Re: [PATCH v1 2/2] riscv: dts: starfive: add the missing monitor core

[Icenowy Zheng <uwu@icenowy.me>]

在 2022-07-13星期三的 15:21 +0000，Conor.Dooley@microchip.com写道：
> On 13/07/2022 16:12, Icenowy Zheng wrote:
> > 在 2022-07-13星期三的 15:09 +0000，Conor.Dooley@microchip.com写道：
> > > 
> > > 
> > > On 13/07/2022 16:02, Icenowy Zheng wrote:
> > > > 在 2022-07-13星期三的 14:55 +0000，Conor.Dooley@microchip.com写道：
> > > > > On 13/07/2022 15:26, Icenowy Zheng wrote:
> > > > > > 
> > > > > > 在 2022-07-11星期一的 19:43 +0100，Conor Dooley写道：
> > > > > > > From: Conor Dooley <conor.dooley@microchip.com>
> > > > > > > 
> > > > > > > The JH7100 has a 32 bit monitor core that is missing from
> > > > > > > the
> > > > > > > device
> > > > > > > tree. Add it (and its cpu-map entry) to more accurately
> > > > > > > reflect
> > > > > > > the
> > > > > > > actual topology of the SoC.
> > > > > > > 
> > > > > > > Signed-off-by: Conor Dooley <conor.dooley@microchip.com>
> > > > > > > ---
> > > > > > >  arch/riscv/boot/dts/starfive/jh7100.dtsi | 21
> > > > > > > +++++++++++++++++++++
> > > > > > >  1 file changed, 21 insertions(+)
> > > > > > > 
> > > > > > > diff --git a/arch/riscv/boot/dts/starfive/jh7100.dtsi
> > > > > > > b/arch/riscv/boot/dts/starfive/jh7100.dtsi
> > > > > > > index c617a61e26e2..92fce5b66d3d 100644
> > > > > > > --- a/arch/riscv/boot/dts/starfive/jh7100.dtsi
> > > > > > > +++ b/arch/riscv/boot/dts/starfive/jh7100.dtsi
> > > > > > > @@ -67,6 +67,23 @@ cpu1_intc: interrupt-controller {
> > > > > > >                         };
> > > > > > >                 };
> > > > > > > 
> > > > > > > +               E24: cpu@2 {
> > > > > > > +                       compatible = "sifive,e24",
> > > > > > > "riscv";
> > > > > > > +                       reg = <2>;
> > > > > > > +                       device_type = "cpu";
> > > > > > > +                       i-cache-block-size = <32>;
> > > > > > > +                       i-cache-sets = <256>;
> > > > > > > +                       i-cache-size = <16384>;
> > > > > > > +                       riscv,isa = "rv32imafc";
> > > > > > > +                       status = "disabled";
> > > > > > > +
> > > > > > > +                       cpu2_intc: interrupt-controller {
> > > > > > > +                               compatible = "riscv,cpu-
> > > > > > > intc";
> > > > > > > +                               interrupt-controller;
> > > > > > > +                               #interrupt-cells = <1>;
> > > > > > > +                       };
> > > > > > > +               };
> > > > > > > +
> > > > > > >                 cpu-map {
> > > > > > >                         cluster0 {
> > > > > > >                                 core0 {
> > > > > > > @@ -76,6 +93,10 @@ core0 {
> > > > > > >                                 core1 {
> > > > > > >                                         cpu = <&U74_1>;
> > > > > > >                                 };
> > > > > > > +
> > > > > > > +                               core2 {
> > > > > > > +                                       cpu = <&E24>;
> > > > > > > +                               };
> > > > > > 
> > > > > > Sorry but I think this change makes the topology more
> > > > > > inaccurate.
> > > > > > 
> > > > > > The E24 core is very independent, just another CPU core
> > > > > > connected
> > > > > > the
> > > > > > same bus -- even no coherency (E24 takes AHB, which is not
> > > > > > coherency-
> > > > > > sensible). Even the TAP of it is independent with the U74
> > > > > > TAP.
> > > > > > 
> > > > > > And by default it does not boot any proper code (if a
> > > > > > debugger
> > > > > > is
> > > > > > attached, it will discover that the E24 is in consistently
> > > > > > fault at
> > > > > > 0x0
> > > > > > (mtvec is 0x0 and when fault it jumps to 0x0 and fault
> > > > > > again),
> > > > > > until
> > > > > > its clock is just shutdown by Linux cleaning up unused
> > > > > > clocks.)
> > > > > > 
> > > > > > Personally I think it should be implemented as a remoteproc
> > > > > > instead.
> > > > > 
> > > > > Maybe I am missing something, but I don't quite get what the
> > > > > detail
> > > > > of how we access this in code has to do with the devicetree?
> > > > > It is added here in a disabled state, and will not be used by
> > > > > Linux.
> > > > > The various SiFive SoCs & SiFive corecomplex users that have
> > > > > a
> > > > > hart
> > > > > not capable of running Linux also have that hart documented
> > > > > in
> > > > > the
> > > > > devicetree.
> > > > > To me, what we are choosing to do with this hart does not
> > > > > really
> > > > > matter very much, since this is a description of what the
> > > > > hardware
> > > > > actually looks like.
> > > > 
> > > > The E24 is not in the core complex at all. It's just a dedicate
> > > > CPU
> > > > connected to another bus (well as I saw the document says the
> > > > E24
> > > > bus
> > > > is maximum 2G, I doubt whether it's the same bus with the U74
> > > > one).
> > > > 
> > > > The U74 MC only allows S5 management cores to be part of it,
> > > > not
> > > > E24.
> > > 
> > > So is the correct topology more like:
> > > cpu-map {
> > >         cluster0 {
> > >                 core0 {
> > >                         cpu = <&U74_0>;
> > >                 };
> > >                 core1 {
> > >                         cpu = <&U74_1>;
> > >                 };
> > >         };
> > >         cluster1 {
> > >                 core0 {
> > >                         cpu = <&E24>;
> > >                 };
> > >         };
> > > };
> > 
> > Considering E24 seems to see a total different bus connected to it,
> > I
> > don't think it even proper to add it to cpus node.
> 
> Well, it is a CPU is it not? How one is supposed to document that a
> CPU is not attached to the same buses I do not know however.

I don't think this kind of CPUs should exist in /cpus, they should just
be seen as peripherals as the main system. The speciality of FU[57]40's
management core is that they're in the same core complex with the CPU
cores that run Linux, just cores with a different capability that we
could not expand Linux to them.

> 
> > 
> > And I don't think it has a hart id of 2, as your node describes.
> 
> Do you have any idea what it would be then?

As I asked one of my friend who has JTAG access to JH7110, the hart id
is 0, the same with the first hart in U74-MC.