On Thu, Nov 13, 2014 at 12:33:41AM +0000, Elliott, Robert (Server Storage)
wrote:
> A queue_type of "none" disables the SCSI midlayer automatic queue depth
> adjustments based on TASK SET FULL vs. other status values, while
> "simple" enables them. That's the feature where TASK SET FULL reception
> causes the queue depth to be lowered, while several GOOD status values
> in a row causes it to be increased.
That's not what it's supposed to do, and I faily to see how it would do
that. Pick a random change_queue_type implementation before my patch:
static int
lpfc_change_queue_type(struct scsi_device *sdev, int tag_type)
{
if (sdev->tagged_supported) {
scsi_set_tag_type(sdev, tag_type);
if (tag_type)
scsi_activate_tcq(sdev, sdev->queue_depth);
else
scsi_deactivate_tcq(sdev, sdev->queue_depth);
} else
tag_type = 0;
return tag_type;
}
So if sdev->tagged_supported is not set it's literally a no-op.
Else if we want to move to none we first do the scsi_set_tag_type call
which ends up setting sdev->ordered_tags and sdev->simple_tags to 0.
Then we call scsi_deactivate_tcq which for a non-mq devices calls
calls blk_queue_free_tags which just clears the QUEUE_FLAG_QUEUED
flag, which causes the block layer to not generated tags anymore, and
causes blk_rq_tagged to be false for every request submitted after
this point. I then also calls scsi_adjust_queue_depth which clears
sdev->ordered_tags and sdev->simple_tags a second time, but otherwise is
a no-op as we pass in the current depth.
Can you explain in what exact scenario you used the attribute to disable
the queue full tracking?
> That logic doesn't work very well right now, so that sysfs file
> might be the best way to keep controlling the automatic adjustment
> feature even if it no longer controls whether tags exist.
I'd rather introduce a sysfs attribute to control that specific behavior
instead of reling on weird accidental side effects.
> One problem is it adjusts the queue depths for all devices sharing
> the same target number; we talked about letting the host choose
> the policy (http://marc.info/?l=linux-scsi&m=140440001619574&w=2).
I've got some patches for queue full tracking pending, but tackling the
scope qualifier is still on the todo list. I really need someone to
sign up helping to test it, writing the code is fairly easy
.
> Even if that is corrected, the ramp-up/ramp-down algorithms
> and controls are not quite right for the full range of device
> types (SSDs to HDDs). With SSDs, I saw the queue_depth:
> * ramp down by about 10 every 2 seconds
> * with the default queue_ramp_up_period of 120000: never ramp up
> * with a queue_ramp_up_period of 120: ramp up by 16 ever 2 seconds
>
> At 1M IOPS, 2 seconds is far too long to decide anything; some
> algorithm changes are needed.
Agreed, the static timing seems wrong. A count of successfull
completions might be a better idea.
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
