IPC Unix domain socket / Windows named pipe

[eed3si9n]

Fixes #3723

In addition to TCP, this adds support for IPC (interprocess communication) using Unix domain socket and Windows named pipe. The Java socket implementation for IPC is lifted from facebook/nailgun, which is released under Apache v2.

The use of Unix domain socket has performance and security benefits.

Here's screenshot of this running on a Windows virtual machine:

serverConnectionType

This introduces a new setting key called serverConnectionType, which is set to ConnectionType.Local by default:

serverConnectionType := ConnectionType.Local,

The other option is ConnectionType.Tcp.

active.json

To communicate to the clients about the connection type, it uses URI with local: as the protocol.

{"uri":"local:///Users/xxx/.sbt/1.0/server/0845deda85cb41abdb9f/sock"}

[eed3si9n]

Travis passed the test. Codacy problems are coming from the Nailgun code's variable names.

[dwijnand]

Why does this include a bunch of Nailgun source code?

[eed3si9n]

@dwijnand The only available nailgun artifact on Maven Central is 0.9.1, and the code that I am borrowing from them is fairly new stuff for nailgun 2.x.

[dwijnand]

Please could you:

1. create an issue requesting this nailgun classes be released somewhere, and
2. add somewhere a reference to what commit sha these source files come from (for future reference)

[dwijnand]

Hmm, maybe it's best to introduce this enum value when it's supported.

[eed3si9n]

@dwijnand facebookarchive/nailgun#109

[dwijnand]

Is there an issue about having a binary release on Maven Central?

[eed3si9n]

Is there an issue about having a binary release on Maven Central?

Binary release for Nailgun? 0.9.1 from 2012 is published, but there hasn't been public release yet since. The interesting bits that I am after were committed in 2017. It could mean that 0.9.2 is not yet out, or ppl are releasing in-house as app.

[dwijnand]

Binary release for Nailgun?

Either of nailgun or a subset of nailgun that includes the classes we're importing here.

I see you've created facebookarchive/nailgun#108 (comment) - I'd've preferred a standalone, more easily findable, issue, but I'm not going to make that get in the way of this landing.

[jameskoch2]

Would there be any security benefit to applying these same Sockets to ForkTests? Reading the code it seems like there's a very small window between ForkTests opening a loopback port and accepting at-most-one connections from the child process. So benefit seems minimal, at a glance, but would like to hear other opinions.

[eed3si9n]

I think any inter-process communication is worth reviewing to see either Unix domain socket and/or server is applicable.

[jameskoch2]

I filed #3938 after closer inspection of the NamedPipes code, because by my reading it looks like it would still allow other users on the same machine to snoop on traffic.