Hi Geert,

On Monday 04 August 2014 13:28:32 Geert Uytterhoeven wrote:

On Wed, Jul 30, 2014 at 2:06 AM, Laurent Pinchart wrote:

quoted

The third option would require storing the clocks lists in device drivers.
I believe this is our best option, as a trade-off between simplicity and
versatility. Drivers that use runtime PM already need to enable it
explicitly when probing devices. Passing a list of clock names to runtime
PM at that point wouldn't complicate drivers much. When the clocks list
isn't SoC- dependent it could be stored as static information. Otherwise
it could be derived from DT (or any other source of hardware description)
using C code, offering all the versatility we need.

The only drawback of this solution I can think of right now is that the
runtime PM core couldn't manage device clocks before probing the device.
Specifically device clocks couldn't be managed if no driver is loaded for
that device. I somehow recall that someone raised this as being a
problem, but I can't remember why.

Perhaps you're thinking of clocks that were enabled (by the boot loader or
implicit reset state) before running Linux, and aren't disabled?

That wasn't the reason, I know that clk_disable_unused() takes care of that 
problem (provided the clock drivers behave correctly, the commit you mention 
below shows that's not always the case, but that's an unrelated issue).

That was fixed by commit bb178da701382a230e26d90cf94e8a24b280e0d9
("clk: shmobile: mstp: Fix the is_enabled() operation").

-- 
Regards,

Laurent Pinchart