[PATCH 3/3] ARM: EXYNOS5: save CLK_TOP_SRC3 register before powergating

[thomas.abraham@linaro.org (Thomas Abraham)]

On 27 November 2012 17:52, Prasanna Kumar <prasanna.ps@samsung.com> wrote:
> From: Prasanna Kumar <prasanna.ps@samsung.com>
>
> This patch adds a software workaround to the hardware
> problem found in exynos5 while powergating.
>
> It is observed that CLK_TOP_SRC3 register gets modified if
> the G-Scaler/MFC devices are power gated. The clock for G-Scaler gets
> set to XXTI which results in the device running very slow .
> A big drop in performance is noticed whilerunning the video.
> This issue also occurs while powergating MFC.
>
> The value of clock source register is restored once the powergating
> operation is completed.
>
> Signed-off-by: Prasanna Kumar <prasanna.ps@samsung.com>
> ---
>  arch/arm/mach-exynos/pm_domains.c |   27 +++++++++++++++++++++++++++
>  1 files changed, 27 insertions(+), 0 deletions(-)
>
> diff --git a/arch/arm/mach-exynos/pm_domains.c b/arch/arm/mach-exynos/pm_domains.c
> index 9f1351d..955cbe3 100644
> --- a/arch/arm/mach-exynos/pm_domains.c
> +++ b/arch/arm/mach-exynos/pm_domains.c
> @@ -24,6 +24,8 @@
>
>  #include <mach/regs-pmu.h>
>  #include <plat/devs.h>
> +#include <mach/regs-clock.h>
> +#include <plat/cpu.h>
>
>  /*
>   * Exynos specific wrapper around the generic power domain
> @@ -41,6 +43,7 @@ static int exynos_pd_power(struct generic_pm_domain *domain, bool power_on)
>         void __iomem *base;
>         u32 timeout, pwr;
>         char *op;
> +       u32 tmp = 0;
>
>         pd = container_of(domain, struct exynos_pm_domain, pd);
>         base = pd->base;
> @@ -48,6 +51,23 @@ static int exynos_pd_power(struct generic_pm_domain *domain, bool power_on)
>         pwr = power_on ? S5P_INT_LOCAL_PWR_EN : 0;
>         __raw_writel(pwr, base);
>
> +       /*
> +        *It is found that the CLK SRC register in exynos5
> +        *gets modified when power domain of gsc/mfc/isp/disp1
> +        *is powered off.This happens only after the system is
> +        *suspended and resumed and not before that.
> +        *The following fix adresses this hardware issue.
> +        *It saves the value of clock source register and
> +        *resores it later
> +        */
> +
> +       if (soc_is_exynos5250()) {
> +               if (!power_on) {
> +                       /* save clock source register */
> +                       tmp = __raw_readl(EXYNOS5_CLKSRC_TOP3);
> +               }
> +       }

Does the value of EXYNOS5_CLKSRC_TOP3 register change as soon as the
G-Scaler/MFC devices are power gated?

- If yes, the value of CLKSRC register should be saved before the
power domain register is programmed.

- If not, as mentioned in the comment, if this issue occurs during a
suspend-resume cycle, the value of this register can be saved and
restored in the clock driver code itself.

The other thing that needs to be relooked into here is the use of
EXYNOS5_CLKSRC_TOP3 register address. Currently, the clock registers
are statically io-remapped. With the upcoming migration to common
clock for Exynos5, the clock registers will not be statically
io-remapped but instead remapped during clock initialization. So the
use of EXYNOS5_CLKSRC_TOP3 will not hold and so we need to relook into
this.

> +
>         /* Wait max 1ms */
>         timeout = 10;
>
> @@ -61,6 +81,13 @@ static int exynos_pd_power(struct generic_pm_domain *domain, bool power_on)
>                 cpu_relax();
>                 usleep_range(80, 100);
>         }
> +
> +       if (soc_is_exynos5250()) {

We could use the of_machine_is_compatible api here instead of
soc_is_exynos5250() macro.

> +               if (!power_on) {
> +                       /* restore clock source register */
> +                       __raw_writel(tmp, EXYNOS5_CLKSRC_TOP3);
> +               }
> +       }
>         return 0;
>  }


Thanks,
Thomas.