This document briefly describes the work done to support Java (and other Language Server Protocol-based completion engines).

Overview

The original PR implemented native support in ycmd for the Java language, based on jdt.ls. In summary, the following key features were added:

• Installation of jdt.ls (built from source with build.py)
• Management of the jdt.ls server instance, projects etc.
• A generic (ish) implementation of a Language Server Protocol client so far as is required for jdt.ls (easily extensible to other engines)
• Support for the following Java semantic engine features:
  • Semantic code completion, including automatic imports
  • As-you-type diagnostics
  • GoTo including GoToReferences
  • FixIt
  • RefactorRename
  • GetType
  • GetDoc

See the trello board for a more complete picture.

Overall design/goals

Key goals:

1. Support Java in ycmd and YCM; make it good enough to replace eclim and javacomplete2 for most people
2. Make it possible/easy to support other lsp servers in future (but, don't suffer from yagni); prove that this works.

An overview of the objects involved can be seen on [this card][design]. In short:

• 2 classes implement the language server protocol in the language_server_completer.py module:
• LanguageServerConnection - an abstraction of the communication with the server, which may be over stdio or any number of TCP/IP ports (or a domain socket, etc.). Only a single implementation is included (stdio), but implementations for TCP/IP were written originally and dropped in favour of stdio's simplicity.
• LanguageServerCompleter - an abstract base for any completer based on LSP, which implements as much standard functionality as possible including completions, diagnostics, goto, fixit, rename, etc.
• The java_completer itself implements the LanguageServerCompleter, boots the jdt.ls server, and instantiates a LanguageServerConnection for communication with jdt.ls.

The overall plan and some general discussion around the project can be found on the trello board I used for development.

Threads, and why we need them

LSP is by its nature an asynchronous protocol. There are request-reply like requests and unsolicited notifications. Receipt of the latter is mandatory, so we cannot rely on their being a bottle thread executing a client request.

So we need a message pump and dispatch thread. This is actually the LanguageServerConnection, which implements thread. It's main method simply listens on the socket/stream and despatches complete messages to the LanguageServerCompleter. It does this:

• For requests: similarly to the TypeScript completer, using python event objects, wrapped in our Response class
• For notifications: via a synchronised queue. More on this later.

A representation of this is on the "Requests and notifications" page of the [design][], including a rough sketch of the thread interaction.

Some handling is done in the message pump.

While it is perhaps regrettable to do general processing directly in the message pump, there are certain notifications which we have to handle immediately when we get them, such as:

• Initialisation messages
• Diagnostics

In these cases, we allow some code to be executed inline within the message pump thread, as there is no other thread guaranteed to execute. These are handled by callback functions and state is protected mutexes.

Startup sequence

See the 'initialisation sequence' tab on the [design][] for a bit of background.

In standard LSP, the initialisation sequence consists of an initialise request-reply, followed by us sending the server an initialised notification. We must not send any other requests until this has completed.

An additional wrinkle is that jdt.ls, being based on eclipse has a whole other initialisation sequence during which time it is not fully functional, so we have to determine when that has completed too. This is done by jdt.ls-specific messages and controls the ServerIsReady response.

In order for none of these shenanigans to block the user, we must do them all asynchronously, effectively in the message pump thread. In addition, we must queue up any file contents changes during this period to ensure the server is up to date when we start processing requests proper.

This is unfortunately complicated, but there were early issues with really bad UI blocking that we just had to get rid of.

Completion

Language server protocol requires that the client can apply textEdits, rather than just simple text. This is not an optional feature, but ycmd clients do not have this ability.

The protocol, however, restricts that the edit must include the original requested completion position, so we can perform some simple text manipulation to apply the edit to the current line and determine the completion start column based on that.

In particular, the jdt.ls server returns textEdits that replace the entered text for import completions, which is one of the most useful completions.

We do this super inefficiently by attempting to normalise the TextEdits into insertion_texts with the same start_codepoint. This is necessary particularly due to the way that eclipse returns import completions for packages.

We also include support for "additionalTextEdits" which allow automatic insertion of, e.g., import statements when selecting completion items. These are sent on the completion response as an additional completer data item called 'fixits'. The client applies the same logic as a standard FixIt once the selected completion item is inserted.

Diagnostics

Diagnostics in LSP are delivered asynchronously via notifications. Normally, we would use the OnFileReadyToParse response to supply diagnostics, but due to the lag between refreshing files and receiving diagnostics, this leads to a horrible user experience where the diagnostics always lag one edit behind.

To resolve this, we use the long-polling mechanism added here (ReceiveMessages request) to return diagnostics to the client asynchronously.

We deliver asynchronous diagnostics to the client in the same way that the language server does, i.e. per-file. The client then fans them out or does whatever makes sense for the client. This is necessary because it isn't possible to know when we have received all diagnostics, and combining them into a single message was becoming clunky and error prone.

In order to be relatively compatible with other clients, we also return diagnostics on the file-ready-to-parse event, even though they might be out of date wrt the code. The client is responsible for ignoring these diagnostics when it handles the asynchronously delivered ones. This requires that we hold the "latest" diagnostics for a file. As it turns out, this is also required for FixIts.

Projects

jdt.ls is based on eclipse. It is in fact an eclipse plugin. So it requires an eclipse workspace. We try and hide this by creating an ad-hoc workspace for each ycmd instance. This prevents the possibility of multiple "eclipse" instances using the same workspace, but can lead to unreasonable startup times for large projects.

The jdt.ls team strongly suggest that we should re-use a workspace based on the hash of the "project directory" (essentially the dir containing the project file: .project, pom.xml or build.gradle). They also say, however, that eclipse frequently corrupts its workspace.

So we have a hidden switch to re-use a workspace as the jdt.ls devs suggest. In testing at work, this was mandatory due to a slow SAN, but at home, startup time is not an issue, even for large projects. I think we'll just have to see how things go to decide which one we want to keep.

Subcommands

GetDoc/GetType

There is no GetType in LSP. There's only "hover". The hover response is hilariously server-specific, so in the base LanguageServerCompleter we just provide the ability to get the hover response and JavaCompleter extracts the appropriate info from there. Thanks to @bstaletic for this!

FixIt

FixIts are implemented as code actions, and require the diagnostic they relate to to be send from us to the server, rather than just a position. We use the stored diags and find the nearest one based on the request_data.

What's worse is that the LSP provides no documentation for what the "Code action" response should be, and it is 100% implementation-specific. They just have this command abstraction which is basically "do a thing".

From what I've seen, most servers just end up with either a WorkspaceEdit or a series of TextEdits, which is fine for us as that's what ycmd's protocol looks like.

The solution is that we have a callback into the JavaCompleter to handle the (custom) java.apply.workspaceEdit "command".

GoToReferences

Annoyingly, jdt.ls sometimes returns references to .class files within jar archives using a custom jdt:// protocol. We can't handle that, so we have to dodge and weave so that we don't crash.

Stopping the server

Much like the initialisation sequence, the LSP shutdown sequence is a bit fiddly. 2 things are required:

1. A shutdown request-reply. The server tides up and prepares to die!
2. An exit notification. We tell the server to die.

This isn't so bad, but jdt.ls is buggy and actually dies without responding to the shutdown request. So we have a bunch of code to handle that and to ensure that the server dies eventually, as it had a habit of getting stuck running, particularly if we threw an exception.