There is a systemic bug in BZ2File where the GIL is released to perform
compression work, and any other thread calling into BZ2File will
deadlock. We noticed in the write method, but inspection of the code
makes it clear that its systemic.

The problem is pretty simple. Say you have two threads and one bz2file
object. One calls write(), the other calls (say) seek(), but it could be
write() or other methods too. Now, its pretty clear that the question
'should these two threads get serialised' could be contentious. So lets
put that aside by saying 'raising an exception or serialising in
arbitrary order would be ok'.

What happens today is:
t1:bz2file.write
   bz2file.lock.acquire
   gil-release
   bz2compression starts
t2:gil-acquired
   bz2file.seek
   bz2file.lock.acquire(wait=1)  <- this thread is stuck now, and has
the GIL
t1:bz2compression finishes
   gil.acquire <- this thread is stuck now, waiting for the GIL

If any owner of the bz2file object lock will release the GIL, *all*
routines that attempt to lock the bz2file object have to release the GIL
if they can't get the lock - blocking won't work. I'm not familiar
enough with the python threading API to know whether its safe to call
without the GIL. If its not then clearly it can't be used to work with
getting the GIL, and lower layer locks should be used.

Thanks, nice catch.

On Sun, 2009-10-25 at 22:00 +0000, Antoine Pitrou wrote:
> Antoine Pitrou <pitrou@free.fr> added the comment:
> 
> Thanks, nice catch.

Yeah :).

> versions: +Python 2.6, Python 2.7, Python 3.1, Python 3.2

Python 2.5 is also affected - its what we're running on the server that
broke :) 

-Rob

> Python 2.5 is also affected - its what we're running on the server that
> broke :) 

Yes, but it doesn't receive any bug fixes anymore -- only security
fixes.

On Sun, 2009-10-25 at 22:27 +0000, Antoine Pitrou wrote:
> Antoine Pitrou <pitrou@free.fr> added the comment:
> 
> > Python 2.5 is also affected - its what we're running on the server that
> > broke :) 
> 
> Yes, but it doesn't receive any bug fixes anymore -- only security
> fixes.

Ok, we'll work around the issue there then ;)

-Rob

Here is a patch.

On Mon, 2009-10-26 at 19:23 +0000, Antoine Pitrou wrote:
> Antoine Pitrou <pitrou@free.fr> added the comment:
> 
> Here is a patch.

Looks fine to me assuming that the locking functions can be used outside
the GIL.

On the test side, the case I supplied was low noise for me - I'd
hesitate to do as much compression as you are (50 times more) unless you
saw it spuriously pass a lot - the nature of the deadlock isn't
dependent on races so much as simple scheduling - as long as the seocnd
thread is scheduled before the first thread completes compressing the
deadlock will occur.

-Rob

> Looks fine to me assuming that the locking functions can be used outside
> the GIL.

Yes, they can. Actually, even the GIL uses them. :-)

> On the test side, the case I supplied was low noise for me - I'd
> hesitate to do as much compression as you are (50 times more) unless you
> saw it spuriously pass a lot - the nature of the deadlock isn't
> dependent on races so much as simple scheduling - as long as the seocnd
> thread is scheduled before the first thread completes compressing the
> deadlock will occur.

Well, your test case often succeeded here, so I decided on a more
aggressive variation.

On Mon, 2009-10-26 at 21:27 +0000, Antoine Pitrou wrote:
> Antoine Pitrou <pitrou@free.fr> added the comment:

> Well, your test case often succeeded here, so I decided on a more
> aggressive variation.

fair enough, if its needed - its needed :)

-Rob

Fixed in r75818 (trunk), r75819 (2.6), r75820 (py3k), r75821 (3.1). Thanks!