[cffi-devel] Re: CFFI/CLISP issues with saved images

[Hoehle, Joerg-Cyril]

Luís Oliveira wrote:
>I was just testing CFFI/CLISP by loading the testsuite, dumping an  
>image, loading it and running the tests. CLISP crashes. There are two  
>problems:
>
>- The foreign library (in this case libtest.so) is not being  
>reloaded. [...]I recall CLISP's ffi:def-call-out taking care of this but
Hmm, did you try an absolute path? I suspect the library was not found.
I'd naively have expected this to work in the context of cffi as well as with direct uses of the FFI.


>- The callbacks become NULL pointers, thus all callback.* tests fail.  
>Any suggestions for fixing that? Randomly placing load-time-value  
>forms in %defcallback didn't seem to work. :-)
[replying without having thought much about this issue].
CLISP's image is not like an Emacs dump of memory. It doesn't preserve addresses.  It doesn't preserve malloc()'ed stuff, which I believe callbacks are (or parts thereof).
Actually, I would *not* expect any implementation to be able to reproduce the exact same address of anything when started from an image. CMUCL seems an exception, where there are additional constraints (like on Solaris, getting the linker to load a shared object file at a known location).

Similarly, any foreign library that stored a callback-pointer *must* *not* expect such a pointer to be valid when restarting. That library may itself be at a different address when reloaded, how could the callback not change?

IMHO, it might work in a scenario on Linux with naive tests where you exit and immediately restart, but what if you upgrade your system in between etc.?

Instead, my mental model is that callbacks must be re-established dynamically when an application starts.
-- which makes me think that the code for free'ing an old callback upon redefinition in cffi should add a ext:validp test before free'ing. I just added a check in my branch.

That's what I meant in my mail "unnamed callback" from 2005-12-19, where I wrote:
>(defun my-application
>  (let (x y)
>    ...code...
>    (register-callback 'close-window

I'd be happy to discuss this further.  Restarting from one's images is being neglected in the Lisp community these days IMHO.  Smalltalk implementations seems to behave better.

Regards,
	Jörg Höhle.