Re: [RFC] Migrating net/sched to new module interface

[Rusty Russell]

In message <[e-mail]> you write:
> [e-mail] wrote:
> > Somewhat overdone.
> 
> I think it would be nice to introduce in 2.7 a shutdowncall
> (*) function class for modules that works like exitcall, but
> with the following differences:
> 
>  - does not return before the module has really de-registered
>    itself everywhere, including synchronization with any
>    callbacks, etc.
>  - has a return code, and can fail if it would have to sleep
>    for a possibly long time
> 
> Before calling the shutdown function, all symbols exported by
> the module are hidden, and after the shutdown functions returns,
> the module can be unloaded.

This already happens.  This is why all accesses to the module are
protected by try_module_get().

I've analyzed dozens of "here's my implementation idea" mails over the
last two years.  Here's the executive summary:

1) It's simply not a good idea to force 1600 modules to change, no
   matter what timescale.  And changing it in a way that is *more*,
   not *less* complex is even worse.

2) It's bad enough to force the interfaces to change: at least the
   primitive they are to use is one many of them are already using,
   and is very simple to understand.

Rusty.
PS.  The *implementation* flaw in your scheme: someone starts using a
     module as you try to deregister it.  Either you re-register the
     module (ie. you can never unload security modules), or you leave
     it half unloaded (even worse).
--
  Anyone who quotes me in their sig is an idiot. -- Rusty Russell.