On 11/8/23 09:11, Krzysztof Kozlowski wrote:
On 08/11/2023 08:16, Michal Simek wrote:
On 11/7/23 22:36, Conor Dooley wrote:
+CC Palmer
On Tue, Nov 07, 2023 at 01:38:15PM +0100, Krzysztof Kozlowski wrote:
On 07/11/2023 13:09, Michal Simek wrote:
On 11/7/23 12:27, Krzysztof Kozlowski wrote:
On 07/11/2023 12:09, Michal Simek wrote:
On 11/6/23 18:07, Conor Dooley wrote:
On Mon, Nov 06, 2023 at 12:53:40PM +0100, Michal Simek wrote:
+description: |
+ AMD boards with MicroBlaze V SOC
+
+properties:
+ $nodename:
+ const: '/'
+ compatible:
+ oneOf:
+ - description: AMD MicroBlaze V
+ items:
+ - const: amd,mbv
You don't actually list any boards here, but instead permit having only
the SoC compatible and no board one. The SoC compatible is also
incredibly generic. Personally I don't think this binding makes any
sense as it appears to exist as a catch all for anything using your
new cores in any combination.
I think I need to define any string for compatibility because it is standard
property. Because this is soft core it can be added to any board with AMD/Xilinx
chip. I don't have really an option to list all boards.
Why? Either there is a product with this soft-core or there is not. It
cannot be both.
I am doing basic enablement. I am not making product. Product will be done by
our customers using this core.
There will be thousands of different configurations done by customers which will
have products with it. Also there could be hundreds configurations done on the
same board.
If this is the same board, then why there is compatible for it?
Does it make sense to have board related compatible string like this if this
evaluation board is used by a lot of customers?
"amd,kcu105-mbv-ABC-vXYZ", "amd,kcu105-mbv", "amd,mbv"
I miss the point what is the hardware. Evaluation board is the hardware.
If someone changes it and makes a new product, it is a new product.
To me, this does actually make (some) sense.
The first compatible is "soc" + board + design.
The second is "soc" + board.
The final one is the "soc".
I say "soc" though, because it is not a single soc - it could be any
configuration of these soft AMD cores on an FPGA in any quality,
possibly set up heterogeneously too. I don't think trying to define a
generic compatible for it like this makes sense as the soc part does
not come close to identifying a specific device.
I am fine to skip it but it is required property which has to be defined somehow.
Until someone actually creates a product that ships with this, I don't
think it makes sense to try and define a binding.
I disagree with this description. I am coming from FPGA world where it is too
late to start to deal with dt binding when you have a product.
You are saying release the product and then start to deal with dt binding. When
product is released there must be stable dt binding.
I meant that creating a binding for something which is not and will not
be a product does not bring any benefits. Why do we even care to
document it? Who requires it? I don't. I don't see DTS or driver, no
need for compatible.
That's why entire discussion starts with DTS (and/or driver).
We have dt description for soft IPs like uartlite
Documentation/devicetree/bindings/serial/xlnx,opb-uartlite.yaml
We have 16550 compatible IP with
Documentation/devicetree/bindings/serial/8250.yaml
Simple ethernet core
Documentation/devicetree/bindings/net/xlnx,emaclite.yaml
Axi ethernet
Documentation/devicetree/bindings/net/xlnx,axi-ethernet.yaml
Adi clock generator
Documentation/devicetree/bindings/clock/adi,axi-clkgen.yaml
Adi fan control
Documentation/devicetree/bindings/hwmon/adi,axi-fan-control.yam
Adi adcs
Documentation/devicetree/bindings/iio/adc/adi,axi-adc.yaml
and much more.
They are IPs from vendor catalogs. We can talk if it is a product (definitely
yes if you need to buy it for your design). But all of them fit to the same
category that you are composing your HW design with them.
All of them as standalone can't run. You will never create a product with just
uartlite IP. You need to add cpu, clocks, reset logic and others around to make
a product out of it.
Our mental model is HW designer create new IP, we are writing driver for it,
customers can buy it (or get it for free) and use it.
They put it to their design, create custom board and sell it as a product.
And in this particular case HW designed create risc-v compatible CPU.
I expect this should fine
https://lore.kernel.org/r/d442d916204d26f82c1c3a924a4cdfb117960e1b.1699270661.git.michal.simek@xxxxxxx
And discussion what we are having is pretty much about how to share the view on
the system.
I spoke to Palmer a bit about this, and what I think would make sense is
if you had some sort of "reference design" bitstream that people could
download and run on xyz AMD devkit. If that existed, then we could
document that configuration etc. Otherwise you're in the same spot that
a lot of IP vendor stuff is, where without there being something that
qualifies as "real hardware" using the core, it doesn't make sense to
try and create bindings etc. It's the same for the various people in
the RISC-V community that created their own CPUs that they run on FPGAs.
Aren't all ARM FVP models enabled by SW before soc vendors put them to a real
chip? Is there any real product available at that time?
FVP also finished one. They do not claim they added compatible for a SoC
or CPU. And that's my impression here.
Are these real chips?
compatible = "arm,foundation-aarch64", "arm,vexpress";
compatible = "arm,fvp-base-revc", "arm,vexpress";
FVP are Fixed Virtual Platforms. Pretty much emulators similar to QEMU.
I will try to find out if there is any official plan for releasing any reference
design against any evaluation board with commitment to supporting it.
Or I can define qemu one.
"amd,qemu-mbv", "amd,mbv"
QEMU is not hardware, so not.
I am still trying to wrap my head around it. In qemu we are actually going to
create model for this configuration but based on what you are saying here we
shouldn't really have DT which describes it.
That's why we likely end up in situation that qemu create DT description self,
put it to memory and u-boot/kernel will consume it. The only difference is going
to be that DT will be used but won't be checked against dt-schema.
I personally prefer to have DT pass dt-schema checking and tell qemu guys, this
is what qemu should generate.
But if you think that this is wrong approach I will let them generate whatever
they want and will just check functionality. It means u-boot won't have DT,
Linux won't have DT and I am done.
Sorry, I am confused now. Are we talking about real hardware or QEMU SW
model? Your description clearly said:
"AMD boards with MicroBlaze V SOC"
so QEMU is not a board. Board has a physical form, a shape. Usually flat.
Let me describe what we do for all our SOCs but Microblaze is the best example here.
Customers open design tools (right know Vivado) and design their system there.
Choose cpu and it's configuration like barrel shifter, divider, multiplicator,
size of caches. Then put there interrupt controller, timer, consoles, ethernet,
spi, i2c, etc. For all IPs you need to choose mmio base address and connect them
to any interrupt line you like.
You normally target a board, evaluation platforms or just standalone chips which
you use on your custom boards.
And build bitstream (configuration for FPGA) and also going over our device tree
generators which generate DT for describing the system.
Very old example is for example visible here
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/microblaze/boot/dts/system.dts?h=v6.6
Another example is mb-v description mentioned in previous thread.
The reason is simple with a lot of IPs in the design none will be able to get
description right in connection to addresses and especially interrupt numbers.
It means at this stage you have bistream for your board and you have DTS
(without board specific information like i2c devices, ethernet phy, etc)
For 10+ years our qemu is taking input as DTB and create qemu model based on it.
It means you say via DT I want this cpu core, this timer at 0x..., interrupt at
0x..., uart at 0x..., etc. and qemu generates model for it. Pretty much the same
DT can be consumed by SW to run it on the model.
We reached the state that you have qemu model which reflects your design choice
and at the same time you have hardware for your board.
It means same DT describe qemu configuration and also hardware.
I think customers should be adding their compatible string in front of generic one.
To what? To evaluation board? Why?
Years ago I have done the same thing with Microblaze where compatible is defined
as xlnx,microblaze only.
Again, what is the use of such binding?
When customer take this soft core, put IPs around and
create a product they should extend it to be for example like this.
"xyz,my-product-1.0", "xlnx,microblaze";
So there is a product, not evaluation board.
And over all of years I have never seen any single customer to try to push dt
description for any Microblaze based product.
I am happy to change it to something else but not sure to what.
Alone this compatible does not bring you anything.
I don't agree with it. It is standard property and I have to define it somehow.
The property is already defined, you do not have to define it. What you
define here is the value for compatible. Why do you need to define it
somehow? Who asks for that?
If not, I get an error.
.../xilinx-mbv32.dtb: /: 'compatible' is a required property
So you have a board? The patches must be linked to each other, e.g.
preferred way is to send them in one patchset.
There were patches sent to U-Boot for an example configuration:
https://lore.kernel.org/u-boot/d488b7016e0d1b1324c64d8a8b2f033851aab6c6.1699271804.git.michal.simek@xxxxxxx/
thx.
And it tells me that this risc-v compatible core runs on AMD fpga and it is
compatible with it.
Basically, it provides no more specific information than the cpu node
does. From me, it's a NAK for a compatible like this that that permits
using it in isolation for any core configuration and combination.
I hope you agree that if we ignore compatible string for the whole DT that DT
describes HW is very reasonable way. It means you know what CPU is, what timer,
console and intc are.
It clearly shows how MicroBlaze-V AMD design will be done for others who are
interested on using it.
It also shows and explain what drivers needs to be enabled in defconfigs.
In connection to above u-boot. I can simply remove DT patch from that series (or
move it to readme) and just say these IPs will be used that's why I am enabling
them and point them to public announcement. And configure that DT will be passed
via address in DDR and that's it.
But I simply think that DT is providing better view on the system.
We will of course extending our device tree generators to provide DT description
directly from the tool as a starting point for customers.
You called the file:
"AMD boards with MicroBlaze V SOC", which means "AMD MicroBlaze V" is
the name of the SoC.
It is official name for risc-v compatible core implementation for AMD/Xilinx
FPGAs. Strictly speaking just name of cpu core.
SoC is cpu with other components around.
DT is describing one of the configuration with MicroBlaze V cpu core with
related IP (timer, intc, serial) which together create SOC.
World boards is likely what should be fixed. Maybe design should be used instead.
Now you create one compatible, not a list, for "AMD MicroBlaze V", so
only for the SoC. That's the discussion about. You cannot create a
final, real and working hardware with this SoC alone. To make it
working, you need to put it into some board.
I can agree with this description.
Now, for SW models,
probably you can... but then how do you run it?
I can put this sw compatible design to thousands of boards but I can't say which
one. As of today this SOC configuration/design is built and running on kcu105.
https://www.xilinx.com/products/boards-and-kits/kcu105.html
But in very easy way it can run on ac701
https://www.xilinx.com/products/boards-and-kits/ek-a7-ac701-g.html
If under QEMU then show
me the QEMU machine file - I bet there are some specifics for this given
configuration. This means you have specific configuration, not a generic
SoC.
But I want to show it as reference only to see how that DTs will look like.
And again we are at the same spot as with hardware - you could not
run the SW model without configuring it, thus you cannot run a "SoC".
And in our case we use DT for SW model configuration too.
Thanks,
Michal
