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):
224 if isinstance(result, cmdlib.ResultWithJobs):
226 job_submission = self._cbs.SubmitManyJobs(result.jobs)
229 result = result.other
231 assert constants.JOB_IDS_KEY not in result, \
232 "Key '%s' found in additional return values" % constants.JOB_IDS_KEY
234 result[constants.JOB_IDS_KEY] = job_submission
238 def _ExecLU(self, lu):
239 """Logical Unit execution sequence.
242 write_count = self.context.cfg.write_count
244 hm = HooksMaster(self.rpc.call_hooks_runner, lu)
245 h_results = hm.RunPhase(constants.HOOKS_PHASE_PRE)
246 lu.HooksCallBack(constants.HOOKS_PHASE_PRE, h_results,
249 if getattr(lu.op, "dry_run", False):
250 # in this mode, no post-hooks are run, and the config is not
251 # written (as it might have been modified by another LU, and we
252 # shouldn't do writeout on behalf of other threads
253 self.LogInfo("dry-run mode requested, not actually executing"
255 return lu.dry_run_result
258 result = self._ProcessResult(lu.Exec(self.Log))
259 h_results = hm.RunPhase(constants.HOOKS_PHASE_POST)
260 result = lu.HooksCallBack(constants.HOOKS_PHASE_POST, h_results,
263 # FIXME: This needs locks if not lu_class.REQ_BGL
264 if write_count != self.context.cfg.write_count:
269 def _LockAndExecLU(self, lu, level, calc_timeout, priority):
270 """Execute a Logical Unit, with the needed locks.
272 This is a recursive function that starts locking the given level, and
273 proceeds up, till there are no more locks to acquire. Then it executes the
274 given LU and its opcodes.
277 adding_locks = level in lu.add_locks
278 acquiring_locks = level in lu.needed_locks
279 if level not in locking.LEVELS:
281 self._cbs.NotifyStart()
283 result = self._ExecLU(lu)
285 elif adding_locks and acquiring_locks:
286 # We could both acquire and add locks at the same level, but for now we
287 # don't need this, so we'll avoid the complicated code needed.
288 raise NotImplementedError("Can't declare locks to acquire when adding"
291 elif adding_locks or acquiring_locks:
292 lu.DeclareLocks(level)
293 share = lu.share_locks[level]
296 assert adding_locks ^ acquiring_locks, \
297 "Locks must be either added or acquired"
301 needed_locks = lu.needed_locks[level]
303 acquired = self._AcquireLocks(level, needed_locks, share,
304 calc_timeout(), priority)
307 add_locks = lu.add_locks[level]
308 lu.remove_locks[level] = add_locks
311 self.context.glm.add(level, add_locks, acquired=1, shared=share)
312 except errors.LockError:
313 raise errors.OpPrereqError(
314 "Couldn't add locks (%s), probably because of a race condition"
315 " with another job, who added them first" % add_locks,
321 lu.acquired_locks[level] = acquired
323 result = self._LockAndExecLU(lu, level + 1, calc_timeout, priority)
325 if level in lu.remove_locks:
326 self.context.glm.remove(level, lu.remove_locks[level])
328 if self.context.glm.is_owned(level):
329 self.context.glm.release(level)
332 result = self._LockAndExecLU(lu, level + 1, calc_timeout, priority)
336 def ExecOpCode(self, op, cbs, timeout=None, priority=None):
337 """Execute an opcode.
339 @type op: an OpCode instance
340 @param op: the opcode to be executed
341 @type cbs: L{OpExecCbBase}
342 @param cbs: Runtime callbacks
343 @type timeout: float or None
344 @param timeout: Maximum time to acquire all locks, None for no timeout
345 @type priority: number or None
346 @param priority: Priority for acquiring lock(s)
347 @raise LockAcquireTimeout: In case locks couldn't be acquired in specified
351 if not isinstance(op, opcodes.OpCode):
352 raise errors.ProgrammerError("Non-opcode instance passed"
355 lu_class = self.DISPATCH_TABLE.get(op.__class__, None)
357 raise errors.OpCodeUnknown("Unknown opcode")
360 calc_timeout = lambda: None
362 calc_timeout = utils.RunningTimeout(timeout, False).Remaining
366 # Acquire the Big Ganeti Lock exclusively if this LU requires it,
367 # and in a shared fashion otherwise (to prevent concurrent run with
369 self._AcquireLocks(locking.LEVEL_CLUSTER, locking.BGL,
370 not lu_class.REQ_BGL, calc_timeout(),
373 lu = lu_class(self, op, self.context, self.rpc)
375 assert lu.needed_locks is not None, "needed_locks not set by LU"
378 return self._LockAndExecLU(lu, locking.LEVEL_INSTANCE, calc_timeout,
382 self.context.cfg.DropECReservations(self._ec_id)
384 self.context.glm.release(locking.LEVEL_CLUSTER)
388 def Log(self, *args):
389 """Forward call to feedback callback function.
393 self._cbs.Feedback(*args)
395 def LogStep(self, current, total, message):
396 """Log a change in LU execution progress.
399 logging.debug("Step %d/%d %s", current, total, message)
400 self.Log("STEP %d/%d %s" % (current, total, message))
402 def LogWarning(self, message, *args, **kwargs):
403 """Log a warning to the logs and the user.
405 The optional keyword argument is 'hint' and can be used to show a
406 hint to the user (presumably related to the warning). If the
407 message is empty, it will not be printed at all, allowing one to
411 assert not kwargs or (len(kwargs) == 1 and "hint" in kwargs), \
412 "Invalid keyword arguments for LogWarning (%s)" % str(kwargs)
414 message = message % tuple(args)
416 logging.warning(message)
417 self.Log(" - WARNING: %s" % message)
419 self.Log(" Hint: %s" % kwargs["hint"])
421 def LogInfo(self, message, *args):
422 """Log an informational message to the logs and the user.
426 message = message % tuple(args)
427 logging.info(message)
428 self.Log(" - INFO: %s" % message)
431 """Returns the current execution context ID.
435 raise errors.ProgrammerError("Tried to use execution context id when"
440 class HooksMaster(object):
443 This class distributes the run commands to the nodes based on the
446 In order to remove the direct dependency on the rpc module, the
447 constructor needs a function which actually does the remote
448 call. This will usually be rpc.call_hooks_runner, but any function
449 which behaves the same works.
452 def __init__(self, callfn, lu):
456 self.pre_env = self._BuildEnv(constants.HOOKS_PHASE_PRE)
458 if self.lu.HPATH is None:
461 nodes = map(frozenset, self.lu.BuildHooksNodes())
463 (self.pre_nodes, self.post_nodes) = nodes
465 def _BuildEnv(self, phase):
466 """Compute the environment and the target nodes.
468 Based on the opcode and the current node list, this builds the
469 environment for the hooks and the target node list for the run.
472 if phase == constants.HOOKS_PHASE_PRE:
474 elif phase == constants.HOOKS_PHASE_POST:
475 prefix = "GANETI_POST_"
477 raise AssertionError("Unknown phase '%s'" % phase)
481 if self.lu.HPATH is not None:
482 lu_env = self.lu.BuildHooksEnv()
484 assert not compat.any(key.upper().startswith(prefix) for key in lu_env)
485 env.update(("%s%s" % (prefix, key), value)
486 for (key, value) in lu_env.items())
488 if phase == constants.HOOKS_PHASE_PRE:
489 assert compat.all((key.startswith("GANETI_") and
490 not key.startswith("GANETI_POST_"))
493 elif phase == constants.HOOKS_PHASE_POST:
494 assert compat.all(key.startswith("GANETI_POST_") for key in env)
495 assert isinstance(self.pre_env, dict)
497 # Merge with pre-phase environment
498 assert not compat.any(key.startswith("GANETI_POST_")
499 for key in self.pre_env)
500 env.update(self.pre_env)
502 raise AssertionError("Unknown phase '%s'" % phase)
506 def _RunWrapper(self, node_list, hpath, phase, phase_env):
507 """Simple wrapper over self.callfn.
509 This method fixes the environment before doing the rpc call.
515 "PATH": "/sbin:/bin:/usr/sbin:/usr/bin",
516 "GANETI_HOOKS_VERSION": constants.HOOKS_VERSION,
517 "GANETI_OP_CODE": self.op.OP_ID,
518 "GANETI_DATA_DIR": constants.DATA_DIR,
519 "GANETI_HOOKS_PHASE": phase,
520 "GANETI_HOOKS_PATH": hpath,
524 env["GANETI_OBJECT_TYPE"] = self.lu.HTYPE
527 env["GANETI_CLUSTER"] = cfg.GetClusterName()
528 env["GANETI_MASTER"] = cfg.GetMasterNode()
531 assert not (set(env) & set(phase_env)), "Environment variables conflict"
532 env.update(phase_env)
534 # Convert everything to strings
535 env = dict([(str(key), str(val)) for key, val in env.iteritems()])
537 assert compat.all(key == "PATH" or key.startswith("GANETI_")
540 return self.callfn(node_list, hpath, phase, env)
542 def RunPhase(self, phase, nodes=None):
543 """Run all the scripts for a phase.
545 This is the main function of the HookMaster.
547 @param phase: one of L{constants.HOOKS_PHASE_POST} or
548 L{constants.HOOKS_PHASE_PRE}; it denotes the hooks phase
549 @param nodes: overrides the predefined list of nodes for the given phase
550 @return: the processed results of the hooks multi-node rpc call
551 @raise errors.HooksFailure: on communication failure to the nodes
552 @raise errors.HooksAbort: on failure of one of the hooks
555 if phase == constants.HOOKS_PHASE_PRE:
557 nodes = self.pre_nodes
559 elif phase == constants.HOOKS_PHASE_POST:
561 nodes = self.post_nodes
562 env = self._BuildEnv(phase)
564 raise AssertionError("Unknown phase '%s'" % phase)
567 # empty node list, we should not attempt to run this as either
568 # we're in the cluster init phase and the rpc client part can't
569 # even attempt to run, or this LU doesn't do hooks at all
572 results = self._RunWrapper(nodes, self.lu.HPATH, phase, env)
574 msg = "Communication Failure"
575 if phase == constants.HOOKS_PHASE_PRE:
576 raise errors.HooksFailure(msg)
578 self.lu.LogWarning(msg)
582 for node_name in results:
583 res = results[node_name]
589 self.lu.LogWarning("Communication failure to node %s: %s",
593 for script, hkr, output in res.payload:
594 if hkr == constants.HKR_FAIL:
595 if phase == constants.HOOKS_PHASE_PRE:
596 errs.append((node_name, script, output))
599 output = "(no output)"
600 self.lu.LogWarning("On %s script %s failed, output: %s" %
601 (node_name, script, output))
603 if errs and phase == constants.HOOKS_PHASE_PRE:
604 raise errors.HooksAbort(errs)
608 def RunConfigUpdate(self):
609 """Run the special configuration update hook
611 This is a special hook that runs only on the master after each
612 top-level LI if the configuration has been updated.
615 phase = constants.HOOKS_PHASE_POST
616 hpath = constants.HOOKS_NAME_CFGUPDATE
617 nodes = [self.lu.cfg.GetMasterNode()]
618 self._RunWrapper(nodes, hpath, phase, self.pre_env)