Re: [PATCH v3] dt-bindings: riscv: deprecate riscv,isa

[Rob Herring <robh@kernel.org>]

On Mon, 26 Jun 2023 11:10:46 +0100, Conor Dooley wrote:
> intro
> =====
> 
> When the RISC-V dt-bindings were accepted upstream in Linux, the base
> ISA etc had yet to be ratified. By the ratification of the base ISA,
> incompatible changes had snuck into the specifications - for example the
> Zicsr and Zifencei extensions were spun out of the base ISA.
> 
> Fast forward to today, and the reason for this patch.
> Currently the riscv,isa dt property permits only a specific subset of
> the ISA string - in particular it excludes version numbering.
> With the current constraints, it is not possible to discern whether
> "rv64i" means that the hart supports the fence.i instruction, for
> example.
> Future systems may choose to implement their own instruction fencing,
> perhaps using a vendor extension, or they may not implement the optional
> counter extensions. Software needs a way to determine this.
> 
> versioning schemes
> ==================
> 
> "Use the extension versions that are described in the ISA manual" you
> may say, and it's not like this has not been considered.
> Firstly, software that parses the riscv,isa property at runtime will
> need to contain a lookup table of some sort that maps arbitrary versions
> to versions it understands. There is not a consistent application of
> version number applied to extensions, with a higgledy-piggledy
> collection of tags, "bare" and versioned documents awaiting the reader
> on the "recently ratified extensions" page:
> https://wiki.riscv.org/display/HOME/Recently+Ratified+Extensions
> 
> 	As an aside, and this is reflected in the patch too, since many
> 	extensions have yet to appear in a release of the ISA specs,
> 	they are defined by commits in their respective "working draft"
> 	repositories.
> 
> Secondly, there is an issue of backwards compatibility, whereby allowing
> numbers in the ISA string, some parsers may be broken. This would
> require an additional property to be created to even use the versions in
> this manner.
> 
> ~boolean properties~ string array property
> ==========================================
> 
> If a new property is needed, the whole approach may as well be looked at
> from the bottom up. A string with limited character choices etc is
> hardly the best approach for communicating extension information to
> software.
> 
> Switching to using properties that are defined on a per extension basis,
> allows us to define explicit meanings for the DT representation of each
> extension - rather than the current situation where different operating
> systems or other bits of software may impart different meanings to
> characters in the string.
> Clearly the best source of meanings is the specifications themselves,
> this just provides us the ability to choose at what point in time the
> meaning is set. If an extension changes incompatibility in the future,
> a new property will be required.
> 
> Off-list, some of the RVI folks have committed to shoring up the wording
> in either the ISA specifications, the riscv-isa-manual or
> so that in the future, modifications to and additions or removals of
> features will require a new extension. Codifying that assertion
> somewhere would make it quite unlikely that compatibility would be
> broken, but we have the tools required to deal with it, if & when it
> crops up.
> It is in our collective interest, as consumers of extension meanings, to
> define a scheme that enforces compatibility.
> 
> The use of individual properties, rather than elements in a single
> string, will also permit validation that the properties have a meaning,
> as well as potentially reject mutually exclusive combinations, or
> enforce dependencies between extensions. That would not have be possible
> with the current dt-schema infrastructure for arbitrary strings, as we
> would need to add a riscv,isa parser to dt-validate!
> That's not implemented in this patch, but rather left as future work (for
> the brave, or the foolish).
> 
> acpi
> ====
> 
> The current ACPI ECR is based on having a single ISA string unfortunately,
> but ideally ACPI will move to another method, perhaps GUIDs, that give
> explicit meaning to extensions.
> 
> parser simplicity
> =================
> 
> Many systems that parse DT at runtime already implement an function that
> can check for the presence of a string in an array of string, as it is
> similar to the process for parsing a list of compatible strings, so a
> bunch of new, custom, DT parsing should not be needed.
> Getting rid of "riscv,isa" parsing would be a nice simplification, but
> unfortunately for backwards compatibility with old dtbs, existing
> parsers may not be removable - which may greatly simplify
> dt parsing code. In Linux, for example, checking for whether a hart
> supports an extension becomes as simple as:
> 	of_property_match_string(node, "riscv,isa-extensions", "zicbom")
> 
> vendor extensions
> =================
> 
> Compared to riscv,isa, this proposed scheme promotes vendor extensions,
> oft touted as the strength of RISC-V, to first-class citizens.
> At present, extensions are defined as meaning what the RISC-V ISA
> specifications say they do. There is no realistic way of using that
> interface to provide cross-platform definitions for what vendor
> extensions mean. Vendor extensions may also have even less consistency
> than RVI do in terms of versioning, or no care about backwards
> compatibility.
> The new property allows us to assign explicit meanings on a per vendor
> extension basis, backed up by a description of their meanings.
> 
> fin
> ===
> 
> Create a new file to store the extension meanings and a new
> riscv,isa-base property to replace the aspect of riscv,isa that is
> not represented by the new property - the base ISA implemented by a hart.
> 
> As a starting point, add properties for extensions currently used in
> Linux.
> 
> Finally, mark riscv,isa as deprecated, as removing support for it in
> existing programs would be an ABI break.
> 
> CC: Palmer Dabbelt <palmer@dabbelt.com>
> CC: Paul Walmsley <paul.walmsley@sifive.com>
> CC: Rob Herring <robh+dt@kernel.org>
> CC: Krzysztof Kozlowski <krzysztof.kozlowski+dt@linaro.org>
> CC: Alistair Francis <alistair.francis@wdc.com>
> CC: Andrew Jones <ajones@ventanamicro.com>
> CC: Anup Patel <apatel@ventanamicro.com>
> CC: Atish Patra <atishp@atishpatra.org>
> CC: Jessica Clarke <jrtc27@jrtc27.com>
> CC: Rick Chen <rick@andestech.com>
> CC: Leo <ycliang@andestech.com>
> CC: Oleksii <oleksii.kurochko@gmail.com>
> CC: linux-riscv@lists.infradead.org
> CC: qemu-riscv@nongnu.org
> CC: u-boot@lists.denx.de
> CC: devicetree@vger.kernel.org
> CC: linux-kernel@vger.kernel.org
> Reviewed-by: Palmer Dabbelt <palmer@rivosinc.com>
> Acked-by: Palmer Dabbelt <palmer@rivosinc.com>
> Signed-off-by: Conor Dooley <conor.dooley@microchip.com>
> ---
> Changes in v3:
> - Per Rob's suggestion, switch to an array of strings. Cuts down on the
>   size, compared to booleans. It has a standard mechanism for parsing
>   (you need to parse arrays of strings for compatibles). It still allows
>   for having a limited set of explicitly defined properties - so the
>   advantages over a free-form string still apply.
> - Pick up Palmer's Ack and Review (although I expect that he will be the
>   one to apply this).
> ---
>  .../devicetree/bindings/riscv/cpus.yaml       |  43 ++-
>  .../devicetree/bindings/riscv/extensions.yaml | 245 ++++++++++++++++++
>  2 files changed, 265 insertions(+), 23 deletions(-)
>  create mode 100644 Documentation/devicetree/bindings/riscv/extensions.yaml
> 

Reviewed-by: Rob Herring <robh@kernel.org>