It has been pointed out by the USB people that the mid-layer doesn't
obey its own online flag.

The attached patch should fix this.  However, there are a few caveats to
offlining (read that as devices should still be prepared to process
commands).

1. Any special command will still be accepted (that's a command either
via the SCSI_IOCTL_SEND_COMMAND, or an internally generated command).
2. Outstanding already processed commands in the queue (i.e. commands
which have already been through the upper layer drivers but needed
requeuing for some reason like QUEUE_FULL or device busy).

I'm willing to consider changing 2., it just requires more speciallised
logic to distinguish between a command that has been prepared by the
upper level drivers and a command sent via 1.

However, not that LLDs may not assume they will receive no commands just
because scsi_device->online is zero.