Re: broken local_t on i386

[Andi Kleen <ak@suse.de>]

On Monday 12 June 2006 18:54, Christoph Lameter wrote:
> On Mon, 12 Jun 2006, Andi Kleen wrote:
> 
> > > #define __get_cpu_var(var) (*RELOC_HIDE(&per_cpu__##var, __my_cpu_offset()))
> > 
> > It is also affected by your race. The inc would only be atomic if the counter
> > was in the PDA, but standard per cpu data isn't. So it has to follow 
> > a pointer and then it could already have switched.
> 
> I thought the above would refer to a PDA memory area that is specially 
> mapped by each processor? That is the only thing that would get this 
> working right because we would map to a different PDA if the process 
> would be mapped to a different processor.

It does, but the per cpu data that everybody uses doesn't reside in the PDA
because it wasn't possible to make this work with binutils

It would require a relocation relative to another symbol which isn't
really supported.

At some point I considered using runtime patching to work around
this limitation, but it would be some work and relatively complex.

So the PDA just contains a pointer to the real per CPU area and it's
added. Unfortunately it's three instructions or so and not atomic 
(mov, add, reference) 

> 
> > Fix would be to disable preemption. I don't think it needs cli/sti
> > on non preemptible kernels.
> 
> Yuck. The advantage of local.t was that it does not need any of these 
> tricks. What is the point of local.t if one needs to disable preemption?

No atomic operations. Preemption just requires to increase a counter
in thread info.

Also on non preemptive kernels - which are the majority - it's a single
instruction on x86. I guess preempt users can live with a bit more
overhead ... 

-Andi