Message-ID: <712ffab1-5b25-f8bb-79a0-e2c0693e4c35@starfivetech.com>
Date:   Fri, 2 Jun 2023 17:42:42 +0800
From:   Xingyu Wu <xingyu.wu@...rfivetech.com>
To:     Conor Dooley <conor.dooley@...rochip.com>,
        Torsten Duwe <duwe@....de>
CC:     <linux-riscv@...ts.infradead.org>, <devicetree@...r.kernel.org>,
        <yanhong.wang@...rfivetech.com>,
        Michael Turquette <mturquette@...libre.com>,
        Stephen Boyd <sboyd@...nel.org>,
        Krzysztof Kozlowski <krzysztof.kozlowski+dt@...aro.org>,
        Philipp Zabel <p.zabel@...gutronix.de>,
        Conor Dooley <conor@...nel.org>,
        Emil Renner Berthing <kernel@...il.dk>,
        "Rob Herring" <robh+dt@...nel.org>,
        Paul Walmsley <paul.walmsley@...ive.com>,
        Palmer Dabbelt <palmer@...belt.com>,
        Albert Ou <aou@...s.berkeley.edu>,
        "Hal Feng" <hal.feng@...rfivetech.com>,
        William Qiu <william.qiu@...rfivetech.com>,
        <linux-kernel@...r.kernel.org>, <linux-clk@...r.kernel.org>,
        <u-boot@...ts.denx.de>
Subject: Re: [PATCH v4 1/7] dt-bindings: clock: Add StarFive JH7110 PLL clock
 generator

On 2023/5/26 20:23, Conor Dooley wrote:
> On Fri, May 26, 2023 at 09:34:32AM +0200, Torsten Duwe wrote:
>> On Wed, 24 May 2023 11:19:48 +0100
>> Conor Dooley <conor.dooley@...rochip.com> wrote:
>> 
>> > On Wed, May 24, 2023 at 05:00:02PM +0800, Xingyu Wu wrote:
>> > > On 2023/5/23 19:28, Conor Dooley wrote:
>> > > > On Tue, May 23, 2023 at 01:10:06PM +0200, Torsten Duwe wrote:
>> > > >> On Tue, 23 May 2023 09:28:39 +0100
>> > > >> Conor Dooley <conor.dooley@...rochip.com> wrote:
>> > > >> 
>> > > >> > On Tue, May 23, 2023 at 10:56:43AM +0800, Xingyu Wu wrote:
>> > > >> > > On 2023/5/19 22:16, Conor Dooley wrote:
>> > > >> > > > On Fri, May 19, 2023 at 03:57:33PM +0200, Torsten Duwe
>> > > >> > > > wrote:
>> > > >> > > >> On Fri, May 12, 2023 at 10:20:30AM +0800, Xingyu Wu wrote:
>> [...]
>> 
>> > > >> > > Because PLL driver is separated from SYSCRG drivers in
>> > > >> > > Linux, the numbering starts from 0. But in Uboot, the PLL
>> > > >> > > driver is included in the SYSCRG driver, and the number
>> > > >> > > follows the SYSCRG.
>> > > >> > 
>> > > >> > Unfortunately, how you choose to construct your drivers has
>> > > >> > nothing to do with this.
>> 
>> Exactly. As I wrote (quote below), the PLLx frequencies are controlled
>> by the I/O block SYS_SYSCON (starting there at offset 0x18), according
>> to the public datasheets. All(?) other clocks are derived from those in
>> the *_CRG units. That *is* the hardware to be described, in *the* (one
>> and only!) DT. U-Boot, and any OS, are free to reorganise their driver
>> framework around that, but the hardware description is quite clear.
> 
> The dt-binding that is in this series specifies that the pll clock
> controller is a child of the syscon:
> https://lore.kernel.org/linux-riscv/20230512022036.97987-1-xingyu.wu@starfivetech.com/T/#Z2e.:..:20230512022036.97987-6-xingyu.wu::40starfivetech.com:1soc:starfive:starfive::2cjh7110-syscon.yaml
> 
> That seems correct to me & U-Boot's devicetree is not compliant.
> 
>> > > >> > These defines/numbers appear in the dts and are part of the DT
>> > > >> > ABI. The same dts is supposed to work for Linux & U-Boot.
>> > > >> 
>> > > >> The JH7110 has 6 blocks of 64k iomem in that functional area:
>> > > >> {SYS,STG,AON} x {CRG,SYSCON}. None of these has 190 clocks.
>> > > >> The good news: the current DTS, as proposed here and in U-Boot
>> > > >> master, provides nodes for all 6 entities. The bad news is that
>> > > >> the clock assignments to those nodes and their numbering is
>> > > >> messed up.
>> > > >> 
>> > > >> AFAICT PLL{0,1,2} _are_ generated in SYS_SYSCON and thus U-Boot
>> > > >> gets it wrong, in addition to the erroneous DTS.
>> > > > 
>> > > > The numbers are kinda hocus-pocus anyway, they are just made up
>> > > > since the clock numbering usually isn't something with a nice TRM
>> > > > to go and reference (unlike interrupts which usually are
>> > > > documented in that way). It is very helpful to make them aligned
>> > > > some register/bit positions or, but that is not required.
>> > > > IOW U-Boot is not wrong per se to use 190 instead of 0, but it is
>> > > > wrong to have different numbers in both places.
>> 
>> U-Boot reuses the Common Clock Framework from Linux, and I'm not sure
>> whether the clock IDs need to be unique in order for the appropriate
>> clock to be found.
> 
> Unique within the clock controller, otherwise it is impossible to tell
> the difference between <&cctrl 1> and <&cctrl 1> apart! (The same
> follows even with increased #clock-cells, something must be unique).
> That's besides the point of this particular issue though.
> 
>> But that would be the only restriction, if it
>> applies. Even then, each driver could register a clock with its own,
>> arbitrarily chosen base offset with the CCF, so each CRG unit could
>> still have its own clocks enumerated starting with 0 in the DTB.
>> 
>> > > > It sounds like you're saying that (and I have not looked) the
>> > > > U-Boot dts actually has structural difference w.r.t. what
>> > > > provides which clock? If so, that'll need to be fixed
>> > > > independently of the numbering problem.
>> 
>> > > 
>> > > Oh, unfortunately, the 7110 can not support to mix the uboot dtb
>> > > and linux dtb up.
>> > 
>> > What does "cannot support" mean? It's normal and desirable for the
>> 
>> IMHO "desirable" is too weak.
> 
> Yeah, agreed. I just don't like being prescriptive about what happens in
> projects that I do not maintain things for I guess.
> 
>> > same dtb to be usable for both. The Linux kernel's dt-bindings are
>> > used for multiple projects, not just Linux - it'd be silly for
>> > U-Boot, FreeBSD etc etc to go off and each have their open set of
>> > (incompatible) bindings.
>> > 
>> > > If boot the Linux and should use the linux dtb instead of the uboot
>> > > dtb. Because all clock ids and reset ids in Linux and Uboot are
>> > > different include PLL, and some modules can work in Linux but not
>> > > in uboot.
>> [...]
>> > 
>> > > I suggest to boot Linux with its own linux dtb.
>> 
>> This is a fragile band-aid, to be used only as a last resort. It
>> creates more problems than it solves. Your DTB will then match your
>> kernel, but whether it describes the actual hardware is a game of
>> chance. Doesn't the VisionFive2 have an RPi connector... ?
>> 
>> One of the IMO few valid use cases of adding a DTB to the kernel
>> at boot is OpenWRT, when you build an OS Image for a particular piece
>> of hardware you have at hand.
>> 
>> > I suggest to make sure that you can use the same dtb for both.
>> 
>> Interestingly enough, U-Boot already has the PLL driver in a separate
>> file. I have a half-baked patch here that moves the sys_syscon DT
>> matching into that file...
> 
> If you have patches that fix the devicetree & drivers in U-Boot, please
> post them. I don't really care at all which set of arbitrary numbers are
> chosen (as long as there is one and one only) but it looks like U-Boot's
> devicetree has an incorrect description of the clock controllers.
> 

Thank you for your suggestions. I will discuss with my partners how to solve this problem.

Best regards,
Xingyu Wu

Powered by blists - more mailing lists