447 } else {
448 $exit_code = ($status >> 8);
449 }
450
451 $self->{results}->{$uid} = {
452 sid => $w->{sid},
453 state => $w->{state},
454 exit_code => $exit_code,
455 };
456
457 my $ss = $self->{service_status};
458
459 # compute hash of valid/existing uids
460 my $valid_uids = {};
461 foreach my $sid (keys %$ss) {
462 my $sd = $ss->{$sid};
463 next if !$sd->{uid};
464 $valid_uids->{$sd->{uid}} = 1;
465 }
466
467 my $results = {};
468 foreach my $id (keys %{$self->{results}}) {
469 next if !$valid_uids->{$id};
470 $results->{$id} = $self->{results}->{$id};
471 }
472 $self->{results} = $results;
473 }
474
475 1;