linux/drivers/base/arch_topology.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Arch specific cpu topology information
   4 *
   5 * Copyright (C) 2016, ARM Ltd.
   6 * Written by: Juri Lelli, ARM Ltd.
   7 */
   8
   9#include <linux/acpi.h>
  10#include <linux/cpu.h>
  11#include <linux/cpufreq.h>
  12#include <linux/device.h>
  13#include <linux/of.h>
  14#include <linux/slab.h>
  15#include <linux/string.h>
  16#include <linux/sched/topology.h>
  17#include <linux/cpuset.h>
  18
  19DEFINE_PER_CPU(unsigned long, freq_scale) = SCHED_CAPACITY_SCALE;
  20
  21void arch_set_freq_scale(struct cpumask *cpus, unsigned long cur_freq,
  22                         unsigned long max_freq)
  23{
  24        unsigned long scale;
  25        int i;
  26
  27        scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq;
  28
  29        for_each_cpu(i, cpus)
  30                per_cpu(freq_scale, i) = scale;
  31}
  32
  33DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
  34
  35void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
  36{
  37        per_cpu(cpu_scale, cpu) = capacity;
  38}
  39
  40static ssize_t cpu_capacity_show(struct device *dev,
  41                                 struct device_attribute *attr,
  42                                 char *buf)
  43{
  44        struct cpu *cpu = container_of(dev, struct cpu, dev);
  45
  46        return sprintf(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id));
  47}
  48
  49static void update_topology_flags_workfn(struct work_struct *work);
  50static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
  51
  52static DEVICE_ATTR_RO(cpu_capacity);
  53
  54static int register_cpu_capacity_sysctl(void)
  55{
  56        int i;
  57        struct device *cpu;
  58
  59        for_each_possible_cpu(i) {
  60                cpu = get_cpu_device(i);
  61                if (!cpu) {
  62                        pr_err("%s: too early to get CPU%d device!\n",
  63                               __func__, i);
  64                        continue;
  65                }
  66                device_create_file(cpu, &dev_attr_cpu_capacity);
  67        }
  68
  69        return 0;
  70}
  71subsys_initcall(register_cpu_capacity_sysctl);
  72
  73static int update_topology;
  74
  75int topology_update_cpu_topology(void)
  76{
  77        return update_topology;
  78}
  79
  80/*
  81 * Updating the sched_domains can't be done directly from cpufreq callbacks
  82 * due to locking, so queue the work for later.
  83 */
  84static void update_topology_flags_workfn(struct work_struct *work)
  85{
  86        update_topology = 1;
  87        rebuild_sched_domains();
  88        pr_debug("sched_domain hierarchy rebuilt, flags updated\n");
  89        update_topology = 0;
  90}
  91
  92static u32 capacity_scale;
  93static u32 *raw_capacity;
  94
  95static int free_raw_capacity(void)
  96{
  97        kfree(raw_capacity);
  98        raw_capacity = NULL;
  99
 100        return 0;
 101}
 102
 103void topology_normalize_cpu_scale(void)
 104{
 105        u64 capacity;
 106        int cpu;
 107
 108        if (!raw_capacity)
 109                return;
 110
 111        pr_debug("cpu_capacity: capacity_scale=%u\n", capacity_scale);
 112        for_each_possible_cpu(cpu) {
 113                pr_debug("cpu_capacity: cpu=%d raw_capacity=%u\n",
 114                         cpu, raw_capacity[cpu]);
 115                capacity = (raw_capacity[cpu] << SCHED_CAPACITY_SHIFT)
 116                        / capacity_scale;
 117                topology_set_cpu_scale(cpu, capacity);
 118                pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
 119                        cpu, topology_get_cpu_scale(cpu));
 120        }
 121}
 122
 123bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu)
 124{
 125        static bool cap_parsing_failed;
 126        int ret;
 127        u32 cpu_capacity;
 128
 129        if (cap_parsing_failed)
 130                return false;
 131
 132        ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz",
 133                                   &cpu_capacity);
 134        if (!ret) {
 135                if (!raw_capacity) {
 136                        raw_capacity = kcalloc(num_possible_cpus(),
 137                                               sizeof(*raw_capacity),
 138                                               GFP_KERNEL);
 139                        if (!raw_capacity) {
 140                                cap_parsing_failed = true;
 141                                return false;
 142                        }
 143                }
 144                capacity_scale = max(cpu_capacity, capacity_scale);
 145                raw_capacity[cpu] = cpu_capacity;
 146                pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
 147                        cpu_node, raw_capacity[cpu]);
 148        } else {
 149                if (raw_capacity) {
 150                        pr_err("cpu_capacity: missing %pOF raw capacity\n",
 151                                cpu_node);
 152                        pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
 153                }
 154                cap_parsing_failed = true;
 155                free_raw_capacity();
 156        }
 157
 158        return !ret;
 159}
 160
 161#ifdef CONFIG_CPU_FREQ
 162static cpumask_var_t cpus_to_visit;
 163static void parsing_done_workfn(struct work_struct *work);
 164static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
 165
 166static int
 167init_cpu_capacity_callback(struct notifier_block *nb,
 168                           unsigned long val,
 169                           void *data)
 170{
 171        struct cpufreq_policy *policy = data;
 172        int cpu;
 173
 174        if (!raw_capacity)
 175                return 0;
 176
 177        if (val != CPUFREQ_NOTIFY)
 178                return 0;
 179
 180        pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
 181                 cpumask_pr_args(policy->related_cpus),
 182                 cpumask_pr_args(cpus_to_visit));
 183
 184        cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
 185
 186        for_each_cpu(cpu, policy->related_cpus) {
 187                raw_capacity[cpu] = topology_get_cpu_scale(cpu) *
 188                                    policy->cpuinfo.max_freq / 1000UL;
 189                capacity_scale = max(raw_capacity[cpu], capacity_scale);
 190        }
 191
 192        if (cpumask_empty(cpus_to_visit)) {
 193                topology_normalize_cpu_scale();
 194                schedule_work(&update_topology_flags_work);
 195                free_raw_capacity();
 196                pr_debug("cpu_capacity: parsing done\n");
 197                schedule_work(&parsing_done_work);
 198        }
 199
 200        return 0;
 201}
 202
 203static struct notifier_block init_cpu_capacity_notifier = {
 204        .notifier_call = init_cpu_capacity_callback,
 205};
 206
 207static int __init register_cpufreq_notifier(void)
 208{
 209        int ret;
 210
 211        /*
 212         * on ACPI-based systems we need to use the default cpu capacity
 213         * until we have the necessary code to parse the cpu capacity, so
 214         * skip registering cpufreq notifier.
 215         */
 216        if (!acpi_disabled || !raw_capacity)
 217                return -EINVAL;
 218
 219        if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL))
 220                return -ENOMEM;
 221
 222        cpumask_copy(cpus_to_visit, cpu_possible_mask);
 223
 224        ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
 225                                        CPUFREQ_POLICY_NOTIFIER);
 226
 227        if (ret)
 228                free_cpumask_var(cpus_to_visit);
 229
 230        return ret;
 231}
 232core_initcall(register_cpufreq_notifier);
 233
 234static void parsing_done_workfn(struct work_struct *work)
 235{
 236        cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
 237                                         CPUFREQ_POLICY_NOTIFIER);
 238        free_cpumask_var(cpus_to_visit);
 239}
 240
 241#else
 242core_initcall(free_raw_capacity);
 243#endif
 244