compressor.js

Jump To … +

compressor.js connection.js endpoint.js flow.js framer.js http.js index.js stream.js

compressor.js
¶

The implementation of the HTTP/2 Header Compression spec is separated from the 'integration' part which handles HEADERS and PUSH_PROMISE frames. The compression itself is implemented in the first part of the file, and consists of three classes: HeaderTable, HeaderSetDecompressor and HeaderSetCompressor. The two latter classes are Transform Stream subclasses that operate in object mode. These transform chunks of binary data into [name, value] pairs and vice versa, and store their state in HeaderTable instances.

The 'integration' part is also implemented by two Transform Stream subclasses that operate in object mode: the Compressor and the Decompressor. These provide a layer between the framer and the connection handling component.
```
exports.HeaderTable = HeaderTable;
exports.HeaderSetCompressor = HeaderSetCompressor;
exports.HeaderSetDecompressor = HeaderSetDecompressor;
exports.Compressor = Compressor;
exports.Decompressor = Decompressor;

var TransformStream = require('stream').Transform;
var assert = process.env.HTTP2_ASSERT ? require('assert') : function noop() {};
var util = require('util');
```
¶

Header compression
¶

The HeaderTable class

The Header Table is a component used to associate headers to index values. It is basically an ordered list of [name, value] pairs, so it's implemented as a subclass of Array.

function HeaderTable(log, table, limit) {
  var self = table.map(entryFromPair);
  self._log = log;
  self._limit = limit || DEFAULT_HEADER_TABLE_LIMIT;
  self._size = tableSize(self);
  self.add = HeaderTable.prototype.add;
  return self;
}

¶

There are few more sets that are needed for the compression/decompression process that are all subsets of the Header Table, and are implemented as flags on header table entries:
- Reference Set: contains a group of headers used as a reference for the differential encoding of a new set of headers. (reference flag)
- Emitted headers: the headers that are already emitted as part of the current decompression process (not part of the spec, emitted flag)
- Headers to be kept: headers that should not be removed as the last step of the encoding process (not part of the spec, keep flag)
```
function entryFromPair(pair) {
  var entry = pair.slice();
  entry.reference = false;
  entry.emitted = false;
  entry.keep = false;
  entry._size = size(entry);
  return entry;
}
```
¶

The encoder decides how to update the header table and as such can control how much memory is used by the header table. To limit the memory requirements on the decoder side, the header table size is bounded.
- The default header table size limit is 4096 bytes.
- The size of an entry is defined as follows: the size of an entry is the sum of its name's length in bytes, of its value's length in bytes and of 32 bytes.
- The size of a header table is the sum of the size of its entries.
```
var DEFAULT_HEADER_TABLE_LIMIT = 4096;

function size(entry) {
  return new Buffer(entry[0] + entry[1], 'utf8').length + 32;
}

function tableSize(table) {
  var size = 0;
  for (var i = 0; i < table.length; i++) {
    size += table[i]._size;
  }
  return size;
}
```

The add(index, entry) can be used to manage the header table:

if index is Infinite it pushes the new entry at the end of the table
otherwise, it replaces the entry with the given index with the new entry
before doing such a modification, it has to be ensured that the header table size will stay lower than or equal to the header table size limit. To achieve this, repeatedly, the first entry of the header table is removed, until enough space is available for the modification.

HeaderTable.prototype.add = function(index, entry) {
  var limit = this._limit - entry._size;
  var droppedEntries = [];

  while ((this._size > limit) && (this.length > 0)) {
    var dropped = this.shift();
    this._size -= dropped._size;
    droppedEntries.push(dropped);
  }

  if (this._size <= limit) {
    index -= droppedEntries.length;
    if (index < 0) {
      this.unshift(entry);
    } else {
      this.splice(index, 1, entry); // this is like push() if index is Infinity
    }
    this._size += entry._size;
  }

  return droppedEntries;
};

¶

Initial header tables

Initial request table

HeaderTable.initialRequestTable  = [
  [ ':scheme'                     , 'http'  ],
  [ ':scheme'                     , 'https' ],
  [ ':host'                       , ''      ],
  [ ':path'                       , '/'     ],
  [ ':method'                     , 'get'   ],
  [ 'accept'                      , ''      ],
  [ 'accept-charset'              , ''      ],
  [ 'accept-encoding'             , ''      ],
  [ 'accept-language'             , ''      ],
  [ 'cookie'                      , ''      ],
  [ 'if-modified-since'           , ''      ],
  [ 'user-agent'                  , ''      ],
  [ 'referer'                     , ''      ],
  [ 'authorization'               , ''      ],
  [ 'allow'                       , ''      ],
  [ 'cache-control'               , ''      ],
  [ 'connection'                  , ''      ],
  [ 'content-length'              , ''      ],
  [ 'content-type'                , ''      ],
  [ 'date'                        , ''      ],
  [ 'expect'                      , ''      ],
  [ 'from'                        , ''      ],
  [ 'if-match'                    , ''      ],
  [ 'if-none-match'               , ''      ],
  [ 'if-range'                    , ''      ],
  [ 'if-unmodified-since'         , ''      ],
  [ 'max-forwards'                , ''      ],
  [ 'proxy-authorization'         , ''      ],
  [ 'range'                       , ''      ],
  [ 'via'                         , ''      ]
];

A HeaderSetCompressor instance is a transform stream that can be used to compress a single header set. Its input is a stream of [name, value] pairs and its output is a stream of binary data chunks.

It is a real streaming compressor, since it does not wait until the header set is complete.

The compression algorithm is (intentionally) not specified by the spec. Therefore, the current compression algorithm can probably be improved in the future.

util.inherits(HeaderSetCompressor, TransformStream);
function HeaderSetCompressor(log, table) {
  TransformStream.call(this, { objectMode: true });

  this._log = log.child({ component: 'compressor' });
  this._table = table;
  this.push = TransformStream.prototype.push.bind(this);
}

HeaderSetCompressor.prototype.send = function send(rep) {
  this._log.trace({ key: rep.name, value: rep.value, index: rep.index },
                  'Emitting header representation');

  if (!rep.chunks) {
    rep.chunks = HeaderSetCompressor.header(rep);
  }
  rep.chunks.forEach(this.push);
};

_transform is the implementation of the corresponding virtual function of the TransformStream class. It processes the input headers one by one:

HeaderSetCompressor.prototype._transform = function _transform(pair, encoding, callback) {
  var name = pair[0].toLowerCase();
  var value = pair[1];
  var entry, rep;

tries to find full (name, value) or name match in the header table

  var nameMatch = -1, fullMatch = -1;
  for (var index = 0; index < this._table.length; index++) {
    entry = this._table[index];
    if (entry[0] === name) {
      if (entry[1] === value) {
        fullMatch = index;
        break;
      } else if (nameMatch === -1) {
        nameMatch = index;
      }
    }
  }

if there's full match, it will be an indexed representation (or more than one) depending on its presence in the reference, the emitted and the keep set

  if (fullMatch !== -1) {
    rep = { name: fullMatch, value: fullMatch, index: -1 };

    if (!entry.reference) {
      this.send(rep);
      entry.reference = true;
      entry.emitted = true;
    }

    else if (entry.keep) {
      this.send(rep);
      this.send(rep);
      this.send(rep);
      this.send(rep);
      entry.keep = false;
      entry.emitted = true;
    }

    else if (entry.emitted) {
      this.send(rep);
      this.send(rep);
    }

    else {
      entry.keep = true;
    }
  }

otherwise, it will be a literal representation (with a name index if there's a name match)

  else {
    entry = entryFromPair(pair);
    entry.emitted = true;

    var insertIndex;
    if (entry._size > this._table._limit / 2) {
      insertIndex = -1;
    } else if (nameMatch !== -1) {
      insertIndex = nameMatch;
    } else {
      insertIndex = Infinity;
    }

    if (insertIndex !== -1) {
      entry.reference = true;
      var droppedEntries = this._table.add(insertIndex, entry);
      for (index = 0; index < droppedEntries.length; index++) {
        var dropped = droppedEntries[index];
        if (dropped.keep) {
          rep = { name: index, value: index, index: -1 };
          this.send(rep);
          this.send(rep);
        }
      }
    }

    this.send({ name: (nameMatch !== -1) ? nameMatch : name, value: value, index: insertIndex });
  }

  callback();
};

¶

_flush is the implementation of the corresponding virtual function of the TransformStream class. It gets called when there's no more header to compress. The final step:
```
HeaderSetCompressor.prototype._flush = function _flush(callback) {
```

removing entries from the header set that are not marked to be kept or emitted

  for (var index = 0; index < this._table.length; index++) {
    var entry = this._table[index];
    if (entry.reference && !entry.keep && !entry.emitted) {
      this.send({ name: index, value: index, index: -1 });
      entry.reference = false;
    }
    entry.keep = false;
    entry.emitted = false;
  }

  callback();
};

¶

Detailed Format

Integer representation

The algorithm to represent an integer I is as follows:

If I < 2^N - 1, encode I on N bits
Else, encode 2^N - 1 on N bits and do the following steps:
1. Set I to (I - (2^N - 1)) and Q to 1
2. While Q > 0
  1. Compute Q and R, quotient and remainder of I divided by 2^7
  2. If Q is strictly greater than 0, write one 1 bit; otherwise, write one 0 bit
  3. Encode R on the next 7 bits
  4. I = Q

HeaderSetCompressor.integer = function writeInteger(I, N) {
  var limit = Math.pow(2,N) - 1;
  if (I < limit) {
    return [new Buffer([I])];
  }

  var bytes = [];
  if (N !== 0) {
    bytes.push(limit);
  }
  I -= limit;

  var Q = 1, R;
  while (Q > 0) {
    Q = Math.floor(I / 128);
    R = I % 128;

    if (Q > 0) {
      R += 128;
    }
    bytes.push(R);

    I = Q;
  }

  return [new Buffer(bytes)];
};

The inverse algorithm:

Set I to the number coded on the lower N bits of the first byte
If I is smaller than 2^N - 1 then return I
Else the number is encoded on more than one byte, so do the following steps:
1. Set M to 0
2. While returning with I
  1. Let B be the next byte (the first byte if N is 0)
  2. Read out the lower 7 bits of B and multiply it with 2^M
  3. Increase I with this number
  4. Increase M by 7
  5. Return I if the most significant bit of B is 0

HeaderSetDecompressor.integer = function readInteger(buffer, N) {
  var limit = Math.pow(2,N) - 1;

  var I = buffer[buffer.cursor] & limit;
  if (N !== 0) {
    buffer.cursor += 1;
  }

  if (I === limit) {
    var M = 0;
    do {
      I += (buffer[buffer.cursor] & 127) << M;
      M += 7;
      buffer.cursor += 1;
    } while (buffer[buffer.cursor - 1] & 128);
  }

  return I;
};

String literal representation

Literal strings can represent header names or header values. They are encoded in two parts:

The string length, defined as the number of bytes needed to store its UTF-8 representation, is represented as an integer with a zero bits prefix. If the string length is strictly less than 128, it is represented as one byte.
The string value represented as a list of UTF-8 characters.

HeaderSetCompressor.string = function writeString(str) {
  var encodedString = new Buffer(str, 'utf8');
  var encodedLength = HeaderSetCompressor.integer(encodedString.length, 0);
  return encodedLength.concat(encodedString);
};

HeaderSetDecompressor.string = function readString(buffer) {
  var length = HeaderSetDecompressor.integer(buffer, 0);
  var str = buffer.toString('utf8', buffer.cursor, buffer.cursor + length);
  buffer.cursor += length;
  return str;
};

Header represenations

The JavaScript object representation is described near the HeaderTable.prototype.execute() method definition.

All binary header representations start with a prefix signaling the representation type and an index represented using prefix coded integers:

  0   1   2   3   4   5   6   7
+---+---+---+---+---+---+---+---+
| 1 |        Index (7+)         |  Indexed Representation
+---+---------------------------+

+---+---+---+---+---+---+---+---+
| 0 | 1 | 1 |    Index (5+)     |  Literal w/o Indexing
+---+---+---+-------------------+

+---+---+---+---+---+---+---+---+
| 0 | 1 | 0 |    Index (5+)     |  Literal w/ Incremental Indexing
+---+---+---+-------------------+

+---+---+---+---+---+---+---+---+
| 0 | 0 |      Index (6+)       |  Literal w/ Substitution Indexing
+---+---+-----------------------+

The Indexed Representation consists of the 1-bit prefix and the Index that is represented as a 7-bit prefix coded integer and nothing else.

After the first bits, all literal representations specify the header name, either as a pointer to the Header Table (Index) or a string literal. When the string literal representation is used, the Index is set to 0 and the string literal starts at the second byte.

When using Substitution Indexing, a new index comes next represented as a 0-bit prefix integer, specifying the record in the Header Table that needs to be replaced.

For all literal representations, the specification of the header value comes next. It is always represented as a string.

var representations = {
  indexed             : { prefix: 7, pattern: 0x80 },
  literal             : { prefix: 5, pattern: 0x60 },
  literalIncremental  : { prefix: 5, pattern: 0x40 },
  literalSubstitution : { prefix: 6, pattern: 0x00 }
};

HeaderSetCompressor.header = function writeHeader(header) {
  var representation, buffers = [];

  if (typeof header.value === 'number') {
    representation = representations.indexed;
  } else if (header.index === -1) {
    representation = representations.literal;
  } else if (header.index === Infinity) {
    representation = representations.literalIncremental;
  } else {
    representation = representations.literalSubstitution;
  }

  if (representation === representations.indexed) {
    buffers.push(HeaderSetCompressor.integer(header.value, representation.prefix));

  } else {
    if (typeof header.name === 'number') {
      buffers.push(HeaderSetCompressor.integer(header.name + 1, representation.prefix));
    } else {
      buffers.push(HeaderSetCompressor.integer(0, representation.prefix));
      buffers.push(HeaderSetCompressor.string(header.name));
    }

    if (representation === representations.literalSubstitution) {
      buffers.push(HeaderSetCompressor.integer(header.index, 0));
    }

    buffers.push(HeaderSetCompressor.string(header.value));
  }

  buffers[0][0][0] |= representation.pattern;

  return Array.prototype.concat.apply([], buffers); // array of arrays of buffers -> array of buffers
};

HeaderSetDecompressor.header = function readHeader(buffer) {
  var representation, header = {};

  var firstByte = buffer[buffer.cursor];
  if (firstByte & 0x80) {
    representation = representations.indexed;
  } else if (firstByte & 0x40) {
    if (firstByte & 0x20) {
      representation = representations.literal;
    } else {
      representation = representations.literalIncremental;
    }
  } else {
    representation = representations.literalSubstitution;
  }

  if (representation === representations.indexed) {
    header.value = header.name = HeaderSetDecompressor.integer(buffer, representation.prefix);
    header.index = -1;

  } else {
    header.name = HeaderSetDecompressor.integer(buffer, representation.prefix) - 1;
    if (header.name === -1) {
      header.name = HeaderSetDecompressor.string(buffer);
    }

    if (representation === representations.literalSubstitution) {
      header.index = HeaderSetDecompressor.integer(buffer, 0);
    } else if (representation === representations.literalIncremental) {
      header.index = Infinity;
    } else {
      header.index = -1;
    }

    header.value = HeaderSetDecompressor.string(buffer);
  }

  return header;
};

¶

Integration with HTTP/2
¶

This section describes the interaction between the compressor/decompressor and the rest of the HTTP/2 implementation. The Compressor and the Decompressor makes up a layer between the framer and the connection handling component. They let most frames pass through, except HEADERS and PUSH_PROMISE frames. They convert the frames between these two representations:
```
{                                   {
 type: 'HEADERS',                    type: 'HEADERS',
 flags: {},                          flags: {},
 stream: 1,               <===>      stream: 1,
 headers: {                          data: Buffer
  N1: 'V1',                         }
  N2: ['V1', 'V2', ...],
  // ...
 }
}
```
There are possibly several binary frame that belong to a single non-binary frame.
```
var MAX_HTTP_PAYLOAD_SIZE = 16383;
```
¶

The Compressor class

The Compressor transform stream is basically stateless.

util.inherits(Compressor, TransformStream);
function Compressor(type, log) {
  TransformStream.call(this, { objectMode: true });

  this._log = log.child({ component: 'compressor' });

  assert((type === 'REQUEST') || (type === 'RESPONSE'));
  var initialTable = (type === 'REQUEST') ? HeaderTable.initialRequestTable
                                          : HeaderTable.initialResponseTable;
  this._table = new HeaderTable(this._log, initialTable);
}

compress takes a header set, and compresses it using a new HeaderSetCompressor stream instance. This means that from now on, the advantages of streaming header encoding are lost, but the API becomes simpler.

Compressor.prototype.compress = function compress(headers) {
  var compressor = new HeaderSetCompressor(this._log, this._table);
  for (var name in headers) {
    var value = headers[name];
    if (value instanceof Array) {
      for (var i = 0; i< value.length; i++) {
        compressor.write([String(name), String(value[i])]);
      }
    } else {
      compressor.write([String(name), String(value)]);
    }
  }
  compressor.end();

  var chunk, chunks = [];
  while (chunk = compressor.read()) {
    chunks.push(chunk);
  }
  return concat(chunks);
};

When a frame arrives

Compressor.prototype._transform = function _transform(frame, encoding, done) {

and it is a HEADERS or PUSH_PROMISE frame
- it generates a header block using the compress method
- cuts the header block into chunks that are not larger than MAX_HTTP_PAYLOAD_SIZE
- for each chunk, it pushes out a chunk frame that is identical to the original, except the data property which holds the given chunk, the type of the frame which is always CONTINUATION except for the first frame, and the END_HEADERS/END_PUSH_STREAM flag that marks the last frame and the END_STREAM flag which is always false before the end

  if (frame.type === 'HEADERS' || frame.type === 'PUSH_PROMISE') {
    var buffer = this.compress(frame.headers);

    var chunks = cut(buffer, MAX_HTTP_PAYLOAD_SIZE);

    for (var i = 0; i < chunks.length; i++) {
      var chunkFrame;
      var first = (i === 0);
      var last = (i === chunks.length - 1);

      if (first) {
        chunkFrame = util._extend({}, frame);
        chunkFrame.flags = util._extend({}, frame.flags);
        chunkFrame.flags['END_' + frame.type] = last;
      } else {
        chunkFrame = {
          type: 'CONTINUATION',
          flags: { END_HEADERS: last },
          stream: frame.stream
        };
      }
      if (chunkFrame.type !== 'PUSH_PROMISE') {
        chunkFrame.flags.END_STREAM = last && frame.END_STREAM;
      }
      chunkFrame.data = chunks[i];

      this.push(chunkFrame);
    }
  }

¶
- otherwise, the frame is forwarded without taking any action
```
  else {
    this.push(frame);
  }

  done();
};
```
¶

The Decompressor class

The Decompressor is a stateful transform stream, since it has to collect multiple frames first, and the decoding comes after unifying the payload of those frames.

If there's a frame in progress, this._inProgress is true. The frames are collected in this._frames, and the type of the frame and the stream identifier is stored in this._type and this._stream respectively.

util.inherits(Decompressor, TransformStream);
function Decompressor(type, log) {
  TransformStream.call(this, { objectMode: true });

  this._log = log.child({ component: 'compressor' });

  assert((type === 'REQUEST') || (type === 'RESPONSE'));
  var initialTable = (type === 'REQUEST') ? HeaderTable.initialRequestTable
                                          : HeaderTable.initialResponseTable;
  this._table = new HeaderTable(this._log, initialTable);

  this._inProgress = false;
  this._base = undefined;
}

decompress takes a full header block, and decompresses it using a new HeaderSetDecompressor stream instance. This means that from now on, the advantages of streaming header decoding are lost, but the API becomes simpler.

Decompressor.prototype.decompress = function decompress(block) {
  var decompressor = new HeaderSetDecompressor(this._log, this._table);
  decompressor.end(block);

  var headers = {};
  var pair;
  while (pair = decompressor.read()) {
    var name = pair[0];
    var value = pair[1];
    if (name in headers) {
      if (headers[name] instanceof Array) {
        headers[name].push(value);
      } else {
        headers[name] = [headers[name], value];
      }
    } else {
      headers[name] = value;
    }
  }

  return headers;
};

When a frame arrives

Decompressor.prototype._transform = function _transform(frame, encoding, done) {

and the collection process is already _inProgress, the frame is simply stored, except if it's an illegal frame

  if (this._inProgress) {
    if ((frame.type !== 'CONTINUATION') || (frame.stream !== this._base.stream)) {
      this._log.error('A series of HEADER frames were not continuous');
      this.emit('error', 'PROTOCOL_ERROR');
      return;
    }
    this._frames.push(frame);
  }

and the collection process is not _inProgress, but the new frame's type is HEADERS or PUSH_PROMISE, a new collection process begins

  else if ((frame.type === 'HEADERS') || (frame.type === 'PUSH_PROMISE')) {
    this._inProgress = true;
    this._base = frame;
    this._frames = [frame];
  }

¶
- otherwise, the frame is forwarded without taking any action
```
  else {
    this.push(frame);
  }
```

When the frame signals that it's the last in the series, the header block chunks are concatenated, the headers are decompressed, and a new frame gets pushed out with the decompressed headers.

  if (this._inProgress && (frame.flags.END_HEADERS || frame.flags.END_PUSH_PROMISE)) {
    var buffer = concat(this._frames.map(function(frame) {
      return frame.data;
    }));
    try {
      var headers = this.decompress(buffer);
    } catch(error) {
      this._log.error({ err: error }, 'Header decompression error');
      this.emit('error', 'COMPRESSION_ERROR');
      return;
    }
    this.push(util._extend({ headers: headers }, this._base));
    this._inProgress = false;
  }

  done();
};

¶

Helper functions

Concatenate an array of buffers into a new buffer

function concat(buffers) {
  var size = 0;
  for (var i = 0; i < buffers.length; i++) {
    size += buffers[i].length;
  }

  var concatenated = new Buffer(size);
  for (var cursor = 0, j = 0; j < buffers.length; cursor += buffers[j].length, j++) {
    buffers[j].copy(concatenated, cursor);
  }

  return concatenated;
}

Cut buffer into chunks not larger than size

function cut(buffer, size) {
  var chunks = [];
  var cursor = 0;
  do {
    var chunkSize = Math.min(size, buffer.length - cursor);
    chunks.push(buffer.slice(cursor, cursor + chunkSize));
    cursor += chunkSize;
  } while(cursor < buffer.length);
  return chunks;
}

compressor.js

Header compression

The HeaderTable class

Initial header tables

Initial request table

Initial response table

The HeaderSetDecompressor class

The HeaderSetCompressor class

Detailed Format

Integer representation

String literal representation

Header represenations

Integration with HTTP/2

The Compressor class

The Decompressor class

Helper functions