Repository navigation
Conversation
|
Tested on a J604 (MacBook Pro 14", M4, DCP firmware 26.6.2) at With no With |
|
No-regression row from a 13" M4 MacBook Air (J713, Mac16,12) on the macOS 27.2 (26B123) stub. This covers the DART side only, not the display or GPU. Setup: kernel 7.3.0-bj713+ = integration/aurora 0abc2f3 + this PR (3f59f0e) + linux#2205, booted with the image's default command line (no Result: 6 boots (one first boot plus 5 warm reboots). Every boot had 0 Not covered: the display and GPU paths this PR is mainly for. On 27.2 the internal DCP is still refused (firmware-27 path needs Assisted-by: LLM (Claude:Opus 5.5) |
|
J715 row for this change (MacBook Air 15-inch, M4, Mac16,13; requested in #2663). Build
Result: pass
|
|
Blockers: 2.
Also: the commit message says kexec and kdump are unavailable "unless both are set to 0". The radio DART ( Rows still needed: J713 on 26.6.2 at this head with the default command line; s2idle/resume with the adopted DARTs (integration now carries 870eb63 on that path); and the 27.2 J713/J715 probe log lines for both internal display DARTs. |
3f59f0e to
28ae172
Compare
|
Note 🤖 Gemini 4 Neon responding on behalf of erikwb Addressed all review items in
Verified with arm64 |
|
Re-checked at 28ae172.
Nit: the dev_err_probe at apple-dart.c:5393 still says "retained J713 display"; it applies to every M4 Mac. Still needed at this head, as the body lists: a J713 26.6.2 boot with the default command line (native display and GPU), s2idle/resume with the adopted DARTs, and the 27.2 J713/J715 probe lines showing both internal display DARTs left untouched. Assisted-by: LLM (Claude:Opus 5.5) |
|
Merge-train adversarial review (MERGETRAIN-LINUX2) at 28ae172: BLOCKERS. Not merging until the blockers below are fixed or answered; the details have file:line and a failure case. Full review#2189 iommu: apple-dart: bring up the M4 internal display and GPU without boot parameters @28ae1721VERDICT: BLOCKERS
Nits: Missing rows:
|
DjDeveloperr
left a comment
There was a problem hiding this comment.
Review at 28ae17213fee: changes requested.
CI is green (7/7), commit attribution is clean (Signed-off-by: only), removing the j713_dcp/j713_display module parameters is the right direction, and the Kconfig/kexec text updates look good.
Blockers:
- Firmware-version gate on the default boot path (§1) & scope mismatch (
drivers/iommu/apple-dart.c:5379-5386,drivers/iommu/apple-dart-dcp.h:1041):apple_dart_probe()gates default internal display DART adoption onapple_dart_dcp_firmware(owner)(apple,firmware-version == <26 6 2>), right abovej713_dcpext/j713_dispextwhich already gate on a loader-set DT handoff property (apple,t8132-dcpext-firmware-ready). Per pre-PR checklist §1:- Either replace the version compare with a loader-set handoff/capability property (landing the Aura side first with 26.6.2 fallback per §1 chicken-and-egg), or
- If
apple_dart_dcp_firmware()is kept strictly as a temporary hardware-safety guard matchingapple_dart_dcp_validate()until the loader property lands, narrow the PR title and commit/body to 26.6.2 only — on the fleet's 27.2 firmware stub this change leaves both internal display DARTs untouched and brings up neither display nor GPU.
- Watchdog precondition rationale (§3): The removed module parameters existed because "userspace must arm its watchdog before probing SID23". Removing them makes SID23 adoption unconditional at early boot; please document in the commit message why the watchdog precondition no longer applies (so a regression doesn't silently wedge boot).
Non-blocking nits:
drivers/iommu/apple-dart.c:5389:dev_err_probemessage still says"retained J713 display"rather than T8132.drivers/iommu/Kconfig:APPLE_DART_J713_DCPhelp says "Say Y" withoutdefault y(relies onasahi.config).
Hardware rows still needed at the final head (§5):
- All existing J713/J604/J715 runs are from the earlier
3f59f0e1head before theof_find_node_by_path("dcp")probe-gate rewrite and param removal. Please run on this head: J713 26.6.2 default-cmdline boot (with 0 SELinux denials, 0 failed units, 0 new kernel errors),s2idle/resume with adopted DARTs onintegration/aurora(with870eb638), and 27.2 J713/J715 probe logs confirming"Leaving inherited display DART untouched".
Assisted-by: LLM (Gemini 4 Neon)
28ae172 to
552b5c8
Compare
|
Note 🤖 Gemini 4 Neon responding on behalf of erikwb Addressed both blockers (and the
Verified with arm64 |
|
Re-read at 552b5c8. Scoping this to the 26.6.2 firmware and keeping Assisted-by: LLM (Claude:Opus 5.5) |
|
J715 row at Build
Result: pass (one boot)
Not covered:
I can't explain one line yet. At 21 s the DCP logs brightness 0 and |
…6.6.2 firmware On T8132 (M4) Macs running 26.6.2 firmware, the internal display DARTs that iBoot leaves live (the DCP's SID23 at 0x30a340000 and the scanout/PIODMA frontend at 0x30a300000) were adopted only when the command line set apple_dart.j713_dcp=1 and apple_dart.j713_display=1. No distribution sets them by default, so an installed 26.6.2 system came up on simpledrm: appledrm never bound the DCP, and the GPU never started. Gate internal display DART adoption directly on the internal DCP devicetree node instead of the board/OS-version helper or opt-in module parameters: require that the "dcp" node is available, compatible with "apple,t8132-dcp", and carries apple,firmware-version 26.6.2 via apple_dart_dcp_firmware(). That check stays strictly as a hardware-safety guard matching apple_dart_dcp_validate() until a loader-published internal DCP handoff property lands in Aura first and the kernel keys on it (tracked in aurora-silicon#2621): the retained-DART path assumes the inherited table and DMA layout of the 26.6.2 DCP firmware, so any other firmware (including 27.2, which is out of scope here) or disabled DCP leaves both internal display DARTs untouched (-ENODEV) without pinning their tables, reserving the SID23 window, or resetting the live DARTs. Remove the runtime-writable apple_dart.j713_dcp and apple_dart.j713_display module parameters, and update the APPLE_DART_J713_DCP Kconfig help text to reflect that it is no longer an opt-in bring-up profile. Those parameters originally guarded early READY bring-up with a requirement that userspace arm a watchdog before probing SID23, when an unvalidated probe risked resetting or corrupting a live locked display DART and wedging the SoC. That precondition no longer applies to the adopted path: - probe performs no DART register or page-table writes: preserve_display skips apple_dart_hw_reset() and apple_dart_drain_stale_exceptions(), and apple_dart_dcp_import() validates the locked T8110 registers (TCR, ENABLE_STREAMS, TTBR) and inherited root/leaf PTEs read-only against ADT reserved-memory bounds and excluded Linux RAM, refusing any mismatch or instability without mutating hardware; - probe refuses adoption (-EIO) if any display DART fault is already pending or masked, attaches the shared fault IRQ only once both display DARTs pass validation, quarantines the faulting DART's borrowed streams on fault without issuing FLUSH_UNLOCK, and quarantines both display peers before releasing the shared IRQ; - a refused DART leaves DCP probe refused (via the IOMMU-group check in drm/apple) so simpledrm keeps the panel safely. Adopting the display DARTs holds the display kexec blocker for the rest of the boot, as it already did with the parameters set; on T8132 Macs the managed radio DART also holds its own kexec blocker. Link: aurora-silicon/aura#462 Link: aurora-silicon#2621 Signed-off-by: Erik Bourget <erik@ebourget.net>
552b5c8 to
0f35583
Compare
|
Note 🤖 Gemini 4 Neon responding on behalf of erikwb Rebased onto
J713 26.6.2 boot + |
Problem
On T8132 (M4) Macs running 26.6.2 firmware, a freshly installed system comes up on
simpledrmonly: no native display, and the GPU never starts. The kernel adopted the internal display DARTs that iBoot leaves live (0x30a340000for DCP SID23 and0x30a300000for scanout/PIODMA SID0/SID16) only when the command line setapple_dart.j713_dcp=1 apple_dart.j713_display=1, and no distribution sets them by default (J713 report: #1704; J604 behaves the same; distro workarounds in aurora-silicon/alarm-aurora#307 and aurora-silicon/fedora-aurora#348).Change
apple_dart_probe()directly on the"dcp"devicetree node instead ofapple_t8132_firmware_path()or opt-in module parameters: require that"dcp"is available, compatible with"apple,t8132-dcp", and carriesapple,firmware-version<26 6 2>viaapple_dart_dcp_firmware(). Because Aura does not yet publish an internal DCP handoff property (and per §1 chicken-and-egg the loader side must land first),apple_dart_dcp_firmware()stays strictly as a temporary hardware-safety guard matchingapple_dart_dcp_validate(), with the Aura handoff property and kernel transition tracked in kboot: T8132: publish an internal DCP handoff property so the kernel can drop the 26.6.2 firmware-version check aura#462 and T8132 display: key the DART and DCP firmware-layout choices on each DCP's handoff, not on version/address/board checks #2621: the retained-DART path assumes the inherited table and DMA layout of the 26.6.2 DCP firmware, so any other firmware (including the 27.2 fleet stub, which is out of scope for this PR) or disabled DCP leaves both internal display DARTs untouched (-ENODEV) without pinning tables, reserving the SID23 window, or resetting the live DARTs.apple_dart.j713_dcpandapple_dart.j713_displaymodule parameters and their-ENODEVopt-out branch. Those parameters originated in initial READY bring-up (52d846f6b8d3), when probing a live locked SID23 DART risked resetting or corrupting hardware state and wedging the SoC unless userspace armed a watchdog first. That precondition no longer applies to the adopted path:dart->preserve_displayskipsapple_dart_hw_reset()andapple_dart_drain_stale_exceptions(), andapple_dart_dcp_import()validates the locked T8110 registers (TCR,ENABLE_STREAMS,TTBR) and inherited root/leaf PTEs read-only against ADT reserved-memory bounds and excluded Linux RAM, refusing any mismatch or instability without mutating hardware.-EIO) if any fault is already pending or masked (apple_dart_fault_pending()/DART_T8110_ERROR_MASK), attaches the shared fault IRQ only once both display DARTs are admitted, quarantines the faulting DART's borrowed streams on fault without issuingFLUSH_UNLOCK, and quarantines both display peers before releasing the shared IRQ.drm/applerefuses DCP probe on the missing IOMMU group (beaae8e35b23), keepingsimpledrmactive without faulting.APPLE_DART_J713_DCPKconfig prompt and help text to describe retained T8132 display DART support rather than an opt-in bring-up profile to leave disabled.0x30a300000,0x30a340000) on 26.6.2 firmware. External display DARTs (j713_dcpext,j713_dispext), 27.2 boots, and non-T8132 Macs are unchanged.Testing
At head
0f35583d8414(rebased ontointegration/auroraa9626a1487ad; code diff identical to552b5c89f5e9):LLVM=1 W=1build ofdrivers/iommu/apple-dart.owithCONFIG_APPLE_DART=y+CONFIG_APPLE_DART_J713_DCP=y(0 errors, 0 warnings).tools/testing/m4-dcp-errors/module_guard.py: PASS.run-host-tests.py):hardware-defaults-tests(1/1 PASS) anddartsuites (16/16 PASS).scripts/checkpatch.pl --git origin/integration/aurora..HEAD: 0 errors, 0 warnings.Hardware runs:
Mac16,13, MacBook Air 15-inch M4, OS 26.6.2, system 27.2, tested by @kazuhidet at552b5c89, identical code diff):integration/aurora0abc2f385+ this PR's change (stage 1 -> aura, default command line with noapple_dart.*args and/sys/module/apple_dart/parameters/absent): PASS (one boot).30a340000"Preserving DCP SID23 root ...; no writes",30a300000"Preserving DCP SID0/SID16 root ..." and "Inherited SID0/SID16 frontend admitted"), 4 external display DARTs untouched, 0 DART faults.appledrm26.6.2: "KMS scanout activated" at 5.46 s (2880x1864 oneDP-1), GPUaurora-g16READY at 5.96 s,sddm+ desktop + keyboard/trackpad normal; failed units and kernel errors unchanged vs baseline (no SELinux on install; s2idle/resume and 27.2 OS firmware not covered).3f59f0e1cbd9, default command line on 26.6.2):Mac16,12, 26.6.2): SID23 root preserved with no writes,appledrmbound the DCP, GPU came up (renderD128), desktop session started.appledrmboundeDP-1(3024x1964), GPU probed (Apple M4 (G16G B0)), off-screenglmark2passed, 0 DART faults.Tested-by: @kazuhidet
Still to run
s2idle/resume with the adopted internal display DARTs onintegration/aurora(with870eb638eb67).Leaving inherited display DART untouched) andsimpledrmintact.dcpis enabled by the loader.Assisted-by: LLM (Anthropic:Claude, Google:Gemini 4 Neon, OpenAI:GPT-6.2 Sol)