Re: [PATCH v12 09/11] x86/tsc: prepare for early sched_clock

From: Pavel Tatashin
Date: Fri Jun 29 2018 - 05:00:36 EST


On 18-06-29 09:30:10, Thomas Gleixner wrote:
> On Thu, 28 Jun 2018, Pavel Tatashin wrote:
> > On Thu, Jun 28, 2018 at 11:23 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> > Hi Thomas,
> >
> > In addition to above, we have xen hvm:
> >
> > setup_arch()
> > ...
> > init_hypervisor_platform();
> > x86_init.hyper.init_platform();
> > xen_hvm_guest_init()
> > xen_hvm_init_time_ops();
>
> Duh. Missed that completely.
>
> > ...
> > tsc_early_delay_calibrate();
> > tsc_khz = x86_platform.calibrate_tsc(); == xen_tsc_khz()
> > ...
> >
> > Which works early.
> >
> > So, what should we do with xen, which seems to be the only platform
> > that would provide different tsc frequency early and late, because of
> > different calibration method?
>
> Fix it? I have no idea why XEN has two variants of the scheme and I neither
> have a clue why the KVM clock stuff is late.

kvm clock I have fixed in my series.
http://lkml.kernel.org/r/20180621212518.19914-5-pasha.tatashin@xxxxxxxxxx

It was late because, it depended on memblock, which is initialized later,
after e820 ranges are read.

So, after my series kvm clock will work right after
init_hypervisor_platform, the same as with most other hypervisors.

With xen, it is a little more challenging: I do not have a hardware
configured to run xen to test it.

On the other hand, most likely that tsc_early_delay_calibrate() now
fails to calibrate tsc on xen, and calibrates correctly later in
tsc_init(). I doubt, that we can use CPUID method to determine tsc
frequency on xen cpu. Anyways, that would be accidental behavior.

To make xen clock available early, we need HYPERVISOR_shared_info. So, at
least that part of xen_setup_shared_info() must be called/available from
init_hypervisor_platform():

void xen_setup_shared_info(void)
{
set_fixmap(FIX_PARAVIRT_BOOTMAP, xen_start_info->shared_info);

HYPERVISOR_shared_info =
(struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);

xen_setup_mfn_list_list();
...

Which, I have no idea if it is possible, or safe to do.

For now, in my series, I would like check in tsc_early_init() if we are
running xen, and set tsc_khz to 0 if so. Later in tsc_init() we will get
a proper calibration. The question is how to make this check least
intrusive.

Pavel