| Message ID | 20210119174557.227318-1-angelogioacchino.delregno@somainline.org |
|---|---|
| Headers | show |
| Series | cpufreq-qcom-hw: Implement full OSM programming | expand |
The CPUFREQ-HW driver is intended to be used only for CPUFREQ HW designs where the firmware programs the look up tables. Suggestion is to separate out the driver where the programming is managed by high level OS. On 1/19/2021 11:15 PM, AngeloGioacchino Del Regno wrote: > On new SoCs (SDM845 onwards) the Operating State Manager (OSM) is > being programmed in the bootloader and write-protected by the > hypervisor, leaving to the OS read-only access to some of its > registers (in order to read the Lookup Tables and also some > status registers) and write access to the p-state register, for > for the OS to request a specific performance state to trigger a > DVFS switch on the CPU through the OSM hardware. > > On old SoCs though (MSM8998, SDM630/660 and variants), the > bootloader will *not* initialize the OSM (and the CPRh, as it > is a requirement for it) before booting the OS, making any > request to trigger a performance state change ineffective, as > the hardware doesn't have any Lookup Table, nor is storing any > parameter to trigger a DVFS switch. In this case, basically all > of the OSM registers are *not* write protected for the OS, even > though some are - but write access is granted through SCM calls. > > This commit introduces support for OSM programming, which has to > be done on these old SoCs that were distributed (almost?) always > with a bootloader that does not do any CPRh nor OSM init before > booting the kernel. > In order to program the OSM on these SoCs, it is necessary to > fullfill a "special" requirement: the Core Power Reduction > Hardened (CPRh) hardware block must be initialized, as the OSM > is "talking" to it in order to perform the Voltage part of DVFS; > here, we are calling initialization of this through Linux generic > power domains, specifically by requesting a genpd attach from the > qcom-cpufreq-hw driver, which will give back voltages associated > to each CPU frequency that has been declared in the OPPs, scaled > and interpolated with the previous one, and will also give us > parameters for the Array Power Mux (APM) and mem-acc, in order > for this driver to be then able to generate the Lookup Tables > that will be finally programmed to the OSM hardware. > > After writing the parameters to the OSM and enabling it, all the > programming work will never happen anymore until a OS reboot, so > all of the allocations and "the rest" will be disposed-of: this > is done mainly to leave the code that was referred only to the > new SoCs intact, as to also emphasize on the fact that the OSM > HW is, in the end, the exact same; apart some register offsets > that are slightly different, the entire logic is the same. > > This also adds the parameters to support CPU scaling on SDM630 > and MSM8998. > > Signed-off-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@somainline.org> > --- > drivers/cpufreq/qcom-cpufreq-hw.c | 1240 ++++++++++++++++++++++++++++- > 1 file changed, 1208 insertions(+), 32 deletions(-) > > diff --git a/drivers/cpufreq/qcom-cpufreq-hw.c b/drivers/cpufreq/qcom-cpufreq-hw.c > index acc645b85e79..a92959bb7b50 100644 > --- a/drivers/cpufreq/qcom-cpufreq-hw.c > +++ b/drivers/cpufreq/qcom-cpufreq-hw.c > @@ -1,33 +1,256 @@ > // SPDX-License-Identifier: GPL-2.0 > /* > * Copyright (c) 2018, The Linux Foundation. All rights reserved. > + * > + * OSM hardware initial programming > + * Copyright (C) 2020, AngeloGioacchino Del Regno > + * <angelogioacchino.delregno@somainline.org> > */ > > #include <linux/bitfield.h> > #include <linux/cpufreq.h> > +#include <linux/delay.h> > #include <linux/init.h> > #include <linux/interconnect.h> > #include <linux/kernel.h> > #include <linux/module.h> > #include <linux/of_address.h> > #include <linux/of_platform.h> > +#include <linux/pm_domain.h> > #include <linux/pm_opp.h> > #include <linux/slab.h> > +#include <linux/qcom_scm.h> > > #define LUT_MAX_ENTRIES 40U > -#define LUT_SRC GENMASK(31, 30) > +#define LUT_SRC_845 GENMASK(31, 30) > +#define LUT_SRC_8998 GENMASK(27, 26) > +#define LUT_PLL_DIV GENMASK(25, 24) > #define LUT_L_VAL GENMASK(7, 0) > #define LUT_CORE_COUNT GENMASK(18, 16) > +#define LUT_VOLT_VC GENMASK(21, 16) > #define LUT_VOLT GENMASK(11, 0) > -#define CLK_HW_DIV 2 > #define LUT_TURBO_IND 1 > +#define OSM_BOOT_TIME_US 5 > + > +#define CYCLE_COUNTER_CLK_RATIO GENMASK(5, 1) > +#define OSM_XO_RATIO_VAL (10 - 1) > +#define CYCLE_COUNTER_USE_XO_EDGE BIT(8) > + > +/* FSM Boost Control */ > +#define CC_BOOST_EN BIT(0) > +#define PS_BOOST_EN BIT(1) > +#define DCVS_BOOST_EN BIT(2) > +#define BOOST_TIMER_REG_HI GENMASK(31, 16) > +#define BOOST_TIMER_REG_LO GENMASK(15, 0) > + > +#define PLL_WAIT_LOCK_TIME_NS 2000 > +#define SAFE_FREQ_WAIT_NS 1000 > +#define DEXT_DECREMENT_WAIT_NS 200 > + > +#define BOOST_SYNC_DELAY 5 > + > +#define HYSTERESIS_UP_MASK GENMASK(31, 16) > +#define HYSTERESIS_DN_MASK GENMASK(15, 0) > +#define HYSTERESIS_CC_NS 200 > +#define HYSTERESIS_LLM_NS 65535 > + > +/* FSM Droop Control */ > +#define PC_RET_EXIT_DROOP_EN BIT(3) > +#define WFX_DROOP_EN BIT(4) > +#define DCVS_DROOP_EN BIT(5) > +#define DROOP_TIMER1 GENMASK(31, 16) > +#define DROOP_TIMER0 GENMASK(15, 0) > +#define DROOP_CTRL_VAL (BIT(3) | BIT(17) | BIT(31)) > +#define DROOP_TIMER_NS 100 > +#define DROOP_WAIT_RELEASE_TIMER_NS 50 > +#define DROOP_RELEASE_TIMER_NS 1 > + > +/* PLL Override Control */ > +#define PLL_OVERRIDE_DROOP_EN BIT(0) > + > +/* Sequencer */ > +#define SEQUENCER_REG(base, n) (base + (n * 4)) > +#define SEQ_APM_THRESH_VC 15 > +#define SEQ_APM_THRESH_PREVC 31 > +#define SEQ_MEM_ACC_LVAL 32 > +#define SEQ_MEM_ACC_0 55 > +#define SEQ_APM_CROSSOVER_VC 72 > +#define SEQ_APM_PARAM 76 > +#define SEQ_MEM_ACC_CROSSOVER_VC 88 > +#define SEQ_MEM_ACC_MAX_LEVELS 4 > +#define SEQ_MEMACC_REG(base, n) SEQUENCER_REG(base, SEQ_MEM_ACC_0 + n) > + > +/* ACD */ > +#define ACD_WRITE_CTL_UPDATE_EN BIT(0) > +#define ACD_WRITE_CTL_SELECT_SHIFT 1 > + > +/** > + * struct qcom_cpufreq_soc_setup_data - Register offsets for OSM setup > + * > + * @reg_osm_sequencer: OSM Sequencer (used to get physical address) > + * @reg_override: Override parameters > + * @reg_spare: Spare parameters (MEMACC-to-VC) > + * @reg_cc_zero_behav: Virtual Corner for cluster power collapse > + * @reg_spm_cc_hyst: DCVS-CC Wait time for frequency inc/decrement > + * @reg_spm_cc_dcvs_dis: DCVS-CC en/disable control > + * @reg_spm_core_ret_map: Treat cores in retention as active/inactive > + * @reg_llm_freq_vote_hyst: DCVS-LLM Wait time for frequency inc/decrement > + * @reg_llm_volt_vote_hyst: DCVS-LLM Wait time for voltage inc/decrement > + * @reg_llm_intf_dcvs_dis: DCVS-LLM en/disable control > + * @reg_seq1: Sequencer extra register > + * @reg_pdn_fsm_ctrl: Boost and Droop FSMs en/disable control > + * @reg_cc_boost_timer: CC-Boost FSM wait first timer register > + * @reg_dcvs_boost_timer: DCVS-Boost FSM wait first timer register > + * @reg_ps_boost_timer: PS-Boost FSM wait first timer register > + * @boost_timer_reg_len: Length of boost timer registers > + * @reg_boost_sync_delay: PLL signal timing control for Boost > + * @reg_droop_ctrl: Droop control value > + * @reg_droop_release_ctrl: Wait for Droop release > + * @reg_droop_unstall_ctrl: Wait for Droop unstall > + * @reg_droop_wait_release_ctrl: Time to wait for state release > + * @reg_droop_timer_ctrl: Droop timer > + * @reg_droop_sync_delay: PLL signal timing control for Droop > + * @reg_pll_override: PLL Droop Override en/disable control > + * @reg_cycle_counter: OSM CPU cycle counter > + * > + * This structure holds the register offsets that are used to set-up > + * the Operating State Manager (OSM) parameters, when it is not (or > + * not entirely) configured from the bootloader and TrustZone. > + * > + * Acronyms used in this documentation: > + * CC = Core Count > + * PS = Power-Save > + * VC = Virtual Corner > + * LLM = Limits Load Management > + * DCVS = Dynamic Clock and Voltage Scaling > + */ > +struct qcom_cpufreq_soc_setup_data { > + /* OSM phys register offsets */ > + u16 reg_osm_sequencer; > + > + /* Frequency domain register offsets */ > + u16 reg_override; > + u16 reg_spare; > + u16 reg_cc_zero_behav; > + u16 reg_spm_cc_hyst; > + u16 reg_spm_cc_dcvs_dis; > + u16 reg_spm_core_ret_map; > + u16 reg_llm_freq_vote_hyst; > + u16 reg_llm_volt_vote_hyst; > + u16 reg_llm_intf_dcvs_dis; > + u16 reg_seq1; > + u16 reg_pdn_fsm_ctrl; > + u16 reg_cc_boost_timer; > + u16 reg_dcvs_boost_timer; > + u16 reg_ps_boost_timer; > + u16 boost_timer_reg_len; > + u16 reg_boost_sync_delay; > + u16 reg_droop_ctrl; > + u16 reg_droop_release_ctrl; > + u16 reg_droop_unstall_ctrl; > + u16 reg_droop_wait_release_ctrl; > + u16 reg_droop_timer_ctrl; > + u16 reg_droop_sync_delay; > + u16 reg_pll_override; > + u16 reg_cycle_counter; > +}; > + > +/** > + * struct qcom_cpufreq_soc_acd_data - Adaptive Clock Distribution data > + * > + * @tl_delay_reg: Tunable-Length Delay (TLD) register offset > + * @acd_ctrl_reg: Control Register (CR) register offset > + * @softstart_reg: Soft Start Control Register (SSCR) register offset > + * @ext_intf_reg: External interface configuration register offset > + * @auto_xfer_reg: Auto Register-Transfer register offset > + * @auto_xfer_cfg_reg: Auto Register-Transfer Configuration reg offset > + * @auto_xfer_ctl_reg: Auto Register-Transfer Control register offset > + * @auto_xfer_sts_reg: Auto Register-Transfer Status register offset > + * @dcvs_sw_reg: Software DCVS register offset > + * @gfmux_cfg_reg: Glitch-Free MUX configuration register offset > + * @write_ctl_reg: Write Control register > + * @write_sts_reg: Write Status register > + * @tl_delay_val: Tunable-Length Delay (TLD) value > + * @acd_ctrl_val: Control Register (CR) value > + * @softstart_val: Soft Start Control Register (SSCR) value > + * @ext_intf0_val: Initial external interface configuration value > + * @ext_intf1_val: Final external interface configuration value > + * @auto_xfer_val: Auto-register Transfer Control value > + * > + * This structure holds the register offsets (from the ACD iospace base) > + * and the parameters that are required to configure the OSM to > + * initialize the Adaptive Clock Distribution (ACD) system. > + */ > +struct qcom_cpufreq_soc_acd_data { > + u8 tl_delay_reg; > + u8 acd_ctrl_reg; > + u8 softstart_reg; > + u8 ext_intf_reg; > + u8 auto_xfer_reg; > + u8 auto_xfer_cfg_reg; > + u8 auto_xfer_ctl_reg; > + u8 auto_xfer_sts_reg; > + u8 dcvs_sw_reg; > + u8 gfmux_cfg_reg; > + u8 write_ctl_reg; > + u8 write_sts_reg; > + u32 tl_delay_val; > + u32 acd_ctrl_val; > + u32 softstart_val; > + u32 ext_intf0_val; > + u32 ext_intf1_val; > + u32 auto_xfer_val; > +}; > + > +/** > + * struct qcom_cpufreq_hw_params - Operating State Manager (OSM) Parameters > + * > + * @volt_lut_val: Value composed of: virtual corner (vc) and voltage in mV. > + * @freq_lut_val: Value composed of: core count, clock source and output > + * frequency in MHz. > + * @override_val: PLL parameters that the OSM uses to override the previous > + * setting coming from the bootloader, or when uninitialized. > + * @spare_val: Spare register, used by both this driver and the OSM HW > + * to identify MEM-ACC levels in relation to virtual corners. > + * @acc_thresh: MEM-ACC threshold level > + * > + * This structure holds the parameters to write to the OSM registers for > + * one "Virtual Corner" (VC), or one Performance State (p-state). > + */ > +struct qcom_cpufreq_hw_params { > + u32 volt_lut_val; > + u32 freq_lut_val; > + u32 override_val; > + u32 spare_val; > + u32 acc_thresh; > +}; > > +/** > + * struct qcom_cpufreq_soc_data - SoC specific register offsets of the OSM > + * > + * @reg_enable: OSM enable status > + * @reg_index: Index of the Virtual Corner > + * @reg_freq_lut: Frequency Lookup Table > + * @reg_freq_lut_src_mask: Frequency Lookup Table clock-source mask > + * @reg_volt_lut: Voltage Lookup Table > + * @reg_perf_state: Performance State request register > + * @lut_row_size: Lookup Table row size > + * @clk_hw_div: Divider for "alternate" OSM clock-source > + * @uses_tz: OSM already set-up and protected by TrustZone > + * @setup_regs: Register offsets for OSM setup > + */ > struct qcom_cpufreq_soc_data { > u32 reg_enable; > + u32 reg_index; > u32 reg_freq_lut; > + u32 reg_freq_lut_src_mask; > u32 reg_volt_lut; > u32 reg_perf_state; > - u8 lut_row_size; > + u8 lut_row_size; > + u8 clk_hw_div; > + bool uses_tz; > + const struct qcom_cpufreq_soc_setup_data setup_regs; > + const struct qcom_cpufreq_soc_acd_data acd_data; > }; > > struct qcom_cpufreq_data { > @@ -35,9 +258,17 @@ struct qcom_cpufreq_data { > const struct qcom_cpufreq_soc_data *soc_data; > }; > > -static unsigned long cpu_hw_rate, xo_rate; > +static const char *cprh_genpd_names[] = { "cprh", NULL }; > +static u64 cpu_hw_rate, xo_rate; > static bool icc_scaling_enabled; > > +/** > + * qcom_cpufreq_set_bw - Set interconnect bandwidth > + * @policy: CPUFreq policy structure > + * @freq_khz: CPU Frequency in KHz > + * > + * Returns: Zero for success, otherwise negative value on errors > + */ > static int qcom_cpufreq_set_bw(struct cpufreq_policy *policy, > unsigned long freq_khz) > { > @@ -59,6 +290,20 @@ static int qcom_cpufreq_set_bw(struct cpufreq_policy *policy, > return ret; > } > > +/** > + * qcom_cpufreq_update_opp - Update CPU OPP tables > + * @policy: CPUFreq policy structure > + * @freq_khz: CPU Frequency for OPP entry in KHz > + * @volt: CPU Voltage for OPP entry in microvolts > + * > + * The CPU frequencies and voltages are being read from the Operating > + * State Manager (OSM) and the related OPPs, read from DT, need to be > + * updated to reflect what the hardware will set for each p-state. > + * If there is no OPP table specified in DT, then this function will > + * add dynamic ones. > + * > + * Returns: Zero for success, otherwise negative value on errors > + */ > static int qcom_cpufreq_update_opp(struct device *cpu_dev, > unsigned long freq_khz, > unsigned long volt) > @@ -79,6 +324,17 @@ static int qcom_cpufreq_update_opp(struct device *cpu_dev, > return dev_pm_opp_enable(cpu_dev, freq_hz); > } > > +/** > + * qcom_cpufreq_hw_target_index - Set frequency/voltage > + * @policy: CPUFreq policy structure > + * @index: Performance state index to be set > + * > + * This function sends a request to the Operating State Manager > + * to set a Performance State index, so, to set frequency and > + * voltage for the target CPU/cluster. > + * > + * Returns: Always zero > + */ > static int qcom_cpufreq_hw_target_index(struct cpufreq_policy *policy, > unsigned int index) > { > @@ -94,6 +350,12 @@ static int qcom_cpufreq_hw_target_index(struct cpufreq_policy *policy, > return 0; > } > > +/** > + * qcom_cpufreq_hw_get - Get current Performance State from OSM > + * @cpu: CPU number > + * > + * Returns: Current CPU/Cluster frequency or zero > + */ > static unsigned int qcom_cpufreq_hw_get(unsigned int cpu) > { > struct qcom_cpufreq_data *data; > @@ -127,6 +389,644 @@ static unsigned int qcom_cpufreq_hw_fast_switch(struct cpufreq_policy *policy, > return policy->freq_table[index].frequency; > } > > +static void qcom_cpufreq_hw_boost_setup(void __iomem *timer0_addr, u32 len) > +{ > + u32 val; > + > + /* timer_reg0 */ > + val = FIELD_PREP(BOOST_TIMER_REG_LO, PLL_WAIT_LOCK_TIME_NS); > + val |= FIELD_PREP(BOOST_TIMER_REG_HI, SAFE_FREQ_WAIT_NS); > + writel_relaxed(val, timer0_addr); > + > + /* timer_reg1 */ > + val = FIELD_PREP(BOOST_TIMER_REG_LO, PLL_WAIT_LOCK_TIME_NS); > + val |= FIELD_PREP(BOOST_TIMER_REG_HI, PLL_WAIT_LOCK_TIME_NS); > + writel_relaxed(val, timer0_addr + len); > + > + /* timer_reg2 */ > + val = FIELD_PREP(BOOST_TIMER_REG_LO, DEXT_DECREMENT_WAIT_NS); > + writel_relaxed(val, timer0_addr + (2 * len)); > +} > + > +/* > + * qcom_cpufreq_gen_params - Generate parameters to send to the hardware > + * @cpu_dev: CPU device > + * @data: SoC specific register offsets > + * @hw_tbl: Pointer to return the array of parameters > + * @apm_vc: APM Virtual Corner crossover number, returned to the caller > + * @acc_vc: MEMACC Virtual Corner crossover number, returned to the caller > + * @cpu_count: Number of CPUs in the frequency domain > + * > + * This function allocates a 'qcom_cpufreq_hw_params' parameters table, > + * fills it and returns it to the consumer, ready to get sent to the HW. > + * Since the APM threshold is just one > + * Freeing the table after usage is left to the caller. > + * > + * Returns: Number of allocated (and filled) elements in the table, > + * otherwise negative value on errors. > + */ > +static int qcom_cpufreq_gen_params(struct device *cpu_dev, > + struct qcom_cpufreq_data *data, > + struct qcom_cpufreq_hw_params **hw_tbl, > + int *apm_vc, int *acc_vc, int cpu_count) > +{ > + struct platform_device *pdev = cpufreq_get_driver_data(); > + const struct qcom_cpufreq_soc_data *soc_data = data->soc_data; > + struct device **opp_virt_dev; > + struct opp_table *genpd_opp_table; > + struct dev_pm_opp *apm_opp, *acc_opp; > + unsigned long apm_uV, acc_uV, rate; > + int i, gpd_opp_cnt, ret; > + u8 last_acc_corner = 0, prev_spare = 0; > + > + /* > + * Install the OPP table that we get from DT here already, > + * otherwise it gets really messy as we'd have to manually walk > + * through the entire OPP DT, manually find frequencies and > + * also manually exclude supported-hw (for speed-binning). > + * This also makes it easier for the genpd to add info in it. > + */ > + ret = dev_pm_opp_of_add_table(cpu_dev); > + if (ret) { > + dev_err(&pdev->dev, "Cannot install CPU OPP table: %d\n", ret); > + return ret; > + } > + > + /* Get a handle to the genpd virtual device */ > + genpd_opp_table = dev_pm_opp_attach_genpd(cpu_dev, cprh_genpd_names, > + &opp_virt_dev); > + if (IS_ERR(genpd_opp_table)) { > + ret = PTR_ERR(genpd_opp_table); > + if (ret != -EPROBE_DEFER) > + dev_err(&pdev->dev, > + "Could not attach to pm_domain: %d\n", ret); > + goto opp_dispose; > + } > + > + /* Get the count of available OPPs coming from the power domain */ > + gpd_opp_cnt = dev_pm_opp_get_opp_count(cpu_dev); > + if (gpd_opp_cnt < 2) { > + ret = gpd_opp_cnt > 0 ? -EINVAL : gpd_opp_cnt; > + goto detach_gpd; > + } > + > + /* Find the APM threshold disabled OPP and "annotate" the voltage */ > + apm_opp = dev_pm_opp_find_freq_exact(cpu_dev, 0, false); > + if (IS_ERR(apm_opp)) { > + ret = -EINVAL; > + goto detach_gpd; > + } > + > + /* If we get no APM voltage, the system is going to be unstable */ > + apm_uV = dev_pm_opp_get_voltage(apm_opp); > + if (apm_uV == 0) { > + ret = -EINVAL; > + goto detach_gpd; > + } > + *apm_vc = -1; > + > + /* > + * Find the ACC threshold disabled OPP and "annotate" the voltage: > + * this is optional, as not every SoC needs it. > + */ > + acc_opp = dev_pm_opp_find_freq_exact(cpu_dev, 1, false); > + if (IS_ERR(acc_opp)) { > + acc_uV = ULONG_MAX; > + *acc_vc = U8_MAX; > + } else { > + /* > + * If the ACC OPP is present, this means that we *have* to > + * set the relative corner on the OSM: if we cannot get the > + * voltage at that point, the CPU will inevitably freeze. > + */ > + acc_uV = dev_pm_opp_get_voltage(acc_opp); > + if (acc_uV == 0) { > + ret = -EINVAL; > + goto detach_gpd; > + } > + *acc_vc = -1; > + } > + > + *hw_tbl = devm_kmalloc_array(&pdev->dev, gpd_opp_cnt, > + sizeof(**hw_tbl), GFP_KERNEL); > + if (!hw_tbl) { > + ret = -ENOMEM; > + goto detach_gpd; > + } > + > + for (i = 0, rate = 0; ; rate++, i++) { > + struct qcom_cpufreq_hw_params *entry = *hw_tbl + i; > + struct dev_pm_opp *genpd_opp; > + struct device_node *np; > + u32 pll_div, millivolts, f_src; > + > + /* Find the next enabled OPP's frequency (ignores APM/ACC) */ > + genpd_opp = dev_pm_opp_find_freq_ceil(cpu_dev, &rate); > + if (IS_ERR(genpd_opp)) > + break; > + > + /* Get mandatory and optional properties from the OPP DT */ > + np = dev_pm_opp_get_of_node(genpd_opp); > + if (!np) > + break; > + > + if (of_property_read_u32(np, "qcom,pll-override", > + &entry->override_val)) { > + of_node_put(np); > + return -EINVAL; > + } > + > + if (of_property_read_u32(np, "qcom,spare-data", > + &entry->spare_val)) > + entry->spare_val = 0; > + > + if (of_property_read_u32(np, "qcom,pll-div", &pll_div)) > + pll_div = 0; > + > + of_node_put(np); > + > + /* Get voltage in microvolts, then convert to millivolts */ > + millivolts = dev_pm_opp_get_voltage(genpd_opp); > + if (millivolts >= apm_uV && *apm_vc < 0) > + *apm_vc = i; > + if (millivolts >= acc_uV && *acc_vc < 0) > + *acc_vc = i; > + > + millivolts /= 1000; > + > + if (millivolts < 150 || millivolts > 1400) { > + dev_err(&pdev->dev, > + "Read invalid voltage: %u.\n", millivolts); > + return -EINVAL; > + } > + > + /* In the OSM firmware, "Virtual Corner" levels start from 0 */ > + entry->volt_lut_val = FIELD_PREP(LUT_VOLT_VC, i); > + entry->volt_lut_val |= FIELD_PREP(LUT_VOLT, millivolts); > + > + /* Only the first frequency can have alternate source */ > + f_src = i ? 1 : 0; > + f_src <<= ffs(soc_data->reg_freq_lut_src_mask) - 1; > + entry->freq_lut_val = f_src | div_u64(rate, xo_rate); > + entry->freq_lut_val |= FIELD_PREP(LUT_CORE_COUNT, cpu_count); > + > + /* > + * PLL divider is not always 0 and there is no way to determine > + * it automatically, as setting this value higher than DIV1 > + * will make the OSM HW to effectively set the PLL at 2-4x > + * the CPU frequency and then divide the CPU clock by this div, > + * so this value is effectively used as both a multiplier and > + * divider. > + * This value cannot be calculated because it depends on > + * manual calibration and is (most probably) used to choose > + * a PLL frequency that gives the least possible jitter. > + */ > + entry->freq_lut_val |= FIELD_PREP(LUT_PLL_DIV, pll_div); > + > + /* > + * MEM-ACC Virtual Corner threshold voltage: set the > + * acc_thresh to the "next different" spare_val, otherwise > + * set it to an invalid value, so that we can recognize it > + * and exclude it later, when this set of data will be used > + * for programming the OSM. > + */ > + if (entry->spare_val != prev_spare) { > + prev_spare = entry->spare_val; > + entry->acc_thresh = last_acc_corner; > + } else { > + entry->acc_thresh = SEQ_MEM_ACC_MAX_LEVELS + 1; > + } > + > + dev_dbg(&pdev->dev, > + "[%d] freq=0x%x volt=0x%x override=0x%x spare=0x%x\n", > + i, entry->freq_lut_val, entry->volt_lut_val, > + entry->override_val, entry->spare_val); > + dev_pm_opp_put(genpd_opp); > + } > + > + /* > + * If we've got a customized mem-acc corner but we couldn't > + * find any suitable crossover, or the corner is less than > + * the minimum amount of required corners for mem-acc scaling, > + * the values are not valid, hence fall back to LUT values. > + */ > + if (acc_uV != ULONG_MAX && *acc_vc < SEQ_MEM_ACC_MAX_LEVELS - 1) { > + dev_dbg(&pdev->dev, > + "MEM-ACC corner: invalid values VC%d %luuV\n", > + *acc_vc, acc_uV); > + *acc_vc = U8_MAX; > + } > + > + /* > + * If we have probed less params than what we need, then the > + * OPP table that we got from the genpd is malformed for some > + * reason: in this case, do not apply the table to the HW. > + */ > + if (i < gpd_opp_cnt) { > + dev_err(&pdev->dev, "Got bad OPP table from power domain.\n"); > + ret = -EINVAL; > + goto detach_gpd; > + } > + > +detach_gpd: > + dev_pm_opp_detach_genpd(genpd_opp_table); > +opp_dispose: > + /* > + * Now that we're totally done with it, dispose of all the dynamic > + * OPPs in the table: like this, at the end of the OSM configuration > + * we are leaving it like it was magically configured by the TZ or > + * by the bootloader and using the rest of the driver as if this > + * programming phase has never happened in the OS, like new SoCs. > + * This also makes us able to not modify a single bit in the basic > + * OSM frequency request logic that was meant for the newest SoCs. > + */ > + dev_pm_opp_remove_all_dynamic(cpu_dev); > + return ret < 0 ? ret : i; > +} > + > +static inline u32 qcom_cpufreq_acd_regbit(u8 acd_reg_offset) > +{ > + return BIT((acd_reg_offset / 4)); > +} > + > +static int qcom_cpufreq_hw_acd_write_autoxfer(struct qcom_cpufreq_data *data, > + void __iomem *acd_base, u32 val) > +{ > + const struct qcom_cpufreq_soc_data *sdata = data->soc_data; > + const struct qcom_cpufreq_soc_acd_data *aregs = &sdata->acd_data; > + u32 regval; > + int retry = 3; > + > + writel_relaxed(val, acd_base + aregs->auto_xfer_cfg_reg); > + > + /* (Clear, then) Set AUTOXFER START */ > + writel_relaxed(0, acd_base + aregs->auto_xfer_reg); > + writel_relaxed(1, acd_base + aregs->auto_xfer_reg); > + > + /* Poll for status: if the first bit is set the transfer is done. */ > + do { > + regval = readl_relaxed(acd_base + aregs->auto_xfer_sts_reg); > + if (regval & BIT(0)) > + break; > + udelay(1); > + } while (--retry); > + > + if (retry <= 0) > + return -ETIMEDOUT; > + return 0; > +} > + > +static int qcom_cpufreq_hw_acd_write_xfer(struct qcom_cpufreq_data *data, > + void __iomem *acd_base, u8 reg, > + u32 val) > +{ > + const struct qcom_cpufreq_soc_data *sdata = data->soc_data; > + const struct qcom_cpufreq_soc_acd_data *aregs = &sdata->acd_data; > + int retry = 2; > + u32 regval; > + > + /* Write to the register, then initiate manual transfer */ > + writel_relaxed(val, acd_base + reg); > + > + /* > + * Make sure that the write goes through, otherwise the local > + * transfer process will wrongly copy the old value. > + */ > + wmb(); > + > + /* Clear write control register */ > + writel_relaxed(0, acd_base + aregs->write_ctl_reg); > + > + regval = (reg / 4) << ACD_WRITE_CTL_SELECT_SHIFT; > + regval |= ACD_WRITE_CTL_UPDATE_EN; > + writel_relaxed(regval, acd_base + aregs->write_ctl_reg); > + > + /* Wait until ACD Local Transfer is done */ > + do { > + regval = readl_relaxed(acd_base + aregs->write_sts_reg); > + if (regval & qcom_cpufreq_acd_regbit(reg)) > + break; > + udelay(1); > + } while (--retry); > + > + if (retry <= 0) > + return -ETIMEDOUT; > + return 0; > +} > + > +/** > + * qcom_cpufreq_hw_acd_init - Initialize ACD params in the OSM > + * @cpu_dev: CPU device > + * @policy: CPUFreq policy structure > + * @index: Instance number (CPU cluster number) > + * > + * On some SoCs it is required to send the ACD configuration parameters > + * to the OSM. This function takes the parameters from the SoC specific > + * configuration and writes them only if a "osm-acdN" iospace has been > + * declared (hence, it's present). > + * > + * Returns: Zero for success, otherwise negative number on error. > + */ > +static int qcom_cpufreq_hw_acd_init(struct device *cpu_dev, > + struct cpufreq_policy *policy, > + int index) > +{ > + struct platform_device *pdev = cpufreq_get_driver_data(); > + struct qcom_cpufreq_data *ddata = policy->driver_data; > + const struct qcom_cpufreq_soc_data *sdata = ddata->soc_data; > + const struct qcom_cpufreq_soc_acd_data *aregs = &sdata->acd_data; > + const char *acd_resname; > + void __iomem *acd_base; > + u32 rmask; > + int ret; > + > + acd_resname = kasprintf(GFP_KERNEL, "osm-acd%d", index); > + if (!acd_resname) > + return -ENOMEM; > + > + acd_base = devm_platform_ioremap_resource_byname(pdev, acd_resname); > + kfree(acd_resname); > + if (IS_ERR(acd_base)) { > + dev_vdbg(cpu_dev, "Skipping ACD initialization.\n"); > + return 0; > + } > + > + writel_relaxed(aregs->tl_delay_val, acd_base + aregs->tl_delay_reg); > + writel_relaxed(aregs->acd_ctrl_val, acd_base + aregs->acd_ctrl_reg); > + writel_relaxed(aregs->softstart_val, acd_base + aregs->softstart_reg); > + writel_relaxed(aregs->ext_intf0_val, acd_base + aregs->ext_intf_reg); > + writel_relaxed(aregs->auto_xfer_val, > + acd_base + aregs->auto_xfer_ctl_reg); > + /* > + * Before initiating register auto transfer, make sure that all > + * the writes go through, otherwise bad values will be transferred. > + */ > + wmb(); > + > + rmask = qcom_cpufreq_acd_regbit(aregs->acd_ctrl_reg) | > + qcom_cpufreq_acd_regbit(aregs->tl_delay_reg) | > + qcom_cpufreq_acd_regbit(aregs->softstart_reg) | > + qcom_cpufreq_acd_regbit(aregs->ext_intf_reg); > + ret = qcom_cpufreq_hw_acd_write_autoxfer(ddata, acd_base, rmask); > + if (ret) > + return ret; > + > + /* Switch CPUSS clock source to ACD clock */ > + ret = qcom_cpufreq_hw_acd_write_xfer(ddata, acd_base, > + aregs->gfmux_cfg_reg, 1); > + if (ret) > + return ret; > + > + /* (Set, then) Clear DCVS_SW */ > + ret = qcom_cpufreq_hw_acd_write_xfer(ddata, acd_base, > + aregs->dcvs_sw_reg, 1); > + if (ret) > + return ret; > + ret = qcom_cpufreq_hw_acd_write_xfer(ddata, acd_base, > + aregs->dcvs_sw_reg, 0); > + if (ret) > + return ret; > + > + /* Wait for clock switch time */ > + udelay(1); > + > + /* Program the final ACD external interface */ > + ret = qcom_cpufreq_hw_acd_write_xfer(ddata, acd_base, > + aregs->ext_intf_reg, > + aregs->ext_intf1_val); > + if (ret) > + return ret; > + > + /* Initiate transfer of the final ACD value */ > + rmask |= qcom_cpufreq_acd_regbit(aregs->gfmux_cfg_reg); > + writel_relaxed(rmask, acd_base + aregs->auto_xfer_cfg_reg); > + > + /* Wait for ACD to stabilize. Same wait as the OSM boot time... */ > + udelay(OSM_BOOT_TIME_US); > + return ret; > +} > + > +/** > + * qcom_cpufreq_hw_write_lut - Write Lookup Table params to the OSM > + * @cpu_dev: CPU device > + * @policy: CPUFreq policy structure > + * @cpu_count: Number of CPUs in the frequency domain > + * @index: Instance number (CPU cluster number) > + * > + * Program all the Lookup Table (LUT) entries and related thresholds > + * to the Operating State Manager on platforms where the same hasn't > + * been done already by the bootloader or TrustZone before booting > + * the operating system's kernel; > + * On these platforms, write access to the OSM is (obviously) not > + * blocked by the hypervisor. > + * > + * Returns: Zero for success, otherwise negative number on error. > + */ > +static int qcom_cpufreq_hw_write_lut(struct device *cpu_dev, > + struct cpufreq_policy *policy, > + int cpu_count, int index) > +{ > + struct platform_device *pdev = cpufreq_get_driver_data(); > + struct qcom_cpufreq_data *ddata = policy->driver_data; > + const struct qcom_cpufreq_soc_data *sdata = ddata->soc_data; > + const struct qcom_cpufreq_soc_setup_data *sregs = &sdata->setup_regs; > + struct qcom_cpufreq_hw_params *hw_tbl; > + struct resource *osm_rsrc; > + const char *osm_resname; > + u32 sreg, seq_addr, acc_lval = 0, last_spare = 1; > + int apm_vc = INT_MAX, acc_vc = U8_MAX, acc_idx = 0; > + int acc_val[SEQ_MEM_ACC_MAX_LEVELS]; > + int i, ret, num_entries; > + > + osm_resname = kasprintf(GFP_KERNEL, "osm-domain%d", index); > + if (!osm_resname) > + return -ENOMEM; > + > + /* > + * On some SoCs the OSM is not getting programmed from bootloader > + * and needs to be done here: in this case, we need to retrieve > + * the base physical address for the "Sequencer", so we will get > + * the OSM base phys and apply the sequencer offset. > + * > + * Note: We are not remapping this iospace because we are really > + * sending the physical address through SCM calls later. > + */ > + osm_rsrc = platform_get_resource_byname(pdev, IORESOURCE_MEM, > + osm_resname); > + kfree(osm_resname); > + > + if (!osm_rsrc) > + return -ENODEV; > + > + seq_addr = osm_rsrc->start + sregs->reg_osm_sequencer; > + > + num_entries = qcom_cpufreq_gen_params(cpu_dev, ddata, &hw_tbl, > + &apm_vc, &acc_vc, cpu_count); > + if (num_entries < 0) > + return num_entries; > + > + for (i = 0; i < LUT_MAX_ENTRIES; i++) { > + struct qcom_cpufreq_hw_params *entry; > + int pos = i * sdata->lut_row_size; > + > + /* > + * If we have reached the end of the params table, write > + * the last valid entry until the end of the OSM table. > + */ > + if (i < num_entries) > + entry = &hw_tbl[i]; > + else > + entry = &hw_tbl[num_entries - 1]; > + > + writel_relaxed(i, ddata->base + sdata->reg_index + pos); > + > + writel_relaxed(entry->volt_lut_val, > + ddata->base + sdata->reg_volt_lut + pos); > + > + writel_relaxed(entry->freq_lut_val, > + ddata->base + sdata->reg_freq_lut + pos); > + > + writel_relaxed(entry->override_val, > + ddata->base + sregs->reg_override + pos); > + > + writel_relaxed(entry->spare_val, > + ddata->base + sregs->reg_spare + pos); > + > + dev_dbg(cpu_dev, > + "Writing [%d] v:0x%x f:0x%x ovr:0x%x s:0x%x\n", i, > + entry->volt_lut_val, entry->freq_lut_val, > + entry->override_val, entry->spare_val); > + > + /* > + * MEM-ACC Virtual Corner threshold voltage: this gets set > + * as the pairs of corners in which there is a transition > + * between one MEM-ACC level and the next one. > + * > + * Notes: The spare_val can never be zero; > + * The first spare_val is always 1; > + * The maximum number of pairs is two (four registers). > + * > + * Example: (C = Corner Level - M = MEM-ACC Level) > + * C0 M1 - C1 M1 - C2 M2 - C3 M2 - C4 M2 - C5 M3 > + * Pairs: 1-2, 4-5 > + */ > + if (entry->spare_val <= last_spare || > + acc_idx >= SEQ_MEM_ACC_MAX_LEVELS - 1) > + continue; > + > + /* Standard mem-acc pairs using spare_val LUT crossovers */ > + last_spare = entry->spare_val; > + acc_val[acc_idx] = i - 1; > + acc_idx++; > + acc_val[acc_idx] = i; > + acc_idx++; > + } > + > + /* Sanity check: we *must* have two mem-acc crossovers (four values) */ > + if (acc_idx < SEQ_MEM_ACC_MAX_LEVELS - 1) > + return -EINVAL; > + > + /* > + * Customized mem-acc corners, if any; in this case, the last corner > + * in the external (CPRh) LUT is this one, placed after the APM one. > + */ > + if (acc_vc != U8_MAX) { > + sreg = SEQUENCER_REG(seq_addr, SEQ_MEM_ACC_CROSSOVER_VC); > + ret = qcom_scm_io_writel(sreg, num_entries + 1); > + if (ret) > + return ret; > + > + /* > + * At the price of very-slightly higher power consumption, > + * switch the ACC at one corner lower than what we've found, > + * as this seems to be needed on at least some MSM8998 chips > + * to achieve full system stability > + */ > + acc_vc--; > + > + /* Change only if we have to move the corner down */ > + if (acc_vc < acc_val[3]) { > + acc_val[2] = acc_vc - 1; > + acc_val[3] = acc_vc; > + } > + > + /* If needed, sanitize previously stored vals from the LUT */ > + if (acc_val[2] <= acc_val[1]) > + acc_val[1] = acc_val[2] - 1; > + if (acc_val[1] <= acc_val[0]) > + acc_val[0] = acc_val[1] - 1; > + } > + > + for (i = 0; i < SEQ_MEM_ACC_MAX_LEVELS; i++) { > + ret = qcom_scm_io_writel(SEQ_MEMACC_REG(seq_addr, i), > + acc_val[i]); > + if (ret) > + return ret; > + } > + dev_dbg(cpu_dev, "Wrote MEM-ACC Pairs: [%u-%u] [%u-%u]\n", > + acc_val[0], acc_val[1], acc_val[2], acc_val[3]); > + > + /* > + * Program the L_VAL of the first corner requesting MEM-ACC > + * voltage level 3 to the right sequencer register > + */ > + acc_lval = FIELD_GET(LUT_L_VAL, hw_tbl[acc_val[3]].freq_lut_val); > + ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_MEM_ACC_LVAL), > + acc_lval); > + if (ret) { > + dev_dbg(cpu_dev, "Cannot send memacc l_val\n"); > + return ret; > + } > + dev_dbg(cpu_dev, "MEM-ACC L-Val is %u\n", acc_lval); > + > + /* > + * Array Power Mux threshold level: the first virtual corner > + * that requires a switch sequence of the APM from MX to APC. > + */ > + if (apm_vc == INT_MAX) > + apm_vc = LUT_MAX_ENTRIES - 1; > + > + /* > + * APM crossover virtual corner refers to CPRh: there, the APM corner > + * is always appended to the table (so, at the end of it, right after > + * the cluster dvfs entries). > + */ > + writel(num_entries, ddata->base + sregs->reg_seq1); > + ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_APM_CROSSOVER_VC), > + num_entries); > + if (ret) > + return ret; > + > + ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_APM_THRESH_VC), > + apm_vc); > + if (ret) > + return ret; > + > + ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_APM_THRESH_PREVC), > + apm_vc - 1); > + if (ret) > + return ret; > + > + ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_APM_PARAM), > + (0x39 | apm_vc << 6)); > + if (ret) > + return ret; > + > + dev_dbg(cpu_dev, "Wrote APM Pair: [%u-%u]\n", apm_vc - 1, apm_vc); > + return ret; > +} > + > +/** > + * qcom_cpufreq_hw_read_lut - Read Lookup Table from the OSM > + * @cpu_dev: CPU device > + * @policy: CPUFreq policy structure > + * > + * The Operating State Manager Lookup Table can always be read, even > + * in case it was pre-programmed by the bootloader or by TrustZone. > + * Read the LUT from it in order to build OPPs containing DVFS info. > + * > + * Returns: Zero for success, otherwise negative number on errors. > + */ > static int qcom_cpufreq_hw_read_lut(struct device *cpu_dev, > struct cpufreq_policy *policy) > { > @@ -166,7 +1066,9 @@ static int qcom_cpufreq_hw_read_lut(struct device *cpu_dev, > for (i = 0; i < LUT_MAX_ENTRIES; i++) { > data = readl_relaxed(drv_data->base + soc_data->reg_freq_lut + > i * soc_data->lut_row_size); > - src = FIELD_GET(LUT_SRC, data); > + src = data & soc_data->reg_freq_lut_src_mask; > + src >>= ffs(soc_data->reg_freq_lut_src_mask) - 1; > + > lval = FIELD_GET(LUT_L_VAL, data); > core_count = FIELD_GET(LUT_CORE_COUNT, data); > > @@ -175,17 +1077,21 @@ static int qcom_cpufreq_hw_read_lut(struct device *cpu_dev, > volt = FIELD_GET(LUT_VOLT, data) * 1000; > > if (src) > - freq = xo_rate * lval / 1000; > + freq = xo_rate * lval; > else > - freq = cpu_hw_rate / 1000; > + freq = cpu_hw_rate; > + freq /= 1000; > > if (freq != prev_freq && core_count != LUT_TURBO_IND) { > if (!qcom_cpufreq_update_opp(cpu_dev, freq, volt)) { > table[i].frequency = freq; > - dev_dbg(cpu_dev, "index=%d freq=%d, core_count %d\n", i, > - freq, core_count); > + dev_dbg(cpu_dev, > + "index=%d freq=%d, core_count %d\n", > + i, freq, core_count); > } else { > - dev_warn(cpu_dev, "failed to update OPP for freq=%d\n", freq); > + dev_warn(cpu_dev, > + "failed to update OPP for freq=%d\n", > + freq); > table[i].frequency = CPUFREQ_ENTRY_INVALID; > } > > @@ -205,18 +1111,19 @@ static int qcom_cpufreq_hw_read_lut(struct device *cpu_dev, > * as the boost frequency > */ > if (prev->frequency == CPUFREQ_ENTRY_INVALID) { > - if (!qcom_cpufreq_update_opp(cpu_dev, prev_freq, volt)) { > + if (!qcom_cpufreq_update_opp(cpu_dev, prev_freq, > + volt)) { > prev->frequency = prev_freq; > prev->flags = CPUFREQ_BOOST_FREQ; > } else { > - dev_warn(cpu_dev, "failed to update OPP for freq=%d\n", > + dev_warn(cpu_dev, > + "can't update OPP for freq=%u\n", > freq); > } > } > > break; > } > - > prev_freq = freq; > } > > @@ -227,10 +1134,18 @@ static int qcom_cpufreq_hw_read_lut(struct device *cpu_dev, > return 0; > } > > -static void qcom_get_related_cpus(int index, struct cpumask *m) > +/* > + * qcom_get_related_cpus - Get mask of CPUs in the same frequency domain > + * @index: CPU number > + * @m: Returned CPU mask > + * > + * Returns: Count of CPUs inserted in the cpumask or negative number for error. > + */ > +static int qcom_get_related_cpus(int index, struct cpumask *m) > { > struct device_node *cpu_np; > struct of_phandle_args args; > + int count = 0; > int cpu, ret; > > for_each_possible_cpu(cpu) { > @@ -245,34 +1160,237 @@ static void qcom_get_related_cpus(int index, struct cpumask *m) > if (ret < 0) > continue; > > - if (index == args.args[0]) > + if (index == args.args[0]) { > cpumask_set_cpu(cpu, m); > + count++; > + } > } > + > + return count > 0 ? count : -EINVAL; > } > > static const struct qcom_cpufreq_soc_data qcom_soc_data = { > .reg_enable = 0x0, > .reg_freq_lut = 0x110, > + .reg_freq_lut_src_mask = LUT_SRC_845, > .reg_volt_lut = 0x114, > .reg_perf_state = 0x920, > .lut_row_size = 32, > + .clk_hw_div = 2, > + .uses_tz = true, > +}; > + > +static const struct qcom_cpufreq_soc_data msm8998_soc_data = { > + .reg_enable = 0x4, > + .reg_index = 0x150, > + .reg_freq_lut = 0x154, > + .reg_freq_lut_src_mask = LUT_SRC_8998, > + .reg_volt_lut = 0x158, > + .reg_perf_state = 0xf10, > + .lut_row_size = 32, > + .clk_hw_div = 1, > + .uses_tz = false, > + .setup_regs = { > + /* Physical offset for sequencer scm calls */ > + .reg_osm_sequencer = 0x300, > + > + /* Frequency domain offsets */ > + .reg_override = 0x15c, > + .reg_spare = 0x164, > + .reg_cc_zero_behav = 0x0c, > + .reg_spm_cc_hyst = 0x1c, > + .reg_spm_cc_dcvs_dis = 0x20, > + .reg_spm_core_ret_map = 0x24, > + .reg_llm_freq_vote_hyst = 0x2c, > + .reg_llm_volt_vote_hyst = 0x30, > + .reg_llm_intf_dcvs_dis = 0x34, > + .reg_seq1 = 0x48, > + .reg_pdn_fsm_ctrl = 0x70, > + .reg_cc_boost_timer = 0x74, > + .reg_dcvs_boost_timer = 0x84, > + .reg_ps_boost_timer = 0x94, > + .boost_timer_reg_len = 0x4, > + .reg_boost_sync_delay = 0xa0, > + .reg_droop_ctrl = 0xa4, > + .reg_droop_release_ctrl = 0xa8, > + .reg_droop_unstall_ctrl = 0xac, > + .reg_droop_wait_release_ctrl = 0xb0, > + .reg_droop_timer_ctrl = 0xb8, > + .reg_droop_sync_delay = 0xbc, > + .reg_pll_override = 0xc0, > + .reg_cycle_counter = 0xf00, > + }, > + .acd_data = { > + .acd_ctrl_reg = 0x4, > + .tl_delay_reg = 0x8, > + .softstart_reg = 0x28, > + .ext_intf_reg = 0x30, > + .dcvs_sw_reg = 0x34, > + .gfmux_cfg_reg = 0x3c, > + .auto_xfer_cfg_reg = 0x80, > + .auto_xfer_reg = 0x84, > + .auto_xfer_ctl_reg = 0x88, > + .auto_xfer_sts_reg = 0x8c, > + .write_ctl_reg = 0x90, > + .write_sts_reg = 0x94, > + .tl_delay_val = 38417, > + .acd_ctrl_val = 0x2b5ffd, > + .softstart_val = 0x501, > + .ext_intf0_val = 0x2cf9ae8, > + .ext_intf1_val = 0x2cf9afe, > + .auto_xfer_val = 0x15, > + }, > }; > > static const struct qcom_cpufreq_soc_data epss_soc_data = { > .reg_enable = 0x0, > .reg_freq_lut = 0x100, > + .reg_freq_lut_src_mask = LUT_SRC_845, > .reg_volt_lut = 0x200, > .reg_perf_state = 0x320, > .lut_row_size = 4, > + .clk_hw_div = 2, > + .uses_tz = true, > }; > > static const struct of_device_id qcom_cpufreq_hw_match[] = { > { .compatible = "qcom,cpufreq-hw", .data = &qcom_soc_data }, > + { .compatible = "qcom,cpufreq-hw-8998", .data = &msm8998_soc_data }, > { .compatible = "qcom,cpufreq-epss", .data = &epss_soc_data }, > {} > }; > MODULE_DEVICE_TABLE(of, qcom_cpufreq_hw_match); > > +/** > + * qcom_cpufreq_hw_osm_setup - Setup and enable the OSM > + * @cpu_dev: CPU device > + * @policy: CPUFreq policy structure > + * @cpu_count: Number of CPUs in the frequency domain > + * > + * On some platforms, the Operating State Manager (OSM) is not getting > + * programmed by the bootloader, nor by TrustZone before booting the OS > + * and its register space is not write-protected by the hypervisor. > + * In this case, to achieve CPU DVFS, it is needed to program it from > + * the OS itself, which includes setting LUT and all the various tunables > + * that are required for it to manage the CPU frequencies and voltages > + * on its own. > + * Calling this function on a platform that had the OSM set-up by TZ > + * will result in a hypervisor fault with system reboot in most cases. > + * > + * Returns: Zero for success, otherwise negative number on errors. > + */ > +static int qcom_cpufreq_hw_osm_setup(struct device *cpu_dev, > + struct cpufreq_policy *policy, > + int cpu_count, int index) > +{ > + struct qcom_cpufreq_data *drv_data = policy->driver_data; > + const struct qcom_cpufreq_soc_setup_data *setup_regs; > + u32 val; > + int ret; > + > + ret = qcom_cpufreq_hw_write_lut(cpu_dev, policy, cpu_count, index); > + if (ret) > + return ret; > + > + setup_regs = &drv_data->soc_data->setup_regs; > + > + /* Set OSM to XO clock ratio and use XO edge for the cycle counter */ > + val = FIELD_PREP(CYCLE_COUNTER_CLK_RATIO, OSM_XO_RATIO_VAL); > + val |= CYCLE_COUNTER_USE_XO_EDGE; > + > + /* Enable the CPU cycle counter */ > + val |= BIT(0); > + writel_relaxed(val, drv_data->base + setup_regs->reg_cycle_counter); > + > + /* CoreCount DCVS Policy: Wait time for frequency inc/decrement */ > + val = FIELD_PREP(HYSTERESIS_UP_MASK, HYSTERESIS_CC_NS); > + val |= FIELD_PREP(HYSTERESIS_DN_MASK, HYSTERESIS_CC_NS); > + writel_relaxed(val, drv_data->base + setup_regs->reg_spm_cc_hyst); > + > + /* Set the frequency index 0 and override for cluster power collapse */ > + writel_relaxed(BIT(0), drv_data->base + setup_regs->reg_cc_zero_behav); > + > + /* Treat cores in retention as active */ > + writel_relaxed(0, drv_data->base + setup_regs->reg_spm_core_ret_map); > + > + /* Enable CoreCount based DCVS */ > + writel_relaxed(0, drv_data->base + setup_regs->reg_spm_cc_dcvs_dis); > + > + /* CoreCount DCVS-LLM Policy: Wait time for frequency inc/decrement */ > + val = FIELD_PREP(HYSTERESIS_UP_MASK, HYSTERESIS_LLM_NS); > + val |= FIELD_PREP(HYSTERESIS_DN_MASK, HYSTERESIS_LLM_NS); > + writel_relaxed(val, drv_data->base + setup_regs->reg_llm_freq_vote_hyst); > + > + /* CoreCount DCVS-LLM Policy: Wait time for voltage inc/decrement */ > + val = FIELD_PREP(HYSTERESIS_UP_MASK, HYSTERESIS_LLM_NS); > + val |= FIELD_PREP(HYSTERESIS_DN_MASK, HYSTERESIS_LLM_NS); > + writel_relaxed(val, drv_data->base + setup_regs->reg_llm_volt_vote_hyst); > + > + /* Enable LLM frequency+voltage voting */ > + writel_relaxed(0, drv_data->base + setup_regs->reg_llm_intf_dcvs_dis); > + > + /* Setup Boost FSM Timers */ > + qcom_cpufreq_hw_boost_setup(drv_data->base + > + setup_regs->reg_cc_boost_timer, > + setup_regs->boost_timer_reg_len); > + qcom_cpufreq_hw_boost_setup(drv_data->base + > + setup_regs->reg_dcvs_boost_timer, > + setup_regs->boost_timer_reg_len); > + qcom_cpufreq_hw_boost_setup(drv_data->base + > + setup_regs->reg_ps_boost_timer, > + setup_regs->boost_timer_reg_len); > + > + /* PLL signal timing control for Boost */ > + writel_relaxed(BOOST_SYNC_DELAY, > + drv_data->base + setup_regs->reg_boost_sync_delay); > + > + /* Setup WFx and PC/RET droop unstall */ > + val = FIELD_PREP(DROOP_TIMER1, DROOP_TIMER_NS); > + val |= FIELD_PREP(DROOP_TIMER0, DROOP_TIMER_NS); > + writel_relaxed(val, drv_data->base + setup_regs->reg_droop_unstall_ctrl); > + > + /* Setup WFx and PC/RET droop wait-to-release */ > + val = FIELD_PREP(DROOP_TIMER1, DROOP_WAIT_RELEASE_TIMER_NS); > + val |= FIELD_PREP(DROOP_TIMER0, DROOP_WAIT_RELEASE_TIMER_NS); > + writel_relaxed(val, > + drv_data->base + setup_regs->reg_droop_wait_release_ctrl); > + > + /* PLL signal timing control for Droop */ > + writel_relaxed(1, drv_data->base + setup_regs->reg_droop_sync_delay); > + > + /* Setup DCVS timers */ > + writel_relaxed(DROOP_RELEASE_TIMER_NS, > + drv_data->base + setup_regs->reg_droop_release_ctrl); > + writel_relaxed(DROOP_TIMER_NS, > + drv_data->base + setup_regs->reg_droop_timer_ctrl); > + > + /* Setup Droop control */ > + val = readl_relaxed(drv_data->base + setup_regs->reg_droop_ctrl); > + val |= DROOP_CTRL_VAL; > + writel_relaxed(val, drv_data->base + setup_regs->reg_droop_ctrl); > + > + /* Enable CC-Boost, DCVS-Boost, PS-Boost, WFx, PC/RET, DCVS FSM */ > + val = readl_relaxed(drv_data->base + setup_regs->reg_pdn_fsm_ctrl); > + val |= CC_BOOST_EN | PS_BOOST_EN | DCVS_BOOST_EN; > + val |= WFX_DROOP_EN | PC_RET_EXIT_DROOP_EN | DCVS_DROOP_EN; > + writel_relaxed(val, drv_data->base + setup_regs->reg_pdn_fsm_ctrl); > + > + /* Enable PLL Droop Override */ > + val = PLL_OVERRIDE_DROOP_EN; > + writel_relaxed(val, drv_data->base + setup_regs->reg_pll_override); > + > + /* Initialize the Adaptive Clock Distribution */ > + ret = qcom_cpufreq_hw_acd_init(cpu_dev, policy, index); > + if (ret) > + return ret; > + > + /* We're ready: enable the OSM and give it time to boot (5uS) */ > + writel_relaxed(1, drv_data->base + drv_data->soc_data->reg_enable); > + udelay(OSM_BOOT_TIME_US); > + > + return ret; > +} > + > static int qcom_cpufreq_hw_cpu_init(struct cpufreq_policy *policy) > { > struct platform_device *pdev = cpufreq_get_driver_data(); > @@ -282,7 +1400,8 @@ static int qcom_cpufreq_hw_cpu_init(struct cpufreq_policy *policy) > struct device *cpu_dev; > void __iomem *base; > struct qcom_cpufreq_data *data; > - int ret, index; > + const char *fdom_resname; > + int cpu_count, index, ret; > > cpu_dev = get_cpu_device(policy->cpu); > if (!cpu_dev) { > @@ -303,9 +1422,14 @@ static int qcom_cpufreq_hw_cpu_init(struct cpufreq_policy *policy) > > index = args.args[0]; > > - base = devm_platform_ioremap_resource(pdev, index); > + fdom_resname = kasprintf(GFP_KERNEL, "freq-domain%d", index); > + if (!fdom_resname) > + return -ENOMEM; > + > + base = devm_platform_ioremap_resource_byname(pdev, fdom_resname); > if (IS_ERR(base)) > return PTR_ERR(base); > + kfree(fdom_resname); > > data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL); > if (!data) { > @@ -315,22 +1439,31 @@ static int qcom_cpufreq_hw_cpu_init(struct cpufreq_policy *policy) > > data->soc_data = of_device_get_match_data(&pdev->dev); > data->base = base; > + policy->driver_data = data; > > - /* HW should be in enabled state to proceed */ > - if (!(readl_relaxed(base + data->soc_data->reg_enable) & 0x1)) { > - dev_err(dev, "Domain-%d cpufreq hardware not enabled\n", index); > - ret = -ENODEV; > - goto error; > - } > - > - qcom_get_related_cpus(index, policy->cpus); > + cpu_count = qcom_get_related_cpus(index, policy->cpus); > if (!cpumask_weight(policy->cpus)) { > dev_err(dev, "Domain-%d failed to get related CPUs\n", index); > ret = -ENOENT; > goto error; > } > > - policy->driver_data = data; > + if (!data->soc_data->uses_tz) { > + ret = qcom_cpufreq_hw_osm_setup(cpu_dev, policy, > + cpu_count, index); > + if (ret) { > + dev_err(dev, "Cannot setup the OSM for CPU%d: %d\n", > + policy->cpu, ret); > + goto error; > + } > + } > + > + /* HW should be in enabled state to proceed */ > + if (!(readl_relaxed(base + data->soc_data->reg_enable) & 0x1)) { > + dev_err(dev, "Domain-%d cpufreq hardware not enabled\n", index); > + ret = -ENODEV; > + goto error; > + } > > ret = qcom_cpufreq_hw_read_lut(cpu_dev, policy); > if (ret) { > @@ -355,6 +1488,7 @@ static int qcom_cpufreq_hw_cpu_init(struct cpufreq_policy *policy) > > return 0; > error: > + policy->driver_data = NULL; > devm_iounmap(dev, base); > return ret; > } > @@ -395,9 +1529,51 @@ static struct cpufreq_driver cpufreq_qcom_hw_driver = { > > static int qcom_cpufreq_hw_driver_probe(struct platform_device *pdev) > { > + const struct qcom_cpufreq_soc_data *soc_data; > + struct device_node *pd_node; > + struct platform_device *pd_dev; > struct device *cpu_dev; > struct clk *clk; > - int ret; > + int clk_div, ret; > + > + cpu_dev = get_cpu_device(0); > + if (!cpu_dev) > + return -EPROBE_DEFER; > + > + soc_data = of_device_get_match_data(&pdev->dev); > + if (!soc_data) > + return -EINVAL; > + > + if (!soc_data->uses_tz) { > + /* > + * When the OSM is not pre-programmed from TZ, we will > + * need to program the sequencer through SCM calls. > + */ > + if (!qcom_scm_is_available()) > + return -EPROBE_DEFER; > + > + /* > + * If there are no power-domains, OSM programming cannot be > + * performed, as in that case, we wouldn't know where to take > + * the params from... > + */ > + pd_node = of_parse_phandle(cpu_dev->of_node, > + "power-domains", 0); > + if (!pd_node) { > + ret = PTR_ERR(pd_node); > + dev_err(cpu_dev, "power domain not found: %d\n", ret); > + return ret; > + } > + > + /* > + * If the power domain device is not registered yet, then > + * defer probing this driver until that is available. > + */ > + pd_dev = of_find_device_by_node(pd_node); > + if (!pd_dev || !pd_dev->dev.driver || > + !device_is_bound(&pd_dev->dev)) > + return -EPROBE_DEFER; > + } > > clk = clk_get(&pdev->dev, "xo"); > if (IS_ERR(clk)) > @@ -410,16 +1586,16 @@ static int qcom_cpufreq_hw_driver_probe(struct platform_device *pdev) > if (IS_ERR(clk)) > return PTR_ERR(clk); > > - cpu_hw_rate = clk_get_rate(clk) / CLK_HW_DIV; > + clk_div = soc_data->clk_hw_div; > + if (clk_div == 0) > + clk_div++; > + > + cpu_hw_rate = clk_get_rate(clk) / clk_div; > clk_put(clk); > > cpufreq_qcom_hw_driver.driver_data = pdev; > > /* Check for optional interconnect paths on CPU0 */ > - cpu_dev = get_cpu_device(0); > - if (!cpu_dev) > - return -EPROBE_DEFER; > - > ret = dev_pm_opp_of_find_icc_paths(cpu_dev, NULL); > if (ret) > return ret; > -- QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, hosted by The Linux Foundation. --
Il 20/01/21 19:25, Taniya Das ha scritto: > The CPUFREQ-HW driver is intended to be used only for CPUFREQ HW designs > where the firmware programs the look up tables. > > Suggestion is to separate out the driver where the programming is > managed by high level OS. > The CPUFREQ HW designs are effectively Operating State Manager, with CPR-Hardened, SAWv4.x (etc) and they basically are "the same" between one another. For how I see it, since this code is made to drive the OSM states, this driver is simply incomplete in the current state, as it misses the ability to program the hardware in case it wasn't programmed by TZ, which is effectively the case with more than just one SoC. In any case, I think that the "layer of separation" between a OSM that was pre-programmed by the TZ/bootloader (anything pre-HLOS) and one that needs to be programmed by, in this case, Linux, is pretty much evident: for how I have engineered the code, the already existant part of the driver is *untouched*. One bool variable is checked for the OSM programming, if that's true then it will do the entire programming - otherwise, the flow will be like nothing was ever added. Separating this out as another driver, managing the same hardware, in the same iospace(s), would only be confusing for developers, making them think that we're dealing with "a different piece", or something that is assisting "that chip" (oversimplifying, please don't be picky), or something that is "special" while, in reality... it's.. not. It's just the same HW, after all. > On 1/19/2021 11:15 PM, AngeloGioacchino Del Regno wrote: >> On new SoCs (SDM845 onwards) the Operating State Manager (OSM) is >> being programmed in the bootloader and write-protected by the >> hypervisor, leaving to the OS read-only access to some of its >> registers (in order to read the Lookup Tables and also some >> status registers) and write access to the p-state register, for >> for the OS to request a specific performance state to trigger a >> DVFS switch on the CPU through the OSM hardware. >> >> On old SoCs though (MSM8998, SDM630/660 and variants), the >> bootloader will *not* initialize the OSM (and the CPRh, as it >> is a requirement for it) before booting the OS, making any >> request to trigger a performance state change ineffective, as >> the hardware doesn't have any Lookup Table, nor is storing any >> parameter to trigger a DVFS switch. In this case, basically all >> of the OSM registers are *not* write protected for the OS, even >> though some are - but write access is granted through SCM calls. >> >> This commit introduces support for OSM programming, which has to >> be done on these old SoCs that were distributed (almost?) always >> with a bootloader that does not do any CPRh nor OSM init before >> booting the kernel. >> In order to program the OSM on these SoCs, it is necessary to >> fullfill a "special" requirement: the Core Power Reduction >> Hardened (CPRh) hardware block must be initialized, as the OSM >> is "talking" to it in order to perform the Voltage part of DVFS; >> here, we are calling initialization of this through Linux generic >> power domains, specifically by requesting a genpd attach from the >> qcom-cpufreq-hw driver, which will give back voltages associated >> to each CPU frequency that has been declared in the OPPs, scaled >> and interpolated with the previous one, and will also give us >> parameters for the Array Power Mux (APM) and mem-acc, in order >> for this driver to be then able to generate the Lookup Tables >> that will be finally programmed to the OSM hardware. >> >> After writing the parameters to the OSM and enabling it, all the >> programming work will never happen anymore until a OS reboot, so >> all of the allocations and "the rest" will be disposed-of: this >> is done mainly to leave the code that was referred only to the >> new SoCs intact, as to also emphasize on the fact that the OSM >> HW is, in the end, the exact same; apart some register offsets >> that are slightly different, the entire logic is the same. >> >> This also adds the parameters to support CPU scaling on SDM630 >> and MSM8998. >> >> Signed-off-by: AngeloGioacchino Del Regno >> <angelogioacchino.delregno@somainline.org> >> --- >> Â drivers/cpufreq/qcom-cpufreq-hw.c | 1240 ++++++++++++++++++++++++++++- >> Â 1 file changed, 1208 insertions(+), 32 deletions(-) >> >> diff --git a/drivers/cpufreq/qcom-cpufreq-hw.c >> b/drivers/cpufreq/qcom-cpufreq-hw.c >> index acc645b85e79..a92959bb7b50 100644 >> --- a/drivers/cpufreq/qcom-cpufreq-hw.c >> +++ b/drivers/cpufreq/qcom-cpufreq-hw.c >> @@ -1,33 +1,256 @@ >> Â // SPDX-License-Identifier: GPL-2.0 >> Â /* >> Â Â * Copyright (c) 2018, The Linux Foundation. All rights reserved. >> + * >> + * OSM hardware initial programming >> + * Copyright (C) 2020, AngeloGioacchino Del Regno >> + *Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â <angelogioacchino.delregno@somainline.org> >> Â Â */ >> Â #include <linux/bitfield.h> >> Â #include <linux/cpufreq.h> >> +#include <linux/delay.h> >> Â #include <linux/init.h> >> Â #include <linux/interconnect.h> >> Â #include <linux/kernel.h> >> Â #include <linux/module.h> >> Â #include <linux/of_address.h> >> Â #include <linux/of_platform.h> >> +#include <linux/pm_domain.h> >> Â #include <linux/pm_opp.h> >> Â #include <linux/slab.h> >> +#include <linux/qcom_scm.h> >> Â #define LUT_MAX_ENTRIESÂ Â Â Â Â Â Â Â Â Â Â 40U >> -#define LUT_SRCÂ Â Â Â Â Â Â Â Â Â Â Â Â Â Â GENMASK(31, 30) >> +#define LUT_SRC_845Â Â Â Â Â Â Â Â Â Â Â GENMASK(31, 30) >> +#define LUT_SRC_8998Â Â Â Â Â Â Â Â Â Â Â GENMASK(27, 26) >> +#define LUT_PLL_DIVÂ Â Â Â Â Â Â Â Â Â Â GENMASK(25, 24) >> Â #define LUT_L_VALÂ Â Â Â Â Â Â Â Â Â Â GENMASK(7, 0) >> Â #define LUT_CORE_COUNTÂ Â Â Â Â Â Â Â Â Â Â GENMASK(18, 16) >> +#define LUT_VOLT_VCÂ Â Â Â Â Â Â Â Â Â Â GENMASK(21, 16) >> Â #define LUT_VOLTÂ Â Â Â Â Â Â Â Â Â Â GENMASK(11, 0) >> -#define CLK_HW_DIVÂ Â Â Â Â Â Â Â Â Â Â 2 >> Â #define LUT_TURBO_INDÂ Â Â Â Â Â Â Â Â Â Â 1 >> +#define OSM_BOOT_TIME_USÂ Â Â Â Â Â Â 5 >> + >> +#define CYCLE_COUNTER_CLK_RATIOÂ Â Â Â Â Â Â GENMASK(5, 1) >> +#define OSM_XO_RATIO_VALÂ Â Â Â Â Â Â (10 - 1) >> +#define CYCLE_COUNTER_USE_XO_EDGEÂ Â Â BIT(8) >> + >> +/* FSM Boost Control */ >> +#define CC_BOOST_ENÂ Â Â Â Â Â Â Â Â Â Â BIT(0) >> +#define PS_BOOST_ENÂ Â Â Â Â Â Â Â Â Â Â BIT(1) >> +#define DCVS_BOOST_ENÂ Â Â Â Â Â Â Â Â Â Â BIT(2) >> +#define BOOST_TIMER_REG_HIÂ Â Â Â Â Â Â GENMASK(31, 16) >> +#define BOOST_TIMER_REG_LOÂ Â Â Â Â Â Â GENMASK(15, 0) >> + >> +#define PLL_WAIT_LOCK_TIME_NSÂ Â Â Â Â Â Â 2000 >> +#define SAFE_FREQ_WAIT_NSÂ Â Â Â Â Â Â 1000 >> +#define DEXT_DECREMENT_WAIT_NSÂ Â Â Â Â Â Â 200 >> + >> +#define BOOST_SYNC_DELAYÂ Â Â Â Â Â Â 5 >> + >> +#define HYSTERESIS_UP_MASKÂ Â Â Â Â Â Â GENMASK(31, 16) >> +#define HYSTERESIS_DN_MASKÂ Â Â Â Â Â Â GENMASK(15, 0) >> +#define HYSTERESIS_CC_NSÂ Â Â Â Â Â Â 200 >> +#define HYSTERESIS_LLM_NSÂ Â Â Â Â Â Â 65535 >> + >> +/* FSM Droop Control */ >> +#define PC_RET_EXIT_DROOP_ENÂ Â Â Â Â Â Â BIT(3) >> +#define WFX_DROOP_ENÂ Â Â Â Â Â Â Â Â Â Â BIT(4) >> +#define DCVS_DROOP_ENÂ Â Â Â Â Â Â Â Â Â Â BIT(5) >> +#define DROOP_TIMER1Â Â Â Â Â Â Â Â Â Â Â GENMASK(31, 16) >> +#define DROOP_TIMER0Â Â Â Â Â Â Â Â Â Â Â GENMASK(15, 0) >> +#define DROOP_CTRL_VALÂ Â Â Â Â Â Â Â Â Â Â (BIT(3) | BIT(17) | BIT(31)) >> +#define DROOP_TIMER_NSÂ Â Â Â Â Â Â Â Â Â Â 100 >> +#define DROOP_WAIT_RELEASE_TIMER_NSÂ Â Â 50 >> +#define DROOP_RELEASE_TIMER_NSÂ Â Â Â Â Â Â 1 >> + >> +/* PLL Override Control */ >> +#define PLL_OVERRIDE_DROOP_ENÂ Â Â Â Â Â Â BIT(0) >> + >> +/* Sequencer */ >> +#define SEQUENCER_REG(base, n)Â Â Â Â Â Â Â (base + (n * 4)) >> +#define SEQ_APM_THRESH_VCÂ Â Â Â Â Â Â 15 >> +#define SEQ_APM_THRESH_PREVCÂ Â Â Â Â Â Â 31 >> +#define SEQ_MEM_ACC_LVALÂ Â Â Â Â Â Â 32 >> +#define SEQ_MEM_ACC_0Â Â Â Â Â Â Â Â Â Â Â 55 >> +#define SEQ_APM_CROSSOVER_VCÂ Â Â Â Â Â Â 72 >> +#define SEQ_APM_PARAMÂ Â Â Â Â Â Â Â Â Â Â 76 >> +#define SEQ_MEM_ACC_CROSSOVER_VCÂ Â Â 88 >> +#define SEQ_MEM_ACC_MAX_LEVELSÂ Â Â Â Â Â Â 4 >> +#define SEQ_MEMACC_REG(base, n)Â Â Â Â Â Â Â SEQUENCER_REG(base, >> SEQ_MEM_ACC_0 + n) >> + >> +/* ACD */ >> +#define ACD_WRITE_CTL_UPDATE_ENÂ Â Â Â Â Â Â BIT(0) >> +#define ACD_WRITE_CTL_SELECT_SHIFTÂ Â Â 1 >> + >> +/** >> + * struct qcom_cpufreq_soc_setup_data - Register offsets for OSM setup >> + * >> + * @reg_osm_sequencer:Â Â Â Â Â OSM Sequencer (used to get physical address) >> + * @reg_override:Â Â Â Â Â Â Â Â Â Â Override parameters >> + * @reg_spare:Â Â Â Â Â Â Â Â Â Â Â Â Â Spare parameters (MEMACC-to-VC) >> + * @reg_cc_zero_behav:Â Â Â Â Â Virtual Corner for cluster power collapse >> + * @reg_spm_cc_hyst:Â Â Â Â Â Â Â DCVS-CC Wait time for frequency >> inc/decrement >> + * @reg_spm_cc_dcvs_dis:Â Â Â DCVS-CC en/disable control >> + * @reg_spm_core_ret_map:Â Â Treat cores in retention as active/inactive >> + * @reg_llm_freq_vote_hyst: DCVS-LLM Wait time for frequency >> inc/decrement >> + * @reg_llm_volt_vote_hyst: DCVS-LLM Wait time for voltage inc/decrement >> + * @reg_llm_intf_dcvs_dis:Â DCVS-LLM en/disable control >> + * @reg_seq1:Â Â Â Â Â Â Â Â Â Â Â Â Â Â Sequencer extra register >> + * @reg_pdn_fsm_ctrl:Â Â Â Â Â Â Boost and Droop FSMs en/disable control >> + * @reg_cc_boost_timer:Â Â Â Â CC-Boost FSM wait first timer register >> + * @reg_dcvs_boost_timer:Â Â DCVS-Boost FSM wait first timer register >> + * @reg_ps_boost_timer:Â Â Â Â PS-Boost FSM wait first timer register >> + * @boost_timer_reg_len:Â Â Â Length of boost timer registers >> + * @reg_boost_sync_delay:Â Â PLL signal timing control for Boost >> + * @reg_droop_ctrl:Â Â Â Â Â Â Â Â Droop control value >> + * @reg_droop_release_ctrl: Wait for Droop release >> + * @reg_droop_unstall_ctrl: Wait for Droop unstall >> + * @reg_droop_wait_release_ctrl: Time to wait for state release >> + * @reg_droop_timer_ctrl:Â Â Droop timer >> + * @reg_droop_sync_delay:Â Â PLL signal timing control for Droop >> + * @reg_pll_override:Â Â Â Â Â Â PLL Droop Override en/disable control >> + * @reg_cycle_counter:Â Â Â Â Â OSM CPU cycle counter >> + * >> + * This structure holds the register offsets that are used to set-up >> + * the Operating State Manager (OSM) parameters, when it is not (or >> + * not entirely) configured from the bootloader and TrustZone. >> + * >> + * Acronyms used in this documentation: >> + * CC = Core Count >> + * PS = Power-Save >> + * VC = Virtual Corner >> + * LLM = Limits Load Management >> + * DCVS = Dynamic Clock and Voltage Scaling >> + */ >> +struct qcom_cpufreq_soc_setup_data { >> +Â Â Â /* OSM phys register offsets */ >> +Â Â Â u16 reg_osm_sequencer; >> + >> +Â Â Â /* Frequency domain register offsets */ >> +Â Â Â u16 reg_override; >> +Â Â Â u16 reg_spare; >> +Â Â Â u16 reg_cc_zero_behav; >> +Â Â Â u16 reg_spm_cc_hyst; >> +Â Â Â u16 reg_spm_cc_dcvs_dis; >> +Â Â Â u16 reg_spm_core_ret_map; >> +Â Â Â u16 reg_llm_freq_vote_hyst; >> +Â Â Â u16 reg_llm_volt_vote_hyst; >> +Â Â Â u16 reg_llm_intf_dcvs_dis; >> +Â Â Â u16 reg_seq1; >> +Â Â Â u16 reg_pdn_fsm_ctrl; >> +Â Â Â u16 reg_cc_boost_timer; >> +Â Â Â u16 reg_dcvs_boost_timer; >> +Â Â Â u16 reg_ps_boost_timer; >> +Â Â Â u16 boost_timer_reg_len; >> +Â Â Â u16 reg_boost_sync_delay; >> +Â Â Â u16 reg_droop_ctrl; >> +Â Â Â u16 reg_droop_release_ctrl; >> +Â Â Â u16 reg_droop_unstall_ctrl; >> +Â Â Â u16 reg_droop_wait_release_ctrl; >> +Â Â Â u16 reg_droop_timer_ctrl; >> +Â Â Â u16 reg_droop_sync_delay; >> +Â Â Â u16 reg_pll_override; >> +Â Â Â u16 reg_cycle_counter; >> +}; >> + >> +/** >> + * struct qcom_cpufreq_soc_acd_data - Adaptive Clock Distribution data >> + * >> + * @tl_delay_reg:Â Â Â Â Â Tunable-Length Delay (TLD) register offset >> + * @acd_ctrl_reg:Â Â Â Â Â Control Register (CR) register offset >> + * @softstart_reg:Â Â Â Â Soft Start Control Register (SSCR) register >> offset >> + * @ext_intf_reg:Â Â Â Â Â External interface configuration register offset >> + * @auto_xfer_reg:Â Â Â Â Auto Register-Transfer register offset >> + * @auto_xfer_cfg_reg: Auto Register-Transfer Configuration reg offset >> + * @auto_xfer_ctl_reg: Auto Register-Transfer Control register offset >> + * @auto_xfer_sts_reg: Auto Register-Transfer Status register offset >> + * @dcvs_sw_reg:Â Â Â Â Â Â Software DCVS register offset >> + * @gfmux_cfg_reg:Â Â Â Â Glitch-Free MUX configuration register offset >> + * @write_ctl_reg:Â Â Â Â Write Control register >> + * @write_sts_reg:Â Â Â Â Write Status register >> + * @tl_delay_val:Â Â Â Â Â Tunable-Length Delay (TLD) value >> + * @acd_ctrl_val:Â Â Â Â Â Control Register (CR) value >> + * @softstart_val:Â Â Â Â Soft Start Control Register (SSCR) value >> + * @ext_intf0_val:Â Â Â Â Initial external interface configuration value >> + * @ext_intf1_val:Â Â Â Â Final external interface configuration value >> + * @auto_xfer_val:Â Â Â Â Auto-register Transfer Control value >> + * >> + * This structure holds the register offsets (from the ACD iospace base) >> + * and the parameters that are required to configure the OSM to >> + * initialize the Adaptive Clock Distribution (ACD) system. >> + */ >> +struct qcom_cpufreq_soc_acd_data { >> +Â Â Â u8 tl_delay_reg; >> +Â Â Â u8 acd_ctrl_reg; >> +Â Â Â u8 softstart_reg; >> +Â Â Â u8 ext_intf_reg; >> +Â Â Â u8 auto_xfer_reg; >> +Â Â Â u8 auto_xfer_cfg_reg; >> +Â Â Â u8 auto_xfer_ctl_reg; >> +Â Â Â u8 auto_xfer_sts_reg; >> +Â Â Â u8 dcvs_sw_reg; >> +Â Â Â u8 gfmux_cfg_reg; >> +Â Â Â u8 write_ctl_reg; >> +Â Â Â u8 write_sts_reg; >> +Â Â Â u32 tl_delay_val; >> +Â Â Â u32 acd_ctrl_val; >> +Â Â Â u32 softstart_val; >> +Â Â Â u32 ext_intf0_val; >> +Â Â Â u32 ext_intf1_val; >> +Â Â Â u32 auto_xfer_val; >> +}; >> + >> +/** >> + * struct qcom_cpufreq_hw_params - Operating State Manager (OSM) >> Parameters >> + * >> + * @volt_lut_val: Value composed of: virtual corner (vc) and voltage >> in mV. >> + * @freq_lut_val: Value composed of: core count, clock source and output >> + *Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â frequency in MHz. >> + * @override_val: PLL parameters that the OSM uses to override the >> previous >> + *Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â setting coming from the bootloader, or when >> uninitialized. >> + * @spare_val:Â Â Â Spare register, used by both this driver and the >> OSM HW >> + *Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â to identify MEM-ACC levels in relation to virtual >> corners. >> + * @acc_thresh:Â Â MEM-ACC threshold level >> + * >> + * This structure holds the parameters to write to the OSM registers for >> + * one "Virtual Corner" (VC), or one Performance State (p-state). >> + */ >> +struct qcom_cpufreq_hw_params { >> +Â Â Â u32 volt_lut_val; >> +Â Â Â u32 freq_lut_val; >> +Â Â Â u32 override_val; >> +Â Â Â u32 spare_val; >> +Â Â Â u32 acc_thresh; >> +}; >> +/** >> + * struct qcom_cpufreq_soc_data - SoC specific register offsets of >> the OSM >> + * >> + * @reg_enable:Â Â Â Â Â Â Â Â Â Â Â OSM enable status >> + * @reg_index:Â Â Â Â Â Â Â Â Â Â Â Â Index of the Virtual Corner >> + * @reg_freq_lut:Â Â Â Â Â Â Â Â Â Frequency Lookup Table >> + * @reg_freq_lut_src_mask: Frequency Lookup Table clock-source mask >> + * @reg_volt_lut:Â Â Â Â Â Â Â Â Â Voltage Lookup Table >> + * @reg_perf_state:Â Â Â Â Â Â Â Performance State request register >> + * @lut_row_size:Â Â Â Â Â Â Â Â Â Lookup Table row size >> + * @clk_hw_div:Â Â Â Â Â Â Â Â Â Â Â Divider for "alternate" OSM clock-source >> + * @uses_tz:Â Â Â Â Â Â Â Â Â Â Â Â Â Â OSM already set-up and protected by TrustZone >> + * @setup_regs:Â Â Â Â Â Â Â Â Â Â Â Register offsets for OSM setup >> + */ >> Â struct qcom_cpufreq_soc_data { >> Â Â Â Â Â u32 reg_enable; >> +Â Â Â u32 reg_index; >> Â Â Â Â Â u32 reg_freq_lut; >> +Â Â Â u32 reg_freq_lut_src_mask; >> Â Â Â Â Â u32 reg_volt_lut; >> Â Â Â Â Â u32 reg_perf_state; >> -Â Â Â u8 lut_row_size; >> +Â Â Â u8Â lut_row_size; >> +Â Â Â u8Â clk_hw_div; >> +Â Â Â bool uses_tz; >> +Â Â Â const struct qcom_cpufreq_soc_setup_data setup_regs; >> +Â Â Â const struct qcom_cpufreq_soc_acd_data acd_data; >> Â }; >> Â struct qcom_cpufreq_data { >> @@ -35,9 +258,17 @@ struct qcom_cpufreq_data { >> Â Â Â Â Â const struct qcom_cpufreq_soc_data *soc_data; >> Â }; >> -static unsigned long cpu_hw_rate, xo_rate; >> +static const char *cprh_genpd_names[] = { "cprh", NULL }; >> +static u64 cpu_hw_rate, xo_rate; >> Â static bool icc_scaling_enabled; >> +/** >> + * qcom_cpufreq_set_bw - Set interconnect bandwidth >> + * @policy:Â Â CPUFreq policy structure >> + * @freq_khz: CPU Frequency in KHz >> + * >> + * Returns: Zero for success, otherwise negative value on errors >> + */ >> Â static int qcom_cpufreq_set_bw(struct cpufreq_policy *policy, >> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â unsigned long freq_khz) >> Â { >> @@ -59,6 +290,20 @@ static int qcom_cpufreq_set_bw(struct >> cpufreq_policy *policy, >> Â Â Â Â Â return ret; >> Â } >> +/** >> + * qcom_cpufreq_update_opp - Update CPU OPP tables >> + * @policy:Â Â CPUFreq policy structure >> + * @freq_khz: CPU Frequency for OPP entry in KHz >> + * @volt:Â Â Â Â CPU Voltage for OPP entry in microvolts >> + * >> + * The CPU frequencies and voltages are being read from the Operating >> + * State Manager (OSM) and the related OPPs, read from DT, need to be >> + * updated to reflect what the hardware will set for each p-state. >> + * If there is no OPP table specified in DT, then this function will >> + * add dynamic ones. >> + * >> + * Returns: Zero for success, otherwise negative value on errors >> + */ >> Â static int qcom_cpufreq_update_opp(struct device *cpu_dev, >> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â unsigned long freq_khz, >> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â unsigned long volt) >> @@ -79,6 +324,17 @@ static int qcom_cpufreq_update_opp(struct device >> *cpu_dev, >> Â Â Â Â Â return dev_pm_opp_enable(cpu_dev, freq_hz); >> Â } >> +/** >> + * qcom_cpufreq_hw_target_index - Set frequency/voltage >> + * @policy:Â Â CPUFreq policy structure >> + * @index:Â Â Â Performance state index to be set >> + * >> + * This function sends a request to the Operating State Manager >> + * to set a Performance State index, so, to set frequency and >> + * voltage for the target CPU/cluster. >> + * >> + * Returns: Always zero >> + */ >> Â static int qcom_cpufreq_hw_target_index(struct cpufreq_policy *policy, >> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â unsigned int index) >> Â { >> @@ -94,6 +350,12 @@ static int qcom_cpufreq_hw_target_index(struct >> cpufreq_policy *policy, >> Â Â Â Â Â return 0; >> Â } >> +/** >> + * qcom_cpufreq_hw_get - Get current Performance State from OSM >> + * @cpu:Â Â Â Â Â CPU number >> + * >> + * Returns: Current CPU/Cluster frequency or zero >> + */ >> Â static unsigned int qcom_cpufreq_hw_get(unsigned int cpu) >> Â { >> Â Â Â Â Â struct qcom_cpufreq_data *data; >> @@ -127,6 +389,644 @@ static unsigned int >> qcom_cpufreq_hw_fast_switch(struct cpufreq_policy *policy, >> Â Â Â Â Â return policy->freq_table[index].frequency; >> Â } >> +static void qcom_cpufreq_hw_boost_setup(void __iomem *timer0_addr, >> u32 len) >> +{ >> +Â Â Â u32 val; >> + >> +Â Â Â /* timer_reg0 */ >> +Â Â Â val = FIELD_PREP(BOOST_TIMER_REG_LO, PLL_WAIT_LOCK_TIME_NS); >> +Â Â Â val |= FIELD_PREP(BOOST_TIMER_REG_HI, SAFE_FREQ_WAIT_NS); >> +Â Â Â writel_relaxed(val, timer0_addr); >> + >> +Â Â Â /* timer_reg1 */ >> +Â Â Â val = FIELD_PREP(BOOST_TIMER_REG_LO, PLL_WAIT_LOCK_TIME_NS); >> +Â Â Â val |= FIELD_PREP(BOOST_TIMER_REG_HI, PLL_WAIT_LOCK_TIME_NS); >> +Â Â Â writel_relaxed(val, timer0_addr + len); >> + >> +Â Â Â /* timer_reg2 */ >> +Â Â Â val = FIELD_PREP(BOOST_TIMER_REG_LO, DEXT_DECREMENT_WAIT_NS); >> +Â Â Â writel_relaxed(val, timer0_addr + (2 * len)); >> +} >> + >> +/* >> + * qcom_cpufreq_gen_params - Generate parameters to send to the hardware >> + * @cpu_dev:Â Â CPU device >> + * @data:Â Â Â Â Â SoC specific register offsets >> + * @hw_tbl:Â Â Â Pointer to return the array of parameters >> + * @apm_vc:Â Â Â APM Virtual Corner crossover number, returned to the >> caller >> + * @acc_vc:Â Â Â MEMACC Virtual Corner crossover number, returned to >> the caller >> + * @cpu_count: Number of CPUs in the frequency domain >> + * >> + * This function allocates a 'qcom_cpufreq_hw_params' parameters table, >> + * fills it and returns it to the consumer, ready to get sent to the HW. >> + * Since the APM threshold is just one >> + * Freeing the table after usage is left to the caller. >> + * >> + * Returns: Number of allocated (and filled) elements in the table, >> + *Â Â Â Â Â Â Â Â Â otherwise negative value on errors. >> + */ >> +static int qcom_cpufreq_gen_params(struct device *cpu_dev, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â struct qcom_cpufreq_data *data, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â struct qcom_cpufreq_hw_params **hw_tbl, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â int *apm_vc, int *acc_vc, int cpu_count) >> +{ >> +Â Â Â struct platform_device *pdev = cpufreq_get_driver_data(); >> +Â Â Â const struct qcom_cpufreq_soc_data *soc_data = data->soc_data; >> +Â Â Â struct device **opp_virt_dev; >> +Â Â Â struct opp_table *genpd_opp_table; >> +Â Â Â struct dev_pm_opp *apm_opp, *acc_opp; >> +Â Â Â unsigned long apm_uV, acc_uV, rate; >> +Â Â Â int i, gpd_opp_cnt, ret; >> +Â Â Â u8 last_acc_corner = 0, prev_spare = 0; >> + >> +Â Â Â /* >> +Â Â Â Â * Install the OPP table that we get from DT here already, >> +Â Â Â Â * otherwise it gets really messy as we'd have to manually walk >> +Â Â Â Â * through the entire OPP DT, manually find frequencies and >> +Â Â Â Â * also manually exclude supported-hw (for speed-binning). >> +Â Â Â Â * This also makes it easier for the genpd to add info in it. >> +Â Â Â Â */ >> +Â Â Â ret = dev_pm_opp_of_add_table(cpu_dev); >> +Â Â Â if (ret) { >> +Â Â Â Â Â Â Â dev_err(&pdev->dev, "Cannot install CPU OPP table: %d\n", ret); >> +Â Â Â Â Â Â Â return ret; >> +Â Â Â } >> + >> +Â Â Â /* Get a handle to the genpd virtual device */ >> +Â Â Â genpd_opp_table = dev_pm_opp_attach_genpd(cpu_dev, cprh_genpd_names, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â &opp_virt_dev); >> +Â Â Â if (IS_ERR(genpd_opp_table)) { >> +Â Â Â Â Â Â Â ret = PTR_ERR(genpd_opp_table); >> +Â Â Â Â Â Â Â if (ret != -EPROBE_DEFER) >> +Â Â Â Â Â Â Â Â Â Â Â dev_err(&pdev->dev, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â "Could not attach to pm_domain: %d\n", ret); >> +Â Â Â Â Â Â Â goto opp_dispose; >> +Â Â Â } >> + >> +Â Â Â /* Get the count of available OPPs coming from the power domain */ >> +Â Â Â gpd_opp_cnt = dev_pm_opp_get_opp_count(cpu_dev); >> +Â Â Â if (gpd_opp_cnt < 2) { >> +Â Â Â Â Â Â Â ret = gpd_opp_cnt > 0 ? -EINVAL : gpd_opp_cnt; >> +Â Â Â Â Â Â Â goto detach_gpd; >> +Â Â Â } >> + >> +Â Â Â /* Find the APM threshold disabled OPP and "annotate" the voltage */ >> +Â Â Â apm_opp = dev_pm_opp_find_freq_exact(cpu_dev, 0, false); >> +Â Â Â if (IS_ERR(apm_opp)) { >> +Â Â Â Â Â Â Â ret = -EINVAL; >> +Â Â Â Â Â Â Â goto detach_gpd; >> +Â Â Â } >> + >> +Â Â Â /* If we get no APM voltage, the system is going to be unstable */ >> +Â Â Â apm_uV = dev_pm_opp_get_voltage(apm_opp); >> +Â Â Â if (apm_uV == 0) { >> +Â Â Â Â Â Â Â ret = -EINVAL; >> +Â Â Â Â Â Â Â goto detach_gpd; >> +Â Â Â } >> +Â Â Â *apm_vc = -1; >> + >> +Â Â Â /* >> +Â Â Â Â * Find the ACC threshold disabled OPP and "annotate" the voltage: >> +Â Â Â Â * this is optional, as not every SoC needs it. >> +Â Â Â Â */ >> +Â Â Â acc_opp = dev_pm_opp_find_freq_exact(cpu_dev, 1, false); >> +Â Â Â if (IS_ERR(acc_opp)) { >> +Â Â Â Â Â Â Â acc_uV = ULONG_MAX; >> +Â Â Â Â Â Â Â *acc_vc = U8_MAX; >> +Â Â Â } else { >> +Â Â Â Â Â Â Â /* >> +Â Â Â Â Â Â Â Â * If the ACC OPP is present, this means that we *have* to >> +Â Â Â Â Â Â Â Â * set the relative corner on the OSM: if we cannot get the >> +Â Â Â Â Â Â Â Â * voltage at that point, the CPU will inevitably freeze. >> +Â Â Â Â Â Â Â Â */ >> +Â Â Â Â Â Â Â acc_uV = dev_pm_opp_get_voltage(acc_opp); >> +Â Â Â Â Â Â Â if (acc_uV == 0) { >> +Â Â Â Â Â Â Â Â Â Â Â ret = -EINVAL; >> +Â Â Â Â Â Â Â Â Â Â Â goto detach_gpd; >> +Â Â Â Â Â Â Â } >> +Â Â Â Â Â Â Â *acc_vc = -1; >> +Â Â Â } >> + >> +Â Â Â *hw_tbl = devm_kmalloc_array(&pdev->dev, gpd_opp_cnt, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â sizeof(**hw_tbl), GFP_KERNEL); >> +Â Â Â if (!hw_tbl) { >> +Â Â Â Â Â Â Â ret = -ENOMEM; >> +Â Â Â Â Â Â Â goto detach_gpd; >> +Â Â Â } >> + >> +Â Â Â for (i = 0, rate = 0; ; rate++, i++) { >> +Â Â Â Â Â Â Â struct qcom_cpufreq_hw_params *entry = *hw_tbl + i; >> +Â Â Â Â Â Â Â struct dev_pm_opp *genpd_opp; >> +Â Â Â Â Â Â Â struct device_node *np; >> +Â Â Â Â Â Â Â u32 pll_div, millivolts, f_src; >> + >> +Â Â Â Â Â Â Â /* Find the next enabled OPP's frequency (ignores APM/ACC) */ >> +Â Â Â Â Â Â Â genpd_opp = dev_pm_opp_find_freq_ceil(cpu_dev, &rate); >> +Â Â Â Â Â Â Â if (IS_ERR(genpd_opp)) >> +Â Â Â Â Â Â Â Â Â Â Â break; >> + >> +Â Â Â Â Â Â Â /* Get mandatory and optional properties from the OPP DT */ >> +Â Â Â Â Â Â Â np = dev_pm_opp_get_of_node(genpd_opp); >> +Â Â Â Â Â Â Â if (!np) >> +Â Â Â Â Â Â Â Â Â Â Â break; >> + >> +Â Â Â Â Â Â Â if (of_property_read_u32(np, "qcom,pll-override", >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â &entry->override_val)) { >> +Â Â Â Â Â Â Â Â Â Â Â of_node_put(np); >> +Â Â Â Â Â Â Â Â Â Â Â return -EINVAL; >> +Â Â Â Â Â Â Â } >> + >> +Â Â Â Â Â Â Â if (of_property_read_u32(np, "qcom,spare-data", >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â &entry->spare_val)) >> +Â Â Â Â Â Â Â Â Â Â Â entry->spare_val = 0; >> + >> +Â Â Â Â Â Â Â if (of_property_read_u32(np, "qcom,pll-div", &pll_div)) >> +Â Â Â Â Â Â Â Â Â Â Â pll_div = 0; >> + >> +Â Â Â Â Â Â Â of_node_put(np); >> + >> +Â Â Â Â Â Â Â /* Get voltage in microvolts, then convert to millivolts */ >> +Â Â Â Â Â Â Â millivolts = dev_pm_opp_get_voltage(genpd_opp); >> +Â Â Â Â Â Â Â if (millivolts >= apm_uV && *apm_vc < 0) >> +Â Â Â Â Â Â Â Â Â Â Â *apm_vc = i; >> +Â Â Â Â Â Â Â if (millivolts >= acc_uV && *acc_vc < 0) >> +Â Â Â Â Â Â Â Â Â Â Â *acc_vc = i; >> + >> +Â Â Â Â Â Â Â millivolts /= 1000; >> + >> +Â Â Â Â Â Â Â if (millivolts < 150 || millivolts > 1400) { >> +Â Â Â Â Â Â Â Â Â Â Â dev_err(&pdev->dev, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â "Read invalid voltage: %u.\n", millivolts); >> +Â Â Â Â Â Â Â Â Â Â Â return -EINVAL; >> +Â Â Â Â Â Â Â } >> + >> +Â Â Â Â Â Â Â /* In the OSM firmware, "Virtual Corner" levels start from 0 */ >> +Â Â Â Â Â Â Â entry->volt_lut_val = FIELD_PREP(LUT_VOLT_VC, i); >> +Â Â Â Â Â Â Â entry->volt_lut_val |= FIELD_PREP(LUT_VOLT, millivolts); >> + >> +Â Â Â Â Â Â Â /* Only the first frequency can have alternate source */ >> +Â Â Â Â Â Â Â f_src = i ? 1 : 0; >> +Â Â Â Â Â Â Â f_src <<= ffs(soc_data->reg_freq_lut_src_mask) - 1; >> +Â Â Â Â Â Â Â entry->freq_lut_val = f_src | div_u64(rate, xo_rate); >> +Â Â Â Â Â Â Â entry->freq_lut_val |= FIELD_PREP(LUT_CORE_COUNT, cpu_count); >> + >> +Â Â Â Â Â Â Â /* >> +Â Â Â Â Â Â Â Â * PLL divider is not always 0 and there is no way to determine >> +Â Â Â Â Â Â Â Â * it automatically, as setting this value higher than DIV1 >> +Â Â Â Â Â Â Â Â * will make the OSM HW to effectively set the PLL at 2-4x >> +Â Â Â Â Â Â Â Â * the CPU frequency and then divide the CPU clock by this div, >> +Â Â Â Â Â Â Â Â * so this value is effectively used as both a multiplier and >> +Â Â Â Â Â Â Â Â * divider. >> +Â Â Â Â Â Â Â Â * This value cannot be calculated because it depends on >> +Â Â Â Â Â Â Â Â * manual calibration and is (most probably) used to choose >> +Â Â Â Â Â Â Â Â * a PLL frequency that gives the least possible jitter. >> +Â Â Â Â Â Â Â Â */ >> +Â Â Â Â Â Â Â entry->freq_lut_val |= FIELD_PREP(LUT_PLL_DIV, pll_div); >> + >> +Â Â Â Â Â Â Â /* >> +Â Â Â Â Â Â Â Â * MEM-ACC Virtual Corner threshold voltage: set the >> +Â Â Â Â Â Â Â Â * acc_thresh to the "next different" spare_val, otherwise >> +Â Â Â Â Â Â Â Â * set it to an invalid value, so that we can recognize it >> +Â Â Â Â Â Â Â Â * and exclude it later, when this set of data will be used >> +Â Â Â Â Â Â Â Â * for programming the OSM. >> +Â Â Â Â Â Â Â Â */ >> +Â Â Â Â Â Â Â if (entry->spare_val != prev_spare) { >> +Â Â Â Â Â Â Â Â Â Â Â prev_spare = entry->spare_val; >> +Â Â Â Â Â Â Â Â Â Â Â entry->acc_thresh = last_acc_corner; >> +Â Â Â Â Â Â Â } else { >> +Â Â Â Â Â Â Â Â Â Â Â entry->acc_thresh = SEQ_MEM_ACC_MAX_LEVELS + 1; >> +Â Â Â Â Â Â Â } >> + >> +Â Â Â Â Â Â Â dev_dbg(&pdev->dev, >> +Â Â Â Â Â Â Â Â Â Â Â "[%d] freq=0x%x volt=0x%x override=0x%x spare=0x%x\n", >> +Â Â Â Â Â Â Â Â Â Â Â i, entry->freq_lut_val, entry->volt_lut_val, >> +Â Â Â Â Â Â Â Â Â Â Â entry->override_val, entry->spare_val); >> +Â Â Â Â Â Â Â dev_pm_opp_put(genpd_opp); >> +Â Â Â } >> + >> +Â Â Â /* >> +Â Â Â Â * If we've got a customized mem-acc corner but we couldn't >> +Â Â Â Â * find any suitable crossover, or the corner is less than >> +Â Â Â Â * the minimum amount of required corners for mem-acc scaling, >> +Â Â Â Â * the values are not valid, hence fall back to LUT values. >> +Â Â Â Â */ >> +Â Â Â if (acc_uV != ULONG_MAX && *acc_vc < SEQ_MEM_ACC_MAX_LEVELS - 1) { >> +Â Â Â Â Â Â Â dev_dbg(&pdev->dev, >> +Â Â Â Â Â Â Â Â Â Â Â "MEM-ACC corner: invalid values VC%d %luuV\n", >> +Â Â Â Â Â Â Â Â Â Â Â *acc_vc, acc_uV); >> +Â Â Â Â Â Â Â *acc_vc = U8_MAX; >> +Â Â Â } >> + >> +Â Â Â /* >> +Â Â Â Â * If we have probed less params than what we need, then the >> +Â Â Â Â * OPP table that we got from the genpd is malformed for some >> +Â Â Â Â * reason: in this case, do not apply the table to the HW. >> +Â Â Â Â */ >> +Â Â Â if (i < gpd_opp_cnt) { >> +Â Â Â Â Â Â Â dev_err(&pdev->dev, "Got bad OPP table from power domain.\n"); >> +Â Â Â Â Â Â Â ret = -EINVAL; >> +Â Â Â Â Â Â Â goto detach_gpd; >> +Â Â Â } >> + >> +detach_gpd: >> +Â Â Â dev_pm_opp_detach_genpd(genpd_opp_table); >> +opp_dispose: >> +Â Â Â /* >> +Â Â Â Â * Now that we're totally done with it, dispose of all the dynamic >> +Â Â Â Â * OPPs in the table: like this, at the end of the OSM configuration >> +Â Â Â Â * we are leaving it like it was magically configured by the TZ or >> +Â Â Â Â * by the bootloader and using the rest of the driver as if this >> +Â Â Â Â * programming phase has never happened in the OS, like new SoCs. >> +Â Â Â Â * This also makes us able to not modify a single bit in the basic >> +Â Â Â Â * OSM frequency request logic that was meant for the newest SoCs. >> +Â Â Â Â */ >> +Â Â Â dev_pm_opp_remove_all_dynamic(cpu_dev); >> +Â Â Â return ret < 0 ? ret : i; >> +} >> + >> +static inline u32 qcom_cpufreq_acd_regbit(u8 acd_reg_offset) >> +{ >> +Â Â Â return BIT((acd_reg_offset / 4)); >> +} >> + >> +static int qcom_cpufreq_hw_acd_write_autoxfer(struct >> qcom_cpufreq_data *data, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â void __iomem *acd_base, u32 val) >> +{ >> +Â Â Â const struct qcom_cpufreq_soc_data *sdata = data->soc_data; >> +Â Â Â const struct qcom_cpufreq_soc_acd_data *aregs = &sdata->acd_data; >> +Â Â Â u32 regval; >> +Â Â Â int retry = 3; >> + >> +Â Â Â writel_relaxed(val, acd_base + aregs->auto_xfer_cfg_reg); >> + >> +Â Â Â /* (Clear, then) Set AUTOXFER START */ >> +Â Â Â writel_relaxed(0, acd_base + aregs->auto_xfer_reg); >> +Â Â Â writel_relaxed(1, acd_base + aregs->auto_xfer_reg); >> + >> +Â Â Â /* Poll for status: if the first bit is set the transfer is done. */ >> +Â Â Â do { >> +Â Â Â Â Â Â Â regval = readl_relaxed(acd_base + aregs->auto_xfer_sts_reg); >> +Â Â Â Â Â Â Â if (regval & BIT(0)) >> +Â Â Â Â Â Â Â Â Â Â Â break; >> +Â Â Â Â Â Â Â udelay(1); >> +Â Â Â } while (--retry); >> + >> +Â Â Â if (retry <= 0) >> +Â Â Â Â Â Â Â return -ETIMEDOUT; >> +Â Â Â return 0; >> +} >> + >> +static int qcom_cpufreq_hw_acd_write_xfer(struct qcom_cpufreq_data >> *data, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â void __iomem *acd_base, u8 reg, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â u32 val) >> +{ >> +Â Â Â const struct qcom_cpufreq_soc_data *sdata = data->soc_data; >> +Â Â Â const struct qcom_cpufreq_soc_acd_data *aregs = &sdata->acd_data; >> +Â Â Â int retry = 2; >> +Â Â Â u32 regval; >> + >> +Â Â Â /* Write to the register, then initiate manual transfer */ >> +Â Â Â writel_relaxed(val, acd_base + reg); >> + >> +Â Â Â /* >> +Â Â Â Â * Make sure that the write goes through, otherwise the local >> +Â Â Â Â * transfer process will wrongly copy the old value. >> +Â Â Â Â */ >> +Â Â Â wmb(); >> + >> +Â Â Â /* Clear write control register */ >> +Â Â Â writel_relaxed(0, acd_base + aregs->write_ctl_reg); >> + >> +Â Â Â regval = (reg / 4) << ACD_WRITE_CTL_SELECT_SHIFT; >> +Â Â Â regval |= ACD_WRITE_CTL_UPDATE_EN; >> +Â Â Â writel_relaxed(regval, acd_base + aregs->write_ctl_reg); >> + >> +Â Â Â /* Wait until ACD Local Transfer is done */ >> +Â Â Â do { >> +Â Â Â Â Â Â Â regval = readl_relaxed(acd_base + aregs->write_sts_reg); >> +Â Â Â Â Â Â Â if (regval & qcom_cpufreq_acd_regbit(reg)) >> +Â Â Â Â Â Â Â Â Â Â Â break; >> +Â Â Â Â Â Â Â udelay(1); >> +Â Â Â } while (--retry); >> + >> +Â Â Â if (retry <= 0) >> +Â Â Â Â Â Â Â return -ETIMEDOUT; >> +Â Â Â return 0; >> +} >> + >> +/** >> + * qcom_cpufreq_hw_acd_init - Initialize ACD params in the OSM >> + * @cpu_dev:Â Â CPU device >> + * @policy:Â Â Â CPUFreq policy structure >> + * @index:Â Â Â Â Instance number (CPU cluster number) >> + * >> + * On some SoCs it is required to send the ACD configuration parameters >> + * to the OSM. This function takes the parameters from the SoC specific >> + * configuration and writes them only if a "osm-acdN" iospace has been >> + * declared (hence, it's present). >> + * >> + * Returns: Zero for success, otherwise negative number on error. >> + */ >> +static int qcom_cpufreq_hw_acd_init(struct device *cpu_dev, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â struct cpufreq_policy *policy, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â int index) >> +{ >> +Â Â Â struct platform_device *pdev = cpufreq_get_driver_data(); >> +Â Â Â struct qcom_cpufreq_data *ddata = policy->driver_data; >> +Â Â Â const struct qcom_cpufreq_soc_data *sdata = ddata->soc_data; >> +Â Â Â const struct qcom_cpufreq_soc_acd_data *aregs = &sdata->acd_data; >> +Â Â Â const char *acd_resname; >> +Â Â Â void __iomem *acd_base; >> +Â Â Â u32 rmask; >> +Â Â Â int ret; >> + >> +Â Â Â acd_resname = kasprintf(GFP_KERNEL, "osm-acd%d", index); >> +Â Â Â if (!acd_resname) >> +Â Â Â Â Â Â Â return -ENOMEM; >> + >> +Â Â Â acd_base = devm_platform_ioremap_resource_byname(pdev, acd_resname); >> +Â Â Â kfree(acd_resname); >> +Â Â Â if (IS_ERR(acd_base)) { >> +Â Â Â Â Â Â Â dev_vdbg(cpu_dev, "Skipping ACD initialization.\n"); >> +Â Â Â Â Â Â Â return 0; >> +Â Â Â } >> + >> +Â Â Â writel_relaxed(aregs->tl_delay_val, acd_base + aregs->tl_delay_reg); >> +Â Â Â writel_relaxed(aregs->acd_ctrl_val, acd_base + aregs->acd_ctrl_reg); >> +Â Â Â writel_relaxed(aregs->softstart_val, acd_base + >> aregs->softstart_reg); >> +Â Â Â writel_relaxed(aregs->ext_intf0_val, acd_base + >> aregs->ext_intf_reg); >> +Â Â Â writel_relaxed(aregs->auto_xfer_val, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â acd_base + aregs->auto_xfer_ctl_reg); >> +Â Â Â /* >> +Â Â Â Â * Before initiating register auto transfer, make sure that all >> +Â Â Â Â * the writes go through, otherwise bad values will be transferred. >> +Â Â Â Â */ >> +Â Â Â wmb(); >> + >> +Â Â Â rmask = qcom_cpufreq_acd_regbit(aregs->acd_ctrl_reg) | >> +Â Â Â Â Â Â Â qcom_cpufreq_acd_regbit(aregs->tl_delay_reg) | >> +Â Â Â Â Â Â Â qcom_cpufreq_acd_regbit(aregs->softstart_reg) | >> +Â Â Â Â Â Â Â qcom_cpufreq_acd_regbit(aregs->ext_intf_reg); >> +Â Â Â ret = qcom_cpufreq_hw_acd_write_autoxfer(ddata, acd_base, rmask); >> +Â Â Â if (ret) >> +Â Â Â Â Â Â Â return ret; >> + >> +Â Â Â /* Switch CPUSS clock source to ACD clock */ >> +Â Â Â ret = qcom_cpufreq_hw_acd_write_xfer(ddata, acd_base, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â aregs->gfmux_cfg_reg, 1); >> +Â Â Â if (ret) >> +Â Â Â Â Â Â Â return ret; >> + >> +Â Â Â /* (Set, then) Clear DCVS_SW */ >> +Â Â Â ret = qcom_cpufreq_hw_acd_write_xfer(ddata, acd_base, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â aregs->dcvs_sw_reg, 1); >> +Â Â Â if (ret) >> +Â Â Â Â Â Â Â return ret; >> +Â Â Â ret = qcom_cpufreq_hw_acd_write_xfer(ddata, acd_base, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â aregs->dcvs_sw_reg, 0); >> +Â Â Â if (ret) >> +Â Â Â Â Â Â Â return ret; >> + >> +Â Â Â /* Wait for clock switch time */ >> +Â Â Â udelay(1); >> + >> +Â Â Â /* Program the final ACD external interface */ >> +Â Â Â ret = qcom_cpufreq_hw_acd_write_xfer(ddata, acd_base, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â aregs->ext_intf_reg, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â aregs->ext_intf1_val); >> +Â Â Â if (ret) >> +Â Â Â Â Â Â Â return ret; >> + >> +Â Â Â /* Initiate transfer of the final ACD value */ >> +Â Â Â rmask |= qcom_cpufreq_acd_regbit(aregs->gfmux_cfg_reg); >> +Â Â Â writel_relaxed(rmask, acd_base + aregs->auto_xfer_cfg_reg); >> + >> +Â Â Â /* Wait for ACD to stabilize. Same wait as the OSM boot time... */ >> +Â Â Â udelay(OSM_BOOT_TIME_US); >> +Â Â Â return ret; >> +} >> + >> +/** >> + * qcom_cpufreq_hw_write_lut - Write Lookup Table params to the OSM >> + * @cpu_dev:Â Â CPU device >> + * @policy:Â Â Â CPUFreq policy structure >> + * @cpu_count: Number of CPUs in the frequency domain >> + * @index:Â Â Â Â Instance number (CPU cluster number) >> + * >> + * Program all the Lookup Table (LUT) entries and related thresholds >> + * to the Operating State Manager on platforms where the same hasn't >> + * been done already by the bootloader or TrustZone before booting >> + * the operating system's kernel; >> + * On these platforms, write access to the OSM is (obviously) not >> + * blocked by the hypervisor. >> + * >> + * Returns: Zero for success, otherwise negative number on error. >> + */ >> +static int qcom_cpufreq_hw_write_lut(struct device *cpu_dev, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â struct cpufreq_policy *policy, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â int cpu_count, int index) >> +{ >> +Â Â Â struct platform_device *pdev = cpufreq_get_driver_data(); >> +Â Â Â struct qcom_cpufreq_data *ddata = policy->driver_data; >> +Â Â Â const struct qcom_cpufreq_soc_data *sdata = ddata->soc_data; >> +Â Â Â const struct qcom_cpufreq_soc_setup_data *sregs = >> &sdata->setup_regs; >> +Â Â Â struct qcom_cpufreq_hw_params *hw_tbl; >> +Â Â Â struct resource *osm_rsrc; >> +Â Â Â const char *osm_resname; >> +Â Â Â u32 sreg, seq_addr, acc_lval = 0, last_spare = 1; >> +Â Â Â int apm_vc = INT_MAX, acc_vc = U8_MAX, acc_idx = 0; >> +Â Â Â int acc_val[SEQ_MEM_ACC_MAX_LEVELS]; >> +Â Â Â int i, ret, num_entries; >> + >> +Â Â Â osm_resname = kasprintf(GFP_KERNEL, "osm-domain%d", index); >> +Â Â Â if (!osm_resname) >> +Â Â Â Â Â Â Â return -ENOMEM; >> + >> +Â Â Â /* >> +Â Â Â Â * On some SoCs the OSM is not getting programmed from bootloader >> +Â Â Â Â * and needs to be done here: in this case, we need to retrieve >> +Â Â Â Â * the base physical address for the "Sequencer", so we will get >> +Â Â Â Â * the OSM base phys and apply the sequencer offset. >> +Â Â Â Â * >> +Â Â Â Â * Note: We are not remapping this iospace because we are really >> +Â Â Â Â *Â Â Â Â Â Â sending the physical address through SCM calls later. >> +Â Â Â Â */ >> +Â Â Â osm_rsrc = platform_get_resource_byname(pdev, IORESOURCE_MEM, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â osm_resname); >> +Â Â Â kfree(osm_resname); >> + >> +Â Â Â if (!osm_rsrc) >> +Â Â Â Â Â Â Â return -ENODEV; >> + >> +Â Â Â seq_addr = osm_rsrc->start + sregs->reg_osm_sequencer; >> + >> +Â Â Â num_entries = qcom_cpufreq_gen_params(cpu_dev, ddata, &hw_tbl, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â &apm_vc, &acc_vc, cpu_count); >> +Â Â Â if (num_entries < 0) >> +Â Â Â Â Â Â Â return num_entries; >> + >> +Â Â Â for (i = 0; i < LUT_MAX_ENTRIES; i++) { >> +Â Â Â Â Â Â Â struct qcom_cpufreq_hw_params *entry; >> +Â Â Â Â Â Â Â int pos = i * sdata->lut_row_size; >> + >> +Â Â Â Â Â Â Â /* >> +Â Â Â Â Â Â Â Â * If we have reached the end of the params table, write >> +Â Â Â Â Â Â Â Â * the last valid entry until the end of the OSM table. >> +Â Â Â Â Â Â Â Â */ >> +Â Â Â Â Â Â Â if (i < num_entries) >> +Â Â Â Â Â Â Â Â Â Â Â entry = &hw_tbl[i]; >> +Â Â Â Â Â Â Â else >> +Â Â Â Â Â Â Â Â Â Â Â entry = &hw_tbl[num_entries - 1]; >> + >> +Â Â Â Â Â Â Â writel_relaxed(i, ddata->base + sdata->reg_index + pos); >> + >> +Â Â Â Â Â Â Â writel_relaxed(entry->volt_lut_val, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â ddata->base + sdata->reg_volt_lut + pos); >> + >> +Â Â Â Â Â Â Â writel_relaxed(entry->freq_lut_val, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â ddata->base + sdata->reg_freq_lut + pos); >> + >> +Â Â Â Â Â Â Â writel_relaxed(entry->override_val, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â ddata->base + sregs->reg_override + pos); >> + >> +Â Â Â Â Â Â Â writel_relaxed(entry->spare_val, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â ddata->base + sregs->reg_spare + pos); >> + >> +Â Â Â Â Â Â Â dev_dbg(cpu_dev, >> +Â Â Â Â Â Â Â Â Â Â Â "Writing [%d] v:0x%x f:0x%x ovr:0x%x s:0x%x\n", i, >> +Â Â Â Â Â Â Â Â Â Â Â entry->volt_lut_val, entry->freq_lut_val, >> +Â Â Â Â Â Â Â Â Â Â Â entry->override_val, entry->spare_val); >> + >> +Â Â Â Â Â Â Â /* >> +Â Â Â Â Â Â Â Â * MEM-ACC Virtual Corner threshold voltage: this gets set >> +Â Â Â Â Â Â Â Â * as the pairs of corners in which there is a transition >> +Â Â Â Â Â Â Â Â * between one MEM-ACC level and the next one. >> +Â Â Â Â Â Â Â Â * >> +Â Â Â Â Â Â Â Â * Notes: The spare_val can never be zero; >> +Â Â Â Â Â Â Â Â *Â Â Â Â Â Â Â The first spare_val is always 1; >> +Â Â Â Â Â Â Â Â *Â Â Â Â Â Â Â The maximum number of pairs is two (four registers). >> +Â Â Â Â Â Â Â Â * >> +Â Â Â Â Â Â Â Â * Example: (C = Corner Level - M = MEM-ACC Level) >> +Â Â Â Â Â Â Â Â *Â Â Â Â Â Â Â Â Â C0 M1 - C1 M1 - C2 M2 - C3 M2 - C4 M2 - C5 M3 >> +Â Â Â Â Â Â Â Â *Â Â Â Â Â Â Â Â Â Pairs: 1-2, 4-5 >> +Â Â Â Â Â Â Â Â */ >> +Â Â Â Â Â Â Â if (entry->spare_val <= last_spare || >> +Â Â Â Â Â Â Â Â Â Â Â acc_idx >= SEQ_MEM_ACC_MAX_LEVELS - 1) >> +Â Â Â Â Â Â Â Â Â Â Â continue; >> + >> +Â Â Â Â Â Â Â /* Standard mem-acc pairs using spare_val LUT crossovers */ >> +Â Â Â Â Â Â Â last_spare = entry->spare_val; >> +Â Â Â Â Â Â Â acc_val[acc_idx] = i - 1; >> +Â Â Â Â Â Â Â acc_idx++; >> +Â Â Â Â Â Â Â acc_val[acc_idx] = i; >> +Â Â Â Â Â Â Â acc_idx++; >> +Â Â Â } >> + >> +Â Â Â /* Sanity check: we *must* have two mem-acc crossovers (four >> values) */ >> +Â Â Â if (acc_idx < SEQ_MEM_ACC_MAX_LEVELS - 1) >> +Â Â Â Â Â Â Â return -EINVAL; >> + >> +Â Â Â /* >> +Â Â Â Â * Customized mem-acc corners, if any; in this case, the last corner >> +Â Â Â Â * in the external (CPRh) LUT is this one, placed after the APM one. >> +Â Â Â Â */ >> +Â Â Â if (acc_vc != U8_MAX) { >> +Â Â Â Â Â Â Â sreg = SEQUENCER_REG(seq_addr, SEQ_MEM_ACC_CROSSOVER_VC); >> +Â Â Â Â Â Â Â ret = qcom_scm_io_writel(sreg, num_entries + 1); >> +Â Â Â Â Â Â Â if (ret) >> +Â Â Â Â Â Â Â Â Â Â Â return ret; >> + >> +Â Â Â Â Â Â Â /* >> +Â Â Â Â Â Â Â Â * At the price of very-slightly higher power consumption, >> +Â Â Â Â Â Â Â Â * switch the ACC at one corner lower than what we've found, >> +Â Â Â Â Â Â Â Â * as this seems to be needed on at least some MSM8998 chips >> +Â Â Â Â Â Â Â Â * to achieve full system stability >> +Â Â Â Â Â Â Â Â */ >> +Â Â Â Â Â Â Â acc_vc--; >> + >> +Â Â Â Â Â Â Â /* Change only if we have to move the corner down */ >> +Â Â Â Â Â Â Â if (acc_vc < acc_val[3]) { >> +Â Â Â Â Â Â Â Â Â Â Â acc_val[2] = acc_vc - 1; >> +Â Â Â Â Â Â Â Â Â Â Â acc_val[3] = acc_vc; >> +Â Â Â Â Â Â Â } >> + >> +Â Â Â Â Â Â Â /* If needed, sanitize previously stored vals from the LUT */ >> +Â Â Â Â Â Â Â if (acc_val[2] <= acc_val[1]) >> +Â Â Â Â Â Â Â Â Â Â Â acc_val[1] = acc_val[2] - 1; >> +Â Â Â Â Â Â Â if (acc_val[1] <= acc_val[0]) >> +Â Â Â Â Â Â Â Â Â Â Â acc_val[0] = acc_val[1] - 1; >> +Â Â Â } >> + >> +Â Â Â for (i = 0; i < SEQ_MEM_ACC_MAX_LEVELS; i++) { >> +Â Â Â Â Â Â Â ret = qcom_scm_io_writel(SEQ_MEMACC_REG(seq_addr, i), >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â acc_val[i]); >> +Â Â Â Â Â Â Â if (ret) >> +Â Â Â Â Â Â Â Â Â Â Â return ret; >> +Â Â Â } >> +Â Â Â dev_dbg(cpu_dev, "Wrote MEM-ACC Pairs: [%u-%u] [%u-%u]\n", >> +Â Â Â Â Â Â Â Â acc_val[0], acc_val[1], acc_val[2], acc_val[3]); >> + >> +Â Â Â /* >> +Â Â Â Â * Program the L_VAL of the first corner requesting MEM-ACC >> +Â Â Â Â * voltage level 3 to the right sequencer register >> +Â Â Â Â */ >> +Â Â Â acc_lval = FIELD_GET(LUT_L_VAL, hw_tbl[acc_val[3]].freq_lut_val); >> +Â Â Â ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_MEM_ACC_LVAL), >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â acc_lval); >> +Â Â Â if (ret) { >> +Â Â Â Â Â Â Â dev_dbg(cpu_dev, "Cannot send memacc l_val\n"); >> +Â Â Â Â Â Â Â return ret; >> +Â Â Â } >> +Â Â Â dev_dbg(cpu_dev, "MEM-ACC L-Val is %u\n", acc_lval); >> + >> +Â Â Â /* >> +Â Â Â Â * Array Power Mux threshold level: the first virtual corner >> +Â Â Â Â * that requires a switch sequence of the APM from MX to APC. >> +Â Â Â Â */ >> +Â Â Â if (apm_vc == INT_MAX) >> +Â Â Â Â Â Â Â apm_vc = LUT_MAX_ENTRIES - 1; >> + >> +Â Â Â /* >> +Â Â Â Â * APM crossover virtual corner refers to CPRh: there, the APM >> corner >> +Â Â Â Â * is always appended to the table (so, at the end of it, right >> after >> +Â Â Â Â * the cluster dvfs entries). >> +Â Â Â Â */ >> +Â Â Â writel(num_entries, ddata->base + sregs->reg_seq1); >> +Â Â Â ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, >> SEQ_APM_CROSSOVER_VC), >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â num_entries); >> +Â Â Â if (ret) >> +Â Â Â Â Â Â Â return ret; >> + >> +Â Â Â ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_APM_THRESH_VC), >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â apm_vc); >> +Â Â Â if (ret) >> +Â Â Â Â Â Â Â return ret; >> + >> +Â Â Â ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, >> SEQ_APM_THRESH_PREVC), >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â apm_vc - 1); >> +Â Â Â if (ret) >> +Â Â Â Â Â Â Â return ret; >> + >> +Â Â Â ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_APM_PARAM), >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â (0x39 | apm_vc << 6)); >> +Â Â Â if (ret) >> +Â Â Â Â Â Â Â return ret; >> + >> +Â Â Â dev_dbg(cpu_dev, "Wrote APM Pair: [%u-%u]\n", apm_vc - 1, apm_vc); >> +Â Â Â return ret; >> +} >> + >> +/** >> + * qcom_cpufreq_hw_read_lut - Read Lookup Table from the OSM >> + * @cpu_dev:Â Â CPU device >> + * @policy:Â Â Â CPUFreq policy structure >> + * >> + * The Operating State Manager Lookup Table can always be read, even >> + * in case it was pre-programmed by the bootloader or by TrustZone. >> + * Read the LUT from it in order to build OPPs containing DVFS info. >> + * >> + * Returns: Zero for success, otherwise negative number on errors. >> + */ >> Â static int qcom_cpufreq_hw_read_lut(struct device *cpu_dev, >> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â struct cpufreq_policy *policy) >> Â { >> @@ -166,7 +1066,9 @@ static int qcom_cpufreq_hw_read_lut(struct device >> *cpu_dev, >> Â Â Â Â Â for (i = 0; i < LUT_MAX_ENTRIES; i++) { >> Â Â Â Â Â Â Â Â Â data = readl_relaxed(drv_data->base + soc_data->reg_freq_lut + >> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â i * soc_data->lut_row_size); >> -Â Â Â Â Â Â Â src = FIELD_GET(LUT_SRC, data); >> +Â Â Â Â Â Â Â src = data & soc_data->reg_freq_lut_src_mask; >> +Â Â Â Â Â Â Â src >>= ffs(soc_data->reg_freq_lut_src_mask) - 1; >> + >> Â Â Â Â Â Â Â Â Â lval = FIELD_GET(LUT_L_VAL, data); >> Â Â Â Â Â Â Â Â Â core_count = FIELD_GET(LUT_CORE_COUNT, data); >> @@ -175,17 +1077,21 @@ static int qcom_cpufreq_hw_read_lut(struct >> device *cpu_dev, >> Â Â Â Â Â Â Â Â Â volt = FIELD_GET(LUT_VOLT, data) * 1000; >> Â Â Â Â Â Â Â Â Â if (src) >> -Â Â Â Â Â Â Â Â Â Â Â freq = xo_rate * lval / 1000; >> +Â Â Â Â Â Â Â Â Â Â Â freq = xo_rate * lval; >> Â Â Â Â Â Â Â Â Â else >> -Â Â Â Â Â Â Â Â Â Â Â freq = cpu_hw_rate / 1000; >> +Â Â Â Â Â Â Â Â Â Â Â freq = cpu_hw_rate; >> +Â Â Â Â Â Â Â freq /= 1000; >> Â Â Â Â Â Â Â Â Â if (freq != prev_freq && core_count != LUT_TURBO_IND) { >> Â Â Â Â Â Â Â Â Â Â Â Â Â if (!qcom_cpufreq_update_opp(cpu_dev, freq, volt)) { >> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â table[i].frequency = freq; >> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â dev_dbg(cpu_dev, "index=%d freq=%d, core_count %d\n", i, >> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â freq, core_count); >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â dev_dbg(cpu_dev, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â "index=%d freq=%d, core_count %d\n", >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â i, freq, core_count); >> Â Â Â Â Â Â Â Â Â Â Â Â Â } else { >> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â dev_warn(cpu_dev, "failed to update OPP for >> freq=%d\n", freq); >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â dev_warn(cpu_dev, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â "failed to update OPP for freq=%d\n", >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â freq); >> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â table[i].frequency = CPUFREQ_ENTRY_INVALID; >> Â Â Â Â Â Â Â Â Â Â Â Â Â } >> @@ -205,18 +1111,19 @@ static int qcom_cpufreq_hw_read_lut(struct >> device *cpu_dev, >> Â Â Â Â Â Â Â Â Â Â Â Â Â Â * as the boost frequency >> Â Â Â Â Â Â Â Â Â Â Â Â Â Â */ >> Â Â Â Â Â Â Â Â Â Â Â Â Â if (prev->frequency == CPUFREQ_ENTRY_INVALID) { >> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â if (!qcom_cpufreq_update_opp(cpu_dev, prev_freq, >> volt)) { >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â if (!qcom_cpufreq_update_opp(cpu_dev, prev_freq, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â volt)) { >> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â prev->frequency = prev_freq; >> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â prev->flags = CPUFREQ_BOOST_FREQ; >> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â } else { >> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â dev_warn(cpu_dev, "failed to update OPP for >> freq=%d\n", >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â dev_warn(cpu_dev, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â "can't update OPP for freq=%u\n", >> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â freq); >> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â } >> Â Â Â Â Â Â Â Â Â Â Â Â Â } >> Â Â Â Â Â Â Â Â Â Â Â Â Â break; >> Â Â Â Â Â Â Â Â Â } >> - >> Â Â Â Â Â Â Â Â Â prev_freq = freq; >> Â Â Â Â Â } >> @@ -227,10 +1134,18 @@ static int qcom_cpufreq_hw_read_lut(struct >> device *cpu_dev, >> Â Â Â Â Â return 0; >> Â } >> -static void qcom_get_related_cpus(int index, struct cpumask *m) >> +/* >> + * qcom_get_related_cpus - Get mask of CPUs in the same frequency domain >> + * @index: CPU number >> + * @m:Â Â Â Â Returned CPU mask >> + * >> + * Returns: Count of CPUs inserted in the cpumask or negative number >> for error. >> + */ >> +static int qcom_get_related_cpus(int index, struct cpumask *m) >> Â { >> Â Â Â Â Â struct device_node *cpu_np; >> Â Â Â Â Â struct of_phandle_args args; >> +Â Â Â int count = 0; >> Â Â Â Â Â int cpu, ret; >> Â Â Â Â Â for_each_possible_cpu(cpu) { >> @@ -245,34 +1160,237 @@ static void qcom_get_related_cpus(int index, >> struct cpumask *m) >> Â Â Â Â Â Â Â Â Â if (ret < 0) >> Â Â Â Â Â Â Â Â Â Â Â Â Â continue; >> -Â Â Â Â Â Â Â if (index == args.args[0]) >> +Â Â Â Â Â Â Â if (index == args.args[0]) { >> Â Â Â Â Â Â Â Â Â Â Â Â Â cpumask_set_cpu(cpu, m); >> +Â Â Â Â Â Â Â Â Â Â Â count++; >> +Â Â Â Â Â Â Â } >> Â Â Â Â Â } >> + >> +Â Â Â return count > 0 ? count : -EINVAL; >> Â } >> Â static const struct qcom_cpufreq_soc_data qcom_soc_data = { >> Â Â Â Â Â .reg_enable = 0x0, >> Â Â Â Â Â .reg_freq_lut = 0x110, >> +Â Â Â .reg_freq_lut_src_mask = LUT_SRC_845, >> Â Â Â Â Â .reg_volt_lut = 0x114, >> Â Â Â Â Â .reg_perf_state = 0x920, >> Â Â Â Â Â .lut_row_size = 32, >> +Â Â Â .clk_hw_div = 2, >> +Â Â Â .uses_tz = true, >> +}; >> + >> +static const struct qcom_cpufreq_soc_data msm8998_soc_data = { >> +Â Â Â .reg_enable = 0x4, >> +Â Â Â .reg_index = 0x150, >> +Â Â Â .reg_freq_lut = 0x154, >> +Â Â Â .reg_freq_lut_src_mask = LUT_SRC_8998, >> +Â Â Â .reg_volt_lut = 0x158, >> +Â Â Â .reg_perf_state = 0xf10, >> +Â Â Â .lut_row_size = 32, >> +Â Â Â .clk_hw_div = 1, >> +Â Â Â .uses_tz = false, >> +Â Â Â .setup_regs = { >> +Â Â Â Â Â Â Â /* Physical offset for sequencer scm calls */ >> +Â Â Â Â Â Â Â .reg_osm_sequencer = 0x300, >> + >> +Â Â Â Â Â Â Â /* Frequency domain offsets */ >> +Â Â Â Â Â Â Â .reg_override = 0x15c, >> +Â Â Â Â Â Â Â .reg_spare = 0x164, >> +Â Â Â Â Â Â Â .reg_cc_zero_behav = 0x0c, >> +Â Â Â Â Â Â Â .reg_spm_cc_hyst = 0x1c, >> +Â Â Â Â Â Â Â .reg_spm_cc_dcvs_dis = 0x20, >> +Â Â Â Â Â Â Â .reg_spm_core_ret_map = 0x24, >> +Â Â Â Â Â Â Â .reg_llm_freq_vote_hyst = 0x2c, >> +Â Â Â Â Â Â Â .reg_llm_volt_vote_hyst = 0x30, >> +Â Â Â Â Â Â Â .reg_llm_intf_dcvs_dis = 0x34, >> +Â Â Â Â Â Â Â .reg_seq1 = 0x48, >> +Â Â Â Â Â Â Â .reg_pdn_fsm_ctrl = 0x70, >> +Â Â Â Â Â Â Â .reg_cc_boost_timer = 0x74, >> +Â Â Â Â Â Â Â .reg_dcvs_boost_timer = 0x84, >> +Â Â Â Â Â Â Â .reg_ps_boost_timer = 0x94, >> +Â Â Â Â Â Â Â .boost_timer_reg_len = 0x4, >> +Â Â Â Â Â Â Â .reg_boost_sync_delay = 0xa0, >> +Â Â Â Â Â Â Â .reg_droop_ctrl = 0xa4, >> +Â Â Â Â Â Â Â .reg_droop_release_ctrl = 0xa8, >> +Â Â Â Â Â Â Â .reg_droop_unstall_ctrl = 0xac, >> +Â Â Â Â Â Â Â .reg_droop_wait_release_ctrl = 0xb0, >> +Â Â Â Â Â Â Â .reg_droop_timer_ctrl = 0xb8, >> +Â Â Â Â Â Â Â .reg_droop_sync_delay = 0xbc, >> +Â Â Â Â Â Â Â .reg_pll_override = 0xc0, >> +Â Â Â Â Â Â Â .reg_cycle_counter = 0xf00, >> +Â Â Â }, >> +Â Â Â .acd_data = { >> +Â Â Â Â Â Â Â .acd_ctrl_reg = 0x4, >> +Â Â Â Â Â Â Â .tl_delay_reg = 0x8, >> +Â Â Â Â Â Â Â .softstart_reg = 0x28, >> +Â Â Â Â Â Â Â .ext_intf_reg = 0x30, >> +Â Â Â Â Â Â Â .dcvs_sw_reg = 0x34, >> +Â Â Â Â Â Â Â .gfmux_cfg_reg = 0x3c, >> +Â Â Â Â Â Â Â .auto_xfer_cfg_reg = 0x80, >> +Â Â Â Â Â Â Â .auto_xfer_reg = 0x84, >> +Â Â Â Â Â Â Â .auto_xfer_ctl_reg = 0x88, >> +Â Â Â Â Â Â Â .auto_xfer_sts_reg = 0x8c, >> +Â Â Â Â Â Â Â .write_ctl_reg = 0x90, >> +Â Â Â Â Â Â Â .write_sts_reg = 0x94, >> +Â Â Â Â Â Â Â .tl_delay_val = 38417, >> +Â Â Â Â Â Â Â .acd_ctrl_val = 0x2b5ffd, >> +Â Â Â Â Â Â Â .softstart_val = 0x501, >> +Â Â Â Â Â Â Â .ext_intf0_val = 0x2cf9ae8, >> +Â Â Â Â Â Â Â .ext_intf1_val = 0x2cf9afe, >> +Â Â Â Â Â Â Â .auto_xfer_val = 0x15, >> +Â Â Â }, >> Â }; >> Â static const struct qcom_cpufreq_soc_data epss_soc_data = { >> Â Â Â Â Â .reg_enable = 0x0, >> Â Â Â Â Â .reg_freq_lut = 0x100, >> +Â Â Â .reg_freq_lut_src_mask = LUT_SRC_845, >> Â Â Â Â Â .reg_volt_lut = 0x200, >> Â Â Â Â Â .reg_perf_state = 0x320, >> Â Â Â Â Â .lut_row_size = 4, >> +Â Â Â .clk_hw_div = 2, >> +Â Â Â .uses_tz = true, >> Â }; >> Â static const struct of_device_id qcom_cpufreq_hw_match[] = { >> Â Â Â Â Â { .compatible = "qcom,cpufreq-hw", .data = &qcom_soc_data }, >> +Â Â Â { .compatible = "qcom,cpufreq-hw-8998", .data = &msm8998_soc_data }, >> Â Â Â Â Â { .compatible = "qcom,cpufreq-epss", .data = &epss_soc_data }, >> Â Â Â Â Â {} >> Â }; >> Â MODULE_DEVICE_TABLE(of, qcom_cpufreq_hw_match); >> +/** >> + * qcom_cpufreq_hw_osm_setup - Setup and enable the OSM >> + * @cpu_dev:Â Â CPU device >> + * @policy:Â Â Â CPUFreq policy structure >> + * @cpu_count: Number of CPUs in the frequency domain >> + * >> + * On some platforms, the Operating State Manager (OSM) is not getting >> + * programmed by the bootloader, nor by TrustZone before booting the OS >> + * and its register space is not write-protected by the hypervisor. >> + * In this case, to achieve CPU DVFS, it is needed to program it from >> + * the OS itself, which includes setting LUT and all the various >> tunables >> + * that are required for it to manage the CPU frequencies and voltages >> + * on its own. >> + * Calling this function on a platform that had the OSM set-up by TZ >> + * will result in a hypervisor fault with system reboot in most cases. >> + * >> + * Returns: Zero for success, otherwise negative number on errors. >> + */ >> +static int qcom_cpufreq_hw_osm_setup(struct device *cpu_dev, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â struct cpufreq_policy *policy, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â int cpu_count, int index) >> +{ >> +Â Â Â struct qcom_cpufreq_data *drv_data = policy->driver_data; >> +Â Â Â const struct qcom_cpufreq_soc_setup_data *setup_regs; >> +Â Â Â u32 val; >> +Â Â Â int ret; >> + >> +Â Â Â ret = qcom_cpufreq_hw_write_lut(cpu_dev, policy, cpu_count, index); >> +Â Â Â if (ret) >> +Â Â Â Â Â Â Â return ret; >> + >> +Â Â Â setup_regs = &drv_data->soc_data->setup_regs; >> + >> +Â Â Â /* Set OSM to XO clock ratio and use XO edge for the cycle >> counter */ >> +Â Â Â val = FIELD_PREP(CYCLE_COUNTER_CLK_RATIO, OSM_XO_RATIO_VAL); >> +Â Â Â val |= CYCLE_COUNTER_USE_XO_EDGE; >> + >> +Â Â Â /* Enable the CPU cycle counter */ >> +Â Â Â val |= BIT(0); >> +Â Â Â writel_relaxed(val, drv_data->base + setup_regs->reg_cycle_counter); >> + >> +Â Â Â /* CoreCount DCVS Policy: Wait time for frequency inc/decrement */ >> +Â Â Â val = FIELD_PREP(HYSTERESIS_UP_MASK, HYSTERESIS_CC_NS); >> +Â Â Â val |= FIELD_PREP(HYSTERESIS_DN_MASK, HYSTERESIS_CC_NS); >> +Â Â Â writel_relaxed(val, drv_data->base + setup_regs->reg_spm_cc_hyst); >> + >> +Â Â Â /* Set the frequency index 0 and override for cluster power >> collapse */ >> +Â Â Â writel_relaxed(BIT(0), drv_data->base + >> setup_regs->reg_cc_zero_behav); >> + >> +Â Â Â /* Treat cores in retention as active */ >> +Â Â Â writel_relaxed(0, drv_data->base + >> setup_regs->reg_spm_core_ret_map); >> + >> +Â Â Â /* Enable CoreCount based DCVS */ >> +Â Â Â writel_relaxed(0, drv_data->base + setup_regs->reg_spm_cc_dcvs_dis); >> + >> +Â Â Â /* CoreCount DCVS-LLM Policy: Wait time for frequency >> inc/decrement */ >> +Â Â Â val = FIELD_PREP(HYSTERESIS_UP_MASK, HYSTERESIS_LLM_NS); >> +Â Â Â val |= FIELD_PREP(HYSTERESIS_DN_MASK, HYSTERESIS_LLM_NS); >> +Â Â Â writel_relaxed(val, drv_data->base + >> setup_regs->reg_llm_freq_vote_hyst); >> + >> +Â Â Â /* CoreCount DCVS-LLM Policy: Wait time for voltage inc/decrement */ >> +Â Â Â val = FIELD_PREP(HYSTERESIS_UP_MASK, HYSTERESIS_LLM_NS); >> +Â Â Â val |= FIELD_PREP(HYSTERESIS_DN_MASK, HYSTERESIS_LLM_NS); >> +Â Â Â writel_relaxed(val, drv_data->base + >> setup_regs->reg_llm_volt_vote_hyst); >> + >> +Â Â Â /* Enable LLM frequency+voltage voting */ >> +Â Â Â writel_relaxed(0, drv_data->base + >> setup_regs->reg_llm_intf_dcvs_dis); >> + >> +Â Â Â /* Setup Boost FSM Timers */ >> +Â Â Â qcom_cpufreq_hw_boost_setup(drv_data->base + >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â setup_regs->reg_cc_boost_timer, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â setup_regs->boost_timer_reg_len); >> +Â Â Â qcom_cpufreq_hw_boost_setup(drv_data->base + >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â setup_regs->reg_dcvs_boost_timer, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â setup_regs->boost_timer_reg_len); >> +Â Â Â qcom_cpufreq_hw_boost_setup(drv_data->base + >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â setup_regs->reg_ps_boost_timer, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â setup_regs->boost_timer_reg_len); >> + >> +Â Â Â /* PLL signal timing control for Boost */ >> +Â Â Â writel_relaxed(BOOST_SYNC_DELAY, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â drv_data->base + setup_regs->reg_boost_sync_delay); >> + >> +Â Â Â /* Setup WFx and PC/RET droop unstall */ >> +Â Â Â val = FIELD_PREP(DROOP_TIMER1, DROOP_TIMER_NS); >> +Â Â Â val |= FIELD_PREP(DROOP_TIMER0, DROOP_TIMER_NS); >> +Â Â Â writel_relaxed(val, drv_data->base + >> setup_regs->reg_droop_unstall_ctrl); >> + >> +Â Â Â /* Setup WFx and PC/RET droop wait-to-release */ >> +Â Â Â val = FIELD_PREP(DROOP_TIMER1, DROOP_WAIT_RELEASE_TIMER_NS); >> +Â Â Â val |= FIELD_PREP(DROOP_TIMER0, DROOP_WAIT_RELEASE_TIMER_NS); >> +Â Â Â writel_relaxed(val, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â drv_data->base + >> setup_regs->reg_droop_wait_release_ctrl); >> + >> +Â Â Â /* PLL signal timing control for Droop */ >> +Â Â Â writel_relaxed(1, drv_data->base + >> setup_regs->reg_droop_sync_delay); >> + >> +Â Â Â /* Setup DCVS timers */ >> +Â Â Â writel_relaxed(DROOP_RELEASE_TIMER_NS, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â drv_data->base + setup_regs->reg_droop_release_ctrl); >> +Â Â Â writel_relaxed(DROOP_TIMER_NS, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â drv_data->base + setup_regs->reg_droop_timer_ctrl); >> + >> +Â Â Â /* Setup Droop control */ >> +Â Â Â val = readl_relaxed(drv_data->base + setup_regs->reg_droop_ctrl); >> +Â Â Â val |= DROOP_CTRL_VAL; >> +Â Â Â writel_relaxed(val, drv_data->base + setup_regs->reg_droop_ctrl); >> + >> +Â Â Â /* Enable CC-Boost, DCVS-Boost, PS-Boost, WFx, PC/RET, DCVS FSM */ >> +Â Â Â val = readl_relaxed(drv_data->base + setup_regs->reg_pdn_fsm_ctrl); >> +Â Â Â val |= CC_BOOST_EN | PS_BOOST_EN | DCVS_BOOST_EN; >> +Â Â Â val |= WFX_DROOP_EN | PC_RET_EXIT_DROOP_EN | DCVS_DROOP_EN; >> +Â Â Â writel_relaxed(val, drv_data->base + setup_regs->reg_pdn_fsm_ctrl); >> + >> +Â Â Â /* Enable PLL Droop Override */ >> +Â Â Â val = PLL_OVERRIDE_DROOP_EN; >> +Â Â Â writel_relaxed(val, drv_data->base + setup_regs->reg_pll_override); >> + >> +Â Â Â /* Initialize the Adaptive Clock Distribution */ >> +Â Â Â ret = qcom_cpufreq_hw_acd_init(cpu_dev, policy, index); >> +Â Â Â if (ret) >> +Â Â Â Â Â Â Â return ret; >> + >> +Â Â Â /* We're ready: enable the OSM and give it time to boot (5uS) */ >> +Â Â Â writel_relaxed(1, drv_data->base + drv_data->soc_data->reg_enable); >> +Â Â Â udelay(OSM_BOOT_TIME_US); >> + >> +Â Â Â return ret; >> +} >> + >> Â static int qcom_cpufreq_hw_cpu_init(struct cpufreq_policy *policy) >> Â { >> Â Â Â Â Â struct platform_device *pdev = cpufreq_get_driver_data(); >> @@ -282,7 +1400,8 @@ static int qcom_cpufreq_hw_cpu_init(struct >> cpufreq_policy *policy) >> Â Â Â Â Â struct device *cpu_dev; >> Â Â Â Â Â void __iomem *base; >> Â Â Â Â Â struct qcom_cpufreq_data *data; >> -Â Â Â int ret, index; >> +Â Â Â const char *fdom_resname; >> +Â Â Â int cpu_count, index, ret; >> Â Â Â Â Â cpu_dev = get_cpu_device(policy->cpu); >> Â Â Â Â Â if (!cpu_dev) { >> @@ -303,9 +1422,14 @@ static int qcom_cpufreq_hw_cpu_init(struct >> cpufreq_policy *policy) >> Â Â Â Â Â index = args.args[0]; >> -Â Â Â base = devm_platform_ioremap_resource(pdev, index); >> +Â Â Â fdom_resname = kasprintf(GFP_KERNEL, "freq-domain%d", index); >> +Â Â Â if (!fdom_resname) >> +Â Â Â Â Â Â Â return -ENOMEM; >> + >> +Â Â Â base = devm_platform_ioremap_resource_byname(pdev, fdom_resname); >> Â Â Â Â Â if (IS_ERR(base)) >> Â Â Â Â Â Â Â Â Â return PTR_ERR(base); >> +Â Â Â kfree(fdom_resname); >> Â Â Â Â Â data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL); >> Â Â Â Â Â if (!data) { >> @@ -315,22 +1439,31 @@ static int qcom_cpufreq_hw_cpu_init(struct >> cpufreq_policy *policy) >> Â Â Â Â Â data->soc_data = of_device_get_match_data(&pdev->dev); >> Â Â Â Â Â data->base = base; >> +Â Â Â policy->driver_data = data; >> -Â Â Â /* HW should be in enabled state to proceed */ >> -Â Â Â if (!(readl_relaxed(base + data->soc_data->reg_enable) & 0x1)) { >> -Â Â Â Â Â Â Â dev_err(dev, "Domain-%d cpufreq hardware not enabled\n", index); >> -Â Â Â Â Â Â Â ret = -ENODEV; >> -Â Â Â Â Â Â Â goto error; >> -Â Â Â } >> - >> -Â Â Â qcom_get_related_cpus(index, policy->cpus); >> +Â Â Â cpu_count = qcom_get_related_cpus(index, policy->cpus); >> Â Â Â Â Â if (!cpumask_weight(policy->cpus)) { >> Â Â Â Â Â Â Â Â Â dev_err(dev, "Domain-%d failed to get related CPUs\n", index); >> Â Â Â Â Â Â Â Â Â ret = -ENOENT; >> Â Â Â Â Â Â Â Â Â goto error; >> Â Â Â Â Â } >> -Â Â Â policy->driver_data = data; >> +Â Â Â if (!data->soc_data->uses_tz) { >> +Â Â Â Â Â Â Â ret = qcom_cpufreq_hw_osm_setup(cpu_dev, policy, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â cpu_count, index); >> +Â Â Â Â Â Â Â if (ret) { >> +Â Â Â Â Â Â Â Â Â Â Â dev_err(dev, "Cannot setup the OSM for CPU%d: %d\n", >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â policy->cpu, ret); >> +Â Â Â Â Â Â Â Â Â Â Â goto error; >> +Â Â Â Â Â Â Â } >> +Â Â Â } >> + >> +Â Â Â /* HW should be in enabled state to proceed */ >> +Â Â Â if (!(readl_relaxed(base + data->soc_data->reg_enable) & 0x1)) { >> +Â Â Â Â Â Â Â dev_err(dev, "Domain-%d cpufreq hardware not enabled\n", index); >> +Â Â Â Â Â Â Â ret = -ENODEV; >> +Â Â Â Â Â Â Â goto error; >> +Â Â Â } >> Â Â Â Â Â ret = qcom_cpufreq_hw_read_lut(cpu_dev, policy); >> Â Â Â Â Â if (ret) { >> @@ -355,6 +1488,7 @@ static int qcom_cpufreq_hw_cpu_init(struct >> cpufreq_policy *policy) >> Â Â Â Â Â return 0; >> Â error: >> +Â Â Â policy->driver_data = NULL; >> Â Â Â Â Â devm_iounmap(dev, base); >> Â Â Â Â Â return ret; >> Â } >> @@ -395,9 +1529,51 @@ static struct cpufreq_driver >> cpufreq_qcom_hw_driver = { >> Â static int qcom_cpufreq_hw_driver_probe(struct platform_device *pdev) >> Â { >> +Â Â Â const struct qcom_cpufreq_soc_data *soc_data; >> +Â Â Â struct device_node *pd_node; >> +Â Â Â struct platform_device *pd_dev; >> Â Â Â Â Â struct device *cpu_dev; >> Â Â Â Â Â struct clk *clk; >> -Â Â Â int ret; >> +Â Â Â int clk_div, ret; >> + >> +Â Â Â cpu_dev = get_cpu_device(0); >> +Â Â Â if (!cpu_dev) >> +Â Â Â Â Â Â Â return -EPROBE_DEFER; >> + >> +Â Â Â soc_data = of_device_get_match_data(&pdev->dev); >> +Â Â Â if (!soc_data) >> +Â Â Â Â Â Â Â return -EINVAL; >> + >> +Â Â Â if (!soc_data->uses_tz) { >> +Â Â Â Â Â Â Â /* >> +Â Â Â Â Â Â Â Â * When the OSM is not pre-programmed from TZ, we will >> +Â Â Â Â Â Â Â Â * need to program the sequencer through SCM calls. >> +Â Â Â Â Â Â Â Â */ >> +Â Â Â Â Â Â Â if (!qcom_scm_is_available()) >> +Â Â Â Â Â Â Â Â Â Â Â return -EPROBE_DEFER; >> + >> +Â Â Â Â Â Â Â /* >> +Â Â Â Â Â Â Â Â * If there are no power-domains, OSM programming cannot be >> +Â Â Â Â Â Â Â Â * performed, as in that case, we wouldn't know where to take >> +Â Â Â Â Â Â Â Â * the params from... >> +Â Â Â Â Â Â Â Â */ >> +Â Â Â Â Â Â Â pd_node = of_parse_phandle(cpu_dev->of_node, >> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â "power-domains", 0); >> +Â Â Â Â Â Â Â if (!pd_node) { >> +Â Â Â Â Â Â Â Â Â Â Â ret = PTR_ERR(pd_node); >> +Â Â Â Â Â Â Â Â Â Â Â dev_err(cpu_dev, "power domain not found: %d\n", ret); >> +Â Â Â Â Â Â Â Â Â Â Â return ret; >> +Â Â Â Â Â Â Â } >> + >> +Â Â Â Â Â Â Â /* >> +Â Â Â Â Â Â Â Â * If the power domain device is not registered yet, then >> +Â Â Â Â Â Â Â Â * defer probing this driver until that is available. >> +Â Â Â Â Â Â Â Â */ >> +Â Â Â Â Â Â Â pd_dev = of_find_device_by_node(pd_node); >> +Â Â Â Â Â Â Â if (!pd_dev || !pd_dev->dev.driver || >> +Â Â Â Â Â Â Â Â Â Â Â !device_is_bound(&pd_dev->dev)) >> +Â Â Â Â Â Â Â Â Â Â Â return -EPROBE_DEFER; >> +Â Â Â } >> Â Â Â Â Â clk = clk_get(&pdev->dev, "xo"); >> Â Â Â Â Â if (IS_ERR(clk)) >> @@ -410,16 +1586,16 @@ static int qcom_cpufreq_hw_driver_probe(struct >> platform_device *pdev) >> Â Â Â Â Â if (IS_ERR(clk)) >> Â Â Â Â Â Â Â Â Â return PTR_ERR(clk); >> -Â Â Â cpu_hw_rate = clk_get_rate(clk) / CLK_HW_DIV; >> +Â Â Â clk_div = soc_data->clk_hw_div; >> +Â Â Â if (clk_div == 0) >> +Â Â Â Â Â Â Â clk_div++; >> + >> +Â Â Â cpu_hw_rate = clk_get_rate(clk) / clk_div; >> Â Â Â Â Â clk_put(clk); >> Â Â Â Â Â cpufreq_qcom_hw_driver.driver_data = pdev; >> Â Â Â Â Â /* Check for optional interconnect paths on CPU0 */ >> -Â Â Â cpu_dev = get_cpu_device(0); >> -Â Â Â if (!cpu_dev) >> -Â Â Â Â Â Â Â return -EPROBE_DEFER; >> - >> Â Â Â Â Â ret = dev_pm_opp_of_find_icc_paths(cpu_dev, NULL); >> Â Â Â Â Â if (ret) >> Â Â Â Â Â Â Â Â Â return ret; >> >
On Wed 20 Jan 12:25 CST 2021, Taniya Das wrote: > The CPUFREQ-HW driver is intended to be used only for CPUFREQ HW designs > where the firmware programs the look up tables. > It's obvious that this is the intended target for the current version of the driver, but what are your technical arguments for keeping it that way? > Suggestion is to separate out the driver where the programming is managed by > high level OS. > Can you please elaborate on the benefits of this approach? PS. Please don't top-post on LKML. Regards, Bjorn > On 1/19/2021 11:15 PM, AngeloGioacchino Del Regno wrote: > > On new SoCs (SDM845 onwards) the Operating State Manager (OSM) is > > being programmed in the bootloader and write-protected by the > > hypervisor, leaving to the OS read-only access to some of its > > registers (in order to read the Lookup Tables and also some > > status registers) and write access to the p-state register, for > > for the OS to request a specific performance state to trigger a > > DVFS switch on the CPU through the OSM hardware. > > > > On old SoCs though (MSM8998, SDM630/660 and variants), the > > bootloader will *not* initialize the OSM (and the CPRh, as it > > is a requirement for it) before booting the OS, making any > > request to trigger a performance state change ineffective, as > > the hardware doesn't have any Lookup Table, nor is storing any > > parameter to trigger a DVFS switch. In this case, basically all > > of the OSM registers are *not* write protected for the OS, even > > though some are - but write access is granted through SCM calls. > > > > This commit introduces support for OSM programming, which has to > > be done on these old SoCs that were distributed (almost?) always > > with a bootloader that does not do any CPRh nor OSM init before > > booting the kernel. > > In order to program the OSM on these SoCs, it is necessary to > > fullfill a "special" requirement: the Core Power Reduction > > Hardened (CPRh) hardware block must be initialized, as the OSM > > is "talking" to it in order to perform the Voltage part of DVFS; > > here, we are calling initialization of this through Linux generic > > power domains, specifically by requesting a genpd attach from the > > qcom-cpufreq-hw driver, which will give back voltages associated > > to each CPU frequency that has been declared in the OPPs, scaled > > and interpolated with the previous one, and will also give us > > parameters for the Array Power Mux (APM) and mem-acc, in order > > for this driver to be then able to generate the Lookup Tables > > that will be finally programmed to the OSM hardware. > > > > After writing the parameters to the OSM and enabling it, all the > > programming work will never happen anymore until a OS reboot, so > > all of the allocations and "the rest" will be disposed-of: this > > is done mainly to leave the code that was referred only to the > > new SoCs intact, as to also emphasize on the fact that the OSM > > HW is, in the end, the exact same; apart some register offsets > > that are slightly different, the entire logic is the same. > > > > This also adds the parameters to support CPU scaling on SDM630 > > and MSM8998. > > > > Signed-off-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@somainline.org> > > --- > > drivers/cpufreq/qcom-cpufreq-hw.c | 1240 ++++++++++++++++++++++++++++- > > 1 file changed, 1208 insertions(+), 32 deletions(-) > > > > diff --git a/drivers/cpufreq/qcom-cpufreq-hw.c b/drivers/cpufreq/qcom-cpufreq-hw.c > > index acc645b85e79..a92959bb7b50 100644 > > --- a/drivers/cpufreq/qcom-cpufreq-hw.c > > +++ b/drivers/cpufreq/qcom-cpufreq-hw.c > > @@ -1,33 +1,256 @@ > > // SPDX-License-Identifier: GPL-2.0 > > /* > > * Copyright (c) 2018, The Linux Foundation. All rights reserved. > > + * > > + * OSM hardware initial programming > > + * Copyright (C) 2020, AngeloGioacchino Del Regno > > + * <angelogioacchino.delregno@somainline.org> > > */ > > #include <linux/bitfield.h> > > #include <linux/cpufreq.h> > > +#include <linux/delay.h> > > #include <linux/init.h> > > #include <linux/interconnect.h> > > #include <linux/kernel.h> > > #include <linux/module.h> > > #include <linux/of_address.h> > > #include <linux/of_platform.h> > > +#include <linux/pm_domain.h> > > #include <linux/pm_opp.h> > > #include <linux/slab.h> > > +#include <linux/qcom_scm.h> > > #define LUT_MAX_ENTRIES 40U > > -#define LUT_SRC GENMASK(31, 30) > > +#define LUT_SRC_845 GENMASK(31, 30) > > +#define LUT_SRC_8998 GENMASK(27, 26) > > +#define LUT_PLL_DIV GENMASK(25, 24) > > #define LUT_L_VAL GENMASK(7, 0) > > #define LUT_CORE_COUNT GENMASK(18, 16) > > +#define LUT_VOLT_VC GENMASK(21, 16) > > #define LUT_VOLT GENMASK(11, 0) > > -#define CLK_HW_DIV 2 > > #define LUT_TURBO_IND 1 > > +#define OSM_BOOT_TIME_US 5 > > + > > +#define CYCLE_COUNTER_CLK_RATIO GENMASK(5, 1) > > +#define OSM_XO_RATIO_VAL (10 - 1) > > +#define CYCLE_COUNTER_USE_XO_EDGE BIT(8) > > + > > +/* FSM Boost Control */ > > +#define CC_BOOST_EN BIT(0) > > +#define PS_BOOST_EN BIT(1) > > +#define DCVS_BOOST_EN BIT(2) > > +#define BOOST_TIMER_REG_HI GENMASK(31, 16) > > +#define BOOST_TIMER_REG_LO GENMASK(15, 0) > > + > > +#define PLL_WAIT_LOCK_TIME_NS 2000 > > +#define SAFE_FREQ_WAIT_NS 1000 > > +#define DEXT_DECREMENT_WAIT_NS 200 > > + > > +#define BOOST_SYNC_DELAY 5 > > + > > +#define HYSTERESIS_UP_MASK GENMASK(31, 16) > > +#define HYSTERESIS_DN_MASK GENMASK(15, 0) > > +#define HYSTERESIS_CC_NS 200 > > +#define HYSTERESIS_LLM_NS 65535 > > + > > +/* FSM Droop Control */ > > +#define PC_RET_EXIT_DROOP_EN BIT(3) > > +#define WFX_DROOP_EN BIT(4) > > +#define DCVS_DROOP_EN BIT(5) > > +#define DROOP_TIMER1 GENMASK(31, 16) > > +#define DROOP_TIMER0 GENMASK(15, 0) > > +#define DROOP_CTRL_VAL (BIT(3) | BIT(17) | BIT(31)) > > +#define DROOP_TIMER_NS 100 > > +#define DROOP_WAIT_RELEASE_TIMER_NS 50 > > +#define DROOP_RELEASE_TIMER_NS 1 > > + > > +/* PLL Override Control */ > > +#define PLL_OVERRIDE_DROOP_EN BIT(0) > > + > > +/* Sequencer */ > > +#define SEQUENCER_REG(base, n) (base + (n * 4)) > > +#define SEQ_APM_THRESH_VC 15 > > +#define SEQ_APM_THRESH_PREVC 31 > > +#define SEQ_MEM_ACC_LVAL 32 > > +#define SEQ_MEM_ACC_0 55 > > +#define SEQ_APM_CROSSOVER_VC 72 > > +#define SEQ_APM_PARAM 76 > > +#define SEQ_MEM_ACC_CROSSOVER_VC 88 > > +#define SEQ_MEM_ACC_MAX_LEVELS 4 > > +#define SEQ_MEMACC_REG(base, n) SEQUENCER_REG(base, SEQ_MEM_ACC_0 + n) > > + > > +/* ACD */ > > +#define ACD_WRITE_CTL_UPDATE_EN BIT(0) > > +#define ACD_WRITE_CTL_SELECT_SHIFT 1 > > + > > +/** > > + * struct qcom_cpufreq_soc_setup_data - Register offsets for OSM setup > > + * > > + * @reg_osm_sequencer: OSM Sequencer (used to get physical address) > > + * @reg_override: Override parameters > > + * @reg_spare: Spare parameters (MEMACC-to-VC) > > + * @reg_cc_zero_behav: Virtual Corner for cluster power collapse > > + * @reg_spm_cc_hyst: DCVS-CC Wait time for frequency inc/decrement > > + * @reg_spm_cc_dcvs_dis: DCVS-CC en/disable control > > + * @reg_spm_core_ret_map: Treat cores in retention as active/inactive > > + * @reg_llm_freq_vote_hyst: DCVS-LLM Wait time for frequency inc/decrement > > + * @reg_llm_volt_vote_hyst: DCVS-LLM Wait time for voltage inc/decrement > > + * @reg_llm_intf_dcvs_dis: DCVS-LLM en/disable control > > + * @reg_seq1: Sequencer extra register > > + * @reg_pdn_fsm_ctrl: Boost and Droop FSMs en/disable control > > + * @reg_cc_boost_timer: CC-Boost FSM wait first timer register > > + * @reg_dcvs_boost_timer: DCVS-Boost FSM wait first timer register > > + * @reg_ps_boost_timer: PS-Boost FSM wait first timer register > > + * @boost_timer_reg_len: Length of boost timer registers > > + * @reg_boost_sync_delay: PLL signal timing control for Boost > > + * @reg_droop_ctrl: Droop control value > > + * @reg_droop_release_ctrl: Wait for Droop release > > + * @reg_droop_unstall_ctrl: Wait for Droop unstall > > + * @reg_droop_wait_release_ctrl: Time to wait for state release > > + * @reg_droop_timer_ctrl: Droop timer > > + * @reg_droop_sync_delay: PLL signal timing control for Droop > > + * @reg_pll_override: PLL Droop Override en/disable control > > + * @reg_cycle_counter: OSM CPU cycle counter > > + * > > + * This structure holds the register offsets that are used to set-up > > + * the Operating State Manager (OSM) parameters, when it is not (or > > + * not entirely) configured from the bootloader and TrustZone. > > + * > > + * Acronyms used in this documentation: > > + * CC = Core Count > > + * PS = Power-Save > > + * VC = Virtual Corner > > + * LLM = Limits Load Management > > + * DCVS = Dynamic Clock and Voltage Scaling > > + */ > > +struct qcom_cpufreq_soc_setup_data { > > + /* OSM phys register offsets */ > > + u16 reg_osm_sequencer; > > + > > + /* Frequency domain register offsets */ > > + u16 reg_override; > > + u16 reg_spare; > > + u16 reg_cc_zero_behav; > > + u16 reg_spm_cc_hyst; > > + u16 reg_spm_cc_dcvs_dis; > > + u16 reg_spm_core_ret_map; > > + u16 reg_llm_freq_vote_hyst; > > + u16 reg_llm_volt_vote_hyst; > > + u16 reg_llm_intf_dcvs_dis; > > + u16 reg_seq1; > > + u16 reg_pdn_fsm_ctrl; > > + u16 reg_cc_boost_timer; > > + u16 reg_dcvs_boost_timer; > > + u16 reg_ps_boost_timer; > > + u16 boost_timer_reg_len; > > + u16 reg_boost_sync_delay; > > + u16 reg_droop_ctrl; > > + u16 reg_droop_release_ctrl; > > + u16 reg_droop_unstall_ctrl; > > + u16 reg_droop_wait_release_ctrl; > > + u16 reg_droop_timer_ctrl; > > + u16 reg_droop_sync_delay; > > + u16 reg_pll_override; > > + u16 reg_cycle_counter; > > +}; > > + > > +/** > > + * struct qcom_cpufreq_soc_acd_data - Adaptive Clock Distribution data > > + * > > + * @tl_delay_reg: Tunable-Length Delay (TLD) register offset > > + * @acd_ctrl_reg: Control Register (CR) register offset > > + * @softstart_reg: Soft Start Control Register (SSCR) register offset > > + * @ext_intf_reg: External interface configuration register offset > > + * @auto_xfer_reg: Auto Register-Transfer register offset > > + * @auto_xfer_cfg_reg: Auto Register-Transfer Configuration reg offset > > + * @auto_xfer_ctl_reg: Auto Register-Transfer Control register offset > > + * @auto_xfer_sts_reg: Auto Register-Transfer Status register offset > > + * @dcvs_sw_reg: Software DCVS register offset > > + * @gfmux_cfg_reg: Glitch-Free MUX configuration register offset > > + * @write_ctl_reg: Write Control register > > + * @write_sts_reg: Write Status register > > + * @tl_delay_val: Tunable-Length Delay (TLD) value > > + * @acd_ctrl_val: Control Register (CR) value > > + * @softstart_val: Soft Start Control Register (SSCR) value > > + * @ext_intf0_val: Initial external interface configuration value > > + * @ext_intf1_val: Final external interface configuration value > > + * @auto_xfer_val: Auto-register Transfer Control value > > + * > > + * This structure holds the register offsets (from the ACD iospace base) > > + * and the parameters that are required to configure the OSM to > > + * initialize the Adaptive Clock Distribution (ACD) system. > > + */ > > +struct qcom_cpufreq_soc_acd_data { > > + u8 tl_delay_reg; > > + u8 acd_ctrl_reg; > > + u8 softstart_reg; > > + u8 ext_intf_reg; > > + u8 auto_xfer_reg; > > + u8 auto_xfer_cfg_reg; > > + u8 auto_xfer_ctl_reg; > > + u8 auto_xfer_sts_reg; > > + u8 dcvs_sw_reg; > > + u8 gfmux_cfg_reg; > > + u8 write_ctl_reg; > > + u8 write_sts_reg; > > + u32 tl_delay_val; > > + u32 acd_ctrl_val; > > + u32 softstart_val; > > + u32 ext_intf0_val; > > + u32 ext_intf1_val; > > + u32 auto_xfer_val; > > +}; > > + > > +/** > > + * struct qcom_cpufreq_hw_params - Operating State Manager (OSM) Parameters > > + * > > + * @volt_lut_val: Value composed of: virtual corner (vc) and voltage in mV. > > + * @freq_lut_val: Value composed of: core count, clock source and output > > + * frequency in MHz. > > + * @override_val: PLL parameters that the OSM uses to override the previous > > + * setting coming from the bootloader, or when uninitialized. > > + * @spare_val: Spare register, used by both this driver and the OSM HW > > + * to identify MEM-ACC levels in relation to virtual corners. > > + * @acc_thresh: MEM-ACC threshold level > > + * > > + * This structure holds the parameters to write to the OSM registers for > > + * one "Virtual Corner" (VC), or one Performance State (p-state). > > + */ > > +struct qcom_cpufreq_hw_params { > > + u32 volt_lut_val; > > + u32 freq_lut_val; > > + u32 override_val; > > + u32 spare_val; > > + u32 acc_thresh; > > +}; > > +/** > > + * struct qcom_cpufreq_soc_data - SoC specific register offsets of the OSM > > + * > > + * @reg_enable: OSM enable status > > + * @reg_index: Index of the Virtual Corner > > + * @reg_freq_lut: Frequency Lookup Table > > + * @reg_freq_lut_src_mask: Frequency Lookup Table clock-source mask > > + * @reg_volt_lut: Voltage Lookup Table > > + * @reg_perf_state: Performance State request register > > + * @lut_row_size: Lookup Table row size > > + * @clk_hw_div: Divider for "alternate" OSM clock-source > > + * @uses_tz: OSM already set-up and protected by TrustZone > > + * @setup_regs: Register offsets for OSM setup > > + */ > > struct qcom_cpufreq_soc_data { > > u32 reg_enable; > > + u32 reg_index; > > u32 reg_freq_lut; > > + u32 reg_freq_lut_src_mask; > > u32 reg_volt_lut; > > u32 reg_perf_state; > > - u8 lut_row_size; > > + u8 lut_row_size; > > + u8 clk_hw_div; > > + bool uses_tz; > > + const struct qcom_cpufreq_soc_setup_data setup_regs; > > + const struct qcom_cpufreq_soc_acd_data acd_data; > > }; > > struct qcom_cpufreq_data { > > @@ -35,9 +258,17 @@ struct qcom_cpufreq_data { > > const struct qcom_cpufreq_soc_data *soc_data; > > }; > > -static unsigned long cpu_hw_rate, xo_rate; > > +static const char *cprh_genpd_names[] = { "cprh", NULL }; > > +static u64 cpu_hw_rate, xo_rate; > > static bool icc_scaling_enabled; > > +/** > > + * qcom_cpufreq_set_bw - Set interconnect bandwidth > > + * @policy: CPUFreq policy structure > > + * @freq_khz: CPU Frequency in KHz > > + * > > + * Returns: Zero for success, otherwise negative value on errors > > + */ > > static int qcom_cpufreq_set_bw(struct cpufreq_policy *policy, > > unsigned long freq_khz) > > { > > @@ -59,6 +290,20 @@ static int qcom_cpufreq_set_bw(struct cpufreq_policy *policy, > > return ret; > > } > > +/** > > + * qcom_cpufreq_update_opp - Update CPU OPP tables > > + * @policy: CPUFreq policy structure > > + * @freq_khz: CPU Frequency for OPP entry in KHz > > + * @volt: CPU Voltage for OPP entry in microvolts > > + * > > + * The CPU frequencies and voltages are being read from the Operating > > + * State Manager (OSM) and the related OPPs, read from DT, need to be > > + * updated to reflect what the hardware will set for each p-state. > > + * If there is no OPP table specified in DT, then this function will > > + * add dynamic ones. > > + * > > + * Returns: Zero for success, otherwise negative value on errors > > + */ > > static int qcom_cpufreq_update_opp(struct device *cpu_dev, > > unsigned long freq_khz, > > unsigned long volt) > > @@ -79,6 +324,17 @@ static int qcom_cpufreq_update_opp(struct device *cpu_dev, > > return dev_pm_opp_enable(cpu_dev, freq_hz); > > } > > +/** > > + * qcom_cpufreq_hw_target_index - Set frequency/voltage > > + * @policy: CPUFreq policy structure > > + * @index: Performance state index to be set > > + * > > + * This function sends a request to the Operating State Manager > > + * to set a Performance State index, so, to set frequency and > > + * voltage for the target CPU/cluster. > > + * > > + * Returns: Always zero > > + */ > > static int qcom_cpufreq_hw_target_index(struct cpufreq_policy *policy, > > unsigned int index) > > { > > @@ -94,6 +350,12 @@ static int qcom_cpufreq_hw_target_index(struct cpufreq_policy *policy, > > return 0; > > } > > +/** > > + * qcom_cpufreq_hw_get - Get current Performance State from OSM > > + * @cpu: CPU number > > + * > > + * Returns: Current CPU/Cluster frequency or zero > > + */ > > static unsigned int qcom_cpufreq_hw_get(unsigned int cpu) > > { > > struct qcom_cpufreq_data *data; > > @@ -127,6 +389,644 @@ static unsigned int qcom_cpufreq_hw_fast_switch(struct cpufreq_policy *policy, > > return policy->freq_table[index].frequency; > > } > > +static void qcom_cpufreq_hw_boost_setup(void __iomem *timer0_addr, u32 len) > > +{ > > + u32 val; > > + > > + /* timer_reg0 */ > > + val = FIELD_PREP(BOOST_TIMER_REG_LO, PLL_WAIT_LOCK_TIME_NS); > > + val |= FIELD_PREP(BOOST_TIMER_REG_HI, SAFE_FREQ_WAIT_NS); > > + writel_relaxed(val, timer0_addr); > > + > > + /* timer_reg1 */ > > + val = FIELD_PREP(BOOST_TIMER_REG_LO, PLL_WAIT_LOCK_TIME_NS); > > + val |= FIELD_PREP(BOOST_TIMER_REG_HI, PLL_WAIT_LOCK_TIME_NS); > > + writel_relaxed(val, timer0_addr + len); > > + > > + /* timer_reg2 */ > > + val = FIELD_PREP(BOOST_TIMER_REG_LO, DEXT_DECREMENT_WAIT_NS); > > + writel_relaxed(val, timer0_addr + (2 * len)); > > +} > > + > > +/* > > + * qcom_cpufreq_gen_params - Generate parameters to send to the hardware > > + * @cpu_dev: CPU device > > + * @data: SoC specific register offsets > > + * @hw_tbl: Pointer to return the array of parameters > > + * @apm_vc: APM Virtual Corner crossover number, returned to the caller > > + * @acc_vc: MEMACC Virtual Corner crossover number, returned to the caller > > + * @cpu_count: Number of CPUs in the frequency domain > > + * > > + * This function allocates a 'qcom_cpufreq_hw_params' parameters table, > > + * fills it and returns it to the consumer, ready to get sent to the HW. > > + * Since the APM threshold is just one > > + * Freeing the table after usage is left to the caller. > > + * > > + * Returns: Number of allocated (and filled) elements in the table, > > + * otherwise negative value on errors. > > + */ > > +static int qcom_cpufreq_gen_params(struct device *cpu_dev, > > + struct qcom_cpufreq_data *data, > > + struct qcom_cpufreq_hw_params **hw_tbl, > > + int *apm_vc, int *acc_vc, int cpu_count) > > +{ > > + struct platform_device *pdev = cpufreq_get_driver_data(); > > + const struct qcom_cpufreq_soc_data *soc_data = data->soc_data; > > + struct device **opp_virt_dev; > > + struct opp_table *genpd_opp_table; > > + struct dev_pm_opp *apm_opp, *acc_opp; > > + unsigned long apm_uV, acc_uV, rate; > > + int i, gpd_opp_cnt, ret; > > + u8 last_acc_corner = 0, prev_spare = 0; > > + > > + /* > > + * Install the OPP table that we get from DT here already, > > + * otherwise it gets really messy as we'd have to manually walk > > + * through the entire OPP DT, manually find frequencies and > > + * also manually exclude supported-hw (for speed-binning). > > + * This also makes it easier for the genpd to add info in it. > > + */ > > + ret = dev_pm_opp_of_add_table(cpu_dev); > > + if (ret) { > > + dev_err(&pdev->dev, "Cannot install CPU OPP table: %d\n", ret); > > + return ret; > > + } > > + > > + /* Get a handle to the genpd virtual device */ > > + genpd_opp_table = dev_pm_opp_attach_genpd(cpu_dev, cprh_genpd_names, > > + &opp_virt_dev); > > + if (IS_ERR(genpd_opp_table)) { > > + ret = PTR_ERR(genpd_opp_table); > > + if (ret != -EPROBE_DEFER) > > + dev_err(&pdev->dev, > > + "Could not attach to pm_domain: %d\n", ret); > > + goto opp_dispose; > > + } > > + > > + /* Get the count of available OPPs coming from the power domain */ > > + gpd_opp_cnt = dev_pm_opp_get_opp_count(cpu_dev); > > + if (gpd_opp_cnt < 2) { > > + ret = gpd_opp_cnt > 0 ? -EINVAL : gpd_opp_cnt; > > + goto detach_gpd; > > + } > > + > > + /* Find the APM threshold disabled OPP and "annotate" the voltage */ > > + apm_opp = dev_pm_opp_find_freq_exact(cpu_dev, 0, false); > > + if (IS_ERR(apm_opp)) { > > + ret = -EINVAL; > > + goto detach_gpd; > > + } > > + > > + /* If we get no APM voltage, the system is going to be unstable */ > > + apm_uV = dev_pm_opp_get_voltage(apm_opp); > > + if (apm_uV == 0) { > > + ret = -EINVAL; > > + goto detach_gpd; > > + } > > + *apm_vc = -1; > > + > > + /* > > + * Find the ACC threshold disabled OPP and "annotate" the voltage: > > + * this is optional, as not every SoC needs it. > > + */ > > + acc_opp = dev_pm_opp_find_freq_exact(cpu_dev, 1, false); > > + if (IS_ERR(acc_opp)) { > > + acc_uV = ULONG_MAX; > > + *acc_vc = U8_MAX; > > + } else { > > + /* > > + * If the ACC OPP is present, this means that we *have* to > > + * set the relative corner on the OSM: if we cannot get the > > + * voltage at that point, the CPU will inevitably freeze. > > + */ > > + acc_uV = dev_pm_opp_get_voltage(acc_opp); > > + if (acc_uV == 0) { > > + ret = -EINVAL; > > + goto detach_gpd; > > + } > > + *acc_vc = -1; > > + } > > + > > + *hw_tbl = devm_kmalloc_array(&pdev->dev, gpd_opp_cnt, > > + sizeof(**hw_tbl), GFP_KERNEL); > > + if (!hw_tbl) { > > + ret = -ENOMEM; > > + goto detach_gpd; > > + } > > + > > + for (i = 0, rate = 0; ; rate++, i++) { > > + struct qcom_cpufreq_hw_params *entry = *hw_tbl + i; > > + struct dev_pm_opp *genpd_opp; > > + struct device_node *np; > > + u32 pll_div, millivolts, f_src; > > + > > + /* Find the next enabled OPP's frequency (ignores APM/ACC) */ > > + genpd_opp = dev_pm_opp_find_freq_ceil(cpu_dev, &rate); > > + if (IS_ERR(genpd_opp)) > > + break; > > + > > + /* Get mandatory and optional properties from the OPP DT */ > > + np = dev_pm_opp_get_of_node(genpd_opp); > > + if (!np) > > + break; > > + > > + if (of_property_read_u32(np, "qcom,pll-override", > > + &entry->override_val)) { > > + of_node_put(np); > > + return -EINVAL; > > + } > > + > > + if (of_property_read_u32(np, "qcom,spare-data", > > + &entry->spare_val)) > > + entry->spare_val = 0; > > + > > + if (of_property_read_u32(np, "qcom,pll-div", &pll_div)) > > + pll_div = 0; > > + > > + of_node_put(np); > > + > > + /* Get voltage in microvolts, then convert to millivolts */ > > + millivolts = dev_pm_opp_get_voltage(genpd_opp); > > + if (millivolts >= apm_uV && *apm_vc < 0) > > + *apm_vc = i; > > + if (millivolts >= acc_uV && *acc_vc < 0) > > + *acc_vc = i; > > + > > + millivolts /= 1000; > > + > > + if (millivolts < 150 || millivolts > 1400) { > > + dev_err(&pdev->dev, > > + "Read invalid voltage: %u.\n", millivolts); > > + return -EINVAL; > > + } > > + > > + /* In the OSM firmware, "Virtual Corner" levels start from 0 */ > > + entry->volt_lut_val = FIELD_PREP(LUT_VOLT_VC, i); > > + entry->volt_lut_val |= FIELD_PREP(LUT_VOLT, millivolts); > > + > > + /* Only the first frequency can have alternate source */ > > + f_src = i ? 1 : 0; > > + f_src <<= ffs(soc_data->reg_freq_lut_src_mask) - 1; > > + entry->freq_lut_val = f_src | div_u64(rate, xo_rate); > > + entry->freq_lut_val |= FIELD_PREP(LUT_CORE_COUNT, cpu_count); > > + > > + /* > > + * PLL divider is not always 0 and there is no way to determine > > + * it automatically, as setting this value higher than DIV1 > > + * will make the OSM HW to effectively set the PLL at 2-4x > > + * the CPU frequency and then divide the CPU clock by this div, > > + * so this value is effectively used as both a multiplier and > > + * divider. > > + * This value cannot be calculated because it depends on > > + * manual calibration and is (most probably) used to choose > > + * a PLL frequency that gives the least possible jitter. > > + */ > > + entry->freq_lut_val |= FIELD_PREP(LUT_PLL_DIV, pll_div); > > + > > + /* > > + * MEM-ACC Virtual Corner threshold voltage: set the > > + * acc_thresh to the "next different" spare_val, otherwise > > + * set it to an invalid value, so that we can recognize it > > + * and exclude it later, when this set of data will be used > > + * for programming the OSM. > > + */ > > + if (entry->spare_val != prev_spare) { > > + prev_spare = entry->spare_val; > > + entry->acc_thresh = last_acc_corner; > > + } else { > > + entry->acc_thresh = SEQ_MEM_ACC_MAX_LEVELS + 1; > > + } > > + > > + dev_dbg(&pdev->dev, > > + "[%d] freq=0x%x volt=0x%x override=0x%x spare=0x%x\n", > > + i, entry->freq_lut_val, entry->volt_lut_val, > > + entry->override_val, entry->spare_val); > > + dev_pm_opp_put(genpd_opp); > > + } > > + > > + /* > > + * If we've got a customized mem-acc corner but we couldn't > > + * find any suitable crossover, or the corner is less than > > + * the minimum amount of required corners for mem-acc scaling, > > + * the values are not valid, hence fall back to LUT values. > > + */ > > + if (acc_uV != ULONG_MAX && *acc_vc < SEQ_MEM_ACC_MAX_LEVELS - 1) { > > + dev_dbg(&pdev->dev, > > + "MEM-ACC corner: invalid values VC%d %luuV\n", > > + *acc_vc, acc_uV); > > + *acc_vc = U8_MAX; > > + } > > + > > + /* > > + * If we have probed less params than what we need, then the > > + * OPP table that we got from the genpd is malformed for some > > + * reason: in this case, do not apply the table to the HW. > > + */ > > + if (i < gpd_opp_cnt) { > > + dev_err(&pdev->dev, "Got bad OPP table from power domain.\n"); > > + ret = -EINVAL; > > + goto detach_gpd; > > + } > > + > > +detach_gpd: > > + dev_pm_opp_detach_genpd(genpd_opp_table); > > +opp_dispose: > > + /* > > + * Now that we're totally done with it, dispose of all the dynamic > > + * OPPs in the table: like this, at the end of the OSM configuration > > + * we are leaving it like it was magically configured by the TZ or > > + * by the bootloader and using the rest of the driver as if this > > + * programming phase has never happened in the OS, like new SoCs. > > + * This also makes us able to not modify a single bit in the basic > > + * OSM frequency request logic that was meant for the newest SoCs. > > + */ > > + dev_pm_opp_remove_all_dynamic(cpu_dev); > > + return ret < 0 ? ret : i; > > +} > > + > > +static inline u32 qcom_cpufreq_acd_regbit(u8 acd_reg_offset) > > +{ > > + return BIT((acd_reg_offset / 4)); > > +} > > + > > +static int qcom_cpufreq_hw_acd_write_autoxfer(struct qcom_cpufreq_data *data, > > + void __iomem *acd_base, u32 val) > > +{ > > + const struct qcom_cpufreq_soc_data *sdata = data->soc_data; > > + const struct qcom_cpufreq_soc_acd_data *aregs = &sdata->acd_data; > > + u32 regval; > > + int retry = 3; > > + > > + writel_relaxed(val, acd_base + aregs->auto_xfer_cfg_reg); > > + > > + /* (Clear, then) Set AUTOXFER START */ > > + writel_relaxed(0, acd_base + aregs->auto_xfer_reg); > > + writel_relaxed(1, acd_base + aregs->auto_xfer_reg); > > + > > + /* Poll for status: if the first bit is set the transfer is done. */ > > + do { > > + regval = readl_relaxed(acd_base + aregs->auto_xfer_sts_reg); > > + if (regval & BIT(0)) > > + break; > > + udelay(1); > > + } while (--retry); > > + > > + if (retry <= 0) > > + return -ETIMEDOUT; > > + return 0; > > +} > > + > > +static int qcom_cpufreq_hw_acd_write_xfer(struct qcom_cpufreq_data *data, > > + void __iomem *acd_base, u8 reg, > > + u32 val) > > +{ > > + const struct qcom_cpufreq_soc_data *sdata = data->soc_data; > > + const struct qcom_cpufreq_soc_acd_data *aregs = &sdata->acd_data; > > + int retry = 2; > > + u32 regval; > > + > > + /* Write to the register, then initiate manual transfer */ > > + writel_relaxed(val, acd_base + reg); > > + > > + /* > > + * Make sure that the write goes through, otherwise the local > > + * transfer process will wrongly copy the old value. > > + */ > > + wmb(); > > + > > + /* Clear write control register */ > > + writel_relaxed(0, acd_base + aregs->write_ctl_reg); > > + > > + regval = (reg / 4) << ACD_WRITE_CTL_SELECT_SHIFT; > > + regval |= ACD_WRITE_CTL_UPDATE_EN; > > + writel_relaxed(regval, acd_base + aregs->write_ctl_reg); > > + > > + /* Wait until ACD Local Transfer is done */ > > + do { > > + regval = readl_relaxed(acd_base + aregs->write_sts_reg); > > + if (regval & qcom_cpufreq_acd_regbit(reg)) > > + break; > > + udelay(1); > > + } while (--retry); > > + > > + if (retry <= 0) > > + return -ETIMEDOUT; > > + return 0; > > +} > > + > > +/** > > + * qcom_cpufreq_hw_acd_init - Initialize ACD params in the OSM > > + * @cpu_dev: CPU device > > + * @policy: CPUFreq policy structure > > + * @index: Instance number (CPU cluster number) > > + * > > + * On some SoCs it is required to send the ACD configuration parameters > > + * to the OSM. This function takes the parameters from the SoC specific > > + * configuration and writes them only if a "osm-acdN" iospace has been > > + * declared (hence, it's present). > > + * > > + * Returns: Zero for success, otherwise negative number on error. > > + */ > > +static int qcom_cpufreq_hw_acd_init(struct device *cpu_dev, > > + struct cpufreq_policy *policy, > > + int index) > > +{ > > + struct platform_device *pdev = cpufreq_get_driver_data(); > > + struct qcom_cpufreq_data *ddata = policy->driver_data; > > + const struct qcom_cpufreq_soc_data *sdata = ddata->soc_data; > > + const struct qcom_cpufreq_soc_acd_data *aregs = &sdata->acd_data; > > + const char *acd_resname; > > + void __iomem *acd_base; > > + u32 rmask; > > + int ret; > > + > > + acd_resname = kasprintf(GFP_KERNEL, "osm-acd%d", index); > > + if (!acd_resname) > > + return -ENOMEM; > > + > > + acd_base = devm_platform_ioremap_resource_byname(pdev, acd_resname); > > + kfree(acd_resname); > > + if (IS_ERR(acd_base)) { > > + dev_vdbg(cpu_dev, "Skipping ACD initialization.\n"); > > + return 0; > > + } > > + > > + writel_relaxed(aregs->tl_delay_val, acd_base + aregs->tl_delay_reg); > > + writel_relaxed(aregs->acd_ctrl_val, acd_base + aregs->acd_ctrl_reg); > > + writel_relaxed(aregs->softstart_val, acd_base + aregs->softstart_reg); > > + writel_relaxed(aregs->ext_intf0_val, acd_base + aregs->ext_intf_reg); > > + writel_relaxed(aregs->auto_xfer_val, > > + acd_base + aregs->auto_xfer_ctl_reg); > > + /* > > + * Before initiating register auto transfer, make sure that all > > + * the writes go through, otherwise bad values will be transferred. > > + */ > > + wmb(); > > + > > + rmask = qcom_cpufreq_acd_regbit(aregs->acd_ctrl_reg) | > > + qcom_cpufreq_acd_regbit(aregs->tl_delay_reg) | > > + qcom_cpufreq_acd_regbit(aregs->softstart_reg) | > > + qcom_cpufreq_acd_regbit(aregs->ext_intf_reg); > > + ret = qcom_cpufreq_hw_acd_write_autoxfer(ddata, acd_base, rmask); > > + if (ret) > > + return ret; > > + > > + /* Switch CPUSS clock source to ACD clock */ > > + ret = qcom_cpufreq_hw_acd_write_xfer(ddata, acd_base, > > + aregs->gfmux_cfg_reg, 1); > > + if (ret) > > + return ret; > > + > > + /* (Set, then) Clear DCVS_SW */ > > + ret = qcom_cpufreq_hw_acd_write_xfer(ddata, acd_base, > > + aregs->dcvs_sw_reg, 1); > > + if (ret) > > + return ret; > > + ret = qcom_cpufreq_hw_acd_write_xfer(ddata, acd_base, > > + aregs->dcvs_sw_reg, 0); > > + if (ret) > > + return ret; > > + > > + /* Wait for clock switch time */ > > + udelay(1); > > + > > + /* Program the final ACD external interface */ > > + ret = qcom_cpufreq_hw_acd_write_xfer(ddata, acd_base, > > + aregs->ext_intf_reg, > > + aregs->ext_intf1_val); > > + if (ret) > > + return ret; > > + > > + /* Initiate transfer of the final ACD value */ > > + rmask |= qcom_cpufreq_acd_regbit(aregs->gfmux_cfg_reg); > > + writel_relaxed(rmask, acd_base + aregs->auto_xfer_cfg_reg); > > + > > + /* Wait for ACD to stabilize. Same wait as the OSM boot time... */ > > + udelay(OSM_BOOT_TIME_US); > > + return ret; > > +} > > + > > +/** > > + * qcom_cpufreq_hw_write_lut - Write Lookup Table params to the OSM > > + * @cpu_dev: CPU device > > + * @policy: CPUFreq policy structure > > + * @cpu_count: Number of CPUs in the frequency domain > > + * @index: Instance number (CPU cluster number) > > + * > > + * Program all the Lookup Table (LUT) entries and related thresholds > > + * to the Operating State Manager on platforms where the same hasn't > > + * been done already by the bootloader or TrustZone before booting > > + * the operating system's kernel; > > + * On these platforms, write access to the OSM is (obviously) not > > + * blocked by the hypervisor. > > + * > > + * Returns: Zero for success, otherwise negative number on error. > > + */ > > +static int qcom_cpufreq_hw_write_lut(struct device *cpu_dev, > > + struct cpufreq_policy *policy, > > + int cpu_count, int index) > > +{ > > + struct platform_device *pdev = cpufreq_get_driver_data(); > > + struct qcom_cpufreq_data *ddata = policy->driver_data; > > + const struct qcom_cpufreq_soc_data *sdata = ddata->soc_data; > > + const struct qcom_cpufreq_soc_setup_data *sregs = &sdata->setup_regs; > > + struct qcom_cpufreq_hw_params *hw_tbl; > > + struct resource *osm_rsrc; > > + const char *osm_resname; > > + u32 sreg, seq_addr, acc_lval = 0, last_spare = 1; > > + int apm_vc = INT_MAX, acc_vc = U8_MAX, acc_idx = 0; > > + int acc_val[SEQ_MEM_ACC_MAX_LEVELS]; > > + int i, ret, num_entries; > > + > > + osm_resname = kasprintf(GFP_KERNEL, "osm-domain%d", index); > > + if (!osm_resname) > > + return -ENOMEM; > > + > > + /* > > + * On some SoCs the OSM is not getting programmed from bootloader > > + * and needs to be done here: in this case, we need to retrieve > > + * the base physical address for the "Sequencer", so we will get > > + * the OSM base phys and apply the sequencer offset. > > + * > > + * Note: We are not remapping this iospace because we are really > > + * sending the physical address through SCM calls later. > > + */ > > + osm_rsrc = platform_get_resource_byname(pdev, IORESOURCE_MEM, > > + osm_resname); > > + kfree(osm_resname); > > + > > + if (!osm_rsrc) > > + return -ENODEV; > > + > > + seq_addr = osm_rsrc->start + sregs->reg_osm_sequencer; > > + > > + num_entries = qcom_cpufreq_gen_params(cpu_dev, ddata, &hw_tbl, > > + &apm_vc, &acc_vc, cpu_count); > > + if (num_entries < 0) > > + return num_entries; > > + > > + for (i = 0; i < LUT_MAX_ENTRIES; i++) { > > + struct qcom_cpufreq_hw_params *entry; > > + int pos = i * sdata->lut_row_size; > > + > > + /* > > + * If we have reached the end of the params table, write > > + * the last valid entry until the end of the OSM table. > > + */ > > + if (i < num_entries) > > + entry = &hw_tbl[i]; > > + else > > + entry = &hw_tbl[num_entries - 1]; > > + > > + writel_relaxed(i, ddata->base + sdata->reg_index + pos); > > + > > + writel_relaxed(entry->volt_lut_val, > > + ddata->base + sdata->reg_volt_lut + pos); > > + > > + writel_relaxed(entry->freq_lut_val, > > + ddata->base + sdata->reg_freq_lut + pos); > > + > > + writel_relaxed(entry->override_val, > > + ddata->base + sregs->reg_override + pos); > > + > > + writel_relaxed(entry->spare_val, > > + ddata->base + sregs->reg_spare + pos); > > + > > + dev_dbg(cpu_dev, > > + "Writing [%d] v:0x%x f:0x%x ovr:0x%x s:0x%x\n", i, > > + entry->volt_lut_val, entry->freq_lut_val, > > + entry->override_val, entry->spare_val); > > + > > + /* > > + * MEM-ACC Virtual Corner threshold voltage: this gets set > > + * as the pairs of corners in which there is a transition > > + * between one MEM-ACC level and the next one. > > + * > > + * Notes: The spare_val can never be zero; > > + * The first spare_val is always 1; > > + * The maximum number of pairs is two (four registers). > > + * > > + * Example: (C = Corner Level - M = MEM-ACC Level) > > + * C0 M1 - C1 M1 - C2 M2 - C3 M2 - C4 M2 - C5 M3 > > + * Pairs: 1-2, 4-5 > > + */ > > + if (entry->spare_val <= last_spare || > > + acc_idx >= SEQ_MEM_ACC_MAX_LEVELS - 1) > > + continue; > > + > > + /* Standard mem-acc pairs using spare_val LUT crossovers */ > > + last_spare = entry->spare_val; > > + acc_val[acc_idx] = i - 1; > > + acc_idx++; > > + acc_val[acc_idx] = i; > > + acc_idx++; > > + } > > + > > + /* Sanity check: we *must* have two mem-acc crossovers (four values) */ > > + if (acc_idx < SEQ_MEM_ACC_MAX_LEVELS - 1) > > + return -EINVAL; > > + > > + /* > > + * Customized mem-acc corners, if any; in this case, the last corner > > + * in the external (CPRh) LUT is this one, placed after the APM one. > > + */ > > + if (acc_vc != U8_MAX) { > > + sreg = SEQUENCER_REG(seq_addr, SEQ_MEM_ACC_CROSSOVER_VC); > > + ret = qcom_scm_io_writel(sreg, num_entries + 1); > > + if (ret) > > + return ret; > > + > > + /* > > + * At the price of very-slightly higher power consumption, > > + * switch the ACC at one corner lower than what we've found, > > + * as this seems to be needed on at least some MSM8998 chips > > + * to achieve full system stability > > + */ > > + acc_vc--; > > + > > + /* Change only if we have to move the corner down */ > > + if (acc_vc < acc_val[3]) { > > + acc_val[2] = acc_vc - 1; > > + acc_val[3] = acc_vc; > > + } > > + > > + /* If needed, sanitize previously stored vals from the LUT */ > > + if (acc_val[2] <= acc_val[1]) > > + acc_val[1] = acc_val[2] - 1; > > + if (acc_val[1] <= acc_val[0]) > > + acc_val[0] = acc_val[1] - 1; > > + } > > + > > + for (i = 0; i < SEQ_MEM_ACC_MAX_LEVELS; i++) { > > + ret = qcom_scm_io_writel(SEQ_MEMACC_REG(seq_addr, i), > > + acc_val[i]); > > + if (ret) > > + return ret; > > + } > > + dev_dbg(cpu_dev, "Wrote MEM-ACC Pairs: [%u-%u] [%u-%u]\n", > > + acc_val[0], acc_val[1], acc_val[2], acc_val[3]); > > + > > + /* > > + * Program the L_VAL of the first corner requesting MEM-ACC > > + * voltage level 3 to the right sequencer register > > + */ > > + acc_lval = FIELD_GET(LUT_L_VAL, hw_tbl[acc_val[3]].freq_lut_val); > > + ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_MEM_ACC_LVAL), > > + acc_lval); > > + if (ret) { > > + dev_dbg(cpu_dev, "Cannot send memacc l_val\n"); > > + return ret; > > + } > > + dev_dbg(cpu_dev, "MEM-ACC L-Val is %u\n", acc_lval); > > + > > + /* > > + * Array Power Mux threshold level: the first virtual corner > > + * that requires a switch sequence of the APM from MX to APC. > > + */ > > + if (apm_vc == INT_MAX) > > + apm_vc = LUT_MAX_ENTRIES - 1; > > + > > + /* > > + * APM crossover virtual corner refers to CPRh: there, the APM corner > > + * is always appended to the table (so, at the end of it, right after > > + * the cluster dvfs entries). > > + */ > > + writel(num_entries, ddata->base + sregs->reg_seq1); > > + ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_APM_CROSSOVER_VC), > > + num_entries); > > + if (ret) > > + return ret; > > + > > + ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_APM_THRESH_VC), > > + apm_vc); > > + if (ret) > > + return ret; > > + > > + ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_APM_THRESH_PREVC), > > + apm_vc - 1); > > + if (ret) > > + return ret; > > + > > + ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_APM_PARAM), > > + (0x39 | apm_vc << 6)); > > + if (ret) > > + return ret; > > + > > + dev_dbg(cpu_dev, "Wrote APM Pair: [%u-%u]\n", apm_vc - 1, apm_vc); > > + return ret; > > +} > > + > > +/** > > + * qcom_cpufreq_hw_read_lut - Read Lookup Table from the OSM > > + * @cpu_dev: CPU device > > + * @policy: CPUFreq policy structure > > + * > > + * The Operating State Manager Lookup Table can always be read, even > > + * in case it was pre-programmed by the bootloader or by TrustZone. > > + * Read the LUT from it in order to build OPPs containing DVFS info. > > + * > > + * Returns: Zero for success, otherwise negative number on errors. > > + */ > > static int qcom_cpufreq_hw_read_lut(struct device *cpu_dev, > > struct cpufreq_policy *policy) > > { > > @@ -166,7 +1066,9 @@ static int qcom_cpufreq_hw_read_lut(struct device *cpu_dev, > > for (i = 0; i < LUT_MAX_ENTRIES; i++) { > > data = readl_relaxed(drv_data->base + soc_data->reg_freq_lut + > > i * soc_data->lut_row_size); > > - src = FIELD_GET(LUT_SRC, data); > > + src = data & soc_data->reg_freq_lut_src_mask; > > + src >>= ffs(soc_data->reg_freq_lut_src_mask) - 1; > > + > > lval = FIELD_GET(LUT_L_VAL, data); > > core_count = FIELD_GET(LUT_CORE_COUNT, data); > > @@ -175,17 +1077,21 @@ static int qcom_cpufreq_hw_read_lut(struct device *cpu_dev, > > volt = FIELD_GET(LUT_VOLT, data) * 1000; > > if (src) > > - freq = xo_rate * lval / 1000; > > + freq = xo_rate * lval; > > else > > - freq = cpu_hw_rate / 1000; > > + freq = cpu_hw_rate; > > + freq /= 1000; > > if (freq != prev_freq && core_count != LUT_TURBO_IND) { > > if (!qcom_cpufreq_update_opp(cpu_dev, freq, volt)) { > > table[i].frequency = freq; > > - dev_dbg(cpu_dev, "index=%d freq=%d, core_count %d\n", i, > > - freq, core_count); > > + dev_dbg(cpu_dev, > > + "index=%d freq=%d, core_count %d\n", > > + i, freq, core_count); > > } else { > > - dev_warn(cpu_dev, "failed to update OPP for freq=%d\n", freq); > > + dev_warn(cpu_dev, > > + "failed to update OPP for freq=%d\n", > > + freq); > > table[i].frequency = CPUFREQ_ENTRY_INVALID; > > } > > @@ -205,18 +1111,19 @@ static int qcom_cpufreq_hw_read_lut(struct device *cpu_dev, > > * as the boost frequency > > */ > > if (prev->frequency == CPUFREQ_ENTRY_INVALID) { > > - if (!qcom_cpufreq_update_opp(cpu_dev, prev_freq, volt)) { > > + if (!qcom_cpufreq_update_opp(cpu_dev, prev_freq, > > + volt)) { > > prev->frequency = prev_freq; > > prev->flags = CPUFREQ_BOOST_FREQ; > > } else { > > - dev_warn(cpu_dev, "failed to update OPP for freq=%d\n", > > + dev_warn(cpu_dev, > > + "can't update OPP for freq=%u\n", > > freq); > > } > > } > > break; > > } > > - > > prev_freq = freq; > > } > > @@ -227,10 +1134,18 @@ static int qcom_cpufreq_hw_read_lut(struct device *cpu_dev, > > return 0; > > } > > -static void qcom_get_related_cpus(int index, struct cpumask *m) > > +/* > > + * qcom_get_related_cpus - Get mask of CPUs in the same frequency domain > > + * @index: CPU number > > + * @m: Returned CPU mask > > + * > > + * Returns: Count of CPUs inserted in the cpumask or negative number for error. > > + */ > > +static int qcom_get_related_cpus(int index, struct cpumask *m) > > { > > struct device_node *cpu_np; > > struct of_phandle_args args; > > + int count = 0; > > int cpu, ret; > > for_each_possible_cpu(cpu) { > > @@ -245,34 +1160,237 @@ static void qcom_get_related_cpus(int index, struct cpumask *m) > > if (ret < 0) > > continue; > > - if (index == args.args[0]) > > + if (index == args.args[0]) { > > cpumask_set_cpu(cpu, m); > > + count++; > > + } > > } > > + > > + return count > 0 ? count : -EINVAL; > > } > > static const struct qcom_cpufreq_soc_data qcom_soc_data = { > > .reg_enable = 0x0, > > .reg_freq_lut = 0x110, > > + .reg_freq_lut_src_mask = LUT_SRC_845, > > .reg_volt_lut = 0x114, > > .reg_perf_state = 0x920, > > .lut_row_size = 32, > > + .clk_hw_div = 2, > > + .uses_tz = true, > > +}; > > + > > +static const struct qcom_cpufreq_soc_data msm8998_soc_data = { > > + .reg_enable = 0x4, > > + .reg_index = 0x150, > > + .reg_freq_lut = 0x154, > > + .reg_freq_lut_src_mask = LUT_SRC_8998, > > + .reg_volt_lut = 0x158, > > + .reg_perf_state = 0xf10, > > + .lut_row_size = 32, > > + .clk_hw_div = 1, > > + .uses_tz = false, > > + .setup_regs = { > > + /* Physical offset for sequencer scm calls */ > > + .reg_osm_sequencer = 0x300, > > + > > + /* Frequency domain offsets */ > > + .reg_override = 0x15c, > > + .reg_spare = 0x164, > > + .reg_cc_zero_behav = 0x0c, > > + .reg_spm_cc_hyst = 0x1c, > > + .reg_spm_cc_dcvs_dis = 0x20, > > + .reg_spm_core_ret_map = 0x24, > > + .reg_llm_freq_vote_hyst = 0x2c, > > + .reg_llm_volt_vote_hyst = 0x30, > > + .reg_llm_intf_dcvs_dis = 0x34, > > + .reg_seq1 = 0x48, > > + .reg_pdn_fsm_ctrl = 0x70, > > + .reg_cc_boost_timer = 0x74, > > + .reg_dcvs_boost_timer = 0x84, > > + .reg_ps_boost_timer = 0x94, > > + .boost_timer_reg_len = 0x4, > > + .reg_boost_sync_delay = 0xa0, > > + .reg_droop_ctrl = 0xa4, > > + .reg_droop_release_ctrl = 0xa8, > > + .reg_droop_unstall_ctrl = 0xac, > > + .reg_droop_wait_release_ctrl = 0xb0, > > + .reg_droop_timer_ctrl = 0xb8, > > + .reg_droop_sync_delay = 0xbc, > > + .reg_pll_override = 0xc0, > > + .reg_cycle_counter = 0xf00, > > + }, > > + .acd_data = { > > + .acd_ctrl_reg = 0x4, > > + .tl_delay_reg = 0x8, > > + .softstart_reg = 0x28, > > + .ext_intf_reg = 0x30, > > + .dcvs_sw_reg = 0x34, > > + .gfmux_cfg_reg = 0x3c, > > + .auto_xfer_cfg_reg = 0x80, > > + .auto_xfer_reg = 0x84, > > + .auto_xfer_ctl_reg = 0x88, > > + .auto_xfer_sts_reg = 0x8c, > > + .write_ctl_reg = 0x90, > > + .write_sts_reg = 0x94, > > + .tl_delay_val = 38417, > > + .acd_ctrl_val = 0x2b5ffd, > > + .softstart_val = 0x501, > > + .ext_intf0_val = 0x2cf9ae8, > > + .ext_intf1_val = 0x2cf9afe, > > + .auto_xfer_val = 0x15, > > + }, > > }; > > static const struct qcom_cpufreq_soc_data epss_soc_data = { > > .reg_enable = 0x0, > > .reg_freq_lut = 0x100, > > + .reg_freq_lut_src_mask = LUT_SRC_845, > > .reg_volt_lut = 0x200, > > .reg_perf_state = 0x320, > > .lut_row_size = 4, > > + .clk_hw_div = 2, > > + .uses_tz = true, > > }; > > static const struct of_device_id qcom_cpufreq_hw_match[] = { > > { .compatible = "qcom,cpufreq-hw", .data = &qcom_soc_data }, > > + { .compatible = "qcom,cpufreq-hw-8998", .data = &msm8998_soc_data }, > > { .compatible = "qcom,cpufreq-epss", .data = &epss_soc_data }, > > {} > > }; > > MODULE_DEVICE_TABLE(of, qcom_cpufreq_hw_match); > > +/** > > + * qcom_cpufreq_hw_osm_setup - Setup and enable the OSM > > + * @cpu_dev: CPU device > > + * @policy: CPUFreq policy structure > > + * @cpu_count: Number of CPUs in the frequency domain > > + * > > + * On some platforms, the Operating State Manager (OSM) is not getting > > + * programmed by the bootloader, nor by TrustZone before booting the OS > > + * and its register space is not write-protected by the hypervisor. > > + * In this case, to achieve CPU DVFS, it is needed to program it from > > + * the OS itself, which includes setting LUT and all the various tunables > > + * that are required for it to manage the CPU frequencies and voltages > > + * on its own. > > + * Calling this function on a platform that had the OSM set-up by TZ > > + * will result in a hypervisor fault with system reboot in most cases. > > + * > > + * Returns: Zero for success, otherwise negative number on errors. > > + */ > > +static int qcom_cpufreq_hw_osm_setup(struct device *cpu_dev, > > + struct cpufreq_policy *policy, > > + int cpu_count, int index) > > +{ > > + struct qcom_cpufreq_data *drv_data = policy->driver_data; > > + const struct qcom_cpufreq_soc_setup_data *setup_regs; > > + u32 val; > > + int ret; > > + > > + ret = qcom_cpufreq_hw_write_lut(cpu_dev, policy, cpu_count, index); > > + if (ret) > > + return ret; > > + > > + setup_regs = &drv_data->soc_data->setup_regs; > > + > > + /* Set OSM to XO clock ratio and use XO edge for the cycle counter */ > > + val = FIELD_PREP(CYCLE_COUNTER_CLK_RATIO, OSM_XO_RATIO_VAL); > > + val |= CYCLE_COUNTER_USE_XO_EDGE; > > + > > + /* Enable the CPU cycle counter */ > > + val |= BIT(0); > > + writel_relaxed(val, drv_data->base + setup_regs->reg_cycle_counter); > > + > > + /* CoreCount DCVS Policy: Wait time for frequency inc/decrement */ > > + val = FIELD_PREP(HYSTERESIS_UP_MASK, HYSTERESIS_CC_NS); > > + val |= FIELD_PREP(HYSTERESIS_DN_MASK, HYSTERESIS_CC_NS); > > + writel_relaxed(val, drv_data->base + setup_regs->reg_spm_cc_hyst); > > + > > + /* Set the frequency index 0 and override for cluster power collapse */ > > + writel_relaxed(BIT(0), drv_data->base + setup_regs->reg_cc_zero_behav); > > + > > + /* Treat cores in retention as active */ > > + writel_relaxed(0, drv_data->base + setup_regs->reg_spm_core_ret_map); > > + > > + /* Enable CoreCount based DCVS */ > > + writel_relaxed(0, drv_data->base + setup_regs->reg_spm_cc_dcvs_dis); > > + > > + /* CoreCount DCVS-LLM Policy: Wait time for frequency inc/decrement */ > > + val = FIELD_PREP(HYSTERESIS_UP_MASK, HYSTERESIS_LLM_NS); > > + val |= FIELD_PREP(HYSTERESIS_DN_MASK, HYSTERESIS_LLM_NS); > > + writel_relaxed(val, drv_data->base + setup_regs->reg_llm_freq_vote_hyst); > > + > > + /* CoreCount DCVS-LLM Policy: Wait time for voltage inc/decrement */ > > + val = FIELD_PREP(HYSTERESIS_UP_MASK, HYSTERESIS_LLM_NS); > > + val |= FIELD_PREP(HYSTERESIS_DN_MASK, HYSTERESIS_LLM_NS); > > + writel_relaxed(val, drv_data->base + setup_regs->reg_llm_volt_vote_hyst); > > + > > + /* Enable LLM frequency+voltage voting */ > > + writel_relaxed(0, drv_data->base + setup_regs->reg_llm_intf_dcvs_dis); > > + > > + /* Setup Boost FSM Timers */ > > + qcom_cpufreq_hw_boost_setup(drv_data->base + > > + setup_regs->reg_cc_boost_timer, > > + setup_regs->boost_timer_reg_len); > > + qcom_cpufreq_hw_boost_setup(drv_data->base + > > + setup_regs->reg_dcvs_boost_timer, > > + setup_regs->boost_timer_reg_len); > > + qcom_cpufreq_hw_boost_setup(drv_data->base + > > + setup_regs->reg_ps_boost_timer, > > + setup_regs->boost_timer_reg_len); > > + > > + /* PLL signal timing control for Boost */ > > + writel_relaxed(BOOST_SYNC_DELAY, > > + drv_data->base + setup_regs->reg_boost_sync_delay); > > + > > + /* Setup WFx and PC/RET droop unstall */ > > + val = FIELD_PREP(DROOP_TIMER1, DROOP_TIMER_NS); > > + val |= FIELD_PREP(DROOP_TIMER0, DROOP_TIMER_NS); > > + writel_relaxed(val, drv_data->base + setup_regs->reg_droop_unstall_ctrl); > > + > > + /* Setup WFx and PC/RET droop wait-to-release */ > > + val = FIELD_PREP(DROOP_TIMER1, DROOP_WAIT_RELEASE_TIMER_NS); > > + val |= FIELD_PREP(DROOP_TIMER0, DROOP_WAIT_RELEASE_TIMER_NS); > > + writel_relaxed(val, > > + drv_data->base + setup_regs->reg_droop_wait_release_ctrl); > > + > > + /* PLL signal timing control for Droop */ > > + writel_relaxed(1, drv_data->base + setup_regs->reg_droop_sync_delay); > > + > > + /* Setup DCVS timers */ > > + writel_relaxed(DROOP_RELEASE_TIMER_NS, > > + drv_data->base + setup_regs->reg_droop_release_ctrl); > > + writel_relaxed(DROOP_TIMER_NS, > > + drv_data->base + setup_regs->reg_droop_timer_ctrl); > > + > > + /* Setup Droop control */ > > + val = readl_relaxed(drv_data->base + setup_regs->reg_droop_ctrl); > > + val |= DROOP_CTRL_VAL; > > + writel_relaxed(val, drv_data->base + setup_regs->reg_droop_ctrl); > > + > > + /* Enable CC-Boost, DCVS-Boost, PS-Boost, WFx, PC/RET, DCVS FSM */ > > + val = readl_relaxed(drv_data->base + setup_regs->reg_pdn_fsm_ctrl); > > + val |= CC_BOOST_EN | PS_BOOST_EN | DCVS_BOOST_EN; > > + val |= WFX_DROOP_EN | PC_RET_EXIT_DROOP_EN | DCVS_DROOP_EN; > > + writel_relaxed(val, drv_data->base + setup_regs->reg_pdn_fsm_ctrl); > > + > > + /* Enable PLL Droop Override */ > > + val = PLL_OVERRIDE_DROOP_EN; > > + writel_relaxed(val, drv_data->base + setup_regs->reg_pll_override); > > + > > + /* Initialize the Adaptive Clock Distribution */ > > + ret = qcom_cpufreq_hw_acd_init(cpu_dev, policy, index); > > + if (ret) > > + return ret; > > + > > + /* We're ready: enable the OSM and give it time to boot (5uS) */ > > + writel_relaxed(1, drv_data->base + drv_data->soc_data->reg_enable); > > + udelay(OSM_BOOT_TIME_US); > > + > > + return ret; > > +} > > + > > static int qcom_cpufreq_hw_cpu_init(struct cpufreq_policy *policy) > > { > > struct platform_device *pdev = cpufreq_get_driver_data(); > > @@ -282,7 +1400,8 @@ static int qcom_cpufreq_hw_cpu_init(struct cpufreq_policy *policy) > > struct device *cpu_dev; > > void __iomem *base; > > struct qcom_cpufreq_data *data; > > - int ret, index; > > + const char *fdom_resname; > > + int cpu_count, index, ret; > > cpu_dev = get_cpu_device(policy->cpu); > > if (!cpu_dev) { > > @@ -303,9 +1422,14 @@ static int qcom_cpufreq_hw_cpu_init(struct cpufreq_policy *policy) > > index = args.args[0]; > > - base = devm_platform_ioremap_resource(pdev, index); > > + fdom_resname = kasprintf(GFP_KERNEL, "freq-domain%d", index); > > + if (!fdom_resname) > > + return -ENOMEM; > > + > > + base = devm_platform_ioremap_resource_byname(pdev, fdom_resname); > > if (IS_ERR(base)) > > return PTR_ERR(base); > > + kfree(fdom_resname); > > data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL); > > if (!data) { > > @@ -315,22 +1439,31 @@ static int qcom_cpufreq_hw_cpu_init(struct cpufreq_policy *policy) > > data->soc_data = of_device_get_match_data(&pdev->dev); > > data->base = base; > > + policy->driver_data = data; > > - /* HW should be in enabled state to proceed */ > > - if (!(readl_relaxed(base + data->soc_data->reg_enable) & 0x1)) { > > - dev_err(dev, "Domain-%d cpufreq hardware not enabled\n", index); > > - ret = -ENODEV; > > - goto error; > > - } > > - > > - qcom_get_related_cpus(index, policy->cpus); > > + cpu_count = qcom_get_related_cpus(index, policy->cpus); > > if (!cpumask_weight(policy->cpus)) { > > dev_err(dev, "Domain-%d failed to get related CPUs\n", index); > > ret = -ENOENT; > > goto error; > > } > > - policy->driver_data = data; > > + if (!data->soc_data->uses_tz) { > > + ret = qcom_cpufreq_hw_osm_setup(cpu_dev, policy, > > + cpu_count, index); > > + if (ret) { > > + dev_err(dev, "Cannot setup the OSM for CPU%d: %d\n", > > + policy->cpu, ret); > > + goto error; > > + } > > + } > > + > > + /* HW should be in enabled state to proceed */ > > + if (!(readl_relaxed(base + data->soc_data->reg_enable) & 0x1)) { > > + dev_err(dev, "Domain-%d cpufreq hardware not enabled\n", index); > > + ret = -ENODEV; > > + goto error; > > + } > > ret = qcom_cpufreq_hw_read_lut(cpu_dev, policy); > > if (ret) { > > @@ -355,6 +1488,7 @@ static int qcom_cpufreq_hw_cpu_init(struct cpufreq_policy *policy) > > return 0; > > error: > > + policy->driver_data = NULL; > > devm_iounmap(dev, base); > > return ret; > > } > > @@ -395,9 +1529,51 @@ static struct cpufreq_driver cpufreq_qcom_hw_driver = { > > static int qcom_cpufreq_hw_driver_probe(struct platform_device *pdev) > > { > > + const struct qcom_cpufreq_soc_data *soc_data; > > + struct device_node *pd_node; > > + struct platform_device *pd_dev; > > struct device *cpu_dev; > > struct clk *clk; > > - int ret; > > + int clk_div, ret; > > + > > + cpu_dev = get_cpu_device(0); > > + if (!cpu_dev) > > + return -EPROBE_DEFER; > > + > > + soc_data = of_device_get_match_data(&pdev->dev); > > + if (!soc_data) > > + return -EINVAL; > > + > > + if (!soc_data->uses_tz) { > > + /* > > + * When the OSM is not pre-programmed from TZ, we will > > + * need to program the sequencer through SCM calls. > > + */ > > + if (!qcom_scm_is_available()) > > + return -EPROBE_DEFER; > > + > > + /* > > + * If there are no power-domains, OSM programming cannot be > > + * performed, as in that case, we wouldn't know where to take > > + * the params from... > > + */ > > + pd_node = of_parse_phandle(cpu_dev->of_node, > > + "power-domains", 0); > > + if (!pd_node) { > > + ret = PTR_ERR(pd_node); > > + dev_err(cpu_dev, "power domain not found: %d\n", ret); > > + return ret; > > + } > > + > > + /* > > + * If the power domain device is not registered yet, then > > + * defer probing this driver until that is available. > > + */ > > + pd_dev = of_find_device_by_node(pd_node); > > + if (!pd_dev || !pd_dev->dev.driver || > > + !device_is_bound(&pd_dev->dev)) > > + return -EPROBE_DEFER; > > + } > > clk = clk_get(&pdev->dev, "xo"); > > if (IS_ERR(clk)) > > @@ -410,16 +1586,16 @@ static int qcom_cpufreq_hw_driver_probe(struct platform_device *pdev) > > if (IS_ERR(clk)) > > return PTR_ERR(clk); > > - cpu_hw_rate = clk_get_rate(clk) / CLK_HW_DIV; > > + clk_div = soc_data->clk_hw_div; > > + if (clk_div == 0) > > + clk_div++; > > + > > + cpu_hw_rate = clk_get_rate(clk) / clk_div; > > clk_put(clk); > > cpufreq_qcom_hw_driver.driver_data = pdev; > > /* Check for optional interconnect paths on CPU0 */ > > - cpu_dev = get_cpu_device(0); > > - if (!cpu_dev) > > - return -EPROBE_DEFER; > > - > > ret = dev_pm_opp_of_find_icc_paths(cpu_dev, NULL); > > if (ret) > > return ret; > > > > -- > QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member > of Code Aurora Forum, hosted by The Linux Foundation. > > --
On 20-01-21, 13:05, Bjorn Andersson wrote: > On Wed 20 Jan 12:25 CST 2021, Taniya Das wrote: > > > The CPUFREQ-HW driver is intended to be used only for CPUFREQ HW designs > > where the firmware programs the look up tables. > > > > It's obvious that this is the intended target for the current version of > the driver, but what are your technical arguments for keeping it that > way? > > > Suggestion is to separate out the driver where the programming is managed by > > high level OS. > > > > Can you please elaborate on the benefits of this approach? > > PS. Please don't top-post on LKML. Taniya, Can you please respond back to this ? We are waiting for merging this patchset.. Bjorn, can you or someone else please review this patch ? -- viresh
On 19-01-21, 18:45, AngeloGioacchino Del Regno wrote: > ** > ** NOTE: To "view the full picture", please look at the following > ** patch series: > ** https://patchwork.kernel.org/project/linux-arm-msm/list/?series=413355 > ** This is a subset of that series. > ** > > Changes in v4: > - Huge patch series has been split for better reviewability, > as suggested by Bjorn > - Rebased code on top of 266991721c15 ("cpufreq: qcom-hw: enable boost > support") Bjorn, what am I supposed to do with patchset ? Is there anyone who can give this some reviews from qcom team ?
Hi Bjorn, On 1/21/2021 12:35 AM, Bjorn Andersson wrote: > On Wed 20 Jan 12:25 CST 2021, Taniya Das wrote: > >> The CPUFREQ-HW driver is intended to be used only for CPUFREQ HW designs >> where the firmware programs the look up tables. >> > > It's obvious that this is the intended target for the current version of > the driver, but what are your technical arguments for keeping it that > way? > Technically the HW we are trying to program here differs in terms of clocking, the LUT definitions and many more. It will definitely make debugging much more troublesome if we try to accomodate multiple versions of CPUFREQ-HW in the same code. Thus to keep it simple, easy to read, debug, the suggestion is to keep it with "v1" tag as the OSM version we are trying to put here is from OSM1.0. >> Suggestion is to separate out the driver where the programming is managed by >> high level OS. >> > > Can you please elaborate on the benefits of this approach? > > PS. Please don't top-post on LKML. > > Regards, > Bjorn > >> On 1/19/2021 11:15 PM, AngeloGioacchino Del Regno wrote: >>> On new SoCs (SDM845 onwards) the Operating State Manager (OSM) is >>> being programmed in the bootloader and write-protected by the >>> hypervisor, leaving to the OS read-only access to some of its >>> registers (in order to read the Lookup Tables and also some >>> status registers) and write access to the p-state register, for >>> for the OS to request a specific performance state to trigger a >>> DVFS switch on the CPU through the OSM hardware. >>> >>> On old SoCs though (MSM8998, SDM630/660 and variants), the >>> bootloader will *not* initialize the OSM (and the CPRh, as it >>> is a requirement for it) before booting the OS, making any >>> request to trigger a performance state change ineffective, as >>> the hardware doesn't have any Lookup Table, nor is storing any >>> parameter to trigger a DVFS switch. In this case, basically all >>> of the OSM registers are *not* write protected for the OS, even >>> though some are - but write access is granted through SCM calls. >>> >>> This commit introduces support for OSM programming, which has to >>> be done on these old SoCs that were distributed (almost?) always >>> with a bootloader that does not do any CPRh nor OSM init before >>> booting the kernel. >>> In order to program the OSM on these SoCs, it is necessary to >>> fullfill a "special" requirement: the Core Power Reduction >>> Hardened (CPRh) hardware block must be initialized, as the OSM >>> is "talking" to it in order to perform the Voltage part of DVFS; >>> here, we are calling initialization of this through Linux generic >>> power domains, specifically by requesting a genpd attach from the >>> qcom-cpufreq-hw driver, which will give back voltages associated >>> to each CPU frequency that has been declared in the OPPs, scaled >>> and interpolated with the previous one, and will also give us >>> parameters for the Array Power Mux (APM) and mem-acc, in order >>> for this driver to be then able to generate the Lookup Tables >>> that will be finally programmed to the OSM hardware. >>> >>> After writing the parameters to the OSM and enabling it, all the >>> programming work will never happen anymore until a OS reboot, so >>> all of the allocations and "the rest" will be disposed-of: this >>> is done mainly to leave the code that was referred only to the >>> new SoCs intact, as to also emphasize on the fact that the OSM >>> HW is, in the end, the exact same; apart some register offsets >>> that are slightly different, the entire logic is the same. >>> >>> This also adds the parameters to support CPU scaling on SDM630 >>> and MSM8998. >>> >>> Signed-off-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@somainline.org> >>> --- >>> drivers/cpufreq/qcom-cpufreq-hw.c | 1240 ++++++++++++++++++++++++++++- >>> 1 file changed, 1208 insertions(+), 32 deletions(-) >>> >>> diff --git a/drivers/cpufreq/qcom-cpufreq-hw.c b/drivers/cpufreq/qcom-cpufreq-hw.c >>> index acc645b85e79..a92959bb7b50 100644 >>> --- a/drivers/cpufreq/qcom-cpufreq-hw.c >>> +++ b/drivers/cpufreq/qcom-cpufreq-hw.c >>> @@ -1,33 +1,256 @@ >>> // SPDX-License-Identifier: GPL-2.0 >>> /* >>> * Copyright (c) 2018, The Linux Foundation. All rights reserved. >>> + * >>> + * OSM hardware initial programming >>> + * Copyright (C) 2020, AngeloGioacchino Del Regno >>> + * <angelogioacchino.delregno@somainline.org> >>> */ >>> #include <linux/bitfield.h> >>> #include <linux/cpufreq.h> >>> +#include <linux/delay.h> >>> #include <linux/init.h> >>> #include <linux/interconnect.h> >>> #include <linux/kernel.h> >>> #include <linux/module.h> >>> #include <linux/of_address.h> >>> #include <linux/of_platform.h> >>> +#include <linux/pm_domain.h> >>> #include <linux/pm_opp.h> >>> #include <linux/slab.h> >>> +#include <linux/qcom_scm.h> >>> #define LUT_MAX_ENTRIES 40U >>> -#define LUT_SRC GENMASK(31, 30) >>> +#define LUT_SRC_845 GENMASK(31, 30) >>> +#define LUT_SRC_8998 GENMASK(27, 26) >>> +#define LUT_PLL_DIV GENMASK(25, 24) >>> #define LUT_L_VAL GENMASK(7, 0) >>> #define LUT_CORE_COUNT GENMASK(18, 16) >>> +#define LUT_VOLT_VC GENMASK(21, 16) >>> #define LUT_VOLT GENMASK(11, 0) >>> -#define CLK_HW_DIV 2 >>> #define LUT_TURBO_IND 1 >>> +#define OSM_BOOT_TIME_US 5 >>> + >>> +#define CYCLE_COUNTER_CLK_RATIO GENMASK(5, 1) >>> +#define OSM_XO_RATIO_VAL (10 - 1) >>> +#define CYCLE_COUNTER_USE_XO_EDGE BIT(8) >>> + >>> +/* FSM Boost Control */ >>> +#define CC_BOOST_EN BIT(0) >>> +#define PS_BOOST_EN BIT(1) >>> +#define DCVS_BOOST_EN BIT(2) >>> +#define BOOST_TIMER_REG_HI GENMASK(31, 16) >>> +#define BOOST_TIMER_REG_LO GENMASK(15, 0) >>> + >>> +#define PLL_WAIT_LOCK_TIME_NS 2000 >>> +#define SAFE_FREQ_WAIT_NS 1000 >>> +#define DEXT_DECREMENT_WAIT_NS 200 >>> + >>> +#define BOOST_SYNC_DELAY 5 >>> + >>> +#define HYSTERESIS_UP_MASK GENMASK(31, 16) >>> +#define HYSTERESIS_DN_MASK GENMASK(15, 0) >>> +#define HYSTERESIS_CC_NS 200 >>> +#define HYSTERESIS_LLM_NS 65535 >>> + >>> +/* FSM Droop Control */ >>> +#define PC_RET_EXIT_DROOP_EN BIT(3) >>> +#define WFX_DROOP_EN BIT(4) >>> +#define DCVS_DROOP_EN BIT(5) >>> +#define DROOP_TIMER1 GENMASK(31, 16) >>> +#define DROOP_TIMER0 GENMASK(15, 0) >>> +#define DROOP_CTRL_VAL (BIT(3) | BIT(17) | BIT(31)) >>> +#define DROOP_TIMER_NS 100 >>> +#define DROOP_WAIT_RELEASE_TIMER_NS 50 >>> +#define DROOP_RELEASE_TIMER_NS 1 >>> + >>> +/* PLL Override Control */ >>> +#define PLL_OVERRIDE_DROOP_EN BIT(0) >>> + >>> +/* Sequencer */ >>> +#define SEQUENCER_REG(base, n) (base + (n * 4)) >>> +#define SEQ_APM_THRESH_VC 15 >>> +#define SEQ_APM_THRESH_PREVC 31 >>> +#define SEQ_MEM_ACC_LVAL 32 >>> +#define SEQ_MEM_ACC_0 55 >>> +#define SEQ_APM_CROSSOVER_VC 72 >>> +#define SEQ_APM_PARAM 76 >>> +#define SEQ_MEM_ACC_CROSSOVER_VC 88 >>> +#define SEQ_MEM_ACC_MAX_LEVELS 4 >>> +#define SEQ_MEMACC_REG(base, n) SEQUENCER_REG(base, SEQ_MEM_ACC_0 + n) >>> + >>> +/* ACD */ >>> +#define ACD_WRITE_CTL_UPDATE_EN BIT(0) >>> +#define ACD_WRITE_CTL_SELECT_SHIFT 1 >>> + >>> +/** >>> + * struct qcom_cpufreq_soc_setup_data - Register offsets for OSM setup >>> + * >>> + * @reg_osm_sequencer: OSM Sequencer (used to get physical address) >>> + * @reg_override: Override parameters >>> + * @reg_spare: Spare parameters (MEMACC-to-VC) >>> + * @reg_cc_zero_behav: Virtual Corner for cluster power collapse >>> + * @reg_spm_cc_hyst: DCVS-CC Wait time for frequency inc/decrement >>> + * @reg_spm_cc_dcvs_dis: DCVS-CC en/disable control >>> + * @reg_spm_core_ret_map: Treat cores in retention as active/inactive >>> + * @reg_llm_freq_vote_hyst: DCVS-LLM Wait time for frequency inc/decrement >>> + * @reg_llm_volt_vote_hyst: DCVS-LLM Wait time for voltage inc/decrement >>> + * @reg_llm_intf_dcvs_dis: DCVS-LLM en/disable control >>> + * @reg_seq1: Sequencer extra register >>> + * @reg_pdn_fsm_ctrl: Boost and Droop FSMs en/disable control >>> + * @reg_cc_boost_timer: CC-Boost FSM wait first timer register >>> + * @reg_dcvs_boost_timer: DCVS-Boost FSM wait first timer register >>> + * @reg_ps_boost_timer: PS-Boost FSM wait first timer register >>> + * @boost_timer_reg_len: Length of boost timer registers >>> + * @reg_boost_sync_delay: PLL signal timing control for Boost >>> + * @reg_droop_ctrl: Droop control value >>> + * @reg_droop_release_ctrl: Wait for Droop release >>> + * @reg_droop_unstall_ctrl: Wait for Droop unstall >>> + * @reg_droop_wait_release_ctrl: Time to wait for state release >>> + * @reg_droop_timer_ctrl: Droop timer >>> + * @reg_droop_sync_delay: PLL signal timing control for Droop >>> + * @reg_pll_override: PLL Droop Override en/disable control >>> + * @reg_cycle_counter: OSM CPU cycle counter >>> + * >>> + * This structure holds the register offsets that are used to set-up >>> + * the Operating State Manager (OSM) parameters, when it is not (or >>> + * not entirely) configured from the bootloader and TrustZone. >>> + * >>> + * Acronyms used in this documentation: >>> + * CC = Core Count >>> + * PS = Power-Save >>> + * VC = Virtual Corner >>> + * LLM = Limits Load Management >>> + * DCVS = Dynamic Clock and Voltage Scaling >>> + */ >>> +struct qcom_cpufreq_soc_setup_data { >>> + /* OSM phys register offsets */ >>> + u16 reg_osm_sequencer; >>> + >>> + /* Frequency domain register offsets */ >>> + u16 reg_override; >>> + u16 reg_spare; >>> + u16 reg_cc_zero_behav; >>> + u16 reg_spm_cc_hyst; >>> + u16 reg_spm_cc_dcvs_dis; >>> + u16 reg_spm_core_ret_map; >>> + u16 reg_llm_freq_vote_hyst; >>> + u16 reg_llm_volt_vote_hyst; >>> + u16 reg_llm_intf_dcvs_dis; >>> + u16 reg_seq1; >>> + u16 reg_pdn_fsm_ctrl; >>> + u16 reg_cc_boost_timer; >>> + u16 reg_dcvs_boost_timer; >>> + u16 reg_ps_boost_timer; >>> + u16 boost_timer_reg_len; >>> + u16 reg_boost_sync_delay; >>> + u16 reg_droop_ctrl; >>> + u16 reg_droop_release_ctrl; >>> + u16 reg_droop_unstall_ctrl; >>> + u16 reg_droop_wait_release_ctrl; >>> + u16 reg_droop_timer_ctrl; >>> + u16 reg_droop_sync_delay; >>> + u16 reg_pll_override; >>> + u16 reg_cycle_counter; >>> +}; >>> + >>> +/** >>> + * struct qcom_cpufreq_soc_acd_data - Adaptive Clock Distribution data >>> + * >>> + * @tl_delay_reg: Tunable-Length Delay (TLD) register offset >>> + * @acd_ctrl_reg: Control Register (CR) register offset >>> + * @softstart_reg: Soft Start Control Register (SSCR) register offset >>> + * @ext_intf_reg: External interface configuration register offset >>> + * @auto_xfer_reg: Auto Register-Transfer register offset >>> + * @auto_xfer_cfg_reg: Auto Register-Transfer Configuration reg offset >>> + * @auto_xfer_ctl_reg: Auto Register-Transfer Control register offset >>> + * @auto_xfer_sts_reg: Auto Register-Transfer Status register offset >>> + * @dcvs_sw_reg: Software DCVS register offset >>> + * @gfmux_cfg_reg: Glitch-Free MUX configuration register offset >>> + * @write_ctl_reg: Write Control register >>> + * @write_sts_reg: Write Status register >>> + * @tl_delay_val: Tunable-Length Delay (TLD) value >>> + * @acd_ctrl_val: Control Register (CR) value >>> + * @softstart_val: Soft Start Control Register (SSCR) value >>> + * @ext_intf0_val: Initial external interface configuration value >>> + * @ext_intf1_val: Final external interface configuration value >>> + * @auto_xfer_val: Auto-register Transfer Control value >>> + * >>> + * This structure holds the register offsets (from the ACD iospace base) >>> + * and the parameters that are required to configure the OSM to >>> + * initialize the Adaptive Clock Distribution (ACD) system. >>> + */ >>> +struct qcom_cpufreq_soc_acd_data { >>> + u8 tl_delay_reg; >>> + u8 acd_ctrl_reg; >>> + u8 softstart_reg; >>> + u8 ext_intf_reg; >>> + u8 auto_xfer_reg; >>> + u8 auto_xfer_cfg_reg; >>> + u8 auto_xfer_ctl_reg; >>> + u8 auto_xfer_sts_reg; >>> + u8 dcvs_sw_reg; >>> + u8 gfmux_cfg_reg; >>> + u8 write_ctl_reg; >>> + u8 write_sts_reg; >>> + u32 tl_delay_val; >>> + u32 acd_ctrl_val; >>> + u32 softstart_val; >>> + u32 ext_intf0_val; >>> + u32 ext_intf1_val; >>> + u32 auto_xfer_val; >>> +}; >>> + >>> +/** >>> + * struct qcom_cpufreq_hw_params - Operating State Manager (OSM) Parameters >>> + * >>> + * @volt_lut_val: Value composed of: virtual corner (vc) and voltage in mV. >>> + * @freq_lut_val: Value composed of: core count, clock source and output >>> + * frequency in MHz. >>> + * @override_val: PLL parameters that the OSM uses to override the previous >>> + * setting coming from the bootloader, or when uninitialized. >>> + * @spare_val: Spare register, used by both this driver and the OSM HW >>> + * to identify MEM-ACC levels in relation to virtual corners. >>> + * @acc_thresh: MEM-ACC threshold level >>> + * >>> + * This structure holds the parameters to write to the OSM registers for >>> + * one "Virtual Corner" (VC), or one Performance State (p-state). >>> + */ >>> +struct qcom_cpufreq_hw_params { >>> + u32 volt_lut_val; >>> + u32 freq_lut_val; >>> + u32 override_val; >>> + u32 spare_val; >>> + u32 acc_thresh; >>> +}; >>> +/** >>> + * struct qcom_cpufreq_soc_data - SoC specific register offsets of the OSM >>> + * >>> + * @reg_enable: OSM enable status >>> + * @reg_index: Index of the Virtual Corner >>> + * @reg_freq_lut: Frequency Lookup Table >>> + * @reg_freq_lut_src_mask: Frequency Lookup Table clock-source mask >>> + * @reg_volt_lut: Voltage Lookup Table >>> + * @reg_perf_state: Performance State request register >>> + * @lut_row_size: Lookup Table row size >>> + * @clk_hw_div: Divider for "alternate" OSM clock-source >>> + * @uses_tz: OSM already set-up and protected by TrustZone >>> + * @setup_regs: Register offsets for OSM setup >>> + */ >>> struct qcom_cpufreq_soc_data { >>> u32 reg_enable; >>> + u32 reg_index; >>> u32 reg_freq_lut; >>> + u32 reg_freq_lut_src_mask; >>> u32 reg_volt_lut; >>> u32 reg_perf_state; >>> - u8 lut_row_size; >>> + u8 lut_row_size; >>> + u8 clk_hw_div; >>> + bool uses_tz; >>> + const struct qcom_cpufreq_soc_setup_data setup_regs; >>> + const struct qcom_cpufreq_soc_acd_data acd_data; >>> }; >>> struct qcom_cpufreq_data { >>> @@ -35,9 +258,17 @@ struct qcom_cpufreq_data { >>> const struct qcom_cpufreq_soc_data *soc_data; >>> }; >>> -static unsigned long cpu_hw_rate, xo_rate; >>> +static const char *cprh_genpd_names[] = { "cprh", NULL }; >>> +static u64 cpu_hw_rate, xo_rate; >>> static bool icc_scaling_enabled; >>> +/** >>> + * qcom_cpufreq_set_bw - Set interconnect bandwidth >>> + * @policy: CPUFreq policy structure >>> + * @freq_khz: CPU Frequency in KHz >>> + * >>> + * Returns: Zero for success, otherwise negative value on errors >>> + */ >>> static int qcom_cpufreq_set_bw(struct cpufreq_policy *policy, >>> unsigned long freq_khz) >>> { >>> @@ -59,6 +290,20 @@ static int qcom_cpufreq_set_bw(struct cpufreq_policy *policy, >>> return ret; >>> } >>> +/** >>> + * qcom_cpufreq_update_opp - Update CPU OPP tables >>> + * @policy: CPUFreq policy structure >>> + * @freq_khz: CPU Frequency for OPP entry in KHz >>> + * @volt: CPU Voltage for OPP entry in microvolts >>> + * >>> + * The CPU frequencies and voltages are being read from the Operating >>> + * State Manager (OSM) and the related OPPs, read from DT, need to be >>> + * updated to reflect what the hardware will set for each p-state. >>> + * If there is no OPP table specified in DT, then this function will >>> + * add dynamic ones. >>> + * >>> + * Returns: Zero for success, otherwise negative value on errors >>> + */ >>> static int qcom_cpufreq_update_opp(struct device *cpu_dev, >>> unsigned long freq_khz, >>> unsigned long volt) >>> @@ -79,6 +324,17 @@ static int qcom_cpufreq_update_opp(struct device *cpu_dev, >>> return dev_pm_opp_enable(cpu_dev, freq_hz); >>> } >>> +/** >>> + * qcom_cpufreq_hw_target_index - Set frequency/voltage >>> + * @policy: CPUFreq policy structure >>> + * @index: Performance state index to be set >>> + * >>> + * This function sends a request to the Operating State Manager >>> + * to set a Performance State index, so, to set frequency and >>> + * voltage for the target CPU/cluster. >>> + * >>> + * Returns: Always zero >>> + */ >>> static int qcom_cpufreq_hw_target_index(struct cpufreq_policy *policy, >>> unsigned int index) >>> { >>> @@ -94,6 +350,12 @@ static int qcom_cpufreq_hw_target_index(struct cpufreq_policy *policy, >>> return 0; >>> } >>> +/** >>> + * qcom_cpufreq_hw_get - Get current Performance State from OSM >>> + * @cpu: CPU number >>> + * >>> + * Returns: Current CPU/Cluster frequency or zero >>> + */ >>> static unsigned int qcom_cpufreq_hw_get(unsigned int cpu) >>> { >>> struct qcom_cpufreq_data *data; >>> @@ -127,6 +389,644 @@ static unsigned int qcom_cpufreq_hw_fast_switch(struct cpufreq_policy *policy, >>> return policy->freq_table[index].frequency; >>> } >>> +static void qcom_cpufreq_hw_boost_setup(void __iomem *timer0_addr, u32 len) >>> +{ >>> + u32 val; >>> + >>> + /* timer_reg0 */ >>> + val = FIELD_PREP(BOOST_TIMER_REG_LO, PLL_WAIT_LOCK_TIME_NS); >>> + val |= FIELD_PREP(BOOST_TIMER_REG_HI, SAFE_FREQ_WAIT_NS); >>> + writel_relaxed(val, timer0_addr); >>> + >>> + /* timer_reg1 */ >>> + val = FIELD_PREP(BOOST_TIMER_REG_LO, PLL_WAIT_LOCK_TIME_NS); >>> + val |= FIELD_PREP(BOOST_TIMER_REG_HI, PLL_WAIT_LOCK_TIME_NS); >>> + writel_relaxed(val, timer0_addr + len); >>> + >>> + /* timer_reg2 */ >>> + val = FIELD_PREP(BOOST_TIMER_REG_LO, DEXT_DECREMENT_WAIT_NS); >>> + writel_relaxed(val, timer0_addr + (2 * len)); >>> +} >>> + >>> +/* >>> + * qcom_cpufreq_gen_params - Generate parameters to send to the hardware >>> + * @cpu_dev: CPU device >>> + * @data: SoC specific register offsets >>> + * @hw_tbl: Pointer to return the array of parameters >>> + * @apm_vc: APM Virtual Corner crossover number, returned to the caller >>> + * @acc_vc: MEMACC Virtual Corner crossover number, returned to the caller >>> + * @cpu_count: Number of CPUs in the frequency domain >>> + * >>> + * This function allocates a 'qcom_cpufreq_hw_params' parameters table, >>> + * fills it and returns it to the consumer, ready to get sent to the HW. >>> + * Since the APM threshold is just one >>> + * Freeing the table after usage is left to the caller. >>> + * >>> + * Returns: Number of allocated (and filled) elements in the table, >>> + * otherwise negative value on errors. >>> + */ >>> +static int qcom_cpufreq_gen_params(struct device *cpu_dev, >>> + struct qcom_cpufreq_data *data, >>> + struct qcom_cpufreq_hw_params **hw_tbl, >>> + int *apm_vc, int *acc_vc, int cpu_count) >>> +{ >>> + struct platform_device *pdev = cpufreq_get_driver_data(); >>> + const struct qcom_cpufreq_soc_data *soc_data = data->soc_data; >>> + struct device **opp_virt_dev; >>> + struct opp_table *genpd_opp_table; >>> + struct dev_pm_opp *apm_opp, *acc_opp; >>> + unsigned long apm_uV, acc_uV, rate; >>> + int i, gpd_opp_cnt, ret; >>> + u8 last_acc_corner = 0, prev_spare = 0; >>> + >>> + /* >>> + * Install the OPP table that we get from DT here already, >>> + * otherwise it gets really messy as we'd have to manually walk >>> + * through the entire OPP DT, manually find frequencies and >>> + * also manually exclude supported-hw (for speed-binning). >>> + * This also makes it easier for the genpd to add info in it. >>> + */ >>> + ret = dev_pm_opp_of_add_table(cpu_dev); >>> + if (ret) { >>> + dev_err(&pdev->dev, "Cannot install CPU OPP table: %d\n", ret); >>> + return ret; >>> + } >>> + >>> + /* Get a handle to the genpd virtual device */ >>> + genpd_opp_table = dev_pm_opp_attach_genpd(cpu_dev, cprh_genpd_names, >>> + &opp_virt_dev); >>> + if (IS_ERR(genpd_opp_table)) { >>> + ret = PTR_ERR(genpd_opp_table); >>> + if (ret != -EPROBE_DEFER) >>> + dev_err(&pdev->dev, >>> + "Could not attach to pm_domain: %d\n", ret); >>> + goto opp_dispose; >>> + } >>> + >>> + /* Get the count of available OPPs coming from the power domain */ >>> + gpd_opp_cnt = dev_pm_opp_get_opp_count(cpu_dev); >>> + if (gpd_opp_cnt < 2) { >>> + ret = gpd_opp_cnt > 0 ? -EINVAL : gpd_opp_cnt; >>> + goto detach_gpd; >>> + } >>> + >>> + /* Find the APM threshold disabled OPP and "annotate" the voltage */ >>> + apm_opp = dev_pm_opp_find_freq_exact(cpu_dev, 0, false); >>> + if (IS_ERR(apm_opp)) { >>> + ret = -EINVAL; >>> + goto detach_gpd; >>> + } >>> + >>> + /* If we get no APM voltage, the system is going to be unstable */ >>> + apm_uV = dev_pm_opp_get_voltage(apm_opp); >>> + if (apm_uV == 0) { >>> + ret = -EINVAL; >>> + goto detach_gpd; >>> + } >>> + *apm_vc = -1; >>> + >>> + /* >>> + * Find the ACC threshold disabled OPP and "annotate" the voltage: >>> + * this is optional, as not every SoC needs it. >>> + */ >>> + acc_opp = dev_pm_opp_find_freq_exact(cpu_dev, 1, false); >>> + if (IS_ERR(acc_opp)) { >>> + acc_uV = ULONG_MAX; >>> + *acc_vc = U8_MAX; >>> + } else { >>> + /* >>> + * If the ACC OPP is present, this means that we *have* to >>> + * set the relative corner on the OSM: if we cannot get the >>> + * voltage at that point, the CPU will inevitably freeze. >>> + */ >>> + acc_uV = dev_pm_opp_get_voltage(acc_opp); >>> + if (acc_uV == 0) { >>> + ret = -EINVAL; >>> + goto detach_gpd; >>> + } >>> + *acc_vc = -1; >>> + } >>> + >>> + *hw_tbl = devm_kmalloc_array(&pdev->dev, gpd_opp_cnt, >>> + sizeof(**hw_tbl), GFP_KERNEL); >>> + if (!hw_tbl) { >>> + ret = -ENOMEM; >>> + goto detach_gpd; >>> + } >>> + >>> + for (i = 0, rate = 0; ; rate++, i++) { >>> + struct qcom_cpufreq_hw_params *entry = *hw_tbl + i; >>> + struct dev_pm_opp *genpd_opp; >>> + struct device_node *np; >>> + u32 pll_div, millivolts, f_src; >>> + >>> + /* Find the next enabled OPP's frequency (ignores APM/ACC) */ >>> + genpd_opp = dev_pm_opp_find_freq_ceil(cpu_dev, &rate); >>> + if (IS_ERR(genpd_opp)) >>> + break; >>> + >>> + /* Get mandatory and optional properties from the OPP DT */ >>> + np = dev_pm_opp_get_of_node(genpd_opp); >>> + if (!np) >>> + break; >>> + >>> + if (of_property_read_u32(np, "qcom,pll-override", >>> + &entry->override_val)) { >>> + of_node_put(np); >>> + return -EINVAL; >>> + } >>> + >>> + if (of_property_read_u32(np, "qcom,spare-data", >>> + &entry->spare_val)) >>> + entry->spare_val = 0; >>> + >>> + if (of_property_read_u32(np, "qcom,pll-div", &pll_div)) >>> + pll_div = 0; >>> + >>> + of_node_put(np); >>> + >>> + /* Get voltage in microvolts, then convert to millivolts */ >>> + millivolts = dev_pm_opp_get_voltage(genpd_opp); >>> + if (millivolts >= apm_uV && *apm_vc < 0) >>> + *apm_vc = i; >>> + if (millivolts >= acc_uV && *acc_vc < 0) >>> + *acc_vc = i; >>> + >>> + millivolts /= 1000; >>> + >>> + if (millivolts < 150 || millivolts > 1400) { >>> + dev_err(&pdev->dev, >>> + "Read invalid voltage: %u.\n", millivolts); >>> + return -EINVAL; >>> + } >>> + >>> + /* In the OSM firmware, "Virtual Corner" levels start from 0 */ >>> + entry->volt_lut_val = FIELD_PREP(LUT_VOLT_VC, i); >>> + entry->volt_lut_val |= FIELD_PREP(LUT_VOLT, millivolts); >>> + >>> + /* Only the first frequency can have alternate source */ >>> + f_src = i ? 1 : 0; >>> + f_src <<= ffs(soc_data->reg_freq_lut_src_mask) - 1; >>> + entry->freq_lut_val = f_src | div_u64(rate, xo_rate); >>> + entry->freq_lut_val |= FIELD_PREP(LUT_CORE_COUNT, cpu_count); >>> + >>> + /* >>> + * PLL divider is not always 0 and there is no way to determine >>> + * it automatically, as setting this value higher than DIV1 >>> + * will make the OSM HW to effectively set the PLL at 2-4x >>> + * the CPU frequency and then divide the CPU clock by this div, >>> + * so this value is effectively used as both a multiplier and >>> + * divider. >>> + * This value cannot be calculated because it depends on >>> + * manual calibration and is (most probably) used to choose >>> + * a PLL frequency that gives the least possible jitter. >>> + */ >>> + entry->freq_lut_val |= FIELD_PREP(LUT_PLL_DIV, pll_div); >>> + >>> + /* >>> + * MEM-ACC Virtual Corner threshold voltage: set the >>> + * acc_thresh to the "next different" spare_val, otherwise >>> + * set it to an invalid value, so that we can recognize it >>> + * and exclude it later, when this set of data will be used >>> + * for programming the OSM. >>> + */ >>> + if (entry->spare_val != prev_spare) { >>> + prev_spare = entry->spare_val; >>> + entry->acc_thresh = last_acc_corner; >>> + } else { >>> + entry->acc_thresh = SEQ_MEM_ACC_MAX_LEVELS + 1; >>> + } >>> + >>> + dev_dbg(&pdev->dev, >>> + "[%d] freq=0x%x volt=0x%x override=0x%x spare=0x%x\n", >>> + i, entry->freq_lut_val, entry->volt_lut_val, >>> + entry->override_val, entry->spare_val); >>> + dev_pm_opp_put(genpd_opp); >>> + } >>> + >>> + /* >>> + * If we've got a customized mem-acc corner but we couldn't >>> + * find any suitable crossover, or the corner is less than >>> + * the minimum amount of required corners for mem-acc scaling, >>> + * the values are not valid, hence fall back to LUT values. >>> + */ >>> + if (acc_uV != ULONG_MAX && *acc_vc < SEQ_MEM_ACC_MAX_LEVELS - 1) { >>> + dev_dbg(&pdev->dev, >>> + "MEM-ACC corner: invalid values VC%d %luuV\n", >>> + *acc_vc, acc_uV); >>> + *acc_vc = U8_MAX; >>> + } >>> + >>> + /* >>> + * If we have probed less params than what we need, then the >>> + * OPP table that we got from the genpd is malformed for some >>> + * reason: in this case, do not apply the table to the HW. >>> + */ >>> + if (i < gpd_opp_cnt) { >>> + dev_err(&pdev->dev, "Got bad OPP table from power domain.\n"); >>> + ret = -EINVAL; >>> + goto detach_gpd; >>> + } >>> + >>> +detach_gpd: >>> + dev_pm_opp_detach_genpd(genpd_opp_table); >>> +opp_dispose: >>> + /* >>> + * Now that we're totally done with it, dispose of all the dynamic >>> + * OPPs in the table: like this, at the end of the OSM configuration >>> + * we are leaving it like it was magically configured by the TZ or >>> + * by the bootloader and using the rest of the driver as if this >>> + * programming phase has never happened in the OS, like new SoCs. >>> + * This also makes us able to not modify a single bit in the basic >>> + * OSM frequency request logic that was meant for the newest SoCs. >>> + */ >>> + dev_pm_opp_remove_all_dynamic(cpu_dev); >>> + return ret < 0 ? ret : i; >>> +} >>> + >>> +static inline u32 qcom_cpufreq_acd_regbit(u8 acd_reg_offset) >>> +{ >>> + return BIT((acd_reg_offset / 4)); >>> +} >>> + >>> +static int qcom_cpufreq_hw_acd_write_autoxfer(struct qcom_cpufreq_data *data, >>> + void __iomem *acd_base, u32 val) >>> +{ >>> + const struct qcom_cpufreq_soc_data *sdata = data->soc_data; >>> + const struct qcom_cpufreq_soc_acd_data *aregs = &sdata->acd_data; >>> + u32 regval; >>> + int retry = 3; >>> + >>> + writel_relaxed(val, acd_base + aregs->auto_xfer_cfg_reg); >>> + >>> + /* (Clear, then) Set AUTOXFER START */ >>> + writel_relaxed(0, acd_base + aregs->auto_xfer_reg); >>> + writel_relaxed(1, acd_base + aregs->auto_xfer_reg); >>> + >>> + /* Poll for status: if the first bit is set the transfer is done. */ >>> + do { >>> + regval = readl_relaxed(acd_base + aregs->auto_xfer_sts_reg); >>> + if (regval & BIT(0)) >>> + break; >>> + udelay(1); >>> + } while (--retry); >>> + >>> + if (retry <= 0) >>> + return -ETIMEDOUT; >>> + return 0; >>> +} >>> + >>> +static int qcom_cpufreq_hw_acd_write_xfer(struct qcom_cpufreq_data *data, >>> + void __iomem *acd_base, u8 reg, >>> + u32 val) >>> +{ >>> + const struct qcom_cpufreq_soc_data *sdata = data->soc_data; >>> + const struct qcom_cpufreq_soc_acd_data *aregs = &sdata->acd_data; >>> + int retry = 2; >>> + u32 regval; >>> + >>> + /* Write to the register, then initiate manual transfer */ >>> + writel_relaxed(val, acd_base + reg); >>> + >>> + /* >>> + * Make sure that the write goes through, otherwise the local >>> + * transfer process will wrongly copy the old value. >>> + */ >>> + wmb(); >>> + >>> + /* Clear write control register */ >>> + writel_relaxed(0, acd_base + aregs->write_ctl_reg); >>> + >>> + regval = (reg / 4) << ACD_WRITE_CTL_SELECT_SHIFT; >>> + regval |= ACD_WRITE_CTL_UPDATE_EN; >>> + writel_relaxed(regval, acd_base + aregs->write_ctl_reg); >>> + >>> + /* Wait until ACD Local Transfer is done */ >>> + do { >>> + regval = readl_relaxed(acd_base + aregs->write_sts_reg); >>> + if (regval & qcom_cpufreq_acd_regbit(reg)) >>> + break; >>> + udelay(1); >>> + } while (--retry); >>> + >>> + if (retry <= 0) >>> + return -ETIMEDOUT; >>> + return 0; >>> +} >>> + >>> +/** >>> + * qcom_cpufreq_hw_acd_init - Initialize ACD params in the OSM >>> + * @cpu_dev: CPU device >>> + * @policy: CPUFreq policy structure >>> + * @index: Instance number (CPU cluster number) >>> + * >>> + * On some SoCs it is required to send the ACD configuration parameters >>> + * to the OSM. This function takes the parameters from the SoC specific >>> + * configuration and writes them only if a "osm-acdN" iospace has been >>> + * declared (hence, it's present). >>> + * >>> + * Returns: Zero for success, otherwise negative number on error. >>> + */ >>> +static int qcom_cpufreq_hw_acd_init(struct device *cpu_dev, >>> + struct cpufreq_policy *policy, >>> + int index) >>> +{ >>> + struct platform_device *pdev = cpufreq_get_driver_data(); >>> + struct qcom_cpufreq_data *ddata = policy->driver_data; >>> + const struct qcom_cpufreq_soc_data *sdata = ddata->soc_data; >>> + const struct qcom_cpufreq_soc_acd_data *aregs = &sdata->acd_data; >>> + const char *acd_resname; >>> + void __iomem *acd_base; >>> + u32 rmask; >>> + int ret; >>> + >>> + acd_resname = kasprintf(GFP_KERNEL, "osm-acd%d", index); >>> + if (!acd_resname) >>> + return -ENOMEM; >>> + >>> + acd_base = devm_platform_ioremap_resource_byname(pdev, acd_resname); >>> + kfree(acd_resname); >>> + if (IS_ERR(acd_base)) { >>> + dev_vdbg(cpu_dev, "Skipping ACD initialization.\n"); >>> + return 0; >>> + } >>> + >>> + writel_relaxed(aregs->tl_delay_val, acd_base + aregs->tl_delay_reg); >>> + writel_relaxed(aregs->acd_ctrl_val, acd_base + aregs->acd_ctrl_reg); >>> + writel_relaxed(aregs->softstart_val, acd_base + aregs->softstart_reg); >>> + writel_relaxed(aregs->ext_intf0_val, acd_base + aregs->ext_intf_reg); >>> + writel_relaxed(aregs->auto_xfer_val, >>> + acd_base + aregs->auto_xfer_ctl_reg); >>> + /* >>> + * Before initiating register auto transfer, make sure that all >>> + * the writes go through, otherwise bad values will be transferred. >>> + */ >>> + wmb(); >>> + >>> + rmask = qcom_cpufreq_acd_regbit(aregs->acd_ctrl_reg) | >>> + qcom_cpufreq_acd_regbit(aregs->tl_delay_reg) | >>> + qcom_cpufreq_acd_regbit(aregs->softstart_reg) | >>> + qcom_cpufreq_acd_regbit(aregs->ext_intf_reg); >>> + ret = qcom_cpufreq_hw_acd_write_autoxfer(ddata, acd_base, rmask); >>> + if (ret) >>> + return ret; >>> + >>> + /* Switch CPUSS clock source to ACD clock */ >>> + ret = qcom_cpufreq_hw_acd_write_xfer(ddata, acd_base, >>> + aregs->gfmux_cfg_reg, 1); >>> + if (ret) >>> + return ret; >>> + >>> + /* (Set, then) Clear DCVS_SW */ >>> + ret = qcom_cpufreq_hw_acd_write_xfer(ddata, acd_base, >>> + aregs->dcvs_sw_reg, 1); >>> + if (ret) >>> + return ret; >>> + ret = qcom_cpufreq_hw_acd_write_xfer(ddata, acd_base, >>> + aregs->dcvs_sw_reg, 0); >>> + if (ret) >>> + return ret; >>> + >>> + /* Wait for clock switch time */ >>> + udelay(1); >>> + >>> + /* Program the final ACD external interface */ >>> + ret = qcom_cpufreq_hw_acd_write_xfer(ddata, acd_base, >>> + aregs->ext_intf_reg, >>> + aregs->ext_intf1_val); >>> + if (ret) >>> + return ret; >>> + >>> + /* Initiate transfer of the final ACD value */ >>> + rmask |= qcom_cpufreq_acd_regbit(aregs->gfmux_cfg_reg); >>> + writel_relaxed(rmask, acd_base + aregs->auto_xfer_cfg_reg); >>> + >>> + /* Wait for ACD to stabilize. Same wait as the OSM boot time... */ >>> + udelay(OSM_BOOT_TIME_US); >>> + return ret; >>> +} >>> + >>> +/** >>> + * qcom_cpufreq_hw_write_lut - Write Lookup Table params to the OSM >>> + * @cpu_dev: CPU device >>> + * @policy: CPUFreq policy structure >>> + * @cpu_count: Number of CPUs in the frequency domain >>> + * @index: Instance number (CPU cluster number) >>> + * >>> + * Program all the Lookup Table (LUT) entries and related thresholds >>> + * to the Operating State Manager on platforms where the same hasn't >>> + * been done already by the bootloader or TrustZone before booting >>> + * the operating system's kernel; >>> + * On these platforms, write access to the OSM is (obviously) not >>> + * blocked by the hypervisor. >>> + * >>> + * Returns: Zero for success, otherwise negative number on error. >>> + */ >>> +static int qcom_cpufreq_hw_write_lut(struct device *cpu_dev, >>> + struct cpufreq_policy *policy, >>> + int cpu_count, int index) >>> +{ >>> + struct platform_device *pdev = cpufreq_get_driver_data(); >>> + struct qcom_cpufreq_data *ddata = policy->driver_data; >>> + const struct qcom_cpufreq_soc_data *sdata = ddata->soc_data; >>> + const struct qcom_cpufreq_soc_setup_data *sregs = &sdata->setup_regs; >>> + struct qcom_cpufreq_hw_params *hw_tbl; >>> + struct resource *osm_rsrc; >>> + const char *osm_resname; >>> + u32 sreg, seq_addr, acc_lval = 0, last_spare = 1; >>> + int apm_vc = INT_MAX, acc_vc = U8_MAX, acc_idx = 0; >>> + int acc_val[SEQ_MEM_ACC_MAX_LEVELS]; >>> + int i, ret, num_entries; >>> + >>> + osm_resname = kasprintf(GFP_KERNEL, "osm-domain%d", index); >>> + if (!osm_resname) >>> + return -ENOMEM; >>> + >>> + /* >>> + * On some SoCs the OSM is not getting programmed from bootloader >>> + * and needs to be done here: in this case, we need to retrieve >>> + * the base physical address for the "Sequencer", so we will get >>> + * the OSM base phys and apply the sequencer offset. >>> + * >>> + * Note: We are not remapping this iospace because we are really >>> + * sending the physical address through SCM calls later. >>> + */ >>> + osm_rsrc = platform_get_resource_byname(pdev, IORESOURCE_MEM, >>> + osm_resname); >>> + kfree(osm_resname); >>> + >>> + if (!osm_rsrc) >>> + return -ENODEV; >>> + >>> + seq_addr = osm_rsrc->start + sregs->reg_osm_sequencer; >>> + >>> + num_entries = qcom_cpufreq_gen_params(cpu_dev, ddata, &hw_tbl, >>> + &apm_vc, &acc_vc, cpu_count); >>> + if (num_entries < 0) >>> + return num_entries; >>> + >>> + for (i = 0; i < LUT_MAX_ENTRIES; i++) { >>> + struct qcom_cpufreq_hw_params *entry; >>> + int pos = i * sdata->lut_row_size; >>> + >>> + /* >>> + * If we have reached the end of the params table, write >>> + * the last valid entry until the end of the OSM table. >>> + */ >>> + if (i < num_entries) >>> + entry = &hw_tbl[i]; >>> + else >>> + entry = &hw_tbl[num_entries - 1]; >>> + >>> + writel_relaxed(i, ddata->base + sdata->reg_index + pos); >>> + >>> + writel_relaxed(entry->volt_lut_val, >>> + ddata->base + sdata->reg_volt_lut + pos); >>> + >>> + writel_relaxed(entry->freq_lut_val, >>> + ddata->base + sdata->reg_freq_lut + pos); >>> + >>> + writel_relaxed(entry->override_val, >>> + ddata->base + sregs->reg_override + pos); >>> + >>> + writel_relaxed(entry->spare_val, >>> + ddata->base + sregs->reg_spare + pos); >>> + >>> + dev_dbg(cpu_dev, >>> + "Writing [%d] v:0x%x f:0x%x ovr:0x%x s:0x%x\n", i, >>> + entry->volt_lut_val, entry->freq_lut_val, >>> + entry->override_val, entry->spare_val); >>> + >>> + /* >>> + * MEM-ACC Virtual Corner threshold voltage: this gets set >>> + * as the pairs of corners in which there is a transition >>> + * between one MEM-ACC level and the next one. >>> + * >>> + * Notes: The spare_val can never be zero; >>> + * The first spare_val is always 1; >>> + * The maximum number of pairs is two (four registers). >>> + * >>> + * Example: (C = Corner Level - M = MEM-ACC Level) >>> + * C0 M1 - C1 M1 - C2 M2 - C3 M2 - C4 M2 - C5 M3 >>> + * Pairs: 1-2, 4-5 >>> + */ >>> + if (entry->spare_val <= last_spare || >>> + acc_idx >= SEQ_MEM_ACC_MAX_LEVELS - 1) >>> + continue; >>> + >>> + /* Standard mem-acc pairs using spare_val LUT crossovers */ >>> + last_spare = entry->spare_val; >>> + acc_val[acc_idx] = i - 1; >>> + acc_idx++; >>> + acc_val[acc_idx] = i; >>> + acc_idx++; >>> + } >>> + >>> + /* Sanity check: we *must* have two mem-acc crossovers (four values) */ >>> + if (acc_idx < SEQ_MEM_ACC_MAX_LEVELS - 1) >>> + return -EINVAL; >>> + >>> + /* >>> + * Customized mem-acc corners, if any; in this case, the last corner >>> + * in the external (CPRh) LUT is this one, placed after the APM one. >>> + */ >>> + if (acc_vc != U8_MAX) { >>> + sreg = SEQUENCER_REG(seq_addr, SEQ_MEM_ACC_CROSSOVER_VC); >>> + ret = qcom_scm_io_writel(sreg, num_entries + 1); >>> + if (ret) >>> + return ret; >>> + >>> + /* >>> + * At the price of very-slightly higher power consumption, >>> + * switch the ACC at one corner lower than what we've found, >>> + * as this seems to be needed on at least some MSM8998 chips >>> + * to achieve full system stability >>> + */ >>> + acc_vc--; >>> + >>> + /* Change only if we have to move the corner down */ >>> + if (acc_vc < acc_val[3]) { >>> + acc_val[2] = acc_vc - 1; >>> + acc_val[3] = acc_vc; >>> + } >>> + >>> + /* If needed, sanitize previously stored vals from the LUT */ >>> + if (acc_val[2] <= acc_val[1]) >>> + acc_val[1] = acc_val[2] - 1; >>> + if (acc_val[1] <= acc_val[0]) >>> + acc_val[0] = acc_val[1] - 1; >>> + } >>> + >>> + for (i = 0; i < SEQ_MEM_ACC_MAX_LEVELS; i++) { >>> + ret = qcom_scm_io_writel(SEQ_MEMACC_REG(seq_addr, i), >>> + acc_val[i]); >>> + if (ret) >>> + return ret; >>> + } >>> + dev_dbg(cpu_dev, "Wrote MEM-ACC Pairs: [%u-%u] [%u-%u]\n", >>> + acc_val[0], acc_val[1], acc_val[2], acc_val[3]); >>> + >>> + /* >>> + * Program the L_VAL of the first corner requesting MEM-ACC >>> + * voltage level 3 to the right sequencer register >>> + */ >>> + acc_lval = FIELD_GET(LUT_L_VAL, hw_tbl[acc_val[3]].freq_lut_val); >>> + ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_MEM_ACC_LVAL), >>> + acc_lval); >>> + if (ret) { >>> + dev_dbg(cpu_dev, "Cannot send memacc l_val\n"); >>> + return ret; >>> + } >>> + dev_dbg(cpu_dev, "MEM-ACC L-Val is %u\n", acc_lval); >>> + >>> + /* >>> + * Array Power Mux threshold level: the first virtual corner >>> + * that requires a switch sequence of the APM from MX to APC. >>> + */ >>> + if (apm_vc == INT_MAX) >>> + apm_vc = LUT_MAX_ENTRIES - 1; >>> + >>> + /* >>> + * APM crossover virtual corner refers to CPRh: there, the APM corner >>> + * is always appended to the table (so, at the end of it, right after >>> + * the cluster dvfs entries). >>> + */ >>> + writel(num_entries, ddata->base + sregs->reg_seq1); >>> + ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_APM_CROSSOVER_VC), >>> + num_entries); >>> + if (ret) >>> + return ret; >>> + >>> + ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_APM_THRESH_VC), >>> + apm_vc); >>> + if (ret) >>> + return ret; >>> + >>> + ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_APM_THRESH_PREVC), >>> + apm_vc - 1); >>> + if (ret) >>> + return ret; >>> + >>> + ret = qcom_scm_io_writel(SEQUENCER_REG(seq_addr, SEQ_APM_PARAM), >>> + (0x39 | apm_vc << 6)); >>> + if (ret) >>> + return ret; >>> + >>> + dev_dbg(cpu_dev, "Wrote APM Pair: [%u-%u]\n", apm_vc - 1, apm_vc); >>> + return ret; >>> +} >>> + >>> +/** >>> + * qcom_cpufreq_hw_read_lut - Read Lookup Table from the OSM >>> + * @cpu_dev: CPU device >>> + * @policy: CPUFreq policy structure >>> + * >>> + * The Operating State Manager Lookup Table can always be read, even >>> + * in case it was pre-programmed by the bootloader or by TrustZone. >>> + * Read the LUT from it in order to build OPPs containing DVFS info. >>> + * >>> + * Returns: Zero for success, otherwise negative number on errors. >>> + */ >>> static int qcom_cpufreq_hw_read_lut(struct device *cpu_dev, >>> struct cpufreq_policy *policy) >>> { >>> @@ -166,7 +1066,9 @@ static int qcom_cpufreq_hw_read_lut(struct device *cpu_dev, >>> for (i = 0; i < LUT_MAX_ENTRIES; i++) { >>> data = readl_relaxed(drv_data->base + soc_data->reg_freq_lut + >>> i * soc_data->lut_row_size); >>> - src = FIELD_GET(LUT_SRC, data); >>> + src = data & soc_data->reg_freq_lut_src_mask; >>> + src >>= ffs(soc_data->reg_freq_lut_src_mask) - 1; >>> + >>> lval = FIELD_GET(LUT_L_VAL, data); >>> core_count = FIELD_GET(LUT_CORE_COUNT, data); >>> @@ -175,17 +1077,21 @@ static int qcom_cpufreq_hw_read_lut(struct device *cpu_dev, >>> volt = FIELD_GET(LUT_VOLT, data) * 1000; >>> if (src) >>> - freq = xo_rate * lval / 1000; >>> + freq = xo_rate * lval; >>> else >>> - freq = cpu_hw_rate / 1000; >>> + freq = cpu_hw_rate; >>> + freq /= 1000; >>> if (freq != prev_freq && core_count != LUT_TURBO_IND) { >>> if (!qcom_cpufreq_update_opp(cpu_dev, freq, volt)) { >>> table[i].frequency = freq; >>> - dev_dbg(cpu_dev, "index=%d freq=%d, core_count %d\n", i, >>> - freq, core_count); >>> + dev_dbg(cpu_dev, >>> + "index=%d freq=%d, core_count %d\n", >>> + i, freq, core_count); >>> } else { >>> - dev_warn(cpu_dev, "failed to update OPP for freq=%d\n", freq); >>> + dev_warn(cpu_dev, >>> + "failed to update OPP for freq=%d\n", >>> + freq); >>> table[i].frequency = CPUFREQ_ENTRY_INVALID; >>> } >>> @@ -205,18 +1111,19 @@ static int qcom_cpufreq_hw_read_lut(struct device *cpu_dev, >>> * as the boost frequency >>> */ >>> if (prev->frequency == CPUFREQ_ENTRY_INVALID) { >>> - if (!qcom_cpufreq_update_opp(cpu_dev, prev_freq, volt)) { >>> + if (!qcom_cpufreq_update_opp(cpu_dev, prev_freq, >>> + volt)) { >>> prev->frequency = prev_freq; >>> prev->flags = CPUFREQ_BOOST_FREQ; >>> } else { >>> - dev_warn(cpu_dev, "failed to update OPP for freq=%d\n", >>> + dev_warn(cpu_dev, >>> + "can't update OPP for freq=%u\n", >>> freq); >>> } >>> } >>> break; >>> } >>> - >>> prev_freq = freq; >>> } >>> @@ -227,10 +1134,18 @@ static int qcom_cpufreq_hw_read_lut(struct device *cpu_dev, >>> return 0; >>> } >>> -static void qcom_get_related_cpus(int index, struct cpumask *m) >>> +/* >>> + * qcom_get_related_cpus - Get mask of CPUs in the same frequency domain >>> + * @index: CPU number >>> + * @m: Returned CPU mask >>> + * >>> + * Returns: Count of CPUs inserted in the cpumask or negative number for error. >>> + */ >>> +static int qcom_get_related_cpus(int index, struct cpumask *m) >>> { >>> struct device_node *cpu_np; >>> struct of_phandle_args args; >>> + int count = 0; >>> int cpu, ret; >>> for_each_possible_cpu(cpu) { >>> @@ -245,34 +1160,237 @@ static void qcom_get_related_cpus(int index, struct cpumask *m) >>> if (ret < 0) >>> continue; >>> - if (index == args.args[0]) >>> + if (index == args.args[0]) { >>> cpumask_set_cpu(cpu, m); >>> + count++; >>> + } >>> } >>> + >>> + return count > 0 ? count : -EINVAL; >>> } >>> static const struct qcom_cpufreq_soc_data qcom_soc_data = { >>> .reg_enable = 0x0, >>> .reg_freq_lut = 0x110, >>> + .reg_freq_lut_src_mask = LUT_SRC_845, >>> .reg_volt_lut = 0x114, >>> .reg_perf_state = 0x920, >>> .lut_row_size = 32, >>> + .clk_hw_div = 2, >>> + .uses_tz = true, >>> +}; >>> + >>> +static const struct qcom_cpufreq_soc_data msm8998_soc_data = { >>> + .reg_enable = 0x4, >>> + .reg_index = 0x150, >>> + .reg_freq_lut = 0x154, >>> + .reg_freq_lut_src_mask = LUT_SRC_8998, >>> + .reg_volt_lut = 0x158, >>> + .reg_perf_state = 0xf10, >>> + .lut_row_size = 32, >>> + .clk_hw_div = 1, >>> + .uses_tz = false, >>> + .setup_regs = { >>> + /* Physical offset for sequencer scm calls */ >>> + .reg_osm_sequencer = 0x300, >>> + >>> + /* Frequency domain offsets */ >>> + .reg_override = 0x15c, >>> + .reg_spare = 0x164, >>> + .reg_cc_zero_behav = 0x0c, >>> + .reg_spm_cc_hyst = 0x1c, >>> + .reg_spm_cc_dcvs_dis = 0x20, >>> + .reg_spm_core_ret_map = 0x24, >>> + .reg_llm_freq_vote_hyst = 0x2c, >>> + .reg_llm_volt_vote_hyst = 0x30, >>> + .reg_llm_intf_dcvs_dis = 0x34, >>> + .reg_seq1 = 0x48, >>> + .reg_pdn_fsm_ctrl = 0x70, >>> + .reg_cc_boost_timer = 0x74, >>> + .reg_dcvs_boost_timer = 0x84, >>> + .reg_ps_boost_timer = 0x94, >>> + .boost_timer_reg_len = 0x4, >>> + .reg_boost_sync_delay = 0xa0, >>> + .reg_droop_ctrl = 0xa4, >>> + .reg_droop_release_ctrl = 0xa8, >>> + .reg_droop_unstall_ctrl = 0xac, >>> + .reg_droop_wait_release_ctrl = 0xb0, >>> + .reg_droop_timer_ctrl = 0xb8, >>> + .reg_droop_sync_delay = 0xbc, >>> + .reg_pll_override = 0xc0, >>> + .reg_cycle_counter = 0xf00, >>> + }, >>> + .acd_data = { >>> + .acd_ctrl_reg = 0x4, >>> + .tl_delay_reg = 0x8, >>> + .softstart_reg = 0x28, >>> + .ext_intf_reg = 0x30, >>> + .dcvs_sw_reg = 0x34, >>> + .gfmux_cfg_reg = 0x3c, >>> + .auto_xfer_cfg_reg = 0x80, >>> + .auto_xfer_reg = 0x84, >>> + .auto_xfer_ctl_reg = 0x88, >>> + .auto_xfer_sts_reg = 0x8c, >>> + .write_ctl_reg = 0x90, >>> + .write_sts_reg = 0x94, >>> + .tl_delay_val = 38417, >>> + .acd_ctrl_val = 0x2b5ffd, >>> + .softstart_val = 0x501, >>> + .ext_intf0_val = 0x2cf9ae8, >>> + .ext_intf1_val = 0x2cf9afe, >>> + .auto_xfer_val = 0x15, >>> + }, >>> }; >>> static const struct qcom_cpufreq_soc_data epss_soc_data = { >>> .reg_enable = 0x0, >>> .reg_freq_lut = 0x100, >>> + .reg_freq_lut_src_mask = LUT_SRC_845, >>> .reg_volt_lut = 0x200, >>> .reg_perf_state = 0x320, >>> .lut_row_size = 4, >>> + .clk_hw_div = 2, >>> + .uses_tz = true, >>> }; >>> static const struct of_device_id qcom_cpufreq_hw_match[] = { >>> { .compatible = "qcom,cpufreq-hw", .data = &qcom_soc_data }, >>> + { .compatible = "qcom,cpufreq-hw-8998", .data = &msm8998_soc_data }, >>> { .compatible = "qcom,cpufreq-epss", .data = &epss_soc_data }, >>> {} >>> }; >>> MODULE_DEVICE_TABLE(of, qcom_cpufreq_hw_match); >>> +/** >>> + * qcom_cpufreq_hw_osm_setup - Setup and enable the OSM >>> + * @cpu_dev: CPU device >>> + * @policy: CPUFreq policy structure >>> + * @cpu_count: Number of CPUs in the frequency domain >>> + * >>> + * On some platforms, the Operating State Manager (OSM) is not getting >>> + * programmed by the bootloader, nor by TrustZone before booting the OS >>> + * and its register space is not write-protected by the hypervisor. >>> + * In this case, to achieve CPU DVFS, it is needed to program it from >>> + * the OS itself, which includes setting LUT and all the various tunables >>> + * that are required for it to manage the CPU frequencies and voltages >>> + * on its own. >>> + * Calling this function on a platform that had the OSM set-up by TZ >>> + * will result in a hypervisor fault with system reboot in most cases. >>> + * >>> + * Returns: Zero for success, otherwise negative number on errors. >>> + */ >>> +static int qcom_cpufreq_hw_osm_setup(struct device *cpu_dev, >>> + struct cpufreq_policy *policy, >>> + int cpu_count, int index) >>> +{ >>> + struct qcom_cpufreq_data *drv_data = policy->driver_data; >>> + const struct qcom_cpufreq_soc_setup_data *setup_regs; >>> + u32 val; >>> + int ret; >>> + >>> + ret = qcom_cpufreq_hw_write_lut(cpu_dev, policy, cpu_count, index); >>> + if (ret) >>> + return ret; >>> + >>> + setup_regs = &drv_data->soc_data->setup_regs; >>> + >>> + /* Set OSM to XO clock ratio and use XO edge for the cycle counter */ >>> + val = FIELD_PREP(CYCLE_COUNTER_CLK_RATIO, OSM_XO_RATIO_VAL); >>> + val |= CYCLE_COUNTER_USE_XO_EDGE; >>> + >>> + /* Enable the CPU cycle counter */ >>> + val |= BIT(0); >>> + writel_relaxed(val, drv_data->base + setup_regs->reg_cycle_counter); >>> + >>> + /* CoreCount DCVS Policy: Wait time for frequency inc/decrement */ >>> + val = FIELD_PREP(HYSTERESIS_UP_MASK, HYSTERESIS_CC_NS); >>> + val |= FIELD_PREP(HYSTERESIS_DN_MASK, HYSTERESIS_CC_NS); >>> + writel_relaxed(val, drv_data->base + setup_regs->reg_spm_cc_hyst); >>> + >>> + /* Set the frequency index 0 and override for cluster power collapse */ >>> + writel_relaxed(BIT(0), drv_data->base + setup_regs->reg_cc_zero_behav); >>> + >>> + /* Treat cores in retention as active */ >>> + writel_relaxed(0, drv_data->base + setup_regs->reg_spm_core_ret_map); >>> + >>> + /* Enable CoreCount based DCVS */ >>> + writel_relaxed(0, drv_data->base + setup_regs->reg_spm_cc_dcvs_dis); >>> + >>> + /* CoreCount DCVS-LLM Policy: Wait time for frequency inc/decrement */ >>> + val = FIELD_PREP(HYSTERESIS_UP_MASK, HYSTERESIS_LLM_NS); >>> + val |= FIELD_PREP(HYSTERESIS_DN_MASK, HYSTERESIS_LLM_NS); >>> + writel_relaxed(val, drv_data->base + setup_regs->reg_llm_freq_vote_hyst); >>> + >>> + /* CoreCount DCVS-LLM Policy: Wait time for voltage inc/decrement */ >>> + val = FIELD_PREP(HYSTERESIS_UP_MASK, HYSTERESIS_LLM_NS); >>> + val |= FIELD_PREP(HYSTERESIS_DN_MASK, HYSTERESIS_LLM_NS); >>> + writel_relaxed(val, drv_data->base + setup_regs->reg_llm_volt_vote_hyst); >>> + >>> + /* Enable LLM frequency+voltage voting */ >>> + writel_relaxed(0, drv_data->base + setup_regs->reg_llm_intf_dcvs_dis); >>> + >>> + /* Setup Boost FSM Timers */ >>> + qcom_cpufreq_hw_boost_setup(drv_data->base + >>> + setup_regs->reg_cc_boost_timer, >>> + setup_regs->boost_timer_reg_len); >>> + qcom_cpufreq_hw_boost_setup(drv_data->base + >>> + setup_regs->reg_dcvs_boost_timer, >>> + setup_regs->boost_timer_reg_len); >>> + qcom_cpufreq_hw_boost_setup(drv_data->base + >>> + setup_regs->reg_ps_boost_timer, >>> + setup_regs->boost_timer_reg_len); >>> + >>> + /* PLL signal timing control for Boost */ >>> + writel_relaxed(BOOST_SYNC_DELAY, >>> + drv_data->base + setup_regs->reg_boost_sync_delay); >>> + >>> + /* Setup WFx and PC/RET droop unstall */ >>> + val = FIELD_PREP(DROOP_TIMER1, DROOP_TIMER_NS); >>> + val |= FIELD_PREP(DROOP_TIMER0, DROOP_TIMER_NS); >>> + writel_relaxed(val, drv_data->base + setup_regs->reg_droop_unstall_ctrl); >>> + >>> + /* Setup WFx and PC/RET droop wait-to-release */ >>> + val = FIELD_PREP(DROOP_TIMER1, DROOP_WAIT_RELEASE_TIMER_NS); >>> + val |= FIELD_PREP(DROOP_TIMER0, DROOP_WAIT_RELEASE_TIMER_NS); >>> + writel_relaxed(val, >>> + drv_data->base + setup_regs->reg_droop_wait_release_ctrl); >>> + >>> + /* PLL signal timing control for Droop */ >>> + writel_relaxed(1, drv_data->base + setup_regs->reg_droop_sync_delay); >>> + >>> + /* Setup DCVS timers */ >>> + writel_relaxed(DROOP_RELEASE_TIMER_NS, >>> + drv_data->base + setup_regs->reg_droop_release_ctrl); >>> + writel_relaxed(DROOP_TIMER_NS, >>> + drv_data->base + setup_regs->reg_droop_timer_ctrl); >>> + >>> + /* Setup Droop control */ >>> + val = readl_relaxed(drv_data->base + setup_regs->reg_droop_ctrl); >>> + val |= DROOP_CTRL_VAL; >>> + writel_relaxed(val, drv_data->base + setup_regs->reg_droop_ctrl); >>> + >>> + /* Enable CC-Boost, DCVS-Boost, PS-Boost, WFx, PC/RET, DCVS FSM */ >>> + val = readl_relaxed(drv_data->base + setup_regs->reg_pdn_fsm_ctrl); >>> + val |= CC_BOOST_EN | PS_BOOST_EN | DCVS_BOOST_EN; >>> + val |= WFX_DROOP_EN | PC_RET_EXIT_DROOP_EN | DCVS_DROOP_EN; >>> + writel_relaxed(val, drv_data->base + setup_regs->reg_pdn_fsm_ctrl); >>> + >>> + /* Enable PLL Droop Override */ >>> + val = PLL_OVERRIDE_DROOP_EN; >>> + writel_relaxed(val, drv_data->base + setup_regs->reg_pll_override); >>> + >>> + /* Initialize the Adaptive Clock Distribution */ >>> + ret = qcom_cpufreq_hw_acd_init(cpu_dev, policy, index); >>> + if (ret) >>> + return ret; >>> + >>> + /* We're ready: enable the OSM and give it time to boot (5uS) */ >>> + writel_relaxed(1, drv_data->base + drv_data->soc_data->reg_enable); >>> + udelay(OSM_BOOT_TIME_US); >>> + >>> + return ret; >>> +} >>> + >>> static int qcom_cpufreq_hw_cpu_init(struct cpufreq_policy *policy) >>> { >>> struct platform_device *pdev = cpufreq_get_driver_data(); >>> @@ -282,7 +1400,8 @@ static int qcom_cpufreq_hw_cpu_init(struct cpufreq_policy *policy) >>> struct device *cpu_dev; >>> void __iomem *base; >>> struct qcom_cpufreq_data *data; >>> - int ret, index; >>> + const char *fdom_resname; >>> + int cpu_count, index, ret; >>> cpu_dev = get_cpu_device(policy->cpu); >>> if (!cpu_dev) { >>> @@ -303,9 +1422,14 @@ static int qcom_cpufreq_hw_cpu_init(struct cpufreq_policy *policy) >>> index = args.args[0]; >>> - base = devm_platform_ioremap_resource(pdev, index); >>> + fdom_resname = kasprintf(GFP_KERNEL, "freq-domain%d", index); >>> + if (!fdom_resname) >>> + return -ENOMEM; >>> + >>> + base = devm_platform_ioremap_resource_byname(pdev, fdom_resname); >>> if (IS_ERR(base)) >>> return PTR_ERR(base); >>> + kfree(fdom_resname); >>> data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL); >>> if (!data) { >>> @@ -315,22 +1439,31 @@ static int qcom_cpufreq_hw_cpu_init(struct cpufreq_policy *policy) >>> data->soc_data = of_device_get_match_data(&pdev->dev); >>> data->base = base; >>> + policy->driver_data = data; >>> - /* HW should be in enabled state to proceed */ >>> - if (!(readl_relaxed(base + data->soc_data->reg_enable) & 0x1)) { >>> - dev_err(dev, "Domain-%d cpufreq hardware not enabled\n", index); >>> - ret = -ENODEV; >>> - goto error; >>> - } >>> - >>> - qcom_get_related_cpus(index, policy->cpus); >>> + cpu_count = qcom_get_related_cpus(index, policy->cpus); >>> if (!cpumask_weight(policy->cpus)) { >>> dev_err(dev, "Domain-%d failed to get related CPUs\n", index); >>> ret = -ENOENT; >>> goto error; >>> } >>> - policy->driver_data = data; >>> + if (!data->soc_data->uses_tz) { >>> + ret = qcom_cpufreq_hw_osm_setup(cpu_dev, policy, >>> + cpu_count, index); >>> + if (ret) { >>> + dev_err(dev, "Cannot setup the OSM for CPU%d: %d\n", >>> + policy->cpu, ret); >>> + goto error; >>> + } >>> + } >>> + >>> + /* HW should be in enabled state to proceed */ >>> + if (!(readl_relaxed(base + data->soc_data->reg_enable) & 0x1)) { >>> + dev_err(dev, "Domain-%d cpufreq hardware not enabled\n", index); >>> + ret = -ENODEV; >>> + goto error; >>> + } >>> ret = qcom_cpufreq_hw_read_lut(cpu_dev, policy); >>> if (ret) { >>> @@ -355,6 +1488,7 @@ static int qcom_cpufreq_hw_cpu_init(struct cpufreq_policy *policy) >>> return 0; >>> error: >>> + policy->driver_data = NULL; >>> devm_iounmap(dev, base); >>> return ret; >>> } >>> @@ -395,9 +1529,51 @@ static struct cpufreq_driver cpufreq_qcom_hw_driver = { >>> static int qcom_cpufreq_hw_driver_probe(struct platform_device *pdev) >>> { >>> + const struct qcom_cpufreq_soc_data *soc_data; >>> + struct device_node *pd_node; >>> + struct platform_device *pd_dev; >>> struct device *cpu_dev; >>> struct clk *clk; >>> - int ret; >>> + int clk_div, ret; >>> + >>> + cpu_dev = get_cpu_device(0); >>> + if (!cpu_dev) >>> + return -EPROBE_DEFER; >>> + >>> + soc_data = of_device_get_match_data(&pdev->dev); >>> + if (!soc_data) >>> + return -EINVAL; >>> + >>> + if (!soc_data->uses_tz) { >>> + /* >>> + * When the OSM is not pre-programmed from TZ, we will >>> + * need to program the sequencer through SCM calls. >>> + */ >>> + if (!qcom_scm_is_available()) >>> + return -EPROBE_DEFER; >>> + >>> + /* >>> + * If there are no power-domains, OSM programming cannot be >>> + * performed, as in that case, we wouldn't know where to take >>> + * the params from... >>> + */ >>> + pd_node = of_parse_phandle(cpu_dev->of_node, >>> + "power-domains", 0); >>> + if (!pd_node) { >>> + ret = PTR_ERR(pd_node); >>> + dev_err(cpu_dev, "power domain not found: %d\n", ret); >>> + return ret; >>> + } >>> + >>> + /* >>> + * If the power domain device is not registered yet, then >>> + * defer probing this driver until that is available. >>> + */ >>> + pd_dev = of_find_device_by_node(pd_node); >>> + if (!pd_dev || !pd_dev->dev.driver || >>> + !device_is_bound(&pd_dev->dev)) >>> + return -EPROBE_DEFER; >>> + } >>> clk = clk_get(&pdev->dev, "xo"); >>> if (IS_ERR(clk)) >>> @@ -410,16 +1586,16 @@ static int qcom_cpufreq_hw_driver_probe(struct platform_device *pdev) >>> if (IS_ERR(clk)) >>> return PTR_ERR(clk); >>> - cpu_hw_rate = clk_get_rate(clk) / CLK_HW_DIV; >>> + clk_div = soc_data->clk_hw_div; >>> + if (clk_div == 0) >>> + clk_div++; >>> + >>> + cpu_hw_rate = clk_get_rate(clk) / clk_div; >>> clk_put(clk); >>> cpufreq_qcom_hw_driver.driver_data = pdev; >>> /* Check for optional interconnect paths on CPU0 */ >>> - cpu_dev = get_cpu_device(0); >>> - if (!cpu_dev) >>> - return -EPROBE_DEFER; >>> - >>> ret = dev_pm_opp_of_find_icc_paths(cpu_dev, NULL); >>> if (ret) >>> return ret; >>> >> >> -- >> QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member >> of Code Aurora Forum, hosted by The Linux Foundation. >> >> -- -- QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, hosted by The Linux Foundation. --
On Mon 12 Apr 00:11 CDT 2021, Viresh Kumar wrote: > On 19-01-21, 18:45, AngeloGioacchino Del Regno wrote: > > ** > > ** NOTE: To "view the full picture", please look at the following > > ** patch series: > > ** https://patchwork.kernel.org/project/linux-arm-msm/list/?series=413355 > > ** This is a subset of that series. > > ** > > > > Changes in v4: > > - Huge patch series has been split for better reviewability, > > as suggested by Bjorn > > - Rebased code on top of 266991721c15 ("cpufreq: qcom-hw: enable boost > > support") > > Bjorn, what am I supposed to do with patchset ? Is there anyone who can give > this some reviews from qcom team ? > Sorry Viresh, I've been postponing reviewing this a few times too many. I'll take the time in the coming days to look through it properly, including Taniya's feedback. Regards, Bjorn
On 12-04-21, 15:01, Taniya Das wrote: > Technically the HW we are trying to program here differs in terms of > clocking, the LUT definitions and many more. It will definitely make > debugging much more troublesome if we try to accomodate multiple versions of > CPUFREQ-HW in the same code. > > Thus to keep it simple, easy to read, debug, the suggestion is to keep it > with "v1" tag as the OSM version we are trying to put here is from OSM1.0. That is a valid point and is always a case with so many drivers. What I am concerned about is how much code is common across versions, if it is 5-70%, or more, then we should definitely share, arrange to have callbacks or ops per version and call them in a generic fashion instead of writing a new driver. This is what's done across drivers/frameworks, etc.
On 4/13/2021 9:19 AM, Viresh Kumar wrote: > On 12-04-21, 15:01, Taniya Das wrote: >> Technically the HW we are trying to program here differs in terms of >> clocking, the LUT definitions and many more. It will definitely make >> debugging much more troublesome if we try to accommodate multiple versions of >> CPUFREQ-HW in the same code. >> >> Thus to keep it simple, easy to read, debug, the suggestion is to keep it >> with "v1" tag as the OSM version we are trying to put here is from OSM1.0. > > That is a valid point and is always a case with so many drivers. What > I am concerned about is how much code is common across versions, if it > is 5-70%, or more, then we should definitely share, arrange to have > callbacks or ops per version and call them in a generic fashion instead > of writing a new driver. This is what's done across > drivers/frameworks, etc. > The code sharing here between versions should be very minimal as most portion of the code here in V1 would focus on programming to prepare the LUT to be further read by the driver, the programming in itself is huge for v1. I am okay if you move the v1 in a different file and invoke based on version. -- QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, hosted by The Linux Foundation. --
On Wed 14 Apr 20:31 CDT 2021, Taniya Das wrote: > > On 4/13/2021 9:19 AM, Viresh Kumar wrote: > > On 12-04-21, 15:01, Taniya Das wrote: > > > Technically the HW we are trying to program here differs in terms of > > > clocking, the LUT definitions and many more. It will definitely make > > > debugging much more troublesome if we try to accommodate multiple versions of > > > CPUFREQ-HW in the same code. > > > > > > Thus to keep it simple, easy to read, debug, the suggestion is to keep it > > > with "v1" tag as the OSM version we are trying to put here is from OSM1.0. > > > > That is a valid point and is always a case with so many drivers. What > > I am concerned about is how much code is common across versions, if it > > is 5-70%, or more, then we should definitely share, arrange to have > > callbacks or ops per version and call them in a generic fashion instead > > of writing a new driver. This is what's done across > > drivers/frameworks, etc. > > > > The code sharing here between versions should be very minimal as most > portion of the code here in V1 would focus on programming to prepare the LUT > to be further read by the driver, the programming in itself is huge for v1. > I am okay if you move the v1 in a different file and invoke based on > version. > The initialization of the hardware certainly is a large chunk of code, but once initialized it's the same driver. Your argument that this new code makes it harder to debug things doesn't seem relevant, as the old support doesn't seem to affect the code paths used on the modern version. Creating a new driver and picking it by compatible is certainly possible to do, but looking at Angelo's patch it seems like it would contain 100% of what's in qcom-cpufreq-hw today. As such, I like the approach suggested in this series. Regards, Bjorn
** ** NOTE: To "view the full picture", please look at the following ** patch series: ** https://patchwork.kernel.org/project/linux-arm-msm/list/?series=413355 ** This is a subset of that series. ** Changes in v4: - Huge patch series has been split for better reviewability, as suggested by Bjorn - Rebased code on top of 266991721c15 ("cpufreq: qcom-hw: enable boost support") Changes in v3: - Fixed a test robot build failure for ARCH=arm - Fixed dt_binding_check YAML doc issues Changes in v2: - Rebased dt-binding on top of Manivannan's patches - Added MSM8998 to cpufreq-dt-platdev blacklist - Implemented dynamic Memory Accelerator corners support, needed by MSM8998 - Implemented ACD programming, needed by MSM8998 Tested on the following smartphones: - Sony Xperia XA2 (SDM630) - Sony Xperia XA2 Ultra (SDM630) - Sony Xperia 10 (SDM630) - Sony Xperia XZ Premium (MSM8998) - F(x)Tec Pro 1 (MSM8998)