[RFC PATCH] dt-bindings: riscv: Add YAML documentation for PMU

[Anup Patel]

On Tue, Jul 28, 2020 at 8:35 AM Zong Li <zong.li at sifive.com> wrote:
>
> On Mon, Jul 27, 2020 at 9:13 PM Anup Patel <anup at brainfault.org> wrote:
> >
> > On Mon, Jul 27, 2020 at 1:57 PM Zong Li <zong.li at sifive.com> wrote:
> > >
> > > Add device tree bindings for performance monitor unit. It passes the
> > > dt_binding_check verification.
> > >
> > > Signed-off-by: Zong Li <zong.li at sifive.com>
> > > ---
> > >  .../devicetree/bindings/riscv/pmu.yaml        | 71 +++++++++++++++++++
> > >  1 file changed, 71 insertions(+)
> > >  create mode 100644 Documentation/devicetree/bindings/riscv/pmu.yaml
> > >
> > > diff --git a/Documentation/devicetree/bindings/riscv/pmu.yaml b/Documentation/devicetree/bindings/riscv/pmu.yaml
> > > new file mode 100644
> > > index 000000000000..0c49039a5d3b
> > > --- /dev/null
> > > +++ b/Documentation/devicetree/bindings/riscv/pmu.yaml
> > > @@ -0,0 +1,71 @@
> > > +# SPDX-License-Identifier: GPL-2.0
> > > +%YAML 1.2
> > > +---
> > > +$id: http://devicetree.org/schemas/riscv/pmu.yaml#
> > > +$schema: http://devicetree.org/meta-schemas/core.yaml#
> > > +
> > > +title: RISC-V Performance Monitor Units
> > > +
> > > +maintainers:
> > > +  - Zong Li <zong.li at sifive.com>
> > > +  - Paul Walmsley <paul.walmsley at sifive.com>
> > > +  - Palmer Dabbelt <palmer at dabbelt.com>
> > > +
> > > +properties:
> > > +  compatible:
> > > +    items:
> > > +      - const: riscv,pmu
> > > +
> > > +  riscv,width-hpmcntr:
> > > +    description: The width of hpmcounter CSRs. Default is 64.
> > > +    $ref: /schemas/types.yaml#/definitions/uint32
> > > +
> > > +  riscv,n-hpmcntr:
> > > +    description: The number of hpmcounter CSRs. Default is zero.
> > > +    $ref: /schemas/types.yaml#/definitions/uint32
> > > +
> > > +  riscv,hw-event-map:
> > > +    description: The mapping of generic hardware events to values of hpmcounter.
> > > +      The key is the encoding of generic hardware events, and the value is the
> > > +      actual value which is implemented by platform. If there is no a key-value
> > > +      pair for specific generic hardware event, view the generic hardware event
> > > +      as not supported. CYCLE and INSTRET be mapped by default, so we shouldn't
> > > +      list PERF_COUNT_HW_CPU_CYCLES and PERF_COUNT_HW_INSTRUCTIONS here.
> > > +    $ref: /schemas/types.yaml#/definitions/uint32-array
> > > +
> > > +  riscv,hw-cache-event-map:
> > > +    description: The mapping of generic hardware cache events to values of
> > > +      hpmcounter. The key is encoding of generic hardware cache events, and the
> > > +      value is the actual value which is implemented by platform. If there is no
> > > +      a key-value pair for specific generic hardware cache event, view the
> > > +      generic hardware cache event as not supported.
> > > +    $ref: /schemas/types.yaml#/definitions/uint32-array
> > > +
> > > +  riscv,hpmcntr-of-event:
> > > +    description: The mapping of platform hardware events to allowed hmpcounters.
> > > +      The key is the platform hardware event, and the value is the bitmap for
> > > +      hmpcounters which support this event. If there is no a key-value pair for
> > > +      specific platform hardware events, view the platform hardware events as
> > > +      supported by all hpmcounters.
> > > +    $ref: /schemas/types.yaml#/definitions/uint32-array
> > > +
> > > +required:
> > > +  - compatible
> > > +
> > > +additionalProperties: false
> > > +
> > > +examples:
> > > +  - |
> > > +    pmu {
> > > +      compatible = "riscv,pmu";
> > > +      riscv,width-hpmcntr = <40>;
> > > +      riscv,n-hpmcntr = <2>;
> > > +      riscv,hw-event-map = <0x3 0x0202
> > > +                            0x4 0x4000>;
> > > +      riscv,hw-cache-event-map = <0x010201 0x0102
> > > +                                  0x010204 0x0802>;
> > > +      riscv,hpmcntr-of-event = <0x100 0x18
> > > +                                0x400 0x10>;
> > > +    };
> > > +
> > > +...
> > > --
> > > 2.27.0
> > >
> >
> > I don't see the point of sending DT bindings documents
> > until the SBI PMU extension is defined and accepted
> > by everyone.
> >
>
> It seems to me that it doesn't conflict with the SBI PMU extension.
> SBI PMU extension is the interface for communication with the lower
> privilege levels, we still can come out the DT in parallel, and make
> sure what information is needed by m-mode firmware. OTOH, we also need
> a Linux driver which has to parse the dtb when we build Linux as nommu
> and run it on m-mode privilege, so if we could define the DT for PMU
> node, we can go the driver for nommu first (the implementation would
> also consider HW counter, SW counter together). I think you have got
> the picture about m-mode firmware, it would be great to start to
> collect the idea of DT.

The RISC-V PMU DT binding is not just for Linux. Even M-mode
runtime firmware (OpenSBI) will be using this binding.

The DT bindings you are proposing is conflicting with the
expected RISC-V PMU DT bindings:
1. The event-map should be mapping SBI PMU event_idx
to HARDWARE HPMEVENT value whereas this binding
defines event-map as mapping Linux perf event types to
HPMEVENT value.
2. Each HPMCOUNTER can support a different set of events
so this DT bindings won't work for it.
3. The SBI PMU driver can have optional edge-triggered
per-HART interrupts for overflow detection. This DT bindings
does not consider this as well.

The SBI PMU event_idx definition is not yet finalized. Please
wait for SBI PMU extension to be finalized.

Please avoid adding DT bindings without SBI PMU extension
being finalized.

Regards,
Anup