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