]>
Commit | Line | Data |
---|---|---|
5f095798 DM |
1 | package PVE::HA::LRM; |
2 | ||
3 | # Local Resource Manager | |
4 | ||
5 | use strict; | |
6 | use warnings; | |
c4a221bc DM |
7 | use Data::Dumper; |
8 | use POSIX qw(:sys_wait_h); | |
5f095798 DM |
9 | |
10 | use PVE::SafeSyslog; | |
11 | use PVE::Tools; | |
12 | use PVE::HA::Tools; | |
13 | ||
14 | # Server can have several states: | |
15 | ||
16 | my $valid_states = { | |
ec911edd | 17 | wait_for_agent_lock => "waiting for agent lock", |
0bba8f60 | 18 | active => "got agent_lock", |
5f095798 DM |
19 | lost_agent_lock => "lost agent_lock", |
20 | }; | |
21 | ||
22 | sub new { | |
23 | my ($this, $haenv) = @_; | |
24 | ||
25 | my $class = ref($this) || $this; | |
26 | ||
27 | my $self = bless { | |
28 | haenv => $haenv, | |
29 | status => { state => 'startup' }, | |
c4a221bc DM |
30 | workers => {}, |
31 | results => {}, | |
067cdf33 | 32 | shutdown_request => 0, |
5f095798 DM |
33 | }, $class; |
34 | ||
b0bf08a9 | 35 | $self->set_local_status({ state => 'wait_for_agent_lock' }); |
5f095798 DM |
36 | |
37 | return $self; | |
38 | } | |
39 | ||
40 | sub shutdown_request { | |
41 | my ($self) = @_; | |
42 | ||
43 | $self->{shutdown_request} = 1; | |
44 | } | |
45 | ||
46 | sub get_local_status { | |
47 | my ($self) = @_; | |
48 | ||
49 | return $self->{status}; | |
50 | } | |
51 | ||
52 | sub set_local_status { | |
53 | my ($self, $new) = @_; | |
54 | ||
55 | die "invalid state '$new->{state}'" if !$valid_states->{$new->{state}}; | |
56 | ||
57 | my $haenv = $self->{haenv}; | |
58 | ||
59 | my $old = $self->{status}; | |
60 | ||
61 | # important: only update if if really changed | |
62 | return if $old->{state} eq $new->{state}; | |
63 | ||
0bba8f60 | 64 | $haenv->log('info', "status change $old->{state} => $new->{state}"); |
5f095798 DM |
65 | |
66 | $new->{state_change_time} = $haenv->get_time(); | |
67 | ||
68 | $self->{status} = $new; | |
69 | } | |
70 | ||
71 | sub get_protected_ha_agent_lock { | |
72 | my ($self) = @_; | |
73 | ||
74 | my $haenv = $self->{haenv}; | |
75 | ||
76 | my $count = 0; | |
77 | my $starttime = $haenv->get_time(); | |
78 | ||
79 | for (;;) { | |
80 | ||
81 | if ($haenv->get_ha_agent_lock()) { | |
82 | if ($self->{ha_agent_wd}) { | |
83 | $haenv->watchdog_update($self->{ha_agent_wd}); | |
84 | } else { | |
85 | my $wfh = $haenv->watchdog_open(); | |
86 | $self->{ha_agent_wd} = $wfh; | |
87 | } | |
88 | return 1; | |
89 | } | |
90 | ||
91 | last if ++$count > 5; # try max 5 time | |
92 | ||
93 | my $delay = $haenv->get_time() - $starttime; | |
94 | last if $delay > 5; # for max 5 seconds | |
95 | ||
96 | $haenv->sleep(1); | |
97 | } | |
98 | ||
99 | return 0; | |
100 | } | |
101 | ||
067cdf33 DM |
102 | sub fenced_service_count { |
103 | my ($self) = @_; | |
104 | ||
105 | my $haenv = $self->{haenv}; | |
106 | ||
107 | my $nodename = $haenv->nodename(); | |
108 | ||
109 | my $ss = $self->{service_status}; | |
110 | ||
111 | my $count = 0; | |
112 | ||
113 | foreach my $sid (keys %$ss) { | |
114 | my $sd = $ss->{$sid}; | |
115 | next if !$sd->{node}; | |
116 | next if $sd->{node} ne $nodename; | |
117 | my $req_state = $sd->{state}; | |
118 | next if !defined($req_state); | |
119 | if ($req_state eq 'fence') { | |
120 | $count++; | |
121 | next; | |
122 | } | |
123 | } | |
124 | ||
125 | return $count; | |
126 | } | |
546e2f1f DM |
127 | |
128 | sub active_service_count { | |
129 | my ($self) = @_; | |
130 | ||
131 | my $haenv = $self->{haenv}; | |
132 | ||
133 | my $nodename = $haenv->nodename(); | |
134 | ||
135 | my $ss = $self->{service_status}; | |
136 | ||
137 | my $count = 0; | |
138 | ||
139 | foreach my $sid (keys %$ss) { | |
140 | my $sd = $ss->{$sid}; | |
141 | next if !$sd->{node}; | |
142 | next if $sd->{node} ne $nodename; | |
143 | my $req_state = $sd->{state}; | |
144 | next if !defined($req_state); | |
145 | next if $req_state eq 'stopped'; | |
146 | ||
147 | $count++; | |
148 | } | |
149 | ||
150 | return $count; | |
151 | } | |
067cdf33 | 152 | |
5f095798 DM |
153 | sub do_one_iteration { |
154 | my ($self) = @_; | |
155 | ||
156 | my $haenv = $self->{haenv}; | |
157 | ||
158 | my $status = $self->get_local_status(); | |
159 | my $state = $status->{state}; | |
160 | ||
067cdf33 DM |
161 | my $ms = $haenv->read_manager_status(); |
162 | $self->{service_status} = $ms->{service_status} || {}; | |
163 | ||
164 | my $fence_request = $self->fenced_service_count(); | |
165 | ||
5f095798 DM |
166 | # do state changes first |
167 | ||
168 | my $ctime = $haenv->get_time(); | |
169 | ||
b0bf08a9 | 170 | if ($state eq 'wait_for_agent_lock') { |
5f095798 | 171 | |
546e2f1f | 172 | my $service_count = $self->active_service_count(); |
5f095798 | 173 | |
067cdf33 | 174 | if (!$fence_request && $service_count && $haenv->quorate()) { |
0bba8f60 DM |
175 | if ($self->get_protected_ha_agent_lock()) { |
176 | $self->set_local_status({ state => 'active' }); | |
5f095798 DM |
177 | } |
178 | } | |
179 | ||
180 | } elsif ($state eq 'lost_agent_lock') { | |
181 | ||
067cdf33 | 182 | if (!$fence_request && $haenv->quorate()) { |
0bba8f60 DM |
183 | if ($self->get_protected_ha_agent_lock()) { |
184 | $self->set_local_status({ state => 'active' }); | |
5f095798 DM |
185 | } |
186 | } | |
187 | ||
0bba8f60 | 188 | } elsif ($state eq 'active') { |
5f095798 | 189 | |
067cdf33 DM |
190 | if ($fence_request) { |
191 | $haenv->log('err', "node need to be fenced - releasing agent_lock\n"); | |
192 | $self->set_local_status({ state => 'lost_agent_lock'}); | |
193 | } elsif (!$self->get_protected_ha_agent_lock()) { | |
5f095798 DM |
194 | $self->set_local_status({ state => 'lost_agent_lock'}); |
195 | } | |
196 | } | |
197 | ||
198 | $status = $self->get_local_status(); | |
199 | $state = $status->{state}; | |
200 | ||
201 | # do work | |
202 | ||
203 | if ($state eq 'wait_for_agent_lock') { | |
204 | ||
205 | return 0 if $self->{shutdown_request}; | |
206 | ||
207 | $haenv->sleep(5); | |
208 | ||
0bba8f60 | 209 | } elsif ($state eq 'active') { |
5f095798 DM |
210 | |
211 | my $startime = $haenv->get_time(); | |
212 | ||
213 | my $max_time = 10; | |
214 | ||
215 | my $shutdown = 0; | |
216 | ||
217 | # do work (max_time seconds) | |
218 | eval { | |
219 | # fixme: set alert timer | |
220 | ||
221 | if ($self->{shutdown_request}) { | |
222 | ||
223 | # fixme: request service stop or relocate ? | |
224 | ||
546e2f1f | 225 | my $service_count = $self->active_service_count(); |
5f095798 DM |
226 | |
227 | if ($service_count == 0) { | |
228 | ||
229 | if ($self->{ha_agent_wd}) { | |
230 | $haenv->watchdog_close($self->{ha_agent_wd}); | |
231 | delete $self->{ha_agent_wd}; | |
232 | } | |
233 | ||
234 | $shutdown = 1; | |
235 | } | |
c4a221bc | 236 | } else { |
c4a221bc DM |
237 | |
238 | $self->manage_resources(); | |
067cdf33 | 239 | |
5f095798 DM |
240 | } |
241 | }; | |
242 | if (my $err = $@) { | |
243 | $haenv->log('err', "got unexpected error - $err"); | |
244 | } | |
245 | ||
246 | return 0 if $shutdown; | |
247 | ||
248 | $haenv->sleep_until($startime + $max_time); | |
249 | ||
250 | } elsif ($state eq 'lost_agent_lock') { | |
251 | ||
252 | # Note: watchdog is active an will triger soon! | |
253 | ||
254 | # so we hope to get the lock back soon! | |
255 | ||
256 | if ($self->{shutdown_request}) { | |
257 | ||
546e2f1f | 258 | my $service_count = $self->active_service_count(); |
5f095798 | 259 | |
546e2f1f | 260 | if ($service_count > 0) { |
5f095798 | 261 | $haenv->log('err', "get shutdown request in state 'lost_agent_lock' - " . |
546e2f1f | 262 | "detected $service_count running services"); |
5f095798 | 263 | |
546e2f1f | 264 | } else { |
5f095798 | 265 | |
546e2f1f | 266 | # all services are stopped, so we can close the watchdog |
5f095798 | 267 | |
546e2f1f DM |
268 | if ($self->{ha_agent_wd}) { |
269 | $haenv->watchdog_close($self->{ha_agent_wd}); | |
270 | delete $self->{ha_agent_wd}; | |
271 | } | |
272 | ||
273 | return 0; | |
5f095798 | 274 | } |
5f095798 DM |
275 | } |
276 | ||
b0bf08a9 DM |
277 | $haenv->sleep(5); |
278 | ||
5f095798 DM |
279 | } else { |
280 | ||
281 | die "got unexpected status '$state'\n"; | |
282 | ||
283 | } | |
284 | ||
285 | return 1; | |
286 | } | |
287 | ||
c4a221bc DM |
288 | sub manage_resources { |
289 | my ($self) = @_; | |
290 | ||
291 | my $haenv = $self->{haenv}; | |
292 | ||
293 | my $nodename = $haenv->nodename(); | |
294 | ||
c4a221bc DM |
295 | my $ss = $self->{service_status}; |
296 | ||
297 | foreach my $sid (keys %$ss) { | |
298 | my $sd = $ss->{$sid}; | |
299 | next if !$sd->{node}; | |
300 | next if !$sd->{uid}; | |
301 | next if $sd->{node} ne $nodename; | |
302 | my $req_state = $sd->{state}; | |
303 | next if !defined($req_state); | |
c4a221bc | 304 | eval { |
e88469ba | 305 | $self->queue_resource_command($sid, $sd->{uid}, $req_state, $sd->{target}); |
c4a221bc DM |
306 | }; |
307 | if (my $err = $@) { | |
f31b7e94 | 308 | $haenv->log('err', "unable to run resource agent for '$sid' - $err"); # fixme |
c4a221bc DM |
309 | } |
310 | } | |
311 | ||
f31b7e94 | 312 | my $starttime = $haenv->get_time(); |
c4a221bc DM |
313 | |
314 | # start workers | |
315 | my $max_workers = 4; | |
316 | ||
6dbf93a0 | 317 | my $sc = $haenv->read_service_config(); |
f31b7e94 DM |
318 | |
319 | while (($haenv->get_time() - $starttime) < 5) { | |
c4a221bc DM |
320 | my $count = $self->check_active_workers(); |
321 | ||
322 | foreach my $sid (keys %{$self->{workers}}) { | |
323 | last if $count >= $max_workers; | |
324 | my $w = $self->{workers}->{$sid}; | |
6dbf93a0 DM |
325 | my $cd = $sc->{$sid}; |
326 | if (!$cd) { | |
f31b7e94 | 327 | $haenv->log('err', "missing resource configuration for '$sid'"); |
6dbf93a0 DM |
328 | next; |
329 | } | |
c4a221bc | 330 | if (!$w->{pid}) { |
f31b7e94 DM |
331 | if ($haenv->can_fork()) { |
332 | my $pid = fork(); | |
333 | if (!defined($pid)) { | |
334 | $haenv->log('err', "fork worker failed"); | |
335 | $count = 0; last; # abort, try later | |
336 | } elsif ($pid == 0) { | |
337 | # do work | |
338 | my $res = -1; | |
339 | eval { | |
340 | $res = $haenv->exec_resource_agent($sid, $cd, $w->{state}, $w->{target}); | |
341 | }; | |
342 | if (my $err = $@) { | |
343 | $haenv->log('err', $err); | |
344 | POSIX::_exit(-1); | |
345 | } | |
346 | POSIX::_exit($res); | |
347 | } else { | |
348 | $count++; | |
349 | $w->{pid} = $pid; | |
350 | } | |
351 | } else { | |
c4a221bc DM |
352 | my $res = -1; |
353 | eval { | |
6dbf93a0 | 354 | $res = $haenv->exec_resource_agent($sid, $cd, $w->{state}, $w->{target}); |
c4a221bc DM |
355 | }; |
356 | if (my $err = $@) { | |
f31b7e94 DM |
357 | $haenv->log('err', $err); |
358 | } | |
359 | $self->resource_command_finished($sid, $w->{uid}, $res); | |
c4a221bc DM |
360 | } |
361 | } | |
362 | } | |
363 | ||
364 | last if !$count; | |
365 | ||
f31b7e94 | 366 | $haenv->sleep(1); |
c4a221bc DM |
367 | } |
368 | } | |
369 | ||
370 | # fixme: use a queue an limit number of parallel workers? | |
371 | sub queue_resource_command { | |
e88469ba | 372 | my ($self, $sid, $uid, $state, $target) = @_; |
c4a221bc DM |
373 | |
374 | if (my $w = $self->{workers}->{$sid}) { | |
375 | return if $w->{pid}; # already started | |
376 | # else, delete and overwrite queue entry with new command | |
377 | delete $self->{workers}->{$sid}; | |
378 | } | |
379 | ||
380 | $self->{workers}->{$sid} = { | |
381 | sid => $sid, | |
382 | uid => $uid, | |
383 | state => $state, | |
384 | }; | |
e88469ba DM |
385 | |
386 | $self->{workers}->{$sid}->{target} = $target if $target; | |
c4a221bc DM |
387 | } |
388 | ||
389 | sub check_active_workers { | |
390 | my ($self) = @_; | |
391 | ||
392 | # finish/count workers | |
393 | my $count = 0; | |
394 | foreach my $sid (keys %{$self->{workers}}) { | |
395 | my $w = $self->{workers}->{$sid}; | |
396 | if (my $pid = $w->{pid}) { | |
397 | # check status | |
398 | my $waitpid = waitpid($pid, WNOHANG); | |
399 | if (defined($waitpid) && ($waitpid == $pid)) { | |
400 | $self->resource_command_finished($sid, $w->{uid}, $?); | |
401 | } else { | |
402 | $count++; | |
403 | } | |
404 | } | |
405 | } | |
406 | ||
407 | return $count; | |
408 | } | |
409 | ||
410 | sub resource_command_finished { | |
411 | my ($self, $sid, $uid, $status) = @_; | |
412 | ||
413 | my $haenv = $self->{haenv}; | |
414 | ||
415 | my $w = delete $self->{workers}->{$sid}; | |
416 | return if !$w; # should not happen | |
417 | ||
418 | my $exit_code = -1; | |
419 | ||
420 | if ($status == -1) { | |
0f70400d | 421 | $haenv->log('err', "resource agent $sid finished - failed to execute"); |
c4a221bc | 422 | } elsif (my $sig = ($status & 127)) { |
0f70400d | 423 | $haenv->log('err', "resource agent $sid finished - got signal $sig"); |
c4a221bc DM |
424 | } else { |
425 | $exit_code = ($status >> 8); | |
c4a221bc DM |
426 | } |
427 | ||
428 | $self->{results}->{$uid} = { | |
429 | sid => $w->{sid}, | |
430 | state => $w->{state}, | |
431 | exit_code => $exit_code, | |
432 | }; | |
433 | ||
434 | my $ss = $self->{service_status}; | |
435 | ||
436 | # compute hash of valid/existing uids | |
437 | my $valid_uids = {}; | |
438 | foreach my $sid (keys %$ss) { | |
439 | my $sd = $ss->{$sid}; | |
440 | next if !$sd->{uid}; | |
441 | $valid_uids->{$sd->{uid}} = 1; | |
442 | } | |
443 | ||
444 | my $results = {}; | |
445 | foreach my $id (keys %{$self->{results}}) { | |
446 | next if !$valid_uids->{$id}; | |
447 | $results->{$id} = $self->{results}->{$id}; | |
448 | } | |
449 | $self->{results} = $results; | |
450 | ||
451 | $haenv->write_lrm_status($results); | |
452 | } | |
453 | ||
5f095798 | 454 | 1; |