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

[Doug Anderson]

Hi,

On Wed, Aug 27, 2014 at 7:58 PM, Olof Johansson <olof at lixom.net> wrote:
> On Wed, Aug 27, 2014 at 5:56 PM, Stephen Boyd <sboyd at codeaurora.org> wrote:
>> On 08/27/14 15:33, Olof Johansson wrote:
>>> On Wed, Aug 27, 2014 at 3:26 PM, Stephen Boyd <sboyd at codeaurora.org> wrote:
>>>
>>>> Is there any reason why the virtual counter can't be read? Maybe we're
>>>> the hyp and we need to make sure we don't use the virtual timer so that
>>>> the guest can use it, but that doesn't have any effect on the usage of
>>>> the virtual counter for the clocksource.
>>> There are several cases where virtual is unusable -- in particular it
>>> might not have been configured properly (i.e. the phys/virt offset is
>>> at a bad value).
>>>
>>
>> Any specifics? It would be nice to say so in the commit text so that
>> others using such devices know they need this patch. I'm guessing the
>> firmware can't be fixed?

Even if we could change things to use a virtual timer in some cases,
Sonny's patch still fixes a bug.  The code as written right now makes
pretenses about supporting the physical timer, but it doesn't work.
That should be fixed.


> Yeah, there are a few. The big.LITTLE on the Chromebook 2 models have
> this issue, due to the A7 cluster having an incorrect offset
> programmed. However, arch timers aren't supported on that SoC in the
> first place, so it's not a problem in reality.
>
> The other known platform is rk3288. It has products out in the wild
> where firmware updates are unlikely.

One other reason is that (I'm told) that the virtual offset is lost in
certain power down conditions (powering down a core, going into S3,
etc).  When we power back up the offset is effectively reset to a
random value.  That means we need something to reprogram the virtual
timer offset whenever we power things back up.

If we've got a hypervisor then the hypervisor will definitely be
involved in powering things back up and it can reset the virtual
offset.  ...but forcing systems to implement a hypervisor (or somehow
adding an interface for the kernel to call back into firmware) is a
huge effort and it means more hard-to-update code sitting in firmware.

Note: having the virtual offset initted to a random value seems like
an unfortunate design choice for the virtual timer offset
(guaranteeing it was initted to 0 would have avoided the problem), but
that's what we seem to have.


-Doug