org.apache.xml.dtm.ref
Class CoroutineManager

java.lang.Object
  extended by org.apache.xml.dtm.ref.CoroutineManager

public class CoroutineManager
extends java.lang.Object

Support the coroutine design pattern.

A coroutine set is a very simple cooperative non-preemptive multitasking model, where the switch from one task to another is performed via an explicit request. Coroutines interact according to the following rules:

Coroutines can be thought of as falling somewhere between pipes and subroutines. Like call/return, there is an explicit flow of control from one coroutine to another. Like pipes, neither coroutine is actually "in charge", and neither must exit in order to transfer control to the other.

One classic application of coroutines is in compilers, where both the parser and the lexer are maintaining complex state information. The parser resumes the lexer to process incoming characters into lexical tokens, and the lexer resumes the parser when it has reached a point at which it has a reliably interpreted set of tokens available for semantic processing. Structuring this as call-and-return would require saving and restoring a considerable amount of state each time. Structuring it as two tasks connected by a queue may involve higher overhead (in systems which can optimize the coroutine metaphor), isn't necessarily as clear in intent, may have trouble handling cases where data flows in both directions, and may not handle some of the more complex cases where more than two coroutines are involved.

Most coroutine systems also provide a way to pass data between the source and target of a resume operation; this is sometimes referred to as "yielding" a value. Others rely on the fact that, since only one member of a coroutine set is running at a time and does not lose control until it chooses to do so, data structures may be directly shared between them with only minimal precautions.

"Note: This should not be taken to mean that producer/consumer problems should be always be done with coroutines." Queueing is often a better solution when only two threads of execution are involved and full two-way handshaking is not required. It's a bit difficult to find short pedagogical examples that require coroutines for a clear solution.

The fact that only one of a group of coroutines is running at a time, and the control transfer between them is explicit, simplifies their possible interactions, and in some implementations permits them to be implemented more efficiently than general multitasking. In some situations, coroutines can be compiled out entirely; in others, they may only require a few instructions more than a simple function call.

This version is built on top of standard Java threading, since that's all we have available right now. It's been encapsulated for code clarity and possible future optimization.

(Two possible approaches: wait-notify based and queue-based. Some folks think that a one-item queue is a cleaner solution because it's more abstract -- but since coroutine _is_ an abstraction I'm not really worried about that; folks should be able to switch this code without concern.)

%TBD% THIS SHOULD BE AN INTERFACE, to facilitate building other implementations... perhaps including a true coroutine system someday, rather than controlled threading. Arguably Coroutine itself should be an interface much like Runnable, but I think that can be built on top of this.


Field Summary
(package private) static int ANYBODY
           
(package private)  java.util.BitSet m_activeIDs
          "Is this coroutine ID number already in use" lookup table.
(package private)  int m_nextCoroutine
          Internal field used to confirm that the coroutine now waking up is in fact the one we intended to resume.
(package private) static int m_unreasonableId
          Limit on the coroutine ID numbers accepted.
(package private)  java.lang.Object m_yield
          Internal field used to hold the data being explicitly passed from one coroutine to another during a co_resume() operation.
(package private) static int NOBODY
           
 
Constructor Summary
CoroutineManager()
           
 
Method Summary
 java.lang.Object co_entry_pause(int thisCoroutine)
          In the standard coroutine architecture, coroutines are identified by their method names and are launched and run up to their first yield by simply resuming them; its's presumed that this recognizes the not-already-running case and does the right thing.
 void co_exit_to(java.lang.Object arg_object, int thisCoroutine, int toCoroutine)
          Make the ID available for reuse and terminate this coroutine, transferring control to the specified coroutine.
 void co_exit(int thisCoroutine)
          Terminate this entire set of coroutines.
 int co_joinCoroutineSet(int coroutineID)
          Each coroutine in the set managed by a single CoroutineManager is identified by a small positive integer.
 java.lang.Object co_resume(java.lang.Object arg_object, int thisCoroutine, int toCoroutine)
          Transfer control to another coroutine which has already been started and is waiting on this CoroutineManager.
 
Methods inherited from class java.lang.Object
clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait
 

Field Detail

m_activeIDs

java.util.BitSet m_activeIDs
"Is this coroutine ID number already in use" lookup table. Currently implemented as a bitset as a compromise between compactness and speed of access, but obviously other solutions could be applied.


m_unreasonableId

static final int m_unreasonableId
Limit on the coroutine ID numbers accepted. I didn't want the in-use table to grow without bound. If we switch to a more efficient sparse-array mechanism, it may be possible to raise or eliminate this boundary.

See Also:
Constant Field Values

m_yield

java.lang.Object m_yield
Internal field used to hold the data being explicitly passed from one coroutine to another during a co_resume() operation. (Of course implicit data sharing may also occur; one of the reasons for using coroutines is that you're guaranteed that none of the other coroutines in your set are using shared structures at the time you access them.) %REVIEW% It's been proposed that we be able to pass types of data other than Object -- more specific object types, or lighter-weight primitives. This would seem to create a potential explosion of "pass x recieve y back" methods (or require fracturing resume into two calls, resume-other and wait-to-be-resumed), and the weight issue could be managed by reusing a mutable buffer object to contain the primitive (remember that only one coroutine runs at a time, so once the buffer's set it won't be walked on). Typechecking objects is interesting from a code-robustness point of view, but it's unclear whether it makes sense to encapsulate that in the coroutine code or let the callers do it, since it depends on RTTI either way. Restricting the parameters to objects implementing a specific CoroutineParameter interface does _not_ seem to be a net win; applications can do so if they want via front-end code, but there seem to be too many use cases involving passing an existing object type that you may not have the freedom to alter and may not want to spend time wrapping another object around.


NOBODY

static final int NOBODY
See Also:
Constant Field Values

ANYBODY

static final int ANYBODY
See Also:
Constant Field Values

m_nextCoroutine

int m_nextCoroutine
Internal field used to confirm that the coroutine now waking up is in fact the one we intended to resume. Some such selection mechanism is needed when more that two coroutines are operating within the same group.

Constructor Detail

CoroutineManager

public CoroutineManager()
Method Detail

co_joinCoroutineSet

public int co_joinCoroutineSet(int coroutineID)

Each coroutine in the set managed by a single CoroutineManager is identified by a small positive integer. This brings up the question of how to manage those integers to avoid reuse... since if two coroutines use the same ID number, resuming that ID could resume either. I can see arguments for either allowing applications to select their own numbers (they may want to declare mnemonics via manefest constants) or generating numbers on demand. This routine's intended to support both approaches.

%REVIEW% We could use an object as the identifier. Not sure it's a net gain, though it would allow the thread to be its own ID. Ponder.

Parameters:
coroutineID - If >=0, requests that we reserve this number. If <0, requests that we find, reserve, and return an available ID number.
Returns:
If >=0, the ID number to be used by this coroutine. If <0, an error occurred -- the ID requested was already in use, or we couldn't assign one without going over the "unreasonable value" mark

co_entry_pause

public java.lang.Object co_entry_pause(int thisCoroutine)
                                throws java.lang.NoSuchMethodException
In the standard coroutine architecture, coroutines are identified by their method names and are launched and run up to their first yield by simply resuming them; its's presumed that this recognizes the not-already-running case and does the right thing. We seem to need a way to achieve that same threadsafe run-up... eg, start the coroutine with a wait. %TBD% whether this makes any sense...

Parameters:
thisCoroutine - the identifier of this coroutine, so we can recognize when we are being resumed.
Throws:
java.lang.NoSuchMethodException - if thisCoroutine isn't a registered member of this group. %REVIEW% whether this is the best choice.

co_resume

public java.lang.Object co_resume(java.lang.Object arg_object,
                                  int thisCoroutine,
                                  int toCoroutine)
                           throws java.lang.NoSuchMethodException
Transfer control to another coroutine which has already been started and is waiting on this CoroutineManager. We won't return from this call until that routine has relinquished control. %TBD% What should we do if toCoroutine isn't registered? Exception?

Parameters:
arg_object - A value to be passed to the other coroutine.
thisCoroutine - Integer identifier for this coroutine. This is the ID we watch for to see if we're the ones being resumed.
toCoroutine - Integer identifier for the coroutine we wish to invoke.
Throws:
java.lang.NoSuchMethodException - if toCoroutine isn't a registered member of this group. %REVIEW% whether this is the best choice.

co_exit

public void co_exit(int thisCoroutine)
Terminate this entire set of coroutines. The others will be deregistered and have exceptions thrown at them. Note that this is intended as a panic-shutdown operation; under normal circumstances a coroutine should always end with co_exit_to() in order to politely inform at least one of its partners that it is going away. %TBD% This may need significantly more work. %TBD% Should this just be co_exit_to(,,CoroutineManager.PANIC)?

Parameters:
thisCoroutine - Integer identifier for the coroutine requesting exit.

co_exit_to

public void co_exit_to(java.lang.Object arg_object,
                       int thisCoroutine,
                       int toCoroutine)
                throws java.lang.NoSuchMethodException
Make the ID available for reuse and terminate this coroutine, transferring control to the specified coroutine. Note that this returns immediately rather than waiting for any further coroutine traffic, so the thread can proceed with other shutdown activities.

Parameters:
arg_object - A value to be passed to the other coroutine.
thisCoroutine - Integer identifier for the coroutine leaving the set.
toCoroutine - Integer identifier for the coroutine we wish to invoke.
Throws:
java.lang.NoSuchMethodException - if toCoroutine isn't a registered member of this group. %REVIEW% whether this is the best choice.