when the driver goes into an error state (waiting for fifo to unmess itself or
other things), where the user just kills the hang process, nouveau decides to
suspend the gpu allthough some kworker are inside some while loops and waiting
for conditions to be true.

Because the card went offline reading stuff through mmio hangs and leaves the
kernel in a really strange state, where usb/network/... can just go corrupt.

That's why the nouveau driver should add more points where the card is marked
as in use, so that runpm won't suspend the card allthough it really should not.