Timers broken in 0.10.34

[apendua]

After updating to 0.10.34 my app keeps throwing this error:

timers.js:432
    timeSinceLastActive = now - cur._monotonicStartTime;
                                   ^
TypeError: Cannot read property '_monotonicStartTime' of null
    at Timer.unrefTimeout [as ontimeout] (timers.js:432:36)

well, I looked into your source code here:

https://github.com/joyent/node/blob/v0.10.34-release/lib/timers.js#L430

and it seems that something is going wrong there, right?

  var cur = unrefList._idlePrev;
  while (cur != unrefList) {
    timeSinceLastActive = now - cur._monotonicStartTime;

    if (timeSinceLastActive < cur._idleTimeout) {

Well I think you forgot to make unrefList a Timer object here:

https://github.com/joyent/node/blob/v0.10.34-release/lib/timers.js#L508

it's just a guess based on the fact that _idlePrev is only initialized in the Timer constructor.

[apendua]

BTW, don't guys test your software before releasing it to public?

[micnic]

do you have a minimal code to produce this exception?

[apendua]

@micnic Good question. I am currently at work, but will try to help you with that later this day.

[misterdjules]

Thanks @apendua for reporting the issue, I'm investigating. Please let us know when you find the code that reproduces it.

[apendua]

@misterdjules Currently working on this one.

[ploer]

Just wanted to add that I started seeing this one in production as well on upgrade to 0.10.34; haven't yet been able to find a simple reproduction but will keep looking.

[ploer]

The following code demonstrates the same error for me:

var timers = require('timers');

var foo = new function() {
  this._onTimeout = function() {
    console.log("in foo._onTimeout");
  }
}();

var bar = new function() {
  this._onTimeout = function() {
    console.log("in bar._onTimeout");
    timers.unenroll(foo);
  }
}();

timers.enroll(bar, 1000);
timers._unrefActive(bar);

timers.enroll(foo, 3000);
timers._unrefActive(foo);

setTimeout(function() {}, 5000);

Output:

ploer-macbook-4:timer_test ploer$ node t2.js
unrefList initialized
unrefTimer initialized
unrefTimer scheduled
unrefList append to end
unrefList append to end
unrefTimer fired
unreftimer firing timeout
in bar._onTimeout
unenroll
unreftimer firing timeout
in foo._onTimeout

timers.js:432
    timeSinceLastActive = now - cur._monotonicStartTime;
                                   ^
TypeError: Cannot read property '_monotonicStartTime' of null
    at Timer.unrefTimeout [as ontimeout] (timers.js:432:36)
ploer-macbook-4:timer_test ploer$ 

[ploer]

And under 0.10.33, this same code appears to work fine:

ploer-macbook-4:timer_test ploer$ node --version
v0.10.33
ploer-macbook-4:timer_test ploer$ node t2.js
unrefList initialized
unrefTimer initialized
unrefList empty
unrefTimer scheduled
unrefList find where we can insert
unrefList append to end
unrefTimer fired
unreftimer firing timeout
in bar._onTimeout
unenroll
unrefList is empty
timeout callback 5000
now: 104385711
5000 list empty
ploer-macbook-4:timer_test ploer$ 

[ploer]

As a side note, with the new list traversal, the following lines at timers.js:457 look like they might be a problem if they ever triggered:

    if (!cur._onTimeout) continue;

    if (domain && domain._disposed)
      continue;

It doesn't look to me like cur has been advanced down the list at this point, so it seems like it might just loop indefinitely on either of these conditions.

[ploer]

The source of the bug appears to be that when you call itemToDelete._onTimeout(), any node in the list can be mutated, so holding open a list iteration and continuing it after this call is unsafe.

With the current list implementation, it seems like to be safe, you would need to do something like scan the full list, and build a local list of all of the items that look like they're ready to fire, and only then iterate over that list and call each item's _onTimeout() (and I guess check as you iterated to see if any of them had become unenrolled in the meantime).

This issue appears to have been introduced in 934bfe2.

[misterdjules]

@ploer Thanks for the minimal reproduction. Yes, the block of code that early-outs in the loop seems problematic, we may need to open a separate issue for that.

[misterdjules]

I'm working on a fix, I'll keep you posted. Thank you again!

[trevnorris]

@apendua

BTW, don't guys test your software before releasing it to public?

I appreciate you taking the time to report this issue, but blatantly snide remarks like this aren't welcome. If you'd taken the time to look at the commit where this code was changed (934bfe2) you'd see there was one additional test added. Not including all the other additional timers tests that are already there.