arm64: Fix the condition used to test for AP initialization

[markj]

Summary

Testing pc_curpcb != NULL is racy, since on arm64 its initially set in
pmap_switch(), where the bootstack is still in use. Take a different
approach: wait for APs to acknowledge an interrupt before freeing the
bootstacks.

Reported by: mmel

[andrew]

Could we not move setting the stack pointer earlier in cpu_throw? If it was before pmap_switch we will have switched away from the boot stack before curpcb is set.

[markj]

In D35435#803606, @andrew wrote:

Could we not move setting the stack pointer earlier in cpu_throw? If it was before pmap_switch we will have switched away from the boot stack before curpcb is set.

I think that could work too, but isn't this method simpler and less tied to the internals of cpu_throw? This approach should work on all platforms, not just arm64.

[mmel]

It solved my problem and it boots without any problems on all my arm64 boards. Thanks.

[markj]

@kib do you see any problem with adopting the same approach on other platforms, i.e., x86 and riscv?

[kib]

I think this should be fine on x86.

[markj]

Use the same approach on x86 and riscv.

[mhorne]

sys/riscv/riscv/mp_machdep.c

213 ↗

(On Diff #106899)

Do we need this here too?

[markj]

• Remove some more unneeded assertions.
• Use acquire semantics for a load to pair with a release store.