On Android, we run the test suite embedded in a standard app, in order to be representative of the environment where Python is most likely to be used. This means there's a native thread that uses sigwait to listen for SIGUSR1 in a loop, and responds by triggering the Java garbage collector and logging a message. This is only used for debugging purposes, and in most cases it doesn't matter because any signal handler will take priority over it.

However, in ThreadSignals.test_signals there's a background thread which sends both SIGUSR1 and SIGUSR2, and then immediately exits. So if the background thread is unavailable because it’s exited, and the main thread is unavailable because it’s already processing one of the signals (see complete_signal in kernel/signal.c), then the other signal may be delivered to Android’s sigwait loop instead of the test's own handler. This happens about 1 time out of 4.

This PR fixes that by sending the signals using raise_signal (which is directed at the current thread) rather than os.kill (which is directed at the whole process). Not only does this avoid the above scenario, it also strengthens the test by verifying that a signal which is delivered to a background thread, and not merely sent by it, still runs the Python-level handler on the main thread.

Since raise "shall not return until after the signal handler does", this also allows for the removal of a whole block of code which waits for the signal to arrive.

• Issue: Meta-issue: support of the android platform #71052