handle_wrap: expose {un}refed check to JS

[Fishrock123]

Pull Request check-list

• Does make -j8 test (UNIX) or vcbuild test nosign (Windows) pass with
  this change (including linting)?
• Is the commit message formatted according to CONTRIBUTING.md?

Affected core subsystem(s)

handle_wrap

Description of change

Give tools a consistent way to check if a handle is unrefed that works for all of our APIs that expose handles. E.g. would be very helpful to see which handles are umrefed in @AndreasMadsen's https://github.com/AndreasMadsen/dprof

(Some APIs, like timers, do not normally expose handles and are not addressed at that level here)

Refs: #5828
Refs: #5827

R= @trevnorris
cc @thlorenz

[Fishrock123]

idk if it's better to have this be isUnrefed?

[cjihrig]

It probably should be isUnrefed(), but isRefed() seems a lot simpler to think about.

[mscdex]

I think isRefed() sounds better ("is this holding the event loop open?").

[trevnorris]

This is not the way it's done. This has to be declared in every child class: https://github.com/nodejs/node/blob/v5.9.0/src/tcp_wrap.cc#L87

Otherwise you're setting an object property on an already constructed object.

[Fishrock123]

oh that kinda sucks, alright.

[Fishrock123]

(does this needs tests..?)

[Fishrock123]

Also: is it ok to return nothing of the handle is not alive?

[trevnorris]

returning nothing == returning undefined. but the JS signature returns bool. The question this is trying to answer is "Can this handle hold open the event loop?" So if !IsAlive() means should return false.

[Fishrock123]

Noted, thanks.

[trevnorris]

So basically this can be boiled down to:

bool refed = IsAlive(wrap) && (wrap->flags_ & kUnref) == 0;
args.GetReturnValue().Set(Boolean::New(env->isolate(), refed);

[thlorenz]

Yes, I think this needs tests. Ideally you'd pass tests similar to the ones in the PRs you referenced.
At that point we can also safely let this PR supersede those.

[mscdex]

Yes, tests are needed.

[Fishrock123]

Does anyone have advice on how I can make a more generic test for this rather than testing everywhere that we expose a handle?

[thlorenz]

@Fishrock123 I guess you could test on C++ level (do we have C++ level tests?) and just create a handle and interact with it directly.
However I also believe that you should test this from a bit higher level for each case as well, at least for timers, sockets, pipes, etc.

[jasnell]

Does it have to be a method? Is it possible to have it as a read-only property? .refed

A quick test case would be worthwhile also.

[Fishrock123]

@jasnell I guess this makes it explicit that you cannot just set it, but I could possibly change it otherwise.

[jasnell]

@Fishrock123 ... that makes sense. The additional () isn't too difficult to type ;-). LGTM with test cases added.

[trevnorris]

The implementation is incomplete. As noted in #5834 (comment) the method needs to be placed on the FunctionTemplate of the child class. This is tedious, but haven't found a better way to automate propagation of these JS methods to the C++ child classes. @bnoordhuis had an idea a while back, but can't remember what it was. Ben, you recall anything about this?

The test is simple enough:

const assert = require('assert');
const net = require('net');

const server = net.createServer(() => {}).listen(common.PORT);
server.unref();
assert.equal(false, server._handle.isRefed());

Remember this won't cover timer cases. To make the API parallel may need to implement a JS version that can be called on the JS timer handle instances.

[AndreasMadsen]

@Fishrock123 thanks for doing this. It would also be useful for just checking what keeps a process alive at any given time.

[Fishrock123]

Updated with a test, some things:

I realized that some _handles we expose directly extend AsyncWrap instead of HandleWrap, and this currently does not cover those. To cover those that would mean moving the c++ check into AsyncWrap I think, and I'm not sure if that is possible/acceptable.

I'm not sure if it is possible to test spawn a net.Socket that uses a Pipe handle without making a child process, so I've just gone ahead and directly made a Pipe.

TTY's cannot be re-refed for some reason. The method does not exist in c++, and does not work if it is added.

[Fishrock123]

CI: https://ci.nodejs.org/job/node-test-pull-request/2025/

[Fishrock123]

Hmm, getting this on Windows vs2015 and vcbt2015:

not ok 86 test-handle-wrap-isrefed.js
# events.js:155
# throw er; // Unhandled 'error' event
# ^
# 
# Error: read EISDIR
# at exports._errnoException (util.js:893:11)
# at TTY.onread (net.js:552:26)

Anyone know how to fix this test for that?

  const assert = makeAssert('isRefed() not working on tty_wrap');
  const ReadStream = require('tty').ReadStream;
  const tty = new ReadStream();
  assert(Object.getPrototypeOf(tty._handle).hasOwnProperty('isRefed'), true);
  assert(tty._handle.isRefed(), true);
  tty.unref();
  assert(tty._handle.isRefed(), false);

[cjihrig]

Do you need to pass a fd to the ReadStream() constructor?

Once the CI is happy, LGTM.

[trevnorris]

This depends on some non-obvious implementation details. If the timer is unref'd then the JS object will be moved to a TimerWrap that also isn't ref'd, correct?

[Fishrock123]

See

node/lib/timers.js

Lines 460 to 485 in d3a7534

	Timeout.prototype.unref = function() {
	if (this._handle) {
	this._handle.unref();
	} else if (typeof this._onTimeout === 'function') {
	var now = TimerWrap.now();
	if (!this._idleStart) this._idleStart = now;
	var delay = this._idleStart + this._idleTimeout - now;
	if (delay < 0) delay = 0;
	// Prevent running cb again when unref() is called during the same cb
	if (this._called && !this._repeat) {
	exports.unenroll(this);
	return;
	}
	var handle = reuse(this);
	this._handle = handle || new TimerWrap();
	this._handle.owner = this;
	this._handle[kOnTimeout] = unrefdHandle;
	this._handle.start(delay, 0);
	this._handle.domain = this.domain;
	this._handle.unref();
	}
	return this;
	};

timer.unref() assigns a ._handle for that specific timer, and is unlikely to change in the short term.

[trevnorris]

I realize that. What I failed to say is that a note of this should be placed in timer_wrap.cc where the method is added so future generations can immediately see why this non-obvious implementation detail actually works.

[Fishrock123]

Where exactly do you mean? Isn't the same thing for the un(ref) methods on timer_wrap?

[trevnorris]

My point is it may be non-obvious why setTimeout(()=>{}, 100)._handle === undefined but setTimeout(()=>{}, 100).unref()._handle links to a Timer instance. So using isRefed() is conditional to the state of timer. So not having a uniform way to access that data from the public object instance could be confusing.

[Fishrock123]

Sure, but the regular timers implementation will also benefit from this. e.g.

node/lib/timers.js

Lines 210 to 214 in cf94929

	if (list._unrefed === true) {
	delete unrefedLists[msecs];
	} else {
	delete refedLists[msecs];
	}

So I'm not sure what I would say in handle_wrap

[Fishrock123]

Ci with fd 0 for TTY: https://ci.nodejs.org/job/node-test-pull-request/2030/

[Fishrock123]

@cjihrig still the same thing on those windows builds/platforms..

[cjihrig]

Crap. The only real use of tty.ReadStream I found was in src/node.js.

[Fishrock123]

maybe cc @bnoordhuis

[Fishrock123]

@nodejs/lts think we can get this in the next minor LTS release?

[jasnell]

Should be possible. It's just not yet clear when that next minor LTS release will be.

[Fishrock123]

OK in hindsight this behavior regarding isAlive() is a huge mistake and this should not ship as is.

This needs to be accurate enough for timers to use too, and currently it isn't with either the following patches:

diff --git a/lib/timers.js b/lib/timers.js
index dc2506e..3a57d86 100644
--- a/lib/timers.js
+++ b/lib/timers.js
@@ -133,8 +133,6 @@ function insert(item, unrefed) {
     list = new TimersList(msecs, unrefed);
     L.init(list);
     list._timer._list = list;
-
-    if (unrefed === true) list._timer.unref();
     list._timer.start(msecs, 0);

     lists[msecs] = list;
@@ -149,7 +147,7 @@ function TimersList(msecs, unrefed) {
   this._idleNext = null; // Create the list with the linkedlist properties to
   this._idlePrev = null; // prevent any unnecessary hidden class changes.
   this._timer = new TimerWrap();
-  this._unrefed = unrefed;
+  if (unrefed === true) this._timer.unref();
   this.msecs = msecs;
 }

@@ -207,7 +205,7 @@ function listOnTimeout() {
   debug('%d list empty', msecs);
   assert(L.isEmpty(list));
   this.close();
-  if (list._unrefed === true) {
+  if (list._timer.isRefed() === false) {
     delete unrefedLists[msecs];
   } else {
     delete refedLists[msecs];

diff --git a/lib/timers.js b/lib/timers.js
index dc2506e..cdc0d1c 100644
--- a/lib/timers.js
+++ b/lib/timers.js
@@ -133,8 +133,6 @@ function insert(item, unrefed) {
     list = new TimersList(msecs, unrefed);
     L.init(list);
     list._timer._list = list;
-
-    if (unrefed === true) list._timer.unref();
     list._timer.start(msecs, 0);

     lists[msecs] = list;
@@ -149,7 +147,7 @@ function TimersList(msecs, unrefed) {
   this._idleNext = null; // Create the list with the linkedlist properties to
   this._idlePrev = null; // prevent any unnecessary hidden class changes.
   this._timer = new TimerWrap();
-  this._unrefed = unrefed;
+  if (unrefed === true) this._timer.unref();
   this.msecs = msecs;
 }

@@ -207,10 +205,10 @@ function listOnTimeout() {
   debug('%d list empty', msecs);
   assert(L.isEmpty(list));
   this.close();
-  if (list._unrefed === true) {
-    delete unrefedLists[msecs];
-  } else {
+  if (list._timer.isRefed() === true) {
     delete refedLists[msecs];
+  } else {
+    delete unrefedLists[msecs];
   }
 }

[Fishrock123]

So... it does work if I close after the delete (duh), but the behavior is still pretty confusing..

Will probably remove the isAlive() check. (Which can be exposed separately if necessary)

[MylesBorins]

@Fishrock123 you may want to chime into the v6 thread before the rc is cut if you don't think this should land

[Fishrock123]

For posterity, landing this requires this commit chain:

• 7d8882b (this PR's commit)
• f938ef7 src: fix -Wunused-variable compiler warning #6129
• 9bb5a5e handle_wrap: IsRefed -> Unrefed, no isAlive check #6204
• cad1a62 src: use libuv's refcounting directly  #6395
• a58d483 src: use libuv's refcounting directly  #6395
• 4282405 src: use libuv's refcounting directly  #6395
• 96a9439 Revert "handle_wrap: IsRefed -> Unrefed, no isAlive check" #6546
• fe4837a Revert "handle_wrap: IsRefed -> Unrefed, no isAlive check" #6546
• 2c480bd src,tools: rewrite check-imports and fix linter errors #6105