[logback-dev] IllegalMonitorStateException inAppenderAttachableImpl.appendLoopOnAppenders()

[Ceki Gulcu]

Hi Joern,

By "hair splitting", I was not referring to Maartens' message but to mine.

Anyway, I think you have made the case for moving the r.lock() invocation out of 
the try block very convincingly. I'll commit the relevant change shortly.

Thank you for your highly valuable input.

Joern Huxhorn wrote:
> Hi Ceki,
> 
> On 03.02.2009, at 17:54, Ceki Gulcu wrote:
> 
>>
>>
>> Hi Maarteen,
>>
>> Maarten Bosteels wrote:
>>> Hello,
>>> I think moving r.lock(); outside of the try block WILL help, because 
>>> the original exception will no longer be hidden.
>>
> 
> Even though I hate being a smart ass and I also don't like to say 
> something like "See, I said so", I have to do just that.
> 
> This is exactly what I meant when I wrote
> "It's generally a very bad idea to save a bit of standard code just 
> because a possible problem isn't imaginable at the moment. I'm not even 
> sure if the code without try {} finally {} will execute faster or if 
> there won't be any difference at all. "
> in http://jira.qos.ch/browse/LBCORE-67 and is also the prime reason why 
> I submitted the ticket in the first place.
> 
> The documentation of the Lock interface at 
> http://java.sun.com/javase/6/docs/api/java/util/concurrent/locks/Lock.html clearly 
> states:
> <quote>
> With this increased flexibility comes additional responsibility. The 
> absence of block-structured locking removes the automatic release of 
> locks that occurs with synchronized  methods and statements. In most 
> cases, the following idiom should be used:
> 
> Lock l = ...;
> l.lock();
> try {
>         // access the resource protected by this lock
> } finally {
>         l.unlock();
> }
> </quote>
> The "should" is only there because there might be an obscure case where 
> a different pattern would probably be helpful - and this isn't such a case.
> 
>> Moving r.lock() outside the try block MAY help because the finally
>> block which throws a related but distinct exception will no longer be
>> invoked. The second exception effectively hides the first
>> exception. However, this is true if r.lock() throws an exception
>> *visible* in logback code. I tend to think, for reasons explained in
>> my previous post, that somewhere within r.lock() call there is a
>> RuntimeException thrown but which is absorbed somewhere within the
>> r.lock() invocation. So, there is a good chance (but less than 100%)
>> that displacing the r.lock() call to outside the try block will *not*
>> reveal anything new.
> 
> The documentation of the lock() method at 
> http://java.sun.com/javase/6/docs/api/java/util/concurrent/locks/Lock.html#lock() states: 
> 
> <quote>
> Implementation Considerations
> 
> A Lock implementation may be able to detect erroneous use of the lock, 
> such as an invocation that would cause deadlock, and may throw an 
> (unchecked) exception in such circumstances. The circumstances and the 
> exception type must be documented by that Lock implementation.
> </quote>
> 
> Furthermore, the documentation of ReentrantReadWriteLock (the one used 
> in AppenderAttachableImpl) at 
> http://java.sun.com/javase/6/docs/api/java/util/concurrent/locks/ReentrantReadWriteLock.html does 
> actually document a case where exactly such a scenario happens:
> <quote>
> Implementation Notes
> 
> This lock supports a maximum of 65535 recursive write locks and 65535 
> read locks. Attempts to exceed these limits result in Error throws from 
> locking methods.
> </quote>
> 
> While this most likely isn't the case here it certainly illustrates one 
> thing:
> Just stick to the pattern.
> 
>> Having said that, if for one reason or another, the r.lock() call
>> throws an exception, which it certainly does, even if it is not
>> visible, there is arguably no point in calling r.unlock(). Calling
>> r.lock() outside the try block MAY reveal the original exception
>> without having a downside.
> 
> It will.
> 
>> On the other hand, if calling r.unlock has
>> a positive side-effect even if r.lock() fails, then moving r.lock()
>> outside the try block may not be such a good idea after all.
> 
> It won't have a positive effect because it would be - in fact - an 
> undocumented behavior and that is *never* something to desire.
> 
> lock() does only throw an unchecked exception if it couldn't acquire the 
> lock (for one reason or another) while unlock() throws the exception 
> mentioned in the subject if the lock is not held (see 
> http://java.sun.com/javase/6/docs/api/java/util/concurrent/locks/ReentrantReadWriteLock.WriteLock.html#unlock() ). 
> 
> 
>> To summarize, although not guaranteed, there is a good chance that
>> moving the r.lock() call outside the try block will reveal new
>> information. However, this change may be problematic in certain cases,
>> which at this time we can only imagine but not pinpoint to. In
>> Rumsfeld-speak, moving r.lock has a known but uncertain positive
>> effect and uncertain and unknown negative effects.
>>
> 
> Everything is perfectly documented and defined. This is mainly caused by 
> the fact that java.util.concurrent was created by Doug Lea and this guy 
> really knows what he's doing. His book on concurrency is quite excellent.
> Beside that, I dislike Rumsfeld ;)
> 
>> Alternatively, we could invoke r.unlock in a safe manner, that is
>> within a try/catch block (embedded within finally). This would not
>> obfuscate the first exception and would ensure that r.unlock is always
>> invoked.
>>
>> I am not sure splitting hairs as demonstrated above gets us any closer
>> to a solution. Zoltan, do you need any specific help from my side?
> 
> I fail to see the hairsplitting in Maartens message and I agree with him.
> 
> Instead of invoking unlock and circumventing the built in security 
> measures using try{...}catch(Throwable){} we should just use the Lock as 
> defined in the API.
> 
> Sorry for being nosy,
> Joern.


-- 
Ceki Gülcü
Logback: The reliable, generic, fast and flexible logging framework for Java.
http://logback.qos.ch