Repository navigation
Windows: fs applies POSIX semantics where libuv differs (6 measured divergences) #10408
Description
Five of the 22 Windows-specific gap failures identified in #10403 are one
cluster: Perry's fs applies POSIX semantics where libuv deliberately behaves
differently on Windows. Each is a small, independent mapping fix, and all six
divergences below are measured against the pinned oracle (Node 26.5.1, Windows
11 / MSVC).
Measured divergences
| fixture | operation | Node | Perry |
|---|---|---|---|
test_gap_fs_errprop_2735plus |
symlink to a dangling target |
EPERM |
EIO |
test_gap_fs_errprop_2735plus |
truncate on a directory |
EINVAL |
EACCES |
test_gap_fs_errprop2_2745plus |
symlinkSync unprivileged |
throws EPERM (errno -4048) and exits |
does not throw |
test_gap_fs_fd_2749 |
fchown on a closed fd |
OK |
EBADF fchown |
test_gap_fs_write_error_propagation |
write to an invalid fd (sync, cb, string and buffer) |
EBADF write |
EACCES write |
test_gap_9574_unlink_waits_for_parked_write |
unlink while a write is parked |
unlink=ENOENT final=exists |
unlink=ok final=missing |
Notes on the individual cases
-
fchownreturningOKon a closed fd is not a Node bug.chown/fchown
have no meaning on Windows, so libuv no-ops them and never reaches an fd
validity check. Perry validates the descriptor first and reportsEBADF. -
symlinkrequires privilege on Windows (admin, or Developer Mode).
Unprivileged, libuv reportsEPERM; Perry reportsEIOin one fixture and
does not fail at all in the other, so the two fixtures likely reach different
code paths and both want checking. -
unlinkof an open file is the most semantically loaded one. Windows
refuses to remove a file with an open handle; Node surfaces that and the file
survives, while Perry removes it. That is a real behavioural difference, not
just an errno label — code that relies on the file still existing after a
failed unlink will diverge. -
writeto an invalid fd wantsEBADF, notEACCES— Perry appears to be
reporting the Windows access error rather than mapping the descriptor failure.
Related
Same family as the err.errno mapping fixed in PR #10403, where libuv's errno
is the negated OS errno on Linux/macOS but its own -4xxx space on Windows.
These are the fs-side equivalents: the code rather than the number, and in
two cases the operation's success/failure itself.
Reproduce: PERRY_SKIP_BUILD=1 PERRY_BIN=… PERRY_RUNTIME_DIR=… ./run_parity_tests.sh --filter test_gap_fs_
then diff test-parity/output/node/<t>.txt against
test-parity/output/perry/<t>.txt. Clear target/perry-auto-* first (#10404).
Activity