On 06/18/2014 05:14 PM, Thierry Reding wrote:

On Wed, Jun 18, 2014 at 04:09:06PM -0600, Stephen Warren wrote:

quoted

On 06/18/2014 04:03 PM, Thierry Reding wrote:

...

quoted

quoted

From what I remember, Mike was fairly strongly opposing the idea of
virtual clocks, but what you're proposing here sounds like it would
assume the existence of virtual clocks. clk_set_rate() per client
doesn't work with the current API as I understand it.

Or perhaps what you're proposing isn't about the individual clocks at
all but rather about a mechanism to express constraints for a set of
clocks?

This doesn't have anything to do with virtual clocks. As you mention,
it's just about constraints.

One user of clock "cpu" wants min rate 216MHz. Another wants max rate
1GHz. cpufreq will request some rate between the 2, or be capped to
those limits. These set of imposed constraints would need to be stored
per client of the clock, not per HW clock, since many clients could set
different max rates (e.g. thermal throttle 1.5GHz due to temperature,
CPU policy 1GHz due to the user selecting low CPU power, etc.)

Similarly for audio, of there are N clients of 1 clock/PLL, and they
each want the PLL to run at a different rate, something needs to detect
that and deny it.

I'm wondering how this should work with the current API. Could the clock
core be modified to return a per-client struct clk * that references the
hardware clock internally? Or do we need to add a new API?

I would assume the we can just change struct clk and hide the details
from any driver. Hopefully only clock-core and clock-drivers would need
any changes.