Skip to content

arm64: dts: apple: t6041: link the GPU to the CLPC - #2173

Closed
AxxzyWasTaken wants to merge 1 commit into
aurora-silicon:integration/aurorafrom
AxxzyWasTaken:pr/t6041-gpu-clpc-link
Closed

AxxzyWasTaken wants to merge 1 commit into
aurora-silicon:integration/aurorafrom
AxxzyWasTaken:pr/t6041-gpu-clpc-link

Conversation

@AxxzyWasTaken

@AxxzyWasTaken AxxzyWasTaken commented Nov 17, 2026 •

Copy link
Copy Markdown
Contributor

Depends on #1369 (draft until it lands). It doesn't need #1369's commits to apply.

The T6040 GPU node has apple,clpc = <&clpc>, but the T6041 one doesn't. So apple_clpc_gpu_get() finds no CLPC for the M4 Max GPU, and the GPU's work never reaches the CLPC. With the upstream DT, #1369's busy-share floor and light effort do nothing on T6041. This adds the link as t6040.dtsi has it.

Why it waits for #1369: T6041 uses the T6040 CLPC table today. With this link alone, the M4 Max GPU would get the M4 Pro's GPU floor (DCS 5 / fabric 2 while GPU work is in flight). That combination never ran on the Max. #1369 gives T6041 its own table.

All three T6041 boards (J575c, J614c, J616c) enable the PMC that the CLPC sits on, so the GPU won't wait for a CLPC that never probes.

Testing

  • My J616C (32-core M4 Max) has booted with this link since 30 Oct: the bring-up DTB adds it to the GPU node. It has run with clpc/g17: busy-share GPU memory floor from the DT #1369's code in my daily kernel. The DTB is mine; this commit was not booted as-is.
  • This branch: the J616c, J614c and J575c DTBs build. The decompiled GPU nodes have apple,clpc pointing at apple,t6041-clpc. No new dtc warnings with W=1 (30 before and after). I couldn't run dtbs_check (no dtschema here); the AGX binding already lists apple,clpc.
  • Not tested: the J614c and the J575c.

Assisted-by: LLM (Anthropic:Claude Opus 5.5)

@joske

joske commented Nov 18, 2026

Copy link
Copy Markdown
Contributor

Tested on the J614c (16-core CPU / 40-core GPU M4 Max), together with #1369, on top of #1664.

With this, the GPU is linked to the CLPC: 28a600000.gpu shows up as a consumer of the CLPC device, and the live tree has apple,clpc on gpu@288000000. It boots without new warnings, the desktop (GNOME) and video playback work, and the GPU firmware loads as before.

🤖 Generated with Claude Code

@RyanTheTide

Copy link
Copy Markdown
Contributor

Thanks. This is the same link t6040.dtsi already has, placed in the SoC dtsi, so every T6041 board gets it. We built the J575c, J614c and J616c trees at 0abc2f385e54 and at this head. All three build with the same warnings as before, and the decompiled GPU node now points at the T6041 CLPC.

The code looks good to us. Your plan to land it after #1369, so that the M4 Max never runs with the M4 Pro's GPU floor, makes sense. When you take it out of draft, a short row on the exact head with #1369 would complete it. On a J616c or J614c: the GPU listed as a CLPC consumer, the GPU floor engaged while glmark2 runs, and ten minutes without GPU faults. joske's J614c result already covers most of that.

@AxxzyWasTaken

Copy link
Copy Markdown
Contributor Author

The row on the exact heads, on my J616C (binned T6041).

Kernel: 7.3.0-j616c-s2rR = #2175 head 4c4ebec (on #1664 c767947, integration/aurora 391fac6) + #1369 head f4adbb5 + this PR's head 90ee582. Both merged without conflicts. I used the tree's own t6041-j616c.dtb (50c5b341). One-shot boot, 0 failed units.

  • GPU as a CLPC consumer: the GPU node has apple,clpc (phandle 0xca). At idle the CLPC status shows host_gpu busy: 5 permille, floor: 0x00000000, reason_bits 0x0.
  • GPU floor under glmark2: I ran glmark2-es2-wayland --off-screen --run-forever for 600 s (60 scenes) and sampled the status every 5 s. 113 of 116 samples had the GPU bit set in the vote:
    • mostly floor: 0x25, reason_bits 0x3 (CPU + GPU);
    • near the end floor: 0x2, reason_bits 0x2 (GPU alone), with host_gpu busy: 466 permille.
  • Faults: none in the ten minutes. 0 GPU fault, recovery, timeout or lost-device lines in dmesg before and after.

I'll take this out of draft once #1369 lands, so that the M4 Max never runs with the M4 Pro's GPU floor.

Disclosure: prepared with an AI tool (the commits carry Assisted-by: LLM); tested as described.

The T6040 GPU node names the CLPC with apple,clpc; the T6041 one does
not, so the CLPC driver never links the M4 Max GPU and the GPU's work
never reaches the CLPC. Add the link as t6040.dtsi has it. All three
T6041 boards (J575c, J614c, J616c) enable the PMC the CLPC sits on.

Signed-off-by: AxxzyWasTaken <axxzytte@gmail.com>
@AxxzyWasTaken

Copy link
Copy Markdown
Contributor Author

Closing as superseded by #1951 (commit 5f3bca941c97, already merged in integration/aurora), which adds apple,clpc = <&clpc>; to gpu@288000000 in t6041.dtsi.

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