Move creating gpu blobs to m1n1. #466

WhatAmISupposedToPutHere · 2025-05-11T16:54:14Z

Instead of doing gpu initialization partially in m1n1 and partially in kernel, build all the blobs in m1n1 and put them in the device tree to be picked up by the kernel. Most of the code here is copied and pasted kernel code, with some fixups

svenpeter42 · 2025-05-13T16:38:45Z

I vaguely remember that bsd has some issues(?) with building rust code. we should check if this would create additional pain for them and then potentially allow building without rust (which means no gpu support then ofc).

svenpeter42 · 2025-05-13T17:10:15Z

looks like this also breaks the CI because https://src.fedoraproject.org/rpms/m1n1/raw/main/f/m1n1-rust-deps.patch doesn't apply anymore. no idea where that file comes from though or why we need it.

@davide125 @Conan-Kudo

davide125 · 2025-05-13T17:19:16Z

It's here: https://src.fedoraproject.org/rpms/m1n1/blob/rawhide/f/m1n1-rust-deps.patch . We need it because in Fedora we build against the system packages, not the vendored ones.

svenpeter42 · 2025-05-13T17:59:23Z

<kettenis> regarding m1n1 and rust on OpenBSD; yes that is a huge problem since you can't easily build your own standard library for a cross compiler

let's not break m1n1 stage2 builds for all openbsd users. so we either need to convert this all to C (ugh...) or allow builds without rust and gpu support and make sure we don't shot ourselves in the foot with that

svenpeter42

Lina's sign-off is missing, my other comment is optional and can be ignored. I also didn't really review the rust code because it's been copied from the kernel and we know that it works

src/kboot.c

WhatAmISupposedToPutHere · 2025-05-13T21:55:33Z

Lina's sign-off is missing, my other comment is optional and can be ignored. I also didn't really review the rust code because it's been copied from the kernel and we know that it works

It is not a 1 to 1 copy, but i have a special kernel build that compares the kernel blob with m1n1 one, and those matched on t6000

svenpeter42 · 2025-05-13T22:09:24Z

Let's see if jannau wants to review all this magic since he's more familiar with both rust and the gpu driver but I'm fine with merging it

jannau

First review of the C code

src/kboot.c

src/kboot_gpu.c

jannau · 2025-05-14T22:13:55Z

src/kboot_gpu.c

+int32_t rust_gpu_initdata_size(uint32_t compat_maj, uint32_t compat_min, size_t *data_a,
+                               size_t *data_b, size_t *globals);
+
+int32_t rust_fill_gpu_initdata(size_t perf_state_table_count, size_t perf_state_count,


can we use a struct for this?

We can, i just find structs to be a bit inconvenient to keep in sync with rust/c

but keeping 22 function arguments in sync is convenient? I find the the function argument variant harder to validate and would assume it's more likely to attract bugs. We could use bindgen.

no, they are both terrible tbh. i'll look into bindgen

jannau · 2025-05-14T22:15:01Z

src/kboot_gpu.c

+                               size_t n_perf_states_afr, const struct aux_perf_state *pstates_afr,
+                               uint32_t is_j180, uint32_t base_pstate, uint64_t uat_ttb_base,
+                               uint32_t num_cores, uint32_t compat_maj, uint32_t compat_min,
+                               uint32_t gpu_rev_id, void *data_a, void *data_b, void *globals);


I'd feel more comfortable if we pass the size data_a, data_b and globals as well.

The caller is supposed to allocate exactly as much as rust_gpu_initdata_size asked for, so not sure what value that adds?

src/kboot.c

jannau's going to take care of this one since he actually knows both the gpu driver and rust

Booting with Linux 6.17-rc1 device trees fails due to missing "operating-points-v2" properties. Check for the downstream only "apple,agx-t*" compatible strings and disable the gpu node if those are not found. This serves as temporary fix until AsahiLinux#466 is merged. Signed-off-by: Janne Grunau <j@jannau.net>

Booting with Linux 6.17-rc1 device trees fails due to missing "operating-points-v2" properties. Check for the downstream only "apple,agx-t*" compatible strings and disable the gpu node if those are not found. This serves as temporary fix until #466 is merged. Signed-off-by: Janne Grunau <j@jannau.net>

Booting with Linux 6.17-rc1 device trees fails due to missing "operating-points-v2" properties. Check for the downstream only "apple,agx-t*" compatible strings and disable the gpu node if those are not found. This serves as temporary fix until AsahiLinux#466 is merged. Signed-off-by: Janne Grunau <j@jannau.net>

jannau

With fixes from https://github.com/jannau/m1n1/commits/bootloader-cal-blobs/ GPU initdata matches on j274, j375c, j375d, j473, j474s, j493 and j180d. I've pushed verification changes to asahi-wip. Missing devices to verify are M1 Pro/Max laptops and M2 Max (laptop and studio).

The branch contains fixes for some of the review comments and depends on #499. Feel free to cherry-pick and squash.

Besides the review comments I'd like to have the code in kboot_gpu.c split into 3 parts:

common
downstream compatibles
upstream compatibles

but that can happen after this PR is merged

Makefile

.gitignore

src/kboot.c

src/kboot.h

Makefile

WhatAmISupposedToPutHere · 2025-09-23T22:54:15Z

m1 pro is verified on my side

from_path_trace inserts an extra root node, which when roundripped via C resutls in a None entry, but breaks using it directly from rust. Replace it with a None Signed-off-by: Sasha Finkelstein <fnkl.kernel@gmail.com>

Instead of doing gpu initialization partially in m1n1 and partially in kernel, build all the blobs in m1n1 and put them in the device tree to be picked up by the kernel. Most of the code here is copied and pasted kernel code, with some fixups Signed-off-by: Asahi Lina <lina@asahilina.net> Co-authored-by: Sasha Finkelstein <fnkl.kernel@gmail.com> Signed-off-by: Sasha Finkelstein <fnkl.kernel@gmail.com>

Signed-off-by: Sasha Finkelstein <fnkl.kernel@gmail.com>

The reserved-memory nodes have a page size aligned size. Signed-off-by: Janne Grunau <j@jannau.net>

This mirrors the DT in the downstream AsahiLinux kernel. j274 misses the "gpu-perf-base-pstate" property in the ADT. j474s uses 1 which is unusual for non-battery powered devices. Using the same values as the downstream DT simplifies comparing Linux and m1n1 initialized GPU initdata. Signed-off-by: Janne Grunau <j@jannau.net>

Based on the downstream Linux DT. Values are in the ADT so they could be read from there but that would be a larger refactoring. Signed-off-by: Janne Grunau <j@jannau.net>

WhatAmISupposedToPutHere requested review from jannau and svenpeter42 as code owners May 11, 2025 16:54

WhatAmISupposedToPutHere force-pushed the bootloader-cal-blobs branch 3 times, most recently from 339d779 to e90dd1d Compare May 12, 2025 09:08

svenpeter42 reviewed May 13, 2025

View reviewed changes

src/kboot.c Show resolved Hide resolved

WhatAmISupposedToPutHere force-pushed the bootloader-cal-blobs branch from 4e61475 to bd07db6 Compare May 13, 2025 21:54

svenpeter42 previously approved these changes May 13, 2025

View reviewed changes

jannau reviewed May 14, 2025

View reviewed changes

WhatAmISupposedToPutHere force-pushed the bootloader-cal-blobs branch 2 times, most recently from 3d6d5de to b131656 Compare May 15, 2025 22:05

WhatAmISupposedToPutHere force-pushed the bootloader-cal-blobs branch 3 times, most recently from 881a53b to 55c9824 Compare May 31, 2025 15:26

WhatAmISupposedToPutHere force-pushed the bootloader-cal-blobs branch 2 times, most recently from c2ae886 to e0ccac2 Compare June 13, 2025 20:48

jannau mentioned this pull request Aug 13, 2025

kboot_gpu: Fix booting with v6.17-rc1 upstream DT #486

Merged

WhatAmISupposedToPutHere force-pushed the bootloader-cal-blobs branch 2 times, most recently from aeb9983 to 46e70a4 Compare August 22, 2025 19:25

jannau reviewed Sep 23, 2025

View reviewed changes

WhatAmISupposedToPutHere force-pushed the bootloader-cal-blobs branch from 46e70a4 to 495680b Compare September 23, 2025 22:52

WhatAmISupposedToPutHere force-pushed the bootloader-cal-blobs branch from 495680b to 2eaed13 Compare September 23, 2025 23:00

WhatAmISupposedToPutHere and others added 6 commits September 24, 2025 09:54

adt: Fix using get_reg_container from rust side

ca0a3a7

from_path_trace inserts an extra root node, which when roundripped via C resutls in a None entry, but breaks using it directly from rust. Replace it with a None Signed-off-by: Sasha Finkelstein <fnkl.kernel@gmail.com>

Silence dead code warning on dt_transfer_virtios

6cb90a9

Signed-off-by: Sasha Finkelstein <fnkl.kernel@gmail.com>

kboot_gpu: Save exact hwdata/globals size for starlight-debug

852c9ef

The reserved-memory nodes have a page size aligned size. Signed-off-by: Janne Grunau <j@jannau.net>

rust: gpu: Override some M1/M2 Max Studio HW config

c8ae45f

Based on the downstream Linux DT. Values are in the ADT so they could be read from there but that would be a larger refactoring. Signed-off-by: Janne Grunau <j@jannau.net>

WhatAmISupposedToPutHere force-pushed the bootloader-cal-blobs branch from 2eaed13 to c8ae45f Compare September 24, 2025 07:56

Uh oh!

Move creating gpu blobs to m1n1. #466

Are you sure you want to change the base?

Move creating gpu blobs to m1n1. #466

Uh oh!

Conversation

WhatAmISupposedToPutHere commented May 11, 2025

Uh oh!

svenpeter42 commented May 13, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

svenpeter42 commented May 13, 2025

Uh oh!

davide125 commented May 13, 2025

Uh oh!

svenpeter42 commented May 13, 2025

Uh oh!

svenpeter42 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

WhatAmISupposedToPutHere commented May 13, 2025

Uh oh!

svenpeter42 commented May 13, 2025

Uh oh!

jannau left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

jannau May 14, 2025

Choose a reason for hiding this comment

Uh oh!

WhatAmISupposedToPutHere May 14, 2025

Choose a reason for hiding this comment

Uh oh!

jannau May 14, 2025

Choose a reason for hiding this comment

Uh oh!

WhatAmISupposedToPutHere May 14, 2025

Choose a reason for hiding this comment

Uh oh!

jannau May 14, 2025

Choose a reason for hiding this comment

Uh oh!

WhatAmISupposedToPutHere May 14, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

jannau left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

WhatAmISupposedToPutHere commented Sep 23, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

svenpeter42 commented May 13, 2025 •

edited

Loading