...

quoted

It's not a problem to change this patchset. The problem is that if
you'll grep mainline for 'pm_runtime_disable', you will find that there
are a lot of drivers in a potential trouble.

Let's start by fixing this patchset, please - then we can consider
what to do with the other cases separately.

Yeah, should be better to discuss it separately.

...

quoted

 void __pm_runtime_disable(struct device *dev, bool check_resume)
 {
+       flush_work(&dev->power.work);
+

What about the latency this may introduce? I am not sure that is
acceptable here!?

I'm not aware about any code which relies on the original 'cancelling'
behaviour, perhaps Rafael should have more insight.

...

quoted

The sysfs rpm-forbid is a separate problem and it's less troublesome
since it requires root privileges. It's also not something that
userspace touches casually. For now I don't know what could be done
about it.

As I said, the common method to address this problem is to run the
following sequence:

pm_runtime_get_sync()
"power off the device"
pm_runtime_disable()
pm_runtime_put_noidle()

This works even if user space, via sysfs, has triggered a call to
pm_runtime_forbid(). Or doesn't it?

If you don't like it, pm_runtime_force_suspend() should work too, at
least for your cases, I believe.

I'll update the patches, thank you.