arch/arm64/kernel/topology.c

   1 /*
   2  * arch/arm64/kernel/topology.c
   3  *
   4  * Copyright (C) 2011,2013,2014 Linaro Limited.
   5  *
   6  * Based on the arm32 version written by Vincent Guittot in turn based on
   7  * arch/sh/kernel/topology.c
   8  *
   9  * This file is subject to the terms and conditions of the GNU General Public
  10  * License.  See the file "COPYING" in the main directory of this archive
  11  * for more details.
  12  */
  13
  14 #include <linux/acpi.h>
  15 #include <linux/cpu.h>
  16 #include <linux/cpumask.h>
  17 #include <linux/init.h>
  18 #include <linux/percpu.h>
  19 #include <linux/node.h>
  20 #include <linux/nodemask.h>
  21 #include <linux/of.h>
  22 #include <linux/sched.h>
  23 #include <linux/slab.h>
  24 #include <linux/string.h>
  25 #include <linux/cpufreq.h>
  26
  27 #include <asm/cpu.h>
  28 #include <asm/cputype.h>
  29 #include <asm/topology.h>
  30
  31 static DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
  32 static DEFINE_MUTEX(cpu_scale_mutex);
  33
  34 unsigned long arch_scale_cpu_capacity(struct sched_domain *sd, int cpu)
  35 {
  36         return per_cpu(cpu_scale, cpu);
  37 }
  38
  39 static void set_capacity_scale(unsigned int cpu, unsigned long capacity)
  40 {
  41         per_cpu(cpu_scale, cpu) = capacity;
  42 }
  43
  44 #ifdef CONFIG_PROC_SYSCTL
  45 static ssize_t cpu_capacity_show(struct device *dev,
  46                                  struct device_attribute *attr,
  47                                  char *buf)
  48 {
  49         struct cpu *cpu = container_of(dev, struct cpu, dev);
  50
  51         return sprintf(buf, "%lu\n",
  52                         arch_scale_cpu_capacity(NULL, cpu->dev.id));
  53 }
  54
  55 static ssize_t cpu_capacity_store(struct device *dev,
  56                                   struct device_attribute *attr,
  57                                   const char *buf,
  58                                   size_t count)
  59 {
  60         struct cpu *cpu = container_of(dev, struct cpu, dev);
  61         int this_cpu = cpu->dev.id, i;
  62         unsigned long new_capacity;
  63         ssize_t ret;
  64
  65         if (count) {
  66                 ret = kstrtoul(buf, 0, &new_capacity);
  67                 if (ret)
  68                         return ret;
  69                 if (new_capacity > SCHED_CAPACITY_SCALE)
  70                         return -EINVAL;
  71
  72                 mutex_lock(&cpu_scale_mutex);
  73                 for_each_cpu(i, &cpu_topology[this_cpu].core_sibling)
  74                         set_capacity_scale(i, new_capacity);
  75                 mutex_unlock(&cpu_scale_mutex);
  76         }
  77
  78         return count;
  79 }
  80
  81 static DEVICE_ATTR_RW(cpu_capacity);
  82
  83 static int register_cpu_capacity_sysctl(void)
  84 {
  85         int i;
  86         struct device *cpu;
  87
  88         for_each_possible_cpu(i) {
  89                 cpu = get_cpu_device(i);
  90                 if (!cpu) {
  91                         pr_err("%s: too early to get CPU%d device!\n",
  92                                __func__, i);
  93                         continue;
  94                 }
  95                 device_create_file(cpu, &dev_attr_cpu_capacity);
  96         }
  97
  98         return 0;
  99 }
 100 subsys_initcall(register_cpu_capacity_sysctl);
 101 #endif
 102
 103 static u32 capacity_scale;
 104 static u32 *raw_capacity;
 105 static bool cap_parsing_failed;
 106
 107 static void __init parse_cpu_capacity(struct device_node *cpu_node, int cpu)
 108 {
 109         int ret;
 110         u32 cpu_capacity;
 111
 112         if (cap_parsing_failed)
 113                 return;
 114
 115         ret = of_property_read_u32(cpu_node,
 116                                    "capacity-dmips-mhz",
 117                                    &cpu_capacity);
 118         if (!ret) {
 119                 if (!raw_capacity) {
 120                         raw_capacity = kcalloc(num_possible_cpus(),
 121                                                sizeof(*raw_capacity),
 122                                                GFP_KERNEL);
 123                         if (!raw_capacity) {
 124                                 pr_err("cpu_capacity: failed to allocate memory for raw capacities\n");
 125                                 cap_parsing_failed = true;
 126                                 return;
 127                         }
 128                 }
 129                 capacity_scale = max(cpu_capacity, capacity_scale);
 130                 raw_capacity[cpu] = cpu_capacity;
 131                 pr_debug("cpu_capacity: %s cpu_capacity=%u (raw)\n",
 132                         cpu_node->full_name, raw_capacity[cpu]);
 133         } else {
 134                 if (raw_capacity) {
 135                         pr_err("cpu_capacity: missing %s raw capacity\n",
 136                                 cpu_node->full_name);
 137                         pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
 138                 }
 139                 cap_parsing_failed = true;
 140                 kfree(raw_capacity);
 141         }
 142 }
 143
 144 static void normalize_cpu_capacity(void)
 145 {
 146         u64 capacity;
 147         int cpu;
 148
 149         if (!raw_capacity || cap_parsing_failed)
 150                 return;
 151
 152         pr_debug("cpu_capacity: capacity_scale=%u\n", capacity_scale);
 153         mutex_lock(&cpu_scale_mutex);
 154         for_each_possible_cpu(cpu) {
 155                 pr_debug("cpu_capacity: cpu=%d raw_capacity=%u\n",
 156                          cpu, raw_capacity[cpu]);
 157                 capacity = (raw_capacity[cpu] << SCHED_CAPACITY_SHIFT)
 158                         / capacity_scale;
 159                 set_capacity_scale(cpu, capacity);
 160                 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
 161                         cpu, arch_scale_cpu_capacity(NULL, cpu));
 162         }
 163         mutex_unlock(&cpu_scale_mutex);
 164 }
 165
 166 #ifdef CONFIG_CPU_FREQ
 167 static cpumask_var_t cpus_to_visit;
 168 static bool cap_parsing_done;
 169 static void parsing_done_workfn(struct work_struct *work);
 170 static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
 171
 172 static int
 173 init_cpu_capacity_callback(struct notifier_block *nb,
 174                            unsigned long val,
 175                            void *data)
 176 {
 177         struct cpufreq_policy *policy = data;
 178         int cpu;
 179
 180         if (cap_parsing_failed || cap_parsing_done)
 181                 return 0;
 182
 183         switch (val) {
 184         case CPUFREQ_NOTIFY:
 185                 pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
 186                                 cpumask_pr_args(policy->related_cpus),
 187                                 cpumask_pr_args(cpus_to_visit));
 188                 cpumask_andnot(cpus_to_visit,
 189                                cpus_to_visit,
 190                                policy->related_cpus);
 191                 for_each_cpu(cpu, policy->related_cpus) {
 192                         raw_capacity[cpu] = arch_scale_cpu_capacity(NULL, cpu) *
 193                                             policy->cpuinfo.max_freq / 1000UL;
 194                         capacity_scale = max(raw_capacity[cpu], capacity_scale);
 195                 }
 196                 if (cpumask_empty(cpus_to_visit)) {
 197                         normalize_cpu_capacity();
 198                         kfree(raw_capacity);
 199                         pr_debug("cpu_capacity: parsing done\n");
 200                         cap_parsing_done = true;
 201                         schedule_work(&parsing_done_work);
 202                 }
 203         }
 204         return 0;
 205 }
 206
 207 static struct notifier_block init_cpu_capacity_notifier = {
 208         .notifier_call = init_cpu_capacity_callback,
 209 };
 210
 211 static int __init register_cpufreq_notifier(void)
 212 {
 213         /*
 214          * on ACPI-based systems we need to use the default cpu capacity
 215          * until we have the necessary code to parse the cpu capacity, so
 216          * skip registering cpufreq notifier.
 217          */
 218         if (!acpi_disabled || cap_parsing_failed)
 219                 return -EINVAL;
 220
 221         if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL)) {
 222                 pr_err("cpu_capacity: failed to allocate memory for cpus_to_visit\n");
 223                 return -ENOMEM;
 224         }
 225         cpumask_copy(cpus_to_visit, cpu_possible_mask);
 226
 227         return cpufreq_register_notifier(&init_cpu_capacity_notifier,
 228                                          CPUFREQ_POLICY_NOTIFIER);
 229 }
 230 core_initcall(register_cpufreq_notifier);
 231
 232 static void parsing_done_workfn(struct work_struct *work)
 233 {
 234         cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
 235                                          CPUFREQ_POLICY_NOTIFIER);
 236 }
 237
 238 #else
 239 static int __init free_raw_capacity(void)
 240 {
 241         kfree(raw_capacity);
 242
 243         return 0;
 244 }
 245 core_initcall(free_raw_capacity);
 246 #endif
 247
 248 static int __init get_cpu_for_node(struct device_node *node)
 249 {
 250         struct device_node *cpu_node;
 251         int cpu;
 252
 253         cpu_node = of_parse_phandle(node, "cpu", 0);
 254         if (!cpu_node)
 255                 return -1;
 256
 257         for_each_possible_cpu(cpu) {
 258                 if (of_get_cpu_node(cpu, NULL) == cpu_node) {
 259                         parse_cpu_capacity(cpu_node, cpu);
 260                         of_node_put(cpu_node);
 261                         return cpu;
 262                 }
 263         }
 264
 265         pr_crit("Unable to find CPU node for %s\n", cpu_node->full_name);
 266
 267         of_node_put(cpu_node);
 268         return -1;
 269 }
 270
 271 static int __init parse_core(struct device_node *core, int cluster_id,
 272                              int core_id)
 273 {
 274         char name[10];
 275         bool leaf = true;
 276         int i = 0;
 277         int cpu;
 278         struct device_node *t;
 279
 280         do {
 281                 snprintf(name, sizeof(name), "thread%d", i);
 282                 t = of_get_child_by_name(core, name);
 283                 if (t) {
 284                         leaf = false;
 285                         cpu = get_cpu_for_node(t);
 286                         if (cpu >= 0) {
 287                                 cpu_topology[cpu].cluster_id = cluster_id;
 288                                 cpu_topology[cpu].core_id = core_id;
 289                                 cpu_topology[cpu].thread_id = i;
 290                         } else {
 291                                 pr_err("%s: Can't get CPU for thread\n",
 292                                        t->full_name);
 293                                 of_node_put(t);
 294                                 return -EINVAL;
 295                         }
 296                         of_node_put(t);
 297                 }
 298                 i++;
 299         } while (t);
 300
 301         cpu = get_cpu_for_node(core);
 302         if (cpu >= 0) {
 303                 if (!leaf) {
 304                         pr_err("%s: Core has both threads and CPU\n",
 305                                core->full_name);
 306                         return -EINVAL;
 307                 }
 308
 309                 cpu_topology[cpu].cluster_id = cluster_id;
 310                 cpu_topology[cpu].core_id = core_id;
 311         } else if (leaf) {
 312                 pr_err("%s: Can't get CPU for leaf core\n", core->full_name);
 313                 return -EINVAL;
 314         }
 315
 316         return 0;
 317 }
 318
 319 static int __init parse_cluster(struct device_node *cluster, int depth)
 320 {
 321         char name[10];
 322         bool leaf = true;
 323         bool has_cores = false;
 324         struct device_node *c;
 325         static int cluster_id __initdata;
 326         int core_id = 0;
 327         int i, ret;
 328
 329         /*
 330          * First check for child clusters; we currently ignore any
 331          * information about the nesting of clusters and present the
 332          * scheduler with a flat list of them.
 333          */
 334         i = 0;
 335         do {
 336                 snprintf(name, sizeof(name), "cluster%d", i);
 337                 c = of_get_child_by_name(cluster, name);
 338                 if (c) {
 339                         leaf = false;
 340                         ret = parse_cluster(c, depth + 1);
 341                         of_node_put(c);
 342                         if (ret != 0)
 343                                 return ret;
 344                 }
 345                 i++;
 346         } while (c);
 347
 348         /* Now check for cores */
 349         i = 0;
 350         do {
 351                 snprintf(name, sizeof(name), "core%d", i);
 352                 c = of_get_child_by_name(cluster, name);
 353                 if (c) {
 354                         has_cores = true;
 355
 356                         if (depth == 0) {
 357                                 pr_err("%s: cpu-map children should be clusters\n",
 358                                        c->full_name);
 359                                 of_node_put(c);
 360                                 return -EINVAL;
 361                         }
 362
 363                         if (leaf) {
 364                                 ret = parse_core(c, cluster_id, core_id++);
 365                         } else {
 366                                 pr_err("%s: Non-leaf cluster with core %s\n",
 367                                        cluster->full_name, name);
 368                                 ret = -EINVAL;
 369                         }
 370
 371                         of_node_put(c);
 372                         if (ret != 0)
 373                                 return ret;
 374                 }
 375                 i++;
 376         } while (c);
 377
 378         if (leaf && !has_cores)
 379                 pr_warn("%s: empty cluster\n", cluster->full_name);
 380
 381         if (leaf)
 382                 cluster_id++;
 383
 384         return 0;
 385 }
 386
 387 static int __init parse_dt_topology(void)
 388 {
 389         struct device_node *cn, *map;
 390         int ret = 0;
 391         int cpu;
 392
 393         cn = of_find_node_by_path("/cpus");
 394         if (!cn) {
 395                 pr_err("No CPU information found in DT\n");
 396                 return 0;
 397         }
 398
 399         /*
 400          * When topology is provided cpu-map is essentially a root
 401          * cluster with restricted subnodes.
 402          */
 403         map = of_get_child_by_name(cn, "cpu-map");
 404         if (!map) {
 405                 cap_parsing_failed = true;
 406                 goto out;
 407         }
 408
 409         ret = parse_cluster(map, 0);
 410         if (ret != 0)
 411                 goto out_map;
 412
 413         normalize_cpu_capacity();
 414
 415         /*
 416          * Check that all cores are in the topology; the SMP code will
 417          * only mark cores described in the DT as possible.
 418          */
 419         for_each_possible_cpu(cpu)
 420                 if (cpu_topology[cpu].cluster_id == -1)
 421                         ret = -EINVAL;
 422
 423 out_map:
 424         of_node_put(map);
 425 out:
 426         of_node_put(cn);
 427         return ret;
 428 }
 429
 430 /*
 431  * cpu topology table
 432  */
 433 struct cpu_topology cpu_topology[NR_CPUS];
 434 EXPORT_SYMBOL_GPL(cpu_topology);
 435
 436 const struct cpumask *cpu_coregroup_mask(int cpu)
 437 {
 438         return &cpu_topology[cpu].core_sibling;
 439 }
 440
 441 static void update_siblings_masks(unsigned int cpuid)
 442 {
 443         struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
 444         int cpu;
 445
 446         /* update core and thread sibling masks */
 447         for_each_possible_cpu(cpu) {
 448                 cpu_topo = &cpu_topology[cpu];
 449
 450                 if (cpuid_topo->cluster_id != cpu_topo->cluster_id)
 451                         continue;
 452
 453                 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
 454                 if (cpu != cpuid)
 455                         cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
 456
 457                 if (cpuid_topo->core_id != cpu_topo->core_id)
 458                         continue;
 459
 460                 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
 461                 if (cpu != cpuid)
 462                         cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
 463         }
 464 }
 465
 466 void store_cpu_topology(unsigned int cpuid)
 467 {
 468         struct cpu_topology *cpuid_topo = &cpu_topology[cpuid];
 469         u64 mpidr;
 470
 471         if (cpuid_topo->cluster_id != -1)
 472                 goto topology_populated;
 473
 474         mpidr = read_cpuid_mpidr();
 475
 476         /* Uniprocessor systems can rely on default topology values */
 477         if (mpidr & MPIDR_UP_BITMASK)
 478                 return;
 479
 480         /* Create cpu topology mapping based on MPIDR. */
 481         if (mpidr & MPIDR_MT_BITMASK) {
 482                 /* Multiprocessor system : Multi-threads per core */
 483                 cpuid_topo->thread_id  = MPIDR_AFFINITY_LEVEL(mpidr, 0);
 484                 cpuid_topo->core_id    = MPIDR_AFFINITY_LEVEL(mpidr, 1);
 485                 cpuid_topo->cluster_id = MPIDR_AFFINITY_LEVEL(mpidr, 2) |
 486                                          MPIDR_AFFINITY_LEVEL(mpidr, 3) << 8;
 487         } else {
 488                 /* Multiprocessor system : Single-thread per core */
 489                 cpuid_topo->thread_id  = -1;
 490                 cpuid_topo->core_id    = MPIDR_AFFINITY_LEVEL(mpidr, 0);
 491                 cpuid_topo->cluster_id = MPIDR_AFFINITY_LEVEL(mpidr, 1) |
 492                                          MPIDR_AFFINITY_LEVEL(mpidr, 2) << 8 |
 493                                          MPIDR_AFFINITY_LEVEL(mpidr, 3) << 16;
 494         }
 495
 496         pr_debug("CPU%u: cluster %d core %d thread %d mpidr %#016llx\n",
 497                  cpuid, cpuid_topo->cluster_id, cpuid_topo->core_id,
 498                  cpuid_topo->thread_id, mpidr);
 499
 500 topology_populated:
 501         update_siblings_masks(cpuid);
 502 }
 503
 504 static void __init reset_cpu_topology(void)
 505 {
 506         unsigned int cpu;
 507
 508         for_each_possible_cpu(cpu) {
 509                 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
 510
 511                 cpu_topo->thread_id = -1;
 512                 cpu_topo->core_id = 0;
 513                 cpu_topo->cluster_id = -1;
 514
 515                 cpumask_clear(&cpu_topo->core_sibling);
 516                 cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
 517                 cpumask_clear(&cpu_topo->thread_sibling);
 518                 cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
 519         }
 520 }
 521
 522 void __init init_cpu_topology(void)
 523 {
 524         reset_cpu_topology();
 525
 526         /*
 527          * Discard anything that was parsed if we hit an error so we
 528          * don't use partial information.
 529          */
 530         if (of_have_populated_dt() && parse_dt_topology())
 531                 reset_cpu_topology();
 532 }