On Thu, 11 Jan 2024 23:54:13 GMT, Kevin Rushforth <k...@openjdk.org> wrote:
>> As described in the JBS bug, there is a long-standing deadlock that happens >> on macOS between the AWT EDT and the JavaFX Application thread (which on >> macOS is the AppKit thread) when processing Input Method Events (IME) in a >> WebView node in a JFXPanel. >> >> This PR fixes the deadlock by adding `"AWTRunLoopMode"` to the list of modes >> that will accept the calls to `performSelectorOnMainThread` used by >> `_submitForLaterInvocation` and `_invokeAndWait`. These two native methods >> are the macOS implemention `runLater` and `runAndWait`, respectively, and >> are used to run Java Runnables on the JavaFX Application Thread. >> >> This deadlock is happening much more often on recent macOS versions, and in >> macOS 14, pressing CAPS LOCK is sufficient to trigger IME calls. >> >> The OS calls the AWT IME methods on the AppKit thread, which is also the >> JavaFX Application Thread. The AWT IME callback methods, for example >> `characterIndexForPoint`, call invokeAndWait to run the IME handler on the >> AWT Event thread. In the case of JFXPanel, the registered IME handler is in >> JavaFX code and we often need to run something on the JavaFX Application >> Thread to avoid threading problems or crashes. See >> [JDK-8196011](https://bugs.openjdk.org/browse/JDK-8196011) and >> [JDK-8322703](https://bugs.openjdk.org/browse/JDK-8322703). >> >> A similar deadlock was observed while trying to fix a threading problem in >> JFXPanel's handling of the IME calls as described in >> [JDK-8322784](https://bugs.openjdk.org/browse/JDK-8322784), so it isn't >> necessarily limited to using WebView. >> >> Anton Tarasov @forantar proposed a fix (in a comment in the JBS bug) in AWT >> that would change the invokeAndWait calls used by the IME methods in >> CInputMethod to call `doAWTRunLoop` with the `processEvents` flag set to >> true, which will allow all events to be processed while in `doAWTRunLoop`. >> >> NOTE: I had created Draft PR openjdk/jdk#17290 with Anton's proposed fix, >> expanded to cover all of the cases, but during the discussion it was pointed >> out that `doAWTRunLoop` should already be able to run selectors via >> `performSelectorOnMainThread` -- see [this >> comment](https://github.com/openjdk/jdk/pull/17290#issuecomment-1880001813). >> The reason `doAWTRunLoop` doesn't process our messages is that it runs the >> loop in a mode that only handles messages that include a custom >> `"AWTRunLoopMode"` mode. >> >> Adding `"AWTRunLoopMode"` to the `performSelectorOnMainThread` calls used by >> `_submitForLaterInvocation` and `_invokeAndWa... > > Kevin Rushforth has updated the pull request incrementally with one > additional commit since the last revision: > > Remove reference to JBS bug ID Marked as reviewed by prr (Reviewer). I think there's a bit of additional explanation joining the dots that would be helpful here > The AWT IME callback methods, for example characterIndexForPoint, call > invokeAndWait to run the IME handler on the AWT Event thread. And that means specifically call AWT's LWCToolkit.invokeAndWait() and THAT uses doAWTRunLoop(), which when entered uses the flag 'inAWT' (which will be true) to invoke NSRunLoop (in a loop) in a way that it should ONLY process requests that specify the special mode [[NSRunLoop currentRunLoop] runMode:(inAWT ? [ThreadUtilities javaRunLoopMode] : NSDefaultRunLoopMode) It stays in this loop until it is informed that the thing it was waiting for is done. This won't happen unless FX specifies that mode so its call is eligible to be processed. ------------- PR Review: https://git.openjdk.org/jfx/pull/1327#pullrequestreview-1817052712 PR Comment: https://git.openjdk.org/jfx/pull/1327#issuecomment-1888179598
