Repository navigation
Description
Summary
A real install was recently confused by a leftover Windows EFI partition. That failure is not merely uncovered by our tests — with the current harness it cannot be covered, because every install E2E creates a fresh empty disk and boots without UEFI firmware at all.
Filing this so the gap is tracked rather than rediscovered on the next install.
What the harness actually does
tests/iso/e2e.sh:
- Blank disk, every run.
qemu-img create -f qcow2 "$install_disk"— the guest never sees a partition table it didn't create. - BIOS only.
-machine q35with no-bios OVMFand no-drive if=pflash, so the guest boots SeaBIOS. The kickstart matches:bootloader --location=mbr. - No NVRAM coverage, which follows directly — there is no EFI variable store in the guest to have entries in.
So zerombr and clearpart --all --initlabel are only ever asserted against a disk with nothing on it to clear. The one scenario that bit a user in production — an existing ESP from another OS — is the scenario the harness cannot construct.
Three distinct things are untested as a result:
- An existing ESP being detected, reused, or reformatted.
- Stale
efibootmgrentries pointing at a bootloader that no longer exists. - Any interaction between our partitioning and a pre-existing GPT.
Why it's worth fixing now rather than later
The pieces already exist in open PRs; they're just in different ones:
- feat(bootc-iso): experimental bootc-installer ISO variant (resurrected from #61) #71 sets up a proper OVMF pflash pair with a per-run writable
OVMF_VARScopy. That is a genuine UEFI/NVRAM path — the hard part is done. It still creates a blank 64G qcow2 and hardcodesDISK="/dev/vda", so thediskandencryptionsteps its ownrecipe.jsonexposes never run against a populated disk. - fix(ci): restore ISO rename output and soft-fail arm64 legs #70 adds
qemu-nbdplumbing and proves the disk is genuinely encrypted (blkid | grep -c crypto_LUKS) rather than silently plaintext. That is exactly the machinery needed to pre-seed an image.
Suggested fix: a second matrix leg that pre-seeds a GPT + FAT32 ESP into the qcow2 before boot, using #71's OVMF firmware and #70's qemu-nbd plumbing. Cheap, and it converts "structurally untestable" into "one more leg".
One nit that compounds the problem today: in bootc-iso/justfile,
[[ -f "{{ luks-qemu-disk }}" ]] || qemu-img create …
reuses an existing image. So locally, a dirty disk only ever arises by accident and nondeterministically — the worst of both worlds, since it neither guarantees a clean run nor tests a dirty one.
Related: the tests disappear instead of failing
pull-request.yml gates both E2E jobs behind needs: build-iso-stable. On #69 they report conclusion: skipped — when the build breaks, the install tests don't go red, they vanish. A PR can look green while the only tests that touch a disk never executed.
Same pattern as dakota-iso#137. Worth decoupling so a skipped install test is visibly not a passing one.
Acceptance
- An E2E leg boots under UEFI (OVMF) with a persistent per-run
OVMF_VARS. - That leg starts from a disk pre-seeded with a GPT + FAT32 ESP, not a blank image.
- The run asserts the resulting boot entries are correct and no stale entry survives.
- A skipped install-test job cannot be mistaken for a passing one.
Activity
juanresendiz813 commentedon Nov 18, 2026
Another real-world case for this, from ublue-os/bluefin#5073: a reinstall of regular Bluefin on a MacBook Pro (UEFI) failed with a non-descriptive error while creating the GRUB config, because /EFI/fedora from the previous Bluefin install was still on the ESP. It only worked after deleting /EFI/fedora by hand. So it isn't only another OS's ESP: reinstalling over our own leftover entry fails too. A UEFI test that installs twice onto the same disk would catch it.