Re: [PATCH 0/2] clockevents: Hide CLOCK_EVT_STATE_* from rest of the kernel

[Viresh Kumar <viresh.kumar@linaro.org>]

On 20-05-15, 15:09, Thomas Gleixner wrote:
> On Wed, 20 May 2015, Viresh Kumar wrote:

> They look at clock_event_mode and not at state, right?

Yeah, it was all useful (that's what I thought initially, but not
anymore) only when we migrate some drivers to the new per-state APIs.

> And that way we add the overhead of a full function call to those
> drivers for the interrupt hot path?

Honestly, I didn't realize that this can be a blocker. My bad.

> Of course you should have done that analysis before posting some
> random helper functions.

I did looked at all the drivers few days back, but failed to give a
summary similar to yours. No excuses.

> Lets look how useful these functions are for the various use cases
> 
> #1) Adds function call over head to the timer interrupt
> 
>     Hot path does matter and that function call is a regression. So
>     that's a NONO

> Now explain me how your magic functions help. For most of the cases
> they would be a performance regression. And for the rest they really
> do not matter at all.

They wouldn't help at all in that case.

So, probably we are left with following choices:

- Maintain state internally within the driver. SMP cases need per-cpu
  storage as clkevt devices are per-cpu. Probably that's a NONO as
  well ?

- Use CLK_EVT_STATE_* directly in drivers (similar to the way we use
  CLK_EVT_MODE_* today).

- Write the routines I proposed as macros or inline functions in
  clockchips.h, and use them. Of course that wouldn't stop exposing
  CLK_EVT_STATE_* to rest of the kernel.

- Something else ?

Which one do you suggest ?

-- 
viresh