Skip to content

Conversation

@lidel
Copy link
Member

@lidel lidel commented Oct 9, 2025

No description provided.

gammazero and others added 6 commits October 8, 2025 18:28
* upgrade go-ds-pebble to v0.5.2
* Change pebble option `MaxConcurrentCompactions` to `CompactionConcurrencyRange`

(cherry picked from commit 9370004)
This fixes a panic that happens when reusing a batch go-datastore Batch.

(cherry picked from commit 2b5adee)
* test: add migration tests for Windows and macOS

- add dedicated CI workflow for migration tests on Windows/macOS
- workflow triggers on migration-related file changes only

* build: remove redundant go version checks

- remove GO_MIN_VERSION and check_go_version scripts
- go.mod already enforces minimum version (go 1.25)
- fixes make build on Windows

* fix: windows migration panic by reading config into memory

fixes migration panic on Windows when upgrading from v0.37 to v0.38
by reading the entire config file into memory before performing atomic
operations. this avoids file locking issues on Windows where open files
cannot be renamed.

also fixes:
- TestRepoDir to set USERPROFILE on Windows (not just HOME)
- CLI migration tests to sanitize directory names (remove colons)

minimal fix that solves the "panic: error can't be dealt with
transactionally: Access is denied" error without adding unnecessary
platform-specific complexity.

* fix: set PATH for CLI migration tests in CI

the CLI tests need the built ipfs binary to be in PATH

* fix: use ipfs shutdown for graceful daemon termination in tests

replaces platform-specific signal handling with ipfs shutdown command
which works consistently across all platforms including Windows

* fix: isolate PATH modifications in parallel migration tests

tests running in parallel with t.Parallel() were interfering with each
other through global PATH modifications via os.Setenv(). this caused
tests to download real migration binaries instead of using mocks,
leading to Windows failures due to path separator issues in external tools.

now each test builds its own custom PATH and passes it explicitly to
commands, preventing interference between parallel tests.

* chore: improve error messages in WithBackup

* fix: Windows CI migration test failures

- add .exe extension to mock migration binaries on Windows
- handle repo lock file properly in mock migration binary
- ensure lock is created and removed to prevent conflicts

* refactor: align atomicfile error handling with fs-repo-migrations

- check close error in Abort() before attempting removal
- leave temp file on rename failure for debugging (like fs-repo-15-to-16)
- improves consistency with external migration implementations

* fix: use req.Context in repo migrate to avoid double-lock

The repo migrate command was calling cctx.Context() which has a hidden
side effect: it lazily constructs the IPFS node by calling GetNode(),
which opens the repository and acquires repo.lock. When migrations then
tried to acquire the same lock, it failed with "lock is already held by us"
because go4.org/lock tracks locks per-process in a global map.

The fix uses req.Context instead, which is a plain context.Context with
no side effects. This provides what migrations need (cancellation handling)
without triggering node construction or repo opening.

Context types explained:
- req.Context: Standard Go context for request lifetime, cancellation,
  and timeouts. No side effects.
- cctx.Context(): Kubo-specific method that lazily constructs the full
  IPFS node (opens repo, acquires lock, initializes subsystems). Returns
  the node's internal context.

Why req.Context is correct here:
- Migrations work on raw filesystem (only need ConfigRoot path)
- Command has SetDoesNotUseRepo(true) - doesn't need running node
- Migrations handle their own locking via lockfile.Lock()
- Need cancellation support but not node lifecycle

The bug only appeared with embedded migrations (v16+) because they run
in-process. External migrations (pre-v16) were separate processes, so
each had isolated state. Sequential migrations (forward then backward)
in the same process exposed this latent double-lock issue.

Also adds repo.lock acquisition to RunEmbeddedMigrations to prevent
concurrent migration access, and removes the now-unnecessary daemon
lock check from the migrate command handler.

* fix: use req.Context for migrations and autoconf in daemon startup

daemon.go was incorrectly using cctx.Context() in two critical places:

1. Line 337: migrations call - cctx.Context() triggers GetNode() which
   opens the repo and acquires repo.lock BEFORE migrations run, causing
   "lock is already held by us" errors when migrations try to lock

2. Line 390: autoconf client.Start() - uses context for HTTP timeouts
   and background updater lifecycle, doesn't need node construction

Both now use req.Context (plain Go context) which provides:
- request lifetime and cancellation
- no side effects (doesn't construct node or open repo)
- correct lifecycle for HTTP requests and background goroutines

(cherry picked from commit f4834e7)
keep -dev version from master
@lidel lidel requested a review from a team as a code owner October 9, 2025 00:37
@lidel lidel added the skip/changelog This change does NOT require a changelog entry label Oct 9, 2025
@lidel lidel merged commit 40c027a into master Oct 9, 2025
1 check passed
@lidel lidel deleted the merge-release-v0.38.1 branch October 9, 2025 00:49
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

skip/changelog This change does NOT require a changelog entry

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants