Re: Module unloading

[Andi K. <ak...@su...>]

On Sun, Sep 02, 2001 at 01:58:33AM +0100, John Levon wrote:
> 
> We have to disable module unloading for each release I think.
> 
> Whilst the race window between the incl of the module use count and
> user processes actually leaving the module code area is small, I don't think
> we can take the risk of leaving it in.

If you don't sleep you could at least make it #ifdef __SMP__ and also
add a magic module argument to still allow it.

> 
> Also as far as I understand we could be in much deeper trouble when
> the kernel becomes pre-emptible in 2.5.
> 
> A silver bullet would be nice, but syscall intercepting is such an ugly
> thing to do, we will have to just accept one of its disadvantages.

There is a silver bullet for module unloading actually. Unfortunately it is
a 2.5 thing too.  The trick is to add a quiescent period in unloading that
makes sure every CPU has been through at least one context switch after the
module count went zero. You could in theory implement it yourself with a 
second module; although it fits better in the main code.

-Andi