On 22/08/2024 14:12, Caleb Connolly wrote:
On 22/08/2024 04:59, Simon Glass wrote:
Hi Caleb,
On Wed, 21 Aug 2024 at 14:33, Caleb Connolly caleb.connolly@linaro.org wrote:
On 21/08/2024 20:27, Simon Glass wrote:
Hi Caleb,
On Wed, 21 Aug 2024 at 10:47, Caleb Connolly caleb.connolly@linaro.org wrote:
On 21/08/2024 18:16, Simon Glass wrote:
Hi Caleb,
On Wed, 21 Aug 2024 at 08:49, Caleb Connolly caleb.connolly@linaro.org wrote: > > > > On 21/08/2024 16:37, Simon Glass wrote: >> Hi Caleb, >> >> On Wed, 21 Aug 2024 at 08:11, Caleb Connolly caleb.connolly@linaro.org wrote: >>> >>> Hi Simon, >>>>> +U_BOOT_DRIVER(gcc_sc7280) = { >>>>> + .name = "gcc_sc7280", >>>>> + .id = UCLASS_NOP, >>>>> + .of_match = gcc_sc7280_of_match, >>>>> + .bind = qcom_cc_bind, >>>>> + .flags = DM_FLAG_PRE_RELOC | DM_FLAG_DEFAULT_PD_CTRL_OFF, >>>>> +}; >>>> >>>> This should use driver model, with a UCLASS_CLK and UCLASS_RESET >>> >>> Please refer to qcom_cc_bind() which binds the clock, reset, and power >>> domain drivers. >> >> Gosh, why? Are these devices not in the devicetree? > > They are, the gcc block contains clock, reset, and pd parts. On Linux > this is not an issue because a single device can be multiple different > classes (e..g when you register a reset you do it for a device) whereas > U-Boot requires a device per class. > > e.g. see devm_reset_controller_register() in Linux, it populates a > struct reset_controller_dev which references the struct device created > for the node. In U-Boot you have to create a new device which _is_ the > reset controller.
OK, I see. Rockchip has a CRU (Clock & Reset Unit) which uses syscon to access registers. The clock driver 'owns' the node in U-Boot and it manually binds a reset driver. It isn't great, either.
Looking at drivers/clk/qcom/clock-sdm845.c (for example), I can't actually find "qcom,gcc-sdm845" (for example) in U-Boot, except as a binding. Can you please point me to the node?
It's in dts/upstream/src/arm64/qcom/sdm845.dtsi
Re devm_reset_controller_register(), yes the U-Boot driver model is a lot more regular, e.g. we don't really want drivers creating their own struct udevice. We want all devices to be created automatically by the devicetree and most memory allocations to be done automatically. This helps to reduce code size and execution time. You probably know all this :-)
Yeah, U-Boot's model is simpler for most cases. This makes sense. But it doesn't reflect the reality of DT so well in cases like this.
To a significant degree, the devicetree bindings are created without much thought to efficient operation in U-Boot. I hope that eventually this might change.
I strongly disagree with this mental model. This is the approach I see vendors take in their BSP sources and the result is not pleasant.
DT should (within reason) never be written with the OS in mind. It is an agnostic structure to describe the hardware. I think the new power sequencing subsystem in Linux does a good job at embodying how we should approach consuming DT.
I'm only really involved in mainline and don't really see vendor trees much. An example is where pinctrl has a GPIO controller but it is not mentioned in the devicetree. It would be better for U-Boot to add a subnode for each GPIO controller. In general, if the SoC has a device, it should be in the devicetree.
The concept of a device is an OS one. DT is not "telling the OS how to use the hardware", it is describing the hardware.
This distinction is important because it's the only way to ensure that future OS changes can be done regardless of the DT. And also of course because different OS's will have different ideas of how to model devices (case in point).
The GCC block on Qualcomm platforms is a single hardware block. The datasheets and hardware programming guides describe it as such. The clocks, resets, and GDSCs are all entwined at the hardware level. They also have overlapping register addresses.
The further away DT gets from describing the hardware in favour of simplifying the OS, the more likely we are to start running into issues with fitting hardware changes into our arbitrary model.
I completely agree with everything you are saying, but you don't go far enough. We should additionally require that all hardware has a description in the devicetree. See for example the GPIO controller I mentioned. When Linux wants it, it gets it, when it doesn't, it isn't there. Sorry to have to say it, but that's not right.
If it's only used in U-Boot, that's justification enough to add the node upstream?
Part of this difference (between U-Boot and Linux) comes about because Linux device setup is fairly manual, whereas U-Boot tries to put all of that in common DM code. Whenever you are including dm/device-internal.h that is often a sign that the binding is causing issues.
To me this indicates an inability for U-Boot's DM to handle complicated devices. I don't think U-Boot should dictate the design of devicetree.
I don't know how else to describe this. This issue has been litigated over and over again on the kernel mailing list. Every time someone suggests changing DT because of a limitation in Linux they are (rightfully) shut down. If these changes were accepted then it would be impossible for DT to be an OS agnostic hardware description, because it would be full of OS specific hacks.
Given that an OS has no size limitations so can afford to do just about anything to deal with one-off cases in each SoC, sure. But let's face it, it isn't project-agnostic. I could provide dozens of examples where the bindings are a pain for U-Boot. Even the use of strings in some of the SoCs' pinctrl bindings is painful. My point is that there are many ways to model and describe the hardware and taking more account of all users would be a big step forward.
Sure, I can understand wanting to prioritize size/speed. I think livetree would help a lot here.
I'm happy for you to change my mind.
Anyway, with the devicetree we have, I wonder how we could do this better?
Some ideas:
- Allow DM to bind multiple devices to each devicetree node, perhaps
as a Kconfig option (to avoid time penalty for all boards), or by requiring a new DM_FLAG_MULTI_NODE flag. The devices would then be independent, with no common parent device
This would make sharing match data hard, and likely cause issues with generic compatible strings.
You would have to repeat the same compatible string in each driver.
For generic compatible strings we could a) worry about it later or b) restrict this technique to only nodes with a single compatible string, or c) use the driver flag as mentioned
- As 1 but have DM create a parent 'UCLASS_MULTI' device
automatically. I am thinking that we should have a new uclass for these things, or rename the NOP uclass. This option would allow easy access to the parent device, if needed.
This is the current approach, we just bind the clk/reset/pd drivers explicitly, allowing us to create and share common data. I don't believe there is a sensible way to do this generically.
How come? What is qcom_cc_bind() doing which DM couldn't?
Maybe DM could look at the "#clock-cells", "#reset-cells", and "#power-domain-cells" properties and match on compatible string + uclass? I think that could work.
Hmmm you mean when it sees those in the node it knows the additional uclasses it is allowed to use?
Yes, the "#reset-cells" property literally means "this is a reset controller".
Well technically it's more "here is the number of parameters I expect for phandles to my node as a reset controller", but since it's required to support phandles for reset handles, having the #<>-cells can be an hint that the node is supposed to act as a reset controller since some other nodes references the current node via a reset phandle.
Since won't solve node that can act as provides _and_ other services, like it's common to have like a mmc controller that can also act as clock controller, you'll have the #clock-cells, but no #mmc-cells since no other nodes depends on it.
You also have the case for firmware nodes, where the link between users and consumers is not described in DT but only in code, like the qcom scm or the amlogic securemonitor.
So I won't base myself on the #<>-cells property.
Neil
- Implement devices which are in more than one uclass. There would
still be a primary uclass, but others can be added too. This would involve declaring a list of secondary uclasses in the U_BOOT_DRIVER() macro. We would then have a struct dmtag_node for each secondary uclass, containing the ID and the uclass pointer. Certain functions would need to be updated to support this, and again it could be behind a Kconfig.
Many device classes in U-Boot rely on going from a struct udevice to some uclass specific data or ops. I have always found this to be a bit odd, though simpler to deal with than Linux.
In U-Boot the uclass is a stronger concept, e.g. you can generically iterate through all devices in a uclass, and all devices have one.
What do you think?
I think if we're to try and solve this at all, the Linux model is by far the most sensible. It is already tried and tested, and would have the huge bonus of simplifying driver porting.
barebox went with this approach and it seems to have worked out quite well for them.
What is the Linux model, in this sense? Whenever I see barebox I wonder why we can't fold whatever new features it needs into U-Boot...perhaps the code bases have converged too much...?
barebox is livetree only (like Linux) as I understand it, and I think it aims to be literally copy/paste compatible with Linux. I would love for U-Boot to adopt this approach.
OK. We do have code-size restrictions though.
All that being said, while it's taken me some time to get my head around "the U-Boot way", I think there is still value in the simplicity of U-Boot's approach. I also think the solution we've ended up with (after many iterations I might add) in clock-qcom is clean, simple, and easy to understand; though I do agree that U-Boot's DM is definitely hitting the limit of what complexity it can handle.
Well I would like to tidy this up in DM, so let's figure out which option makes the most sense...once I understand what you mean by 'Linux model' above.
I mean where a single struct device can be multiple classes, so you create a device and then create a reset controller and associate the device with it.
It's the bit about having one driver manually creating devices that I very much want to avoid. If the devicetree really does (fully) describe hardware, it shouldn't be necessary.
For sure
I would honestly be much more interested in seeing early init get cleaned up, OF_LIVE becoming the default, and the ofnode abstraction going away.
So far as I can tell, you are always going to have a flat tree, even if only before relocation or in SPL. How would we get around that? Also, what don't you like about ofnode?
Yeah, you start with a flat tree and then should build a live one as early as possible.
The problem with ofnode is that it is more or less restricted to the subset of operations fdt supports. It lacks some of the fancier features of a live tree. It also means that every use of the of_* API has to be preceded by a check that we're actually running the live tree, otherwise some kind of hard bail.
Indeed. In principle, ofnode could support anything, but the cost might be high when implemented in the flat tree. People do add new functions from time to time.
Would be great to use of_* API by default, for capable platforms anyway.
Given that Qualcomm is only using U-Boot as a second-stage loader so far, (please, not for long!!) everything looks quite different. But most platforms use U-Boot from SoC-boot-ROM handoff, so the constraints are different (tighter).
For U-Boot SPL yeah I expect the constraints to be different. We're getting a bit closer to the metal on the rb3gen2 (now it's just bootROM -> SBL1 -> (tz -> hyp) -> U-Boot, without going through edk2 first haha).
You should teach rpi to do that too! At the moment I think it boots edk2 before U-Boot.
I hope we'll get to do U-Boot SPL on some Qualcomm platforms eventually.
Anyway, certainly OF_LIVE being the default would be good. I have often wondered if we can (at build-time) convert the devicetree into a 'live' version, where the pointers are replaced by integers, such that the early U-Boot code can easily compute the pointer value for each node. It should make the unflattening much faster. For pre-relocation and SPL, since we know the load address, we can (I am pretty sure) have Binman put a full, live tree in the image and avoid the unflattening code at all. Relocating a livetree is fairly easy too.
This might be a nice optimisation? I think it would be acceptable to just read the memory layout -> enable the MMU -> build the live tree. This would probably be fast enough.
And of course, the great thing we have on Qualcomm platforms is that we can run the same U-Boot binary on every sdm845 and newer platform.
That's nice, and a good demo of devicetree done right.
If I don't hear any strong preference from you I am likely to have a go (in the fullness of time) at implementing the easiest DM-based solution to get rid of that binding function you have. We can iterate on it in the review in any case.
That makes sense. I assume in the mean time it's ok for me to take this patch?
Kind regards,
Regards, Simon