Re: [PATCH] [v2] media: i2c: mt9m114: use fsleep() in place of udelay()

[Laurent Pinchart]

On Wed, Dec 13, 2023 at 01:56:48PM +0200, Tomi Valkeinen wrote:
> On 13/12/2023 13:48, Laurent Pinchart wrote:
> > On Wed, Dec 13, 2023 at 01:40:54PM +0200, Tomi Valkeinen wrote:
> >> On 13/12/2023 13:23, Arnd Bergmann wrote:
> >>> From: Arnd Bergmann <arnd@xxxxxxxx>
> >>>
> >>> With clang-16, building without COMMON_CLK triggers a range check on
> >>> udelay() because of a constant division-by-zero calculation:
> >>>
> >>> ld.lld: error: undefined symbol: __bad_udelay
> >>>>>> referenced by mt9m114.c
> >>>>>>                 drivers/media/i2c/mt9m114.o:(mt9m114_power_on) in archive vmlinux.a
> >>>
> >>> In this configuration, the driver already fails to probe, before
> >>> this function gets called, so it's enough to suppress the assertion.
> >>>
> >>> Do this by using fsleep(), which turns long delays into sleep() calls
> >>> in place of the link failure.
> >>>
> >>> This is probably a good idea regardless to avoid overly long dynamic
> >>> udelay() calls on a slow clock.
> >>>
> >>> Cc: Sakari Ailus <sakari.ailus@xxxxxxxxxxxxxxx>
> >>> Fixes: 24d756e914fc ("media: i2c: Add driver for onsemi MT9M114 camera sensor")
> >>> Signed-off-by: Arnd Bergmann <arnd@xxxxxxxx>
> >>> ---
> >>>    drivers/media/i2c/mt9m114.c | 2 +-
> >>>    1 file changed, 1 insertion(+), 1 deletion(-)
> >>>
> >>> diff --git a/drivers/media/i2c/mt9m114.c b/drivers/media/i2c/mt9m114.c
> >>> index 0a22f328981d..68adaecaf481 100644
> >>> --- a/drivers/media/i2c/mt9m114.c
> >>> +++ b/drivers/media/i2c/mt9m114.c
> >>> @@ -2116,7 +2116,7 @@ static int mt9m114_power_on(struct mt9m114 *sensor)
> >>>    		duration = DIV_ROUND_UP(2 * 50 * 1000000, freq);
> >>>    
> >>>    		gpiod_set_value(sensor->reset, 1);
> >>> -		udelay(duration);
> >>> +		fsleep(duration);
> >>>    		gpiod_set_value(sensor->reset, 0);
> >>>    	} else {
> >>>    		/*
> >>
> >> I think this is fine, so:
> >>
> >> Reviewed-by: Tomi Valkeinen <tomi.valkeinen@xxxxxxxxxxxxxxxx>
> >>
> >> But: If we don't have COMMON_CLK (or rather, I think, HAVE_CLK), the
> >> freq will be zero at compile time. So won't the compiler give a warning
> >> for the DIV_ROUND_UP() call?
> >>
> >> Interestingly, for me, this doesn't give a div-by-zero warning:
> >>
> >> 	int x;
> >> 	int y = 0;
> >> 	x = DIV_ROUND_UP(10, y);
> >>
> >> but this does:
> >>
> >> 	int x;
> >> 	const int y = 0;
> >> 	x = DIV_ROUND_UP(10, y);
> >>
> >> And looks like this gives the warning too:
> >>
> >> 	int x;
> >> 	const int y = 0;
> >> 	if (y)
> >> 		x = DIV_ROUND_UP(10, y);
> >>
> >> So, I think, the code in the driver could fail to compile at some later
> >> point, if the compiler warnings are improved (?), or if someone adds a
> >> 'const' in front of 'long freq = clk_get_rate(sensor->clk);' line.
> >>
> >> Maybe worry about that if it actually happens =).
> > 
> > Maybe :-) I would be tempted to make VIDEO_CAMERA_SENSOR depend on
> > COMMON_CLK.
> 
> I think HAVE_CLK would be more correct.

Indeed. Only arch/m68k (for Coldfire), arch/mips (for BCM63xx and
Lantiq) and arch/sh (for pre-DT platforms) still make use of
HAVE_LEGACY_CLK. It takes a long time to address the last users of a
legacy API, but it seems we could get there one day.

-- 
Regards,

Laurent Pinchart