[PATCH 0/3] Dynamic CPU frequency switching for the HiFive

[Palmer Dabbelt]

On Mon, 06 Jul 2020 01:15:42 PDT (-0700), schwab at suse.de wrote:
> On Jul 03 2020, Yash Shah wrote:
>
>> Yes, you are right. The userspace governor is the only one supported.
>
> That doesn't make sense to me.  How is the userspace governor different
> from any other governor?

It's not.

The issue here is that pretty much everything is on one clock, including all
the cores and the simple IO devices.  These IO devices have programmable clock
dividers that need to mesh up with the input clock rate in order to operate
correctly.  There's no way to do that atomicly WRT the device's clock, so we
have to pause device IO while messing with the clocks.

I bet there's some issue with 4cbd7814bbd5 ("tty: sifive: Finish transmission
before changing the clock") that's causing the serial to get corrupted, and I
guess I'd start with all the issues I pointed out in the patch text :).
Presumably we need to fix the SPI driver as well, in much the same fashion
(though hopefully with fewer hardware surprises to work around).  I'd be way
more worried about ChipLink, as I don't know how that's clocked.  IIRC Ethernet
has its own clock so it should be safe.

I'd be very skeptical about turning on dynamic frequency scaling for the HiFive
Unleashed, as it's probably just not worth chasing the long tail of bugs.  Even
allowing the userspace governor is a bit sketchy, but given that it's a very
early dev board I assume these sorts of problems are expected.

If we do ever turn dynamic frequency scaling on it's not going to default to
allowing 1.4GHz, from what I saw speeds much over 1GHz don't work for most
people.