On 20-01-16, 16:20, Stephen Boyd wrote:

Ok, first off, I don't understand why the regulator and clock
pointers are in the struct device_opp instead of the struct
device_list_opp. I thought we wanted to make it possible for two
devices to share the same OPP table (device_opp), but have
physically different clocks and regulators (non opp-shared case)?
If we put the clock and regulator handles in the device_opp then
the opp table is limited to one device or a set of devices that
all share the same clock and regulators.

Not exactly. We wanted devices (with separate rails) to share the OPP
tables ONLY IN DT. So that the same thing isn't required to be copied
multiple times in DT, for example in case of Krait.

But the code isn't supposed to shared device_opp structure at all.
There is one device_opp structure for a groups of devices that share
their OPPs and their clock/voltage rails. opp_device_list is
representing individual devices that share the same device_opp thing.

And this works pretty well and I don't think we should touch it now.

BTW, these structure names confuse me all the time. I have to
rename dev_pm_opp to opp_entry, device_opp to opp_table, and
device_list_opp to opp_device_list in my head for anything to
make sense.

Sure, we can (should) rename them after this series goes in.

Gripes aside, the clock and regulator pointers should never be
'put' and go away until the device driver that's using the
dev_pm_opp_set_rate() API has called dev_pm_opp_remove().

Right.

So any
concerns about that happening during an OPP switch aren't the
concern of the OPP framework, but the concern of the consumer
drivers having the proper locking and tear down sequences.

I am fine with that view as well.. It simplifies lots of things for me
:)

-- 
viresh