Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 1 addition & 18 deletions .changeset/ebusy-retries.md
Original file line number Diff line number Diff line change
Expand Up @@ -2,21 +2,4 @@
"watchpack": patch
---

fix: retry `fs.lstat` on transient `EBUSY` errors instead of flagging the
file as removed (fixes #223, #44).

On Windows it is common for anti-virus scanners, indexers or the editor
itself to briefly hold an exclusive handle on a file that has just
changed. Before this change the watcher would receive the `fs.watch`
event, call `lstat`, get back `EBUSY`, and fall through to `setMissing`
— causing a spurious `remove` event and in some cases leaving the
watcher unable to see further changes for that file until the directory
was re-scanned.

`DirectoryWatcher` now retries `lstat` up to three times (100 ms apart)
before giving up, and does not emit a remove when the only reason the
file could not be stat'd was `EBUSY`.

The retry count is controlled by the `WATCHPACK_RETRIES` environment
variable (default: `3`; set to `0` or `"false"` to disable retrying and
restore the previous behaviour).
fix: retry `fs.lstat` on transient `EBUSY` errors instead of emitting a spurious `remove` event (fixes #223, #44). The retry count is controlled by the `WATCHPACK_RETRIES` environment variable (default: `3`; set to `0` or `"false"` to disable retrying).
8 changes: 1 addition & 7 deletions .changeset/perf-ignored-reduceplan.md
Original file line number Diff line number Diff line change
Expand Up @@ -2,10 +2,4 @@
"watchpack": patch
---

perf: skip the path-separator replacement when the input has no backslash
(benchmarks measure ~35–45% less time for `ignored` matchers on POSIX paths),
fast-path single-element `ignored` arrays, and make `reducePlan`'s selection
loop walk only structurally valid candidates with an early exit when the
ideal reduction is found (measured ~20–40% faster on medium and large
plans). Adds a tinybench suite under `bench/` and a CodSpeed GitHub Actions
workflow so future regressions are caught automatically.
perf: speed up `ignored` matchers (~35–45% faster on POSIX paths) and `reducePlan` (~20–40% faster on medium and large plans). Also adds a tinybench suite under `bench/` and a CodSpeed GitHub Actions workflow to catch future regressions.
30 changes: 1 addition & 29 deletions .changeset/silence-permission-scan-errors.md
Original file line number Diff line number Diff line change
Expand Up @@ -2,32 +2,4 @@
"watchpack": patch
---

fix: don't log "Watchpack Error (initial scan)" for unreadable entries
inside a watched parent directory (fixes #187).

Webpack registers every ancestor of a watched file as a watched
directory (so `/mnt/c/Users/me/proj` causes watchpack to scan `/mnt/c`,
`/mnt`, `/`). When such a parent contains entries the current process
can't `lstat` — `pagefile.sys` / `hiberfil.sys` on WSL, `/efi` on Linux
when the EFI partition isn't mounted, protected paths on Node ≥22.17 on
Windows where libuv now reports `EINVAL` instead of `EACCES` — the
initial scan would print:

Watchpack Error (initial scan): Error: EACCES: permission denied, lstat '/mnt/c/pagefile.sys'
Watchpack Error (initial scan): Error: EINVAL: invalid argument, lstat 'C:\\hiberfil.sys'
Watchpack Error (initial scan): Error: ENODEV: no such device, lstat '/efi'

These entries aren't actually being watched (only their sibling, e.g.
`/mnt/c/Users`, is) so the log was harmless but very noisy and sent a
lot of users on wild goose chases.

`DirectoryWatcher#doScan` now treats `EACCES` / `ENODEV` (and `EINVAL`
on Windows) the same way it already treats `EPERM` / `ENOENT` /
`EBUSY`: the offending entry is recorded as missing and the scan
continues silently. The same set is applied to the `readdir` error path
on the watched directory itself, so an unreadable mount point is
treated as removed instead of logged.

No public API change. If you were relying on the error appearing on
stderr, set the impacted entry up as an explicit watch so a real failure
on it surfaces through the existing `error` event instead.
fix: don't log "Watchpack Error (initial scan)" for unreadable entries inside a watched parent directory, e.g. `pagefile.sys` on WSL or `/efi` on Linux (fixes #187). `EACCES` / `ENODEV` (and `EINVAL` on Windows) errors are now handled like `EPERM` / `ENOENT` / `EBUSY`: the entry is recorded as missing and the scan continues silently.
22 changes: 1 addition & 21 deletions .changeset/symlink-cycle-guard.md
Original file line number Diff line number Diff line change
Expand Up @@ -2,24 +2,4 @@
"watchpack": patch
---

fix: prevent unbounded watcher growth when a symlinked directory points
back to one of its own ancestors (cycle protection for the
`followSymlinks: true` symlink-descent path).

The recent fixes for #190 / #231 made `DirectoryWatcher` follow
symlinked directories whose realpath lives outside the watched real
directory, registering them as nested watched directories. That logic
short-circuits when the symlink target is a sibling in the same parent
(`dirname(realPath) === this.path`), but it does not catch the case
where the target is an _ancestor_ of the symlink itself — for example
`a/b/loop -> ..`. In that case `readdir` followed the symlink, found
the original tree again, and a new `DirectoryWatcher` was created at
each recursion level until the path exceeded `PATH_MAX` (locally
observed: ~1500 watchers within 2 s, ~2500 within 2.5 s).

`DirectoryWatcher` now computes `path.relative(realPath, itemPath)`
before descending; if the relative path doesn't start with `..` and
isn't absolute (i.e. the symlink target is at-or-above the symlink in
the directory tree), the symlink is recorded as a plain entry instead
of being descended into. Behaviour for symlinks pointing outside the
watched tree (the case #231 fixes) is unchanged.
fix: prevent unbounded watcher growth when a symlinked directory points back to one of its own ancestors (e.g. `a/b/loop -> ..`) with `followSymlinks: true`. Such symlinks are now recorded as plain entries instead of being descended into; symlinks pointing outside the watched tree (#231) behave as before.
Loading