I've observed this behaviour as well on the 2GB wide Quantum
Atlas. A test I performed, with tagged queuing enabled and with the maximum
number of tagged SCSI commands set to one, resulted in a QUEUE_FULL status
every 38 I/O's when doing a one block I/O operation. Setting the FUA bit
within the 10-byte SCSI SCB "fixed" the problem but at a substantial cost
to overall I/O throughput. As a result, I think a switch to a Seagate
Barracuda may be in order - hopefully, it does not exhibit the same sort
of behaviour.
QUEUE FULL is a perfectly valid status for a drive to return. If a host
adapter's firmware or the driver's sequencer logic does not handle this
condition, then it ought not to be using tagged queuing at all. Assuming that
any specific number of outstanding commands will always be accepted is broken.
Leonard