On 11/19/24 12:01 AM, Damien Le Moal wrote:
On 11/19/24 09:27, Bart Van Assche wrote:
This patch series improves small write IOPS by a factor of four (+300%) for
zoned UFS devices on my test setup with an UFSHCI 3.0 controller. Although
you are probably busy because the merge window is open, please take a look
at this patch series when you have the time. This patch series is organized
as follows:
- Bug fixes for existing code at the start of the series.
- The write pipelining support implementation comes after the bug fixes.
Impressive improvements but the changes are rather invasive. Have you tried
simpler solution like forcing unplugging a zone write plug from the driver once
a command is passed to the driver and the driver did not reject it ? It seems
like this would make everything simpler on the block layer side. But I am not
sure if the performance gains would be the same.
(replying to an email from two months ago)
Hi Damien,
Here is a strong reason why the simpler solution mentioned above is not
sufficient: if anything goes wrong with the communication between UFS
host controller and UFS device (e.g. a command timeout or a power mode
transition fails) then the SCSI error handler is activated. This results
in ufshcd_err_handler() being called. That function resets both the host
controller and the UFS device (ufshcd_reset_and_restore()). At that time
multiple commands may be outstanding.
In other words, submitting UFS commands in order is not sufficient. An
approach is needed that is compatible with the SCSI error handler and
also that ensures that commands are resubmitted in LBA order per zone
after the SCSI error handler has completed.
The statistics I have access to show that the UFS error handler is
activated infrequently or never on any single device but also that it is
activated a nontrivial number of times across the entire device
population.
Thanks,
Bart.
