Currently, all I/O requests are queued internally, and a number of
threads will pick I/O requests of the queue and send them to CTL with
a synchronous CTL_IO ioctl. On the other hand, TMF requests are sent
to CTL immediately using the same synchronous ioctl.

Besides being unworkable for non-CTL backends such as for SCSI
passthrough, this simple approach may easily run into situations
where a TMF request operating on a particular I/O request is sent
to CTL while it is still on our queue and thus unknown to CTL.
In addition, for target and/or LUN resets we should really clear
our queue and return all outstanding I/O requests with a proper
status.

This is currently also under review in illumos: https://code.illumos.org/c/illumos-gate/+/4424