Preemption of userspace PCI device model

[mills-ng]

I am using libvfio-user to model a PCI device for QEMU. Occasionally, the PCI device model process gets preempted by the host, resulting in the kernel driver running in QEMU thinking that the device has timed out. The kernel driver then begins timeout recovery logic, only for the host to resume the PCI device model process. This results in a non-sensical PCI device state within the kernel driver.

I think the correct solution to this is to harden the timeout logic in the kernel driver code, but I was wondering if anyone knows of a way to have QEMU pend for vfio device operations rather than asynchronously continue kernel execution. A single iowrite32() call in a kernel driver would realistically never take such an excessive amount of time to complete that it would cause a driver timeout.

[tmakatos]

This is vfio-user client side question, so it's better to ask in qemu-devel, @jlevon is the maintainer.

[tmakatos]

Oh and BTW, we'd love to know what kind of device you're modelling!

[mills-ng]

Linking qemu-devel post in case it ever gets a response.

I'm modeling a PCIe device with several UARTS, ADCs, and NAND.