amdgpu: cannot set modified EDID binaries. "Requesting EDID firmware ... failed (err=-2)"

[soupglasses]

Describe the bug

When using drm.edid_firmware on NixOS, amdgpu fails to find the relevant firmware file specified, even if it exists. Loading the EDID manually from /run/current-system/firmware works correctly.

% journalctl -b0 | grep edid
Jan 09 06:05:22 desktop kernel: Command line: initrd=\efi\nixos\3r63xc80969amx1mldhh18jn39807i4r-initrd-linux-6.6.8-initrd.efi init=/nix/store/ix3vqm06g4s8s97yb2givdx0yas5zwaq-nixos-system-desktop-24.05.20240109.dirty/init drm.edid_firmware=edid/MSI-1462.bin mitigations=auto split_lock_detect=off loglevel=4
Jan 09 06:05:22 desktop kernel: Kernel command line: initrd=\efi\nixos\3r63xc80969amx1mldhh18jn39807i4r-initrd-linux-6.6.8-initrd.efi init=/nix/store/ix3vqm06g4s8s97yb2givdx0yas5zwaq-nixos-system-desktop-24.05.20240109.dirty/init drm.edid_firmware=edid/MSI-1462.bin mitigations=auto split_lock_detect=off loglevel=4
Jan 09 06:05:22 desktop kernel: amdgpu 0000:03:00.0: Direct firmware load for edid/MSI-1462.bin failed with error -2
Jan 09 06:05:22 desktop kernel: amdgpu 0000:03:00.0: [drm] *ERROR* [CONNECTOR:121:DP-2] Requesting EDID firmware "edid/MSI-1462.bin" failed (err=-2)

Steps To Reproduce

NixOS configuration should have the following set somewhere:

{pkgs, lib, ...}: let
  edid-custom = pkgs.runCommandNoCC "edid-custom" {} ''
    mkdir -p $out/lib/firmware/edid
    base64 -d > "$out/lib/firmware/edid/custom.bin" <<EOF
    <insert your base64 encoded EDID file here `base64 < /sys/class/drm/card0-.../edid`>
    EOF
  '';
in {
  boot.kernelParams = ["drm.edid_firmware=edid/custom.bin"];
  hardware.firmware = [edid-custom];
}

My EDID file looks like this for comparison:

AP///////wA2aWIUAQAAABMdAQSlUCF4O44FrU8zsCYNUFSvzwCBgIGolQCVKLMAsyjR6NHAQ9Bw
oNCgKVAwIDUAHU0xAAAaAAAA/QAwZKKiAgEAAAAAAAAAAAAA/ABNU0kgTUFHMzQxQ1EKgvBwUNCg
UVAIIPgACLBBAAAaATICAxh1S5CfhZQEEwMSAQIRIwkHB4MBAAACOoDQcjgtQBAsNYAdTTEAABj+
fHCg0KApUDAgNQAdTTEAABhWXgCgoKApUDAgNQAdTTEAAB46nQCgoKApUDBANQAdTTEAAB6uYIAY
cTgtQDAsNQAdTTEAABgAAAAAAAAAAAAAAAAAEg==

1. Rebuild and switch configuration.
2. Reboot.
3. See edid errors in journalctl -b0.

Expected behavior

Applies the EDID file correctly on boot.

Applying it manually, for example as sudo bash -c "unxz -dc /run/current-system/firmware/edid/custom.bin.xz > /sys/kernel/debug/dri/0/DP-2/edid_override" to apply it to DP-2 works correctly. This can work as a pseudo workaround, but requires you to disconnect and reconnect your monitor to apply the modified edid file.

Additional context

Open Discourse about the issue: https://discourse.nixos.org/t/copying-custom-edid/31593

Attempted to look for upstream issues in amdgpu, but nothing seemed relevant. Google only shows other NixOS users having this issue that i can find.

Notify maintainers

Unsure, @RaitoBezarius as latest editor of stage-1.nix.
@flokli @nazarewk may have insight as maintainers of edid-generator.nix.

Metadata

Please run nix-shell -p nix-info --run "nix-info -m" and paste the result.

% nix-shell -p nix-info --run "nix-info -m"

 - system: `"x86_64-linux"`
 - host os: `Linux 6.6.8-zen1, NixOS, 24.05 (Uakari), 24.05.20240109.dirty`
 - multi-user?: `yes`
 - sandbox: `yes`
 - version: `nix-env (Nix) 2.18.1`
 - nixpkgs: `/nix/store/yxld2qw84ka591fay1xkh5jflbgxjqa1-source`

---

Add a 👍 reaction to issues you find important.

[nixos-discourse]

This issue has been mentioned on NixOS Discourse. There might be relevant details there:

https://discourse.nixos.org/t/copying-custom-edid/31593/24

[nazarewk]

Not sure it is the same issue, but I had to resort to some ugly hacks that I had no time to straigthen out:
https://github.com/nazarewk-iac/nix-configs/blob/48e93a44b9cca863868294db38036d3debd980da/modules/hardware/edid/default.nix

I'm on full-amd (CPU & GPU) myself and at the end got it to work just fine eventually.

As far as I remember I had 2 issues:

1. hardware.firmware gets compressed by default and it cannot be loaded because of that
2. hardware.firmware does not make /lib/firmware/edid available to the initrd, this is most likely your issue

I think this is the critical part to fix point 2

    # TODO: remove this "fix" at some point?
    #   see https://github.com/NixOS/nixpkgs/blob/5e4c2ada4fcd54b99d56d7bd62f384511a7e2593/nixos/modules/system/boot/systemd/initrd.nix#L383-L385
    boot.initrd.systemd.contents."/lib/firmware".source =
      let
        # "/lib/modules".source = "${modulesClosure}/lib/modules";
        modulesClosure = lib.strings.removeSuffix "/lib/modules" config.boot.initrd.systemd.contents."/lib/modules".source;

        joined = pkgs.symlinkJoin {
          name = "kdn-initrd-firmware";
          paths = [ modulesClosure generated ];
        };
      in
      lib.mkForce "${joined}/lib/firmware";

[nazarewk]

thanks for motivating me to finally dig into the issue and upstream changes :)

[nazarewk]

note (for now?) you will need to add compressFirmware = false; to pkgs.runCommandNoCC "edid-custom" {} to make it work, otherwise it will not be included in initrd.

[soupglasses]

Applied https://github.com/NixOS/nixpkgs/commit/11e9e9bb70f8d09494ccbfbc7a6192bf733356ec.patch to my nixpkgs. It did not help. Also doing compressFirmware = false; together with it also did not help.

The boot.initrd.systemd.contents workaround also seems to not work, and does not even generate a new binary (likely because they are equivalent).

Running binwalk --verbose --term --rm against the built -initrd.efi only shows amdgpu and GenuineIntel being added to the initrd. Amdgpu comes from a linux-6.6.8-modules-shrunk which I find curious.

[nazarewk]

indeed, i must have tested it the wrong way

[soupglasses]

@nazarewk I found something interesting! While the boot fails:

% journalctl -b0 | grep edid
Jan 09 14:10:13 desktop kernel: Command line: initrd=\efi\nixos\00grk6kdy8rmcsga2lpl070hfg7w2s4m-initrd-linux-6.6.8-initrd.efi init=/nix/store/pkaql82r8ga0iad8048mg4c1j9qhhxs6-nixos-system-desktop-24.05.20240109.dirty/init drm.edid_firmware=edid/msi1462.bin mitigations=auto split_lock_detect=off loglevel=4
Jan 09 14:10:13 desktop kernel: Kernel command line: initrd=\efi\nixos\00grk6kdy8rmcsga2lpl070hfg7w2s4m-initrd-linux-6.6.8-initrd.efi init=/nix/store/pkaql82r8ga0iad8048mg4c1j9qhhxs6-nixos-system-desktop-24.05.20240109.dirty/init drm.edid_firmware=edid/msi1462.bin mitigations=auto split_lock_detect=off loglevel=4
Jan 09 14:10:13 desktop kernel: amdgpu 0000:03:00.0: Direct firmware load for edid/msi1462.bin failed with error -2
Jan 09 14:10:13 desktop kernel: amdgpu 0000:03:00.0: [drm] *ERROR* [CONNECTOR:121:DP-2] Requesting EDID firmware "edid/msi1462.bin" failed (err=-2)

Disconnecting and reconnecting the display after boot has finished does apply the custom EDID file without logging to journalctl.

[soupglasses]

https://github.com/nazarewk/nixpkgs/commit/44b70ebd07798d94130e87cc3edd948fd1f8aebe.patch

Does function correctly for me and allows it to apply without reconnecting my monitor after boot. Thank you for figuring it out! ❤️

[nazarewk]

Disconnecting and reconnecting the display after boot has finished does apply the custom EDID file without logging to journalctl.

I am quite sure it "just works" because on a live system it's available at ls -la /run/current-system/firmware/edid/ while still missing from initrd (during boot), it gave me quite a headache when originally debugging.

[nazarewk]

there is also second PR removing need to pass compressFirmware #279828

[soupglasses]

Here's my commit with a functional EDID loaded from initrd 🎉
soupglasses/coturnix@2c22165

[RaitoBezarius]

I can review PRs but I cannot help much here, apologies.