[ofa-general] Re: [PATCH 2/3] IB/srp: allow user to control host queue length

[David Dillow]

On Wed, 2008-01-23 at 19:22 -0800, Roland Dreier wrote:
> So thinking about this some more... does it make more sense to leave
> our can_queue alone, use scsi_adjust_queue_depth() to limit the queue
> to the initial request limit,

I don't think we can use scsi_adjust_queue_depth() -- that's used per
scsi_device (ie, LUN), and is not the host's overall queue depth. When
not using tagged queuing, it looks to be equivalent to cmds_per_lun.

>  and then add a .change_queue_depth
> method so that the user can adjust the queue depth via the standard
> scsi sysfs entry?

I like the idea of being able to tune this on the fly via sysfs, but I'm
not sure what to call it, since the mid-layer already registers a
"can_queue" file which is read-only. Other than that, it looks possible
to just update can_queue at will, as long as we make sure it is >= 1 and
we can accept a few extra commands in flight until the queue drains
below the new setting. Ideas?

There's another reason to have some ability to change can_queue, either
dynamically or at add-target time -- there are some devices out there
that report a slightly inflated initial credit limit, which lets us
drive into unspecified territory. I expect there to be a fix released
soon, but not everyone will be able to upgrade and will need to limited
the queue length manually to stay out of trouble.
-- 
Dave Dillow
National Center for Computational Science
Oak Ridge National Laboratory
(865) 241-6602 office