RISC-V U-Boot for HiFive Unleashed Board is available on GitHub

1,286 views
Skip to first unread message

Padmarao Begari

unread,
Oct 11, 2018, 7:10:24 AM10/11/18
to RISC-V SW Dev
RISC-V U-Boot(64-bit) for HiFive Unleashed Board is available on GitHub
                    

This repository contains the RISC-V U-Boot created by Microsemi a Microchip company for RISC-V core on HiFive Unleashed 540 device. The HiFive Unleashed Board is used for HiFive U-Boot development

Regards
Padmarao

Christoph Hellwig

unread,
Oct 11, 2018, 9:24:58 AM10/11/18
to Padmarao Begari, RISC-V SW Dev
Can you explain a bit more where U-Boot fits into the picture? Is it
started as an S-mode payload, or have you ported M-mode code to U-boot?

atish patra

unread,
Oct 12, 2018, 2:52:30 AM10/12/18
to h...@lst.de, padma...@gmail.com, RISC-V SW Dev
I think the uboot runs in M mode and acts as a FSBL in this case.
The boot flow still require bbl image (bbl+vmlinux) to boot Linux.

At least. that's what I understood from the makefile changes. Padmarao can correct me if I am wrong.
sgdisk --clear                                                               \
           --new=1:2048:4095   --change-name=1:uboot      --typecode=1:$(FSBL)   \
           --new=2:4096:69631  --change-name=2:bootloader --typecode=2:$(BBL)   \
           --new=3:264192:     --change-name=3:root       --typecode=3:$(LINUX) \

--
You received this message because you are subscribed to the Google Groups "RISC-V SW Dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sw-dev+un...@groups.riscv.org.
To post to this group, send email to sw-...@groups.riscv.org.
Visit this group at https://groups.google.com/a/groups.riscv.org/group/sw-dev/.
To view this discussion on the web visit https://groups.google.com/a/groups.riscv.org/d/msgid/sw-dev/20181011132455.GA7242%40lst.de.


--
Regards,
Atish

Padmarao Begari

unread,
Oct 12, 2018, 3:35:54 AM10/12/18
to RISC-V SW Dev, h...@lst.de, padma...@gmail.com
The SiFive FSBL is replaced with U-Boot and stores in SD card and runs in M-mode at present.
The u-boot requires bbl image (bbl+vmlinux) to boot Linux at present.

Our next step is to change M-mode to S-mode in u-boot and removing bbl from  Linux image (bbl+vmlinux). 

Regards
Padmarao

atish patra

unread,
Oct 12, 2018, 3:48:51 AM10/12/18
to padma...@gmail.com, RISC-V SW Dev, h...@lst.de
Hi Padmarao,
Thanks for the confirmation. 

On Fri, Oct 12, 2018 at 12:35 AM Padmarao Begari <padma...@gmail.com> wrote:
The SiFive FSBL is replaced with U-Boot and stores in SD card and runs in M-mode at present.
The u-boot requires bbl image (bbl+vmlinux) to boot Linux at present.

Our next step is to change M-mode to S-mode in u-boot and removing bbl from  Linux image (bbl+vmlinux). 


Awesome. Does that mean you are planning to implement SBI as well in uboot or bbl continue to exist in M mode ?
Can you please explain the boot flow design if u-boot runs in S-mode ?

Regards,
Atish
 

Anup Patel

unread,
Oct 12, 2018, 3:58:28 AM10/12/18
to padma...@gmail.com, sw-...@groups.riscv.org, Christoph Hellwig
Yes, thats better.

U-boot should run in S-mode only so that we can later on use same
U-boot as bootloader for Guest/VM.

In fact, it should also be possible run U-boot on QEMU Virt machine
using BBL (i.e. using bbl+uboot).

Regards,
Anup
> To view this discussion on the web visit https://groups.google.com/a/groups.riscv.org/d/msgid/sw-dev/dfaca48a-b9b5-4408-94a2-c26aa8523874%40groups.riscv.org.

Anup Patel

unread,
Oct 12, 2018, 4:05:16 AM10/12/18
to atis...@gmail.com, padma...@gmail.com, sw-...@groups.riscv.org, Christoph Hellwig
On Fri, Oct 12, 2018 at 1:18 PM atish patra <atis...@gmail.com> wrote:
>
> Hi Padmarao,
> Thanks for the confirmation.
>
> On Fri, Oct 12, 2018 at 12:35 AM Padmarao Begari <padma...@gmail.com> wrote:
>>
>> The SiFive FSBL is replaced with U-Boot and stores in SD card and runs in M-mode at present.
>> The u-boot requires bbl image (bbl+vmlinux) to boot Linux at present.
>>
>> Our next step is to change M-mode to S-mode in u-boot and removing bbl from Linux image (bbl+vmlinux).
>>
>
> Awesome. Does that mean you are planning to implement SBI as well in uboot or bbl continue to exist in M mode ?
> Can you please explain the boot flow design if u-boot runs in S-mode ?

Please no M-mode code in U-boot or any general purpose bootloader.

Instead of replacing SiFive FSBL with U-Boot, I suggest to boot
bbl+uboot.bin from SiFive FSBL

U-boot can use SBI services just like any other S-mode software.

U-boot should boot RAW Linux image (instead of vmlinux) using bootm.

The role of last stage bootloader (such as U-boot, UEFI, Grub, etc) is
to support booting protocol for
variety of OSes (Linux, BSD, Vxworks, etc). There should be no
restriction on what last stage
bootloader will be except that it should run in S-mode. For any M-mode
facilities, it should use
SBI calls just like Linux.

Regards,
Anup

Christoph Hellwig

unread,
Oct 12, 2018, 4:08:42 AM10/12/18
to Anup Patel, atis...@gmail.com, padma...@gmail.com, sw-...@groups.riscv.org, Christoph Hellwig
On Fri, Oct 12, 2018 at 01:35:02PM +0530, Anup Patel wrote:
> > Awesome. Does that mean you are planning to implement SBI as well in uboot or bbl continue to exist in M mode ?
> > Can you please explain the boot flow design if u-boot runs in S-mode ?
>
> Please no M-mode code in U-boot or any general purpose bootloader.

Indeed, that is what I feared :(

> Instead of replacing SiFive FSBL with U-Boot, I suggest to boot
> bbl+uboot.bin from SiFive FSBL

That would be more useful indeed.

ron minnich

unread,
Oct 12, 2018, 11:57:53 AM10/12/18
to Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, sw-...@groups.riscv.org
I'm surprised to see this idea but in any event coreboot replaces bbl and has m mode code and I don't expect that to change. I'm done with trusting vendor supplied fsbl no matter what the source.

--
You received this message because you are subscribed to the Google Groups "RISC-V SW Dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sw-dev+un...@groups.riscv.org.
To post to this group, send email to sw-...@groups.riscv.org.
Visit this group at https://groups.google.com/a/groups.riscv.org/group/sw-dev/.

Paul Walmsley

unread,
Oct 12, 2018, 12:48:08 PM10/12/18
to Anup Patel, atis...@gmail.com, padma...@gmail.com, sw-...@groups.riscv.org, Christoph Hellwig
Many embedded users will have no desire to run any external firmware
beyond an open-source bootloader and the kernel.

No one should be discouraging that use-case, as it reflects the needs of
a significant chunk of the user base.

That's worked fine on ARM for many years, and it's the same approach
that folks should keep in mind for RISC-V, even if people might not
understand why others want that.


- Paul

Paul Walmsley

unread,
Oct 12, 2018, 12:48:28 PM10/12/18
to ron minnich, Christoph Hellwig, padma...@gmail.com, Anup Patel, atis...@gmail.com, sw-...@groups.riscv.org

Yes - for RISC-V U-boot, it's best to follow the example set by the ARM64 U-boot folks and support booting in any mode:

https://github.com/theopolis/u-boot/blob/master/doc/README.arm64


- Paul

kr...@berkeley.edu

unread,
Oct 12, 2018, 1:42:29 PM10/12/18
to ron minnich, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, sw-...@groups.riscv.org

Ron,

Please don't confuse the desire to separate M-mode functionality from
S-mode functionality for the purposes of making code more portable,
with a desire by vendors to lock down M-mode.

Having a clean separation of concerns between abstraction layers is
important for engineering reasons, and is completely orthogonal to
whether each abstraction layer is open or not.

We want to enable platforms that don't trust the OS or even
bootloaders, as much as we want to allow platforms that don't trust
vendor firmware. We also want to make it easier to build emulations
and flexible virtualization systems.

My comments are not meant to read on any of the proposals below, but
just on the sentiment that somehow specifying this abstraction
interface will prevent open implementations of the abstraction
interface, or prevent different implementations that ignore the
abstraction interface.

Krste
| CAP6exYKr_uHoX1FvKSTnPB8gPkYzYY0LOKB%2BbvTvYiyV%3DcfgfA%40mail.gmail.com.

ron minnich

unread,
Oct 12, 2018, 4:13:25 PM10/12/18
to kr...@berkeley.edu, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, sw-...@groups.riscv.org, Jonathan Neuschäfer, Philipp Hug
Krste, thanks, I was only reacting to this sentence:
"Please no M-mode code in U-boot or any general purpose bootloader."

U-boot and coreboot live in the same "place" in the boot path, usually called firmware. Some people call them bootloaders, although I do not.

Since the first port in 2014, coreboot has always started in M-mode, and has always provided an SBI implementation, and will continue to do so. 

Further, we're looking at the following model for things coreboot loads (payloads)
- payload can be loaded in M mode and if it is, is allowed to replace existing M-mode code with its own (this includes Linux)
  coreboot will not write protect or lock away the M mode code area
- payload can be loaded in S mode and if it is, we assume it uses coreboot's M-Mode code. 
  coreboot might write protect the M mode code area but will never revoke execute or write permissions.
  revoking read access to M-mode is just another instance of security through obscurity and it worries me. But vendors do this
  type of thing all the time (See: SMM)

I was a bit concerned by several comments on this thread implying firmware running in M-mode is not desirable. 
I would certainly be worried should a given riscv implementation and mask rom make it impossible for coreboot to start in M-mode.

Thanks

ron

alist...@gmail.com

unread,
Oct 12, 2018, 5:12:08 PM10/12/18
to RISC-V SW Dev, kr...@berkeley.edu, h...@lst.de, an...@brainfault.org, atis...@gmail.com, padma...@gmail.com, j.neus...@gmx.net, phi...@hug.cx
I agree that we shouldn't trust vendor firmware. It is always buggy at best.

The problem that I see with starting coreboot or u-boot (the last stage bootloaders) in M mode is that we then end up with duplicated work between the projects.

This is where a model like Arm Trusted Firmware fits in well. We can see something like this:
 1. CPU boots up and ROM does it's thing
 2. ROM boots into an open source firmware (ATF equivalent) in M mode
    - This will end up being pretty SoC specific
 3. The firmware boots into u-boot or core boot in S mode (or H mode when that exists)
    - Hopefully this can then be as generic as possible
 4. From there we boot the system

The advantage I see with this is that we only need to implement the M mode functionality once and then coreboot, u-boot and anything else can utilise it. It also keeps bigger later stage boot loaders out of M mode which has a mild security benefit of smaller attack surfaces.

The other aim is to allow a single RISC-V image to boot on a variety of different SoCs. The more generic we can make the last stage boot loaders the more achievable this is.

Alistair

Tommy Murphy

unread,
Oct 12, 2018, 5:40:52 PM10/12/18
to RISC-V SW Dev, kr...@berkeley.edu, h...@lst.de, an...@brainfault.org, atis...@gmail.com, padma...@gmail.com, j.neus...@gmx.net, phi...@hug.cx
On Friday, 12 October 2018 22:12:08 UTC+1, alist...@gmail.com wrote:
I agree that we shouldn't trust vendor firmware. It is always buggy at best.
 
But once the source is open/available surely one can see what's going on and address problems/bugs as needed? 

Declaration/disclaimer: I work for Microchip/Microsemi.

ron minnich

unread,
Oct 12, 2018, 6:30:34 PM10/12/18
to Tommy Murphy, RISC-V SW Dev, kr...@berkeley.edu, h...@lst.de, an...@brainfault.org, atis...@gmail.com, padma...@gmail.com, j.neus...@gmx.net, phi...@hug.cx
On Fri, Oct 12, 2018 at 2:40 PM Tommy Murphy <tommy_...@hotmail.com> wrote:
On Friday, 12 October 2018 22:12:08 UTC+1, alist...@gmail.com wrote:
I agree that we shouldn't trust vendor firmware. It is always buggy at best.
 
But once the source is open/available surely one can see what's going on and address problems/bugs as needed? 

40 years of experience shows otherwise. The only thing that works is if everyone always has the ability to compile and build their firmware, and from more than one source.

A very similar argument to the one we use for open source kernels, in fact.

ron 

ron minnich

unread,
Oct 12, 2018, 6:36:27 PM10/12/18
to alist...@gmail.com, RISC-V SW Dev, kr...@berkeley.edu, h...@lst.de, an...@brainfault.org, atis...@gmail.com, padma...@gmail.com, j.neus...@gmx.net, phi...@hug.cx
On Fri, Oct 12, 2018 at 2:12 PM <alist...@gmail.com> wrote:
I agree that we shouldn't trust vendor firmware. It is always buggy at best.

The problem that I see with starting coreboot or u-boot (the last stage bootloaders) in M mode is that we then end up with duplicated work between the projects.

Sure. There's lots of duplicated work between kernels (BSD, Linux, ...); compilers (clang, gcc); browsers, shells, editors ...

nobody has ever made a convincing argument that this is a problem. And it's not a problem for firmware.
 

This is where a model like Arm Trusted Firmware fits in well. We can see something like this:

... 
(interested text removed :-)
 
The advantage I see with this is that we only need to implement the M mode functionality once and then coreboot, u-boot and anything else can utilise it. It also keeps bigger later stage boot loaders out of M mode which has a mild security benefit of smaller attack surfaces.


I don't see any advantage. If M-mode is so complex that we can only have one workable implementation, something's gone horribly wrong in RISC-V.

We also learned the value of multiple  firmware implementations back in the linuxbios days, over and over again, particularly on the early opteron: an AMD engineer told us the linuxbios hypertransport enumeration code was better than what AMD had. And we knew this from experience: companies had opteron mainboards that would not boot if socket 0 was empty, and linuxbios had no trouble with that at all.

So, no, software monoculture is not a good idea for firmware, as it is not a good idea anywhere else.

ron 

alist...@gmail.com

unread,
Oct 12, 2018, 6:43:44 PM10/12/18
to RISC-V SW Dev, alist...@gmail.com, kr...@berkeley.edu, h...@lst.de, an...@brainfault.org, atis...@gmail.com, padma...@gmail.com, j.neus...@gmx.net, phi...@hug.cx
I'm not saying that we have to have one implementation. I just see a lot of value in separating what runs in M mode and what runs in S mode.

If someone wanted to they can run something that handles both, nothing is stopping them from doing that. There is no enforcement of mono-culture or of vendor firmware.

I just see a lot of value in separating the firmware from the last stage bootloader. As part of this though we would need to come up with a firmware option that can be used for those who want to.

Alistair
 

ron 

ron minnich

unread,
Oct 12, 2018, 7:02:27 PM10/12/18
to alist...@gmail.com, RISC-V SW Dev, kr...@berkeley.edu, h...@lst.de, an...@brainfault.org, atis...@gmail.com, padma...@gmail.com, j.neus...@gmx.net, phi...@hug.cx
On Fri, Oct 12, 2018 at 3:43 PM <alist...@gmail.com> wrote:


I just see a lot of value in separating the firmware from the last stage bootloader. As part of this though we would need to come up with a firmware option that can be used for those who want to.




I guess I don't understand the distinction. On chromebooks, for example, coreboot is the firmware and the last stage bootloader. And certainly u-boot on many systems is the firmware and last stage bootloader.  This is an area with lots of variety :-)

But all that said, if we're not closing off firmware and even kernels from having M-mode components they provide, I'm fine.


Luke Kenneth Casson Leighton

unread,
Oct 12, 2018, 7:23:10 PM10/12/18
to alist...@gmail.com, RISC-V SW Dev, Krste Asanovic, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx
On Fri, Oct 12, 2018 at 10:12 PM <alist...@gmail.com> wrote:

> The problem that I see with starting coreboot or u-boot (the last stage bootloaders)
> in M mode is that we then end up with duplicated work between the projects.

yes. that has not stopped either project from continuing. it also
hasn't stopped libreboot, or grub, or the system which allows a linux
kernel to actually *be* the bootloader (which runs an initramfs and
calls kexec).

each of these bootloaders has different fans and userbases. they
don't necessarily talk to each other, and that's just how it is.

expecting each independent team to cooperate, particularly when some
of them are hard forks of other bootloader projects, is not realistic.

> This is where a model like Arm Trusted Firmware fits in well.

you mean, ARM _treacherous_ firmware? the "trust" you refer to is
directed at *manufacturers*, providing the specific goal of
*betraying* users. "Trust Us. We Will HELP You To Prevent And
Prohibit Users From Owning Their Devices".

>We can see something like this:
> 1. CPU boots up and ROM does it's thing
> 2. ROM boots into an open source firmware (ATF equivalent) in M mode
> - This will end up being pretty SoC specific

yes. and may not fit into the available memory, which may be
absolutely tiny.

the ROM bootloader, which is best made about as ridiculously dumb and
tiny as can possibly be made, is usually just about capable of loading
from a single (16k) sector of a Micro-SD or eMMC device, or 16k from
an SPI device, or 16k from an ONFI NAND, and so on. that sub-16k
executable the "first stage" bootloader.

16K is the "normal" amount of available SRAM that SoCs are prepared to
allocate for these first-stage bootloaders. that 1st stage basically
does things like "initialise the DDR3 RAM, set up IRQs"... oh and it
happens to initialise the
device-from-which-the-next-phase-is-to-be-loaded.

in u-boot, this first stage bootloader is called the "SPL", and it's
compiled from u-boot with some special #defines that *MASSIVELY* cut
down u-boot to the absolute, absolute bare minimum (surprise: under
16k).

all of this is typically entirely hidden from end-users. you don't
have to know. you don't have to care. people have *no idea* that
their system goes through a four, five or even six stage boot process
to get to an OS. even intel x86 systems.

basically i'm saying that "one size does not fit all", that creating
standards in this area is not only unrealistic but also completely
undesirable it is so heavily application-specific, and that
restricting vendors through standardisation to be not permitted to run
from the absolute lowest possible level (M-Mode) will almost certainly
cause a massive back-lash.

*on the other hand*, if there exists *informal reference source code*
out there, that's absolutely fantastic, as it will cut down massively
on development NREs. example ROM bootloaders, for example, would be
fantastic.

why would it be fantastic? because the entire process is so
proprietary, so secret, that very few people actually know what's
involved. full libre-licensed source code - preferably with
libre-licensed overview documents - would completely change how SoCs
are developed.

l.

alist...@gmail.com

unread,
Oct 12, 2018, 7:45:29 PM10/12/18
to RISC-V SW Dev, alist...@gmail.com, kr...@berkeley.edu, h...@lst.de, an...@brainfault.org, atis...@gmail.com, padma...@gmail.com, j.neus...@gmx.net, phi...@hug.cx
On Friday, October 12, 2018 at 4:23:10 PM UTC-7, Luke Kenneth Casson Leighton wrote:
On Fri, Oct 12, 2018 at 10:12 PM <alist...@gmail.com> wrote:

> The problem that I see with starting coreboot or u-boot (the last stage bootloaders)
> in M mode is that we then end up with duplicated work between the projects.

 yes.  that has not stopped either project from continuing.  it also
hasn't stopped libreboot, or grub, or the system which allows a linux
kernel to actually *be* the bootloader (which runs an initramfs and
calls kexec).

 each of these bootloaders has different fans and userbases.  they
don't necessarily talk to each other, and that's just how it is.

 expecting each independent team to cooperate, particularly when some
of them are hard forks of other bootloader projects, is not realistic.

> This is where a model like Arm Trusted Firmware fits in well.

 you mean, ARM _treacherous_ firmware?  the "trust" you refer to is
directed at *manufacturers*, providing the specific goal of
*betraying* users.   "Trust Us.  We Will HELP You To Prevent And
Prohibit Users From Owning Their Devices".

That seems a little harsh. ATF is open source and you can compile it yourself from the source to deploy. You are also free to re-write it if need be.
The risk that I see with just letting everyone go their own way is that we will end up with a similar situation to the ARM world.

That is that every SoC vendor does their own thing. This can result in a less open ecosystem as every vendor writes their own boot code and keeps it closed.

The other problem I see is a fragmented boot process. Where some boards use coreboot and some use u-boot and some use something else. We will then get to a point where every board is different. x86 has this pretty much sorted as I can boot my Linux USB stick on every PC I own. ARM is the exact opposite of this. I would like to see RISC-V not go down this fragmentation route. One way to do this is just to have a "best" way to setup the boot process. Then SoC vendors will hopefully opt to follow it.

A mix-mash of ways seems like a dangerous path to lead to fragmentation and confusion and that is what I'm afraid of.

Alistair
 

l.

Luke Kenneth Casson Leighton

unread,
Oct 12, 2018, 7:58:01 PM10/12/18
to alist...@gmail.com, RISC-V SW Dev, Krste Asanovic, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx
---
crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68

On Sat, Oct 13, 2018 at 12:45 AM <alist...@gmail.com> wrote:

> That seems a little harsh. ATF is open source and you can
> compile it yourself from the source to deploy. You are also free to re-write it if need be.

superb (and thank you for the additional, missing, qualifying
information). then by definition it's not treacherous.

>> why would it be fantastic? because the entire process is so
>> proprietary, so secret, that very few people actually know what's
>> involved. full libre-licensed source code - preferably with
>> libre-licensed overview documents - would completely change how SoCs
>> are developed.
>
>
> The risk that I see with just letting everyone go their own way is that we will end up with a similar situation to the ARM world.
>
> That is that every SoC vendor does their own thing. This can
> result in a less open ecosystem as every vendor writes their own
> boot code and keeps it closed.

there's a reason for that, despite in many cases the code that
vendors being used actually being GPLv2-licensed (and thus they
release criminally-infringing products). i'll back-track and post on
that from a different point in the thread.

> The other problem I see is a fragmented boot process.

this is just an unavoidable reflection of the disparate
hardware-level requirements and design constraints that implementors
face.

> x86 has this pretty much sorted as I can boot my
> Linux USB stick on every PC I own.

yah. have you tried booting it when the SPI bootloader chip on the
motherboard is empty, corrupted, or missing? have you tried asking
intel for information about how the first stage bootloader process
actually works? try those things... it's extremely illuminating.

> One way to do this is just to have a "best" way to setup
> the boot process. Then SoC vendors will hopefully opt to follow it.

i specifically avoided the use of the word "best", as that implies
that "other means and methods are worst". however you've pretty much
said exactly what i said, just in different words.

> A mix-mash of ways seems like a dangerous path to lead
> to fragmentation and confusion and that is what I'm afraid of.

like i said: fragmentation, here, is just unavoidable, and efforts to
standardise *WILL* mean with stone-walling. you cannot possibly
expect completely different hardware to have anything in common at the
*boot initialisation* level, and we know from experience that attempts
at abstraction result in bloat.

bloat when you have only 16K to play with... the first thing that
gets cut is: attempts at abstraction.

l.

ron minnich

unread,
Oct 12, 2018, 8:07:28 PM10/12/18
to alist...@gmail.com, RISC-V SW Dev, kr...@berkeley.edu, h...@lst.de, an...@brainfault.org, atis...@gmail.com, padma...@gmail.com, j.neus...@gmx.net, phi...@hug.cx
On Fri, Oct 12, 2018 at 4:45 PM <alist...@gmail.com> wrote:


That seems a little harsh. ATF is open source and you can compile it yourself from the source to deploy. You are also free to re-write it if need be.

I agree. ARM has been very good about ATF, in fact it's a model of how it can be done. I'm sorry to see this type of message in what has been an interesting technical discussion. 

 
A mix-mash of ways seems like a dangerous path to lead to fragmentation and confusion and that is what I'm afraid of.


I don't see that multiple implementations necessarily lead to this case, but you make a fair point. 

Anyway, my question has been answered: the things I want will remain possible, so I'm happy.

Thanks everyone.

ron

alist...@gmail.com

unread,
Oct 12, 2018, 8:16:56 PM10/12/18
to RISC-V SW Dev, alist...@gmail.com, kr...@berkeley.edu, h...@lst.de, an...@brainfault.org, atis...@gmail.com, padma...@gmail.com, j.neus...@gmx.net, phi...@hug.cx
To a certain extent it's unavoidable, but if RISC-V is ever going to take off we are going to have to at least somewhat deal with it. Booting ARM boards today is such a pain that I suspect 90% of people give up and just take the prebuilt binaries from the SoC vendor. It is a hassle and if we tackle this early on we might be able to reduce this for RISC-V.

At least the first step is are guide lines of what should be implemented and where. Just to have everyone on the same page is very helpful.

For example saying that the first stage boot loader must start in M (obviously) and the last stage in S. That means that you can have one binary that does both, or two binaries separately. Even a very basic and general best practices will help push vendors to do the right thing and make it easier for developers.This can even be done just using current implementations as reference. It just hopefully stops someone in the future doing something super crazy that doesn't work well with anyone else.

I still prefer the layered approach that ATF follows, but any standardisation is good, so we can drop that out.

Alistair

Luke Kenneth Casson Leighton

unread,
Oct 12, 2018, 8:36:05 PM10/12/18
to Krste Asanovic, ron minnich, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, RISC-V SW Dev
---
crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68

On Fri, Oct 12, 2018 at 6:42 PM <kr...@berkeley.edu> wrote:
>
>
> Ron,
>
> Please don't confuse the desire to separate M-mode functionality from
> S-mode functionality for the purposes of making code more portable,
> with a desire by vendors to lock down M-mode.
>
> Having a clean separation of concerns between abstraction layers is
> important for engineering reasons, and is completely orthogonal to
> whether each abstraction layer is open or not.
>
> We want to enable platforms that don't trust the OS or even
> bootloaders, as much as we want to allow platforms that don't trust
> vendor firmware. We also want to make it easier to build emulations
> and flexible virtualization systems.

these are reasonable statements and a good reflection of a wide range
of goals and requirements from many different sources.

the mention of abstraction however has me twitching, slightly, where
(as mentioned in another reply), the design constraints are so
disparate, and the available memory constraints so insanely low, that
putting in any kind of abstraction is almost certainly likely to fail,
as the first thing that vendors will do is look at the hardware cost
of doing so, and reject it immediately.

chain-loading on the other hand - multi-stage bootloading - not a
problem, and is extremely common:

* http://linux-sunxi.org/EGON#eGON.BRM and you can see a link to the
BROM page as well. following the reverse-engineering that is
described, there, u-boot now supports the "eGON" header format for
allwinner SoCs, such that u-boot's "SPL" first stage bootloader is
recognised.

* http://opensource.rock-chips.com/wiki_U-Boot - here you see "u-boot
SPL" which is that 1st-stage bootloader previously mentioned. the
RK3288 is a bit annoying as the 1st stage bootloader has to be CRC/RC4
encrypted and signed. and that's in the boot ROM so is completely
impossible to circumvent. or fix.

where confusion might arise from what you wrote, krste, is that
normally, u-boot's first-stage SPL (and likely the coreboot
equivalent) *absolutely* has to be run in M-Mode. where else would
the DDR3 timings, the PLL, IRQs and other absolutely essential
peripherals be initialised?

and, that being the case, if u-boot or coreboot were operating in
S-Mode, it would be impossible for those hardware-level
initialisations to be properly carried out. more to the point: what
got u-boot / coreboot *into* S-Mode in the first place??

:)

so, logically, when you refer to a desire to run "bootloaders" in
S-Mode we may *deduce* that you are NOT referring to the ENTIRE boot
process, that instead that you may be referring to the SECOND phase of
u-boot / coreboot. i.e. definitely *not* the (16-or-so-K) first phase
(known as SPL mode for u-boot)... [unless the processor's bootrom has
already done so... which would be hugely risky from a design
perspective]

here - the second phase - it _might_ be reasonable (and possible) to
run from S-Mode. binary sizes are 600k to 2MB (estimated), in main
(DDR) memory. it just depends on whether there was enough spare SRAM
on the processor for a first-stage boot loader with S-Mode
initialisation to *actually fit*.

*that may not actually be the case*... and you'd be looking instead
at triple-stage loading.

really i just want to make people aware, this area is quite involved,
and the number of people in the world who know what's going on *and we
can communicate with because they're not under NDA* is very small: the
members of the u-boot team, plus the members of the coreboot team, and
libreboot as well.

getting their input here - and making absolutely sure that they're
properly compensated for their time (reminder to Foundation Members:
please read section 2.3 of your Member Charter Agreement) - is
absolutely essential.


one other thing that has me concerned: most DDR3/4 vendors FORCE
customers to sign ridiculously-restrictive NDAs, that do not even
permit them to release the source code of the DDR3/4 initialisation
sequence.

the *only* way for fabless semiconductors to release useful product
(i.e. product that will actually boot) is:

(a) release a proprietary first stage bootloader [and then wait for
libre software engineers to reverse-engineer it. in their "spare"
time. unpaid.]

(b) criminally infringe copyright law.

Allwinner and many other Chinese Fabless semiconductor companies
systematically choose option (b). Western corporations, who
understand a little bit more about Copyright Law, typically choose
option (a).

i suspect that even SiFive has had to face this situation. and may
even be under such a restrictive NDA that they can't even talk about
it. anyone at SiFive who can confirm that, would be nice to hear from
you. of course... if we _don't_ hear from you... :)

l.

Luke Kenneth Casson Leighton

unread,
Oct 12, 2018, 8:48:27 PM10/12/18
to alist...@gmail.com, RISC-V SW Dev, Krste Asanovic, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx
On Sat, Oct 13, 2018 at 1:16 AM <alist...@gmail.com> wrote:

> To a certain extent it's unavoidable, but if RISC-V is ever going to take off we are going to have to at least somewhat deal with it. Booting ARM boards today is such a pain that I suspect 90% of people give up and just take the prebuilt binaries from the SoC vendor. It is a hassle and if we tackle this early on we might be able to reduce this for RISC-V.

no. really. please. try to understand and accept the implications,
here. if you want RISC-V to be in the embedded world, and in multiple
SoCs, and in desktops, and in high-end servers, you have to accept the
reality of the absolutely massive fundamental hardware differences
between all those different scenarios, where the bootloader and
bootloader choice absolutely necessarily reflects that.

if you start talking about UEFI to low-cost tablet vendors for
example they will just laugh at you. and if you even mention ACPI to
them as a required standard as part of a boot process, they'd be quite
likely to punch your lights out (and i would cheer and congratulate
them).

x86 does not have this "problem" that ARM has because x86 has FAILED
to enter the ARM-dominated SoC market. despite intel bribing vendors,
sellers, OEMs *and customers* to buy their products, they *still could
not get market share*. yes, i've been talking to a business
development manager for an ARM-based fabless semi company, who
regularly was asked, "how much are YOU going to pay US to use your
SoC, because that's what intel does".

all that having been said: i concur that *if* full source code is
available (and available early enough), chances are indeed high that
fabless semi companies will just cookie-cut the available solutions.

unfortunately, the DDR3/4 hard macro vendors (who typically provide
proprietary initialisation source code under insane NDAs) are going to
get in the way, there.

so that's down to the early adopters - SiFive in particular as the
high-profile one - to take responsibility for that, and tell the
DDR3/4 interface vendors in no uncertain terms that proprietary
licensing on the DRAM initialisation is not acceptable.

l.

Samuel Falvo II

unread,
Oct 12, 2018, 8:57:17 PM10/12/18
to Luke Kenneth Casson Leighton, alist...@gmail.com, sw-...@groups.riscv.org, Krste Asanovic, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx
On Fri, Oct 12, 2018 at 4:23 PM Luke Kenneth Casson Leighton
<lk...@lkcl.net> wrote:
> basically i'm saying that "one size does not fit all", that creating
> standards in this area is not only unrealistic but also completely
> undesirable it is so heavily application-specific, and that
> restricting vendors through standardisation to be not permitted to run
> from the absolute lowest possible level (M-Mode) will almost certainly
> cause a massive back-lash.
>
> *on the other hand*, if there exists *informal reference source code*
> out there, that's absolutely fantastic, as it will cut down massively
> on development NREs. example ROM bootloaders, for example, would be
> fantastic.

I'm going to once again put forth the idea that it was a mistake for
the RISC-V community to settle on an ITU-style standards setting
process. While it may prove advantageous in some circumstances, I
think it sets a bad precedent overall where viable implementations
come second to the desire for an up-front, all-inclusive, frequently
self-inconsistent standard with which everyone must conform. See also
CORBA.

If standards are genuinely desirable, then I **strongly** encourage
the use of IETF-inspired standards track processes. I cannot
emphasize this enough. The process is 100% compatible with the idea
that implementation-specific boot-loaders are not only possible, but
even expected. IETF's processes say you need at least two (more are
preferred) independent, working, and interoperable implementations for
your (bootstrap) proposal to be even *considered* for RFC status.
This is a recognition that standards evolve bottom-up, and aren't
dictated top-down. It really doesn't matter how many RFCs related to
a topic exists (for example, there exists no fewer than four remote
service invokation mechanisms: XDR, NDR, XML-RPC, and SOAP); the
market will eventually sort out which one becomes dominant. And when
one (or a couple) becomes dominant, you get to say things like, "This
system complies with RFC-12345 bootstrapping standards," or "Our
latest model now supports RFC-12345."

--
Samuel A. Falvo II

Luke Kenneth Casson Leighton

unread,
Oct 12, 2018, 8:58:06 PM10/12/18
to alist...@gmail.com, RISC-V SW Dev, Krste Asanovic, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx
---
crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68

On Sat, Oct 13, 2018 at 1:16 AM <alist...@gmail.com> wrote:

> For example saying that the first stage boot loader
> must start in M (obviously) and the last stage in S.
> That means that you can have one binary that does both,

sorry, i missed this: no, you can't. really, you can't. the
processor is typically booted up with the DDR3/4 RAM *COMPLETELY
DISABLED*. as in; ***SWITCHED OFF***.

the only way to get it up and running is to turn it on, then do some
scans to find out what DRAM (and what type of DRAM) is there (and in
which banks), then tell both DRAM and Controller to start doing
impedance matching... it's *horrendously* complex, and most first
stage bootloaders do the absolute bare minimum (get the DDR3/4 RAM
running at say 200-300mhz) then leave the rest to second-stage.

the only way to load programs that do that is to have *SRAM* on-board
the processor, which, given that you need a minimum of 16k,
realistically, it's quite a lot to put down in a processor, so you
want to minimise the amount. you could hypothetically put the DRAM
initialisation into ROM however you would be taking a serious risk to
do so, as future DRAM from third party vendors could be incompatible,
ending the life of the processor immediately.

> or two binaries separately.

or, as i mentioned in the message to krste, three. first stage
M-Mode and initialising DRAM (and other critical peripherals). second
stage S-Mode initialisation. *THIRD* stage "main" boot loader, what
most people think of as "u-boot" (or coreboot).

all done as chain-loaders.

there is one _hell_ of a lot going on to get a processor up and running.

l.

Anup Patel

unread,
Oct 13, 2018, 2:45:08 AM10/13/18
to alist...@gmail.com, sw-...@groups.riscv.org, kr...@berkeley.edu, Christoph Hellwig, atish patra, Padmarao Begari, j.neus...@gmx.net, phi...@hug.cx
I totally agree with Alistair's comments.

The intentions behind having last stage bootloader run only in S-mode are:
1. Re-use it for VS-mode. In other words, it should be possible to
re-use last stage bootloader under a Guest/VM.
2. Platform specific services for CPU hotplug, Reboot, Shutdown, Power
management, Image authentication, etc (i.e. SBI calls) should be
provided by platform specific M-mode firmware

We certainly don't want #2 to be replicated across bootloaders (Uboot,
Coreboot, UEFI, etc). The best people to maintain code related to #2
are SOC vendors themselves. Of course, code related to #2 has to open
source so that people get chance to fix bugs themselves and also have
transparency.

It is also important that we don't want separate bootloader for
Guest/VM (i.e. VS-mode) so last stage bootloader should consider
Guest/VM requirements (i.e. #1 above)

Both above points are not cooked-up ideas, these are based on failures
and learning from ARM world. There is not point of discussion if
people want to re-invent wheel here.

For RISC-V ecosystem to succeed, all good practices from other
architectures should be considered from start itself.

It is better to take the pain now and get booting flow standardised
for RISC-V systems. It will be very difficult in future when we have
lot of RISC-V systems with there own booting flow.

Regards,
Anup

Luke Kenneth Casson Leighton

unread,
Oct 13, 2018, 4:33:25 AM10/13/18
to Anup Patel, alist...@gmail.com, RISC-V SW Dev, Krste Asanovic, Christoph Hellwig, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx
---
crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68

On Sat, Oct 13, 2018 at 7:45 AM Anup Patel <an...@brainfault.org> wrote:

> On Sat, Oct 13, 2018 at 5:46 AM <alist...@gmail.com> wrote:

> > I still prefer the layered approach that ATF follows, but any standardisation is good, so we can drop that out.
> >
> > Alistair
> >
>
> I totally agree with Alistair's comments.
>
> The intentions behind having last stage bootloader run only in S-mode are:
> 1. Re-use it for VS-mode. In other words, it should be possible to
> re-use last stage bootloader under a Guest/VM.

yes, this is a reasonable aim.

just bear in mind that each boot system being developed usually finds
it easier to go through their own chain-loading system and on
subsequently to a full OS. to add support for anything other than
that actually requires speciic effort (which obviously some
corporation has to pay for).

any kind of inter-project cross-collaboration - opportunities for
standardisation - is usually an informal affair that's completely by
accident, from a slow "organic" growth.

for example, UEFI was forced onto teams through x86 BIOSes adopting
it. consequently, YEARS after it became mainstream in the x86 world,
if you now google "u-boot UEFI" it is actually possible to do, and
there is in fact a talk about "marrying u-boot, uEFI and grub2":

https://www.cnx-software.com/2016/08/11/u-boot-now-supports-uefi-on-32-bit-and-64-bit-arm-platforms/

https://www.youtube.com/watch?v=qJAkJ3nmWgM

> 2. Platform specific services for CPU hotplug, Reboot, Shutdown, Power
> management, Image authentication, etc (i.e. SBI calls) should be
> provided by platform specific M-mode firmware
>
> We certainly don't want #2 to be replicated across bootloaders (Uboot,
> Coreboot, UEFI, etc).

apologies for putting it this way, anup: i appreciate that you're not
*actually* saying this: what you've written may very unfortunately be
interpreted as an unpaid "demand" - an order that must be obeyed - to
the u-boot and coreboot developers (and to the grub developers who
have an implementation of UEFI) that they absolutely must follow and
do precisely and exactly what you have said.

the closest that you can get is as follows:

(1) create a reasonable and rational logical argument and, in NEUTRAL
TERMS, propose it (separately) to the u-boot developers and to the
coreboot developers, and the grub developers.

(2) include an offer of compensation and payment for services as part
of the contact with those respective teams

(3) if there is no take-up (no interest) implement it yourself and
*demonstrate* that it's a good idea.

you cannot in *any way* expect sovereign libre project developers -
many of whom are chronically underfunded and absolutely sick and tired
of MULTIPLE BILLION DOLLAR corporations PERSISTENTLY spongeing off of
their expertise and time - to drop everything and "do what you want".

i know you *didn't* make demands of software libre teams: please
however be extremely careful how you put things, ok?

> The best people to maintain code related to #2
> are SOC vendors themselves. Of course, code related to #2 has to open
> source so that people get chance to fix bugs themselves and also have
> transparency.

indeed. and the upstream developers adequately compensated and paid
for their time. this does not "go without saying". it's something
that corporations have to actively take responsibility for. you made
money? PAY the people on whose success that money is critically
dependent!

i'll mention the words "shellshock", "heartbleed" and leave it at that.


> It is also important that we don't want separate bootloader for
> Guest/VM (i.e. VS-mode) so last stage bootloader should consider
> Guest/VM requirements (i.e. #1 above)
>
> Both above points are not cooked-up ideas, these are based on failures
> and learning from ARM world.

indeed... so that means that a full and comprehensive analysis *of*
the ARM world is needed. all of it. right from the Cortex M-series,
through the ARM9, ARM11, Cortex A8, A7/15, to the A53 and beyond.

then studying the way that PowerPC mirrored x86, and why MIPS
architectures do not, for example.

in other words, it's not just about saying "ARM was a failure because
the boot world is a mess", you have to understand **WHY** it got that
way in the first place.

and in doing so, you may learn that there are extremely good reasons why.


> There is not point of discussion if
> people want to re-invent wheel here.

equally there is no point in discussion if there is a failure to
recognise the reality of quite how disparate the hardware world really
is.

also, again: i'm obliged to remind you that you don't get to tell
software libre developers "what to do". if they want to reinvent
wheels as spheres, squares or other shapes, that's entirely their
prerogative, and you cannot in any way stop them.

secondly: i know you are not (i'm wording it deliberately this way by
way of illustration) - are you *really* proposing, for example, that a
low-cost RISC-V tablet processor design companyy be FORCED to use UEFI
boot, for example, because it "Shall Become A Standard That Is To Be
Forced Onto People In the RISC-V World"?


> For RISC-V ecosystem to succeed, all good practices from other
> architectures should be considered from start itself.

absolutely. except i would not make it *absolutely* conditional that
"RISC-V will *FAIL* if all good practices are not considered", which
is unfortunately a possible implication of the wording that you've
chosen here.


> It is better to take the pain now and get booting flow standardised
> for RISC-V systems.

anup: you have jumped the gun, here. it is very strange. in one
sentence you say "RISC-V should look at good practices from other
architectures", and then *IMMEDIATELY* jump to, "and that process is
ABSOLUTELY going to result in standardisation for RISC-V boot
procedures".

that sounds like the proposed review of good practices from other
architectures is a loaded, foregone conclusion. you see how that
could be bad?

> It will be very difficult in future when we have
> lot of RISC-V systems with there own booting flow.

nooo, the future will reflect what individual diverse fabless
semiconductor companies with completely and utterly different hardware
and customer requirements *need* from a boot flow.

basically what i would like to get across to people here, is, that
u-boot, coreboot, libreboot and the grub developers have been at this
for a lot longer than RISC-V has been around. their expertise and
approach actually takes precedence over that of RISC-V! u-boot has to
deal with thousands of disparate systems (only a tiny handful of which
are *developer* boards with RISC-V cores)

RISC-V is the *newcomer* here. people may *want* boot
standardisation in RISC-V in order to reduce development and
maintenance costs (a desirable goal), however starting from a position
of assuming or implying that the teams who manage those projects must
be "stupid" by having gone their own separate ways and having
"reinvented wheels", is not going to fly.

also a reminder to RISC-V Foundation Members of section 2.2:
https://content.riscv.org/wp-content/uploads/2016/01/RISC-V-Foundation-Membership-Agreement-8Dec2016.pdf

"During the term of its membership in Foundation, Member is expected
to support the free and open design, development and improvement of
the RISC-V ISA, together with its software and hardware ecosystem for
use in all computing devices."

l.

Anup Patel

unread,
Oct 13, 2018, 5:42:29 AM10/13/18
to lk...@lkcl.net, alist...@gmail.com, sw-...@groups.riscv.org, kr...@berkeley.edu, Christoph Hellwig, atish patra, Padmarao Begari, j.neus...@gmx.net, phi...@hug.cx
I tried to explain my view in best possible way. If this is taken as
unpaid demand then that was not the intention.

>
> the closest that you can get is as follows:
>
> (1) create a reasonable and rational logical argument and, in NEUTRAL
> TERMS, propose it (separately) to the u-boot developers and to the
> coreboot developers, and the grub developers.
>
> (2) include an offer of compensation and payment for services as part
> of the contact with those respective teams
>
> (3) if there is no take-up (no interest) implement it yourself and
> *demonstrate* that it's a good idea.
>
> you cannot in *any way* expect sovereign libre project developers -
> many of whom are chronically underfunded and absolutely sick and tired
> of MULTIPLE BILLION DOLLAR corporations PERSISTENTLY spongeing off of
> their expertise and time - to drop everything and "do what you want".
>
> i know you *didn't* make demands of software libre teams: please
> however be extremely careful how you put things, ok?

I see this getting derailed. The whole discussion about what should be
correct boot flow.

The "should" in my comment should not be interpreted unpaid demand. The
"should" here is the sense "what I think correct way is".

>
> > The best people to maintain code related to #2
> > are SOC vendors themselves. Of course, code related to #2 has to open
> > source so that people get chance to fix bugs themselves and also have
> > transparency.
>
> indeed. and the upstream developers adequately compensated and paid
> for their time. this does not "go without saying". it's something
> that corporations have to actively take responsibility for. you made
> money? PAY the people on whose success that money is critically
> dependent!
>
> i'll mention the words "shellshock", "heartbleed" and leave it at that.
>
>
> > It is also important that we don't want separate bootloader for
> > Guest/VM (i.e. VS-mode) so last stage bootloader should consider
> > Guest/VM requirements (i.e. #1 above)
> >
> > Both above points are not cooked-up ideas, these are based on failures
> > and learning from ARM world.
>
> indeed... so that means that a full and comprehensive analysis *of*
> the ARM world is needed. all of it. right from the Cortex M-series,
> through the ARM9, ARM11, Cortex A8, A7/15, to the A53 and beyond.
>
> then studying the way that PowerPC mirrored x86, and why MIPS
> architectures do not, for example.

Yes, of course.

People should certainly make suggestions/recommendations based on
their experiences with other architectures.

(Note: Please don't mis-interpret "should" here.)

>
> in other words, it's not just about saying "ARM was a failure because
> the boot world is a mess", you have to understand **WHY** it got that
> way in the first place.
>
> and in doing so, you may learn that there are extremely good reasons why.

Yes, of course.

I don't see people asking "WHY" were things done in a particular way
in some ABC architecture.

>
>
> > There is not point of discussion if
> > people want to re-invent wheel here.
>
> equally there is no point in discussion if there is a failure to
> recognise the reality of quite how disparate the hardware world really
> is.
>
> also, again: i'm obliged to remind you that you don't get to tell
> software libre developers "what to do". if they want to reinvent
> wheels as spheres, squares or other shapes, that's entirely their
> prerogative, and you cannot in any way stop them.

Again, please don't mis-interpret my statement.

>
> secondly: i know you are not (i'm wording it deliberately this way by
> way of illustration) - are you *really* proposing, for example, that a
> low-cost RISC-V tablet processor design companyy be FORCED to use UEFI
> boot, for example, because it "Shall Become A Standard That Is To Be
> Forced Onto People In the RISC-V World"?

I am not supporting any particular bootloader here.

I am trying to suggest what should be correct boot flow.

(Please don't mis-interpret "should" in above sentence)

>
>
> > For RISC-V ecosystem to succeed, all good practices from other
> > architectures should be considered from start itself.
>
> absolutely. except i would not make it *absolutely* conditional that
> "RISC-V will *FAIL* if all good practices are not considered", which
> is unfortunately a possible implication of the wording that you've
> chosen here.

I am done clarifying your mis-interpretations. Where have I implied
RISC-V will FAIL.
It was really interesting to see how simple comments can be
heavily mis-interpreted.

Regards,
Anup

Luke Kenneth Casson Leighton

unread,
Oct 13, 2018, 5:47:29 AM10/13/18
to Anup Patel, alist...@gmail.com, RISC-V SW Dev, Krste Asanovic, Christoph Hellwig, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx
On Sat, Oct 13, 2018 at 10:42 AM Anup Patel <an...@brainfault.org> wrote:

> It was really interesting to see how simple comments can be
> heavily mis-interpreted.

indeed it is... and not by me. i used phrases that indicated that
they *could* (third person) be misinterpreted.

l.

Luke Kenneth Casson Leighton

unread,
Oct 13, 2018, 6:07:37 AM10/13/18
to Anup Patel, alist...@gmail.com, RISC-V SW Dev, Krste Asanovic, Christoph Hellwig, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx
On Sat, Oct 13, 2018 at 10:42 AM Anup Patel <an...@brainfault.org> wrote:
>
> The whole discussion about what should be
> correct boot flow.

for which sub-design architecture? are you referring to the correct
boot flow for embedded systems with below 3MB of (internal) RAM? for
embedded systems with below 48k of (hard-coded, on-board) RAM?
low-cost smartphone / tablet SoC market where the mass-volume target
price of a 4-core 1.5ghz SoC is around USD $3, for use in products
where the OEM price will be under $30? chromebook/netbook style
systems? desktop systems? high-end server systems?

all of these have wildly disparate boot flow requirements, yet you
consistently refer to "one" boot flow, and also use the word
"correct", as if there shall be one and only one possible option. if
you had said, "the whole discussion is about what should be the best
recommended boot flow for each platform", i would have agreed with you
entirely.

apologies for saying this, anup: i feel that it is disingenuous to
claim that the "conversation has been derailed" when i am simply
pointing out that the points that you raised are incomplete,
oversimplified, and indicative of a lack of knowledge of this
extremely comprehensive and diverse area.

> I don't see people asking "WHY" were things done in a particular way
> in some ABC architecture.

i have. as a reverse-engineer and embedded systems engineer who has
over 20 years experience of board bring-up on dozens of systems (many
of them with zero knowledge) i've studied dozens of different
architectures.

i have indeed asked the question "why" for example is PowerPC such a
uniform and easy architecture to work with, when the 1000+ licensees
of ARM do not have it easy at all (one clue is in the "1000+").

l.

Luke Kenneth Casson Leighton

unread,
Oct 13, 2018, 6:31:09 AM10/13/18
to padma...@gmail.com, RISC-V SW Dev
On Thu, Oct 11, 2018 at 12:10 PM Padmarao Begari <padma...@gmail.com> wrote:
>
> RISC-V U-Boot(64-bit) for HiFive Unleashed Board is available on GitHub
>
> https://github.com/Microsemi-SoC-IP/HiFive_U-Boot
>
> This repository contains the RISC-V U-Boot created by Microsemi a Microchip company for RISC-V core on HiFive Unleashed 540 device. The HiFive Unleashed Board is used for HiFive U-Boot development

padmarao, hi, as the discussion has changed topic, i thought it
important to say thank you and recognise the work that you and
microsemi have put in here. as someone who has worked on u-boot board
bring-up i know it is very detailed work, so thank you.

l.

Anup Patel

unread,
Oct 13, 2018, 7:04:12 AM10/13/18
to lk...@lkcl.net, alist...@gmail.com, sw-...@groups.riscv.org, kr...@berkeley.edu, Christoph Hellwig, atish patra, Padmarao Begari, j.neus...@gmx.net, phi...@hug.cx
On Sat, Oct 13, 2018 at 3:37 PM Luke Kenneth Casson Leighton
<lk...@lkcl.net> wrote:
>
> On Sat, Oct 13, 2018 at 10:42 AM Anup Patel <an...@brainfault.org> wrote:
> >
> > The whole discussion about what should be
> > correct boot flow.
>
> for which sub-design architecture? are you referring to the correct
> boot flow for embedded systems with below 3MB of (internal) RAM? for
> embedded systems with below 48k of (hard-coded, on-board) RAM?
> low-cost smartphone / tablet SoC market where the mass-volume target
> price of a 4-core 1.5ghz SoC is around USD $3, for use in products
> where the OEM price will be under $30? chromebook/netbook style
> systems? desktop systems? high-end server systems?

Yap, ATF-like boot flow work perfectly fine for internal SRAM as
low as 128KB

I have seen this in both embedded and high-end server systems.

>
> all of these have wildly disparate boot flow requirements, yet you
> consistently refer to "one" boot flow, and also use the word
> "correct", as if there shall be one and only one possible option. if
> you had said, "the whole discussion is about what should be the best
> recommended boot flow for each platform", i would have agreed with you
> entirely.

Yes, we can still come-up with boot flow where certain components
of boot chain are configurable as well as optional.

>
> apologies for saying this, anup: i feel that it is disingenuous to
> claim that the "conversation has been derailed" when i am simply
> pointing out that the points that you raised are incomplete,
> oversimplified, and indicative of a lack of knowledge of this
> extremely comprehensive and diverse area.

I think so because whole discussion was about why M-mode
code should not be there in last stage bootloader and it was not
about describing a detailed boot flow.

May be I should have directly proposed a detailed boot flow to
satisfy your queries.

>
> > I don't see people asking "WHY" were things done in a particular way
> > in some ABC architecture.
>
> i have. as a reverse-engineer and embedded systems engineer who has
> over 20 years experience of board bring-up on dozens of systems (many
> of them with zero knowledge) i've studied dozens of different
> architectures.
>
> i have indeed asked the question "why" for example is PowerPC such a
> uniform and easy architecture to work with, when the 1000+ licensees
> of ARM do not have it easy at all (one clue is in the "1000+").
>

Thanks for asking why.

I have provided enough justification for why i think M-mode code should
not be part of last stage bootloader.

Regards,
Anup

Luke Kenneth Casson Leighton

unread,
Oct 13, 2018, 8:29:53 AM10/13/18
to Anup Patel, alist...@gmail.com, RISC-V SW Dev, Krste Asanovic, Christoph Hellwig, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx
On Sat, Oct 13, 2018 at 12:04 PM Anup Patel <an...@brainfault.org> wrote:

> May be I should have directly proposed a detailed boot flow to
> satisfy your queries.

clarity always helps.

> > i have indeed asked the question "why" for example is PowerPC such a
> > uniform and easy architecture to work with, when the 1000+ licensees
> > of ARM do not have it easy at all (one clue is in the "1000+").
> >
>
> Thanks for asking why.
>
> I have provided enough justification for why i think M-mode code should
> not be part of last stage bootloader.

nggh, anup, y'know... i face a lot of shit on this list, as you've
probably witnessed, and many people correct my mistakes, sometimes
kind and informative ways, and at other times in really quite rude
ways, which is extremely distressing, so, knowing how that feels, i'm
in absolutely no way going to do the same to you, ever, so please bear
that in mind, ok?

i do feel compelled to point out that i don't feel that you have
provided enough justifiction for why M-Mode should not be part of the
last stage bootloader. it is i feel pretty easy to come up with a
scenario (embedded barely-fast-enough low-cost tablets where the cost
of every storage component is counted down to the last $0.01) where
the cost of the extra code in terms of performance and storage
capacity of such a [non-M-mode last-stage bootloader] proposal would
result in its outright rejection by OEMs.

that said: the *option* for last stage bootloaders to not run in
M-Mode: i would agree, yes, that is indeed a desirable characteristic.

l.

Tommy Murphy

unread,
Oct 13, 2018, 8:35:37 AM10/13/18
to RISC-V SW Dev
Huh? You're saying that if the sources are available you *can't* see and fix bugs if necessary? Doesn't make any sense to me...

Samuel Falvo II

unread,
Oct 13, 2018, 8:48:04 AM10/13/18
to Tommy Murphy, sw-...@groups.riscv.org
On Sat, Oct 13, 2018 at 5:35 AM Tommy Murphy <tommy_...@hotmail.com> wrote:
> Huh? You're saying that if the sources are available you *can't* see and fix bugs if necessary? Doesn't make any sense to me...

You did not quote, even partially, the person you're responding to, so
I'm going to assume you are responding to me.

The sources to UEFI are readily available on Github (for at least one
vendor, at least); despite this, you don't see me reflashing my PC
every time it gets updated. First, I don't have the development
toolchain that ASUS, American Megatrends, et. al. use. Second, I
don't have the proprietary extensions my firmware vendor has lumped
into the sources (including but not limited to the proprietary
motherboard-specific blobs). Third, it's largely unnecessary, for
when Linux boots, it largely obsoletes and renders UEFI moot. Because
of (3), then, clearly I'm not referring to bootloader-related stuff.
I'm talking about M-mode code for things like emulating FPU
instructions, or emulating a hardware page table walker, or other
kinds of things for which M-mode was actually intended for: the kinds
of things which frequently reside in "opaque, binary blobs" as they're
colloquially referred to these days. The kinds of things which
hardware vendors generally doesn't distribute sources to.

Padmarao Begari

unread,
Oct 13, 2018, 8:56:39 AM10/13/18
to Luke Kenneth Casson Leighton, RISC-V SW Dev
Thank you Luke...👍

Sent from my iPhone

Jonathan Neuschäfer

unread,
Oct 13, 2018, 11:36:20 AM10/13/18
to Luke Kenneth Casson Leighton, Krste Asanovic, ron minnich, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, RISC-V SW Dev
On Sat, Oct 13, 2018 at 01:35:29AM +0100, Luke Kenneth Casson Leighton wrote:
[...]
> so, logically, when you refer to a desire to run "bootloaders" in
> S-Mode we may *deduce* that you are NOT referring to the ENTIRE boot
> process, that instead that you may be referring to the SECOND phase of
> u-boot / coreboot.

True.

To elaborate a bit more on the different parts being discussed: I see
three somewhat separable concerns in this discussion, that a RISC-V
boot/firmware "stack" needs to address:

(a) Low-level hardware initialization, for example of the DRAM
(b) Run-time callable firmware, which (on RISC-V) implements the SBI
(c) A full-featured boot loader which can read different storage/network
media and file systems, in order to have flexibility in loading the OS.

Some examples of how these concerns are implemented:

- SiFive's FSBL implements (a), and bbl implements (b). The original
boot stack from SiFive does not implement concern (c).
- coreboot's focus is on (a), and it happens to implement* (b) on RISC-V
because that's necessary for running linux. coreboot itself is not,
and has never been a good boot loader (concern (c)). It's just good
enough to load a "payload" from its boot medium (i.e. SPI flash).
[ And on x86, coreboot provides ACPI tables, which may be considered
run-time callable firmware, and an SMI handler. ]
- ATF in the ARM world focuses on (b), but also implements (a)
sometimes, AFAIUI.
- On chromebooks, (a) is implemented by coreboot, and (c) is implemented
by [Depthcharge]. On ARM chromebooks, (b) is usually implemented by ATF.
- On coreboot+TianoCore platforms, coreboot implements (a), and
TianoCore implements (b) and (c).
- u-boot usually implements (a) and (c), and I guess it can also
implement (b)
- The [LinuxBoot] project focuses on (c), by reusing Linux's vast
support for storage/network media and file systems


Personally, I'd welcome a central or popular implementation of the SBI
(and thus, concern (b)), which can be used with coreboot or u-boot SPL,
or the vendor's proprietary thing for concern (a), and u-boot or
LinuxBoot (once kexec is ready) or something else for concern (c).
Maybe a slightly stripped-down bbl can fit that mold.

And, personally, once these three concerns are disentangled, I don't see
why a pure boot loader (concern (c)) shouldn't run in S-mode.
[ If something must run in M-mode, it's probably not a pure boot loader. ]


Thanks,
Jonathan Neuschäfer

(* In a patch that's still under review. Development/review has been
somewhat slow, sorry about that.)

[Depthcharge]: https://chromium.googlesource.com/chromiumos/platform/depthcharge/
[LinuxBoot]: https://www.linuxboot.org/
signature.asc

ron minnich

unread,
Oct 13, 2018, 12:39:46 PM10/13/18
to Tommy Murphy, RISC-V SW Dev
On Sat, Oct 13, 2018 at 5:35 AM Tommy Murphy <tommy_...@hotmail.com> wrote:
Huh? You're saying that if the sources are available you *can't* see and fix bugs if necessary? Doesn't make any sense to me...



tl;dr  sorry, this got too long :-(

you betcha. It makes all kinds of sense. This is the common misunderstanding about firmware: people expect that firmware problems are as easily fixed as in kernels. It's a whole higher level of pain.

There are many cases where source code is released, but the vendor has found out ways to make it impossible to build or install it. Either the build toolchain is not available, or the device is locked down in a way that it can't be reflashed, or some critical piece of "core IP" is left out. Just look at the last 20 years of Intel cpus. Or the Motorola phones, ca 12 years ago: all the source you want, no way you ever get to install it. Or look at BootGuard: you can build all you want, but you don't get to sign it, so if you use your build you can't have security. Or the one-time fuses in chipsets that let you set up Root Of Trust *once*, meaning among other things it's very hard to resell systems, since the eventual end user can't change the keys (I'm involved in this mess right now). There are so many barriers.

Source availability does not imply that you can build or fix things. Not at all. And, there's no guarantee that the supplied binary will match the supplied source, so just seeing source is of no real value. We've seen that too. 

The only real guarantee is that you run what you build.

But back to M mode.

M mode can do anything, without limit. For that reason, it must always be possible to audit, rebuild and install M-mode code from source. This could be made impossible with a few key decisions at the chipset level. Vendors over the decades have shown a tendency to make those decisions, as part of "adding value" (their phrase) or "vendor lockin" (the customer's phrase). 

The RISCV community has a chance to get this right, but we could also get it wrong, and I'd rather see us get it right. When vendors start arguing for common implementations,I worry that we're about to get it wrong, because frequently "common implementation" ends up meaning "baked in and unchangeable implementation." Because that's happened so many times. See: "adding value."

Further, given that there are still at least 1 BILLION x86 systems out there with all kinds of embedded exploits in runtime firmware, I now strongly believe that M mode code should be installed by the kernel, post boot, since that gets around a big problem: nobody ever updates firmware, even when the worst exploits appear, because they don't know, don't care, or it's almost impossible (a German friend at the BSI explained that Intel's mitigation software for one ME exploit won't run under Windows 10, due to W10 security checks; they had to downgrade systems to Windows 7, run the update, then reinstall WIndows 10. Care to guess how many people will do this?). 

A kernel can supply its own M mode code, so the path to critical security fixes is a reboot, not a reflash. Much simpler. One billion times more likely to happen. 

So not only do I argue that firmware should supply its own M mode code, so should kernels, and that argument is supported by decades of watching ODMs get firmware security wrong. I would list all the problems the LinuxBoot community has found in UEFI in just the last few months, but we don't have time. Let's just say it's turning out there's a lot of money in auditing x86 firmware binaries for exploits; they are being found; and let's further say that many x86 system vendors are pretty clueless about the firmware they're shipping.

This is not an argument against a standardized boot flow. This is an argument for ensuring that kernels can guarantee properties of their runtime environment, and recognizing that environment includes M mode. Further, multiple implementations frequently strengthen, not weaken, standards. See, for example, TCP/IP. Many implementations, all interoperable. Long ago (1980s) there were arguments for one TCP/IP code base, for the exact same reasons I'm seeing for a single M mode code base. It made no sense back then either.

Kernels should always have the option to not share a machine with a binary that is provided by a hardware vendor. Hence it should always be possible for a kernel to supply its own M mode code.

ron

kr...@berkeley.edu

unread,
Oct 13, 2018, 1:17:53 PM10/13/18
to Samuel Falvo II, Luke Kenneth Casson Leighton, alist...@gmail.com, sw-...@groups.riscv.org, Krste Asanovic, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx

>>>>> On Fri, 12 Oct 2018 18:00:14 -0700, Samuel Falvo II <sam....@gmail.com> said:
| I'm going to once again put forth the idea that it was a mistake for
| the RISC-V community to settle on an ITU-style standards setting
| process. While it may prove advantageous in some circumstances, I
| think it sets a bad precedent overall where viable implementations
| come second to the desire for an up-front, all-inclusive, frequently
| self-inconsistent standard with which everyone must conform. See also
| CORBA.

RISC-V standard setting process is for the ISA and closely related
issues. Software layers above can operate standards procedure as they
usually do for each software ecosystem. Some coordination between
these layers is obviously useful, but does not have to be heavyweight
(see how the calling convention was established as an example).

Setting standards for what will be baked into silicon after huge
investment is very different that setting software standards where
systems can support multiple variants and be quickly updated in the
field.

Krste

Luke Kenneth Casson Leighton

unread,
Oct 13, 2018, 1:48:15 PM10/13/18
to ron minnich, Tommy Murphy, RISC-V SW Dev
On Sat, Oct 13, 2018 at 5:39 PM ron minnich <rmin...@gmail.com> wrote:

> Or look at BootGuard: you can build all you want,
> but you don't get to sign it, so if you use your build
> you can't have security. Or the one-time fuses in chipsets
> that let you set up Root Of Trust *once*, meaning among
> other things it's very hard to resell systems, since the
> eventual end user can't change the keys (I'm involved
> in this mess right now).

nothing wrong with that, right? you get to create jobs by throwing
away the old device and buying a new one. that's how capitalism
works, right?

https://news.slashdot.org/story/18/10/12/233230/tech-suffers-from-lack-of-humanities-says-mozilla-head
(and other links, this is only just the latest one)

Luke Kenneth Casson Leighton

unread,
Oct 13, 2018, 2:11:52 PM10/13/18
to Krste Asanovic, Samuel Falvo II, alist...@gmail.com, RISC-V SW Dev, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx
On Sat, Oct 13, 2018 at 6:17 PM <kr...@berkeley.edu> wrote:

> Setting standards for what will be baked into silicon after huge
> investment is very different that setting software standards where
> systems can support multiple variants and be quickly updated in the
> field.

absolutely, i mentioned the example of the DDR3/4 initialisation
sequence as just one of the examples which make it absolutely
essential to keep the boot ROM to the absolute bare minimum, and using
as generic a range of protocols as can possibly be gotten away with.

expecting a ROM-based bootloader to run SATA, PCIe, IDE or other
complex bus is absolutely asking for trouble. a boot peripheral has
to be *absolutely* dead simple, and implementable - FULLY - in a few
*HUNDREDS* of lines of code. not thousands. not tens of thousands.
a few HUNDRED lines of code.

then, also, the actual boot media has to be incredibly popular and
unlikely to change. SPI, QuadSPI, SDMMC (which has fallback to SPI),
eMMC (being ultimately all based on SPI) are the usual ones that have
tens of thousands of devices available for decades, from dozens of
manufacturers.

any processor that offers those protocols as part of its boot ROM, and
can load at least one 16k block from it, is going to be pretty safe
for decades of life. in the embedded world, I2C, UART and USB1.1 (if
the SoC actually *has* USB) are typically added as well.

*all* of those protocols are world-wide popular, stable, decades-old,
not controlled by any one corporation... implication: they're not
going away any time soon.

now for an example of where a massive company got it wrong... without
doing anything wrong.

nintendo's gameboy was a nightmare scenario (for them). to cut down
the cost they didn't use a separate SPI/QSPI bootloader IC (because
they can actually be very expensive, relatively speaking, and a
mass-volume product is extremely price-sensitive on the BOM).

so they booted directly from NAND.

TSSOP-48 ONFI NAND ICs change so fast relatively speaking that the
*entire gameboy product* had to be end-of-lifed ... why? because the
HARD-CODED bootloader in its SoC's ROM had a series of fixed
parameters, and the NAND ICs supporting those exact and precise
parameters were *no longer being manufactured* [1]

so that's what products - and fabless semiconductor companies - can
face if they get the design of the bootloader wrong.

l.

[1] those parameters include voltage, block size, checksum size,
speed, density... and no, the JEDEC standards are not properly
supported by manufacturers. every single NAND IC that has different
parameters is completely and utterly incompatible. it's pretty
insane.

Luke Kenneth Casson Leighton

unread,
Oct 13, 2018, 2:29:18 PM10/13/18
to ron minnich, Tommy Murphy, RISC-V SW Dev
On Sat, Oct 13, 2018 at 5:39 PM ron minnich <rmin...@gmail.com> wrote:

> Vendors over the decades have shown a tendency to make those
> decisions, as part of "adding value" (their phrase)

"trust"

> or "vendor lockin" (the customer's phrase).

"treachery".

> So not only do I argue that firmware should supply its own
> M mode code, so should kernels, and that argument is supported
> by decades of watching ODMs get firmware security wrong.

interesting. and insightful.

> This is not an argument against a standardized boot flow.

understood

> This is an argument for ensuring that kernels can guarantee
> properties of their runtime environment, and recognizing that
> environment includes M mode.

so what you're saying is: once one phase of a bootloader only permits
applications (including the next phases in a bootloader) to be run in
S-Mode (only), the only guarantees that can be made about security are
those of the S-Mode "guest"... which is to say... absolutely no
security guarantees at all.

[historically, from the examples you gave, and many more that would
be far too numerous to list, we've seen manufacturers persistently and
consistently fail for literally decades to even *comprehend* security.
because it costs money to get it right].

that in turn results in people throwing product out and buying new
ones... with no guarantee that the new product is not designed along
the exact same deeply-flawed design paradigm as the old one they just
put into landfill.

> Kernels should always have the option to not share a machine
> with a binary that is provided by a hardware vendor. Hence it
> should always be possible for a kernel to supply its own M mode code.

so this is a very insightful and reasonable argument for ensuring
that a boot flow does *not* preclude the *possibility* of running in
M-Mode.

realistically the cynic in me says that this won't stop vendors from
creating RISC-V processors that lock out M-Mode... they're not going
to learn... however if there's at least some products released early
enough, leading the way and showing that successful product does *not*
have to be treacherous to its users...

l.

Samuel Falvo II

unread,
Oct 13, 2018, 2:39:08 PM10/13/18
to Krste Asanovic, Luke Kenneth Casson Leighton, alist...@gmail.com, sw-...@groups.riscv.org, Christoph Hellwig, Anup Patel, atish patra, Padmarao Begari, Jonathan Neuschäfer, phi...@hug.cx
On Sat, Oct 13, 2018 at 10:17 AM <kr...@berkeley.edu> wrote:
> Setting standards for what will be baked into silicon after huge
> investment is very different that setting software standards where
> systems can support multiple variants and be quickly updated in the
> field.

Hi Krste,

This is correct, which is why I clarified that ITU-style has value in
certain circumstances. I'm just expressing a fear that this approach
will bleed over, if you will, into the software world where things
that can be allowed to be made more reactive to market forces *should*
be. Apologies in advance if I'm not (or still not) expressing myself
clearly.

For example, there's a large concern over standardizing the boot flow
for everything from wristwatches to weather-predicting supercomputer
clusters and literally everything in between. This is a noble goal,
but I also feel it's something of a fools' errand, for reasons that
both Ron and Luke expressed far more gracefully than I ever could.
More frequently than not, when we end up in debates like these, a lot
of time and effort is spent justifying one's position with logic and
reason but not necessarily with concrete implementations to make
objective comparisons against. I'm much more inclined to agree with
Ron and Luke because both have their names in code repositories that I
have ready access to and whose projects are steeped in the
bootstrapping process for a plurality of architectures, not just x86
and not just ARM, and certainly not for any NDA-covered projects that
nobody can talk about.

This is why I advocate for an implementation-based, bottom-up
standards track process. I believe working code and field-deployments
are far more valuable than speculation, however well-reasoned they may
be. Except for code signing and the semantics of the GPT, much of the
benefit of UEFI existed long before it came along. Per-partition
filesystems, for example, existed in OS/2's "micro-filesystem
drivers", for instance, and these existed with a BIOS barely more
sophisticated than CP/M's. OpenFirmware gave us both an interactive
boot-time shell as well as the venerable Device Tree standard we still
use today.

EFI technically lost in the market; the only reason we got UEFI today
is because Intel and Microsoft effectively forced it up on us by
mandating all new motherboards ship with it, and Windows refused to
boot on a system without it past some cut-off date. In case few
remember, OpenFirmware was actually the leading alternative to PC BIOS
back then, and was gaining up-take *fast* in a variety of different
markets, including data-center server racks.

Bluntly, I don't want to see another UEFI happen.

However, if the community does prefer a top-down approach to
establishing some manner of standards, then I'd like to raise a point
as an independent observer. This conversation has two conflicting
points to it: the ARM world is an abject failure because its
bootstrapping conventions are a total mess, and the ARM world is to be
taken as something of a model because ARM Ltd. defined something
called ATF and is the way of the future. If I'm reading this thread
correctly, these are mutually exclusive viewpoints, both of which are
factoring into the desire for a unified boot flow for the RISC-V
community (hence my mention of self-inconsistent when referring to
top-down standards). Before we get into standardizing any kind of
boot flows, I think we need to quantitatively define what "failure"
means in the context of a processor architecture and ISA which has all
but thoroughly dominated the RISC market and has done so without the
help of trusted firmware for the majority of its existence.

Karsten Merker

unread,
Oct 13, 2018, 4:32:19 PM10/13/18
to alist...@gmail.com, kr...@berkeley.edu, RISC-V SW Dev, h...@lst.de, an...@brainfault.org, atis...@gmail.com, padma...@gmail.com, j.neus...@gmx.net, phi...@hug.cx
On Fri, Oct 12, 2018 at 02:12:08PM -0700, alist...@gmail.com wrote:

> I agree that we shouldn't trust vendor firmware. It is always buggy at best.
>
> The problem that I see with starting coreboot or u-boot (the last stage
> bootloaders) in M mode is that we then end up with duplicated work between
> the projects.
>
> This is where a model like Arm Trusted Firmware fits in well. We can see
> something like this:
> 1. CPU boots up and ROM does it's thing
> 2. ROM boots into an open source firmware (ATF equivalent) in M mode
> - This will end up being pretty SoC specific
> 3. The firmware boots into u-boot or core boot in S mode (or H mode when
> that exists)
> - Hopefully this can then be as generic as possible
> 4. From there we boot the system
>
> The advantage I see with this is that we only need to implement the M mode
> functionality once and then coreboot, u-boot and anything else can utilise
> it. It also keeps bigger later stage boot loaders out of M mode which has a
> mild security benefit of smaller attack surfaces.
>
> The other aim is to allow a single RISC-V image to boot on a variety of
> different SoCs. The more generic we can make the last stage boot loaders
> the more achievable this is.

Hello,

I'd like to chime in here with some thoughts about
firmware/bootloader interfaces from a Linux distribution's point
of view. This isn't strictly about M-mode vs. S-mode, but
nonetheless related. I've changed the subject of the email to
show that this is somewhat of a side-thread, but IMHO important
nonetheless in the context of the existing discussion.

As an ideal, Linux distributions would like to see a situation
where (provided that the hardware is technicially capable of
running the distribution), a single installer image can be used
on all systems and the resulting installed system "just works".
Having to build per-machine-type installer images is a nightmare
and doesn't scale - I have been doing that for ARM-based systems
for quite some time, so I can speak from first-hand experience
here.

In reality unfortunately this ideal isn't achieved anywhere, not
even on x86 which tends to be commonly mentioned as the platform
where "things just work". Before laying out what I would like to
see standardized on RISC-V, I'll try to describe some of the
experiences that Debian developers have made on x86 and ARM as
those are usually the two platforms that come up as examples in
discussions related to boot interfaces. This hopefully helps to
explain why I would like to see certain things standardized in a
certain way.

- x86 has classic BIOS and UEFI, and the latter in 32bit and
64bit variants, sometimes in severely broken hybrid forms
(64bit CPUs combined with 32bit-only UEFI implementations that
require dirty hacks to boot a 64bit OS).

Even when declaring the classic BIOS to be legacy and just
concentrating on UEFI with ACPI, there are still lots of broken
implementations in the wild that require tons of workarounds on
multiple levels, be it the plethora of in-kernel workarounds
for broken ACPI implementations or the broken boot handling in
quite a number of existing UEFI implementations that make the
theoretical advantages of having a proper bootloader interface
defined in the UEFI specification often moot in practice. Add
to that the problem that in practice many UEFI and ACPI
implementations don't do what the spec says but instead do what
works with Windows. As a consequence, the Linux kernel had to
move to emulating the Windows behaviour instead of doing things
the way that the spec actually provides because otherwise it
wouldn't work properly on real-world systems. And then there
are of course such great things as real-world UEFI
implementations that kill themselves (as in converting a
working Laptop to a paperweight) as soon one actually uses the
interfaces that the UEFI spec provides but that aren't commonly
used by Windows. Remember the "Linux kills Laptops" uproar?
In fact that was "broken UEFI kills Laptops" - Linux called a
UEFI function to write a crash dump into a UEFI NVRAM variable
in a perfectly spec-compliant way, but a number of existing
UEFI implementations didn't properly handle large writes
resulting in a tight loop on UEFI initialization at the next
reboot, turning the Laptop into a paperweight. Windows doesn't
do large writes by default, so the problem usually didn't show
up under Windows. You can probably imagine the reactions that
Linux distributions faced with people claiming "You have killed
my laptop!".

- On armhf (32bit) platforms, there hasn't traditionally been a
standard firmware interface for a long time. The SoC bootroom
usually loads a bitstream from a vendor-specific location
(either from SPI NOR flash or from a vendor-defined MMC block
and executes that, whatever there is. In the embedded space,
vendors have usually forked u-boot and hacked in support for
their SoC, defining a vendor-specific method for loading the
kernel, and left it at being incompatible with the rest of the
world.

In mainline u-boot there have been efforts to standardize the
boot interface that is provided to the later stages since IIRC
2014. This goes under the label of "distro bootcommand" and
provides a vendor-neutral boot interface to e.g. Linux
distributions. With this, one can boot the same installer
image on all armhf platforms that use a u-boot with "distro
bootcommand" support. All platforms that are newly added to
mainline u-boot are expected to implement "distro bootcommand"
support and many existing platforms have been converted to it
in the meantime.

On arm64 there are two worlds - the embedded space, using
u-boot+device-tree (usually combined with the "ARM Trusted
Firmware" (ATF)), and the server space with either UEFI+ACPI or
UEFI+device-tree. Pure enterprise distributions like RHEL
usually only target UEFI+ACPI on arm64 and ignore everything
else. Community distributions like Debian usually target all
of the aforementioned options: u-boot+device-tree,
UEFI+device-tree and UEFI+ACPI.

ATF is published under BSD license by ARM, but there are two
big problems with ATF in practice:

- There are two incompatible ATF interface definitions, an
older and a newer one. Unfortunately some SoC vendors use
the old one while others use the new one :-(.

- The "upstream" ATF originally didn't have support for any
hardware besides some reference systems from ARM (similar to
the situation with Tianocore in the UEFI world). SoC vendors
usually used the BSD-licensed ATF as the base for a
closed-source fork that included hardware support for their
specific SoC and didn't contribute back anything to the
upstream ATF, and of course also often haven't pulled in any
later ATF bug fixes. In the meantime, open-source developers
have contributed (partially reverse-engineered) support for
some arm64 platforms to upstream ATF, but that are still
rather few.

The problem that Linux distributions have with the
closed-source vendor forks of ATF is that they a) are often
broken and b) aren't stored in some form of onboard memory
(like an SPI NOR flash) but are shipped by the vendors as
part of their SD card boot images. In consequence, Linux
distributions would have to include those proprietary ATF
builds in their boot images, which they often can't for legal
(nobody except the SoC vendor has the right to distribute the
proprietary code) and/or policy reasons (Debian doesn't ship
any software that isn't Open Source as part of the main
distribution). Exactly the same problem exists with the code
that initializes the DRAM controller. Like with ATF the
actual implementation is usually proprietary (at least until
some open-source developer has managed to reverse-engineer
the way the DRAM controller needs to be set up) and in
contrast to PCs, where the DRAM initialization is always done
by code that is stored in an onboard flash chip, on arm
platforms the DRAM init code is often shipped as part of the
SD card boot image. In consequence, the corresponding
systems cannot be supported by Debian.


Alex Graf from SUSE has been working on making it possible to
install distributions that expect UEFI, but can use device-tree
instead of ACPI (which is a configuration that is covered by
the UEFI spec) on systems using u-boot as their boot firmware
by adding a UEFI emulation layer on top of u-boot. The
important point hereby is that this layer implements the
UEFI boot services, but not the UEFI runtime services, i.e.
it can be used to load a UEFI-GRUB, but it cannot be used
with operating systems that expect to be able to call
UEFI functions during runtime, like writing to UEFI NVRAM
variables, so this is not a replacement for a full UEFI
implementation.


For the "low-level" part of system firmware (i.e. the part that
handles stuff like initial voltage regulator setup, DRAM
initialization, etc.), I would like to see the following things
standardized for RISC-V as part of the unix platform spec:

* The platform vendor must provide DRAM initialization code
with a standardized interface in an on-board memory, so that
an OS provider doesn't have to ship the code as part of the
OS image, but can call it when needed.

The reason for this is outlined above in the section about
ATF - most DRAM init code is only provided under NDA by the
DRAM controller IP vendor, so OS providers may not be able to
use it if it isn't already available on the system.

* Define a standard partition table layout that is taken into
account by anything on the system, including the boot
firmware. My proposal would be GPT because it's flexible,
supported by about everything and the kitchen sink, and
supports large volumes (which is a problem with the classic
BIOS partition table format).

* If a SoC supports loading its raw firmware bitstream from a
block device (typically SD/eMMC) by loading a certain amount
of bytes from a certain offset, define _one_ standard offset
and length where it is to be loaded from, making sure that it
is compatible with the chosen partition table layout. The
reason for that is that without such a definition, one would
again risk incompatibilities with OS, userland applications
and possible alternative firmwares.

SoCs can usually boot their firmware from multiple places,
selectable by external pins or in a fixed order. When using
a fixed order, usually removable storage (SD card) has
highest priority, then comes eMMC, then SPI NOR flash. This
is really nice because it allows to easily "unbrick" a broken
system, e.g. if flashing a new firmware into the onboard NOR
flash has failed and the system doesn't come up anymore.
Code is only loaded from a device if it has a valid header
(usually some magic number and a CRC), so if there is no
valid header on removeable media because nobody has installed
a firmware image there, the boot process simply continues
with the in-flash vendor code. If there is a valid firmware
image on the block device, it gets loaded and executed. This
allows either to fix the broken NOR flash contents or just
use this firmware as a flexible boot-time replacement for the
firmware in the NOR flash. Usually the area the SoC loads
its firmware from is somewhere after the partition table but
below 1MB (where normally the first partition starts for
BIOS-style and GPT-style partition tables). In the arm
world, a number of SoCs have placed their firmware location
as a 32kB block at 8k offset from the beginning of the
device. Unfortunately while this works for systems using
BIOS-style partition tables, it breaks for systems using
GPT-style partition tables, because those are larger. The
maximum length of the firmware boot block needs to be
specified as well, because firmware like u-boot usually uses
the space between the first-stage boot image and the start of
the first partition for its own (non-SPL) code and data, so
it needs to know from which offset onwards blocks can be used
without interfering.


For the "high-level" part of boot firmware (the part that
implements the "user-facing" part of the boot process and
implements methods to boot from SATA disks, network interfaces,
provide system configuration options, etc.), finding a reasonable
standard that covers all use cases is really difficult.

I can understand the "simply go with UEFI and ACPI" argument that
Richard Jones had made in his talk during the Barcelona workshop,
but only when limited to the premise that it has been made under,
i.e. "RedHat is only interested in the server market and nothing
else".

One can surely with good reason argue "UEFI+ACPI is used on x86
and arm64, do the same for RISC-V" as there is existing support
for UEFI-based systems in many operating systems and bootloaders.
On the other hand, there is a lot to be said against UEFI+ACPI,
including that it is way too complex (not only for the embedded
Linux use case, but also too complex for reasonably creating
secure and bug-free implementations in the server space), has a
large attack surface (see Ron Minnichs talks), and reality shows
that things that are fine in theory are way too often broken in
real-world implementations (which I believe is also to a certain
amount the fault of UEFI's enormous complexity).

ACPI is a no-go from my point of view, I've seen way too many
broken ACPI implementations in the PC world that cannot be fixed
by the user and in practice usually don't get fixed by the
manufacturers. Things are different if you "live" in the
high-end server market (both on x86 and on ARM), but for
commodity hardware my experience has been rather devastating.
If one would want to standardize on using UEFI at all, then IMHO
only in the form of UEFI+device-tree.

For the embedded Linux use case, UEFI+ACPI isn't even worth
a single thought - from my experience with the embedded Linux
world on ARM platforms I am sure that there isn't the chance of
a snowball in hell that platform vendors for the embedded Linux
market would be willing or capable of providing a reasonable
UEFI+ACPI implementation for their systems.

U-boot on the other hand is a rather nice firmware and bootloader
for embedded systems, but it is definitely not designed for the
server world. For the embedded space, u-boot can surely provide
a reasonably easy and platform-neutral interface for booting an
operating system.

The obvious compromise could indeed be an interface standard that
defines some stripped-down UEFI boot services as the common
denominator. This would allow hardware vendors to choose between
various existing solutions, i.e. between using either some
Tianocore-based UEFI implementation with their own HAL, using
coreboot for the lower layer with a Tianocore-based payload
providing the UEFI interface to the outside, or using a
u-boot-based implementation with u-boot's UEFI emulation layer.
The latter would provide the standardized interface for embedded
use cases without significant cost (the code size increase due to
the emulation layer is rather negligible).

Comments welcome :).

Regards,
Karsten
--
Gem. Par. 28 Abs. 4 Bundesdatenschutzgesetz widerspreche ich der Nutzung
sowie der Weitergabe meiner personenbezogenen Daten für Zwecke der
Werbung sowie der Markt- oder Meinungsforschung.

kr...@berkeley.edu

unread,
Oct 13, 2018, 5:29:53 PM10/13/18
to Samuel Falvo II, Krste Asanovic, Luke Kenneth Casson Leighton, alist...@gmail.com, sw-...@groups.riscv.org, Christoph Hellwig, Anup Patel, atish patra, Padmarao Begari, Jonathan Neuschäfer, phi...@hug.cx

[ Not particularly responding to Sam, but more to whole thread ]

In this thread, as is common when engaging a wide community of smart
experienced people, many of the disagreements are a result of people
having different, even if possibly overlapping, subsets of the
computing universe in mind. It would be helpful if everyone qualified
their position with the systems they're assuming (microcontroller with
no external DRAM, or servers with 8 sockets and terabytes of DRAM) and
especially if we could introduce a common taxonomy/vocabulary, so
people would be less likely to talk past each other.

I think we can all agree that reducing unnecessary duplication of
effort is a good thing, and both standards and open-source code are
great and complementary ways to achieve that goal.

I think we _should_ all agree that there is necessary fragmentation
because different platforms have different needs. While some systems
might never trust vendor-supplied M-mode software, there are other
systems that would never trust a Linux kernel to install/execute
M-mode code. Similarly, in structuring code bases, some people might
want their Linux to know nothing of what's underneath to support
flexible virtualization, while others want deep control of the
hardware, e.g., for real-time concerns. Other systems will never run
any Unix, but might still need to run many of the same steps during
boot and need similar runtime support afterwards.

The big issue underlying most peoples' positions is that structuring a
single set of interfaces and a single code base to support all these
different use cases is not easy and perhaps not possible with current
system programming tools (beyond "copy-and-paste" reuse), and so
they'd like to make sure their use-case is the one supported by the
shared software effort.

I would hope the community could work towards figuring out how to
share most of the code even if each piece of code runs in different
places in a given boot/runtime system.

I would also ask to separate out concerns about "binary blob" device
drivers from defining boot flow/runtime components, as that is a
mostly orthogonal problem to be solved.

RISC-V should reduce the concerns about monopolistic hardware vendor
policies, as there is no monopoly and there can be many independent
hardware vendors catering to different markets, and using very
different CPU designs (instead of many different vendors using the
same few CPU designs). Because there will be many more RISC-V vendors
than other architectures, there is a greater danger of fragmentation,
so it is important to try to develop a small set of standards with
supporting open implementations, and hopefully open compliance suites,
that cover the common use cases.

Krste

kr...@berkeley.edu

unread,
Oct 13, 2018, 5:37:29 PM10/13/18
to Karsten Merker, RISC-V SW Dev

Karsten,

Thanks for the detailed writeup of past experience. I'm sure many
will find this very helpful to review. Palmer is starting up an
effort to define RISC-V Unix platform standards and hopefully this can
feed into that.

Krste


>>>>> On Sat, 13 Oct 2018 22:32:14 +0200, Karsten Merker <mer...@debian.org> said:
[..]

Luke Kenneth Casson Leighton

unread,
Oct 13, 2018, 7:12:53 PM10/13/18
to Karsten Merker, alist...@gmail.com, Krste Asanovic, RISC-V SW Dev, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx
from what you write below (i read down then replied bottom-up), i
agree entirely. really appreciated your input here, karsten. you
have first-hand knowledge of the areas i've covered at various
different points over the past 20 years, and haven't really been able
to descrribe to people here quite what the reality is, and you manage
to do that extremely well.

> As an ideal, Linux distributions would like to see a situation
> where (provided that the hardware is technicially capable of
> running the distribution), a single installer image can be used
> on all systems and the resulting installed system "just works".

i know... *sigh*.

> Having to build per-machine-type installer images is a nightmare
> and doesn't scale - I have been doing that for ARM-based systems
> for quite some time,

i _did_ try to warn people, back when devicetree was touted as "the
solution!!!! to all the ARM world disparate pain!!!!!", that with
literally tens of thousands of sensor and peripheral ICs used in
products, all that would happen with devicetree is that the
programming would move *from* c source to *to* devicetree source code.

the experience of PowerPC having so successfully adopted device-tree
(because it's near-identical to x86 in terms of usage, i.e. PCIe-based
motherboards) was unfortunately quoted as a "counter-example", and
nobody listened.

whoops.

so, my sincere apologies: i know the pain that you describe, all too
well, and by the very nature of what SoCs are for, what hardware
interfaces they have (non-self-describing bus architectures like UART
and SPI) and what they do, it's really *really* just not "fixable".

i cannot emphasise that enough.


> so I can speak from first-hand experience here.

cool. that's very helpful, karsten. i have a hybrid of first-hand
experience and second-hand knowledge in the areas you comprehensively
describe, below, so can chip in and give insights and also point out
where things might work, and where they definitely won't (and why).

> - x86 has classic BIOS and UEFI, and the latter in 32bit and
> 64bit variants, sometimes in severely broken hybrid forms
> (64bit CPUs combined with 32bit-only UEFI implementations that
> require dirty hacks to boot a 64bit OS).

whoops.

> - On armhf (32bit) platforms, there hasn't traditionally been a
> standard firmware interface for a long time. The SoC bootroom
> usually loads a bitstream from a vendor-specific location
> (either from SPI NOR flash or from a vendor-defined MMC block
> and executes that, whatever there is.

yep. allwinner's bootrom has been reverse-engineered, and is clearly
evidently based on u-boot. it checks for SPI, and eMMC, and SDMMC, on
various different ports, and, finally gives up and drops into USB-OTG
"Front-end Loader" mode.

> In the embedded space,
> vendors have usually forked u-boot and hacked in support for
> their SoC, defining a vendor-specific method for loading the
> kernel, and left it at being incompatible with the rest of the
> world.

yep. and it works extremely well for them. many chinese fabless
semiconductor companies are now billion-dollar operations by virtue of
this "strategy".

> For the "low-level" part of system firmware (i.e. the part that
> handles stuff like initial voltage regulator setup, DRAM
> initialization, etc.), I would like to see the following things
> standardized for RISC-V as part of the unix platform spec:
>
> * The platform vendor must provide DRAM initialization code
> with a standardized interface in an on-board memory, so that
> an OS provider doesn't have to ship the code as part of the
> OS image, but can call it when needed.

unfortunately they can't do that, for reasons i mentioned earlier, in
other messages: there may be some cross-over.

DRAM initialisation is far too complex, there's not enough
standardisation (DDR2 is different from DDR3 is different from LPDDR2
is different from LPDDR3 is different from DDR4 is different from
LPDDR4 and that's not taking the speeds, layouts (x16, x8, 2x8, x32,
2x16) into account.

you can't possibly predict whether any given *THIRD PARTY* RAM chip
supplier - in the near to immediate future - is going to work.
fabless semi companies actually have to provide a "compatibility list"
of combinations that they've tested. these lists run to over a
hundred entries!

also, if the ROM gets it wrong, on the impedance matching, because
there's some tolerance in the PCB layout due to tight board size,
thickness and number of layers constraints, the ODM will just write
the entire SoC off as a total loss and pick another one.

whilst it would indeed solve the libre-licensing issues, it's just
far, far too risky to consider.

> The reason for this is outlined above in the section about
> ATF - most DRAM init code is only provided under NDA by the
> DRAM controller IP vendor, so OS providers may not be able to
> use it if it isn't already available on the system.

i know. it's a f*****g nuisance. it takes months if not years for
reverse-engineered libre support for various hardware to be available,
by which time the product's dead as far as the manufacturer is
concerned.

this is why i said it will take fabless semi companies putting their
foot down and saying, "um, no, sorry" on this. they can't _possibly_
expect companies to violate the GPL with u-boot and the linux kernel.
despite many of their customers being chinese OEMs who don't give a
flying fuck about Copyright Law, "well our Chinese customers don't
have a problem, so why do you" isn't an acceptable answer!


> * Define a standard partition table layout that is taken into
> account by anything on the system, including the boot
> firmware. My proposal would be GPT because it's flexible,
> supported by about everything and the kitchen sink, and
> supports large volumes (which is a problem with the classic
> BIOS partition table format).

that's almost certainly not going to fly on a first-stage bootloader
(cf: u-boot SPL). as you know they're often looking at under 16k.
loading from a specific raw sector (offset) is about as good as it
gets.

plus, don't forget: with NAND and NOR, you don't have the luxury of
knowing if the IC has built-in error-correction. or, more to the
poin, you *absolutely must not* assume that it has.

this is why for example the allwinner bootrom just does a for-loop,
increasing the offset by fixed amounts, requesting a single block and
looking for the signaure "eGON" at the front. if success, execute.

basically, a standard partition at the front of a raw NAND flash
device is a sure-fire guaranteed way to have a high percentage of
returns on customer product.

> * If a SoC supports loading its raw firmware bitstream from a
> block device (typically SD/eMMC) by loading a certain amount
> of bytes from a certain offset, define _one_ standard offset
> and length where it is to be loaded from, making sure that it
> is compatible with the chosen partition table layout. The
> reason for that is that without such a definition, one would
> again risk incompatibilities with OS, userland applications
> and possible alternative firmwares.

nggggh this *might* be ok to do. i hesitate to say it's a good idea
because of the fact that some raw NAND / NOR devices have different
block sizes, and it may hypothetically be better for an SoC to read on
the boundary of the NAND flash IC, *not* a specific standards-defined
size that doesn't take the possible ICs into account.

TSSOP-48 NAND ICs are a frickin nuisance, however unfortunately the
reduced cost (and sizes they are available in) is a compelling
argument for their use over eMMC. eMMC i haven't been able to easily
find anything less than 4GB, whereas Winbond does 128k TSSOP-48 NAND
ICs and will continue to do so for a long time yet.


> SoCs can usually boot their firmware from multiple places,
> selectable by external pins or in a fixed order. When using
> a fixed order, usually removable storage (SD card) has
> highest priority, then comes eMMC, then SPI NOR flash. This
> is really nice because it allows to easily "unbrick" a broken
> system, e.g. if flashing a new firmware into the onboard NOR
> flash has failed and the system doesn't come up anymore.

*sigh* the RK3288 got this horribly wrong, sadly. the first boot
device is hard-fixed in the boot ROM as eMMC. consequently, if the
eMMC is ever corrupted, you're f*****d. brick. and no way to
recover.

on the *other* hand, if that eMMC is *empty* (zeros), *then* it will
boot from SD/MMC no problem. but.. write a single byte to it, and
you're done. landfill.

so getting this order right is *important*.

> Code is only loaded from a device if it has a valid header
> (usually some magic number and a CRC), so if there is no
> valid header on removeable media because nobody has installed
> [ ...]
> BIOS-style and GPT-style partition tables). In the arm
> world, a number of SoCs have placed their firmware location
> as a 32kB block at 8k offset from the beginning of the
> device. Unfortunately while this works for systems using
> BIOS-style partition tables, it breaks for systems using
> GPT-style partition tables, because those are larger.

yep... see above as to a possible reason why they do that. although
to be honest, it's probably more down to the fact that BIOS-style
partition tables are dead simple but also "were there first".

i have to say: maybe the people who developed the GPT-style
partitions should have consulted people in the embedded industry
before going ahead with such a layout??

also, if you only *have* a 4MB SPI NOR boot iC, if you *absolutely*
have to have a GPT-style partition table (because it's part of the
"standard"), that's ... that's... 25% of the frickin SPI NOR IC
*wasted* on a frickin *partition* table for god's sake! what OEM in
their right mind would *deliberately* design any product that was so
cost-uncompetitive, when $0.05 on the BOM could mean the difference
between market success and failure??


> The
> maximum length of the firmware boot block needs to be
> specified as well, because firmware like u-boot usually uses
> the space between the first-stage boot image and the start of
> the first partition for its own (non-SPL) code and data, so
> it needs to know from which offset onwards blocks can be used
> without interfering.

yep. seen that hard-coded into vendor releases. works great. not
nice to maintain, but "it works".

> For the "high-level" part of boot firmware (the part that
> implements the "user-facing" part of the boot process and
> implements methods to boot from SATA disks, network interfaces,
> provide system configuration options, etc.), finding a reasonable
> standard that covers all use cases is really difficult.

thank god. someone who actually understands the reality of booting
in the embedded space (and more).

> I can understand the "simply go with UEFI and ACPI" argument that
> Richard Jones had made in his talk during the Barcelona workshop,
> but only when limited to the premise that it has been made under,
> i.e. "RedHat is only interested in the server market and nothing
> else".

yyep. there aren't enough people around who actually understand how
completely and utterly different embedded SoCs really are. "surely
the x86 style solution should be enough for everybody, right???"

errr... no :)

> One can surely with good reason argue "UEFI+ACPI is used on x86
> and arm64, do the same for RISC-V" as there is existing support
> for UEFI-based systems in many operating systems and bootloaders.
> On the other hand, there is a lot to be said against UEFI+ACPI,

it's funny... the lesson that comes through loud and clear from what
you write, here, karsten, is every single implementation of a "boot
standard" that's been described has consistently been a total
nightmare. ATF, UEFI, ACPI, *every* single attempt at a standard is
bug-ridden security-unconscious bloatware!

and yet... people have been complaining "the ARM embedded world is
hell! surely it doesn't have to be that complex!!! surely it can be
fixed with a standard!!!!"

by contrast, vendor-proliferated u-boot... um... "just works"...

there's a lesson there. i'm not sure what it *is* at this point, but
there's a lesson there :)


> ACPI is a no-go from my point of view, I've seen way too many
> broken ACPI implementations in the PC world that cannot be fixed
> by the user and in practice usually don't get fixed by the
> manufacturers. Things are different if you "live" in the
> high-end server market (both on x86 and on ARM), but for
> commodity hardware my experience has been rather devastating.
> If one would want to standardize on using UEFI at all, then IMHO
> only in the form of UEFI+device-tree.
>
> For the embedded Linux use case, UEFI+ACPI isn't even worth
> a single thought - from my experience with the embedded Linux
> world on ARM platforms I am sure that there isn't the chance of
> a snowball in hell that platform vendors for the embedded Linux
> market would be willing or capable of providing a reasonable
> UEFI+ACPI implementation for their systems.

nope. or, yup, i agree: not a chance. you can't tell ingenic, for
example, who produce $2.50 1ghz MIPS processors, and now smartwatch
processors with embedded 32MB LPDDR2 RAM, all-in at under $3.50, that
they absolutely have to use ACPI. being chinese you won't find out
their true reaction: the fact that they're under enormous size and
cost pressure in the smartwatch market from their customers should
allow you to infer their response, without needing to contact them.

> U-boot on the other hand is a rather nice firmware and bootloader
> for embedded systems, but it is definitely not designed for the
> server world. For the embedded space, u-boot can surely provide
> a reasonably easy and platform-neutral interface for booting an
> operating system.

despite - or in fact _because_ of - the massive forking and
proliferation. it's a double-edged sword.

> The obvious compromise could indeed be an interface standard that
> defines some stripped-down UEFI boot services as the common
> denominator.

from experience of witnessing the proliferation of u-boot (and
understanding why that is) i'd honestly say that even this fills me
with... what's the word... alarm / skepticism / cynicism / realism /
other.

to stand a chance in the embedded space, a proposal has to be
justified in terms of code size, boot speed, and, ultimately, sale
price of the end product.

anything that increases the final sale price (not of the SoC, not of
the product to the users, but the product - the *full* product - to
the *OEMs*) is automatically going to be discared.

so, if part of the proposal to include UEFI is a mandatory
requirement that imposes an extra $1.50 on the BOM because certain
types of non-UEFI-compliant external boot peripherals are also
mandatory, and they're more expensive, or if the fabless semi company
has to shell out for hardware encryption RTL, significantly increasing
NREs in the process, you can kiss any such proposal's acceptance
goodbye.

> This would allow hardware vendors to choose between
> various existing solutions, i.e. between using either some
> Tianocore-based UEFI implementation with their own HAL, using
> coreboot for the lower layer with a Tianocore-based payload
> providing the UEFI interface to the outside, or using a
> u-boot-based implementation with u-boot's UEFI emulation layer.
> The latter would provide the standardized interface for embedded
> use cases without significant cost (the code size increase due to
> the emulation layer is rather negligible).

that's good... the real key question is: does it impose on the BOM?
does it prevent and prohibit the OEMs from choosing specific boot
storage devices? is there a hardware-related protocol that is an
absolute requirement to support u-boot's UEFI layer? is there a
hidden cost in terms of the amount or format of storage that UEFI
requires, which would impact the sale price of the product?

also, now that i've read the bits about GPT, is UEFI *only* possible
on top of GPT? and: is there expected to be encryption or signing,
*without* which firmware is not loadable? because if so, that's a
"no" because it imposes a hardware (and development) cost on the
vendor.

i trust that people reading this now have had their eyes opened a bit,
thanks to these questions, spurred by karsten's extremely valuable
response. the embedded (RV32, <48k and <1MB) world hasn't even begun
to be discussed in depth, either, yet.

i trust that people now see why i said that restrictive
standardisation in the area of boot loading and boot initialisation is
a near-impossible task, and why it is actually in most cases even
undesirable.

it really is far more complex and comprehensive than most people ever
dream it could possibly get.... and yet there are good reasons why it
is that way.

l.

Samuel Falvo II

unread,
Oct 13, 2018, 8:06:51 PM10/13/18
to Luke Kenneth Casson Leighton, Karsten Merker, alist...@gmail.com, Krste Asanovic, sw-...@groups.riscv.org, Christoph Hellwig, Anup Patel, atish patra, Padmarao Begari, Jonathan Neuschäfer, phi...@hug.cx
Sorry, this was originally intended to be sent to the list as a whole,
not to one person.

On Sat, Oct 13, 2018 at 4:12 PM Luke Kenneth Casson Leighton
<lk...@lkcl.net> wrote:
> yep... see above as to a possible reason why they do that. although
> to be honest, it's probably more down to the fact that BIOS-style
> partition tables are dead simple but also "were there first".
>
> i have to say: maybe the people who developed the GPT-style
> partitions should have consulted people in the embedded industry
> before going ahead with such a layout??
>
> also, if you only *have* a 4MB SPI NOR boot iC, if you *absolutely*
> have to have a GPT-style partition table (because it's part of the
> "standard"), that's ... that's... 25% of the frickin SPI NOR IC
> *wasted* on a frickin *partition* table for god's sake! what OEM in
> their right mind would *deliberately* design any product that was so
> cost-uncompetitive, when $0.05 on the BOM could mean the difference
> between market success and failure??

BIOS-style partition tables are both an incredibly short-sighted
design as well as hardly the limit of the technology of the day. The
Commodore-Amiga supported more than 4 partitions, larger partition
sizes, and even OS/2-like per-partition filesystems back in 1987. I
refer you to their Rigid Disk Block structure, which formed the head
of a linked list of partitions, each defined by a single-sector
Partition Block. Since it used linked lists, you could have as many
partitions as you wanted. Partition blocks also stored preferred
device names and volume names as well, allowing the Kickstart ROM's
built-in boot loader to provide a Grub-like boot menu as well.

RDB/PBs have 32-bit limitations these days, so they're technically
obsolete as implemented on Kickstart. However, the spirit of the
technology is sound, and I believe might work better for an embedded
environment than GPT would, particularly if you're concerned with the
ratio of storage space to partition metadata.

Luke Kenneth Casson Leighton

unread,
Oct 13, 2018, 9:21:52 PM10/13/18
to Krste Asanovic, Samuel Falvo II, alist...@gmail.com, RISC-V SW Dev, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx
On Sat, Oct 13, 2018 at 10:29 PM <kr...@berkeley.edu> wrote:

> In this thread, as is common when engaging a wide community of smart
> experienced people, many of the disagreements are a result of people
> having different, even if possibly overlapping, subsets of the
> computing universe in mind. It would be helpful if everyone qualified
> their position with the systems they're assuming (microcontroller with
> no external DRAM, or servers with 8 sockets and terabytes of DRAM) and
> especially if we could introduce a common taxonomy/vocabulary, so
> people would be less likely to talk past each other.

from so many different backgrounds, that's quite challenging. i feel
that if people genuinely share the same goal, they'll work through
issues, to find the best ways to successfully communicate.

> I think we can all agree that reducing unnecessary duplication of
> effort is a good thing,

it's... desirable, absolutely. whether time and cost constraints on
NREs make fabless semi companies actually take pre-emptive *action*,
as opposed to just forking a codebase, throwing it at some engineers
and saying "get this into shape before the VC funding runs out, or
you're fired", is another matter.

> and both standards and open-source code are
> great and complementary ways to achieve that goal.

this is where i feel that taking an early lead, and *getting it
right*, will be so absolutely critical. identifying the failures
(proprietary binary-only distribution of "open" source firmware,
resulting in the persistent endemic security-related nightmares that
ron described so well) and the reasons why (use of BSD licenses by the
original authors because of corporate sponsorship - google f****g
about re-implementing perfectly good GNU libraries for android is by
far and above the absolute worst offender here) will help lay a path
that is followed if only because it's cheaper for SoC fabless semi
companies to do so.

> I think we _should_ all agree that there is necessary fragmentation
> because different platforms have different needs.

basically. yes. there's really no point trying to assert otherwise.
standards *only work* where and when there is common ground. if there
*is* no common ground...


> While some systems
> might never trust vendor-supplied M-mode software, there are other
> systems that would never trust a Linux kernel to install/execute
> M-mode code.

yyeahh i think the case that ron gave could be equally well applied
for {insert any OS or kernel} as it does for linux.

> Similarly, in structuring code bases, some people might
> want their Linux to know nothing of what's underneath to support
> flexible virtualization, while others want deep control of the
> hardware, e.g., for real-time concerns.

indeed.

and for those incredibly stupidly-designed smartphones i
reverse-engineered back in 2003, running WinCE, where you could set
the GSM Radio ROM to levels so high that the tiny phone could heat up
to dangerous levels *just by changing an unprotected
memory-addressable globally-accessible memory register*, because the
frickin baseband DSP + ARM11 core were sharing the same (global)
memory in order to save cost, having OS virtualisation that subdivides
such designs so that an iommu shuts out the main OS from the DSP and
vice-versa is REALLY important.

[the only reason those phones got FCC approval is because they
neglected to inform the FCC of the way that they were designed].

i mention this scenario because it's not just about "oh my laptop got
bricked", whilst that's inconvenient, it doesn't threaten your life.
having virtualisation can actually ensure that people don't get harmed
or killed, due to software bugs. or virus attacks.


> The big issue underlying most peoples' positions is that structuring a
> single set of interfaces and a single code base to support all these
> different use cases is not easy and perhaps not possible with current
> system programming tools (beyond "copy-and-paste" reuse),

this is pretty much exactly what every single chinese fabless
semiconductor company does. and many of the smaller korean ones. it
works well for them.

> and so
> they'd like to make sure their use-case is the one supported by the
> shared software effort.

indeed... and whilst there are more... "diplomatic" ways to put this,
it has to be said (no matter how many hackles it raises) that it would
not be good for such companies to expect either the free software
community or the RISC-V Foundation to "pick up the tab".

> I would hope the community could work towards figuring out how to
> share most of the code even if each piece of code runs in different
> places in a given boot/runtime system.

even with no actual coordination, it will "be" figured out. each
mainline repository is ultimately the canonical (atomic) location
around which, in an organic fashion, development will coalesce over
time.

the real question is: are the companies making millions to hundreds to
billions of dollars in profit adequately compensating the free
software developers on whom their profits critically depend? in 100%
of china fabless semi companies, the answer's a resounding "no".

let's do a little bit better than that, eh?

> I would also ask to separate out concerns about "binary blob" device
> drivers from defining boot flow/runtime components, as that is a
> mostly orthogonal problem to be solved.

mmmm... the way i see it, it is and it isn't. if any given hardware
doesn't e.g. require DDR3/4, then it's not a problem. memory being
pretty critical, it's a good example of a high-priority "orthogonal"
problem that needs to be solved.

now, as karsten pointed out: debian *will not* distribute proprietary
firmware blobs, no matter how much a company pleads. if that
proprietary DRAM initialisation blob is absolutely essential to
booting: tough.

and, i think everyone would agree that having debian as a supported
first-class citizen for RISC-V hardware is a pretty important
strategic goal, for the entire RISC-V community, given the sheer
number of packages that are supported, in such a stable fashion, *NOT*
controlled by any one corporation (such that even paranoid governments
tend to use it as the basis for secure OSes).

if it's not clear, ask any engineer in your team who is currently
desperately struggling with buildroot what it would be worth to them,
and how much time it would save, if they could only "just get debian
RISC-V up and running on the hardware [or emulator]" instead.

given that RISC-V has paved the way for an open debate and a much
more open process than other architectures, the "problems that were
swept under the rug" by Intel, AMD and ARM licencees on DRAM
initialisation, has now been brought to light, and has to be faced,
and tackled.

this is one of those systemic logical chains that i am good at
spotting, that other people tend to miss, where there are many
inter-related cross-correlations that don't *necessarily* involve
easy-to-explain technical issues.

and everything that my intuition and logical reasoning can get its
hands on is telling me that getting libre licenses for DRAM
initialisation source code out of Mentor, Cadence, Synopsys and other
DRAM vendors is about the highest possible priority "orthogonal"
strategic issue that places a major stumbling block to successful
RISC-V adoption as there could possibly be.

scouring the source code of reverse-engineered projects like the
allwinner u-boot and RK3288 u-boot sources, looking for DRAM
initialisation sequences (they're definitely there), identifying the
DRAM vendor and then having a quiet word with them and saying "um,
this is already out there, it's basically exactly what's needed,
there's absolutely no secrets here, you can even compare it with other
vendors source, here *that* is, you both do exactly the same thing,
now please for god's sake stop messing about", would be a good
starting point, once translated to appropriate diplomatic language.


> RISC-V should reduce the concerns about monopolistic hardware vendor
> policies, as there is no monopoly and there can be many independent
> hardware vendors catering to different markets, and using very
> different CPU designs (instead of many different vendors using the
> same few CPU designs).

this is one aspect of RISC-V that is absolutely tascinating,
fantastic, and inspiring. i don't think there's ever been a point in
computing history where so many disparate parties - libre software and
hardware engineers - multi-billion-dollar corporations - universities
- independent engineers - have ever collaborated even indirectly,
before.

yes linux is the largest software project in the world however it's
more a case of linus and a core team "curating" a set of disparate
vendor-specific contributions. RISC-V is about creating a revolution
in the *entire computing eco-system*. you began that, krste, and it's
something that should be recognised and appreciated.

> Because there will be many more RISC-V vendors
> than other architectures, there is a greater danger of fragmentation,
> so it is important to try to develop a small set of standards with
> supporting open implementations, and hopefully open compliance suites,
> that cover the common use cases.

agreed. the common use cases, mirroring the platform specs, is
probably the best most realistic option, and getting it *right* will
be critical here, as i feel that, realistically, the majority of
fabless semi companies will simply cookie-cut whatever's available.

that being the case, i would strongly, strongly advocate that in each
of those use-cases, early adopters come forward and make themselves
known (publicly, *not* privately, as this is about *collaboration and
standardisation*), with a view to being the guinea-pigs for everyone
else to poke sticks at and help see if standardisation in each of the
identified areas is in fact genuinely desirable and possible.

it is however incredibly important to accept that this really is a
massively disparate area of computing fraught with problems (as i,
karsten, ron and others have been graphically illustrating), and that
it's *perfectly ok* to agree that, actually, standardisation might not
be the best thing after all.

on the "use cases", i'd like to propose the Libre RISC-V SoC as one of
the candidates (the one with the RISC-V-based 3D GPU and VPU), as
literally by definition it absolutely *has* to be entirely open and
transparent, and at the same time it is specifically designed to enter
the 100-million-units-and-above low-to-mid-range tablet / netbook /
smartphone / chromebook / etc. market. documentation starts here:
https://libre-riscv.org/shakti/m_class/

for this particular SoC, u-boot is the highest preferred candidate,
and the bootloader / boot-sequence model that it will be following is
the one adopted by the allwinner a20: the "unbrickable" design of boot
sequence (that karsten described) that also has UART and JTAG
multiplexed with the primary SD/MMC interface pins, so that end-user
products in development do not require disassembly to debug: just plug
in a MicroSD breakout adapter and power it up.

little tricks like this are actually really important to properly
document. they don't sound like much, however when such an SoC is
used as part of the product lifecycle they massively simplify
development, and greatly reduce a *product's* NREs (people tend to
forget that a $3.50 quad-core 1.5ghz SoC is below 5% of the total BOM
for some of these mass-volume products!)

one other team i would suggest is the Shakti Group. i spent some time
with them after the Chennai Conference, so i know that they use
QuadSPI as the boot ICs, and that, fascinatingly, the use of Bluespec
allowed them to actually compile the entire processor's HDL into a
fast-running executable, with the RISC-V boot ROM binary actually then
executed *by* the compiled HDL, including an emulation of an actual
real-world QuadSPI external IC.

this allowed them to successfully develop and test proposed ROM
firmware *entirely through software emulation of the SoC*, all without
even needing an FPGA board, in a reasonable amount of time. again:
little tricks like this save hugely on development NREs, and are
absolutely worth documenting, so that other teams can try various
proposals out *WITHOUT* having to actually spend USD $5m and upwards
on actual silicon. or USD $85m on licensing Cadence or Mentor
Graphics tools.

i'll leave it at that.

l.

Luke Kenneth Casson Leighton

unread,
Oct 14, 2018, 5:34:53 AM10/14/18
to Jonathan Neuschäfer, Krste Asanovic, ron minnich, Christoph Hellwig, Anup Patel, atish patra, Padmarao Begari, RISC-V SW Dev
---
crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68

On Sat, Oct 13, 2018 at 4:36 PM Jonathan Neuschäfer
<j.neus...@gmx.net> wrote:
>
> On Sat, Oct 13, 2018 at 01:35:29AM +0100, Luke Kenneth Casson Leighton wrote:
> [...]
> > so, logically, when you refer to a desire to run "bootloaders" in
> > S-Mode we may *deduce* that you are NOT referring to the ENTIRE boot
> > process, that instead that you may be referring to the SECOND phase of
> > u-boot / coreboot.
>
> True.
>
> To elaborate a bit more on the different parts being discussed: I see
> three somewhat separable concerns in this discussion, that a RISC-V
> boot/firmware "stack" needs to address:
>
> (a) Low-level hardware initialization, for example of the DRAM
> (b) Run-time callable firmware, which (on RISC-V) implements the SBI
> (c) A full-featured boot loader which can read different storage/network
> media and file systems, in order to have flexibility in loading the OS.
>
> Some examples of how these concerns are implemented:
>
> - SiFive's FSBL implements (a),

ok, so from what i've heard, that someone mentioned that SiFive's
FSBL is proprietary (someone correct me if i read it wrong), it's
almost certainly the case that they've been caught out by the DRAM
licensor slamming them with an extreme NDA.

normally, because an x86 system would be shipped with a BIOS where
it's flat-out impossible to contact the vendor and complain about the
hell that people are being put through because of the extreme NDAs of
DRAM interface licensors, nobody does anything.

... in this case, with SiFive, we're watching closely, and...
*click*... YIPPEE! WELL DONE SIFIVE!

https://github.com/sifive/freedom-u540-c000-bootloader/tree/master/fsbl

rock on! DRAM initialisation source code, oh yeah, oh yeah :)


> - The [LinuxBoot] project focuses on (c), by reusing Linux's vast
> support for storage/network media and file systems

ok! cool! _this_ is the one i was referring to about actually
booting an *actual* linux kernel *as* the last-stage bootloader, then
getting that to "kexec" the "real" one:
https://www.linuxboot.org/page/faq/

> Personally, I'd welcome a central or popular implementation of the SBI

ah, right, i keep forgetting: SBL. with serial UART "remote boot"
systems it is pretty much absolutely essential that for lights-out and
remote recovery, a system be *dual* console bootable or better
switchable between *multiple* potential serial ports and the main
console.

in addition, u-boot can have multiple boot consoles, just like linux
can have "console=/dev/ttyS0,vga" and so on... the purpose of that
being to have, again, the console I/O from u-boot sent to multiple
simultaneous low-level abstracted devices (SBI consoles PLURAL).

in addition, it is pretty normal on embedded systems to have the main
linux console on a serial port ("console=/dev/ttyS0" init cmd option),
which is not just "nice to have" it's absolutely essential for
debugging and recovery, as the JTAG port can require special expertise
or dismantlement of the device to get at.

in addition to *that*, it's quite common to have dmesg go to the
serial port, where the screen is displaying "nice pretty warm and
fuzzy images" during boot.

for all of these, SBI really really does need to have the option of
not just the one console, but *multiple* consoles.

it would just so happen that, in addition, spike would then *happen*
to be able to use one of those serial consoles as a pppd network
tunnel.

so there are a hell of a lot of reasons why it would be a really
really good idea for SBI to have multiple serial device support.

thoughts invited.

l.

Karsten Merker

unread,
Oct 14, 2018, 9:54:47 AM10/14/18
to Luke Kenneth Casson Leighton, Karsten Merker, alist...@gmail.com, Krste Asanovic, RISC-V SW Dev, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx
On Sun, Oct 14, 2018 at 12:12:10AM +0100, Luke Kenneth Casson Leighton wrote:
> On Sat, Oct 13, 2018 at 9:32 PM Karsten Merker <mer...@debian.org> wrote:

[...]

> > Having to build per-machine-type installer images is a nightmare
> > and doesn't scale - I have been doing that for ARM-based systems
> > for quite some time,
>
> i _did_ try to warn people, back when devicetree was touted as "the
> solution!!!! to all the ARM world disparate pain!!!!!", that with
> literally tens of thousands of sensor and peripheral ICs used in
> products, all that would happen with devicetree is that the
> programming would move *from* c source to *to* devicetree source code.

That isn't correct - there is no programming in the device-tree.
Device-tree is a purely descriptive format of which hardware is
there and how it is interconnected. It doesn't contain any
executable elements.

Of course there needs to be code that implements a driver for
each peripherhal device, but those drivers can be built in a
generic fashion. Before device-tree, on arm platforms one had to
build a separate kernel image for each and every single arm board
out there, because the structure of the hardware (like: there is
a UART at address x, it is of type 16550A, it uses 32bit-wide
registers instead of 8bit-wide registers, its interrupt line is
connected to input 12 of the interrupt controller of type y at at
address z and the interrupt is level-triggered) was hard-coded
into the kernel image. No single kernel could be booted on
another device type than the one that it had been built for.
Device-tree has made it possible to have exactly one generic
kernel image which gets passed the hardware description in form
of the device-tree and can therefore run on all systems - of
course provided that it contains drivers for the general type of
peripherals that are used in the system.

That we still have the problem of having to build
per-machine-type installer images for quite a number of systems
in the ARM world isn't the fault of device-tree, quite the
opposite - device-tree has made the situation a lot better. The
reason we still have to build per-machine-type images is that
stuff like DRAM init code and ATF builds are still per
machine-type and aren't stored by the vendors in flash memory on
the board but have to be included into the boot images. That's
why I am proposing to make sure that RISC-V doesn't make the same
mistake and mandates that such code must be provided in onboard
flash memory.
You have probably misunderstood me - I was talking about the
mainboard level, not the SoC level. The code needs to be placed
into an on-board (as in "on-the-mainboard") flash, not into
memory inside the SoC. The board manufacturer has all the
information and can provide exactly the DRAM init code that is
required for the particular board.

> > * Define a standard partition table layout that is taken into
> > account by anything on the system, including the boot
> > firmware. My proposal would be GPT because it's flexible,
> > supported by about everything and the kitchen sink, and
> > supports large volumes (which is a problem with the classic
> > BIOS partition table format).
>
> that's almost certainly not going to fly on a first-stage bootloader
> (cf: u-boot SPL). as you know they're often looking at under 16k.
> loading from a specific raw sector (offset) is about as good as it
> gets.

Loading from a defined offset it perfectly fine - I just want
that offset defined in a way that it doesn't _conflict_ with the
chosen partition table format (in this case GPT). I don't expect
the SPL to actually parse the GPT, just to behave in a way that
doesn't make it incompatible with using GPT for other parts of
the system.

> plus, don't forget: with NAND and NOR, you don't have the luxury of
> knowing if the IC has built-in error-correction. or, more to the
> poin, you *absolutely must not* assume that it has.

AFAIK that is true for "naked" NAND, but not for SPI NOR flash.
AIUI, standard SPI NOR-Flash delivers either an access error or a
guaranteed correct read, but never silent bit errors. I might be
wrong here, though, as this isn't exactly my field of expertise.
Perhaps somebody on the list with experience in this field can
enlighten me?

> this is why for example the allwinner bootrom just does a for-loop,
> increasing the offset by fixed amounts, requesting a single block and
> looking for the signaure "eGON" at the front. if success, execute.

AIUI it does that because it indeed supports booting from "naked"
NAND flash that doesn't have any error correction / error
detection and can (and usually has) a number of defective cells
right out of the fab.

> basically, a standard partition at the front of a raw NAND flash
> device is a sure-fire guaranteed way to have a high percentage of
> returns on customer product.

I'm not talking about raw NAND here, but about stuff like
SD/eMMC. Just as a sidenode: based on my experience in the ARM
world I would consider using raw NAND instead of eMMC a design
error nowadays. There is probably a reason why board
manufacturers targeting the embedded ARM Linux market have
largely abandoned designs with raw NAND and instead tend to use
eMMC since quite a while.

> also, if you only *have* a 4MB SPI NOR boot iC, if you *absolutely*
> have to have a GPT-style partition table (because it's part of the
> "standard"), that's ... that's... 25% of the frickin SPI NOR IC
> *wasted* on a frickin *partition* table for god's sake! what OEM in
> their right mind would *deliberately* design any product that was so
> cost-uncompetitive, when $0.05 on the BOM could mean the difference
> between market success and failure??

Again, I have been talking about things like SD/eMMC, not about
small first-stage loader NOR flash that is usually used read-only
and without any partitioning scheme. There is probably no reason
to have a GPT in the typical NOR flash bootrom.

Things are different for eMMC though. There are quite a number
of designs in the embedded ARM Linux world that don't have an SPI
NOR flash but have a reasonably large eMMC (usually in the range
of 8GB). In those systems, the firmware (including the SPL that
handles the DRAM init) and kernel+userland are on the same eMMC
chip and need to coexist, so it is important to make sure that
they don't trample on each other's feet. For that, defining a
standard layout definitely makes sense.

> > The obvious compromise could indeed be an interface standard that
> > defines some stripped-down UEFI boot services as the common
> > denominator.
>
> from experience of witnessing the proliferation of u-boot (and
> understanding why that is) i'd honestly say that even this fills me
> with... what's the word... alarm / skepticism / cynicism / realism /
> other.
>
> to stand a chance in the embedded space, a proposal has to be
> justified in terms of code size, boot speed, and, ultimately, sale
> price of the end product.
>
> anything that increases the final sale price (not of the SoC, not of
> the product to the users, but the product - the *full* product - to
> the *OEMs*) is automatically going to be discared.
>
> so, if part of the proposal to include UEFI is a mandatory
> requirement that imposes an extra $1.50 on the BOM because certain
> types of non-UEFI-compliant external boot peripherals are also
> mandatory, and they're more expensive, or if the fabless semi company
> has to shell out for hardware encryption RTL, significantly increasing
> NREs in the process, you can kiss any such proposal's acceptance
> goodbye.

I can't follow you. Nobody is talking about a full EFI
implementation for embedded system and I have no idea which kind
of additional "boot peripherals" you might be referring to. We
are talking about a small emulation stub on top of u-boot that
adapts the UEFI boot services calling convention to the already
existing u-boot functions. The only potential increase in BOM
costs would be a (at most) few hundred kBytes of code size
increase for the main u-boot binary.

> also, now that i've read the bits about GPT, is UEFI *only* possible
> on top of GPT?

AFAIK UEFI requires GPT.

> and: is there expected to be encryption or signing,
> *without* which firmware is not loadable?

No, why should it?

Luke Kenneth Casson Leighton

unread,
Oct 14, 2018, 11:09:16 AM10/14/18
to Karsten Merker, alist...@gmail.com, Krste Asanovic, RISC-V SW Dev, Christoph Hellwig, Anup Patel, atish patra, Padmarao Begari, Jonathan Neuschäfer, phi...@hug.cx
---
crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68

On Sun, Oct 14, 2018 at 2:41 PM Karsten Merker <mer...@debian.org> wrote:

> > i _did_ try to warn people, back when devicetree was touted as "the
> > solution!!!! to all the ARM world disparate pain!!!!!", that with
> > literally tens of thousands of sensor and peripheral ICs used in
> > products, all that would happen with devicetree is that the
> > programming would move *from* c source to *to* devicetree source code.
>
> That isn't correct - there is no programming in the device-tree.
> Device-tree is a purely descriptive format of which hardware is
> there and how it is interconnected. It doesn't contain any
> executable elements.

... apologies, i was oversimplifying.

> Of course there needs to be code that implements a driver for
> each peripherhal device, but those drivers can be built in a
> generic fashion.

... if they are in fact generic. the key point being: the massively
disparate exclusive use of specific special one-off peripheral ICs
across the breadth of the embedded world makes it a moot point to even
bother to device-tree-ify such c code.

> Before device-tree, on arm platforms one had to
> build a separate kernel image for each and every single arm board
> out there, because the structure of the hardware (like: there is
> a UART at address x, it is of type 16550A,

ah yes, if it's a generally well-used common peripheral, yes absolutely.

> The
> reason we still have to build per-machine-type images is that
> stuff like DRAM init code and ATF builds are still per
> machine-type and aren't stored by the vendors in flash memory on
> the board but have to be included into the boot images. That's
> why I am proposing to make sure that RISC-V doesn't make the same
> mistake and mandates that such code must be provided in onboard
> flash memory.

... let me take a look further below, to find out why, after reading
what i wrote, you believe that it would be a mistake. certainly, i
can categorically say that any kind of "mandating" is absolutely out
of the question. the strongest we can possibly hope for is
"recommended practices", and even then there's absolutely no guarantee
they'll be followed.
ah then we *might* be talking about the same ting.
> The code needs to be placed
> into an on-board (as in "on-the-mainboard") flash,

ok, i see where the confusion is: then it should be categorised as
off-chip, or "external", or, yes, on-mainboard. "on-board" in SoC
terminology can mean "on the PCB - the board - on which the SoC die
sits".

> not into memory inside the SoC.

... define "memory" [sorry!]. do you mean ROM, do you mean in-SoC
NAND/NOR/EEPROM, do you mean in-SoC SRAM?

> The board manufacturer has all the
> information and can provide exactly the DRAM init code that is
> required for the particular board.

yyyyes. this doesn't strictly solve the absolute strictest
definition of "libre", however i *believe* that having an external
[on-motherboard] NAND/NOR/EEPROM containing proprietary firmware blobs
is actually RYF Endorseable. it's an aspect of RYF Endorsement that
many strict libre advocates *really* don't like, however it is much
more realistic and pragmatic.

i say that however the only people permitted to say for sure are the
holders of the RYF Certification Mark, and that's the FSF.

> Loading from a defined offset it perfectly fine - I just want
> that offset defined in a way that it doesn't _conflict_ with the
> chosen partition table format (in this case GPT). I don't expect
> the SPL to actually parse the GPT, just to behave in a way that
> doesn't make it incompatible with using GPT for other parts of
> the system.

understood. it's something that should be on the proposed
"recommended best practices" list [we *really* cannot make it
"mandatory": we just absolutely do not have that right to tell third
party manufacturers with far more experience in the field what to do,
here].

> > plus, don't forget: with NAND and NOR, you don't have the luxury of
> > knowing if the IC has built-in error-correction. or, more to the
> > point, you *absolutely must not* assume that it has.
>
> AFAIK that is true for "naked" NAND, but not for SPI NOR flash.
> AIUI, standard SPI NOR-Flash delivers either an access error or a
> guaranteed correct read, but never silent bit errors. I might be
> wrong here, though, as this isn't exactly my field of expertise.
> Perhaps somebody on the list with experience in this field can
> enlighten me?

NAND i know can have errors, and to cope with this, many NAND
manufacturers make the first N Mbytes out of SLC NAND
(single-level-cell, which is far more reliable), and the rest MLC
(multi-level-cell).

NOR, i honestly have no idea: i've never used it, as it's way waaay
more expensive for what it is.

> > this is why for example the allwinner bootrom just does a for-loop,
> > increasing the offset by fixed amounts, requesting a single block and
> > looking for the signaure "eGON" at the front. if success, execute.
>
> AIUI it does that because it indeed supports booting from "naked"
> NAND flash that doesn't have any error correction / error
> detection and can (and usually has) a number of defective cells
> right out of the fab.

yeah :) they're cheap for a reason!

> > basically, a standard partition at the front of a raw NAND flash
> > device is a sure-fire guaranteed way to have a high percentage of
> > returns on customer product.
>
> I'm not talking about raw NAND here, but about stuff like
> SD/eMMC.

ok, yes: eMMC would prooobably be fine... they have spare-block
swapping built-in to the firmware, usually (just like on SSDs).

SD... hmmm, i've had low-cost SD cards fail on me before...

> Just as a sidenode: based on my experience in the ARM
> world I would consider using raw NAND instead of eMMC a design
> error nowadays.

yyeah, you will have seen the "nightmare scenario" i outlined in an
earlier message, which caught out even Nintendo.

> There is probably a reason why board
> manufacturers targeting the embedded ARM Linux market have
> largely abandoned designs with raw NAND and instead tend to use
> eMMC since quite a while.

except for ultra-ultra-low-cost applications. Winbond know their
stuff, and, unlike a lot of other NAND manufacturers who produce
faster, higher-capacity NAND, Winbond produce stable
much-lower-capacity TSSOP-48 ICs at around the 128mb mark, and it's
used a lot in mass-volume "openwrt" style products.

> > also, if you only *have* a 4MB SPI NOR boot iC, if you *absolutely*
> > have to have a GPT-style partition table (because it's part of the
> > "standard"), that's ... that's... 25% of the frickin SPI NOR IC
> > *wasted* on a frickin *partition* table for god's sake! what OEM in
> > their right mind would *deliberately* design any product that was so
> > cost-uncompetitive, when $0.05 on the BOM could mean the difference
> > between market success and failure??
>
> Again, I have been talking about things like SD/eMMC, not about
> small first-stage loader NOR flash that is usually used read-only
> and without any partitioning scheme. There is probably no reason
> to have a GPT in the typical NOR flash bootrom.

yehyeh.

> Things are different for eMMC though. There are quite a number
> of designs in the embedded ARM Linux world that don't have an SPI
> NOR flash but have a reasonably large eMMC (usually in the range
> of 8GB).

yes. it's pretty damn hard to even *source* anything below 4GB eMMC.

> In those systems, the firmware (including the SPL that
> handles the DRAM init) and kernel+userland are on the same eMMC
> chip and need to coexist, so it is important to make sure that
> they don't trample on each other's feet. For that, defining a
> standard layout definitely makes sense.

i'd tend to agree, here, as it's a reliable media, similar to an SSD in nature.

> > so, if part of the proposal to include UEFI is a mandatory
> > requirement that imposes an extra $1.50 on the BOM because certain
> > types of non-UEFI-compliant external boot peripherals are also
> > mandatory, and they're more expensive, or if the fabless semi company
> > has to shell out for hardware encryption RTL, significantly increasing
> > NREs in the process, you can kiss any such proposal's acceptance
> > goodbye.
>
> I can't follow you. Nobody is talking about a full EFI
> implementation for embedded system and I have no idea which kind
> of additional "boot peripherals" you might be referring to.

i was concerned that if there was some form of mandatory encryption
or... something else (hand-waving, here, i don't know enough to say),
that only certain peripherals from certain specific manufacturers
would even stock such specialist encrypted-NAND /
encrypted-UEFI-enabled-whatever.

however, even if specialist NAND/NOR/eMMC isn't required, if
encryption or digital signing *is* needed, as part of the 1st stage
SPL boot, that's expensive, right there, as it will almost certainly
need the chip designer to license an accelerated custom hardware RTL,
to get boot time down.

> We
> are talking about a small emulation stub on top of u-boot that
> adapts the UEFI boot services calling convention to the already
> existing u-boot functions. The only potential increase in BOM
> costs would be a (at most) few hundred kBytes of code size
> increase for the main u-boot binary.

a few *hundred*?? *shocked*!! so, that's not fitting in the 16k
first stage SPL, then :) also, a few hundred kb is starting to hit a
threshold where cost-sensitive embedded ODMs will take notice [and rip
it out].

> > also, now that i've read the bits about GPT, is UEFI *only* possible
> > on top of GPT?
>
> AFAIK UEFI requires GPT.

ok.

> > and: is there expected to be encryption or signing,
> > *without* which firmware is not loadable?
>
> No, why should it?

you will almost certainly know more about this than i do: i was under
the impression that UEFI was the layer where encrypted (or
digitally-signed) bootloaders sit. and that this is where
"Treacherous" Computing comes into play.

Samsung's Exynos SoCs are bitches, here, as are many of the TI range
of ARM SoCs. these contain e-fuses which, once blown, there's
absolutely no way in hell you can change the SPL digital signature
stored in the SoC's on-die EEPROM. you *literally* have to de-solder
the entire SoC and replace it with one that hasn't had its e-fuses
blown.

do you (or anyone else) know any of the details, there?

l.

Karsten Merker

unread,
Oct 14, 2018, 12:07:20 PM10/14/18
to Luke Kenneth Casson Leighton, Karsten Merker, alist...@gmail.com, Krste Asanovic, RISC-V SW Dev, Christoph Hellwig, Anup Patel, atish patra, Padmarao Begari, Jonathan Neuschäfer, phi...@hug.cx
On Sun, Oct 14, 2018 at 04:08:32PM +0100, Luke Kenneth Casson Leighton wrote:

> > > so, if part of the proposal to include UEFI is a mandatory
> > > requirement that imposes an extra $1.50 on the BOM because certain
> > > types of non-UEFI-compliant external boot peripherals are also
> > > mandatory, and they're more expensive, or if the fabless semi company
> > > has to shell out for hardware encryption RTL, significantly increasing
> > > NREs in the process, you can kiss any such proposal's acceptance
> > > goodbye.
> >
> > I can't follow you. Nobody is talking about a full EFI
> > implementation for embedded system and I have no idea which kind
> > of additional "boot peripherals" you might be referring to.
>
> i was concerned that if there was some form of mandatory encryption
> or... something else (hand-waving, here, i don't know enough to say),
> that only certain peripherals from certain specific manufacturers
> would even stock such specialist encrypted-NAND /
> encrypted-UEFI-enabled-whatever.
>
> however, even if specialist NAND/NOR/eMMC isn't required, if
> encryption or digital signing *is* needed, as part of the 1st stage
> SPL boot, that's expensive, right there, as it will almost certainly
> need the chip designer to license an accelerated custom hardware RTL,
> to get boot time down.

There is no need for any crypto, neither at the SPL stage nor at
the main u-boot stage. I have no idea where you take that idea
from.

> > We
> > are talking about a small emulation stub on top of u-boot that
> > adapts the UEFI boot services calling convention to the already
> > existing u-boot functions. The only potential increase in BOM
> > costs would be a (at most) few hundred kBytes of code size
> > increase for the main u-boot binary.
>
> a few *hundred*?? *shocked*!! so, that's not fitting in the 16k
> first stage SPL, then :) also, a few hundred kb is starting to hit a
> threshold where cost-sensitive embedded ODMs will take notice [and rip
> it out].

Please read again what I wrote: "the _main_ u-boot binary". The
UEFI boot services interface is implemented by the main u-boot
binary that runs from DRAM, not by the SPL that runs from SRAM.
The SPL has absolutely nothing to do with providing UEFI boot
services.

> > > and: is there expected to be encryption or signing,
> > > *without* which firmware is not loadable?
> >
> > No, why should it?
>
> you will almost certainly know more about this than i do: i was under
> the impression that UEFI was the layer where encrypted (or
> digitally-signed) bootloaders sit. and that this is where
> "Treacherous" Computing comes into play.

There is "UEFI secure boot", but that is a purely optional
element of the UEFI specification. Microsoft mandates that
implementations that want to get some form of certification from
Microsoft implement it, which is why PC mainboard manufacturers
do it, but that's nothing that concerns us here - one can
perfectly implement the UEFI boot services without any support
for "UEFI secure boot".

> Samsung's Exynos SoCs are bitches, here, as are many of the TI range
> of ARM SoCs. these contain e-fuses which, once blown, there's
> absolutely no way in hell you can change the SPL digital signature
> stored in the SoC's on-die EEPROM. you *literally* have to de-solder
> the entire SoC and replace it with one that hasn't had its e-fuses
> blown.

That's a completely orthogonal topic to UEFI. Those SoCs require
that the first-stage loader that is loaded directly from the SoCs
internal boot ROM is signed by a specific key, but that doesn't
have anything to do with any later stages in the boot process.

Luke Kenneth Casson Leighton

unread,
Oct 14, 2018, 12:32:04 PM10/14/18
to Karsten Merker, alist...@gmail.com, Krste Asanovic, RISC-V SW Dev, Christoph Hellwig, Anup Patel, atish patra, Padmarao Begari, Jonathan Neuschäfer, phi...@hug.cx
---
crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68

On Sun, Oct 14, 2018 at 4:54 PM Karsten Merker <mer...@debian.org> wrote:

> There is no need for any crypto, neither at the SPL stage nor at
> the main u-boot stage. I have no idea where you take that idea
> from.

[you kindly identify it below, as "UEFI secure boot"]

> > > We
> > > are talking about a small emulation stub on top of u-boot that
> > > adapts the UEFI boot services calling convention to the already
> > > existing u-boot functions. The only potential increase in BOM
> > > costs would be a (at most) few hundred kBytes of code size
> > > increase for the main u-boot binary.
> >
> > a few *hundred*?? *shocked*!! so, that's not fitting in the 16k
> > first stage SPL, then :) also, a few hundred kb is starting to hit a
> > threshold where cost-sensitive embedded ODMs will take notice [and rip
> > it out].
>
> Please read again what I wrote: "the _main_ u-boot binary".

apologies, the thread's pretty long: i won't be able to absorb everything.

> The
> UEFI boot services interface is implemented by the main u-boot
> binary that runs from DRAM, not by the SPL that runs from SRAM.
> The SPL has absolutely nothing to do with providing UEFI boot
> services.

indeed... that was just an aside: the size is however going to be an
issue for cost-sensitive embedded deployments.

> > you will almost certainly know more about this than i do: i was under
> > the impression that UEFI was the layer where encrypted (or
> > digitally-signed) bootloaders sit. and that this is where
> > "Treacherous" Computing comes into play.
>
> There is "UEFI secure boot", but that is a purely optional
> element of the UEFI specification.

okaay, fantastic: sorry i'm not totally familiar with the terminology
in UEFI. so yes, that's what i was referring to.

> Microsoft mandates that
> implementations that want to get some form of certification from
> Microsoft implement it, which is why PC mainboard manufacturers
> do it, but that's nothing that concerns us here - one can
> perfectly implement the UEFI boot services without any support
> for "UEFI secure boot".

ok, cool.

> > Samsung's Exynos SoCs are bitches, here, as are many of the TI range
> > of ARM SoCs. these contain e-fuses which, once blown, there's
> > absolutely no way in hell you can change the SPL digital signature
> > stored in the SoC's on-die EEPROM. you *literally* have to de-solder
> > the entire SoC and replace it with one that hasn't had its e-fuses
> > blown.
>
> That's a completely orthogonal topic to UEFI. Those SoCs require
> that the first-stage loader that is loaded directly from the SoCs
> internal boot ROM is signed by a specific key, but that doesn't
> have anything to do with any later stages in the boot process.

it's still worth discussing... and unfortunately providing as an
option, as, despite the Treachery abuse such schemes will get, there
are genuine (mission-critical, life-threatening) scenarios where it's
part of a wider safety strategy.

my understanding is that there was some discussion years ago about
digital-signing of linux kernels. phil hands (long-standing ftp uk
debian mirror maintainer) told me a story about a vendor saying
something to the effect, "well you can always send the binary kernel
to us and we'll sign it for you".

i distinctly recall quite a lot of fuss when UEFI secure boot came
out, where it was feared that linux systems would be completely and
utterly locked out from being installable.

secondly: the first Microsoft Surface Laptop that actually had an ARM
core in it: i believe it was a first-generation NVidia Tegra. they're
impossible to reverse-engineer. as in, there's no point. why?
because from that initial first-stage loader, there's a
hardware-backed chain of Treachery (i can't bring myself to call it
"Trust") that actively prevents and prohibits all and any efforts to
replace the OS.

even microsoft end-users couldn't put 3rd party apps on it.

anyway... digression, there. apologies.

l.

ron minnich

unread,
Oct 14, 2018, 2:21:26 PM10/14/18
to Luke Kenneth Casson Leighton, Jonathan Neuschäfer, Krste Asanovic, Christoph Hellwig, Anup Patel, atish patra, Padmarao Begari, RISC-V SW Dev
I love serial ports and have no objection to having one. There's almost no better way to bring up a new board than to have a default, enabled, configured serial port ready and waiting at power-on/reset with 115200 baud rate and standard rs232 signaling levels (and, frequently nowadays, the board manufacturers screw this up too: invalid baud rates, ttl signaling levels, and complex/non-enabled/weirdly configured serial port setup are common).

But that said, there's all kinds of precedent for lights out recovery that does not need even one serial port. It's not required, and sometimes not desired, and often not even possible to use it.

There's several 10s of millions of Chromebooks out there today that are a QED that serial ports are not mandatory for recovery from the occasional bad firmware flash experience. It takes some doing but it can be done. I built many HPC systems back in the day with LinuxBIOS where our recovery path was the network, not serial. 

Serial ports are nice for bringup, and not always nice for recovery, and sometimes quite the wrong way to go.

ron

Michael Clark

unread,
Oct 14, 2018, 5:12:32 PM10/14/18
to alist...@gmail.com, RISC-V SW Dev, kr...@berkeley.edu, h...@lst.de, an...@brainfault.org, atis...@gmail.com, padma...@gmail.com, j.neus...@gmx.net, phi...@hug.cx


On 13/10/2018, at 12:45 PM, alist...@gmail.com wrote:

On Friday, October 12, 2018 at 4:23:10 PM UTC-7, Luke Kenneth Casson Leighton wrote:
On Fri, Oct 12, 2018 at 10:12 PM <alist...@gmail.com> wrote:

> The problem that I see with starting coreboot or u-boot (the last stage bootloaders)
> in M mode is that we then end up with duplicated work between the projects.

 yes.  that has not stopped either project from continuing.  it also
hasn't stopped libreboot, or grub, or the system which allows a linux
kernel to actually *be* the bootloader (which runs an initramfs and
calls kexec).

 each of these bootloaders has different fans and userbases.  they
don't necessarily talk to each other, and that's just how it is.

 expecting each independent team to cooperate, particularly when some
of them are hard forks of other bootloader projects, is not realistic.

> This is where a model like Arm Trusted Firmware fits in well.

 you mean, ARM _treacherous_ firmware?  the "trust" you refer to is
directed at *manufacturers*, providing the specific goal of
*betraying* users.   "Trust Us.  We Will HELP You To Prevent And
Prohibit Users From Owning Their Devices".

That seems a little harsh. ATF is open source and you can compile it yourself from the source to deploy. You are also free to re-write it if need be.
 

>We can see something like this:
>  1. CPU boots up and ROM does it's thing
>  2. ROM boots into an open source firmware (ATF equivalent) in M mode
>     - This will end up being pretty SoC specific

 yes.  and may not fit into the available memory, which may be
absolutely tiny.

the ROM bootloader, which is best made about as ridiculously dumb and
tiny as can possibly be made, is usually just about capable of loading
from a single (16k) sector of a Micro-SD or eMMC device, or 16k from
an SPI device, or 16k from an ONFI NAND, and so on.  that sub-16k
executable the "first stage" bootloader.

16K is the "normal" amount of available SRAM that SoCs are prepared to
allocate for these first-stage bootloaders.   that 1st stage basically
does things like "initialise the DDR3 RAM, set up IRQs"... oh and it
happens to initialise the
device-from-which-the-next-phase-is-to-be-loaded.

in u-boot, this first stage bootloader is called the "SPL", and it's
compiled from u-boot with some special #defines that *MASSIVELY* cut
down u-boot to the absolute, absolute bare minimum (surprise: under
16k).

all of this is typically entirely hidden from end-users.  you don't
have to know.  you don't have to care.  people have *no idea* that
their system goes through a four, five or even six stage boot process
to get to an OS.  even intel x86 systems.

basically i'm saying that "one size does not fit all", that creating
standards in this area is not only unrealistic but also completely
undesirable it is so heavily application-specific, and that
restricting vendors through standardisation to be not permitted to run
from the absolute lowest possible level (M-Mode) will almost certainly
cause a massive back-lash.

*on the other hand*, if there exists *informal reference source code*
out there, that's absolutely fantastic, as it will cut down massively
on development NREs.  example ROM bootloaders, for example, would be
fantastic.

why would it be fantastic? because the entire process is so
proprietary, so secret, that very few people actually know what's
involved.  full libre-licensed source code - preferably with
libre-licensed overview documents - would completely change how SoCs
are developed.

The risk that I see with just letting everyone go their own way is that we will end up with a similar situation to the ARM world.

That is that every SoC vendor does their own thing. This can result in a less open ecosystem as every vendor writes their own boot code and keeps it closed.

The other problem I see is a fragmented boot process. Where some boards use coreboot and some use u-boot and some use something else. We will then get to a point where every board is different. x86 has this pretty much sorted as I can boot my Linux USB stick on every PC I own. ARM is the exact opposite of this. I would like to see RISC-V not go down this fragmentation route. One way to do this is just to have a "best" way to setup the boot process. Then SoC vendors will hopefully opt to follow it.

A mix-mash of ways seems like a dangerous path to lead to fragmentation and confusion and that is what I'm afraid of.

The way to solve that is platform-level conformance. But folk may not like the platform e.g. me. I’d like EBBR i.e. slimmed down UEFI with device-tree and FAT32 partitions for EFI/BOOTRV64.EFI

If one actually spends time to read the open UEFI spec, then one finds it provides everything the SBI provides (Timers, Console, Boot Tables which may be device-tree or ACPI, Boot partition file access e.g. FAT on SD SPI like lowRISC/UCAM’s bbl fork which has BSD licensed FAT and SPI drivers).

It’s a standard. MS will be happy too. Beautiful. End of story. Coreboot does their own thing. U-boot is a EFI payload that gives you a menu (or they do their own thing but are not platform standard compliant like they aren’t on arm; just defacto open source in a fragmented ecosystem which officially has EBBR in the embedded space and UEFI/ACPI in the server space and ‘n’ different bootloaders in the IoT/mobile space).

PS. I can tell you writing a EFI boot loader with gnu-efi on Linux is way nicer than that multiboot rubbish, Ubuntu does capsule updates for my Dell XPS 14, Intel devices let me change the Secure Boot root of trust and QEMU has OVMF.

UEFI Secure Boot is technically quite nice and can be implemented securely. The only complexity is the hierarchy of keys/certs which is a result of the platform/OEM/OS-vendors being different parties. It would be nice to at minimum just require a root cert, and optional intermediates in the trust store. The minimum case should require one+ certs/intermediates whose owners have signing authority for firmware. That’s however could occur in a simpler OVMF alternative that doesn’t have the bugs that were present in tiannocore, perhaps due to its size. e.g. a smaller and well-audited implementation of EBBR/UEFI would be very attractive to someone who prefers to adopt an accepted industry standard vs some open source project that doesn’t implement a well documented standard.

Michael Clark

unread,
Oct 14, 2018, 5:35:45 PM10/14/18
to Luke Kenneth Casson Leighton, alist...@gmail.com, RISC-V SW Dev, Krste Asanovic, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx


> On 13/10/2018, at 1:47 PM, Luke Kenneth Casson Leighton <lk...@lkcl.net> wrote:
>
>> On Sat, Oct 13, 2018 at 1:16 AM <alist...@gmail.com> wrote:
>>
>> To a certain extent it's unavoidable, but if RISC-V is ever going to take off we are going to have to at least somewhat deal with it. Booting ARM boards today is such a pain that I suspect 90% of people give up and just take the prebuilt binaries from the SoC vendor. It is a hassle and if we tackle this early on we might be able to reduce this for RISC-V.
>
> no. really. please. try to understand and accept the implications,
> here. if you want RISC-V to be in the embedded world, and in multiple
> SoCs, and in desktops, and in high-end servers, you have to accept the
> reality of the absolutely massive fundamental hardware differences
> between all those different scenarios, where the bootloader and
> bootloader choice absolutely necessarily reflects that.
>
> if you start talking about UEFI to low-cost tablet vendors for
> example they will just laugh at you. and if you even mention ACPI to
> them as a required standard as part of a boot process, they'd be quite
> likely to punch your lights out (and i would cheer and congratulate
> them).

Well they are uninformed.

i.e. have not seen how trivial it is tool declare efi_main(), access console, timers and device-tree, load another file from the boot device, be that a simple cpio or zip in ROM

I could code in a few minutes as I used ZIP filesystems on my android/iOS/macOS/windows/Linux OpenGL ES UI framework, as these are used for asset filesystems everywhere incl. most games (renamed .PAK or whatever). With UEFI/EBBR it’s just some simple C code. No assembly trampolines and low level M-mode code required at all.

Right we need a GUI. EGL is the appropriate standard here. i.e. is the khronos standard for getting a GL context on an embedded device that has a simple GPU. If one wants to paint a simple logo on an LCD, then framebuffer dimensions and pixel format and memory aperture in device-tree would suffice.

One doesn’t need to mandate Secure Boot. A slim device-tree EBBR subset could probably be implemented in ~64K-128K, much smaller than Linux which is likely going to be ~500K. It’s only FUD that makes folk think UEFI is complex. A well-defined subset could be extremely simple. I might demonstrate give one weekend (or two)...

> x86 does not have this "problem" that ARM has because x86 has FAILED
> to enter the ARM-dominated SoC market. despite intel bribing vendors,
> sellers, OEMs *and customers* to buy their products, they *still could
> not get market share*. yes, i've been talking to a business
> development manager for an ARM-based fabless semi company, who
> regularly was asked, "how much are YOU going to pay US to use your
> SoC, because that's what intel does".
>
> all that having been said: i concur that *if* full source code is
> available (and available early enough), chances are indeed high that
> fabless semi companies will just cookie-cut the available solutions.
>
> unfortunately, the DDR3/4 hard macro vendors (who typically provide
> proprietary initialisation source code under insane NDAs) are going to
> get in the way, there.
>
> so that's down to the early adopters - SiFive in particular as the
> high-profile one - to take responsibility for that, and tell the
> DDR3/4 interface vendors in no uncertain terms that proprietary
> licensing on the DRAM initialisation is not acceptable.
>
> l.
>
> --
> You received this message because you are subscribed to the Google Groups "RISC-V SW Dev" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to sw-dev+un...@groups.riscv.org.
> To post to this group, send email to sw-...@groups.riscv.org.
> Visit this group at https://groups.google.com/a/groups.riscv.org/group/sw-dev/.
> To view this discussion on the web visit https://groups.google.com/a/groups.riscv.org/d/msgid/sw-dev/CAPweEDyLSq0i%3D7VsY5QC_W5T_eyyuyJr5W4Y4fEN7RpHrU-JpA%40mail.gmail.com.

Luke Kenneth Casson Leighton

unread,
Oct 14, 2018, 5:50:38 PM10/14/18
to Michael Clark, alist...@gmail.com, RISC-V SW Dev, Krste Asanovic, Christoph Hellwig, Anup Patel, atish patra, Padmarao Begari, Jonathan Neuschäfer, phi...@hug.cx
On Sun, Oct 14, 2018 at 10:35 PM 'Michael Clark' via RISC-V SW Dev
<sw-...@groups.riscv.org> wrote:

> Well they are uninformed.

i feel it's probably a case of, "not well-known in mainline". if
it's really as simple as you describe that's fantastic. the proof will
be in if, *after* an implementation hits mainline, and has a proper
push through the u-boot (or whatever/wherever) channels, take-up
slowly starts to drift into real-world products. that'll take a
while.

l.

Michael Clark

unread,
Oct 14, 2018, 6:01:30 PM10/14/18
to Paul Walmsley, ron minnich, Christoph Hellwig, padma...@gmail.com, Anup Patel, atis...@gmail.com, sw-...@groups.riscv.org


On 13/10/2018, at 5:48 AM, Paul Walmsley <paul.w...@sifive.com> wrote:

Yes - for RISC-V U-boot, it's best to follow the example set by the ARM64 U-boot folks and support booting in any mode:

https://github.com/theopolis/u-boot/blob/master/doc/README.arm64


Hi Paul,

Interesting perspective. I must say I am just an observer here.

One ask. Can folk preserve the ability to load an ELF kernel. The symbols can be very useful for debugging. QEMU loads ELF and possibly flat. I think load_rom checks for ELF and otherwise loads as flat. There is also a use case for .hex. We have been using ELF as bbl is ELF iirc. I have used ELF symbols for symbolicated tracing in rv8 and QEMU. Of course symbols can be very useful for hackers too and must be separated in production  (guessing there is an inverse strip that just keeps symbols/file/line and DWARF metadata). Is there a technical reason to prefer flat over ELF? 

ELF headers are tiny and they are used for modules in the kernel and likely will be used in modular boot loaders, even when modules are in firmware.

We should also consider the output of `file` on a RISC-V Supervisor image. This works on other architectures.

I have to note that ELF or other formats are essential for relocatable images for No-MMU systems, where text+data need to be loaded at addresses that may not be known at compile time. I know we do not have RISC-V FD-PIC yet, where the ELF container is necessary. While flat images are fine in some circumstances, they are not in many others. vmlinux.gz seems nice and could be decompressed using stream decompression from FLASH unlike other decompression schemes that use more RAM (bzImage on x86). We also use ELF images when loading via JTAG. Flat images can always be extracted when necessary for binaries that are compatible. Note ELF also has segment protection metadata (R,W^X) as a “general” mechanism.

PE and TE are other formats that may crop up. TE being a simpler version of PE used for relocatable drivers in UEFI firmware. PE being the ubiquitous COFF derivative used by Windows.

I have to say, a reduced subset of UEFI like that present in GNU EFI and OVMF is actually quite a clean abstraction and UEFI even supports device-tree vs ACPI (as in the UEFI subset defined in arm’s EBBR which I was made aware of by Stephan O’Rear on IRC). UEFI has quite nice services APIs and as we know, FAT32 is somewhat ubiquitous if one needs to copy firmware onto SD cards and what not. Windows users don’t have `dd`.

Also +1 for Windows PE / Nano on RISC-V. We have PE object file codes already (they are something cute like 0x5032, 0x5064 and 0x5128 iirc).

My only nit with UEFI is having cdecl on x86 which requires thunk wrappers when calling from x64/Itanium ABI objects and the use of PE format instead of ELF. I do hope Microsoft decide to use the RISC-V ABI. BTW Anyone know what is sizeof(long) on Aarch64 Windows? I suspect it is 4. i.e. -mabi=llp64

Now that MS bundles Clang with VS, we could use ELF and lp64 for the UEFI variant on RISC-V perhaps. I like that idea.

Michael

On 10/12/18 8:57 AM, ron minnich wrote:
I'm surprised to see this idea but in any event coreboot replaces bbl and has m mode code and I don't expect that to change. I'm done with trusting vendor supplied fsbl no matter what the source.

On Fri, Oct 12, 2018, 1:08 AM Christoph Hellwig <h...@lst.de> wrote:
On Fri, Oct 12, 2018 at 01:35:02PM +0530, Anup Patel wrote:
> > Awesome. Does that mean you are planning to implement SBI as well in uboot or bbl continue to exist in M mode ?
> > Can you please explain the boot flow design if u-boot runs in S-mode ?
>
> Please no M-mode code in U-boot or any general purpose bootloader.

Indeed, that is what I feared :(

> Instead of replacing SiFive FSBL with U-Boot, I suggest to boot
> bbl+uboot.bin from SiFive FSBL

That would be more useful indeed. 

--
You received this message because you are subscribed to the Google Groups "RISC-V SW Dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sw-dev+un...@groups.riscv.org.
To post to this group, send email to sw-...@groups.riscv.org.
Visit this group at https://groups.google.com/a/groups.riscv.org/group/sw-dev/.
--
You received this message because you are subscribed to the Google Groups "RISC-V SW Dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sw-dev+un...@groups.riscv.org.
To post to this group, send email to sw-...@groups.riscv.org.
Visit this group at https://groups.google.com/a/groups.riscv.org/group/sw-dev/.

--
You received this message because you are subscribed to the Google Groups "RISC-V SW Dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sw-dev+un...@groups.riscv.org.
To post to this group, send email to sw-...@groups.riscv.org.
Visit this group at https://groups.google.com/a/groups.riscv.org/group/sw-dev/.

ron minnich

unread,
Oct 14, 2018, 6:42:26 PM10/14/18
to Michael Clark, Paul Walmsley, Christoph Hellwig, padma...@gmail.com, Anup Patel, atis...@gmail.com, sw-...@groups.riscv.org
So, I apologize for anything I did to make this thread longer. I think we're all in agreement that we don't agree about the boot solutions, but we do agree that things must be completely open, buildable, and flashable. And that's the only thing I really care about.


Michael Clark

unread,
Oct 14, 2018, 8:04:48 PM10/14/18
to Karsten Merker, alist...@gmail.com, kr...@berkeley.edu, RISC-V SW Dev, h...@lst.de, an...@brainfault.org, atis...@gmail.com, padma...@gmail.com, j.neus...@gmx.net, phi...@hug.cx
Hi Karsten.

It would be very nice if you posted your commentary on a wiki page. This is a very good summary of the many concerns.

I largely agree. I disagree in some respects. i.e. it may be more costly and bug prone to take virtualized server system assets that target PCI and them use some as yet unknown undocumented standard. With respects to 32-bit UEFI on 64-bit systems, well that is legacy, and while a distro needs to support old hardware, including BIOS, modern 64-bit UEFI is more uniform. I am also aware there are issues with OEMs that have Secure Boot enabled meaning folk need the Ubuntu chain loader. This however can’t be prevented as an ISA can’t prevent uses where vendors have content license agreements that require these kinds of mechanisms (e.g. HDCP for displays). We have GNU EFI but we can’t prevent BSD EFI and open plus proprietary derivatives.

I need to learn more about u-boot. Perhaps it is more standardised than I think, however it comes to mind as an implementation. Padmaro, good work with the u-boot port! If it doesn’t preclude loading RISC-V UEFI payloads, then I suspect I will really like it.

I sort of agree that boot loaders should start in S-mode but I can also see a need for M-mode virtualization, where earlier firmware can run, but that is a different matter altogether.

Michael
> --
> You received this message because you are subscribed to the Google Groups "RISC-V SW Dev" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to sw-dev+un...@groups.riscv.org.
> To post to this group, send email to sw-...@groups.riscv.org.
> Visit this group at https://groups.google.com/a/groups.riscv.org/group/sw-dev/.
> To view this discussion on the web visit https://groups.google.com/a/groups.riscv.org/d/msgid/sw-dev/20181013203214.f4fzo6hz4nroejfd%40excalibur.cnev.de.

ron minnich

unread,
Oct 14, 2018, 8:23:11 PM10/14/18
to Karsten Merker, alist...@gmail.com, kr...@berkeley.edu, RISC-V SW Dev, h...@lst.de, an...@brainfault.org, atis...@gmail.com, padma...@gmail.com, j.neus...@gmx.net, phi...@hug.cx
Karsten, interesting note. Thanks. 

I was sorry not to see coreboot in your note anywhere, as every single 64-bit ARM chromebook sold uses coreboot, as do a fair number of other embedded ARM systems. Coreboot on 64-bit ARM does make use of ATF. 

I've worked with UEFI on and off since its creation 20 years ago, and I've become involved with it in the last two years as we work with servers and LinuxBoot, and while I certainly don't object to people wanting to be able to implement UEFI and ACPI for their own purposes, standardizing on UEFI and/or ACPI is not a good idea IMHO. 

There are too many of us working too hard at too many companies to remove as much UEFI as possible from our servers, and seeing it reappear on RISCV as a standard would be distressing to say the least.

ron

Luke Kenneth Casson Leighton

unread,
Oct 15, 2018, 2:29:02 AM10/15/18
to ron minnich, Michael Clark, paul.w...@sifive.com, Christoph Hellwig, Padmarao Begari, Anup Patel, atish patra, RISC-V SW Dev
On Sun, Oct 14, 2018 at 11:42 PM ron minnich <rmin...@gmail.com> wrote:

> So, I apologize for anything I did to make this thread longer.

summary analysis is always useful.

> I think we're all in agreement that we don't agree about the
> boot solutions,

i did say it's far too diverse! :) realistically, RISC-V, which is a
newcomer to this game, just has to accept that.

> but we do agree that things must be completely open, buildable,
> and flashable.

absolutely. that's the one consistent mistake that is persistently
occurring thanks to pathological spongeing corporate behaviour, and
the software libre community is expected to pick up the tab.

it's absolutely fundamentally in violation of systemic laws that
corporate organisations exploit free software developer goodwill.

l.

Luke Kenneth Casson Leighton

unread,
Oct 15, 2018, 3:53:28 AM10/15/18
to Michael Clark, Karsten Merker, alist...@gmail.com, Krste Asanovic, RISC-V SW Dev, Christoph Hellwig, Anup Patel, atish patra, Padmarao Begari, Jonathan Neuschäfer, phi...@hug.cx
---
crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68

On Mon, Oct 15, 2018 at 1:04 AM 'Michael Clark' via RISC-V SW Dev
<sw-...@groups.riscv.org> wrote:
>
> Hi Karsten.
>
> It would be very nice if you posted your commentary on a wiki page.

indeed: we know that mailing lists are a complete failure for
maintaining structured information.

> I need to learn more about u-boot. Perhaps it is more standardised
> than I think,

de-facto standard more than anything. pretty much every single
china-clone tablet in the world uses it. every china clone
smartwatch. it's also the top bootloader of choice for smartphones,
and for openwrt-based routers, networked cameras etc. etc.

basically its deployment is absolutely massive quantities... you just
don't get to hear about it because runs for about one to two seconds
behind usually GPL-violating products, where the manufacturer takes
what they can get, sells it via an OEM/ODM distribution chain as a
"cookie-cut" binary-only "solution".

l.

ron minnich

unread,
Oct 15, 2018, 5:01:26 PM10/15/18
to Paul Walmsley, Michael Clark, Christoph Hellwig, padma...@gmail.com, Anup Patel, atis...@gmail.com, sw-...@groups.riscv.org


On Mon, Oct 15, 2018 at 1:21 PM Paul Walmsley <paul.w...@sifive.com> wrote:
I'd submit that the Important criterion is that the hardware or platform
specifications don't foreclose either approach.


On 10/14/18 3:42 PM, ron minnich wrote:
> So, I apologize for anything I did to make this thread longer. I think
> we're all in agreement that we don't agree about the boot solutions,
> but we do agree that things must be completely open, buildable, and
> flashable. And that's the only thing I really care about.
>

I don't see an inconsistency between what you are saying and what I just said?
 

Karsten Merker

unread,
Oct 16, 2018, 2:50:51 PM10/16/18
to Michael Clark, Karsten Merker, alist...@gmail.com, kr...@berkeley.edu, RISC-V SW Dev, h...@lst.de, an...@brainfault.org, atis...@gmail.com, padma...@gmail.com, j.neus...@gmx.net, phi...@hug.cx
On Mon, Oct 15, 2018 at 01:04:37PM +1300, 'Michael Clark' via RISC-V SW Dev wrote:

[Thoughts about firmware/bootloader interfaces from a Linux
distribution's point of view]

> I largely agree. I disagree in some respects. i.e. it may be
> more costly and bug prone to take virtualized server system
> assets that target PCI and them use some as yet unknown
> undocumented standard.

I'm very sorry but I don't understand what you mean with
the last sentence. Could you elaborate that a bit?

Karsten Merker

unread,
Oct 16, 2018, 3:43:51 PM10/16/18
to Michael Clark, Luke Kenneth Casson Leighton, alist...@gmail.com, RISC-V SW Dev, Krste Asanovic, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx
On Mon, Oct 15, 2018 at 10:34:54AM +1300, 'Michael Clark' via RISC-V SW Dev wrote:
> > On 13/10/2018, at 1:47 PM, Luke Kenneth Casson Leighton <lk...@lkcl.net> wrote:
> >
> > if you start talking about UEFI to low-cost tablet vendors for
> > example they will just laugh at you. and if you even mention ACPI to
> > them as a required standard as part of a boot process, they'd be quite
> > likely to punch your lights out (and i would cheer and congratulate
> > them).
>
> Well they are uninformed.
>
> i.e. have not seen how trivial it is tool declare efi_main(),
> access console, timers and device-tree, load another file from
> the boot device, be that a simple cpio or zip in ROM
>
> I could code in a few minutes as I used ZIP filesystems on my
> android/iOS/macOS/windows/Linux OpenGL ES UI framework, as
> these are used for asset filesystems everywhere incl. most
> games (renamed .PAK or whatever). With UEFI/EBBR it’s just
> some simple C code. No assembly trampolines and low level
> M-mode code required at all.

Hello Michael,

possibly there has been a misunderstanding here. If I understand
the paragraph above correctly, you are arguing that the UEFI API
is easy to use and not complex from the viewpoint of a _consumer_
of the UEFI API, but the argument upwards in this thread wasn't
about using the UEFI API as a consumer, but about platform
vendors having to implement the UEFI+ACPI stack that provides
this API to a consumer, i.e. two different things.

Luke Kenneth Casson Leighton

unread,
Oct 16, 2018, 4:01:53 PM10/16/18
to Samuel Falvo II, Krste Asanovic, alist...@gmail.com, RISC-V SW Dev, Christoph Hellwig, Anup Patel, atish patra, Padmarao Begari, Jonathan Neuschäfer, phi...@hug.cx
On Sat, Oct 13, 2018 at 7:39 PM Samuel Falvo II <sam....@gmail.com> wrote:

> On Sat, Oct 13, 2018 at 10:17 AM <kr...@berkeley.edu> wrote:
> > Setting standards for what will be baked into silicon after huge
> > investment is very different that setting software standards where
> > systems can support multiple variants and be quickly updated in the
> > field.
>
> Hi Krste,
>
> This is correct, which is why I clarified that ITU-style has value in
> certain circumstances. I'm just expressing a fear that this approach
> will bleed over, if you will, into the software world where things
> that can be allowed to be made more reactive to market forces *should*
> be.

this is actually already happening.

the "when they're done, ready and completed in full and not before"
policy of releasing software modifications necessarily associated with
upcoming RV Standards is preventing and prohibiting all and any
discussion, involvement, contributions, reviews and understanding of
the *upstream* software projects.

RVV's LLVM modifications is one specific example. the developer who
has been tasked with adding RVV support to LLVM has not been given
authorisation by the RV Working Group to work in public.
consequently, neither SV nor ARM with the SVE extension, both of which
require the exact same variable-length vector data structure
"intrinsic" to be added to LLVM's "Intermediary Representation", may
get involved, comment, collaborate, discuss, review, understand, or
otherwise *actually* do what the entire purpose of "open software
development" is supposed to be about.

this in turn means that it is *PERSONALLY* costing me money - not a
corporation, not a university - i **PERSONALLY** am having to shell
out a fuck-load of money where i have absolutely no clue when or if i
am going to be able to get any more, in order to keep myself and my
family fed and a roof over our heads, to **DUPLICATE** the
functionality that the RVV LLVM developer is *ALREADY WORKING ON*,
because it's not publicly available.

this is not collaboration. it's not being "open", and it's why
RISC-V may not be categorised as an "Open" Architecture.

l.

Bin Meng

unread,
Oct 26, 2018, 12:46:46 PM10/26/18
to an...@brainfault.org, padma...@gmail.com, sw-...@groups.riscv.org, h...@lst.de
On Fri, Oct 12, 2018 at 3:58 PM Anup Patel <an...@brainfault.org> wrote:
>
> Yes, thats better.
>
> U-boot should run in S-mode only so that we can later on use same
> U-boot as bootloader for Guest/VM.
>
> In fact, it should also be possible run U-boot on QEMU Virt machine
> using BBL (i.e. using bbl+uboot).
>

FYI

U-Boot support for QEMU RISC-V Virt machine is now in u-boot/master,
and will be officially in the U-Boot v2018.11 release.

M-mode only, no SBI implementation yet. The plan is to boot Linux
directly without BBL.

Regards,
Bin

ron minnich

unread,
Oct 26, 2018, 2:21:13 PM10/26/18
to Bin Meng, an...@brainfault.org, padma...@gmail.com, sw-...@groups.riscv.org, h...@lst.de
On Fri, Oct 26, 2018 at 9:46 AM Bin Meng <bmen...@gmail.com> wrote:

M-mode only, no SBI implementation yet. The plan is to boot Linux
directly without BBL.


Good stuff, congratulations! Will you consider having Linux supply its own SBI implementation? That would be really neat, since then both coreboot and u-boot could let Linux manage SBI setup and use. We would then have the  "one common" implementation, save it is supplied by Linux, not firmware. 

ron 

atish patra

unread,
Oct 26, 2018, 8:47:23 PM10/26/18
to rmin...@gmail.com, bmen...@gmail.com, an...@brainfault.org, Padmarao Begari, sw-...@groups.riscv.org, h...@lst.de
@bmeng Awesome. Looking forward to your work separating BBL & Linux kernel.


--
You received this message because you are subscribed to the Google Groups "RISC-V SW Dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sw-dev+un...@groups.riscv.org.
To post to this group, send email to sw-...@groups.riscv.org.
Visit this group at https://groups.google.com/a/groups.riscv.org/group/sw-dev/.

atish patra

unread,
Oct 27, 2018, 1:37:37 PM10/27/18
to rmin...@gmail.com, bmen...@gmail.com, an...@brainfault.org, Padmarao Begari, sw-...@groups.riscv.org, h...@lst.de
As per my offline discussion with Ron, the here is one possible approach to implement SBI without duplicating effort across boot loaders or firmwares.
 
We can have a separate open project(with BSD or MIT Licensed) that just provides only SBI implementation(something like libsbi.a) and doesn't involve in booting/passing DT like BBL.
Any bootloader (i.e. coreboot/uboot) can choose to exist either in S or M mode and can include libsbi.a. Same goes with any vendor specific firmware.
In this way, we can avoid SBI fragmentation as well.

Any thoughts?

On Fri, Oct 26, 2018 at 11:21 AM ron minnich <rmin...@gmail.com> wrote:
--
You received this message because you are subscribed to the Google Groups "RISC-V SW Dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sw-dev+un...@groups.riscv.org.
To post to this group, send email to sw-...@groups.riscv.org.
Visit this group at https://groups.google.com/a/groups.riscv.org/group/sw-dev/.
To view this discussion on the web visit https://groups.google.com/a/groups.riscv.org/d/msgid/sw-dev/CAP6exYJGHN0JX87FP1-p_HWyh9j%2BGxJD0uYHm5wVBVZfvSbb9A%40mail.gmail.com.


--
Regards,
Atish

Luke Kenneth Casson Leighton

unread,
Oct 27, 2018, 2:49:02 PM10/27/18
to atish patra, ron minnich, bmen...@gmail.com, Anup Patel, Padmarao Begari, sw-...@groups.riscv.org, Christoph Hellwig
On Sat, Oct 27, 2018 at 6:37 PM atish patra <atis...@gmail.com> wrote:

> As per my offline discussion with Ron, the here is one possible approach to implement SBI without duplicating effort across boot loaders or firmwares.
>
> We can have a separate open project(with BSD or MIT Licensed) that just provides only SBI implementation(something like libsbi.a) and doesn't involve in booting/passing DT like BBL.
> Any bootloader (i.e. coreboot/uboot) can choose to exist either in S or M mode and can include libsbi.a. Same goes with any vendor specific firmware.
> In this way, we can avoid SBI fragmentation as well.
>
> Any thoughts?

sounds great. proposals to extend SBI with patches to e.g. include
more than one serial console (for debug purposes, power management,
PPPD networking-over-serial, ttyS0-3 redirection and virtualisation
which is an essential and little-known part of the u-boot development
and management processes and procedures), would only need to submit
patches to one location rather than several fragmented ones.

l.

Michael Clark

unread,
Oct 27, 2018, 3:17:58 PM10/27/18
to Karsten Merker, Luke Kenneth Casson Leighton, alist...@gmail.com, RISC-V SW Dev, Krste Asanovic, Christoph Hellwig, Anup Patel, atis...@gmail.com, padma...@gmail.com, Jonathan Neuschäfer, phi...@hug.cx
Yes.

That said I don’t necessarily think the EBBR subset is that onerous to implement as a service provider. Perhaps a more full UEFI implementation. The thing is, I read the spec and it provides most of what I need from SBI plus fills many gaps, for early boot, like timers and interrupts.

SBI on the other hand has runtime functions and doesn’t delineate boot services. The expectation is that Linux has a driver for the interrupt controller (Timers, IPIs) and doesn’t use SBI. This will likely be solved when virtualization is addresses in the PLIC and the CLIC with the VPLIC and VCLIC (if those names hold).

My text regards ACPI was a bit garbled. There is a risk of porting working software for PCI based systems that use ACPI to some unknown solution or trying to overload something like device-tree with some as yet unknown representation of some function that it was not designed for. device-tree is certainly okay for simple static configurations.

There is also much duplication in the stack, such as UEFI byte code and AML. UEFI byte code I believe is largely unused. AML is used at runtime. The x86 kernel developers likely know which bits and pieces they find useful vs awkward.

I think the models to look at are arm’s EBBR and SBBR. Using modern elements like GPT are no brainers.

Michael Clark

unread,
Oct 27, 2018, 3:58:52 PM10/27/18
to ron minnich, Bin Meng, an...@brainfault.org, padma...@gmail.com, sw-...@groups.riscv.org, h...@lst.de
We have to think about this a little as there are already some boundaries that need to be more carefully defined. e.g. the PLIC currently exposes M-mode apertures to the kernel and due to the stride, PMP cannot be used to protect them and the PLIC has no SBI interface. I think we need to evolve the interrupt controller specifications, such that M-mode maintains control either with software-based delegation or hardware-based accelerated interrupt virtualization so S-mode has autonomous control over a slice of the interrupt controller.

This is being addressed in the CLIC spec to some degree, initially for MCU class systems. The CLIC has a distinct M-mode aperture and an optional S-mode aperture as well as the ability to configure which mode has control over which interrupts. The model still needs more testing, when there are Linux class systems that have the CLIC. There are currently no Linux drivers for the CLIC. There also needs to be evolution of the PLIC. I would like to see a possibility for a variable number of timers so that S-mode can have direct control over a hardware timer and M-mode doesn’t need to worry about interrupt starvation. Of course we need to maintain the ability to scale down to systems with a single timer and software based delegation (as present).

We probably want to keep SBI as it is in the short-term which means the bootloaders need a stay-resident monitor that handles SBI ecalls and instruction emulation (misaligned loads, stores, softfloat, etc). I know misaligned load stores is handled by Linux kernel on some platforms. The code in BBL is BSD licensed so I guess BBL can be canabalized.

The point being, if SBI code moves to Linux, then there is no SBI, however the issue is the M-mode S-mode privilege boundary. i.e. Linux can’t provide an M-mode payload to the boot loader to run with higher privileges; as that would be a gaping privilege escalation hole, similarly to the still currently open hole where a malicious supervisor can map the memory at mtvec, overwrite the code and escalate to M-mode (this is well known, and it was known before Don Bailey did a POC as we did not have PMP at that time). Now we can just starve M-mode from platform interrupts. We should avoid ugly PC style hacks where the monitor does things like constantly writing back the PLIC M-mode registers using the timer interrupt it still controls.

BBL is currently likely still the most featureful boot loader (as if SBI is not implemented then is emulation implemented?). In fact, with a small change to protect 0x80000000 - 0x80200000 (on RV64) with PMP, it will be the most secure boot loader. This is likely a 10 line patch.

I think now is probably a good time to make the 10 line patch for BBL and it can be the baseline for security. If the other boot loaders are not implementing SBI then what has happened to the M-mode monitor?

Fortunately this is happening pre-production so we can talk publicly about the security (we still need KASLR and ASLR in Linux too)...

~mc

Luke Kenneth Casson Leighton

unread,
Oct 27, 2018, 7:04:58 PM10/27/18
to Michael Clark, ron minnich, bmen...@gmail.com, Anup Patel, Padmarao Begari, sw-...@groups.riscv.org, Christoph Hellwig
On Sat, Oct 27, 2018 at 8:58 PM 'Michael Clark' via RISC-V SW Dev
<sw-...@groups.riscv.org> wrote:

> This is being addressed in the CLIC spec to some degree,
> initially for MCU class systems.

i've not seen any announcement or discussion of that, anywhere, on
any of the riscv lists, and i've been on these lists for nearly a
year.

on which what publicly-accessible forum is the CLIC being actively
discussed? i would like to see people who have a huge amount of
experience in the open hardware and software communities feel that
they are welcome to actively participate.

l.

ron minnich

unread,
Oct 27, 2018, 9:13:45 PM10/27/18
to Michael Clark, Bin Meng, Anup Patel, Padmarao Begari, sw-...@groups.riscv.org, Christoph Hellwig
On Sat, Oct 27, 2018 at 12:58 PM Michael Clark <michae...@mac.com> wrote:

> The point being, if SBI code moves to Linux, then there is no SBI, however the issue is the M-mode S-mode privilege boundary. i.e. Linux can’t provide an M-mode payload to the boot loader to run with higher privileges; as that would be a gaping privilege escalation hole, ...

I've never understood the argument that kernel access to M mode is a
priv escalation. Let's just say the kernel replaces M mode with some
other code. What did it do? Attack itself? The argument that it is a
priv escalation makes no sense to me.

From my point of view, the existence of runtime firmware that the
kernel can not view or change is the real security issue, and events
of the last 10 years have strengthened that view.

The abililty of a kernel to change the M mode code was, in my view, a
real step forward from systems like x86 system management mode, with
its hidden code running on the main CPU in SMM or UEFI.

Now if you want to argue that a kernel should be able to install
something like SMM or M mode code, then lock it down so it can't be
changed, that's not so bad, and for SMM I've done this proof of
concept: https://docs.google.com/presentation/d/1ECTPzrt4hT37Ef729GAXh0o9lMvx4MJrkJxN88-k9Ps/edit?usp=sharing

I've never viewed the RISCV architectural extensions (such as PMP)
that would allow ODMs to create SMM-like code a step forward for RISCV
security. The work that I and others are doing on LinuxBoot is aimed
at disabling, removing, or replacing runtime firmware blobs, be it on
the main CPU or attach CPUs such as ME or PSP. Again, if the kernel
can use PMP to lock down regions it wants to protect, fine; if vendors
are using PMP to lock in M mode code that a kernel can not read or
change, that's a very bad idea.

At many companies, runtime firmware such as UEFI or SMM is living on
borrowed time. I would not like to see it reappear on RISCV, but that
seems to be the goal in some cases.

thanks

ron

Michael Clark

unread,
Oct 27, 2018, 10:30:58 PM10/27/18
to ron minnich, Bin Meng, Anup Patel, Padmarao Begari, sw-...@groups.riscv.org, Christoph Hellwig


On 28/10/2018, at 2:13 PM, ron minnich <rmin...@gmail.com> wrote:

On Sat, Oct 27, 2018 at 12:58 PM Michael Clark <michae...@mac.com> wrote:

The point being, if SBI code moves to Linux, then there is no SBI, however the issue is the M-mode S-mode privilege boundary. i.e. Linux can’t provide an M-mode payload to the boot loader to run with higher privileges; as that would be a gaping privilege escalation hole,  ...

I've never understood the argument that kernel access to M mode is a
priv escalation. Let's just say the kernel replaces M mode with some
other code. What did it do? Attack itself? The argument that it is a
priv escalation makes no sense to me.

Of course it’s a privilege escalation. Linux is a Supervisor and BBL is the monitor. M-mode and the PMP mechanism is how a secure world equivalent is implemented on RISC-V.

It’s a ten line patch to bbl. I will test and submit a patch at some point. Certainly bbl was not originally designed as a secure monitor but it is trivial for it to ensure it has control over M-mode. Whether it’s accepted or not is another issue. I can maintain a fork.

I understand there are folk who want systems where there is control over what runs in M-mode for guaranteed interrupt latency purposes (RTOS with VIrtual Memory?), but privilege separation is required to implement context switching for VMMs. It is a privilege escalation for S-mode to gain control of M-mode, unless that is a documented feature.

Of course, in that case Linux might as well start in M-mode, so it’s not a firmware / bootloader bug, rather a Linux bug that allows escalation from S-mode to M-mode.

From my point of view, the existence of runtime firmware that the
kernel can not view or change is the real security issue, and events
of the last 10 years have strengthened that view.

The abililty of a kernel to change the M mode code was, in my view, a
real step forward from systems like x86 system management mode, with
its hidden code running on the main CPU in SMM or UEFI.

Sure, it’s great that the person deploying a RISC-V system has this level of control, but that’s besides the point. Clearly mapping the memory under the trap vector of the monitor is a “back door”. BBL not labelled as a secure monitor and this is a relatively well known vector but it’s also relatively trivial to fix.

I’m quite folk will want to deploy their own monitors and will like the ability to do this, but I’m also sure that some folk developing VMM’s by nature will not want the Supervisor to break out of its VM. From that POV, there will be folk virtualizing M-mode (it’s not that hard, except for perhaps mstatus.MPRV which is itself an emulation feature that does not need to be present in Virtual M-mode, for whichever virtualization solutions that allow running boot loaders in a VM).

Now if you want to argue that a kernel should be able to install
something like SMM or M mode code, then lock it down so it can't be
changed, that's not so bad, and for SMM I've done this proof of
concept: https://docs.google.com/presentation/d/1ECTPzrt4hT37Ef729GAXh0o9lMvx4MJrkJxN88-k9Ps/edit?usp=sharing

I've never viewed the RISCV architectural extensions (such as PMP)
that would allow ODMs to create SMM-like code a step forward for RISCV
security. The work that I and others are doing on LinuxBoot is aimed
at disabling, removing, or replacing runtime firmware blobs, be it on
the main CPU or attach CPUs such as ME or PSP. Again, if the kernel
can use PMP to lock down regions it wants to protect, fine; if vendors
are using PMP to lock in M mode code that a kernel can not read or
change, that's a very bad idea.

At many companies, runtime firmware such as UEFI or SMM is living on
borrowed time. I would not like to see it reappear on RISCV, but that
seems to be the goal in some cases.

I don’t think we are necessarily in disagreement. The question is “who” has which capability and RISC-V there is a choice for users of deployments to have control over M-mode versus their vendors. The question is whether this should be extended to anyone with insmod permissions.

It can be rigged for it to appear like they can write to mtvec anyhow. M-mode can be emulated in U-mode if one gets SIGILL for CSR instructions. MMIO emulation really needs control over the page tables.

My question about SBI, is why is it needed if  u-boot can provide UEFI with device-tree to the Supervisor. This is already a standard on x86 and arm with EBBR, SBBR, device-tree or ACPI


~mc

Bin Meng

unread,
Oct 27, 2018, 10:31:41 PM10/27/18
to rmin...@gmail.com, an...@brainfault.org, padma...@gmail.com, sw-...@groups.riscv.org, h...@lst.de
Hi Ron,
My plan is to implement SBI in U-Boot. However I do like your idea of
implementing M-mode by Linux. Frankly speaking, I don't quite
understand why RISC-V reference software stack (BBL + Linux) was
initially designed this way. Looking at the SBI functionalities, I see
no reason why these features cannot be provided by kernel directly.
Furthermore, I don't see values of limiting functionalities in the
M-mode. For example, simply setting a timer count has to go through a
context switch between S-mode and M-mode, what's point? It's just an
MMIO or CSR write if we implemented it in S-mode. Having such
implemented in M-mode affects performance a lot. So do other SBI
interfaces.

Regards,
Bin

ron minnich

unread,
Oct 27, 2018, 10:55:18 PM10/27/18
to Michael Clark, Bin Meng, Anup Patel, Padmarao Begari, sw-...@groups.riscv.org, Christoph Hellwig
On Sat, Oct 27, 2018 at 7:30 PM Michael Clark <michae...@mac.com> wrote:

> Of course it’s a privilege escalation. Linux is a Supervisor and BBL is the monitor. M-mode and the PMP mechanism is how a secure world equivalent is implemented on RISC-V.

and there's the difference in our thinking. I don't want a monitor.

But it's ok, since currently it will remain possible for the kernel to
install its own M mode. You and I will each be able to do what we
want.

anyway thanks Michael, I think we don't disagree too, and I look
forward to seeing where your work takes you, it sounds interesting.

ron

Anup Patel

unread,
Oct 28, 2018, 12:40:03 AM10/28/18
to atish patra, rmin...@gmail.com, bmen...@gmail.com, Padmarao Begari, sw-...@groups.riscv.org, Christoph Hellwig
On Sat, Oct 27, 2018 at 11:07 PM atish patra <atis...@gmail.com> wrote:
>
> As per my offline discussion with Ron, the here is one possible approach to implement SBI without duplicating effort across boot loaders or firmwares.
>
> We can have a separate open project(with BSD or MIT Licensed) that just provides only SBI implementation(something like libsbi.a) and doesn't involve in booting/passing DT like BBL.
> Any bootloader (i.e. coreboot/uboot) can choose to exist either in S or M mode and can include libsbi.a. Same goes with any vendor specific firmware.
> In this way, we can avoid SBI fragmentation as well.
>
> Any thoughts?

This seems much better. It will be bootloader's choice to be either in
S-mode or M-mode. Maybe bootloader can have compile-time option for
runtime mode.

I was also thinking that libsbi open project can create both libsbi.a
and sbi.bin where sbi.bin is platform-specific runtime services blob.
Bootloaders that don't want any M-mode code can use sbi.bin in M-mode
while bootloaders that want M-mode code can link to libsbi.a

Regards,
Anup

Michael Clark

unread,
Oct 28, 2018, 4:30:43 AM10/28/18
to Bin Meng, rmin...@gmail.com, an...@brainfault.org, padma...@gmail.com, sw-...@groups.riscv.org, h...@lst.de
You may have issues.

Some of the CLINT local interrupt registers can only be controlled by M mode. You can delegate the traps to go to S-mode but S-mode may not be able to read the mstatus bits.

mtimecmp is okay as it is MMIO and a pending interrupt can be acknowledged (well, not quite acknowledged as it’s a level interrupt) by writing a future timer value. There is also the MMIO msip register for each hart. I would need to check the spec, but this might require M-mode to clear mstatus.MSIP?

Note: SiFive’s CLINT is currently filtered in bbl and it is not intended to be exposed directly to the supervisor rather the SBI is the current paravirtual interface to timer interrupts and IPIs. Local Interrupts are still a work in progress.

You might find the patches to add the non-SBI timer and IPI interrupts won’t be accepted in upstream Linux. Your best bet is to stick with SBI in the medium term until there is a well defined evolutionary path for Supervisor local interrupts.

There is some work to virtualize interrupts. The CLIC has improved virtualization support i.e. it has an S-mode aperture.

Luke Kenneth Casson Leighton

unread,
Oct 28, 2018, 4:44:33 AM10/28/18
to Michael Clark, bmen...@gmail.com, ron minnich, Anup Patel, Padmarao Begari, sw-...@groups.riscv.org, Christoph Hellwig
On Sun, Oct 28, 2018 at 8:30 AM 'Michael Clark' via RISC-V SW Dev
<sw-...@groups.riscv.org> wrote:

> The CLIC has improved virtualization support i.e.
> it has an S-mode aperture.

where is that being publicly discussed in an unrestricted forum so
that the correctness and security of the approach may be evaluated,
particularly by the parties whose input may be needed at a future
point, *particularly* those who may be called on to *deal* with
incorrect or insecure design flaws and implementations [1]?

l.

[1] https://www.zdnet.com/article/linus-torvalds-talks-frankly-about-intel-security-bugs/

Jim Wilson

unread,
Oct 29, 2018, 2:14:50 PM10/29/18
to Luke Kenneth Casson Leighton, Michael Clark, bmen...@gmail.com, ron minnich, an...@brainfault.org, padma...@gmail.com, sw-...@groups.riscv.org, Christoph Hellwig
On Sun, Oct 28, 2018 at 1:44 AM Luke Kenneth Casson Leighton
<lk...@lkcl.net> wrote:
> On Sun, Oct 28, 2018 at 8:30 AM 'Michael Clark' via RISC-V SW Dev
> <sw-...@groups.riscv.org> wrote:
>
> > The CLIC has improved virtualization support i.e.
> > it has an S-mode aperture.
>
> where is that being publicly discussed in an unrestricted forum so
> that the correctness and security of the approach may be evaluated,
> particularly by the parties whose input may be needed at a future
> point, *particularly* those who may be called on to *deal* with
> incorrect or insecure design flaws and implementations [1]?

You can find a copy of the draft CLIC spec here
https://github.com/sifive/clic-spec

The CLIC spec is being discussed in the RISC-V Foundation Fast
Interrupts task group. Any RISC-V Foundation member can join. Though
there has not been much discussion lately.

Jim
Reply all
Reply to author
Forward
0 new messages