hpet_disable() call sites

hpet_disable() call sites

[Jan Beulich]

In c86c7fbc829e27e2a4093f98ded9fbd75e515adb (and subsequently
0c1b2724069951b1902373e688042b2ec382f68f) hpet_disable() gets
called in the shutdown path. The first of them gives HPET enabling
through PCI quirks in conjunction with the use of legacy replacement
as sole reason, yet even outside of the context of either the starting
up kernel has a problem if the HPET is in an unexpected state, in
particular preventing "normal" timer interrupts from occurring (which
was in particular found to be the case during kdump attempts after
Xen was running).

Is there any reason why hpet_disable() should not also be called
from (or some equivalent action be taken, perhaps including clearing
certain bits in the individual counters' configuration registers, which
are apparently - but perhaps wrongly - implied to be clear in e.g.
hpet_set_mode(), in) hpet_enable()?

Thanks, Jan

Re: hpet_disable() call sites

[Thomas Gleixner]

On Tue, 27 Mar 2012, Jan Beulich wrote:
> up kernel has a problem if the HPET is in an unexpected state, in
> particular preventing "normal" timer interrupts from occurring (which
> was in particular found to be the case during kdump attempts after
> Xen was running).

What's Xen doing special with the hpet ?
 
> Is there any reason why hpet_disable() should not also be called
> from (or some equivalent action be taken, perhaps including clearing
> certain bits in the individual counters' configuration registers, which
> are apparently - but perhaps wrongly - implied to be clear in e.g.
> hpet_set_mode(), in) hpet_enable()?

No, there is no particular reason why we don't clear those registers.

Thanks,

	tglx

Re: hpet_disable() call sites

[Jan Beulich]

>>> On 27.03.12 at 14:58, Thomas Gleixner <tglx@linutronix.de> wrote:
> On Tue, 27 Mar 2012, Jan Beulich wrote:
>> up kernel has a problem if the HPET is in an unexpected state, in
>> particular preventing "normal" timer interrupts from occurring (which
>> was in particular found to be the case during kdump attempts after
>> Xen was running).
> 
> What's Xen doing special with the hpet ?

Just like Linux it may use the legacy routing and fsb delivery bits, and
a (native) kdump kernel subsequently coming up would fail the
timer-connected-to-IOAPIC test if either remained enabled.

>> Is there any reason why hpet_disable() should not also be called
>> from (or some equivalent action be taken, perhaps including clearing
>> certain bits in the individual counters' configuration registers, which
>> are apparently - but perhaps wrongly - implied to be clear in e.g.
>> hpet_set_mode(), in) hpet_enable()?
> 
> No, there is no particular reason why we don't clear those registers.

In that case I'll prepare a patch to do so. One related question is
whether use of the HPET should be suppressed when any bit unknown
to the kernel is found set, or whether unknown bits should also be
cleared.

Jan

Re: hpet_disable() call sites

[Thomas Gleixner]

On Tue, 27 Mar 2012, Jan Beulich wrote:
> >>> On 27.03.12 at 14:58, Thomas Gleixner <tglx@linutronix.de> wrote:
> > On Tue, 27 Mar 2012, Jan Beulich wrote:
> >> Is there any reason why hpet_disable() should not also be called
> >> from (or some equivalent action be taken, perhaps including clearing
> >> certain bits in the individual counters' configuration registers, which
> >> are apparently - but perhaps wrongly - implied to be clear in e.g.
> >> hpet_set_mode(), in) hpet_enable()?
> > 
> > No, there is no particular reason why we don't clear those registers.
> 
> In that case I'll prepare a patch to do so. One related question is
> whether use of the HPET should be suppressed when any bit unknown
> to the kernel is found set, or whether unknown bits should also be
> cleared.

Hmm. Good question. We might at least add a printk to alert about it.

Thanks,

	tglx