[Pyrex] [Cython-dev] Callbacks from threads and PyGILState_Ensure/PyGILState_Release

[Robert Bradshaw]

On Sep 13, 2007, at 11:21 PM, Stefan Behnel wrote:

> Greg Ewing wrote:
>> Stefan Behnel wrote:
>>> Robert Bradshaw wrote:
>>>
>>>> 1) Between pyx and pxd files. If we allow them to differ, which  
>>>> takes
>>>> precedence?
>>
>> They would have to be "sufficiently compatible". If the
>> pxd does not annotate, and the pyx annotates as "without
>> GIL", that's acceptable, but not the other way around.
>>
>> However, I'm not sure there would be any benefit in
>> allowing this flexibility, so I would probably just
>> require them to match, at least in a first implementation.
>
> Hmmm, have you actually read my post that you stripped from the  
> quote above?
>
> I was proposing the exact opposite: disallow "with GIL" in .pxd  
> files and in
> function pre-declarations, and only allow them on the function  
> implementation.
> That way, you have them where they are required (and where they  
> make sense)
> and you just can't get it wrong because it's declared at the point  
> closest to
> the code that requires (or does not require) the GIL.
>
> The reasoning is that you are not gaining anything by declaring a  
> function
> signature "with GIL". It might get called from Python code, it  
> might get
> called from Cython/Pyrex code holding or not holding the GIL, and  
> it might get
> called from third party code as a callback. There is no way Pyrex/ 
> Cython can
> always make sure it knows if the GIL is held by the calling code.  
> It's C after
> all, things can get called from all over the place. So Pyrex/Cython  
> should not
> rely on its knowledge about the code it sees and just acquire the  
> GIL inside
> the implementation. Which makes it a clear implementation detail  
> that is in no
> way interesting to the outside world. It's not a signature issue.

I think the "without GIL" idea that Greg is proposing is potentially  
more powerful, probably more subtle, and certainly further from  
implementation that the "with GIL" statement as already implemented.  
I propose we disallow this decelerator in pxd files/forward  
declaration for this case.

- Robert