4 # Copyright (C) 2006, 2007, 2011 Google Inc.
6 # This program is free software; you can redistribute it and/or modify
7 # it under the terms of the GNU General Public License as published by
8 # the Free Software Foundation; either version 2 of the License, or
9 # (at your option) any later version.
11 # This program is distributed in the hope that it will be useful, but
12 # WITHOUT ANY WARRANTY; without even the implied warranty of
13 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 # General Public License for more details.
16 # You should have received a copy of the GNU General Public License
17 # along with this program; if not, write to the Free Software
18 # Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
22 """Module implementing the logic behind the cluster operations
24 This module implements the logic for doing operations in the cluster. There
25 are two kinds of classes defined:
26 - logical units, which know how to deal with their specific opcode only
27 - the processor, which dispatches the opcodes to their logical units
35 from ganeti import opcodes
36 from ganeti import constants
37 from ganeti import errors
38 from ganeti import rpc
39 from ganeti import cmdlib
40 from ganeti import locking
41 from ganeti import utils
42 from ganeti import compat
49 class LockAcquireTimeout(Exception):
50 """Exception to report timeouts on acquiring locks.
55 def _CalculateLockAttemptTimeouts():
56 """Calculate timeouts for lock attempts.
59 result = [constants.LOCK_ATTEMPTS_MINWAIT]
60 running_sum = result[0]
62 # Wait for a total of at least LOCK_ATTEMPTS_TIMEOUT before doing a
64 while running_sum < constants.LOCK_ATTEMPTS_TIMEOUT:
65 timeout = (result[-1] * 1.05) ** 1.25
67 # Cap max timeout. This gives other jobs a chance to run even if
68 # we're still trying to get our locks, before finally moving to a
70 timeout = min(timeout, constants.LOCK_ATTEMPTS_MAXWAIT)
71 # And also cap the lower boundary for safety
72 timeout = max(timeout, constants.LOCK_ATTEMPTS_MINWAIT)
74 result.append(timeout)
75 running_sum += timeout
80 class LockAttemptTimeoutStrategy(object):
81 """Class with lock acquire timeout strategy.
90 _TIMEOUT_PER_ATTEMPT = _CalculateLockAttemptTimeouts()
92 def __init__(self, _time_fn=time.time, _random_fn=random.random):
93 """Initializes this class.
95 @param _time_fn: Time function for unittests
96 @param _random_fn: Random number generator for unittests
101 self._timeouts = iter(self._TIMEOUT_PER_ATTEMPT)
102 self._time_fn = _time_fn
103 self._random_fn = _random_fn
105 def NextAttempt(self):
106 """Returns the timeout for the next attempt.
110 timeout = self._timeouts.next()
111 except StopIteration:
112 # No more timeouts, do blocking acquire
115 if timeout is not None:
116 # Add a small variation (-/+ 5%) to timeout. This helps in situations
117 # where two or more jobs are fighting for the same lock(s).
118 variation_range = timeout * 0.1
119 timeout += ((self._random_fn() * variation_range) -
120 (variation_range * 0.5))
125 class OpExecCbBase: # pylint: disable-msg=W0232
126 """Base class for OpCode execution callbacks.
129 def NotifyStart(self):
130 """Called when we are about to execute the LU.
132 This function is called when we're about to start the lu's Exec() method,
133 that is, after we have acquired all locks.
137 def Feedback(self, *args):
138 """Sends feedback from the LU code to the end-user.
142 def CheckCancel(self):
143 """Check whether job has been cancelled.
147 def SubmitManyJobs(self, jobs):
148 """Submits jobs for processing.
150 See L{jqueue.JobQueue.SubmitManyJobs}.
153 raise NotImplementedError
156 def _LUNameForOpName(opname):
157 """Computes the LU name for a given OpCode name.
160 assert opname.startswith(_OP_PREFIX), \
161 "Invalid OpCode name, doesn't start with %s: %s" % (_OP_PREFIX, opname)
163 return _LU_PREFIX + opname[len(_OP_PREFIX):]
166 def _ComputeDispatchTable():
167 """Computes the opcode-to-lu dispatch table.
170 return dict((op, getattr(cmdlib, _LUNameForOpName(op.__name__)))
171 for op in opcodes.OP_MAPPING.values()
175 class Processor(object):
176 """Object which runs OpCodes"""
177 DISPATCH_TABLE = _ComputeDispatchTable()
179 def __init__(self, context, ec_id):
180 """Constructor for Processor
182 @type context: GanetiContext
183 @param context: global Ganeti context
185 @param ec_id: execution context identifier
188 self.context = context
191 self.rpc = rpc.RpcRunner(context.cfg)
192 self.hmclass = HooksMaster
194 def _AcquireLocks(self, level, names, shared, timeout, priority):
195 """Acquires locks via the Ganeti lock manager.
198 @param level: Lock level
199 @type names: list or string
200 @param names: Lock names
202 @param shared: Whether the locks should be acquired in shared mode
203 @type timeout: None or float
204 @param timeout: Timeout for acquiring the locks
205 @raise LockAcquireTimeout: In case locks couldn't be acquired in specified
210 self._cbs.CheckCancel()
212 acquired = self.context.glm.acquire(level, names, shared=shared,
213 timeout=timeout, priority=priority)
216 raise LockAcquireTimeout()
220 def _ProcessResult(self, result):
221 """Examines opcode result.
223 If necessary, additional processing on the result is done.
226 if isinstance(result, cmdlib.ResultWithJobs):
228 job_submission = self._cbs.SubmitManyJobs(result.jobs)
231 result = result.other
233 assert constants.JOB_IDS_KEY not in result, \
234 "Key '%s' found in additional return values" % constants.JOB_IDS_KEY
236 result[constants.JOB_IDS_KEY] = job_submission
240 def _ExecLU(self, lu):
241 """Logical Unit execution sequence.
244 write_count = self.context.cfg.write_count
246 hm = HooksMaster(self.rpc.call_hooks_runner, lu)
247 h_results = hm.RunPhase(constants.HOOKS_PHASE_PRE)
248 lu.HooksCallBack(constants.HOOKS_PHASE_PRE, h_results,
251 if getattr(lu.op, "dry_run", False):
252 # in this mode, no post-hooks are run, and the config is not
253 # written (as it might have been modified by another LU, and we
254 # shouldn't do writeout on behalf of other threads
255 self.LogInfo("dry-run mode requested, not actually executing"
257 return lu.dry_run_result
260 result = self._ProcessResult(lu.Exec(self.Log))
261 h_results = hm.RunPhase(constants.HOOKS_PHASE_POST)
262 result = lu.HooksCallBack(constants.HOOKS_PHASE_POST, h_results,
265 # FIXME: This needs locks if not lu_class.REQ_BGL
266 if write_count != self.context.cfg.write_count:
271 def _LockAndExecLU(self, lu, level, calc_timeout, priority):
272 """Execute a Logical Unit, with the needed locks.
274 This is a recursive function that starts locking the given level, and
275 proceeds up, till there are no more locks to acquire. Then it executes the
276 given LU and its opcodes.
279 adding_locks = level in lu.add_locks
280 acquiring_locks = level in lu.needed_locks
281 if level not in locking.LEVELS:
283 self._cbs.NotifyStart()
285 result = self._ExecLU(lu)
287 elif adding_locks and acquiring_locks:
288 # We could both acquire and add locks at the same level, but for now we
289 # don't need this, so we'll avoid the complicated code needed.
290 raise NotImplementedError("Can't declare locks to acquire when adding"
293 elif adding_locks or acquiring_locks:
294 lu.DeclareLocks(level)
295 share = lu.share_locks[level]
298 assert adding_locks ^ acquiring_locks, \
299 "Locks must be either added or acquired"
303 needed_locks = lu.needed_locks[level]
305 self._AcquireLocks(level, needed_locks, share,
306 calc_timeout(), priority)
309 add_locks = lu.add_locks[level]
310 lu.remove_locks[level] = add_locks
313 self.context.glm.add(level, add_locks, acquired=1, shared=share)
314 except errors.LockError:
315 raise errors.OpPrereqError(
316 "Couldn't add locks (%s), probably because of a race condition"
317 " with another job, who added them first" % add_locks,
321 result = self._LockAndExecLU(lu, level + 1, calc_timeout, priority)
323 if level in lu.remove_locks:
324 self.context.glm.remove(level, lu.remove_locks[level])
326 if self.context.glm.is_owned(level):
327 self.context.glm.release(level)
330 result = self._LockAndExecLU(lu, level + 1, calc_timeout, priority)
334 def ExecOpCode(self, op, cbs, timeout=None, priority=None):
335 """Execute an opcode.
337 @type op: an OpCode instance
338 @param op: the opcode to be executed
339 @type cbs: L{OpExecCbBase}
340 @param cbs: Runtime callbacks
341 @type timeout: float or None
342 @param timeout: Maximum time to acquire all locks, None for no timeout
343 @type priority: number or None
344 @param priority: Priority for acquiring lock(s)
345 @raise LockAcquireTimeout: In case locks couldn't be acquired in specified
349 if not isinstance(op, opcodes.OpCode):
350 raise errors.ProgrammerError("Non-opcode instance passed"
353 lu_class = self.DISPATCH_TABLE.get(op.__class__, None)
355 raise errors.OpCodeUnknown("Unknown opcode")
358 calc_timeout = lambda: None
360 calc_timeout = utils.RunningTimeout(timeout, False).Remaining
364 # Acquire the Big Ganeti Lock exclusively if this LU requires it,
365 # and in a shared fashion otherwise (to prevent concurrent run with
367 self._AcquireLocks(locking.LEVEL_CLUSTER, locking.BGL,
368 not lu_class.REQ_BGL, calc_timeout(),
371 lu = lu_class(self, op, self.context, self.rpc)
373 assert lu.needed_locks is not None, "needed_locks not set by LU"
376 return self._LockAndExecLU(lu, locking.LEVEL_INSTANCE, calc_timeout,
380 self.context.cfg.DropECReservations(self._ec_id)
382 self.context.glm.release(locking.LEVEL_CLUSTER)
386 def Log(self, *args):
387 """Forward call to feedback callback function.
391 self._cbs.Feedback(*args)
393 def LogStep(self, current, total, message):
394 """Log a change in LU execution progress.
397 logging.debug("Step %d/%d %s", current, total, message)
398 self.Log("STEP %d/%d %s" % (current, total, message))
400 def LogWarning(self, message, *args, **kwargs):
401 """Log a warning to the logs and the user.
403 The optional keyword argument is 'hint' and can be used to show a
404 hint to the user (presumably related to the warning). If the
405 message is empty, it will not be printed at all, allowing one to
409 assert not kwargs or (len(kwargs) == 1 and "hint" in kwargs), \
410 "Invalid keyword arguments for LogWarning (%s)" % str(kwargs)
412 message = message % tuple(args)
414 logging.warning(message)
415 self.Log(" - WARNING: %s" % message)
417 self.Log(" Hint: %s" % kwargs["hint"])
419 def LogInfo(self, message, *args):
420 """Log an informational message to the logs and the user.
424 message = message % tuple(args)
425 logging.info(message)
426 self.Log(" - INFO: %s" % message)
429 """Returns the current execution context ID.
433 raise errors.ProgrammerError("Tried to use execution context id when"
438 class HooksMaster(object):
441 This class distributes the run commands to the nodes based on the
444 In order to remove the direct dependency on the rpc module, the
445 constructor needs a function which actually does the remote
446 call. This will usually be rpc.call_hooks_runner, but any function
447 which behaves the same works.
450 def __init__(self, callfn, lu):
454 self.pre_env = self._BuildEnv(constants.HOOKS_PHASE_PRE)
456 if self.lu.HPATH is None:
459 nodes = map(frozenset, self.lu.BuildHooksNodes())
461 (self.pre_nodes, self.post_nodes) = nodes
463 def _BuildEnv(self, phase):
464 """Compute the environment and the target nodes.
466 Based on the opcode and the current node list, this builds the
467 environment for the hooks and the target node list for the run.
470 if phase == constants.HOOKS_PHASE_PRE:
472 elif phase == constants.HOOKS_PHASE_POST:
473 prefix = "GANETI_POST_"
475 raise AssertionError("Unknown phase '%s'" % phase)
479 if self.lu.HPATH is not None:
480 lu_env = self.lu.BuildHooksEnv()
482 assert not compat.any(key.upper().startswith(prefix) for key in lu_env)
483 env.update(("%s%s" % (prefix, key), value)
484 for (key, value) in lu_env.items())
486 if phase == constants.HOOKS_PHASE_PRE:
487 assert compat.all((key.startswith("GANETI_") and
488 not key.startswith("GANETI_POST_"))
491 elif phase == constants.HOOKS_PHASE_POST:
492 assert compat.all(key.startswith("GANETI_POST_") for key in env)
493 assert isinstance(self.pre_env, dict)
495 # Merge with pre-phase environment
496 assert not compat.any(key.startswith("GANETI_POST_")
497 for key in self.pre_env)
498 env.update(self.pre_env)
500 raise AssertionError("Unknown phase '%s'" % phase)
504 def _RunWrapper(self, node_list, hpath, phase, phase_env):
505 """Simple wrapper over self.callfn.
507 This method fixes the environment before doing the rpc call.
513 "PATH": "/sbin:/bin:/usr/sbin:/usr/bin",
514 "GANETI_HOOKS_VERSION": constants.HOOKS_VERSION,
515 "GANETI_OP_CODE": self.op.OP_ID,
516 "GANETI_DATA_DIR": constants.DATA_DIR,
517 "GANETI_HOOKS_PHASE": phase,
518 "GANETI_HOOKS_PATH": hpath,
522 env["GANETI_OBJECT_TYPE"] = self.lu.HTYPE
525 env["GANETI_CLUSTER"] = cfg.GetClusterName()
526 env["GANETI_MASTER"] = cfg.GetMasterNode()
529 assert not (set(env) & set(phase_env)), "Environment variables conflict"
530 env.update(phase_env)
532 # Convert everything to strings
533 env = dict([(str(key), str(val)) for key, val in env.iteritems()])
535 assert compat.all(key == "PATH" or key.startswith("GANETI_")
538 return self.callfn(node_list, hpath, phase, env)
540 def RunPhase(self, phase, nodes=None):
541 """Run all the scripts for a phase.
543 This is the main function of the HookMaster.
545 @param phase: one of L{constants.HOOKS_PHASE_POST} or
546 L{constants.HOOKS_PHASE_PRE}; it denotes the hooks phase
547 @param nodes: overrides the predefined list of nodes for the given phase
548 @return: the processed results of the hooks multi-node rpc call
549 @raise errors.HooksFailure: on communication failure to the nodes
550 @raise errors.HooksAbort: on failure of one of the hooks
553 if phase == constants.HOOKS_PHASE_PRE:
555 nodes = self.pre_nodes
557 elif phase == constants.HOOKS_PHASE_POST:
559 nodes = self.post_nodes
560 env = self._BuildEnv(phase)
562 raise AssertionError("Unknown phase '%s'" % phase)
565 # empty node list, we should not attempt to run this as either
566 # we're in the cluster init phase and the rpc client part can't
567 # even attempt to run, or this LU doesn't do hooks at all
570 results = self._RunWrapper(nodes, self.lu.HPATH, phase, env)
572 msg = "Communication Failure"
573 if phase == constants.HOOKS_PHASE_PRE:
574 raise errors.HooksFailure(msg)
576 self.lu.LogWarning(msg)
580 for node_name in results:
581 res = results[node_name]
587 self.lu.LogWarning("Communication failure to node %s: %s",
591 for script, hkr, output in res.payload:
592 if hkr == constants.HKR_FAIL:
593 if phase == constants.HOOKS_PHASE_PRE:
594 errs.append((node_name, script, output))
597 output = "(no output)"
598 self.lu.LogWarning("On %s script %s failed, output: %s" %
599 (node_name, script, output))
601 if errs and phase == constants.HOOKS_PHASE_PRE:
602 raise errors.HooksAbort(errs)
606 def RunConfigUpdate(self):
607 """Run the special configuration update hook
609 This is a special hook that runs only on the master after each
610 top-level LI if the configuration has been updated.
613 phase = constants.HOOKS_PHASE_POST
614 hpath = constants.HOOKS_NAME_CFGUPDATE
615 nodes = [self.lu.cfg.GetMasterNode()]
616 self._RunWrapper(nodes, hpath, phase, self.pre_env)