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/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 #include <linux/cpumask.h>
  19 #include <linux/init.h>
  20 #include <linux/percpu.h>
  21 #include <linux/sched.h>
  22 #include <linux/smp.h>
  23
  24 DEFINE_PER_CPU(unsigned long, freq_scale) = SCHED_CAPACITY_SCALE;
  25
  26 void arch_set_freq_scale(struct cpumask *cpus, unsigned long cur_freq,
  27                          unsigned long max_freq)
  28 {
  29         unsigned long scale;
  30         int i;
  31
  32         scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq;
  33
  34         for_each_cpu(i, cpus)
  35                 per_cpu(freq_scale, i) = scale;
  36 }
  37
  38 DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
  39
  40 void topology_set_cpu_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", topology_get_cpu_scale(cpu->dev.id));
  52 }
  53
  54 static void update_topology_flags_workfn(struct work_struct *work);
  55 static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
  56
  57 static DEVICE_ATTR_RO(cpu_capacity);
  58
  59 static int register_cpu_capacity_sysctl(void)
  60 {
  61         int i;
  62         struct device *cpu;
  63
  64         for_each_possible_cpu(i) {
  65                 cpu = get_cpu_device(i);
  66                 if (!cpu) {
  67                         pr_err("%s: too early to get CPU%d device!\n",
  68                                __func__, i);
  69                         continue;
  70                 }
  71                 device_create_file(cpu, &dev_attr_cpu_capacity);
  72         }
  73
  74         return 0;
  75 }
  76 subsys_initcall(register_cpu_capacity_sysctl);
  77
  78 static int update_topology;
  79
  80 int topology_update_cpu_topology(void)
  81 {
  82         return update_topology;
  83 }
  84
  85 /*
  86  * Updating the sched_domains can't be done directly from cpufreq callbacks
  87  * due to locking, so queue the work for later.
  88  */
  89 static void update_topology_flags_workfn(struct work_struct *work)
  90 {
  91         update_topology = 1;
  92         rebuild_sched_domains();
  93         pr_debug("sched_domain hierarchy rebuilt, flags updated\n");
  94         update_topology = 0;
  95 }
  96
  97 static u32 capacity_scale;
  98 static u32 *raw_capacity;
  99
 100 static int free_raw_capacity(void)
 101 {
 102         kfree(raw_capacity);
 103         raw_capacity = NULL;
 104
 105         return 0;
 106 }
 107
 108 void topology_normalize_cpu_scale(void)
 109 {
 110         u64 capacity;
 111         int cpu;
 112
 113         if (!raw_capacity)
 114                 return;
 115
 116         pr_debug("cpu_capacity: capacity_scale=%u\n", capacity_scale);
 117         for_each_possible_cpu(cpu) {
 118                 pr_debug("cpu_capacity: cpu=%d raw_capacity=%u\n",
 119                          cpu, raw_capacity[cpu]);
 120                 capacity = (raw_capacity[cpu] << SCHED_CAPACITY_SHIFT)
 121                         / capacity_scale;
 122                 topology_set_cpu_scale(cpu, capacity);
 123                 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
 124                         cpu, topology_get_cpu_scale(cpu));
 125         }
 126 }
 127
 128 bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu)
 129 {
 130         static bool cap_parsing_failed;
 131         int ret;
 132         u32 cpu_capacity;
 133
 134         if (cap_parsing_failed)
 135                 return false;
 136
 137         ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz",
 138                                    &cpu_capacity);
 139         if (!ret) {
 140                 if (!raw_capacity) {
 141                         raw_capacity = kcalloc(num_possible_cpus(),
 142                                                sizeof(*raw_capacity),
 143                                                GFP_KERNEL);
 144                         if (!raw_capacity) {
 145                                 cap_parsing_failed = true;
 146                                 return false;
 147                         }
 148                 }
 149                 capacity_scale = max(cpu_capacity, capacity_scale);
 150                 raw_capacity[cpu] = cpu_capacity;
 151                 pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
 152                         cpu_node, raw_capacity[cpu]);
 153         } else {
 154                 if (raw_capacity) {
 155                         pr_err("cpu_capacity: missing %pOF raw capacity\n",
 156                                 cpu_node);
 157                         pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
 158                 }
 159                 cap_parsing_failed = true;
 160                 free_raw_capacity();
 161         }
 162
 163         return !ret;
 164 }
 165
 166 #ifdef CONFIG_CPU_FREQ
 167 static cpumask_var_t cpus_to_visit;
 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 (!raw_capacity)
 180                 return 0;
 181
 182         if (val != CPUFREQ_CREATE_POLICY)
 183                 return 0;
 184
 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
 189         cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
 190
 191         for_each_cpu(cpu, policy->related_cpus) {
 192                 raw_capacity[cpu] = topology_get_cpu_scale(cpu) *
 193                                     policy->cpuinfo.max_freq / 1000UL;
 194                 capacity_scale = max(raw_capacity[cpu], capacity_scale);
 195         }
 196
 197         if (cpumask_empty(cpus_to_visit)) {
 198                 topology_normalize_cpu_scale();
 199                 schedule_work(&update_topology_flags_work);
 200                 free_raw_capacity();
 201                 pr_debug("cpu_capacity: parsing done\n");
 202                 schedule_work(&parsing_done_work);
 203         }
 204
 205         return 0;
 206 }
 207
 208 static struct notifier_block init_cpu_capacity_notifier = {
 209         .notifier_call = init_cpu_capacity_callback,
 210 };
 211
 212 static int __init register_cpufreq_notifier(void)
 213 {
 214         int ret;
 215
 216         /*
 217          * on ACPI-based systems we need to use the default cpu capacity
 218          * until we have the necessary code to parse the cpu capacity, so
 219          * skip registering cpufreq notifier.
 220          */
 221         if (!acpi_disabled || !raw_capacity)
 222                 return -EINVAL;
 223
 224         if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL))
 225                 return -ENOMEM;
 226
 227         cpumask_copy(cpus_to_visit, cpu_possible_mask);
 228
 229         ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
 230                                         CPUFREQ_POLICY_NOTIFIER);
 231
 232         if (ret)
 233                 free_cpumask_var(cpus_to_visit);
 234
 235         return ret;
 236 }
 237 core_initcall(register_cpufreq_notifier);
 238
 239 static void parsing_done_workfn(struct work_struct *work)
 240 {
 241         cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
 242                                          CPUFREQ_POLICY_NOTIFIER);
 243         free_cpumask_var(cpus_to_visit);
 244 }
 245
 246 #else
 247 core_initcall(free_raw_capacity);
 248 #endif
 249
 250 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
 251 /*
 252  * This function returns the logic cpu number of the node.
 253  * There are basically three kinds of return values:
 254  * (1) logic cpu number which is > 0.
 255  * (2) -ENODEV when the device tree(DT) node is valid and found in the DT but
 256  * there is no possible logical CPU in the kernel to match. This happens
 257  * when CONFIG_NR_CPUS is configure to be smaller than the number of
 258  * CPU nodes in DT. We need to just ignore this case.
 259  * (3) -1 if the node does not exist in the device tree
 260  */
 261 static int __init get_cpu_for_node(struct device_node *node)
 262 {
 263         struct device_node *cpu_node;
 264         int cpu;
 265
 266         cpu_node = of_parse_phandle(node, "cpu", 0);
 267         if (!cpu_node)
 268                 return -1;
 269
 270         cpu = of_cpu_node_to_id(cpu_node);
 271         if (cpu >= 0)
 272                 topology_parse_cpu_capacity(cpu_node, cpu);
 273         else
 274                 pr_info("CPU node for %pOF exist but the possible cpu range is :%*pbl\n",
 275                         cpu_node, cpumask_pr_args(cpu_possible_mask));
 276
 277         of_node_put(cpu_node);
 278         return cpu;
 279 }
 280
 281 static int __init parse_core(struct device_node *core, int package_id,
 282                              int core_id)
 283 {
 284         char name[10];
 285         bool leaf = true;
 286         int i = 0;
 287         int cpu;
 288         struct device_node *t;
 289
 290         do {
 291                 snprintf(name, sizeof(name), "thread%d", i);
 292                 t = of_get_child_by_name(core, name);
 293                 if (t) {
 294                         leaf = false;
 295                         cpu = get_cpu_for_node(t);
 296                         if (cpu >= 0) {
 297                                 cpu_topology[cpu].package_id = package_id;
 298                                 cpu_topology[cpu].core_id = core_id;
 299                                 cpu_topology[cpu].thread_id = i;
 300                         } else if (cpu != -ENODEV) {
 301                                 pr_err("%pOF: Can't get CPU for thread\n", t);
 302                                 of_node_put(t);
 303                                 return -EINVAL;
 304                         }
 305                         of_node_put(t);
 306                 }
 307                 i++;
 308         } while (t);
 309
 310         cpu = get_cpu_for_node(core);
 311         if (cpu >= 0) {
 312                 if (!leaf) {
 313                         pr_err("%pOF: Core has both threads and CPU\n",
 314                                core);
 315                         return -EINVAL;
 316                 }
 317
 318                 cpu_topology[cpu].package_id = package_id;
 319                 cpu_topology[cpu].core_id = core_id;
 320         } else if (leaf && cpu != -ENODEV) {
 321                 pr_err("%pOF: Can't get CPU for leaf core\n", core);
 322                 return -EINVAL;
 323         }
 324
 325         return 0;
 326 }
 327
 328 static int __init parse_cluster(struct device_node *cluster, int depth)
 329 {
 330         char name[10];
 331         bool leaf = true;
 332         bool has_cores = false;
 333         struct device_node *c;
 334         static int package_id __initdata;
 335         int core_id = 0;
 336         int i, ret;
 337
 338         /*
 339          * First check for child clusters; we currently ignore any
 340          * information about the nesting of clusters and present the
 341          * scheduler with a flat list of them.
 342          */
 343         i = 0;
 344         do {
 345                 snprintf(name, sizeof(name), "cluster%d", i);
 346                 c = of_get_child_by_name(cluster, name);
 347                 if (c) {
 348                         leaf = false;
 349                         ret = parse_cluster(c, depth + 1);
 350                         of_node_put(c);
 351                         if (ret != 0)
 352                                 return ret;
 353                 }
 354                 i++;
 355         } while (c);
 356
 357         /* Now check for cores */
 358         i = 0;
 359         do {
 360                 snprintf(name, sizeof(name), "core%d", i);
 361                 c = of_get_child_by_name(cluster, name);
 362                 if (c) {
 363                         has_cores = true;
 364
 365                         if (depth == 0) {
 366                                 pr_err("%pOF: cpu-map children should be clusters\n",
 367                                        c);
 368                                 of_node_put(c);
 369                                 return -EINVAL;
 370                         }
 371
 372                         if (leaf) {
 373                                 ret = parse_core(c, package_id, core_id++);
 374                         } else {
 375                                 pr_err("%pOF: Non-leaf cluster with core %s\n",
 376                                        cluster, name);
 377                                 ret = -EINVAL;
 378                         }
 379
 380                         of_node_put(c);
 381                         if (ret != 0)
 382                                 return ret;
 383                 }
 384                 i++;
 385         } while (c);
 386
 387         if (leaf && !has_cores)
 388                 pr_warn("%pOF: empty cluster\n", cluster);
 389
 390         if (leaf)
 391                 package_id++;
 392
 393         return 0;
 394 }
 395
 396 static int __init parse_dt_topology(void)
 397 {
 398         struct device_node *cn, *map;
 399         int ret = 0;
 400         int cpu;
 401
 402         cn = of_find_node_by_path("/cpus");
 403         if (!cn) {
 404                 pr_err("No CPU information found in DT\n");
 405                 return 0;
 406         }
 407
 408         /*
 409          * When topology is provided cpu-map is essentially a root
 410          * cluster with restricted subnodes.
 411          */
 412         map = of_get_child_by_name(cn, "cpu-map");
 413         if (!map)
 414                 goto out;
 415
 416         ret = parse_cluster(map, 0);
 417         if (ret != 0)
 418                 goto out_map;
 419
 420         topology_normalize_cpu_scale();
 421
 422         /*
 423          * Check that all cores are in the topology; the SMP code will
 424          * only mark cores described in the DT as possible.
 425          */
 426         for_each_possible_cpu(cpu)
 427                 if (cpu_topology[cpu].package_id == -1)
 428                         ret = -EINVAL;
 429
 430 out_map:
 431         of_node_put(map);
 432 out:
 433         of_node_put(cn);
 434         return ret;
 435 }
 436 #endif
 437
 438 /*
 439  * cpu topology table
 440  */
 441 struct cpu_topology cpu_topology[NR_CPUS];
 442 EXPORT_SYMBOL_GPL(cpu_topology);
 443
 444 const struct cpumask *cpu_coregroup_mask(int cpu)
 445 {
 446         const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
 447
 448         /* Find the smaller of NUMA, core or LLC siblings */
 449         if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
 450                 /* not numa in package, lets use the package siblings */
 451                 core_mask = &cpu_topology[cpu].core_sibling;
 452         }
 453         if (cpu_topology[cpu].llc_id != -1) {
 454                 if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
 455                         core_mask = &cpu_topology[cpu].llc_sibling;
 456         }
 457
 458         return core_mask;
 459 }
 460
 461 void update_siblings_masks(unsigned int cpuid)
 462 {
 463         struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
 464         int cpu;
 465
 466         /* update core and thread sibling masks */
 467         for_each_online_cpu(cpu) {
 468                 cpu_topo = &cpu_topology[cpu];
 469
 470                 if (cpuid_topo->llc_id == cpu_topo->llc_id) {
 471                         cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling);
 472                         cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling);
 473                 }
 474
 475                 if (cpuid_topo->package_id != cpu_topo->package_id)
 476                         continue;
 477
 478                 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
 479                 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
 480
 481                 if (cpuid_topo->core_id != cpu_topo->core_id)
 482                         continue;
 483
 484                 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
 485                 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
 486         }
 487 }
 488
 489 static void clear_cpu_topology(int cpu)
 490 {
 491         struct cpu_topology *cpu_topo = &cpu_topology[cpu];
 492
 493         cpumask_clear(&cpu_topo->llc_sibling);
 494         cpumask_set_cpu(cpu, &cpu_topo->llc_sibling);
 495
 496         cpumask_clear(&cpu_topo->core_sibling);
 497         cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
 498         cpumask_clear(&cpu_topo->thread_sibling);
 499         cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
 500 }
 501
 502 void __init reset_cpu_topology(void)
 503 {
 504         unsigned int cpu;
 505
 506         for_each_possible_cpu(cpu) {
 507                 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
 508
 509                 cpu_topo->thread_id = -1;
 510                 cpu_topo->core_id = -1;
 511                 cpu_topo->package_id = -1;
 512                 cpu_topo->llc_id = -1;
 513
 514                 clear_cpu_topology(cpu);
 515         }
 516 }
 517
 518 void remove_cpu_topology(unsigned int cpu)
 519 {
 520         int sibling;
 521
 522         for_each_cpu(sibling, topology_core_cpumask(cpu))
 523                 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
 524         for_each_cpu(sibling, topology_sibling_cpumask(cpu))
 525                 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
 526         for_each_cpu(sibling, topology_llc_cpumask(cpu))
 527                 cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling));
 528
 529         clear_cpu_topology(cpu);
 530 }
 531
 532 __weak int __init parse_acpi_topology(void)
 533 {
 534         return 0;
 535 }
 536
 537 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
 538 void __init init_cpu_topology(void)
 539 {
 540         reset_cpu_topology();
 541
 542         /*
 543          * Discard anything that was parsed if we hit an error so we
 544          * don't use partial information.
 545          */
 546         if (parse_acpi_topology())
 547                 reset_cpu_topology();
 548         else if (of_have_populated_dt() && parse_dt_topology())
 549                 reset_cpu_topology();
 550 }
 551 #endif