[PATCH] clocksource: arch_timer: Fix code to use physical timers when requested

[Lorenzo Pieralisi]

On Thu, Sep 04, 2014 at 06:01:27PM +0100, Sonny Rao wrote:

[...]

> > If an OS is booted at PL2 it can access the physical counters, and
> > should do so in case something like KVM will be used later. The OS can
> > write to CNTVOFF at PL2, and if it sets CNTVOFF to zero the physical and
> > virtual counters are equivalent. Thus it can use the virtual counters
> > and doesn't need to have additional code in several places (including
> > the VDSO) where it needs to choose to read which counters to read.
> >
> > The problem only exists where PL2 exists and the firmware/bootloader
> > skipped PL2 without initialising the necessary PL2 state. This is in
> > general a stupid thing to do; it introduces a problem that need not
> > exist and throws away the option of using the features PL2 provides.
> > This is a firmware/bootloader bug.
> 
> Well it's not quite that simple, this is actually an issue with the
> hardware that the CNTVOFF comes up with different values on different
> cores.  This happens not only at boot, but any time the core is
> powered on, which could include deep sleep or CPU hotplug and suspend
> to ram.  The firmware may not be involved in all these cases, so we
> cannot rely on it to fix this problem.

And why isn't firmware involved in those cases if it _is_ involved in
cold boot ? Resuming from low-power means resuming the machine/core as it
was when it was running before power down, anything that deviates from that
behaviour is a programming bug.

Lorenzo