backend,disk: make partition ops work on busy block devices#411
Merged
Conversation
Contributor
Author
|
@deitch could you take a look when you get a chance? This splits the two block-device fixes (non-exclusive whole-disk open + BLKPG re-read fallback) out of my downstream resizer work into go-diskfs, now with a root-gated loop-device regression test for the BLKPG path. Thanks! |
eriknordmark
force-pushed
the
block-device-partition-ops
branch
from
45a14f8 to
da6fccc
Compare
Two fixes needed to drive e2fsck/resize2fs on the partitions of a live
block device (surfaced running partitionresizer against a real /dev/nbd0,
which the file-based tests cannot exercise):
- file.OpenFromPathWithExclusive lets a caller open a device read-write
WITHOUT O_EXCL while still recording the path. OpenFromPath keeps its
O_EXCL default. Holding the whole disk O_EXCL makes the kernel reject
O_EXCL opens of its child partitions ("device is in use"), so a tool
that shells out to e2fsck/resize2fs cannot hold the disk exclusively.
- ReReadPartitionTable falls back to per-partition BLKPG reconciliation
when BLKRRPART fails (EBUSY during the udev re-probe a table write
triggers, or a mounted sibling partition). BLKPG adds/removes/recreates
only the changed entries via ioctls, with no whole-disk exclusive
re-read and no dependency on external tools (partx). Build-verified;
the partx-based predecessor was validated end-to-end on /dev/nbd0,
the BLKPG path still needs the same on-device validation.
Signed-off-by: eriknordmark <erik@zededa.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Exercise the BLKPG re-read fallback on a real block device, which the file-based tests structurally cannot reach. The test builds a GPT image, exposes it via losetup, and mounts one partition so a whole-disk BLKRRPART is guaranteed to fail EBUSY; it asserts that precondition directly, then rewrites the table to shrink a different partition and confirms the kernel picked up the change through per-partition BLKPG while the mounted sibling was left untouched. The device is opened non-exclusively via OpenFromPathWithExclusive, the same way a consumer that shells out to e2fsck/resize2fs must. Requires root; skipped otherwise. Signed-off-by: eriknordmark <erik@zededa.com> Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
eriknordmark
force-pushed
the
block-device-partition-ops
branch
from
da6fccc to
3910579
Compare
When BLKRRPART fails with EBUSY -- a sibling partition is mounted, i.e. we are repartitioning the disk we booted from -- and the BLKPG per-partition fallback also cannot reconcile the in-use partition, the on-disk table that Partition() already wrote is nonetheless committed. Wrap that case in the new exported ErrReReadDeferred sentinel so a caller can errors.Is it and reboot to apply the table on the next boot, instead of treating a committed-but-not-yet-live table as a hard failure. Non-EBUSY re-read failures keep returning the existing hard error. Signed-off-by: eriknordmark <erik@zededa.com> Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The O_EXCL whole-disk open default makes the kernel refuse the O_EXCL open that resize2fs/e2fsck do on a child partition, so a consumer that holds the disk open while running those tools is blocked. The file.OpenFromPathWithExclusive non-exclusive open fixes this, but only the BLKPG re-read fallback had block-device coverage. Add a root-gated loop-device regression test that proves both directions in one run: a parent held O_EXCL blocks the child's O_EXCL open with EBUSY, while a non-exclusive parent leaves it openable. A regular file cannot reach this (O_EXCL is a no-op and there are no child partition nodes), so the test is skipped without root. Signed-off-by: eriknordmark <erik@zededa.com> Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
eriknordmark
added a commit
to eriknordmark/partitionresizer
that referenced
this pull request
Jun 18, 2026
When a partition-table commit reaches the disk but the kernel cannot be made to re-read it live -- the boot disk is busy because the partition being changed is mounted, as when repartitioning the disk we booted from -- go-diskfs now reports disk.ErrReReadDeferred. The table is already on disk, so translate that sentinel at each commit site into the exported ErrRebootToApply, letting a caller reboot to apply the table on the next boot instead of treating a committed-but-not-yet-live table as a failure. Pin go-diskfs to the fork commit that adds ErrReReadDeferred via a replace directive; the replace is temporary and will be dropped once diskfs/go-diskfs#411 merges. Signed-off-by: eriknordmark <erik@zededa.com> Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Collaborator
|
You have been pushing commits. Is this one ready to be reviewed? |
Contributor
Author
Yes. No more commits planned. |
eriknordmark
added a commit
to eriknordmark/eve
that referenced
this pull request
Jun 18, 2026
Vendor the go-diskfs and partitionresizer changes the offline repartition relies on: a non-exclusive block-device open (so the resizer can shell out to e2fsck/resize2fs on child partitions), per-partition BLKPG re-read when BLKRRPART is busy, and the ErrReReadDeferred / ErrRebootToApply sentinels that let a busy boot disk be repartitioned and applied on the next boot. Both deps are pinned to fork commits via replace, pending the upstream PRs diskfs/go-diskfs#411 and diskfs/partitionresizer#15. Signed-off-by: eriknordmark <erik@zededa.com> Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Collaborator
|
Overall, it looks good. Happy to have the extra test and handle the case. I am a little concerned that this is the first test that requires root to run, but it handles it with the skip correctly. So I am good with it. |
deitch
approved these changes
Jun 18, 2026
eriknordmark
added a commit
to eriknordmark/partitionresizer
that referenced
this pull request
Jun 19, 2026
Bump github.com/diskfs/go-diskfs to the 2026-06-18 master tip, which now carries the ErrReReadDeferred change from diskfs/go-diskfs#411, and remove the temporary `replace => github.com/eriknordmark/go-diskfs` directive that was added to pin the pre-merge fork commit. The replace was always meant to be dropped once #411 landed in master; it had inadvertently stayed in tree. go mod tidy also prunes the stale go.sum entries left from the fork pin. Signed-off-by: eriknordmark <erik@zededa.com> Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
eriknordmark
added a commit
to eriknordmark/partitionresizer
that referenced
this pull request
Jun 19, 2026
Bump github.com/diskfs/go-diskfs to the 2026-06-18 master tip, which now carries the ErrReReadDeferred change from diskfs/go-diskfs#411, and remove the temporary `replace => github.com/eriknordmark/go-diskfs` directive. go mod tidy also prunes the stale go.sum entries left from the fork pin. The replace was only ever meant to pin the pre-merge fork commit and be dropped once #411 landed in master. It was added in diskfs#15, where Claude broke its own rule: a temporary fork-pinning replace must be dropped before the PR merges, never carried into main. diskfs#15 merged with it still in tree, leaving upstream main depending on a personal fork; this commit removes it. Signed-off-by: eriknordmark <erik@zededa.com> Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Two fixes needed to run partition tools (
e2fsck/resize2fs) against thepartitions of a live block device. The file-based tests cannot reach either
path, so both were surfaced running a downstream resizer against a real
/dev/nbd0.file.OpenFromPathWithExclusivelets a caller open a device read-writewithout
O_EXCLwhile still recording the path (OpenFromPathkeeps itsO_EXCLdefault). Holding the whole diskO_EXCLmakes the kernel rejectO_EXCLopens of its child partitions ("device is in use"), so a tool thatshells out to
e2fsck/resize2fscannot hold the disk exclusively.ReReadPartitionTablefalls back to per-partition BLKPGreconciliation when
BLKRRPARTfails — EBUSY during the udev re-probe atable write triggers, or a mounted sibling partition. BLKPG adds/removes/
recreates only the changed entries via ioctls, with no whole-disk exclusive
re-read and no dependency on external tools (
partx).On a regular file
O_EXCLis a no-op and there are no partition nodes tore-read, which is why the existing file-based tests never exercise either
path. Two root-gated loop-device tests cover the fixes:
disk/disk_blockdev_linux_test.goexercises the BLKPG fallback: it mountsone partition to force
BLKRRPARTEBUSY, shrinks another, and verifies thekernel picks up the change while the mounted sibling is left untouched.
disk/disk_openexclusive_linux_test.goexercises the non-exclusive open: aparent held
O_EXCLblocks anO_EXCLopen of its child partition (thee2fsck/resize2fscase) with EBUSY, while a non-exclusive parent leavesthe child openable.