By default bind interrupts to the set of CPUs in the domain they are connected to

[jeff]

Summary

This changes the default behavior from round-robin between all CPUs to round-robin among all CPUs in the identified domain. This can be overridden by binding the interrupts directly to the CPUs of choice as you would otherwise. I think this code should eventually be pulled up into an MI interface anyway but I didn't want to tackle that as part of this project.

I moved the vm_ndomains definition to make it easier to include. vm_phys.h requires vm_page.h which gets you into a half dozen vm headers just to use one int. So I put it in vm.h for simplicity.

[kib]

x86/x86/intr_machdep.c

514 ↗

(On Diff #40714)

Don't you need to initialize current_cpu for each domain to some CPU belonging to that domain ? Otherwise, the first call to intr_next_cpu() for each domain would return apic_id of BSP. Or I am missing something there.

707 ↗

(On Diff #40714)

And there.

[jeff]

x86/x86/intr_machdep.c

514 ↗

(On Diff #40714)

You are right. I will fix this.

[jhb]

x86/x86/nexus.c

592 ↗

(On Diff #40714)

I would sort of prefer the logic for how to handle devices without a domain to be centralized in intr_add_handler() (e.g. set to -1 here and then have intr_add_handler() set to 0) in case we want to modify that in the future to do the legacy thing of round-robin among all CPUs, but that is a minor nit.