On 1/18/23 14:46, Bart Van Assche wrote:
On 1/18/23 14:14, Keith Busch wrote:
On Tue, Jan 03, 2023 at 11:53:37AM -0800, Bart Van Assche wrote:
This algorithm hurts performance for UFS devices because UFS devices
have multiple logical units. One of these logical units (WLUN) is used
to submit control commands, e.g. START STOP UNIT. If any request is
submitted to the WLUN, the queue depth is reduced from 31 to 15 or
lower for data LUNs.
Can you give the WLUN it's own tagset instead?
Hi Keith,
Does that mean creating an additional tagset for the WLUN and statically
allocating all tags from the WLUN tagset that are used by other LUNs?
That approach would have the following disadvantages:
- The maximum queue depth for other LUNs would be reduced from 31 to 30.
This would have a small but noticeable negative performance impact.
- The code removed by this patch negatively impacts performance of all
SCSI hosts with two or more data LUNs and also of all NVMe controllers
that export two or more namespaces if there are significant
differences in the number of I/O operations per second for different
LUNs/namespaces. This is why I think that this should be solved in the
block layer instead of modifying each block driver individually.
(replying to my own e-mail)
Hi Keith,
An additional concern about the approach of using multiple tag sets per
SCSI host: this approach is not compatible with any code that iterates
over tag sets. Examples of functions that iterate over tag sets are
blk_mq_tagset_busy_iter(), scsi_host_busy(),
scsi_host_complete_all_commands() and scsi_host_busy_iter().
scsi_host_busy() is used by e.g. scsi_host_queue_ready() and
scsi_eh_wakeup(). So I think this approach would break the SCSI core in
a subtle way.
Bart.
