Skip to content

arm64: dts: apple: t8142: CPU idle, PMU and cluster power gate on every M5 (on #2325) - #2328

Open
RyanTheTide wants to merge 4 commits into
integration/aurorafrom
dtfill/t8142-cpu-idle-default
Open

RyanTheTide wants to merge 4 commits into
integration/aurorafrom
dtfill/t8142-cpu-idle-default

Conversation

@RyanTheTide

Copy link
Copy Markdown
Contributor

Problem

On the MacBook Pro 14-inch M5 (J704), the boot P core never powers down in WFI, perf has no hardware counters, and cpuidle-psci has no idle state under the loader's PSCI monitor. The CPU cluster power-gate enables, the CPU PMU and the PSCI retention state were described only for the MacBook Airs:

  • cluster_pwrgate existed only in t8142-j813.dts, so the J815 lacked it too.
  • cpu_pmu was enabled per board.
  • The idle states lived in the Air-only t8142-j813-psci-idle.dtsi.

These are CPU and PMGR facts of the T8142: the J704, J813 and J815 ADTs have the same /cpus and the same pmgr.

Change

  • Move cluster_pwrgate into t8142.dtsi, as t8140.dtsi already does.
  • Enable cpu_pmu in t8142.dtsi.
  • Rename t8142-j813-psci-idle.dtsi to t8142-psci-idle.dtsi and include it from t8142.dtsi. cpuidle-psci still only probes when a /psci node is present.
  • Update the file name in the comment in cpuidle-apple.c.

Testing

cpp + dtc before and after, decompiled trees compared node by node:

  • J813: unchanged.
  • J815: gains cluster_pwrgate.
  • J704: gains cluster_pwrgate, the PMU, the idle-states node and cpu-idle-states on all ten CPUs.

dt-validate -m reports nothing new on any of the three boards.

Still to run

  • @term-guy (J704), booted through the loader's PSCI monitor:
    • /sys/devices/system/cpu/cpu*/cpuidle/state*/usage climbs on every CPU.
    • perf stat -e cycles,instructions sleep 1 counts.
    • The cluster power-gate driver probes.
    • Three s2idle cycles pass. Deep sleep stays off: the J704 has no sleep controller node yet.
  • @m0d16l14n1 (J815): cpuidle residency on the boot core before and after.
Details

Assisted-by: LLM (Claude:Opus 5.5)

…ery M5

The Neural Engine (with its three DARTs and ANE_SYS power state), the
video encoder (two DARTs, VENC power tree) and both JPEG engines (DART,
JPG power state) are SoC blocks: the J704, J813 and J815 ADTs give the
same addresses, interrupts and power gates for ane, dart-ane, ave,
dart-ave, jpeg0, jpeg1 and dart-jpeg. Only the J815 enabled all three
and the J813 only the encoder (twice); the J704 had none of them.

Turn them on in t8142-peripherals.dtsi and drop the per-board enables.
The ANE and the encoder run firmware that iBoot preloads: the boot
loader hands the preload over by compatible and disables the node when
the ADT has none, so a board without a preload keeps them off. Every
DART keeps its engine's power domain. The JPEG engines need no firmware.

J815: no change in the built tree beyond status "okay" becoming implicit.
J813: gains the Neural Engine and both JPEG engines. J704: gains all
three.

The ANE and JPEG enables on J815 come from m0d16l14n1's
33d8f13 and c3c1417.

Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
…ry M5

The CPU cluster power-gate enables, the CPU PMU and the PSCI retention
idle state are CPU/PMGR facts of the T8142, not of a board: the J704,
J813 and J815 ADTs have the same /cpus and the same pmgr. They were
only described for the MacBook Airs (cluster_pwrgate in t8142-j813.dts
alone, cpu_pmu enabled per board, the idle states in the Air-only
t8142-j813-psci-idle.dtsi), so on the MacBook Pro 14" (J704) the boot
P core never powered down in WFI, perf had no hardware counters and
cpuidle-psci had no state under aura's PSCI monitor.

Move cluster_pwrgate into t8142.dtsi (as t8140.dtsi already has it),
enable cpu_pmu there, and rename t8142-j813-psci-idle.dtsi to
t8142-psci-idle.dtsi, included from t8142.dtsi. Without a /psci node
cpuidle-psci does not probe, as before.

J813: unchanged. J815: gains cluster_pwrgate. J704: gains
cluster_pwrgate, the PMU and the retention idle state.

cluster_pwrgate is gig3m's (c58ab8c).

Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
@RyanTheTide

Copy link
Copy Markdown
Contributor Author

No row has covered this one yet. gig3m's J813 row on #2325 ran with this PR on top, but this change leaves the J813 tree unchanged, so that row only shows it doesn't break the J813 (including two deep sleeps). term-guy's J704 rows on #2325 and #2329 were on a kernel without it.

Still open, as listed above:

  • @term-guy (J704, through the loader's PSCI monitor): cpuidle usage climbing on every CPU, perf stat -e cycles,instructions sleep 1, the cluster power-gate probe, and three s2idle cycles.
  • @m0d16l14n1 (J815): cpuidle residency on the boot core before and after.

@m0d16l14n1

Copy link
Copy Markdown
Contributor

J815 (MacBook Air 15" M5) row for the boot-core cpuidle residency. Kernel: integration/aurora 2c35f5094de1 + this head 95dda1cb24eb (with #2325 1f4da1ec68ee), #2329, #2330, #2383, #2384, #2385 at their current heads.

Before and after. On the J815 this PR does not change the device tree: the J815 DTB with it is identical, node for node, to our previous J815 DTB without it, apart from phandle numbers. (In the old tree /pmu and dart-aop had an explicit status = "okay"; now they have no status property, which means the same.) So a "before" boot runs the same description. The row below is the "after".

Residency. 30 s with the desktop idle, deltas of cpuidle/stateN/{usage,time} (apple_idle driver):

CPU state0 (WFI) state1 (index 1)
cpu0 (boot core) 2 entries, 0.0% 14144 entries, 98.5%
all 10 CPUs 0.0% each 195531 entries, mean 91.7%

On T8142, index 1 is the second WFI entry by design (the driver keeps index 1 for the s2idle path, and CYC_OVRD is locked from M4 on), so it is named WFI, not CPU PD. The governor uses it for nearly all idle time on every core. No cpuidle or PMU errors in the boot.

@term-guy

Copy link
Copy Markdown
Contributor

J704 (MacBook Pro 14" M5) row. macOS 27.2 (26B123) stub. Kernel: this head 95dda1cb24eb (with #2325 1f4da1ec68ee), plus the J704 GPU (#2093 at f41fdbe5) and our Fedora install's J704 DTS on top. Booted through the installed aura.

No PSCI monitor here. The installed aura starts the CPUs with spin-table (secondaries confined at EL2), so there is no /psci node and cpuidle-psci does not probe. The idle driver is apple_idle, as in m0d16l14n1's J815 row. The tree's idle-states/cpu-retention and cpu-idle-states are present on all ten CPUs.

cpuidle. 30 s with the desktop idle, deltas of cpuidle/stateN/usage and residency:

CPU state0 (WFI) state1 (WFI)
cpu0 (boot core) +0 +2321, 99.8%
cpu1-cpu9 +0 to +6 +397 to +2037 each, 99.7-100%

Every CPU climbs.

PMU. Before this PR the J704 had no hardware counters (The cycles event is not supported). With it: hw perfevents: enabled with apple_t8142_pmu PMU driver, 10 (0,000003ff) counters available, and perf stat -e cycles,instructions sleep 1 counts 841,913 cycles and 2,012,280 instructions.

Cluster power gate. The driver ran at boot (apple_cluster_pwrgate.dyndbg=+p):

/soc/power-gate@380724000: power gate 0 0x80000000, now 0x80000000
/soc/power-gate@380724000: power gate 1 0x0, now 0x80000000

Read back through /dev/mem: both words are 0x80000000. On the kernel without this PR, the P word (+0x60, the boot core's) is 0x00000000, as the description says.

s2idle. rtcwake cannot wake the J704 (the SMC RTC is not a wakeup source), so each wake was a key press on the internal keyboard. The right-port USB gadget refuses to suspend after a resume (dwc3 41a280000.usb: dwc3_plat_suspend returns -11, on the kernel without this PR too), so the three-cycle run unbinds it first.

  • Three cycles in one boot: three s2idle entries and exits, each woken by a key, no device refusing to suspend.
  • Single cycles on three more boots of this kernel: two woke on the key. In the first of them, the boot also ran the cpuidle and perf checks above. It entered s2idle, did not wake on the key, and the Mac reset itself about 3.5 minutes later. The journal stops before the suspend, so there is nothing more to go on. An exact repeat of that boot woke normally, and so did a boot of this kernel with only the power gate disabled. So that is one hang in six s2idle entries on this kernel, not reproduced. I can't tie it to this PR, but I wanted it on record.

Deep sleep stays off: no sleep controller node on the J704 yet.

Unrelated to this PR, seen in the same boots: with #2325's AVE enabled, the J704 on 27.2 logs apple-ave ...: unsupported firmware (text CRC 24eb4625), since the driver knows only the 26A428 encoder firmware.

AI-assisted, tested by me on the hardware.

Conflicts: t8142-j813.dts, t8142-j815.dts (took integration: #1806 moved the board enables to t8142-j813-j815.dtsi and cluster_pwrgate there); dropped the now-redundant venc/ANE/JPEG enables from t8142-j813-j815.dtsi.

Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
Via the updated dtfill/t8142-media-default (#2325). No textual conflicts; t8142-j813-j815.dtsi: dropped the cluster_pwrgate copy integration added there (now in t8142.dtsi), t8142-j813-j815-sleep.dtsi comment points at t8142.dtsi.

Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
@RyanTheTide

Copy link
Copy Markdown
Contributor Author

Merged integration/aurora (d1b022d, via the updated #2325 branch); no textual conflicts, dropped the duplicate cluster_pwrgate integration added to t8142-j813-j815.dtsi (now in t8142.dtsi); t8142-j704/j813/j815 dtbs and cpuidle-apple.o build (J813/J815 differ from #2325 only by the implicit pmu status=okay).

@RyanTheTide
RyanTheTide marked this pull request as ready for review November 22, 2026 10:34
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants