Mark Lord wrote:

To maximize throughput, some kind of host-queuing would be needed,
or just have the driver sit in a tight loop, starting the next I/O
immediately when the previous one finishes.  Linux isn't that quick (yet).

I was talking on IRC with Tejun just recently.  There are several 
controllers (and/or "situations") like this, where some amount of host 
queueing would permit greater throughput, even when NCQ is not 
supported.  sata_sx4 is the most dramatic example, where host queueing 
could potentially increase speed by a factor of 10 or more, since it is 
penalized by an awful two-irq-per-command (w/ a per-host bottleneck to 
boot) setup.  Silicon Image has a "command buffer".  And overall, I 
designed ->qc_prep() hook separate from ->qc_issue() to enable the 
prepartion of multiple commands such that it only takes a simple "go" 
I/O to start a transaction, immediately after the previous one ends.

	Jeff