Re: [RFC PATCH 1/4] memory: tegra124-emc: Add EMC driver

[Thierry Reding]

On Wed, Jun 18, 2014 at 04:33:39PM -0600, Stephen Warren wrote:
> On 06/18/2014 04:19 PM, StÃphane Marchesin wrote:
> > On Wed, Jun 18, 2014 at 3:00 PM, Thierry Reding
> > <thierry.reding@xxxxxxxxx> wrote:
> >> On Wed, Jun 18, 2014 at 07:23:47PM +0200, Tomeu Vizoso wrote:
> >>> On 06/17/2014 06:15 PM, Stephen Warren wrote:
> >>>> On 06/17/2014 06:16 AM, Tomeu Vizoso wrote:
> >>>>> On 06/16/2014 10:02 PM, Stephen Warren wrote:
> >>>>>> On 06/16/2014 07:35 AM, Tomeu Vizoso wrote:
> >>>>>>> +#ifdef CONFIG_TEGRA124_EMC
> >>>>>>> +int tegra124_emc_reserve_bandwidth(unsigned int consumer, unsigned
> >>>>>>> long rate);
> >>>>>>> +void tegra124_emc_set_floor(unsigned long freq);
> >>>>>>> +void tegra124_emc_set_ceiling(unsigned long freq);
> >>>>>>> +#else
> >>>>>>> +int tegra124_emc_reserve_bandwidth(unsigned int consumer, unsigned
> >>>>>>> long rate)
> >>>>>>> +{ return -ENODEV; }
> >>>>>>> +void tegra124_emc_set_floor(unsigned long freq)
> >>>>>>> +{ return; }
> >>>>>>> +void tegra124_emc_set_ceiling(unsigned long freq)
> >>>>>>> +{ return; }
> >>>>>>> +#endif
> >>>>>>
> >>>>>> I'll repeat what I said off-list so that we can have the whole
> >>>>>> conversation on the list:
> >>>>>>
> >>>>>> That looks like a custom Tegra-specific API. I think it'd be much better
> >>>>>> to integrate this into the common clock framework as a standard clock
> >>>>>> constraints API. There are other use-cases for clock constraints besides
> >>>>>> EMC scaling (e.g. some in audio on Tegra, and I'm sure many on other
> >>>>>> SoCs too).
> >>>>>
> >>>>> Yes, I wrote a bit in the cover letter about our requirements and how
> >>>>> they map to the CCF. Could you please comment on that?
> >>>>
> >>>> My comments remain the same. I believe this is something that belongs in
> >>>> the clock driver, or at the least, some API that takes a struct clock as
> >>>> its parameter, so that drivers can use the existing DT clock lookup
> >>>> mechanism.
> >>>
> >>> Ok, let me put this strawman here to see if I have gotten close to what you
> >>> have in mind:
> >>>
> >>> * add per-client accounting (Rabin's patches referenced before)
> >>>
> >>> * add clk_set_floor, to be used by cpufreq, load stats, etc.
> >>>
> >>> * add clk_set_ceiling, to be used by battery drivers, thermal, etc.
> >>>
> >>> * an EMC driver would collect bandwidth and latency requests from consumers
> >>> and call clk_set_floor on the EMC clock.
> >>>
> >>> * the EMC driver would also register for rate change notifications in the
> >>> EMC clock and would update the latency allowance registers at that point.
> >>
> >> Latency allowance registers are part of the MC rather than the EMC. So I
> >> think we have two options: a) have a unified driver for MC and EMC or b)
> >> provide two parts of the API in two drivers.
> >>
> >> Or perhaps c), create a generic framework that both MC and EMC can
> >> register with (bandwidth for EMC, latency for MC).
> > 
> > Is there any motivation for keeping MC and EMC separate? In my mind,
> > the solution was always to handle those together.
> 
> Well, they are documented as being separate HW modules in the TRM.
> 
> I know there's an interlock in HW so that when the EMC clock is changed,
> the EMC registers can flip atomically to a new configuration.
> 
> I'm not aware of any similar HW interlock between MC and EMC registers.
> That would imply that very tight co-ordination shouldn't be required.
> 
> Do the MC latency allowance registers /really/ need to be *very tightly*
> managed whenever the EMC clock is changed, or is it just a matter of it
> being a good idea to update EMC clock and MC latency allowance registers
> at roughly the same time?

I guess depending on the timing you could get FIFO underflows if the
registers aren't updated promptly enough. Once the programming takes
effect things should be able to catch up again, but it's possible that
there could be glitches.

> Even if there's some co-ordination required,
> maybe it can be handled rather like cpufreq notifications; use clock
> pre-rate change notifications to set MC up in a way that'll work at both
> old/new EMC clocks, and then clock post-rate notifications to the final
> MC configuration?

Yes, I think something like that should work. As I understand it, the
latency allowance is dependent on the EMC frequency, so in case where
the EMC frequency is increased, adjusting in a post-rate notifier should
be fine. When the EMC frequency is decreased, then programming the
latency allowance registers in a pre-rate notifier should allow glitch-
free transition.

Note that this is all purely theoretical knowledge, so I have no idea if
it'll actually work like that in practice.

Thierry

Attachment: pgp6KiuFRlaTJ.pgp
Description: PGP signature