The implementation of our life cycle design has the following characteristics:
- It uses UUIDs as the object identities for nodes to avoid object reincarnation problems.
- When
destroy
is called on a node, the node needs to destroy itself and inform its parent directory that it has been destroyed (because the parent directory is the node's factory and also acts as a collection manager for child nodes).
Note that, in contrast to the initial version, the entire implementation resides in a FilesystemI
package instead of being part of the Filesystem
package. Doing this is not essential, but is a little cleaner because it keeps the implementation in a package that is separate from the Slice-generated package.
On this page:
Object Life Cycle Changes for the NodeI
Class in Java
Our DirectoryI
and FileI
servants derive from a common NodeI
base interface. This interface is not essential, but useful because it allows us to treat servants of type DirectoryI
and FileI
polymorphically:
package FilesystemI; public interface NodeI { com.zeroc.Ice.Identity id(); }
The only method is the id
method, which returns the identity of the corresponding node.
Object Life Cycle Changes for the DirectoryI
Class in Java
As in the initial version, the DirectoryI
class derives from the generated base class _DirectoryDisp
. In addition, the class implements the NodeI
interface. DirectoryI
must implement each of the Slice operations, leading to the following outline:
package FilesystemI; import com.zeroc.Ice.*; import Filesystem.*; public class DirectoryI implements Directory, NodeI { public Identity id(); public synchronized String name(Current c); public synchronized NodeDesc[] list(Current c); public synchronized NodeDesc find(String name, Current c) throws NoSuchName; public synchronized FilePrx createFile(String name, Current c) throws NameInUse; public synchronized DirectoryPrx createDirectory(String name, Current c) throws NameInUse; public void destroy(Current c) throws PermissionDenied; // ... }
To support the implementation, we also require a number of methods and data members:
package FilesystemI; import com.zeroc.Ice.*; import Filesystem.*; public class DirectoryI implements Directory, NodeI { // ... public DirectoryI(); public DirectoryI(String name, DirectoryI parent); public synchronized void removeEntry(String name); private String _name; // Immutable private DirectoryI _parent; // Immutable private Identity _id; // Immutable private boolean _destroyed; private java.util.Map<String, NodeI> _contents; }
The _name
and _parent
members store the name of this node and a reference to the node's parent directory. (The root directory's _parent
member is null.) Similarly, the _id
member stores the identity of this directory. The _name
, _parent
, and _id
members are immutable once they have been initialized by the constructor. The _destroyed
member prevents a race condition; to interlock access to _destroyed
(as well as the _contents
member) we can use synchronized methods (as for the name
method), or use a synchronized(this)
block.
The _contents
map records the contents of a directory: it stores the name of an entry, together with a reference to the child node.
Here are the two constructors for the class:
public DirectoryI() { this("/", null); } public DirectoryI(String name, DirectoryI parent) { _name = name; _parent = parent; _id = new Identity(); _destroyed = false; _contents = new java.util.HashMap<>(); _id.name = parent == null ? "RootDir" : java.util.UUID.randomUUID().toString(); }
The first constructor is a convenience function to create the root directory with the fixed identity "RootDir" and a null parent.
The real constructor initializes the _name
, _parent
, _id
, _destroyed
, and _contents
members. Note that nodes other than the root directory use a UUID as the object identity.
The removeEntry
method is called by the child to remove itself from its parent's _contents
map:
public synchronized void removeEntry(String name) { _contents.remove(name); }
The implementation of the Slice name operation simply returns the name of the node, but also checks whether the node has been destroyed:
public synchronized String name(Current c) { if(_destroyed) { throw new ObjectNotExistException(); } return _name; }
Note that this method is synchronized, so the _destroyed
member cannot be accessed concurrently.
Here is the destroy
member function for directories:
public void destroy(Current c) throws PermissionDenied { if(_parent == null) { throw new PermissionDenied("Cannot destroy root directory"); } synchronized(this) { if(_destroyed) { throw new ObjectNotExistException(); } if(_contents.size() != 0) { throw new PermissionDenied("Cannot destroy non-empty directory"); } c.adapter.remove(id()); _destroyed = true; } _parent.removeEntry(_name); }
The code first prevents destruction of the root directory and then checks whether this directory was destroyed previously. It then acquires the lock and checks that the directory is empty. Finally, destroy
removes the Active Servant Map (ASM) entry for the destroyed directory and removes itself from its parent's _contents
map. Note that we call removeEntry
outside the synchronization to avoid deadlocks.
The createDirectory
implementation acquires the lock before checking whether the directory already contains a node with the given name (or an invalid empty name). If not, it creates a new servant, adds it to the ASM and the _contents
map, and returns its proxy:
public synchronized DirectoryPrx createDirectory(String name, Current c) throws NameInUse { if(_destroyed) { throw new ObjectNotExistException(); } if(name.length() == 0 || _contents.containsKey(name)) { throw new NameInUse(name); } DirectoryI d = new DirectoryI(name, this); ObjectPrx node = c.adapter.add(d, d.id()); _contents.put(name, d); return DirectoryPrx.uncheckedCast(node); }
The createFile
implementation is identical, except that it creates a file instead of a directory:
public synchronized FilePrx createFile(String name, Current c) throws NameInUse { if(_destroyed) { throw new ObjectNotExistException(); } if(name.length() == 0 || _contents.containsKey(name)) { throw new NameInUse(name); } FileI f = new FileI(name, this); ObjectPrx node = c.adapter.add(f, f.id()); _contents.put(name, f); return FilePrx.uncheckedCast(node); }
Here is the implementation of list
:
public synchronized NodeDesc[] list(Current c) { if(_destroyed) { throw new ObjectNotExistException(); } NodeDesc[] ret = new NodeDesc[_contents.size()]; java.util.Iterator<java.util.Map.Entry<String, NodeI> > pos = _contents.entrySet().iterator(); for(int i = 0; i < _contents.size(); ++i) { java.util.Map.Entry<String, NodeI> e = pos.next(); NodeI p = e.getValue(); ret[i] = new NodeDesc(); ret[i].name = e.getKey(); ret[i].type = p instanceof FileI ? NodeType.FileType : NodeType.DirType; ret[i].proxy = NodePrx.uncheckedCast(c.adapter.createProxy(p.id())); } return ret; }
After acquiring the lock, the code iterates over the directory's contents and adds a NodeDesc
structure for each entry to the returned array.
The find
operation proceeds along similar lines:
public synchronized NodeDesc find(String name, Current c) throws NoSuchName { if(_destroyed) { throw new ObjectNotExistException(); } NodeI p = _contents.get(name); if(p == null) { throw new NoSuchName(name); } NodeDesc d = new NodeDesc(); d.name = name; d.type = p instanceof FileI ? NodeType.FileType : NodeType.DirType; d.proxy = NodePrx.uncheckedCast(c.adapter.createProxy(p.id())); return d; }
Object Life Cycle Changes for the FileI
Class in Java
The FileI
class is similar to the DirectoryI
class. The data members store the name, parent, and identity of the file, as well as the _destroyed
flag and the contents of the file (in the _lines
member). The constructor initializes these members:
package FilesystemI; import com.zeroc.Ice.*; import Filesystem.*; import FilesystemI.*; public class FileI implements File, NodeI { // ... public FileI(String name, DirectoryI parent) { _name = name; _parent = parent; _destroyed = false; _id = new Identity(); _id.name = java.util.UUID.randomUUID().toString(); } private String _name; private DirectoryI _parent; private boolean _destroyed; private Identity _id; private String[] _lines; }
The implementation of the remaining member functions of the FileI
class is trivial, so we present all of them here:
public synchronized String name(Current c) { if(_destroyed) { throw new ObjectNotExistException(); } return _name; } public Identity id() { return _id; } public synchronized String[] read(Current c) { if(_destroyed) { throw new ObjectNotExistException(); } return _lines; } public synchronized void write(String[] text, Current c) { if(_destroyed) { throw new ObjectNotExistException(); } _lines = (String[])text.clone(); } public void destroy(Current c) { synchronized(this) { if(_destroyed) { throw new ObjectNotExistException(); } c.adapter.remove(id()); _destroyed = true; } _parent.removeEntry(_name); }
Object Life Cycle Concurrency Issues in Java
The preceding implementation is provably deadlock free. All methods hold only one lock at a time, so they cannot deadlock with each other or themselves. While the locks are held, the methods do not call other methods that acquire locks, so any potential deadlock can only arise by concurrent calls to another mutating method, either on the same node or on different nodes. For concurrent calls on the same node, deadlock is impossible because such calls are strictly serialized on the instance; for concurrent calls to destroy
on different nodes, each node locks itself, releases itself again, and then acquires and releases a lock on its parent (by calling removeEntry
), also making deadlock impossible.