Skip to content

arm64: dts: apple: t6050: describe the GPU on J714c, J873s, K114s, K116c - #2276

Draft
RyanTheTide wants to merge 4 commits into
integration/aurorafrom
spec-m5neo/t6050-gpu-dt
Draft

RyanTheTide wants to merge 4 commits into
integration/aurorafrom
spec-m5neo/t6050-gpu-dt

Conversation

@RyanTheTide

@RyanTheTide RyanTheTide commented Nov 19, 2026 •

Copy link
Copy Markdown
Contributor

Problem

J714c, J873s, K114s and K116c inherit the inventory GPU node from t6050-peripherals.dtsi (sgx, compatible apple,agx-t6050). drm/asahi only matches gpu,t6050 and apple,agx-g17s (drivers/gpu/drm/asahi/driver.rs), and the aura GPU handoff only takes apple,agx-g17s (main, kboot_gpu.c) or apple,agx-g17c under /soc@2300000000 (aura#281). So the GPU never binds on these four boards.

Change

  • J714c and K116c (M5 Max) get the J716c GPU description. It moves out of t6050-j716c-gpu-panel.dtsi unchanged into t6050-gpu-g17c.dtsi. A board with no GPU bus of its own defines T6050_GPU_BUS, and t6050-peripherals.dtsi then puts the GPU PMGR under /soc@2300000000, the same as on J716c. J714c already runs its PMP, so its GPU stays on the PMP's GFX report entry (pmp0_gfx), as before.
  • J873s and K114s (M5 Pro) get the K116s/K114c G17S description. Its shared part moves into t6050-gpu-g17s.dtsi, and the K114c and K116s files keep only their notes and their one board property.

Nothing in this is per board. The driver picks its configuration from the GPU ID registers (hw/t6050.rs config()). The loader fills the reservations and operating points from the ADT the Mac boots with, and it checks the firmware build before it enables anything. Until then everything stays disabled, so a loader that refuses leaves these boards as they are today.

Testing

  • DTBs for all 11 T6050 boards, built before and after with kbuild's cpp and dtc flags: J714s, J716c, J716s, J775c, J775d, K114c and K116s are byte-identical. Only J714c, J873s, K114s and K116c change. No dtc warnings.
  • dt-schema (dt-validate 2026.9, bindings of the base commit): K114s and J873s have no warnings before or after. J714c and K116c each gain the same four warnings that J716c already has today: the G17C block names its CS/AFR tables opp-table-gpu-cs/opp-table-gpu-afr, which the opp-v2 node-name pattern rejects. Renaming them would change the J716c DTB, so I left that for a follow-up. The loader finds them by phandle (apple,cs-opp, apple,afr-opp), not by name.

Still to run

None of the four boards has been booted with this. Owners, could you post a boot row? dmesg | grep -i -e asahi -e agx and the aura log line FDT: GPU: ....

Details

Per board:

Board Part (model string) Description used Loader path Kernel path
J714c M5 Max t6050-gpu-g17c.dtsi (J716c), GPU on pmp0_gfx aura#281 (/soc@2300000000/gpu@2302600000, apple,agx-g17c) apple,agx-g17s fallback, ID registers select CONFIG_MAX
K116c M5 Max t6050-gpu-g17c.dtsi (J716c) aura#281 same
J873s M5 Pro t6050-gpu-g17s.dtsi (K116s/K114c) aura main dt_set_gpu_t6050 (alias gpu, apple,agx-g17s, firmware digest) apple,agx-g17s, ID registers select CONFIG
K114s M5 Pro t6050-gpu-g17s.dtsi same same

What these boards still depend on is the firmware build, not the board. The kernel admits 26A428 (27.0) for G17S and G17C. 27.2 is in #1173/#1650 (G17S) and #2185 (G17C). aura main has digests for 26A428 and 26B8125 (G17S). An unknown build is refused and the GPU stays disabled.

Two things are shared per variant rather than read per part. The driver's M5 Pro/Max controller defaults and perf tables are the J714/J716C ones, applied on every part (#2080). Fused-off cores are covered by the per-variant core layout. Both are the same today for K116s/K114c (Pro tables) and J716c (Max tables).

On aura main, a G17C tree without a gpu alias (J714c, K116c after this change, and J716c today) skips the GPU handoff and the GPU stays disabled. aura#281 adds the path-based Max handoff.

dtc before/after, normalized diff (phandles resolved to paths): the J873s/K114s GPU nodes match K116s's minus its PMP link, and the J714c/K116c /soc@2300000000 subtree matches J716c's minus J716c's sleep properties (apple,sleep-ps-auto, GFX always-on from t6050-j716c-sleep-full.dtsi).

Assisted-by: LLM (Claude:Opus 5.5)

t6050-k114c-gpu.dtsi and t6050-k116s-gpu.dtsi describe the GPU in the same
G17S layout: the gpu alias, six loader-filled reservations, the GFX and GFX1
mailboxes in place of the inventory ASC nodes, the always-on GPX and AFR
states and the sgx override. Only one property differs per board (K114c
keeps GFX on across s2idle, K116s puts the GPU on its PMP report entry).

Move the common part into t6050-gpu-g17s.dtsi so other M5 Pro boards can use
it. Both boards keep their own notes and their one board property. The K116s
ADT reference values move to the file header unchanged.

No functional change: the K114c and K116s DTBs are byte-identical before and
after.

Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
The M5 Max GPU description in t6050-j716c-gpu-panel.dtsi (the GPU node under
/soc@2300000000, its mailboxes, reservations, operating point tables and
GPU power states) has nothing J716c-specific in it: every value is either
the SoC's or left zero for the loader to fill from the ADT.

Move that half into t6050-gpu-g17c.dtsi and include it from the J716c file,
which keeps the internal panel half. Let a board that has no GPU bus of its
own ask t6050-peripherals.dtsi for one with T6050_GPU_BUS, so the GPU PMGR
lands under /soc@2300000000 where the loader looks for it, as on J716c.

No functional change: no board defines T6050_GPU_BUS yet and the J716c DTB is
byte-identical before and after.

Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
These four boards inherit the inventory GPU node from t6050-peripherals.dtsi,
compatible "apple,agx-t6050". Neither drm/asahi (which matches "gpu,t6050"
and "apple,agx-g17s") nor the loader's GPU handoff (which looks for
"apple,agx-g17s", or "apple,agx-g17c" under /soc@2300000000) accepts it, so
the GPU never binds on them.

Describe the GPU as on the booted boards of the same part:
 - J714c and K116c (M5 Max): t6050-gpu-g17c.dtsi, as J716c, on its own
   /soc@2300000000 bus (T6050_GPU_BUS). J714c already runs its PMP, so its
   GPU stays on the PMP's GFX report entry as before.
 - J873s and K114s (M5 Pro): t6050-gpu-g17s.dtsi, as K116s and J714s.

Nothing here is per board: the driver picks its configuration from the GPU's
ID registers, the loader fills the reservations and operating points from
the ADT the Mac boots with and checks the firmware build before it enables
anything. Everything stays disabled until it does, so a loader that refuses
leaves these boards as they are now.

No other board's DTB changes. Not yet booted on any of the four.

Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
opp-v2-base.yaml allows one suffix after opp-table, so
opp-table-gpu-afr and opp-table-gpu-cs fail CHECK_DTBS. J716c already
carried those warnings; sharing the block with J714c and K116c would
add the same four on each board. Rename the nodes to opp-table-afr and
opp-table-cs. Nothing looks them up by name: drm/asahi and the loader
follow the GPU node's apple,afr-opp and apple,cs-opp phandles.

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

Copy link
Copy Markdown
Contributor Author

Blockers: 2.

The split itself is clean: rebuilt on current integration/aurora, J714s, J716s, J775c, J775d, K114c and K116s are byte-identical, J716c differs only by the two OPP table names, and the extra dtc warnings on the four boards are the unit-address ones kbuild already suppresses for J716c/K116s.

  1. dt-checks fails on the head for a real reason, not the runner: t6050-j873s: new absent DT path: /soc/gpu-coproc@2302e00000. t6050-gpu-g17s.dtsi:64-65 deletes &gfx_asc/&gfx1_asc, but arch/arm64/boot/dts/apple/adt-coverage/t6050-j873s.txt:1450-1451 still maps /arm-io/gfx1-asc and its nub to that node. Point both lines at /soc/mailbox@2302e00000 (where gfx1_mbox now lives on J873s), and map gfx-asc (1448) to /soc/mailbox@2302600000 while you are there.
  2. J873s, K114s and K116c get a GPU that the loader can bind but no GFX handling across s2idle. Every booted T6050 GPU has one: K114c sets apple,sleep-retained on &ps_gfx (t6050-k114c-gpu.dtsi:19) because genpd gating GFX under the live firmware hung the machine in pm_test=platform; J714s/K116s put the GPU on the PMP's GFX entry (t6050-k116s-gpu.dtsi:51); J716c keeps GFX apple,always-on (t6050-j716c-sleep-full.dtsi:4401). None of the three new boards runs the PMP, and t6050-gpu-g17s.dtsi / t6050-gpu-g17c.dtsi carry none of these. Since the live-firmware behaviour belongs to the GPU, not the board, put apple,sleep-retained on &ps_gfx in both shared files (K114c then drops its copy) and let a board with a running PMP override with pmp0_gfx as J714c already does.
  3. Not a blocker, but worth settling here: the same M5 Max GPU now has two descriptions. K114c (Max) uses the G17S layout and gets the GPU from aura main today; J714c and K116c (t6050-j714c.dts:260, t6050-k116c.dts:28) get the J716c G17C layout, which aura main skips (no gpu alias) until aura#281 lands. Pick one layout per part. If G17C is the target, say K114c moves with it; if not, give K116c the K114c layout so it works on today's loader.

HW row: on the head, K114s and J873s (aura main) and K116c (aura#281): FDT: GPU: handoff line, drm/asahi probe, a Vulkan smoke, then 3x s2idle with the GPU bound and the smoke again after resume; 0 GPU faults, 0 AVC, 0 failed units, no new kernel errors.

Assisted-by: LLM (Claude:Opus 5.5)

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.

1 participant