Skip to content

usockets(windows): keep a closed poll alive until the outer tick and libuv are done with it - #39643

Merged
cirospaciari merged 5 commits into
mainfrom
farm/47550167/win-poll-free-protocol
Sep 11, 2026
Merged

cirospaciari merged 5 commits into
mainfrom
farm/47550167/win-poll-free-protocol

Conversation

@robobun

@robobun robobun commented Aug 19, 2026 •

Copy link
Copy Markdown
Collaborator

Problem

  • Windows crashes after a handler returns: us_socket_is_closed or us_internal_socket_follow_adopted at 0xFFFFFFFFFFFFFFFF (BUN-43AH, BUN-4NC8), us_internal_socket_close_raw on a freed socket (BUN-442H, BUN-442Y, BUN-4NP3), a garbage poll_cb called from uv__fast_poll_process_poll_req (BUN-442Z), mimalloc free-list crashes, and the test/bake/deinitialization.test.ts failures on the Windows lanes. The 1.4.0 release (34cbb9a40) reports this family from bun test on Windows (as of 2026-08-21: BUN-442H 12 events, BUN-4NC8 6, BUN-442Y 5, BUN-4NP3 2, BUN-442Z 1) and the TLS forms BUN-4MNY, BUN-4NFZ, BUN-4NJG, BUN-4NKV and BUN-4QEY, one each. BUN-4QEV, BUN-4QHN and BUN-4QQ6 are the reused-handle arm described in the notes.
  • loop.c:450 frees closed sockets only at tick_depth <= 1. The libuv backend never counts it, so a handler that waits for a promise (a nested uv_run) frees the socket the outer dispatch still reads (loop.c:746).
  • The nested run also finishes closing that socket's uv_poll_t under libuv's outer uv__fast_poll_process_poll_req frame. That frame then queues the endgame again, and close_cb_free_poll runs twice. uv_run is documented as not reentrant, so the misuse is ours.

Fix

  • us_loop_run and us_loop_pump count tick_depth, as the POSIX backend does.
  • While a poll_cb frame is on the stack (poll_cb_depth), us_poll_stop only disarms the handle. The outermost frame calls uv_close on its way back into libuv, which libuv supports, and then closes the socket (us_internal_poll_close_fd), so libuv sees the same order as before. libuv is unchanged.
  • us_poll_free and close_cb_free_poll record who ran first (released, uv_closed). The second one frees both blocks. The old data = NULL handshake leaked when the callback ran first (usockets(win): free the poll when us_poll_free runs after the uv close callback #37105). A stopped poll is never re-armed or re-initialized.
  • Verified: the new case in test/js/bun/net/socket.test.ts. On a Windows debug build of main its child segfaults at 0xFFFFFFFFFFFFFFFF (the BUN-43AH signature), and the release canary crashes too. It passes with the fix (Windows x64 and arm64 debug, Linux ASAN). appcontainer.test.ts gets a step for the socket order. Notes have the rest.

Background

  • On Windows a socket is two blocks: the us_socket_t (it starts with the us_poll_t) and a libuv uv_poll_t. libuv's in-flight AFD requests live inside the uv_poll_t, so it has to live until the close callback. A closed socket waits on loop->data.closed_head until us_internal_loop_post frees it, and loop.c reads it after its handlers return.
  • Endgame: uv_close cancels the requests. When the last one completes, libuv queues the endgame, which unlinks the handle and runs the close callback. uv__fast_poll_process_poll_req checks for it right after poll_cb returns.
  • Nested tick: wait_for_promise (expect().resolves, auto-install) runs uv_run inside the handler.
Notes

Symbolized fail-before (Windows x64 debug build of main a35696478d, a standalone copy of the new test's fixture. The test itself fails on that build with the same fault address):

panic(main thread): Segmentation fault at address 0xFFFFFFFFFFFFFFFF
us_internal_socket_follow_adopted   packages/bun-usockets/src/internal/internal.h:357
us_internal_dispatch_ready_poll     packages/bun-usockets/src/loop.c:746   (after us_dispatch_data returned)
poll_cb                             packages/bun-usockets/src/eventing/libuv.c:165
uv__fast_poll_process_poll_req      vendor/libuv/src/win/poll.c:233
uv__process_reqs / uv_run           vendor/libuv/src/win/core.c
us_loop_run                         packages/bun-usockets/src/eventing/libuv.c:417

The freed socket is filled with mimalloc's debug poison, so flags.adopted reads as set and prev is followed. The open() variant crashes the same way at loop.c:556. The same fixture without the nested wait passes on that build. The release canary (1.4.0-canary.1+32e87032b) passes the first case and crashes in the second, at 0xFFFFFFFFFFFFFFFF in one run and 0x1AF0000002A in another, which is the heap corruption from the double free.

How the three Sentry shapes follow. In release, mi_free overwrites the first word of the freed uv_poll_t, which is data. The rest stays intact, so the outer frame sees events == 0, CLOSING, and no requests in flight, and queues the endgame again. The second close_cb_free_poll frees h->data, now the free-list link to another freed block, and h itself again. Later allocations alias (a garbage poll_cb, BUN-442Z, or us_poll_start_rc faulting at 0 in deinitialization.test.ts), or the freed socket is reused and the outer dispatch runs the error close on it (BUN-442Y), or follow_adopted reads a reused block (BUN-43AH).

Socket order, found by CI. The first push closed the socket in close_raw before the deferred uv_close. appcontainer.test.ts failed on the Windows 11 arm64 lane with exit 0xC0000008 (STATUS_INVALID_HANDLE), and failed 10 of 10 runs on an arm64 machine with that build against 5 of 5 passes with main's libuv.c on the same machine. GetProcessMitigationPolicy(ProcessStrictHandleCheckPolicy) inside the container reports 0x3, so a call on a closed handle raises instead of failing, and uv__poll_close cancels the in-flight request with an ioctl on the socket. A probe with four steps (listener only, serve + fetch, terminate from data(), terminate from a timer with the peer closed by its own dispatch) died in every step that closes a socket from a dispatch. With us_internal_poll_close_fd the four steps and the test pass (10 of 10 runs on arm64, and on x64). The test now also closes a socket from its own data() handler, which is the shortest path to this order. Outside a nested tick, the later socket close is not observable from JS: the loop does not run again before the handler's dispatch ends. Inside one, a handler that closes its socket and then waits for the peer to notice now waits until the handler returns. That wait was already unreliable on Windows (the peer's completion is often in the outer run's batch) and then corrupted the heap.

The reused-handle arm. Between the inner run's endgame and the outer frame's return, us_create_poll can hand the freed uv_poll_t block to a new poll (same size class, most recent free). Then the second close_cb_free_poll runs on a live handle: h->data is the new poll's us_socket_t, so that socket is freed with no close path, and the new uv_poll_t is freed under its pending request. The JS wrapper of that socket still reports open. BUN-4QHN (ws.send(): us_socket_group_ext on the freed socket's group), BUN-4QQ6 (socket.readyState: SSL_get_shutdown on its ssl) and BUN-4QEV (uv__process_poll_req reading the freed handle for a completed request, poll.c:574) are that arm. With one uv_close per handle, issued after its last callback frame, the second callback no longer exists.

Why the hand-off is needed together with tick_depth. With tick_depth alone, every socket closed during a nested tick with no live frame finishes closing in the inner run, and the deferred us_poll_free then met a handle whose close callback had already run: the old code handed it data and leaked both blocks. #37105 fixes that case on its own with a marker in data. This PR covers it with the two bits.

Defensive branches with no current caller: us_poll_free on a poll that was never stopped, us_poll_start_rc on a registered poll (mask change, or UV_EBADF once stopped), us_poll_change after stop. Every current caller creates, starts, and later stops a poll exactly once. The uv_poll_init_socket failure path now goes through us_poll_stop. uv_poll_stop on a half-initialized handle only clears events (checked against uv__poll_set). That path was reasoned about, not run: it needs a --socketFaultInjection=on build, which rebuilds every Rust crate.

Suites run on the Windows x64 debug build with the fix, all green: test/js/bun/net/socket.test.ts (86 pass), tcp-server, socket-retention, socket-syscall-fault, test/js/bun/udp/udp_socket.test.ts, test/js/bun/websocket/websocket-server.test.ts, test/js/node/net/node-net-server.test.ts, test/js/node/tls/node-tls-server.test.ts, test/js/node/http/node-http.test.ts, test/js/bun/http/serve.test.ts (294 pass), test/bake/deinitialization.test.ts 3 of 3 (the unfixed release canary hangs for 60 s and is killed on the same machine). test/js/web/fetch/fetch.test.ts had 14 timeouts of (with gc) body tests on this slow debug box. The two socket tests among them pass in isolation. Per the analysis in #33018, deinitialization.test.ts closes the socket from a nested poll_cb on the same handle, so it also covers poll_cb_depth > 1.

Sentry groups on the 1.4.0 release (34cbb9a40, all Windows): BUN-4NC8 is follow_adopted (internal.h:357) under loop.c:746, the frames of the fail-before above. BUN-442H is close_raw at socket.c:292, the low-prio unlink writing through s->prev of a reused block, reached from the eof or error close at the end of the dispatch (BUN-442Y and BUN-4NP3 are the same call: the unlink at context.c:235 and :240 writing through the reused block's prev and next, and us_internal_disable_sweep_timer through its group). BUN-4MNY is ssl_retry_parked_write (openssl.c:2103, s->group read as NULL) in the tail of us_internal_ssl_on_data after us_dispatch_data returned: the TLS form of the same read of a freed socket. BUN-4NFZ is the other exit of that tail: us_internal_ssl_close (openssl.c:1927, again s->group NULL) from the ssl_close call at openssl.c:2324 after the data callback. BUN-4NJG and BUN-4NKV are the keylog and session flushes in the same tail (openssl.c:2430 to :351, and :2429 to :426): s->ssl read from the reused block, so SSL_get_ex_data returns garbage. Both flushes are guarded by !s->ssl || is_closed, which only a reused block gets past. The four TLS groups are the four helpers of that tail, in whichever order a given reused block fails. With the closed list left alone until the outermost tick, ssl_gone, is_closed and s->ssl read the real state there and those tails return.

Limit. tick_depth counts the ticks Bun itself runs (us_loop_run, us_loop_pump). A native addon that calls uv_run() on this loop from inside a handler is not counted. The socket whose own callback is on the stack is still safe there, because its uv_close and frees are tied to poll_cb_depth, but a socket that another frame holds (an accepted socket closed from its own open() while the accept loop still reads it, or a block retired by adoption) is not. Counting the frames of the callbacks libuv runs for us (poll_cb, prepare_cb, timer_cb, async_cb) would cover that entry as well. None of the crashes above needed it, so it is left for a follow-up rather than re-verifying this PR for it.

Release-build numbers for test/bake/deinitialization.test.ts are in the comments below: on one Windows Server 2019 machine the fixture fails 8 of 8 runs on main (6 hangs, 2 segfaults) and passes 20 of 20 with this diff.

Related. #33018 was an earlier attempt at this bug. It patched libuv's post-callback check instead of moving the uv_close. #38024 has the same bug class in the c-ares poll and still carries that patch. Deferring its uv_close past the outermost callback frame would remove the need for it there too. The test fixture waits on a timer on purpose: a nested run cannot wait for an event of the peer socket, because that completion may sit in the outer run's IOCP batch.


no test proof · iteration 4 · platform-specific test(s) that do not run on this machine, deferring to CI, which covers all platforms: test/js/bun/windows/appcontainer.test.ts, test/js/bun/net/socket.test.ts

@coderabbitai

coderabbitai Bot commented Aug 19, 2026 •

Copy link
Copy Markdown
Contributor

Review Change StackReview Change Stack

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

  • @coderabbitai resume to resume automatic reviews.
  • @coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

  • ▶️ Resume reviews
  • 🔍 Trigger review

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info
⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Essentials

Run ID: 8b858052-b382-44e3-9e2f-52b6055e04ec

📥 Commits

Reviewing files that changed from the base of the PR and between 9e1b23a and 9f56f1f.

📒 Files selected for processing (1)
  • test/js/bun/net/socket.test.ts

Included review availability: Your plan provides up to 10 included reviews per hour; 1 remains after this review.


Walkthrough

Changes

This change adds backend-aware poll descriptor closure. It expands libuv poll lifecycle handling for callback reentrancy, deferred cleanup, and stopped handles. Regression tests cover closure during data() and open() callbacks.

Socket poll lifecycle

Layer / File(s) Summary
Poll lifecycle and backend closure
packages/bun-usockets/src/internal/eventing/libuv.h, packages/bun-usockets/src/internal/internal.h, packages/bun-usockets/src/internal/loop_data.h, packages/bun-usockets/src/eventing/libuv.c, packages/bun-usockets/src/eventing/epoll_kqueue.c
Poll state now tracks stopping, release, closure, deferred descriptor handling, and nested callbacks. Libuv poll operations coordinate handle cleanup and always subscribe to disconnect events.
Socket close-path wiring
packages/bun-usockets/src/socket.c, packages/bun-usockets/src/udp.c, packages/bun-usockets/src/context.c
Socket, UDP, and listening descriptor closure paths now use us_internal_poll_close_fd.
Socket closure regression coverage
test/js/bun/net/socket.test.ts, test/js/bun/windows/appcontainer.test.ts
Tests cover socket closure during data() and open() callbacks, including repeated subprocess execution and Windows AppContainer behavior.

Suggested reviewers: jarred-sumner, cirospaciari

Priority: ⬆️ High

Merge Risk: ⚪ Minimal · up to 9f56f

The added socket regression test does not introduce a current merge-blocking risk.

🚥 Pre-merge checks | ✅ 4
✅ Passed checks (4 passed)
Check name Status Explanation
Title check ✅ Passed The title clearly describes the main Windows libuv poll-lifetime fix and the condition that keeps closed polls alive until the outer tick and libuv complete. It is specific and related to the changese…
Description check ✅ Passed The description explains the problem, fix, background, limitations, and verification results. It does not use the exact template headings, but it provides the required information, including how the c…
Linked Issues check ✅ Passed Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check ✅ Passed Check skipped because no linked issues were found for this pull request.

Comment @coderabbitai help to get the list of available commands.

@robobun

robobun commented Aug 19, 2026 •

Copy link
Copy Markdown
Collaborator Author

Status: merged as bb8aa37afa on 2026-09-11 01:24 UTC, then reverted three minutes later by #42258 (d0aafac5d3). No reason is recorded on either PR. The main build on the merge commit (114076) had not run yet: the revert push cancelled it. The squash commit held exactly this PR's diff (10 files, checked against the PR head), and the PR's last build (114040) passed every lane. A reland branch with identical content is pushed as robobun/47550167/reland-win-poll-lifetime. It is not opened as a PR until the reason for the revert is known.

Reproduced on Windows x64 with the new case in test/js/bun/net/socket.test.ts. A data() handler calls socket.terminate() and then expect(Bun.sleep(20)).resolves, which ticks the loop inside the dispatch.

  • Debug build of main (a35696478d): the child process segfaults at 0xFFFFFFFFFFFFFFFF. The frames are us_internal_socket_follow_adopted <- us_internal_dispatch_ready_poll (loop.c:746) <- poll_cb <- uv__fast_poll_process_poll_req (poll.c:233), the same frames as BUN-43AH.
  • Release canary 1.4.0-canary.1+32e87032b: the same fixture crashes too, with a different fault address on each run.
  • With this branch: the case passes on Windows x64 and arm64 debug builds and on a Linux ASAN build. test/bake/deinitialization.test.ts passes on both Windows builds. The unfixed canary hangs on it for 60 s on the same x64 machine.

The first push (876db250) closed the socket before the deferred uv_close. CI caught that with test/js/bun/windows/appcontainer.test.ts on the Windows 11 arm64 lane (exit 0xC0000008): AppContainers run with strict handle checks, and libuv's cancel ioctl hit the closed socket. It failed 10 of 10 runs on an arm64 machine and passed 5 of 5 there with main's libuv.c. f91d44edb8 closes the socket right after the uv_close instead (us_internal_poll_close_fd), which is the order libuv always had. The test passes 10 of 10 on arm64 and passes on x64, and it now has a step for this order. Details are in the PR notes.

No libuv patch. The uv_close for a poll whose callback is on the stack is issued by the outermost poll_cb frame instead.

@claude claude Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I reviewed this PR and the automated bug-hunt found no issues. Because it reworks lifetime/ownership of uv_poll_t blocks in the Windows usockets backend — memory-safety-critical native code with subtle ordering across nested uv_run — a human look is still worthwhile.

What was reviewed:

  • Traced the released/uv_closed hand-off across all us_poll_stop/us_poll_free call sites and the us_poll_start_rc init-failure path — each block is freed exactly once on every path.
  • Checked us_poll_resize: the memcpy carries poll_cb_depth/stopped to the new block and poll_cb re-reads p->data after dispatch, so the decrement and deferred uv_close land on the live block; the old block's stale counter is unreachable (uv_p == NULL → fast-path free).
  • Confirmed us_internal_loop_data_t layout is unchanged (comment-only edit), so the Rust mirror needs no update; us_poll_t has no Rust mirror.
  • The early-return in poll_cb before the depth increment runs no JS, so it cannot race us_poll_stop.
Extended reasoning...

Overview

This PR fixes a Windows-only use-after-free / double-free in the usockets libuv backend when a socket handler re-enters the event loop (e.g. expect(promise).resolves, auto-install). It touches packages/bun-usockets/src/eventing/libuv.c (poll lifecycle: poll_cb, us_poll_stop, us_poll_free, us_poll_start_rc, close_cb_free_poll, us_loop_run/us_loop_pump), adds four state fields to us_poll_t in libuv.h, updates a comment in loop_data.h, and adds a subprocess regression test to test/js/bun/net/socket.test.ts.

The mechanism: (1) count tick_depth on the libuv backend so us_internal_loop_post defers freeing closed sockets past nested ticks (POSIX parity); (2) defer uv_close on a poll until the outermost poll_cb frame for that handle unwinds, so libuv's outer uv__fast_poll_process_poll_req never sees a handle whose endgame already ran; (3) replace the old data = NULL handshake between us_poll_stop and close_cb_free_poll with two explicit bits (released, uv_closed) so whichever runs second frees both blocks — the old scheme leaked when the callback ran first.

Security risks

None user-facing. This is internal event-loop plumbing; no untrusted input parsing, no auth/crypto. The risk class is memory safety (UAF, double-free, leak) in a hot path that every Windows socket goes through.

Level of scrutiny

High. Per the repo's review guidance, native memory safety is the most-blocked category. This change rewrites the ownership contract between two heap blocks (us_poll_t and its uv_poll_t) across three release points (us_poll_stop, us_poll_free, close_cb_free_poll) and a re-entrant uv_run. The reasoning in the PR description is thorough and the code reads correctly to me, but the failure mode of a mistake here is heap corruption on every Windows user's machine, and one defensive branch (the uv_poll_init_socket half-init failure path) was reasoned about rather than exercised. That warrants a maintainer's eyes.

Other factors

  • I traced every us_poll_stop/us_poll_free/us_poll_start_rc caller in loop.c, socket.c, context.c, and udp.c; each follows create → start → stop → free, matching the new one-way stopped contract. The us_poll_start_rc "already registered" branch and us_poll_change after stop are defensive with no current caller (as the PR notes).
  • us_poll_resize interaction checks out: memcpy copies the new fields to the replacement block, p->data is repointed, and poll_cb re-reads it after dispatch; the old block's uv_p = NULL sends its later us_poll_free down the fast path so its stale poll_cb_depth is never consulted.
  • loop_data.h change is comment-only; the Rust mirror in src/uws_sys/InternalLoopData.rs needs no update. No Rust mirror of us_poll_t exists.
  • Test follows harness conventions (subprocess, tempDir, bunEnv, drains both pipes concurrently, asserts a combined object). The 20 ms Bun.sleep inside the fixture is the deliberate mechanism (drive nested ticks), not a wait-for-condition, and is commented as such.

@robobun

robobun commented Aug 19, 2026 •

Copy link
Copy Markdown
Collaborator Author
Updated 6:18 PM PT - Sep 10th, 2026

✅ @robobun, your commit 9f56f1f8765cc56b30f64db19210e8f3d36ca70a passed in Build #114040! 🎉


🧪   To try this PR locally:

bunx bun-pr 39643

That installs a local version of the PR into your bun-39643 executable, so you can run:

bun-39643 --bun

@claude claude Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I reviewed this PR and didn't find any bugs. Because it redesigns the two-block ownership handoff for every Windows socket close (a memory-safety-critical path where getting it wrong trades one UAF for another) and introduces a five-bit state machine (stopped/close_fd/uv_closed/released/poll_cb_depth), a human look would still be worthwhile — particularly to sign off on this shape versus the earlier libuv-patch approach from #33018.

Checked: us_poll_resize copies poll_cb_depth via memcpy before p->data is repointed, so the epilogue's re-read lands on the right counter; us_socket_detach (stop without close_fd) keeps the fd open through the deferred uv_close; the us_poll_start_rc init-failure path composes with the caller's us_poll_free via released; the four bsd_close_socket(us_poll_fd(...)) sites that follow a stop are all converted, and the remaining one in loop.c:862 is epoll-gated.

Extended reasoning...

Overview

This PR fixes Windows crashes (BUN-43AH/442Y/442Z, deinitialization.test.ts) caused by nested uv_run freeing a socket the outer dispatch still holds, and by uv_close running under an outer uv__fast_poll_process_poll_req frame that then queues the endgame twice. The fix has three parts: (1) us_loop_run/us_loop_pump now count tick_depth so us_internal_loop_post defers the closed-socket free to the outermost tick, matching POSIX; (2) a per-poll poll_cb_depth counter defers uv_close to the outermost poll_cb frame's epilogue, so libuv sees a supported close-from-inside-callback instead of a close from a nested run; (3) the old data = NULL handshake between us_poll_stop/us_poll_free/close_cb_free_poll is replaced by explicit released/uv_closed bits so whichever runs second frees both blocks (the old scheme leaked when the callback ran first). A new us_internal_poll_close_fd helper defers closesocket past the deferred uv_close so AppContainer strict-handle-checks don't kill the process on the cancel ioctl. Call sites in context.c, socket.c, udp.c are updated; epoll_kqueue.c gets a pass-through implementation.

Security risks

None identified. This is internal event-loop lifecycle management with no user-controlled input reaching the changed paths.

Level of scrutiny

High. This is the most-blocked category in Bun's review guide (native memory safety): every acquisition/release pairing on the Windows socket-close path is being reorganized, and the failure mode of a mistake is exactly the UAF/double-free class being fixed. The state machine has five interacting bits plus a depth counter, and correctness depends on invariants that span us_poll_stop → us_internal_poll_close_fd → poll_cb epilogue → close_cb_free_poll → us_poll_free, with us_poll_resize cutting across. The PR description is exceptionally thorough (symbolized fail-before, mechanism for each Sentry shape, ordering rationale, defensive-branch inventory) and the author ran the relevant Windows suites, but the changed code cannot be exercised on this Linux environment and the design choice (per-poll depth counter + deferred close vs. #33018's libuv patch) is one a maintainer should ratify.

Other factors

I traced the us_poll_resize interaction: memcpy copies poll_cb_depth after the outer frame's increment, and the epilogue re-reads p->data to land on the relocated block, so nested dispatch + adopt + close balances correctly. us_socket_detach correctly does not set close_fd, so the fd survives the deferred uv_close for handoff. The one remaining us_poll_stop not followed by us_internal_poll_close_fd at loop.c:862 is under #ifdef LIBUS_USE_EPOLL. The us_poll_start_rc early-return for an already-registered handle and the us_poll_change guard on stopped are defensive (no current caller per the PR notes) but harmless. The new socket.test.ts case is a spawned crash repro with a snapshot on {stdout, exitCode}; the appcontainer.test.ts addition correctly notes the exit code is the real assertion. The PR notes that the uv_poll_init_socket failure path was reasoned about but not run (needs a fault-injection build) — that path now routes through us_poll_stop, which is a behavior change worth a maintainer's eye.

@robobun

robobun commented Aug 20, 2026

Copy link
Copy Markdown
Collaborator Author

Data point from the test/bake/deinitialization.test.ts failures on the Windows lanes (35 of the last 400 builds). They are this bug.

Symbolized CI crash. Build 101365 (Windows x64, 1c8d14b8f, Segmentation fault at address 0x0), symbolized with that build's bun-profile.pdb:

us_internal_socket_after_resolve        packages/bun-usockets/src/context.c:755   (c->group is NULL, c->closed is 0)
us_internal_drain_pending_dns_resolve   packages/bun-usockets/src/loop.c:349
us_internal_loop_post                   packages/bun-usockets/src/loop.c:440
uv__check_invoke / uv_run
TestCommand::run

The crash comes right after the closeActiveConnections sendAnyRequests websocket=1 case passes. The us_connecting_socket_t on the ready list belongs to the next case (websocket=8, eight connects to localhost). A stray free of a live block is what the second close_cb_free_poll in this PR's notes produces, and the reduction below crashes at the same point.

How the fixture gets there. The test body resumes inside the HMR client socket's poll_cb (the onopen microtask). The un-awaited expect(fetch(...)).rejects then runs a nested uv_run from that frame. Inside it the plugin calls server.stop(true) and the server closes the connection. The inner run dispatches the same poll again (poll_cb_depth 2) and closes the socket. The fetch rejection comes from the HTTP thread a few iterations later, so the inner run also processes the AFD cancellation and runs the endgame. Both blocks are freed under the outer uv__fast_poll_process_poll_req frame, which then queues the endgame again. Whether the close lands inside the inner run is a race. That is why CI sees it in about 1 of 10 builds per lane, and why the shape varies (0x0, 0xFFFFFFFFFFFFFFFF, Option::unwrap() on a None value, 60 s hangs). A copy of the two fixture cases hangs or crashes with the un-awaited matchers and passes with them awaited.

Reduction without the dev server. The Bun.sleep(100) keeps the inner run going long enough to run the endgame. The console.log calls are part of it: which block the stray free hits depends on the heap layout, and a variant without them passed 3 of 3 runs on the canary.

import { expect, test } from "bun:test";

test("peer closes the socket during a nested tick inside its own poll callback", async () => {
  const server = Bun.serve({
    port: 0,
    fetch(req, server) {
      if (server.upgrade(req)) return;
      return new Response("no");
    },
    websocket: { message() {} },
  });

  const { promise: done, resolve: finish } = Promise.withResolvers<void>();
  const ws = new WebSocket(`ws://localhost:${server.port}/`);
  ws.onopen = () => {
    const closed = new Promise<number>(resolve => (ws.onclose = e => resolve(e.code))).then(code =>
      Bun.sleep(100).then(() => code),
    );
    setTimeout(() => server.stop(true), 10);
    // Not awaited on purpose: runs a nested event loop tick inside this poll callback.
    expect(closed).resolves.toBe(1006);
    console.log("nested tick returned, readyState =", ws.readyState);
    finish();
  };
  await done;
  console.log("onopen returned");

  // Like the fixture's next case: allocate a batch of fresh connecting sockets.
  await using server2 = Bun.serve({
    port: 0,
    fetch(req, server) {
      if (server.upgrade(req)) return;
      return new Response("no");
    },
    websocket: { message() {} },
  });
  const opens: Promise<void>[] = [];
  const sockets: WebSocket[] = [];
  for (let i = 0; i < 8; i++) {
    const { promise, resolve, reject } = Promise.withResolvers<void>();
    const s = new WebSocket(`ws://localhost:${server2.port}/`);
    s.onopen = () => resolve();
    s.onclose = () => reject(new Error(`socket ${i} closed before open`));
    sockets.push(s);
    opens.push(promise);
  }
  await Promise.all(opens);
  console.log("all 8 follow-up sockets opened");
  const closes = sockets.map(s => new Promise<void>(r => (s.onclose = () => r())));
  for (const s of sockets) s.close();
  await Promise.all(closes);
});

Results on one Windows x64 machine:

build result
canary 1.4.0-canary.1+32e87032b 4 of 4 runs crash, exit code 3: two at 0x0, two at 0xFFFFFFFFFFFFFFFF
debug build of main cfa9f8e15b 3 of 3 runs crash at 0xFFFFFFFFFFFFFFFF
the same debug build with this PR merged 5 of 5 runs pass
the same build, test/bake/deinitialization.test.ts 10 of 10 runs pass (the unmodified canary hangs on it, 2 of 2 runs)

This reduction closes the socket from a second poll_cb frame on the same handle. The test in this PR closes it from the first frame. It may be worth adding next to it.

alii added a commit that referenced this pull request Aug 20, 2026
@alii

alii commented Aug 20, 2026

Copy link
Copy Markdown
Member

For reference from #39488: the test/bake/deinitialization.test.ts Windows x64 segfaults on that branch (builds 101180 and 101478, both at the start of the websocket=8 case) symbolise to heap corruption in the client us_socket_t size class surfacing under us_internal_drain_pending_dns_resolve, i.e. the family this PR fixes. I briefly carried the tick_depth-only subset there and backed it out in favour of this landing whole.

@robobun

robobun commented Aug 20, 2026

Copy link
Copy Markdown
Collaborator Author

Thanks, that shape fits. The second close_cb_free_poll frees whatever block the free-list link in the dead uv_poll_t points at, so the stray free lands on a live block of the same size class, and the next user of that block crashes. A us_connecting_socket_t or client us_socket_t on the DNS ready list is one of the first users in that fixture. tick_depth on its own does not stop that: it keeps the outer dispatch from reading freed memory, but the inner run still finishes the close and the outer frame still queues the endgame again. So backing the subset out was right, and this PR has to land as one piece.

Current state: f91d44edb8. The first push closed the socket before the deferred uv_close, and appcontainer.test.ts caught it on the Windows 11 arm64 lane (strict handle checks are on inside an AppContainer, so libuv's cancel ioctl on the closed socket ended the process). The close paths now close the socket right after the uv_close, the order libuv had before. Build 101247 on this head is green on every lane, and deinitialization.test.ts passes on the x64 and arm64 debug builds here.

@robobun

robobun commented Aug 20, 2026

Copy link
Copy Markdown
Collaborator Author

Release build evidence for the test/bake/deinitialization.test.ts part of this PR. The Windows failure of that test was handed to me as a separate bug (seen on #34654, Buildkite builds 99642 and 101711, Segmentation fault at address 0x0 after the first 8 cases pass). It is this bug, and this PR fixes it. I am not opening a second PR.

A/B on one machine (Windows Server 2019 x64, CPU without AVX, so the build is the baseline variant like the CI lane). Both binaries are local --profile=release builds from the same tree and toolchain. The only difference is this PR's diff (f91d44edb8 merged onto main 01c4e2fd6d). Each run is the child fixture, bun test ./test.ts in test/bake/fixtures/deinitialization, with the env that bunEnv sets.

  • main 01c4e2fd6d without this PR: 8 of 8 runs fail. 6 runs hang with the main thread at 100% CPU. 2 runs die with panic(main thread): Segmentation fault, at address 0x0 once and 0x270 once. Every run fails after the 8th (pass) line, in the flags: websocket=8 case. The CI canary of the same commit (1.4.0-canary.1+01c4e2fd6) fails 6 of 6 runs the same way (5 hangs, 1 segfault at 0x26C).
  • main 01c4e2fd6d with this PR: the fixture passes 20 of 20 runs (10 tests each, exit 0). The outer test/bake/deinitialization.test.ts passes 3 of 3 runs.

The PR body lists a debug build for this test. The numbers above cover the release build, which is the build that hangs or crashes on the CI lanes.

@robobun

robobun commented Aug 20, 2026

Copy link
Copy Markdown
Collaborator Author

Verified this branch (f91d44edb8) against one more signature of the same bug: BUN-4MNY, a Windows x64 crash in ssl_retry_parked_write at the tail of us_internal_ssl_on_data (openssl.c:2423), reported on 1.4.0.

Repro: a Bun.connect client writes in open(), calls terminate() in its first data(), and then runs a nested loop from that handler (Bun.build() with a plugin whose setup() returns a promise that settles after 20 timer hops). On a Windows x64 debug build of main (01c4e2fd6d) it segfaults at 0xFFFFFFFFFFFFFFFF on every run: the TCP variant in us_internal_socket_follow_adopted under loop.c:746, the TLS variant in ssl_flush_pending_session under openssl.c:2429, the same tail as the Sentry stack. On this branch both variants pass 15 of 15 runs. The fixture is test/js/bun/net/socket-close-in-nested-loop-fixture.ts on farm/54233a81/libuv-nested-tick-free, in case the TLS tail is wanted as a second case here. The rest of that branch is superseded by this PR.

One note on where tick_depth is counted. us_loop_run and us_loop_pump cover a nested run that Bun enters. An addon that holds the loop from napi_get_uv_event_loop and calls uv_run() from inside a handler (BUN-4MNY has an addon loaded) runs check_cb at depth 1 and frees the closed list. poll_cb_depth still protects the socket whose own callback is on the stack, but not a socket another frame holds, for example an accepted socket closed from its own open() while the accept loop still reads it (loop.c:556). Counting in the callbacks libuv runs for the loop instead (poll_cb, prepare_cb, check_cb, timer_cb, async_cb, with check_cb's own increment making the outermost loop_post depth 1) covers both entry paths, and is a small delta on top of this design.

@robobun

robobun commented Aug 21, 2026

Copy link
Copy Markdown
Collaborator Author

State after the merge of main (ca25925ba5): the diff itself is unchanged from f91d44edb8, and build 102732 is running on the merged head.

The notes in the description now also record one limit. An addon that calls uv_run() on this loop from inside a handler is not counted by tick_depth. The socket whose own callback is on the stack is still covered by poll_cb_depth, a socket that another frame holds is not. None of the crashes listed here needed that, so it is left for a follow-up (counting the frames of the callbacks libuv runs for us would cover it).

@claude claude Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I reviewed this PR and didn't find any bugs. Because it redesigns the Windows poll lifecycle state machine (five new lifecycle bits plus poll_cb_depth, deferred uv_close/fd-close ordering, and a new released/uv_closed free handshake) across every socket close path, a human look from someone with libuv/usockets ownership would still be worthwhile.

What was reviewed:

  • The released/uv_closed handshake in close_cb_free_poll/us_poll_free — each block is freed exactly once regardless of which runs first; the old data = NULL leak (#37105) is covered.
  • poll_cb re-reading p->data after dispatch so a mid-dispatch us_poll_resize carries poll_cb_depth/stopped to the new block, and the retired block's uv_p = NULL short-circuits us_poll_stop/us_poll_free.
  • us_internal_poll_close_fd on epoll/kqueue is a no-op wrapper over bsd_close_socket, and the kqueue branch of close_raw (which skips us_poll_stop) still closes immediately there.
  • tick_depth now brackets us_loop_run/us_loop_pump, so check_cb → us_internal_loop_post's tick_depth <= 1 gate holds for nested uv_run; no Rust-side layout mirror of us_poll_t exists to fall out of sync.
Extended reasoning...

Overview

This PR fixes a family of Windows-only use-after-free / double-free crashes in bun-usockets' libuv backend that surface when a socket handler drives a nested uv_run (e.g. expect(promise).resolves, auto-install). It touches eventing/libuv.c (the core lifecycle rewrite), internal/eventing/libuv.h (five new bitfields plus poll_cb_depth on us_poll_t), epoll_kqueue.c/internal.h (a pass-through us_internal_poll_close_fd), loop_data.h (comment only), and rewires four close sites in context.c/socket.c/udp.c from bsd_close_socket(us_poll_fd(p)) to the new helper. Two tests are added: a subprocess fixture in socket.test.ts that reproduced the BUN-43AH segfault on main, and an AppContainer step guarding the strict-handle-check ordering.

Security risks

None identified. This is internal event-loop lifecycle plumbing with no user-controlled input reaching the new state; the changes narrow, rather than widen, the window in which freed memory can be touched.

Level of scrutiny

High. This is native memory-safety code that changes when uv_close, the socket closesocket(), and the two heap frees run relative to poll-callback frames and nested uv_runs. The design introduces a small state machine (stopped/close_fd/uv_closed/released/poll_cb_depth) whose invariants must hold across adoption/resize, the uv_poll_init_socket failure path, us_socket_detach, and UDP. The PR body documents each invariant and the CI iteration that caught the fd-before-uv_close ordering (AppContainer 0xC0000008), and the A/B numbers on deinitialization.test.ts are compelling — but the design choices (defer uv_close to the outermost poll_cb frame rather than patching libuv; count tick_depth at us_loop_run/us_loop_pump rather than in every libuv callback) are exactly the kind a maintainer should ratify.

Other factors

  • The bug-hunting pass found nothing. I traced the free protocol in both orders, the resize path (memcpy carries the counter; the retired block's uv_p = NULL makes us_poll_stop/us_poll_free no-ops on it), the kqueue close_raw branch (still closes immediately since epoll/kqueue's us_internal_poll_close_fd is just bsd_close_socket), and us_socket_detach (defers uv_close, leaves close_fd = 0, so the fd survives as intended).
  • us_internal_loop_post at loop.c:446 already gates the closed-list free on tick_depth <= 1; this PR wires the libuv entry points into that counter, matching us_loop_run_bun_tick on POSIX. I confirmed there is no Rust #[repr(C)] mirror of us_poll_t that the added fields would desync.
  • The PR notes an acknowledged limit (native addons calling uv_run directly are not counted by tick_depth) and defers it; that is a design call worth a maintainer's nod.
  • Build 101247 was green on all lanes at f91d44edb8; the diff is unchanged since. alii's comment corroborates this fixes the #39488 Windows failures.

@robobun

robobun commented Aug 21, 2026

Copy link
Copy Markdown
Collaborator Author

Cross reference: #39910 (opened 2026-08-21, CI green) fixes the same bug. It also counts tick_depth around uv_run in us_loop_run and us_loop_pump, and it also stops us_poll_stop from starting the uv_close. It issues the uv_close from us_poll_free instead of counting poll_cb frames, so it is smaller (+118/-55 against +295/-65 here). Neither description references the other.

I built both heads (063e4c0f7 here, dd7070d64 there) on a Windows Server 2019 x64 machine. The baseline is the canary build of 40ef8113d, the main commit both are based on:

The full comparison is in #39910 (comment). It lists the hunks and tests in this PR that #39910 does not have: the us_internal_poll_close_fd close order and the poll_cb_depth it depends on, the released / uv_closed bits (only this design needs them), the defensive guards after a stop, the two socket.test.ts cases and the appcontainer step, and the Sentry groups. #39910 has a fixture that closes from the client side with allocation churn, which this PR does not have. Both PRs also cover #37105.

I am not closing either PR. One of the two should land and the other should close. If #39910 lands first, the close order hunks and the appcontainer step from here should follow it as a rebased follow-up, or AppContainer processes regress from main.

@claude claude Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I reviewed this PR and the automated bug hunt found no issues. Given the scope — a redesign of the libuv poll lifecycle (deferred uv_close via poll_cb_depth, the released/uv_closed two-block hand-off, deferred fd close for the AppContainer ordering) — a human look is still warranted, and there's a competing design in #39910 to decide between.

Checked: the four bsd_close_socket → us_internal_poll_close_fd call sites all follow us_poll_stop (or its kqueue equivalent), so the epoll/kqueue path is a straight rename. us_poll_resize copies the new bitfields and poll_cb_depth into the replacement block, and poll_cb re-reads p->data after dispatch, so a mid-dispatch adopt keeps the counter and deferred-close state consistent. The h->type != UV_POLL guard in us_poll_stop/us_poll_free relies on us_create_poll's new us_calloc zeroing, which is in place.

Extended reasoning...

Overview

This PR reworks the Windows/libuv poll lifecycle in packages/bun-usockets to fix a family of use-after-free crashes (BUN-43AH, BUN-442H/Y/Z, BUN-4NC8, BUN-4NP3, and four TLS variants) that occur when a socket handler synchronously waits for a promise, which re-enters uv_run while the outer uv__fast_poll_process_poll_req frame still holds the handle. The fix has three interlocking pieces: (1) tick_depth is now counted around uv_run in us_loop_run/us_loop_pump so us_internal_loop_post defers freeing the closed list to the outermost tick (POSIX parity); (2) us_poll_stop only disarms the handle while poll_cb_depth > 0, and the outermost poll_cb frame issues uv_close on its way back into libuv, which is the reentrancy shape libuv supports; (3) us_poll_free and close_cb_free_poll coordinate via released/uv_closed bits so whichever runs second frees both blocks. A new us_internal_poll_close_fd helper defers the socket close until after the deferred uv_close is issued, preserving the order libuv's cancel ioctl needs (AppContainers with strict handle checks die otherwise — caught by CI on the first push). Four call sites in context.c/socket.c/udp.c are rewired through the helper; on epoll/kqueue it is a straight bsd_close_socket wrapper.

Security risks

None identified. This is event-loop lifecycle and memory-ownership code, not input parsing or auth/crypto. The change closes UAF/double-free bugs rather than introducing new attack surface.

Level of scrutiny

High. This is native C in the event loop, the "memory safety (the most-blocked category)" section of the review rules. It redesigns two-block ownership across an asynchronous libuv close callback, adds five state fields to us_poll_t, and changes when uv_close and closesocket run relative to nested uv_run frames. The PR description is exceptionally thorough (symbolized fail-before, A/B on release builds, Sentry group mapping, the AppContainer regression it caught and fixed mid-PR), and CI is green on the merged head. But the design has an acknowledged limit (native addons calling uv_run directly), one defensive path was "reasoned about, not run" (the fault-injection failure branch), and #39910 offers a smaller competing design that this PR's author has A/B-tested and shown to regress the AppContainer case — that trade-off is a design decision a maintainer should make.

Other factors

  • No prior human review on the thread; the only non-bot comment (from alii) is a cross-reference confirming this fixes their branch's crashes too.
  • Tests: a new subprocess-spawning case in socket.test.ts that reproduces the crash on unfixed builds (verified in the PR notes), and a new closeInHandler step in appcontainer.test.ts that guards the socket-close ordering. Both are Windows-relevant and cannot run on this reviewer's platform.
  • The us_poll_t struct grows by ~5 bytes of bitfields plus an unsigned int; us_create_poll initializes all five new fields explicitly, and us_poll_resize's memcpy carries them to the replacement block.
  • I traced us_poll_resize against the new poll_cb epilogue: the epilogue re-reads wp = p->data after dispatch, and resize sets p->uv_p = NULL on the retired block while pointing uv_p->data at the new one, so the deferred uv_close and close_fd land on the live block. The retired block's own us_poll_free (from us_internal_free_closed_sockets) takes the !h fast path.
  • The h->type == UV_POLL re-start guard in us_poll_start_rc and the p->stopped guard in us_poll_change are noted as having no current caller; they are defensive and look correct against the state machine described in the header comments.

@robobun

robobun commented Aug 22, 2026

Copy link
Copy Markdown
Collaborator Author

For whoever merges: #39910 took the same two steps (tick_depth, and no uv_close from under the poll callback) but issued the uv_close from us_poll_free, after close_raw has closed the socket. That is the order the AppContainer lane rejects (strict handle checks make libuv's cancel ioctl on the closed socket fatal, see the notes above), so this PR issues it from the outermost poll_cb frame and closes the socket right after, the order main has. The side by side is in the #39910 thread, and that PR is now closed.

Build 103029 on the rebased head 174085f6ee: the only red test is test/cli/install/bun-audit.test.ts, which fails on main since the 1.4.1 version bump (normalizeBunSnapshot rewrites the <1.4.1 range in the mime advisory) and is reported for a fix there. Everything else passed, or passed on retry.

@robobun

robobun commented Aug 23, 2026 •

Copy link
Copy Markdown
Collaborator Author

Two more downstream symptoms for the Sentry list in the description, both from one Windows bun test session on 1.4.0 (34cbb9a):

  • BUN-4Q7W (2026-08-22 17:23 UTC): segfault in JSC::CodeCacheMap::canPruneQuickly at 0x1C7BFA80004, from new Function. The fault address is m_table - 12 (the WTF HashTable key count), so m_table was 0x1C7BFA80010: a 64 KiB arena slice start plus the 16 byte table header. That is exactly where mimalloc v3 places a singleton block with MI_PAGE_META_IS_SEPARATED, so the pointer is the real table. Its first OS page is decommitted. The table's page was freed and purged while the CodeCache still pointed at it. A stale mi_free of any address inside a singleton page frees the whole page, which is what a second close_cb_free_poll run does once the old small page's slices have been reused.
  • BUN-4Q6H (17:25 UTC, the same suite run again): segfault in the SourceCodeKey bucket read of the same map at 0x21C26CF46F0. The table header was readable and the bucket index was out of range, so tableSizeMask had been overwritten by another owner of that memory.

On Windows _mi_prim_decommit is VirtualFree(MEM_DECOMMIT), so a read through a stale pointer faults. On Linux and macOS it is madvise, the pages stay readable, and the same corruption stays silent. That is why this family only reports from Windows.

The same user, with an identical tag set (TestCommand, http_server, fetch, spawn, Bun.stdin), hit BUN-4NP3 (us_internal_socket_group_unlink_socket on a freed socket) 13 minutes before, and BUN-442Z the same morning. Windows release builds link WebKit against the same mimalloc as uSockets (USE_EXTERNAL_MIMALLOC=ON), so JSC's FastMalloc objects share the heap with us_socket_t and uv_poll_t.

Also from that session: BUN-4NA3 (5 events, fault inside mimalloc's free list walk under WindowsNamedPipe::on_read_alloc) and BUN-4P1X (2 events, uWS::LoopData::findCorkSlot). In total 13 crashes across 8 issues in 17 hours, all heap corruption shaped.

@Jarred-Sumner Jarred-Sumner left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Read end to end against libuv's win/poll.c / uv__process_reqs. I think this is correct and is the right shape; notes for the merger:

Why the deferral is sufficient. The original double endgame needs a uv__fast_poll_process_poll_req(A, req1) frame on the stack (it cleared submitted_events_1 before calling poll_cb) while an inner uv_run processes A's cancelled req2, sees CLOSING && submitted == 0, and runs the endgame; the outer frame then queues it again. With uv_close deferred until A's outermost poll_cb returns, no inner run can observe CLOSING for a poll that has a frame on the stack. For a poll B closed at depth 0 from inside someone else's nested run, B has no frame on the stack, and a B request already sitting in the outer run's detached pending list is harmless: each request clears only its own submitted_events_*, so B's endgame is queued exactly once, after both are processed, and close_cb_free_poll running before us_poll_free (outer loop_post at tick_depth <= 1) is the uv_closed = 1 arm.

Second producer fixed in passing. us_poll_start_rc used to memset + uv_poll_init_socket a handle that was already registered (adopt/resize paths reach it with h->type == UV_POLL). That wipes libuv's handle_queue links and the in-flight AFD request state; the next completion lands in a handle libuv no longer tracks. The new "already registered → just uv_poll_start the mask; stopped → EBADF" branch closes that.

Checked: us_poll_resize memcpys the header so stopped/close_fd/uv_closed/released/poll_cb_depth move with the block and poll_cb re-reads p->data after dispatch; the context.c:715 failure path (bsd_close_socket between the uv_close issued inside start_rc and us_poll_free) keeps the close-handle-then-close-socket order the AppContainer lane needs; sizeof(struct us_poll_t) is unchanged on LLP64 (the flag byte and the unsigned fit in the padding after poll_type); Bun's own direct uv_run calls are either on spawnSync's isolated loop or in worker/loop teardown, so the uncounted-nested-run limit in the notes really is addons only.

Nit, non-blocking: us_poll_stop returns early for a never-registered handle without setting stopped; nothing depends on it today, but setting it there too would make "stopped ⇒ never re-armed" hold unconditionally.

This is one of the two identified producers of the Windows corrupted-free-list family (BUN-40BH and siblings; the other, #39897, is merged), so it's worth landing promptly.

@robobun

robobun commented Aug 23, 2026

Copy link
Copy Markdown
Collaborator Author

Thanks for the read. The nit is in 375711fac0: us_poll_stop sets stopped before the h->type != UV_POLL return, and us_poll_start_rc now refuses a stopped poll before its registered-or-not branch (before, it only checked the flag for a registered handle, so the invariant needed both). No other change. On a Windows x64 debug build of that commit the new socket.test.ts case, appcontainer.test.ts, udp_socket, deinitialization and node-tls-server pass.

One detail on the second producer, for the record: us_poll_start_rc is reached with a registered handle only through the fault-injection path today (every adopt or resize keeps the handle and goes through us_poll_change), so that branch is a guard rather than a fix for a crash we saw. The memset it replaces was still wrong for the reason you give.

robobun added a commit that referenced this pull request Aug 26, 2026
…ed sleeps

test/bake/deinitialization.test.ts now pipes the child's output and
snapshots the per-case results and summary instead of checking only the
exit code. The fixture waits for the HMR clients' close events instead of
sleeping 250 ms in the plugin and after each abrupt stop, and polls the
deinit count with 10 ms turns. The child suite takes 0.09 s instead of
2.1 s on a release build.

Adds a timer-free fixture that closes a WebSocket from the server side
while a synchronous `expect().resolves` runs a nested event loop inside
the client's open callback. On a Windows debug build of main it faults at
0xDFDFDFDFDFDFDFF9 in us_internal_socket_follow_adopted, the same frames
as the deinitialization hang on the Windows CI lanes. The fix for that is
tracked in #40021 and #39643; this branch is test-only.

@claude claude Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code review found no issues

No high-confidence issues detected in this change.

@claude claude Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code review found no issues

No high-confidence issues detected in this change.

On kqueue the client's connect completion is a one-shot write filter. It
can arrive in the same kevent batch as the accept, and the nested tick in
open() drops the rest of that batch, so the connect promise never settles
and the fixture timed out on the darwin lane. Only the server side is
under test, so the client is left to itself.

@claude claude Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code review found no issues

No high-confidence issues detected in this change.

@cirospaciari
cirospaciari merged commit bb8aa37 into main Sep 11, 2026
6 checks passed
@cirospaciari
cirospaciari deleted the farm/47550167/win-poll-free-protocol branch September 11, 2026 01:24
cirospaciari added a commit that referenced this pull request Sep 11, 2026
Jarred-Sumner added a commit that referenced this pull request Oct 20, 2026
… and CLI (#46021)

### What does this PR do?

Hardening. This collects the open pull requests that fix a crash in Bun
into one change, and fixes about 20 more crashes that had no pull
request.

"Crash" here means a segfault, a Rust panic, an abort, a failed
assertion, memory corruption that a sanitizer reports, a native stack
overflow, or an exception that ends the process and that the program
can't catch. With this PR those cases throw a normal error, print a
normal CLI error and exit 1, or just work.

306 pull requests are included: 227 whole, and 79 cut down to the crash
fix where the rest was a refactor, a feature or a behavior change. Each
one was checked against current `main` first. 45 more turned out to be
fixed already, and 17 were duplicates. The full list is at the bottom.

#### Crashes you could hit in normal use

- **`bun build` with macros and a `.env` file crashed at random.** In a
repro with many macro modules, more than a third of the builds crashed.
Now 3,000 of 3,000 produce the right bundle. (#45960)
- **`bun run` and `bunx` aborted if a signal arrived while the child was
being spawned.** (#38926, #45965)
- **`bun build --compile ./index.js ./public/*` panicked** if an
embedded file's name wasn't valid UTF-8.
- **A truncated or corrupted compiled executable** died with SIGBUS or a
panic at startup. It now prints an error. (#31787, #38251, #41360)
- **Dev server:** several `Option::unwrap()` panics after a failed or
deleted import (#38046, #44744, #44746), a segfault when an `onLoad`
plugin throws while another bundle is queued, a request that hung and
then panicked with `bundle active`, and `server.stop()` that never
resolved after a client aborted a waiting request.
- **`Bun.serve`:** `server.fetch(url, { headers: new Headers() })`
followed by a GC segfaulted (#40888). Routes were used after free after
`server.reload()` with `tls.serverName` (#38636).
- **`text()` and `json()` on a stream of string chunks** could abort the
process, a regression in 1.4.0. (#44269)
- **`Bun.Transpiler` with `exports.eliminate` or `exports.replace`**
aborted on `export default 1` and on most re-export clauses. (#44137)
- **A decorated `declare` field in a class expression** with
`experimentalDecorators` panicked. (#38095)
- **YAML and JSON5 files with a key that isn't a string** panicked with
`--no-bundle` and in `Bun.Transpiler`. (#41827)
- **`Bun.FileSystemRouter`:** reading `.params` segfaulted when a
rejected route candidate left its parameters behind. (#39290, #39291)
- **Native addons that call `uv_async_*`, `uv_timer_*`, `uv_idle_*`,
`uv_prepare_*`, `uv_check_*`, `uv_queue_work` or `uv_cwd`** aborted with
"unsupported uv function" on Linux and macOS. They're implemented now.
(#36609, #39652, #39892)
- **`--watch` and `--hot`** panicked when the system was out of inotify
instances or file descriptors, or when the reload `exec` failed. They
print an error now. (#39629, #39975, #44837)
- **Paths longer than the path buffer** panicked in `bun install`, the
resolver, the bundler, `--compile`, `tsconfig.json` `paths` and
`baseUrl`, `package.json` `main` and `sideEffects`, `bun run --filter`,
`bun pm diff`, source maps and snapshots. They're ordinary errors now,
or they work. (#36340, #37529, #39104, #39650, #43067)
- **A Worker that creates a `ShadowRealm` and imports
`node:worker_threads` in it** segfaulted.
- **`structuredClone` or `postMessage` of an X25519 key** gave an object
that aborted the process when read. (#45624)
- **`node:stream` web adapters and `node:http2`** could end the process
with an exception that no handler could catch. (#45076, #45975)
- **Windows:** a watcher panic in a burst of file changes (#44344), a
segfault after the machine resumes from sleep (#33015), crashes in the
shell's pipe reader (#39638, #43606), and addons that import from
`node.exe` without the delay-load hook (#35618).

Most of the rest needs unusual input: strings near the 2 GiB limit,
buffers that another thread or a getter resizes or detaches in the
middle of a call, `worker.terminate()` at a bad moment, running out of
memory or file descriptors, very deep nesting, or hostile arguments to
`bun:ffi`, `node:vm` and `bun:sqlite`.

#### Behavior changes

Things that didn't crash before and behave differently now:

- `new` on native functions that aren't constructors throws a
`TypeError`, as in Node: `util.types.*`, `buffer.isAscii`, `isUtf8`,
`transcode`, `resolveObjectURL`, `bun:jsc` exports, `cc()` symbols. It
used to return the call's result. They also stop printing as `[class
hostname]`. (#34575)
- `` $`cmd` `` throws `ERR_INVALID_ARG_VALUE` if an environment
variable's name or value contains a NUL byte. It used to cut the value
at the NUL. (#44981)
- `cc()` functions throw a `TypeError` for arguments that used to reach
C as garbage, and `read.*(ptr, NaN)` throws. (#37989, #38057, #40773)
- As PID 1, `bun` exits with 128 + the signal number on Ctrl+C or
`docker stop`. (#38877)
- A `NODE_MODULE` addon that's already loaded by the main thread throws
`ERR_DLOPEN_FAILED` when a live Worker loads it too, as in Node.
(#39968)
- An exception from a `node:vm` context reaches `uncaughtException`
handlers. (#39992) `vm.constants.DONT_CONTEXTIFY` returns the real
global proxy. (#34623)
- `createPrivateKey` rejects an EC JWK whose private value doesn't match
its public point. (#33752)
- `@dec static {}` is a syntax error, as in TypeScript and esbuild.
(#44883)
- `clearTimeout` ignores 0 and negative ids, and timer ids no longer
wrap. (#44771)
- HEAD for a `Response` whose body is an S3-backed `Blob` reports the
length that GET sends. (#44734)
- `bun:sqlite` and `node:sqlite`: if `iterate()` left a write unfinished
on a connection that has a hook, a user function or an authorizer, the
write is finished on the connection's next call instead of during
garbage collection. If the database is gone too, it's rolled back.
- `bun test --bail` runs exit listeners and closes SQLite databases.
(#39010)
- Dev server: a request for a client script waits while a bundle is in
flight (#42733), query strings reach framework routes (#33232), and a
request gets a 500 instead of hanging when plugins fail (#36812).

#### File formats

- `.jsc` bytecode files get a 24-byte footer so they can be verified
before decoding. Files from an older Bun are ignored and the source is
parsed instead. (#39396)
- `node:vm` `cachedData` gets an 8-byte checksum. A blob from an older
Bun is rejected, which the API allows. (#32839)
- The built-in modules section of a compiled executable changes from
format 1 to 2 and carries an engine stamp. (#44907)

#### Performance

Measured in instructions per operation on release builds of this branch
and its base, with an A/A run for the noise floor (0.0 to 0.2 %).

No change: `Bun.serve`, `node:http`, WebSocket echo and publish, `bun
build` of 3,807 modules, `bun run` with 2,000 imports, `readdir` /
`Glob` / `fs.cp` over 23,689 entries, startup and RSS. The binary grows
by 176 KB (0.17 %).

| Faster | Instructions |
| --- | --- |
| WebSocket echo of a 1 MB non-ASCII text message | −72 % |
| `Bun.inspect` of a large non-ASCII string | −39 % |
| `Bun.markdown.html` on large non-ASCII input | −33 % |
| `` $`cmd` `` with 300 environment variables | −6 % |
| `fs.readSync` | −5 % |
| `TextEncoder.encodeInto`, short ASCII strings | −4 % |
| async `crypto.pbkdf2`, 16-byte inputs | −17 % |

| Slower | Instructions | Why |
| --- | --- | --- |
| `cc()` calls | +5 to +51 per call | Argument count and type checks,
and a check for `close()` (#37989, #38057, #43802, #44921). `dlopen()`
calls don't change |
| `process.env.X = v` | 38 → 593 per write | Writes aren't cached. Reads
don't change (#38871) |
| async `crypto.pbkdf2`, inputs of 1 KB or more | +4 to +9 % | Password
and salt are copied (#40554) |
| `bun build` with an object `--define` used thousands of times | +4.5 %
| The object is copied per use (#45969) |

#### Not included

- Fixes that need a WebKit change: #39967, #42534, #42982, #43764,
#43786, #44226. #42185 too: a typed array that has no `ArrayBuffer` yet
can't be pinned, so `view.buffer.transfer()` during an async call can
still free its bytes. Creating the `ArrayBuffer` up front costs 7 to 12
% on every async `fs` call, so the fix belongs in JSC: an `ArrayBuffer`
created while the view is borrowed should start out pinned.
- #45442 (`calc()` sums as a flat list): it changes minified output for
real stylesheets to survive a `calc()` with 25,000 terms. It deserves
its own review.
- #40023 (Windows event loop redesign), #29928 (Windows exit path),
#40534, #43963, #39571, #42904, #42834, #36355, #43266: refactors or
features whose crash is already covered or isn't a crash.
- The pull requests marked *(crash fix only)* below should stay open for
their other half.

### How did you verify your code works?

- **Tests.** Every fix has a test that fails on `main` (a release or an
ASAN build) and passes on this branch, except where that isn't
practical: code that only runs on Windows or macOS, inputs that need
several GB of memory, and a few rare races. Those were checked by hand
or by reading.
- **Stored repro scripts.** 230 crash repros ran on release and ASAN
builds of this branch and of `main`, side by side. Release: 88 no longer
crash. ASAN: 106 no longer crash. None crashes that didn't before. What
still crashes needs WebKit, is SQLite's own bug, or isn't a crash. The
ASAN number is from a later commit than the release number.
- **Dev server.** Because one fix makes requests wait, 2,171 generated
programs ran random interleavings of edits, build errors, failing
plugins, reloads, stops and disconnects with about 120,000 requests. No
request hung, was answered twice or got a stale body. `main` hung in 82
of 1,161 programs. This also found the four dev server bugs listed
above.
- **Path helpers.** 336,000 boundary cases with buffers of exactly the
computed size, and 12,003 random calls through 30 JS APIs on an ASAN
build, compared with `main`. No sanitizer report, and the results match
wherever `main` doesn't panic.
- **What worked on 1.4.3 still works.** Reviewers compared behavior with
1.4.3 and took out what changed it: for example an output directory
layout change for 9 or more entry points, a `server.reload()` change,
and a changed error message for `import("file:///50%deposit.ts")`.
- **Not run locally:** the Windows, macOS and FreeBSD code paths. CI
compiles and tests them.

### Included pull requests

**Bundler, transpiler, CSS, macros** (33)

- #33881 bundler: allocate wrapper_ref for CommonJS modules with no
non-hoistable statements
- #35055 Bun.build plugins: handle filter RegExp flags beyond i/m/v
- #37769 bundler: fail the build instead of panicking when the bundle
thread cannot create its waker
- #37784 bundler: copy inlined env values into the define arena, and
give each Bun.build its own env copy *(crash fix only)*
- #37858 bundler: fail the build instead of panicking when a plugin
answer string cannot be converted
- #38046 bun build --server-components: require the bun target and
reject directives without a framework *(crash fix only)*
- #38095 js_parser: fix panic on decorated declare/abstract fields in
class expressions with experimentalDecorators
- #38966 bundler: keep deep CSS, export star and composes graphs from
overflowing the linker's stack
- #38984 bundler: stop an export star ambiguity that loops back into a
re-export chain from overflowing the stack
- #39113 css: reject stylesheets of 2 GiB or more instead of aborting on
an int cast
- #39396 bundler: frame .jsc bytecode sidecars and verify them before
decoding
- #39855 bundler: join the work still on the pools before the bundle is
torn down *(crash fix only)*
- #41157 dev server: leave out an html link to a cached json module
instead of panicking, and mark a new placeholder file stale
- #41801 css: decode ill-formed UTF-8 to U+FFFD before tokenizing
*(crash fix only)*
- #41827 bundler: handle YAML and JSON5 keys that are not strings
instead of panicking
- #42242 js_parser: compress symbol link chains when following merged
declarations
- #42387 bundler: check receivers on the Bun.build plugin object and
freeze its filters during a build
- #42413 Bounds-check the InternalSourceMap window reader
- #42657 macros: keep visited values alive so a reused address does not
get another value's Expr
- #43330 transpiler: do not print an empty `var` for a data file with no
named export
- #43355 bundler: fix "The memory limit has been exceeded." on large
HTML entries
- #43658 react-compiler: fix skipped functions and build aborts for dead
code in a try body *(crash fix only)*
- #44137 js_parser: fix four crashes of exports.eliminate and
exports.replace
- #44467 js_parser: replace skipped expressions with E::Missing and skip
argument checks after a stack overflow
- #44662 js_parser: walk binary chains on the heap and check the stack
when rewriting decorated #private accesses
- #44712 json5: reject sources of 2 GiB or more before parsing instead
of aborting on an int cast
- #44864 js_parser: walk conditional chains in import()/require() and
define keys in a loop
- #44883 js_parser: reject a decorator on a class static block, keep
decorators out of the body of `declare global` *(crash fix only)*
- #45045 Fix a segfault in new Bun.Transpiler(): re-point options.log
after the constructor moves its Log
- #45604 css: keep a url() in an at-rule prelude as written, and never
print the resolved path of a url() *(crash fix only)*
- #45748 bundler: put --banner/--footer in the bytecode ESM module
record (flags from a parse, entries from JSC) *(crash fix only)*
- #45960 macros: give each parse-worker macro VM its own env loader
- #45969 js_parser: copy object and array defines per use, and copy the
list of a lone JSX spread object *(crash fix only)*

**`bun build --compile` and standalone executables** (3)

- #31787 Don't crash at startup when a standalone executable's embedded
graph is corrupted *(crash fix only)*
- #38251 compile: exit with an error instead of SIGBUS when a standalone
executable is truncated
- #41360 standalone: check every offset in the embedded module graph
before reading it

**Resolver, module loading, plugins, paths** (25)

- #32269 Bun.plugin: validate the namespace without a shared regex
- #33183 bun:jsc: filter internal cells out of getProtectedObjects()
- #34575 Don't register native functions as their own constructors
- #36340 paths: size the thread-local join output buffers to
MAX_PATH_BYTES on Windows
- #37037 mock.module: reject async factory results that are not objects
- #37065 bun:jsc: propagate JSONParse exceptions from
samplingProfilerStackTraces
- #37529 resolver: don't abort on a package.json sideEffects entry
longer than the join buffer
- #38179 node:module: throw instead of crashing when module._compile is
set to a non-function
- #38273 bun test: report a test file or preload the loader cannot
resolve instead of exiting; keep the JUnit report valid U… *(crash fix
only)*
- #38732 paths: compute the common path of 9+ inputs with the same loop
as fewer *(crash fix only)*
- #39228 paths: strip C:\\ down to the drive root in
without_trailing_slash_windows_path *(crash fix only)*
- #39650 bun run --filter: stop panicking on workspace paths longer than
the path buffer
- #40186 resolver: stop deeply nested exports targets from overflowing
the stack
- #40258 resolver: publish Entry.abs_path through the per-entry mutex
- #42183 node:module: reject a _resolveFilename override result that
contains a null byte
- #42401 node:module: accept any `this` in Module.prototype.require
*(crash fix only)*
- #42405 Bun.plugin: transpile a copy of typed array onLoad contents
- #42414 node:module: end the Module._stat path at the first null byte
- #42685 resolver: strip the trailing separator inside bust_dir_cache
- #42735 Throw a RangeError instead of aborting when require.resolve
paths or Worker preload outgrows its Vector
- #43067 Report a path that does not fit a path buffer instead of
aborting *(crash fix only)*
- #44410 Fix a crash when code from a node:vm context reads
require.cache or require.extensions
- #44466 Trace the value an object or data module exports from its
JSSourceCode
- #45044 Bun.plugin: give onLoad's args.path its own bytes
- #45117 cjs: do not call an `__esModule` getter during the annotation
lookup

**Dev server, `bun bake`, routers, `--watch` and `--hot`** (28)

- #28793 Stop file watcher thread before Global.exit tears down memory
*(crash fix only)*
- #32143 bake: fix "Runtime file not found" panic in production builds
of custom frameworks
- #32443 watcher: don't free the Watcher in the spawn-to-start window
*(crash fix only)*
- #33232 bake: normalize the HTTP request-target before matching
framework routes
- #36812 bake: respond 500 and release both RequestContext refs when
on_plugins_rejected / OOM drains a deferred framework r…
- #37039 bake: report per-file print failures in the dev server instead
of serving empty modules
- #37276 hot_reloader: take entries_mutex for watcher-thread directory
cache probes
- #37444 bake: unregister Bake::SourceProvider from the source map table
when it is destroyed
- #39290 FileSystemRouter: match / against a root-level optional
catch-all route *(crash fix only)*
- #39291 FileSystemRouter: order dynamic routes by per-segment
precedence like Next.js *(crash fix only)*
- #39629 watch: report a failed watcher init as an error instead of a
crash
- #41185 bake: check for exceptions in the production build's module
loader hooks and helpers
- #41340 bake: skip the directory watch when the directory cannot be
opened
- #41617 Decide a directory watch event from the watchlist, not from the
resolver's cached listing *(crash fix only)*
- #42547 bake: re-arm the memory visualizer timer when it fires
- #42733 dev server: make a client script request wait for the bundle in
flight
- #42828 bake: serve routes from a router root outside the project root
instead of aborting or answering 404
- #42831 bake: resolve app.root and router roots through one checked
join
- #42836 bake: report route scan errors instead of aborting, and fail
bun build --app on them *(crash fix only)*
- #42840 dev server: map a router type's server entry point to the root
route of that type
- #42862 Bun.serve: return before a pending app.plugins setup() settles
*(crash fix only)*
- #42920 Do not create a mimalloc heap per FileSystemRouter, reload() or
:nth-child(An-B) selector
- #42934 dev server: register a client component without evaluating it
- #43329 dev server: check the route bundle index of a client script URL
before Index::init
- #44249 fs.watch: read each directory and close it before descending
- #44344 Windows: fix a watcher panic and lost events in a burst of file
changes, and a --hot that stops watching a deleted…
- #44744 dev server: answer a cached-file lookup only for a graph that
bundled the file
- #44746 dev server: keep an importer in its own graph when an import
fails or is deleted

**`Bun.serve`, sockets, HTTP, WebSocket, HTTP/2, QUIC, DNS** (28)

- #42667 Reland "usockets(windows): keep a closed poll alive until the
outer tick and libuv are done with it" (#39643)
- #33015 fix(windows): segfault in uv__wake_all_loops after system
resume *(crash fix only)*
- #35761 node:quic: detach lsquic ctx pointers before the finalize-order
sweep at worker VM teardown
- #37082 Bun.serve: terminate FIFO/pipe file responses at EOF
- #38024 dns: keep the c-ares uv_poll_t alive while its poll callback is
on the stack (Windows)
- #38636 serve: fix route use-after-free after reload() with
tls.serverName *(crash fix only)*
- #40888 server.fetch: copy a Headers argument instead of adopting the
wrapper's pointer
- #41337 Update lsquic to 4.9.4 and free unsent stateless resets on
destroy *(crash fix only)*
- #41384 usockets: hold a tick level for the whole of close_all
- #41412 Windows: wait for close completions when closing a worker's uv
loop
- #41516 node:http2: handle an empty byte range in respondWithFile and
respondWithFD
- #41521 udp: report every binaryType from the getter, reject a dgram
bind of the other family *(crash fix only)*
- #42188 net: do not assert a two-character host in SocketAddress.parse
- #42210 libuv(windows): keep a pipe's completed read counted until its
read callback returns
- #42334 fetch, Bun.serve: throw a RangeError for Headers that total 4
GiB or more *(crash fix only)*
- #43738 usockets: close a socket whose resume the kernel refuses from
the loop, not from the caller
- #44338 node:http: response trailers travel with the end call of the
response *(crash fix only)*
- #44433 bake: make the parked-request list non-owning and look up the
DevServer when a plugin load settles
- #44533 node:quic: capture the handshake while the conn is alive, set
the application at session creation
- #44734 Bun.serve: answer HEAD for an S3-backed Blob body without an S3
stat
- #44751 node:quic: close() waits only for opened streams, and lsquic
sends no GOAWAY before HTTP/3 starts *(crash fix only)*
- #44897 node:http2: throw instead of panic when the native
setNextStreamID() misses its argument
- #45076 node:http2: let close() and destroy() win over the refusal of a
request
- #45341 uws: AsyncSocket::write takes a size_t length, so res.end()
sends a body of 2 GiB or more whole
- #45522 node:http2: throw for an ALTSVC frame that is too long and for
a pushPromise id that is not a stream id
- #45676 node:quic: do not reset a stream lsquic has already finished
- #45822 node:quic: name a stream in queued session events by a
per-session key, not by its address
- #45858 Bun.serve WebSocket: a message that does not fit in memory
fails one connection, not the server *(crash fix only)*

**Streams, Blob, S3, files, `node:fs`** (28)

- #33845 io: propagate IoRequestLoop init failure instead of aborting
*(crash fix only)*
- #36365 prompt: drop debug_assert that rejects answers containing CR
- #37090 Deliver FIFO bodies whose writer connects after the reader
opens *(crash fix only)*
- #38039 node:fs: stop fs.copyFile(src, "") from segfaulting on Windows
- #38653 shell(cp): reject copying a directory into itself; walk cp
trees without recursion
- #39207 Convert scope-object receivers in the chaining functions and
inspect.custom fallbacks
- #39390 Blob deserialize: reject over-long paths, treat size >=
MAX_SIZE as unknown, make slice() non-panicking
- #39639 fetch: drop the upload and drain hop refs at VM teardown
*(crash fix only)*
- #40008 inspect hooks: copy user options with putDirectMayBeIndex
*(crash fix only)*
- #40462 Windows: FileSink.start({ path }) on a live writer: hand the
writer a libuv fd and close the old source *(crash fix only)*
- #41440 node:fs: pass the buffer to the read/readv/write/writev
callback on error
- #41584 Walk the timer pairing heap without recursion
- #41608 native-readable: accept a stream with no native handle
- #41710 JSSink write(): convert the chunk before resolving the native
sink *(crash fix only)*
- #42711 webstreams: bound the native recursion through a chain of tees
or text streams
- #43264 fetch: reject a file: URL whose path has an encoded NUL
- #43709 s3: fix double free of a multipart part's bytes on worker
terminate
- #43950 streams: link readers and writers last, reject their promises
as handled
- #44036 PipeReader: label a chunk Eof only when the reader is at its
end
- #44181 streams: throw instead of aborting when a direct stream's text
sink cannot take a string chunk
- #44195 fs: validate POSIX directory records in one parser, report a
malformed record as EIO *(crash fix only)*
- #44269 streams: join the string chunks of a text consumer with one
allocation
- #44490 watch(windows): report a failed child spawn as an error instead
of a panic *(crash fix only)*
- #44954 process.stdout/stderr with no FileSink: write synchronously,
reset after a failed write *(crash fix only)*
- #45004 streams: leave a stream alone when a consumer fails to lock it
- #45521 FileSink: throw ERR_ILLEGAL_CONSTRUCTOR from the constructor
that script reaches by writer.constructor
- #45738 s3: own the Blob store across the calls that run user JS
- #45975 node:stream: do not touch the web controller after the web side
is done

**Node-API, V8 API, `bun:ffi`, `process.dlopen`** (15)

- #36609 napi: implement uv_cwd on POSIX
- #37989 bun:ffi: stop cc() from turning non-numeric pointer arguments
into garbage pointers
- #38057 bun:ffi: make cc() reject non-numeric i64/u64 arguments instead
of passing garbage
- #38425 bun:ffi: throw on an empty cc() source array instead of
panicking
- #39652 napi: implement uv_async_t, uv_queue_work and uv_default_loop
on posix
- #39805 Share one handle scope implementation between Node-API and the
V8 API
- #39892 napi: implement uv_idle_t, uv_prepare_t, uv_check_t and
uv_timer_t on posix
- #40144 bun:ffi: return undefined for a NULL napi_value from cc() and
throw the scheduled napi exception
- #40773 bun:ffi: add the full byteOffset in read.*, do not truncate it
to int32
- #41704 napi: hand the thrown value to
napi_get_and_clear_last_exception after napi_run_script
- #42960 process.dlopen: run a V8 addon's init function after dlopen()
returns
- #43797 Fix debug assertion on an empty error message from a V8 addon
or SQLite *(crash fix only)*
- #43802 bun:ffi: throw a TypeError when a cc() symbol is called after
its library's close()
- #43940 napi: fail with ERR_BUFFER_TOO_LARGE instead of aborting above
the ArrayBuffer limit *(crash fix only)*
- #44921 bun:ffi: pad a short call to a cc() symbol with undefined

**`node:vm`, workers, process, timers** (19)

- #34623 node:vm: run DONT_CONTEXTIFY contexts directly against the real
global
- #32839 node:vm: reject invalid cachedData instead of crashing
- #36921 error printer: clear the pending exception when formatting a
non-Error uncaught value throws
- #37270 error printer: one render per error, [Circular] under the key
that closes the cycle
- #38740 bun:test: give each VM its own fake clock *(crash fix only)*
- #39864 Report a refused HTTP, bundle, IO or process waiter thread as
an error instead of a crash *(crash fix only)*
- #39992 process: deliver exceptions reported against a node:vm context
to uncaughtException
- #41654 node:vm: give each SourceTextModule request the import
attributes of the declaration that created it *(crash fix only)*
- #41739 Do not forward the inherited IPC channel to spawned scripts and
tools
- #41769 node:vm: stop compiling accepted cachedData, report each
cachedData decode to the GC *(crash fix only)*
- #42045 bun:test fake timers: never move the fake clock backwards
- #42308 process.execve: throw instead of aborting for a string past the
string limits
- #42734 node:vm: report cachedData of 2 GiB or more as rejected instead
of aborting
- #42843 process.dlopen: reject a filename of PATH_MAX or more before
dlopen()
- #43500 process: accept SIGSTKFLT, SIGPOLL and SIGPWR, and emit every
name of a signal number *(crash fix only)*
- #44486 worker: stop the start sequence when a handler it calls stops
the worker
- #44771 timers: allocate ids from one 64-bit counter that never wraps
- #44837 --hot: defer a reload during the preloads, and report a reload
that cannot start
- #44991 Fix a use-after-free in the VM ref-string intern for a string
over the length limit

**Shell, spawn, CLI, `bun test`, `bun install`** (31)

- #30642 Fail fast at startup when simdutf has no usable implementation
for this CPU
- #34223 Fix four crashes: HTMLRewriter disturbed body, S3 locked-stream
upload, ConnectionsList closed parser, shell brace… *(crash fix only)*
- #34884 Guard the console and test-runner formatters against native
stack overflow on deeply nested values *(crash fix only)*
- #36584 process.chdir: publish an immutable cwd slice so no thread sees
a torn cwd
- #37135 bun test --parallel: hold the JSC API lock while draining the
worker's final IPC backlog
- #37527 shell(mv): report ENAMETOOLONG instead of panicking when a
source name does not fit the path buffers
- #37528 cli: stop aborting on absolute script paths longer than 1024
bytes on Windows
- #38162 shell(cp): report ENAMETOOLONG instead of panicking on operands
longer than the path buffers
- #38205 shell(windows): report ENAMETOOLONG for operands that do not
fit the path buffer
- #39010 bun test: exit through the VM teardown path when --bail stops a
run
- #39104 bun test: report over-long coverage dir, JUnit outfile and
bunfig root instead of panicking *(crash fix only)*
- #39638 shell(cat): serve a second cat on the same stdin instead of
hanging or panicking
- #39641 usockets: report an out-of-descriptors event loop init as an
error, not a crash *(crash fix only)*
- #40181 shell: stop deeply nested if clauses from overflowing the
parser stack
- #40885 shell: throw instead of panicking when a ReadableStream is used
as a redirect
- #41155 ipc: send a queue item larger than 2 GiB in capped write
requests
- #42298 bun:test: print String objects with 16-bit storage correctly
- #42577 bun:test: root the return values toHaveReturnedWith collects
- #43606 shell(windows): report a failed pipe reader start through
on_reader_error instead of leaking the subprocess *(crash fix only)*
- #43838 node:cluster: read a worker's `ack` and `addressType` in the
primary only when they are int32
- #44172 node:cluster: refuse a descriptor of the wrong kind in the
primary *(crash fix only)*
- #44185 node:cluster: one handle for each descriptor that workers name
*(crash fix only)*
- #44491 bun test: run every callback when native code leaves an
exception pending *(crash fix only)*
- #44694 shell mv: refuse to copy a directory into itself on the
cross-device fallback
- #44752 shell: take the command-substitution capture and drop its NUL
bytes
- #44981 shell: reject NUL bytes in .cwd() and .env() at the native
setters
- #45206 ini: reject a source of 2 GiB or more in Parser::parse instead
of aborting on an int cast *(crash fix only)*
- #45348 bun test: fix a segfault in .not.toThrow(Class) when
Symbol.hasInstance accepts a thrown string, symbol or BigInt
- #45649 shell: replace the output counters in mkdir, cp, ls and touch
with the OutputTask queue
- #45792 bun test: SIGKILL a spawnSync child that is still running 500
ms after the test's timeout *(crash fix only)*
- #45910 create: own the remote-template response buffers instead of
parking them in the CLI arena

**Web APIs, Buffer, encoding, URL, crypto, zlib, SQLite, Redis and the
rest** (96)

- #29709 console: guard JSX/Proxy recursion in Bun.inspect (stack
overflow SIGSEGV)
- #30817 console.takeHeapSnapshot: clear prior snapshots before building
- #30956 spawnSync: make signal-forwarding register/unregister safe for
concurrent callers *(crash fix only)*
- #31505 node:zlib: reject invalid brotli flush values at validation
*(crash fix only)*
- #31735 node:path: fix win32 drive-relative resolve on non-Windows
hosts *(crash fix only)*
- #32085 Fix crash when RedisClient onclose/onconnect is set to a
non-callable
- #32240 Fix BroadcastChannel custom inspect crash with numeric option
keys
- #33181 Reject interior null bytes in Bun.mmap, Glob.scan, and
FileSystemRouter paths *(crash fix only)*
- #33752 node:crypto: validate EC JWK private scalar matches public
point in createPrivateKey
- #34976 spyOn: install and restore through one writer that throws when
the object refuses *(crash fix only)*
- #35562 structuredClone: purify NaN in deserialized Date timestamp so
getTime never returns a forged JSValue
- #35618 napi(windows): load addons that import node.exe without
win_delay_load_hook
- #35757 Coerce later arguments before capturing input buffers in
Bun.Transpiler and Bun.RedisClient
- #37252 Report UTF-16 to UTF-8 encode counts through size_t so exactly
2^32 written bytes doesn't wrap to 0
- #37338 node:module: build require.cache without user-writable globals
and propagate its exceptions *(crash fix only)*
- #37340 redis: accept unsubscribe acks that arrive outside subscriber
mode
- #37370 Error.appendStackTrace: make appending an error's trace to
itself a no-op
- #37443 node:crypto: fix null deref when worker.terminate() lands
during an EC JWK private key import
- #37448 jsc: check for exceptions before using a stored
constructEmptyArray result *(crash fix only)*
- #38009 MessagePort: fix a heap-use-after-free in contextDestroyed(),
release ref() when the last 'message' listener is rem… *(crash fix
only)*
- #38469 napi: settle or refuse async work queued during VM teardown
instead of taking a ticket
- #38708 Handle 16-bit module sources in the error preview, the
new-position fix-up and the REPL *(crash fix only)*
- #38821 process.env: make a failed env build throw instead of aborting
(Windows 0xC0000409 on Bun.$ / Bun.sql / process.env…
- #38871 process.env: coerce to string on every execution of an
assignment, not just the first
- #38877 cli: exit 128 + signo when bun cannot die from a re-raised
signal (bun as container PID 1)
- #38926 Clear the inherited signal mask at startup and exec children
with an empty one *(crash fix only)*
- #39382 Make a failed util.inspect load throw from custom inspect
functions instead of aborting
- #39481 Always create the error when a native error message exceeds the
string length limit
- #39651 performance: keep the user timing entry counts exact after
clearMarks(name) and clearMeasures(name)
- #39775 Turn seccomp SIGSYS traps into ENOSYS so the syscall fallbacks
run on Android
- #39968 process.dlopen: replay a non-context-aware V8 addon only while
no other live VM owns it
- #39975 watch: report a failed reload exec as an error instead of a
panic *(crash fix only)*
- #40346 Give Glob's native scan methods their own private names
- #40554 node:crypto: copy password and salt for async pbkdf2 and scrypt
- #40567 url: reject URL-encoded input over the WTF decode limit instead
of aborting
- #40577 url: throw a RangeError instead of aborting when
URLSearchParams or FormData outgrows its Vector
- #41073 Store array-index property names as indexed properties in
native object builders
- #41137 Write the sampling profiler report at exit from Bun, not JSC's
atexit hook
- #41365 webview: fail only the view when WebKit lacks
initWithDirectory:
- #41574 Decode a private copy of a JS buffer in the UTF-8 decoders
*(crash fix only)*
- #41652 console: read a BigInt depth option by value in console.dir
- #41749 Startup: reopen fd 0-2 as /dev/null only when the fd is closed
- #41819 Report an incoherent BUN_JSC_* option combination as a
configuration error, not a crash *(crash fix only)*
- #42158 webview: reject malformed Chrome replies instead of crashing
- #42179 shell: clamp a `> ${buf}` redirect to the target buffer's live
length
- #42187 node:crypto: throw ERR_OUT_OF_RANGE for >2 GiB buffer arguments
instead of aborting
- #42191 URLPattern: throw instead of aborting when a pattern string
passes the string limit
- #42192 Copy WebAssembly.Memory bytes for native borrows that outlive
the call *(crash fix only)*
- #42198 Make Bun.$, Bun.sql, Bun.postgres and Bun.SQL custom value
getters so reifying the Bun object passes exception chec…
- #42203 node:fs: fs.read into a WebAssembly.Memory view writes into the
block a grow freed
- #42214 Headers: throw a RangeError instead of aborting when the
set-cookie list cannot grow
- #42218 Headers: throw instead of aborting when a value passes the
string length limit
- #42220 URLPattern: throw instead of aborting when the token list
cannot grow
- #42225 Throw ERR_STRING_TOO_LONG from Bun.wrapAnsi instead of aborting
when the output passes the string length limit
- #42226 Throw instead of aborting when a MIME string does not fit in a
string *(crash fix only)*
- #42237 Bun.Cookie: throw instead of aborting when the Set-Cookie
string passes the string length limit
- #42249 Copy a shared input before libdeflate compresses it in
Bun.gzipSync and Bun.deflateSync
- #42253 bun test: walk nested describe scopes without native recursion
- #42259 Throw instead of aborting when a domainToASCII host passes the
string length limit
- #42310 Parse a private copy of a shared or resizable buffer in
Bun.Transpiler, Bun.markdown, Bun.YAML and Bun.TOML *(crash fix only)*
- #42312 YAML, JSON5, XML stringify: throw instead of aborting when the
output passes the string length limit
- #42318 Bound libdeflate's emit path so a racing input cannot overrun
the output
- #42338 inspector: check the type of process.argv in getModuleGraph
- #42364 node:perf_hooks: do not truncate the monitorEventLoopDelay
resolution to int32
- #42412 transpiler cache: reject an entry with no sourcemap section
- #42479 worker_threads: do not abort when
markAsUntransferable/markAsUncloneable gets a WebAssembly GC reference
- #42517 EventTarget: skip a listener object that has no realm
- #42518 fs.Stats: make the date accessors define the property like
Object.defineProperty *(crash fix only)*
- #42575 Stop natives from writing onto a foreign `this` with putDirect:
process.ppid, require.cache, StringDecoder, the nat… *(crash fix only)*
- #42631 JSStringView: pin the cell that owns the borrowed characters
- #42660 bun_core: read /dev/urandom when getrandom(2) answers ENOSYS or
EPERM
- #42780 node:crypto: treat a name too long to convert to UTF-8 as
unknown instead of aborting
- #42782 node:crypto: throw ERR_OUT_OF_RANGE for a passphrase larger
than INT_MAX bytes
- #42856 url: do not abort when a host is too long for the IDNA
conversion buffers
- #42868 sqlite, console, process: do not abort on a string too long to
convert to UTF-8 *(crash fix only)*
- #42941 Bun.sliceAnsi: throw instead of aborting when the result passes
the string length limit
- #43058 Bun.serve: fix crash and cross-request 500 when inspecting a
non-Response return value throws
- #43904 JSC bindings: build property keys from a BunString through
Bun::toIdentifier
- #44124 Bun.Image: decode WebP into a Bun-owned canvas and read ICCP
without the demuxer *(crash fix only)*
- #44190 Convert scope-object receivers before Buffer read*/write* and
vm.Script describe them
- #44211 Stop getsockname(out), useFakeTimers() and the stream source
pull() from writing onto a caller's object with putDirect
- #44295 bun test --coverage: append a final newline to a text that
lacks one *(crash fix only)*
- #44360 rust-argon2: return OutOfMemory instead of aborting when the
memory cost cannot be allocated
- #44421 node:zlib: free zstd's workers before the dictionary they read
- #44521 util.parseArgs: pick the default args start from a per-VM eval
flag, as Node v26 does *(crash fix only)*
- #44740 structured clone: tolerate a zero-sized shared
WebAssembly.Memory in computeMemoryCost()
- #44759 HTMLRewriter: park a suspended handler's token before the
output flush
- #44907 Builtins section format 2: an engine stamp says which buns
compile its modules
- #45188 Reject an oversized WebCrypto cipher output instead of aborting
- #45354 Throw for a `file:` URL whose path is not UTF-8 on Linux and
macOS, through one `fileSystemPath()` accessor *(crash fix only)*
- #45608 ipc: refuse an over-long message before the receive buffer
stores it
- #45624 webcrypto: write the key type into the X25519 structured-clone
record
- #45625 node:sqlite: refuse session.close() and Symbol.dispose inside a
SQLite callback
- #45653 glob: a group with `?` or a class matches when an alternative
takes 64 bytes or more
- #45820 Bun.file, S3Client.file and Bun.write no longer keep a buffer
path alive forever
- #45832 `bun check`: fix a panic on `export *` without a module
specifier, and a false error for one in parentheses

### Already fixed on `main`, or duplicates of a pull request above

Already fixed: #27323 #30401 #31011 #31206 #31566 #31626 #31695 #31702
#31945 #32773 #34154 #34697 #34744 #34964 #35655 #35816 #36020 #36181
#36296 #36342 #36581 #36602 #36807 #36818 #36857 #37172 #37235 #37239
#37267 #37420 #37731 #37914 #38113 #38119 #38509 #38514 #38540 #38975
#39410 #40749 #41163 #41171 #42108 #43085 #43230

Duplicates: #45654 #29986 #33378 #36895 #37070 #37639 #38014 #39021
#41266 #42250 #42951 #43238 #44259 #45335 #45565 #45726 #45965

Fixes #5019
Fixes #6106
Fixes #10690
Fixes #10886
Fixes #13566
Fixes #14745
Fixes #15551
Fixes #20454
Fixes #21310
Fixes #29145
Fixes #30454
Fixes #30613
Fixes #30766
Fixes #32103
Fixes #32142
Fixes #32183
Fixes #32212
Fixes #39060
Fixes #42296
Fixes #42931
Fixes #43671
Fixes #44201
Fixes #44735
Fixes #44783
Fixes #45198
Fixes #45327
Fixes #45491
Fixes #45721

---------

Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
cevheri added a commit to libredb/libredb-studio that referenced this pull request Oct 28, 2026
…ts, which crashes Bun 1.4.2 on Windows

bun:test's expect(promise).rejects waits for a pending promise by running
the event loop inside the caller. A case that starts right after an await
settled in a socket callback of the fake API runs that nested loop inside
the socket's dispatch, and Bun 1.4.2's Windows backend frees the socket
under it: a segfault in us_internal_ssl_close or
us_internal_dispatch_ready_poll, 7 of 20 runs of this file on
windows-latest (oven-sh/bun#39643, relanded in Bun 1.4.4 by #42667).

Every case now awaits the promise and asserts on what it rejected with.
"the harness" shows the nested loop and holds the file to it.
cevheri added a commit to libredb/libredb-studio that referenced this pull request Oct 28, 2026
…ts, which crashes Bun 1.4.2 on Windows

bun:test's expect(promise).rejects waits for a pending promise by running
the event loop inside the caller. A case that starts right after an await
settled in a socket callback of the fake API runs that nested loop inside
the socket's dispatch, and Bun 1.4.2's Windows backend frees the socket
under it: a segfault in us_internal_ssl_close or
us_internal_dispatch_ready_poll, 7 of 20 runs of this file on
windows-latest (oven-sh/bun#39643, relanded in Bun 1.4.4 by #42667).

Every case now awaits the promise and asserts on what it rejected with.
"the harness" shows the nested loop and holds the file to it.
cevheri added a commit to libredb/libredb-studio that referenced this pull request Oct 28, 2026
…ts, which crashes Bun 1.4.2 on Windows

bun:test's expect(promise).rejects waits for a pending promise by running
the event loop inside the caller. A case that starts right after an await
settled in a socket callback of the fake API runs that nested loop inside
the socket's dispatch, and Bun 1.4.2's Windows backend frees the socket
under it: a segfault in us_internal_ssl_close or
us_internal_dispatch_ready_poll, 7 of 20 runs of this file on
windows-latest (oven-sh/bun#39643, relanded in Bun 1.4.4 by #42667).

Every case now awaits the promise and asserts on what it rejected with.
"the harness" shows the nested loop and holds the file to it.
cevheri added a commit to libredb/libredb-studio that referenced this pull request Oct 28, 2026
…ts, which crashes Bun 1.4.2 on Windows

bun:test's expect(promise).rejects waits for a pending promise by running
the event loop inside the caller. A case that starts right after an await
settled in a socket callback of the fake API runs that nested loop inside
the socket's dispatch, and Bun 1.4.2's Windows backend frees the socket
under it: a segfault in us_internal_ssl_close or
us_internal_dispatch_ready_poll, 7 of 20 runs of this file on
windows-latest (oven-sh/bun#39643, relanded in Bun 1.4.4 by #42667).

Every case now awaits the promise and asserts on what it rejected with.
"the harness" shows the nested loop and holds the file to it.
cevheri added a commit to libredb/libredb-studio that referenced this pull request Oct 28, 2026
…ts, which crashes Bun 1.4.2 on Windows (#2029)

bun:test's expect(promise).rejects waits for a pending promise by running
the event loop inside the caller. A case that starts right after an await
settled in a socket callback of the fake API runs that nested loop inside
the socket's dispatch, and Bun 1.4.2's Windows backend frees the socket
under it: a segfault in us_internal_ssl_close or
us_internal_dispatch_ready_poll, 7 of 20 runs of this file on
windows-latest (oven-sh/bun#39643, relanded in Bun 1.4.4 by #42667).

Every case now awaits the promise and asserts on what it rejected with.
"the harness" shows the nested loop and holds the file to it.
cevheri added a commit to libredb/libredb-studio that referenced this pull request Oct 28, 2026
Bun 1.4.2 on Windows segfaults when bun:test's expect(promise).rejects
runs a nested event loop inside a socket dispatch (oven-sh/bun#39643,
relanded in 1.4.4 by oven-sh/bun#42667). #2029 worked around it in one
test file; moving the pin removes the cause for every file.

The packageManager pin, every setup-bun bun-version and npm install of
bun in the workflows, the three Dockerfile deps stages, the dev
container and the contributor docs move together. Notes that record a
measurement taken on 1.4.2 keep that version.
cevheri added a commit to libredb/libredb-studio that referenced this pull request Oct 28, 2026
* chore(toolchain): pin Bun 1.4.4

Bun 1.4.2 on Windows segfaults when bun:test's expect(promise).rejects
runs a nested event loop inside a socket dispatch (oven-sh/bun#39643,
relanded in 1.4.4 by oven-sh/bun#42667). #2029 worked around it in one
test file; moving the pin removes the cause for every file.

The packageManager pin, every setup-bun bun-version and npm install of
bun in the workflows, the three Dockerfile deps stages, the dev
container and the contributor docs move together. Notes that record a
measurement taken on 1.4.2 keep that version.

* chore(rabbitmq): run the fixture seed on Bun 1.4.4

The seed one-shot runs on the repository's own Bun, so its image follows the packageManager pin. The digest is the oven/bun:1.4.4 index digest from Docker Hub.

* test(rabbitmq): let the server finish its handshake before the dial test destroys the socket

A TLS client destroyed in the same tick as its secureConnect resets the connection before the server reads the client's Finished, so the server never completes its side. Node does that, and Bun does from 1.4.4, so the test waits for the server's side of each handshake.

* test(sqlite): bun:sqlite answers declaredTypes before the rows from Bun 1.4.4

Bun 1.4.2 threw until the statement had run; 1.4.4 answers before and after with the same answer. The bridge still reads it after the rows, so an older Bun keeps working. The provider doc records the re-measure on the new pin.

* test(http): a TLS record corrupted after the headers is a TLS failure under Bun 1.4.4 too

Bun 1.4.4 emits the TLS layer's code before the reset, as Node does, so the Node-only expectation becomes the expectation for both runtimes.

* test(prometheus): follow Bun 1.4.4's fetch errors

Bun 1.4.4's fetch reports a refused port as ECONNREFUSED, where 1.4.2 said ConnectionRefused, and no longer quotes a refused header value in its message. The transport already read both codes and never copied the message, so only the controls and the wording change.

* fix(etcd): name a client-certificate alert in BoringSSL's words

From Bun 1.4.4 grpc-js carries the server's alert as TLSV1_ALERT_CERTIFICATE_REQUIRED or TLSV1_ALERT_UNKNOWN_CA, where 1.4.2 said only Failed to connect. The TLS cause table matched OpenSSL's spelling only, so under Bun the refusal read as a failure with no cause named, beside the cause itself.

* fix(spark): name a client-certificate alert in BoringSSL's words

Bun 1.4.4 reports the TLS proxy's alert through grpc-js in BoringSSL's capitals; the classifier now reads them as Node's, so a missing or untrusted client certificate gets its own sentence under Bun too.

* fix(milvus): name a client-certificate alert in BoringSSL's words

Bun 1.4.4 reports the server's alert through grpc-js in BoringSSL's capitals; the TLS cause table now reads them as Node's.

* fix(oxia): name a client-certificate alert in BoringSSL's words

The same TLS cause table as etcd and Milvus, for the same Bun 1.4.4 texts.

* docs(providers): say Bun names TLS alerts from 1.4.4, and pin the lazy declaredTypes read

The Spark error table, and the etcd and Milvus runtime flag comments, still said Bun reports no TLS alert; from 1.4.4 it does, and the flag only words a TLS failure that names no cause.
A stand-in bun statement that refuses declaredTypes until it has run now keeps the bun adapter reading it after the rows, as a Bun before 1.4.4 needs.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

5 participants