Hi Dmitry,
On 11/11/20 10:14 AM, Dmitry Osipenko wrote:
> Add devfreq support to the Tegra20 EMC driver. Memory utilization
> statistics will be periodically polled from the memory controller and
> appropriate minimum clock rate will be selected by the devfreq governor.
>
> Signed-off-by: Dmitry Osipenko <digetx@xxxxxxxxx>
> ---
> drivers/memory/tegra/Kconfig | 3 +-
> drivers/memory/tegra/tegra20-emc.c | 90 ++++++++++++++++++++++++++++++
> 2 files changed, 92 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/memory/tegra/Kconfig b/drivers/memory/tegra/Kconfig
> index ac3dfe155505..8cc1ec5be443 100644
> --- a/drivers/memory/tegra/Kconfig
> +++ b/drivers/memory/tegra/Kconfig
> @@ -12,7 +12,8 @@ config TEGRA20_EMC
> tristate "NVIDIA Tegra20 External Memory Controller driver"
> default y
> depends on TEGRA_MC && ARCH_TEGRA_2x_SOC
> - select PM_OPP
> + select DEVFREQ_GOV_SIMPLE_ONDEMAND
> + select PM_DEVFREQ
> help
> This driver is for the External Memory Controller (EMC) found on
> Tegra20 chips. The EMC controls the external DRAM on the board.
> diff --git a/drivers/memory/tegra/tegra20-emc.c b/drivers/memory/tegra/tegra20-emc.c
> index d01b556a6d06..b9cd965980e2 100644
> --- a/drivers/memory/tegra/tegra20-emc.c
> +++ b/drivers/memory/tegra/tegra20-emc.c
> @@ -8,6 +8,7 @@
> #include <linux/clk.h>
> #include <linux/clk/tegra.h>
> #include <linux/debugfs.h>
> +#include <linux/devfreq.h>
> #include <linux/err.h>
> #include <linux/interconnect-provider.h>
> #include <linux/interrupt.h>
> @@ -102,6 +103,10 @@
>
> #define EMC_FBIO_CFG5_DRAM_WIDTH_X16 BIT(4)
>
> +#define EMC_PWR_GATHER_CLEAR (1 << 8)
> +#define EMC_PWR_GATHER_DISABLE (2 << 8)
> +#define EMC_PWR_GATHER_ENABLE (3 << 8)
> +
> static const u16 emc_timing_registers[] = {
> EMC_RC,
> EMC_RFC,
> @@ -157,6 +162,7 @@ struct emc_timing {
> };
>
> enum emc_rate_request_type {
> + EMC_RATE_DEVFREQ,
> EMC_RATE_DEBUG,
> EMC_RATE_ICC,
> EMC_RATE_TYPE_MAX,
> @@ -193,6 +199,8 @@ struct tegra_emc {
>
> /* protect shared rate-change code path */
> struct mutex rate_lock;
> +
> + struct devfreq_simple_ondemand_data ondemand_data;
> };
>
> static irqreturn_t tegra_emc_isr(int irq, void *data)
> @@ -1003,6 +1011,87 @@ static int tegra_emc_init_clk(struct tegra_emc *emc)
> return 0;
> }
>
> +static int tegra_emc_devfreq_target(struct device *dev, unsigned long *freq,
> + u32 flags)
> +{
> + struct tegra_emc *emc = dev_get_drvdata(dev);
> + struct dev_pm_opp *opp;
> + unsigned long rate;
> +
> + opp = devfreq_recommended_opp(dev, freq, flags);
> + if (IS_ERR(opp)) {
> + dev_err(dev, "failed to find opp for %lu Hz\n", *freq);
> + return PTR_ERR(opp);
> + }
> +
> + rate = dev_pm_opp_get_freq(opp);
> + dev_pm_opp_put(opp);
> +
> + return emc_set_min_rate(emc, rate, EMC_RATE_DEVFREQ);
> +}
> +
> +static int tegra_emc_devfreq_get_dev_status(struct device *dev,
> + struct devfreq_dev_status *stat)
> +{
> + struct tegra_emc *emc = dev_get_drvdata(dev);
> +
> + /* freeze counters */
> + writel_relaxed(EMC_PWR_GATHER_DISABLE, emc->regs + EMC_STAT_CONTROL);
> +
> + /*
> + * busy_time: number of clocks EMC request was accepted
> + * total_time: number of clocks PWR_GATHER control was set to ENABLE
> + */
> + stat->busy_time = readl_relaxed(emc->regs + EMC_STAT_PWR_COUNT);
> + stat->total_time = readl_relaxed(emc->regs + EMC_STAT_PWR_CLOCKS);
> + stat->current_frequency = clk_get_rate(emc->clk);
> +
> + /* clear counters and restart */
> + writel_relaxed(EMC_PWR_GATHER_CLEAR, emc->regs + EMC_STAT_CONTROL);
> + writel_relaxed(EMC_PWR_GATHER_ENABLE, emc->regs + EMC_STAT_CONTROL);
> +
> + return 0;
> +}
> +
> +static struct devfreq_dev_profile tegra_emc_devfreq_profile = {
> + .polling_ms = 30,
> + .target = tegra_emc_devfreq_target,
> + .get_dev_status = tegra_emc_devfreq_get_dev_status,
> +};
> +
> +static int tegra_emc_devfreq_init(struct tegra_emc *emc)
> +{
> + struct devfreq *devfreq;
> +
> + /*
> + * PWR_COUNT is 1/2 of PWR_CLOCKS at max, and thus, the up-threshold
> + * should be less than 50. Secondly, multiple active memory clients
> + * may cause over 20% of lost clock cycles due to stalls caused by
> + * competing memory accesses. This means that threshold should be
> + * set to a less than 30 in order to have a properly working governor.
> + */
> + emc->ondemand_data.upthreshold = 20;
> +
> + /*
> + * Reset statistic gathers state, select global bandwidth for the
> + * statistics collection mode and set clocks counter saturation
> + * limit to maximum.
> + */
> + writel_relaxed(0x00000000, emc->regs + EMC_STAT_CONTROL);
> + writel_relaxed(0x00000000, emc->regs + EMC_STAT_LLMC_CONTROL);
> + writel_relaxed(0xffffffff, emc->regs + EMC_STAT_PWR_CLOCK_LIMIT);
> +
> + devfreq = devm_devfreq_add_device(emc->dev, &tegra_emc_devfreq_profile,
> + DEVFREQ_GOV_SIMPLE_ONDEMAND,
> + &emc->ondemand_data);
> + if (IS_ERR(devfreq)) {
> + dev_err(emc->dev, "failed to initialize devfreq: %pe", devfreq);
> + return PTR_ERR(devfreq);
> + }
> +
> + return 0;
> +}
> +
> static int tegra_emc_probe(struct platform_device *pdev)
> {
> struct device_node *np;
> @@ -1058,6 +1147,7 @@ static int tegra_emc_probe(struct platform_device *pdev)
> tegra_emc_rate_requests_init(emc);
> tegra_emc_debugfs_init(emc);
> tegra_emc_interconnect_init(emc);
> + tegra_emc_devfreq_init(emc);
>
> /*
> * Don't allow the kernel module to be unloaded. Unloading adds some
>
Reviewed-by: Chanwoo Choi <cw00.choi@xxxxxxxxxxx>
--
Best Regards,
Chanwoo Choi
Samsung Electronics
_______________________________________________
dri-devel mailing list
dri-devel@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/dri-devel
