On 02/06/23 13:54, Peter Zijlstra wrote:
> On Wed, May 24, 2023 at 05:40:47PM +0100, Valentin Schneider wrote:
>>
>> So this bit sent me on a little spelunking session :-)
>>
>> From a control flow perspective the initialization isn't required, but then
>> I looked into the comment and found it comes from the
>> arch_timer_read_counter() definition... Which itself doesn't get used by
>> sched_clock() until the sched_clock_register() below!
>>
>> So AFAICT that comment was true as of
>>
>> 220069945b29 ("clocksource: arch_timer: Add support for memory mapped timers")
>>
>> but not after a commit that came 2 months later:
>>
>> 65cd4f6c99c1 ("arch_timer: Move to generic sched_clock framework")
>>
>> which IIUC made arm/arm64 follow the default approach of using the
>> jiffy-based sched_clock() before probing DT/ACPI and registering a "proper"
>> sched_clock.
>>
>> All of that to say: the comment about arch_timer_read_counter() vs early
>> sched_clock() doesn't apply anymore, but I think we need to keep its
>> initalization around for stuff like get_cycles(). This initialization here
>> should be OK to put to the bin, though.
>
> Something like the below folded in then?
>
Much better, thank you!
