Change sys.write to async in tsserver

[zhengbli]

This is related to
microsoft/TypeScript-Sublime-Plugin#379
and
#2758 (comment)

fs.writeSync truncates a long line, and will throw an exception of EAGAIN: Resource not available if re trying right away. Changing the writing behavior to async solves the problem.

[RyanCavanaugh]

This seems pretty dangerous as a change -- how do we know we don't have any consumers of write that are assuming it's synchronous (e.g. that you could call write and then invoke a program on the next line that consumes that file)?

[zhengbli]

That is a legitimate concern. As the problem currently only affects tsserver, it might be safer to overwrite the definition of ts.sys.write within server.ts.

[mhegazy]

I do not think this is the right thing to do. we have had that in the past, and we run into issues not draining the stream before the process ends. if this is for ts server purposes, we can override the implementation in the server.ts

[zhengbli]

I moved the changes to tsserver.ts, and add some mechanism so that one write only starts when last write finished.

[vladima]

why do you need to copy the list (given that messagesToWrite is effectively a queue so items are added\removed at known positions)? You can do better with messagesToWrite.shift

    let pending: string[] = [];
    function queueMessage(s: string) {
        pending.push(s);
        if (pending.length === 1) {
            drain();
        }
    }

    function drain() {
        Debug.assert(pending.length > 0);
        writeBuffer(new Buffer(pending[0], "utf8"), 0);
    }

    function writeBuffer(buffer: Buffer, offset: number) {
        const toWrite = buffer.length - offset;
        fs.write(1, buffer, offset, toWrite, undefined, (err, written, buffer) => {
            if (toWrite > written) {
                writeBuffer(buffer, offset + written);
            }
            else {
                Debug.assert(pending.length > 0);
                pending.shift();
                if (pending.length > 0) {
                    drain();
                }
            }
        });
    }

[zhengbli]

@vladima @mhegazy I updated the PR according to vlad's comment, which is shorted and uses shift instead. Do you have other comments?

[mhegazy]

👍

[vladima]

do we care about errors here?

[zhengbli]

We are retrying instead of handling the error anyway, it may only be used for printing error messages I think.

[vladima]

LGTM