crypto: cap GCM IV length at 128 bytes in createCipheriv - #34092
Conversation
Node.js (via OpenSSL 3's GCM provider) rejects GCM IVs longer than 1024 bits with ERR_CRYPTO_INVALID_IV. BoringSSL's EVP_CTRL_AEAD_SET_IVLEN has no such cap, so Bun accepted arbitrarily long IVs, turning an attacker-controlled IV length into unbounded GHASH work inside the constructor and producing ciphertext Node cannot decipher. Enforce the same 128-byte cap at the constructor for GCM ciphers.
|
Updated 5:32 PM PT - Jul 13th, 2026
✅ @robobun, your commit 94450969278d6af060a6a66dd7ca1cfa279c12ab passed in 🧪 To try this PR locally: bunx bun-pr 34092That installs a local version of the PR into your bun-34092 --bun |
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: ASSERTIVE Plan: Pro Run ID: 📒 Files selected for processing (3)
WalkthroughChangesGCM cipher construction now rejects IVs longer than 128 bytes. Tests cover the 128-byte boundary, oversized IVs across AES-GCM variants, and the corresponding Node compatibility limit. GCM IV Validation
Suggested reviewers: 🚥 Pre-merge checks | ✅ 4✅ Passed checks (4 passed)
Comment |
There was a problem hiding this comment.
Beyond the inline nit, I also checked that ivView cannot be null at the new GCM check — GCM's getIvLength() is non-zero, so the earlier !ivView && expectedIvLen != 0 guard has already returned ERR_CRYPTO_INVALID_IV; the ASSERT(ivView) is safe and mirrors the ChaCha20-Poly1305 block directly above.
Extended reasoning...
The added block follows the exact shape of the adjacent isChaCha20Poly1305() check, and the only non-obvious question is whether ivView can be null there. It cannot: for GCM ciphers cipher.getIvLength() returns 12, so a null ivView is rejected by the !ivView && expectedIvLen != 0 branch a few lines earlier. The change strictly tightens validation to match Node/OpenSSL 3, so I'm not approving only because it's in the crypto path — the code itself looks correct.
#39445) Stacked on #36463 (the base branch is that PR's branch, so the diff here is only the additions). Merging this into #36463 adds the behavior changes listed below; #36463 itself now covers the #38333 install batch, the optional-peer correction, and the TOML / `bun init` fixes, so this PR no longer touches those. ### Problem - These 1.3 to 1.4 behavior changes are not in the guide at `701b3e2a0`: - MySQL: the first `caching_sha2_password` connection over plain TCP is refused unless `allowPublicKeyRetrieval: true` (#31129; 1.3.14 requested the key automatically, `MySQLConnection.zig` in the 1.3.14 tag). SQL `tls` / `ssl` options now require TLS instead of falling back to plaintext, and `?ssl=` / `?ssl-mode=` are read (`shared.ts` 1.3.14 only read `?sslmode=`; #37669). - Install: `~/.npmrc` fallback when `XDG_CONFIG_HOME` is set (#36289), credentials in `--registry` / env / bunfig object URLs are sent and outrank same-host `.npmrc` tokens (#38796, #38824), `bun outdated` exits 1 on fetch failures (#38809), new `dedupe` / `up` commands shadow scripts of those names and `bun feedback` is removed (#38333, #38444), `workspace:` ranges inside registry packages (#37669), isolated store entry names (#39014). - Runtime: `module.enableCompileCache()` / `NODE_COMPILE_CACHE` implemented (#34660), `require()` / `import` not-found messages (#34660), `AbortError` message without the period (#39277; 1.3.14's `BunCommonStrings.h` has the period), GCM IV length (#34092), `mkdtemp("")` (#34908), vm options (#38381), `server.reload` (#38697), ICU 75/73 to 78 (#38013), Compression stream chunking (#38695), `Bun.SQL` sqlite bindings (#35950). - Bundler: `splitting` with `cjs` / `iife` is an error (#32685), block-scoped `enum` lowers to `let` (#34249), exports emitted ascending instead of descending (#35957; `doStep5.zig` in 1.3.14 used `sortDesc`), minified `$` (#35668). - The TOML integer bullet did not say what the limit or the fix is. ### Fix - Adds a MySQL public key section (plus a summary table row), a TLS note under the `PGSSLMODE` section, an `.npmrc` / credentials addendum to the `bunfig.toml` section, a `module.enableCompileCache()` section, and the rest as bullets in the existing lists. - `docs/pm/overrides.mdx`: one-line change adding a pointer to this guide in the existing `lockfileVersion` 3 limitation. (The base branch briefly had a duplicate "Nested overrides" section; it removed that itself in `8257d01acb`, and this PR was rebased over it.) - Verification: each runtime claim was run against `1.4.0-canary.1+8326d1bd3` (22 commits behind main; contains every change referenced), and each install or bundler claim was checked against the source on main, with the 1.3 side taken from the `bun-v1.3.14` tag where the PR body did not state it. The `/runtime/sql#mysql` and `/upgrade-to-1.4` links resolve. `prettier --check` passes. ### Not included on purpose - Lifecycle scripts no longer receiving `npm_package_name` / `npm_package_version` / `npm_package_json` / `npm_config_local_prefix` during `bun install`, and transitive `"*"` ranges no longer deduplicating onto the root's version: regressions with open fixes (#36690, #38110, #38770). They need either the fixes or a guide line before release. - Postgres `sslmode=prefer` / `allow` (including `PGSSLMODE=prefer`, which 1.4 newly reads) hangs until the connection timeout against a server without SSL because nothing sends the startup message after the `N` reply. Same code in 1.3.14; filed as a bug instead of documented. <details> <summary>Commands used to verify the runtime claims</summary> ``` timers/promises setTimeout with an aborted signal # "The operation was aborted" bun req.cjs # Cannot find module ... Require stack: bun b.mjs (import() of a missing package / relative file) # Cannot find package 'x' imported from /path, ERR_MODULE_NOT_FOUND bun a_static.mjs (unhandled static import) # printed line still: Cannot find package 'x' from '/path' process.versions.icu # 78.3 createCipheriv("aes-128-gcm", key, Buffer.alloc(129)) # ERR_CRYPTO_INVALID_IV DecompressionStream of a 1 MiB gzip member # 16 chunks of 65536 bytes new SQL("sqlite://:memory:") with ${[1,2]} / ${new Date()} # Binding expected ... fs.mkdtempSync("") # EINVAL vm.runInThisContext("1", []) # ERR_INVALID_ARG_TYPE NODE_COMPILE_CACHE=/tmp/cc bun cc.cjs # creates /tmp/cc/v1.4.0-x86_64-<sha>-<uid> NODE_DISABLE_COMPILE_CACHE=1 + enableCompileCache() # status 3 (DISABLED) bun dedupe / bun up with package.json scripts of those names # built-in command runs bun feedback # Script not found "feedback" Bun.build({ splitting: true, format: "cjs" }) # Code splitting is currently only supported ... bun build of a function-scoped enum and import * as ns # let Color; exports a, m, z new SQL({ url: "postgres://...", tls: true }) on a non-TLS server # ERR_POSTGRES_TLS_NOT_AVAILABLE Bun.TOML.parse("a = 9007199254740993") # Integer cannot be losslessly represented ... ``` </details> <details> <summary>Previous revision</summary> The first revision of this PR (`3c5611454a`) also rewrote the package manager section for #38333 / #38853 (nested overrides and `lockfileVersion: 3`, the optional-peer correction, `bun update`, `bunfig.toml` over `.npmrc`, `--filter`) and fixed the TOML date and `bun init` lines. #36463 picked those up in its own commits the same day, so this PR was rebased onto its new head and reduced to the items above. </details> <!-- robobun:evidence:begin --> --- **no test proof** · iteration 0 · docs-only change; test-proof not applicable <!-- robobun:evidence:end -->
#39445) Stacked on #36463 (the base branch is that PR's branch, so the diff here is only the additions). Merging this into #36463 adds the behavior changes listed below; #36463 itself now covers the #38333 install batch, the optional-peer correction, and the TOML / `bun init` fixes, so this PR no longer touches those. ### Problem - These 1.3 to 1.4 behavior changes are not in the guide at `701b3e2a0`: - MySQL: the first `caching_sha2_password` connection over plain TCP is refused unless `allowPublicKeyRetrieval: true` (#31129; 1.3.14 requested the key automatically, `MySQLConnection.zig` in the 1.3.14 tag). SQL `tls` / `ssl` options now require TLS instead of falling back to plaintext, and `?ssl=` / `?ssl-mode=` are read (`shared.ts` 1.3.14 only read `?sslmode=`; #37669). - Install: `~/.npmrc` fallback when `XDG_CONFIG_HOME` is set (#36289), credentials in `--registry` / env / bunfig object URLs are sent and outrank same-host `.npmrc` tokens (#38796, #38824), `bun outdated` exits 1 on fetch failures (#38809), new `dedupe` / `up` commands shadow scripts of those names and `bun feedback` is removed (#38333, #38444), `workspace:` ranges inside registry packages (#37669), isolated store entry names (#39014). - Runtime: `module.enableCompileCache()` / `NODE_COMPILE_CACHE` implemented (#34660), `require()` / `import` not-found messages (#34660), `AbortError` message without the period (#39277; 1.3.14's `BunCommonStrings.h` has the period), GCM IV length (#34092), `mkdtemp("")` (#34908), vm options (#38381), `server.reload` (#38697), ICU 75/73 to 78 (#38013), Compression stream chunking (#38695), `Bun.SQL` sqlite bindings (#35950). - Bundler: `splitting` with `cjs` / `iife` is an error (#32685), block-scoped `enum` lowers to `let` (#34249), exports emitted ascending instead of descending (#35957; `doStep5.zig` in 1.3.14 used `sortDesc`), minified `$` (#35668). - The TOML integer bullet did not say what the limit or the fix is. ### Fix - Adds a MySQL public key section (plus a summary table row), a TLS note under the `PGSSLMODE` section, an `.npmrc` / credentials addendum to the `bunfig.toml` section, a `module.enableCompileCache()` section, and the rest as bullets in the existing lists. - `docs/pm/overrides.mdx`: one-line change adding a pointer to this guide in the existing `lockfileVersion` 3 limitation. (The base branch briefly had a duplicate "Nested overrides" section; it removed that itself in `8257d01acb`, and this PR was rebased over it.) - Verification: each runtime claim was run against `1.4.0-canary.1+8326d1bd3` (22 commits behind main; contains every change referenced), and each install or bundler claim was checked against the source on main, with the 1.3 side taken from the `bun-v1.3.14` tag where the PR body did not state it. The `/runtime/sql#mysql` and `/upgrade-to-1.4` links resolve. `prettier --check` passes. ### Not included on purpose - Lifecycle scripts no longer receiving `npm_package_name` / `npm_package_version` / `npm_package_json` / `npm_config_local_prefix` during `bun install`, and transitive `"*"` ranges no longer deduplicating onto the root's version: regressions with open fixes (#36690, #38110, #38770). They need either the fixes or a guide line before release. - Postgres `sslmode=prefer` / `allow` (including `PGSSLMODE=prefer`, which 1.4 newly reads) hangs until the connection timeout against a server without SSL because nothing sends the startup message after the `N` reply. Same code in 1.3.14; filed as a bug instead of documented. <details> <summary>Commands used to verify the runtime claims</summary> ``` timers/promises setTimeout with an aborted signal # "The operation was aborted" bun req.cjs # Cannot find module ... Require stack: bun b.mjs (import() of a missing package / relative file) # Cannot find package 'x' imported from /path, ERR_MODULE_NOT_FOUND bun a_static.mjs (unhandled static import) # printed line still: Cannot find package 'x' from '/path' process.versions.icu # 78.3 createCipheriv("aes-128-gcm", key, Buffer.alloc(129)) # ERR_CRYPTO_INVALID_IV DecompressionStream of a 1 MiB gzip member # 16 chunks of 65536 bytes new SQL("sqlite://:memory:") with ${[1,2]} / ${new Date()} # Binding expected ... fs.mkdtempSync("") # EINVAL vm.runInThisContext("1", []) # ERR_INVALID_ARG_TYPE NODE_COMPILE_CACHE=/tmp/cc bun cc.cjs # creates /tmp/cc/v1.4.0-x86_64-<sha>-<uid> NODE_DISABLE_COMPILE_CACHE=1 + enableCompileCache() # status 3 (DISABLED) bun dedupe / bun up with package.json scripts of those names # built-in command runs bun feedback # Script not found "feedback" Bun.build({ splitting: true, format: "cjs" }) # Code splitting is currently only supported ... bun build of a function-scoped enum and import * as ns # let Color; exports a, m, z new SQL({ url: "postgres://...", tls: true }) on a non-TLS server # ERR_POSTGRES_TLS_NOT_AVAILABLE Bun.TOML.parse("a = 9007199254740993") # Integer cannot be losslessly represented ... ``` </details> <details> <summary>Previous revision</summary> The first revision of this PR (`3c5611454a`) also rewrote the package manager section for #38333 / #38853 (nested overrides and `lockfileVersion: 3`, the optional-peer correction, `bun update`, `bunfig.toml` over `.npmrc`, `--filter`) and fixed the TOML date and `bun init` lines. #36463 picked those up in its own commits the same day, so this PR was rebased onto its new head and reduced to the items above. </details> <!-- robobun:evidence:begin --> --- **no test proof** · iteration 0 · docs-only change; test-proof not applicable <!-- robobun:evidence:end -->
What does this PR do?
crypto.createCipherivandcreateDecipherivaccepted GCM IVs of any length (tested up to 64 MiB). Node.js rejects GCM IVs longer than 128 bytes withERR_CRYPTO_INVALID_IVbecause OpenSSL 3's GCM provider enforcesGCM_IV_MAX_SIZE = 1024 / 8. BoringSSL has no such cap, soEVP_CTRL_AEAD_SET_IVLENsilently accepted the oversized IV.Two consequences:
createDecipherivthrows).Reproduction
ERR_CRYPTO_INVALID_IVERR_CRYPTO_INVALID_IVERR_CRYPTO_INVALID_IVERR_CRYPTO_INVALID_IVERR_CRYPTO_INVALID_IVERR_CRYPTO_INVALID_IVAuth tags for the 12- and 128-byte rows are byte-identical across Node and Bun, isolating the divergence to the missing upper bound.
How did you verify your code works?
bun bd test test/js/bun/crypto/cipheriv-decipheriv.test.ts(new test fails on main, passes with fix)bun bd test/js/node/test/parallel/test-crypto-cipheriv-decipheriv.jsbun bd test/js/node/test/parallel/test-crypto-authenticated.jsbun bd test/js/node/test/parallel/test-crypto-authenticated-stream.js[stamp-90s] gate passed · iteration 0 · 3 files touched
fails on main (without fix)
passes on PR (with fix)
diff hotspot
gate history · 1 passed · 0 rejected · iteration 0
evidence per changed file