On 01/06/2013 06:53 PM, Eric W. Biederman wrote:
...
> Yes. On x86 we have had the generic equivalent of disable_nonboot_cpus
> in the machine_shutdown for a long time.
It looks like the x86 implementation does a bit more than
disable_nonboot_cpus():
disable_nonboot_cpus():
find first cpu in online cpu mask
disable everything else
x86's machine_shutdown():
default to rebooting on cpu 0
if user specified a different cpu, override default
bring that cpu online
disable everything else
So, x86 always kexec's on a specific CPU, whereas if we use
disable_nonboot_cpus() on ARM, we'll end up kexec'ing on whichever is
the first online CPU, which might not be the actual boot CPU, and can vary.
On Tegra this doesn't (currently?) matter since CPU 0 can't be taken
offline due to our CPU hotplug driver disallowing it. But, perhaps other
SoCs or future Tegra versions don't/won't have this restriction, so the
difference will be material.
Should the x86 code be lifted into the implementation of
disable_nonboot_cpus()?
For the record, here's what I can tell about what the various
arch-specific machine_shutdown() do:
ARM, ARM64: calls smp_send_stop()
-> disable_nonboot_cpus() would be correct
IA64: shuts down all CPUs except the current one
-> disable_nonboot_cpus() would be correct
Microblaze: nothing (but no SMP support?)
-> disable_nonboot_cpus() would be irrelevant, but fine
MIPS: machine-specific:
Cavium Octeon: Shuts down CPUs, waits until num_online_cpus()==1
Not sure which CPU isn't shut down though; the current one?
Others: Nothing (but at least some have SMP support)
-> disable_nonboot_cpus() would be a behaviour change
PPC: machine-specific
Only implementations either aren't for SMP, or do nothing (but
presumably many have SMP support)
-> disable_nonboot_cpus() would be a behaviour change
SH: smp_send_stop()
-> disable_nonboot_cpus() would be correct
S390: nothing (but appears to have SMP support)
-> disable_nonboot_cpus() would be a behaviour change
Tile: nothing (but bans kexec unless no SMP or only 1 CPU online)
-> disable_nonboot_cpus() would be irrelevant, but fine, and perhaps
even allow removal of the kexec ban for SMP?
So at least I don't see anything that would particularly indicate that
having the kexec generic/core code call disable_nonboot_cpus() would
cause problems; many architectures have done something like that
themselves. That said, it certainly would cause some behavioural
differences on some big platforms like PPC...
