Parallelize LUCreateInstance
[ganeti-local] / lib / mcpu.py
index 26512bc..7bb4866 100644 (file)
@@ -34,23 +34,22 @@ from ganeti import constants
 from ganeti import errors
 from ganeti import rpc
 from ganeti import cmdlib
-from ganeti import config
 from ganeti import ssconf
 from ganeti import logger
+from ganeti import locking
+
 
 class Processor(object):
   """Object which runs OpCodes"""
   DISPATCH_TABLE = {
     # Cluster
-    opcodes.OpInitCluster: cmdlib.LUInitCluster,
     opcodes.OpDestroyCluster: cmdlib.LUDestroyCluster,
     opcodes.OpQueryClusterInfo: cmdlib.LUQueryClusterInfo,
-    opcodes.OpClusterCopyFile: cmdlib.LUClusterCopyFile,
-    opcodes.OpRunClusterCommand: cmdlib.LURunClusterCommand,
     opcodes.OpVerifyCluster: cmdlib.LUVerifyCluster,
-    opcodes.OpMasterFailover: cmdlib.LUMasterFailover,
     opcodes.OpDumpClusterConfig: cmdlib.LUDumpClusterConfig,
     opcodes.OpRenameCluster: cmdlib.LURenameCluster,
+    opcodes.OpVerifyDisks: cmdlib.LUVerifyDisks,
+    opcodes.OpSetClusterParams: cmdlib.LUSetClusterParams,
     # node lu
     opcodes.OpAddNode: cmdlib.LUAddNode,
     opcodes.OpQueryNodes: cmdlib.LUQueryNodes,
@@ -66,104 +65,147 @@ class Processor(object):
     opcodes.OpStartupInstance: cmdlib.LUStartupInstance,
     opcodes.OpRebootInstance: cmdlib.LURebootInstance,
     opcodes.OpDeactivateInstanceDisks: cmdlib.LUDeactivateInstanceDisks,
-    opcodes.OpAddMDDRBDComponent: cmdlib.LUAddMDDRBDComponent,
-    opcodes.OpRemoveMDDRBDComponent: cmdlib.LURemoveMDDRBDComponent,
     opcodes.OpReplaceDisks: cmdlib.LUReplaceDisks,
     opcodes.OpFailoverInstance: cmdlib.LUFailoverInstance,
     opcodes.OpConnectConsole: cmdlib.LUConnectConsole,
     opcodes.OpQueryInstances: cmdlib.LUQueryInstances,
     opcodes.OpQueryInstanceData: cmdlib.LUQueryInstanceData,
-    opcodes.OpSetInstanceParms: cmdlib.LUSetInstanceParms,
+    opcodes.OpSetInstanceParams: cmdlib.LUSetInstanceParams,
+    opcodes.OpGrowDisk: cmdlib.LUGrowDisk,
     # os lu
     opcodes.OpDiagnoseOS: cmdlib.LUDiagnoseOS,
     # exports lu
     opcodes.OpQueryExports: cmdlib.LUQueryExports,
     opcodes.OpExportInstance: cmdlib.LUExportInstance,
+    opcodes.OpRemoveExport: cmdlib.LURemoveExport,
     # tags lu
     opcodes.OpGetTags: cmdlib.LUGetTags,
     opcodes.OpSearchTags: cmdlib.LUSearchTags,
     opcodes.OpAddTags: cmdlib.LUAddTags,
     opcodes.OpDelTags: cmdlib.LUDelTags,
+    # test lu
+    opcodes.OpTestDelay: cmdlib.LUTestDelay,
+    opcodes.OpTestAllocator: cmdlib.LUTestAllocator,
     }
 
-  def __init__(self, feedback=None):
+  def __init__(self, context):
     """Constructor for Processor
 
     Args:
      - feedback_fn: the feedback function (taking one string) to be run when
                     interesting events are happening
     """
-    self.cfg = None
-    self.sstore = None
-    self._feedback_fn = feedback
-
-  def ExecOpCode(self, op):
-    """Execute an opcode.
+    self.context = context
+    self._feedback_fn = None
+    self.exclusive_BGL = False
 
-    Args:
-     - cfg: the configuration in which we execute this opcode
-     - opcode: the opcode to be executed
+  def _ExecLU(self, lu):
+    """Logical Unit execution sequence.
 
     """
-    if not isinstance(op, opcodes.OpCode):
-      raise errors.ProgrammerError("Non-opcode instance passed"
-                                   " to ExecOpcode")
-
-    lu_class = self.DISPATCH_TABLE.get(op.__class__, None)
-    if lu_class is None:
-      raise errors.OpCodeUnknown("Unknown opcode")
-
-    if lu_class.REQ_CLUSTER and self.cfg is None:
-      self.cfg = config.ConfigWriter()
-      self.sstore = ssconf.SimpleStore()
-    if self.cfg is not None:
-      write_count = self.cfg.write_count
-    else:
-      write_count = 0
-    lu = lu_class(self, op, self.cfg, self.sstore)
+    write_count = self.context.cfg.write_count
     lu.CheckPrereq()
     hm = HooksMaster(rpc.call_hooks_runner, self, lu)
-    hm.RunPhase(constants.HOOKS_PHASE_PRE)
-    result = lu.Exec(self._feedback_fn)
-    hm.RunPhase(constants.HOOKS_PHASE_POST)
-    if lu.cfg is not None:
-      # we use lu.cfg and not self.cfg as for init cluster, self.cfg
-      # is None but lu.cfg has been recently initialized in the
-      # lu.Exec method
-      if write_count != lu.cfg.write_count:
+    h_results = hm.RunPhase(constants.HOOKS_PHASE_PRE)
+    lu.HooksCallBack(constants.HOOKS_PHASE_PRE, h_results,
+                     self._feedback_fn, None)
+    try:
+      result = lu.Exec(self._feedback_fn)
+      h_results = hm.RunPhase(constants.HOOKS_PHASE_POST)
+      result = lu.HooksCallBack(constants.HOOKS_PHASE_POST, h_results,
+                                self._feedback_fn, result)
+    finally:
+      # FIXME: This needs locks if not lu_class.REQ_BGL
+      if write_count != self.context.cfg.write_count:
         hm.RunConfigUpdate()
 
     return result
 
-  def ChainOpCode(self, op):
-    """Chain and execute an opcode.
+  def _LockAndExecLU(self, lu, level):
+    """Execute a Logical Unit, with the needed locks.
 
-    This is used by LUs when they need to execute a child LU.
+    This is a recursive function that starts locking the given level, and
+    proceeds up, till there are no more locks to acquire. Then it executes the
+    given LU and its opcodes.
+
+    """
+    adding_locks = level in lu.add_locks
+    acquiring_locks = level in lu.needed_locks
+    if level not in locking.LEVELS:
+      result = self._ExecLU(lu)
+    elif adding_locks and acquiring_locks:
+      # We could both acquire and add locks at the same level, but for now we
+      # don't need this, so we'll avoid the complicated code needed.
+      raise NotImplementedError(
+        "Can't declare locks to acquire when adding others")
+    elif adding_locks or acquiring_locks:
+      lu.DeclareLocks(level)
+      share = lu.share_locks[level]
+      if acquiring_locks:
+        needed_locks = lu.needed_locks[level]
+        lu.acquired_locks[level] = self.context.glm.acquire(level,
+                                                            needed_locks,
+                                                            shared=share)
+      else: # adding_locks
+        add_locks = lu.add_locks[level]
+        lu.remove_locks[level] = add_locks
+        try:
+          self.context.glm.add(level, add_locks, acquired=1, shared=share)
+        except errors.LockError:
+          raise errors.OpPrereqError(
+            "Coudn't add locks (%s), probably because of a race condition"
+            " with another job, who added them first" % add_locks)
+      try:
+        try:
+          if adding_locks:
+            lu.acquired_locks[level] = add_locks
+          result = self._LockAndExecLU(lu, level + 1)
+        finally:
+          if level in lu.remove_locks:
+            self.context.glm.remove(level, lu.remove_locks[level])
+      finally:
+        if self.context.glm.is_owned(level):
+          self.context.glm.release(level)
+    else:
+      result = self._LockAndExecLU(lu, level + 1)
+
+    return result
+
+  def ExecOpCode(self, op, feedback_fn):
+    """Execute an opcode.
 
     Args:
-     - opcode: the opcode to be executed
+      op: the opcode to be executed
 
     """
     if not isinstance(op, opcodes.OpCode):
       raise errors.ProgrammerError("Non-opcode instance passed"
                                    " to ExecOpcode")
 
+    self._feedback_fn = feedback_fn
     lu_class = self.DISPATCH_TABLE.get(op.__class__, None)
     if lu_class is None:
       raise errors.OpCodeUnknown("Unknown opcode")
 
-    if lu_class.REQ_CLUSTER and self.cfg is None:
-      self.cfg = config.ConfigWriter()
-      self.sstore = ssconf.SimpleStore()
-    #do_hooks = lu_class.HPATH is not None
-    lu = lu_class(self, op, self.cfg, self.sstore)
-    lu.CheckPrereq()
-    #if do_hooks:
-    #  hm = HooksMaster(rpc.call_hooks_runner, self, lu)
-    #  hm.RunPhase(constants.HOOKS_PHASE_PRE)
-    result = lu.Exec(self._feedback_fn)
-    #if do_hooks:
-    #  hm.RunPhase(constants.HOOKS_PHASE_POST)
+    if lu_class.REQ_WSSTORE:
+      sstore = ssconf.WritableSimpleStore()
+    else:
+      sstore = ssconf.SimpleStore()
+
+    # Acquire the Big Ganeti Lock exclusively if this LU requires it, and in a
+    # shared fashion otherwise (to prevent concurrent run with an exclusive LU.
+    self.context.glm.acquire(locking.LEVEL_CLUSTER, [locking.BGL],
+                             shared=not lu_class.REQ_BGL)
+    try:
+      self.exclusive_BGL = lu_class.REQ_BGL
+      lu = lu_class(self, op, self.context, sstore)
+      lu.ExpandNames()
+      assert lu.needed_locks is not None, "needed_locks not set by LU"
+      result = self._LockAndExecLU(lu, locking.LEVEL_INSTANCE)
+    finally:
+      self.context.glm.release(locking.LEVEL_CLUSTER)
+      self.exclusive_BGL = False
+
     return result
 
   def LogStep(self, current, total, message):
@@ -258,6 +300,12 @@ class HooksMaster(object):
 
     This is the main function of the HookMaster.
 
+    Args:
+      phase: the hooks phase to run
+
+    Returns:
+      the result of the hooks multi-node rpc call
+
     """
     if not self.node_list[phase]:
       # empty node list, we should not attempt to run this as either
@@ -281,6 +329,7 @@ class HooksMaster(object):
             errs.append((node_name, script, output))
       if errs:
         raise errors.HooksAbort(errs)
+    return results
 
   def RunConfigUpdate(self):
     """Run the special configuration update hook