]>
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 | $haenv->sleep(1); | |
97 | } | |
98 | ||
99 | return 0; | |
100 | } | |
101 | ||
546e2f1f DM |
102 | sub active_service_count { |
103 | my ($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 | next if $req_state eq 'stopped'; | |
120 | ||
121 | $count++; | |
122 | } | |
123 | ||
124 | return $count; | |
125 | } | |
067cdf33 | 126 | |
5f095798 DM |
127 | sub do_one_iteration { |
128 | my ($self) = @_; | |
129 | ||
130 | my $haenv = $self->{haenv}; | |
131 | ||
132 | my $status = $self->get_local_status(); | |
133 | my $state = $status->{state}; | |
134 | ||
067cdf33 DM |
135 | my $ms = $haenv->read_manager_status(); |
136 | $self->{service_status} = $ms->{service_status} || {}; | |
137 | ||
49777d09 | 138 | my $fence_request = PVE::HA::Tools::count_fenced_services($self->{service_status}, $haenv->nodename()); |
067cdf33 | 139 | |
5f095798 DM |
140 | # do state changes first |
141 | ||
142 | my $ctime = $haenv->get_time(); | |
143 | ||
b0bf08a9 | 144 | if ($state eq 'wait_for_agent_lock') { |
5f095798 | 145 | |
546e2f1f | 146 | my $service_count = $self->active_service_count(); |
5f095798 | 147 | |
067cdf33 | 148 | if (!$fence_request && $service_count && $haenv->quorate()) { |
0bba8f60 DM |
149 | if ($self->get_protected_ha_agent_lock()) { |
150 | $self->set_local_status({ state => 'active' }); | |
5f095798 DM |
151 | } |
152 | } | |
153 | ||
154 | } elsif ($state eq 'lost_agent_lock') { | |
155 | ||
067cdf33 | 156 | if (!$fence_request && $haenv->quorate()) { |
0bba8f60 DM |
157 | if ($self->get_protected_ha_agent_lock()) { |
158 | $self->set_local_status({ state => 'active' }); | |
5f095798 DM |
159 | } |
160 | } | |
161 | ||
0bba8f60 | 162 | } elsif ($state eq 'active') { |
5f095798 | 163 | |
067cdf33 DM |
164 | if ($fence_request) { |
165 | $haenv->log('err', "node need to be fenced - releasing agent_lock\n"); | |
166 | $self->set_local_status({ state => 'lost_agent_lock'}); | |
167 | } elsif (!$self->get_protected_ha_agent_lock()) { | |
5f095798 DM |
168 | $self->set_local_status({ state => 'lost_agent_lock'}); |
169 | } | |
170 | } | |
171 | ||
172 | $status = $self->get_local_status(); | |
173 | $state = $status->{state}; | |
174 | ||
175 | # do work | |
176 | ||
177 | if ($state eq 'wait_for_agent_lock') { | |
178 | ||
179 | return 0 if $self->{shutdown_request}; | |
180 | ||
181 | $haenv->sleep(5); | |
182 | ||
0bba8f60 | 183 | } elsif ($state eq 'active') { |
5f095798 DM |
184 | |
185 | my $startime = $haenv->get_time(); | |
186 | ||
187 | my $max_time = 10; | |
188 | ||
189 | my $shutdown = 0; | |
190 | ||
191 | # do work (max_time seconds) | |
192 | eval { | |
193 | # fixme: set alert timer | |
194 | ||
195 | if ($self->{shutdown_request}) { | |
196 | ||
197 | # fixme: request service stop or relocate ? | |
198 | ||
546e2f1f | 199 | my $service_count = $self->active_service_count(); |
5f095798 DM |
200 | |
201 | if ($service_count == 0) { | |
202 | ||
203 | if ($self->{ha_agent_wd}) { | |
204 | $haenv->watchdog_close($self->{ha_agent_wd}); | |
205 | delete $self->{ha_agent_wd}; | |
206 | } | |
207 | ||
208 | $shutdown = 1; | |
209 | } | |
c4a221bc | 210 | } else { |
c4a221bc DM |
211 | |
212 | $self->manage_resources(); | |
067cdf33 | 213 | |
5f095798 DM |
214 | } |
215 | }; | |
216 | if (my $err = $@) { | |
217 | $haenv->log('err', "got unexpected error - $err"); | |
218 | } | |
219 | ||
220 | return 0 if $shutdown; | |
221 | ||
222 | $haenv->sleep_until($startime + $max_time); | |
223 | ||
224 | } elsif ($state eq 'lost_agent_lock') { | |
225 | ||
226 | # Note: watchdog is active an will triger soon! | |
227 | ||
228 | # so we hope to get the lock back soon! | |
229 | ||
230 | if ($self->{shutdown_request}) { | |
231 | ||
546e2f1f | 232 | my $service_count = $self->active_service_count(); |
5f095798 | 233 | |
546e2f1f | 234 | if ($service_count > 0) { |
5f095798 | 235 | $haenv->log('err', "get shutdown request in state 'lost_agent_lock' - " . |
546e2f1f | 236 | "detected $service_count running services"); |
5f095798 | 237 | |
546e2f1f | 238 | } else { |
5f095798 | 239 | |
546e2f1f | 240 | # all services are stopped, so we can close the watchdog |
5f095798 | 241 | |
546e2f1f DM |
242 | if ($self->{ha_agent_wd}) { |
243 | $haenv->watchdog_close($self->{ha_agent_wd}); | |
244 | delete $self->{ha_agent_wd}; | |
245 | } | |
246 | ||
247 | return 0; | |
5f095798 | 248 | } |
5f095798 DM |
249 | } |
250 | ||
b0bf08a9 DM |
251 | $haenv->sleep(5); |
252 | ||
5f095798 DM |
253 | } else { |
254 | ||
255 | die "got unexpected status '$state'\n"; | |
256 | ||
257 | } | |
258 | ||
259 | return 1; | |
260 | } | |
261 | ||
c4a221bc DM |
262 | sub manage_resources { |
263 | my ($self) = @_; | |
264 | ||
265 | my $haenv = $self->{haenv}; | |
266 | ||
267 | my $nodename = $haenv->nodename(); | |
268 | ||
c4a221bc DM |
269 | my $ss = $self->{service_status}; |
270 | ||
271 | foreach my $sid (keys %$ss) { | |
272 | my $sd = $ss->{$sid}; | |
273 | next if !$sd->{node}; | |
274 | next if !$sd->{uid}; | |
275 | next if $sd->{node} ne $nodename; | |
276 | my $req_state = $sd->{state}; | |
277 | next if !defined($req_state); | |
c4a221bc | 278 | eval { |
e88469ba | 279 | $self->queue_resource_command($sid, $sd->{uid}, $req_state, $sd->{target}); |
c4a221bc DM |
280 | }; |
281 | if (my $err = $@) { | |
f31b7e94 | 282 | $haenv->log('err', "unable to run resource agent for '$sid' - $err"); # fixme |
c4a221bc DM |
283 | } |
284 | } | |
285 | ||
f31b7e94 | 286 | my $starttime = $haenv->get_time(); |
c4a221bc DM |
287 | |
288 | # start workers | |
289 | my $max_workers = 4; | |
290 | ||
6dbf93a0 | 291 | my $sc = $haenv->read_service_config(); |
f31b7e94 DM |
292 | |
293 | while (($haenv->get_time() - $starttime) < 5) { | |
c4a221bc DM |
294 | my $count = $self->check_active_workers(); |
295 | ||
296 | foreach my $sid (keys %{$self->{workers}}) { | |
297 | last if $count >= $max_workers; | |
298 | my $w = $self->{workers}->{$sid}; | |
6dbf93a0 DM |
299 | my $cd = $sc->{$sid}; |
300 | if (!$cd) { | |
f31b7e94 | 301 | $haenv->log('err', "missing resource configuration for '$sid'"); |
6dbf93a0 DM |
302 | next; |
303 | } | |
c4a221bc | 304 | if (!$w->{pid}) { |
f31b7e94 DM |
305 | if ($haenv->can_fork()) { |
306 | my $pid = fork(); | |
307 | if (!defined($pid)) { | |
308 | $haenv->log('err', "fork worker failed"); | |
309 | $count = 0; last; # abort, try later | |
310 | } elsif ($pid == 0) { | |
311 | # do work | |
312 | my $res = -1; | |
313 | eval { | |
314 | $res = $haenv->exec_resource_agent($sid, $cd, $w->{state}, $w->{target}); | |
315 | }; | |
316 | if (my $err = $@) { | |
317 | $haenv->log('err', $err); | |
318 | POSIX::_exit(-1); | |
319 | } | |
320 | POSIX::_exit($res); | |
321 | } else { | |
322 | $count++; | |
323 | $w->{pid} = $pid; | |
324 | } | |
325 | } else { | |
c4a221bc DM |
326 | my $res = -1; |
327 | eval { | |
6dbf93a0 | 328 | $res = $haenv->exec_resource_agent($sid, $cd, $w->{state}, $w->{target}); |
c4a221bc DM |
329 | }; |
330 | if (my $err = $@) { | |
f31b7e94 DM |
331 | $haenv->log('err', $err); |
332 | } | |
333 | $self->resource_command_finished($sid, $w->{uid}, $res); | |
c4a221bc DM |
334 | } |
335 | } | |
336 | } | |
337 | ||
338 | last if !$count; | |
339 | ||
f31b7e94 | 340 | $haenv->sleep(1); |
c4a221bc DM |
341 | } |
342 | } | |
343 | ||
344 | # fixme: use a queue an limit number of parallel workers? | |
345 | sub queue_resource_command { | |
e88469ba | 346 | my ($self, $sid, $uid, $state, $target) = @_; |
c4a221bc DM |
347 | |
348 | if (my $w = $self->{workers}->{$sid}) { | |
349 | return if $w->{pid}; # already started | |
350 | # else, delete and overwrite queue entry with new command | |
351 | delete $self->{workers}->{$sid}; | |
352 | } | |
353 | ||
354 | $self->{workers}->{$sid} = { | |
355 | sid => $sid, | |
356 | uid => $uid, | |
357 | state => $state, | |
358 | }; | |
e88469ba DM |
359 | |
360 | $self->{workers}->{$sid}->{target} = $target if $target; | |
c4a221bc DM |
361 | } |
362 | ||
363 | sub check_active_workers { | |
364 | my ($self) = @_; | |
365 | ||
366 | # finish/count workers | |
367 | my $count = 0; | |
368 | foreach my $sid (keys %{$self->{workers}}) { | |
369 | my $w = $self->{workers}->{$sid}; | |
370 | if (my $pid = $w->{pid}) { | |
371 | # check status | |
372 | my $waitpid = waitpid($pid, WNOHANG); | |
373 | if (defined($waitpid) && ($waitpid == $pid)) { | |
374 | $self->resource_command_finished($sid, $w->{uid}, $?); | |
375 | } else { | |
376 | $count++; | |
377 | } | |
378 | } | |
379 | } | |
380 | ||
381 | return $count; | |
382 | } | |
383 | ||
384 | sub resource_command_finished { | |
385 | my ($self, $sid, $uid, $status) = @_; | |
386 | ||
387 | my $haenv = $self->{haenv}; | |
388 | ||
389 | my $w = delete $self->{workers}->{$sid}; | |
390 | return if !$w; # should not happen | |
391 | ||
392 | my $exit_code = -1; | |
393 | ||
394 | if ($status == -1) { | |
0f70400d | 395 | $haenv->log('err', "resource agent $sid finished - failed to execute"); |
c4a221bc | 396 | } elsif (my $sig = ($status & 127)) { |
0f70400d | 397 | $haenv->log('err', "resource agent $sid finished - got signal $sig"); |
c4a221bc DM |
398 | } else { |
399 | $exit_code = ($status >> 8); | |
c4a221bc DM |
400 | } |
401 | ||
402 | $self->{results}->{$uid} = { | |
403 | sid => $w->{sid}, | |
404 | state => $w->{state}, | |
405 | exit_code => $exit_code, | |
406 | }; | |
407 | ||
408 | my $ss = $self->{service_status}; | |
409 | ||
410 | # compute hash of valid/existing uids | |
411 | my $valid_uids = {}; | |
412 | foreach my $sid (keys %$ss) { | |
413 | my $sd = $ss->{$sid}; | |
414 | next if !$sd->{uid}; | |
415 | $valid_uids->{$sd->{uid}} = 1; | |
416 | } | |
417 | ||
418 | my $results = {}; | |
419 | foreach my $id (keys %{$self->{results}}) { | |
420 | next if !$valid_uids->{$id}; | |
421 | $results->{$id} = $self->{results}->{$id}; | |
422 | } | |
423 | $self->{results} = $results; | |
424 | ||
425 | $haenv->write_lrm_status($results); | |
426 | } | |
427 | ||
5f095798 | 428 | 1; |