Zombie FileHandler locks can exhaust all available log file locks.

[Daniel Fuchs]

Hi Jason,

Looking at the diff for 6244047 - I see that, as you pointed
out, the unwanted behavior described comes from the fact that
the new code is using CREATE_NEW - which prevents the 'zombie
lock files' from being reused.

I am not an expert in file locks - but I wonder whether we
could revert to using CREATE instead: wouldn't tryLock() later
tell us if the file is used by another entity?

Another possibility would be to combine CREATE_NEW and
DELETE_ON_CLOSE, which according to StandardOpenOptions will
attempt to delete the file on JVM termination if close()
hasn't been called.
This probably wouldn't help in case on VM crash,
but it would help in the case demonstrated by your test below.
I have however some reluctance because I see that we call
FileChannel.close() in the case were the lock can't be obtained,
and I'm not sure what that would do...
Also StandardOpenOptions has some warning about using
DELETE_ON_CLOSE for files that can be opened concurrently
by other entities - so I'm not sure it would be appropriate.

The last possibility I see would be to change the lock HashMap to
take instances of a subclass of WeakReference<FileHandler> as
values (instead of String), and add some code that attempts to
close & remove the lock file when the FileHandler is no longer referenced.
Again - this will probably not help in the case of crash, and also
adds the question on when the weak reference queue should be polled to
ensure that the no longer referenced FileChannel are closed in a
timely fashion.

All in all - I feel our best options would be to revert to using
CREATE, or use CREATE_NEW + DELETE_ON_CLOSE, or not do anything
and live with the issue.
Hopefully some nio experts will chime in ;-)

On another track - we could also make MAX configurable - but that
would just be shifting the issue around - wouldn't it?

best regards,

-- daniel

On 6/20/14 10:54 PM, Jason Mehrens wrote:
> Daniel, Jim,
>
>
> In JDK8 the FileHandler file locking was changed to use FileChannel.open with CREATE_NEW. If the file exists (locked or not) the FileHandler will rotate due to safety concerns about writing to the same log file. The FileHandler has an upper bound of 100 as the number of file locks that can be attempted to be acquired. Given the right pattern, enough time, and enough failures it seems at it is possible for a program to end up in a state where the FileHandler is surrounded by zombie lock files refuses log or perform a clean up action. (A.K.A Farmer Rick Grimes). This means that the lck files have to manually be deleted or the FileHandler will just fail with an IOException every time it is created.
>
>
> A simple test to emulate crashing is to run (JDK7 vs. JDK8) the following code twice without deleting the log and lck files:
>
>
> public static void main(String[] args) throws Exception {
>     System.out.println(System.getProperty("java.runtime.version"));
>     ReferenceQueue<FileHandler> q = new ReferenceQueue<>();
>
>     for (int i=0; i<100; i++) {
>       WeakReference<FileHandler> h = new WeakReference<>(
>       new FileHandler("./test_%u.log", 10000, 2, true), q);
>       while (q.poll() != h) {
>           System.runFinalization();
>           System.gc();
>           System.runFinalization();
>           Thread.yield();
>       }
>    }
>
> }
>
>
> I understand that if you can't trust that the file locking always works then there isn't much that can be done. Leaving the number of locks as unbounded isn't really an option either. Seems like there should be a way to identify zombie lock files (ATOMIC_MOVE) and delete them.  Any thoughts on this?
>
> The source discussion on this can be found at http://stackoverflow.com/questions/24321098/is-java-util-logging-filehandler-in-java-8-broken
>
>
> Regards,
>
> Jason 		 	   		
>