Fix bug where connection did not return the write buffer to the pool
after a write error. Add test for the same.
Rename messsageWriter.fatal method to endMessage and consolidate all
message cleanup code there. This ensures that the buffer is returned to
pool on all code paths.
Rename Conn.prepMessage to beginMessage for symmetry with endMessage.
Move some duplicated code at calls to prepMessage to beginMessage.
Bonus improvement: Adjust message and buffer size in TestWriteBufferPool
to test that pool works with fragmented messages.
Add WriteBufferPool to Dialer and Upgrader. This field specifies a pool
to use for write operations on a connection. Use of the pool can reduce
memory use when there is a modest write volume over a large number of
connections.
Use larger of hijacked buffer and buffer allocated for connection (if
any) as buffer for building handshake response. This decreases possible
allocations when building the handshake response.
Modify bufio reuse test to call Upgrade instead of the internal
newConnBRW. Move the test from conn_test.go to server_test.go because
it's a serer test.
Update newConn and newConnBRW:
- Move the bufio "hacks" from newConnBRW to separate functions and call
these functions directly from Upgrade.
- Rename newConn to newTestConn and move to conn_test.go. Shorten
argument list to common use case.
- Rename newConnBRW to newConn.
- Add pool code to newConn.
The documentation sometimes used the term "frame" when referring to
single frame control messages. Use the term "message" for consistency
in the documentation and to hide a detail that most application
programmers do not need to know about.
Return empty message for CloseNoStatusReceived. This status indicates
that the message is empty, so make it so. Because it's illegal to send
CloseNoStatusReceived, this change should not break a correct
application.
Reuse the buffer backing the bufio.Writer returned from hijack if that
buffer is large enough to be generally useful and
Upgrader.WriteBufferSize == 0.
Update the logic for reusing bufio.Reader returned from hijack to match
the logic for bufio.Reader: The buffer backing the reader must be
sufficiently large to be generally useful and Upgrader.ReadBufferSize ==
0.
Improve the documentation for ReadBufferSize and WriterBufferSize in
Dialer and Upgrader.
Change the default compression level to 1. This level is faster and uses
less memory.
Add Conn.SetCompressionLevel API to allow applications to tune
compression on a per message basis.
- Do not fail NextWriter when close of previous writer fails.
- Replace closeSent field with mutex protected writeErr. Set writeErr on
any error writing to underlying network connection. Check and return
writeErr before attempting to write to network connection. Check
writeErr in NextWriter so application can detect failed connection
before attempting to write.
- Do not close underlying network connection on error.
- Move message writing state and method flushFrame from Conn to
messageWriter. This makes error code paths (and the code in general)
easier to understand.
- Add messageWriter field err to latch errors in messageWriter.
Bonus: Improve test coverage.
Add newCompressionWriter and newDecompressionReader fields to Conn. When
not nil, these functions are used to create a compression/decompression
wrapper around an underlying message writer/reader.
Add code to set and check for RSV1 frame header bit.
Add functions compressNoContextTakeover and decompressNoContextTakeover
for creating no context takeover wrappers around an underlying message
writer/reader.
Work remaining:
- Add fields to Dialer and Upgrader for specifying compression options.
- Add compression negotiation to Dialer and Upgrader.
- Add function to enable/disable write compression:
// EnableWriteCompression enables and disables write compression of
// subsequent text and binary messages. This function is a noop if
// compression was not negotiated with the peer.
func (c *Conn) EnableWriteCompression(enable bool) {
c.enableWriteCompression = enable
}
Update documentation to explicitly state that applications must break out of a
read loop on error.
Detect application read loops spinning on a failed connection and panic.
Detect concurrent writes and panic. The detection is best-effort.
Update documentation to state that connections respond to close frames.
- Use new closeError type for reporting close frames to the application.
- Use closeError with code 1006 when the peer closes connection without
sending a close frame. The error io.ErrUnexpectedEOF was used
previously. This change helps developers distinguish abnormal closure
and an unexpected EOF in the JSON parser.
- Modify data message reader to return io.ErrUnexpectedEOF if a close
message is received before the final frame of the message.
- Modify NextReader to return io.ErrUnexpectedEOF if underlying
connection returns io.EOF before a close message.
Tariq Ibrahim