Skip to content

drm/asahi: T8140 G17P on macOS 27.2 firmware - #1997

Draft
RyanTheTide wants to merge 4 commits into
integration/aurorafrom
fw272/g17p-26b123
Draft

RyanTheTide wants to merge 4 commits into
integration/aurorafrom
fw272/g17p-26b123

Conversation

@RyanTheTide

@RyanTheTide RyanTheTide commented Nov 15, 2026 •

Copy link
Copy Markdown
Contributor

Problem. drm/asahi knows only the 25G83 G17P image, so the Neo's GPU does not start on macOS 27.2 (26B123) firmware.

Change. Identify both the 25G83 and the 26B123 J700 G17P images from their normalized firmware TEXT. 25G83 stays on its existing graph; the exact 26B123 image selects the common 26B123 graph. Describe the J700 27.2 HW data, performance rows, register map, Region C, policy and status-config layouts. The Region C source is now a board data field rather than a chip-id check (same encoded bytes). The shared G17 26B123 InitData relocation from #1683 is carried with its original authorship.

Testing. arm64 Rust kernel build with drivers/gpu/drm/asahi/.kunitconfig; Asahi KUnit 337 tests, 335 passed, 2 skipped, 0 failed. New fw-id test covers both J700 builds and newest-first selection; new layout tests cover the root word, policy +0x60, status-config +0x8, shared-cluster +0x80, page totals and J700 Region C offsets. Merged with current integration/aurora (a duplicated J813_26B123 from the textual merge was removed).

Still to run.

  • Neo on 26B123: drm/asahi probe, Vulkan mustpass chunk subset unchanged against the 25G83 baseline.
  • With drm: add GPU cooling v2 and J700 thermal bindings #1065 (GPU cooling): drm: add GPU cooling v2 and J700 thermal bindings #1065 now treats the packed 26B123 graph as having no live power layout (ab10dc1). The J700 still probes on 26B123 and logs that it has no GPU cooling device. Merge-tested: drm: add GPU cooling v2 and J700 thermal bindings #1065 head 00e0a70 + this head merge with no conflicts; the probe path was read through on the merged tree. Neo row: on 26B123 with both PRs, drm/asahi probes, the log has the "no live power state" line, and there is no GPU cooling device. Cooling on 26B123 is a follow-up. The in-tree layout tables already place the inputs: RegionC max power at +0x88 (the common graph shifts +0x38 onward by four; see this PR's Region C test), and the PPM enable/budget at status config +0x128/+0x12c (unshifted by 26B123). It also needs build_m5 to keep the status-config pointer, V26B123.live_power set to that layout, and a Neo readback row.

Assisted-by: LLM (Claude:Opus 5.5)

sithglan and others added 4 commits November 16, 2026 00:37
The J813 runs the common 26A428 InitData graph, the J873's interface. macOS
27.2 (26B123) ships a new G17G build (RTKit 3514.41.1), which never
acknowledges that graph: the secondary's InitData times out.

The 26B123 interface differs from 26A428 in three places, the same three
as the J873's G18G build of this release:

- both roles expect the InitData root version word with bit 45 set,
  0x0c8b_6339_83a0_04c0;
- the policy grows by 0x60 and the status config by 8; the other 22
  allocations keep their sizes. Everything after the policy in the shared
  cluster moves by 0x80; the page total and the private cluster (0x6d8
  still packs to 0x700) stay as they are;
- policy fields +0x2f0c..+0x2f17 move by 0x10 and +0x2fb4..+0x6d9f by
  0x18, with a new word at +0x2f28; status-config fields +0x59c..+0x5b3
  move by 4 and +0x5b4 to the end by 8, with a new word at +0x5b8 and
  +0x59c left zero. The new words carry a new accelerator power-controller
  field, which the generic power and performance setup sets to 5000 and
  the G17G one leaves alone.

The HW data, the roots, the main configs, status B and region C keep
their field offsets.

Add the 26B123 profile and graph. The 26A428 encoders run unchanged and
their policy and status config are moved in place by bands; a nonzero
byte no band carries is refused. Compile-time checks keep the bands
ascending and inside the new sizes. The board for it is J813_26B123; no
configuration selects it yet, and the 26A428 encoding is unchanged.

Signed-off-by: Thomas Glanzmann <thomas@glanzmann.de>
Identify the current 25G83 and macOS 27.2 26B123 J700 images from both roles’ normalized TEXT. Keep 25G83 on its existing graph and select the common 26B123 graph only for the exact newer image.

Describe the J700 hardware data, performance rows, register map and Region C in the packed graph. Reuse the shared 27.2 policy and status-config relocation, with tests for both firmware identities and the changed layouts.

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
Resolve fwid.rs: keep the T8140 test next to the new T8142 one.
Both sides added J813_26B123 to g17/fw/m5/initdata.rs; keep one.

Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
encode_common26_board_region_c() picked the 25G76-era region C layout
by chip ID (0x8140 -> J700 G17P, else J813 G17G). Which layout a board
starts from follows its firmware generation, so carry it in the Board
data (region_c_source) instead of testing the chip ID. No change in
the encoded bytes for any board.

Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
RyanTheTide added a commit that referenced this pull request Nov 19, 2026
admit() refused the probe with EOPNOTSUPP when the GPU node has
#cooling-cells but the selected firmware has no live power layout. The
cooling device is optional: the firmware die limiter and the board's
SMC thermal zones protect the GPU without it, as they did before this
series. Refusing the probe turns a missing optional feature into no GPU.
Log it once and continue without a cooling device.

Decide from the firmware layout data actually selected, not from the
variant alone. The common layout names objects of the G17P and G16S
25G83 graphs. The packed M5-profile graphs (boot_m5) place RegionC and
the power state elsewhere and never attach, so a variant that inherits
live_power from V26A428 while its build selects a packed graph (the
J700 26B123 build in #1997) has no live power
state. Before, such a build passed admit() and then ran without a
cooling device and without a word in the log.

cooling_words() now reports an unbuilt power state as None, checked
before RegionC is resolved, and attach() warns and continues. A
malformed #cooling-cells, invalid mappings and a zero initialized
maximum still fail initialization.

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

2 participants