Re: [pyOpenSSL] Python core dumps when using pyOpenSSL 0.8.0 and threads

[Jean-Paul C. <ex...@di...>]

On Wed, 1 Apr 2009 15:26:43 -0500, Jean-Paul Calderone <ex...@di...> wrote:
>
> [snip]
>
>I wasn't actually very confident that the patch was correct, but now that
>I've actually written all that up and traced through the various levels of
>thread control APIs in CPython, I'm relatively confident that the patch is
>doing the right thing.  I also learned that there are some APIs in CPython
>which do much of what pyOpenSSL is doing.  They aren't available in Python
>2.3 though, so I'm not quite ready to switch over to them.

Shortly after sending this message, I realized that it is also a perfectly
valid explanation for why the current trunk code also works correctly and
does not segfault.  However, the current trunk code does segfault, so
something must be wrong. ;)

After a few hours with gdb, I found the problem.  Python threads identifiers
are only unique for the lifetime of the thread.  As threads are destroyed
and others created, the identifiers may be re-used.  Since thread identifiers
are used as TLS keys, this means that thread A may set a TLS key then exit
and a later thread B may then be able to see the value for that TLS key.
This means that without the delete, thread B won't save the correct new
PyThreadState*, thus causing the bug.

Jean-Paul