[DjDeveloperr]

Runs the PMP on the M5 Pro (J714s) and tells it when the GPU is on, so the PMP votes DRAM (DCS) and fabric performance for GPU load as it does under macOS. Follows #1324 and #1326, which added the PMC, CLPC, PMP and PMP-report support for T6050 with everything disabled on the boards.

One commit, arm64: dts: apple: t6050-j714s: Run the PMP and report the GPU to it:

• t6050-peripherals.dtsi: the GPU's request in the PMP report (AGX, soc-device 16, so report@f), on the GPU's own power state (ps_gfx), as T6041 does. The PMP is told the GPU is on only while it is powered. A request with the GPU off crashes the T6041 PMP.

• t6050-j714s.dts: enables these together:

  • the PMP with its boot arguments (ADT /chosen board-id 8, dram-vendor-id 255, /arm-io/pmgr pmc 1, pmc-pmgr 0x3f), its mailbox, DART and power states;
  • the PMP report;
  • the PMC's agent 0 (apple,static-voters = <0 1 3>, apple,pmp so the PMC waits for the PMP);
  • sgx taking the report entry as its power domain.

  Because they are enabled together, the GPU can't wait on a disabled report. The CLPC keeps its vote; the PMC takes the higher of the two.

• t6050-j714s-pmp-tunables.dtsi (new): the 148 iop-pmp0-nub firmware properties from the J714sAP 27.0 ADT, raw bytes, as t8142-j813-pmp-tunables.dtsi does for J813.

Two follow-ups for the first dt-checks run, which found three new warnings on t6050-j714s.dtb:

• dt-bindings: soc: apple: pmp: Type the firmware's segment names as a string. The ADT value is a NUL-terminated string. A DTB has no types, so dt-validate decodes it as a string and the uint8-array apple,tunable- pattern rejected it, already on t8142-j813 in the base run. The property now has its own string type and is kept out of the bytes pattern. The DTB bytes don't change.
• arm64: dts: apple: t6050: Name the PMP DART's regions: "dart", "dapf", from the ADT dart-pmp0 instance list (DART, then DAPF; the node also carries dapf-instance-0). No driver looks them up by name.

Results on J714sAP

Default SSD boot with this DT (2026-11-09; the 7.3.0-m5-merge kernel with only the DTB from this commit; installed over SSH, then rolled back to the default 7.3.0-m5-argb):

• Boot and desktop: the set comes up over SSH and Hyprland runs. No oops/BUG/WARN/SError; error lines and deferred devices are the same as on the default set.
• PMP and GPU: apple_pmp: io.cpp:61: PMP started; the GPU binds through pmp-gfx (genpd gfx -> pmp-gfx, both on). apple-pmp-report 284240000.pmp-report: PMP running (status 0x1), requests 0x100004f000: MACC0/1, PACC, DISPINT, ANS and now AGX. SOC-DEV-PS-ACK reads back the same 0x4f000 / 0x10.
• DCS / fabric votes under gpubench (CLPC on; µs per submission, median):

gpubench	default set: PMP agent	with this DT: PMP agent (honoured)	CLPC	µs default / this DT
compute.heavy alu d8	iBoot's 5/1, not honoured	5 / 2	6 / 3	267.5 / 267.5
render.heavy fill 1080p d8	iBoot's 5/1, not honoured	5 / 3	6 / 3	126.1 / 125.8
compute.heavy bw d8	iBoot's 5/1, not honoured	5 / 2	6 / 3	489.2 / 488.0
copy.bw buf2buf d4	iBoot's 5/1, not honoured	5 / 2	6 / 3	489.2 / 488.3
compute.light d4	iBoot's 5/1, not honoured	5 / 1-3	6 / 3	40.2 / 34.7
submit.empty d1	iBoot's 5/1, not honoured	0 / 0	6 / 3	1.156 / 1.158
idle after the GPU work	iBoot's 5/1, not honoured	0 / 0	0 / 0

Timings and SMC power are unchanged, as expected: with both voting the PMC takes the CLPC's higher DCS 6.

PMP alone (RAM-boot kit, same DT content; the CLPC's PMC enable bits cleared for one test): the PMC honours the PMP's DCS 5 / fabric 3. Bandwidth-bound compute runs at 541.3 µs and 27.4 W, against 486.1 µs and 29.8 W with the CLPC's DCS 6. A 12-thread CPU memcpy gets no PMP vote: the PMP follows the GPU, not CPU memory traffic. Without the GPU's request the PMP acknowledges the other requests and drops iBoot's static vote to 0, but never votes for GPU load.

Testing

• All 11 T6050 board DTBs build with dtc; J714s has no new warnings (only the existing simple_bus_reg on pmp@284e00000).
• Locally at the head, with dtschema 2026.9 (niced, -j6):
  • make CHECK_DTBS=y apple/t6050-j714s.dtb apple/t8142-j813.dtb: no schema warnings on J714s; J813 drops its two segment-names lines and keeps only its existing ones.
  • make dt_binding_check DT_SCHEMA_FILES=soc/apple/apple,pmp.yaml: clean.
• The bindings cover every other property used: the report entry's power-domains, the PMP's board properties, and the PMC's apple,pmp.
• On the M5, booted from the internal SSD: the merge-retest kernel 7.3.0-m5-merge (which carries soc: apple: pmp: start the T6050 PMP on bare metal #1326's report driver), with only the DTB built from this commit, installed over SSH and then rolled back to the default set.

apple,dram-vendor-id is this unit's ADT value, like J813's; a loader that copies /chosen overrides it.

🤖 Generated with Claude Code

[DjDeveloperr]

Merging: validated on the M5 Pro (J714sAP) with CI green (dt-checks, host-tests).

Hardware evidence (2026-11-09):

• Default SSD boot with this DT (14:05-14:09, the 7.3.0-m5-merge kernel with only this commit's DTB, installed over SSH, then rolled back to 7.3.0-m5-argb):

  • The set boots to Hyprland. No oops/BUG/WARN/SError, and no new error lines or deferred devices.
  • The PMP starts. The GPU binds through pmp-gfx under gfx. PMP running (status 0x1), requests 0x100004f000 (now including AGX); SOC-DEV-PS-ACK reads back 0x4f000.
  • Under gpubench the PMP's vote follows the GPU and the PMC honours it: DCS 5 / fabric 2-3 on heavy work, 1-3 on light submissions, 0 on empty submissions and at idle. On the default set it sits at iBoot's unhonoured 5/1.
  • GPU timings and SMC power are unchanged (bandwidth 489.2 -> 488.0 µs, empty submissions 1.156 -> 1.158 µs), since the CLPC's DCS 6 stays the higher vote.

• PMP alone (RAM-boot kit, CLPC's PMC enables cleared for one test): the PMC honours the PMP's DCS 5 / fabric 3. Bandwidth-bound work runs at 541 µs and 27.4 W, against 486 µs and 29.8 W at the CLPC's DCS 6.

• All 11 T6050 board DTBs build, and J714s has no new dtc warnings.

• The first dt-checks run found three schema warnings from the newly enabled nodes, fixed on the branch:

  • the PMP DART's reg-names ("dart", "dapf", from the ADT instance list);
  • apple,tunable-segment-names, which the binding now types as a string, since the ADT value is a NUL-terminated string. This also clears J813's two existing warnings.

  Local CHECK_DTBS on J714s/J813 and dt_binding_check on apple,pmp.yaml are clean.

  The hardware run used 397dfec's DTB. The follow-ups add only reg-names, which no driver reads, and a binding change, so the evidence holds for the head.

The M5's default set stays 7.3.0-m5-argb; prev is the tested PMP-DT set.

🤖 Generated with Claude Code