[PATCH 19/22] OMAP: hsmmc: implement clock switcher

[Grazvydas Ignotas]

On Thu, May 5, 2011 at 2:51 PM, Adrian Hunter <adrian.hunter at nokia.com> wrote:
> From: Andy Shevchenko <ext-andriy.shevchenko at nokia.com>
>
> There are 3 new platform data methods which should help us to do a clock
> switching when notification is happened or request is started.
>
> The purpose of the patch is to avoid high frequency of MMC controller on low
> OPPs due to an HW bug in OMAP 3630.
>
> The algorithm:
>  - the PM governor switches the clock of L3 (and therefore L4) bus on demand
>  - the MMC controller clock should follow the change
>
> We have considered two OPPs for L3/L4 bus. Thus, we have corresponding flow:
>  - OPP1 -> OPP2 (in case of abort - OPP1)
>  - OPP2 -> OPP1 (in case of abort - OPP2)
>
> During idle the MMC gates functional clock and we don't care about the
> frequency. Most important to get proper solution when notification comes during
> request. Then:
>  - in case of OPP1 -> OPP2 we update frequency limit and it will be used for
>   new coming requests (it assumes the current frequency of the controller is
>   lower then new one)
>  - otherwise we wait until no device has used higher frequency then upcoming
>   one
>
> Known issues and limitations:
>  - originally a clock notifier was used for the core L4 iclk but for upstream
>   Adrian changed to a cpufreq notifier where we assume OPP1 below 400MHz and
>   OPP2 above 400MHz
>
> Signed-off-by: Andy Shevchenko <ext-andriy.shevchenko at nokia.com>
> Signed-off-by: Adrian Hunter <adrian.hunter at nokia.com>
> ---
>  arch/arm/mach-omap2/hsmmc.c           |  180 ++++++++++++++++++++++++++++++++-
>  arch/arm/plat-omap/include/plat/mmc.h |    8 ++
>  2 files changed, 187 insertions(+), 1 deletions(-)
>
> diff --git a/arch/arm/mach-omap2/hsmmc.c b/arch/arm/mach-omap2/hsmmc.c
> index 6b97fae..34cba89 100644
> --- a/arch/arm/mach-omap2/hsmmc.c
> +++ b/arch/arm/mach-omap2/hsmmc.c

<snip>

> +
> +static int hsmmc_clk_notifier(struct notifier_block *nb, unsigned long event,
> +                             void *data)
> +{
> +       struct cpufreq_freqs *freqs = data;
> +       unsigned int threshold = 400000; /* between opp1 and opp2 */
> +
> +       switch (event) {
> +       case CPUFREQ_PRECHANGE:
> +               if (freqs->new < threshold && freqs->old > threshold) {

What if we go from 400000 to, say, 200000?

> +                       /* opp2 -> opp1 */
> +                       hsmmc_max_freq = HSMMC_MAX_FREQ >> 1;
> +
> +                       /* Timeout is 1 sec */
> +                       if (!wait_event_timeout(hsmmc_max_freq_wq,
> +                                               hsmmc_max_freq_ok(),
> +                                               msecs_to_jiffies(1000)))
> +                               pr_err("MMC violates maximum frequency "
> +                                      "constraint\n");
> +               }
> +               break;
> +       case CPUFREQ_POSTCHANGE:
> +               if (freqs->old < threshold && freqs->new > threshold) {

Same here, you could go from 200000 to 400000 and then 600000 and code
would never notice it.

> +                       /* opp1 -> opp2 */
> +                       hsmmc_max_freq = HSMMC_MAX_FREQ;
> +               }