Re: [PATCH V4 1/2] dt-bindings: riscv: add MMU Standard Extensions support for Svpbmt

[Heiko Stübner]

Am Mittwoch, 1. Dezember 2021, 09:15:18 CET schrieb Atish Patra:
> On Tue, Nov 30, 2021 at 7:06 PM Tsukasa OI <research_trasio@xxxxxxxxxxxx> wrote:
> >
> > On 2021/12/01 10:21, Atish Patra wrote:
> > > On Tue, Nov 30, 2021 at 8:13 AM Jessica Clarke <jrtc27@xxxxxxxxxx> wrote:
> > >>
> > >> On 30 Nov 2021, at 15:01, Philipp Tomsich <philipp.tomsich@xxxxxxxx> wrote:
> > >>>
> > >>> We did touch on this in our coordination call a few weeks ago: the
> > >>> grouping under mmu and the bool-entries were chosen because of their
> > >>> similarity to other extensions (i.e. for Zb[abcs] there could/should
> > >>> be a bool-entry under each cpu-node — for some Zv* entries a subnode
> > >>> might be needed with further parameters).
> > >>>
> > >>> The string-based approach (as in the originally proposed "mmu-type=")
> > >>> would like not scale with the proliferation of small & modular
> > >>> extensions.
> > >>
> > >> I don’t see why the Sv* extensions need to be under an mmu node then,
> > >> unless the intent is that every extension be grouped under a sub-node
> > >> (which doesn’t seem viable due to extensions like Zbk*, unless you
> > >> group by Ss, Sv and Z)?
> > >>
> > >
> > > It shouldn't be. All the ISA extensions (i.e. standard, supervisor & hypervisor)
> > > with prefix S,Z,H should be kept separate in a separate node for easy
> > > parsing.
> >
> > "Easy parsing" is not quite convincing.
> 
> The device tree need to carry a very long "riscv,isa" string. The
> parser need to parse
> that string in memory as well.
> 
> >
> > There's a reason other than that I made RFC PATCH to parse
> > multi-letter extensions:
> >
> > v1: <http://lists.infradead.org/pipermail/linux-riscv/2021-November/010252.html>
> > v2: <http://lists.infradead.org/pipermail/linux-riscv/2021-November/010350.html>
> >
> 
> It's on my todo list to review the series. I think we can work
> together to propose a better framework for riscv isa extensions.
> 
> > (note: those patches will break RISC-V KVM because of possible ISA
> >        Manual inconsistency and discussion/resolution needed)
> >
> > (...continued below...)
> >
> > >
> > > "riscv,isa" dt property will not scale at all. Just look at the few
> > > extensions that were ratified this year
> > > and Linux kernel needs to support them.
> > >
> > > "Sscofpmf", "Svpbmt", "Zicbom"
> > >
> > >> Also, what is going to happen to the current riscv,isa? Will that
> > >> continue to exist and duplicate the info, or will kernels be required
> > >> to reconstruct the string themselves if they want to display it to
> > >> users?
> > >>
> 
> Sorry. I missed this question earlier. See my answer below.
> 
> > >
> > > This is my personal preference:
> > > riscv,isa will continue to base Standard ISA extensions that have
> > > single letter extensions.
> > >
> > > This new DT node will encode all the non-single letter extensions.
> > > I am not sure if it should include some provisions for custom
> > > extensions starting with X because
> > > that will be platform specific.
> > >
> > > Again, this is just my personal preference. I will try to send a patch
> > > soon so that we can initiate a broader
> > > discussion of the scheme and agree/disagree on something.
> >
> > For supervisor-only extensions like "Svpbmt", new DT node would be a
> > reasonable solution (and I would not directly object about that node).
> >
> > However, there's many multi-letter extensions that are useful for
> > user mode.  Because "riscv,isa" is exposed via sysfs and procfs
> > (/proc/cpuinfo), it can be really helpful to have multi-letter
> 
> Irrespective of the method chosen to parse the device tree in kernel,
> we need to provide the extension information to the userspace.
> 
> This is what I have in mind. An individual row with comma separated
> extension names for each type of extensions (Ss, Sv, Sh)
> after the base extension (rv64imafdc) in /proc/cpuinfo output. I am
> open to other ideas as well.
>
> isa   rv64imafdc
> isa-ext-Sv Svpbmt
> isa-ext-Ss Sscofpmf
> isa-ext-Sh <hypervisor related extensions>
> isa-ext-Z   Zicbom
> 
> We can even explicitly name the extensions after isa-ext. However, it
> may be necessary and too long.

Aren't other architectures just using a flags [x86] or features [arm64]
line in cpuinfo to expose the available additional cpu features
as a space-separated list?

So you could also just do something similar like
isa:   rv64imafdc
isa-ext: Svpbmt Sscofpmf foo bar

That would make a nice compromise between length and readability
by users I guess?


Heiko

> I guess you prefer to directly print the entire "riscv,isa" string in
> "isa" row in /proc/cpuinfo output.
> It is probably okay with the current number of extensions available
> today. However, it will become so long string
> in the future that it has to be broken into multiple lines.
> 
> > extensions.  Also, current version of Spike, a RISC-V ISA Simulator
> > puts all multi-letter extensions in "riscv,isa" and I thought this is
> > intended.
> >
> > My preference:
> > (1) Allow having multi-letter extensions and versions in "riscv,isa"
> > (2) Adding new DT node for supervisor-related extensions would be
> >     reasonable (but I don't strongly agree/disagree).
> >
> > Thanks,
> > Tsukasa
> >
> > >
> > >
> > >
> > >> As a FreeBSD developer I’m obviously not a part of many of these
> > >> discussions, but what the Linux community imposes as the device tree
> > >> bindings has a real impact on us.
> > >>
> > >> Jess
> > >>
> > >>> On Tue, 30 Nov 2021 at 14:59, Jessica Clarke <jrtc27@xxxxxxxxxx> wrote:
> > >>>>
> > >>>> On 30 Nov 2021, at 13:27, Heiko Stübner <heiko@xxxxxxxxx> wrote:
> > >>>>>
> > >>>>> Hi,
> > >>>>>
> > >>>>> Am Dienstag, 30. November 2021, 14:17:41 CET schrieb Jessica Clarke:
> > >>>>>> On 30 Nov 2021, at 12:07, Heiko Stübner <heiko@xxxxxxxxx> wrote:
> > >>>>>>>
> > >>>>>>> Am Montag, 29. November 2021, 13:06:23 CET schrieb Heiko Stübner:
> > >>>>>>>> Am Montag, 29. November 2021, 09:54:39 CET schrieb Heinrich Schuchardt:
> > >>>>>>>>> On 11/29/21 02:40, wefu@xxxxxxxxxx wrote:
> > >>>>>>>>>> From: Wei Fu <wefu@xxxxxxxxxx>
> > >>>>>>>>>>
> > >>>>>>>>>> Previous patch has added svpbmt in arch/riscv and add "riscv,svpmbt"
> > >>>>>>>>>> in the DT mmu node. Update dt-bindings related property here.
> > >>>>>>>>>>
> > >>>>>>>>>> Signed-off-by: Wei Fu <wefu@xxxxxxxxxx>
> > >>>>>>>>>> Co-developed-by: Guo Ren <guoren@xxxxxxxxxx>
> > >>>>>>>>>> Signed-off-by: Guo Ren <guoren@xxxxxxxxxx>
> > >>>>>>>>>> Cc: Anup Patel <anup@xxxxxxxxxxxxxx>
> > >>>>>>>>>> Cc: Palmer Dabbelt <palmer@xxxxxxxxxxx>
> > >>>>>>>>>> Cc: Rob Herring <robh+dt@xxxxxxxxxx>
> > >>>>>>>>>> ---
> > >>>>>>>>>> Documentation/devicetree/bindings/riscv/cpus.yaml | 10 ++++++++++
> > >>>>>>>>>> 1 file changed, 10 insertions(+)
> > >>>>>>>>>>
> > >>>>>>>>>> diff --git a/Documentation/devicetree/bindings/riscv/cpus.yaml b/Documentation/devicetree/bindings/riscv/cpus.yaml
> > >>>>>>>>>> index aa5fb64d57eb..9ff9cbdd8a85 100644
> > >>>>>>>>>> --- a/Documentation/devicetree/bindings/riscv/cpus.yaml
> > >>>>>>>>>> +++ b/Documentation/devicetree/bindings/riscv/cpus.yaml
> > >>>>>>>>>> @@ -63,6 +63,16 @@ properties:
> > >>>>>>>>>>      - riscv,sv48
> > >>>>>>>>>>      - riscv,none
> > >>>>>>>>>>
> > >>>>>>>>>> +  mmu:
> > >>>>>>>>>
> > >>>>>>>>> Shouldn't we keep the items be in alphabetic order, i.e. mmu before
> > >>>>>>>>> mmu-type?
> > >>>>>>>>>
> > >>>>>>>>>> +    description:
> > >>>>>>>>>> +      Describes the CPU's MMU Standard Extensions support.
> > >>>>>>>>>> +      These values originate from the RISC-V Privileged
> > >>>>>>>>>> +      Specification document, available from
> > >>>>>>>>>> +      https://riscv.org/specifications/
> > >>>>>>>>>> +    $ref: '/schemas/types.yaml#/definitions/string'
> > >>>>>>>>>> +    enum:
> > >>>>>>>>>> +      - riscv,svpmbt
> > >>>>>>>>>
> > >>>>>>>>> The privileged specification has multiple MMU related extensions:
> > >>>>>>>>> Svnapot, Svpbmt, Svinval. Shall they all be modeled in this enum?
> > >>>>>>>>
> > >>>>>>>> I remember in some earlier version some way back there was the
> > >>>>>>>> suggestion of using a sub-node instead and then adding boolean
> > >>>>>>>> properties for the supported extensions.
> > >>>>>>>>
> > >>>>>>>> Aka something like
> > >>>>>>>>   mmu {
> > >>>>>>>>           riscv,svpbmt;
> > >>>>>>>>   };
> > >>>>>>>
> > >>>>>>> For the record, I'm talking about the mail from september
> > >>>>>>> https://lore.kernel.org/linux-riscv/CAAeLtUChjjzG+P8yg45GLZMJy5UR2K5RRBoLFVZhtOaZ5pPtEA@xxxxxxxxxxxxxx/
> > >>>>>>>
> > >>>>>>> So having a sub-node would make adding future extensions
> > >>>>>>> way nicer.
> > >>>>>>
> > >>>>>> Svpbmt is just an ISA extension, and should be treated like any other.
> > >>>>>> Let’s not invent two different ways of representing that in the device
> > >>>>>> tree.
> > >>>>>
> > >>>>> Heinrich asked how the other extensions should be handled
> > >>>>> (Svnapot, Svpbmt, Svinval), so what do you suggest to do with these?
> > >>>>
> > >>>> Whatever is done for Zb[abcs], Zk*, Zv*, Zicbo*, etc. There may not be
> > >>>> a concrete plan for that yet, but that means you should speak with the
> > >>>> people involved with such extensions and come up with something
> > >>>> appropriate together.
> > >>>>
> > >>>> Jess
> > >>>>
> > >>
> > >>
> > >> _______________________________________________
> > >> linux-riscv mailing list
> > >> linux-riscv@xxxxxxxxxxxxxxxxxxx
> > >> http://lists.infradead.org/mailman/listinfo/linux-riscv
> > >
> > >
> > >
> > > --
> > > Regards,
> > > Atish
> > >
> > > _______________________________________________
> > > linux-riscv mailing list
> > > linux-riscv@xxxxxxxxxxxxxxxxxxx
> > > http://lists.infradead.org/mailman/listinfo/linux-riscv
> > >
> 
> 
> 
>