diff --git a/src/stream_base.cc b/src/stream_base.cc
index bbb95bf86158ee..bb46ea1febbca0 100644
--- a/src/stream_base.cc
+++ b/src/stream_base.cc
@@ -374,8 +374,9 @@ void EmitToJSStreamListener::OnStreamRead(ssize_t nread, const uv_buf_t& buf) {
   }
 
   CHECK_LE(static_cast<size_t>(nread), buf.len);
+  char* base = Realloc(buf.base, nread);
 
-  Local<Object> obj = Buffer::New(env, buf.base, nread).ToLocalChecked();
+  Local<Object> obj = Buffer::New(env, base, nread).ToLocalChecked();
   stream->CallJSOnreadMethod(nread, obj);
 }
 
diff --git a/test/sequential/test-net-bytes-per-incoming-chunk-overhead.js b/test/sequential/test-net-bytes-per-incoming-chunk-overhead.js
new file mode 100644
index 00000000000000..8f766e8c7a4106
--- /dev/null
+++ b/test/sequential/test-net-bytes-per-incoming-chunk-overhead.js
@@ -0,0 +1,41 @@
+// Flags: --expose-gc
+'use strict';
+
+const common = require('../common');
+const assert = require('assert');
+const net = require('net');
+
+// Tests that, when receiving small chunks, we do not keep the full length
+// of the original allocation for the libuv read call in memory.
+
+let client;
+let baseRSS;
+const receivedChunks = [];
+const N = 250000;
+
+const server = net.createServer(common.mustCall((socket) => {
+  baseRSS = process.memoryUsage().rss;
+
+  socket.setNoDelay(true);
+  socket.on('data', (chunk) => {
+    receivedChunks.push(chunk);
+    if (receivedChunks.length < N) {
+      client.write('a');
+    } else {
+      client.end();
+      server.close();
+    }
+  });
+})).listen(0, common.mustCall(() => {
+  client = net.connect(server.address().port);
+  client.setNoDelay(true);
+  client.write('hello!');
+}));
+
+process.on('exit', () => {
+  global.gc();
+  const bytesPerChunk =
+    (process.memoryUsage().rss - baseRSS) / receivedChunks.length;
+  // We should always have less than one page (usually ~ 4 kB) per chunk.
+  assert(bytesPerChunk < 512, `measured ${bytesPerChunk} bytes per chunk`);
+});