[PATCH v2] clocksource: arch_timer: Allow the device tree to specify the physical timer

Will,

On Thu, Sep 11, 2014 at 10:07 AM, Will Deacon [off-list ref] wrote:

On Thu, Sep 11, 2014 at 05:59:53PM +0100, Doug Anderson wrote:

quoted

On Thu, Sep 11, 2014 at 9:47 AM, Will Deacon [off-list ref] wrote:

quoted

I'd say `Only supported for ARM' to better match what we've done. Probably
also worth mentioning that this relies on the hypervisor/firmware having set
CNTHCTL.PL1PCEN and CNTHCTL.EL1PCTEN (but assumedly made a mess of CNTVOFF
;) if you want to boot on the non-secure side (e.g. as a guest).

Note that the reset value of CNTHCTL.PL1PCEN and CNTHCTL.PL1PCTEN are
both 1 in my version of the ARM ARM.  On the other hand CNTVOFF is
documented to have an UNKNOWN reset value.  If only ARM had guaranteed
that CNTVOFF started out as 0 (which seems like it would have been
sensible) we wouldn't be in this mess.  :-/

I'm afraid we went the opposite way -- in ARMv8 there are a tiny handful
of EL3 registers that are well-defined out of reset, then the rest of the
system is UNKNOWN. The hardware guys prefer that and it can also be useful
for very low-level debugging (system crashes, do a reset, read out the
state).

Well, I guess if that's the way it is then software will have to deal
with it.  It sure seems like having things default to some state (even
if it's zero) at reset is really nice.  Otherwise you get things where
one 1 of every 10 times you boot the system it behaves differently and
that's hard to track down.  ...of course, then it maybe forces you to
think about really initting all bits in software and that's not
terrible.


-Doug