drm/nouveau: Support devfreq for Tegra

Using pmu counters for usage stats. This enables dynamic frequency
scaling on all of the currently supported Tegra gpus.

The register offsets are valid for gk20a, gm20b, gp10b, and gv11b. If
support is added for ga10b, this will need rearchitected.

Signed-off-by: Aaron Kling <webgeek1234@gmail.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
[fixed tab alignment in gk20a_devfreq_target()]
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://lore.kernel.org/r/20250906-gk20a-devfreq-v2-1-0217f53ee355@gmail.com

drivers/gpu/drm/nouveau/Kconfig

@@ -28,6 +28,7 @@ config DRM_NOUVEAU
 	select THERMAL if ACPI && X86
 	select ACPI_VIDEO if ACPI && X86
 	select SND_HDA_COMPONENT if SND_HDA_CORE
+	select PM_DEVFREQ if ARCH_TEGRA
 	help
 	  Choose this option for open-source NVIDIA support.
 

drivers/gpu/drm/nouveau/include/nvkm/core/tegra.h

@@ -9,6 +9,8 @@ struct nvkm_device_tegra {
 	struct nvkm_device device;
 	struct platform_device *pdev;
 
+	void __iomem *regs;
+
 	struct reset_control *rst;
 	struct clk *clk;
 	struct clk *clk_ref;

drivers/gpu/drm/nouveau/nouveau_platform.c

@@ -21,6 +21,8 @@
  */
 #include "nouveau_platform.h"
 
+#include <nvkm/subdev/clk/gk20a_devfreq.h>
+
 static int nouveau_platform_probe(struct platform_device *pdev)
 {
 	const struct nvkm_device_tegra_func *func;
@@ -40,6 +42,21 @@ static void nouveau_platform_remove(struct platform_device *pdev)
 	nouveau_drm_device_remove(drm);
 }
 
+#ifdef CONFIG_PM_SLEEP
+static int nouveau_platform_suspend(struct device *dev)
+{
+	return gk20a_devfreq_suspend(dev);
+}
+
+static int nouveau_platform_resume(struct device *dev)
+{
+	return gk20a_devfreq_resume(dev);
+}
+
+static SIMPLE_DEV_PM_OPS(nouveau_pm_ops, nouveau_platform_suspend,
+			 nouveau_platform_resume);
+#endif
+
 #if IS_ENABLED(CONFIG_OF)
 static const struct nvkm_device_tegra_func gk20a_platform_data = {
 	.iommu_bit = 34,
@@ -81,6 +98,9 @@ struct platform_driver nouveau_platform_driver = {
 	.driver = {
 		.name = "nouveau",
 		.of_match_table = of_match_ptr(nouveau_platform_match),
+#ifdef CONFIG_PM_SLEEP
+		.pm = &nouveau_pm_ops,
+#endif
 	},
 	.probe = nouveau_platform_probe,
 	.remove = nouveau_platform_remove,

drivers/gpu/drm/nouveau/nvkm/engine/device/tegra.c

@@ -259,6 +259,10 @@ nvkm_device_tegra_new(const struct nvkm_device_tegra_func *func,
 	tdev->func = func;
 	tdev->pdev = pdev;
 
+	tdev->regs = devm_platform_ioremap_resource(pdev, 0);
+	if (IS_ERR(tdev->regs))
+		return PTR_ERR(tdev->regs);
+
 	if (func->require_vdd) {
 		tdev->vdd = devm_regulator_get(&pdev->dev, "vdd");
 		if (IS_ERR(tdev->vdd)) {

drivers/gpu/drm/nouveau/nvkm/subdev/clk/Kbuild

@@ -11,6 +11,7 @@ nvkm-y += nvkm/subdev/clk/gk104.o
 nvkm-y += nvkm/subdev/clk/gk20a.o
 nvkm-y += nvkm/subdev/clk/gm20b.o
 nvkm-y += nvkm/subdev/clk/gp10b.o
+nvkm-$(CONFIG_PM_DEVFREQ) += nvkm/subdev/clk/gk20a_devfreq.o
 
 nvkm-y += nvkm/subdev/clk/pllnv04.o
 nvkm-y += nvkm/subdev/clk/pllgt215.o

drivers/gpu/drm/nouveau/nvkm/subdev/clk/gk20a.c

@@ -23,6 +23,7 @@
  *
  */
 #include "priv.h"
+#include "gk20a_devfreq.h"
 #include "gk20a.h"
 
 #include <core/tegra.h>
@@ -589,6 +590,10 @@ gk20a_clk_init(struct nvkm_clk *base)
 		return ret;
 	}
 
+	ret = gk20a_devfreq_init(base, &clk->devfreq);
+	if (ret)
+		return ret;
+
 	return 0;
 }
 

drivers/gpu/drm/nouveau/nvkm/subdev/clk/gk20a.h

@@ -118,6 +118,7 @@ struct gk20a_clk {
 	const struct gk20a_clk_pllg_params *params;
 	struct gk20a_pll pll;
 	u32 parent_rate;
+	struct gk20a_devfreq *devfreq;
 
 	u32 (*div_to_pl)(u32);
 	u32 (*pl_to_div)(u32);

drivers/gpu/drm/nouveau/nvkm/subdev/clk/gk20a_devfreq.c

@@ -0,0 +1,320 @@
+// SPDX-License-Identifier: MIT
+#include <linux/clk.h>
+#include <linux/math64.h>
+#include <linux/platform_device.h>
+#include <linux/pm_opp.h>
+
+#include <drm/drm_managed.h>
+
+#include <subdev/clk.h>
+
+#include "nouveau_drv.h"
+#include "nouveau_chan.h"
+#include "priv.h"
+#include "gk20a_devfreq.h"
+#include "gk20a.h"
+#include "gp10b.h"
+
+#define PMU_BUSY_CYCLES_NORM_MAX		1000U
+
+#define PWR_PMU_IDLE_COUNTER_TOTAL		0U
+#define PWR_PMU_IDLE_COUNTER_BUSY		4U
+
+#define PWR_PMU_IDLE_COUNT_REG_OFFSET		0x0010A508U
+#define PWR_PMU_IDLE_COUNT_REG_SIZE		16U
+#define PWR_PMU_IDLE_COUNT_MASK			0x7FFFFFFFU
+#define PWR_PMU_IDLE_COUNT_RESET_VALUE		(0x1U << 31U)
+
+#define PWR_PMU_IDLE_INTR_REG_OFFSET		0x0010A9E8U
+#define PWR_PMU_IDLE_INTR_ENABLE_VALUE		0U
+
+#define PWR_PMU_IDLE_INTR_STATUS_REG_OFFSET	0x0010A9ECU
+#define PWR_PMU_IDLE_INTR_STATUS_MASK		0x00000001U
+#define PWR_PMU_IDLE_INTR_STATUS_RESET_VALUE	0x1U
+
+#define PWR_PMU_IDLE_THRESHOLD_REG_OFFSET	0x0010A8A0U
+#define PWR_PMU_IDLE_THRESHOLD_REG_SIZE		4U
+#define PWR_PMU_IDLE_THRESHOLD_MAX_VALUE	0x7FFFFFFFU
+
+#define PWR_PMU_IDLE_CTRL_REG_OFFSET		0x0010A50CU
+#define PWR_PMU_IDLE_CTRL_REG_SIZE		16U
+#define PWR_PMU_IDLE_CTRL_VALUE_MASK		0x3U
+#define PWR_PMU_IDLE_CTRL_VALUE_BUSY		0x2U
+#define PWR_PMU_IDLE_CTRL_VALUE_ALWAYS		0x3U
+#define PWR_PMU_IDLE_CTRL_FILTER_MASK		(0x1U << 2)
+#define PWR_PMU_IDLE_CTRL_FILTER_DISABLED	0x0U
+
+#define PWR_PMU_IDLE_MASK_REG_OFFSET		0x0010A504U
+#define PWR_PMU_IDLE_MASK_REG_SIZE		16U
+#define PWM_PMU_IDLE_MASK_GR_ENABLED		0x1U
+#define PWM_PMU_IDLE_MASK_CE_2_ENABLED		0x200000U
+
+/**
+ * struct gk20a_devfreq - Device frequency management
+ */
+struct gk20a_devfreq {
+	/** @devfreq: devfreq device. */
+	struct devfreq *devfreq;
+
+	/** @regs: Device registers. */
+	void __iomem *regs;
+
+	/** @gov_data: Governor data. */
+	struct devfreq_simple_ondemand_data gov_data;
+
+	/** @busy_time: Busy time. */
+	ktime_t busy_time;
+
+	/** @total_time: Total time. */
+	ktime_t total_time;
+
+	/** @time_last_update: Last update time. */
+	ktime_t time_last_update;
+};
+
+static struct gk20a_devfreq *dev_to_gk20a_devfreq(struct device *dev)
+{
+	struct nouveau_drm *drm = dev_get_drvdata(dev);
+	struct nvkm_subdev *subdev = nvkm_device_subdev(drm->nvkm, NVKM_SUBDEV_CLK, 0);
+	struct nvkm_clk *base = nvkm_clk(subdev);
+
+	switch (drm->nvkm->chipset) {
+	case 0x13b: return gp10b_clk(base)->devfreq; break;
+	default: return gk20a_clk(base)->devfreq; break;
+	}
+}
+
+static void gk20a_pmu_init_perfmon_counter(struct gk20a_devfreq *gdevfreq)
+{
+	u32 data;
+
+	// Set pmu idle intr status bit on total counter overflow
+	writel(PWR_PMU_IDLE_INTR_ENABLE_VALUE,
+	       gdevfreq->regs + PWR_PMU_IDLE_INTR_REG_OFFSET);
+
+	writel(PWR_PMU_IDLE_THRESHOLD_MAX_VALUE,
+	       gdevfreq->regs + PWR_PMU_IDLE_THRESHOLD_REG_OFFSET +
+	       (PWR_PMU_IDLE_COUNTER_TOTAL * PWR_PMU_IDLE_THRESHOLD_REG_SIZE));
+
+	// Setup counter for total cycles
+	data = readl(gdevfreq->regs + PWR_PMU_IDLE_CTRL_REG_OFFSET +
+		     (PWR_PMU_IDLE_COUNTER_TOTAL * PWR_PMU_IDLE_CTRL_REG_SIZE));
+	data &= ~(PWR_PMU_IDLE_CTRL_VALUE_MASK | PWR_PMU_IDLE_CTRL_FILTER_MASK);
+	data |= PWR_PMU_IDLE_CTRL_VALUE_ALWAYS | PWR_PMU_IDLE_CTRL_FILTER_DISABLED;
+	writel(data, gdevfreq->regs + PWR_PMU_IDLE_CTRL_REG_OFFSET +
+		     (PWR_PMU_IDLE_COUNTER_TOTAL * PWR_PMU_IDLE_CTRL_REG_SIZE));
+
+	// Setup counter for busy cycles
+	writel(PWM_PMU_IDLE_MASK_GR_ENABLED | PWM_PMU_IDLE_MASK_CE_2_ENABLED,
+	       gdevfreq->regs + PWR_PMU_IDLE_MASK_REG_OFFSET +
+	       (PWR_PMU_IDLE_COUNTER_BUSY * PWR_PMU_IDLE_MASK_REG_SIZE));
+
+	data = readl(gdevfreq->regs + PWR_PMU_IDLE_CTRL_REG_OFFSET +
+		     (PWR_PMU_IDLE_COUNTER_BUSY * PWR_PMU_IDLE_CTRL_REG_SIZE));
+	data &= ~(PWR_PMU_IDLE_CTRL_VALUE_MASK | PWR_PMU_IDLE_CTRL_FILTER_MASK);
+	data |= PWR_PMU_IDLE_CTRL_VALUE_BUSY | PWR_PMU_IDLE_CTRL_FILTER_DISABLED;
+	writel(data, gdevfreq->regs + PWR_PMU_IDLE_CTRL_REG_OFFSET +
+		     (PWR_PMU_IDLE_COUNTER_BUSY * PWR_PMU_IDLE_CTRL_REG_SIZE));
+}
+
+static u32 gk20a_pmu_read_idle_counter(struct gk20a_devfreq *gdevfreq, u32 counter_id)
+{
+	u32 ret;
+
+	ret = readl(gdevfreq->regs + PWR_PMU_IDLE_COUNT_REG_OFFSET +
+		    (counter_id * PWR_PMU_IDLE_COUNT_REG_SIZE));
+
+	return ret & PWR_PMU_IDLE_COUNT_MASK;
+}
+
+static void gk20a_pmu_reset_idle_counter(struct gk20a_devfreq *gdevfreq, u32 counter_id)
+{
+	writel(PWR_PMU_IDLE_COUNT_RESET_VALUE, gdevfreq->regs + PWR_PMU_IDLE_COUNT_REG_OFFSET +
+					       (counter_id * PWR_PMU_IDLE_COUNT_REG_SIZE));
+}
+
+static u32 gk20a_pmu_read_idle_intr_status(struct gk20a_devfreq *gdevfreq)
+{
+	u32 ret;
+
+	ret = readl(gdevfreq->regs + PWR_PMU_IDLE_INTR_STATUS_REG_OFFSET);
+
+	return ret & PWR_PMU_IDLE_INTR_STATUS_MASK;
+}
+
+static void gk20a_pmu_clear_idle_intr_status(struct gk20a_devfreq *gdevfreq)
+{
+	writel(PWR_PMU_IDLE_INTR_STATUS_RESET_VALUE,
+	       gdevfreq->regs + PWR_PMU_IDLE_INTR_STATUS_REG_OFFSET);
+}
+
+static void gk20a_devfreq_update_utilization(struct gk20a_devfreq *gdevfreq)
+{
+	ktime_t now, last;
+	u64 busy_cycles, total_cycles;
+	u32 norm, intr_status;
+
+	now = ktime_get();
+	last = gdevfreq->time_last_update;
+	gdevfreq->total_time = ktime_us_delta(now, last);
+
+	busy_cycles = gk20a_pmu_read_idle_counter(gdevfreq, PWR_PMU_IDLE_COUNTER_BUSY);
+	total_cycles = gk20a_pmu_read_idle_counter(gdevfreq, PWR_PMU_IDLE_COUNTER_TOTAL);
+	intr_status = gk20a_pmu_read_idle_intr_status(gdevfreq);
+
+	gk20a_pmu_reset_idle_counter(gdevfreq, PWR_PMU_IDLE_COUNTER_BUSY);
+	gk20a_pmu_reset_idle_counter(gdevfreq, PWR_PMU_IDLE_COUNTER_TOTAL);
+
+	if (intr_status != 0UL) {
+		norm = PMU_BUSY_CYCLES_NORM_MAX;
+		gk20a_pmu_clear_idle_intr_status(gdevfreq);
+	} else if (total_cycles == 0ULL || busy_cycles > total_cycles) {
+		norm = PMU_BUSY_CYCLES_NORM_MAX;
+	} else {
+		norm = (u32)div64_u64(busy_cycles * PMU_BUSY_CYCLES_NORM_MAX,
+				total_cycles);
+	}
+
+	gdevfreq->busy_time = div_u64(gdevfreq->total_time * norm, PMU_BUSY_CYCLES_NORM_MAX);
+	gdevfreq->time_last_update = now;
+}
+
+static int gk20a_devfreq_target(struct device *dev, unsigned long *freq,
+				u32 flags)
+{
+	struct nouveau_drm *drm = dev_get_drvdata(dev);
+	struct nvkm_subdev *subdev = nvkm_device_subdev(drm->nvkm, NVKM_SUBDEV_CLK, 0);
+	struct nvkm_clk *base = nvkm_clk(subdev);
+	struct nvkm_pstate *pstates = base->func->pstates;
+	int nr_pstates = base->func->nr_pstates;
+	int i, ret;
+
+	for (i = 0; i < nr_pstates - 1; i++)
+		if (pstates[i].base.domain[nv_clk_src_gpc] * GK20A_CLK_GPC_MDIV >= *freq)
+			break;
+
+	ret = nvkm_clk_ustate(base, pstates[i].pstate, 0);
+	ret |= nvkm_clk_ustate(base, pstates[i].pstate, 1);
+	if (ret) {
+		nvkm_error(subdev, "cannot update clock\n");
+		return ret;
+	}
+
+	*freq = pstates[i].base.domain[nv_clk_src_gpc] * GK20A_CLK_GPC_MDIV;
+
+	return 0;
+}
+
+static int gk20a_devfreq_get_cur_freq(struct device *dev, unsigned long *freq)
+{
+	struct nouveau_drm *drm = dev_get_drvdata(dev);
+	struct nvkm_subdev *subdev = nvkm_device_subdev(drm->nvkm, NVKM_SUBDEV_CLK, 0);
+	struct nvkm_clk *base = nvkm_clk(subdev);
+
+	*freq = nvkm_clk_read(base, nv_clk_src_gpc) * GK20A_CLK_GPC_MDIV;
+
+	return 0;
+}
+
+static void gk20a_devfreq_reset(struct gk20a_devfreq *gdevfreq)
+{
+	gk20a_pmu_reset_idle_counter(gdevfreq, PWR_PMU_IDLE_COUNTER_BUSY);
+	gk20a_pmu_reset_idle_counter(gdevfreq, PWR_PMU_IDLE_COUNTER_TOTAL);
+	gk20a_pmu_clear_idle_intr_status(gdevfreq);
+
+	gdevfreq->busy_time = 0;
+	gdevfreq->total_time = 0;
+	gdevfreq->time_last_update = ktime_get();
+}
+
+static int gk20a_devfreq_get_dev_status(struct device *dev,
+					struct devfreq_dev_status *status)
+{
+	struct nouveau_drm *drm = dev_get_drvdata(dev);
+	struct gk20a_devfreq *gdevfreq = dev_to_gk20a_devfreq(dev);
+
+	gk20a_devfreq_get_cur_freq(dev, &status->current_frequency);
+
+	gk20a_devfreq_update_utilization(gdevfreq);
+
+	status->busy_time = ktime_to_ns(gdevfreq->busy_time);
+	status->total_time = ktime_to_ns(gdevfreq->total_time);
+
+	gk20a_devfreq_reset(gdevfreq);
+
+	NV_DEBUG(drm, "busy %lu total %lu %lu %% freq %lu MHz\n",
+		 status->busy_time, status->total_time,
+		 status->busy_time / (status->total_time / 100),
+		 status->current_frequency / 1000 / 1000);
+
+	return 0;
+}
+
+static struct devfreq_dev_profile gk20a_devfreq_profile = {
+	.timer = DEVFREQ_TIMER_DELAYED,
+	.polling_ms = 50,
+	.target = gk20a_devfreq_target,
+	.get_cur_freq = gk20a_devfreq_get_cur_freq,
+	.get_dev_status = gk20a_devfreq_get_dev_status,
+};
+
+int gk20a_devfreq_init(struct nvkm_clk *base, struct gk20a_devfreq **gdevfreq)
+{
+	struct nvkm_device *device = base->subdev.device;
+	struct nouveau_drm *drm = dev_get_drvdata(device->dev);
+	struct nvkm_device_tegra *tdev = device->func->tegra(device);
+	struct nvkm_pstate *pstates = base->func->pstates;
+	int nr_pstates = base->func->nr_pstates;
+	struct gk20a_devfreq *new_gdevfreq;
+	int i;
+
+	new_gdevfreq = drmm_kzalloc(drm->dev, sizeof(struct gk20a_devfreq), GFP_KERNEL);
+	if (!new_gdevfreq)
+		return -ENOMEM;
+
+	new_gdevfreq->regs = tdev->regs;
+
+	for (i = 0; i < nr_pstates; i++)
+		dev_pm_opp_add(base->subdev.device->dev,
+			       pstates[i].base.domain[nv_clk_src_gpc] * GK20A_CLK_GPC_MDIV, 0);
+
+	gk20a_pmu_init_perfmon_counter(new_gdevfreq);
+	gk20a_devfreq_reset(new_gdevfreq);
+
+	gk20a_devfreq_profile.initial_freq =
+		nvkm_clk_read(base, nv_clk_src_gpc) * GK20A_CLK_GPC_MDIV;
+
+	new_gdevfreq->gov_data.upthreshold = 45;
+	new_gdevfreq->gov_data.downdifferential = 5;
+
+	new_gdevfreq->devfreq = devm_devfreq_add_device(device->dev,
+							&gk20a_devfreq_profile,
+							DEVFREQ_GOV_SIMPLE_ONDEMAND,
+							&new_gdevfreq->gov_data);
+	if (IS_ERR(new_gdevfreq->devfreq))
+		return PTR_ERR(new_gdevfreq->devfreq);
+
+	*gdevfreq = new_gdevfreq;
+
+	return 0;
+}
+
+int gk20a_devfreq_resume(struct device *dev)
+{
+	struct gk20a_devfreq *gdevfreq = dev_to_gk20a_devfreq(dev);
+
+	if (!gdevfreq || !gdevfreq->devfreq)
+		return 0;
+
+	return devfreq_resume_device(gdevfreq->devfreq);
+}
+
+int gk20a_devfreq_suspend(struct device *dev)
+{
+	struct gk20a_devfreq *gdevfreq = dev_to_gk20a_devfreq(dev);
+
+	if (!gdevfreq || !gdevfreq->devfreq)
+		return 0;
+
+	return devfreq_suspend_device(gdevfreq->devfreq);
+}

drivers/gpu/drm/nouveau/nvkm/subdev/clk/gk20a_devfreq.h

@@ -0,0 +1,24 @@
+/* SPDX-License-Identifier: MIT */
+#ifndef __GK20A_DEVFREQ_H__
+#define __GK20A_DEVFREQ_H__
+
+#include <linux/devfreq.h>
+
+struct gk20a_devfreq;
+
+#if defined(CONFIG_PM_DEVFREQ)
+int gk20a_devfreq_init(struct nvkm_clk *base, struct gk20a_devfreq **devfreq);
+
+int gk20a_devfreq_resume(struct device *dev);
+int gk20a_devfreq_suspend(struct device *dev);
+#else
+static inline int gk20a_devfreq_init(struct nvkm_clk *base, struct gk20a_devfreq **devfreq)
+{
+	return 0;
+}
+
+static inline int gk20a_devfreq_resume(struct device dev) { return 0; }
+static inline int gk20a_devfreq_suspend(struct device *dev) { return 0; }
+#endif /* CONFIG_PM_DEVFREQ */
+
+#endif /* __GK20A_DEVFREQ_H__ */

drivers/gpu/drm/nouveau/nvkm/subdev/clk/gm20b.c

@@ -27,6 +27,7 @@
 #include <core/tegra.h>
 
 #include "priv.h"
+#include "gk20a_devfreq.h"
 #include "gk20a.h"
 
 #define GPCPLL_CFG_SYNC_MODE	BIT(2)
@@ -869,6 +870,10 @@ gm20b_clk_init(struct nvkm_clk *base)
 		return ret;
 	}
 
+	ret = gk20a_devfreq_init(base, &clk->devfreq);
+	if (ret)
+		return ret;
+
 	return 0;
 }
 

drivers/gpu/drm/nouveau/nvkm/subdev/clk/gp10b.c

@@ -5,6 +5,7 @@
 #include <core/tegra.h>
 
 #include "priv.h"
+#include "gk20a_devfreq.h"
 #include "gk20a.h"
 #include "gp10b.h"
 
@@ -23,6 +24,10 @@ gp10b_clk_init(struct nvkm_clk *base)
 		return ret;
 	}
 
+	ret = gk20a_devfreq_init(base, &clk->devfreq);
+	if (ret)
+		return ret;
+
 	return 0;
 }
 

drivers/gpu/drm/nouveau/nvkm/subdev/clk/gp10b.h

@@ -5,6 +5,7 @@
 struct gp10b_clk {
 	/* currently applied parameters */
 	struct nvkm_clk base;
+	struct gk20a_devfreq *devfreq;
 	struct clk *clk;
 	u32 rate;