[m3nu]

Description

SchedulerState._announce_job queued jobs_changed with QTimer.singleShot(0, self.scheduler.jobs_changed.emit). The pending call does not keep the scheduler alive. If the scheduler is freed before the event loop delivers the call, emit runs on a freed QObject: a segfault, or AttributeError: 'VortaScheduler' does not have a signal with the signature jobs_changed() when the memory has been reused.

_announce_job now queues the signal with QMetaObject.invokeMethod(self.scheduler, 'jobs_changed', QueuedConnection). The event is posted to the scheduler itself, so Qt discards it if the scheduler is deleted first.

Related Issue

Root cause of #2595. Also the macOS failure of #2621.

Motivation and Context

The app's own scheduler lives as long as the app, so users are very unlikely to hit this. The unit tests hit it on macOS: init_db teardown drops the backup_finished_event connection that keeps each test-created VortaScheduler alive, while one of its announcements is still queued. A later GC pass frees the scheduler, and the next test's qtbot.waitSignal delivers the call. That's why the crash always showed up inside a later test's event loop and moved with GC timing (test_recording_a_skip_emits_jobs_changed in #2595, test_create_backup_records_a_running_job in #2621, 2 of 2 attempts each).

The Python traceback only showed the event loop. A native crash report from a diagnostic CI run (.ips from ~/Library/Logs/DiagnosticReports) showed the main thread in sendPostedEvents → PyQtSlotProxy::unislot → PyQtSlot::call → cfunction_call, faulting at 0x45. A standalone script reproduces it 3 of 3 times:

o = Obj()  # QObject with a reference cycle
QtCore.QTimer.singleShot(0, o.sig.emit)
del o; gc.collect()
app.exec()  # segfault

The same script with invokeMethod doesn't crash, and still delivers the signal once when the object is alive.

How Has This Been Tested?

• New tests/unit/test_scheduler.py::test_announced_job_outlives_its_scheduler queues an announcement on a scheduler, frees the scheduler and spins the event loop. 3 of 3 passed with the fix. 3 of 3 failed without it (one segfault, two of the AttributeError above).
• tests/unit locally on macOS, Python 3.12, Borg 1.4.3: 626 passed, 8 skipped.
• Return consistent parent() rows and bounds-check index() in FileTreeModel #2621's commit plus this fix on the macOS CI job that had crashed twice: 580 passed, 7 skipped (run 43716244818).
• make lint hooks pass on the changed files.

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)

Checklist:

• I have read the CONTRIBUTING guide.
• My code follows the code style of this project.
• My change requires a change to the documentation.
• I have updated the documentation accordingly.
• I have added tests to cover my changes.
• All new and existing tests passed.

I provide my contribution under the terms of the license of this repository and I affirm the Developer Certificate of Origin.