Related: koverstreet/bcachefs#429
May be best to compile as built-in for now.

Ready for review.

cc @davidak

We may want to wait on merging actually, I'm not sure if it's a problem with 5.18 in general or if it is because we aren't patching anymore, but I'm getting the kernel not being able to work around duplicate NSIDs for NVME drives.

Patching problem is with block/bio.c, no idea why I didn't see that before.
Not sure how to properly work around this, but working on it.

I'm out of my depth again. Can't figure anything out even with any derivation of:

{
  name = "bcachefs-${commit}";

  patch = fetchpatch {
    name = "bcachefs-${commit}.diff";
    url = "https://evilpiepirate.org/git/bcachefs.git/rawdiff/?id=${commit}&id2=v${lib.versions.majorMinor kernel.version}";
    sha256 = diffHash;

    postFetch = ''
      ${pkgs.buildPackages.patchutils}/bin/filterdiff -x 'a/block/bio.c' "$out" > "$tmpfile"
      mv "$tmpfile" "$out"
    '';
};

Looks like the duplicate NSID error is a problem with 5.18 in general.

At this point, we may want to consider getting rid of the linux bcachefs testing package and waiting for it to go into mainline.
AFAIK all the other bcachefs stuff (bcachefs-tools, tests, tasks) is fine, but it seems to be near-impossible to continue supporting this.

I managed to get your PR working when I changed it to the latest Kent has and compiled it as built in instead of a module. Not sure whether this is helpful but I would be glad if you would keep this PR open and we would keep up with the effort however futile it seems.

Thanks for doing this work on this PR!

Last time I tried it was an issue with patching. IIRC compiling it as built-in was required now anyway due to a bug.

This worked for me

{ lib
, pkgs
, fetchpatch
, kernel
, date ? "2022-07-03"
, commit ? "04962c7b97e783d4b705616a858af141cc6ba878"
, diffHash ? "sha256-zsieai/29Ry4LsJQuwMaxtdYdIJdEJCusg016ovUM2Q="
, kernelPatches # must always be defined in bcachefs' all-packages.nix entry because it's also a top-level attribute supplied by callPackage
, argsOverride ? {}
, ...
} @ args:

# NOTE: bcachefs-tools should be updated simultaneously to preserve compatibility
(kernel.override ( args // {
  argsOverride = {
    version = "${kernel.version}-bcachefs-unstable-${date}";

    extraMeta = {
      branch = "master";
      maintainers = with lib.maintainers; [ davidak Madouura ];
    };
  } // argsOverride;

  kernelPatches = [
    {
      name = "bcachefs-${commit}";

      patch = fetchpatch {
        name = "bcachefs-${commit}.diff";
        url = "https://evilpiepirate.org/git/bcachefs.git/rawdiff/?id=${commit}&id2=v${lib.versions.majorMinor kernel.version}";
        sha256 = diffHash;

        postFetch = ''
          ${pkgs.buildPackages.patchutils}/bin/filterdiff -x 'a/block/bio.c' "$out" > "$tmpfile"
          mv "$tmpfile" "$out"
        '';
      };

      extraConfig = "BCACHEFS_FS y";
    }

    {
      # Needed due to patching failure otherwise
      name = "linux-bcachefs-bio.c-fix";
      patch = ./linux-bcachefs-bio.c-fix.patch;
    }
  ] ++ kernelPatches;
}))

The only changes to your PR are the compiling in and a never commit from Kent's repo.

I'm unsure as to why the patching is working now, but it is.
I'll build, run the tests, and hopefully it'll be ready for review in an hour.

error: builder for '/nix/store/6qkiyxdrj4p2p76qsxlmrcx3v9qg8n6k-linux-5.18.13-bcachefs-unstable-2022-07-25.drv' failed with exit code 2;
       last 10 log lines:
       >   CC [M]  drivers/gpu/drm/nouveau/nouveau_led.o
       >   CC [M]  drivers/gpu/drm/nouveau/nvc0_fence.o
       >   CC [M]  drivers/gpu/drm/nouveau/nv84_fence.o
       >   CC [M]  drivers/gpu/drm/nouveau/nv50_fence.o
       >   CC [M]  drivers/gpu/drm/nouveau/nv17_fence.o
       >   CC [M]  drivers/gpu/drm/nouveau/nv10_fence.o
       >   LD [M]  drivers/gpu/drm/amd/amdgpu/amdgpu.o
       >   LD [M]  drivers/gpu/drm/nouveau/nouveau.o
       >   AR      drivers/gpu/built-in.a
       >   AR      drivers/built-in.a
       For full logs, run 'nix log /nix/store/6qkiyxdrj4p2p76qsxlmrcx3v9qg8n6k-linux-5.18.13-bcachefs-unstable-2022-07-25.drv'.

Last time this was a problem, it was solved by a kernel update. Not sure what to do in the meantime.

Turns out I was compiling a previous source derivation, diffHash didn't give me an error.
Let's see how it goes now.

Builds, tests succeed. Ready for review.
I did have to change from metadata replicas to replicas (metadata replicas and data replicas) due to the bcachefsMulti test failing every third restart. Seems to be a write error leading to a trans path overflow, but that's on kent's side.
The reason why this change works is because the disks are now in a RAID 1-like mode, and can cover for the error.

I tested this PR on my machine and it builds and works nicely.

@ofborg eval

Looks like 5.19 fixed the weirdness.
Tests pass, had to temporarily remove "zfs" from nixos/modules/profile/base.nix to bypass ZFS not being updated to 5.19 yet.

@SuperSandro2000 Are we good for merge?

Successfully created backport PR #194905 for release-22.05.