Detecting threading problems faster

[Andy Goryachev]

One should probably use System.nanoTime() instead of .currentTimeMillis() ...

-andy

From: openjfx-dev <openjfx-dev-retn at openjdk.org> on behalf of Thiago Milczarek Sayão <thiago.sayao at gmail.com>
Date: Thursday, August 8, 2024 at 03:47
To: John Hendrikx <john.hendrikx at gmail.com>
Cc: openjfx-dev at openjdk.org <openjfx-dev at openjdk.org>
Subject: Re: Detecting threading problems faster
Nice! I will steal it if you don't mind.

Em qua., 7 de ago. de 2024 às 15:55, John Hendrikx <john.hendrikx at gmail.com<mailto:john.hendrikx at gmail.com>> escreveu:

For detecting long running tasks on the FX thread, there are some other options which you can do as a user (but perhaps we can support it directly within FX).  I use this kind of code to detect long running things on the FX thread:

/**

* Adds a slow event warning whenever an event takes more than 10 ms to process. Note

* that time spent in nested event loops cannot be properly taken into account as time

* spent in nested event loops will be part of the event that triggered it giving false

* positives. In order for this time to be accurately reflected, the methods to enter

* a nested event loop in this class should be used instead of the ones in {@link Platform}.

*

* @param scene a Scene to which to add the slow event warning detection, cannot be null

*/

public static void addSlowEventWarning(Scene scene) {

final EventDispatcher eventDispatcher = scene.getEventDispatcher();

scene.setEventDispatcher(new EventDispatcher() {

private ScheduledFuture<StackTraceElement[]> future;

@Override

public Event dispatchEvent(Event event, EventDispatchChain tail) {

if(future != null) {

future.cancel(false);

}

long startTime = System.currentTimeMillis();

TIME_TRACKER.enterNested(startTime); // nesting can happen in two ways, an event triggering another event, or when a nested event loop is entered

Event returnedEvent = eventDispatcher.dispatchEvent(event, tail);

long endTime = System.currentTimeMillis();

long timeSpentInNested = TIME_TRACKER.exitNested(endTime);

if(timeSpentInNested > 10) {

long total = endTime - startTime;

LOGGER.warning("Slow Event (self/total: " + timeSpentInNested + "/" + total + " ms @ level " + TIME_TRACKER.getCurrentLevel() + "): " + event);

}

return returnedEvent;

}

});

}

--John
On 05/08/2024 17:17, Thiago Milczarek Sayão wrote:
Hi,

Interesting idea. We have this problem specially when Junior developers touch the code.

The other way around would be nice too - if some I/O task executes on the FX thread.

This can make the OS think the application hanged and offer to kill it, since it won't respond to "pings". And I/O tasks processing time may vary between installations. Also causes "white screens" since it blocks painting.

-- Thiago.

Em seg., 5 de ago. de 2024 às 11:59, Kevin Rushforth <kevin.rushforth at oracle.com<mailto:kevin.rushforth at oracle.com>> escreveu:
Wouldn't it be better to implement this check in assert to avoid any impact in production?

No. Using an assert in a case like this is an anti-pattern. A call to assert in a library such as JavaFX is only appropriate for checking an invariant in internal logic. If we are going to go down this route of doing a thread check when mutating properties of "live" nodes, we will throw the same IllegalStateException that is currently thrown by some methods on Stage and Scene.

As for the proposal itself, adding this check is an interesting idea. We considered doing this back in the JDK 7 (JavaFX 2) time frame, but decided not to pursue it then. I think the idea is worth further discussion. I would limit any thread checking to setting the property. It would be too restrictive (and largely unnecessary) to prevent reading a property from the application thread.

The things to consider would be:

1. What is the performance hit of doing this check on the setting of every property?
2. What is the effect on bound properties?
3. How intrusive is it in the code?
4. Should we add a property to enable / disable the thread check, possibly a three- or four-valued property (allow|warn|debug?|deny), as was recently done in JEP 471 for sun.misc.Unsafe memory access methods. If so, what should the default be?

My quick take is that if this can be done in a minimally intrusive manner with low overhead, we should consider pursing this. As for 4, my preference would be to add a three- or four-valued system property to control the check, with "warn" as the default initially, changing the default to "disallow" in a subsequent version. This would, of course, require a lot of testing.

-- Kevin

On 8/4/2024 8:40 PM, quizynox wrote:
Hello,

Wouldn't it be better to implement this check in assert to avoid any impact in production?

пн, 5 авг. 2024 г. в 03:30, John Hendrikx <john.hendrikx at gmail.com<mailto:john.hendrikx at gmail.com>>:
Hi list,

I know of quite some bugs and users that have been bitten by the
threading model used by JavaFX.  Basically, anything directly or
indirectly linked to an active Scene must be accessed on the FX thread.
However, as FX also allows manipulating nodes and properties before
they're displayed, there can be no "hard" check everywhere to ensure we
are on the FX thread (specifically, in properties).

Now, I think this situation is annoying, as a simple mistake where a
Platform.runLater wrapper was forgotten usually results in programs
operating mostly flawlessly, but then fail in mysterious and random and
hard to reproduce ways.  The blame is often put on FX as the resulting
exceptions will almost never show the user code which was the actual
culprit.  It can result in FX being perceived as unstable or buggy.

So I've been thinking if there isn't something we can do to detect these
bugs originating from user code much earlier, similar to the
`ConcurrentModificationException` the collection classes do when
accessed in nested or concurrent contexts.

I think it may be possible to have properties check whether they're part
of an active scene without too much of an performance impact, possibly
even behind a switch. It would work like this:

Properties involved with Nodes will have an associated bean instance
(`getBean`).  This is an object, but we could check here if this
instance implements an interface:

      if (getBean() instanceof MayBePartOfSceneGraph x) {
            if (x.isPartOfActiveScene() && !isOnFxThread()) {
                 throw new IllegalStateException("Property must only be
used from the FX Application Thread");
            }
      }

This check could be done on every set of the property, and potentially
on every get as well.  It should be relatively cheap, but will expose
problematic code patterns at a much earlier stage.  There's a chance
that this will "break" some programs that seemed to be behaving
correctly as well, so we may want to put it behind a switch until such
programs (or libraries) can be fixed.

What do you all think?

--John

(*) Names of methods/interfaces are only used for illustration purposes,
we can think of good names if this moves forward.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/openjfx-dev/attachments/20240808/3df44dc7/attachment-0001.htm>