On 19/07/2016 at 16:56:56 -0700, Andrey Smirnov wrote :

quoted

quoted

I don't see any value in doing that, could you give me a realistic
example of a scenario in which a user would want to spend some of
uptime with RTC oscillator fault detection/glitch filtering disabled
and then enable it?

Well, the issue is not being dynamic, it is differentiating between
hardware description and user configuration. Configuration must not be in
DT.

Why? And I don't mean in a generic sense, but in this particular case.
What is gained by not having this bit of configuration, whose only
consumer is the driver, in the device tree file?

Because configuration doesn't belong to DT. DT is about hardware
description, not configuration.

quoted

And this choice is definitively not hardware related (as opposed to
the trickle charging that depends on the battery that is used on the
board).

There's most certainly plenty of precedents of non hardware-related in
device tree, first two that come to mind are "chosen" node and
"local-mac-address" property and, granted, those might be
exceptions/legacy bindings that are just there to stay, but even if
you look at RTC subsystem rtc-cmos.txt, atmel,at91sam-rtc.txt and
possibly rtc-st-lpc.txt are providing bindings that are not exactly
hardware related.

Rtc-cmos.txt is especially noticeable example since it literally does
what I am trying to do -- allows the user to specify initial values to
certain registers that would be initialized by the driver.

Well, the RTC subsystem has been left unmaintained for a while and it is
not because we made mistakes in the past that we have to make them
again.

rtc-st-lpc is only hardware related. The mode it is used in is board
dependant. And I have a plan to change how the gpbr register is
allocated for the at91 RTT. I agree that rtc-cmos is an example of what
not to do.

quoted

Well, it doesn't leak abstraction as long as all the RTC that are able
to disable the oscillator failure detection use the same ABI.

Correct me if I am wrong, but no such, at least semi-standardized, ABI
exist as of today, correct? If that is so, then what you are saying is
that the abstraction doesn't leak as long as you put it inside of a
new abstraction that doesn't leak. I am not arguing that it is
impossible to create a new one that would allow to hide hardware
differences, I am positive it is, what I am arguing is that to do so
is a lot of work for as far as I can see for no gain.

You are correct, that ABI doesn't exist and I'm asking to create it.
That is how kernel development happens.

quoted

quoted

quoted

 - on subsequent reboots, the mode is kept alongside the time and date

This assumes that your bootloader leaves those mode bits alone.

Well, if that is not the case, the bootloader as to be fixed anyway and
silently changing the configuration back using DT is probably worse.

How so? Consider the following two scenarios with assumption that the
bootloader is broken:

  - Bits set wrong by bootloader, then corrected by kernel, device is
powered off RTC consumes expected amount of current

  - Bits set wrong by bootloader, kernel does nothing, device is
powered off RTC consumes more than anticipated and we drain the power
storage device and loose time.

What do you you think former is worse than latter?

Whether is is done in the kernel or in userspace doesn't change much
to that use case.



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