The modernization of PCIe hotplug in Linux

I've been struggling with this in my own adventures with using an eGPU over Thunderbolt. Unfortunately, I expect that PCIe-level hotplug improvements will only be half of the battle, as kernel modules for certain device types (certainly GPUs) respond violently to having their backing devices disappear, even if it happens cleanly at the PCIe level. Manually unloading the kernel modules before removing the device sometimes works, but it's often the case that they refuse to be unloaded because they're in-use by other kernel modules or userland. Perhaps there's a systematic way to hunt down and remove anything blocking the module unloading to prepare for disconnection, but I'm not convinced that's the right answer for user experience.

For PCIe hotplug (including removal) to "just work," kernel modules are going to need to get a lot more accustomed to an inversion of dependency removal. As suggested in "surprise removal," the traditional expectation has been an orderly retirement of consumers before their corresponding producers (to use the API terminology). We now have to accommodate the opposite: surprise electrical removal, followed by surprise PCIe device removal, followed by a kernel module surprising its consumers with an inability to service them, followed by graceful handling on downward. The ability to expose practically any PCIe device over Thunderbolt has changed the scope of hot removal support from a few types of server-style components (e.g. NVMe) to literally anything that may run over PCIe.

I would love to reach the point where I can safely and consistently unplug a Thunderbolt eGPU from my Linux machine -- without shutting it down first.