Re: mapping interrupts to a particular cpu

On Wednesday 31 January 2007 22:33, Alfred Perlstein wrote:
> Through source inspection it appears that we don't have interrupts
> running in the context of the actual interrupt handler unless they
> are truly "FAST".  (correct me if I'm wrong).
> 
> That makes my decision a lot easier.  I will just use bind_cpu.
> 
> If so, I will just use bind_cpu.

What is bind_cpu, sched_bind maybe?

If you want to bind interrupts to CPUs, look at 
http://www.FreeBSD.org/~jhb/patches/intr_bind.patch  It's tested and works on 
i386 and amd64 and lets you bind interrupts to specific CPUs.  There is a 
sysarch() to do it, so you can write a trivial userland app.  It doesn't 
currently provide a way for userland to ask which interrupts are bound to 
which CPUs.  I'm mostly waiting for folks to benchmark it to make sure it 
isn't a pessimization.

> * Alfred Perlstein <[email protected]> [070131 19:18] wrote:
> > [[jhb cc'd cause he's Mr Apic from what I understand.]]
> > 
> > I'm trying to figure out pinning interrupts to a particular cpu.
> > 
> > If it can be done, if it's worth it and how.
> > 
> > Any generic pointers one could give to help me do this?  I'd like to
> > bind network interfaces to a particular cpu.
> > 
> > I guess the questions boil down to:
> > 
> > 1) Are we at the point that an interrupt for something such as an
> >    ethernet device actually runs under interrupt context?  
> >    (or do we always switch to an ithread (for !FAST_INTR))
> >    (I already know about bind_cpu which can be done inside interrupt
> >     routine (I think?))
> > 
> > Anyhow, if we have the actual interrupts running directly...
> > 2) How can I bind an interrupt source to a cpu?
> > 3) Is there a way to determine which cpu is best to wire an interrupt to?
> > 
> 

-- 
John Baldwin
_______________________________________________
[email protected] mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-smp
To unsubscribe, send any mail to "[email protected]"