/dev/disk0 vs /dev/mmc0

[David Jander]

On Tue, 08 Oct 2013 16:11:16 +0200
Lucas Stach <l.stach at pengutronix.de> wrote:

> Am Dienstag, den 08.10.2013, 15:47 +0200 schrieb David Jander:
> > Dear Lucas,
> > 
> > On Tue, 08 Oct 2013 11:39:24 +0200
> > Lucas Stach <l.stach at pengutronix.de> wrote:
> > 
> > > Am Dienstag, den 08.10.2013, 11:13 +0200 schrieb David Jander:
> > > [...]
> > > > > > You are right. Sorry for the confusion. There were aliases in
> > > > > > barebox's version of imx53.dtsi which went missing somehow in my
> > > > > > tree. I have been merging stuff from mainline Linux....
> > > > > > That leads me to the question: What is the intended relationship
> > > > > > between barebox and Linux kernel DTS files? Which one is supposed
> > > > > > to be the "master" or "upstream" version? Aren't they supposed to
> > > > > > be kept in sync somehow? At least the basic SoC-specific .dtsi
> > > > > > files?
> > > > > > 
> > > > > Missing other options we regard the Linux DTs as the upstream.
> > > > > Ideally DTs should be stored in a separate repo, so they don't get
> > > > > broken unnecessarily.
> > > > > 
> > > > > We will try to keep the Barebox DTs in sync with Linux as much as
> > > > > possible. But even then there is no guarantee that Barebox and Linux
> > > > > DTs don't drift away from each other over time.
> > > > > If a binding gets broken (let's hope that everyone gets their act
> > > > > together and don't do this anymore), it will require code changes and
> > > > > maintenance work to get things in sync again.
> > > > > 
> > > > > Another real-life use-case where you might end up with different DTs
> > > > > is if you update your Linux kernel more regularly than the Barebox
> > > > > on your device. So please if ever possible don't boot the Linux
> > > > > kernel with the builtin Barebox DT, but design your boot concept in
> > > > > a way that you always deliver a DTB fitting your kernel.
> > > > 
> > > > Our boot-scheme checks for a DTB on the boot-medium. If it finds one
> > > > matching the board, it is used, otherwise it boots with barebox's
> > > > built-in DTB, and that should be the default situation, unless it
> > > > becomes necessary to overrule the internal DTB for whatever
> > > > regrettable reason.
> > > > 
> > > Yeah that's the conclusion you may be inclined to take if you try to
> > > live in an ideal world. (Including a dot-shaped earth in a vacuum.)
> > 
> > Aren't we living on a dot-shaped earth in a vacuum?
> > 
> > > Experience with the device tree shows that there are a lot more of those
> > > regrettable reasons than one might imagine. So in order to get around
> > > all those real world problems always boot your kernel with a devicetree
> > > matching exactly that kernel version. Take this as a well-meaning advise
> > > from someone who hit the hard wall of the real-world device tree more
> > > than one.
> > 
> > Don't get me wrong, I have had enough cases of device-trees changing and
> > kernels being incompatible with them. I have hit the hard wall enough times
> > already, but that doesn't mean it hasn't always been for pretty regrettable
> > reasons. Open Firmware and Device-Trees are both good ideas and we are not
> > using them (yet) as they are supposed to be used. Saying that we shouldn't
> > just because we haven't been able to make it work as intended sounds more
> > like an excuse to me than a hard fact.... ;-)
> > Make the problem be the exception, not the rule.
> 
> Without doubt we all should always treat DTs the way we would like them
> to be used in the ideal case, but don't depend on those assumptions for
> your real products.
> 
> Using the builtin DTB normally, but having a way to override it when
> needed sounds like a reasonable compromise to make.

Thanks.

> > > > I have had strong discussions (almost flame-war) long time ago on alkml
> > > > about the issue of hardware initialization. IMNSHO, hardware
> > > > initialization that has to do with the electrical layout of the board
> > > > (PAD settings, drive-strength, etc...) must always be done in the
> > > > boot-loader. Back in the time where there was no DT for ARM, Linux
> > > > kernel board support code was a bloody mess with this sort of HW
> > > > initialization everywhere, and everywhere done differently. Only the
> > > > hardware designer know what drive-strength and other PAD settings need
> > > > to be chosen for each PAD. That is where your design is based on.
> > > > Kernel (-developers) do not know, and should not need to care. Now
> > > > there is this "invention" of putting all this information in the DT.
> > > > This actually seems like a good compromise between what I said above
> > > > and the fact that the kernel sometimes needs to reconfigure PADs due
> > > > to PM, pluggable devices and such. But still the knowledge about PAD
> > > > settings (and what hardware the board happens to have for that matter)
> > > > belongs to the boot-loader. Having more than one source for this is
> > > > dangerous.
> > > > 
> > > No, that is utterly wrong. Linux should always know everything about
> > > your hardware and _not_ expect anything to be set up by the bootloader.
> > 
> > Huh? AFAIK, it has never been like that. PC's have BIOS (or UEFI firmware
> > for that matter) that takes care of signal integrity setup of the PC's
> > chipset, motherboard, CPU and RAM. Next, do you want to leave DDR timing
> > setup to the Linux kernel?? You are kidding, right?
> > 
> And those BIOS settings are so regularly wrong that the kernel has to
> ship a lot of cludge just to work around or correct that bootloader
> setup.

I don't think a BIOS has ever been wrong about chipset drive-strength and
slew-rate settings in a way the Linux kernel could do anything about it....

> I agree that the bootloader should set up things needed for the boot,
> but in many cases the kernel needs the same knowledge anyway. Your DDR
> timing is a pretty prime example for this. The bootloader has to set it
> up initially, but the Linux kernel has to change timings dynamically
> (including calibration in some situations) when doing dynamic memory
> clock scaling.

For DRAM clock scaling, only the clock divider or PLL should ever be
changed. Any other parameters should stay the same. If this has really been
done, I'd like to see it...
DRAM calibration settings have to do with signal propagation on the PCB. That
has nothing to do with clock speeds or -settings.

> CPU bringup is the same matter, Linux needs to know how to do it anyway
> as we need it for CPU hotplug when saving power.

(Un)fortunately we don't have ACPI on ARM ;-)
But then again, CPU bringup is still very different from PAD settings.

> Just because there is a real small amount of static setup that could be
> done exclusively in the bootloader, doesn't mean we should always force
> it down into the bootloader and tell Linux to depend on it.

Yes we should! This "real small" amount as you call it is something you can't
understand even if you had all the documentation of the chips placed on the
board, because it is part of the total hardware design, including the board
layout. And sometimes a crucial part of it I might add. The most ugly thing is
that other settings might even work correctly 95% of the time, so it can be
potentially very hard to find a bug related to "thinking I know what I am
doing by defining some seemingly arbitrary drive-strength settings for this
PAD", or even copying settings from a complete other hardware design based on
the same chip.

Things like this in a dtsi file like imx53.dtsi:

	pinctrl_fec_1: fecgrp-1 {
		fsl,pins = <
			MX53_PAD_FEC_MDC__FEC_MDC	 0x80000000
			MX53_PAD_FEC_MDIO__FEC_MDIO	 0x80000000
			MX53_PAD_FEC_REF_CLK__FEC_TX_CLK 0x80000000
			MX53_PAD_FEC_RX_ER__FEC_RX_ER    0x80000000
			MX53_PAD_FEC_CRS_DV__FEC_RX_DV   0x80000000
			MX53_PAD_FEC_RXD1__FEC_RDATA_1   0x80000000
			MX53_PAD_FEC_RXD0__FEC_RDATA_0   0x80000000
			MX53_PAD_FEC_TX_EN__FEC_TX_EN    0x80000000
			MX53_PAD_FEC_TXD1__FEC_TDATA_1   0x80000000
			MX53_PAD_FEC_TXD0__FEC_TDATA_0   0x80000000
		>;
	};

... are conceptually wrong. They suggest one can universally apply
pinctrl_fec_1 if the FEC happens to be connected to the MX53_PAD_FEC_... pads.
Wrong! Because these settings also sneakily overrule the PAD settings for
drive-strength, slew-rate, pull-/keepers, etc....
There is nothing wrong with having this option, but it may apply only to one
hardware design, and be mostly useless to the rest. Ok, maybe I exaggerate a
little, but I assume you understand what I mean by "conceptually wrong".

The only semantically correct solution I have been able to come up with for
this, is to actually generate this DT in the boot-loader.

> > > There are some corner cases like chip errata fixes here, but generally
> > > don't assume the bootloader to set things up for you.
> > 
> > Another case of making the problem the rule instead of the exception....
> > The fact that board manufacturers don't get their act together to make
> > decent firmware (bootloader) is no excuse for saying it shouldn't be done
> > like that.
> > 
> > > The bootloader
> > > should only do the minimal needed init for loading the kernel and
> > > booting. The kernel is your hardware abstraction for your running
> > > system, not the bootloader.
> > 
> > That is ok. What I am saying is that settings that have strictly to do with
> > _hardware_ constraints should never be bothered by the Linux kernel.
> > Otherwise you would need to recompile the kernel (or change the
> > device-tree) for potentially every single hardware revision of one board
> > model. If I place other RAM chips on the board for instance, or if there
> > is a layout revision, or a peripheral chip that has had a die revision
> > requiring different drive-strength and such changes.... that information
> > should be taken care of by the boot-loader.
> > 
> > Configuring peripherals, setting MUX registers and such is a different
> > story. That should (and many times must) be done in the kernel.
> > 
> > > Ideally both the bootloader and kernel would source their needed
> > > information from a shared DT that's decoupled from kernel development
> > > and stable, once again reality strikes in here. I agree with you that
> > > not a kernel developer, but board designer should write the DT and we
> > > might actually get this separation in a clean way, if we manage to split
> > > out the DTs from the kernel some day.
> > 
> > Well, I am a board designer, and I am trying to do just that. Pardon me for
> > being an idealist sometimes ;-)
> > 
> We can all do our share to better the overall situation by being
> idealist, but avoid basing your decision for real-world products on
> assumptions that are only true for an ideal world. :)

I don't do that of course. But with 99% of all ARM boards having broken
firmware we shouldn't loose our horizon while developing either Barebox or the
Linux kernel :-)

> > Funny enough, some of my gripes about the state of ARM-Linux initially were
> > echoed years later by Linus himself in his famous e-mail about the mess
> > that was ARM-Linux board support before DT came along. Back then, when I
> > asked why ARM wouldn't benefit from OF-device-trees just like PowerPC and
> > Sparc did, I was being laughed at. Unfortunately a few conceptual things
> > have been implemented the wrong way (like PAD setup), but hey! Not bad ;-)

Ok, now can we both please go back to work?

Best regards,

-- 
David Jander
Protonic Holland.