On Wed, 2023-05-31 at 05:48 -0700, Bart Van Assche wrote:
>
> External email : Please do not click links or open attachments until
> you have verified the sender or the content.
> On 5/30/23 18:54, Stanley Chu wrote:
> > In addition, some benchmark data shows that the performance can be
> > improved by using fewer CQs to aggregate the interrupt handling of
> > completion requests.
>
> What has been measured? IOPS only or both IOPS and latency?
>
> How big is the difference? A few percent or more?
> For which number of SQs and which number of CQs has performance data
> been measured?
>
Comparing 8-8 to 8-1 mapping, it shows few percents improvement.
>
> Would the following work instead of introducing a new vop?
> - Introduce a new capability flag, e.g. UFSHCD_CAP_SINGLE_CQ.
> - Set that flag from inside ufs_mtk_init().
> - Modify the UFS core driver such that the number of completion
> queues
> depends on the UFSHCD_CAP_SINGLE_CQ flag.
>
According to spec, driver is free to assign any SQ to CQ mapping. I am
not sure if it's ideal to constrain mapping to specific kind.
> > Therefore, we would like to introduce a vop to allow the host to
> > configure it accordingly.
>
> We do not accept new vops upstream without a user. Where is the
> implementation of the new .config_cqid() callback?
>
Yes, please refer to
"[PATCH v2 3/3] scsi: ufs: ufs-mediatek: Add MCQ support for MTK
platform"
+static int ufs_mtk_config_cqid(struct ufs_hba *hba)
+{
+ struct ufs_hw_queue *hwq;
+ int i;
+
+ for (i = 0; i < hba->nr_hw_queues; i++) {
+ hwq = &hba->uhq[i];
+ hwq->cqid = 3;
+ }
+
+ return 0;
+}
Po-Wen
