Re: [PATCH v7 7/8] scsi: ufs-qcom: Delay specific time before gate ref clk

[Can Guo <cang@codeaurora.org>]

On 2020-02-07 10:10, Bjorn Andersson wrote:
> On Thu 06 Feb 17:09 PST 2020, Can Guo wrote:
> 
>> On 2020-02-07 04:33, Bjorn Andersson wrote:
>> > On Thu 06 Feb 00:33 PST 2020, Can Guo wrote:
>> >
>> > > After enter hibern8, as UFS JEDEC ver 3.0 requires, a specific
>> > > gating wait
>> > > time is required before disable the device reference clock. If it is
>> > > not
>> > > specified, use the old delay.
>> > >
>> > > Signed-off-by: Can Guo <cang@codeaurora.org>
>> > > Reviewed-by: Asutosh Das <asutoshd@codeaurora.org>
>> > > Reviewed-by: Hongwu Su <hongwus@codeaurora.org>
>> > > ---
>> > >  drivers/scsi/ufs/ufs-qcom.c | 22 +++++++++++++++++++---
>> > >  1 file changed, 19 insertions(+), 3 deletions(-)
>> > >
>> > > diff --git a/drivers/scsi/ufs/ufs-qcom.c b/drivers/scsi/ufs/ufs-qcom.c
>> > > index 85d7c17..39eefa4 100644
>> > > --- a/drivers/scsi/ufs/ufs-qcom.c
>> > > +++ b/drivers/scsi/ufs/ufs-qcom.c
>> > > @@ -833,6 +833,8 @@ static int ufs_qcom_bus_register(struct
>> > > ufs_qcom_host *host)
>> > >
>> > >  static void ufs_qcom_dev_ref_clk_ctrl(struct ufs_qcom_host *host,
>> > > bool enable)
>> > >  {
>> > > +	unsigned long gating_wait;
>> > > +
>> > >  	if (host->dev_ref_clk_ctrl_mmio &&
>> > >  	    (enable ^ host->is_dev_ref_clk_enabled)) {
>> > >  		u32 temp = readl_relaxed(host->dev_ref_clk_ctrl_mmio);
>> > > @@ -845,11 +847,25 @@ static void ufs_qcom_dev_ref_clk_ctrl(struct
>> > > ufs_qcom_host *host, bool enable)
>> > >  		/*
>> > >  		 * If we are here to disable this clock it might be immediately
>> > >  		 * after entering into hibern8 in which case we need to make
>> > > -		 * sure that device ref_clk is active at least 1us after the
>> > > +		 * sure that device ref_clk is active for specific time after
>> > >  		 * hibern8 enter.
>> > >  		 */
>> > > -		if (!enable)
>> > > -			udelay(1);
>> > > +		if (!enable) {
>> > > +			gating_wait = host->hba->dev_info.clk_gating_wait_us;
>> > > +			if (!gating_wait) {
>> >
>> > Afaict this can't happen, because in patch 6 you check for gating_wait
>> > being 0 and if so set it to 0xff.
>> >
>> 
>> Sorry, I was intended to give clk_gating_wait_us values only if it is
>> a UFS3.0 device. I will revise patch 6/8.
>> 
> 
> Okay, sounds good.
> 
>> > > +				udelay(1);
>> > > +			} else {
>> > > +				/*
>> > > +				 * bRefClkGatingWaitTime defines the minimum
>> > > +				 * time for which the reference clock is
>> > > +				 * required by device during transition from
>> > > +				 * HS-MODE to LS-MODE or HIBERN8 state. Give it
>> > > +				 * more time to be on the safe side.
>> > > +				 */
>> > > +				gating_wait += 10;
>> > > +				usleep_range(gating_wait, gating_wait + 10);
>> >
>> > I presume there's no strong requirement on the max, so how about using a
>> > substantially larger max - say 1k, or 10k - to allow the usleep_range()
>> > to do it's job?
>> >
>> >
>> > PS. Please include linux-arm-msm@ on all the patches in the series, not
>> > just two of them.
>> >
>> > Regards,
>> > Bjorn
>> >
>> 
>> bRefClkGatingWaitTime, as vendor defined in their device attribute is
>> usually
>> around 50~100, 1k or 10k delay makes it too large. usleep_range() 
>> works well
>> so long as the delay is within (10us - 20ms), so I added 10 to make 
>> sure it
>> is
>> above 10us.
>> 
> 
> I meant specifically the second parameter, i.e:
>   usleep_range(bRefClkGatingWaitTime + 10, bRefClkGatingWaitTime + 
> 1000);
> 
> As you're not guaranteed an upper bound of this sleep anyway you might
> as well give usleep_range() a window of a millisecond (or more) to give
> it the flexibility of matching other timer events.
> 
> The only drawback with this is that you might "waste" a millisecond.
> 
> Regards,
> Bjorn
> 

Hi Bjorn,

Device ref clk gate/ungate can happen very frequently if clk gating is
enabled. The "wasted" milliseconds might be accumulated to affect
read/write latency, i.e:

If read/write requests come, clk ungate kicks start, it flushes ongoing
gate work if any, if we "waste" a millisecond here, the read/write
requests will be delayed a millisecond. I am not sure how much it may
impact the overall performance in a long term test.

Thanks,

Can Guo.

>> SLEEPING FOR ~USECS OR SMALL MSECS ( 10us - 20ms):
>> 	* Use usleep_range
>> https://www.kernel.org/doc/Documentation/timers/timers-howto.txt
>> 
>> Thanks,
>> 
>> Can Guo.
>> 
>> > > +			}
>> > > +		}
>> > >
>> > >  		writel_relaxed(temp, host->dev_ref_clk_ctrl_mmio);
>> > >
>> > > --
>> > > The Qualcomm Innovation Center, Inc. is a member of the Code Aurora
>> > > Forum,
>> > > a Linux Foundation Collaborative Project