On Fri, 1 Mar 2024 18:44:17 GMT, Marius Hanl <mh...@openjdk.org> wrote:
>> This PR fixes the dialog freeze problem once and for all. >> >> This one is a bit tricky to understand, here is how it works: >> This bug happens on every platform, although the implementation of nested >> event loops differs on every platform. >> E.g. on Linux we use `gtk_main` and `gtk_main_quit`, on Windows and Mac we >> have an own implementation of a nested event loop (while loop), controlled >> by a boolean flag. >> >> Funny enough, the reason why this bug happens is always the same: Timing. >> >> 1. When we hide a dialog, `_leaveNestedEventLoop` is called. >> 2. This will call native code to get out of the nested event loop, e.g. on >> Windows we try to break out of the while loop with a boolean flag, on Linux >> we call `gtk_main_quit`. >> 3. Now, if we immediately open a new dialog, we enter a new nested event >> loop via `_enterNestedEventLoop`, as a consequence we do not break out of >> the while loop on Windows (the flag is set back again, the while loop is >> still running), and we do not return from `gtk_main` on Linux. >> 4. And this will result in the Java code never returning and calling >> `notifyLeftNestedEventLoop`, which we need to recover the UI. >> >> So it is actually not trivial to fix this problem, and we can not really do >> anything on the Java side. We may can try to wait until one more frame has >> run so that things will hopefully be right, but that sounds rather hacky. >> >> I therefore analyzed, if we even need to return from >> `_enterNestedEventLoop`. Turns out, we don't need to. >> There is a return value which we actually do not use (it does not have any >> meaning to us, other that it is used inside an assert statement). >> Without the need of a return value, we also do not need to care when >> `_enterNestedEventLoop` is returning - instead we cleanup and call >> `notifyLeftNestedEventLoop` in `_leaveNestedEventLoop`, after the native >> code was called. >> >> Lets see if this is the right approach (for all platforms). >> Testing appreciated. >> # >> - [x] Tested on Windows >> - [x] Tested on Linux >> - [x] Tested on Mac >> - [ ] Tested on iOS (although the nested event loop code is the same as for >> Mac) (I would appreciate if someone can do this as I have no access to an >> iOS device) >> - [ ] Adjust copyright >> - [ ] Write Systemtest > > Marius Hanl has updated the pull request with a new target base due to a > merge or a rebase. The incremental webrev excludes the unrelated changes > brought in by the merge/rebase. The pull request contains four additional > commits since the last revision: > > - Merge branch 'master' of https://github.com/openjdk/jfx into > JDK-8285893-dialog-freezing-🥶 > - Merge remote-tracking branch 'openjfx/master' into > JDK-8285893-dialog-freezing-🥶 > - JDK-8285893: Decrement nestedEventLoopCounter in leaveNestedEventLoop > - JDK-8285893: Hiding dialog and showing new one causes dialog to be frozen The first `runLater` block is being invoked by the first nested event loop; that loop is on the call stack. In that block you call `exitNestedEventLoop` but don't allow the stack to unwind before calling `enterNestedEventLoop`. So on the stack the new event loop is indeed nested inside the previous one (the previous one is marked as LEAVING but it's still doing something). Since this is a stack there's only one way to unwind it which is in reverse order. The output I would expect would be: enter1 exit1 enter2 exit2 456 123 There is that bug in the invokeLaterDispatcher which will prevent the second `runLater` block from executing. And the documentation doesn't really address this case but it should. By the way, when I grab the code in the original bug report and run it agains this PR the innermost modal is not blocked and everything seems fine. But if I press the Close button exceptions start getting thrown. This is happening on both Mac and Windows. ------------- PR Comment: https://git.openjdk.org/jfx/pull/1324#issuecomment-1973953042
