On Thursday 26 July 2007 16:55, Linus Torvalds wrote:
...
That three-liner will crash ACPI+SMP-HOTPLUG_CPU kernels on resume.
While cpu0 is in a known state when the power goes out,
without HOTPLUG_CPU the other cpus (and the memory they touch)
are in an indeterminate state.
Yes, we could invent a new mechanism to offline the other CPUS
before suspend and online them upon resume,
but that is what the more general HOTPLUG_CPU code does for us already.
Indeed, that is pretty much _all_ that HOTPLUG_CPU code does on X86 --
as we don't have any physical hotplug support today beneath
this the logical hotplug support -- you could call it CONFIG_CPU_OFFLINE_ONLINE...
Re: warts, I agree.
My question is why the HOTPLUG_CPU=y code is any different.
When I compile HOTPLUG_CPU out of an x86_64 kernel, the kernel shrinks
by only 18KB, which on a kernel that has ACPI+SMP doesn't seem
like such a big wart.
Yes, now that you brought it up, I think it would be just dandy if
HOTPLUG_CPU simply got folded into SMP -- for I see little to no benefit
to having it as its own config option.
But on the assumption that you are not swayed (when was the last time you were?)
and you still feel strongly we should be able to exclude ACPI_SLEEP and HOTPLUG_CPU
from ACPI+SMP kernels, I'll send you a patch do to that properly.
It will largely restores things to how we had them in 2.6.22.
It looks like a step backwards to me, but you may see it differently.
-Len
-
