Re: [PATCH] Remove module reference counting.

From: Rusty Russell <rusty@rustcorp.com.au>
To: Gianni Tedesco <gianni@scaramanga.co.uk>
Cc: davem@redhat.com, arjanv@redhat.com, torvalds@transmeta.com,
	greg@kroah.com, linux-kernel@vger.kernel.org,
	alan@lxorguk.ukuu.org.uk
Subject: Re: [PATCH] Remove module reference counting.
Date: Sun, 27 Jul 2003 06:18:05 +1000	[thread overview]
Message-ID: <20030727184118.3D7872C003@lists.samba.org> (raw)
In-Reply-To: Your message of "25 Jul 2003 23:43:16 +0100." <1059172995.16255.6.camel@sherbert>

In message <1059172995.16255.6.camel@sherbert> you write:
> Wasn't the idea once banded about of a 2-stage unload that went
> something like:
> 
> 1. ->cleanup() - unregister IRQ handlers, timers, etc.
> 2. Quiesce the system
> 3. Safe to unload
> 
> surely if nothing is registered and all CPUs do a voluntary schedule()
> then there can be no chance of calling back in to the module.

Yes, I implemented this, even.  The problem is that the desired module
semantics are "unload if it'll work, otherwise do nothing and fail".
This means that you either get halfway through the cleanup function
and back out (which leaves the race where some interfaces to your
module is MIA for a while), or you hang forever if things are in use.

It's this "atomically check refcount and deregister" that
try_module_get() gives us, by effectively unregistering all the
modules' interfaces at once.

Rusty.
--
  Anyone who quotes me in their sig is an idiot. -- Rusty Russell.