On Tue, Dec 13, 2016 at 11:08 PM, Maxime Ripard
[off-list ref] wrote:

Hi,

On Mon, Dec 12, 2016 at 12:16:07PM +0000, Andre Przywara wrote:

quoted

Hi Chen-Yu,

thanks for the answer!

On 12/12/16 04:41, Chen-Yu Tsai wrote:

quoted

On Mon, Dec 12, 2016 at 7:54 AM, Andr? Przywara [off-list ref] wrote:

quoted

Hi,

I was observing that the new sunxi-ng clock code apparently explicitly
turns off _all_ clocks that are not used or needed. I find this rather
unfortunate, as I wanted to use the THS temperature sensor from ARM
Trusted Firmware to implement emergency shutdown or DVFS throttling.
That works until the clock framework finds the clock (as enumerated in
ccu-sun50i-a64.c) and obviously explicitly clears bit 31 in the THS mod
clock register and bit 8 in the respective clock gate register.
Turning them manually back on via /dev/mem or removing the THS clock
from the sunxi-ng driver fixes this for me.

This was not happening with the old Allwinner clocks, since the kernel
wouldn't even know about it if there was no driver and the clock wasn't
mentioned in the DT.

Now with sunxi-ng even though the THS clock is not actually referenced
or used in the DT, the kernel turns it off. I would expect that upon
entering the kernel all unneeded clocks are turned off anyway, so there
is no need to mess with clocks that have no user, but are enumerated in
the ccu driver.

I can't say that's absolutely true (wink).

quoted

I wonder if this kills simplefb as well, for instance, since I believe
that U-Boot needs to turn on certain clocks and relies on them staying up.

The simplefb bindings takes clocks and regulators expressly for the
purpose of keeping them enabled.

Right, I should have checked this before ...

quoted

quoted

So my questions:
1) Is this expected behaviour?

Yes.

quoted

2) If yes, should it really be?
3) If yes, shouldn't there be way to explicitly tell Linux to leave that
clock alone, preferably via DT? Although the sunxi-ng clocks take
control over the whole CCU unit, I wonder if it should really mess with
clocks there are not referenced in the DT.

As it owns the whole CCU unit, why not? And how would it know if some
clock is referenced or not, unless going through the whole device tree?

I was hoping that it just provides clocks to any users (drivers) and
wouldn't bother with tinkering with every clock unless explicitly being
asked for (by a driver being pointed to a specific clock through DT).
So it would need to iterate through anything - neither the whole DT nor
it's own list of clocks.

quoted

Furthermore, nothing prevents another device driver from referencing
said clock and turning it off when it's not in use. Think THS driver
with runtime PM.

I am totally OK with that: Any potential Linux THS driver can take over,
if the DT references this device and the clock.
My point is that atm there is no such driver and so the clocks framework
shouldn't turn that clock off.

You could turn that exact argument the other way though. If there's no
user in the system, why should we waste power and leave it enabled?

quoted

quoted

Are you also mapping the THS to secure only? Otherwise nothing would
prevent Linux from also claiming it.

Unfortunately the THS is always unsecure. And even if it could be
switched, after a recent IRC discussion I came to believe that those
secure peripherals features only works when the secure boot feature is
used, which requires to blow an efuse and thus is not easily doable on
most boards and also irreversible.
Also I am not sure whether this security feature actually extends to the
mod clocks and the bus reset and clock gates bits.

But I was relying on that Linux doesn't touch hardware that's not
referenced in the DT, so if firmware uses the THS, any Linux THS node
would need to go - or the other way around: if there is a Linux THS
node, firmware backs off.

It's not just about node though, but also based on the kernel
configuration. If the kernel didn't have a THS driver compiled (or not
loaded), then if you want to implement such a behaviour, you should
also keep the THS driver in the firmware.

quoted

quoted

quoted

Maybe there is some way to reference those clocks via some dummy driver
or DT node to avoid this behaviour? Is there any prior art in this respect?

If you want a clock to not be disabled by anyone, adding CLK_IS_CRITICAL
to its flags is the proper option. This is done in the clk driver though.

Yes, I was thinking about that, but it indeed sounds like a hack to
follow this.

You also have the option to add a clock-critical property.

This is not supported by the clk core though. Rather, the clk core just
provides the helper function of_clk_detect_critical() to set the flag.
We don't support it either. Furthermore, the function's comment says:

    Do not use this function. It exists only for legacy Device Tree
    bindings, such as the one-clock-per-node style that are outdated.
    Those bindings typically put all clock data into .dts and the Linux
    driver has no clock data, thus making it impossible to set this flag
    correctly from the driver. Only those drivers may call
    of_clk_detect_critical from their setup functions.

ChenYu

Keep in mind that just preventing it from shutting down at boot gives
no warranty that the clock will remain enabled. Other clocks in the
same sub-tree might do a reparenting or a disable that would lead to
that clock being modified or disabled too as a side effect.

Maxime

--
Maxime Ripard, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com