Re: [PATCH v2] ARM: Drop fixed 200 Hz timer requirement from Samsung platforms

[Krzysztof Kozlowski]

On Fri, Nov 18, 2016 at 11:43:14AM +0100, Sylwester Nawrocki wrote:
> On 11/18/2016 09:46 AM, Arnd Bergmann wrote:
> > On Friday, November 18, 2016 9:16:58 AM CET Krzysztof Kozlowski wrote:
> >> > All Samsung platforms, including the Exynos, are selecting HZ_FIXED with
> >> > 200 Hz.  Unfortunately in case of multiplatform image this affects also
> >> > other platforms when Exynos is enabled.
> >> > 
> >> > This looks like an very old legacy code, dating back to initial
> >> > upstreaming of S3C24xx.  Probably it was required for s3c24xx timer
> >> > driver, which was removed in commit ad38bdd15d5b ("ARM: SAMSUNG: Remove
> >> > unused plat-samsung/time.c").
> >> > 
> >> > Since then, this fixed 200 Hz spread everywhere, including out-of-tree
> >> > Samsung kernels (SoC vendor's and Tizen's).  I believe this choice
> >> > was rather an effect of coincidence instead of conscious choice.
> >> > 
> >> > Exynos uses its own MCT or arch timer and can work with all HZ values.
> >> > Older platforms use newer Samsung PWM timer driver which should handle
> >> > down to 100 Hz.
> >> > 
> >> > Few perf mem and sched tests on Odroid XU3 board (Exynos5422, 4x Cortex
> >> > A7, 4x Cortex A15) show no regressions when switching from 200 Hz to
> >> > other values.
> >> > 
> >> > Reported-by: Lee Jones <lee.jones@xxxxxxxxxx>
> >> > [Dropping 200_HZ from S3C/S5P suggested by Arnd]
> >> > Reported-by: Arnd Bergmann <arnd@xxxxxxxx>
> >> > Signed-off-by: Krzysztof Kozlowski <krzk@xxxxxxxxxx>
> >> > Cc: Kukjin Kim <kgene@xxxxxxxxxx>
> >> > Tested-by: Javier Martinez Canillas <javier@xxxxxxxxxxxxxxx>
> >> > 
> > Acked-by: Arnd Bergmann <arnd@xxxxxxxx>
> > 
> > Maybe add a paragraph about the specific problem:
> > 
> > "On s3c24xx, the PWM counter is only 16 bit wide, and with the
> > typical 12MHz input clock that overflows every 5.5ms. This works
> > with HZ=200 or higher but not with HZ=100 which needs a 10ms
> > interval between ticks. On Later chips (S3C64xx, S5P and EXYNOS),
> > the counter is 32 bits and does not have this problem.
> > The new samsung_pwm_timer driver solves the problem by scaling
> > the input clock by a factor of 50 on s3c24xx, which makes it
> > less accurate but allows HZ=100 as well as CONFIG_NO_HZ with
> > fewer wakeups".
> 
> I've tested on S3C2440 SoC based board and I didn't notice any
> issues with HZ=100.
> 
> Clock frequencies look a bit different because AFAIU MPLL
> clock is mostly used as a root clock. The 12 MHz oscillator clock
> is used a root clock for the MPLL.
> 
> refclk:    12000 kHz
> mpll:     405000 kHz
> upll:      48000 kHz
> fclk:     405000 kHz
> hclk:     101250 kHz
> pclk:      50625 kHz
> 
> So frequency of the timer block's source clock (PCLK) is 50.625 MHz.
> This is further divided by 50 in the prescaler as you pointed out.
> 
> So the 16-bit is clocked with 1012500 Hz clock. I added some printks
> to verify this.
> 
> Here is boot log for HZ=200: http://pastebin.com/JuWZdYwh
> and HZ=100 http://pastebin.com/HnDnBfhc
> 
> samsung_clocksource_init:351 pclk: 50625000, timer clock_rate: 1012500
> sched_clock: 16 bits at 1012kHz, resolution 987ns, wraps every 32362962ns

Thanks for tests! I really appreciate it.

> I just don't understand why the log says timer overflow is every 32.362 ms
> and not twice this value (65536 * 1/1012500).

The answer could be at kernel/time/clocksource.c:
 * NOTE: This function includes a safety margin of 50%, in other words, we
 * return half the number of nanoseconds the hardware counter can technically
 * cover.

Best regards,
Krzysztof