Repository navigation
Conversation
…ealed record The write BIO of a usockets TLS socket refused a record that the kernel did not take while another socket owned the loop's spill slot. BoringSSL had already sealed that record and kept it for a retry with the same bytes, and us_internal_ssl_writev returned a count without it. The next write of other data then sent the stale record and lost as many accepted bytes, and a smaller next write failed with SSL_R_BAD_WRITE_RETRY. Each TLS connection now has its own queue for ciphertext that the kernel did not take, and the BIO takes every record. The spill slot of the loop is removed. The batching rule stays: while one connection holds the rest of a batch, the other sockets send record by record. A close_notify that the kernel does not take is queued, and the FIN follows it. A close that waits for queued ciphertext ends at the socket's timeout.
Collaborator
Author
|
Updated 6:08 AM PT - Oct 3rd, 2026
❌ @robobun, your commit 4f60a4f has 2 failures in
🧪 To try this PR locally: bunx bun-pr 44529That installs a local version of the PR into your bun-44529 --bun |
Collaborator
Author
|
Status of #44529: draft. The self-review has not finished yet. How the problem was reproduced:
|
Collaborator
|
Superseded by #44618, which consolidates the open TLS pull requests. This fix and its tests are in there, either as written, rewritten smaller, or merged with the other PRs that patched the same cause (see the "By area" list in that PR). Closing in favor of it. |
Jarred-Sumner
added a commit
that referenced
this pull request
Oct 6, 2026
…s every record it sealed (#44529) While another TLS socket owned the loop's spill slot, the write BIO refused what the kernel did not take. BoringSSL had already sealed that record and kept it for a retry, and us_internal_ssl_writev did not count it. The next write of other data flushed the old record and dropped as many accepted bytes (silent stream corruption); a shorter one, or any write after setMaxSendFragment(), failed with BAD_WRITE_RETRY and the socket went fatal with no event. The BIO now takes every record. Only what the kernel refused is copied, to the connection (us_ssl_rare_t), and freed when it drains. The loop-wide rule stays: while one connection holds the rest of a batch flush nobody batches, so userspace holds one flush unit per loop plus at most the rest of one record per other stalled connection. That record used to sit, whole, in BoringSSL's own per-connection write buffer. 16 stalled writers on one loop: 340378 B before (94288 slot + 15 x 16406), 275968 B + 76 B per connection now. Nothing is allocated or copied while the kernel keeps up: 200 writes of 64 KiB, 203 send() calls and 0 spills, as before. us_ssl_rare_t 56 -> 64 bytes, loop_ssl_data 392 -> 368, us_socket_t 80 -> 80 (one bit of its padding). Gone with the slot: its owner, the relocation hook for it, and the "slot is another socket's, so this connection dies" branch. A close now waits behind the spill of every stalled connection, not only the slot owner's (16 peers that stop reading, destroy() on all: 9 fds stay open, 1 before). The next commit bounds that wait. tls-low-prio-queue-fixture.ts: a refused flight no longer leaves the next byte unread, so the burst that fails the parked handshakes is a fatal alert.
Jarred-Sumner
added a commit
that referenced
this pull request
Oct 7, 2026
…s every record it sealed (#44529) While another TLS socket owned the loop's spill slot, the write BIO refused what the kernel did not take. BoringSSL had already sealed that record and kept it for a retry, and us_internal_ssl_writev did not count it. The next write of other data flushed the old record and dropped as many accepted bytes (silent stream corruption); a shorter one, or any write after setMaxSendFragment(), failed with BAD_WRITE_RETRY and the socket went fatal with no event. The BIO now takes every record. Only what the kernel refused is copied, to the connection (us_ssl_rare_t), and freed when it drains. The loop-wide rule stays: while one connection holds the rest of a batch flush nobody batches, so userspace holds one flush unit per loop plus at most the rest of one record per other stalled connection. That record used to sit, whole, in BoringSSL's own per-connection write buffer. 16 stalled writers on one loop: 340378 B before (94288 slot + 15 x 16406), 275968 B + 76 B per connection now. Nothing is allocated or copied while the kernel keeps up: 200 writes of 64 KiB, 203 send() calls and 0 spills, as before. us_ssl_rare_t 56 -> 64 bytes, loop_ssl_data 392 -> 368, us_socket_t 80 -> 80 (one bit of its padding). Gone with the slot: its owner, the relocation hook for it, and the "slot is another socket's, so this connection dies" branch. A close now waits behind the spill of every stalled connection, not only the slot owner's (16 peers that stop reading, destroy() on all: 9 fds stay open, 1 before). The next commit bounds that wait. tls-low-prio-queue-fixture.ts: a refused flight no longer leaves the next byte unread, so the burst that fails the parked handshakes is a fatal alert.
Jarred-Sumner
added a commit
that referenced
this pull request
Oct 8, 2026
…s every record it sealed (#44529) While another TLS socket owned the loop's spill slot, the write BIO refused what the kernel did not take. BoringSSL had already sealed that record and kept it for a retry, and us_internal_ssl_writev did not count it. The next write of other data flushed the old record and dropped as many accepted bytes (silent stream corruption); a shorter one, or any write after setMaxSendFragment(), failed with BAD_WRITE_RETRY and the socket went fatal with no event. The BIO now takes every record. Only what the kernel refused is copied, to the connection (us_ssl_rare_t), and freed when it drains. The loop-wide rule stays: while one connection holds the rest of a batch flush nobody batches, so userspace holds one flush unit per loop plus at most the rest of one record per other stalled connection. That record used to sit, whole, in BoringSSL's own per-connection write buffer. 16 stalled writers on one loop: 340378 B before (94288 slot + 15 x 16406), 275968 B + 76 B per connection now. Nothing is allocated or copied while the kernel keeps up: 200 writes of 64 KiB, 203 send() calls and 0 spills, as before. us_ssl_rare_t 56 -> 64 bytes, loop_ssl_data 392 -> 368, us_socket_t 80 -> 80 (one bit of its padding). Gone with the slot: its owner, the relocation hook for it, and the "slot is another socket's, so this connection dies" branch. A close now waits behind the spill of every stalled connection, not only the slot owner's (16 peers that stop reading, destroy() on all: 9 fds stay open, 1 before). The next commit bounds that wait. tls-low-prio-queue-fixture.ts: a refused flight no longer leaves the next byte unread, so the burst that fails the parked handshakes is a fatal alert.
Jarred-Sumner
added a commit
that referenced
this pull request
Oct 10, 2026
…s every record it sealed (#44529) While another TLS socket owned the loop's spill slot, the write BIO refused what the kernel did not take. BoringSSL had already sealed that record and kept it for a retry, and us_internal_ssl_writev did not count it. The next write of other data flushed the old record and dropped as many accepted bytes (silent stream corruption); a shorter one, or any write after setMaxSendFragment(), failed with BAD_WRITE_RETRY and the socket went fatal with no event. The BIO now takes every record. Only what the kernel refused is copied, to the connection (us_ssl_rare_t), and freed when it drains. The loop-wide rule stays: while one connection holds the rest of a batch flush nobody batches, so userspace holds one flush unit per loop plus at most the rest of one record per other stalled connection. That record used to sit, whole, in BoringSSL's own per-connection write buffer. 16 stalled writers on one loop: 340378 B before (94288 slot + 15 x 16406), 275968 B + 76 B per connection now. Nothing is allocated or copied while the kernel keeps up: 200 writes of 64 KiB, 203 send() calls and 0 spills, as before. us_ssl_rare_t 56 -> 64 bytes, loop_ssl_data 392 -> 368, us_socket_t 80 -> 80 (one bit of its padding). Gone with the slot: its owner, the relocation hook for it, and the "slot is another socket's, so this connection dies" branch. A close now waits behind the spill of every stalled connection, not only the slot owner's (16 peers that stop reading, destroy() on all: 9 fds stay open, 1 before). The next commit bounds that wait. tls-low-prio-queue-fixture.ts: a refused flight no longer leaves the next byte unread, so the burst that fails the parked handshakes is a fatal alert.
Jarred-Sumner
added a commit
that referenced
this pull request
Oct 10, 2026
…ps, WebSocket, SQL) (#44618) ### What does this PR do? Consolidates the open TLS pull requests into one. Each was reproduced on `main` and, for `node:*` behavior, on Node v26.3.0 first. About a third are ported as written, the rest are rewritten smaller or merged into one fix where several PRs patched the same cause. One commit per fix, so it can be read commit by commit. Fixes #43520, fixes #31396, fixes #43635, fixes #37193, fixes #43846, fixes #17932, fixes #41061, fixes #36887, fixes #31810, fixes #35240, fixes #32234, fixes #44365, fixes #43807, fixes #42280, fixes #44517. Addresses #41856 (SNI and `servername`; not `checkServerIdentity` for SQL), #24845 (the spin is gone, shown with fault injection on Linux; not run on macOS), #19754 (node-fetch forwards the agent's TLS options; the Kubernetes client itself was not run). #### The ones that matter most | | On `main` | PRs | |---|---|---| | Client certificate disclosure | `https.request()` with a client certificate sends it to a server it then refuses (wrong name, `checkServerIdentity`, `destroy()` in `'secureConnect'`, `terminate()` in `handshake`). A server can force it with a junk record behind its Finished | #43946 | | False `authorized` | Over a Duplex, `secureConnect` with `authorized === true` for a peer that failed the key proof; `secureConnect` for a plaintext peer with `rejectUnauthorized: false` | #44422, #32929 | | Cleartext https | `https.createServer()` without a usable key/cert answers plain HTTP | #41672, #33539 | | Revoked client certificates | An https mTLS server never sees `crl`, so a revoked client is `authorized` | #41641 | | Pooled sockets | Requests with different client certificates or CAs share an `https.Agent` socket and session | #42498 | | Silent plaintext | `tls: [...]` given to `Bun.listen` / `Bun.connect` is plain TCP | #41490 | | Server weakened by a client knob | `NODE_TLS_REJECT_UNAUTHORIZED=0` turns off a server's client-certificate enforcement | #35245 | | Pins never checked | `WebSocket` never calls `tls.checkServerIdentity` and ignores `tls.serverName` | #41648 | | `verify-full` dropped | `PGSSLMODE=verify-*` is lost next to a `TLS_*` URL variable; `tls: true` sends no SNI | #44498 | | Crashes | use-after-free from `destroy()` in `ALPNCallback` over a Duplex; `abort()` on a late `setSession()`; SIGABRT in `fetch` with an https proxy from the environment and a `Bun.file()` body | #44462, #41671, #44458 | | Stream corruption | A TLS `write()` can lose 16 KiB it reported as written while another socket on the loop is stalled | #44529 | | Hangs and spins | 100% CPU on a failing `send()`; a fatal `SSL_write` leaves the socket open forever; `idleTimeout` never sheds a TLS client that ignores `close_notify` | #34510, #38176, #42336 | | Wrong certificate (regression since 1.3.14) | Connections accepted before `stop()` / `close()` get the default certificate and skip their entry's `requestCert` / `ca` | #42355 | | Quadratic Duplex / proxy tunnel | Reading one chunk over a Duplex, CPU: 8 MB 0.88 s → 0.14 s, 16 MB 3.18 s → 0.23 s, 32 MB 11.75 s → 0.39 s; `fetch` upload through CONNECT: 1.7 s → 0.18 s (debug build) | #44464 | #### By area - **fd engine, write path** (`openssl.c`, `socket.c`): #42352, #34510 + #38176 + #42336 as one change, #44529, #44458, #44192. A rejected `send()` ends the write side only and closes at the next writable event unless the peer's bytes are still queued (a 413 sent before a reset is still read). No new per-socket state. Also, on kqueue, **a FIN no longer ends a socket that waits in the low-priority queue** (`loop.c`): with more than 5 TLS handshakes at once, a client that ended right after its handshake could be reset and its server socket report `socket hang up`, because the eof that the sentinel read knote reports was acted on ahead of the unread Finished. That is on `main` too (the macOS entry for `node-tls-server.test.ts` in `test/flaky-tests.txt`: 7 of 48 recent builds of other branches), and this branch made it likelier (6 of 8 builds), since Finished now leaves in one segment with the close_notify. - **Error reporting, both engines**: #44422, #32929, #44516, #37094, #41272 + #42324 + #44223 as one change, #44021, #37472, #43946, #33630. One channel: a fatal error on an established session is reported, then **the engine closes the connection itself**, whatever the owner does with the report. `test/js/bun/net/tls-fatal-error-closes.test.ts` asserts closed-and-nothing-delivered for every owner (node:tls, `Bun.connect`, `Bun.listen`, `fetch` direct and through CONNECT, `Bun.serve`, `WebSocket` direct and through a proxy, Postgres, MySQL, Valkey, Duplex). - **Duplex engine** (`SSLWrapper`, `UpgradedDuplex`): #44462, #43529, #42332, #44464. #43877 + #44394 were in and are **out again**, see "Worth a look" 5. - **node:tls wrap lifecycle** (`net.ts`, `tls.ts`): #38007, #38058, #38028 + #38122 + #38076 as one change (six copies of the attach code become two helpers), #38311, #39008, #38154, #42340 + #42343 + #42339 + #42453 as one change, #43791, #42425, #44085, #42683, #39088, #39040, #40375, and what was still real of #36534. - **SNI, ALPN, server contexts**: #43080, #42050, #37195 + #43849 as one change (**one** SNI matcher for TCP and HTTP/3), #42355, #42285, #33253, part of #37896, part of #37013. A `tls.Server` has one `SSL_CTX`. - **Verification and options**: #44738, #41490, #37005 + the cwd pin of #40984, #31811, #43982, #33483 + #35245, #41810, #32235, #44441, #38092. - **node:tls API and CA store**: #41671, #38145, #32824, #43594, #39997, #41696, #33534, #34748, #42991, #42996, #42970. - **node:https, Agent, `ws`, node-fetch**: #41672 (https half), #41641, #38261, #42498, #44346, #35609, #31397, #42325. - **WebSocket client**: #41648, #37487 + #43048 as one change. - **SQL, Redis**: #33666, #41711, #44498, part of #42054. - **Tests only**: #41426, #40040, #44395, #44016, #37860, #40591, #44440, #41424. Found on the way and fixed here: an upload that a TLS 1.2 server interrupts with a renegotiation never completes on `main` (0 of 32 runs over `https.request`, `fetch`, `node:tls` and `Bun.connect`: the renegotiation ClientHello lands inside an application record that is still unsent, or the socket gets no `drain` again) and completes here, with two tests from robobun; the fix for #40653 (final flight and first write in one segment) stopped working whenever another TLS socket on the loop was stalled, on `main` too; the `tls.Server` prototype pinned the last server constructed and every `SSL_CTX` it owned; `Object.create(process.env).NODE_TLS_REJECT_UNAUTHORIZED = "0"` turned verification off process-wide once a `SHARE_ENV` worker existed; two debug panics when wrapping a shut-down or still-connecting socket; a `fetch` POST through a proxy sent its headers twice when the origin renegotiated; `BlockList` ignored IPv6 zone ids; a test now ties `root_certs.der` to `certdata.txt`. #### Behavior changes - **A server's `ca` without `requestCert` no longer asks for a client certificate** (`Bun.serve`, `Bun.listen`, HTTP/3, node:tls). It matches the docs and Node. On `main` such a server refused clients with no certificate but served any unrelated self-signed one, so it was never authentication. **Set `requestCert: true` to require a certificate.** A matrix test pins that `requestCert: true` still refuses no certificate and an untrusted one on 8 kinds of server, TLS 1.2 and 1.3, with `NODE_TLS_REJECT_UNAUTHORIZED` unset and `0`. - `NODE_TLS_REJECT_UNAUTHORIZED=0` no longer relaxes a server. - `Bun.connect` / `Bun.listen` hear of a fatal TLS error after the handshake through `error(socket, err)`. With no `error` handler the socket just closes. - HTTP/3 server names match like TCP: `*.` covers exactly one label, case is ignored, a trailing dot is ignored, the last registration of a name wins. - `requestCert` on node:https is `=== true`, as in Node. - An array where a generated options dictionary is expected throws (`tls: []`, `jest.useFakeTimers([])`). - `key` / `cert` arrays serve every identity. A client that can use both gets ECDSA, where `main` served whichever pair came last. - `ecdhCurve` is forwarded by node:https, `ws` and node-fetch now, so a group BoringSSL lacks (`X448`) throws there as it already does in `tls.createServer`. - A wrapped socket's error is re-emitted on the TLS socket as in Node, so `raw.destroy(err)` with a listener on `raw` only is uncaught, as in Node. - `sql.options.tls` is always an object, never `true`. `RedisClient` sends SNI. - `tls: { secureContext }` alone asks for TLS on `Bun.listen` / `Bun.connect` (it was plain TCP), and a value that is not a `SecureContext` throws. The context is served as it is: the `requestCert` / `rejectUnauthorized` it was created with hold whatever the options next to it say, and `requestCert` in the options over a context that does not ask throws at `listen()`. - `tls.DEFAULT_CIPHERS` reaches every client once assigned (`fetch`, `WebSocket`, `Bun.connect`, `RedisClient`, `Bun.SQL`, `S3Client`, proxy tunnels) and servers again. A list that selects no cipher throws `ERR_SSL_NO_CIPHER_MATCH` at the assignment. `fetch.preconnect()` dials nothing after an assignment. - The warning for an unreadable `NODE_EXTRA_CA_CERTS` is Node's one line, without the `warn:` prefix. - `BUN_CONFIG_WS_CLOSE_TIMEOUT` (default 30 s): how long a `WebSocket` client waits for the server to close the connection after the closing handshake. #### Worth a look in review 1. **#44529**: the kernel-refused remainder of a TLS write moves from the loop's one slot onto the connection (in the existing rare struct), so the write BIO never refuses a sealed record. Nothing is allocated on an unstalled path (200 writes: 0 appends, same `send()` count as `main`), memory with 16 stalled writers is lower than on `main` (276 KB vs 340 KB, which `main` holds inside BoringSSL's buffers), `us_socket_t` stays 80 bytes. It needs a bound on how long a deferred close waits, or a peer that stops reading pins the fd past `destroy()`: `US_SSL_CLOSE_AFTER_SPILL_TIMEOUT` is a fixed 10 s, not re-armed on progress. Separate commits, but the fix that keeps the client certificate off the wire beside a stalled socket builds on them. 2. **The default name check of node:tls also runs inside the handshake**, so a wrong-name server gets no client certificate on TLS 1.2 either. JS still runs it after every successful handshake, so a difference between the two matchers can only refuse. Error objects are byte-identical. 3. **#44441** widens trust by design: a self-issued leaf whose `keyUsage` lacks `keyCertSign` (`dotnet dev-certs`) is its own anchor when the store holds a byte-identical copy. No BoringSSL change. Expired pin, same subject with another key, wrong EKU and a pinned intermediate are tested to fail. 4. **#32235** only adds Ed25519 and ECDSA P-521 to the verify list. A captured ClientHello shows `main`'s list with the two inserted; `rsa_pkcs1_sha1` stays. 5. **A stream that a TLS socket wraps, when that TLS socket closes.** An earlier state of this branch lost data here while CI was green (found by #44709's report): with the peer closing first, 4 of 8 MiB arrived with TLS in TLS, 4 of 32 MiB on the http2 `emit("connection")` path, and a `write()` with no `'error'` listener ended the process. Three Node-parity changes only hold together: destroying the wrapped stream at the close (#38028 + #38122 + #38076, #38154) is safe only if every write has really completed (#43877), which in turn needs Node's handling of the peer's close_notify, which needs half-open sockets that the GC can collect. So: - #43877 + #44394 are reverted and reopened. A write over a stream completes once the stream has taken the ciphertext, as on `main`. - Until the verdict on the peer lets the session through, the application cannot have written over it. There the wrapped stream is destroyed as in Node, with the sessions below it. That keeps the release of the connection after a failed handshake, a rejected certificate and an early `destroy()`. The same for an http2 socket the application never got, and for `resetAndDestroy()`. - After that it is `main`'s teardown: a `net.Socket` only gets the engine's `end()`, closes at its peer's FIN, keeps its own timeout and reports its own errors. Any other stream is destroyed with the TLS socket. The regular suites cannot see any of this (999 files were green on every broken variant), so it was steered by eleven seeded differential fuzzers run on this build, `main`, Node v26.3.0 and the earlier state: close, `end()`, `destroy()`, `destroySoon()`, resets, hung and half-open peers, paused writers, timeouts, two and three sessions deep, over TCP and over Duplexes, before, at and after the handshake, and http2 requests. See "How did you verify". #### Known limits - `fetch` with a `checkServerIdentity` function still sends the client certificate (not the request) to a server the function refuses. On TLS 1.2 any verdict a JS callback gives is too late, as in Node. - `addContext()` / `SNICallback` still do not apply to a server-side socket on the stream engine (`emit("connection", duplex)`, TLS in TLS, unflushed writes, named pipes), as on `main`. - A CA bundled in a pfx extends an explicit `ca` only, for `ws` / node-fetch / `WebSocket`: the native `ca` can only replace the default store, and that store keeps `SSL_CERT_FILE` / `SSL_CERT_DIR`. - P-521 leaves work on TLS 1.3 only. TLS 1.2 needs secp521r1 in every ClientHello (`it.todo`). - Once `tls.DEFAULT_CIPHERS` is assigned, `fetch(url, { protocol: "http3" })` is `HTTP3Unsupported`, as with an explicit `ciphers`. - `addCACert()` by hand does not extend the chains of a context with several identities. - A throwing `ALPNCallback` sends `no_application_protocol` on both engines. Node sends nothing and its client sees `ECONNRESET`. - TLS in TLS, peer FIN while the outer handshake runs: the inner socket gets one `write EPIPE`, where Node gives `ECONNRESET` (`main` gives it no error at all). - `@SECLEVEL` in `ciphers` is dropped by the `ws` / node-fetch shims, which used to ignore `ciphers`. node:tls keeps throwing `ERR_SSL_INVALID_COMMAND`. - Beside a stalled TLS socket only the first record (16 KiB) of the first write leaves with the handshake flight. The rest goes record by record, which is what bounds the memory of stalled writers. - After a fatal error on an established session the socket emits `'error'` and then `'close'`. Node emits `'error'` and leaves the socket open. - A paused reader whose own write the kernel rejects loses what it had not read yet, with an `EPIPE`, as on Node. `main` reports no error there and delivers it. - On `main` too: a `Bun.listen` socket without `allowHalfOpen` that has unsent ciphertext when the client's `shutdown()` arrives loses that ciphertext (32 KiB), and over plain TCP `end()` with the peer still sending is a close over unread input, so a reset. - Differences from both `main` and Node that the differential runs below found and that stay, all with a peer that aborts: `ECONNRESET` instead of a clean `'end'` after the socket's own `'finish'` when the peer destroyed with unread data; under TLS 1.2, a zero-length `write()` followed by `destroy()` in `'secureConnection'` leaves the client without `'secureConnect'` (a plain `destroy()` there matches Node); a TLS 1.2 client that destroys in `'secureConnect'` gets no `'session'`; a `ClientRequest` whose handshake fails with an alert emits `'error'` and `'close'` but no `'finish'` (`writableFinished` is true). - Once `tls.DEFAULT_CIPHERS` is assigned, `fetch.preconnect()` opens nothing: `fetch()` then uses a context of its own, and a socket warmed under the default one would never be picked up. - A TLS `send()` that the kernel refuses outside a `write()` call (the drain of unsent ciphertext) is reported with the close, as `read EPIPE` / `read ECONNRESET`. Node says `write EPIPE`. `main` does not report it at all. - Once the application has a session over a `net.Socket` (TLS in TLS, http2 `emit("connection")`), a peer that never sends its FIN holds that socket after the TLS socket closed, as on `main`. Node destroys it. Two tests of #38154 are `todo` for this. Closing it any earlier (at its `'finish'`, say) makes the kernel drop what it has not sent yet as soon as the peer's close_notify arrives. - Plaintext that was queued on a socket before it was wrapped (STARTTLS with a backlog) is dropped when the TLS socket is destroyed, or its handshake fails, before the session is accepted. Node drops it too, except on `destroySoon()`. `main` sends it. - Over a stream that is no `net.Socket`, `end()` can still cut what that stream has buffered, and there is no backpressure, both as on `main` (#43877). - `tls.secureContext` (the undocumented door node:tls uses) is not read by a Windows named pipe listener, which builds its context from the options. On `upgradeTLS({ isServer: true })` the options next to it are the policy, as with Node's `SetVerifyMode`. - `selectServerName()` rebuilds the name tree per ClientHello for injected sockets of a server with `addContext()` entries: 0.4 µs for 1 entry, 3.7 µs for 10, 41 µs for 100, against 631–1111 µs for a handshake. #### Not included Left open, because they need a decision or are not TLS: #43877 + #44394 (see "Worth a look" 5; #43874 stays open with them), #38548, #38591 (both shrink who is trusted), #41589 (`verify-full` vs `NODE_TLS_REJECT_UNAUTHORIZED=0`), #37197, #41706, #43216, #33487, #33545, #36707, #32435, #37255, #28691, #40275, #30314 (features), #38120 (needs the BoringSSL fork, as did #33517, which the stale bot has closed since), #38529 (needs a Windows measurement), #34342, #38232, #43089, #44454, #40451, #42710, #44527, #38088, #38093, #41898. #37896, #42054 and #37013 stay open for the halves not taken. `http.createServer({ key, cert })` keeps serving TLS on purpose. One open question: `tls: {}` (an object that names no TLS option) is plain TCP on `Bun.listen` / `Bun.connect`, here and on `main`. It is the same trap as `tls: []`, but changing it changes a Bun default, so it is left alone. ### How did you verify your code works? - Every new test fails on `main` for the stated reason and passes here, except guards that pin existing behavior, each shown to fail when its clause is removed. `node:*` tests also pass on Node v26.3.0; the few that cannot say which Node version has the behavior. - 212 test files that touch TLS, sockets, http, http2, fetch, WebSocket, SQL, Valkey and workers: 6154 pass, 2 fail. Both are seen on `main` too: `serve.test.ts` "root range port" (the box runs as root), and `worker_threads.test.ts` "terminate(): nothing of the worker's runs after the request", which is flaky there and passed in the run below. - 58 of those files the way the ASAN lane runs them (LeakSanitizer + `BUN_JSC_validateExceptionChecks`): 58 files, 48 of them with leak checking, 4132 pass, 3 fail. All three also fail on `main`: `serve.test.ts` "root range port", `node-net.test.ts` "should not leak when connect({path}) fails synchronously on a reused handle" (times out under this environment), `worker_threads.test.ts` "process.exit() with a shell cp in flight" (a `ShellCpTask` leak). - 647 vendored `test-tls-*`, `test-https-*`, `test-net-*`, `test-http2-*`: the only two failures also fail on `main`. - The SNI matcher was diffed against both old matchers: 3 seeds × 1.23 M lookups × 3 registration flavours, every difference in one of the intended classes, TCP and HTTP/3 identical on every lookup. - The headline rows were also driven by hand with scripts against this build, `main` and Node v26.3.0: cleartext https, `crl`, `tls: []` / `{ secureContext }`, the `ca` / `requestCert` matrix, the client certificate on a wrong-name server, late `setSession()`, `destroy()` in `ALPNCallback`, `[rsa, ec]` identities with an intermediate from `ca`, `WebSocket` `checkServerIdentity`, a corrupted record, the Duplex read above, `tls.DEFAULT_CIPHERS`. - The `setSession()` guard was checked against the real `abort()` at 43 handshake states. - `bun run rust:check-all`: 12 of 12 targets. `tsc`, oxlint, source lints, prettier, rustfmt, mordant clean. - usockets' `_Nonnull` is compiled out of debug builds, so 105 of those files were also run on a local release ASAN build with the CI runner's environment (92 with leak checking): 4595 pass, 1 fail, `child_process.test.ts` "spawn reports EPERM after dropping privileges", which cannot pass as root and fails on `main` too. - The close of a TLS socket over another stream ("Worth a look" 5): eleven seeded differential fuzzers, 8,424 scenarios compared, each run on a release ASAN build of this branch, on `main`, on Node v26.3.0 and on the earlier state of the branch. Against `main`: - Data that `main` delivers in full is cut in 5 scenarios, and about 150 that `main` cuts arrive in full. Of the 5, in 2 `main` never notices the peer's close and keeps the socket for good, 2 call `end()` on the middle one of three sessions over an in-memory Duplex, and 1 does the same on Node. - No dead timeout, no silent reset and no uncaught error that `main` does not have (4 uncaught errors fewer). - A socket stays open where `main` closes it in 109, and closes where `main` keeps it in 295. 92 of the 109 do the same on Node or on the earlier state (a `destroy()` that an in-memory Duplex does not show its peer, half-open peers). 14 wait for a peer that paused reading and so does not read the FIN (#42332's backpressure, as in Node); the socket's own timeout fires there. 3 are left: one on a 5 ms timer, two with three sessions over an in-memory Duplex. - The earlier state of the branch cut data in 173 of the 400 scenarios of one of them, where `main` cuts none and this cuts none. - 23 new tests pin what they found. Each earlier attempt at this fix fails the ones that describe it, the earlier state of the branch fails 7, and all pass on Node. - After that change: 999 test files on the release ASAN build (20,246 pass; the 11 files that fail need a database, Docker, DNS or a non-root user, or share a temp directory with a parallel run and pass alone), 61 on the debug build. - TLS over a file descriptor (`openssl.c`, the path of `fetch`, `Bun.serve`, `tls.connect`, `Bun.connect`) got the same treatment after the rebase: seeded differential fuzzers on CI's release build of this branch, on `main` and, for `node:*`, on Node v26.3.0. Every runtime also against itself for the noise floor, injected faults and known bugs of `main` as positive controls, and a difference counts only if it shows in 5 of 5 fresh processes. - `node:tls` over TCP: 11,500 scenarios (one connection with Node as the oracle line by line; 2 to 60 connections beside stalled neighbours; raw peers that break the handshake). HTTPS: about 136,000 runs over `Bun.serve` + `fetch`, `node:https`, `node:http2` and `wss://`, also with the two ends in different runtimes. `Bun.connect` / `Bun.listen` / `upgradeTLS`: 11,500 scenarios and 720 slow connections, with writers driven by what `write()` returns, beside up to 6 stalled, dripping, closing or resetting neighbours, and plain TCP as a second oracle. No crash, hang, duplication, reordering or silent truncation, and no change in time or in connection reuse. - They found six things that `main` does better, none of which any test showed. All are fixed, each with a test that fails on the build before: what the peer sent lost behind a rejected `send()` (23 scenarios, and an early HTTPS response lost with only `EPIPE`), the same silently for a paused reader, `server.close()` never calling back on a half-open server after a ClientHello and a reset (17), `closeAllConnections()` taking 12 s with a stalled client, `end()` losing up to 1.3 of 4 MiB that `write()` had reported while the peer still uploads, and `end()` a little after a stall never closing beside other stalled TLS sockets. The last two fixes also deliver the 1 to 2 MiB that `main` loses there, and close the socket that `main` keeps for good without such neighbours. - All of them again after every fix, on CI's release build of it. That caught one regression of a fix itself (a reader stopped for backpressure lost 86,385 bytes, 1 of 6,000 scenarios), fixed too. On the last build: scenarios that lose data where `main` does not 23 → 2, and Node loses it in both, with the same `EPIPE`; `server.close()` that never calls back 17 → 0; connections held 4 → 0; requests that end in an error only where `main` has a response 6 → 0. With a Node server in another process, a request ends in an error only in 8 and 10 of 1,500 scenarios here, 5 and 3 on `main`, 10 with Node as the client. - `Bun.connect` / `Bun.listen` on the last build against `main`, in scenarios: hangs 0 against 1,031, sockets and fds never released 0 against 965, corrupted data 0 against 345, `abort()` 0 against 26 (`setSession()` after the handshake), writers that never close 0 against 101 of 720 connections. No kind of failure shows here and not on `main`. About a third of the slow connections close later than on `main`, in 1 to 16 s instead of at once, waiting for unsent ciphertext or for the peer's close_notify, and 79 more of them deliver all that `write()` reported. RSS and time with 16 to 256 stalled writers are the same. - What they found that `main` does worse: a `WebSocket` that calls `close()` with sends pending loses messages in 81 of 999 scenarios (0 here), 37 server sockets left open, 10 `server.close()` that never call back, 20 write callbacks that never run. - The kqueue fix cannot be run on Linux. The `connectionListener` count test now says what became of a missing connection, which is how the cause was found (`'tlsClientError'` "socket hang up", then `read ECONNRESET` at the client of the same port, after its `'secureConnect'`). On macOS x64 it failed every attempt of the three builds before the fix and passed at the first attempt of the build with it. - Windows and macOS were only run by CI. Four new tests asserted what only the Linux kernel does (a FIN read ahead of a reset, unread bytes surviving a reset, loopback buffer sizes, `fstat()` on a socket) and now say so per platform. --------- Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com> Co-authored-by: robobun <117481402+robobun@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
Bun.connect/Bun.listenTLSwrite()can report N bytes while one more sealed record stays inside BoringSSL, not counted. The next write of other data sends that record and loses as many accepted bytes. A smaller next write fails withSSL_R_BAD_WRITE_RETRYand never drains.BIO_s_custom_write(packages/bun-usockets/src/crypto/openssl.c) refuses a record the kernel does not take. BoringSSL already sealed it, andus_internal_ssl_writevdoes not count it.Fix
test/js/bun/net/socket.test.ts(5 fail without the change), and the TLS fault-injection and backpressure suites.Background
send(). The spill slot held the part the kernel refused, one socket per event loop.Downsides
us_ssl_rare_tgrows from 56 to 64 bytes per accepted TLS connection.Notes
Reproduction (linux x64 loopback, release
1.4.3-canary.1+367d939d9and a debug build of main)Two TLS clients on one loop. The first fills the kernel and stays stalled. The second calls
write(1 MiB)until a write is short, then writes other data.write()did not report, and 16384 accepted bytes never arrivewrite()returns 0, nodrain, the socket is marked fatal with no eventsetMaxSendFragment(512)Mechanism
do_tls_write(BoringSSLssl/s3_pkt.cc) seals the record before it flushes. When the BIO refuses, it keeps the record andpending_write, and expects a retry with the same bytes.SSL_MODE_ACCEPT_MOVING_WRITE_BUFFER, so a retry with another buffer of the same size or more flushes the old record and reports that many bytes of the new buffer as written. A smaller buffer fails withSSL_R_BAD_WRITE_RETRY.node:net,node:tlsandBun.servekeep the unsent remainder and retry exactly it, which is what BoringSSL expects.What changes besides the count
SSL_shutdown,SSL_do_handshakeandSSL_readwrite through the same BIO. Their records (close_notify, handshake flights, alerts) are queued too. The dropped close_notify branch inssl_handle_shutdownis gone: the alert goes out when the kernel takes it, then the FIN.loop.c,us_socket_resume).Bun.connectsocket then stayed open. When the holder armed no timeout and bytes are still unsent, the close arms 10 s, the default idle timeout ofBun.serve. The timeout wheel ticks every 4 s, so that is 8 to 12 s, and each drain progress starts it again.SSL_ERROR_WANT_WRITEarms of the handshake and read drivers withssl_read_wants_write, the foreign-slot kill inssl_flush_write_batch, the spill re-point inus_internal_ssl_socket_relocated.Worst case held in the queue: one flush unit (at most about 147 KB) for the one connection per loop that holds the rest of a batch, and the rest of one record (16 KiB of plaintext) plus a handshake flight or an alert for any other. Main held the same: the record sat inside BoringSSL. No record of the other drivers is unbounded: BoringSSL sends one KeyUpdate reply per write of ours, and client renegotiation is limited to 3 per 600 s.
Measurements (main
519963edc8against this branch)llvm-dwarfdump,llvm-size):us_socket_t80 to 80,us_ssl_rare_t56 to 64,loop_ssl_data384 to 368 bytes..textofopenssl.o22423 to 23121,loop.o5983 to 6079,socket.o4882 to 4896 bytes. A release binary was not built.send()calls, gdb breakpoint hit counts (releasebun-profileof 367d939 against the debug build of this branch): 200 writes of 64 KiB with nothing stalled 204 to 204. Writer beside a stalled socket 193 to 193. Single writer that stalls 27 to 27.ssl_out_queue_append): 0 for the 200 unstalled writes. In the new test (one run): 42256 bytes for the stalled socket (rest of a batch) and 7462 bytes for the socket under test (rest of one record).Bun.serve({ tls, idleTimeout: 2 }): connections still open 30 s after the timeout, 1 of 4 to 0 of 4 (runs of 40 s: 1 on the release build, 2 on this branch).perforvalgrindin the container) and server CPU beside a stalled socket (no release build of the branch).Open question for a maintainer
The usockets rewrite (#34037) records a directive to keep one loop-shared spill slot so that memory stays O(1), and it rejects a spill per socket. This PR keeps the byte bound (the same gate), but the bytes live on the connection. If the buffer itself must stay loop-shared, the alternative is to keep the slot and add a place for one refused record per connection.
Other PRs
end(data)test on release lanes).Not covered
ECONNRESETorEPIPEis still queued, as it was spilled before (tls: report a rejected send() as a write error instead of a clean close #42336).endfor a close_notify and for a bare FIN.Tests run on the debug build (linux x64):
socket.test.ts,socket-syscall-fault.test.ts,tls-syscall-fault.test.ts,node-tls-server.test.ts,node-http-backpressure.test.ts,node-tls-connect.test.ts,node-tls-duplex-end-verify.test.ts. Before the last rebase alsonode-tls-cert,node-tls-context,node-tls-upgrade,renegotiation,node-tls-raw-end,node-tls-socket-allow-half-open-option,node-tls-wrapped-socket-close,tls-connect-socket-churn,tcp-server,socket-retention,tls-reject-before-client-cert,bun-serve-ssl,tls-keepalive,node-http-pinned-write,node-https-checkServerIdentity,fetch.tls,node-http2. Failures seen were 5 s timeouts of tests that pass when run alone on the loaded machine, and one test that needs the public internet.The low-prio queue fixture (
tls-low-prio-queue-fixture.ts) failed sockets through the refused-flight close. Its clients now send a fatal alert instead. With the double-park guard inloop.cremoved it still aborts ongroup->low_prio_count == 0.