🔴 Live TV: denoland/deno/pull/

divybot · 2026-05-19T19:34:33Z

Summary

Implement real node:wasi preview1 compatibility for the 20 WASI tests in the Node compat suite that were previously absent from tests/node_compat/config.jsonc. All 20 now pass on Linux.
Add 20 new entries to config.jsonc for the newly-passing tests.

The implementation work is in ext/node/ops/wasi.rs and ext/node/polyfills/wasi.ts; the suite entries are added at the bottom of config.jsonc. No existing entries were flipped — every test added here is newly green.

What changed in node:wasi

ENOTCAPABLE for sandbox escape. cant_dotdot / symlink_escape expect errno == ENOTCAPABLE (76); we previously returned ERRNO_PERM (63). The starts_with check now maps to NOTCAPABLE.
Symlink loop detection. Probe raw_os_error() for ELOOP so we don't depend on the unstable ErrorKind::FilesystemLoop variant.
path_open / path_filestat_get / path_readlink follow flags. Honor LOOKUPFLAGS_SYMLINK_FOLLOW for path_open / path_filestat_get, and force no-follow on the final component for path_readlink so std::fs::read_link doesn't reject the resolved target.
futimens. fd_filestat_set_times and path_filestat_set_times use the filetime crate, with full FSTFLAGS_ATIM(_NOW) / FSTFLAGS_MTIM(_NOW) semantics.
path_link. Added (std::fs::hard_link) and wired through the JS polyfill so link(2) works.
poll_oneoff. Implements clock subscriptions (relative and abstime) plus immediate fd_read/fd_write events. For stdin, the EVENT_FD_READWRITE_HANGUP bit is set so wasi-libc's poll() shim returns POLLHUP|POLLIN as test-wasi-poll.js requires.
stdio fd routing. The WASI constructor's stdin/stdout/stderr options now do something: fds >= 3 are dup'd into an owned std::fs::File and used in place of the inherited process streams. Closing the user's original fd afterwards doesn't affect ours.
sock_accept. Returns ERRNO_BADF for unknown fds and ERRNO_NOTSOCK for non-socket fds in our table (not ERRNO_NOSYS), matching Node/uvwasi semantics.
ExperimentalWarning on construction. Match Node's "WASI is an experimental feature and might change at any time" warning so the suite's common.expectWarning('ExperimentalWarning', ...) doesn't fail the very first test.
finalizeBindings. Node-only API used by wasi:thread-spawn workers to bind a shared WebAssembly.Memory without re-running _start. test-wasi-pthread needs it.

Runner fixup

The vendored Node suite is missing the empty test/fixtures/wasi/subdir directory (git doesn't track empty dirs). test-wasi-readdir asserts that the preopen has four entries including subdir. The runner now creates it before enumerating tests, with a docstring explaining why.

Test plan

All 25 WASI tests in tests/node_compat/runner/suite/test/wasi/ pass via cargo test --test node_compat 'test-wasi-'
tests/unit_node/wasi_test.ts (added tests for finalizeBindings and the ExperimentalWarning) passes
tools/format.js --check, tools/lint.js --rs, and tools/lint.js --js all clean

Platform-specific notes

The 20 candidate tests are bounded WASI behavior — no host-OS-specific syscalls aside from the dup/_dup paths in stdio routing. The Linux run exercises all 20; macOS should behave identically. Windows path canonicalization / hard_link / symlink behavior is consistent with what we already exercise elsewhere in ext/node/ops/wasi.rs, so I expect the same wins on macOS and Windows but only Linux was verified in this session.

Closes denoland/orchid#150

Bring node:wasi closer to Node's behavior so the WASI subgroup of the Node compat suite passes. Twenty previously-absent tests are now wired up in tests/node_compat/config.jsonc and pass on Linux: test-wasi-cant_dotdot, test-wasi-clock_getres, test-wasi-exitcode, test-wasi-fd_prestat_get_refresh, test-wasi-ftruncate, test-wasi-getentropy, test-wasi-getrusage, test-wasi-gettimeofday, test-wasi-io, test-wasi-main_args, test-wasi-notdir, test-wasi-poll, test-wasi-preopen_populates, test-wasi-pthread, test-wasi-readdir, test-wasi-sock, test-wasi-stat, test-wasi-stdio, test-wasi-symlinks, test-wasi-write_file Implementation highlights: * Sandbox-escape paths now return ERRNO_NOTCAPABLE (76) instead of ERRNO_PERM, matching Node/uvwasi. Symlink loops surface as ERRNO_LOOP via raw_os_error so we don't rely on the unstable ErrorKind::FilesystemLoop variant. * path_open honors LOOKUPFLAGS_SYMLINK_FOLLOW; path_filestat_get and path_readlink stop following the final symlink when callers ask for lstat / readlink semantics, so readlink no longer ENOTCAPABLEs or returns the canonical target. * fd_filestat_set_times and path_filestat_set_times implemented on top of the filetime crate, including FSTFLAGS_*_NOW handling, so futimens() in test-wasi-stat.js works. * path_link added (hardlink), wired through the JS polyfill. * poll_oneoff implemented for clock + fd_read/fd_write subscriptions, including the EVENT_FD_READWRITE_HANGUP bit for stdin so libc poll() returns POLLHUP|POLLIN as test-wasi-poll.js expects. * WASI stdio fd routing: the constructor honors user-supplied stdin/stdout/stderr fds (e.g. from openSync) by duping the host fd into an owned std::fs::File, leaving the user's original fd intact. * sock_accept returns ERRNO_BADF/ERRNO_NOTSOCK instead of ERRNO_NOSYS so accept() on a non-socket fd matches Node's errno. * WASI constructor emits the same ExperimentalWarning Node emits, so the suite's common.expectWarning() helper is satisfied. * WASI.finalizeBindings() polyfill added; used by the test fixture's wasi:thread-spawn workers to bind a shared memory without re-running _start, enabling test-wasi-pthread. The vendored Node suite is missing an empty fixtures/wasi/subdir directory (git can't track empty dirs). The test runner now restores it before enumerating tests so test-wasi-readdir sees the expected four entries. Closes denoland/orchid#150 Co-Authored-By: Divy Srivastava <me@littledivy.com>

test-wasi-symlinks (create_symlink subtest) fails on Windows because the WASI program calls symlink("./input-in-subdir.txt", linkpath) and NTFS reparse points store the target verbatim. Windows then can't resolve `./input-in-subdir.txt` when path_open follows the link, so fopen returns NULL on line 18 of create_symlink.c. Normalize forward slashes to backslashes before calling symlink_file/symlink_dir, and pick symlink_dir for directory targets so resolution doesn't break on directory symlinks either. Co-Authored-By: Divy Srivastava <me@littledivy.com>

…link on Windows After normalizing the symlink target to backslashes in path_symlink so Windows can resolve relative links, path_readlink was returning the backslash form. test-wasi-symlinks's create_symlink subtest stores "./input-in-subdir.txt" and asserts the readlink result matches byte for byte, so the backslash form failed strcmp. Reverse the normalization on Windows when reading the link back so WASI callers see the POSIX-style target they originally stored. Co-Authored-By: Divy Srivastava <me@littledivy.com>

littledivy · 2026-05-20T03:37:29Z

@divybot resolve conflicts

# Conflicts: # Cargo.lock # ext/node/Cargo.toml

Audit fixes so node:wasi can't be used to bypass --allow-read / --allow-write: * path_open now requires Deno write permission when any mutating right is requested (write data, set times, set size, allocate, sync, datasync, plus the path_* mutating rights). RIGHTS_FD_WRITE alone was the previous trigger, but RIGHTS_FD_FILESTAT_SET_TIMES would let a caller utimens a file through a Read-only handle (utimensat succeeds on read-only fds for the file owner). * path_open hands out exactly the requested rights, intersected with the maximum permitted for the entry type. Previously every File got the full RIGHTS_FILE union and every Dir got RIGHTS_DIR regardless of what was asked, so any open implied every right. * Mutating fd_* ops (fd_write, fd_pwrite, fd_filestat_set_size, fd_filestat_set_times, fd_allocate, fd_sync, fd_datasync) and the read ops (fd_read, fd_pread, fd_readdir) now check the fd's stored rights and return ERRNO_NOTCAPABLE when the bit is missing. Stdio and HostFile entries don't carry rights and continue to bypass the check. * make_stdio_entry only accepts user-supplied stdio fds that node:fs has registered in deno_io::FdTable. A raw OS fd inherited from the parent process or from internal Deno opens is rejected so the WASI constructor can't dup arbitrary host fds. * Separated needs_write_perm (Deno gate) from wants_os_write (open mode) so a SET_TIMES-only request doesn't force OS write mode on the underlying file. All 25 WASI node_compat tests continue to pass on Linux. Co-Authored-By: Divy Srivastava <me@littledivy.com>

bartlomieju

Nice work — 20 new WASI compat tests passing is a real win, and the security tightening in the last commit (rights intersection at path_open, RIGHTS_MUTATING gating, FdTable check on user stdio fds) addresses the most obvious permission-bypass concerns up front. A few items worth addressing before merge.

Gate items

CI shows No checks found — needs to be green on all platforms before merge, especially since only Linux was verified locally and there's nontrivial Windows-conditional code (path_symlink normalization, dup_user_fd_to_file, ELOOP mapping).
This PR looks AI-assisted (divybot co-author, the "in this session" phrasing) — could you add an explicit disclosure in the PR body if so?

Inline comments below for the substantive items.

bartlomieju · 2026-05-20T09:30:58Z

+/// return `None` from `fd_rights` and we let those fall through to the
+/// per-op entry-type checks. For File/Dir entries the rights were granted
+/// at path_open time and intersected with the caller's request.
+fn has_right(rights: Option<u64>, needed: u64) -> bool {


fd_rights returns None for Stdin/Stdout/Stderr/HostFile, and has_right treats None as "all rights granted". For inherited stdio that's fine, but for HostFile (a user-supplied fd dup'd in via the constructor) it means none of the per-op rights checks added in this PR actually gate it — e.g. a HostFile bound as stdin will still pass the RIGHTS_FD_WRITE check.

The make_stdio_entry FdTable check ensures the Deno permission was already paid at op_node_open_sync time, so this is probably safe by design — but please call it out explicitly in this comment so a future reader doesn't assume HostFile is rights-gated.

bartlomieju · 2026-05-20T09:30:58Z

+    // SAFETY: libc::dup is always safe to call. It either returns -1 on
+    // error or a fresh fd owned by the caller. We only construct File from
+    // a successful dup, so the resulting File owns a valid OS fd.
+    let new_fd = unsafe { libc::dup(user_fd) };


libc::dup() does not set FD_CLOEXEC, so the dup'd fd will leak across exec* in any process the WASI program spawns. Prefer libc::fcntl(user_fd, F_DUPFD_CLOEXEC, 0) (or dup3 with O_CLOEXEC on Linux) to avoid leaking host fds into child processes.

bartlomieju · 2026-05-20T09:30:58Z

+    {
+      let now = std::time::Instant::now();
+      if due > now {
+        std::thread::sleep(due - now);


nit: std::thread::sleep stalls the JS isolate thread for the full clock subscription duration. WASI's poll is synchronous so the semantics are unavoidable, but the existing comment block focuses on the borrow drop — please add a line noting that this intentionally blocks JS, since it's a real footgun for callers passing long timeouts.

bartlomieju · 2026-05-20T09:30:58Z

+  clock_id: i32,
+  clock_timeout: u64,
+  #[allow(dead_code, reason = "kept for parity with full WASI errno set")]
+  clock_precision: u64,


nit: the reason = "kept for parity with full WASI errno set" is copy-pasted from ERRNO_PERM — it fits the errno constant but not a poll subscription field. Suggest reason = "WASI subscription field, not currently consulted" or similar.

bartlomieju · 2026-05-20T09:30:58Z

@@ -1450,28 +1779,251 @@ impl WasiContext {
  #[fast]
  fn path_filestat_set_times(


_flags is ignored here, which means calls with flags=0 (no LOOKUPFLAGS_SYMLINK_FOLLOW) silently follow symlinks anyway. path_filestat_get was updated to honor the bit — please mirror that here: toggle between std::fs::metadata/symlink_metadata and between filetime::set_file_times/set_symlink_file_times based on the lookupflag.

bartlomieju · 2026-05-20T09:30:58Z

  }

+  #[fast]
+  fn path_link(


nit: every other mutating path op in this PR consults the dirfd's stored rights via has_right; path_link doesn't check RIGHTS_PATH_LINK_SOURCE / RIGHTS_PATH_LINK_TARGET. The Deno check_open call in resolve_preopen_path still gates the OS-level operation, so this isn't a hole — just inconsistent with the rest of the file. Worth adding for symmetry.

bartlomieju · 2026-05-20T09:30:58Z

+        // On Windows, path_symlink rewrote forward slashes to backslashes so
+        // the OS could resolve relative symlinks. Convert them back here so
+        // WASI callers see the POSIX-style target they originally stored
+        // (Node's test-wasi-symlinks asserts the bytes round-trip exactly).


Blanket \\ → / rewrite on Windows assumes the link target was originally POSIX-style. For symlinks created outside WASI in the same preopen tree (or created with Windows-native separators by other code), this will corrupt the bytes returned to the caller. Worth noting in the comment that this is best-effort and biased toward the WASI test suite's round-tripping case.

bartlomieju · 2026-05-20T09:30:58Z

+    this.#memory = memory as WebAssembly.Memory;
+    // finalizeBindings is intentionally idempotent: the same WASI instance
+    // can be bound to multiple wasm instances across threads.
+    this.#started = true;


Setting #started = true here is correct for the thread-spawn worker case, but it also means a subsequent start() / initialize() on the same instance will throw ERR_WASI_ALREADY_STARTED. Could you double-check that matches Node's behavior, and add a short docstring noting the intentional one-way transition? Otherwise the divergence from start/initialize will be surprising.

bartlomieju · 2026-05-20T09:30:58Z

+  assertEquals(view.getUint32(0, true), 0);
+});
+
+Deno.test("[node/wasi] - constructor emits ExperimentalWarning once", async () => {


nit: the test name says "once" but the assertion only checks that any warnings that fire carry the expected message — it doesn't actually verify the once-only behavior (because of the module-scoped flag, as the comment explains). Suggest renaming to "... emits ExperimentalWarning with the expected message" so the name matches what's tested.

* path_filestat_set_times now honors LOOKUPFLAGS_SYMLINK_FOLLOW like path_filestat_get does — switches between metadata/symlink_metadata for the absent-bit lookup and between set_file_times / set_symlink_file_times for the utimensat-equivalent call. Previously flags=0 silently followed. * dup_user_fd_to_file (Unix) now uses fcntl(F_DUPFD_CLOEXEC, 0) instead of plain dup so the dup'd host fd doesn't leak into processes the WASI program spawns. Windows already creates the duplicate handle with bInheritHandle=false. * path_link consults the source/target dirfd rights via has_right for parity with the other mutating path_* ops; added RIGHTS_PATH_LINK_SOURCE / RIGHTS_PATH_LINK_TARGET constants and RIGHTS_PATH_LINK_TARGET to RIGHTS_MUTATING. * fd_rights doc comment: call out that HostFile entries are intentionally not rights-gated because their permission was paid upstream by node:fs.openSync. * path_symlink reverse-normalization comment: note that the Windows backslash->forward-slash rewrite in path_readlink is best-effort and biased toward the WASI round-trip case. * poll_oneoff thread::sleep: note that it intentionally blocks the V8 isolate for the clock subscription duration. * PollSubscription.clock_precision dead_code reason matches the field's role rather than the errno copy-paste. * WASI finalizeBindings docstring: spell out the one-way #started transition shared with start()/initialize(). * Renamed "constructor emits ExperimentalWarning once" test to reflect what it actually checks — the warning carries the WASI message — since the once-only behavior can't be verified across the test process. Co-Authored-By: Divy Srivastava <me@littledivy.com>

divybot

Thanks for the careful review! Pushed 3152fcf addressing everything:

path_filestat_set_times now mirrors path_filestat_get: switches between metadata/symlink_metadata and between filetime::set_file_times/set_symlink_file_times based on LOOKUPFLAGS_SYMLINK_FOLLOW.
dup_user_fd_to_file (Unix) uses fcntl(F_DUPFD_CLOEXEC, 0) instead of plain dup. Windows already calls DuplicateHandle with bInheritHandle=false, so no change needed there.
path_link consults RIGHTS_PATH_LINK_SOURCE / RIGHTS_PATH_LINK_TARGET via has_right for symmetry with the other path_* ops (added both constants, plus RIGHTS_PATH_LINK_TARGET to RIGHTS_MUTATING).
fd_rights doc comment spells out that HostFile entries are intentionally not rights-gated because the permission was paid upstream by node:fs.openSync (and the FdTable check in make_stdio_entry).
path_readlink Windows comment notes the backslash→forward-slash rewrite is best-effort and biased toward the WASI round-trip case.
poll_oneoff thread::sleep comment added a paragraph spelling out that it intentionally blocks the V8 isolate for the clock subscription duration.
clock_precision dead_code reason changed to "WASI subscription field, not currently consulted".
finalizeBindings docstring spells out the one-way #started transition shared with start()/initialize() — Node's wasi_thread_start path has the same behavior.
Test rename: constructor emits ExperimentalWarning once → constructor emits ExperimentalWarning with the expected message.

PR description updated with the AI-assist disclosure and the expanded permissions-gating section. CI is currently green on Linux/macOS/Windows debug + release on the prior head; will rebuild on this commit.

- Implement real `node:wasi` preview1 compatibility for the 20 WASI tests in the Node compat suite that were previously absent from `tests/node_compat/config.jsonc`. All 20 now pass on Linux. - Add 20 new entries to `config.jsonc` for the newly-passing tests. The implementation work is in `ext/node/ops/wasi.rs` and `ext/node/polyfills/wasi.ts`; the suite entries are added at the bottom of `config.jsonc`. No existing entries were flipped — every test added here is newly green. - **ENOTCAPABLE for sandbox escape.** `cant_dotdot` / `symlink_escape` expect `errno == ENOTCAPABLE (76)`; we previously returned `ERRNO_PERM (63)`. The starts_with check now maps to `NOTCAPABLE`. - **Symlink loop detection.** Probe `raw_os_error()` for `ELOOP` so we don't depend on the unstable `ErrorKind::FilesystemLoop` variant. - **path_open / path_filestat_get / path_readlink follow flags.** Honor `LOOKUPFLAGS_SYMLINK_FOLLOW` for path_open / path_filestat_get, and force no-follow on the final component for path_readlink so `std::fs::read_link` doesn't reject the resolved target. - **futimens.** `fd_filestat_set_times` and `path_filestat_set_times` use the `filetime` crate, with full `FSTFLAGS_ATIM(_NOW)` / `FSTFLAGS_MTIM(_NOW)` semantics. - **path_link.** Added (`std::fs::hard_link`) and wired through the JS polyfill so `link(2)` works. - **poll_oneoff.** Implements clock subscriptions (relative and abstime) plus immediate fd_read/fd_write events. For stdin, the `EVENT_FD_READWRITE_HANGUP` bit is set so wasi-libc's poll() shim returns `POLLHUP|POLLIN` as test-wasi-poll.js requires. - **stdio fd routing.** The WASI constructor's `stdin`/`stdout`/`stderr` options now do something: fds >= 3 are dup'd into an owned `std::fs::File` and used in place of the inherited process streams. Closing the user's original fd afterwards doesn't affect ours. - **sock_accept.** Returns `ERRNO_BADF` for unknown fds and `ERRNO_NOTSOCK` for non-socket fds in our table (not `ERRNO_NOSYS`), matching Node/uvwasi semantics. - **ExperimentalWarning on construction.** Match Node's "WASI is an experimental feature and might change at any time" warning so the suite's `common.expectWarning('ExperimentalWarning', ...)` doesn't fail the very first test. - **finalizeBindings.** Node-only API used by `wasi:thread-spawn` workers to bind a shared `WebAssembly.Memory` without re-running `_start`. test-wasi-pthread needs it. The vendored Node suite is missing the empty `test/fixtures/wasi/subdir` directory (git doesn't track empty dirs). test-wasi-readdir asserts that the preopen has four entries including `subdir`. The runner now creates it before enumerating tests, with a docstring explaining why. - [x] All 25 WASI tests in `tests/node_compat/runner/suite/test/wasi/` pass via `cargo test --test node_compat 'test-wasi-'` - [x] `tests/unit_node/wasi_test.ts` (added tests for finalizeBindings and the ExperimentalWarning) passes - [x] `tools/format.js --check`, `tools/lint.js --rs`, and `tools/lint.js --js` all clean The 20 candidate tests are bounded WASI behavior — no host-OS-specific syscalls aside from the dup/_dup paths in stdio routing. The Linux run exercises all 20; macOS should behave identically. Windows path canonicalization / hard_link / symlink behavior is consistent with what we already exercise elsewhere in `ext/node/ops/wasi.rs`, so I expect the same wins on macOS and Windows but only Linux was verified in this session. Closes denoland/orchid#150 --------- Co-authored-by: divybot <divybot@users.noreply.github.com> Co-authored-by: Divy Srivastava <me@littledivy.com>

divybot changed the title ~~node compat: improve WASI preview1 behavior for absent Node tests~~ feat(ext/node): implement node:wasi preview1 compat for Node tests May 19, 2026

divybot and others added 2 commits May 19, 2026 20:00

divybot and others added 2 commits May 20, 2026 03:47

Merge remote-tracking branch 'origin/main' into orch/divybot-150

9bf4674

# Conflicts: # Cargo.lock # ext/node/Cargo.toml

bartlomieju reviewed May 20, 2026

View reviewed changes

divybot commented May 20, 2026

View reviewed changes

littledivy approved these changes May 20, 2026

View reviewed changes

littledivy merged commit ed12991 into main May 20, 2026
136 checks passed

littledivy deleted the orch/divybot-150 branch May 20, 2026 13:12

Navigation Menu

Repository navigation

feat(ext/node): implement node:wasi preview1 compat for Node tests - #34245

feat(ext/node): implement node:wasi preview1 compat for Node tests#34245
littledivy merged 6 commits into
mainfrom
orch/divybot-150

divybot commented May 19, 2026 •

edited

Loading

Uh oh!

littledivy commented May 20, 2026

Uh oh!

bartlomieju left a comment

Uh oh!

bartlomieju May 20, 2026

Uh oh!

bartlomieju May 20, 2026

Uh oh!

bartlomieju May 20, 2026

Uh oh!

bartlomieju May 20, 2026

Uh oh!

bartlomieju May 20, 2026

Uh oh!

bartlomieju May 20, 2026

Uh oh!

bartlomieju May 20, 2026

Uh oh!

bartlomieju May 20, 2026

Uh oh!

bartlomieju May 20, 2026

Uh oh!

divybot left a comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

		@@ -1450,28 +1779,251 @@ impl WasiContext {
		#[fast]
		fn path_filestat_set_times(

Conversation

divybot commented May 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

What changed in node:wasi

Runner fixup

Test plan

Platform-specific notes

Uh oh!

littledivy commented May 20, 2026

Uh oh!

bartlomieju left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

divybot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

divybot commented May 19, 2026 •

edited

Loading