Re: [PATCH] cpufreq: brcmstb-cpufreq: CPUfreq driver for older Broadcom STB SoCs

From: Markus Mayer
Date: Thu Nov 17 2016 - 13:39:53 EST

On 17 November 2016 at 01:02, Viresh Kumar <viresh.kumar@xxxxxxxxxx> wrote:
> On 10-11-16, 15:56, Markus Mayer wrote:
>> From: Markus Mayer <mmayer@xxxxxxxxxxxx>
>>
>> This CPUfreq driver provides basic frequency scaling for older Broadcom
>> STB SoCs that do not use AVS firmware with DVFS support. There is no
>> support for voltage scaling.
>>
>> Signed-off-by: Markus Mayer <mmayer@xxxxxxxxxxxx>
>> ---
>>
>> This patch is based on Rafael's "bleeding-edge" branch.
>>
>> drivers/cpufreq/Kconfig.arm | 12 ++
>> drivers/cpufreq/Makefile | 1 +
>> drivers/cpufreq/brcmstb-cpufreq.c | 407 ++++++++++++++++++++++++++++++++++++++
>> 3 files changed, 420 insertions(+)
>> create mode 100644 drivers/cpufreq/brcmstb-cpufreq.c
>>
>> diff --git a/drivers/cpufreq/Kconfig.arm b/drivers/cpufreq/Kconfig.arm
>> index 920c469..36422af 100644
>> --- a/drivers/cpufreq/Kconfig.arm
>> +++ b/drivers/cpufreq/Kconfig.arm
>> @@ -33,6 +33,18 @@ config ARM_BRCMSTB_AVS_CPUFREQ_DEBUG
>>
>> If in doubt, say N.
>>
>> +config ARM_BRCMSTB_CPUFREQ
>> + tristate "Broadcom STB CPUfreq driver"
>> + depends on ARCH_BRCMSTB || COMPILE_TEST
>> + default y
>> + help
>> + Some Broadcom SoCs offer multiple operating frequencies that CPUfreq
>> + can take advantage of to improve energy efficiency.
>> +
>> + Say Y, if you have a supported Broadcom SoC. If your Broadcom SoC
>> + has AVS firmware with support for frequency and voltage scaling,
>> + say N here and enable ARM_BRCMSTB_AVS_CPUFREQ instead.
>> +
>> config ARM_DT_BL_CPUFREQ
>> tristate "Generic probing via DT for ARM big LITTLE CPUfreq driver"
>> depends on ARM_BIG_LITTLE_CPUFREQ && OF
>> diff --git a/drivers/cpufreq/Makefile b/drivers/cpufreq/Makefile
>> index 1e46c39..23700aa 100644
>> --- a/drivers/cpufreq/Makefile
>> +++ b/drivers/cpufreq/Makefile
>> @@ -52,6 +52,7 @@ obj-$(CONFIG_ARM_BIG_LITTLE_CPUFREQ) += arm_big_little.o
>> obj-$(CONFIG_ARM_DT_BL_CPUFREQ) += arm_big_little_dt.o
>>
>> obj-$(CONFIG_ARM_BRCMSTB_AVS_CPUFREQ) += brcmstb-avs-cpufreq.o
>> +obj-$(CONFIG_ARM_BRCMSTB_CPUFREQ) += brcmstb-cpufreq.o
>> obj-$(CONFIG_ARCH_DAVINCI) += davinci-cpufreq.o
>> obj-$(CONFIG_UX500_SOC_DB8500) += dbx500-cpufreq.o
>> obj-$(CONFIG_ARM_EXYNOS5440_CPUFREQ) += exynos5440-cpufreq.o
>> diff --git a/drivers/cpufreq/brcmstb-cpufreq.c b/drivers/cpufreq/brcmstb-cpufreq.c
>> new file mode 100644
>> index 0000000..419638a
>> --- /dev/null
>> +++ b/drivers/cpufreq/brcmstb-cpufreq.c
>> @@ -0,0 +1,407 @@
>> +/*
>> + * CPU frequency scaling for Broadcom set top box SoCs
>> + *
>> + * Copyright (c) 2016 Broadcom
>> + *
>> + * This program is free software; you can redistribute it and/or
>> + * modify it under the terms of the GNU General Public License as
>> + * published by the Free Software Foundation version 2.
>> + *
>> + * This program is distributed "as is" WITHOUT ANY WARRANTY of any
>> + * kind, whether express or implied; without even the implied warranty
>> + * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
>> + * GNU General Public License for more details.
>> + */
>> +
>> +#include <linux/clk.h>
>> +#include <linux/clk-provider.h>
>> +#include <linux/cpufreq.h>
>> +#include <linux/module.h>
>> +#include <linux/of_address.h>
>> +#include <linux/platform_device.h>
>> +
>> +#define BRCMSTB_CPUFREQ_PREFIX "brcmstb"
>> +#define BRCMSTB_CPUFREQ_NAME BRCMSTB_CPUFREQ_PREFIX "-cpufreq"
>> +
>> +/* We search for these compatible strings. */
>> +#define BRCMSTB_DT_CPU_CLK_CTRL "brcm,brcmstb-cpu-clk-div"
>> +#define BRCMSTB_DT_MEMC_DDR "brcm,brcmstb-memc-ddr"
>> +#define BRCM_AVS_CPU_DATA "brcm,avs-cpu-data-mem"
>> +
>> +/* We also need a few clocks in device tree. These are node names. */
>> +#define BRCMSTB_CLK_MDIV_CH0 "cpu_mdiv_ch0"
>> +#define BRCMSTB_CLK_NDIV_INT "cpu_ndiv_int"
>> +#define BRCMSTB_CLK_SW_SCB "sw_scb"
>> +
>> +#define BRCMSTB_TBL_SAFE_MODE BIT(0)
>> +#define BRCMSTB_REG_SAFE_MODE BIT(4)
>> +
>> +/* This is as low as we'll go in the frequency table. */
>> +#define MIN_CPU_FREQ (100 * 1000) /* in kHz */
>> +
>> +struct private_data {
>> + void __iomem *cpu_clk_ctrl_reg;
>> + struct device *dev;
>> +};
>> +
>> +/* Count the active memory controllers in the system. */
>> +static int count_memory_controllers(void)
>> +{
>> + struct device_node *np = NULL;
>> + int i = 0;
>> +
>> + do {
>> + np = of_find_compatible_node(np, NULL, BRCMSTB_DT_MEMC_DDR);
>> + if (of_device_is_available(np))
>> + i++;
>> + of_node_put(np);
>> + } while (np);
>> +
>> + return i;
>> +}
>> +
>> +static int get_frequencies(const struct cpufreq_policy *policy,
>> + unsigned int *vco_freq, unsigned int *cpu_freq,
>> + unsigned int *scb_freq)
>> +{
>> + struct clk *cpu_ndiv_int, *sw_scb;
>> +
>> + cpu_ndiv_int = __clk_lookup(BRCMSTB_CLK_NDIV_INT);
>> + if (!cpu_ndiv_int)
>> + return -ENODEV;
>> +
>> + sw_scb = __clk_lookup(BRCMSTB_CLK_SW_SCB);
>> + if (!sw_scb)
>> + return -ENODEV;
>> +
>> + /* return frequencies in kHz */
>> + *vco_freq = clk_get_rate(cpu_ndiv_int) / 1000;
>> + *cpu_freq = clk_get_rate(policy->clk) / 1000;
>> + *scb_freq = clk_get_rate(sw_scb) / 1000;
>> +
>> + return 0;
>> +}
>> +
>> +/*
>> + * Safe mode: When set, the CPU's bus unit is being throttled. This is done to
>> + * avoid buffer overflows when the CPU-to-bus-clock ratio is low.
>> + *
>> + * The formula as to what constitutes a low CPU-to-bus-clock ratio takes into
>> + * account the number of memory controllers active in the system and the SCB
>> + * frequency. More memory controllers means safe mode is required starting at
>> + * higher frequencies.
>> + *
>> + * For 1 memory controller, cpu_freq/scb_freq must be greater than or equal to
>> + * 2 to not require safe mode.
>> + *
>> + * For 2 or 3 memory controllers, cpu_freq/scb_freq must be greater than or
>> + * equal 3 to not require safe mode.
>> + */
>> +
>> +static int freq_requires_safe_mode(unsigned int cpu_freq, unsigned int scb_freq,
>> + int num_memc)
>> +{
>> + unsigned int safe_ratio;
>> +
>> + switch (num_memc) {
>> + case 1:
>> + safe_ratio = 2;
>> + break;
>> + case 2:
>> + case 3:
>> + safe_ratio = 3;
>> + break;
>> + default:
>> + return -EINVAL;
>> + }
>> +
>> + return ((cpu_freq / scb_freq) < safe_ratio);
>> +}
>> +
>> +static struct cpufreq_frequency_table *
>> +brcmstb_get_freq_table(const struct cpufreq_policy *policy)
>> +{
>> + unsigned int cpu_freq, vco_freq, scb_freq, mdiv, init_mdiv, f;
>> + struct cpufreq_frequency_table *table;
>> + struct private_data *priv;
>> + int num_memc, ret;
>> + unsigned int i = 0;
>> +
>> + ret = get_frequencies(policy, &vco_freq, &cpu_freq, &scb_freq);
>> + if (ret)
>> + return ERR_PTR(ret);
>> +
>> + priv = policy->driver_data;
>> + num_memc = count_memory_controllers();
>> +
>> + /* Calculate the initial mdiv value. We'll increment mdiv from here. */
>> + init_mdiv = vco_freq / cpu_freq;
>> +
>> + /* Count how many frequencies we'll offer. */
>> + f = cpu_freq;
>> + for (mdiv = init_mdiv; f >= MIN_CPU_FREQ; mdiv++, f = vco_freq / mdiv) {
>> + /* We only want to use "whole" MHz. */
>> + if ((f % 1000) == 0)
>> + i++;
>> + }
>> +
>> + table = devm_kzalloc(priv->dev, (i + 1) * sizeof(*table), GFP_KERNEL);
>> + if (!table)
>> + return ERR_PTR(-ENOMEM);
>> +
>> + /* Now, fill the table. */
>> + f = cpu_freq;
>> + i = 0;
>> + for (mdiv = init_mdiv; f >= MIN_CPU_FREQ; mdiv++, f = vco_freq / mdiv) {
>> + if ((f % 1000) == 0) {
>> + table[i].frequency = f;
>> + ret = freq_requires_safe_mode(f, scb_freq, num_memc);
>> + if (ret < 0)
>> + return ERR_PTR(ret);
>> + if (ret > 0)
>> + table[i].driver_data |= BRCMSTB_TBL_SAFE_MODE;
>> + i++;
>> + }
>> + }
>> + table[i].frequency = CPUFREQ_TABLE_END;
>> +
>> + return table;
>> +}
>> +
>> +static unsigned int brcmstb_cpufreq_get(unsigned int cpu)
>> +{
>> + struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
>> +
>> + return clk_get_rate(policy->clk) / 1000;
>> +}
>
> Rather use cpufreq_generic_get().

Done.

>> +
>> +static int brcmstb_target_index(struct cpufreq_policy *policy,
>> + unsigned int index)
>> +{
>> + struct cpufreq_frequency_table *entry;
>> + struct private_data *priv;
>> + int ret, safe_mode_needed;
>> + u32 reg;
>> +
>> + priv = policy->driver_data;
>> + entry = &policy->freq_table[index];
>> + safe_mode_needed = entry->driver_data & BRCMSTB_TBL_SAFE_MODE;
>> +
>> + reg = readl(priv->cpu_clk_ctrl_reg);
>> + if (safe_mode_needed && !(reg & BRCMSTB_REG_SAFE_MODE)) {
>> + reg |= BRCMSTB_REG_SAFE_MODE;
>> + writel(reg, priv->cpu_clk_ctrl_reg);
>> + }
>> + ret = clk_set_rate(policy->clk, entry->frequency * 1000);
>> + if (!ret && !safe_mode_needed && (reg & BRCMSTB_REG_SAFE_MODE)) {
>> + reg &= ~BRCMSTB_REG_SAFE_MODE;
>> + writel(reg, priv->cpu_clk_ctrl_reg);
>> + }
>> +
>> + return ret;
>> +}
>> +
>> +/*
>> + * All initialization code that we only want to execute once goes here. Setup
>> + * code that can be re-tried on every core (if it failed before) can go into
>> + * brcm_avs_cpufreq_init().
>> + */
>> +static int brcmstb_prepare_init(struct platform_device *pdev)
>> +{
>> + struct private_data *priv;
>> + struct resource *res;
>> + struct device *dev;
>> +
>> + /*
>> + * If the BRCM STB AVS CPUfreq driver is supported, we bail, so that
>> + * the more modern approach implementing DVFS in firmware can be used.
>> + */
>> + if (IS_ENABLED(CONFIG_ARM_BRCM_AVS_CPUFREQ)) {
>> + struct device_node *np;
>> +
>> + np = of_find_compatible_node(NULL, NULL, BRCM_AVS_CPU_DATA);
>> + if (np) {
>> + of_node_put(np);
>> + return -ENXIO;
>> + }
>> + }
>> +
>> + dev = &pdev->dev;
>> +
>> + priv = devm_kzalloc(dev, sizeof(*priv), GFP_KERNEL);
>> + if (!priv)
>> + return -ENOMEM;
>> +
>> + res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
>> + priv->cpu_clk_ctrl_reg = devm_ioremap_resource(dev, res);
>> + if (IS_ERR(priv->cpu_clk_ctrl_reg)) {
>> + dev_err(dev, "couldn't map DT entry %s\n",
>> + BRCMSTB_DT_CPU_CLK_CTRL);
>> + return -ENODEV;
>> + }
>> +
>> + priv->dev = dev;
>> + platform_set_drvdata(pdev, priv);
>> +
>> + return 0;
>> +}
>> +
>> +static int brcmstb_cpu_init(struct cpufreq_policy *policy)
>> +{
>> + struct cpufreq_frequency_table *freq_table;
>> + struct platform_device *pdev;
>> + struct private_data *priv;
>> + struct clk *cpu_mdiv_ch0;
>> + struct device *dev;
>> + int ret;
>> +
>> + cpu_mdiv_ch0 = __clk_lookup(BRCMSTB_CLK_MDIV_CH0);
>> + if (!cpu_mdiv_ch0)
>> + return -ENODEV;
>> +
>> + pdev = cpufreq_get_driver_data();
>> + priv = platform_get_drvdata(pdev);
>> + dev = &pdev->dev;
>> +
>> + policy->clk = cpu_mdiv_ch0;
>> + policy->driver_data = priv;
>> +
>> + freq_table = brcmstb_get_freq_table(policy);
>> + if (IS_ERR(freq_table)) {
>> + ret = PTR_ERR(freq_table);
>> + dev_err(dev, "Couldn't determine frequency table (%d).\n", ret);
>> + if (ret == -EINVAL)
>> + dev_emerg(dev,
>> + "Invalid number of memory controllers -- %d!\n",
>> + count_memory_controllers());
>> + return ret;
>> + }
>> +
>> + ret = cpufreq_table_validate_and_show(policy, freq_table);
>> + if (ret) {
>> + dev_err(dev, "invalid frequency table: %d\n", ret);
>> + return ret;
>> + }
>> +
>> + dev_info(dev, "registered\n");
>> +
>> + /* All cores share the same clock and thus the same policy. */
>> + cpumask_setall(policy->cpus);
>
> You can use cpufreq_generic_init() instead of above two calls.

Done.

>> +
>> + /* We start at the first entry in the frequency table. */
>> + policy->cur = freq_table[0].frequency;
>
> No, you can't update policy->cur by yourself. Its for the core to
> handle it. Though check if you want to use
> CPUFREQ_NEED_INITIAL_FREQ_CHECK.

Yes, that works.

>> +
>> + return 0;
>> +}
>> +
>> +/* Shows the number of memory controllers. */
>> +static ssize_t show_brcmstb_num_memc(struct cpufreq_policy *policy, char *buf)
>> +{
>> + return sprintf(buf, "%u\n", count_memory_controllers());
>> +}
>> +
>> +/* Shows vco_freq, cpu_freq, and scb_freq in kHz. */
>> +static ssize_t show_brcmstb_freqs(struct cpufreq_policy *policy, char *buf)
>> +{
>> + unsigned int vco_freq, cpu_freq, scb_freq;
>> + int ret;
>> +
>> + ret = get_frequencies(policy, &vco_freq, &cpu_freq, &scb_freq);
>> + if (ret)
>> + return sprintf(buf, "<unknown>\n");
>> +
>> + return sprintf(buf, "%u %u %u\n", vco_freq, cpu_freq, scb_freq);
>> +}
>> +
>> +/* Shows the lowest frequency (in kHz) that can be used without "safe mode". */
>> +static ssize_t show_brcmstb_safe_freq(struct cpufreq_policy *policy, char *buf)
>> +{
>> + unsigned int vco_freq, cpu_freq, scb_freq;
>> + unsigned int safe_freq = 0;
>> + int i, num_memc, ret;
>> +
>> + ret = get_frequencies(policy, &vco_freq, &cpu_freq, &scb_freq);
>> + if (ret)
>> + return sprintf(buf, "<unknown>\n");
>> +
>> + num_memc = count_memory_controllers();
>> +
>> + for (i = 0; policy->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
>
> You can use cpufreq_for_each_valid_entry() here.
>
>> + ret = freq_requires_safe_mode(policy->freq_table[i].frequency,
>> + scb_freq, num_memc);
>> + if (ret < 0)
>> + return sprintf(buf, "<error> (%d)\n", ret);
>> + if (ret == 0)
>> + safe_freq = policy->freq_table[i].frequency;
>
> You don't want to break from the loop here ?

No, because I am trying to find the lowest frequency that doesn't
require safe mode and it's looping through the table from highest to
lowest. So there could still be a lower frequency after the current
one that doesn't require safe mode.

What I can do, however, is something like this:

static ssize_t show_brcmstb_safe_freq(struct cpufreq_policy *policy, char *buf)
{
struct cpufreq_frequency_table *entry;
unsigned int safe_freq = 0;

cpufreq_for_each_valid_entry(entry, policy->freq_table) {
if (!(entry->driver_data & BRCMSTB_TBL_SAFE_MODE))
safe_freq = entry->frequency;
}

return sprintf(buf, "%u\n", safe_freq);
}

This is using the existing data from the frequency table rather than
re-generating it on the fly by calling freq_requires_safe_mode().

>> + }
>> +
>> + return sprintf(buf, "%u\n", safe_freq);
>> +}
>> +
>> +cpufreq_freq_attr_ro(brcmstb_num_memc);
>> +cpufreq_freq_attr_ro(brcmstb_freqs);
>> +cpufreq_freq_attr_ro(brcmstb_safe_freq);
>> +
>> +static struct freq_attr *brcmstb_cpufreq_attr[] = {
>> + &cpufreq_freq_attr_scaling_available_freqs,
>> + &brcmstb_num_memc,
>> + &brcmstb_freqs,
>> + &brcmstb_safe_freq,
>> + NULL
>> +};
>> +
>> +static struct cpufreq_driver brcmstb_driver = {
>> + .flags = CPUFREQ_NEED_INITIAL_FREQ_CHECK,
>> + .verify = cpufreq_generic_frequency_table_verify,
>> + .target_index = brcmstb_target_index,
>> + .get = brcmstb_cpufreq_get,
>> + .init = brcmstb_cpu_init,
>
> What about .exit ?

All my allocations are managed (i.e. using devm* functions), so
cleanup should be automatic. Do I still need one?

>> + .attr = brcmstb_cpufreq_attr,
>> + .name = BRCMSTB_CPUFREQ_PREFIX,
>> +};
>> +
>> +static int brcmstb_cpufreq_probe(struct platform_device *pdev)
>> +{
>> + int ret;
>> +
>> + ret = brcmstb_prepare_init(pdev);
>> + if (ret)
>> + return ret;
>> +
>> + brcmstb_driver.driver_data = pdev;
>> +
>> + return cpufreq_register_driver(&brcmstb_driver);
>> +}
>> +
>> +static int brcmstb_cpufreq_remove(struct platform_device *pdev)
>> +{
>> + int ret;
>> +
>> + ret = cpufreq_unregister_driver(&brcmstb_driver);
>> + if (ret)
>> + return ret;
>> +
>> + platform_set_drvdata(pdev, NULL);
>> +
>> + return 0;
>> +}
>> +
>> +static const struct of_device_id brcmstb_cpufreq_match[] = {
>> + { .compatible = BRCMSTB_DT_CPU_CLK_CTRL },
>> + { }
>> +};
>> +MODULE_DEVICE_TABLE(platform, brcmstb_cpufreq_match);
>> +
>> +static struct platform_driver brcmstb_cpufreq_platdrv = {
>
> How is the device getting created for this? Be aware the DT can't have
> a separate node for cpufreq-device ..

It's using for a clock node (brcm,brcmstb-cpu-clk-div) which exists
independently of this driver.

>> + .driver = {
>> + .name = BRCMSTB_CPUFREQ_NAME,
>> + .of_match_table = brcmstb_cpufreq_match,
>> + },
>> + .probe = brcmstb_cpufreq_probe,
>> + .remove = brcmstb_cpufreq_remove,
>> +};
>> +module_platform_driver(brcmstb_cpufreq_platdrv);
>> +
>> +MODULE_AUTHOR("Markus Mayer <mmayer@xxxxxxxxxxxx>");
>> +MODULE_DESCRIPTION("CPUfreq driver for Broadcom STB SoCs");
>> +MODULE_LICENSE("GPL");

Thanks,
-Markus

> --
> viresh