Detecting threading problems faster

[John Hendrikx]

Hi Andy,

Problems that occur due to modifying the scene graph (like 
adding/removing children, or changing certain properties) can be 
incredibly subtle, and almost never point out the code that is the 
actual culprit.  I suspect that the PR below is one of those for 
example, but there are likely many more.

https://github.com/openjdk/jfx/pull/1123

Issues that complain about some kind of odd bug in the FX code (like a 
NPE, ConcurrentModificationException, ArrayIndexOutOfBoundsException, 
etc) could all be caused by user code doing things on the wrong thread, 
but its really hard to diagnose as the stack traces provided will never 
point out the user code that may be the culprit.  I've investigated PR 
1123 extensively, and I can't see any bugs in the FX code at all 
**assuming** it is used from a single thread...

--John

On 05/08/2024 17:01, Andy Goryachev wrote:
>
> John:
>
> Can you cite a bug or give an example of such a problem?
>
> During all my time of actively using javafx (since about 2015) I have 
> never encountered an issue with threading you describe.  It is 
> possible that I use the system trivially and not to the full extent.  
> Both swing and javafx are single threaded by design.  Yes, there might 
> be occasions when one can use multiple threads and it is sort of 
> allowed, but doing so may lead to unpleasant issues down the road, so 
> the question is why would you want to do that?
>
> -andy
>
> *From: *openjfx-dev <openjfx-dev-retn at openjdk.org> on behalf of John 
> Hendrikx <john.hendrikx at gmail.com>
> *Date: *Sunday, August 4, 2024 at 16:30
> *To: *openjfx-dev at openjdk.org <openjfx-dev at openjdk.org>
> *Subject: *Detecting threading problems faster
>
> 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/20240807/59c9f2b4/attachment.htm>