Re: [PATCH v3 1/5] arch: riscv: add support for building DTB files from DT source data

[Palmer Dabbelt]

On Thu, 06 Jun 2019 22:12:05 PDT (-0700), Paul Walmsley wrote:
> On Tue, 4 Jun 2019, Loys Ollivier wrote:
>
> > Always build it ?
> > Any particular reason to drop ARCH_SIFIVE ?
>
> Palmer had some reservations about it, so I dropped it for now.  But then
> as I was thinking about it, I remembered that I also had some reservations
> about it, years ago: that everyone should use CONFIG_SOC_* for this,
> rather than CONFIG_ARCH.  CONFIG_ARCH_* seems better reserved for
> CPU architectures.

Specifically my worry is that "ARCH_SIFIVE" makes it sound like we're adding
SiFive-specific architecture features, and we've been trying really hard to
make sure that the various bits of core software avoid boing vendor specific.
We've had suggestions of adding vendor-specific instructions to the Linux port
with those instructions being conditionally compiled under ARCH_$VENDOR, but
I'd rejected that under the "no vendor-specific stuff" argument.  As such it
doesn't seem fair to go add in an ARCH_SIFIVE for our vendor-specific stuff.

The SOC stuff will, of course, be vendor specific.  In this idealized world
SiFive's SOC support has nothing to do with RISC-V, but of course all of
SiFive's SOCs are RISC-V based so the separation is a bit of pedantry.  That
said, in this case I think getting the name right does make it slightly easier
to espouse this "one kernel can run on all RISC-V systems" philosophy.
Balancing the SiFive and RISC-V stuff can be a bit tricky, which is why I am
sometimes a bit pedantic about these sorts of things.

> If you agree, would you like to send a followup series, based on the DT
> patches, to make the SiFive DT file builds depend on CONFIG_SOC_* instead?

I'd be happy with something like that.  We'd also talked about this selecting
all the SiFive platform drivers.  It should, of course, be possible to select
multiple SOC vendors in a single kernel -- we don't have any other real
hardware right now, but maybe some sort of "CONFIG_SOC_RISCV_VIRT" would be a
good proof of concept?

> Thanks for the comment,
>
> - Paul