perf(websocket): unroll client frame masking

lib/web/websocket/frame.js

@@ -20,6 +20,31 @@ function generateMask () {
   return [buffer[bufIdx++], buffer[bufIdx++], buffer[bufIdx++], buffer[bufIdx++]]
 }
 
+function maskPayload (source, target, targetOffset, maskKey) {
+  const length = source.byteLength
+  const mask0 = maskKey[0]
+  const mask1 = maskKey[1]
+  const mask2 = maskKey[2]
+  const mask3 = maskKey[3]
+  const unrolledEnd = length & ~7
+
+  let i = 0
+  for (; i < unrolledEnd; i += 8) {
+    target[targetOffset + i] = source[i] ^ mask0
+    target[targetOffset + i + 1] = source[i + 1] ^ mask1
+    target[targetOffset + i + 2] = source[i + 2] ^ mask2
+    target[targetOffset + i + 3] = source[i + 3] ^ mask3
+    target[targetOffset + i + 4] = source[i + 4] ^ mask0
+    target[targetOffset + i + 5] = source[i + 5] ^ mask1
+    target[targetOffset + i + 6] = source[i + 6] ^ mask2
+    target[targetOffset + i + 7] = source[i + 7] ^ mask3
+  }
+
+  for (; i < length; ++i) {
+    target[targetOffset + i] = source[i] ^ maskKey[i & 3]
+  }
+}
+
 class WebsocketFrameSend {
   /**
    * @param {Buffer|undefined} data
@@ -71,8 +96,8 @@ class WebsocketFrameSend {
     buffer[1] |= 0x80 // MASK
 
     // mask body
-    for (let i = 0; i < bodyLength; ++i) {
-      buffer[offset + i] = frameData[i] ^ maskKey[i & 3]
+    if (bodyLength !== 0) {
+      maskPayload(frameData, buffer, offset, maskKey)
     }
 
     return buffer
@@ -87,9 +112,7 @@ class WebsocketFrameSend {
     const bodyLength = buffer.length
 
     // mask body
-    for (let i = 0; i < bodyLength; ++i) {
-      buffer[i] ^= maskKey[i & 3]
-    }
+    maskPayload(buffer, buffer, 0, maskKey)
 
     let payloadLength = bodyLength
     let offset = 6

test/websocket/frame.js

@@ -37,3 +37,29 @@ test('Writing 16-bit frame length value at correct offset when buffer has a non-
   t.assert.strictEqual(frame[3], payloadLength & 0xff)
   t.assert.strictEqual(smallBuffer.length, 1) // ensure smallBuffer can't be garbage-collected too soon
 })
+
+test('Masks aligned and unaligned payloads without changing frame semantics', (t) => {
+  for (const length of [0, 1, 3, 4, 5, 7, 8, 9, 125, 126, 127, 65535, 65536]) {
+    const payload = Buffer.alloc(length)
+    for (let i = 0; i < payload.length; ++i) payload[i] = i & 0xff
+    const originalPayload = Buffer.from(payload)
+    const frame = new WebsocketFrameSend(payload).createFrame(opcodes.BINARY)
+    const offset = length > 65535 ? 14 : length > 125 ? 8 : 6
+    const mask = frame.subarray(offset - 4, offset)
+    const decoded = Buffer.alloc(length)
+    for (let i = 0; i < length; ++i) decoded[i] = frame[offset + i] ^ mask[i & 3]
+    t.assert.deepStrictEqual(decoded, originalPayload)
+    t.assert.deepStrictEqual(payload, originalPayload)
+  }
+
+  const text = Buffer.from('héllo websocket 🌍'.repeat(2))
+  const backing = Buffer.alloc(text.length + 1)
+  text.copy(backing, 1)
+  const view = backing.subarray(1)
+  const original = Buffer.from(view)
+  const [head, body] = WebsocketFrameSend.createFastTextFrame(view)
+  const mask = head.subarray(head.length - 4)
+  const decoded = Buffer.alloc(body.length)
+  for (let i = 0; i < body.length; ++i) decoded[i] = body[i] ^ mask[i & 3]
+  t.assert.deepStrictEqual(decoded, original)
+})