If D31188 and D32866 were approved quickly, D30935 would instead merely move this piece to xen_intr_alloc_isrc(). This would reduce code duplication since no architecture would need to call xen_intr_assign_cpu(). Instead each architecture would simply assign ->xi_cpu and this would take care of the assignment.

This is a more advanced version of what I was referring to in my recent comment on D30935.

Now that is a funky line attribution for a diff.

Before ca7af67ac958d3749ebfc4771cecf0cd7525e6a0 there was:

if (error == 0)
        error = intr_event_bind(isrc->xi_intsrc.is_event, cpu);

Here, but that got lost. I believe this is a small commit goof, so I believe doing intr_event_bind() (or xen_arch_intr_event_bind()) before the xen_intr_assign_cpu() is the correct. Otherwise, removing the intr_event_bind()/xen_arch_intr_event_bind() call for all cases is correct.

What did I learn when looking at history? Apparently not having this for PORT channels was a deliberate performance decision. That seems odd, but was apparently found worthwhile.

In such case this segment should have a if (isrc->xi_type != EVTCHN_TYPE_PORT) { added around it. This retains the advantage of doing processor assignments for all interrupts during initialization and in the same location.

Then read again and the commit messages don't state such.

The one for 76acc41fb7c740bc49e2638529b8cc750ff281d5 is more concerned about avoiding the APICs (this retains that). Meanwhile when this was implemented at f229f35db70f9a4bfac12c49e91bd9dd56a4975e there is no reference.

I'm reverting to this was simply missed during implementation for PORT channels.