Home » Modeling » OCL » Non-deterministic error with OCL validation in Xtext project(CompleteOCLEObjectValidator experiences semi-reproducible stack traces. Perhaps race condition during initialization?)
| |
| Re: Non-deterministic error with OCL validation in Xtext project [message #1783916 is a reply to message #1783696] |
Tue, 20 March 2018 10:44   |
 German Vega
Messages: 104
Registered: December 2015
Location: Grenoble, France |
Senior Member |
|
|
Hello
I have a standalone application that uses complete OCL validation, and I got very similar random crashes to yours, with exactly the same kind of exceptions (stdlib not available).
At the end (after some painful debugging) I discovered that my problem was caused by the environment factory (org.eclipse.ocl.pivot.utilities.EnvironmentFactory) used by the validator being detached in the middle of the execution.
After some more painful debugging to try to understand why it was disposed, it boiled down to the fact that the org.eclipse.ocl.pivot.utilities.OCL object used to create the EnvironmentFactory was being garbage collected, and it has a finalize method that detaches the factory.
To make a long history short, the solution was just to be sure to keep a hard reference to the OCL object used to create the factory.
Although I am not validating xtext models, maybe you have a similar problem to mine.
So I think that instead of something like this :
public class StatesJavaValidator extends AbstractStatesJavaValidator {
@Override
public void register(EValidatorRegistrar registrar) {
super.register(registrar);
StatesPackage ePackage = StatesPackage.eINSTANCE;
URI oclURI = URI.createPlatformResourceURI("/StatesProject/model/States.ocl", true);
OCL ocl = OCL.newInstance();
registrar.register(ePackage, new CompleteOCLEObjectValidator(ePackage, oclURI,ocl.getEnvironmentFactory()));
}
}
You'd rather do something more like this :
public class OCLReference {
public static OCL ocl = OCL.newInstance();
public static void dispose() {
ocl.dispose();
}
}
public class StatesJavaValidator extends AbstractStatesJavaValidator {
@Override
public void register(EValidatorRegistrar registrar) {
super.register(registrar);
StatesPackage ePackage = StatesPackage.eINSTANCE;
URI oclURI = URI.createPlatformResourceURI("/StatesProject/model/States.ocl", true);
registrar.register(ePackage, new CompleteOCLEObjectValidator(ePackage, oclURI, OCLReference.ocl.getEnvironmentFactory()));
}
}
And you explicitly dispose the OCL instance at a point in your application when you are sure you do not need validation any longer.
Hope it helps
German Vega
[Updated on: Tue, 20 March 2018 10:49] Report message to a moderator |
|
| |
| Re: Non-deterministic error with OCL validation in Xtext project [message #1783939 is a reply to message #1783920] |
Tue, 20 March 2018 15:11 |
German Vega
Messages: 104
Registered: December 2015
Location: Grenoble, France |
Senior Member |
|
|
Hello Ed,
Well, once I could diagnose the problem, I found that the lifecycle of the OCL instances was well documented in the source code and javadoc. But, I do not recall the documentation about CompleteOCL validation mentions it.
Concerning the bug you opened, you can imagine so many different scenarios for validation (batch/live/on-demand, by resource set, by resource, .....) that is difficult to imagine a lifecycle mechanism to suit them all. In any case, this problem mainly concern standalone users, so perhaps better documenting the issues in the programmer's guide is the only cost-effective deal.
One thing I kept wondering while I was investigating the problem was why it did not happen inside Eclipse, only in my standalone application. So I looked in the source code to see how it was handled when you do "OCl -> load document.." and discovered classes CompleteOCLLoader and OCLAdapter.
They suited perfectly well my application's needs, as they couple the lifecycle of the OCL object to the lifecycle of the resource set. So at the end I completely forgot about CompleteOCLEObjectValidator and instead used CompleteOCLLoader directly. I join the utility class I developed for performing validation, so others can get inspiration if they read this thread.
Perhaps this deserves some documentation in chapter "Unified or Pivot Programmers Guide - Validators"
However, I am not sure this can help in the scenario of "Complete OCL for Xtext Validation". I tried to look at how xtext handles the lifecycle of validators but it is too difficult to understand, as it uses guice injection. That's why I proposed this solution with the static field and the explicit dispose, even if it is far from satisfactory.
The ideal will be to dispose the OCL instance when the xtext validator is no longer required. Maybe if you or Christian Burghard, who seem more knowledgeable in xtext than me, can ask the question in their forum.
Regards
German
|
|
| | | | | |
| Re: Non-deterministic error with OCL validation in Xtext project [message #1784083 is a reply to message #1784071] |
Thu, 22 March 2018 10:35 |
German Vega
Messages: 104
Registered: December 2015
Location: Grenoble, France |
Senior Member |
|
|
Hi Ed,
The problem is not really that xtext editors are introducing some hidden sharing, the bottom line is that all are using the same singleton EMF's EValidator.Registry.INSTANCE for performing validation.
So xtext registers the CompleteOCLEObjectValidator in the validation chain of EValidator.Registry.INSTANCE, and from then on it will be implicitly shared by any application (not only xtext editors) that uses the standard org.eclipse.emf.ecore.util.Diagnostician
Because initialization of CompleteOCLEObjectValidator is lazy, it can really happen in any thread (depending on how it is triggered : from a menu, from a builder, ..... ) and sometimes it can happen concurrently in several threads.
The scenario of having three opened files (you can see them at the right of the first screenshot test.mydsl test2.mydsl test3.mydsl) was just a way to systematically reproduce the problem. The files had nothing in particular.
The steps to reproduce the scenario are the following :
1) define your xtext DSL and add complete OCL validation
2) start a debugged eclipse with your DSL plugins
3) edit several DSL model files in different editors
4) let the editors opened and exit the debugged eclipse
5) put a breakpoint in CompleteOCLEObjectValidator.initialize
6) restart the debugged eclipse
7) the first thing each editor will do on reopening is try to perform a validation in a working thread, and you will get into the situation of the first screenshot
I agree with you that EMF itself is not thread safe, but validation needs just read-only access to the model so it should work .. the problem is all the side-effects that are not really read-only (loading of resources, CS2AS , calculation of the complete OCL model, ...)
I know you really did a great effort and wonderful job to make OCL work in isolated resource sets and as transparently as possible... but there are so many singletons in EMF that multi-thread sharing is almost inevitable ...
|
|
| | | |
Goto Forum:
Current Time: Fri Apr 19 23:19:16 GMT 2024
Powered by FUDForum. Page generated in 0.03399 seconds |
] 
Search
Help
Register
Login
Home
