| Message ID | 20201025221735.3062-50-digetx@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Introduce memory interconnect for NVIDIA Tegra SoCs | expand |
Hi Dmitry, This patch contains the three features as following: 1. Use interconnect interface for controlling the clock instead of controlling it direclty 2. Use EMC_STAT instead of IMC_STAT 3. Change polling_interval and upthreshold for more fast responsiveness I think you need to make the separate patches for each role. But, if it is difficult or not proper to split out 1,2 roles, you can make two patches for 1,2 and 3 roles. Also, if you want to get more responsiveness, you could use delayed timer instead of deferrable timer by editing the devfreq_dev_profile structure. Regards, Chanwoo Choi On Mon, Oct 26, 2020 at 7:21 AM Dmitry Osipenko <digetx@gmail.com> wrote: > > External (EMC) and Internal Memory Controllers (IMC) have a nearly > identical statistics gathering module. This patch switches driver to use > EMC_STAT instead of IMC_STAT and adds device-tree support which brings ICC > support and makes driver to use bandwidth OPPs defined in device-tree. > > The previous tegra20-devfreq variant was depending on presence of both > EMC and IMC drivers simultaneously because it wasn't apparent how to use > EMC_STAT properly back in the day. Dependency on the IMC driver is gone > after this patch. > > The older variant of the devfreq driver also isn't suitable anymore > because EMC got support for interconnect framework and DVFS, hence > tegra20-devfreq shouldn't drive the EMC clock directly, but use OPP > API for issuing memory bandwidth requests. > > The polling interval is changed from 500ms to 30ms in order to improve > responsiveness of the system in general and because EMC clock is now > allowed to go lower than before since display driver supports ICC now > as well. > > The parent EMC device is an MFD device now and tegra20-devfreq its > sub-device. Devfreq driver uses SYSCON API for retrieving regmap of the > EMC registers from the parent device. > > Signed-off-by: Dmitry Osipenko <digetx@gmail.com> > --- > drivers/devfreq/Kconfig | 1 + > drivers/devfreq/tegra20-devfreq.c | 174 +++++++++++++----------------- > 2 files changed, 75 insertions(+), 100 deletions(-) > > diff --git a/drivers/devfreq/Kconfig b/drivers/devfreq/Kconfig > index 0ee36ae2fa79..1bd225e571df 100644 > --- a/drivers/devfreq/Kconfig > +++ b/drivers/devfreq/Kconfig > @@ -126,6 +126,7 @@ config ARM_TEGRA20_DEVFREQ > depends on ARCH_TEGRA_2x_SOC || COMPILE_TEST > depends on COMMON_CLK > select DEVFREQ_GOV_SIMPLE_ONDEMAND > + select MFD_SYSCON > help > This adds the DEVFREQ driver for the Tegra20 family of SoCs. > It reads Memory Controller counters and adjusts the operating > diff --git a/drivers/devfreq/tegra20-devfreq.c b/drivers/devfreq/tegra20-devfreq.c > index fd801534771d..0a36b085d32a 100644 > --- a/drivers/devfreq/tegra20-devfreq.c > +++ b/drivers/devfreq/tegra20-devfreq.c > @@ -7,180 +7,148 @@ > > #include <linux/clk.h> > #include <linux/devfreq.h> > -#include <linux/io.h> > #include <linux/kernel.h> > #include <linux/module.h> > +#include <linux/mfd/syscon.h> > #include <linux/of_device.h> > #include <linux/platform_device.h> > #include <linux/pm_opp.h> > +#include <linux/regmap.h> > #include <linux/slab.h> > > -#include <soc/tegra/mc.h> > - > #include "governor.h" > > -#define MC_STAT_CONTROL 0x90 > -#define MC_STAT_EMC_CLOCK_LIMIT 0xa0 > -#define MC_STAT_EMC_CLOCKS 0xa4 > -#define MC_STAT_EMC_CONTROL 0xa8 > -#define MC_STAT_EMC_COUNT 0xb8 > +#define EMC_STAT_CONTROL 0x160 > +#define EMC_STAT_LLMC_CONTROL 0x178 > +#define EMC_STAT_PWR_CLOCK_LIMIT 0x198 > +#define EMC_STAT_PWR_CLOCKS 0x19c > +#define EMC_STAT_PWR_COUNT 0x1a0 > > -#define EMC_GATHER_CLEAR (1 << 8) > -#define EMC_GATHER_ENABLE (3 << 8) > +#define EMC_PWR_GATHER_CLEAR (1 << 8) > +#define EMC_PWR_GATHER_DISABLE (2 << 8) > +#define EMC_PWR_GATHER_ENABLE (3 << 8) > > struct tegra_devfreq { > + struct devfreq_simple_ondemand_data ondemand_data; > + struct opp_table *opp_table; > struct devfreq *devfreq; > struct clk *emc_clock; > - void __iomem *regs; > + struct regmap *rmap; > }; > > static int tegra_devfreq_target(struct device *dev, unsigned long *freq, > u32 flags) > { > - struct tegra_devfreq *tegra = dev_get_drvdata(dev); > - struct devfreq *devfreq = tegra->devfreq; > struct dev_pm_opp *opp; > - unsigned long rate; > - int err; > + int ret; > > opp = devfreq_recommended_opp(dev, freq, flags); > - if (IS_ERR(opp)) > + 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); > + ret = dev_pm_opp_set_bw(dev, opp); > dev_pm_opp_put(opp); > > - err = clk_set_min_rate(tegra->emc_clock, rate); > - if (err) > - return err; > - > - err = clk_set_rate(tegra->emc_clock, 0); > - if (err) > - goto restore_min_rate; > - > - return 0; > - > -restore_min_rate: > - clk_set_min_rate(tegra->emc_clock, devfreq->previous_freq); > - > - return err; > + return ret; > } > > static int tegra_devfreq_get_dev_status(struct device *dev, > struct devfreq_dev_status *stat) > { > struct tegra_devfreq *tegra = dev_get_drvdata(dev); > + u32 count, clocks; > > - /* > - * EMC_COUNT returns number of memory events, that number is lower > - * than the number of clocks. Conversion ratio of 1/8 results in a > - * bit higher bandwidth than actually needed, it is good enough for > - * the time being because drivers don't support requesting minimum > - * needed memory bandwidth yet. > - * > - * TODO: adjust the ratio value once relevant drivers will support > - * memory bandwidth management. > - */ > - stat->busy_time = readl_relaxed(tegra->regs + MC_STAT_EMC_COUNT); > - stat->total_time = readl_relaxed(tegra->regs + MC_STAT_EMC_CLOCKS) / 8; > - stat->current_frequency = clk_get_rate(tegra->emc_clock); > + /* freeze counters */ > + regmap_write(tegra->rmap, EMC_STAT_CONTROL, EMC_PWR_GATHER_DISABLE); > + > + /* number of clocks when EMC request was accepted */ > + regmap_read(tegra->rmap, EMC_STAT_PWR_COUNT, &count); > + /* total number of clocks while PWR_GATHER control was set to ENABLE */ > + regmap_read(tegra->rmap, EMC_STAT_PWR_CLOCKS, &clocks); > > - writel_relaxed(EMC_GATHER_CLEAR, tegra->regs + MC_STAT_CONTROL); > - writel_relaxed(EMC_GATHER_ENABLE, tegra->regs + MC_STAT_CONTROL); > + /* clear counters and restart */ > + regmap_write(tegra->rmap, EMC_STAT_CONTROL, EMC_PWR_GATHER_CLEAR); > + regmap_write(tegra->rmap, EMC_STAT_CONTROL, EMC_PWR_GATHER_ENABLE); > + > + stat->busy_time = count; > + stat->total_time = clocks; > + stat->current_frequency = clk_get_rate(tegra->emc_clock); > > return 0; > } > > static struct devfreq_dev_profile tegra_devfreq_profile = { > - .polling_ms = 500, > + .polling_ms = 30, > .target = tegra_devfreq_target, > .get_dev_status = tegra_devfreq_get_dev_status, > }; > > -static struct tegra_mc *tegra_get_memory_controller(void) > -{ > - struct platform_device *pdev; > - struct device_node *np; > - struct tegra_mc *mc; > - > - np = of_find_compatible_node(NULL, NULL, "nvidia,tegra20-mc-gart"); > - if (!np) > - return ERR_PTR(-ENOENT); > - > - pdev = of_find_device_by_node(np); > - of_node_put(np); > - if (!pdev) > - return ERR_PTR(-ENODEV); > - > - mc = platform_get_drvdata(pdev); > - if (!mc) > - return ERR_PTR(-EPROBE_DEFER); > - > - return mc; > -} > - > static int tegra_devfreq_probe(struct platform_device *pdev) > { > + struct device_node *emc_np = pdev->dev.parent->of_node; > struct tegra_devfreq *tegra; > - struct tegra_mc *mc; > - unsigned long max_rate; > - unsigned long rate; > int err; > > - mc = tegra_get_memory_controller(); > - if (IS_ERR(mc)) { > - err = PTR_ERR(mc); > - dev_err(&pdev->dev, "failed to get memory controller: %d\n", > - err); > - return err; > - } > - > tegra = devm_kzalloc(&pdev->dev, sizeof(*tegra), GFP_KERNEL); > if (!tegra) > return -ENOMEM; > > /* EMC is a system-critical clock that is always enabled */ > - tegra->emc_clock = devm_clk_get(&pdev->dev, "emc"); > + tegra->emc_clock = devm_get_clk_from_child(&pdev->dev, emc_np, NULL); > if (IS_ERR(tegra->emc_clock)) > return dev_err_probe(&pdev->dev, PTR_ERR(tegra->emc_clock), > "failed to get emc clock\n"); > > - tegra->regs = mc->regs; > - > - max_rate = clk_round_rate(tegra->emc_clock, ULONG_MAX); > + tegra->rmap = device_node_to_regmap(emc_np); > + if (IS_ERR(tegra->rmap)) > + return dev_err_probe(&pdev->dev, PTR_ERR(tegra->rmap), > + "failed to get emc regmap\n"); > > - for (rate = 0; rate <= max_rate; rate++) { > - rate = clk_round_rate(tegra->emc_clock, rate); > + tegra->opp_table = dev_pm_opp_get_opp_table(&pdev->dev); > + if (IS_ERR(tegra->opp_table)) > + return dev_err_probe(&pdev->dev, PTR_ERR(tegra->opp_table), > + "failed to prepare opp table\n"); > > - err = dev_pm_opp_add(&pdev->dev, rate, 0); > - if (err) { > - dev_err(&pdev->dev, "failed to add opp: %d\n", err); > - goto remove_opps; > - } > + err = dev_pm_opp_of_add_table(&pdev->dev); > + if (err) { > + dev_err(&pdev->dev, "failed to add opp table: %d\n", err); > + goto put_table; > } > > + /* > + * 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. > + */ > + tegra->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, tegra->regs + MC_STAT_CONTROL); > - writel_relaxed(0x00000000, tegra->regs + MC_STAT_EMC_CONTROL); > - writel_relaxed(0xffffffff, tegra->regs + MC_STAT_EMC_CLOCK_LIMIT); > + regmap_write(tegra->rmap, EMC_STAT_CONTROL, 0x00000000); > + regmap_write(tegra->rmap, EMC_STAT_LLMC_CONTROL, 0x00000000); > + regmap_write(tegra->rmap, EMC_STAT_PWR_CLOCK_LIMIT, 0xffffffff); > > platform_set_drvdata(pdev, tegra); > > tegra->devfreq = devfreq_add_device(&pdev->dev, &tegra_devfreq_profile, > - DEVFREQ_GOV_SIMPLE_ONDEMAND, NULL); > + DEVFREQ_GOV_SIMPLE_ONDEMAND, > + &tegra->ondemand_data); > if (IS_ERR(tegra->devfreq)) { > err = PTR_ERR(tegra->devfreq); > - goto remove_opps; > + goto put_table; > } > > return 0; > > -remove_opps: > - dev_pm_opp_remove_all_dynamic(&pdev->dev); > +put_table: > + dev_pm_opp_put_opp_table(tegra->opp_table); > > return err; > } > @@ -190,21 +158,27 @@ static int tegra_devfreq_remove(struct platform_device *pdev) > struct tegra_devfreq *tegra = platform_get_drvdata(pdev); > > devfreq_remove_device(tegra->devfreq); > - dev_pm_opp_remove_all_dynamic(&pdev->dev); > + dev_pm_opp_of_remove_table(&pdev->dev); > + dev_pm_opp_put_opp_table(tegra->opp_table); > > return 0; > } > > +static const struct of_device_id tegra_devfreq_of_match[] = { > + { .compatible = "nvidia,tegra20-emc-statistics" }, > + { }, > +}; > + > static struct platform_driver tegra_devfreq_driver = { > .probe = tegra_devfreq_probe, > .remove = tegra_devfreq_remove, > .driver = { > .name = "tegra20-devfreq", > + .of_match_table = tegra_devfreq_of_match, > }, > }; > module_platform_driver(tegra_devfreq_driver); > > -MODULE_ALIAS("platform:tegra20-devfreq"); > MODULE_AUTHOR("Dmitry Osipenko <digetx@gmail.com>"); > MODULE_DESCRIPTION("NVIDIA Tegra20 devfreq driver"); > MODULE_LICENSE("GPL v2"); > -- > 2.27.0 >
01.11.2020 16:31, Chanwoo Choi пишет: > Hi Dmitry, > > This patch contains the three features as following: > 1. Use interconnect interface for controlling the clock instead of > controlling it direclty > 2. Use EMC_STAT instead of IMC_STAT > 3. Change polling_interval and upthreshold for more fast responsiveness > > I think you need to make the separate patches for each role. > But, if it is difficult or not proper to split out 1,2 roles, you can > make two patches for 1,2 and 3 roles. Hello Chanwoo, We will probably move the Tegra20 EMC_STAT devfreq driver into the memory driver and remove the older IMC_STAT driver in v7, like it was suggested by Thierry Reding. This will be a much less invasive code change. > Also, if you want to get more responsiveness, you could use delayed timer > instead of deferrable timer by editing the devfreq_dev_profile structure. Thanks, I'll try the deferrable timer.
01.11.2020 17:12, Dmitry Osipenko пишет: ... > We will probably move the Tegra20 EMC_STAT devfreq driver into the > memory driver and remove the older IMC_STAT driver in v7, like it was > suggested by Thierry Reding. This will be a much less invasive code change. > >> Also, if you want to get more responsiveness, you could use delayed timer >> instead of deferrable timer by editing the devfreq_dev_profile structure. > > Thanks, I'll try the deferrable timer. I took a brief look at the delayed timer and I think the deferrable timer should be more a preferred option because this devfreq drive is more an assistance for the optimal bandwidth selection and it will be more preferred to keep system idling whenever possible. My primary concern is the initial performance lag in a case of multimedia applications. But this will be resolved by hooking up performance voting to all drivers, once we will get around to it.
On 11/3/20 5:08 AM, Dmitry Osipenko wrote: > 01.11.2020 17:12, Dmitry Osipenko пишет: > ... >> We will probably move the Tegra20 EMC_STAT devfreq driver into the >> memory driver and remove the older IMC_STAT driver in v7, like it was >> suggested by Thierry Reding. This will be a much less invasive code change. >> >>> Also, if you want to get more responsiveness, you could use delayed timer >>> instead of deferrable timer by editing the devfreq_dev_profile structure. >> >> Thanks, I'll try the deferrable timer. > > I took a brief look at the delayed timer and I think the deferrable > timer should be more a preferred option because this devfreq drive is > more an assistance for the optimal bandwidth selection and it will be > more preferred to keep system idling whenever possible. > > My primary concern is the initial performance lag in a case of > multimedia applications. But this will be resolved by hooking up > performance voting to all drivers, once we will get around to it. OK. You can choice the type of timer on both probe and via sysfs file on the runtime.
diff --git a/drivers/devfreq/Kconfig b/drivers/devfreq/Kconfig index 0ee36ae2fa79..1bd225e571df 100644 --- a/drivers/devfreq/Kconfig +++ b/drivers/devfreq/Kconfig @@ -126,6 +126,7 @@ config ARM_TEGRA20_DEVFREQ depends on ARCH_TEGRA_2x_SOC || COMPILE_TEST depends on COMMON_CLK select DEVFREQ_GOV_SIMPLE_ONDEMAND + select MFD_SYSCON help This adds the DEVFREQ driver for the Tegra20 family of SoCs. It reads Memory Controller counters and adjusts the operating diff --git a/drivers/devfreq/tegra20-devfreq.c b/drivers/devfreq/tegra20-devfreq.c index fd801534771d..0a36b085d32a 100644 --- a/drivers/devfreq/tegra20-devfreq.c +++ b/drivers/devfreq/tegra20-devfreq.c @@ -7,180 +7,148 @@ #include <linux/clk.h> #include <linux/devfreq.h> -#include <linux/io.h> #include <linux/kernel.h> #include <linux/module.h> +#include <linux/mfd/syscon.h> #include <linux/of_device.h> #include <linux/platform_device.h> #include <linux/pm_opp.h> +#include <linux/regmap.h> #include <linux/slab.h> -#include <soc/tegra/mc.h> - #include "governor.h" -#define MC_STAT_CONTROL 0x90 -#define MC_STAT_EMC_CLOCK_LIMIT 0xa0 -#define MC_STAT_EMC_CLOCKS 0xa4 -#define MC_STAT_EMC_CONTROL 0xa8 -#define MC_STAT_EMC_COUNT 0xb8 +#define EMC_STAT_CONTROL 0x160 +#define EMC_STAT_LLMC_CONTROL 0x178 +#define EMC_STAT_PWR_CLOCK_LIMIT 0x198 +#define EMC_STAT_PWR_CLOCKS 0x19c +#define EMC_STAT_PWR_COUNT 0x1a0 -#define EMC_GATHER_CLEAR (1 << 8) -#define EMC_GATHER_ENABLE (3 << 8) +#define EMC_PWR_GATHER_CLEAR (1 << 8) +#define EMC_PWR_GATHER_DISABLE (2 << 8) +#define EMC_PWR_GATHER_ENABLE (3 << 8) struct tegra_devfreq { + struct devfreq_simple_ondemand_data ondemand_data; + struct opp_table *opp_table; struct devfreq *devfreq; struct clk *emc_clock; - void __iomem *regs; + struct regmap *rmap; }; static int tegra_devfreq_target(struct device *dev, unsigned long *freq, u32 flags) { - struct tegra_devfreq *tegra = dev_get_drvdata(dev); - struct devfreq *devfreq = tegra->devfreq; struct dev_pm_opp *opp; - unsigned long rate; - int err; + int ret; opp = devfreq_recommended_opp(dev, freq, flags); - if (IS_ERR(opp)) + 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); + ret = dev_pm_opp_set_bw(dev, opp); dev_pm_opp_put(opp); - err = clk_set_min_rate(tegra->emc_clock, rate); - if (err) - return err; - - err = clk_set_rate(tegra->emc_clock, 0); - if (err) - goto restore_min_rate; - - return 0; - -restore_min_rate: - clk_set_min_rate(tegra->emc_clock, devfreq->previous_freq); - - return err; + return ret; } static int tegra_devfreq_get_dev_status(struct device *dev, struct devfreq_dev_status *stat) { struct tegra_devfreq *tegra = dev_get_drvdata(dev); + u32 count, clocks; - /* - * EMC_COUNT returns number of memory events, that number is lower - * than the number of clocks. Conversion ratio of 1/8 results in a - * bit higher bandwidth than actually needed, it is good enough for - * the time being because drivers don't support requesting minimum - * needed memory bandwidth yet. - * - * TODO: adjust the ratio value once relevant drivers will support - * memory bandwidth management. - */ - stat->busy_time = readl_relaxed(tegra->regs + MC_STAT_EMC_COUNT); - stat->total_time = readl_relaxed(tegra->regs + MC_STAT_EMC_CLOCKS) / 8; - stat->current_frequency = clk_get_rate(tegra->emc_clock); + /* freeze counters */ + regmap_write(tegra->rmap, EMC_STAT_CONTROL, EMC_PWR_GATHER_DISABLE); + + /* number of clocks when EMC request was accepted */ + regmap_read(tegra->rmap, EMC_STAT_PWR_COUNT, &count); + /* total number of clocks while PWR_GATHER control was set to ENABLE */ + regmap_read(tegra->rmap, EMC_STAT_PWR_CLOCKS, &clocks); - writel_relaxed(EMC_GATHER_CLEAR, tegra->regs + MC_STAT_CONTROL); - writel_relaxed(EMC_GATHER_ENABLE, tegra->regs + MC_STAT_CONTROL); + /* clear counters and restart */ + regmap_write(tegra->rmap, EMC_STAT_CONTROL, EMC_PWR_GATHER_CLEAR); + regmap_write(tegra->rmap, EMC_STAT_CONTROL, EMC_PWR_GATHER_ENABLE); + + stat->busy_time = count; + stat->total_time = clocks; + stat->current_frequency = clk_get_rate(tegra->emc_clock); return 0; } static struct devfreq_dev_profile tegra_devfreq_profile = { - .polling_ms = 500, + .polling_ms = 30, .target = tegra_devfreq_target, .get_dev_status = tegra_devfreq_get_dev_status, }; -static struct tegra_mc *tegra_get_memory_controller(void) -{ - struct platform_device *pdev; - struct device_node *np; - struct tegra_mc *mc; - - np = of_find_compatible_node(NULL, NULL, "nvidia,tegra20-mc-gart"); - if (!np) - return ERR_PTR(-ENOENT); - - pdev = of_find_device_by_node(np); - of_node_put(np); - if (!pdev) - return ERR_PTR(-ENODEV); - - mc = platform_get_drvdata(pdev); - if (!mc) - return ERR_PTR(-EPROBE_DEFER); - - return mc; -} - static int tegra_devfreq_probe(struct platform_device *pdev) { + struct device_node *emc_np = pdev->dev.parent->of_node; struct tegra_devfreq *tegra; - struct tegra_mc *mc; - unsigned long max_rate; - unsigned long rate; int err; - mc = tegra_get_memory_controller(); - if (IS_ERR(mc)) { - err = PTR_ERR(mc); - dev_err(&pdev->dev, "failed to get memory controller: %d\n", - err); - return err; - } - tegra = devm_kzalloc(&pdev->dev, sizeof(*tegra), GFP_KERNEL); if (!tegra) return -ENOMEM; /* EMC is a system-critical clock that is always enabled */ - tegra->emc_clock = devm_clk_get(&pdev->dev, "emc"); + tegra->emc_clock = devm_get_clk_from_child(&pdev->dev, emc_np, NULL); if (IS_ERR(tegra->emc_clock)) return dev_err_probe(&pdev->dev, PTR_ERR(tegra->emc_clock), "failed to get emc clock\n"); - tegra->regs = mc->regs; - - max_rate = clk_round_rate(tegra->emc_clock, ULONG_MAX); + tegra->rmap = device_node_to_regmap(emc_np); + if (IS_ERR(tegra->rmap)) + return dev_err_probe(&pdev->dev, PTR_ERR(tegra->rmap), + "failed to get emc regmap\n"); - for (rate = 0; rate <= max_rate; rate++) { - rate = clk_round_rate(tegra->emc_clock, rate); + tegra->opp_table = dev_pm_opp_get_opp_table(&pdev->dev); + if (IS_ERR(tegra->opp_table)) + return dev_err_probe(&pdev->dev, PTR_ERR(tegra->opp_table), + "failed to prepare opp table\n"); - err = dev_pm_opp_add(&pdev->dev, rate, 0); - if (err) { - dev_err(&pdev->dev, "failed to add opp: %d\n", err); - goto remove_opps; - } + err = dev_pm_opp_of_add_table(&pdev->dev); + if (err) { + dev_err(&pdev->dev, "failed to add opp table: %d\n", err); + goto put_table; } + /* + * 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. + */ + tegra->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, tegra->regs + MC_STAT_CONTROL); - writel_relaxed(0x00000000, tegra->regs + MC_STAT_EMC_CONTROL); - writel_relaxed(0xffffffff, tegra->regs + MC_STAT_EMC_CLOCK_LIMIT); + regmap_write(tegra->rmap, EMC_STAT_CONTROL, 0x00000000); + regmap_write(tegra->rmap, EMC_STAT_LLMC_CONTROL, 0x00000000); + regmap_write(tegra->rmap, EMC_STAT_PWR_CLOCK_LIMIT, 0xffffffff); platform_set_drvdata(pdev, tegra); tegra->devfreq = devfreq_add_device(&pdev->dev, &tegra_devfreq_profile, - DEVFREQ_GOV_SIMPLE_ONDEMAND, NULL); + DEVFREQ_GOV_SIMPLE_ONDEMAND, + &tegra->ondemand_data); if (IS_ERR(tegra->devfreq)) { err = PTR_ERR(tegra->devfreq); - goto remove_opps; + goto put_table; } return 0; -remove_opps: - dev_pm_opp_remove_all_dynamic(&pdev->dev); +put_table: + dev_pm_opp_put_opp_table(tegra->opp_table); return err; } @@ -190,21 +158,27 @@ static int tegra_devfreq_remove(struct platform_device *pdev) struct tegra_devfreq *tegra = platform_get_drvdata(pdev); devfreq_remove_device(tegra->devfreq); - dev_pm_opp_remove_all_dynamic(&pdev->dev); + dev_pm_opp_of_remove_table(&pdev->dev); + dev_pm_opp_put_opp_table(tegra->opp_table); return 0; } +static const struct of_device_id tegra_devfreq_of_match[] = { + { .compatible = "nvidia,tegra20-emc-statistics" }, + { }, +}; + static struct platform_driver tegra_devfreq_driver = { .probe = tegra_devfreq_probe, .remove = tegra_devfreq_remove, .driver = { .name = "tegra20-devfreq", + .of_match_table = tegra_devfreq_of_match, }, }; module_platform_driver(tegra_devfreq_driver); -MODULE_ALIAS("platform:tegra20-devfreq"); MODULE_AUTHOR("Dmitry Osipenko <digetx@gmail.com>"); MODULE_DESCRIPTION("NVIDIA Tegra20 devfreq driver"); MODULE_LICENSE("GPL v2");
External (EMC) and Internal Memory Controllers (IMC) have a nearly identical statistics gathering module. This patch switches driver to use EMC_STAT instead of IMC_STAT and adds device-tree support which brings ICC support and makes driver to use bandwidth OPPs defined in device-tree. The previous tegra20-devfreq variant was depending on presence of both EMC and IMC drivers simultaneously because it wasn't apparent how to use EMC_STAT properly back in the day. Dependency on the IMC driver is gone after this patch. The older variant of the devfreq driver also isn't suitable anymore because EMC got support for interconnect framework and DVFS, hence tegra20-devfreq shouldn't drive the EMC clock directly, but use OPP API for issuing memory bandwidth requests. The polling interval is changed from 500ms to 30ms in order to improve responsiveness of the system in general and because EMC clock is now allowed to go lower than before since display driver supports ICC now as well. The parent EMC device is an MFD device now and tegra20-devfreq its sub-device. Devfreq driver uses SYSCON API for retrieving regmap of the EMC registers from the parent device. Signed-off-by: Dmitry Osipenko <digetx@gmail.com> --- drivers/devfreq/Kconfig | 1 + drivers/devfreq/tegra20-devfreq.c | 174 +++++++++++++----------------- 2 files changed, 75 insertions(+), 100 deletions(-)