RE: [PATCH v2 1/5] scsi: ufs-mediatek: Fix imprecise waiting time for ref-clk control

From: Stanley Chu <stanley.chu@mediatek.com>
To: Avri Altman <Avri.Altman@wdc.com>
Cc: "linux-scsi@vger.kernel.org" <linux-scsi@vger.kernel.org>,
	"martin.petersen@oracle.com" <martin.petersen@oracle.com>,
	"alim.akhtar@samsung.com" <alim.akhtar@samsung.com>,
	"jejb@linux.ibm.com" <jejb@linux.ibm.com>,
	"beanhuo@micron.com" <beanhuo@micron.com>,
	"asutoshd@codeaurora.org" <asutoshd@codeaurora.org>,
	"cang@codeaurora.org" <cang@codeaurora.org>,
	"matthias.bgg@gmail.com" <matthias.bgg@gmail.com>,
	"bvanassche@acm.org" <bvanassche@acm.org>,
	"linux-mediatek@lists.infradead.org" 
	<linux-mediatek@lists.infradead.org>,
	"linux-arm-kernel@lists.infradead.org" 
	<linux-arm-kernel@lists.infradead.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	"kuohong.wang@mediatek.com" <kuohong.wang@mediatek.com>,
	"peter.wang@mediatek.com" <peter.wang@mediatek.com>,
	"chun-hung.wu@mediatek.com" <chun-hung.wu@mediatek.com>,
	"andy.teng@mediatek.com" <andy.teng@mediatek.com>,
	"cc.chou@mediatek.com" <cc.chou@mediatek.com>,
	"chaotian.jing@mediatek.com" <chaotian.jing@mediatek.com>,
	"pengshun.zhao@mediatek.com" <pengshun.zhao@mediatek.com>
Subject: RE: [PATCH v2 1/5] scsi: ufs-mediatek: Fix imprecise waiting time for ref-clk control
Date: Sun, 31 May 2020 21:45:06 +0800	[thread overview]
Message-ID: <1590932706.25636.14.camel@mtkswgap22> (raw)
In-Reply-To: <SN6PR04MB464015BDF84DF7A9779BEB41FC8D0@SN6PR04MB4640.namprd04.prod.outlook.com>

Hi Avri,

On Sun, 2020-05-31 at 07:10 +0000, Avri Altman wrote:
> > 
> > Currently ref-clk control timeout is implemented by Jiffies. However
> > jiffies is not accurate enough thus "false timeout" may happen.
> > 
> > Use more accurate delay mechanism instead, for example, ktime.
> > 
> > Signed-off-by: Stanley Chu <stanley.chu@mediatek.com>
> > Reviewed-by: Andy Teng <andy.teng@mediatek.com>
> Reviewed-by: Avri Altman <avri.altman@wdc.com>
> 

Thanks for your review.

> > 
> >         /* Wait for ack */
> > -       timeout = jiffies + msecs_to_jiffies(REFCLK_REQ_TIMEOUT_MS);
> > +       timeout = ktime_add_us(ktime_get(), REFCLK_REQ_TIMEOUT_US);
> >         do {
> > +               time_checked = ktime_get();
> >                 value = ufshcd_readl(hba, REG_UFS_REFCLK_CTRL);
> > 
> >                 /* Wait until ack bit equals to req bit */
> > @@ -144,7 +145,7 @@ static int ufs_mtk_setup_ref_clk(struct ufs_hba *hba,
> > bool on)
> >                         goto out;
> > 
> >                 usleep_range(100, 200);
> > -       } while (time_before(jiffies, timeout));
> > +       } while (ktime_before(time_checked, timeout));
> Nit: you could get rid of time_checked if you would use ktime_compare(ktime_get(), timeout) > 0
> 
> Thanks,
> Avri

If this context is preempted and scheduled out between ufshcd_readl()
and ktime_compare(ktime_get(), timeout), then the ktime_get() may get a
"timed-out" time even though the last ufshcd_readl() is actually
executed before the "timed-out" time. In this case, false alarm will
show up. Using "time_checked" here could solve above issue.

Thanks,
Stanley Chu