Re: [PATCH] scsi: ufs: core: quiesce request queues before check pending cmds

[Ziqi Chen <quic_ziqichen@quicinc.com>]

On 6/26/2024 12:13 AM, Bart Van Assche wrote:
> On 6/24/24 8:38 PM, Peter Wang (王信友) wrote:
>> But ufshcd_scsi_block_requests usage is correct in SDR mode.
> 
> ufshcd_scsi_block_requests() uses scsi_block_requests(). It is almost
> never correct to use scsi_block_requests() in a blk-mq driver because
> scsi_block_requests() does not wait for ongoing request submission
> calls to complete. scsi_block_requests() is a legacy from the time when
> all request dispatching and queueing was protected by the SCSI host
> lock, a change that was made in 2010 or about 14 years ago. See also
> https://lore.kernel.org/linux-scsi/20101105002409.GA21714@havoc.gtf.org/
> 
>> So, I think ufshcd_wait_for_doorbell_clr should be revise.
>> Check tr_doorbell in SDR mode. (before 8d077ede48c1 do)
>> Check each HWQ's are all empty in MCQ mode. (need think how to do)
>> Make sure all requests is complete, and finish this function' job
>> correctly.
>> Or there still have a gap in ufshcd_wait_for_doorbell_clr.
> 
> ufshcd_wait_for_doorbell_clr() should be removed and 
> ufshcd_clock_scaling_prepare() should use blk_mq_freeze_*().
> See also my patch "ufs: Simplify the clock scaling mechanism
> implementation" from 5 years ago 
> (https://lore.kernel.org/linux-scsi/20191112173743.141503-5-bvanassche@acm.org/).
> 
The defect of blk_mq_freeze_*() is that it would bring in significant 
latency and performance regression. I don't think it is what many people 
want to see.

BRs,
Ziqi

> Best regards,
> 
> Bart.