Skip to content

M3 Pro/Max: G15 as the T6030/T6031 GPU driver, and integration/#358 display and platform union - #148

Closed
DjDeveloperr wants to merge 49 commits into
omacom:integration/aurorafrom
aurora-silicon:m3-pro-rebase
Closed

DjDeveloperr wants to merge 49 commits into
omacom:integration/aurorafrom
aurora-silicon:m3-pro-rebase

Conversation

@DjDeveloperr

@DjDeveloperr DjDeveloperr commented Oct 19, 2026 •

Copy link
Copy Markdown
Collaborator

M3 Pro (J514S, T6030) and M3 Max (J516C, T6031) support on integration/aurora as a union of integration's work and aurora-silicon/linux#358 (Eryk Wieliczko) and #363 (Naeem Malik).

GPU: G15 is the T6030/T6031 driver

  • drivers/gpu/drm/asahi/g15/: Eryk's driver for the firmware the boot loader leaves loaded, plus Naeem's T6031/G15C row.
    • apple,agx-t6030 and apple,agx-t6031 probe straight into it, like g17/.
    • There is no parameter or Kconfig choice between GPU implementations.
  • Removed:
    • integration's T6030/T6031 runtime and "manager" backends (g15_boot/probe/initdata/selftest.rs, t6031_start/knobs.rs, hw/t6030.rs, hw/t6031.rs, the T6030/T6031 rows of the M3 runtime);
    • asahi.m3_backend, the asahi.t6031_* parameters, CONFIG_DRM_ASAHI_G15 and the GPU config fragment.
  • Kept and unchanged:
    • T8122 (J613/J615) keeps integration's M3 runtime and HAL200 runtime; G15 does not support it.
    • M1/M2 and G17P are unchanged. g17/ is byte-identical.
  • Userspace contract: matches aurora-silicon/mesa aurora-wip (52ecc30), which runs G15S/G15C on the native AGX3 path (asahi, hk: G15S (M3 Pro) on the native AGX3 path aurora-silicon/mesa#23, G14X compute preemption: qualify merged M2 Max default and broader hardware #25).
    • Integration UAPI numbering and layouts: render 448 B, compute 200 B, timestamp copies zero.
    • QUEUE_LIMITS 58/2.
    • The kernel-owned VA range [0x7000000000, 0x7100000000).
  • Integration glue:
    • Hooks on drm_gpu::DrmGpu/gpu::Gpu.
    • The crate root is renamed asahi.rs → asahi_core.rs, so g15/glue.o links into asahi.o.
    • The static gpu@292400000/gpu@40a400000 nodes are used.
    • The ADT comes through m1n1's phram export.
  • PMP: the T6030 AGX vote uses apple_pmp_set_device_power(). soc: apple: pmp: let a T6030 consumer vote without a readiness report makes that work without an apple,pmp-report phandle. T6031 needs no vote.
  • Review fixes:
    • a user-triggerable scissor-base overflow panic;
    • a group-allocation ENOMEM no longer stops the whole GPU;
    • closed VMs are released when their handle goes;
    • the power hold and vote are released on an early start failure;
    • only used pass slots are marked dirty.

Display: integration's stack plus #358/#363

Ported from #358:

  • HDMI audio (J514S DCPEXT0/ATC3, SIO channel 0x66):
    • the T6030 SIO binding and channel configuration, with APPLE_SIO_T6030 enabled in asahi.config;
    • the DPA1/SIO DT nodes;
    • the 14.7 external role's audio protocol.
  • DCPAV client: Open(4)/Close(5)/Terminate with EDID validation. Integration sent SetVirtualEDIDMode (6) and never closed.
  • 14.7 panel brightness: in the swap at +0x354/+0x35c; 500-nit maximum, 140-nit default, with a backlight class device.
  • Panel sleep: the panel mailbox is held through system sleep.
  • External link fixes:
    • inactive sinks are reported disconnected until the link is back;
    • drive levels the link asks for are programmed;
    • the AFK reply byte and re-announced AV service are handled;
    • dropped-swap replies, the A437 timing poll after power-on, D101 stride and D300 are handled.
  • Catalogs:
    • the fw14 RGB8 profile, DSC admission and 128 colour IDs in the parser;
    • link-budget filtering of external catalogs;
    • HDMI 2.1 catalogs larger than 1 MiB (8/24 MiB property limits).
  • HDMI converter: a 100 ms converter service runs while the sink is held, alongside integration's pulse fast path. J514S's HPD GPIO sits on the AOP.
  • Colour matrix: offered only where the firmware takes one.
  • KUnit: the present encoding is checked byte for byte; integration's encoder reproduces all of drm/apple: time-stamp T6030 swaps on 14.7 for 120 Hz #358's hardware-qualified CRCs. The external catalog profile and link budget have tests too.

Ported from #363: the J514C (M3 Max 14-inch) board record.

Kept from integration:

  • its handoff contract, callback handling and hotplug model;
  • refusing sleep while an external display is attached;
  • its native DP alt mode, through ATC Type-C and the crossbar, plus USB4 DP tunnels;
  • adoption of a DCPEXT left running.

Platform: integration's base plus #358

  • DRAM/fabric floor (DCS4/FAB2 for the CPU clients, T6030):
    • Lives in pmp-report.c, the T6030 PTD owner, and is applied after the PMP reports ready.
    • Indices come from apple,tunable-ptd-range, not constants.
    • On by default, with no parameter.
    • drm/apple: time-stamp T6030 swaps on 14.7 for 120 Hz #358 measured an 8-thread CPU copy at 28 → 123 GB/s and a Minecraft replay at 36.5 → 55 FPS. Integration had no floor on T6030.
  • PMP readiness: a non-blocking apple_pmp_report_check_ready().
  • pmgr-pwrstate: apple,no-auto-pm is honoured; apple,inherited-on is documented as integration uses it.
  • Display crossbar: drm/apple: time-stamp T6030 swaps on 14.7 for 120 Hz #358's exact register-image checks and runtime-PM hold are merged into integration's t6030_dpin handling, so USB4 DP-IN routes still work. drm/apple: time-stamp T6030 swaps on 14.7 for 120 Hz #358's DPPHY-only T6030 crossbar would have refused them.

Not ported:

Testing

  • Native build (LLVM 18, Rust 1.85), defconfig + asahi.config: Image modules dtbs passes (7.1.12-gee4f0f3ee295). The asahi crate has dead-code warnings only: 26 in integration's untouched T8122 g16_* runtime and 5 in m3_*/t8122_*, 16 from the g15/ PR files, and 11 from the legacy G15V14_8_3 instantiation that the versions macro still generates for the T8122 runtime's firmware structs (cleanup follow-up).
  • g17/ is byte-identical to integration/aurora.
  • The M3 runtime host tests (tools/asahi/m3-runtime-tests.rs) pass 41/41.
  • The T6030/T6031/T8122/T8140 DTBs compile.
  • Mesa side, under the drm-shim:
    • G17P submissions are byte-identical before and after the G15 merge.
    • G15S/G15C ran 200 CTS cases each with no submission refused by the kernel-gate mirror.
  • Not hardware-tested. None of this branch has been booted: GPU boot and the PMP vote, UAT adoption, T6031, the memory floor, HDMI audio, the DCPAV client change, drive levels, brightness and the crossbar checks. drm/apple: time-stamp T6030 swaps on 14.7 for 120 Hz #358/drm/asahi: enable J514C graphics on normal boots #363 were tested on J514S/J516C on aurora-wip; this union needs the same on integration.

🤖 Generated with Claude Code

erykwieliczko and others added 30 commits October 18, 2026 20:18
The M3 Pro (T6030) has the SIO coprocessor too; on the J514S it carries
the built-in HDMI port's audio (DPA1, channel 0x66), as the M1 Pro/Max
SIO does for their HDMI port.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: unchanged]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
…pects

The ADT records each SIO client's channel configuration with the client
(dp-audioN dma-channels: channel, data shape, timeout, FIFO size,
threshold). On the M3 Pro (T6030) the DisplayPort audio channels refill
their FIFO at 0x600 bytes: [0x66, 2, 0, 0x800, 0x600, 0, 0, 0] for DPA1,
the J514S HDMI port's, and the same for DPA0/DPA2. Configure that
threshold for an apple,t6030-sio controller, and publish every period
descriptor whole (its other fields zero), since a descriptor slot may
come back from an earlier transfer.

The boot tables stay the boot loader's (apple,sio-firmware-params), as
the binding has it. Behind APPLE_SIO_T6030, so that other configurations
build the driver unchanged.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: unchanged]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The G15 generation of the asahi GPU driver, added later in this series,
links page tables into the UAT of the GPU's firmware coprocessor, whose
table walker does not snoop the CPU caches. The tables are ordinary RAM
pages that the shared page-table code allocates and writes through the
cacheable linear map; they are not DMA-mapped, so the DMA API has nothing
to sync them with. Each table, and each changed entry, has to be cleaned
to the point of coherency before the firmware can walk it.

Export the existing dcache_clean_poc() to GPL modules for that, instead
of adding a new cache maintenance helper.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: applied unchanged from aurora-silicon/linux PR #358]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The SIO coprocessor (ASC at 0x289c00000, mailbox at +0x8000, interrupts
907..910, stream 0 of the SIO DART) is the DMA controller of the
DisplayPort audio interfaces. As on the other SoCs it stays disabled:
the boot loader reserves its firmware, describes its parameters and
enables it. Its firmware sets up the SPI, UART and audio DMA shims
whichever of them its clients use, so SPI_P, UART_P and AUDIO_P join SIO
as parents of the SIO CPU domain, as t8103, t8112, t600x and t602x
describe them.

The J514S built-in HDMI port's audio is DCPEXT0's audio service with
DPA1, the DisplayPort audio interface linked from DCPEXT0's port@0, whose
samples SIO channel 0x66 carries; DPA1 is in its own power domain. This
is the M1 Pro/Max HDMI port's arrangement (dcpext0, dpaudio1, SIO 0x66).
Enable the SIO mailbox, the alias the boot loader looks the SIO up by,
and DPA1's audio controller.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: the DPA1 audio controller, DCPEXT0's port@0
and the SIO alias and mailbox enable go into t603x-j514s-j516s-hdmi.dtsi
against integration's dcpext node (the PR's t6030-j514s-dcpext.dtsi is
not carried); the t6030.dtsi and t6030-pmgr.dtsi hunks are unchanged]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The J514S HDMI port's audio is DCPEXT0's DP/HDMI audio service (EP 0x29,
DCPAVAudioInterface) with DPA1, whose samples the SIO coprocessor
carries. On DCP firmware 14.7's external role the service numbers its
methods as 13.5 does, but differs from what the driver does for M1/M2:

- An element's ElementData, the link data every link method takes, is
  32 bytes, not 24; the last eight are opaque and must be kept.
- StopLink (0/10) and CompleteLink (0/11) carry the link data as
  PrepareLink and StartLink do, and every link method answers an IOReturn
  at body+48.
- One link per stream start: the SIO channel is configured, PrepareLink
  sent; the cyclic DMA is issued before StartLink; on stop the DMA is
  terminated and drained before StopLink and CompleteLink. A link that
  was only prepared (a format probe) is completed at hw_free, the link of
  a connection that has gone is dropped unsent. No DPA reset follows
  CompleteLink: it silenced every later stream on this machine.
- The catalog of a sink's audio elements can be large (114 elements on a
  TV): it is queried with a 256 KiB capacity.
- An open DCPAV interface is a client the firmware waits for, up to 10 s,
  before it removes a terminated interface, and the sink's link with it.
  So the card follows the connection but the service is not opened with
  it: the PCM opens it (Open, command 4, status and opened Boolean
  checked) and its close closes it (Close, command 5). A Terminate
  report on an open service disconnects the card and closes the instance
  from a work item; an instance that goes away takes its open with it,
  and its successor is never closed for it.
- The open, close and link calls wait for their replies without
  srv_rwsem and plug_lock, which the endpoint's receive path takes when
  an instance is published or removed: the instance the PCM opened is
  kept under a mutex of its own. When the card is disconnected for the
  display or the sink, an open PCM's link is completed on the still open
  service first; otherwise the firmware keeps it prepared and refuses
  every later PrepareLink.
- The role's CEC interface (DCPDPCECInterface) broadcasts the same
  reports on the endpoint; it is matched and its reports dropped.

The external role starts the audio endpoint on the driver's opt-in
(appledrm.hdmi_audio), as the other DCPs do, for a controller whose
port@0 links an audio device, once per session and after the sink's
first EDID: with the endpoint's interfaces published, every CopyEDID made
the firmware rebuild them and re-plug the sink. The audio service is
connected when the display is powered on (A472 on) and disconnected when
it is powered off and before the sink is released.

Everything is behind DRM_APPLE_DCP_T6030_EXT and the 14.7 external role:
other configurations build these files unchanged, and an M1/M2 DCP's
24-byte link data still goes out as the same 64 bytes.

Tested on the J514S HDMI port: PCM streams through the card, DPMS cycles
and firmware re-plugs with a PCM open.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: keyed to integration's native T6030
external pipe, dcp_is_v14_7_external() (DCP_FIRMWARE_V_14_7 with
external_native), in place of the PR's iomfb_is_v14_7_external(), and
gated by a new DRM_APPLE_DCP_T6030_AUDIO option (default y under
DRM_APPLE_AUDIO) in place of DRM_APPLE_DCP_T6030_EXT. The PR's
dcpext-t6030.c lifecycle hunks become: the AV endpoint is started from
dcp_get_modes() once the first EDID is installed
(dcp_external_audio_start(), external_audio_work in dcp.c, disabled
across unbind and system sleep); the audio service is connected after
A472 on and disconnected before A472 off in dcp_v14_external_modeset()
and iomfb_v14_7_poweroff(), and av_service_connect() waits for
iomfb_v14_7_external_powered(); integration's dcp_poweroff() and DPTX
release paths already disconnect the service. The av.c branches drop
the pinned service through the out label, as integration's service
pinning requires. The AV endpoint's RTKit messages were already routed
to dcp->avep by the DCPLink receive path.]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
Prepare the shared layers for a generation whose firmware the boot
loader leaves loaded, whose UAT regions are not System RAM and whose UAT
walker does not snoop the CPU caches (G15, the M3 Pro). Nothing changes
for G13, G14 or G17:

- gpu::Gpu gains default methods, each returning the current behaviour:
  user_range() (the GET_PARAMS VM range, by default the geometry's usable
  range), tracked_submission() (use the VM_BIND batch and SUBMIT
  output-fence paths that file.rs selects for G17 by type),
  soft_faults() (whether SOFT_FAULTS may follow fault_control),
  prepare_new_objects() (populate, clean and pin new user objects before
  any CPU view exists) and bind_batch_limit() (a bound on VM_BIND
  batches, None by default).
- file.rs consults them; gem.rs adds prepare_new_object(), which pins an
  object through its DMA mapping: on a device that is not DMA-coherent
  that cleans the freshly zeroed cacheable alias to the point of
  coherency.
- mmu.rs adds Uat::new_adopted() for a handoff the generation establishes
  itself. It maps the handoff and the TTBAT write-combined from physical
  ranges, names the firmware's kernel root as the upper half of user
  contexts, and keeps the kernel VM below a driver-owned shadow root whose
  window entries the generation links into the firmware's own table.
  Vm::with_tracked_status() attaches status and job lifetime tracking
  without the G17 context views, footprints and residency reclaim, and a
  few accessors serve the generation's VM handling and stream checks.
- Page tables of an adopted UAT clean every new table page before its
  parent names it, every changed table or leaf descriptor after the
  store, and the zeroed root at creation, and end each mutating walk with
  a DSB.
- regs.rs adds SGX accessors at runtime offsets.

The generation-only items are built only with CONFIG_DRM_ASAHI_G15,
which a later patch adds; without it the page-table cleaning is a
constant false, so G13/G14/G17 page-table writes are unchanged.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: the hooks live on drm_gpu::DrmGpu as well
as gpu::Gpu, since file.rs dispatches through the backend trait here;
file.rs recognizes G17 through AsahiData::g17() and tracked_submission()
selects the same paths for G15, including VM_STATUS_MIRROR; the
page-table cleaning reuses this tree's `noncoherent` table publication
(enable_clean_poc() sets it and publishes the root) instead of a
separate clean_poc walk; Uat::new_adopted() builds the integration
UatInner/UatConfig (published-roots config: no M3 owner, legacy PPL
lock, no handoff cache-flush requests) and Vm::new() enables the
cleaning when the UAT asks; the lifetime paths that integration routes
to the M3 owner for non-published VMs (install_driver_mappings,
driver_range_overlaps, unmap_user_ranges, drop_mappings_now) take the
lifetime path for a tracked VM; prepare_new_object() also marks the
object held, as this tree does for pinned objects]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
First part of the G15 generation (T6030, M3 Pro: J514S, J516S) of the
asahi driver: bringing the firmware the boot loader left loaded (RTKit
2419, the macOS 14.7-14.8 stub) to READY. The generation is built from
the patch that adds its UAPI glue and Kconfig option.

- adt: the boot loader's ADT copy (phram "adt"): the GPU regions and
  firmware segments, performance states, admission of their layout.
- identity: the exact firmware build, by the SHA-256 of its loaded TEXT
  with the GKTS entropy bytes zeroed.
- board: the three board templates (HardwareData, Globals, Power) of
  that build, verified by hash at boot, and the DVFS tables rebuilt from
  this machine's ADT.
- uat: the PPL handoff: the window mappings, the PPL lock exchange, the
  startup epoch, and the host window under a RAM shadow root whose L2
  entries are mirrored into the firmware's shared L2 page, with every
  table the firmware can walk cleaned to the point of coherency first.
- object: firmware objects on GEM kernel objects (populated and DMA-mapped
  before any write-combined access), the firmware cache classes, a bump
  allocator with guard pages, and the lower and client aliases.
- initdata: the 49-object initdata graph at its canonical VAs and its
  live words (root, runtime, control).
- channel, event: posting to the submit channels (also behind unconsumed
  records), the device-control ring, draining the receive rings and
  decoding the event ring.
- boot: probe-time admission (board, PMP supplier, DMA mask, ADT, firmware
  identity, MMIO, runtime-PM hold; a DMA-coherent GPU device is refused,
  as its firmware objects would not be cleaned), the RTKit operations on
  the C RTKit API (the only buffer accepted is the inherited one inside
  firmware DATA), the start to READY (PMP AGX vote, hardware identity and
  core masks, cold-only check, RTKit, UAT adoption, initdata, DEVCTL
  0x13 and 9), and an orphan state that retains and drains a session
  whose start failed.
- glue.c: C helpers without Rust bindings (SHA-256, the board check, the
  runtime-PM hold, the module reference, the phram lookup, ioremap_wc).

A firmware session cannot be stopped once started: everything it can
reach is retained until reboot.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
(cherry picked from commit 0154c24)
[port to integration/aurora: the PMP AGX vote goes through this tree's
PMP C bridge, apple_pmp_set_device_power(0x0f, 5, 1), which sends the
same request word the J514S bridge did; the supplier is linked with
apple_pmp_link_device() only when the GPU node names it with apple,pmp,
and a vote while no PMP is registered defers the probe; the board check
is of_machine_is_compatible(j514s|j516s) in glue.c instead of the PR's
pmp-display bridge; RTKit uses the GPU node's single unnamed mailbox,
as the static t6030 node has no mbox-names]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The G15 firmware is classic: one ring per engine (TA, FS, compute) and
ordering expressed by in-ring barrier records that wait on another
queue's stamp. All logical queues and VMs share one persistent compute
graph and one render graph:

- graph: the compute graph, the TA/FS render queues, notifiers and InitBM,
  class-2 queue-info rows, and the waited Globals +0x9b0 = 0 import before
  the queues are first exposed.
- slots: compute slots, render pass slots and clustering buffers, with
  their CLEAN/USED/DIRTY/CLEARING/STALE states scrubbed on a dedicated
  unbound, memory-reclaim workqueue.
- encode, fw/work: the compute, TA and FS work images and microsequence
  records of this firmware build.
- engine: commands of a job are published as prefixes (consecutive
  commands of one engine, or of both in API order), each member in a
  claimed slot, with per-edge barrier types, priming prefixes after a
  recovered abort, submit-ahead behind unconsumed records, compute
  horizons ahead, and at most two flights in flight (the second of the
  same VM). The order is total, which satisfies every valid header
  barrier and EXACT_PRIOR ordinal. Firmware-visible writes keep a fixed
  order: images, thresholds and counters, a barrier, the queue WRITEs, a
  barrier, then the compute/FS/TA records and doorbells.
- poll: the poll pass, under the device lock: drain the receive rings,
  retire flights on their stamps with trailing polls, answer
  parameter-buffer growth, check hangs, publish; a completion hrtimer
  with a one-jiffy backup; the asahi-g15/stats debugfs file and opt-in
  phase timers (debug_flags bit 49).

Fences are signalled after the device lock is dropped. Built with the
generation from the patch that adds its UAPI glue.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
(cherry picked from commit 0a2dcce)
[port to integration/aurora: applied unchanged]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
…licy

- recovery: deadlines for a published prefix without progress
  (g15_deadline_ms) and for a started compute (g15_compute_deadline_ms),
  the compute hang kill through Globals +0x1c/+0x24 and a posted DEVCTL
  0x13 with publication frozen, its waited disarm, the retirement verdict
  of a killed prefix, and the halt/resume handshake answering up to four
  further halts while draining (g15_recovery; 0 makes every halt
  fail-stop). Whatever cannot be recovered latches fail-stop: every VM
  reports EIO, and everything firmware-visible a failed command may still
  reference is retained until reboot.
- pools: the render UMA and parameter-buffer allocators with an uncached
  spare set, their initialisation and the takeover of the spares after an
  abort, and parameter-buffer growth answered from the poll pass with an
  epoch gate (K = 4096).
- power: the firmware idle policy initdata writes into Globals: standby
  2500 us instead of the template's 700 (with 700 us the GPU powers down
  between works about 100 us apart and pays about 300 us at the next
  pickup), minimum confidence 0.5, and the finite-delay branch instead of
  the predictive one, which makes 2-3 % of works under sustained low-duty
  submission pay a power-up.

The pure decisions (receipt comparisons, the compute hang test, the
growth ledger, the idle words) live in core-only submodules. Built with
the generation from the next patch.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
(cherry picked from commit f0f8523)
[port to integration/aurora: applied unchanged]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
apple_pmp_set_device_power() waits for the supplier's readiness report
before sending a power command. On the T6030 the PMP node names no
report (the display gate points the display processors at the report
entries, not the PMP), so apple_pmp_report_wait_supplier_ready() falls
to the T8122 wait and returns -ENODEV: no T6030 consumer could vote.

The G15 GPU generation votes for PMP device 5 (AGX) before it starts
the firmware, through this function. A registered T6030 PMP has
already started its firmware and endpoint (pmp.rs registers after
that), and apple_pmp_send_power_command() waits for the firmware's
acknowledgement, so treat the missing report as satisfied for the
T6030 PMP only. Every other supplier keeps the readiness requirement.

Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
…nk swap

The MacBook Pro (M3 Pro, J514S/J516S) internal panel runs DCP firmware
14.7 through Aurora's DCPLink session (iomfb_v14_7.c), which presented
frames but never set the panel level: integration had no backlight
device for it. The 14.7 swap request carries the level in nits inside
its 0x508-byte prefix: a present byte at +0x354 and the level as an
unaligned IEEE 754 binary64 at +0x35c, both zero when the level does not
change. The block sits in the opaque gap between the background colour
(+0x154) and the first surface record (+0x508) of integration's swap
layout (iomfb_v14_7_swap.h: DCP_V14_SWAP_BACKGROUND, DCP_V14_SWAP_SURFACE),
the same prefix aurora-silicon/linux PR #358 (Eryk Wieliczko) validated
for this panel (its static asserts place bl_nits_present at 0x354 and
bl_nits at 0x35c of its dcp_swap_v14_7).

Policy, as PR #358 has it for this 500-nit mini-LED panel: at most 500
nits (less if the panel's apple,max-brightness says so), 140 nits
presented by the first present after takeover, levels jump (no fade),
and the level goes out only when it changes. It runs on the H17P
backlight state machine (dcp_backlight.c): dcp_backlight_prepare()
reserves the level just before A407 and the swap's result completes it;
a level request kicks a brightness-only present that repeats the last
frame from backlight_work under the RPC lock; CRTC off blanks at 0 nits
(dcp_backlight_dpms) and the next present after CRTC on carries the
level again; a refused present is retried with the policy's backoff.
dcp_backlight.c admits the 14.7 panel next to H17P (a jump instead of a
fade at configuration, the level request path instead of the DAC
tables, no device without an active policy), and the panel's probe
initialises the policy. The backlight device registers from
bl_register_wq once the panel has started.

External 14.7 pipes encode the same zero level block as before: their
swaps are byte-for-byte unchanged, and nothing here runs for them, nor
for M1/M2 or H17P.

The level block encoding (dcp_v14_nits_binary64()), the policy
constants and the dcp_backlight.c changes follow PR #358's
iomfb_v14_7_present.h and dcp_backlight.c; the present path itself is
integration's.

Not yet tested on hardware.

Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
Complete the G15 generation (T6030, M3 Pro) with what userspace sees,
and build it behind the new DRM_ASAHI_G15 option (default n):

- vm: user VMs, each holding an exclusive TTBAT slot (ASIDs 1..63) for its
  lifetime, with the shared graph's client aliases mapped into every one,
  in VA ranges that belong to the kernel in every VM.
- queue, job, fence: logical queues own no firmware state. Each has one
  DRM scheduler with 16 credits; a SUBMIT is one job with one hardware
  fence, handed to the device-global engine, and fences signal in queue
  order with the VM status latched first. QUEUE_DESTROY drains the queue.
- command, gate: SUBMIT parsing and validation, in order: the stream's
  structure, the timestamp handles, the header barriers and EXACT_PRIOR
  ordinals (the semantics of the G17 rules; the engine's total order
  satisfies every valid one), every command's mapped spans with the right
  access, and the CDM/VDM stream gates that admit only known records
  (g15_stream_gate).
- timestamp: timestamp buffers mapped into the firmware's 64 MiB arena in
  the host window, where the firmware writes nanoseconds directly.
- params: GET_PARAMS (feature word 0x3e, QUEUE_LIMITS, identity), VM_BIND
  batches of at most 1 << 20 operations. SOFT_FAULTS, FEW_PRIMITIVES and
  RENDER_RAY_QUERY are never advertised; FRAGMENT_BARRIERS only with
  g15_early_tiling.
- mod: the generation's GPU, its shared state and device lock, and its
  platform driver for apple,agx-t6030. The driver is registered
  separately from the other generations and has no module alias: the M3
  must start its GPU only after the native display is up, so the firmware
  boots when userspace loads the module afterwards. It suppresses its
  bind attributes, and a started session keeps the module and everything
  the firmware can reach until reboot.

The module parameters g15_* carry the few tunables that stay tunable
(load time only): deadlines, the repoll and completion intervals, the
firmware idle standby time and confidence, recovery, the stream gates
and early tiling.

Tested on the J514S with Mesa's asahi Vulkan and Gallium drivers: a
Minecraft Vulkan replay runs at about 161 FPS with identical frames,
the deqp-vk memory model, compute, render and submit subsets show no
failures, and induced compute and render hangs are ended and recovered
without a firmware crash.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
(cherry picked from commit 13a93ec)
[port to integration/aurora: the generation is a backend of the shared
platform driver instead of a second platform driver without module
alias: asahi.m3_backend gains the value g15, and auto starts it on
the validated J514S/J516S boards when CONFIG_DRM_ASAHI_G15 is built
(asahi.m3_backend=runtime keeps the M3 runtime; without the option
auto keeps starting the runtime); AsahiDriver::probe_g15() keeps the
device bound with no DRM device when the probe fails after the
firmware was touched; drm_gpu::Backend::G15 adapts gpu::Gpu to the
DrmGpu file layer, and file.rs keeps the M3 runtime's per-SoC feature
rules off a G15 device; the command stream check requires the
ts_frag_end_copies/ts_end_copies arrays of this tree's UAPI layout to
be empty, as G17P does, since G15 implements no timestamp copies; the
crate root is renamed asahi_core.rs so that g15/glue.o can be linked
into asahi.o; the Kconfig text follows]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The T6031 GPU is the same firmware generation as the T6030's (RTKit 2419,
the same initdata graph and canonical VAs) on a different SoC: four
clusters of ten core slots, so the enabled-core mask is two words
(+0xe01500 and +0xe01504), identity words 0x07041100/0x0021040a, revision
0x12, and the GPU is powered by its power domain alone (there is no PMP
vote on this SoC). Its GPU node is in the kernel device tree, filled by
the boot loader (the UAT regions, the loaded firmware segments as fw-text
and fw-data with their VAs, and the operating points), so the regions do
not come from an ADT export.

Make the SoC-specific parts of the G15 driver a property of the
configuration row instead of module constants: the machine check and the
PMP vote (Platform), the source of the regions (the ADT export or the GPU
node), the initdata placement that differs between firmware builds
(Layout: root version, HardwareData VA, size, cached prefix and fourth
view, the IOMapping table), the board images (Board) and the compute
register values that scale with the cluster count (ComputeRegs). The
identity read fills one core mask per cluster, up to four, and GET_PARAMS
reports as many as the SoC has.

Add the T6031 row matched by apple,agx-t6031. It carries no firmware pin
and no board data yet, so admission prints the measured TEXT identity and
stops before anything firmware-visible. The T6030 row keeps every value it
had; its behaviour is unchanged.

Signed-off-by: Naeem Malik <malik2015naeem@gmail.com>
(cherry picked from commit 7dfc3c2)
[port to integration/aurora: the row is selected by the shared platform
driver (asahi.m3_backend=g15, or auto on the J516C when the generation
is built; every other value and board keeps the T6031 runtime
selection) instead of a second OF match table; both SoCs use the GPU
node's single unnamed mailbox; the T6030 row keeps this tree's PMP
vote through apple_pmp_set_device_power; the T6031 node is
integration's static t6031-gpu.dtsi, which the boot loader fills with
the same region and segment names this row reads]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The GPU firmware TEXT a J516C's macOS 14.7 boot chain leaves loaded
(iBoot-10151.140.19.700.2, RTKit-2419.140.12.release): TEXT 0x64000 and
DATA 0x140000 bytes, the GKTS entropy marker at 0x5cf28, SHA-256 of TEXT
with the 8 entropy bytes zeroed
c26ec89c1150eb6549cf602e012d0c5f49eb756e896d6bb8342cb2c5d98b332c,
measured on the loaded segment of a J516C.

The same machine started through a macOS 14.8.x chain carries a different
stub whose full digest has not been recorded. For that case, and for the
next build of either stub, add g15_accept_text_prefix: an unpinned build
whose digest starts with the given 8 bytes is admitted for that boot with
a warning that prints its full identity, so it can be pinned by the next
kernel without a second measurement boot. Off by default.

Signed-off-by: Naeem Malik <malik2015naeem@gmail.com>
(cherry picked from commit 9609d14)
[port to integration/aurora: the module parameter is added to the
renamed crate root asahi_core.rs]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The T6030 board templates do not fit the T6031: its HardwareData record is
0xb480 bytes (the fourth HardwareDataB view at +0xb400), its root version
word is 0x0c08e21e84820490, its IOMapping table has 17 slots at the T6031
register blocks (Fender at sgx+0xd00000 with a 0x144000 window, eight
0xd8000 memory-cache instances, the clock generator at sgx+0xe1c000, the
ANE doorbell at 0x3c945c000, ...), it has four per-cluster voltage tables,
and hundreds of its words differ from the two-cluster image. Everything
here was read on a J516C: the device tree the boot loader hands over, the
SGX registers and that machine's initdata.

Give the T6031 row its layout and IO table (VAs packed from
0xfffffc2068000000 as the T6030 table is) and take its three board images
from one firmware file, asahi/t6031-j516c-board.bin, pinned by SHA-256 and
built into the kernel with CONFIG_EXTRA_FIRMWARE: the images that
machine's firmware accepted, as the earlier J516C driver uploaded them.
The build checks the file's IOMapping records against the layout, then
writes what it owns: the TTBAT address, the IOMapping VAs, the timestamp
arena, the creation timestamp and the idle policy. The HardwareData
object is placed 16-byte aligned (0xfffffc20406fb580) and mapped
ASC-cached as a whole: the firmware's MTR alarm handler does a 64-bit
atomic in HardwareDataA, which faulted at the T6030 placement on T8122.

The T6030 path (templates, DVFS from the ADT, the S8 request words) is
unchanged.

Signed-off-by: Naeem Malik <malik2015naeem@gmail.com>
(cherry picked from commit 04e10f0)
[port to integration/aurora: applied unchanged]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The compute work image carries four cluster-sized register values that
the two-cluster G15S values get wrong on the four-cluster G15C, where
they fault the first job: the saved-state pointers into the preemption
buffer at +0x2980 instead of +0x1480 (0xd80 plus 0x700 per cluster), UMA
PTC (0xa599) 0x6000400020 instead of 0x13200400020, DUPM (0xd411)
0x3c000000140 instead of 0x200000001 (32 N | 96 N << 32 for N = 10 cores
per cluster) and 0x1a0e9 = 0 instead of 8. The J516C's own driver also
encodes a queue's first command with the saved-state pointers and the
0x12091/0x101d9 words zero; the G15C commands here are all encoded that
way, which is how compute work first completed on a J516C.

The values were read from 38 compute commands on a J516C and verified
by the register-fault analysis of the earlier J516C driver's boots. The
T6030 values are unchanged. The render (TA/FS) values for the G15C are
not known yet and stay the G15S ones.

Signed-off-by: Naeem Malik <malik2015naeem@gmail.com>
(cherry picked from commit 9174c1e)
[port to integration/aurora: applied unchanged]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
On the T6031 the boot loader describes the four UAT regions as
reserved-memory nodes without no-map (uat-ttbs, uat-pagetables,
uat-handoff, uat-pagetables-l2), so the kernel maps them and /proc/iomem
lists them as System RAM. The T6030 admission requires every region to be
outside System RAM (its regions are boot-loader carveouts) and maps them
write-combined, which refused the T6031 GPU at its first region ("GPU
region 0 (ttbs) ... not admissible", J516C boot t86) and would have
failed the write-combined remap of RAM next.

On the device-tree platform, admit a UAT region that lies entirely in
System RAM as well, and map such a region through the kernel's own
mapping (memremap MEMREMAP_WB) instead of write-combined, in the handoff
window and in the adopted UAT: what the shared driver does for the other
generations' handoff and TTBAT regions, and how the earlier J516C driver
reached the same regions. The firmware segments must still be outside
System RAM. Carveouts keep the write-combined mapping, so the T6030 path
is unchanged.

Signed-off-by: Naeem Malik <malik2015naeem@gmail.com>
(cherry picked from commit bfcfee0)
[port to integration/aurora: the ADT platform (T6030) also admits UAT regions
that lie in System RAM, since this tree's static t6030 GPU node has the
boot loader fill the same reserved-memory nodes and their no-map state
is not fixed in the kernel sources; firmware segments still must be
outside System RAM on both platforms]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The macOS 14.8 boot volume's stub loads a different J516C GPU firmware
build than the 14.7 chain: TEXT 0x64000, DATA 0x140000, GKTS 0x5cf28,
SHA-256 1bf2fd3476befe736bc2c070bad45f8be36b68cd11bcff16c9029d6bcda026cf,
first measured through g15_accept_text_prefix. The layout is identical
and nothing else in the T6031 configuration is keyed on the build, so it
is a second entry in the same table. The prefix escape stays.

Signed-off-by: Naeem Malik <malik2015naeem@gmail.com>
(cherry picked from commit e30d702)
[port to integration/aurora: applied unchanged]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
check_render() covers the 8 bytes at isp_scissor_base with scissor + 8.
The base is user input: 0xffff_ffff_ffff_fff8 passes the alignment check
and the addition overflows, which is a panic with RUST_OVERFLOW_CHECKS.
Use checked_add, as the other bases of the same check do, so that the
value is refused with EINVAL instead.

Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
…failure

publish_prefix() grows the compute horizon of a prefix by grouping later
compute-only jobs of the same queue, reserving one group entry per job
with ?. By then the base job's commands have been encoded and its queue
WRITE may have advanced, so publish() treats any error with sent != 0 as
an uncertain post-WRITE state and latches fail-stop: a transient ENOMEM
from that reserve stopped the whole engine.

Nothing of the grouped job has been touched when the reserve runs. End
the horizon instead and publish what was prepared; the job stays pending
for a later prefix.

Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
vm::Registry::prune() drops the registry's references to VMs whose
client handle is closed, but it only ran from reserve(), the next
VM_CREATE. A closed VM's page tables, every object bound in it and its
TTBAT slot therefore stayed alive until some client created another VM.

Prune when the handle goes: file.rs tells the backend through a new
DrmGpu::vm_closed() hook (a no-op for every other generation) from
VM_DESTROY, and a G15 queue prunes when it is destroyed, which also
covers a file close, where the VM handles drop before the queues. A VM's
state still lives until its last queue or job is gone, as before.

Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
…fails early

admit() takes the runtime-PM hold of the GFX domain and start() takes
the PMP AGX vote before the hardware identity, the core masks, the ASC
state, the TTBAT and the UAT self-check are examined. Each of those can
refuse the start before anything firmware-visible is written and before
the session is retained, and the probe then fails (or defers) with the
hold and the vote still taken, so a retry starts from a different state
than the first attempt and the GFX domain never powers down.

Give both back on those paths: the vote through a guard that releases
it unless the session is retained, the hold once Gpu::new() fails while
nothing is retained. A failure after the firmware was touched keeps both
with the retained session, as before.

Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
Slots::release() marks every pass slot in the prefix's pass mask Dirty,
whatever its state. A slot the prefix claimed but never wrote (a
publication that failed before its records went out) is still Clean,
and a slot already scrubbed is not dirty either; marking them Dirty
sends them through the scrubber again for nothing, and loses a state
the scrubber relies on. Guard the transition with the Used state, as
the compute and clustering slots of the same release are guarded by
their own states.

Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The G15 generation of the asahi driver drives the J514S GPU. Enable it in
j514s-rust.config, merged on top of j514s.config, and note in the latter
that userspace loads the GPU driver after the native display is up, as
its platform driver has no module alias.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
(cherry picked from commit 79d6375)
[port to integration/aurora: one fragment, j514s-gpu.config, merged on
top of defconfig and asahi.config; it enables DRM_ASAHI as a module
with DRM_ASAHI_G15 and APPLE_PMP, and documents the asahi.m3_backend
selection, since the generation is started by the shared platform
driver here rather than by a second driver without module alias]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The J514S HDMI audio path needs the T6030 SIO channel configuration.
APPLE_SIO_T6030 only changes behaviour for an apple,t6030-sio
controller, so enable it with the driver.

Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The G15 generation (g15/) is the driver of the T6030 (M3 Pro) and T6031
(M3 Max) GPUs. Until now it was one of several choices: a Kconfig symbol
(DRM_ASAHI_G15) decided whether it was built, and asahi.m3_backend chose
at probe time between it, the serialized M3 runtime, the G15 GpuManager
bring-up backend and nothing, with per-board defaults; a T6031 without
asahi.m3_backend=g15 went to the M3 runtime's asahi.t6031_start path.

Bind the generation directly: apple,agx-t6030 and apple,agx-t6031 get a
ProbeConfig::G15 row with their g15::hw configuration, the same way the
T8140 gets its G17 row, and probe starts probe_g15() for them. Build the
generation whenever DRM_ASAHI is built, as g17/ is, and move what its
Kconfig symbol carried onto DRM_ASAHI: the PMP driver dependency (the
AGX vote goes through apple-pmp's exported bridge, so asahi cannot be
built in while that driver is a module) and the firmware-loader
abstractions (the T6031 board images). The cfg gates in the shared files
go with the symbol.

Drop asahi.m3_backend and its T6030/T6031 selectors. The T8122 keeps its
board gate (the M3 runtime starts only on the M3 MacBook Airs), which no
longer consults the parameter. The J514S GPU config fragment only set
DRM_ASAHI_G15 on top of values asahi.config already carries (DRM_ASAHI=m,
APPLE_PMP=m), so it is removed.

The M3 runtime's T6030/T6031 tables and the GpuManager bring-up backend
are now unreachable; a following change removes them.

Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
… power-on

After every A472 on, the 14.7 external firmware enables the display's
timings asynchronously, and drops a swap that arrives before that: the
A407 reply carries status 0 with its first byte set, and no D589 ever
follows for that swap. The DCPLink client waited the full call timeout
for that completion and then stopped the session, so the first present
after a mode set or a DPMS wake could cost the display until reboot.

Poll the display status (A437 get_block, block 0x6e) after A472 on,
bounded at 2 s, until the pipe, link and timings are up, before the mode
set returns and the first swap goes out; and end a swap the firmware
reports dropped at once, with the old framebuffer kept on screen, the
way a refused swap is. A late D589 for such a swap completes nothing.

The default stride the firmware asks for (D101) follows the mode once
one is set, as the firmware expects, and the runtime levels it publishes
(D300: id 19 the panel level in nits, id 18 its correction) are recorded
for the debugfs status file.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: the PR's dcp_swapped() dropped-swap path,
 iomfb_v14_7_ext_wait_ready() and D101/D300 handling re-implemented on
 the DCPLink client (iomfb_v14_7.c, iomfb_v14_7_link.c) instead of the
 IOMFB template]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
… 1 MiB

An HDMI 2.1 (FRL) sink's TimingElements arrive from the external
processor as one chunked property transfer of more than 1 MiB (1.38 MB
for a 4K120 TV). The 14.7 client refused any property above 1 MiB, so
such a sink was described with no timings at all.

Let an external processor keep up to 8 MiB per raw property and 24 MiB
in all, with the chunked transfer's announced count bounded the same
way; the panel's properties keep their limits.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: the PR's DCP_PROPERTY_RAW_LIMIT/TOTAL for
 the external role, applied to the DCPLink client's raw property store]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
Once the T6030 PMP runs it owns the DRAM controller (DCS) and SoC fabric
performance states, and nothing asks it for more than their lowest
states: an 8-thread CPU copy measures 28 GB/s and the GPU is starved.
The PMP's CPU clients (EACC0/PACC0) are the first two entries of the
SOC-DEV-DVFS range of its device-state table, PTD indices 288 and 289
on the J514S; a request there with the valid bit (61) set asks for a
DCS state (bits 51:48) and a fabric state (bits 35:32), and the
DVFS-STATE entries 8 and 10 publish the firmware's response.

The PMP applies a CPU-client request on its change notification, and a
read of the request entry may consume that notification: a floor
written and read back at once stayed latched without effect. So, from
a delayed work once the PMP has reported ready and acknowledged its
startup requests: read both entries once, clear the floor fields (the
floor is always a fresh transition), write DCS4/FAB2 into both back to
back through the request-set aperture, and confirm on the DVFS-STATE
entries without reading the requests again; retry up to ten times, a
second apart, if the firmware did not respond. With it the DVFS-STATE
entries go from 0x100 to 0x2104 and the CPU copy bandwidth from 28 to
123 GB/s; a Minecraft Vulkan replay went from 36.5 to 55 FPS.

The floor is requested by the report driver that starts the T6030 PMP,
through the PTD aperture it already maps, on the apple,t6030-pmp-v2
report only. The entry indices come from the PMP's own
apple,tunable-ptd-range (the SOC-DEV-DVFS and DVFS-STATE ranges, looked
up by id and name and bounds-checked against the aperture) rather than
being hard-coded; without them no floor is requested and the PMP still
starts. The T6031 table's CPU-client entries have not been read, so the
T6031 report requests no floor.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: the memory floor of 6c43ba3 ("soc:
apple: pmp: Add the T6030 display hold, AGX vote and memory floor"),
moved from pmp-display.c into the PTD report driver that starts the
integration kernel's T6030 PMP, scheduled after the PMP is ready
instead of after the display's acknowledgement, with the entry indices
from apple,tunable-ptd-range and without the j514s_memory_floor module
parameter; a KUnit case covers the range lookup]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
apple_pmp_report_wait_ready() starts the PMP on the first call and
waits for it to acknowledge a report entry. A display driver that has
already asked for the PMP, or that must not start it from the context
it is in, has no way to ask whether the acknowledgement is there yet.

Add apple_pmp_report_check_ready(): the same lookup, but it never
starts the PMP and never waits. It returns 0 once the PMP reports ready
and has acknowledged the entry, -EAGAIN while it has not, the reason
the PMP was not started once that is known, and the same -EPROBE_DEFER,
-ENODEV and -EINVAL cases as the waiting call.

Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
erykwieliczko and others added 18 commits October 18, 2026 21:59
A domain marked apple,no-auto-pm never gets automatic power management
enabled: not when it is found on at probe, not when it is powered on,
and not when a failed power-off is undone. Once powered up it stays at
its requested state until Linux changes it, which a coprocessor CPU
domain whose firmware Linux starts cold needs.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
(cherry picked from commit b968acb12e5e6ea5a8f24cbe0a4e24ce3fc8d6d8)
[port to integration/aurora: applied onto the driver's undo_timeout and
retained_25g83 paths]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The T6030, T6031 and T8122 device trees mark the internal display power
chain (DISP_SYS, DISP_FE, DISP_CPU) with apple,inherited-on, and the
display handoff refuses to add the PMP unless the chain carries it; it
then sets apple,min-state on the chain, which the power state driver
applies, so the PMP cannot power the running display coprocessor down.
The binding did not describe the property.

Document it as the marker it is: the boot loader left the domain on.
It is not apple,firmware-owned; the domain's register is still written.

Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
… the crossbar's power

Each USB-C ATC of the T6030 (M3 Pro) has a display crossbar at its core
+ 0x4c000. The connection sequence of an output ends in a complete
register image, and a crossbar with no connection up shows an exact
idle image (the gates at +0x004/+0x014/+0x024 closed with their status
bits set, everything else clear), so a transition that did not take can
be told from the registers.

After every connection up or down of a T6030 output, check the
output's own image (its source bit in the write-side clock and enable
registers and their status, its bit in the output registers, its fields
in the select registers) and, once no connection is left up, the
crossbar's idle image; report a mismatch with the complete image. A
first connection is also expected to start from the idle image.

The crossbar's power (its ATC's USB domain, through power-domains) is
held through runtime PM from the first select until the last output has
been deselected and the crossbar verified idle; a crossbar that did not
reach the idle image keeps its power, as the hardware may have been
written. Probe writes nothing.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: the register image checks and the runtime
PM hold of a0bf9ef ("mux: apple-display-crossbar: Add the T6030
USB-C crossbars"), merged into the existing T6030 handling of the
apple,t6020-display-crossbar match, which also drives the USB4 DP IN
outputs; the connection sequences were already the same. No new
compatible: a mismatch is reported, not failed, and a route stays
selected while its connection is down, as the display driver and the
resume path of this tree expect]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
…e takes one

A 14.7 pipe sends the CRTC's colour matrix only on a board whose record
names a tested matrix setter (the J613/J615 A421/A420 pair); the T6030
and T6031 panels and every external processor send none. Their CRTCs
still offered the CTM property, so a compositor's night light was taken
and silently never applied, instead of falling back to its own shader.

Enable DRM colour management on a CRTC only where the pipe can apply a
matrix.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: the PR removes the CTM from every 14.7 CRTC
 (read-only j514s_ctm restores it); here the board record's matrix
 setter decides, so the J613/J615 CTM path is unchanged]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The 14.7 panel keeps scanning out through system sleep (soft DPMS blanks
it to the background at 0 nits), and its firmware keeps publishing to
the AP meanwhile. The mailbox interrupt is not suspended, so each of
those messages wakes a CPU out of suspend-to-idle, as the H17P panel's
did before dcp_hold_for_sleep().

Once the brightness work is cancelled and no call is in flight (calls
and notifications run under the RPC lock), hold the firmware's messages
in the mailbox at suspend and let them through again at resume. An
external processor is unchanged: sleep is still refused while it holds a
display.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: the PR gates dcp_hold_for_sleep() on
 dcp_presents_nits(); here the DCPLink client holds RX itself]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
… idle image only

The connection-up image of a T6030 output does not include the clock
status bits at +0x800, +0x808/+0x80c and +0x814/+0x818; only the idle
image does, where they are clear. Expecting them set after a connection
up would report a mismatch on every link up if the hardware does not
reflect the enables there. Check them clear after a connection down,
as the idle image does, and leave them out of the up image.

Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
…AV service

Two wire facts of the DCP 14.7 external role (T6030 DCPEXT), observed
with its AV service and DPTX port on the J514S:

- Its callback replies carry 8 in byte 35 of the EPIC sub-header. The
  AFK layer zeroed that field for every firmware.

- After a replug it announces its AV service again on the channel the
  old instance had. That is a new instance: the old one is gone with the
  sink it served. The AFK layer refused a second announcement on a
  channel that still had a service, so the host kept the dead instance,
  never opened the new one, and read no EDID through it.

Set the byte on the external role's callback replies, and tear the old
AV instance down before the new announcement is taken, on that role
only.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: gated on dcp_is_v14_7_external(), the
 native DCPLink external pipe; integration's handling of the 14.7
 kind-0 replies is kept]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
… for

While it trains a link, the DCP 14.7 external role asks the host to
program the voltage swing and pre-emphasis it chose for every lane
(SetDriveSettings: the lane count at +16, one level pair per lane from
+32), and reads them back (GetDriveSettings). The port handler kept two
words of the request and never touched the PHY, and answered the
read-back with lane 0's pre-emphasis as 0, so the link trained at the
PHY's default levels whatever the firmware had chosen.

On that role, program the levels into a running link through the PHY
(the T6030 ATC PHY takes them) and answer with what was programmed;
report a request the PHY refuses as failed instead of applied; record a
request made while the link is down. The M1/M2 port handler is
unchanged.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: dcpext_link_plan/reply's SET/GET drive
 cases re-implemented in dptxport_call() for the native 14.7 pipe]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
… is back

The DCP 14.7 external role reports a sink that is in standby or on
another input with DPTX callback 20 (InactiveSinkDetected), keeps its
display powered, and brings it back with a link-up once the sink is
active again: it retrains only a powered display's link. The port
handler only acknowledged the callback, so a TV in standby stayed a
connected output whose swaps went nowhere.

Keep the inactive state on the port. While a powered display's sink is
inactive, report its connector disconnected, hold swaps, and leave the
display powered when DRM turns the CRTC off; a link-up, a new catalog
or the sink's release ends the state. During link training the same
callback is routine and an unpowered display's connector is left alone.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: the PR's dcpext_link.inactive, carried on
 struct dptx_port and consumed by the DCPLink client's hotplug, swap
 admission and poweroff paths]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The DCP 14.7 external role answers EDID requests through an AV service
instance its client has opened. The host opened it with command 6, which
is SetVirtualEDIDMode: with the instance's interfaces present it removes
and recreates them, a software unplug of the sink, and it opens nothing.
DCPAV Open is command 4 (a zero 32-byte body; the reply's status at +0
and its opened Boolean at +16 say whether it took).

An open client is one the firmware waits for, up to 10 s, before it
removes the instance: it holds the reply to the sink's HPD low until the
instance's device has closed ("timeout while waiting for DCPDPDevice
close"), and with it the sink's release and the next sink. The host
never closed its client, so an unplug could leave the controller stuck.

Open the instance with DCPAV Open and check its reply; close the client
(DCPAV Close, command 5) before every HPD low and before a release, and
from a work item when the firmware terminates the instance (report group
2, command 1), after which the instance gets no new use. Close goes only
to the instance that was opened; a removed instance took its open with
it. The EDID read through this role is checked block by block (header
and checksums) before it is published.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: dcp_ext_av_work/dcp_ext_av_close and
 dcp_ext_v14_7_av_report re-implemented in dpavservep.c on the
 dcpavserv.opened bookkeeping, the Close hooked into dptxport_set_hpd()
 and dptxport_release_display(); drm_edid_valid() replaces the PR's
 dcpext_edid_object() checks]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
With the G15 generation bound to apple,agx-t6030 and apple,agx-t6031,
the M3 runtime's T6030 and T6031 support and the G15 GpuManager
bring-up backend cannot be reached. Remove them and keep what the T8122
(M3, J613/J615) runtime still uses.

Gone are the T6030 and T6031 SoC rows (m3_soc, hw/t6030, hw/t6031, the
T6030 firmware images, layout and IO-mapping tables), the T6031 start
(t6031_start, t6031_knobs and its asahi.t6031_* parameters, the
Experiment threaded through identification, InitData construction, the
device setup and the runtime's verdicts) and the T6030 GpuManager
backend: g15_probe and its g15_boot, g15_initdata and g15_selftest
modules, the Fender MMU/kick and self-test register helpers, the
firmware health report and gfx-data peek diagnostics, and the
asahi.g15_debug bits that only drove them. The G15 legacy GpuManager
instantiation that the versions macro generates stays, as the T8122
runtime builds its InitData from the same 14.8.3 structures.

The T8122 keeps its HwDataB words (now written out instead of derived
from the T6030 row), start experiment, HAL200 runtime and parameters;
the StopBeforeAsc and M3ResumeAfterFault debug bits stay. The former
T6030 image rows, firmware layout and IO maps live on only as fixtures
of the identification and IO-map host tests.

Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
…displays

The T6030 external processor's TimingElements went through the common
M1/M2 colour selection: no timing colour restrictions, the highest-score
RGB colour up to 10 bits (so a 10-bit colour over the 8-bit one the pipe
is qualified with), interlaced timings taken as progressive, and a
timing with more than 64 colour restriction IDs (an HDMI 2.1 FRL sink
lists 68) dropped.

Take that catalog through the external SDR path with a fw14 profile:
only an 8-bit SDR RGB colour (sRGB or default RGB) is preferred, the
highest score and then the lowest ID; interlaced, split, multi-pipe and
pixel-repeated timings are refused; the unsafe and DSC-required lists
are honoured, up to 128 IDs. A DSC-required 8-bit RGB colour the sink
takes compressed (SupportsDSC) is admitted for a timing that has no
uncompressed colour, never an unsafe one: the firmware compresses it
itself once the host sets the catalog's colour and timing. That is how a
4K TV's 3840x2160 at 100 and 120 Hz, listed with DSC-required colours
only, become available.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: the ctx flags are unconditional and set by
 the DCPLink client; a timing without an 8-bit RGB colour keeps
 integration's common fallback (YCbCr 4:4:4, any SDR, best) instead of
 being refused, so no sink integration admits today loses its modes]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The firmware publishes a sink's whole catalog; nothing checked that the
routed link could carry a timing. A two-lane pin-D route offered 4K60
(14.26 Gb/s over 8.64), which could only fail at the mode set.

Keep the timings the routed lanes carry at the port's highest rate:
uncompressed at 24 bpp within the lanes' 8b/10b payload, or compressed
by the firmware (a DSC timing at 8 bpp within the payload less the 2.4 %
FEC): four lanes of HBR2 (pin C) or HBR3 (the HDMI port) carry 4K120
with DSC, two of HBR2 (pin D) do not. When the firmware's preferred
timing goes, the largest and then fastest remaining one is preferred.
The lanes are what the PHY reported to the firmware for the route
(GetMaxLaneCount); the rate is the HBR3 every T6030 route reports.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: dcpext_catalog_filter() as
 dcp_catalog_budget() in the parser, applied by replace_modes() from the
 parse context; the PR's 3840x2160 and 24..120 Hz caps are not taken,
 as integration admits wider and faster sinks today]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The DP-to-HDMI converter behind the T6030 HDMI port raises IRQ_HPD
service requests the firmware services only when told
(InterruptRequestOccurred, DPTX group 8 command 9). The host forwarded
short HPD pulses, which only a non-sleeping HPD GPIO can time (the
J514S line is on the AOP and sleeps), and otherwise polled for 10 s after
an HBR3 link configuration. Anything later than that, or on an HBR/HBR2
link (a 1080p or 1440p sink), was never serviced: without the service
the firmware never learns of the converter's events, sink metadata and
FRL readiness among them.

Poll every 100 ms from the sink's connection (REQUEST_DISPLAY) until its
release, as the firmware needs; a request still goes out at once on an
HBR3 link configuration and on each timed pulse, and a firmware that
refuses three requests in a row is left alone until the next sink.

This is one small RPC every 100 ms while an HDMI sink is attached; the
earlier window had the same rate for 10 s.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: dcp_ext_service_arm/disarm as
 dcp_hdmi_sink_held() in dcp-hdmi.c, armed from the shared DPTX
 connect/release paths; the pulse fast path is kept]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The T6031 board record admitted the J516C only; on a J514C the DCP was
refused with "no board record for this machine" and the boot framebuffer
stayed. #363 admits the J514C alongside the J516C.

Add a J514C record on the T6031 DCP with the J516C image (the image is
the SoC's; its UUID is checked at probe). The J514C board file names
the J514S panel values (the same chassis), not yet read from a J514C, so
the record takes the panel size from the boot framebuffer like the
T6030 record does, and logs every panel timing and the display clock on
its first boots.

Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
A KUnit suite checks the 14.7 panel's A407 request encoding byte for
byte: the CRC32 of every representative request of the panel's
hardware-qualified profile (a primary frame, the first present at 140
nits, the soft-DPMS blank at 0 nits, a brightness-only re-present,
ARGB8888 at the 500-nit maximum, the background swap, a level of 1),
the wire landmarks of the surface and swap records, and the binary64
levels with their bounds.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: the same reference CRCs against the DCPLink
 client's encoder (iomfb_v14_7_swap.h), which reproduces every one of
 them; the template-specific policy and single-surface cases are not
 taken]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
A KUnit suite builds fw14-shaped TimingElements fixtures (derived from
two captured publications, a Philips monitor and a Sony TV: virtual
colours without a Score, no pipe counts, nested metadata) and checks
the parser's fw14 RGB8 profile (the 8-bit sRGB colour over 10-bit and
AdobeRGB, ties to the lower ID, interlaced, split, multi-pipe and
pixel-repeated variants refused, a DMT timing without a horizontal
precise rate admitted), the DSC admission of the TV's 2160p120 timing
and its budget on pin C, pin D and the HDMI port, and the link budget's
preferred-mode choice when the firmware's preferred timing goes.

Signed-off-by: Eryk Wieliczko <eryk@wieliczko.ninja>
[port to integration/aurora: the catalog cases of dcpext-t6030-test.c
 on dcp_catalog_budget() and the parse context flags; the budget
 expectations follow the ported filter, which has no 3840x2160 or
 24..120 Hz caps, and a case covers the kept common fallback]
Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
The M3 runtime no longer serves T6031, whose GPU is driven by the G15
generation, and t6031_knobs.rs is gone.

Assisted-by: LLM
Signed-off-by: Diljit Singh <hey@djdev.me>
@DjDeveloperr DjDeveloperr changed the title M3 Pro/Max: G15 GPU generation, J514S HDMI audio and panel brightness M3 Pro/Max: G15 as the T6030/T6031 GPU driver, and integration/#358 display and platform union Oct 19, 2026
@chriskearney

chriskearney commented Oct 22, 2026 •

Copy link
Copy Markdown

Thanks for putting this together, Dj. I’m working through the display/audio side and have a couple of SIO fixes worth sharing before the full boot tests finish.

  • APPLE_SIO needs select DMA_VIRTUAL_CHANNELS. With SIO built in and ADMAC off, the kernel fails to link on vchan_*; SIO built in with ADMAC modular also leaves the dependency modular. The usual m/m config hides this. We reproduced the link failure and checked the fix with both T6030 configurations.
  • With APPLE_SIO_T6030=n, the generic apple,sio match still probes the T6030 node, but sio_is_t6030() is false. That runs the earlier descriptor/threshold path on T6030. We now refuse that compatible at probe entry when its support is off, before changing the DMA mask or starting firmware; the M1 path still proceeds. The on/off KUnit tests cover both.

There’s also a loader pairing detail: the new sio alias lets older m1n1 enable this node, but aurora d05ba17d doesn’t export the protocol-10 tables from our m1n1 PR7. We added an early layout check so those older tables leave SIO unavailable cleanly, rather than sending them to the new path. HDMI audio should remain optional in that case.

These are source/build checks so far, not an HDMI-audio hardware pass. Happy to send the small fixes separately if that makes them easier to fold in.

One more reproducible issue in 375161e3’s DPAV client lifecycle: Open releases the state lock while its RPC is in flight. If HPD-low calls Close at that point, Close sees no opened client and returns; the successful Open reply then publishes the client anyway. Terminate during that same window is also ignored because opened != service. I replayed the actual helper bodies from this PR’s ee4f0f3e head: ordinary Open/Close passes, but both interleavings leave one Open and no Close. That’s a host regression test, not a hardware diagnosis. We’re fixing cancellation/instance tracking and checking delayed work against service replacement before carrying this part.

DjDeveloperr pushed a commit to aurora-silicon/linux that referenced this pull request Oct 23, 2026
Add the backend policy hooks, adopted UAT constructor, tracked VM status
and point-of-coherency page-table publication needed by native G15.
Port the public PR148/SiliconForest hooks onto this tree's newer GPUVM,
GEM ownership and DRM RegistrationData interfaces instead of restoring
the removed M3 runtime architecture.

Existing backends keep their user aperture, feature numbering, object
preparation, VM_BIND and SUBMIT paths. Their page-table constructors leave
clean_poc false. Preserve checked_object_size, current SG-table pinning,
G17 residency/context tracking and normal ioctl signatures.

The new feature_flags hook preserves the normal feature additions while
letting G15 report its private ABI's feature numbering. G15's private wire
facade and adapter are separate follow-ups; this commit does not qualify
an enabled native backend on hardware.

Based-on: Eryk Wieliczko's retained-generation shared hooks
Source: omacom/linux-aurora#148
Upstream-commit: 4f2edd6
Source: https://github.com/iconidentify/siliconforest-linux/commit/65cd54ba0da8f93e4530f851382ccb5187ab794c
Generated-by: LLM
Signed-off-by: Chris Kearney <303316+iconidentify@users.noreply.github.com>
(cherry picked from commit 0f36607)
DjDeveloperr pushed a commit to aurora-silicon/linux that referenced this pull request Oct 23, 2026
Keep G15 in the maintained asahi module with shared GEM, MMU, file and
queue infrastructure. Reuse the target RegistrationData API and preserve
normal/G17 registration and teardown. No old M3 runtime files or global
UAPI header selection are restored.

G15 retains its separate no-alias platform adapter and suppresses manual
bind attributes. When the default-off DRM_ASAHI_G15 option is enabled,
remove only T6030/T6031 from the generic adapter's exported table so
module autoload cannot start firmware before native display setup.

Adapt native objects to the target's permanently pinned, GEM-owned SG
table and VMap interfaces. Retain a GEM owner in HeldPages rather than
reintroducing the removed owned-SG wrapper. G15 register mappings are
owned by its retained session; normal Resources keeps its target devres
lifetime. Use only the maintained PMP link/vote API with the exact
T6030 plus J514S/J516S and GPU-compatible checks.

The original source imports/carries precede this adaptation, with author
and supplied signoffs retained. g15/SOURCES.json records the public pins
and carry ledger. Original work is by Eryk Wieliczko, Naeem Malik and
DjDeveloperr; this commit adapts their public code to the current APIs.
Private native-wire integration and the lazy driver-data initialization
fix follow separately; this is not hardware qualification.

Source: omacom/linux-aurora#148
Source: https://github.com/iconidentify/siliconforest-linux/commit/65cd54ba0da8f93e4530f851382ccb5187ab794c
Generated-by: LLM
Signed-off-by: Chris Kearney <303316+iconidentify@users.noreply.github.com>
(cherry picked from commit 3f80156)
DjDeveloperr pushed a commit to aurora-silicon/linux that referenced this pull request Oct 23, 2026
Registering a platform driver may call its probes and still fail later
in driver_add_groups. A second driver registration can also fail after
the first adapter probed. Starting retained G15 firmware in either
registration therefore allowed module-init rollback to detach its power
domain, despite retained DMA objects.

Keep native probes deferred while MODULE_READY is false. Initialize the
cleanup queue and both owned registrations first, then publish readiness
in the final infallible unit initializer. Retry the G15 adapter through
driver_attach; an attach failure is logged and never returned as a
module-init error because an earlier device may already have started.
All later ordinary/deferred/built-in probes observe the same ready gate.

The reprobe helper borrows driver_find's pointer only while the pinned
registration remains owned by DriverModule during initialization. The
API supplies no owned driver reference and no put_driver is performed.
Normal/G17 initialization, unbind and default-off behavior are unchanged.

Extend the actual-pin-init host proof to nine cases, including a late
failure after attempted probing in each registration, and attach failure
after successful startup. The actual C activation helper passes three
cases. Four mutants build and fail at their intended runtime assertions:
eager probe, retained error escaping, missing ready gate and activation
before the last registration. Allocator/registration/GPU side effects are
models; the macro, probe, module initializer, callback dispatch,
set_drvdata and C reprobe bodies are production source. No hardware claim.

Source: omacom/linux-aurora#148
Generated-by: LLM
Signed-off-by: Chris Kearney <303316+iconidentify@users.noreply.github.com>
(cherry picked from commit cbb678e)
@novuon

novuon commented Oct 23, 2026

Copy link
Copy Markdown

J514S / M3 Pro hardware test (2026-10-23)

I tested this on a 14-inch MacBook Pro J514S / M3 Pro (T6030). On one successful boot, a Wayland EGL client rendered on Apple M3 Pro (G15S B1). EGL swap passed with GL error 0.

At the same time, the built-in display ran at 3024×1964 at 120 Hz, and my LG 38WN95C ran at 3840×1600 at 119.982 Hz. Brightness control was available.

This did not test PR #148’s branch. The kernel was 7.2.11-J514S-native-g15, based on Omacom integration/aurora commit 42672ed79dead511ecf65e96d988a98a0f94b5bb (the #177 merge). The test also used a private Mesa runtime and a temporary m1n1 GPU-on handoff.

This was one successful boot. Another boot had a DCP failure and no external display, so cold-boot reliability and sleep/wake remain unverified. I have the relevant logs and exact build identifiers available if they would help.

@DjDeveloperr

Copy link
Copy Markdown
Collaborator Author

Closing: this conflicts with integration/aurora now, which carries the G15 M3 work in another form. The T6030 display and PMP commits that aren't upstream yet can come back as a fresh PR against the current tree if they're still needed.

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.

5 participants