Re: [PATCH v3] dt-bindings: riscv: add SBI PMU event mappings
From: Andrew Jones
Date: Tue Jan 03 2023 - 04:28:28 EST
On Mon, Jan 02, 2023 at 04:55:51PM +0000, Conor Dooley wrote:
> From: Conor Dooley <conor.dooley@xxxxxxxxxxxxx>
>
> The SBI PMU extension requires a firmware to be aware of the event to
> counter/mhpmevent mappings supported by the hardware. OpenSBI may use
> DeviceTree to describe the PMU mappings. This binding is currently
> described in markdown in OpenSBI (since v1.0 in Dec 2021) & used by QEMU
> since v7.2.0.
>
> Import the binding for use while validating dtb dumps from QEMU and
> upcoming hardware (eg JH7110 SoC) that will make use of the event
> mapping.
>
> Link: https://github.com/riscv-software-src/opensbi/blob/master/docs/pmu_support.md
> Link: https://github.com/riscv-non-isa/riscv-sbi-doc/blob/master/riscv-sbi.adoc # Performance Monitoring Unit Extension
> Co-developed-by: Atish Patra <atishp@xxxxxxxxxxxx>
> Signed-off-by: Atish Patra <atishp@xxxxxxxxxxxx>
> Signed-off-by: Conor Dooley <conor.dooley@xxxxxxxxxxxxx>
> ---
> Changes in v3:
> - align descriptions to SBI spec (and fix a misinterpretation of mine)
> - switch to a nested items description, since the descriptions are for
> the elements of each entry, not the entries themselves
>
> Changes in v2:
> - use the schema mechanism for dependancies between properties
> - +CC perf maintainers...
> - move the matrix element descriptions into regular item descriptions
> rather than doing so freeform in the property description
> - drop some description text that no longer applies since changes were
> made to the SBI spec
> - drop mention of the "generic platform" which is OpenSBI specific
> - drop the min/max items from the matrices, they don't appear to be
> needed?
>
> Note:
> OpenSBI is BSD-2-Clause licensed so I am unsure as to whether I can
> submit it with a dual license.
> ---
> .../devicetree/bindings/perf/riscv,pmu.yaml | 160 ++++++++++++++++++
> 1 file changed, 160 insertions(+)
> create mode 100644 Documentation/devicetree/bindings/perf/riscv,pmu.yaml
>
> diff --git a/Documentation/devicetree/bindings/perf/riscv,pmu.yaml b/Documentation/devicetree/bindings/perf/riscv,pmu.yaml
> new file mode 100644
> index 000000000000..87a47e365311
> --- /dev/null
> +++ b/Documentation/devicetree/bindings/perf/riscv,pmu.yaml
> @@ -0,0 +1,160 @@
> +# SPDX-License-Identifier: BSD-2-Clause
> +%YAML 1.2
> +---
> +$id: http://devicetree.org/schemas/perf/riscv,pmu.yaml#
> +$schema: http://devicetree.org/meta-schemas/core.yaml#
> +
> +title: RISC-V SBI PMU events
> +
> +maintainers:
> + - Atish Patra <atishp@xxxxxxxxxxxx>
> +
> +description: |
> + SBI PMU extension supports allow supervisor software to configure, start &
support allows
I'd replace '&' with 'and'
> + stop any performance counter at anytime. Thus, a user can leverage full
> + capability of performance analysis tools such as perf if the SBI PMU
s/full capability/all capabilities/
> + extension is enabled. The OpenSBI implementation makes the following
^ Shouldn't we drop references to OpenSBI?
> + assumptions about the hardware platform:
> +
> + The platform must provide information about PMU event to counter mapping
mappings
> + via device tree or platform specific hooks. Otherwise, the SBI PMU
> + extension will not be enabled.
> +
> + The platforms should provide information about the PMU event selector
> + values that should be encoded in the expected value of MHPMEVENTx while
> + configuring MHPMCOUNTERx for that specific event. This can be done via a
> + device tree or platform specific hooks. The exact value to be written to
> + the MHPMEVENTx is completely dependent on the platform.
> +
> + For information on the SBI specification see the section "Performance
> + Monitoring Unit Extension" of:
> + https://github.com/riscv-non-isa/riscv-sbi-doc/blob/master/riscv-sbi.adoc
> +
> +properties:
> + compatible:
> + const: riscv,pmu
> +
> + riscv,event-to-mhpmevent:
> + $ref: /schemas/types.yaml#/definitions/uint32-matrix
> + description:
> + Represents an ONE-to-ONE mapping between a PMU event and the event
> + selector value that platform expects to be written to the MHPMEVENTx CSR
^ the
> + for that event.
> + The mapping is encoded in an matrix format where each element represents
> + an event.
> + This property shouldn't encode any raw hardware event.
> + items:
> + items:
> + - description: event_idx, a 20-bit wide encoding of the event type and
> + code. Refer to the SBI specification for a complete description of
> + the event types and codes.
> + - description: upper 32 bits of the event selector value for MHPMEVENTx
> + - description: lower 32 bits of the event selector value for MHPMEVENTx
> +
> + riscv,event-to-mhpmcounters:
> + $ref: /schemas/types.yaml#/definitions/uint32-matrix
> + description:
> + Represents a MANY-to-MANY mapping between a range of events and all the
> + MHPMCOUNTERx in a bitmap format that can be used to monitor these range
> + of events. The information is encoded in an matrix format where each
> + element represents a certain range of events and corresponding counters.
> + This property shouldn't encode any raw event.
> + items:
> + items:
> + - description: first event_idx of the range of events
> + - description: last event_idx of the range of events
> + - description: bitmap of MHPMCOUNTERx for this event
> +
> + riscv,raw-event-to-mhpmcounters:
> + $ref: /schemas/types.yaml#/definitions/uint32-matrix
> + description:
> + Represents an ONE-to-MANY or MANY-to-MANY mapping between the rawevent(s)
> + and all the MHPMCOUNTERx in a bitmap format that can be used to monitor
> + that raw event.
> + The encoding of the raw events are platform specific. The information is
> + encoded in a matrix format where each element represents the specific raw
> + event(s).
> + If a platform directly encodes each raw PMU event as a unique ID, the
> + value of variant must be 0xffffffff_ffffffff.
> + items:
> + items:
> + - description:
> + upper 32 invariant bits for the range of events
> + - description:
> + lower 32 invariant bits for the range of events
> + - description:
> + upper 32 bits of the variant bit mask for the range of events
> + - description:
> + lower 32 bits of the variant bit mask for the range of events
> + - description:
> + bitmap of all MHPMCOUNTERx that can monitor the range of events
> +
> +dependencies:
> + "riscv,event-to-mhpmevent": [ "riscv,event-to-mhpmcounters" ]
> + "riscv,event-to-mhpmcounters": [ "riscv,event-to-mhpmevent" ]
> +
> +required:
> + - compatible
> +
> +additionalProperties: false
> +
> +examples:
> + - |
> + pmu {
> + compatible = "riscv,pmu";
> + riscv,event-to-mhpmevent = <0x0000B 0x0000 0x0001>;
> + riscv,event-to-mhpmcounters = <0x00001 0x00001 0x00000001>,
> + <0x00002 0x00002 0x00000004>,
> + <0x00003 0x0000A 0x00000ff8>,
> + <0x10000 0x10033 0x000ff000>;
> + riscv,raw-event-to-mhpmcounters =
> + /* For event ID 0x0002 */
> + <0x0000 0x0002 0xffffffff 0xffffffff 0x00000f8>,
> + /* For event ID 0-4 */
> + <0x0 0x0 0xffffffff 0xfffffff0 0x00000ff0>,
> + /* For event ID 0xffffffff0000000f - 0xffffffff000000ff */
> + <0xffffffff 0x0 0xffffffff 0xffffff0f 0x00000ff0>;
> + };
> +
> + - |
> + /*
> + * For HiFive Unmatched board the encodings can be found here
> + * https://sifive.cdn.prismic.io/sifive/1a82e600-1f93-4f41-b2d8-86ed8b16acba_fu740-c000-manual-v1p6.pdf
I'd add a blank line here in this comment block.
> + * This example also binds standard SBI PMU hardware id's to U74 PMU event
IDs
> + * codes, U74 uses a bitfield for events encoding, so several U74 events
> + * can be bound to single perf id.
^ a ID
> + * See SBI PMU hardware id's in OpenSBI's include/sbi/sbi_ecall_interface.h
IDs
Linux has its own copy of enum sbi_pmu_hw_generic_events_t in
arch/riscv/include/asm/sbi.h which can be referenced.
> + */
> + pmu {
> + compatible = "riscv,pmu";
> + riscv,event-to-mhpmevent =
> + /* SBI_PMU_HW_CACHE_REFERENCES -> Instruction or Data cache/ITIM busy */
> + <0x00003 0x00000000 0x1801>,
> + /* SBI_PMU_HW_CACHE_MISSES -> Instruction or Data cache miss or MMIO access */
> + <0x00004 0x00000000 0x0302>,
> + /* SBI_PMU_HW_BRANCH_INSTRUCTIONS -> Conditional branch retired */
> + <0x00005 0x00000000 0x4000>,
> + /* SBI_PMU_HW_BRANCH_MISSES -> Branch or jump misprediction */
> + <0x00006 0x00000000 0x6001>,
> + /* L1D_READ_MISS -> Data cache miss or MMIO access */
> + <0x10001 0x00000000 0x0202>,
> + /* L1D_WRITE_ACCESS -> Data cache write-back */
> + <0x10002 0x00000000 0x0402>,
> + /* L1I_READ_ACCESS -> Instruction cache miss */
> + <0x10009 0x00000000 0x0102>,
> + /* LL_READ_MISS -> UTLB miss */
> + <0x10011 0x00000000 0x2002>,
> + /* DTLB_READ_MISS -> Data TLB miss */
> + <0x10019 0x00000000 0x1002>,
> + /* ITLB_READ_MISS-> Instruction TLB miss */
> + <0x10021 0x00000000 0x0802>;
> + riscv,event-to-mhpmcounters = <0x00003 0x00006 0x18>,
> + <0x10001 0x10002 0x18>,
> + <0x10009 0x10009 0x18>,
> + <0x10011 0x10011 0x18>,
> + <0x10019 0x10019 0x18>,
> + <0x10021 0x10021 0x18>;
> + riscv,raw-event-to-mhpmcounters = <0x0 0x0 0xffffffff 0xfc0000ff 0x18>,
> + <0x0 0x1 0xffffffff 0xfff800ff 0x18>,
> + <0x0 0x2 0xffffffff 0xffffe0ff 0x18>;
> + };
> --
> 2.39.0
>
Thanks,
drew