fulmine.js 5.1.4 → 5.1.7

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/NOTICE CHANGED
@@ -12,6 +12,10 @@ itself and can be inspected with `git log` and `git shortlog -sn`.
12
12
  Fulmine is not affiliated with, endorsed by, or maintained by the authors of
13
13
  Ultimate Express.
14
14
 
15
+ This product includes code derived from fast-querystring
16
+ (https://github.com/anonrig/fast-querystring), Copyright (c) Yagiz Nizipli,
17
+ licensed under the MIT License, vendored in src/parse-query.js.
18
+
15
19
  As required by section 4(b) of the Apache License, the following are the
16
20
  significant changes made to the original work:
17
21
 
@@ -19,10 +23,27 @@ significant changes made to the original work:
19
23
  - Public API documented, and type checked from those annotations.
20
24
  - Router dispatch reworked so a chain of N middlewares allocates one promise
21
25
  instead of N, and the mount stack is no longer walked on every hop.
22
- - Request bodies are collected into a single buffer instead of being copied
23
- once per chunk and again on concatenation.
26
+ - Request bodies are collected in native code by uWebSockets.js when the
27
+ length is known, with the size limit enforced before any byte reaches
28
+ JavaScript, instead of being copied once per chunk and again on
29
+ concatenation; the body parsers read the few headers they need from the
30
+ raw header entries and bind the caller's async context only when a read
31
+ actually goes asynchronous.
24
32
  - Empty request bodies now produce the same value each Express body parser
25
33
  produces, rather than being skipped.
34
+ - Routes whose path is a literal are registered as native uWebSockets.js
35
+ routes, with their path and method handed to the request as
36
+ registration-time constants.
37
+ - sendFile answers small unchanged files from a stat-validated cache, and
38
+ concurrent reads of the same file are coalesced.
39
+ - The query string parser is vendored from fast-querystring and answers on a
40
+ bare null prototype, matching how Express displays parsed queries.
41
+ - Express 5's own test suite runs against Fulmine as a CI gate, from a
42
+ pinned checkout of Express.
43
+ - The app is callable as a request listener, so http.createServer(app),
44
+ supertest and anything else that invokes an app directly keeps working
45
+ through a node:http shim.
46
+ - Releases are built and published to npm from CI.
26
47
  - Benchmark harness reworked: wrk replaced by autocannon so the suite runs
27
48
  anywhere Node does, NODE_ENV is set, load errors and response validation are
28
49
  no longer discarded, rows bounded by shared work are labelled, and scenarios
package/README.md CHANGED
@@ -55,7 +55,7 @@ to run it yourself.
55
55
 
56
56
  Numbers produced by a project about itself deserve suspicion, so Fulmine also stands in public arenas, run by their own rigs under their own rules:
57
57
 
58
- - **[HttpArena](https://www.http-arena.com/#sort=rps:-1&q=Js)** (the link lands filtered on the JavaScript entries): the saved run measures 5.89 million pipelined requests per second, 1.12 million on baseline and 1.04 million on the json profile, ahead of every JavaScript entry on the board.
58
+ - **[HttpArena](https://www.http-arena.com/#sort=rps:-1&q=Js)** (the link lands filtered on the JavaScript entries): first among the JavaScript entries and second overall across every language on the board. The saved run measures 7.64 million pipelined requests per second, 1.12 million on the json profile, 457 thousand on compressed json and 222 thousand on the Postgres profile.
59
59
  - **[web-frameworks](https://github.com/the-benchmarker/web-frameworks)**: entry merged, numbers arrive with their next published round.
60
60
 
61
61
  More to come as their maintainers take the entries in.
@@ -90,6 +90,31 @@ npx fulmine differences # print the list below and change nothing
90
90
  The command is installed under both `fulmine` and `fulmine.js`. Use `fulmine`: `npx` cannot run a
91
91
  command whose name ends in `.js` on Windows, where it exits without a word.
92
92
 
93
+ ## Docker
94
+
95
+ Two things about µWebSockets.js make a Dockerfile that works for Express fail here, and both have easy answers:
96
+
97
+ - **No Alpine, and no Debian bookworm either.** µWebSockets.js ships prebuilt binaries linked against glibc 2.38 or newer. Alpine images use musl, so the binary does not load at all; `node:22` and `node:22-slim` are Debian bookworm, whose glibc 2.36 fails at startup with `GLIBC_2.38' not found`. Use the trixie variants: `node:22-trixie-slim` and up.
98
+ - **`git` must be there when `npm install` runs.** µWebSockets.js is not on npm; it is installed straight from GitHub (`github:uNetworking/uWebSockets.js`), and npm uses git to fetch it. Full images like `node:22-trixie` have git; `-slim` ones do not.
99
+
100
+ The clean way to satisfy both is a multi-stage build: install with the full image, run with the slim one.
101
+
102
+ ```dockerfile
103
+ FROM node:22-trixie AS build
104
+ WORKDIR /app
105
+ COPY package*.json ./
106
+ RUN npm ci --omit=dev
107
+
108
+ FROM node:22-trixie-slim
109
+ WORKDIR /app
110
+ COPY --from=build /app/node_modules ./node_modules
111
+ COPY . .
112
+ EXPOSE 3000
113
+ CMD ["node", "server.js"]
114
+ ```
115
+
116
+ A single-stage `node:22-trixie-slim` image works too if you `apt-get install -y git ca-certificates` before `npm ci`. Prebuilt binaries exist for x64 and arm64 on Linux, macOS and Windows, so nothing is compiled at install time either way.
117
+
93
118
  ## Differences from Express
94
119
 
95
120
  - `app.listen()` returns the app, not an `http.Server`. There is no node server underneath, so `server.close()`, `server.address()` and anything that attaches itself to a real `http.Server` need a look. `app.close()`, `app.address()` and `app.listening` are there and do what you would expect.
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "fulmine.js",
3
- "version": "5.1.4",
3
+ "version": "5.1.7",
4
4
  "description": "Drop-in Express 5 replacement on uWebSockets.js. Your existing middleware keeps working.",
5
5
  "main": "src/index.js",
6
6
  "bin": {
@@ -128,9 +128,10 @@ class Application extends Router {
128
128
  * @param {any} res
129
129
  * @param {any} app
130
130
  * @param {any} [preset]
131
+ * @param {any} [skipHolder]
131
132
  */
132
- constructor(req, res, app, preset) {
133
- super(req, res, app, preset);
133
+ constructor(req, res, app, preset, skipHolder) {
134
+ super(req, res, app, preset, skipHolder);
134
135
  }
135
136
  };
136
137
  this._response = class extends Response {
@@ -359,6 +360,14 @@ class Application extends Router {
359
360
  this.settings[key] = Array.isArray(value) ? value.map((dir) => path.resolve(dir)) : path.resolve(value);
360
361
  return this;
361
362
  } else if (key === "etag") {
363
+ // an etag arriving after listen would make send consult freshness headers the
364
+ // header-skip routes never copied, so those skips are taken back
365
+ if (value !== false && this._skipPresets?.size) {
366
+ for (const preset of this._skipPresets) {
367
+ preset.skipHeaders = false;
368
+ }
369
+ this._skipPresets.clear();
370
+ }
362
371
  if (typeof value === "function") {
363
372
  this.settings["etag fn"] = value;
364
373
  } else {
@@ -431,10 +440,12 @@ class Application extends Router {
431
440
  * @param {any} res uWS response
432
441
  * @param {any} req uWS request, readable only during this call
433
442
  * @param {any} [preset] a literal registration's constants, see nativePreset in the router
443
+ * @param {any} [skipHolder] where a granted header skip lives, forwarded whole: dropping
444
+ * it here silently turned every skip off, since the native closures call this override
434
445
  * @returns {any} the request, with the response reachable as request.res
435
446
  */
436
- handleRequest(res, req, preset) {
437
- const request = super.handleRequest(res, req, preset);
447
+ handleRequest(res, req, preset, skipHolder) {
448
+ const request = super.handleRequest(res, req, preset, skipHolder);
438
449
  // removal rides the close listener the Response constructor already has, since a second
439
450
  // once() per request measured a tenth of a microsecond on the hot path.
440
451
  // An aborted response only flips its flags without emitting 'close', which is why
@@ -22,6 +22,7 @@ const zlib = require("fast-zlib");
22
22
  const typeis = require("type-is");
23
23
  const qs = require("qs");
24
24
  const parseQuery = require("./parse-query.js");
25
+ const { kGetSafe } = require("./usage.js");
25
26
  const { AsyncResource } = require("async_hooks");
26
27
  const { fastQueryParse, NullObject, asStatError, httpError, memoizeByString } = require("./utils.js");
27
28
 
@@ -289,7 +290,10 @@ function serveStatic(root, options) {
289
290
  options._ownEtag = true;
290
291
 
291
292
  return (req, res, next) => {
292
- next = AsyncResource.bind(next);
293
+ // Not bound here: every path down to sendFile is synchronous, statSync included, so the
294
+ // caller's async context is intact at each of these next() calls. Only sendFile's
295
+ // completion can arrive on a uWS callback that carries no context, and that one
296
+ // continuation is bound where it is handed over.
293
297
 
294
298
  // a file is read, not written: anything but GET and HEAD belongs to whoever comes next, or
295
299
  // is refused outright when this middleware is the last word
@@ -410,11 +414,15 @@ function serveStatic(root, options) {
410
414
 
411
415
  options._stat = stat;
412
416
 
413
- return res.sendFile(_path, options, (e) => {
414
- if (e) {
415
- next(options.fallthrough && FALLTHROUGH_STATUSES.has(e.status) ? undefined : e);
416
- }
417
- });
417
+ return res.sendFile(
418
+ _path,
419
+ options,
420
+ AsyncResource.bind((e) => {
421
+ if (e) {
422
+ next(options.fallthrough && FALLTHROUGH_STATUSES.has(e.status) ? undefined : e);
423
+ }
424
+ })
425
+ );
418
426
  };
419
427
  }
420
428
 
@@ -462,9 +470,11 @@ function createInflate(contentEncoding) {
462
470
  * @param {string} [charsetPolicy] which charsets this parser accepts, as body-parser draws the
463
471
  * lines: "utf" (json, utf-* only), "urlencoded" (utf-8 and iso-8859-1), "any" (anything iconv
464
472
  * knows), or undefined for a parser that never decodes (raw)
473
+ * @param {boolean} [keepsBuffer] whether the collected buffer itself escapes to the application,
474
+ * which rules out handing it a view over uWS memory
465
475
  * @returns {(options?: object) => Function} the middleware factory
466
476
  */
467
- function createBodyParser(defaultType, beforeReturn, checkOptions, charsetPolicy) {
477
+ function createBodyParser(defaultType, beforeReturn, checkOptions, charsetPolicy, keepsBuffer) {
468
478
  return function (options) {
469
479
  // a copy, because everything below writes the parsed values back: with the caller's own
470
480
  // object, altering it after the parser was built would alter the parser
@@ -500,6 +510,10 @@ function createBodyParser(defaultType, beforeReturn, checkOptions, charsetPolicy
500
510
  }
501
511
  if (typeof options.defaultCharset === "undefined") options.defaultCharset = "utf-8";
502
512
 
513
+ // whether the collected bytes escape the collection callback: the raw parser hands the
514
+ // buffer itself to the application, and a verify hook may keep what it is shown
515
+ const copyBody = keepsBuffer || typeof options.verify === "function";
516
+
503
517
  // Whether a content-type is one this parser claims, remembered per parser.
504
518
  //
505
519
  // Only reached when the caller asked for a wildcard or a list, since a plain type takes the
@@ -513,15 +527,18 @@ function createBodyParser(defaultType, beforeReturn, checkOptions, charsetPolicy
513
527
 
514
528
  let additionalMethods;
515
529
 
516
- return (req, res, next) => {
517
- next = AsyncResource.bind(next);
530
+ const parserMiddleware = (req, res, next) => {
531
+ // Not bound yet: every return in this prologue is synchronous, so the caller's async
532
+ // context is still intact and an AsyncResource here would be 1.4 microseconds of
533
+ // nothing. The bind happens below, only once a real read is about to go async.
518
534
 
519
535
  // skip reading body twice
520
536
  if (req.bodyRead) {
521
537
  return next();
522
538
  }
523
539
 
524
- const type = req.headers["content-type"];
540
+ // straight from the raw entries: three headers do not justify building the object
541
+ const type = req._rawHeader("content-type");
525
542
 
526
543
  // req.body is deliberately left undefined until a parser claims the request. That is
527
544
  // what lets a handler tell "nothing parsed this" apart from "the body was empty",
@@ -534,13 +551,13 @@ function createBodyParser(defaultType, beforeReturn, checkOptions, charsetPolicy
534
551
  return next();
535
552
  }
536
553
 
537
- const length = req.headers["content-length"];
554
+ const length = req._rawHeader("content-length");
538
555
 
539
556
  // No content-length and no transfer-encoding means the request carries no body at all,
540
557
  // and a body parser must leave it alone rather than parse nothing into an empty value.
541
558
  // type-is applies this before matching the type, but the simpleType shortcut below
542
559
  // compares strings directly and would otherwise skip the check.
543
- if (req.headers["transfer-encoding"] === undefined && isNaN(length)) {
560
+ if (req._rawHeader("transfer-encoding") === undefined && isNaN(length)) {
544
561
  return next();
545
562
  }
546
563
 
@@ -620,7 +637,8 @@ function createBodyParser(defaultType, beforeReturn, checkOptions, charsetPolicy
620
637
  const abs = [];
621
638
  let inflate;
622
639
  let totalSize = 0;
623
- const contentEncoding = (req.headers["content-encoding"] || "identity").toLowerCase();
640
+ const rawContentEncoding = req._rawHeader("content-encoding");
641
+ const contentEncoding = (rawContentEncoding || "identity").toLowerCase();
624
642
  if (!options.inflate && contentEncoding !== "identity") {
625
643
  return next(
626
644
  bodyError("content encoding unsupported", 415, "encoding.unsupported", {
@@ -629,19 +647,64 @@ function createBodyParser(defaultType, beforeReturn, checkOptions, charsetPolicy
629
647
  );
630
648
  }
631
649
  if (options.inflate) {
632
- inflate = createInflate(req.headers["content-encoding"]);
650
+ inflate = createInflate(rawContentEncoding);
633
651
  if (inflate === false) {
634
652
  return next(
635
653
  bodyError(
636
- 'unsupported content encoding "' + req.headers["content-encoding"] + '"',
654
+ 'unsupported content encoding "' + rawContentEncoding + '"',
637
655
  415,
638
656
  "encoding.unsupported",
639
- { encoding: req.headers["content-encoding"] }
657
+ {
658
+ encoding: rawContentEncoding
659
+ }
640
660
  )
641
661
  );
642
662
  }
643
663
  }
644
664
 
665
+ // From here the body really gets read, and uWS delivers it on native callbacks that
666
+ // carry no async context, so this is the one continuation that has to be bound: an
667
+ // upstream middleware's AsyncLocalStorage must still be there when next runs
668
+ next = AsyncResource.bind(next);
669
+
670
+ // with a known content-length and nothing to decompress, uWS can collect the whole
671
+ // body in native code: one callback instead of one per chunk, the limit enforced
672
+ // before any byte reaches JS, and no copy at all - the parsers turn the bytes into
673
+ // req.body before the callback returns, so a view over uWS's own memory is enough
674
+ if (!req.receivedData && !inflate && !isNaN(length) && Number(length) > 0 && req._res.collectBody) {
675
+ req.bodyRead = true;
676
+ const declared = Number(length);
677
+ req._res.collectBody(options.limit, (body) => {
678
+ if (body === null) {
679
+ // over maxSize: uWS refused it natively
680
+ return next(
681
+ bodyError("request entity too large", 413, "entity.too.large", {
682
+ limit: options.limit,
683
+ received: options.limit
684
+ })
685
+ );
686
+ }
687
+ if (body.byteLength !== declared) {
688
+ return next(
689
+ bodyError("request size did not match content length", 400, "request.size.invalid", {
690
+ expected: declared,
691
+ length: declared,
692
+ received: body.byteLength
693
+ })
694
+ );
695
+ }
696
+ let buf = Buffer.from(body);
697
+ if (copyBody) {
698
+ buf = Buffer.from(buf);
699
+ }
700
+ if (!runVerify(req, res, next, options, buf)) {
701
+ return;
702
+ }
703
+ beforeReturn(req, res, next, options, buf, encoding);
704
+ });
705
+ return;
706
+ }
707
+
645
708
  // uWS neuters its ArrayBuffer after the callback, so every chunk has to be copied out of
646
709
  // it - and then Buffer.concat copied the whole body a second time. when content-length is
647
710
  // known and we aren't inflating, the final size is known up front, so chunks can go
@@ -813,6 +876,14 @@ function createBodyParser(defaultType, beforeReturn, checkOptions, charsetPolicy
813
876
  req.on("end", onEnd);
814
877
  }
815
878
  };
879
+ // A GET without a declared body leaves this middleware through the synchronous
880
+ // no-body exit before anything type- or charset-shaped is read, and a GET that does
881
+ // declare one takes the full header copy in the constructor, so the header-skip
882
+ // analysis may trust it. A type function sees the request itself, so it may not.
883
+ if (typeof options.type !== "function") {
884
+ parserMiddleware[kGetSafe] = true;
885
+ }
886
+ return parserMiddleware;
816
887
  };
817
888
  }
818
889
 
@@ -860,10 +931,17 @@ const json = createBodyParser(
860
931
  "utf"
861
932
  );
862
933
 
863
- const raw = createBodyParser("application/octet-stream", function (req, res, next, options, buf) {
864
- req.body = buf;
865
- next();
866
- });
934
+ const raw = createBodyParser(
935
+ "application/octet-stream",
936
+ function (req, res, next, options, buf) {
937
+ req.body = buf;
938
+ next();
939
+ },
940
+ undefined,
941
+ undefined,
942
+ // req.body is the collected buffer itself, so it must not be a view over uWS memory
943
+ true
944
+ );
867
945
 
868
946
  const text = createBodyParser(
869
947
  "text/plain",
package/src/request.js CHANGED
@@ -200,16 +200,61 @@ module.exports = class Request extends Readable {
200
200
  * @param {any} [preset] a literal native registration's constants: µWS matched the URL byte
201
201
  * for byte against that exact pattern and dispatched by method, so path, method and what
202
202
  * derives from them are known without asking
203
+ * @param {any} [skipHolder] where a granted header skip lives: the preset itself for a
204
+ * literal registration, a holder of its own for a parameterised one
203
205
  */
204
- constructor(req, res, app, preset) {
206
+ constructor(req, res, app, preset, skipHolder) {
205
207
  // the same object every time: Readable reads these options and never writes to them
206
208
  super(READABLE_OPTIONS);
207
209
  this._res = res;
208
210
  this._req = req;
209
211
  this.readable = true;
210
- currentRequest = this;
211
- this._req.forEach(Request.#collectHeader);
212
- currentRequest = null;
212
+ if (skipHolder !== undefined && skipHolder.skipHeaders) {
213
+ // The chain behind this registration provably never reads a header, so instead of
214
+ // copying them all out of uWS the constructor asks for the four that steer the
215
+ // framework itself: body framing, keep-alive, and accept for the error page a
216
+ // throw could still need. A GET that does declare a body is the rare case, and
217
+ // the parsers and the stream want the whole picture, so it takes the full copy.
218
+ const length = req.getHeader("content-length");
219
+ const transferEncoding = req.getHeader("transfer-encoding");
220
+ if ((length !== "" && length !== "0") || transferEncoding !== "") {
221
+ currentRequest = this;
222
+ this._req.forEach(Request.#collectHeader);
223
+ currentRequest = null;
224
+ } else {
225
+ const entries = this.#rawHeadersEntries;
226
+ const connection = req.getHeader("connection");
227
+ if (connection !== "") {
228
+ entries.push("connection", connection);
229
+ if (connection.length === 5 && connection.toLowerCase() === "close") {
230
+ this._connectionClose = true;
231
+ }
232
+ }
233
+ const accept = req.getHeader("accept");
234
+ if (accept !== "") {
235
+ entries.push("accept", accept);
236
+ }
237
+ // send consults freshness whatever the etag setting: if-none-match can be "*"
238
+ // and a handler may set a validator by hand, so the conditional trio has to be
239
+ // really absent rather than merely uncopied
240
+ const ifNoneMatch = req.getHeader("if-none-match");
241
+ if (ifNoneMatch !== "") {
242
+ entries.push("if-none-match", ifNoneMatch);
243
+ }
244
+ const ifModifiedSince = req.getHeader("if-modified-since");
245
+ if (ifModifiedSince !== "") {
246
+ entries.push("if-modified-since", ifModifiedSince);
247
+ }
248
+ const cacheControl = req.getHeader("cache-control");
249
+ if (cacheControl !== "") {
250
+ entries.push("cache-control", cacheControl);
251
+ }
252
+ }
253
+ } else {
254
+ currentRequest = this;
255
+ this._req.forEach(Request.#collectHeader);
256
+ currentRequest = null;
257
+ }
213
258
  this.routeCount = 1;
214
259
  this.key = key++;
215
260
  if (key > 100000) {
@@ -324,6 +369,28 @@ module.exports = class Request extends Readable {
324
369
  });
325
370
  }
326
371
 
372
+ /**
373
+ * One header by its lowercase wire name, straight from the raw entries. The body parsers ask
374
+ * for three of these per request, and materializing the whole headers object for that costs
375
+ * more than all three scans together. Reads the built object instead when it already exists,
376
+ * so joined duplicates come out the same either way.
377
+ *
378
+ * @param {string} name lowercase
379
+ * @returns {string|undefined}
380
+ */
381
+ _rawHeader(name) {
382
+ if (this.#cachedHeaders !== null) {
383
+ return this.#cachedHeaders[name];
384
+ }
385
+ const entries = this.#rawHeadersEntries;
386
+ for (let i = 0, len = entries.length; i < len; i += 2) {
387
+ if (entries[i] === name) {
388
+ return entries[i + 1];
389
+ }
390
+ }
391
+ return undefined;
392
+ }
393
+
327
394
  /**
328
395
  * Whether there is any point still reading the body: once the response is finished or the
329
396
  * connection is gone, uWS has nothing left to hand over.
package/src/router.js CHANGED
@@ -36,6 +36,7 @@ const compileDeclarative = require("./declarative.js");
36
36
  const statuses = require("statuses");
37
37
  const { METHODS } = require("http");
38
38
  const { isNodeRequest, serveNodeRequest } = require("./node-shim.js");
39
+ const { chainSkipsHeaders } = require("./usage.js");
39
40
 
40
41
  // every method the declarative compiler can emit: a patched one must disable compilation, or the
41
42
  // patch would be honoured everywhere but on compiled routes
@@ -691,10 +692,38 @@ function nativePreset(path, method, strict) {
691
692
  endsWithSlash,
692
693
  opPath: endsWithSlash && path !== "/" && !strict ? path.slice(0, -1) : path,
693
694
  isOptions: method === "OPTIONS",
694
- isHead: method === "HEAD"
695
+ isHead: method === "HEAD",
696
+ // set at registration when the whole chain provably never reads a header; mutable,
697
+ // because a middleware added after listen has to be able to take it back
698
+ skipHeaders: false
695
699
  };
696
700
  }
697
701
 
702
+ /**
703
+ * Whether any error middleware exists anywhere under this router, mounted routers and sub-apps
704
+ * included. The header-skip analysis needs the answer to be no: a throw inside an analyzed
705
+ * handler would hand the request to code nobody analyzed.
706
+ *
707
+ * @param {any} router
708
+ * @returns {boolean}
709
+ */
710
+ function hasErrorMiddleware(router) {
711
+ for (const route of router._routes) {
712
+ for (const callback of route.callbacks) {
713
+ // a mounted router or a callable sub-app carries routes of its own; the callable
714
+ // app is also a function, so the routes are looked for first
715
+ if (callback && callback._routes) {
716
+ if (hasErrorMiddleware(callback)) {
717
+ return true;
718
+ }
719
+ } else if (typeof callback === "function" && callback.length >= 4) {
720
+ return true;
721
+ }
722
+ }
723
+ }
724
+ return false;
725
+ }
726
+
698
727
  /**
699
728
  *
700
729
  */
@@ -822,6 +851,12 @@ module.exports = class Router extends EventEmitter {
822
851
  this._paramCallbacks = new Map();
823
852
  this._mountpathCache = new Map();
824
853
  this._routes = [];
854
+ // the native presets allowed to skip the header copy, so a late middleware or an etag
855
+ // arriving after listen can take the permission back; null until one is granted
856
+ /** @type {Set<any>|null} */
857
+ this._skipPresets = null;
858
+ /** @type {boolean|undefined} */
859
+ this._hasErrMwCache = undefined;
825
860
  // an array when mounted on several paths at once, as Express allows
826
861
  /** @type {string|string[]} */
827
862
  this.mountpath = "/";
@@ -1107,6 +1142,16 @@ module.exports = class Router extends EventEmitter {
1107
1142
  }
1108
1143
  this._routes.push(...routes);
1109
1144
 
1145
+ // anything registered after listen invalidates what the header-skip analysis proved:
1146
+ // it could catch a throw or read what a chain never did, so every skip is taken back
1147
+ this._hasErrMwCache = undefined;
1148
+ if (this._skipPresets?.size) {
1149
+ for (const preset of this._skipPresets) {
1150
+ preset.skipHeaders = false;
1151
+ }
1152
+ this._skipPresets.clear();
1153
+ }
1154
+
1110
1155
  return parent;
1111
1156
  }
1112
1157
 
@@ -1338,10 +1383,12 @@ module.exports = class Router extends EventEmitter {
1338
1383
  * @param {any} res uWS response
1339
1384
  * @param {any} req uWS request, readable only during this call
1340
1385
  * @param {any} [preset] a literal registration's constants, see nativePreset
1386
+ * @param {any} [skipHolder] the object a granted header skip lives on: the preset itself
1387
+ * for a literal registration, a holder of its own for a parameterised one
1341
1388
  * @returns {any} the request, with the response reachable as request.res
1342
1389
  */
1343
- handleRequest(res, req, preset) {
1344
- const request = new this._request(req, res, this, preset);
1390
+ handleRequest(res, req, preset, skipHolder) {
1391
+ const request = new this._request(req, res, this, preset, skipHolder);
1345
1392
  const response = new this._response(res, request, this);
1346
1393
  request.res = response;
1347
1394
  response.req = request;
@@ -1417,7 +1464,15 @@ module.exports = class Router extends EventEmitter {
1417
1464
  if (route.path.includes(":")) {
1418
1465
  route.optimizedParams = route.path.match(regExParam).map((p) => p.slice(1));
1419
1466
  }
1420
- const makeHandler = (chain, preset) => {
1467
+ const makeHandler = (chain, preset, skips) => {
1468
+ // the mutable object a granted skip lives on, so a middleware arriving after
1469
+ // listen can take it back: a literal registration's preset doubles as it, and a
1470
+ // parameterised one, which has no preset, gets a holder of its own
1471
+ let skipHolder = preset;
1472
+ if (skipHolder === undefined && skips) {
1473
+ skipHolder = { skipHeaders: true };
1474
+ (this._skipPresets ??= new Set()).add(skipHolder);
1475
+ }
1421
1476
  // all three are registration-time constants: computing them in the handler was a
1422
1477
  // closure and a scan of the chain on every native request.
1423
1478
  // Falling back resumes after the mount, not after the router's leaf: the leaf can have
@@ -1430,7 +1485,7 @@ module.exports = class Router extends EventEmitter {
1430
1485
  // and this one never did. nativeDone and nativeFail defer their epilogues to a
1431
1486
  // microtask, which is where the await used to resume, so the visible order holds
1432
1487
  return (res, req) => {
1433
- const request = this.handleRequest(res, req, preset);
1488
+ const request = this.handleRequest(res, req, preset, skipHolder);
1434
1489
  const response = request.res;
1435
1490
  if (optimizedParams) {
1436
1491
  request.optimizedParams = new NullObject();
@@ -1461,6 +1516,7 @@ module.exports = class Router extends EventEmitter {
1461
1516
  const getChain =
1462
1517
  route.method === "GET" ? optimizedPath.filter((r) => r.all || r.method !== "HEAD") : optimizedPath;
1463
1518
  route.optimizedPath = optimizedPath;
1519
+ const headChain = getChain.length === optimizedPath.length ? getChain : optimizedPath;
1464
1520
 
1465
1521
  // A fully literal registration knows path and method here, so each registration site
1466
1522
  // hands the request constructor its own constants. An "any" registration serves every
@@ -1471,11 +1527,45 @@ module.exports = class Router extends EventEmitter {
1471
1527
  // registering that path here is the only way it could
1472
1528
  const strictHere = Boolean((route.owner ?? this).get("strict routing"));
1473
1529
 
1474
- let fn = makeHandler(getChain, canPreset ? nativePreset(route.path, route.method, strictHere) : undefined);
1530
+ // Whether requests served by this registration may skip the header copy: GET and its
1531
+ // HEAD twins only, the app must not compute etags (send would consult freshness
1532
+ // headers), no error middleware may exist anywhere (a throw hands the request to code
1533
+ // the analysis never saw), and every callback in the chain has to pass the source
1534
+ // analysis in usage.js, whose default answer is no.
1535
+ let getSkips = false;
1536
+ let headSkips = false;
1537
+ if (route.method === "GET" && this.get("etag") === false) {
1538
+ let hasErr = this._hasErrMwCache;
1539
+ if (hasErr === undefined) {
1540
+ hasErr = this._hasErrMwCache = hasErrorMiddleware(this);
1541
+ }
1542
+ if (!hasErr) {
1543
+ // a terminal next() may only fall into the framework's own 404, so no later
1544
+ // route may be able to catch the same path
1545
+ const owner = route.owner ?? this;
1546
+ const noLaterMatch = !owner._isFollowedByAnOverlap.call(owner, route, owner._routes);
1547
+ getSkips = chainSkipsHeaders(getChain, noLaterMatch);
1548
+ headSkips = headChain === getChain ? getSkips : chainSkipsHeaders(headChain, noLaterMatch);
1549
+ }
1550
+ }
1551
+ // remembered so a middleware or setting arriving after listen can take the skips back
1552
+ const makePreset = (path, method, skips) => {
1553
+ const preset = nativePreset(path, method, strictHere);
1554
+ if (skips) {
1555
+ preset.skipHeaders = true;
1556
+ (this._skipPresets ??= new Set()).add(preset);
1557
+ }
1558
+ return preset;
1559
+ };
1560
+
1561
+ let fn = makeHandler(
1562
+ getChain,
1563
+ canPreset ? makePreset(route.path, route.method, getSkips) : undefined,
1564
+ getSkips
1565
+ );
1475
1566
  const jsFn = fn;
1476
1567
 
1477
1568
  let replacedPath = route.path;
1478
- const headChain = getChain.length === optimizedPath.length ? getChain : optimizedPath;
1479
1569
 
1480
1570
  // the response prototype the route will really run under: its own app's, which sees a
1481
1571
  // method patched there or inherited from a parent app, falling back to the registering app
@@ -1505,13 +1595,17 @@ module.exports = class Router extends EventEmitter {
1505
1595
  fn !== jsFn
1506
1596
  ? fn
1507
1597
  : canPreset
1508
- ? makeHandler(getChain, nativePreset(route.path + "/", route.method, strictHere))
1598
+ ? makeHandler(getChain, makePreset(route.path + "/", route.method, getSkips), getSkips)
1509
1599
  : fn;
1510
1600
  this.uwsApp[method](replacedPath + "/", slashFn);
1511
1601
  if (method === "get") {
1512
1602
  this.uwsApp.head(
1513
1603
  replacedPath + "/",
1514
- makeHandler(headChain, canPreset ? nativePreset(route.path + "/", "HEAD", strictHere) : undefined)
1604
+ makeHandler(
1605
+ headChain,
1606
+ canPreset ? makePreset(route.path + "/", "HEAD", headSkips) : undefined,
1607
+ headSkips
1608
+ )
1515
1609
  );
1516
1610
  }
1517
1611
  }
@@ -1519,7 +1613,7 @@ module.exports = class Router extends EventEmitter {
1519
1613
  // its own handler always: the shared one would carry the GET registration's method
1520
1614
  this.uwsApp.head(
1521
1615
  replacedPath,
1522
- makeHandler(headChain, canPreset ? nativePreset(route.path, "HEAD", strictHere) : undefined)
1616
+ makeHandler(headChain, canPreset ? makePreset(route.path, "HEAD", headSkips) : undefined, headSkips)
1523
1617
  );
1524
1618
  }
1525
1619
  }
package/src/usage.js ADDED
@@ -0,0 +1,247 @@
1
+ "use strict";
2
+
3
+ const acorn = require("acorn");
4
+
5
+ // Marks a middleware the analysis may trust on a GET request without reading its source: the
6
+ // body parsers set it, whose prologue only reads body-framing headers and leaves a bodyless
7
+ // GET alone (and a GET that declares a body falls back to the full header copy).
8
+ const kGetSafe = Symbol("fulmine.getSafe");
9
+
10
+ // What a handler may do with `req` and still let the header copy be skipped: members whose
11
+ // reads never reach a header. Anything else, computed access included, keeps the copy.
12
+ // req.res and req.app are deliberately absent: the walk judges only the member directly on
13
+ // the parameter, so anything that can reach another object could reach headers through it.
14
+ const REQ_OK = new Set(["query", "params", "body", "method", "path", "url", "baseUrl", "originalUrl", "route"]);
15
+
16
+ // What a handler may do with `res`: writing the response. Anything that negotiates against
17
+ // request headers (format, redirect, sendFile, jsonp) is deliberately absent.
18
+ const RES_OK = new Set([
19
+ "json",
20
+ "send",
21
+ "end",
22
+ "status",
23
+ "sendStatus",
24
+ "set",
25
+ "setHeader",
26
+ "header",
27
+ "get",
28
+ "type",
29
+ "contentType",
30
+ "append",
31
+ "vary",
32
+ "links",
33
+ "locals",
34
+ "statusCode",
35
+ "writeHead",
36
+ "headersSent",
37
+ "finished",
38
+ "cork"
39
+ ]);
40
+
41
+ // what the analysis can say about one callback
42
+ const NO = 0; // could read headers, or could not be read at all
43
+ const SAFE = 1; // never reads a header, never touches next
44
+ const SAFE_NEXT = 2; // never reads a header, calls next: fine mid-chain, and at the end of
45
+ // the chain only when no later route could catch the fall-through
46
+
47
+ const verdicts = new WeakMap();
48
+
49
+ /**
50
+ * What one callback provably does. The default is NO: any shape this walk does not understand
51
+ * and any alias of req, res or next keeps the header copy. That inversion is what makes
52
+ * source analysis sound to act on.
53
+ *
54
+ * @param {Function} fn
55
+ * @returns {number} NO, SAFE or SAFE_NEXT
56
+ */
57
+ function callbackSkipsHeaders(fn) {
58
+ if (fn[kGetSafe]) {
59
+ return SAFE_NEXT;
60
+ }
61
+ let verdict = verdicts.get(fn);
62
+ if (verdict === undefined) {
63
+ verdict = analyze(fn);
64
+ verdicts.set(fn, verdict);
65
+ }
66
+ return verdict;
67
+ }
68
+
69
+ /** @param {Function} fn @returns {number} */
70
+ function analyze(fn) {
71
+ let code = fn.toString();
72
+ if (code.startsWith("function") || code.startsWith("async function")) {
73
+ code = code.replace(/function *\(/, "function __cb(");
74
+ }
75
+ let tree;
76
+ try {
77
+ tree = acorn.parse(code, { ecmaVersion: "latest" });
78
+ } catch {
79
+ // class methods and native functions do not parse alone, and unread code is unknown code
80
+ return NO;
81
+ }
82
+ let root = /** @type {any} */ (tree.body[0]);
83
+ if (!root) {
84
+ return NO;
85
+ }
86
+ if (root.type === "ExpressionStatement") {
87
+ root = root.expression;
88
+ }
89
+ if (
90
+ root.type !== "FunctionDeclaration" &&
91
+ root.type !== "ArrowFunctionExpression" &&
92
+ root.type !== "FunctionExpression"
93
+ ) {
94
+ return NO;
95
+ }
96
+
97
+ const params = /** @type {any[]} */ (root.params);
98
+ // rest or destructured parameters alias the objects somewhere the walk cannot follow
99
+ for (const p of params) {
100
+ if (p.type !== "Identifier") {
101
+ return NO;
102
+ }
103
+ }
104
+ const reqName = params[0] ? params[0].name : null;
105
+ const resName = params[1] ? params[1].name : null;
106
+ const nextName = params[2] ? params[2].name : null;
107
+
108
+ // Every appearance of the three names in the whole body is judged, nested functions
109
+ // included: an inner binding that shadows one of them only makes this stricter, never
110
+ // looser, so scope tracking is not needed for soundness.
111
+ let ok = true;
112
+ let usesNext = false;
113
+ walk(root.body, null, (node, parent) => {
114
+ if (!ok || node.type !== "Identifier") {
115
+ return;
116
+ }
117
+ const name = node.name;
118
+ if (name === "eval" || name === "arguments") {
119
+ ok = false;
120
+ return;
121
+ }
122
+ if (name !== reqName && name !== resName && name !== nextName) {
123
+ return;
124
+ }
125
+ // being renamed inside a member expression (req.query's `query`) is not a use
126
+ if (parent && parent.type === "MemberExpression" && parent.property === node && !parent.computed) {
127
+ return;
128
+ }
129
+ // a redeclaration as an inner parameter or variable name is not a use either
130
+ if (
131
+ parent &&
132
+ (((parent.type === "FunctionDeclaration" ||
133
+ parent.type === "FunctionExpression" ||
134
+ parent.type === "ArrowFunctionExpression") &&
135
+ parent.params.includes(node)) ||
136
+ (parent.type === "VariableDeclarator" && parent.id === node) ||
137
+ (parent.type === "Property" && parent.key === node && !parent.computed))
138
+ ) {
139
+ return;
140
+ }
141
+ if (name === nextName) {
142
+ // calling next is how a chain advances, and past its end or with an error the
143
+ // request lands in the framework's own final handler, which the constructor's
144
+ // accept pre-read covers. Anything but a direct call aliases the continuation,
145
+ // and an argument that could be the string "route" would leave the chain for
146
+ // routes nobody analyzed, so only shapes that cannot be a string pass.
147
+ if (!parent || parent.type !== "CallExpression" || parent.callee !== node) {
148
+ ok = false;
149
+ return;
150
+ }
151
+ usesNext = true;
152
+ const args = parent.arguments;
153
+ if (args.length === 0) {
154
+ return;
155
+ }
156
+ const arg = args[0];
157
+ if (
158
+ args.length > 1 ||
159
+ (arg.type !== "NewExpression" &&
160
+ arg.type !== "ObjectExpression" &&
161
+ !(arg.type === "Literal" && typeof arg.value !== "string"))
162
+ ) {
163
+ ok = false;
164
+ }
165
+ return;
166
+ }
167
+ if (!parent || parent.type !== "MemberExpression" || parent.object !== node || parent.computed) {
168
+ ok = false;
169
+ return;
170
+ }
171
+ const member = parent.property.name;
172
+ if (name === reqName ? !REQ_OK.has(member) : !RES_OK.has(member)) {
173
+ ok = false;
174
+ }
175
+ });
176
+ return ok ? (usesNext ? SAFE_NEXT : SAFE) : NO;
177
+ }
178
+
179
+ /**
180
+ * Walks every node, handing each its parent. Arrays and nested objects are entered, nothing
181
+ * is interpreted: the judging happens in the visitor.
182
+ *
183
+ * @param {any} node
184
+ * @param {any} parent
185
+ * @param {(node: any, parent: any) => void} visit
186
+ */
187
+ function walk(node, parent, visit) {
188
+ if (!node || typeof node.type !== "string") {
189
+ return;
190
+ }
191
+ visit(node, parent);
192
+ for (const key in node) {
193
+ if (key === "type" || key === "start" || key === "end") {
194
+ continue;
195
+ }
196
+ const value = node[key];
197
+ if (Array.isArray(value)) {
198
+ for (const item of value) {
199
+ if (item && typeof item.type === "string") {
200
+ walk(item, node, visit);
201
+ }
202
+ }
203
+ } else if (value && typeof value.type === "string") {
204
+ walk(value, node, visit);
205
+ }
206
+ }
207
+ }
208
+
209
+ /**
210
+ * Whether a native literal route's whole chain provably never reads a request header, so the
211
+ * request constructor may skip copying them and read the few framing headers directly.
212
+ *
213
+ * A callback that calls next passes anywhere but in the terminal route, where next would fall
214
+ * out of the chain: there it only passes when the caller established that no later route
215
+ * could catch the fall-through.
216
+ *
217
+ * @param {any[]} chain the routes the native handler runs, in order, this route last
218
+ * @param {boolean} allowTerminalNext whether a fall-through past the chain lands only in the
219
+ * framework's own final handler
220
+ * @returns {boolean}
221
+ */
222
+ function chainSkipsHeaders(chain, allowTerminalNext) {
223
+ for (let i = 0; i < chain.length; i++) {
224
+ const entry = chain[i];
225
+ const callbacks = entry.callbacks;
226
+ if (!Array.isArray(callbacks)) {
227
+ return false;
228
+ }
229
+ const terminal = i === chain.length - 1;
230
+ for (const cb of callbacks) {
231
+ if (typeof cb !== "function") {
232
+ return false;
233
+ }
234
+ const verdict = callbackSkipsHeaders(cb);
235
+ if (verdict === NO || (verdict === SAFE_NEXT && terminal && !allowTerminalNext)) {
236
+ return false;
237
+ }
238
+ }
239
+ // a param callback runs code this walk never saw
240
+ if (entry.paramCallbacks && entry.paramCallbacks.size > 0) {
241
+ return false;
242
+ }
243
+ }
244
+ return true;
245
+ }
246
+
247
+ module.exports = { chainSkipsHeaders, callbackSkipsHeaders, kGetSafe };