Re: [RFC] Thread Migration Preemption

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

* Andi Kleen (andi@firstfloor.org) wrote:
> On Fri, Jul 06, 2007 at 10:41:44AM -0400, Mathieu Desnoyers wrote:
> > I haven't thought about making it the default for kernel space
> > preemption, but yes, it would make sense.
> 
> Now it's too late -- getcpu() has infected the kernel everywhere.
> It would have made sense a few years ago.
> 
> 
> > ... getcpu()...
> 
> Hmm ok, although i suspect it's rare to assume that. But understood
> you don't want to audit all getcpu users because of this.
> 
> > using a short instead of an int on modern x86 will cause pipeline stalls
> > due to partial register use.
> 
> Sorry, that's totally bogus. Primarily because the access would be directly
> on memory and there is no partial register tracking there.
> 
> Besides pipeline stall is not the correct description on what would
> happen if you used a register, the worst you get is a single false dependency
> but no pipeline flush.
> 
> Besides the latest x86 cpus (C2, K8) don't have much trouble with these
> false dependencies in general.
> 

Yes, false dependency is what I meant. And hrm, yeah I guess that mostly
movzbl would be used to get the byte from memory and zero-extend, making
sure there is no false dependency.

> > usage, since it is followed by an unsigned long; gcc structure alignment
> 
> 
> > will put padding instead of the integer, which does not buy us anything
> 
> on i386 unsigned long is 4 bytes.
> 

Since we have, on i386:

        int                     preempt_count;  /* 0 => preemptable, <0 => BUG */
        int                     migrate_count;/* 0: can migrate, <0: BUG */
        mm_segment_t            addr_limit;     /* thread address space:

and:

typedef struct {
        unsigned long seg;
} mm_segment_t;

Turning migrate_count into a char would only put padding between
migrate_count and addr_limit. So I still do not see a clear improvement
in memory usage there... but I don't specially care about it being an
integer or a byte though :)

Mathieu


> -Andi

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