Re: [PATCH 04/10] riscv: Add BUILTIN_DTB support

[Sean Anderson <seanga2@gmail.com>]

On 3/5/20 3:18 AM, Atish Patra wrote:
> On Wed, Mar 4, 2020 at 9:14 PM Damien Le Moal <Damien.LeMoal@wdc.com> wrote:
>>
>> On 2020/03/05 13:58, Anup Patel wrote:
>>>
>>>
>>>> -----Original Message-----
>>>> From: Palmer Dabbelt <palmer@dabbelt.com>
>>>> Sent: 05 March 2020 00:59
>>>> To: Damien Le Moal <Damien.LeMoal@wdc.com>
>>>> Cc: linux-riscv@lists.infradead.org; Paul Walmsley
>>>> <paul.walmsley@sifive.com>; Anup Patel <Anup.Patel@wdc.com>
>>>> Subject: Re: [PATCH 04/10] riscv: Add BUILTIN_DTB support
>>>>
>>>> On Wed, 12 Feb 2020 02:34:26 PST (-0800), Damien Le Moal wrote:
>>>>> Enable a kernel builtin dtb for boards not capable of providing a
>>>>> device tree to the kernel.
>>>>
>>>> I'd prefer if we picked a mechanism that allows a single kernel binary to be
>>>> run on multiple systems.  I think there's two use cases here:
>>>
>>> I strongly support "single kernel binary for multiple systems" but it's for
>>> different purpose here.
>>>
>>>>
>>>> * Bootloaders that provide no DTB at all.
>>>> * Bootloaders that provied a DTB that, for some reason, isn't usable.
>>
>> Sure, but as Anup mentions below, the current use case it to not even use any
>> bootloader at all for the K210 since that brings no value at all (in my
>> opinion). More on this below.
>>
>>>>
>>>> Given those constraints, could we do something similar to the early fixups?
>>>> I'm thinking we could build a mapping between a hardware identifier and a
>>>> DTB, then look up the DTB we want to use.  Users that want a kernel that
>>>> only runs on a single device can do so by configuring only a single DTB, users
>>>> that want a more portable kernel can select a bunch -- that's essentially the
>>>> same as how we're treating everything else (for example, the
>>>> CONFIG_SOC_* stuff).
>>>
>>> There is no bootloader on Kendryte K210. The Linux RISC-V NOMMU kernel
>>> boots directly. The BUILTIN_DTB is only applicable to cases where there is
>>> no bootloader before kernel.
>>>
>>> The Linux RISC-V NOMMU will tend be used in cases where:
>>> 1. There is no bootloader and kernel boots directly hence we need
>>> builtin DTB feature.
>>> 2. There is very less RAM so we will have to build kernel specific to
>>> a particular platform with bare minimum drivers. Due to this, we will
>>> have separate defconfig for NOMMU platforms.
>>>
>>> I think point1 can be tackled if we enforce having bootloader (such as U-Boot)
>>> for NOMMU systems and drop this patch.
>>
>> But that would go against point 2 as that will use more memory... By "drop this
>> patch", may be you meant to say "not use this config option" ?
>>
>>> For point2 above, we don't have much alternatives other than reducing
>>> kernel binary size by disabling unwanted drivers.
>>
>> And not using a boot loader. Sean got U-boot working with Kendryte, so it is not
>> that we cannot make it work. It is only that it may be less optimal due to the
>> memory used by the boot loader itself. Unless we can recover it if the kernel
>> relocate itself over it ? Surely doable, but it does sound to me like an
>> overkill for this particular use case, i.e. a tiny, hyper-embedded board where
>> running Linux is probably not the best choice in the first place, at least when
>> looking at real applications. The story is different for "hobbyist" level. My
>> on-going 6 DoF robotic arm project controlled with Linux on K210 is a valid use
>> case after all :)
>>
> 
> Just a thought: How about keeping the loader out of kernel as a
> separate project as a tinyloader ?
> That tinyloader project can host the DTB and pass it to kernel in a1
> register. This tinyloader can be used for
> any M-mode only platforms with memory constraints.  If platform has
> sufficient memory and supports U-boot, users can use that as well.
> That will allow single kernel image to boot all the platforms and we
> can mandate one booting protocol for all platforms.
> Otherwise, we have to specify different booting protocol for
> M-Mode/NoMMU linux and S-mode Linux.
> In future, there may be more platforms with Builtin DTB request as well.

Couldn't this sort of thing be done by SBI? OpenSBI currently has a port
for the K210. The only issue with SBI in general is that there is no way
to set the VM mode, since it is stored in mstatus and not satp on older
priv spec versions. There used to be a proposal related to this, but
they chose to change the location of the VM mode rather than have SBI or
some other bootloader set it. I think one of the ideas was for SBI to
set the mode based off the mmu-type property.

--Sean