Re: [PATCH 2/7] watchdog: imx7ulp: Add explict memory barrier for unlock sequence

[Guenter Roeck <linux@roeck-us.net>]

On Tue, Aug 23, 2022 at 11:10:27AM +0200, Marco Felsch wrote:
> On 22-08-23, Alice Guo (OSS) wrote:
> > 
> > 
> > > -----Original Message-----
> > > From: Guenter Roeck <groeck7@gmail.com> On Behalf Of Guenter Roeck
> > > Sent: Monday, August 22, 2022 10:04 PM
> > > To: Marco Felsch <m.felsch@pengutronix.de>
> > > Cc: Alice Guo (OSS) <alice.guo@oss.nxp.com>; wim@linux-watchdog.org;
> > > shawnguo@kernel.org; s.hauer@pengutronix.de; festevam@gmail.com;
> > > linux-arm-kernel@lists.infradead.org; linux-kernel@vger.kernel.org;
> > > dl-linux-imx <linux-imx@nxp.com>; kernel@pengutronix.de;
> > > linux-watchdog@vger.kernel.org
> > > Subject: Re: [PATCH 2/7] watchdog: imx7ulp: Add explict memory barrier for
> > > unlock sequence
> > > 
> > > On Mon, Aug 22, 2022 at 10:00:10AM +0200, Marco Felsch wrote:
> > > > On 22-08-22, Alice Guo (OSS) wrote:
> > > > > > -----Original Message-----
> > > > > > From: Marco Felsch <m.felsch@pengutronix.de>
> > > > > > Sent: Tuesday, August 16, 2022 2:24 PM
> > > > > > To: Alice Guo (OSS) <alice.guo@oss.nxp.com>
> > > > > > Cc: wim@linux-watchdog.org; linux@roeck-us.net;
> > > > > > shawnguo@kernel.org; s.hauer@pengutronix.de; festevam@gmail.com;
> > > > > > linux-arm-kernel@lists.infradead.org;
> > > > > > linux-kernel@vger.kernel.org; dl-linux-imx <linux-imx@nxp.com>;
> > > > > > kernel@pengutronix.de; linux-watchdog@vger.kernel.org
> > > > > > Subject: Re: [PATCH 2/7] watchdog: imx7ulp: Add explict memory
> > > > > > barrier for unlock sequence
> > > > > >
> > > > > > On 22-08-16, Alice Guo (OSS) wrote:
> > > > > > > From: Jacky Bai <ping.bai@nxp.com>
> > > > > > >
> > > > > > > Add explict memory barrier for the wdog unlock sequence.
> > > > > >
> > > > > > Did you inspected any failures? It's not enough to say what you
> > > > > > did, you need to specify the why as well.
> > > > > >
> > > > > > Regards,
> > > > > >   Marco
> > > > >
> > > > > Hi,
> > > > >
> > > > > Two 16-bit writes of unlocking the Watchdog should be completed within a
> > > certain time. The first mb() is used to ensure that previous instructions are
> > > completed.
> > > > > The second mb() is used to ensure that the unlock sequence cannot be
> > > affected by subsequent instructions. The reason will be added in the commit
> > > log of v2.
> > > >
> > > > Hi,
> > > >
> > > > I know what memory barriers are. My question was, did you see any
> > > > issues? Since the driver is used mainline and no one reported issues.
> > > >
> > > > Also just don't use the *_relaxed() versions is more common, than
> > > > adding
> > > > mb() calls around *_relaxed() versions.
> > > >
> > > 
> > > Agreed with both. The series is a bit short in explaining _why_ the changes are
> > > made.
> > > 
> > > Guenter
> > > 
> > > > Regards,
> > > >   Marco
> > > >
> > > > >
> > 
> > Hi Guenter and Marco,
> > 
> > 1. did you see any issues?
> > This WDOG Timer first appeared in i.MX7ULP, no one report issues
> > probably because few people use i.MX7ULP. This issue was found when we
> > did a stress test on it. When we reconfigure the WDOG Timer, there is
> > a certain probability that it reset. The reason for the error is that
> > when WDOG_CS[CMD32EN] is 0, the unlock sequence is two 16-bit writes
> > (0xC520, 0xD928) to the CNT register within 16 bus clocks, and
> > improper unlock sequence causes the WDOG to reset. Adding mb() is to
> > guarantee that two 16-bit writes are finished within 16 bus clocks.
> 
> After this explanation the whole imx7ulp_wdt_init() seems a bit buggy
> because writel_relaxed() as well as writel() are 32bit access functions.
> So the very first thing to do is to enable the 32-bit mode.
> 
Agreed. This is much better than having extra code to deal with
both 16-bit and 32-bit access.

> Also this is a explanation worth to be added to the commit message ;)
> 

Definitely. Also, the use of mb(), if it should indeed be needed,
would have to be explained in a code comment.

Thanks,
Guenter

> > 2. Also just don't use the *_relaxed() versions is more common, than
> > adding mb() calls around *_relaxed() versions.  Memory barriers cannot
> > be added between two 16-bit writes. I do not know the reason.
> 
> As written above, writel() as well as writel_relaxed() are not 16-bit
> access functions.
> 
> So to me it looks as you need first to ensure that 32-bit access mode is
> enabled. After that you can write drop the to writel_relaxed() functions
> and instead just write:
> 
>    writel(UNLOCK, wdt->base + WDOG_CNT);
> 
> Also why do we need to unlock the watchdog during imx7ulp_wdt_init()?
> This is handled just fine during imx7ulp_wdt_enable() and during
> imx7ulp_wdt_set_timeout(). So just drop those relaxed writes and
> everything should be fine.
> 
> Regards,
>   Marco