Allocate arm64 per-CPU data in the correct domain

[andrew]

Summary

To minimise NUMA traffic allocate the pcpu, dpcpu, and boot stacks in
the correct domain when possible.

[markj]

sys/arm64/arm64/machdep.c

106

Isn't this the same as cpuid_to_pcpu[] now? Do we need __pcpu[] at all?

sys/arm64/arm64/mp_machdep.c

505

639

I think this probably deserves an XXX comment.

[markj]

sys/arm64/arm64/mp_machdep.c

514

FWIW, there is no benefit to allocating domain-local boot stacks, they are used only in the window after the AP starts and before it switches to a thread.

[andrew]

Use cpuid_to_pcpu

[andrew]

sys/arm64/arm64/mp_machdep.c

514

I mostly did it for consistency.

[markj]

sys/arm64/arm64/machdep.c

773

We assume that .bss is allocated from the domain to which the BSP belongs. That's probably true in practice though, at least amd64 makes a similar assumption when it allocates the pcpu area in pmap_bootstrap().