Re: Solaris dom0 system time jumps backwards, bug in dtrace_xpv_getsystim

I wrote

> John wrote,
> 
> > On Mon, Aug 13, 2007 at 04:00:20AM -0700, Jürgen Keil wrote:
> > 
> > > I don't see that the xen hypervisor guarantees that the reference time 
> > > stamp 
> > > is "recent"; in fact
> > 
> > IIRC the calibration is re-done on every CPU once a second, it seems this
> > should be enough (just?) to avoid any overflow. So I'd be interested to know
> > exactly what's happening when there's a delta larger than 2**32

> I think I'll have to do a few more experiments, e.g. if we ever use this
> "goto out;" in xen/arch/x86/time.c local_time_calibration(). I think it
> could explain why local_tsc_stamp hasn't changed:
> 
> 788      /*
> 789       * Weirdness can happen if we lose sync with the platform timer.
> 790       * We could be smarter here: resync platform timer with local timer?
> 791       */
> 792      if ( ((s64)stime_elapsed64 < (EPOCH / 2)) )
> 793          goto out;

Yep, at some point during the Solaris dom0 boot, it starts to use that
"goto out".  Initially, when I'm stopped in "kmdb -d" it doesn't use that
"goto out".  The platform timer value that is returned by calling
read_platform_stime() stops incrementing at some point; 
"stime_elapsed64 = curr_master_stime - prev_master_stime;" is
always less than 0.5 seconds.   My Tecra S1 is using the
PIT platform timer ((XEN) Platform timer is 1.193MHz PIT).


After the platform timer is dead, there are no more vcpu timestamp
information updates.  (And with my change in dtrace_xpv_getsystime(),
dom0 survives this state)


The platform timer gets broken as soon as Solaris dom0 is enumerating
isa devices, using acpica.  This will initialize the acpica subsystem,
and breaks the PIT timer when the notebook is switched into acpi mode
(i.e. when the "acpi_enable" command byte from the ACPI FADT table
is sent to the "smc_cmd" I/O port).

By booting the Solaris dom0 kernel with "-B acpi-user-options=8" I can
boot dom0 into single user mode without breaking the PIT platform timer.


Question is: why does enabling acpi mode break the PIT timer?

It seems this does not happen on all x86 systems, though.  On another
system, ASUS M2NPV-VM, nVidia chipset, AMD64 X2 AM2 cpu,
there is a BIOS setup option to disable the HPET timer, and with this
setup the xen hypervisor is using the PIT platform timer, too.  But on
this box a Solaris dom0 is booting just fine.
 
 
This message posted from opensolaris.org
_______________________________________________
xen-discuss mailing list
xen-discuss@xxxxxxxxxxxxxxx