Re: [PATCH -rt 4/5] use migrate_disable for __local_begin

[Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>]

* Peter Zijlstra (a.p.zijlstra@chello.nl) wrote:
> On Sat, 2007-07-14 at 14:35 -0400, Mathieu Desnoyers wrote:
> > * Peter Zijlstra (a.p.zijlstra@chello.nl) wrote:
> > > On Sat, 2007-07-14 at 13:16 -0400, Mathieu Desnoyers wrote:
> > > > * Peter Zijlstra (a.p.zijlstra@chello.nl) wrote:
> > > > > Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
> > > > > ---
> > > > >  include/asm-i386/local.h   |    7 ++++---
> > > > >  include/asm-x86_64/local.h |    7 ++++---
> > > > >  2 files changed, 8 insertions(+), 6 deletions(-)
> > > > > 
> > > > > Index: linux-2.6/include/asm-i386/local.h
> > > > > ===================================================================
> > > > > --- linux-2.6.orig/include/asm-i386/local.h
> > > > > +++ linux-2.6/include/asm-i386/local.h
> > > > > @@ -197,11 +197,12 @@ static __inline__ long local_sub_return(
> > > > >  #define __local_begin(__flags)					\
> > > > >  {								\
> > > > >  	(__flags) = 0;						\
> > > > > -	preempt_disable();					\
> > > > > +	migrate_disable();					\
> > > > 
> > > > Brrrr. That's wrong. Your non atomic __local*() updates only makes sense
> > > > when preempt_disable/enable() protects them from concurrent threads on
> > > > the same CPU, which is not the case of migrate_disable/enable(). This is
> > > > why I suggest that you use local_begin/end() mapped to
> > > > migrate_disable/enable() for normal local variables, and, if you really
> > > > want a __local_begin/end(), then it should be mapped to
> > > > preempt_disable/enable() and should state that it provides no protection
> > > > against interrupts.
> > > 
> > > Sure, but on -rt it does suffice, this part of the patch is rather WIP.
> > > 
> > > 
> > 
> > Hrm, how can it suffice, I wonder ? migrate_disable() does not protect
> > against other threads on the same CPU, so you could suffer from
> > concurrent updates to the same variables. How is it different in -rt ?
> 
> I thought the idea was that all these local_* operation were atomic wrt
> to the local cpu.
> 

Yes, they are. But your __local_*() variants are not. They require a
full preempt_disable because they do not execute atomically.

So having local_begin/end() -> migrate_disable/enable() is ok.
And __local_begin/end() -> preempt_disable/enable() is ok.

I just want to make the naming and usage consistent with the protection
offered.

Mathieu

> The only difference with -rt is that we generally don't care about
> interrupts.
> 
> Anyway, I'm dropping all this local stuff, and just hard code it right
> into slub.c
> 
> 
> 

-- 
Mathieu Desnoyers
Computer Engineering Ph.D. Student, Ecole Polytechnique de Montreal
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68