Re: [PULL] cpumask: finally make them variable size w/ CPUMASK_OFFSTACK.

[Rusty Russell <rusty@rustcorp.com.au>]

On Wed, 9 May 2012 10:44:53 +0200, Ingo Molnar <mingo@kernel.org> wrote:
> 
> * Rusty Russell <rusty@rustcorp.com.au> wrote:
> 
> > Hi Ingo,
> > 
> >         I finally rebased this on top of your tip tree, and tested it
> > locally.  Some more old-style cpumask usages have crept in, but it's a
> > fairly simple series.
> 
> Cool! Most of it looks pretty sane. I have a question about the 
> gist of the series:
> 
> > commit 898eb73305e2277be91b931c5a75484f8c87ae36
> > Author: Rusty Russell <rusty@rustcorp.com.au>
> > Date:   Wed May 9 15:01:15 2012 +0930
> > 
> >     cpumask: remove struct cpumask definition when CONFIG_CPUMASK_OFFSTACK=y
> >     
> >     We're about to change CONFIG_CPUMASK_OFFSTACK so it only allocate
> >     nr_cpu_ids bits for all cpumasks.  We need to make sure that when
> >     CONFIG_CPUMASK_OFFSTACK is set:
> >     
> >     1) Noone uses the old bitmap ops, which use NR_CPUS bits (use cpumask_*)
> >     2) Noone uses assignment of struct cpumask (use cpumask_copy)
> >     3) Noone passes a struct cpumask (pass a pointer)
> >     4) Noone declares them on the stack (use cpumask_var_t)
> >     
> >     So we finally remove the definition of struct cpumask when
> >     CONFIG_CPUMASK_OFFSTACK=y.  This means that these usages will hit a compile
> >     error the moment that config option is turned on.
> >     
> >     Note that it also means you can't declare a static cpumask.  You
> >     should avoid this anyway (use cpumask_var_t), but there's a
> >     deliberately-ugly workaround for special cases, using DECLARE_BITMAP()
> >     and to_cpumask().
> >     
> >     Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
> >     Cc: Arnd Bergmann <arnd@arndb.de>
> >     Cc: anton@samba.org
> >     Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
> >     Cc: Mike Travis <travis@sgi.com>
> 
> Is there any good reason to not remove it altogether, regardless 
> of whether the OFFSTACK config is set? I mean, triggering build 
> failures for a relatively rarely turned on config option is 
> asking for constant maintenance trouble.

Mainly because I didn't want to disturb the archs which don't care at
all about large cpumasks.  After all, putting a struct cpumask on the
stack is pretty convenient.

But we could add a new arch config which removes it, and set it from
x86.

Cheers,
Rusty.