Re: [RFC] new module format

[Daniel Phillips <phillips@arcor.de>]

On Tuesday 16 July 2002 15:04, Roman Zippel wrote:
> 1. Properly fixing module races: I'm playing with a init/start/stop/exit
> model, this has the advantage that we can stop anyone from reusing a
> module and we only have to wait for remaining users to go away until we
> can safely unload the module.

I'm satisfied that, for filesystems at least, all the module races can be 
solved without adding start/stop, and I will present code in due course.
However, Rusty tells me there are harder cases than filesystems.  At this
point I'm waiting for a specific example.

For filesystems, we rely on the filesystem code itself to know when all users 
have gone away.  If somebody is still executing in a filesystem module after 
all umounts are done, it's a horrible nasty bug.  We might still want to play 
games with checking execution addresses of processes to see if anybody is 
still in a module, but that would just be for debug; sys_delete_module can 
rely on the filesystem's opinion about whether a module is quiescent or not.

Somebody please give me an example of why this same strategy will not
work for all types of modular code.

-- 
Daniel