zlib: do not coalesce multiple .flush() calls

This is an approach to address the issue linked below. Previously,
when `.write()` and `.flush()` calls to a zlib stream were interleaved
synchronously (i.e. without waiting for these operations to finish),
multiple flush calls would have been coalesced into a single flushing
operation.

This patch changes behaviour so that each `.flush()` all corresponds
to one flushing operation on the underlying zlib resource, and the
order of operations is as if the `.flush()` call were a `.write()`
call.

One test had to be removed because it specifically tested the previous
behaviour.

As a drive-by fix, this also makes sure that all flush callbacks are
called. Previously, that was not the case.

Fixes: #28478

PR-URL: #28520
Reviewed-By: Rich Trott <rtrott@gmail.com>
Reviewed-By: Ruben Bridgewater <ruben@bridgewater.de>
Reviewed-By: Luigi Pinca <luigipinca@gmail.com>

lib/zlib.js

@@ -49,6 +49,8 @@ const {
 } = require('buffer');
 const { owner_symbol } = require('internal/async_hooks').symbols;
 
+const kFlushFlag = Symbol('kFlushFlag');
+
 const constants = internalBinding('constants').zlib;
 const {
   // Zlib flush levels
@@ -261,7 +263,6 @@ function ZlibBase(opts, mode, handle, { flush, finishFlush, fullFlush }) {
   this._chunkSize = chunkSize;
   this._defaultFlushFlag = flush;
   this._finishFlushFlag = finishFlush;
-  this._nextFlush = -1;
   this._defaultFullFlushFlag = fullFlush;
   this.once('end', this.close);
   this._info = opts && opts.info;
@@ -308,21 +309,35 @@ ZlibBase.prototype._flush = function(callback) {
 
 // If a flush is scheduled while another flush is still pending, a way to figure
 // out which one is the "stronger" flush is needed.
+// This is currently only used to figure out which flush flag to use for the
+// last chunk.
 // Roughly, the following holds:
 // Z_NO_FLUSH (< Z_TREES) < Z_BLOCK < Z_PARTIAL_FLUSH <
 //     Z_SYNC_FLUSH < Z_FULL_FLUSH < Z_FINISH
 const flushiness = [];
 let i = 0;
-for (const flushFlag of [Z_NO_FLUSH, Z_BLOCK, Z_PARTIAL_FLUSH,
-                         Z_SYNC_FLUSH, Z_FULL_FLUSH, Z_FINISH]) {
+const kFlushFlagList = [Z_NO_FLUSH, Z_BLOCK, Z_PARTIAL_FLUSH,
+                        Z_SYNC_FLUSH, Z_FULL_FLUSH, Z_FINISH];
+for (const flushFlag of kFlushFlagList) {
   flushiness[flushFlag] = i++;
 }
 
 function maxFlush(a, b) {
   return flushiness[a] > flushiness[b] ? a : b;
 }
 
-const flushBuffer = Buffer.alloc(0);
+// Set up a list of 'special' buffers that can be written using .write()
+// from the .flush() code as a way of introducing flushing operations into the
+// write sequence.
+const kFlushBuffers = [];
+{
+  const dummyArrayBuffer = new ArrayBuffer();
+  for (const flushFlag of kFlushFlagList) {
+    kFlushBuffers[flushFlag] = Buffer.from(dummyArrayBuffer);
+    kFlushBuffers[flushFlag][kFlushFlag] = flushFlag;
+  }
+}
+
 ZlibBase.prototype.flush = function(kind, callback) {
   const ws = this._writableState;
 
@@ -337,13 +352,8 @@ ZlibBase.prototype.flush = function(kind, callback) {
   } else if (ws.ending) {
     if (callback)
       this.once('end', callback);
-  } else if (this._nextFlush !== -1) {
-    // This means that there is a flush currently in the write queue.
-    // We currently coalesce this flush into the pending one.
-    this._nextFlush = maxFlush(this._nextFlush, kind);
   } else {
-    this._nextFlush = kind;
-    this.write(flushBuffer, '', callback);
+    this.write(kFlushBuffers[kind], '', callback);
   }
 };
 
@@ -361,9 +371,8 @@ ZlibBase.prototype._transform = function(chunk, encoding, cb) {
   var flushFlag = this._defaultFlushFlag;
   // We use a 'fake' zero-length chunk to carry information about flushes from
   // the public API to the actual stream implementation.
-  if (chunk === flushBuffer) {
-    flushFlag = this._nextFlush;
-    this._nextFlush = -1;
+  if (typeof chunk[kFlushFlag] === 'number') {
+    flushFlag = chunk[kFlushFlag];
   }
 
   // For the last chunk, also apply `_finishFlushFlag`.

test/parallel/test-zlib-flush-write-sync-interleaved.js

@@ -0,0 +1,57 @@
+'use strict';
+const common = require('../common');
+const assert = require('assert');
+const { createGzip, createGunzip, Z_PARTIAL_FLUSH } = require('zlib');
+
+// Verify that .flush() behaves like .write() in terms of ordering, e.g. in
+// a sequence like .write() + .flush() + .write() + .flush() each .flush() call
+// only affects the data written before it.
+// Refs: https://github.com/nodejs/node/issues/28478
+
+const compress = createGzip();
+const decompress = createGunzip();
+decompress.setEncoding('utf8');
+
+const events = [];
+const compressedChunks = [];
+
+for (const chunk of ['abc', 'def', 'ghi']) {
+  compress.write(chunk, common.mustCall(() => events.push({ written: chunk })));
+  compress.flush(Z_PARTIAL_FLUSH, common.mustCall(() => {
+    events.push('flushed');
+    const chunk = compress.read();
+    if (chunk !== null)
+      compressedChunks.push(chunk);
+  }));
+}
+
+compress.end(common.mustCall(() => {
+  events.push('compress end');
+  writeToDecompress();
+}));
+
+function writeToDecompress() {
+  // Write the compressed chunks to a decompressor, one by one, in order to
+  // verify that the flushes actually worked.
+  const chunk = compressedChunks.shift();
+  if (chunk === undefined) return decompress.end();
+  decompress.write(chunk, common.mustCall(() => {
+    events.push({ read: decompress.read() });
+    writeToDecompress();
+  }));
+}
+
+process.on('exit', () => {
+  assert.deepStrictEqual(events, [
+    { written: 'abc' },
+    'flushed',
+    { written: 'def' },
+    'flushed',
+    { written: 'ghi' },
+    'flushed',
+    'compress end',
+    { read: 'abc' },
+    { read: 'def' },
+    { read: 'ghi' }
+  ]);
+});

test/parallel/test-zlib-write-after-flush.js

@@ -39,7 +39,6 @@ for (const [ createCompress, createDecompress ] of [
   gunz.on('data', (c) => output += c);
   gunz.on('end', common.mustCall(() => {
     assert.strictEqual(output, input);
-    assert.strictEqual(gzip._nextFlush, -1);
   }));
 
   // Make sure that flush/write doesn't trigger an assert failure