[Devel] Re: [PATCH 1/3] Introduce cpuid_on_cpu() and cpuid_eax_on_cpu()

[Alexey Dobriyan]

On Mon, Apr 02, 2007 at 02:10:29PM +0200, Andi Kleen wrote:
> On Monday 02 April 2007 13:38, Alexey Dobriyan wrote:
> > They will be used by cpuid driver and powernow-k8 cpufreq driver.
> >
> > With these changes powernow-k8 driver could run correctly on OpenVZ kernels
> > with virtual cpus enabled (SCHED_VCPU).
>
> This means openvz has multiple virtual CPU levels?

Not sure what do you mean.

> One for cpuid/rdmsr and one for the rest of the kernel?

Now I think I do. No, only one VCPU level:

	+---+ +---+ +---+
PCPU	| 0 | | 1 | | 2 |
	+---+ +---+ +---+
	  |	\
	  |	 `---,
	  V	      \
VCPU	+---+ +---+ +---+
	| 0 | | 1 | | 2 |
	+---+ +---+ +---+

PCPU chooses VCPU, chooses task from it's runqueue, or becomes idle.

> Both powernow-k8 and cpuid attempt to schedule
> to the target CPU so they should already run there. But it is some other CPU,
> but when they ask your _on_cpu() functions they suddenly get a "real" CPU?
> Where is the difference between these levels of virtualness?

*_on_cpu functions do some work on given physical CPU.
set_cpus_allowed() in openvz operates on VCPU level, so process doing
set_cpus_allowed() still could be scheduled anywhere. Which horribly
breaks cpufreq drivers. To unbreak them, we need a way to do frequency
changing work on given PCPU whose number comes from /sys/*/cpu/cpu*
hierarchy.

> That sounds quite fragile and will likely break often.  I just rejected a similar
> concept -- virtual nodes and "physical nodes" for similar reasons.

Care to drop a link, so I could compare them? Don't recall something
like that posted on l-k.

> Also it has weird semantics. For example if you have multiple
> virtual CPUs mapping to a single CPU then would the powernow-k8 driver
> try to set the frequency multiple times on the same physical CPU?

If core cpufreq locking is OK, why would it?

> That might
> go wrong actually because the CPU might not be happy to be poked again
> while it is already in a frequency change. Also there is no locking
> so in theory two vcpus might try to change frequency in parallel with
> probably quite bad effects.
>
> I'm sure there are other scenarios with similar problems. e.g. what
> happens with microcode updates etc.?

apply_microcode() looks small enough to convert it to IPIs, but so far
nobody asked for microcode updates in openvz.

> Before adding any hacks like this I think your vcpu concept
> needs to be discussed properly on l-k. For me it doesn't look like it is
> something good right now though.

Andi, I think it all relies on correctness of core cpufreq locking. Core
cpufreq locking is not changed with these patches so number of bugs
should stay the same.

current way				new way
-----------				-------

set_cpus_allowed(cpu);
[frequency change part #1]		[IPI part #1]
			<=== nasty here ===>
[frequency change part #2]		[IPI part #2]
set_cpus_allowed(old)