Repository navigation
TransformStream: survive controller API calls after a native sink tears down the readable - #33781
Conversation
…rs down the readable When a Bun native sink (Bun.serve response body, Bun.spawn stdin) consumes a TransformStream's readable side and the consumer aborts, the sink pump's finally step (rsisFinally / resumableReleaseReader / readDirectStreamCloseImpl) runs clearStreamControllerSlots() on the readable, setting its controllerKind to None and clearing m_controller. The TransformStream still holds that readable via m_readable, so a later controller.error()/enqueue()/terminate() or desiredSize access (from an async transform()/flush() that resumed after the abort, or from a queued sink reaction) walks transformReadableController() which asserts ControllerKind::Default and then dereferences a null controller. On a release build this is a near-null deref (segfault at 0x88); on an assert-enabled build it is a SIGABRT at TransformStreamOperations.cpp:34. A remote client aborting an upload mid-stream is enough to trigger it. Make both transformReadableController() copies null-safe (return nullptr when the readable's controller slot has been cleared) and make every call site skip the readable-side AO on null. The spec AOs would no-op on a terminal readable anyway; the writable-side error/unblock still runs. This mirrors the teeBranch*Controller null-safety added for the tee() sibling of this bug.
|
Warning Review limit reached
Next review available in: 24 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: ASSERTIVE Plan: Pro Run ID: 📒 Files selected for processing (3)
Comment |
|
Found 1 issue this PR may fix:
🤖 Generated with Claude Code |
|
Not linking #31004: that issue is on Bun 1.3.13/1.3.14, which predates the C++ webstreams rewrite (#33193) this PR touches. The |
There was a problem hiding this comment.
I didn't find any bugs — the null-check pattern is applied consistently across all 8 call sites and mirrors #33778 — but this is a crash fix in the C++ WebStreams JSC bindings, so it's worth a human confirming the spec-AO no-op reasoning and exception-scope placement.
Extended reasoning...
Overview
The PR makes transformReadableController() (duplicated in JSTransformStreamDefaultController.cpp and TransformStreamOperations.cpp) return nullptr instead of asserting when the readable's controller slot has been cleared by a Bun native-sink teardown (clearStreamControllerSlots() in BunStreamSource.cpp). All eight call sites are updated to skip the readable-side abstract operation on null while still running the writable-side error/unblock. A subprocess-based regression test exercises desiredSize, enqueue, terminate, and error after Bun.spawn stdin teardown.
Security risks
None identified. This is defensive null-checking on an internal path; no user-controlled input validation, auth, or crypto is involved. The pre-fix behavior was a segfault (DoS), so the change strictly reduces attack surface.
Level of scrutiny
High. This is memory-safety-critical C++ in the JSC WebStreams bindings — a layer where exception-scope discipline, GC lifetime, and spec-AO semantics all interact. The change removes an ASSERT that encoded a (now-false) invariant, and moves RETURN_IF_EXCEPTION calls inside conditionals. While each edit is mechanically simple and the pattern mirrors the recently-landed tee-branch fix in #33778, the correctness depends on the claim that the spec AOs would no-op on a non-Readable stream anyway — which a human familiar with the streams implementation should confirm.
Other factors
- Grep confirms all
transformReadableControllercallers are covered; none were missed. - The test asserts exact per-action outcomes (not just "didn't crash") and checks
signalCode: null, which is the right hang-guard pattern. - WPT streams (1175 pass) and related suites reported green in the description.
- The removed ASSERT previously also checked
readablefor null; the new code dereferencesreadable->m_controllerKindunconditionally. That's fine becausem_readableis set ininitializeTransformStreamand never cleared, but it's the kind of detail a maintainer should glance at.
|
On the two points from the review:
Spec-AO no-op reasoning: the three AOs called through the now-nullable pointer all gate on the readable's state as their first step:
All three Exception-scope placement: each |
…42609) ### Problem - `Bun.serve()` sends a second `Response` around an already sent `ReadableStream` as a `200` with an empty body. The `error` handler does not run. - A native sink consumer (`Bun.serve()`, a `fetch()` upload, `Bun.write()`, `Bun.spawn()` stdin, S3, `HTMLRewriter`) pumps a JS-backed stream in `BunStreamSource.cpp`. At the end `rsisFinally` (`:853`) releases the reader and `directStreamOnClose` (`:662`) drops the lock: `locked === false`. A `type: "direct"` stream, so every async iterable body, is not disturbed either. `RequestContext.rs:3152` checks the lock alone. - Found by a comparison with node v26.3.0, not by a user report. ### Fix - `rsisBegin` and `readDirectStream` call the new `JSReadableStream::markConsumedAsBody()` once the pump holds the stream. The stream stays locked after the pump lets go. - Behaviour change: after such a consumer took a stream, `getReader()`, `tee()`, `pipeTo()` and `cancel()` fail with a `TypeError`. Body mixin methods (#42116) and natively wired sinks (#42477) already do this. Neither is released. - `pipeTo()`, async iteration and `Bun.readableStreamTo*()` do not reach these pumps and still unlock. - Verified: `web-stream-state.test.ts` (30 new tests fail on main), `body.test.ts` (12), `serve-reused-response.test.ts` (2). Self-reviewed: 3 concerns, 2 addressed, 1 out of scope (Notes). ### Background - A JSSink is a native sink (`HTTPResponseSink`, `FileSink`). `JSSink::assign_to_stream` pumps a stream into it. - `readStreamIntoSink` reads through a default reader. `readDirectStream` gives the sink to the `pull()` of a direct stream. - `m_consumedAsBody` (#42116) is a bit that `isReadableStreamLocked()` includes and nothing clears. The fetch spec never releases the reader of a body. <details><summary>Notes</summary> **The rule.** A consumer that takes a stream as a body keeps it locked: the body mixin methods (#42116), `to_any_blob` lifts (#42516), natively wired `ByteStream`/`FileReader` sinks (#42477), and now the two pumps. A reader at the stream level releases its lock as the Streams spec says. Checked on this branch: `pipeTo()`, `pipeThrough()`, `for await`, `getReader()` + `releaseLock()`, `Bun.readableStreamToText()`, `Bun.readableStreamToArrayBuffer()`, `stream.text()`, `stream.bytes()`, `stream.json()`, `stream.blob()` all end with `locked === false`. **Scope of the pumps.** `readStreamIntoSink` and `readDirectStream` have one caller, `assignToStream`, which only `JSSinkController__assignToStream` calls (Rust `JSSink::assign_to_stream`: `HTTPResponseSink` and its TLS and HTTP/3 siblings, `FetchRequestBodySink`, `NetworkSink`, `FileSink`, `RewriterPipe`). **State on main.** The lock state after the hand-off depends on how the pump ended. A clean end and a sink that closes early (client gone, aborted upload, child exited) go through `rsisFinish` and release the reader one microtask after the stream closes. A producer error goes through `rsisAbrupt`, which orphans the reader, so that stream stays locked. A direct stream is never disturbed. **Repro (the user-visible part).** ```js const stream = new ReadableStream({ start(c) { c.enqueue(new TextEncoder().encode("payload")); c.close(); } }); const responses = [new Response(stream), new Response(stream)]; await using server = Bun.serve({ port: 0, fetch: () => responses.shift(), error: e => new Response(e.code, { status: 500 }) }); for (let i = 0; i < 2; i++) { const r = await fetch(server.url); console.log(r.status, JSON.stringify(await r.text())); } // main: 200 "payload", 200 "" // branch: 200 "payload", 500 "ERR_STREAM_CANNOT_PIPE" (Stream already used, please create a new one) ``` For a direct stream `new Response(stream)` per request shows the same on main: `200 "hello"`, `200 ""`, `500`. On this branch the second `new Response(stream)` throws `Body object should not be disturbed or locked`. **Other observable changes.** `request.bodyUsed` after `fetch(request)` with an async iterable body is now `true` (was `false`, the stream was not disturbed). `Bun.readableStreamToText(stream)` on a stream a sink holds rejects with "ReadableStream has already been used" in place of "ReadableStream is locked" (same `ERR_INVALID_STATE`). **Bare streams.** `fetch(url, { body: stream })`, `Bun.write(path, stream)` and `Bun.spawn({ stdin: stream })` keep a bare stream locked too. Natively wired sources already do this for the same calls (`ReadableStream__lockNative` never unlocks, see "errors a Bun.file() stream whose file does not open" from #42477). **node v26.3.0.** `fetch(url, { method: "POST", body: jsStream, duplex: "half" })`: afterwards `locked === true` and `getReader()` throws. The same after an upload that an `AbortSignal` stopped. The other five consumers are Bun APIs with no Node counterpart. **The mark comes after the reader acquisition.** A stream that another reader already holds is not marked. `Bun.serve()`, `fetch()`, `Bun.write()` and `HTMLRewriter` reject such a stream before the pump. `Bun.spawn()` stdin checks only `is_disturbed` (`stdio.rs:388`, `:536`), so a locked stream reaches the pump there. **Out of scope (the concern not addressed).** `rsisFinally` calls `clearStreamControllerSlots` also when the pump never got the lock. `Bun.spawn({ stdin: stream })` with a stream the caller holds through `getReader()` reaches that: the caller's `reader.read()` then never settles. This is the same on main and on this branch. #41532 rejects a locked stdin stream before the pump. **Not covered here.** - A handler that drains or partly reads a stream with its own reader, releases it, and then returns it in a `Response` still gets a `200` with the empty or remaining body. #36110 covers that gate. - Handing the same Node `Readable` or generator object to a second consumer makes a new stream each time. **Changed tests.** Three tests in `bun-write.test.js` (from #42114) read the terminal state through `stream.getReader().closed`. One test in `streams.test.js` (from #33781) polled `ts.readable.locked` to wait for the sink teardown. Both observed the unlock as a means, not as the subject. They use `finished()` from `node:stream/promises` now. The `streams.test.js` test still proves the teardown ran: `desiredSize` reads `null` only after the controller slot is cleared, and reads `0` for a stream that is only closed. **Fail before.** With `src/` from main and `bun bd`: `web-stream-state.test.ts` 30 of 46 fail, `body.test.ts` 12 of 778 fail, `serve-reused-response.test.ts` 2 of 9 fail, and the three `bun-write.test.js` tests fail on the new `locked` assertion. With this branch all pass. On main only the `Bun.serve()` row of "the stdout of a running child" fails: the other four consumers wire a pipe-backed `FileReader` natively. **Built-in JS.** I searched `src/js` for code that touches a stream after a native sink took it. The only `stream.cancel()` on a web stream is `ReadableFromWeb._destroy`, which #42573 handles for the body mixin case. **Earlier reports of the same class.** #7001 (fixed in #7861) and #6860. **Suites run on the debug build.** `test/js/web/streams/{streams,streams-leak,readable-stream-blob-consumed,readable-stream-terminal-barrier-release,transform-stream-leak}`, `test/js/third_party/wpt-streams`, `test/js/web/fetch/{body,body-stream,body-stream-excess,body-clone,body-async-iterator,body-mixin-errors,blob-write,fetch,fetch.stream,fetch-backpressure,fetch-abort-stream-body,fetch-stream-cancel-leak,fetch-redirect,response}`, `test/js/web/request/request`, `test/js/bun/http/{serve,bun-server,serve-reused-response,serve-direct-readable-stream,serve-body-leak,serve-pending-promise-abort-leak,async-iterator-stream,serve-async-stream-client-abort,serve-error-handler-stream,serve-response-stream-sink-leak,serve-stream-body-error,serve-stream-reject-flush-leak}`, `test/js/bun/spawn/{spawn,spawn-stdin-readable-stream}`, `test/js/bun/io/bun-write`, `test/js/workerd/{html-rewriter,html-rewriter-leak}`, `test/js/bun/s3/{s3-stream-cancel-leak,s3-stream-error-gc,s3-upload-stream-gc,s3-write-to-file-sync-close,s3-connection-close}`, `test/js/node/stream/{node-stream,web-stream-state}`, `test/js/node/async_hooks/AsyncLocalStorage`, `test/regression/issue/07001`. The failures that remain also fail with `src/` from main in this container: IPv6, tests that need a non-root user, external hosts, and 5 s timeouts of the debug build. </details> <!-- robobun:evidence:begin --> --- **no test proof** · iteration 0 · platform-specific test(s) that do not run on this machine, deferring to CI, which covers all platforms: test/js/web/streams/streams.test.js, test/js/bun/io/bun-write.test.js <!-- robobun:evidence:end -->
…ven-sh#42609) ### Problem - `Bun.serve()` sends a second `Response` around an already sent `ReadableStream` as a `200` with an empty body. The `error` handler does not run. - A native sink consumer (`Bun.serve()`, a `fetch()` upload, `Bun.write()`, `Bun.spawn()` stdin, S3, `HTMLRewriter`) pumps a JS-backed stream in `BunStreamSource.cpp`. At the end `rsisFinally` (`:853`) releases the reader and `directStreamOnClose` (`:662`) drops the lock: `locked === false`. A `type: "direct"` stream, so every async iterable body, is not disturbed either. `RequestContext.rs:3152` checks the lock alone. - Found by a comparison with node v26.3.0, not by a user report. ### Fix - `rsisBegin` and `readDirectStream` call the new `JSReadableStream::markConsumedAsBody()` once the pump holds the stream. The stream stays locked after the pump lets go. - Behaviour change: after such a consumer took a stream, `getReader()`, `tee()`, `pipeTo()` and `cancel()` fail with a `TypeError`. Body mixin methods (oven-sh#42116) and natively wired sinks (oven-sh#42477) already do this. Neither is released. - `pipeTo()`, async iteration and `Bun.readableStreamTo*()` do not reach these pumps and still unlock. - Verified: `web-stream-state.test.ts` (30 new tests fail on main), `body.test.ts` (12), `serve-reused-response.test.ts` (2). Self-reviewed: 3 concerns, 2 addressed, 1 out of scope (Notes). ### Background - A JSSink is a native sink (`HTTPResponseSink`, `FileSink`). `JSSink::assign_to_stream` pumps a stream into it. - `readStreamIntoSink` reads through a default reader. `readDirectStream` gives the sink to the `pull()` of a direct stream. - `m_consumedAsBody` (oven-sh#42116) is a bit that `isReadableStreamLocked()` includes and nothing clears. The fetch spec never releases the reader of a body. <details><summary>Notes</summary> **The rule.** A consumer that takes a stream as a body keeps it locked: the body mixin methods (oven-sh#42116), `to_any_blob` lifts (oven-sh#42516), natively wired `ByteStream`/`FileReader` sinks (oven-sh#42477), and now the two pumps. A reader at the stream level releases its lock as the Streams spec says. Checked on this branch: `pipeTo()`, `pipeThrough()`, `for await`, `getReader()` + `releaseLock()`, `Bun.readableStreamToText()`, `Bun.readableStreamToArrayBuffer()`, `stream.text()`, `stream.bytes()`, `stream.json()`, `stream.blob()` all end with `locked === false`. **Scope of the pumps.** `readStreamIntoSink` and `readDirectStream` have one caller, `assignToStream`, which only `JSSinkController__assignToStream` calls (Rust `JSSink::assign_to_stream`: `HTTPResponseSink` and its TLS and HTTP/3 siblings, `FetchRequestBodySink`, `NetworkSink`, `FileSink`, `RewriterPipe`). **State on main.** The lock state after the hand-off depends on how the pump ended. A clean end and a sink that closes early (client gone, aborted upload, child exited) go through `rsisFinish` and release the reader one microtask after the stream closes. A producer error goes through `rsisAbrupt`, which orphans the reader, so that stream stays locked. A direct stream is never disturbed. **Repro (the user-visible part).** ```js const stream = new ReadableStream({ start(c) { c.enqueue(new TextEncoder().encode("payload")); c.close(); } }); const responses = [new Response(stream), new Response(stream)]; await using server = Bun.serve({ port: 0, fetch: () => responses.shift(), error: e => new Response(e.code, { status: 500 }) }); for (let i = 0; i < 2; i++) { const r = await fetch(server.url); console.log(r.status, JSON.stringify(await r.text())); } // main: 200 "payload", 200 "" // branch: 200 "payload", 500 "ERR_STREAM_CANNOT_PIPE" (Stream already used, please create a new one) ``` For a direct stream `new Response(stream)` per request shows the same on main: `200 "hello"`, `200 ""`, `500`. On this branch the second `new Response(stream)` throws `Body object should not be disturbed or locked`. **Other observable changes.** `request.bodyUsed` after `fetch(request)` with an async iterable body is now `true` (was `false`, the stream was not disturbed). `Bun.readableStreamToText(stream)` on a stream a sink holds rejects with "ReadableStream has already been used" in place of "ReadableStream is locked" (same `ERR_INVALID_STATE`). **Bare streams.** `fetch(url, { body: stream })`, `Bun.write(path, stream)` and `Bun.spawn({ stdin: stream })` keep a bare stream locked too. Natively wired sources already do this for the same calls (`ReadableStream__lockNative` never unlocks, see "errors a Bun.file() stream whose file does not open" from oven-sh#42477). **node v26.3.0.** `fetch(url, { method: "POST", body: jsStream, duplex: "half" })`: afterwards `locked === true` and `getReader()` throws. The same after an upload that an `AbortSignal` stopped. The other five consumers are Bun APIs with no Node counterpart. **The mark comes after the reader acquisition.** A stream that another reader already holds is not marked. `Bun.serve()`, `fetch()`, `Bun.write()` and `HTMLRewriter` reject such a stream before the pump. `Bun.spawn()` stdin checks only `is_disturbed` (`stdio.rs:388`, `:536`), so a locked stream reaches the pump there. **Out of scope (the concern not addressed).** `rsisFinally` calls `clearStreamControllerSlots` also when the pump never got the lock. `Bun.spawn({ stdin: stream })` with a stream the caller holds through `getReader()` reaches that: the caller's `reader.read()` then never settles. This is the same on main and on this branch. oven-sh#41532 rejects a locked stdin stream before the pump. **Not covered here.** - A handler that drains or partly reads a stream with its own reader, releases it, and then returns it in a `Response` still gets a `200` with the empty or remaining body. oven-sh#36110 covers that gate. - Handing the same Node `Readable` or generator object to a second consumer makes a new stream each time. **Changed tests.** Three tests in `bun-write.test.js` (from oven-sh#42114) read the terminal state through `stream.getReader().closed`. One test in `streams.test.js` (from oven-sh#33781) polled `ts.readable.locked` to wait for the sink teardown. Both observed the unlock as a means, not as the subject. They use `finished()` from `node:stream/promises` now. The `streams.test.js` test still proves the teardown ran: `desiredSize` reads `null` only after the controller slot is cleared, and reads `0` for a stream that is only closed. **Fail before.** With `src/` from main and `bun bd`: `web-stream-state.test.ts` 30 of 46 fail, `body.test.ts` 12 of 778 fail, `serve-reused-response.test.ts` 2 of 9 fail, and the three `bun-write.test.js` tests fail on the new `locked` assertion. With this branch all pass. On main only the `Bun.serve()` row of "the stdout of a running child" fails: the other four consumers wire a pipe-backed `FileReader` natively. **Built-in JS.** I searched `src/js` for code that touches a stream after a native sink took it. The only `stream.cancel()` on a web stream is `ReadableFromWeb._destroy`, which oven-sh#42573 handles for the body mixin case. **Earlier reports of the same class.** oven-sh#7001 (fixed in oven-sh#7861) and oven-sh#6860. **Suites run on the debug build.** `test/js/web/streams/{streams,streams-leak,readable-stream-blob-consumed,readable-stream-terminal-barrier-release,transform-stream-leak}`, `test/js/third_party/wpt-streams`, `test/js/web/fetch/{body,body-stream,body-stream-excess,body-clone,body-async-iterator,body-mixin-errors,blob-write,fetch,fetch.stream,fetch-backpressure,fetch-abort-stream-body,fetch-stream-cancel-leak,fetch-redirect,response}`, `test/js/web/request/request`, `test/js/bun/http/{serve,bun-server,serve-reused-response,serve-direct-readable-stream,serve-body-leak,serve-pending-promise-abort-leak,async-iterator-stream,serve-async-stream-client-abort,serve-error-handler-stream,serve-response-stream-sink-leak,serve-stream-body-error,serve-stream-reject-flush-leak}`, `test/js/bun/spawn/{spawn,spawn-stdin-readable-stream}`, `test/js/bun/io/bun-write`, `test/js/workerd/{html-rewriter,html-rewriter-leak}`, `test/js/bun/s3/{s3-stream-cancel-leak,s3-stream-error-gc,s3-upload-stream-gc,s3-write-to-file-sync-close,s3-connection-close}`, `test/js/node/stream/{node-stream,web-stream-state}`, `test/js/node/async_hooks/AsyncLocalStorage`, `test/regression/issue/07001`. The failures that remain also fail with `src/` from main in this container: IPv6, tests that need a non-root user, external hosts, and 5 s timeouts of the debug build. </details> <!-- robobun:evidence:begin --> --- **no test proof** · iteration 0 · platform-specific test(s) that do not run on this machine, deferring to CI, which covers all platforms: test/js/web/streams/streams.test.js, test/js/bun/io/bun-write.test.js <!-- robobun:evidence:end -->
What
Bun.serveresponding withreq.body.pipeThrough(transform)where the transformer's asynctransform()callscontroller.error()(a streaming validator rejecting a payload) segfaults the whole server process when the client aborts the upload mid-body. The same applies tocontroller.enqueue(),controller.terminate()andcontroller.desiredSize.Repro
Assert build:
Release build:
Segmentation fault at address 0x88.Cause
When a Bun native sink (
Bun.serveresponse body,Bun.spawnstdin) consumes aTransformStream's readable side and the consumer aborts, the sink pump's finally step (rsisFinally/resumableReleaseReader/readDirectStreamCloseImplinBunStreamSource.cpp) runsclearStreamControllerSlots()on the readable, settingm_controllerKind = Noneand clearingm_controller. TheTransformStreamstill holds that readable viam_readable, so a later controller method call (from an asynctransform()/flush()that resumed after the abort, or from a queued sink reaction) walkstransformReadableController(), which assertsControllerKind::Defaultand thenuncheckedDowncasts a null slot. Every caller immediately dereferences the result.This is the TransformStream sibling of the tee-branch crash addressed in #33778.
Fix
Make both
transformReadableController()copies null-safe (returnnullptrwhen the readable's controller slot has been cleared) and make every call site skip the readable-side abstract operation on null:transformStreamError,transformStreamDefaultControllerEnqueue/Terminate, thedesiredSizegetter, and the four sink abort/close reaction handlers. The spec AOs (ReadableStreamDefaultControllerError/Close/Enqueue) would no-op on a non-Readablestream anyway; the writable-side error/unblock still runs. This mirrors theteeBranch*Controllerapproach from #33778.Verification
New test in
test/js/web/streams/streams.test.jspassests.readableas stdin to a child process, kills the child, awaits the sink teardown (ts.readable.locked === false), then exercises all four controller members.Without the fix: SIGABRT on the first action, 3/3 on the debug build.
With the fix:
desiredSizereturnsnull,enqueue()throwsTypeError,terminate()anderror()return normally; exit 0, 3/3.test/js/third_party/wpt-streams/wpt-streams.test.ts: 1175 pass / 0 fail.test/js/web/streams/streams.test.js: no new failures (2 pre-existing timeouts also present on main).test/js/bun/spawn/spawn-stdin-readable-stream.test.ts: 25 pass / 0 fail.test/js/web/streams/transform-stream-leak.test.ts: 3 pass / 0 fail.node:netclient-abort loop (400 iterations): survives.