Fix buffer leak in HttpObjectEncoder when content is always empty

[pderop]

Motivation:

When the HttpObjectEncoder.encodeAndClose method encodes a response which content is always empty (as described in _HttpObjectEncoder. isContentAlwaysEmpty ), then the empty last http content or the empty FullHttpResponse is not closed.

An http response content is always empty for example when status code is INFORMATIONAL, is a 204, 205, 304 ... See for example here

Example scenario:

1. a DefaultHttpResponse is first encoded with status 204, in this case, the HttpObjectEncoder.encodeAndClose method will initialize its state to ST_CONTENT_ALWAYS_EMPTY, see here.

2. next, an EmptyLastHttpContent is encoded. So, the HttpObjectEncoder.encodeAndClose state being in ST_CONTENT_ALWAYS_EMPTY, then the encodeAndClose method switch statement will proceed here directly, which is not closing the EmptyLastHttpContent

Modification:

Move the code which was disposing the message from here to the case on ST_CONTENT_ALWAYS_EMPTY here.
Also added some test cases in HttpResponseEncoderTest.

Result:

Disposing the message to the case on ST_CONTENT_ALWAYS_EMPTY will continue to ensure that the message is closed when the state is ST_CONTENT_NON_CHUNK and when content length is zero (because the case fall-through, see here), and also ensures that any HttpContent is closed when state is ST_CONTENT_ALWAYS_EMPTY.

[pderop]

@chrisvest , @normanmaurer , can you take a look ?
thanks.

[franz1981]

Thanks for this, when I'll port the netty 4.1 version (probably doing things differently really ie using visitors/virtual dispatch) -> see #12708 the test will be our saviour to capture regressions

[pderop]

@franz1981 ,

Thanks for the info; to be honest, I did not take a look at the false sharing issue yet, but I'll try to look into #12708 soon.
(for the moment, I'm investigating some remaining buffer leaks).

[normanmaurer]

LGTM!

[chrisvest]

Nice find!

[pderop]

thanks !