drivers/base/arch_topology.c

   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/arch_topology.h>
  11 #include <linux/cpu.h>
  12 #include <linux/cpufreq.h>
  13 #include <linux/device.h>
  14 #include <linux/of.h>
  15 #include <linux/slab.h>
  16 #include <linux/string.h>
  17 #include <linux/sched/topology.h>
  18 #include <linux/cpuset.h>
  19
  20 DEFINE_PER_CPU(unsigned long, freq_scale) = SCHED_CAPACITY_SCALE;
  21
  22 void arch_set_freq_scale(struct cpumask *cpus, unsigned long cur_freq,
  23                          unsigned long max_freq)
  24 {
  25         unsigned long scale;
  26         int i;
  27
  28         scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq;
  29
  30         for_each_cpu(i, cpus)
  31                 per_cpu(freq_scale, i) = scale;
  32 }
  33
  34 static DEFINE_MUTEX(cpu_scale_mutex);
  35 DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
  36
  37 void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
  38 {
  39         per_cpu(cpu_scale, cpu) = capacity;
  40 }
  41
  42 static ssize_t cpu_capacity_show(struct device *dev,
  43                                  struct device_attribute *attr,
  44                                  char *buf)
  45 {
  46         struct cpu *cpu = container_of(dev, struct cpu, dev);
  47
  48         return sprintf(buf, "%lu\n", topology_get_cpu_scale(NULL, cpu->dev.id));
  49 }
  50
  51 static void update_topology_flags_workfn(struct work_struct *work);
  52 static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
  53
  54 static ssize_t cpu_capacity_store(struct device *dev,
  55                                   struct device_attribute *attr,
  56                                   const char *buf,
  57                                   size_t count)
  58 {
  59         struct cpu *cpu = container_of(dev, struct cpu, dev);
  60         int this_cpu = cpu->dev.id;
  61         int i;
  62         unsigned long new_capacity;
  63         ssize_t ret;
  64
  65         if (!count)
  66                 return 0;
  67
  68         ret = kstrtoul(buf, 0, &new_capacity);
  69         if (ret)
  70                 return ret;
  71         if (new_capacity > SCHED_CAPACITY_SCALE)
  72                 return -EINVAL;
  73
  74         mutex_lock(&cpu_scale_mutex);
  75         for_each_cpu(i, &cpu_topology[this_cpu].core_sibling)
  76                 topology_set_cpu_scale(i, new_capacity);
  77         mutex_unlock(&cpu_scale_mutex);
  78
  79         schedule_work(&update_topology_flags_work);
  80
  81         return count;
  82 }
  83
  84 static DEVICE_ATTR_RW(cpu_capacity);
  85
  86 static int register_cpu_capacity_sysctl(void)
  87 {
  88         int i;
  89         struct device *cpu;
  90
  91         for_each_possible_cpu(i) {
  92                 cpu = get_cpu_device(i);
  93                 if (!cpu) {
  94                         pr_err("%s: too early to get CPU%d device!\n",
  95                                __func__, i);
  96                         continue;
  97                 }
  98                 device_create_file(cpu, &dev_attr_cpu_capacity);
  99         }
 100
 101         return 0;
 102 }
 103 subsys_initcall(register_cpu_capacity_sysctl);
 104
 105 static int update_topology;
 106
 107 int topology_update_cpu_topology(void)
 108 {
 109         return update_topology;
 110 }
 111
 112 /*
 113  * Updating the sched_domains can't be done directly from cpufreq callbacks
 114  * due to locking, so queue the work for later.
 115  */
 116 static void update_topology_flags_workfn(struct work_struct *work)
 117 {
 118         update_topology = 1;
 119         rebuild_sched_domains();
 120         pr_debug("sched_domain hierarchy rebuilt, flags updated\n");
 121         update_topology = 0;
 122 }
 123
 124 static u32 capacity_scale;
 125 static u32 *raw_capacity;
 126
 127 static int free_raw_capacity(void)
 128 {
 129         kfree(raw_capacity);
 130         raw_capacity = NULL;
 131
 132         return 0;
 133 }
 134
 135 void topology_normalize_cpu_scale(void)
 136 {
 137         u64 capacity;
 138         int cpu;
 139
 140         if (!raw_capacity)
 141                 return;
 142
 143         pr_debug("cpu_capacity: capacity_scale=%u\n", capacity_scale);
 144         mutex_lock(&cpu_scale_mutex);
 145         for_each_possible_cpu(cpu) {
 146                 pr_debug("cpu_capacity: cpu=%d raw_capacity=%u\n",
 147                          cpu, raw_capacity[cpu]);
 148                 capacity = (raw_capacity[cpu] << SCHED_CAPACITY_SHIFT)
 149                         / capacity_scale;
 150                 topology_set_cpu_scale(cpu, capacity);
 151                 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
 152                         cpu, topology_get_cpu_scale(NULL, cpu));
 153         }
 154         mutex_unlock(&cpu_scale_mutex);
 155 }
 156
 157 bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu)
 158 {
 159         static bool cap_parsing_failed;
 160         int ret;
 161         u32 cpu_capacity;
 162
 163         if (cap_parsing_failed)
 164                 return false;
 165
 166         ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz",
 167                                    &cpu_capacity);
 168         if (!ret) {
 169                 if (!raw_capacity) {
 170                         raw_capacity = kcalloc(num_possible_cpus(),
 171                                                sizeof(*raw_capacity),
 172                                                GFP_KERNEL);
 173                         if (!raw_capacity) {
 174                                 pr_err("cpu_capacity: failed to allocate memory for raw capacities\n");
 175                                 cap_parsing_failed = true;
 176                                 return false;
 177                         }
 178                 }
 179                 capacity_scale = max(cpu_capacity, capacity_scale);
 180                 raw_capacity[cpu] = cpu_capacity;
 181                 pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
 182                         cpu_node, raw_capacity[cpu]);
 183         } else {
 184                 if (raw_capacity) {
 185                         pr_err("cpu_capacity: missing %pOF raw capacity\n",
 186                                 cpu_node);
 187                         pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
 188                 }
 189                 cap_parsing_failed = true;
 190                 free_raw_capacity();
 191         }
 192
 193         return !ret;
 194 }
 195
 196 #ifdef CONFIG_CPU_FREQ
 197 static cpumask_var_t cpus_to_visit;
 198 static void parsing_done_workfn(struct work_struct *work);
 199 static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
 200
 201 static int
 202 init_cpu_capacity_callback(struct notifier_block *nb,
 203                            unsigned long val,
 204                            void *data)
 205 {
 206         struct cpufreq_policy *policy = data;
 207         int cpu;
 208
 209         if (!raw_capacity)
 210                 return 0;
 211
 212         if (val != CPUFREQ_NOTIFY)
 213                 return 0;
 214
 215         pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
 216                  cpumask_pr_args(policy->related_cpus),
 217                  cpumask_pr_args(cpus_to_visit));
 218
 219         cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
 220
 221         for_each_cpu(cpu, policy->related_cpus) {
 222                 raw_capacity[cpu] = topology_get_cpu_scale(NULL, cpu) *
 223                                     policy->cpuinfo.max_freq / 1000UL;
 224                 capacity_scale = max(raw_capacity[cpu], capacity_scale);
 225         }
 226
 227         if (cpumask_empty(cpus_to_visit)) {
 228                 topology_normalize_cpu_scale();
 229                 schedule_work(&update_topology_flags_work);
 230                 free_raw_capacity();
 231                 pr_debug("cpu_capacity: parsing done\n");
 232                 schedule_work(&parsing_done_work);
 233         }
 234
 235         return 0;
 236 }
 237
 238 static struct notifier_block init_cpu_capacity_notifier = {
 239         .notifier_call = init_cpu_capacity_callback,
 240 };
 241
 242 static int __init register_cpufreq_notifier(void)
 243 {
 244         int ret;
 245
 246         /*
 247          * on ACPI-based systems we need to use the default cpu capacity
 248          * until we have the necessary code to parse the cpu capacity, so
 249          * skip registering cpufreq notifier.
 250          */
 251         if (!acpi_disabled || !raw_capacity)
 252                 return -EINVAL;
 253
 254         if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL)) {
 255                 pr_err("cpu_capacity: failed to allocate memory for cpus_to_visit\n");
 256                 return -ENOMEM;
 257         }
 258
 259         cpumask_copy(cpus_to_visit, cpu_possible_mask);
 260
 261         ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
 262                                         CPUFREQ_POLICY_NOTIFIER);
 263
 264         if (ret)
 265                 free_cpumask_var(cpus_to_visit);
 266
 267         return ret;
 268 }
 269 core_initcall(register_cpufreq_notifier);
 270
 271 static void parsing_done_workfn(struct work_struct *work)
 272 {
 273         cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
 274                                          CPUFREQ_POLICY_NOTIFIER);
 275         free_cpumask_var(cpus_to_visit);
 276 }
 277
 278 #else
 279 core_initcall(free_raw_capacity);
 280 #endif