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