Re: Indeterminism in tarlz's test suite

[Antonio Diaz Diaz]

Hi Florian,

Florian Schmaus wrote:
> testing bad input...tarlz: Exiting with failure status due to previous
> errors.
> 1108(5)

Thank you very much for reporting this.

The problem is in fact a race happening on a test archive with a format 
violation put there on purpose, and only can happen when more than 2 threads 
are being used.

The archive in question contains a directory with a long name requiring an 
extended header, and a global header has been inserted in its own lzip 
member between the extended header and the ustar header like this:

+=================+===============+==============+
| extended header | global header | ustar header |
+=================+===============+==============+

The problem happens when the thread decoding the ustar header finishes 
before the thread decoding the extended header has time to realize that its 
lzip member is incomplete.

I have fixed the problem by making the thread decoding the ustar header 
signal the error in a way that can later be overridden by another thread.

The underlying problem is that the pax format is not really adequate for 
parallel decoding, forcing tarlz to jump through a lot of hoops to achieve 
it. I hope a new tar format can be designed with parallel decoding in mind.

I'll release a fixed version as soon as I test the fix and finish a couple 
more things.

Best regards,
Antonio.