Re: [PATCH v2 3/3] arm64: dts: qcom: apq8016: Add Schneider HMIBSC board DTS

[Sumit Garg]

On Fri, 15 Mar 2024 at 20:01, Stephan Gerhold <stephan@xxxxxxxxxxx> wrote:
>
> On Fri, Mar 15, 2024 at 03:01:27PM +0530, Sumit Garg wrote:
> > On Thu, 14 Mar 2024 at 21:07, Caleb Connolly <caleb.connolly@xxxxxxxxxx> wrote:
> > > On 14/03/2024 15:20, Konrad Dybcio wrote:
> > > > On 3/14/24 14:50, Sumit Garg wrote:
> > > >> On Thu, 14 Mar 2024 at 18:54, Stephan Gerhold <stephan@xxxxxxxxxxx>
> > > >> wrote:
> > > >>>
> > > >>> On Thu, Mar 14, 2024 at 05:26:27PM +0530, Sumit Garg wrote:
> > > >>>> On Thu, 14 Mar 2024 at 16:13, Stephan Gerhold <stephan@xxxxxxxxxxx>
> > > >>>> wrote:
> > > >>>>> On Thu, Mar 14, 2024 at 03:02:31PM +0530, Sumit Garg wrote:
> > > >>>>>> On Thu, 14 Mar 2024 at 14:48, Konrad Dybcio
> > > >>>>>> <konrad.dybcio@xxxxxxxxxx> wrote:
> > > >>>>>>> On 3/14/24 10:04, Sumit Garg wrote:
> > > >>>>>>>> On Wed, 13 Mar 2024 at 18:34, Konrad Dybcio
> > > >>>>>>>> <konrad.dybcio@xxxxxxxxxx> wrote:
> > > >>>>>>>>> On 3/13/24 13:30, Sumit Garg wrote:
> > > >>>>>>>>>> Add Schneider Electric HMIBSC board DTS. The HMIBSC board is
> > > >>>>>>>>>> an IIoT Edge
> > > >>>>>>>>>> Box Core board based on the Qualcomm APQ8016E SoC.
> > > >>>>>>>>>>
> > > >>>>>>>>>> Support for Schneider Electric HMIBSC. Features:
> > > >>>>>>>>>> - Qualcomm Snapdragon 410C SoC - APQ8016 (4xCortex A53, Adreno
> > > >>>>>>>>>> 306)
> > > >>>>>>>>>> - 1GiB RAM
> > > >>>>>>>>>> - 8GiB eMMC, SD slot
> > > >>>>>>>>>> - WiFi and Bluetooth
> > > >>>>>>>>>> - 2x Host, 1x Device USB port
> > > >>>>>>>>>> - HDMI
> > > >>>>>>>>>> - Discrete TPM2 chip over SPI
> > > >>>>>>>>>> - USB ethernet adaptors (soldered)
> > > >>>>>>>>>>
> > > >>>>>>>>>> Co-developed-by: Jagdish Gediya <jagdish.gediya@xxxxxxxxxx>
> > > >>>>>>>>>> Signed-off-by: Jagdish Gediya <jagdish.gediya@xxxxxxxxxx>
> > > >>>>>>>>>> Signed-off-by: Sumit Garg <sumit.garg@xxxxxxxxxx>
> > > >>>>>>>>>> ---
> > > >>>>>>>>>
> > > >>>>>>>>> [...]
> > > >>>>>>>>>
> > > >>>>>>>>>> +     memory@80000000 {
> > > >>>>>>>>>> +             reg = <0 0x80000000 0 0x40000000>;
> > > >>>>>>>>>> +     };
> > > >>>>>>>>>
> > > >>>>>>>>> I'm not sure the entirety of DRAM is accessible..
> > > >>>>>>>>>
> > > >>>>>>>>> This override should be unnecessary, as bootloaders generally
> > > >>>>>>>>> update
> > > >>>>>>>>> the size field anyway.
> > > >>>>>>>>
> > > >>>>>>>> On this board, U-Boot is used as the first stage bootloader
> > > >>>>>>>> (replacing
> > > >>>>>>>> Little Kernel (LK), thanks to Stephan's work). And U-Boot consumes
> > > >>>>>>>> memory range from DT as Linux does but doesn't require any
> > > >>>>>>>> memory to
> > > >>>>>>>> be reserved for U-Boot itself. So apart from reserved memory nodes
> > > >>>>>>>> explicitly described in DT all the other DRAM regions are
> > > >>>>>>>> accessible.
> > > >>>>>>>
> > > >>>>>>> Still, u-boot has code to fetch the size dynamically, no?
> > > >>>>>>>
> > > >>>>>>
> > > >>>>>> No U-Boot being the first stage bootloader fetches size from DT which
> > > >>>>>> is bundled into U-Boot binary.
> > > >>>>>>
> > > >>>>>
> > > >>>>> Back when I added support for using U-Boot as first stage
> > > >>>>> bootloader on
> > > >>>>> DB410c the way it worked is that U-Boot used a fixed amount of DRAM
> > > >>>>> (originally 968 MiB, later 1 GiB since I fixed this in commit
> > > >>>>> 1d667227ea51 ("board: dragonboard410c: Fix PHYS_SDRAM_1_SIZE") [1]).
> > > >>>>> When booting Linux, the Linux DT was dynamically patched with the
> > > >>>>> right
> > > >>>>> amount of DRAM (obtained from SMEM). So if you had e.g. a Geniatech
> > > >>>>> DB4
> > > >>>>> board with 2 GiB DRAM, U-Boot was only using 1 GiB of DRAM, but Linux
> > > >>>>> later got the full 2 GiB patched into its DTB.
> > > >>>>>
> > > >>>>> I didn't have much time for testing U-Boot myself lately but a quick
> > > >>>>> look at the recent changes suggest that Caleb accidentally removed
> > > >>>>> that
> > > >>>>> functionality in the recent cleanup. Specifically, the SMEM-based DRAM
> > > >>>>> size detection was removed in commit 14868845db54 ("board:
> > > >>>>> dragonboard410c: import board code from mach-snapdragon" [2]), the
> > > >>>>> msm_fixup_memory() function does not seem to exist anymore now. :')
> > > >>>>
> > > >>>> Ah now I see the reasoning for that particular piece of code. Is SMEM
> > > >>>> based approach the standardized way used by early stage boot-loaders
> > > >>>> on other Qcom SoCs too?
> > > >>>>
> > > >>>
> > > >>> It is definitely used on all the SoCs that were deployed with LK. I am
> > > >>> not entirely sure about the newer ABL/UEFI-based ones. A quick look at
> > > >>> the ABL source code suggests it is abstracted through an EFI protocol
> > > >>> there (so we cannot see where the information comes from with just the
> > > >>> open-source code). However, in my experience SMEM data structures are
> > > >>> usually kept quite stable (or properly versioned), so it is quite likely
> > > >>> that we could use this approach for all Qualcomm SoCs.
> > > >>>
> > > >>
> > > >> If the SoCs which support this standardized way to dynamic discover
> > > >> DRAM size via SMEM then why do we need to rely on DT at all for those
> > > >> SoCs? Can't U-Boot and Linux have the same driver to fetch DRAM size
> > > >> via SMEM? I am not sure if it's an appropriate way for U-Boot to fixup
> > > >> DT when that information can be discovered dynamically.
> > >
> > > [...]
> > >
> > > The reason I decided to hardcode this in DT for U-Boot is because SMEM
> > > currently relies on the driver model and isn't available early enough.
> > >
> > > Also admittedly I just wasn't that familiar with the U-Boot codebase. I
> > > just wanted to avoid hardcoding this in C code, and given that this was
> > > already supported for the "chainloading from ABL" usecase, just
> > > hardcoding the values was the obvious solution.
> > >
> > > I would definitely be open to revisiting this in U-Boot, having an SMEM
> > > framework that we could use without the driver model which would just
> > > take a base address and then let us interact with SMEM and populate the
> > > dram bank data would be a good improvement, and would let us avoid
> > > hardcoding the memory layout in DT. We'd just need to manually find the
> > > SMEM base address in the FDT as part of "dram_init_banksize()" and
> > > retrieve the data there.
> >
> > These are the similar problems Linux has to deal with too but on Qcom
> > platforms it is rather offloaded to bootloaders to rather implement
> > this. It leads to custom DT modification or board code in U-Boot which
> > is hard to maintain. If we want to implement it properly then
> > corresponding bindings should be upstreamed too regarding how DRAM
> > range is to be discovered via SMEM.
> >
> > >
> > > That all being said, I don't see any reason not to define the memory
> > > layout in DT, it's a hardware feature, DT describes the hardware. The
> > > whole "bootloader will fill this in" implies that the bootloader isn't
> > > also using DT as the source of truth, but now with U-Boot it actually
> > > is, so it's all the more important that DT be accurate ;P
> >
> > I agree DT should be accurate and not a fan of DT fixups. However,
> > when it comes to some hardware configuration being discoverable then
> > IMHO DT isn't the appropriate place for that. For the time being I am
> > fine with the DRAM range to be specified in DT.
> >
> > > >
> > > > You're mixing two things. Linux expects a devicetree where
> > > > /memory/reg[size]
> > > > is valid. Such driver should indeed be (re)implemented in u-boot to provide
> > > > this information.
> >
> > No, I don't think so. We should rather start thinking about the
> > overall stack rather than just being Linux kernel centric. Do you have
> > a generic solution for U-Boot regarding how this should be
> > implemented?
> >
>
> Detecting available memory via /memory in the DT

I would rather call that as "Hardcoding available memory via ...". The
basic DT description [1] says:

"A DTSpec-compliant devicetree describes device information in a
system that cannot necessarily be dynamically detected by a client
program."

[1] https://devicetree-specification.readthedocs.io/en/latest/chapter2-devicetree-basics.html#overview

> or other firmware
> services (such as UEFI GetMemoryMap()) is *the* generic solution used
> everywhere, independent of the hardware (e.g. Qualcomm) and the
> operating system (e.g. Linux or Xen).

That's an appropriate alternate example of how available memory
information can be passed to the operating system. But on non-UEFI
systems, bootloaders are just stuck with DT fixups as the operating
system doesn't support an alternate method to retrieve memory
information.

IMHO, whether it's a UEFI system or a non-UEFI system the DT should be
identical without the need for fixups.

>
> It allows us to implement the board/Qualcomm-specific memory detection
> in a single project,

I suppose in your view U-Boot is the only bootloader project out there
but you should at least check out [2].

[2] https://en.wikipedia.org/wiki/Bootloader

> rather than having extra porting overhead for each
> operating system for something as essential as available memory.
>
> If we implement the SMEM-based memory detection in U-Boot (probably in
> dram_init_banksize() as Caleb suggested) everything else will just work
> without any Qualcomm-specific DT patching in U-Boot, and without any
> special code in the operating system:
>
>  - U-Boot automatically updates the /memory node in generic code (see
>    fdt_fixup_memory_banks() call in arch/arm/lib/bootm-fdt.c). If you
>    are using UEFI for booting, U-Boot will provide the same information
>    via GetMemoryMap(). (The DT spec says /memory should be ignored on
>    UEFI systems.)
>
>  - Linux, Xen, and any other operating system can obtain the memory size
>    via the standard method, and do not need any Qualcomm-specific at all
>    (at least for memory detection).
>
> I don't think there is anything Linux kernel centric for the memory
> detection here. Quite the opposite really. :)

The question here is really about the thinking that people still
consider DT as a way to describe information which should rather be
detected dynamically. The DT fixups seems to be an outcome of this
thinking.

On a non-UEFI system, if people are instead looking for a generic way
to pass information then we should be able to consider firmware
handoff data structures [3] too. However, if we can just have a
reference in DT /memory node to a particular data structure then we
should just be fine with the corresponding driver extracting the
memory information.

[3] https://github.com/FirmwareHandoff/firmware_handoff

-Sumit

>
> Thanks,
> Stephan