[PATCH v15 0/3] perf: Add Raspberry Pi AXI PMU driver
Ian Rogers
irogers at google.com
Mon Nov 2 15:42:58 PST 2026
Motivation & Background
=======================
Currently, Linux lacks a standard perf-API compatible driver for the
Broadcom AXI bus-watcher performance monitor blocks on Raspberry Pi
platforms (BCM2835, BCM2711, and BCM2712). Prior out-of-tree vendor
solutions relied on custom debugfs nodes and ad-hoc kthreads, preventing
integration with standard Linux perf tooling (`perf stat`, `perf list`,
etc.).
Note that this hardware block (`brcm,bcm2835-axiperf`) is distinct from
the Northstar BCM5301X AXI Trace Module (`brcm,bcm5301x-axi-pmu`,
https://lore.kernel.org/r/20261027175124.3518239-1-florian.fainelli@broadcom.com).
Key Architectural Features
==========================
1. Standard Linux Perf Integration:
- Exposes uncore AXI interconnect events via
`/sys/bus/event_source/devices/rpi_axi_pmu/` and documents them in
`Documentation/admin-guide/perf/rpi-axi-pmu.rst`.
- Supports up to 3 concurrently monitored `(bus, filter)` pairs per
monitor (`NUM_BUS_WATCHERS_PER_MONITOR = 3`), with reference counting
so all counter metrics on the same `(monitor, bus, filter)` share a
single hardware watcher without multiplexing.
- Enforces group validation (`rpi_axi_pmu__validate_group()`) to reject
event groups requiring more than 3 watchers per monitor at
`perf_event_open()` time, while supporting time-multiplexing across
more than 3 ungrouped buses.
- Masks counter deltas to each hardware register's bit width (32-bit for
`*trans`, 28-bit for `*wait`/`*max`, and 8-bit for `rpend`) and tracks
per-interval watermarks for level registers (`*max` and `rpend`) so
`perf stat` and `perf stat -I` report accurate peak levels.
2. Hybrid MMIO & Firmware Mailbox Architecture:
- System Monitor (`monitor=0`): Performs fast atomic-safe MMIO reads.
- VideoCore VPU Monitor (`monitor=1`): Polled asynchronously in process
context (`vpu_work`) via a single 44-word bulk Raspberry Pi firmware
mailbox `RPI_FIRMWARE_GET_PERIPH_REG` call across all three VPU bus
watchers. At probe time, `rpi_axi_pmu_vpu_accessible()` verifies that
the firmware can reach the VPU monitor hardware and hides the `vpu_*`
sysfs events if unavailable.
3. Per-SoC & Stepping Support:
- Patch 1 adds the Device Tree binding (`brcm,bcm2835-axiperf`,
`brcm,bcm2711-axiperf`, `brcm,bcm2712-axiperf`) and `MAINTAINERS`
entry.
- Patch 2 adds the driver and admin-guide documentation for BCM2835 and
BCM2711.
- Patch 3 adds BCM2712 (Raspberry Pi 5) C0 and D0 support, including
runtime D0 stepping detection via the bus-watcher AXI ID mask field,
C0 (13-port) and D0 (10-port) System bus tables, 6-bit sparse AXI
master ID filter tables (`BCM2712_FLT_MAX = 64`), and `ratrans`.
Hardware Validation
===================
The bus tables, master ID filters, and register layouts have been validated
against Broadcom register documentation and direct hardware measurements on:
- BCM2835 (Raspberry Pi 1 B+)
- BCM2837 (Raspberry Pi 3 B+)
- BCM2711 B0 (Raspberry Pi 4 B and Raspberry Pi 400)
- BCM2712 C0 (Raspberry Pi 5 rev 1.0)
- BCM2712 D0 (Raspberry Pi 5 rev 1.2 and Raspberry Pi 500)
Example on Raspberry Pi 5 (BCM2712) counting all 11 metrics on the `arm`
bus simultaneously on a single bus watcher:
$ sudo perf stat -a \
-e rpi_axi_pmu/arm_atwait/ -e rpi_axi_pmu/arm_atrans/ \
-e rpi_axi_pmu/arm_amax/ -e rpi_axi_pmu/arm_wwait/ \
-e rpi_axi_pmu/arm_wtrans/ -e rpi_axi_pmu/arm_wmax/ \
-e rpi_axi_pmu/arm_rwait/ -e rpi_axi_pmu/arm_rtrans/ \
-e rpi_axi_pmu/arm_rmax/ -e rpi_axi_pmu/arm_rpend/ \
-e rpi_axi_pmu/arm_ratrans/ \
-- dd if=/dev/zero of=/dev/null bs=1M count=2048
Performance counter stats for 'system wide':
222,085 rpi_axi_pmu/arm_atwait/
21,789,405 rpi_axi_pmu/arm_atrans/
2,196 rpi_axi_pmu/arm_amax/
39,623,624 rpi_axi_pmu/arm_wwait/
43,482,567 rpi_axi_pmu/arm_wtrans/
2,197 rpi_axi_pmu/arm_wmax/
554 rpi_axi_pmu/arm_rwait/
96,264 rpi_axi_pmu/arm_rtrans/
1 rpi_axi_pmu/arm_rmax/
5 rpi_axi_pmu/arm_rpend/
48,148 rpi_axi_pmu/arm_ratrans/
0.208600205 seconds time elapsed
And filtering `per_atrans` by `VPU_UC0` (`filter=1`) and `ARM` (`filter=11`):
$ sudo perf stat -a \
-e rpi_axi_pmu/per_atrans/ \
-e rpi_axi_pmu/per_atrans,filter=1/ \
-e rpi_axi_pmu/per_atrans,filter=11/ \
-- sleep 1
Performance counter stats for 'system wide':
29,393 rpi_axi_pmu/per_atrans/
19,934 rpi_axi_pmu/per_atrans,filter=1/
9,452 rpi_axi_pmu/per_atrans,filter=11/
1.001089440 seconds time elapsed
Tooling & AI Assistance
=======================
Per Documentation/process/coding-assistants.rst and
Documentation/process/generated-content.rst:
- Tools used: Google Antigravity (gemini-3.1-pro), Anthropic Claude, and
Sashiko automated kernel patch review, alongside standard kernel tooling
(scripts/checkpatch.pl, scripts/kernel-doc, dt-doc-validate).
- Summary of prompts and assistance:
- Assisted in converting the downstream Raspberry Pi debugfs AXI monitor
(drivers/perf/raspberrypi_axi_monitor.c) into a Linux perf PMU driver
(drivers/perf/rpi_axi_pmu.c), generating the repetitive
PMU_EVENT_ATTR_STRING sysfs event tables across SoC generations, the
Device Tree binding schema (brcm,bcm2835-axiperf.yaml), and the
admin-guide documentation (Documentation/admin-guide/perf/rpi-axi-pmu.rst).
- Assisted with iterative review and fixes across revisions, including
synchronization between atomic MMIO reads (pmu->lock) and deferred
process-context VPU mailbox polling (vpu_mutex / vpu_work), bulk VPU
mailbox reads, BCM2712 C0/D0 stepping detection and filter masks, and
per-interval watermark tracking for level registers (*max and rpend).
- Portions affected:
Documentation/devicetree/bindings/perf/brcm,bcm2835-axiperf.yaml,
Documentation/admin-guide/perf/rpi-axi-pmu.rst, and
drivers/perf/rpi_axi_pmu.c.
- Testing and verification:
All generated code and tables were reviewed by the authors and tested on
physical BCM2835, BCM2837, BCM2711 B0, BCM2712 C0, and BCM2712 D0
hardware using perf list and perf stat (single-watcher multi-metric
counting, perf stat -I interval watermark tracking, AXI master ID
filtering, VPU firmware accessibility probing, and 10-bus multiplexing).
Changelog
=========
v15:
- Device Tree Bindings & MAINTAINERS (Patch 1):
- Dropped `kernel-list at raspberrypi.com` from `maintainers:`, sorted
`compatible:` alphabetically, and required both `reg` items (`System
monitor registers` and `VPU monitor registers`) across all SoCs per
Krzysztof Kozlowski.
- Added `MAINTAINERS` entry and noted distinction from
`brcm,bcm5301x-axi-pmu`.
- Driver & Hardware Validation Fixes (Patches 2 & 3, co-developed with Dom
Cobley on https://github.com/raspberrypi/linux/pull/7671):
- Added `Documentation/admin-guide/perf/rpi-axi-pmu.rst`.
- Masked `*wait`/`*max` deltas to 28 bits (`GENMASK(27, 0)`) and `rpend`
to 8 bits (`GENMASK(7, 0)`), and implemented per-read-interval
watermark tracking (`counter_is_level()`) for `*max` and `rpend` with
`prev_count` reset under `pmu->lock` in `rpi_axi_pmu_read()`.
- Removed non-existent `ratrans` on BCM2835 and BCM2711, and corrected
the BCM2835 VPU bus table (`0..13` matching BCM2711 plus `14 = sdram`).
- Re-enabled the BCM2712 VPU monitor (`0x7ee00000`) via the firmware
mailbox, added probe-time `rpi_axi_pmu_vpu_accessible()` verification,
and hid `vpu_*` sysfs events via `.is_visible` when firmware support is
unavailable.
- Replaced per-event VPU mailbox reads with a single 44-word bulk read
(`rpi_axi_pmu_vpu_read_watchers()`) with `__le32` endianness
conversions (`cpu_to_le32()` / `le32_to_cpu()`) and `config_gen[]`
tracking, and defaulted `hrtimer_interval_ms` to 1000 ms when the VPU
mailbox is used.
- Used `rpi_firmware_find_node()` to locate the firmware node.
- Corrected BCM2712 sparse AXI master ID filter numbering
(`BCM2712_FLT_MAX = 64`), added BCM2712 D0 runtime stepping detection
(`pmu->axi_id_mask`), D0 System/VPU filter mask encoding, the 10-port
D0 System bus table, and renamed `bstm_top`/`arb3`/`src` to
`rp1`/`sdio_wifi`/`sd_dma`.
v14: https://lore.kernel.org/r/20261002175738.3242646-1-irogers@google.com
- Removed `firmware` phandle property from `brcm,bcm2835-axiperf.yaml` per
Krzysztof Kozlowski.
- Removed `LINUX_VERSION_CODE` checks per Uwe Kleine-Koenig and Florian
Fainelli.
- Separated BCM2835, BCM2711, and BCM2712 bus and filter tables, exposed
all counter metrics per bus without synthetic `Bytes` scaling, and added
bus-watcher reference counting and group validation.
v13: https://lore.kernel.org/r/20260815172712.50119-1-irogers@google.com
- Added `allOf`/`if` block in `brcm,bcm2835-axiperf.yaml` and cleaned up
sysfs attribute definitions and comments.
v12: https://lore.kernel.org/r/20260815082048.4043217-1-irogers@google.com
- Added Device Tree binding patch (`brcm,bcm2835-axiperf.yaml`), split
per-chip `attr_groups`, and returned `-EPROBE_DEFER` when firmware is not
yet ready.
v11: https://lore.kernel.org/r/20260815082048.4043217-1-irogers@google.com
- Added `READ_ONCE()`/`WRITE_ONCE()` on `enabled[idx]` and restored
`flush_work()` in `rpi_axi_pmu__exit()` so queued VPU watcher disables
complete on module unload.
v10: https://lore.kernel.org/r/20260815082048.4043217-1-irogers@google.com
- Removed unused `num_started` state variable from `struct rpi_axi_hw_events`.
v9: https://lore.kernel.org/r/20260813064228.2799872-1-irogers@google.com
- Adjusted PMU and CPU hotplug registration order in `probe()` and
simplified BCM2712 sysfs event visibility.
v8: https://lore.kernel.org/r/20260813064228.2799872-1-irogers@google.com
- Switched `pmu->lock` to `raw_spinlock_t` for `PREEMPT_RT` compatibility
and reduced `RPI_AXI_PMU_TIMER_INTERVAL` from 2000 ms to 100 ms.
v7: https://lore.kernel.org/r/20260813052419.2770841-1-irogers@google.com
- Isolated all BCM2712 definitions into Patch 2 and initialized
`vpu_disable_pending` on watcher allocation.
v6: https://lore.kernel.org/r/20260813043533.2744442-1-irogers@google.com
- Fixed `Kconfig` `RASPBERRYPI_FIRMWARE` dependency, added
`vpu_disable_pending` for VPU watcher teardown, and let the hrtimer
self-terminate when `active_events == 0`.
v5: https://lore.kernel.org/r/20260812230800.2646749-1-irogers@google.com
- Added `monitor_running` tracking, deferred VPU watcher disablement to
`vpu_work`, and protected `event->hw.state` updates under `pmu->lock`.
v4: https://lore.kernel.org/r/20260812052417.2361040-1-irogers@google.com
- Avoided global monitor reset when other watchers are active and fixed
error handling on mailbox read failures.
v3: https://lore.kernel.org/r/20260812002728.2283079-1-irogers@google.com
- Switched hrtimer to `HRTIMER_MODE_REL_SOFT`, set `.suppress_bind_attrs =
true`, and cleaned up probe error unwinding.
v2: https://lore.kernel.org/r/20260811235112.2267241-1-irogers@google.com
- Added `of_device_get_match_data()` chip selection, CPU affinity check in
`event_init()`, and counter wrap delta masking.
v1: https://lore.kernel.org/r/20260811083828.2057695-1-irogers@google.com
- Initial submission.
Ian Rogers (3):
dt-bindings: perf: Add Broadcom Raspberry Pi AXI PMU definition
perf: Add Raspberry Pi BCM2835 AXI PMU driver
perf: Add Raspberry Pi 5 (BCM2712) AXI PMU support
Documentation/admin-guide/perf/index.rst | 1 +
.../admin-guide/perf/rpi-axi-pmu.rst | 164 +
.../bindings/perf/brcm,bcm2835-axiperf.yaml | 50 +
MAINTAINERS | 9 +
drivers/perf/Kconfig | 14 +
drivers/perf/Makefile | 1 +
drivers/perf/rpi_axi_pmu.c | 4145 +++++++++++++++++
7 files changed, 4384 insertions(+)
create mode 100644 Documentation/admin-guide/perf/rpi-axi-pmu.rst
create mode 100644 Documentation/devicetree/bindings/perf/brcm,bcm2835-axiperf.yaml
create mode 100644 drivers/perf/rpi_axi_pmu.c
--
2.56.0.385.gac671c819a-goog
More information about the linux-arm-kernel
mailing list