[TEGRA194_CPUFREQ Patch v3 0/4] Add cpufreq driver for Tegra194

[TEGRA194_CPUFREQ Patch v3 0/4] Add cpufreq driver for Tegra194

[Sumit Gupta]

The patch series adds cpufreq driver for Tegra194 SOC.

v2[2] -> v3
- Set same policy for all cpus in a cluster[Viresh].
- Add compatible string for CPU Complex under cpus node[Thierry].
- Add reference to bpmp node under cpus node[Thierry].
- Bind cpufreq driver to CPU Complex compatible string[Thierry].
- Remove patch to get bpmp data as now using cpus node to get that[Thierry].

v1[1] -> v2:
- Remove cpufreq_lock mutex from tegra194_cpufreq_set_target [Viresh].
- Remove CPUFREQ_ASYNC_NOTIFICATION flag [Viresh].
- Remove redundant _begin|end() call from tegra194_cpufreq_set_target.
- Rename opp_table to freq_table [Viresh].

Sumit Gupta (4):
  dt-bindings: arm: Add t194 ccplex compatible and bpmp property
  arm64: tegra: Add t194 ccplex compatible and bpmp property
  cpufreq: Add Tegra194 cpufreq driver
  arm64: defconfig: Enable CONFIG_ARM_TEGRA194_CPUFREQ

 Documentation/devicetree/bindings/arm/cpus.yaml |   9 +
 arch/arm64/boot/dts/nvidia/tegra194.dtsi        |   2 +
 arch/arm64/configs/defconfig                    |   1 +
 drivers/cpufreq/Kconfig.arm                     |   6 +
 drivers/cpufreq/Makefile                        |   1 +
 drivers/cpufreq/tegra194-cpufreq.c              | 403 ++++++++++++++++++++++++
 6 files changed, 422 insertions(+)
 create mode 100644 drivers/cpufreq/tegra194-cpufreq.c

[1] https://marc.info/?t=157539452300001&r=1&w=2
[2] https://marc.info/?l=linux-tegra&m=158602857106213&w=2 
-- 
2.7.4

[TEGRA194_CPUFREQ Patch v3 1/4] dt-bindings: arm: Add t194 ccplex compatible and bpmp property

[Sumit Gupta]

To do frequency scaling on all CPUs within T194 CPU Complex, we need
to query BPMP for data on valid operating points. Document a compatible
string under 'cpus' node to represent the CPU Complex for binding drivers
like cpufreq which don't have their node or CPU Complex node to bind to.
Also, document a property to point to the BPMP device that can be queried
for all CPUs.

Signed-off-by: Sumit Gupta <sumitg@nvidia.com>
---
 Documentation/devicetree/bindings/arm/cpus.yaml | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/Documentation/devicetree/bindings/arm/cpus.yaml b/Documentation/devicetree/bindings/arm/cpus.yaml
index a018147..737b55e 100644
--- a/Documentation/devicetree/bindings/arm/cpus.yaml
+++ b/Documentation/devicetree/bindings/arm/cpus.yaml
@@ -162,6 +162,7 @@ properties:
       - nvidia,tegra132-denver
       - nvidia,tegra186-denver
       - nvidia,tegra194-carmel
+      - nvidia,tegra194-ccplex
       - qcom,krait
       - qcom,kryo
       - qcom,kryo260
@@ -255,6 +256,14 @@ properties:
 
       where voltage is in V, frequency is in MHz.
 
+  nvidia,bpmp:
+    $ref: '/schemas/types.yaml#/definitions/phandle'
+    descrption: |
+      Specifies the bpmp node that needs to be queried to get
+      operating point data for all CPUs.
+
+      Optional for NVIDIA Tegra194 Carmel CPUs
+
   power-domains:
     $ref: '/schemas/types.yaml#/definitions/phandle-array'
     description:
-- 
2.7.4

[TEGRA194_CPUFREQ Patch v3 2/4] arm64: tegra: Add t194 ccplex compatible and bpmp property

[Sumit Gupta]

On Tegra194, data on valid operating points for the CPUs needs to be
queried from BPMP. In T194, there is no node representing CPU complex.
So, add compatible string to the 'cpus' node instead of using dummy
node to bind cpufreq driver. Also, add reference to the BPMP instance
for the CPU complex.

Signed-off-by: Sumit Gupta <sumitg@nvidia.com>
---
 arch/arm64/boot/dts/nvidia/tegra194.dtsi | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/arch/arm64/boot/dts/nvidia/tegra194.dtsi b/arch/arm64/boot/dts/nvidia/tegra194.dtsi
index 7c9511a..0abf287 100644
--- a/arch/arm64/boot/dts/nvidia/tegra194.dtsi
+++ b/arch/arm64/boot/dts/nvidia/tegra194.dtsi
@@ -1764,6 +1764,8 @@
 	};
 
 	cpus {
+		compatible = "nvidia,tegra194-ccplex";
+		nvidia,bpmp = <&bpmp>;
 		#address-cells = <1>;
 		#size-cells = <0>;
 
-- 
2.7.4

[TEGRA194_CPUFREQ Patch v3 3/4] cpufreq: Add Tegra194 cpufreq driver

[Sumit Gupta]

Add support for CPU frequency scaling on Tegra194. The frequency
of each core can be adjusted by writing a clock divisor value to
a MSR on the core. The range of valid divisors is queried from
the BPMP.

Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Sumit Gupta <sumitg@nvidia.com>
---
 drivers/cpufreq/Kconfig.arm        |   6 +
 drivers/cpufreq/Makefile           |   1 +
 drivers/cpufreq/tegra194-cpufreq.c | 403 +++++++++++++++++++++++++++++++++++++
 3 files changed, 410 insertions(+)
 create mode 100644 drivers/cpufreq/tegra194-cpufreq.c

diff --git a/drivers/cpufreq/Kconfig.arm b/drivers/cpufreq/Kconfig.arm
index 15c1a12..f3d8f09 100644
--- a/drivers/cpufreq/Kconfig.arm
+++ b/drivers/cpufreq/Kconfig.arm
@@ -314,6 +314,12 @@ config ARM_TEGRA186_CPUFREQ
 	help
 	  This adds the CPUFreq driver support for Tegra186 SOCs.
 
+config ARM_TEGRA194_CPUFREQ
+	tristate "Tegra194 CPUFreq support"
+	depends on ARCH_TEGRA && TEGRA_BPMP
+	help
+	  This adds CPU frequency driver support for Tegra194 SOCs.
+
 config ARM_TI_CPUFREQ
 	bool "Texas Instruments CPUFreq support"
 	depends on ARCH_OMAP2PLUS
diff --git a/drivers/cpufreq/Makefile b/drivers/cpufreq/Makefile
index f6670c4..66b5563 100644
--- a/drivers/cpufreq/Makefile
+++ b/drivers/cpufreq/Makefile
@@ -83,6 +83,7 @@ obj-$(CONFIG_ARM_TANGO_CPUFREQ)		+= tango-cpufreq.o
 obj-$(CONFIG_ARM_TEGRA20_CPUFREQ)	+= tegra20-cpufreq.o
 obj-$(CONFIG_ARM_TEGRA124_CPUFREQ)	+= tegra124-cpufreq.o
 obj-$(CONFIG_ARM_TEGRA186_CPUFREQ)	+= tegra186-cpufreq.o
+obj-$(CONFIG_ARM_TEGRA194_CPUFREQ)	+= tegra194-cpufreq.o
 obj-$(CONFIG_ARM_TI_CPUFREQ)		+= ti-cpufreq.o
 obj-$(CONFIG_ARM_VEXPRESS_SPC_CPUFREQ)	+= vexpress-spc-cpufreq.o
 
diff --git a/drivers/cpufreq/tegra194-cpufreq.c b/drivers/cpufreq/tegra194-cpufreq.c
new file mode 100644
index 0000000..8de8000
--- /dev/null
+++ b/drivers/cpufreq/tegra194-cpufreq.c
@@ -0,0 +1,403 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved
+ */
+
+#include <linux/cpu.h>
+#include <linux/cpufreq.h>
+#include <linux/delay.h>
+#include <linux/dma-mapping.h>
+#include <linux/module.h>
+#include <linux/of.h>
+#include <linux/of_platform.h>
+#include <linux/platform_device.h>
+#include <linux/slab.h>
+
+#include <asm/smp_plat.h>
+
+#include <soc/tegra/bpmp.h>
+#include <soc/tegra/bpmp-abi.h>
+
+#define KHZ                     1000
+#define REF_CLK_MHZ             408 /* 408 MHz */
+#define US_DELAY                500
+#define US_DELAY_MIN            2
+#define CPUFREQ_TBL_STEP_HZ     (50 * KHZ * KHZ)
+#define MAX_CNT                 ~0U
+
+/* cpufreq transisition latency */
+#define TEGRA_CPUFREQ_TRANSITION_LATENCY (300 * 1000) /* unit in nanoseconds */
+
+#define LOOP_FOR_EACH_CPU_OF_CLUSTER(cl) for (cpu = (cl * 2); \
+					cpu < ((cl + 1) * 2); cpu++)
+
+enum cluster {
+	CLUSTER0,
+	CLUSTER1,
+	CLUSTER2,
+	CLUSTER3,
+	MAX_CLUSTERS,
+};
+
+struct tegra194_cpufreq_data {
+	void __iomem *regs;
+	size_t num_clusters;
+	struct cpufreq_frequency_table **tables;
+};
+
+struct tegra_cpu_ctr {
+	u32 cpu;
+	u32 delay;
+	u32 coreclk_cnt, last_coreclk_cnt;
+	u32 refclk_cnt, last_refclk_cnt;
+};
+
+struct read_counters_work {
+	struct work_struct work;
+	struct tegra_cpu_ctr c;
+};
+
+static struct workqueue_struct *read_counters_wq;
+
+
+static enum cluster get_cpu_cluster(u8 cpu)
+{
+	return MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu), 1);
+}
+
+/*
+ * Read per-core Read-only system register NVFREQ_FEEDBACK_EL1.
+ * The register provides frequency feedback information to
+ * determine the average actual frequency a core has run at over
+ * a period of time.
+ *	[31:0] PLLP counter: Counts at fixed frequency (408 MHz)
+ *	[63:32] Core clock counter: counts on every core clock cycle
+ *			where the core is architecturally clocking
+ */
+static u64 read_freq_feedback(void)
+{
+	u64 val = 0;
+
+	asm volatile("mrs %0, s3_0_c15_c0_5" : "=r" (val) : );
+
+	return val;
+}
+
+u16 map_freq_to_ndiv(struct mrq_cpu_ndiv_limits_response *nltbl, u32 freq)
+{
+	return DIV_ROUND_UP(freq * nltbl->pdiv * nltbl->mdiv,
+			    nltbl->ref_clk_hz / KHZ);
+}
+
+static inline u32 map_ndiv_to_freq(struct mrq_cpu_ndiv_limits_response
+				   *nltbl, u16 ndiv)
+{
+	return nltbl->ref_clk_hz / KHZ * ndiv / (nltbl->pdiv * nltbl->mdiv);
+}
+
+static void tegra_read_counters(struct work_struct *work)
+{
+	struct read_counters_work *read_counters_work;
+	struct tegra_cpu_ctr *c;
+	u64 val;
+
+	/*
+	 * ref_clk_counter(32 bit counter) runs on constant clk,
+	 * pll_p(408MHz).
+	 * It will take = 2 ^ 32 / 408 MHz to overflow ref clk counter
+	 *              = 10526880 usec = 10.527 sec to overflow
+	 *
+	 * Like wise core_clk_counter(32 bit counter) runs on core clock.
+	 * It's synchronized to crab_clk (cpu_crab_clk) which runs at
+	 * freq of cluster. Assuming max cluster clock ~2000MHz,
+	 * It will take = 2 ^ 32 / 2000 MHz to overflow core clk counter
+	 *              = ~2.147 sec to overflow
+	 */
+	read_counters_work = container_of(work, struct read_counters_work,
+					  work);
+	c = &read_counters_work->c;
+
+	val = read_freq_feedback();
+	c->last_refclk_cnt = lower_32_bits(val);
+	c->last_coreclk_cnt = upper_32_bits(val);
+	udelay(c->delay);
+	val = read_freq_feedback();
+	c->refclk_cnt = lower_32_bits(val);
+	c->coreclk_cnt = upper_32_bits(val);
+}
+
+/*
+ * Return instantaneous cpu speed
+ * Instantaneous freq is calculated as -
+ * -Takes sample on every query of getting the freq.
+ *	- Read core and ref clock counters;
+ *	- Delay for X us
+ *	- Read above cycle counters again
+ *	- Calculates freq by subtracting current and previous counters
+ *	  divided by the delay time or eqv. of ref_clk_counter in delta time
+ *	- Return Kcycles/second, freq in KHz
+ *
+ *	delta time period = x sec
+ *			  = delta ref_clk_counter / (408 * 10^6) sec
+ *	freq in Hz = cycles/sec
+ *		   = (delta cycles / x sec
+ *		   = (delta cycles * 408 * 10^6) / delta ref_clk_counter
+ *	in KHz	   = (delta cycles * 408 * 10^3) / delta ref_clk_counter
+ *
+ * @cpu - logical cpu whose freq to be updated
+ * Returns freq in KHz on success, 0 if cpu is offline
+ */
+static unsigned int tegra194_get_speed_common(u32 cpu, u32 delay)
+{
+	struct read_counters_work read_counters_work;
+	struct tegra_cpu_ctr c;
+	u32 delta_refcnt;
+	u32 delta_ccnt;
+	u32 rate_mhz;
+
+	/*
+	 * udelay() is required to reconstruct cpu frequency over an
+	 * observation window. Using workqueue to call udelay() with
+	 * interrupts enabled.
+	 */
+	read_counters_work.c.cpu = cpu;
+	read_counters_work.c.delay = delay;
+	INIT_WORK_ONSTACK(&read_counters_work.work, tegra_read_counters);
+	queue_work_on(cpu, read_counters_wq, &read_counters_work.work);
+	flush_work(&read_counters_work.work);
+	c = read_counters_work.c;
+
+	if (c.coreclk_cnt < c.last_coreclk_cnt)
+		delta_ccnt = c.coreclk_cnt + (MAX_CNT - c.last_coreclk_cnt);
+	else
+		delta_ccnt = c.coreclk_cnt - c.last_coreclk_cnt;
+	if (!delta_ccnt)
+		return 0;
+
+	/* ref clock is 32 bits */
+	if (c.refclk_cnt < c.last_refclk_cnt)
+		delta_refcnt = c.refclk_cnt + (MAX_CNT - c.last_refclk_cnt);
+	else
+		delta_refcnt = c.refclk_cnt - c.last_refclk_cnt;
+	if (!delta_refcnt) {
+		pr_debug("cpufreq: %d is idle, delta_refcnt: 0\n", cpu);
+		return 0;
+	}
+	rate_mhz = ((unsigned long)(delta_ccnt * REF_CLK_MHZ)) / delta_refcnt;
+
+	return (rate_mhz * KHZ); /* in KHz */
+}
+
+static unsigned int tegra194_get_speed(u32 cpu)
+{
+	return tegra194_get_speed_common(cpu, US_DELAY);
+}
+
+static unsigned int tegra194_fast_get_speed(u32 cpu)
+{
+	return tegra194_get_speed_common(cpu, US_DELAY_MIN);
+}
+
+static int tegra194_cpufreq_init(struct cpufreq_policy *policy)
+{
+	struct tegra194_cpufreq_data *data = cpufreq_get_driver_data();
+	int cl = get_cpu_cluster(policy->cpu);
+	u32 cpu;
+
+	if (cl >= data->num_clusters)
+		return -EINVAL;
+
+	policy->cur = tegra194_fast_get_speed(policy->cpu); /* boot freq */
+
+	/* set same policy for all cpus in a cluster */
+	LOOP_FOR_EACH_CPU_OF_CLUSTER(cl)
+		cpumask_set_cpu(cpu, policy->cpus);
+
+	policy->freq_table = data->tables[cl];
+	policy->cpuinfo.transition_latency = TEGRA_CPUFREQ_TRANSITION_LATENCY;
+
+	return 0;
+}
+
+static void set_cpu_ndiv(void *data)
+{
+	struct cpufreq_frequency_table *tbl = data;
+	u64 ndiv_val = (u64)tbl->driver_data;
+
+	asm volatile("msr s3_0_c15_c0_4, %0" : : "r" (ndiv_val));
+}
+
+static int tegra194_cpufreq_set_target(struct cpufreq_policy *policy,
+				       unsigned int index)
+{
+	struct cpufreq_frequency_table *tbl = policy->freq_table + index;
+
+	on_each_cpu_mask(policy->cpus, set_cpu_ndiv, tbl, true);
+
+	return 0;
+}
+
+static struct cpufreq_driver tegra194_cpufreq_driver = {
+	.name = "tegra194",
+	.flags = CPUFREQ_STICKY | CPUFREQ_CONST_LOOPS |
+		CPUFREQ_NEED_INITIAL_FREQ_CHECK,
+	.verify = cpufreq_generic_frequency_table_verify,
+	.target_index = tegra194_cpufreq_set_target,
+	.get = tegra194_get_speed,
+	.init = tegra194_cpufreq_init,
+	.attr = cpufreq_generic_attr,
+};
+
+static void tegra194_cpufreq_free_resources(void)
+{
+	flush_workqueue(read_counters_wq);
+	destroy_workqueue(read_counters_wq);
+}
+
+static struct cpufreq_frequency_table *
+init_freq_table(struct platform_device *pdev, struct tegra_bpmp *bpmp,
+		unsigned int cluster_id)
+{
+	struct cpufreq_frequency_table *freq_table;
+	struct mrq_cpu_ndiv_limits_response resp;
+	unsigned int num_freqs, ndiv, delta_ndiv;
+	struct mrq_cpu_ndiv_limits_request req;
+	struct tegra_bpmp_message msg;
+	u16 freq_table_step_size;
+	int err, index;
+
+	memset(&req, 0, sizeof(req));
+	req.cluster_id = cluster_id;
+
+	memset(&msg, 0, sizeof(msg));
+	msg.mrq = MRQ_CPU_NDIV_LIMITS;
+	msg.tx.data = &req;
+	msg.tx.size = sizeof(req);
+	msg.rx.data = &resp;
+	msg.rx.size = sizeof(resp);
+
+	err = tegra_bpmp_transfer(bpmp, &msg);
+	if (err)
+		return ERR_PTR(err);
+
+	/*
+	 * Make sure frequency table step is a multiple of mdiv to match
+	 * vhint table granularity.
+	 */
+	freq_table_step_size = resp.mdiv *
+			DIV_ROUND_UP(CPUFREQ_TBL_STEP_HZ, resp.ref_clk_hz);
+
+	dev_dbg(&pdev->dev, "cluster %d: frequency table step size: %d\n",
+		cluster_id, freq_table_step_size);
+
+	delta_ndiv = resp.ndiv_max - resp.ndiv_min;
+
+	if (unlikely(delta_ndiv == 0))
+		num_freqs = 1;
+	else
+		/* We store both ndiv_min and ndiv_max hence the +1 */
+		num_freqs = delta_ndiv / freq_table_step_size + 1;
+
+	num_freqs += (delta_ndiv % freq_table_step_size) ? 1 : 0;
+
+	freq_table = devm_kcalloc(&pdev->dev, num_freqs + 1,
+				  sizeof(*freq_table), GFP_KERNEL);
+	if (!freq_table)
+		return ERR_PTR(-ENOMEM);
+
+	for (index = 0, ndiv = resp.ndiv_min;
+			ndiv < resp.ndiv_max;
+			index++, ndiv += freq_table_step_size) {
+		freq_table[index].driver_data = ndiv;
+		freq_table[index].frequency = map_ndiv_to_freq(&resp, ndiv);
+	}
+
+	freq_table[index].driver_data = resp.ndiv_max;
+	freq_table[index++].frequency = map_ndiv_to_freq(&resp, resp.ndiv_max);
+	freq_table[index].frequency = CPUFREQ_TABLE_END;
+
+	return freq_table;
+}
+
+static int tegra194_cpufreq_probe(struct platform_device *pdev)
+{
+	struct tegra194_cpufreq_data *data;
+	struct tegra_bpmp *bpmp;
+	int err, i;
+
+	data = devm_kzalloc(&pdev->dev, sizeof(*data), GFP_KERNEL);
+	if (!data)
+		return -ENOMEM;
+
+	data->num_clusters = MAX_CLUSTERS;
+	data->tables = devm_kcalloc(&pdev->dev, data->num_clusters,
+				    sizeof(*data->tables), GFP_KERNEL);
+	if (!data->tables)
+		return -ENOMEM;
+
+	platform_set_drvdata(pdev, data);
+
+	bpmp = tegra_bpmp_get(&pdev->dev);
+	if (IS_ERR(bpmp))
+		return PTR_ERR(bpmp);
+
+	read_counters_wq = alloc_workqueue("read_counters_wq", __WQ_LEGACY, 1);
+	if (!read_counters_wq) {
+		dev_err(&pdev->dev, "fail to create_workqueue\n");
+		err = -EINVAL;
+		goto put_bpmp;
+	}
+
+	for (i = 0; i < data->num_clusters; i++) {
+		data->tables[i] = init_freq_table(pdev, bpmp, i);
+		if (IS_ERR(data->tables[i])) {
+			err = PTR_ERR(data->tables[i]);
+			goto put_bpmp;
+		}
+	}
+
+	tegra_bpmp_put(bpmp);
+
+	tegra194_cpufreq_driver.driver_data = data;
+
+	err = cpufreq_register_driver(&tegra194_cpufreq_driver);
+	if (err)
+		goto err_free_res;
+
+	return err;
+
+put_bpmp:
+	tegra_bpmp_put(bpmp);
+err_free_res:
+	tegra194_cpufreq_free_resources();
+	return err;
+}
+
+static int tegra194_cpufreq_remove(struct platform_device *pdev)
+{
+	cpufreq_unregister_driver(&tegra194_cpufreq_driver);
+	tegra194_cpufreq_free_resources();
+
+	return 0;
+}
+
+static const struct of_device_id tegra194_cpufreq_of_match[] = {
+	{ .compatible = "nvidia,tegra194-ccplex", },
+	{ /* sentinel */ }
+};
+MODULE_DEVICE_TABLE(of, tegra194_cpufreq_of_match);
+
+static struct platform_driver tegra194_ccplex_driver = {
+	.driver = {
+		.name = "tegra194-cpufreq",
+		.of_match_table = tegra194_cpufreq_of_match,
+	},
+	.probe = tegra194_cpufreq_probe,
+	.remove = tegra194_cpufreq_remove,
+};
+module_platform_driver(tegra194_ccplex_driver);
+
+MODULE_AUTHOR("Mikko Perttunen <mperttunen@nvidia.com>");
+MODULE_AUTHOR("Sumit Gupta <sumitg@nvidia.com>");
+MODULE_DESCRIPTION("NVIDIA Tegra194 cpufreq driver");
+MODULE_LICENSE("GPL v2");
-- 
2.7.4

[TEGRA194_CPUFREQ Patch v3 4/4] arm64: defconfig: Enable CONFIG_ARM_TEGRA194_CPUFREQ

[Sumit Gupta]

Enable Tegra194 CPU frequency scaling support by default.

Signed-off-by: Sumit Gupta <sumitg@nvidia.com>
---
 arch/arm64/configs/defconfig | 1 +
 1 file changed, 1 insertion(+)

diff --git a/arch/arm64/configs/defconfig b/arch/arm64/configs/defconfig
index f9d378d..385bd35 100644
--- a/arch/arm64/configs/defconfig
+++ b/arch/arm64/configs/defconfig
@@ -91,6 +91,7 @@ CONFIG_ARM_QCOM_CPUFREQ_NVMEM=y
 CONFIG_ARM_QCOM_CPUFREQ_HW=y
 CONFIG_ARM_RASPBERRYPI_CPUFREQ=m
 CONFIG_ARM_TEGRA186_CPUFREQ=y
+CONFIG_ARM_TEGRA194_CPUFREQ=y
 CONFIG_QORIQ_CPUFREQ=y
 CONFIG_ARM_SCPI_PROTOCOL=y
 CONFIG_RASPBERRYPI_FIRMWARE=y
-- 
2.7.4

Re: [TEGRA194_CPUFREQ Patch v3 3/4] cpufreq: Add Tegra194 cpufreq driver

[Sumit Gupta]

Hi,

Thank you for the report.

> 
> Hi Sumit,
> 
> Thank you for the patch! Perhaps something to improve:
> 
> [auto build test WARNING on pm/linux-next]
> [also build test WARNING on robh/for-next arm64/for-next/core v5.8-rc2 next-20200621]
> [If your patch is applied to the wrong git tree, kindly drop us a note.
> And when submitting patch, we suggest to use  as documented in
> https://git-scm.com/docs/git-format-patch]
> 
> url:    https://github.com/0day-ci/linux/commits/Sumit-Gupta/Add-cpufreq-driver-for-Tegra194/20200622-053622
> base:   https://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm.git linux-next
> config: arm64-allyesconfig (attached as .config)
> compiler: aarch64-linux-gcc (GCC) 9.3.0
> reproduce (this is a W=1 build):
>          wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross
>          chmod +x ~/bin/make.cross
>          # save the attached .config to linux build tree
>          COMPILER_INSTALL_PATH=$HOME/0day COMPILER=gcc-9.3.0 make.cross ARCH=arm64
> 
> If you fix the issue, kindly add following tag as appropriate
> Reported-by: kernel test robot <lkp@intel.com>

Sure.
> 
> All warnings (new ones prefixed by >>, old ones prefixed by <<):
> 
>>> drivers/cpufreq/tegra194-cpufreq.c:86:5: warning: no previous prototype for 'map_freq_to_ndiv' [-Wmissing-prototypes]
> 86 | u16 map_freq_to_ndiv(struct mrq_cpu_ndiv_limits_response *nltbl, u32 freq)
> |     ^~~~~~~~~~~~~~~~
> 
> vim +/map_freq_to_ndiv +86 drivers/cpufreq/tegra194-cpufreq.c
> 
>      85
>    > 86  u16 map_freq_to_ndiv(struct mrq_cpu_ndiv_limits_response *nltbl, u32 freq)
>      87  {
>      88          return DIV_ROUND_UP(freq * nltbl->pdiv * nltbl->mdiv,
>      89                              nltbl->ref_clk_hz / KHZ);
>      90  }
>      91
> 

will take care of this in next version.
> ---
> 0-DAY CI Kernel Test Service, Intel Corporation
> https://lists.01.org/hyperkitty/list/kbuild-all@lists.01.org
> 

Re: [TEGRA194_CPUFREQ Patch v3 3/4] cpufreq: Add Tegra194 cpufreq driver

[Viresh Kumar]

On 22-06-20, 03:04, Sumit Gupta wrote:
> diff --git a/drivers/cpufreq/tegra194-cpufreq.c b/drivers/cpufreq/tegra194-cpufreq.c
> new file mode 100644
> index 0000000..8de8000
> --- /dev/null
> +++ b/drivers/cpufreq/tegra194-cpufreq.c
> @@ -0,0 +1,403 @@
> +// SPDX-License-Identifier: GPL-2.0
> +/*
> + * Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved

                    2020

> + */
> +
> +#include <linux/cpu.h>
> +#include <linux/cpufreq.h>
> +#include <linux/delay.h>
> +#include <linux/dma-mapping.h>
> +#include <linux/module.h>
> +#include <linux/of.h>
> +#include <linux/of_platform.h>
> +#include <linux/platform_device.h>
> +#include <linux/slab.h>
> +
> +#include <asm/smp_plat.h>
> +
> +#include <soc/tegra/bpmp.h>
> +#include <soc/tegra/bpmp-abi.h>
> +
> +#define KHZ                     1000
> +#define REF_CLK_MHZ             408 /* 408 MHz */
> +#define US_DELAY                500
> +#define US_DELAY_MIN            2
> +#define CPUFREQ_TBL_STEP_HZ     (50 * KHZ * KHZ)
> +#define MAX_CNT                 ~0U
> +
> +/* cpufreq transisition latency */
> +#define TEGRA_CPUFREQ_TRANSITION_LATENCY (300 * 1000) /* unit in nanoseconds */
> +
> +#define LOOP_FOR_EACH_CPU_OF_CLUSTER(cl) for (cpu = (cl * 2); \
> +					cpu < ((cl + 1) * 2); cpu++)

Both latency and this loop are used only once in the code, maybe just open code
it. Also you should have passed cpu as a parameter to the macro, even if it
works fine without it, for better readability.

> +
> +u16 map_freq_to_ndiv(struct mrq_cpu_ndiv_limits_response *nltbl, u32 freq)

Unused routine

> +{
> +	return DIV_ROUND_UP(freq * nltbl->pdiv * nltbl->mdiv,
> +			    nltbl->ref_clk_hz / KHZ);
> +}

> +static int tegra194_cpufreq_init(struct cpufreq_policy *policy)
> +{
> +	struct tegra194_cpufreq_data *data = cpufreq_get_driver_data();
> +	int cl = get_cpu_cluster(policy->cpu);
> +	u32 cpu;
> +
> +	if (cl >= data->num_clusters)
> +		return -EINVAL;
> +
> +	policy->cur = tegra194_fast_get_speed(policy->cpu); /* boot freq */
> +
> +	/* set same policy for all cpus in a cluster */
> +	LOOP_FOR_EACH_CPU_OF_CLUSTER(cl)
> +		cpumask_set_cpu(cpu, policy->cpus);
> +
> +	policy->freq_table = data->tables[cl];
> +	policy->cpuinfo.transition_latency = TEGRA_CPUFREQ_TRANSITION_LATENCY;
> +
> +	return 0;
> +}

> +static int tegra194_cpufreq_set_target(struct cpufreq_policy *policy,
> +				       unsigned int index)
> +{
> +	struct cpufreq_frequency_table *tbl = policy->freq_table + index;
> +
> +	on_each_cpu_mask(policy->cpus, set_cpu_ndiv, tbl, true);

I am still a bit confused. While setting the frequency you are calling this
routine for each CPU of the policy (cluster). Does that mean that CPUs within a
cluster can actually run at different frequencies at any given point of time ?

If cpufreq terms, a cpufreq policy represents a group of CPUs that change
frequency together, i.e. they share the clk line. If all CPUs in your system can
do DVFS separately, then you must have policy per CPU, instead of cluster.

> +static void tegra194_cpufreq_free_resources(void)
> +{
> +	flush_workqueue(read_counters_wq);

Why is this required exactly? I see that you add the work request and
immediately flush it, then why would you need to do this separately ?

> +	destroy_workqueue(read_counters_wq);
> +}
> +
> +static struct cpufreq_frequency_table *
> +init_freq_table(struct platform_device *pdev, struct tegra_bpmp *bpmp,
> +		unsigned int cluster_id)
> +{
> +	struct cpufreq_frequency_table *freq_table;
> +	struct mrq_cpu_ndiv_limits_response resp;
> +	unsigned int num_freqs, ndiv, delta_ndiv;
> +	struct mrq_cpu_ndiv_limits_request req;
> +	struct tegra_bpmp_message msg;
> +	u16 freq_table_step_size;
> +	int err, index;
> +
> +	memset(&req, 0, sizeof(req));
> +	req.cluster_id = cluster_id;
> +
> +	memset(&msg, 0, sizeof(msg));
> +	msg.mrq = MRQ_CPU_NDIV_LIMITS;
> +	msg.tx.data = &req;
> +	msg.tx.size = sizeof(req);
> +	msg.rx.data = &resp;
> +	msg.rx.size = sizeof(resp);
> +
> +	err = tegra_bpmp_transfer(bpmp, &msg);

So the firmware can actually return different frequency tables for the clusters,
right ? Else you could have received the table only once and used it for all the
CPUs.

> +	if (err)
> +		return ERR_PTR(err);
> +
> +	/*
> +	 * Make sure frequency table step is a multiple of mdiv to match
> +	 * vhint table granularity.
> +	 */
> +	freq_table_step_size = resp.mdiv *
> +			DIV_ROUND_UP(CPUFREQ_TBL_STEP_HZ, resp.ref_clk_hz);
> +
> +	dev_dbg(&pdev->dev, "cluster %d: frequency table step size: %d\n",
> +		cluster_id, freq_table_step_size);
> +
> +	delta_ndiv = resp.ndiv_max - resp.ndiv_min;
> +
> +	if (unlikely(delta_ndiv == 0))
> +		num_freqs = 1;
> +	else
> +		/* We store both ndiv_min and ndiv_max hence the +1 */
> +		num_freqs = delta_ndiv / freq_table_step_size + 1;
> +
> +	num_freqs += (delta_ndiv % freq_table_step_size) ? 1 : 0;
> +
> +	freq_table = devm_kcalloc(&pdev->dev, num_freqs + 1,
> +				  sizeof(*freq_table), GFP_KERNEL);
> +	if (!freq_table)
> +		return ERR_PTR(-ENOMEM);
> +
> +	for (index = 0, ndiv = resp.ndiv_min;
> +			ndiv < resp.ndiv_max;
> +			index++, ndiv += freq_table_step_size) {
> +		freq_table[index].driver_data = ndiv;
> +		freq_table[index].frequency = map_ndiv_to_freq(&resp, ndiv);
> +	}
> +
> +	freq_table[index].driver_data = resp.ndiv_max;
> +	freq_table[index++].frequency = map_ndiv_to_freq(&resp, resp.ndiv_max);
> +	freq_table[index].frequency = CPUFREQ_TABLE_END;
> +
> +	return freq_table;
> +}

-- 
viresh

Re: [TEGRA194_CPUFREQ Patch v3 1/4] dt-bindings: arm: Add t194 ccplex compatible and bpmp property

[Viresh Kumar]

On 22-06-20, 03:04, Sumit Gupta wrote:
> To do frequency scaling on all CPUs within T194 CPU Complex, we need
> to query BPMP for data on valid operating points. Document a compatible
> string under 'cpus' node to represent the CPU Complex for binding drivers
> like cpufreq which don't have their node or CPU Complex node to bind to.
> Also, document a property to point to the BPMP device that can be queried
> for all CPUs.

You shouldn't be putting how linux is going to use this information and entries
shouldn't be made just so cpufreq can bind to a driver.

Though I see that this is a real hardware register which you can use to interact
with the firmware ? And so it makes sense to have it, maybe in different form
though.

I will let Rob explain what would be the right way of doing this though.

> 
> Signed-off-by: Sumit Gupta <sumitg@nvidia.com>
> ---
>  Documentation/devicetree/bindings/arm/cpus.yaml | 9 +++++++++
>  1 file changed, 9 insertions(+)
> 
> diff --git a/Documentation/devicetree/bindings/arm/cpus.yaml b/Documentation/devicetree/bindings/arm/cpus.yaml
> index a018147..737b55e 100644
> --- a/Documentation/devicetree/bindings/arm/cpus.yaml
> +++ b/Documentation/devicetree/bindings/arm/cpus.yaml
> @@ -162,6 +162,7 @@ properties:
>        - nvidia,tegra132-denver
>        - nvidia,tegra186-denver
>        - nvidia,tegra194-carmel
> +      - nvidia,tegra194-ccplex
>        - qcom,krait
>        - qcom,kryo
>        - qcom,kryo260
> @@ -255,6 +256,14 @@ properties:
>  
>        where voltage is in V, frequency is in MHz.
>  
> +  nvidia,bpmp:
> +    $ref: '/schemas/types.yaml#/definitions/phandle'
> +    descrption: |
> +      Specifies the bpmp node that needs to be queried to get
> +      operating point data for all CPUs.
> +
> +      Optional for NVIDIA Tegra194 Carmel CPUs
> +
>    power-domains:
>      $ref: '/schemas/types.yaml#/definitions/phandle-array'
>      description:
> -- 
> 2.7.4

-- 
viresh

Re: [TEGRA194_CPUFREQ Patch v3 3/4] cpufreq: Add Tegra194 cpufreq driver

[Sumit Gupta]

Hi Viresh,

Thank you for the review. please find my reply inline.


>> +++ b/drivers/cpufreq/tegra194-cpufreq.c
>> @@ -0,0 +1,403 @@
>> +// SPDX-License-Identifier: GPL-2.0
>> +/*
>> + * Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved
> 
>                      2020
> 
>> + */
>> +
>> +#include <linux/cpu.h>
>> +#include <linux/cpufreq.h>
>> +#include <linux/delay.h>
>> +#include <linux/dma-mapping.h>
>> +#include <linux/module.h>
>> +#include <linux/of.h>
>> +#include <linux/of_platform.h>
>> +#include <linux/platform_device.h>
>> +#include <linux/slab.h>
>> +
>> +#include <asm/smp_plat.h>
>> +
>> +#include <soc/tegra/bpmp.h>
>> +#include <soc/tegra/bpmp-abi.h>
>> +
>> +#define KHZ                     1000
>> +#define REF_CLK_MHZ             408 /* 408 MHz */
>> +#define US_DELAY                500
>> +#define US_DELAY_MIN            2
>> +#define CPUFREQ_TBL_STEP_HZ     (50 * KHZ * KHZ)
>> +#define MAX_CNT                 ~0U
>> +
>> +/* cpufreq transisition latency */
>> +#define TEGRA_CPUFREQ_TRANSITION_LATENCY (300 * 1000) /* unit in nanoseconds */
>> +
>> +#define LOOP_FOR_EACH_CPU_OF_CLUSTER(cl) for (cpu = (cl * 2); \
>> +                                     cpu < ((cl + 1) * 2); cpu++)
> 
> Both latency and this loop are used only once in the code, maybe just open code
> it. Also you should have passed cpu as a parameter to the macro, even if it
> works fine without it, for better readability.
> 
Ok, i will open code the loop in next version. For latency value, i feel 
named macro makes readability better. So, prefer keeping it.

>> +
>> +u16 map_freq_to_ndiv(struct mrq_cpu_ndiv_limits_response *nltbl, u32 freq)
> 
> Unused routine
> 
Sure, will remove it.

>> +{
>> +     return DIV_ROUND_UP(freq * nltbl->pdiv * nltbl->mdiv,
>> +                         nltbl->ref_clk_hz / KHZ);
>> +}
> 
>> +static int tegra194_cpufreq_init(struct cpufreq_policy *policy)
>> +{
>> +     struct tegra194_cpufreq_data *data = cpufreq_get_driver_data();
>> +     int cl = get_cpu_cluster(policy->cpu);
>> +     u32 cpu;
>> +
>> +     if (cl >= data->num_clusters)
>> +             return -EINVAL;
>> +
>> +     policy->cur = tegra194_fast_get_speed(policy->cpu); /* boot freq */
>> +
>> +     /* set same policy for all cpus in a cluster */
>> +     LOOP_FOR_EACH_CPU_OF_CLUSTER(cl)
>> +             cpumask_set_cpu(cpu, policy->cpus);
>> +
>> +     policy->freq_table = data->tables[cl];
>> +     policy->cpuinfo.transition_latency = TEGRA_CPUFREQ_TRANSITION_LATENCY;
>> +
>> +     return 0;
>> +}
> 
>> +static int tegra194_cpufreq_set_target(struct cpufreq_policy *policy,
>> +                                    unsigned int index)
>> +{
>> +     struct cpufreq_frequency_table *tbl = policy->freq_table + index;
>> +
>> +     on_each_cpu_mask(policy->cpus, set_cpu_ndiv, tbl, true);
> 
> I am still a bit confused. While setting the frequency you are calling this
> routine for each CPU of the policy (cluster). Does that mean that CPUs within a
> cluster can actually run at different frequencies at any given point of time ?
> 
> If cpufreq terms, a cpufreq policy represents a group of CPUs that change
> frequency together, i.e. they share the clk line. If all CPUs in your system can
> do DVFS separately, then you must have policy per CPU, instead of cluster.
> 
T194 supports four CPU clusters, each with two cores. Each CPU cluster 
is capable of running at a specific frequency sourced by respective 
NAFLL to provide cluster specific clocks. Individual cores within a 
cluster write freq in per core register. Cluster h/w forwards the 
max(core0, core1) request to per cluster NAFLL.

>> +static void tegra194_cpufreq_free_resources(void)
>> +{
>> +     flush_workqueue(read_counters_wq);
> 
> Why is this required exactly? I see that you add the work request and
> immediately flush it, then why would you need to do this separately ?
> 
Ya, will remove flush_workqueue().

>> +     destroy_workqueue(read_counters_wq);
>> +}
>> +
>> +static struct cpufreq_frequency_table *
>> +init_freq_table(struct platform_device *pdev, struct tegra_bpmp *bpmp,
>> +             unsigned int cluster_id)
>> +{
>> +     struct cpufreq_frequency_table *freq_table;
>> +     struct mrq_cpu_ndiv_limits_response resp;
>> +     unsigned int num_freqs, ndiv, delta_ndiv;
>> +     struct mrq_cpu_ndiv_limits_request req;
>> +     struct tegra_bpmp_message msg;
>> +     u16 freq_table_step_size;
>> +     int err, index;
>> +
>> +     memset(&req, 0, sizeof(req));
>> +     req.cluster_id = cluster_id;
>> +
>> +     memset(&msg, 0, sizeof(msg));
>> +     msg.mrq = MRQ_CPU_NDIV_LIMITS;
>> +     msg.tx.data = &req;
>> +     msg.tx.size = sizeof(req);
>> +     msg.rx.data = &resp;
>> +     msg.rx.size = sizeof(resp);
>> +
>> +     err = tegra_bpmp_transfer(bpmp, &msg);
> 
> So the firmware can actually return different frequency tables for the clusters,
> right ? Else you could have received the table only once and used it for all the
> CPUs.
> 
Yes, frequency tables are returned per cluster by BPMP firmware. In T194 
SOC, currently same table values are used for all clusters. This might 
change in future.

>> +     if (err)
>> +             return ERR_PTR(err);
>> +
>> +     /*
>> +      * Make sure frequency table step is a multiple of mdiv to match
>> +      * vhint table granularity.
>> +      */
>> +     freq_table_step_size = resp.mdiv *
>> +                     DIV_ROUND_UP(CPUFREQ_TBL_STEP_HZ, resp.ref_clk_hz);
>> +
>> +     dev_dbg(&pdev->dev, "cluster %d: frequency table step size: %d\n",
>> +             cluster_id, freq_table_step_size);
>> +
>> +     delta_ndiv = resp.ndiv_max - resp.ndiv_min;
>> +
>> +     if (unlikely(delta_ndiv == 0))
>> +             num_freqs = 1;
>> +     else
>> +             /* We store both ndiv_min and ndiv_max hence the +1 */
>> +             num_freqs = delta_ndiv / freq_table_step_size + 1;
>> +
>> +     num_freqs += (delta_ndiv % freq_table_step_size) ? 1 : 0;
>> +
>> +     freq_table = devm_kcalloc(&pdev->dev, num_freqs + 1,
>> +                               sizeof(*freq_table), GFP_KERNEL);
>> +     if (!freq_table)
>> +             return ERR_PTR(-ENOMEM);
>> +
>> +     for (index = 0, ndiv = resp.ndiv_min;
>> +                     ndiv < resp.ndiv_max;
>> +                     index++, ndiv += freq_table_step_size) {
>> +             freq_table[index].driver_data = ndiv;
>> +             freq_table[index].frequency = map_ndiv_to_freq(&resp, ndiv);
>> +     }
>> +
>> +     freq_table[index].driver_data = resp.ndiv_max;
>> +     freq_table[index++].frequency = map_ndiv_to_freq(&resp, resp.ndiv_max);
>> +     freq_table[index].frequency = CPUFREQ_TABLE_END;
>> +
>> +     return freq_table;
>> +}
> 
> --
> viresh
> 

Re: [TEGRA194_CPUFREQ Patch v3 1/4] dt-bindings: arm: Add t194 ccplex compatible and bpmp property

[Sumit Gupta]

On 22/06/20 12:52 PM, Viresh Kumar wrote:
> External email: Use caution opening links or attachments
> 
> 
> On 22-06-20, 03:04, Sumit Gupta wrote:
>> To do frequency scaling on all CPUs within T194 CPU Complex, we need
>> to query BPMP for data on valid operating points. Document a compatible
>> string under 'cpus' node to represent the CPU Complex for binding drivers
>> like cpufreq which don't have their node or CPU Complex node to bind to.
>> Also, document a property to point to the BPMP device that can be queried
>> for all CPUs.
> 
> You shouldn't be putting how linux is going to use this information and entries
> shouldn't be made just so cpufreq can bind to a driver.
> 
> Though I see that this is a real hardware register which you can use to interact
> with the firmware ? And so it makes sense to have it, maybe in different form
> though.
> 
CPUFREQ driver doesn't communicate directly with BPMP firmware. It uses 
BPMP node's reference to call api's exported by BPMP driver which 
communicates with BPMP firmware.

> I will let Rob explain what would be the right way of doing this though.
> 
This is already discussed by Thierry with Rob.
Please refer https://marc.info/?l=linux-arm-kernel&m=158999171528418&w=2

>>
>> Signed-off-by: Sumit Gupta <sumitg@nvidia.com>
>> ---
>>   Documentation/devicetree/bindings/arm/cpus.yaml | 9 +++++++++
>>   1 file changed, 9 insertions(+)
>>
>> diff --git a/Documentation/devicetree/bindings/arm/cpus.yaml b/Documentation/devicetree/bindings/arm/cpus.yaml
>> index a018147..737b55e 100644
>> --- a/Documentation/devicetree/bindings/arm/cpus.yaml
>> +++ b/Documentation/devicetree/bindings/arm/cpus.yaml
>> @@ -162,6 +162,7 @@ properties:
>>         - nvidia,tegra132-denver
>>         - nvidia,tegra186-denver
>>         - nvidia,tegra194-carmel
>> +      - nvidia,tegra194-ccplex
>>         - qcom,krait
>>         - qcom,kryo
>>         - qcom,kryo260
>> @@ -255,6 +256,14 @@ properties:
>>
>>         where voltage is in V, frequency is in MHz.
>>
>> +  nvidia,bpmp:
>> +    $ref: '/schemas/types.yaml#/definitions/phandle'
>> +    descrption: |
>> +      Specifies the bpmp node that needs to be queried to get
>> +      operating point data for all CPUs.
>> +
>> +      Optional for NVIDIA Tegra194 Carmel CPUs
>> +
>>     power-domains:
>>       $ref: '/schemas/types.yaml#/definitions/phandle-array'
>>       description:
>> --
>> 2.7.4
> 
> --
> viresh
> 

Re: [TEGRA194_CPUFREQ Patch v3 3/4] cpufreq: Add Tegra194 cpufreq driver

[Viresh Kumar]

On 23-06-20, 10:49, Sumit Gupta wrote:
> Hi Viresh,
> 
> Thank you for the review. please find my reply inline.
> 
> 
> > > +++ b/drivers/cpufreq/tegra194-cpufreq.c
> > > @@ -0,0 +1,403 @@
> > > +// SPDX-License-Identifier: GPL-2.0
> > > +/*
> > > + * Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved
> > 
> >                      2020

You missed this ?

> T194 supports four CPU clusters, each with two cores. Each CPU cluster is
> capable of running at a specific frequency sourced by respective NAFLL to
> provide cluster specific clocks. Individual cores within a cluster write
> freq in per core register. Cluster h/w forwards the max(core0, core1)
> request to per cluster NAFLL.

Okay, this is clear now. Add a comment about this max thing in the
target routine to show why you need to do this on all CPUs.

-- 
viresh