x86/idle: reduce contention on ACPI register accesses

x86/idle: reduce contention on ACPI register accesses

Other than when they're located in I/O port space, accessing them when
in MMIO space (currently) implies usage of some sort of global lock: In
-unstable this would be due to the use of vmap(), is older trees the
necessary locking was introduced by 2ee9cbf9 ("ACPI: fix
acpi_os_map_memory()"). This contention was observed to result in Dom0
kernel soft lockups during the loading of the ACPI processor driver
there on systems with very many CPU cores.

There are a couple of things being done for this:
- re-order elements of an if() condition so that the register access
  only happens when we really need it
- turn off arbitration disabling only when the first CPU leaves C3
  (paralleling how arbitration disabling gets turned on)
- only set the (global) bus master reload flag once (when the first
  target CPU gets processed)

Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Keir Fraser <keir@xen.org>