You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
G15: T6030 (M3 Pro) GPU firmware TEXT varies between Macs on the same boot chain; two J514S builds are refused #247
Two 14-inch M3 Pros (apple,j514s) on edge (linux-aurora 7.2.11.aurora8-1, m1n1-aurora 1.6.1.aurora21-1) have m1n1 switch the GPU on, but the G15 probe refuses their firmware, so there's no render node and they fall back to software rendering. The display (DCP) works on both.
J514S USB4/Thunderbolt and Type-C display support #2 runs the build pinned for J516S, but the pin's board scope refuses it (identity.rs: a qualified build pins only on the machine it names).
J514S USB4/Thunderbolt imported prerequisites #1 runs on the exact boot chain the first entry names (iBoot-10151.140.19.700.2), with a TEXT digest that isn't pinned anywhere.
The boot-chain version alone doesn't determine the TEXT image on T6030. Is the selection per chip revision (the J516S pin is labelled B0)? If so, the pins probably need to be keyed by revision rather than by board, and the B0 pin widened to J514S.
Both users are willing to test. For a single test boot, asahi.g15_accept_text_prefix= exists. The entry comment prefers a full pin, though, and J514S #1's image hasn't been qualified on any board.
Result:
GPU firmware build admitted by g15_accept_text_prefix: TEXT 0x5c000 DATA 0x114000 GKTS 0x58e88 sha256 cd05af6a828b52e4c2fed20086d3d18a7db42e03a8fd6f8354f9f06e0a61363e
M3: firmware identity accepted
board data: ADT performance states, 1380 MHz top (differs from the reference machine)
board data sha256 0f2fcd85a90cfff305c1121736033d36af612ca066b405e13e0b1cd046f55661
T6030 G15S: cores enabled per cluster: [7, 7]
G15: firmware at READY, starting the backend
/dev/dri/renderD128 is present and there are no faults in dmesg. glxinfo reports Apple M3 Pro (G15S B1), OpenGL 4.6, Mesa 26.3.0-devel. Hyprland CPU dropped from roughly 400% on llvmpipe to about 2%.
One observation on the question of what selects the image: my system-fw-version is 26.6.2 (os-fw-version 14.7, iBoot2 iBoot-10151.140.19.700.2, iBoot1 mBoot-18000.161.10). The PR #242 description reports system-fw-version 27.0 for the da26ecca… build on the same iBoot. That might be what separates the two, though two machines don't prove it.
Not tested yet: Vulkan, and sustained load. Happy to boot-test a patch without the override.
TEXT 0x5c000, DATA 0x114000, GKTS 0x58e88 (same as the other two)
Board data sha256
6402a3f24b2399e3a97d9b93fd1d7fc1da07b4fc5091bc04a37d62da7bd453db, ADT performance states, 1380 MHz top
Power
version=0x7031100 counts=0x110209 core-mask=0x7e5bd, cores per cluster [7, 7]
The boot chain is identical to J514S #1's (14.7 stub, iBoot-10151.140.19.700.2, system firmware 26.6.2), but the TEXT digest differs (cd05af6a… vs 411833a7…). That supports your point that the boot chain alone doesn't select the TEXT image on T6030. It's probably the chip revision or something else per unit.
Setup:linux-aurora 7.2.11.aurora10-1, m1n1-aurora 1.6.1.aurora22-1, mesa-aurora / vulkan-aurora 26.3.0_devel.20261029.231476.672657e05e6-1, Hyprland 0.56.2. The override is in /etc/modprobe.d/asahi-g15.conf:
GPU firmware build admitted by g15_accept_text_prefix: TEXT 0x5c000 DATA 0x114000 GKTS 0x58e88 sha256 411833a7ded773feeac4a181552e3ff5582eaf4435f2dd8f620c1ed0b31df1b0
M3: firmware identity accepted
T6030 G15S admitted, ASC=0x0
G15: firmware at READY, starting the backend
/dev/dri/renderD128 is present, and Hyprland/aquamarine and Quickshell render on it. The internal panel runs through DCP at 3024×1964@120. The kernel log shows no G15 faults, timeouts or recoveries so far, but this is early: only a few minutes of uptime when I checked.
A side note for anyone switching an older omarchy-mac (legacy GRUB) install from linux-asahi to linux-aurora: right after the package switch, m1n1's boot.bin carried linux-asahi's t6030-j514s.dtb, not aurora's. I caught and fixed this before rebooting. Booted as-is, G15 would have had no device-tree node, independent of the firmware digest. Details in a separate packaging issue.
G15: probing (T6030 G15S)
unsupported GPU firmware build: TEXT 0x5c000 DATA 0x114000 GKTS 0x58e88 sha256 cd05af6a828b52e4c2fed20086d3d18a7db42e03a8fd6f8354f9f06e0a61363e; to have it supported, report this line and the macOS version of the boot chain that loaded it
supported: macOS 14.7-14.8.x stub firmware (iBoot-10151.140.19.700.2)
supported: J516S B0 RTKit-2419.140.12 (maintained 14.x handoff)
supported: J514S B0 RTKit-2419.140.12 (maintained 14.x handoff)
loaded GFX TEXT identity mismatch: EINVAL
probe with driver asahi-g15 failed with error -22
No renderD128, so Hyprland falls back to llvmpipe (~440% CPU at 3456×2234@120). The same boot chain (14.7 stub, iBoot-10151.140.19.700.2) and the same TEXT digest as J514S #1, on a different board and a different iBoot1 (mBoot-18000.121.3) and system firmware (26.5.2), which fits the observation that the boot chain alone doesn't select the TEXT image.
I haven't tried the g15_accept_text_prefix=0xcd05af6a828b52e4 override yet; happy to boot-test it or a patch that pins this build for J516S.
First J516S (16" M3 Pro) report of the 411833a7… TEXT build, the same digest as the third J514S above. It's refused on this board too, on the same boot chain.
Machine
apple,j516s / apple,t6030, MacBook Pro 16" M3 Pro (Nov 2023)
G15: probing (T6030 G15S)
unsupported GPU firmware build: TEXT 0x5c000 DATA 0x114000 GKTS 0x58e88 sha256 411833a7ded773feeac4a181552e3ff5582eaf4435f2dd8f620c1ed0b31df1b0; to have it supported, report this line and the macOS version of the boot chain that loaded it
supported: macOS 14.7-14.8.x stub firmware (iBoot-10151.140.19.700.2)
supported: J516S B0 RTKit-2419.140.12 (maintained 14.x handoff)
supported: J514S B0 RTKit-2419.140.12 (maintained 14.x handoff)
loaded GFX TEXT identity mismatch: EINVAL
probe with driver asahi-g15 failed with error -22
No renderD128, so the session runs on llvmpipe. This makes two T6030 boards (J514S and J516S) on the same boot chain (14.7 stub, system firmware 26.6.2, mBoot-18000.161.10) loading 411833a7…, while J516S has now been seen with both 411833a7… and cd05af6a…. That fits the per-unit or per-revision selection theory, not a per-board one. Happy to boot-test a pin for this build on J516S.
Something to watch out for on Omarchy Mac installs: these boot a Limine UKI (ENABLE_UKI=yes in /etc/default/limine), so mkinitcpio -P alone doesn't apply the option. It only rebuilds /boot/initramfs-linux-aurora.img. You need limine-mkinitcpio to rebuild the UKI on the ESP.
Result:
GPU firmware build admitted by g15_accept_text_prefix: TEXT 0x5c000 DATA 0x114000 GKTS 0x58e88 sha256 411833a7ded773feeac4a181552e3ff5582eaf4435f2dd8f620c1ed0b31df1b0
M3: firmware identity accepted
T6030 G15S admitted, ASC=0x0
board data: ADT performance states, 1380 MHz top (differs from the reference machine)
board data sha256 b3024f4d8134e56645913588da2f0e637e95a0988aee8cb1740c0371d340c3ce
T6030 G15S: cores enabled per cluster: [9, 9]
M3: firmware accepted owned initdata and device controls
G15: firmware at READY, starting the backend
/dev/dri/renderD128 is present, and Hyprland renders on it on the internal panel (eDP-1, 3456×2234@120). Hyprland CPU dropped from about 440% on llvmpipe to about 3%. The kernel log shows no G15 faults, timeouts or recoveries, but it's early: only a few minutes of uptime when I checked, and I haven't tested Vulkan or sustained load.
Setup: linux-aurora 7.2.11.aurora14-1, m1n1-aurora 1.6.1.aurora22-1, mesa-aurora / vulkan-aurora 26.3.0_devel.20261031.231483.a1969111475-1, Hyprland 0.56.2. Happy to boot-test a pin for this build on J516S.
After applying the override, rebuilding the boot images and rebooting, the kernel log reports that the firmware was admitted and the GPU backend reached READY. The system registers renderD128. I haven't independently verified the active application renderer or tested sustained GPU workloads.
The probe log is identical to the one in the issue body. There's no renderD128; aquamarine logs CDRMRenderer(drm): Can't create renderer, no matching devices found, and Hyprland sits at ~670% CPU on the internal panel (3024×1964@120).
On what selects the image: cd05af6a… has now been seen with system firmware 26.6.2 / mBoot-18000.161.10 (J514S #1) and 26.5.2 / mBoot-18000.121.3 (this machine and the J516S report). So neither system firmware nor iBoot1 seems to select it, which fits the per-unit or per-revision theory. Happy to boot-test a pin for this build on J514S.
Follow-up: the g15_accept_text_prefix=0xcd05af6a828b52e4 override works on this J514S (system firmware 26.5.2, mBoot-18000.121.3), set in /etc/modprobe.d/asahi-g15.conf and rebuilt with limine-mkinitcpio.
GPU firmware build admitted by g15_accept_text_prefix: TEXT 0x5c000 DATA 0x114000 GKTS 0x58e88 sha256 cd05af6a828b52e4c2fed20086d3d18a7db42e03a8fd6f8354f9f06e0a61363e
M3: firmware identity accepted
T6030 G15S admitted, ASC=0x0
T6030 G15S: cores enabled per cluster: [7, 7]
G15: firmware at READY, starting the backend
renderD128 is present and aquamarine renders on Apple M3 Pro (G15S B1), OpenGL ES 3.2 Mesa 26.3.0-devel. Hyprland CPU dropped from ~670% on llvmpipe to ~5%. No G15 faults, timeouts or recoveries in the kernel log so far, but only a few minutes of uptime; Vulkan and sustained load not tested.
With the override, G15 reaches READY, renderD128 appears, and Hyprland drops from about 300–600% CPU on llvmpipe to about 1%.
Testing (all clean):
Vulkan: vulkaninfo and vkcube OK (Borealis, Vulkan 1.4)
glmark2 (fullscreen, 3024×1964): 2887
vkmark: 3628
Basemark Web 3.0 in Brave: ~1150, vs 2386 on macOS on the same machine (about 48%)
SuperTuxKart at maximum settings, fullscreen: about 11 min
There were 0 G15 faults, timeouts or recoveries across two boots (about 2 hours total), including a reboot through macOS. Not yet tested: sustained load longer than about 11 minutes.
@jmayaalv thanks for the J516S report. Since it's the same cd05af6a828b52e4… build that works with the override on J514S, could you boot-test the override on your machine?
Rebuild the boot image so the option is picked up. On Omarchy Mac installs (Limine UKI) mkinitcpio -P alone isn't enough; run limine-mkinitcpio as well (see davidhorak's comment above).
Reboot, then post the G15 lines from the kernel log. We're looking for "GPU firmware build admitted by g15_accept_text_prefix", "firmware identity accepted" and "G15: firmware at READY", plus any faults, timeouts or recoveries.
Tell us whether /dev/dri/renderD128 exists, and what glxinfo reports for the renderer (others got Apple M3 Pro (G15S B1)).
Tell us the Hyprland CPU use now versus the ~440% you saw on llvmpipe.
If you can, also try vulkaninfo/vkcube and some sustained load; nobody has covered J516S cd05af6a… on that yet.
If it works, J516S will need its own pin row (on T6030 it's one row per board), and we can use your boot as the evidence for it.
Drafted with Claude Opus 5.5. I read every line and checked the numbers on my own Mac before I posted it.
@malik-na this bears on the question in your issue body. The fused GPU core mask looks like the selector, not the chip revision.
Two reports here give both the mask and the digest:
Reporter
core-mask
TEXT sha256
okhan1980
0x7e5bd
411833a7…
djfergus
0x6f5bd
cd05af6a…
Same board, same boot chain (os-fw 14.7, system-fw 26.6.2, mBoot-18000.161.10, iBoot-10151.140.19.700.2). Different mask, different digest.
The rest of the thread fits:
cd05af6a… appears on three units (jdlinventures, dtt101, djfergus), so the digest is not unique per unit.
dtt101 saw cd05af6a… on system-fw 26.5.2 and djfergus saw it on 26.6.2, so the firmware version does not select it.
jdlinventures and djfergus share cd05af6a… with different board-data digests (0f2fcd85…, 7d5f24b8…), so the perf-state table does not select it.
Every reporter logs version=0x7031100 counts=0x110209, which weakens the chip-revision theory in the issue body.
If iBoot writes the binning value into TEXT before the handoff, the digest tracks the binning pattern. That gives a small finite set of images and explains all four points.
Two things would settle it.@okhan1980@djfergus your two reports are the whole basis for this. Could you each post your full core-mask line beside your digest? Two different digests behind one mask kill the idea. Several matched pairs confirm it, and the pins can then key on the mask. Could you also post the sha256 of your stub's armfw_g15s.im4p? One payload hash behind two different TEXT digests means iBoot patches the image at load time.
macOS loads a different image
On this Mac, under macOS 27.0.1, arm-io/gfx-asc carries three segments:
Every pinned T6030 build has TEXT 0x5c000 and DATA 0x114000, and adt.rs requires exactly two segments named __TEXT;__DATA. The parse refuses this ADT before it hashes anything. The image therefore belongs to the OS firmware generation that loads it.
That firmware is a file, readable before any boot:
<Preboot>/restore/Firmware/agx/armfw_g15s.im4p
size 3356649
sha256 c82f824c8fd812260be98329c57413a2d2086c93fce8053c90a3ecf090d2f26e
IM4P type "gfxf", rkosftab payload
This machine, and what I cannot give you
Model
MacBook Pro 14-inch, M3 Pro, Nov 2023 (Mac15,6, apple,j514s, apple,t6030)
CPU
11 cores: 5 performance, 6 efficiency
GPU
14 cores (gpu-core-count = 14)
Memory
36 GB
ADT GPU node
compatible = "gpu,t6030"
chip-id / chip-revision
0x6030 / 0x11
ADT perf states
14 entries, 338 MHz lowest, 1380 MHz top
macOS
27.0.1 (26A434), kernel RELEASE_ARM64_T6030
System firmware
mBoot-20457.1.29
The bin matters for the hypothesis. The 14-inch M3 Pro ships as 11 CPU cores with 14 GPU cores, or as 12 with 18. This is the first. Every report in this thread shows cores enabled per cluster: [7, 7], so all of them are the 14-core GPU bin, yet their masks and digests differ. A report from an 18-core unit would say whether the digest splits along the bin, or only along the mask inside one bin.
The model string does not separate the two bins. t6030-j514s.dts sets model = "Apple MacBook Pro (14-inch, M3 Pro, Nov 2023)" with no core count, while t6034-j514m.dts names "14 CPU cores" for the M3 Max. A pin keyed on the model string would cover both M3 Pro bins without distinguishing them.
I have not booted Linux on it, so I have no TEXT digest and no probe log. The numbers above come from the macOS 27 chain and pin nothing. The core-mask correlation rests on two reports that are not mine. I have not checked that an installed stub leaves armfw_g15s.im4p readable, or that the payload is identical across Macs on one macOS build.
Tell me which ADT property to read next and I will post it.
Both were read from <uuid>/restore/Firmware/agx/ in each Preboot volume (the stub's directory is spelled Restore/), with the stub's mounted read-only from macOS. The stub had only that one copy.
@jmayaalv the J516S row for your cd05af6a… build has landed on integration/aurora. It came in with #427, which carried Joel's J516S cd05 commit from #390 and the J516S 4118 row, so a J516S now has its own pin for both cd05af6a… and 411833a7…. It's not in a tagged release yet. The CI package from the merge is kernel-package-5bc7ad11033bafbc46d4c45bcdd61db5 in https://github.com/omacom/linux-aurora/actions/runs/42754745532 (needs a GitHub sign-in to download). If you install it without the g15_accept_text_prefix line, you should see GPU firmware build: J516S cd05 firmware (14.7 boot chain) and no "admitted by g15_accept_text_prefix". Could you post the asahi-g15 lines and say whether renderD128 is there? Thanks!
@LubomirGeorgiev thanks for the core-mask idea. A few cases in the thread don't fit it:
The same cd05af6a… runs with [7, 7] cores per cluster on the J514S reports and [9, 9] on a J516S (drm/asahi: g15: pin the cd05 firmware TEXT on the J516S #390). 411833a7… shows [7, 7] on okhan1980's J514S and [9, 9] on davidhorak's J516S. One digest, different core counts.
So I don't think the mask picks the image. maralcbr's byte-diff plan (repeat the hash over cold boots, diff the TEXT dumps and the Preboot payloads) is the next step: #289 (comment). If you can dump TEXT from a unit, that would help.
Since all three T6030 digests are now pinned for both boards, the title here is out of date. Could a maintainer retitle it, say "G15: what selects the GPU firmware TEXT on T6030/T6031? Different digests on the same boot chain", or close it and keep the investigation on #289? Either is fine with me.
Last thing, for the M3 Max (T6031): digest 669913fb… on the J516C has no row yet. That's tracked in #289. #340 (the third J516C pin, 125f0497…) is approved but conflicted with integration/aurora; I've rebased it locally and will push once Scott confirms the shape.
Drafted with Claude Opus 5. I read every line and reproduced every number myself before I posted it.
The driver does not hash the firmware Apple ships. The boot loader rewrites TEXT first, so the pinned digest records the loader's output, not the firmware. That is why one identical file gives da26ecca…, cd05af6a… and 411833a7… on different Macs.
@malik-na you were right to drop the core-mask idea, and #390 settled it. This replaces it.
I pulled the stub firmware straight out of Apple's asset and measured it. The 14.8.3 OS firmware that upstream asahi-installer names in IPSW_VERSIONS for j514sap holds three GPU firmware files under AssetData/boot/Firmware/agx/. HTTP range requests fetch them in about 1.3 MB.
@djfergus the g15s size and digest are what you read from your own stub. So the file is generic. Every Mac on this stub reads identical bytes, and the payload is not personalised.
The file holds several complete builds
The IM4P payload is rkosftab, a table. Each entry is a whole Mach-O image with its own LC_UUID. The segment layouts are the ones already pinned.
armfw_g15s.im4p, T6030, TEXT 0x5c000, DATA 0x114000, marker 0x58e88:
armfw_g15g.im4p, T8122, TEXT 0x64000, DATA 0xb0000, marker 0x5d54a, holds three entries: a000, b000 and c000. The entry count is not fixed per chip.
The proof that TEXT is rewritten
g15/hw.rs:152 says of the row pinned as measured TEXT da26ecca…:
// The maintained J516S runtime already identifies this B0 image by
// UUID dbf37c40-ead5-3760-9441-995018782955 and uses the same
// 0x0c08e21e83800490 InitData ABI. Pin its complete measured TEXT
// digest here, rather than enabling the bring-up prefix override.
Entry b000 above carries exactly that UUID. Hashed by the rule in identity.rs, with the 8 bytes after the single GKTS\x08\0\0\0 marker zeroed, its TEXT is 06a4e7e0…, not da26ecca….
Same image. File TEXT 06a4e7e0…. Measured TEXT da26ecca….
None of the six file digests above appears anywhere in drivers/gpu/drm/asahi/. The images are the same on every Mac; the patch is not.
The GKTS entropy bytes are not the mechanism either. They are the constant ff0a000000000000 in all seven images across the three files, and identity.rs zeroes them before hashing.
So the pin table cannot converge. Each Mac configuration that patches differently needs a new row, and three entries in armfw_g15g.im4p predict up to three digest families on j504, j613 and j615 per firmware generation before any patch variation.
What I suggest, and what is still open
The LC_UUID sits inside TEXT, at offset 0x4214 in all seven images. It is therefore inside the segment the driver already copies, and it differs per entry. Matching those 16 bytes with the segment sizes and the marker offset would identify the build exactly and admit every unit of a known build, whatever the loader patched elsewhere. That is the identity the comment above already names. @maralcbr this looks like the normalisation you proposed in #289, with a field that is bounded and explained.
Two things I could not measure, both needing a booted Linux, so they belong to whoever can dump TEXT:
which bytes the loader rewrites, and what varies them per unit;
whether the loader also rewrites the UUID region. If it does, matching on the UUID fails. The comment in hw.rs suggests it does not.
On macOS the running system names the entry it was given:
This Mac (apple,j514s, chip-revision 0x11) answers b000 on both macOS 27.0.1 and 27.2.1, across the system firmware change from mBoot-20457.1.29 to mBoot-20457.41.3. The OS changed the file and the digest; the entry stayed.
I have two small scripts for this: one that range-extracts a member from the remote asset, one that prints each entry's layout, UUID and TEXT digest. Would a gist help, or would you rather they came as a PR against tools/asahi/?
@malik-na I agree the title is stale. "What rewrites the GPU firmware TEXT on T6030/T6031" would fit what the thread now shows.
Drafted with Claude Opus 5. I read every line and reproduced every number myself before I posted it.
@ilyasakin can you print 16 bytes from your J514S? #481 has the GPU bound on a kernel you built, so you are set up for it.
The pinned digests are not digests of the firmware Apple ships. Apple's generic armfw_g15s.im4p holds two complete T6030 builds, and neither one's TEXT digest appears under drivers/gpu/drm/asahi/. The boot chain writes into TEXT before the driver hashes it, so each pin records one machine's loaded bytes, not a build. Measurements earlier in this issue; #487 has the same result for T6031.
I proposed matching the build's LC_UUID instead. Its 16-byte value sits inside the hashed TEXT at 0x4214 in all seven 14.x stub images I measured, and covers none of the regions the loader rewrites. Whether the loader leaves those 16 bytes alone is untested.
In drivers/gpu/drm/asahi/g15/identity.rs, in firmware():
let mut copy = copy_physical(adt.text.start, size)?;
+ if let Some(u) = copy.get(0x4214..0x4230) {+ dev_info!(dev, "GPU firmware TEXT+0x4214: {}\n", Hex(u));+ }
Hex and dev_info! are in scope. 0x4214 is far from the GKTS marker, so the zeroing below does not reach it.
Apple's file carries these two UUIDs at that offset:
One of them means the loader leaves the UUID alone, and a UUID match plus text_size, data_size and the marker offset can replace the digest pins on T6030. Anything else means the loader rewrites the UUID too, and this does not work.
Also: which row does your GPU firmware build: line name? That shows how many units map to each pin.
I have no Linux on this Mac, because the T6030 rows are in no release.
The first 16 bytes are dbf37c40-ead5-3760-9441-995018782955, the b000 entry of armfw_g15s.im4p. So the loader leaves the UUID alone on this unit, and this unit maps to the J514S 4118 row.
The bytes after it, 00000000 00000000 888e0500, end in 0x58e88 (LE), which is the GKTS marker offset the T6030 rows pin. That may help as an extra layout check alongside the UUID.
Boot setup: the 14.7 stub boot chain, which is what the row names. Until the 4118 row landed I booted with asahi.g15_accept_text_prefix=0x411833a7… on the command line. That only changes admission, not the loaded bytes, and this boot doesn't use it: there's no "admitted by g15_accept_text_prefix" line.
Activity
handoff: 2026-10-28 user-support findings, omacom#247
jdlinventures commentedon Oct 28, 2026
I'm J514S #1 (cd05af6a…). I tested the asahi.g15_accept_text_prefix override on my machine, and it works for this build.
Config: options asahi g15_accept_text_prefix=0xcd05af6a828b52e4, with mesa-aurora / vulkan-aurora 26.3.0_devel.20261025.231095.7c86130d26c-1, kernel 7.2.11-8-1-ARCH.
Result:
GPU firmware build admitted by g15_accept_text_prefix: TEXT 0x5c000 DATA 0x114000 GKTS 0x58e88 sha256 cd05af6a828b52e4c2fed20086d3d18a7db42e03a8fd6f8354f9f06e0a61363e
M3: firmware identity accepted
board data: ADT performance states, 1380 MHz top (differs from the reference machine)
board data sha256 0f2fcd85a90cfff305c1121736033d36af612ca066b405e13e0b1cd046f55661
T6030 G15S: cores enabled per cluster: [7, 7]
G15: firmware at READY, starting the backend
/dev/dri/renderD128 is present and there are no faults in dmesg. glxinfo reports Apple M3 Pro (G15S B1), OpenGL 4.6, Mesa 26.3.0-devel. Hyprland CPU dropped from roughly 400% on llvmpipe to about 2%.
One observation on the question of what selects the image: my system-fw-version is 26.6.2 (os-fw-version 14.7, iBoot2 iBoot-10151.140.19.700.2, iBoot1 mBoot-18000.161.10). The PR #242 description reports system-fw-version 27.0 for the da26ecca… build on the same iBoot. That might be what separates the two, though two machines don't prove it.
Not tested yet: Vulkan, and sustained load. Happy to boot-test a patch without the override.
okhan1980 commentedon Oct 29, 2026
A third J514S TEXT, on the same boot chain as J514S #1 but with a different digest. The
g15_accept_text_prefixoverride works for it too.apple,j514s/apple,t6030, MacBook Pro 14" M3 Pro (Nov 2023)asahi,os-fw-versionasahi,system-fw-versionasahi,iboot1-versionasahi,iboot2-version411833a7ded773feeac4a181552e3ff5582eaf4435f2dd8f620c1ed0b31df1b06402a3f24b2399e3a97d9b93fd1d7fc1da07b4fc5091bc04a37d62da7bd453db, ADT performance states, 1380 MHz topversion=0x7031100 counts=0x110209 core-mask=0x7e5bd, cores per cluster[7, 7]The boot chain is identical to J514S #1's (14.7 stub,
iBoot-10151.140.19.700.2, system firmware 26.6.2), but the TEXT digest differs (cd05af6a…vs411833a7…). That supports your point that the boot chain alone doesn't select the TEXT image on T6030. It's probably the chip revision or something else per unit.Setup:
linux-aurora 7.2.11.aurora10-1,m1n1-aurora 1.6.1.aurora22-1,mesa-aurora/vulkan-aurora 26.3.0_devel.20261029.231476.672657e05e6-1, Hyprland 0.56.2. The override is in/etc/modprobe.d/asahi-g15.conf:Result:
/dev/dri/renderD128is present, and Hyprland/aquamarine and Quickshell render on it. The internal panel runs through DCP at 3024×1964@120. The kernel log shows no G15 faults, timeouts or recoveries so far, but this is early: only a few minutes of uptime when I checked.A side note for anyone switching an older omarchy-mac (legacy GRUB) install from
linux-asahitolinux-aurora: right after the package switch, m1n1'sboot.bincarried linux-asahi'st6030-j514s.dtb, not aurora's. I caught and fixed this before rebooting. Booted as-is, G15 would have had no device-tree node, independent of the firmware digest. Details in a separate packaging issue.jmayaalv commentedon Nov 5, 2026
First J516S (16" M3 Pro) report: the same TEXT build as J514S #1 (
cd05af6a…) is refused on this board too.apple,j516s/apple,t6030, MacBook Pro 16" M3 Pro (Nov 2023)asahi,os-fw-versionasahi,system-fw-versionasahi,iboot1-versionasahi,iboot2-versioncd05af6a828b52e4c2fed20086d3d18a7db42e03a8fd6f8354f9f06e0a61363eSetup:
linux-aurora 7.2.11.aurora14-1,m1n1-aurora 1.6.1.aurora22-1,mesa-aurora/vulkan-aurora 26.3.0_devel.20261031.231483.a1969111475-1, Hyprland 0.56.2.Kernel log:
No
renderD128, so Hyprland falls back to llvmpipe (~440% CPU at 3456×2234@120). The same boot chain (14.7 stub,iBoot-10151.140.19.700.2) and the same TEXT digest as J514S #1, on a different board and a different iBoot1 (mBoot-18000.121.3) and system firmware (26.5.2), which fits the observation that the boot chain alone doesn't select the TEXT image.I haven't tried the
g15_accept_text_prefix=0xcd05af6a828b52e4override yet; happy to boot-test it or a patch that pins this build for J516S.davidhorak commentedon Nov 6, 2026
First J516S (16" M3 Pro) report of the
411833a7…TEXT build, the same digest as the third J514S above. It's refused on this board too, on the same boot chain.apple,j516s/apple,t6030, MacBook Pro 16" M3 Pro (Nov 2023)asahi,os-fw-versionasahi,system-fw-versionasahi,iboot1-versionasahi,iboot2-versionasahi,m1n1-stage1-version/ stage2411833a7ded773feeac4a181552e3ff5582eaf4435f2dd8f620c1ed0b31df1b0Setup:
linux-aurora 7.2.11.aurora14-1,mesa-aurora 26.3.0_devel.20261031.231483.a1969111475-1, omarchy-mac 0.2.3-2, Omarchy 4.5.0-3.No
renderD128, so the session runs on llvmpipe. This makes two T6030 boards (J514S and J516S) on the same boot chain (14.7 stub, system firmware 26.6.2,mBoot-18000.161.10) loading411833a7…, while J516S has now been seen with both411833a7…andcd05af6a…. That fits the per-unit or per-revision selection theory, not a per-board one. Happy to boot-test a pin for this build on J516S.davidhorak commentedon Nov 6, 2026
Follow-up on the J516S
411833a7…report above: theg15_accept_text_prefixoverride works on this board.Config in
/etc/modprobe.d/asahi-g15.conf:Something to watch out for on Omarchy Mac installs: these boot a Limine UKI (
ENABLE_UKI=yesin/etc/default/limine), somkinitcpio -Palone doesn't apply the option. It only rebuilds/boot/initramfs-linux-aurora.img. You needlimine-mkinitcpioto rebuild the UKI on the ESP.Result:
/dev/dri/renderD128is present, and Hyprland renders on it on the internal panel (eDP-1, 3456×2234@120). Hyprland CPU dropped from about 440% on llvmpipe to about 3%. The kernel log shows no G15 faults, timeouts or recoveries, but it's early: only a few minutes of uptime when I checked, and I haven't tested Vulkan or sustained load.Setup:
linux-aurora 7.2.11.aurora14-1,m1n1-aurora 1.6.1.aurora22-1,mesa-aurora/vulkan-aurora 26.3.0_devel.20261031.231483.a1969111475-1, Hyprland 0.56.2. Happy to boot-test a pin for this build on J516S.drm/asahi: g15: pin the cd05 firmware TEXT on the J516S
paulnunezm commentedon Nov 8, 2026
Confirmed that the
g15_accept_text_prefix=0x411833a7ded773feoverride also works on my machine.After applying the override, rebuilding the boot images and rebooting, the kernel log reports that the firmware was admitted and the GPU backend reached READY. The system registers
renderD128. I haven't independently verified the active application renderer or tested sustained GPU workloads.dtt101 commentedon Nov 9, 2026
Another J514S loading the
cd05af6a…TEXT build, refused on the same boot chain.apple,j514s/apple,t6030, MacBook Pro 14" M3 Pro (Nov 2023)asahi,os-fw-versionasahi,system-fw-versionasahi,iboot1-versionasahi,iboot2-versioncd05af6a828b52e4c2fed20086d3d18a7db42e03a8fd6f8354f9f06e0a61363eSetup: Omarchy 4.5.0-3,
linux-aurora 7.2.11.aurora14-1,m1n1-aurora 1.6.1.aurora22-1,mesa-aurora 26.3.0_devel.20261031.231483.a1969111475-1.The probe log is identical to the one in the issue body. There's no
renderD128; aquamarine logsCDRMRenderer(drm): Can't create renderer, no matching devices found, and Hyprland sits at ~670% CPU on the internal panel (3024×1964@120).On what selects the image:
cd05af6a…has now been seen with system firmware 26.6.2 / mBoot-18000.161.10 (J514S #1) and 26.5.2 / mBoot-18000.121.3 (this machine and the J516S report). So neither system firmware nor iBoot1 seems to select it, which fits the per-unit or per-revision theory. Happy to boot-test a pin for this build on J514S.via Claude Opus 5.5
dtt101 commentedon Nov 9, 2026
Follow-up: the
g15_accept_text_prefix=0xcd05af6a828b52e4override works on this J514S (system firmware 26.5.2, mBoot-18000.121.3), set in/etc/modprobe.d/asahi-g15.confand rebuilt withlimine-mkinitcpio.renderD128is present and aquamarine renders onApple M3 Pro (G15S B1), OpenGL ES 3.2 Mesa 26.3.0-devel. Hyprland CPU dropped from ~670% on llvmpipe to ~5%. No G15 faults, timeouts or recoveries in the kernel log so far, but only a few minutes of uptime; Vulkan and sustained load not tested.via Claude Opus 5.5
djfergus commentedon Nov 10, 2026
Another J514S on
cd05af6a…: theg15_accept_text_prefix=0xcd05af6a828b52e4override works here too.apple,j514s, MacBook Pro 14" M3 Pro (Nov 2023)mBoot-18000.161.10,iBoot-10151.140.19.700.27d5f24b8…, 1380 MHz topcore-mask=0x6f5bd, cores per cluster[7, 7]linux-aurora 7.2.11.aurora14-1,mesa-aurora 26.3.0_devel.20261031With the override, G15 reaches READY,
renderD128appears, and Hyprland drops from about 300–600% CPU on llvmpipe to about 1%.Testing (all clean):
vulkaninfoandvkcubeOK (Borealis, Vulkan 1.4)There were 0 G15 faults, timeouts or recoveries across two boots (about 2 hours total), including a reboot through macOS. Not yet tested: sustained load longer than about 11 minutes.
Happy to boot-test a pin for this build.
drm/asahi: g15: pin J514S cd05 firmware TEXT
drm/asahi: g15: pin the cd05 firmware TEXT on the J516S
drm/asahi: g15: pin the 4118 firmware TEXT on the J514S and J516S
malik-na commentedon Nov 10, 2026
@jmayaalv thanks for the J516S report. Since it's the same
cd05af6a828b52e4…build that works with the override on J514S, could you boot-test the override on your machine?/etc/modprobe.d/asahi-g15.confwith:options asahi g15_accept_text_prefix=0xcd05af6a828b52e4
mkinitcpio -Palone isn't enough; runlimine-mkinitcpioas well (see davidhorak's comment above)./dev/dri/renderD128exists, and whatglxinforeports for the renderer (others gotApple M3 Pro (G15S B1)).vulkaninfo/vkcubeand some sustained load; nobody has covered J516Scd05af6a…on that yet.If it works, J516S will need its own pin row (on T6030 it's one row per board), and we can use your boot as the evidence for it.
LubomirGeorgiev commentedon Nov 11, 2026
Note
Drafted with Claude Opus 5.5. I read every line and checked the numbers on my own Mac before I posted it.
@malik-na this bears on the question in your issue body. The fused GPU core mask looks like the selector, not the chip revision.
Two reports here give both the mask and the digest:
core-mask0x7e5bd411833a7…0x6f5bdcd05af6a…Same board, same boot chain (os-fw 14.7, system-fw 26.6.2,
mBoot-18000.161.10,iBoot-10151.140.19.700.2). Different mask, different digest.The rest of the thread fits:
cd05af6a…appears on three units (jdlinventures, dtt101, djfergus), so the digest is not unique per unit.cd05af6a…on system-fw 26.5.2 and djfergus saw it on 26.6.2, so the firmware version does not select it.cd05af6a…with different board-data digests (0f2fcd85…,7d5f24b8…), so the perf-state table does not select it.version=0x7031100 counts=0x110209, which weakens the chip-revision theory in the issue body.If iBoot writes the binning value into TEXT before the handoff, the digest tracks the binning pattern. That gives a small finite set of images and explains all four points.
Two things would settle it. @okhan1980 @djfergus your two reports are the whole basis for this. Could you each post your full
core-maskline beside your digest? Two different digests behind one mask kill the idea. Several matched pairs confirm it, and the pins can then key on the mask. Could you also post the sha256 of your stub'sarmfw_g15s.im4p? One payload hash behind two different TEXT digests means iBoot patches the image at load time.macOS loads a different image
On this Mac, under macOS 27.0.1,
arm-io/gfx-asccarries three segments:Every pinned T6030 build has TEXT
0x5c000and DATA0x114000, andadt.rsrequires exactly two segments named__TEXT;__DATA. The parse refuses this ADT before it hashes anything. The image therefore belongs to the OS firmware generation that loads it.That firmware is a file, readable before any boot:
This machine, and what I cannot give you
Mac15,6,apple,j514s,apple,t6030)gpu-core-count = 14)compatible = "gpu,t6030"chip-id/chip-revision0x6030/0x11RELEASE_ARM64_T6030mBoot-20457.1.29The bin matters for the hypothesis. The 14-inch M3 Pro ships as 11 CPU cores with 14 GPU cores, or as 12 with 18. This is the first. Every report in this thread shows
cores enabled per cluster: [7, 7], so all of them are the 14-core GPU bin, yet their masks and digests differ. A report from an 18-core unit would say whether the digest splits along the bin, or only along the mask inside one bin.The model string does not separate the two bins.
t6030-j514s.dtssetsmodel = "Apple MacBook Pro (14-inch, M3 Pro, Nov 2023)"with no core count, whilet6034-j514m.dtsnames "14 CPU cores" for the M3 Max. A pin keyed on the model string would cover both M3 Pro bins without distinguishing them.I have not booted Linux on it, so I have no TEXT digest and no probe log. The numbers above come from the macOS 27 chain and pin nothing. The core-mask correlation rests on two reports that are not mine. I have not checked that an installed stub leaves
armfw_g15s.im4preadable, or that the payload is identical across Macs on one macOS build.Tell me which ADT property to read next and I will post it.
djfergus commentedon Nov 12, 2026
@LubomirGeorgiev full lines from this J514S, identical on two separate boots:
I haven't hashed the stub's
armfw_g15s.im4pyet; I'll post it once I can read it from macOS.djfergus commentedon Nov 12, 2026
@LubomirGeorgiev stub firmware from this J514S (TEXT
cd05af6a…,core-mask=0x6f5bd), as promised:armfw_g15s.im4p3063641b3f93f7c29b162af5aa4b17dec063ca9f870614c11fc377db0b5e04d9e9c8a66armfw_g15s.im4p, for comparison4077433ba2cef55be6c1f73674aa62f8b8ebb51e469c29f8836eb68ee011afe33d21e4cBoth were read from
<uuid>/restore/Firmware/agx/in each Preboot volume (the stub's directory is spelledRestore/), with the stub's mounted read-only from macOS. The stub had only that one copy.malik-na commentedon Nov 14, 2026
@jmayaalv the J516S row for your
cd05af6a…build has landed onintegration/aurora. It came in with #427, which carried Joel's J516S cd05 commit from #390 and the J516S 4118 row, so a J516S now has its own pin for bothcd05af6a…and411833a7…. It's not in a tagged release yet. The CI package from the merge iskernel-package-5bc7ad11033bafbc46d4c45bcdd61db5in https://github.com/omacom/linux-aurora/actions/runs/42754745532 (needs a GitHub sign-in to download). If you install it without theg15_accept_text_prefixline, you should seeGPU firmware build: J516S cd05 firmware (14.7 boot chain)and no "admitted by g15_accept_text_prefix". Could you post theasahi-g15lines and say whetherrenderD128is there? Thanks!@LubomirGeorgiev thanks for the core-mask idea. A few cases in the thread don't fit it:
cd05af6a…runs with[7, 7]cores per cluster on the J514S reports and[9, 9]on a J516S (drm/asahi: g15: pin the cd05 firmware TEXT on the J516S #390).411833a7…shows[7, 7]on okhan1980's J514S and[9, 9]on davidhorak's J516S. One digest, different core counts.125f0497…) and mlninja-dev (669913fb…) have the same boot chain (system firmware 26.6.2,mBoot-18000.161.10) and the samecore-mask=0xffffffffff, but different TEXT.So I don't think the mask picks the image. maralcbr's byte-diff plan (repeat the hash over cold boots, diff the TEXT dumps and the Preboot payloads) is the next step: #289 (comment). If you can dump TEXT from a unit, that would help.
Since all three T6030 digests are now pinned for both boards, the title here is out of date. Could a maintainer retitle it, say "G15: what selects the GPU firmware TEXT on T6030/T6031? Different digests on the same boot chain", or close it and keep the investigation on #289? Either is fine with me.
Last thing, for the M3 Max (T6031): digest
669913fb…on the J516C has no row yet. That's tracked in #289. #340 (the third J516C pin,125f0497…) is approved but conflicted withintegration/aurora; I've rebased it locally and will push once Scott confirms the shape.LubomirGeorgiev commentedon Nov 15, 2026
Note
Drafted with Claude Opus 5. I read every line and reproduced every number myself before I posted it.
The driver does not hash the firmware Apple ships. The boot loader rewrites TEXT first, so the pinned digest records the loader's output, not the firmware. That is why one identical file gives
da26ecca…,cd05af6a…and411833a7…on different Macs.@malik-na you were right to drop the core-mask idea, and #390 settled it. This replaces it.
I pulled the stub firmware straight out of Apple's asset and measured it. The 14.8.3 OS firmware that upstream
asahi-installernames inIPSW_VERSIONSforj514sapholds three GPU firmware files underAssetData/boot/Firmware/agx/. HTTP range requests fetch them in about 1.3 MB.armfw_g15s.im4pb3f93f7c29b162af5aa4b17dec063ca9f870614c11fc377db0b5e04d9e9c8a66armfw_g15c.im4pf404aac59e5bd6539e0727745e374bdfcd08032c204ea940f09d74c8dc96f6c6armfw_g15g.im4pea448437f188d1a87d0b4f27ebe7cb3060b2b1d46d995657fa3d1dcd8c144773@djfergus the
g15ssize and digest are what you read from your own stub. So the file is generic. Every Mac on this stub reads identical bytes, and the payload is not personalised.The file holds several complete builds
The IM4P payload is
rkosftab, a table. Each entry is a whole Mach-O image with its ownLC_UUID. The segment layouts are the ones already pinned.armfw_g15s.im4p, T6030, TEXT0x5c000, DATA0x114000, marker0x58e88:LC_UUIDa0000cbb17af-783d-367b-9ced-d5910944731e7975bf4c2b2b1b1548193cca567bf16e2a33d90161b780f7d452b07abbe1802cb000dbf37c40-ead5-3760-9441-99501878295506a4e7e0f3ac34b3845c122ea3bdc65d178c18aca206f88d692158f022903269armfw_g15c.im4p, T6031, TEXT0x64000, DATA0x140000, marker0x5cf28, matching the #289 reports:LC_UUIDa0000f22e68b-3667-3c2e-a823-3315fadd3a6a53342914acdba6e6b178737cf21831639d82a8638c7e4f9266b1bd0bc49b6579b000f9bea4a3-d65b-386d-b884-ecf2df7b159eebbec4aff2032e131efb7efb9f9727250e864f53db6e95f5f21538973cd073c5armfw_g15g.im4p, T8122, TEXT0x64000, DATA0xb0000, marker0x5d54a, holds three entries:a000,b000andc000. The entry count is not fixed per chip.The proof that TEXT is rewritten
g15/hw.rs:152says of the row pinned as measured TEXTda26ecca…:Entry
b000above carries exactly that UUID. Hashed by the rule inidentity.rs, with the 8 bytes after the singleGKTS\x08\0\0\0marker zeroed, its TEXT is06a4e7e0…, notda26ecca….Same image. File TEXT
06a4e7e0…. Measured TEXTda26ecca….None of the six file digests above appears anywhere in
drivers/gpu/drm/asahi/. The images are the same on every Mac; the patch is not.The GKTS entropy bytes are not the mechanism either. They are the constant
ff0a000000000000in all seven images across the three files, andidentity.rszeroes them before hashing.So the pin table cannot converge. Each Mac configuration that patches differently needs a new row, and three entries in
armfw_g15g.im4ppredict up to three digest families onj504,j613andj615per firmware generation before any patch variation.What I suggest, and what is still open
The
LC_UUIDsits inside TEXT, at offset0x4214in all seven images. It is therefore inside the segment the driver already copies, and it differs per entry. Matching those 16 bytes with the segment sizes and the marker offset would identify the build exactly and admit every unit of a known build, whatever the loader patched elsewhere. That is the identity the comment above already names. @maralcbr this looks like the normalisation you proposed in #289, with a field that is bounded and explained.Two things I could not measure, both needing a booted Linux, so they belong to whoever can dump TEXT:
hw.rssuggests it does not.On macOS the running system names the entry it was given:
This Mac (
apple,j514s,chip-revision 0x11) answersb000on both macOS 27.0.1 and 27.2.1, across the system firmware change frommBoot-20457.1.29tomBoot-20457.41.3. The OS changed the file and the digest; the entry stayed.I have two small scripts for this: one that range-extracts a member from the remote asset, one that prints each entry's layout, UUID and TEXT digest. Would a gist help, or would you rather they came as a PR against
tools/asahi/?@malik-na I agree the title is stale. "What rewrites the GPU firmware TEXT on T6030/T6031" would fit what the thread now shows.
LubomirGeorgiev commentedon Nov 17, 2026
Note
Drafted with Claude Opus 5. I read every line and reproduced every number myself before I posted it.
@ilyasakin can you print 16 bytes from your J514S? #481 has the GPU bound on a kernel you built, so you are set up for it.
The pinned digests are not digests of the firmware Apple ships. Apple's generic
armfw_g15s.im4pholds two complete T6030 builds, and neither one's TEXT digest appears underdrivers/gpu/drm/asahi/. The boot chain writes into TEXT before the driver hashes it, so each pin records one machine's loaded bytes, not a build. Measurements earlier in this issue; #487 has the same result for T6031.I proposed matching the build's
LC_UUIDinstead. Its 16-byte value sits inside the hashed TEXT at0x4214in all seven 14.x stub images I measured, and covers none of the regions the loader rewrites. Whether the loader leaves those 16 bytes alone is untested.In
drivers/gpu/drm/asahi/g15/identity.rs, infirmware():Hexanddev_info!are in scope.0x4214is far from the GKTS marker, so the zeroing below does not reach it.Apple's file carries these two UUIDs at that offset:
One of them means the loader leaves the UUID alone, and a UUID match plus
text_size,data_sizeand the marker offset can replace the digest pins on T6030. Anything else means the loader rewrites the UUID too, and this does not work.Also: which row does your
GPU firmware build:line name? That shows how many units map to each pin.I have no Linux on this Mac, because the T6030 rows are in no release.
ilyasakin commentedon Nov 17, 2026
Done, on my J514S (MacBook Pro 14" M3 Pro), with your print added to
firmware()on my daily branch:The first 16 bytes are
dbf37c40-ead5-3760-9441-995018782955, theb000entry ofarmfw_g15s.im4p. So the loader leaves the UUID alone on this unit, and this unit maps to theJ514S 4118row.The bytes after it,
00000000 00000000 888e0500, end in0x58e88(LE), which is the GKTS marker offset the T6030 rows pin. That may help as an extra layout check alongside the UUID.Boot setup: the 14.7 stub boot chain, which is what the row names. Until the 4118 row landed I booted with
asahi.g15_accept_text_prefix=0x411833a7…on the command line. That only changes admission, not the loaded bytes, and this boot doesn't use it: there's no "admitted by g15_accept_text_prefix" line.drm/asahi: g15: pin J514S cd05 firmware TEXT
drm/asahi: g15: pin the cd05 firmware TEXT on the J516S
drm/asahi: g15: pin the 4118 firmware TEXT on the J514S and J516S