this message came up on debian-arm and i figured that it is worthwhile
endeavouring to get across to people why device tree cannot and will
not ever be the solution it was believed to be, in the ARM world.
[just a quick note to david who asked this question on the debian-arm
mailing list: any chance you could use replies with plaintext in
future? converting from HTML to text proved rather awkward and
burdensome, requiring considerable editing. the generally-accepted
formatting rules for international technical mailing lists are
plaintext only and 7-bit characters]
On Sun, May 5, 2013 at 11:14 AM, David Goodenough
<david.goodenough@btconnect.com> wrote:
On Sunday 05 May 2013, Luke Kenneth Casson Leighton wrote:
And I have a question: as the Debian installer takes the arch armhf in
charge, do you think a standard install' from a netboot image will work
?
this has been on my list for a loooong time. as with *all* debian
installer images however you are hampered by the fact that there is no
BIOS - at all - on ARM devices - and therefore it is impossible to
have a "one size fits all" debian installer.
I wonder if the device tree is the answer here. If the box comes with
a DT or one is available on the web then the installer could read it and
know what to install. That and the armmp kernel should solve the problem.
you'd think so, and it's a very good question, to which the answer
could have been and was predicted to be "not a snowbal in hell's
chance", even before work started on device tree, and turns out to
*be* "not a snowball in hell's chance" which i believe people are now
beginning to learn, based on the ultra-low adoption rate of device
tree in the ARM world (side-note: [*0]).
in the past, i've written at some length as to why this is the case,
however the weighting given to my opinions on linux kernel strategic
decision-making is negligeable, so as one associate of mine once
wisely said, "you just gotta let the train wreck happen".
device tree was designed to take the burden off of the linux kernel
due to proliferation of platform-specific hard-coding of support for
peripherals. however it was designed ***WITHOUT*** its advocates
having a full grasp of the sheer overwhelming diversity of the
platforms.
specifically i am referring to linus torvald's complete lack of
understanding of the ARM linux kernel world, as his primary experience
is with x86. in his mind, and the minds of those people who do not
understand how ARM-based boxes are built and linux brought up on them,
*surely* it cannot be all that complicated, *surely* it cannot be as
bad as it is, right?
what they're completely missing is the following:
* the x86 world resolves around standards such as ACPI, BIOSes and
general-purpose dynamic buses.
* ACPI normalises every single piece of hardware from the perspective
of most low-level peripherals.
* the BIOS also helps in that normalisation. DOS INT33 is the
classic one i remember.
* the general-purpose dynamic buses include:
- USB and its speed variants (self-describing peripherals)
- PCI and its derivatives (self-describing peripherals)
- SATA and its speed variants (self-describing peripherals)
exceptions to the above include i2c (unusual, and taken care of by
i2c-sensors, which uses good heuristics to "probe" devices from
userspace) and the ISA bus and its derivatives such as Compact Flash
and IDE. even PCMCIA got sufficient advances to auto-identify devices
from userspace at runtime.
so as a general rule, supporting a new x86-based piece of hardware is
a piece of piss. get datasheet or reverse-engineer, drop it in, it's
got BIOS, ACPI, USB, PCIe, SATA, wow big deal, job done. also as a
general rule, hardware that conforms to x86-motherboard-like layouts
such as the various powerpc architectures are along the same lines.
so here, device tree is a real easy thing to add, and to some extent a
"nice-to-have". i.e. it's not really essential to have device tree on
top of something where 99% of the peripherals can describe themselves
dynamically over their bus architecture when they're plugged in!
now let's look at the ARM world.
* is there a BIOS? no. so all the boot-up procedures including
ultra-low-level stuff like DDR3 RAM timings initialisation, which is
normally the job of the BIOS - must be taken care of BY YOU (usually
in u-boot) and it must be done SPECIFICALLY CUSTOMISED EACH AND EVERY
SINGLE TIME FOR EVERY SINGLE SPECIFIC HARDWARE COMBINATION.
* is there ACPI present? no. so anything related to power
management, fans (if there are any), temperature detection (if there
is any), all of that must be taken care of BY YOU.
* what about the devices? here's where it becomes absolute hell on
earth as far as attempting to "streamline" the linux kernel into a
"one size fits all" monolithic package.
the classic example i give here is the HTC Universal, which was a
device that, after 3 years of dedicated reverse-engineering, finally
had fully-working hardware with the exception of write to its on-board
NAND. the reason for the complexity is in the hardware design, where
not even 110 GPIO pins of the PXA270 were enough to cover all of the
peripherals, so they had to use a custom ASIC with an additional 64
GPIO pins. it turned out that *that* wasn't enough either, so in
desperation the designers used the 16 GPIO pins of the Ericsson 3G
Radio ROM, in order to do basic things like switch on the camera flash
LED.
the point is: each device that's designed using an ARM processor is
COMPLETELY AND UTTERLY DIFFERENT from any other device in the world.
when i say "completely and utterly different", i am not just talking
about the processor, i am not just talking about the GPIO, or even the
buses: i'm talking about the sensors, the power-up mechanisms, the
startup procedures - everything. one device uses GPIO pin 32 for
powering up and resetting a USB hub peripheral, yet for another device
that exact same GPIO pin is used not even as a GPIO but as an
alternate multiplexed function e.g. as RS232 TX pin!
additionally, there are complexities in the bring-up procedure for
devices, where a hardware revision has made a mistake (or made too
many cost savings), and by the skin of their teeth the kernel
developers work out a bring-up procedure. the example i give here is
the one of the HTC Blueangel, where the PXA processor's GPIO was used
(directly) to power the device. unfortunately, there simply wasn't
enough current. but that's ok! why? because what they did was:
* bring up the 3.3v GPIO (power to the GSM chip)
* bring up the 2nd 3.3v GPIO
* pull the GPIO pin connected to the GSM "reset" chip
* wait 5 milliseconds
* **PULL EVERYTHING BACK DOWN AGAIN**
* wait 1 millisecond
* bring up the 1st 3.3v GPIO (again)
* wait 10 milliseconds
* bring up the 2nd 3.3v GPIO (again)
* wait 5 milliseconds
* pull up the "RESET" GPIO
* wait 10 milliseconds
* pull the "RESET" GPIO down
* ***AGAIN*** do the reset GPIO.
this procedure was clearly designed to put enough power into the
capacitors of the on-board GSM chip so that it could start up (and
crash) then try again (crash again), and finally have enough power to
not drain itself beyond its capacity.
... the pointed question is: how the bloody hell are you going to
represent *that* in "device tree"??? and why - even if it was
possible to do - should you burden other platforms with such an insane
boot-up procedure even if they *did* use the exact same chipset?
... but these devices, because they are in a huge market with
ever-changing prices and are simply overwhelmed with choice for
low-level I2C, I2S devices etc, each made in different countries, each
with their NDAs, simply don't use the same peripheral chips. and even
if they did, they certainly don't use them in the same way!
again, the example that i give here is of the Phillips UDA1381 which
was quite a common sound IC used in Compaq iPAQ PDAs (designed by
HTC). so, of course, when HTC did the Himalaya, they used the same
sound IC.
.... did they power it up in the exact same way across both devices?
no.
did they even use the same *interfaces* across both devices?
no.
why not?
because the UDA1381 can be used *either* in I2S mode *or* in SPI mode,
and one [completely independent] team used one mode, and the other
team used the other.
so when it came to looking at the existing uda1381.c source code, and
trying to share that code across both platforms, could i do that?
no.
then - as if that wasn't enough you also have the diversity amongst
the ARM chips themselves. if you look for example at the history of
the development of the S3C6410, then the S5PC100 and the S5PC110, it
just doesn't make any sense... *until* you are aware that my associate
was the global director of R&D at samsung and he instigated the
procedure of having two overlapping teams, one who does development
and the other does testing; they swap every 8-9 months.
basically what happened was that the S3C6410 used FIMD 3D GPU, and so
did the 800mhz S5PC100. by the time the S5PC110 came out (which was
nothing more than a 1ghz jump) it had been flipped over to a
*completely* different 3D engine! changing the 3D GPU in mid-flow on
a CPU series!
and that's just within *one* of the fabless semiconductor companies,
and you have to bear in mind that there are *several hundred* ARM
licensees. when this topic was last raised, someone mentioned that
ARM attempted to standardise on dynamic peripheral publication on the
internal AXI/AHB bus, so that things like device tree or udev could
read it. what happened? some company failed to implement the
standard properly, and it was game over right there.
the best that device tree does in the ARM world is add an extra burden
for device development, because there is so little that can actually
be shared between disparate hardware platforms - so much so that it is
utterly hopeless and pointless for a time-pressured product designer
to even consider going down that route when they *know* it's not going
to be of benefit for them.
you also have to bear in mind that the SoC vendors don't really talk
to each other. you also have to bear in mind that they are usually
overwhelmed by the ignorance of the factories and OEMs that use their
SoCs - a situation that's not helped in many cases by their failure to
provide adequate documentation [but if you're selling 50 million SoCs
a year through android and the SoC is, at $7.50, a small part of the
BOM, why would you care about or even answer requests for adequate
documentation??] - so it's often the SoC vendors that have to write
the linux kernel source code themselves. MStar Semi take this to its
logical GPL-violating extreme by even preventing and prohibiting
*everyone* from gaining access to even the *product* designs. if they
like your idea, they will design it for you - in total secrecy - from
start to finish. and if not, you f*** off.
so at worst, device tree becomes a burden on the product designers
when using ARM processors, because of the sheer overwhelming
diversity. at best... no, there is no "best". device tree just moves
the problem from the linux kernel source code into the device tree
specifications.
possible solutions:
* can a BIOS help? no, because you will never get disparate ARM SoC
licensees to agree to use it. their SoCs are all highly-specialised.
and, why would they add extra layers of complexity and function cals
in a memory-constrained, CPU-constrained and power-constrained
specialist environment?
* can ACPI help? no, because power management is done in a completely
different way, and low-level peripherals are typically connected
directly to the chip and managed directly. ACPI is an x86 solution.
* can splitting CPUs into northbridge architectural designs help? no,
because again that's an x86 solution. and it usually adds 2 watts to
the power budget to drive the 128-bit or 256-bit-wide bus between the
two chips. which is insane, considering that some ARM SoCs are 0.1 to
1 watts. Not Gonna Happen. plus, it doesn't help you with the
existing mess.
* can hardware standardisation help? yes [*1] but that has its own
challenges, as well as additional costs which need to have some sort
of strategic benefit before companies will adopt it [or users].
* can chip-level standardisation help? no - ARM tried that already,
one SoC vendor screwed it up (failed to implement the tables properly)
and that was the end of the standard. plus, it doesn't help with the
existing SoCs nor the NDA situation wrt all the sensors and
peripherals nor the proliferation and localisation and
contact-relationships associated with these hard-wired sensors and
peripherals.
* what about splitting up the linux kernel into "core" and
"peripheral"? that might do it. something along the lines of OSKit -
except done right, and along lines that are acceptable to and
initiated by the key linux kernel developers. the "core" could be as
little as 2% of the entire linux kernel code-base: lib/* arch/*,
kernel/* and so on. the rest (peripherals) done as a gitmodule.
whilst it doesn't *actually* solve the problem, it reduces and
clarifies the development scope on each side of the fence,
so-to-speak, and, critically, separates out the discussions and focus
for each part. perhaps even go further and have forks of the "core"
which deal *exclusively* with each architecture.
i don't know. i don't know - honestly - what the solution is here,
and perhaps there really isn't one - certainly not at the software
level, not with the sheer overwhelming hardware diversity. linus
"ordered" the ARM representatives at that 2007 conference to "go away,
get themselves organised, come back with only one representative" and
that really says it all, the level of naivete and lack of
understanding of the scope, and the fact that each ARM processor
*really is* a completely different processor: they might as well have
completely different instruction sets and it's just damn lucky that
they do.
so i'm attempting to solve this from a different angle: the hardware
level. drastic and i mean *drastic* simplification of the permitted
interfaces, thus allowing hardware to become a series of dynamic buses
plus a series of about 10 device tree descriptions.
so, in answer to your question, david: *yes* device-tree could help,
if the hardware was simple enough and was standardised, and that's
what i'm working towards, but the EOMA standards are the *only* ones
which are simple enough and have taken this problem properly into
consideration. if you don't use hardware standards then the answer
is, sadly i feel, that device tree has absolutely no place in the ARM
world and will only cause harm rather than good [*2]
thoughts on a postcard....
l.
[*0] allwinner use "script.fex". it's an alternative to device-tree
and it's a hell of a lot better. INI config file format.
standardised across *all* hardware that uses the A10. it even allows
setting of the pin mux functions. _and_ there is a GUI application
which helps ODMs to customise the linux kernel for their clients...
WITHOUT RECOMPILING THE LINUX KERNEL AT ALL. the designers of
devicetree could learn a lot from what allwinner have achieved.
[*1] http://elinux.org/Embedded_Open_Modular_Architecture/EOMA-68
[*2] that's not quite true. if there were multiple ARM-based systems
that conformed to x86 hardware standards such as ITX motherboard
layout, those ARM systems would actually work and benefit from
device-tree. if aarm64 starts to get put regularly into x86-style
hardware form-factors, that would work. but right now, and for the
forseeable future, any ARM SoC that's used for "embedded" single-board
purposes, flat-out forget it. and even when aarm64 comes out, there
will still be plenty of cheap 32-bit ARM systems out there.