On 11/07/2017 16:46, Bart Van Assche wrote:
On Tue, 2017-07-11 at 15:14 +0100, John Garry wrote:
On 11/07/2017 14:32, Bart Van Assche wrote:
On Tue, 2017-07-11 at 11:22 +0100, John Garry wrote:
On 10/07/2017 16:50, Bart Van Assche wrote:
Since a fix for the performance regression triggered by this patch will be upstream
soon (see also https://git.kernel.org/pub/scm/linux/kernel/git/axboe/linux-block.git/commit/?h=for-linus&id=32825c45ff8f4cce937ab85b030dc693ceb1aa0a):
FYI, on linux-next 20170711 (which now includes the above patch Bart
mentioned) we see a large performance regression on hisi_sas (LLDD does
not config shost for mq).
fio read mode iops goes from ~833K (scsi_mod.use_blk_mq=n on cmdline) to
~320K
Hello John,
Thanks for the feedback. Is the kernel config with which these measurements
were performed available somewhere?
It is the default arm64 defconfig with the following changes:
CONFIG_ARM_SMMU_V3=n
CONFIG_9P_FS=n
We were getting a compile error in the 9p fs code, so disabled it.
Turning on the IOMMU drops performance across the board for our
platform, so just disabling it for the test.
Hello John,
What block driver controls the block device for which the performance regression
has been observed? How many hardware queues were created by that block driver
(see also /sys/block/*/mq/...)?
Hi Bart,
Here's the shost init for our SCSI LLDD:
http://elixir.free-electrons.com/linux/latest/source/drivers/scsi/hisi_sas/hisi_sas_main.c#L1736
So we don't set hr_hw_queues (which would mean = 0), so this should set
shost->tag_set.nr_hw_queues to 1 in scsi_mq_setup_tags().
FWIW, I can confirm sysfs entry when I get hw access tomorrow.
John
I'm asking this because the number of hardware
queues controls which I/O scheduler is selected as default. From block/elevator.c:
if (q->mq_ops) {
if (q->nr_hw_queues == 1)
e = elevator_get("mq-deadline", false);
if (!e)
return 0;
} else
e = elevator_get(CONFIG_DEFAULT_IOSCHED, false);
Bart.
