Repository navigation
Conversation
scottjones
left a comment
There was a problem hiding this comment.
We reviewed ce30ba3 on top of 396f2d6, which is the current integration/aurora tip. Thanks for the careful write-up and for keeping the naming to where the digest was measured.
Where the digest comes from. It is the driver's own measurement. digital-brew's probe line from the g15_accept_text_prefix path prints the full sha256, and that is the TEXT hash with the GKTS entropy bytes zeroed, the same value a pin compares. The prefix is 0x125f0497c7904fde, and the digest doesn't duplicate either of the other two rows. Both probe lines on #289 give the full digest, MikeK's relayed from Discord in the issue body and digital-brew's first-hand, and the row matches both.
Layout. We see no risk of picking the wrong layout from this row. FirmwareBuild carries only sizes, marker and digest. Layout, root magic, IOMappings, board images and compute registers are all one per-SoC Config, and boot.rs:501 throws away the build identity::firmware returns. So a pinned build and one admitted by g15_accept_text_prefix run exactly the same code after the lookup. The only difference is an info line instead of a warning. That means the evidence on #289 (glxinfo on the M3 Max renderer, GLES compute with 0 mismatches, 10 admitted boots with no faults) already exercised the path this pin turns on. There is no other per-build table that needs an entry. The m3 KNOWN_IMAGES list covers T6030 and T8122 images, not this one.
The open question is the one maralcbr raised: matching sizes and marker don't prove the firmware ABI is the same. GL compute and daily desktop use cover a lot, but not every firmware path. Sleep, fault recovery and OpenCL weren't run on this digest. We're fine with this as the stopgap #289 agreed to, with the byte diff left for later.
Tests. No host test reads the T6031 firmware list. g15-maintained-handoff-tests.py and g15-zero-words-tests.py read the T6030 table in hw.rs, and the selftest under asahi-g15 tests lookup.rs and matching.rs with synthetic builds. We checked this two ways. We changed the marker and the last digest byte of the new row, and selftest run.py, maintained-handoff, zero-words and display-order all still passed. We also dropped the digest compare in matching.rs, and run.py then failed five lookup and matching tests. So the lookup is well covered but the table isn't. The two existing J516C rows have the same gap, so it's not a blocker. A small check that every T6031 row has a 32-byte digest, the J516C sizes and marker and no duplicates would be a cheap follow-up.
Small wording point, optional. The row comment says both Macs have OS firmware 14.7. For MikeK's Mac that comes from reading "14.727.0" as 14.7 plus 27.0, which hasn't been confirmed. Something like "OS firmware 14.7 on the second" would be more exact.
Hardware. The new digest itself has only been tested through the override on the two reporters' Macs. The lab boot (j1031-185713-2fe6) was the pre-rebase commit on a Mac that loads the 14.7 pin, so it shows that nothing regressed for the existing rows, not that the new pin works. Since the pinned and override paths are identical after the lookup, we don't think that is a problem.
What we ran, from a fresh worktree with Rust 1.85: selftest run.py and g15 input-error, queue-credit, maintained-handoff, zero-words and display-order pass. probe-init, retention, pb-growth and stats-drain wouldn't build on this host (two need Rust 1.93.1, two use features that are unstable in 1.85). None of them read t6031.rs. checkpatch --strict is clean, check-commits.py reports Signed-off-by yes and checkpatch clean, and the commit has Naeem's Signed-off-by and an Assisted-by trailer with no agent sign-off. drivers/gpu/drm/asahi compiles with clang LLVM=1 and Rust 1.85 on aarch64, with the new row in the object. We didn't do the final module link, because that needs a full vmlinux build. CI at ce30ba3 is green, including the boot job, though that job doesn't load this firmware. We didn't boot anything or run anything on hardware.
|
Hi Naeem, sorry this one sat. I approved it on 11-03 with CI green, but it never got queued, and it now conflicts with integration/aurora. That's almost certainly #363 (J514C defaults), which merged on 11-07 and reworked the same firmware table in Could you rebase onto current integration/aurora? While you're in there, there's new evidence for this digest on the 14" board. Davy (davycardinaal) has a J514C (M3 Max, T6031; os-fw 14.7, macOS 26.6.2) that loads the same TEXT, 125f0497. With Would you be up for admitting 125f0497 on the J514C as well, bound to Once it's rebased and CI is green I'll re-review and queue it. Thanks again for the careful work on this. |
Two J516C MacBook Pro 16" M3 Max owners (issue #289) load a GPU firmware whose TEXT digest is 125f0497c7904fde... (sha256 125f0497c7904fde87b05a9264a88cf03bdb3572a1715cec91bd0f9b78245544, TEXT 0x64000, DATA 0x140000, marker 0x5cf28). The driver refuses it and the Mac falls back to llvmpipe. Add it as a third FirmwareBuild, written like the two existing J516C entries and still gated to the J516C. The entry name and the comment record where the digest was measured; they do not claim a distinct firmware build. Evidence from #289: with the digest accepted through asahi.g15_accept_text_prefix, glxinfo reports Apple M3 Max (G15C B2), and a GLES compute run shows 0 mismatches and no faults in dmesg. This branch boots on a J516C M3 Max in the Mac lab (job j1031-185713-2fe6, 3 of 3 tests pass, GPU probed). That Mac loads the macOS 14.7 stub firmware, so the new digest itself was not exercised there. Assisted-by: Claude:claude-opus-5-5 Signed-off-by: Naeem Malik <malik2015naeem@gmail.com>
Two 14-inch M3 Max J514C owners load the same GPU firmware TEXT that the previous commit admits on J516C: sha256 125f0497c7904fde87b05a9264a88cf03bdb3572a1715cec91bd0f9b78245544, TEXT 0x64000, DATA 0x140000, marker 0x5cf28. Both run OS firmware 14.7. With the digest accepted through asahi.g15_accept_text_prefix the GPU comes up on both (G15C B2; OpenGL 4.6 accelerated on one, render and compute known-answer tests with 0 wrong on the other). Without it the driver refuses the firmware and the Mac falls back to llvmpipe. A row takes one machine, so add a second row for the same digest bound to apple,j514c. The J516C row stays bound to apple,j516c; the digest is not open to any other T6031 board. The board images come from the Config, not from the row, so the row carries only the firmware identity. The J514C host test counted exactly one J514C row and expected another board to be refused a digest; with one digest bound to two boards, count two rows and expect the first row bound to the asking board instead. Not tested on hardware by this change: the new digest itself was not exercised in the Mac lab, whose J514C loads the already admitted build. Power and thermal qualification of J514C remains open. Assisted-by: Claude:claude-sonnet-5-5 Signed-off-by: Naeem Malik <malik2015naeem@gmail.com>
ce30ba3 to
b33057c
Compare
scottjones
left a comment
There was a problem hiding this comment.
Thanks for the quick rebase, Naeem, and for taking on the J514C row too.
I checked both new rows. Each decodes to TEXT sha256 125f0497…245544, TEXT 0x64000, DATA 0x140000, marker 0x5cf28, one bound to apple,j516c and one to apple,j514c, and no other rows changed. The tests pass, and they fail when the change is broken. Firmware rows carry no power data, so the J514C row runs on the same J516C reference power words the existing J514C row already uses. That's fine.
Approving. A few small asks before it gets queued:
- Retitle the PR to mention J514C, since it now covers both boards.
- The J514C row comment and the commit message say two owners brought the GPU up on J514C. Only Davy's case is verified (idle, OpenGL, 3 boots, 63.8 C). Could you say that, and record his system firmware 26.6.2, the way the existing J514C row records 26.5.2?
A Vulkan and sustained-load run on J514C is still to do, but that's a follow-up for the whole board, not for this PR.
maralcbr
left a comment
There was a problem hiding this comment.
Reviewed at b33057c on top of integration/aurora f5180d8 (merges cleanly; the base moved only in g15/mod.rs and probe-init tests, nothing that touches the firmware table).
The code is fine. Both new rows decode to TEXT sha256 125f0497…245544, TEXT 0x64000, DATA 0x140000, marker 0x5cf28, bound to apple,j516c and apple,j514c. Lookup takes the first full board-compatible match, so the J516C row can't shadow the J514C one, and the test's expected() handles the shared digest correctly. g15-j514c-firmware-tests.py passes here (14/14), and CI is green at this head.
Holding the merge for Scott's two asks, since commit messages are permanent in the merge history:
- Retitle the PR to mention J514C.
- The second commit message says "Two 14-inch M3 Max J514C owners … OpenGL 4.6 accelerated on one, render and compute known-answer tests with 0 wrong on the other", and the J514C row comment says two owners. I can only find one J514C result for this digest: Davy's on omacom/omarchy-mac-pkgs#129 (os-fw 14.7, iBoot2 iBoot-10151.140.19.700.2, system firmware 26.6.2, OpenGL 4.6 accelerated, 3 boots, no Vulkan or load yet). If there's a second report, please link it; otherwise reword the commit message and comment to that one case and record 26.6.2. Please amend the commit rather than add a fix-up on top.
Optional, from the first round: the J516C row comment says OS firmware 14.7 for both Macs, but MikeK's 14.7 is a reading of "14.727.0" that was never confirmed.
Worth noting for the longer-term fix: mlninja-dev's J516C on #289 has the same iBoot1, iBoot2, OS firmware and system firmware as mvanduijker's 125f0497 unit but loads 669913fb, so these digests vary per unit, not per boot-chain version. Fine as a stopgap; it's more evidence for normalising the per-unit fields instead of pinning.
Once the reworded head is up and CI is green, I'll merge it.
|
Carried in #487 with the changes asked for in review (J514C title, message narrowed to the verified J514C case) and Scott's sign-off added. Pushing to this branch was refused, so the carry comes from an omacom branch. Naeem stays the author of both commits. |
Base: omacom/linux-aurora
integration/aurora(396f2d6 at time of writing)Head: malik-na/linux-aurora
g15-289-third-j516c-pin-v2@ ce30ba3Fixes/relates: #289
What
Adds a third
FirmwareBuildto the J516C list indrivers/gpu/drm/asahi/g15/hw/t6031.rs:125f0497c7904fde87b05a9264a88cf03bdb3572a1715cec91bd0f9b782455440x64000, DATA0x140000, marker0x5cf28(same sizes and marker as the two existing entries)machine: None, so it stays gated by the existing J516C layout checks, like the other two entries.The entry name and comment say where the digest was measured. They do not claim it is a distinct firmware build (see maralcbr's wording points on #289 below).
Why
Two J516C MacBook Pro 16" M3 Max owners reported on #289 that the driver refuses the GPU firmware their Mac loads, so the Mac falls back to llvmpipe. Both loaded TEXT
125f0497….Evidence (as posted on #289)
7.2.11-11-1-ARCH(aurora11), refused without override7.2.11-11-90-ARCH(aurora11 + #292 display series; G15 driver unchanged)mBoot-20457.1.29mBoot-20457.41.3iBoot-10151.140.19.700.2iBoot-10151.140.19.700.214.727.0in his paste, read as 14.7 + 27.0 (not confirmed)14.7(system-fwunknown)asahi.g15_accept_text_prefix=0x125f0497c7904fdeglxinfo -B: Apple M3 Max (G15C B2), Accelerated: yes, OpenGL 4.6 core / ES 3.2;vulkaninfo: Apple M3 Max (G15C B2); GLES 3.1 compute: 20 dispatches, 0 mismatches, no GL errors, no GPU messages in dmesg; same TEXT digest in 13 journal boots (3 refused, 10 admitted), no faults in any admitted boot; daily use with internal + one external displayMac 2's probe log line:
GPU firmware build admitted by g15_accept_text_prefix: TEXT 0x64000 DATA 0x140000 GKTS 0x5cf28 sha256 125f0497…5544. Its compute run is GL compute; there is no OpenCL ICD on that install.Boot test of this branch on a J516C
Mac Lab job
j1031-185713-2fe6(M3 Max, J516C) booted the earlier form of this change (6b9cb33, same code, before the rebase and message fix): 3 of 3 tests pass, GPU probed (asahi-g15 …: G15: probing (T6031 G15C), firmware identity accepted, cores enabled[10, 10, 10, 10]).Limit: that Mac loads the macOS 14.7 stub firmware (the log shows it matched the existing "macOS 14.7 stub firmware (iBoot-10151.140.19.700.2), J516C" entry). The new digest was not exercised there; it only shows the branch boots and the GPU still probes for the existing pins. The rebased commit (ce30ba3) itself has not been booted.
Related
maralcbr's comment on #289 about not treating matching versions or sizes as build identity, and the idea of normalising bounded, validated per-unit fields instead of pinning digests (needs maintainer agreement first): #289 (comment). That comment ties it to the T6030 investigation in #247. This PR is the stopgap that comment calls fine once a compute job and a renderer check pass, which Mac 2 provided.
Checks run on ce30ba3
checkpatch.pl --strict: 0 errors, 0 warnings, 0 checkstools/aurora-ci/check-commits.py(base 396f2d6): Signed-off-by yes, checkpatch cleang15-maintained-handoff-tests.pyandg15-zero-words-tests.pypass (full host suite not run)Not tested
125f0497….asahi.g15_accept_text_prefix, not the pin itself.Assisted-by
Commit carries
Assisted-by: Claude:claude-opus-5-5and Naeem's own Signed-off-by; no agent Signed-off-by, perDocumentation/process/coding-assistants.rst.[unverified: MikeK's report is relayed in the first #289 comment, not first-hand; "14.7 + 27.0" for his Mac is that comment's own guess.]