Message-ID: <20170317194235.5a883df8.drivshin@awxrd.com>
Date:   Fri, 17 Mar 2017 19:42:35 -0400
From:   David Rivshin <drivshin@...rd.com>
To:     Grygorii Strashko <grygorii.strashko@...com>
Cc:     <linux-gpio@...r.kernel.org>, <linux-omap@...r.kernel.org>,
        Santosh Shilimkar <ssantosh@...nel.org>,
        Kevin Hilman <khilman@...nel.org>,
        Linus Walleij <linus.walleij@...aro.org>,
        Alexandre Courbot <gnurou@...il.com>,
        <linux-arm-kernel@...ts.infradead.org>,
        <linux-kernel@...r.kernel.org>, <stable@...r.kernel.org>
Subject: Re: [PATCH 1/2] gpio: omap: return error if requested debounce time
 is not possible

On Fri, 17 Mar 2017 16:43:56 -0500
Grygorii Strashko <grygorii.strashko@...com> wrote:

> On 03/17/2017 03:50 PM, David Rivshin wrote:
> > On Fri, 17 Mar 2017 13:54:28 -0500
> > Grygorii Strashko <grygorii.strashko@...com> wrote:
> >  
> >> On 03/17/2017 12:54 PM, David Rivshin wrote:  
> >>> Hi Grygorii,
> >>>
> >>> On Fri, 17 Mar 2017 11:45:56 -0500
> >>> Grygorii Strashko <grygorii.strashko@...com> wrote:
> >>>  
> >>>> On 03/16/2017 07:57 PM, David Rivshin wrote:  
> >>>>> From: David Rivshin <DRivshin@...worx.com>
> >>>>>
> >>>>> omap_gpio_debounce() does not validate that the requested debounce
> >>>>> is within a range it can handle. Instead it lets the register value
> >>>>> wrap silently, and always returns success.
> >>>>>
> >>>>> This can lead to all sorts of unexpected behavior, such as gpio_keys
> >>>>> asking for a too-long debounce, but getting a very short debounce in
> >>>>> practice.
> >>>>>
> >>>>> Fix this by returning -EINVAL if the requested value does not fit into
> >>>>> the register field. If there is no debounce clock available at all,
> >>>>> return -ENOTSUPP.  
> >>>>
> >>>> In general this patch looks good, but there is one thing I'm worry about..
> >>>>  
> >>>>>
> >>>>> Fixes: e85ec6c3047b ("gpio: omap: fix omap2_set_gpio_debounce")
> >>>>> Cc: <stable@...r.kernel.org> # 4.3+
> >>>>> Signed-off-by: David Rivshin <drivshin@...worx.com>
> >>>>> ---
> >>>>>  drivers/gpio/gpio-omap.c | 16 +++++++++++-----
> >>>>>  1 file changed, 11 insertions(+), 5 deletions(-)
> >>>>>
> >>>>> diff --git a/drivers/gpio/gpio-omap.c b/drivers/gpio/gpio-omap.c
> >>>>> index efc85a2..33ec02d 100644
> >>>>> --- a/drivers/gpio/gpio-omap.c
> >>>>> +++ b/drivers/gpio/gpio-omap.c
> >>>>> @@ -208,8 +208,10 @@ static inline void omap_gpio_dbck_disable(struct gpio_bank *bank)
> >>>>>   * OMAP's debounce time is in 31us steps
> >>>>>   *   <debounce time> = (GPIO_DEBOUNCINGTIME[7:0].DEBOUNCETIME + 1) x 31
> >>>>>   * so we need to convert and round up to the closest unit.
> >>>>> + *
> >>>>> + * Return: 0 on success, negative error otherwise.
> >>>>>   */
> >>>>> -static void omap2_set_gpio_debounce(struct gpio_bank *bank, unsigned offset,
> >>>>> +static int omap2_set_gpio_debounce(struct gpio_bank *bank, unsigned offset,
> >>>>>  				    unsigned debounce)
> >>>>>  {
> >>>>>  	void __iomem		*reg;
> >>>>> @@ -218,11 +220,12 @@ static void omap2_set_gpio_debounce(struct gpio_bank *bank, unsigned offset,
> >>>>>  	bool			enable = !!debounce;
> >>>>>
> >>>>>  	if (!bank->dbck_flag)
> >>>>> -		return;
> >>>>> +		return -ENOTSUPP;
> >>>>>
> >>>>>  	if (enable) {
> >>>>>  		debounce = DIV_ROUND_UP(debounce, 31) - 1;
> >>>>> -		debounce &= OMAP4_GPIO_DEBOUNCINGTIME_MASK;
> >>>>> +		if ((debounce & OMAP4_GPIO_DEBOUNCINGTIME_MASK) != debounce)
> >>>>> +			return -EINVAL;  
> >>>>
> >>>> This might cause boot issues as current drivers may expect this op to succeed even if
> >>>> configured value is wrong - just think, may be we can do warn here and use max value as
> >>>> fallback?  
> >>>
> >>> I have not looked through all drivers to be sure, but at least the gpio-keys
> >>> driver requires set_debounce to return an error if it can't satisfy the request.
> >>> In that case gpio-keys will use a software timer instead.
> >>>
> >>>                 if (button->debounce_interval) {
> >>>                         error = gpiod_set_debounce(bdata->gpiod,
> >>>                                         button->debounce_interval * 1000);
> >>>                         /* use timer if gpiolib doesn't provide debounce */
> >>>                         if (error < 0)
> >>>                                 bdata->software_debounce =
> >>>                                                 button->debounce_interval;
> >>>                 }
> >>>
> >>> Also, at least some other GPIO drivers (e.g. gpio-max7760) return -EINVAL in
> >>> such a case. And gpiolib will return -ENOTSUPP if there is no debounce
> >>> callback at all. So I expect all drivers which use gpiod_set_debounce() to
> >>> handle error returns gracefully.
> >>>
> >>> So I certainly understand the concern about backwards compatibility, but I
> >>> think clipping to max is the greater of the evils in this case. Even a
> >>> warning may be too much, because it's not necessarily anything wrong.
> >>> Perhaps an info or debug message would be helpful, though?
> >>>
> >>> If you prefer, I can try to go through all callers of gpiod_set_debounce()
> >>> and see how they'd handle an error return. The handful I've looked through so
> >>> far all behave like gpio-keys. The only ones I'd be particularly concerned
> >>> about are platform-specific drivers which were perhaps never used with other
> >>> gpio drivers. Do you know of that I should pay special attention to?  
> >>
> >> Yeh agree. But the problem here will be not only with drivers itself - it can be wrong data in DT :(
> >> As result, even  gpio-keys driver will just silently switch to software_debounce
> >> without any notification.  
> >
> > I think that switching to software_debounce silently is exactly the
> > intended/desired behavior of gpio-keys (and other drivers). For example,
> > if the DT requests a 20ms debounce on a gpio-key, the existing math
> > resulted in a hardware debounce of just 2ms. With the error return,
> > gpio-keys would silently switch to software_debounce of the requested
> > 20ms (potentially longer if the CPU is busy, but I don't think that's
> > a problem for correctness), exactly what the DT asked for.
> >
> > Of course that would be a change in behavior for any such existing DT,
> > and it's conceivable that the DT for some HW is somehow relying on that
> > previous incorrect behavior. I suspect it's more likely that they are
> > silently broken, and no-one has noticed. A quick search of some in-tree
> > DTs finds most debounce times are 5ms (which has no issue), and then
> > these three examples (all happen to be gpio-keys):
> >   am335x-shc.dts:                 debounce-interval = <1000>;
> >   am335x-shc.dts:                 debounce-interval = <1000>;
> >   omap5-uevm.dts:                 debounce_interval = <50>;
> > The first two currently result in a HW debounce of about 4ms. The
> > third would be 2.5ms, except it's the wrong property name so it
> > does nothing (it gets the default gpio-keys debounce of 5ms).  
> 
> Yep. looks like error in dt. There are mod such DTs actually
> ./arch/arm/boot/dts/atlas7-evb.dts
> ./arch/arm/boot/dts/emev2-kzm9d.dts
> ./arch/arm/boot/dts/kirkwood-pogoplug-series-4.dts
> ./arch/arm/boot/dts/omap5-uevm.dts
> ./arch/arm/boot/dts/ste-snowball.dts

Ah yes, I just grepped for 'debounce' in am* and omap*. I guess
that typo has been copied over from DT to DT. I'm tempted to spin 
a patch correcting the typo, but I have no knowledge of those
boards or HW to test with. Obviously no-one has complained about 
the 5ms vs 50ms debounce so far, so maybe 50ms isn't the correct
number in the first place?

> >
> > Not having seen any of that hardware, I can't say for certain what the
> > true HW requirements are. 1000ms does seem like a long debounce, perhaps
> > the author meant 1ms (1000us) for those buttons? Or perhaps it really
> > needs a 1000ms debounce, and is currently wrong?
> >  
> >>
> >> But agree - max might not be a good choose, so can you add dev_err() below, pls.  
> >
> > Given the above, I personally feel that a dev_err() is undesirable in most
> > cases. If I have a system and matching DT that just happens to need a longer
> > debounce than the GPIO HW is capable of, gpio-keys (etc) does the best it can automatically. I don't consider that there is any error in that case, or
> > anything to be fixed.
> > I can understanding wanting to draw attention to a change in behavior (just
> > in case the DT is incorrect), but I'd personally lean towards dev_info() if
> > anything.
> >
> > That said: if you still prefer dev_err(), I will certainly do so.  
> 
> Fair enough :) thanks.
> 
> Acked-by: Grygorii Strashko <grygorii.strashko@...com>

Just to make sure I don't misunderstand, would you like me to:
A) put in a dev_err() 
B) put in a dev_info() 
C) leave it as-is without any message 
?

I can spin a v2 as early as Monday, depending on the results of discussion 
on the second patch.

> 
> >
> > Tangent:
> > This discussion makes me think that adding a gpiod_get_max_debounce()
> > would allow even better behavior. Then asking for a too-high debounce
> > could be a dev_err() in all gpio drivers, with the expectation that no
> > driver should ask for such. Also, drivers could do something like use
> > max hardware debounce plus a software debounce for the remaining time,
> > in order to avoid CPU overhead on short glitches.
> >

Powered by blists - more mailing lists