Re: cciss updates for 2.6.6xxx [1/2]

[Jeff Garzik <jgarzik@pobox.com>]

mikem@beardog.cca.cpqcorp.net wrote:
> This patch adds per logical device queues to the HP cciss driver. It currently only implements a single lock but when time permits I will provide that funtionality. Thanks to Jeff Garzik for providing some sample code.
> This patch built against 2.6.5. Please consider this for inclusion.


I appreciate the credit but I don't see that it addressed my original 
objection -- the starvation issue.

Do you cap the number of per-array requests a "1024 / n_arrays", or 
something like that?  You mentioned that the hardware has a maximum of 
1024 outstanding commands, for all devices.  The two typical solutions 
are a round-robin queue (see carmel.c) or limiting each array such that 
if all arrays are full of commands, the total outstanding never exceeds 
1024.

This patch may be OK for -mm, I would rather not see it go upstream -- 
it seems to me you are choosing to decrease stability to obtain a 
performance increase.  I think you can increase performance without 
decreasing stability.

	Jeff