Message-ID: <20200515081914.lejcqqwezn3zwpft@gilmour.lan>
Date:   Fri, 15 May 2020 10:19:14 +0200
From:   Maxime Ripard <maxime@...no.tech>
To:     Nicolas Saenz Julienne <nsaenzjulienne@...e.de>
Cc:     Eric Anholt <eric@...olt.net>,
        Tim Gover <tim.gover@...pberrypi.com>,
        Dave Stevenson <dave.stevenson@...pberrypi.com>,
        Stephen Boyd <sboyd@...nel.org>,
        Michael Turquette <mturquette@...libre.com>,
        linux-kernel@...r.kernel.org, dri-devel@...ts.freedesktop.org,
        linux-clk@...r.kernel.org, bcm-kernel-feedback-list@...adcom.com,
        linux-rpi-kernel@...ts.infradead.org,
        Phil Elwell <phil@...pberrypi.com>,
        linux-arm-kernel@...ts.infradead.org
Subject: Re: [PATCH v2 20/91] clk: bcm: rpi: Discover the firmware clocks

Hi Nicolas,

On Mon, May 04, 2020 at 02:05:47PM +0200, Nicolas Saenz Julienne wrote:
> Hi Maxime, as always, thanks for the series!
> Some extra context, and comments below.
> 
> On Fri, 2020-04-24 at 17:34 +0200, Maxime Ripard wrote:
> > The RaspberryPi4 firmware actually exposes more clocks than are currently
> > handled by the driver and we will need to change some of them directly
> > based on the pixel rate for the display related clocks, or the load for the
> > GPU.
> > 
> > This rate change can have a number of side-effects, including adjusting the
> > various PLL voltages or the PLL parents. The firmware will also update
> > those clocks by itself for example if the SoC runs too hot.
> 
> To complete this:
> 
> RPi4's firmware implements DVFS. Clock frequency and SoC voltage are
> correlated, if you can determine all clocks are running below a maximum then it
> should be safe to lower the voltage. Currently, firmware actively monitors arm,
> core, h264, v3d, isp and hevc to determine what voltage can be reduced to, and
> if arm increases any of those clocks behind the firmware's back, when it has a
> lowered voltage, a crash is highly likely.
> 
> As pointed out to me by RPi foundation engineers hsm/pixel aren't currently
> being monitored, as driving a high resolution display also requires a high core
> clock, which already sets an acceptable min voltage threshold. But that might
> change if needed.
> 
> Ultimately, from the DVFS, the safe way to change clocks from arm would be to
> always use the firmware interface, which we're far from doing right now. On the
> other hand, AFAIK not all clocks have a firmware counterpart.
> 
> Note that we could also have a word with the RPi foundation and see if
> disabling DVFS is possible (AFAIK it's not an option right now). Although I
> personally prefer to support this upstream, aside from the obvious benefits to
> battery powered use cases, not consuming power unnecessarily is always big
> plus.
> 
> > In order to make Linux play as nice as possible with those constraints, it
> > makes sense to rely on the firmware clocks as much as possible.
> 
> As I comment above, not as simple as it seems. I suggest, for now, we only
> register the clocks we're going to use and that are likely to be affected by
> DVFS. hsm being a contender here.
> 
> As we'd be settling on a hybrid solution here, which isn't ideal to say the
> least, I'd like to gather some opinions on whether pushing towards a fully
> firmware based solution is something you'd like to see.

Thanks for the summary above, I'll try to adjust that commit log to reflect
this. As for my opinion, I don't really think that an hybrid approach is
practical, since that would leave us with weird interactions between the
firmware (and possibly multiple versinos of it) and the linux driver, and this
would be pretty hard to maintain in the long run.

That leaves us either the MMIO-based driver or the firmware-based one, and here
with what you said above, at the moment, the firmware based one is a clear
winner.

> > Fortunately,t he firmware has an interface to discover the clocks it
> > exposes.
> 
> nit: wrongly placed space
> 
> > Let's use it to discover, register the clocks in the clocks framework and
> > then expose them through the device tree for consumers to use them.
> >
> > Cc: Michael Turquette <mturquette@...libre.com>
> > Cc: Stephen Boyd <sboyd@...nel.org>
> > Cc: linux-clk@...r.kernel.org
> > Signed-off-by: Maxime Ripard <maxime@...no.tech>
> > ---
> >  drivers/clk/bcm/clk-raspberrypi.c          | 104 +++++++++++++++++++---
> 
> [...]
> 
> > +static struct clk_hw *raspberrypi_clk_register(struct raspberrypi_clk *rpi,
> > +					       unsigned int parent,
> > +					       unsigned int id)
> > +{
> > +	struct raspberrypi_clk_data *data;
> > +	struct clk_init_data init = {};
> > +	int ret;
> > +
> > +	if (id == RPI_FIRMWARE_ARM_CLK_ID) {
> > +		struct clk_hw *hw;
> > +
> > +		hw = raspberrypi_register_pllb(rpi);
> > +		if (IS_ERR(hw)) {
> > +			dev_err(rpi->dev, "Failed to initialize pllb, %ld\n",
> > +				PTR_ERR(hw));
> > +			return hw;
> > +		}
> > +
> > +		return raspberrypi_register_pllb_arm(rpi);
> > +	}
> 
> We shouldn't create a special case for pllb. My suggestion here is that we
> revert the commit that removed pllb from the mmio driver, in order to maintain
> a nice view of the clock tree, and register this as a regular fw-clk.
> 
> The only user to this is RPi's the cpufreq driver, which shouldn't mind the
> change as long as you maintain the clk lookup creation.

Ok, I'll change that.

> On that topic, now that the clocks are available in DT we could even move
> raspberrypi-cpufreq's registration there. But that's out of the scope of this
> series.
> 
> > +
> > +	data = devm_kzalloc(rpi->dev, sizeof(*data), GFP_KERNEL);
> > +	if (!data)
> > +		return ERR_PTR(-ENOMEM);
> > +	data->rpi = rpi;
> > +	data->id = id;
> > +
> > +	init.name = devm_kasprintf(rpi->dev, GFP_KERNEL, "fw-clk-%u", id);
> 
> I'd really like more descriptive names here, sadly the firmware interface
> doesn't provide them, but they are set in stone nonetheless. Something like
> 'fw-clk-arm' and 'fw-clk-hsm' comes to mind (SCMI does provide naming BTW).
> 
> See here for a list of all clocks:
> https://github.com/raspberrypi/firmware/wiki/Mailbox-property-interface#clocks

That's a good idea, I'll add that too.

Thanks!
Maxime

Download attachment "signature.asc" of type "application/pgp-signature" (229 bytes)

Powered by blists - more mailing lists