Skip to content

Restore Vulkan on Macs that lost it in the Mesa 26.2 split - #617

Closed
oliverlukschander wants to merge 1 commit into
omacom:quattro-upstreamfrom
oliverlukschander:fix/vulkan-asahi-mesa-split
Closed

oliverlukschander wants to merge 1 commit into
omacom:quattro-upstreamfrom
oliverlukschander:fix/vulkan-asahi-mesa-split

Conversation

@oliverlukschander

@oliverlukschander oliverlukschander commented Sep 27, 2026 •

Copy link
Copy Markdown

Summary

Mesa 26.2 moved the Asahi Vulkan ICD out of mesa into the split package vulkan-asahi. #470 reports two outcomes on Macs that relied on the old provide: the update aborts when something depends on vulkan-driver, or it succeeds and Vulkan is gone.

What quattro-upstream already covers: since 562a1f4, packages/omarchy-mac/README.md names vulkan-asahi as a runtime dependency of the add-on, and the Apple package list leaves it to that dependency. The omarchy-mac recipe in omacom/omarchy-pkgs doesn't declare it yet.

This PR covers the rest, following the rework described for #489 in docs/apple-silicon-quattro-pr-triage.md:

  • migrations/1790490111.sh: detector-gated backfill for Macs that already lost Vulkan. It does nothing off Apple Silicon, when no repository has the package, or while vulkan-asahi is installed or still provided by an older mesa.
  • bin/omarchy-upgrade-to-quattro: the Apple Vulkan check read lspci … Apple, which never lists the Apple GPU (it's a platform device). It now reads the device tree, because this script runs before the detector is installed.

Not in this PR: the recipe dependency itself. The omarchy-mac recipe lives in omacom/omarchy-pkgs, and per the triage doc only a real dependency resolves inside the update's own -Syu, since migrations run after it. Until the recipe declares it, and on Macs that already lost Vulkan, the migration is what brings vulkan-asahi back.

Rebased onto quattro-upstream 351c9fb (2026-09-28). The README sentence this PR added is dropped: 562a1f4 documents the dependency upstream.

Left alone on purpose: migrations/1784401744.sh keeps its lspci gate. Naming the detector there makes test/shell.d/apple-pacman-repo-test.sh fail ("the repository migration sorts before every Apple Silicon migration"), and the new migration covers the backfill.

Refs #470, #489

Testing

  • New test/shell.d/vulkan-asahi-migration-test.sh (4 cases: off Apple Silicon, package unavailable, already installed or provided, installs when missing). Passes.
  • On the branch rebased onto quattro-upstream 351c9fb: the full test/all, packages/omarchy-mac/test/all and the boot package suite, sandboxed on Arch Linux ARM with omarchy-pkgs and omarchy-iso checkouts. packages/omarchy-mac/test/all passes. Five test/shell.d files and one boot package check fail, identically on an unmodified 351c9fb in the same sandbox; nothing new fails.
  • Not tested on hardware: the migration was not run on a Mac. For reference, this Mac14,10 (fresh 4.0.3rc4 install) has mesa 1:26.2.3-1 and vulkan-asahi 1:26.2.3-1 as separate packages.

Investigated and fixed with Claude Code; I can walk through any of it in review.

🤖 Generated with Claude Code

Mesa 26.2 moved the Asahi Vulkan ICD out of mesa into vulkan-asahi.
Macs that relied on the old provide either abort the update (anything
depending on vulkan-driver) or lose Vulkan silently (omacom#470).

Since 562a1f4 the add-on names vulkan-asahi as a runtime dependency,
so Macs get it with omarchy-mac once its recipe carries it. This covers
the rest:

- A detector-gated migration restores vulkan-asahi on Macs that already
  lost it. While an older mesa still provides it, nothing happens.
- The upgrade script's Apple Vulkan check reads the device tree; the
  Apple GPU is a platform device and lspci never lists it.

migrations/1784401744.sh keeps its lspci gate: it is an older upstream
migration, already stamped, and naming the detector there would break
the ordering lint in apple-pacman-repo-test.sh.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@oliverlukschander
oliverlukschander force-pushed the fix/vulkan-asahi-mesa-split branch from 22d30da to 308a7f9 Compare September 28, 2026 11:03
@maralcbr

Copy link
Copy Markdown
Member

Thanks for this, Oliver, and for tracking the Mesa split down so carefully. Since #658, vulkan-asahi is a dependency of the omarchy-mac package (omacom/omarchy-pkgs#672), so every Mac gets it back with its next update, and that replaces the backfill migration. Your finding that the lspci … Apple checks never match is going into omacom#13362, with credit to you. Closing this one in favour of those.

@maralcbr maralcbr closed this Sep 28, 2026
@oliverlukschander

Copy link
Copy Markdown
Author

Thanks, happy it was useful

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants