Repository navigation
kboot: default J514C native display and graphics - #23
chriskearney wants to merge 4 commits into
Conversation
|
J514C hardware validation at 22bd961, paired with the explicitly opted-in Aurora14 kernel in omacom/linux-aurora#363: native display handoff succeeded at 3024×1964/120 Hz; GPU firmware reached READY and all 40 cores were enabled. The installed-package check matched the frozen identity before/booted/after. A subsequent normal installed boot logged in automatically and passed short native OpenGL/Vulkan pixel checks with zero wrong pixels. The tested target now uses this Stage2 profile persistently, with its original recovery loader/UKI retained. Stage1 was not enrolled. A target-local SDDM readiness wait was needed for the framebuffer/VT handoff; it is separate from this loader patch. This supports the experimental J514C build profile, without general board power/thermal or sustained-load qualification. The default build's J516C-only behavior remains unchanged. |
[Medium impact] The PR appears safe to merge; the existing GPU guide issue is non-blocking. Findings
SummaryJ514C now receives the same built-in display and GPU defaults as J516C.
Reviews (4) · Last reviewed commit: "docs: describe normal J514C graphics def..." · Reviewed by Greptile |
| This switch runs the same fail-soft G15 GPU setup that `chosen.asahi,t6030-gpu=1` | ||
| runs on the M3 Pro (`dt_set_gpu_g15()` in `src/kboot_gpu.c`), with the T8122 | ||
| entry of its SoC table. It is off by default: without it a T8122 boot does not | ||
| run the setup at all (there is no dry run), and only the diagnostic switches | ||
| above apply. With it, `asahi,t8122-gpu-handoff-diag` is ignored. |
There was a problem hiding this comment.
This section says T8122 GPU setup is off by default. Later, it says missing firmware and power data prevent hardware access. But this PR adds GPU, experimental power-model, and leakage defaults for J613/J615, and supplies the experimental profile’s firmware and power callbacks. Owners following the guide could wrongly expect no GPU hardware access without an explicit opt-in. Update both claims to explain the new defaults, the experimental path, and how owners can turn it off.
Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!
scottjones
left a comment
There was a problem hiding this comment.
We reviewed the J514C change at 22bd961 against integration/aurora. The change is small and correctly gated. A normal build produces exactly the same board table as before, and the tests catch the gate being broken.
Base
This was opened against main and showed about 54k lines across 332 files. It has since been retargeted to integration/aurora, which fixes that. Its merge-base with integration/aurora is bc79acc (the aurora22 stable). The only commit on top is 22bd961: Makefile, src/kboot.c and tests/python/test_m3_graphics_defaults.py, +40/-2.
The gate
- Makefile:79-80 adds
J514C_BRINGUPtoCFGonly whenJ514C_BRINGUP=1. build/build_cfg.h is rewritten on every build (the cmp/mv rule) and included by kboot.c, so an older build directory can't carry the define over.#ifdef J514C_BRINGUP(src/kboot.c:416) addsapple,j514conly to the T6031 entry, anddt_set_m3_defaults()still requiresapple,t6031before it looks at the board list. - That keeps J516C, the Pro and Air boards, and the T6034 Maxes as they are, and a default build doesn't add J514C. The master switch, an owner's
=0, and the drop-everything-on-a-lost-config-line behaviour still apply to J514C in the candidate build. - A J514C with the candidate m1n1 but a kernel without
asahi.g15_j514c_bringup=1gets the same handoff that the existing manualchosen.asahi,t6031-gpu=1line already gives. The kernel then refuses the GPU, and the display handoff is unchanged.
The failing checks are not from this change
- format-check: the workflow (.github/workflows/main.yml) only runs on pushes and PRs to
main, so it won't run now that the base is integration/aurora. The failing run built the merge with main (7384ff0). Its violations are almost all in code integration/aurora already has: kboot_gpu_t8122.c, atc_phy.c, afk.c, kboot_gpu.c and others. integration/aurora's own kboot.c already has 25 violations under clang-format 20.1.8. The change does add 5 more in the newm3_max_boardsblock (kboot.c:412-418). That's a nit given the base, but easy to fix. - macOS pytest: every failure is
tests/python/test_m3_defaults_optout.pydying with SIGABRT, on every board including J613, J514S and J516C. The same suite already fails on integration/aurora's own push run (121fcbf, macos-14) with the same tests. It's a problem with the macOS harness on this branch, not this change. Both Ubuntu jobs and the native arm64 job pass.
Minor
- The comment above the board table (kboot.c:392) still says the J514C is "not touched". In a candidate build that's no longer true. Worth one line saying so.
- The PR body still says "Hardware qualification is pending; no working GPU claim yet", while the later comment reports a working GPU. Please update the body so the record matches. Same caveat as on linux-aurora#363: this is short pixel-correctness evidence on borrowed J516C power and thermal values, not a power or thermal qualification. Please keep
J514C_BRINGUPout of any channel or release build until the kernel side has a J514C power ceiling of its own.
What we ran (aarch64 host)
pytest tests/python/test_m3_graphics_defaults.py tests/python/test_m3_defaults_optout.py: 146 passed, 1 xfailed.- Mutations, each reverted afterwards:
"apple,j514c"moved outside the#ifdef(always on): 8 failures (test_other_maxes_and_macs_get_no_default,test_the_j514c_manual_line_works_as_before,test_other_boards_keep_their_lines_after_a_lossand others).#ifdefflipped to#ifndef: 9 failures, including the newtest_j514c_candidate_build_keeps_exact_board_soc_and_owner_optout.
- clang-format 20.1.8 (the CI version) on src/kboot.c, at this head and at integration/aurora.
- Read the failing CI logs and compared them with integration/aurora's run 121fcbf.
What we couldn't verify
The hardware result, which we only have from the posted comment. We didn't do a full cross-build of m1n1 (CI build passes).
|
Small correction to our review: a |
The candidate build defaults both T6031 display and GPU switches on J514C. Several comments still say it always requires manual config. Describe the build-only exception at each handoff and firmware setup site, retaining owner values and the normal-build behavior. Cover chosen.asahi,t6031-dcp=0 separately: the explicit display opt-out survives while the independent GPU switch still defaults on. Format the new board table without reformatting inherited source. The two production-body defaults/opt-out suites pass 139 tests, with eight optional updater integration checks skipped. This revision changes comments, table whitespace and tests only; the target keeps tested Stage2 22bd961. Assisted-by: Codex:GPT-6
|
Want your agent to iterate on Greptile's feedback? Try greploops. |
|
Applied at This changes comments, table whitespace and tests only; the target keeps the hardware-tested |
Give exact J514C/T6031 boots the same display and GPU defaults as J516C without a build define or manual boot.bin enable line. Remove the J514C_BRINGUP switch and update each handoff comment accordingly. The defaults remain subordinate to explicit owner settings, master-off and configuration-loss handling; T6034 is unchanged. Exercise no-config Stage2 boots for both T6031 boards, independent display/GPU opt-outs and loss handling. The production-body host suites pass 155 tests, with eight optional updater integration cases skipped. Assisted-by: Codex:GPT-6
scottjones
left a comment
There was a problem hiding this comment.
We re-reviewed 08c65e6, the new default-on commit on top of b920630. The change is clean. J514C now gets the same display and GPU defaults as J516C, and the owner opt-outs, master-off and config-loss handling all still apply to it.
The default-on commit
- J514C is now in
m3_max_boards(src/kboot.c:411) and theJ514C_BRINGUPdefine is gone from the Makefile. The test now asserts it stays gone.dt_set_m3_defaults()still requiresapple,t6031before it looks at the boards, so T6034 Maxes and other boards get nothing. The source comments are updated. - One behaviour change to know about: a J514C owner who set the manual
t6031-gpu=1/dcp=1lines now falls under the lost-config-line rule like J516C. If a line is lost, their graphics stay off instead of keeping the lines (tests/python/test_m3_graphics_defaults.py:572-577). That's the safe direction and matches J516C. - The J516C power/thermal values the kernel borrows for J514C are still the known unqualified item. We've asked on linux-aurora#363 for a sustained-load run on Ryan's Mac before merge.
Still open, minor
- docs/m1-m3-handoff-sources.md:55 and :62 still say "A J514C still needs the line". That's now wrong for every build.
- Small stale test text: the comment at test_m3_graphics_defaults.py:162-163 still lists J514C as "untouched", the one at :571 says the loss rule is "only for the five with defaults", and
test_the_j514c_manual_line_works_as_before(test_m3_defaults_optout.py:515) no longer tests "as before". - omarchy-mac-pkgs#92 still describes J514C as "Experimental J514C login ... outside the supported release". Worth a wording pass there once this lands.
CI
Same pre-existing macOS failure, not this change. macOS-14 pytest fails only in test_m3_defaults_optout.py, which fails the same way on integration/aurora's own run 41718740971. It's now 149 failures instead of 133 because the commit runs more of that file's tests on J514C too. macOS-15 was cancelled when macOS-14 failed. Both Ubuntu jobs and native arm64 pass.
What we ran (aarch64 host)
pytest tests/python/test_m3_graphics_defaults.py tests/python/test_m3_defaults_optout.py: 162 passed, 1 xfailed.- Mutations of
dt_set_m3_defaults(), each caught and then reverted:- J514C owner's
dcp=0overwritten by the default: 5 failures. - All J514C owner lines overwritten: 8 or more failures.
- J514C ignores master-off: 7 or more failures.
- J514C ignores a lost config line: 2 failures.
- J514C removed from the board list: 6 or more failures.
- SoC check removed: 2 failures.
- J514C owner's
- Adding
apple,j514mto the board list fails nothing. That's expected, because the T6031 SoC check already excludes it.
The handoff guide still said J514C needed manual chosen lines despite the normal-build default. Describe both J514C and J516C defaults and owner opt-outs, and align host-test names and comments with the behavior they assert. Assisted-by: Codex:GPT-6
scottjones
left a comment
There was a problem hiding this comment.
We re-reviewed 6560adc. The only new commit since 08c65e6 is docs and test text, with no code change. It fixes the stale "A J514C still needs the line" text in docs/m1-m3-handoff-sources.md, the "untouched" comment, and the "as before" test name we pointed out.
Small leftovers, not blocking: test_m3_graphics_defaults.py still says "the five boards" at :154, :521 and :571. Since J514C was added there are six.
CI
Same pre-existing macOS failure as before. macOS-14 pytest fails only in test_m3_defaults_optout.py (148 failures), and integration/aurora fails the same way. macOS-15 was cancelled when macOS-14 failed. Both Ubuntu jobs and native arm64 pass.
What we ran (aarch64 host)
pytest tests/python/test_m3_graphics_defaults.py tests/python/test_m3_defaults_optout.py: 162 passed, 1 xfailed.- The whole tests/python tree: 410 passed. The same 2 failures and 2 errors also happen on #24's head and need tools this host doesn't have (malformed-ADT runner, Rust FAT parser, stage1 T6032 probe), so this change doesn't cause them.
|
Hardware results from our Mac lab on a J514C (14" M3 Max, 40-core GPU), m1n1 built from 6560adc:
Also in a combined trial with m1n1 #23 + #24 + #25 and linux-aurora AsahiLinux#363 + AsahiLinux#372 + AsahiLinux#388 + AsahiLinux#400 (stage2 v1.6.1-473-gd2b8d44b) on an M3 Air, J504, J514M and MacBook Neo: 0 DART faults on all four, no regressions vs stock apart from audio.microphone-signal on the J514M, which fails on its stock kernel too. The J514C hasn't run the combined build yet. |
J514C/T6031 boots now receive the built-in display and GPU defaults in a normal Stage2 build, just like J516C. Remove the J514C build flag: owners need no special build option or manual boot.bin enable lines. Existing explicit owner settings, independent display/GPU opt-outs, master-off and configuration-loss behavior remain in force. T6034 is excluded and Stage1 enrollment is unchanged.
The paired kernel linux-aurora#363 admits the exact board automatically while retaining native display, firmware identity and hardware checks. The SDDM package follow-up supplies the matching automatic readiness wait.
At current head
08c65e6d5821c750c675c67df00ccd1572fe1505, the production-body defaults/opt-out suites pass 155 tests, with eight optional updater integration checks skipped. The no-config Stage2 path is exercised on both J514C and J516C, including independent display/GPU owner opt-outs; the host harness needs no J514C build define.Historical hardware evidence is at
22bd9616205e3a7c795e47356a5fb80f8b4b2dec, whose explicitly built J514C table is now the normal-build table: native 3024×1964/120 Hz display, firmware READY on all 40 cores, installed identity checks, automatic login and short GLES/Vulkan pixel checks with zero wrong pixels. The live target still uses that tested package. The new default-build revision has host evidence, without a new hardware result; no CTS, sustained-load or power/thermal qualification is claimed. The four J516C reference groups are documented on the kernel PR. No channel or release build has been published by this update.