Re: SMP status ?

[Andrew Doran <ad%netbsd.org@localhost>]

On Thu, Jan 13, 2011 at 07:25:11PM +1100, matthew green wrote:
> 
> i've commited a change to -current that should work around this problem,
> so any stability issues should be gone.  it's just a workaround but it is
> worth updating to since it makes things mostly work, at least on my SS20.
> 
> it still has the problem, but doesn't crash now, and the "savefp null ipi"
> event counters will slowly increase.
> 
> if you're adventurous, please try -current sources and apply this patch:
> 
>       http://www.netbsd.org/~mrg/sparc.kpreempt.diff
> 
> for me it's run for almost 2 hours without a problem reported.  i chose
> the various points to add these new kpreempt_disable() calls by reviewing
> what the x86 platform does for it.  it *probably* contains points where
> preemption is disabled unnecessarily.
> 
> andy, can you have a look?  i really am not sure why this matters, but
> it seems to really help.

Hey Matt. Some thoughts, first whenever we are > IPL_NONE, preemption is off
already (so splvm or whatever does the job).

Since interrupts raise the IPL you've got the same thing there
implicitly while in an interrupt path.

Second the kernel preemption gunk
should do almost nothing on a non-preemptive machine.  So I'd be really
surpised if this is having any effect other than changing timing or
something wacky like code/register/cache layout.  evcounters for preemption
should always report zero on sparc (unless somebody went an implemented
preemption there!).

> 
> 
> 
> .mrg.