[
https://issues.apache.org/jira/browse/COMPRESS-357?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15325939#comment-15325939
]
Sebb commented on COMPRESS-357:
-------------------------------
Unfortunately AFAICT that still does not guarantee that all the variables will
be safely published.
A variable is only safely published if the write happens-before the read.
One way to do this is where both the writer and reader synchronize on the
*same* lock.
Another way I think is a shared volatile.
If thread A updates some fields and then writes the volatile, when thread B
reads the same volatile it will see any such updates.
But if thread A has since updated some fields, thread B could see either the
original or a later value.
> BZip2CompressorOutputStream can affect output stream incorrectly
> ----------------------------------------------------------------
>
> Key: COMPRESS-357
> URL: https://issues.apache.org/jira/browse/COMPRESS-357
> Project: Commons Compress
> Issue Type: Bug
> Components: Compressors
> Affects Versions: 1.9, 1.11
> Environment: multithreaded
> Reporter: Richard Shapiro
> Labels: easyfix
> Fix For: 1.12
>
> Original Estimate: 1h
> Remaining Estimate: 1h
>
> BZip2CompressorOutputStream has an unsynchronized finished() method, and an
> unsynchronized finalize method. Finish checks to see if the output stream is
> null, and if it is not it calls various methods, some of which write to the
> output stream.
> Now, consider something like this sequence.
> BZip2OutputStream s = ...
> ...
> s.close();
> s = null;
> After the s = null, the stream is garbage. At some point the garbage
> collector call finalize(), which calls finish(). But, since the GC may be on
> a different thread, there is no guarantee that the assignment this.out = null
> in finish() has actually been made visible to the GC thread, which results in
> bad data in the output stream.
> This is not a theoretical problem; In a part of a large project I'm working
> on, this happens about 2% of the time.
> The fixes are simple
> 1) synchronize finish() or
>
> 2) don't call finish from finalize().
> A workaround is to derive a class and override the finalize() method.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
