On Fri, 17 Mar 2000, Richard Henderson wrote:
>I figured that might be the case. Andrea's last set of irq patches
>delivered an interrupt to all cpus, only one of which would eventually
>handle it. The rest of them wasted time.
The wasted time is not wasted since it's killing the irq latency (as IA32
does in hardware). If there would be no cons in the current setting I
wouldn't be using it in first place. We can do the CPU selection only in
software and nevertheless it means some CPU cycles has to be spent in the
irq selection if we want that feature.
>I just wrote a little widget to statically distribute the PCI interrupts
>across the active processors. That should help a bit.
Good, I think it's fine to have the per-cpu irq distribution script in
userspace. However I would prefer to have a kind of distribution also in
kernel. That's __init section stuff and everybody can override it anyway
with the userspace proggy later if the kernel heuristic is not optimal.
Andrea
