test: http2 client socket operations

[trivikr]

This commit tests pause, resume, drain operations on client.socket

Refs: #14985

Checklist

• make -j4 test (UNIX), or vcbuild test (Windows) passes
• tests and/or benchmarks are included
• commit message follows commit guidelines

Affected core subsystem(s)

test, http2

[apapirovski]

Thank you for starting this work @trivikr. Some general feedback: I understand that prettier is currently a popular code formatter but it goes against established style within the node project and also has a tendency to really bloat the indent level of code. It also tends to obscure git history for purely stylistic reasons.

[apapirovski]

I don't think much is gained by moving this here instead of having it in a separate file. The server still has to be instantiated separately and the two test cases aren't really connected. We also lose the history of the other file. I would strongly prefer to have the other test restored and have this taken out.

[trivikr]

Restored the code in test-http2-client-socket-destroy.js in new commit

[apapirovski]

I'll leave this up to @jasnell & @mcollina but I don't think this is actually testing what it claims to. Feels like a way to bump up code coverage without truly testing the functionality which is worse than not having a test since it hides potential problem areas.

Unfortunately I don't really have great suggestion for how to properly test it given how the socket is used in h2.

[apapirovski]

Ideally this would be on one line with server.on('stream', common.mustCall((stream) => {. See my general note re: using prettier.

[trivikr]

Moved it to one line in new commit

[apapirovski]

Indent will need to be adjusted throughout to pass the linter.

[trivikr]

Adjusted indentation and confirmed that make lint-js passes in new commit

[apapirovski]

Same as above, indent level is off now. I don't think this will pass the linter.

Nit: this would normally be one line in node req.on('response', common.mustCall(() => {. Same below for end. (Which also reduces the indentation by one level inside.)

[trivikr]

Adjusted indentation and confirmed that make lint-js passes in new commit

[apapirovski]

Reposting this since github incorrectly hides it.

I'll leave this up to @jasnell & @mcollina but I don't think this is actually testing what it claims to. Feels like a way to bump up code coverage without truly testing the functionality which is worse than not having a test since it hides potential problem areas.

Unfortunately I don't really have great suggestion for how to properly test it given how the socket is used in h2.

[trivikr]

I'd tried mocking calls to readStart and readStop the same way as submitSettings is mocked as discussed in #16083 (comment)

However, the call was not reaching the mock. I'll wait for inputs from @jasnell / @mcollina
I agree that current commit doesn't confirm that pause, resume and drain methods were called.

[trivikr]

One solution I can think of is ensuring that operations were called on client.socket and testing value stored in client.socket._handle.reading as follows:

client.socket.on('pause', common.mustCall(() => {
  assert(!client.socket._handle.reading);
}, 2));
client.socket.on('resume', common.mustCall(() => {
  assert(client.socket._handle.reading);
}, 2));
client.socket.on('drain', common.mustCall());

[mcollina]

The socket on the session should really not be used, because it is managed in C++. So whatever this is testing is having no net effect on the flow of data.

@apapirovski do you think we should move your Proxy technique to the session level? It would avoid this to come up. I'm not sure it won't slow things down.

[mcollina]

To be clear, I think this should not land.

[apapirovski]

The code this is dealing with, which I believe @jasnell added in his original implementation, is here:

node/lib/internal/http2/core.js

Lines 2179 to 2201 in 9b3d6a0

	function socketOnResume() {
	if (this._paused)
	return this.pause();
	if (this._handle && !this._handle.reading) {
	this._handle.reading = true;
	this._handle.readStart();
	}
	}
	function socketOnPause() {
	if (this._handle && this._handle.reading) {
	this._handle.reading = false;
	this._handle.readStop();
	}
	}
	function socketOnDrain() {
	const needPause = 0 > this._writableState.highWaterMark;
	if (this._paused && !needPause) {
	this._paused = false;
	this.resume();
	}
	}

I'll let him speak to its purpose but as far as I'm concerned, using the proxy socket and removing this code might be preferable. I can't really think of a use case to pausing or resuming the socket manually, even if there's code to potentially handle it.

When I tested the proxy performance on compat it seemed to have literally no impact so I can't imagine the overhead is that big.

[jasnell]

This code was added defensively with the intent of circling back to see if it's even needed. It was copied nearly verbatim from the existing http module. Given how we are reading and writing data from the socket at the native layer, I doubt that this code would even be exercised in normal cases. Perhaps we should just remove these handlers and see if it makes any difference whatsoever :-)

[apapirovski]

Do you have strong feelings one way or another re: the Proxy socket implementation living on the session instead if in compatibility? It would give us the option of throwing when users try to directly write, read, etc. as we currently do in compatibility mode. (Assuming the performance turns out to be fine, of course.)

(The real socket would still live at kSocket and be used internally.)

[jasnell]

I think that makes sense. At some point in the future it might make sense to have some kind of official escape hatch for getting to the actual socket, but for now, hiding it behind the proxy makes the most sense as long as the proxy is never used internally.

[trivikr]

@mcollina @jasnell @apapirovski
Can one of you create an issue for discussion so that this pull request can be closed without landing?

[trivikr]

Closing as issue is created at #16252