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 static DEFINE_PER_CPU(struct scale_freq_data *, sft_data);
  25 static struct cpumask scale_freq_counters_mask;
  26 static bool scale_freq_invariant;
  27
  28 static bool supports_scale_freq_counters(const struct cpumask *cpus)
  29 {
  30         return cpumask_subset(cpus, &scale_freq_counters_mask);
  31 }
  32
  33 bool topology_scale_freq_invariant(void)
  34 {
  35         return cpufreq_supports_freq_invariance() ||
  36                supports_scale_freq_counters(cpu_online_mask);
  37 }
  38
  39 static void update_scale_freq_invariant(bool status)
  40 {
  41         if (scale_freq_invariant == status)
  42                 return;
  43
  44         /*
  45          * Task scheduler behavior depends on frequency invariance support,
  46          * either cpufreq or counter driven. If the support status changes as
  47          * a result of counter initialisation and use, retrigger the build of
  48          * scheduling domains to ensure the information is propagated properly.
  49          */
  50         if (topology_scale_freq_invariant() == status) {
  51                 scale_freq_invariant = status;
  52                 rebuild_sched_domains_energy();
  53         }
  54 }
  55
  56 void topology_set_scale_freq_source(struct scale_freq_data *data,
  57                                     const struct cpumask *cpus)
  58 {
  59         struct scale_freq_data *sfd;
  60         int cpu;
  61
  62         /*
  63          * Avoid calling rebuild_sched_domains() unnecessarily if FIE is
  64          * supported by cpufreq.
  65          */
  66         if (cpumask_empty(&scale_freq_counters_mask))
  67                 scale_freq_invariant = topology_scale_freq_invariant();
  68
  69         for_each_cpu(cpu, cpus) {
  70                 sfd = per_cpu(sft_data, cpu);
  71
  72                 /* Use ARCH provided counters whenever possible */
  73                 if (!sfd || sfd->source != SCALE_FREQ_SOURCE_ARCH) {
  74                         per_cpu(sft_data, cpu) = data;
  75                         cpumask_set_cpu(cpu, &scale_freq_counters_mask);
  76                 }
  77         }
  78
  79         update_scale_freq_invariant(true);
  80 }
  81 EXPORT_SYMBOL_GPL(topology_set_scale_freq_source);
  82
  83 void topology_clear_scale_freq_source(enum scale_freq_source source,
  84                                       const struct cpumask *cpus)
  85 {
  86         struct scale_freq_data *sfd;
  87         int cpu;
  88
  89         for_each_cpu(cpu, cpus) {
  90                 sfd = per_cpu(sft_data, cpu);
  91
  92                 if (sfd && sfd->source == source) {
  93                         per_cpu(sft_data, cpu) = NULL;
  94                         cpumask_clear_cpu(cpu, &scale_freq_counters_mask);
  95                 }
  96         }
  97
  98         update_scale_freq_invariant(false);
  99 }
 100 EXPORT_SYMBOL_GPL(topology_clear_scale_freq_source);
 101
 102 void topology_scale_freq_tick(void)
 103 {
 104         struct scale_freq_data *sfd = *this_cpu_ptr(&sft_data);
 105
 106         if (sfd)
 107                 sfd->set_freq_scale();
 108 }
 109
 110 DEFINE_PER_CPU(unsigned long, arch_freq_scale) = SCHED_CAPACITY_SCALE;
 111 EXPORT_PER_CPU_SYMBOL_GPL(arch_freq_scale);
 112
 113 void topology_set_freq_scale(const struct cpumask *cpus, unsigned long cur_freq,
 114                              unsigned long max_freq)
 115 {
 116         unsigned long scale;
 117         int i;
 118
 119         if (WARN_ON_ONCE(!cur_freq || !max_freq))
 120                 return;
 121
 122         /*
 123          * If the use of counters for FIE is enabled, just return as we don't
 124          * want to update the scale factor with information from CPUFREQ.
 125          * Instead the scale factor will be updated from arch_scale_freq_tick.
 126          */
 127         if (supports_scale_freq_counters(cpus))
 128                 return;
 129
 130         scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq;
 131
 132         for_each_cpu(i, cpus)
 133                 per_cpu(arch_freq_scale, i) = scale;
 134 }
 135
 136 DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
 137
 138 void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
 139 {
 140         per_cpu(cpu_scale, cpu) = capacity;
 141 }
 142
 143 DEFINE_PER_CPU(unsigned long, thermal_pressure);
 144
 145 void topology_set_thermal_pressure(const struct cpumask *cpus,
 146                                unsigned long th_pressure)
 147 {
 148         int cpu;
 149
 150         for_each_cpu(cpu, cpus)
 151                 WRITE_ONCE(per_cpu(thermal_pressure, cpu), th_pressure);
 152 }
 153
 154 static ssize_t cpu_capacity_show(struct device *dev,
 155                                  struct device_attribute *attr,
 156                                  char *buf)
 157 {
 158         struct cpu *cpu = container_of(dev, struct cpu, dev);
 159
 160         return sysfs_emit(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id));
 161 }
 162
 163 static void update_topology_flags_workfn(struct work_struct *work);
 164 static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
 165
 166 static DEVICE_ATTR_RO(cpu_capacity);
 167
 168 static int register_cpu_capacity_sysctl(void)
 169 {
 170         int i;
 171         struct device *cpu;
 172
 173         for_each_possible_cpu(i) {
 174                 cpu = get_cpu_device(i);
 175                 if (!cpu) {
 176                         pr_err("%s: too early to get CPU%d device!\n",
 177                                __func__, i);
 178                         continue;
 179                 }
 180                 device_create_file(cpu, &dev_attr_cpu_capacity);
 181         }
 182
 183         return 0;
 184 }
 185 subsys_initcall(register_cpu_capacity_sysctl);
 186
 187 static int update_topology;
 188
 189 int topology_update_cpu_topology(void)
 190 {
 191         return update_topology;
 192 }
 193
 194 /*
 195  * Updating the sched_domains can't be done directly from cpufreq callbacks
 196  * due to locking, so queue the work for later.
 197  */
 198 static void update_topology_flags_workfn(struct work_struct *work)
 199 {
 200         update_topology = 1;
 201         rebuild_sched_domains();
 202         pr_debug("sched_domain hierarchy rebuilt, flags updated\n");
 203         update_topology = 0;
 204 }
 205
 206 static DEFINE_PER_CPU(u32, freq_factor) = 1;
 207 static u32 *raw_capacity;
 208
 209 static int free_raw_capacity(void)
 210 {
 211         kfree(raw_capacity);
 212         raw_capacity = NULL;
 213
 214         return 0;
 215 }
 216
 217 void topology_normalize_cpu_scale(void)
 218 {
 219         u64 capacity;
 220         u64 capacity_scale;
 221         int cpu;
 222
 223         if (!raw_capacity)
 224                 return;
 225
 226         capacity_scale = 1;
 227         for_each_possible_cpu(cpu) {
 228                 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
 229                 capacity_scale = max(capacity, capacity_scale);
 230         }
 231
 232         pr_debug("cpu_capacity: capacity_scale=%llu\n", capacity_scale);
 233         for_each_possible_cpu(cpu) {
 234                 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
 235                 capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
 236                         capacity_scale);
 237                 topology_set_cpu_scale(cpu, capacity);
 238                 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
 239                         cpu, topology_get_cpu_scale(cpu));
 240         }
 241 }
 242
 243 bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu)
 244 {
 245         struct clk *cpu_clk;
 246         static bool cap_parsing_failed;
 247         int ret;
 248         u32 cpu_capacity;
 249
 250         if (cap_parsing_failed)
 251                 return false;
 252
 253         ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz",
 254                                    &cpu_capacity);
 255         if (!ret) {
 256                 if (!raw_capacity) {
 257                         raw_capacity = kcalloc(num_possible_cpus(),
 258                                                sizeof(*raw_capacity),
 259                                                GFP_KERNEL);
 260                         if (!raw_capacity) {
 261                                 cap_parsing_failed = true;
 262                                 return false;
 263                         }
 264                 }
 265                 raw_capacity[cpu] = cpu_capacity;
 266                 pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
 267                         cpu_node, raw_capacity[cpu]);
 268
 269                 /*
 270                  * Update freq_factor for calculating early boot cpu capacities.
 271                  * For non-clk CPU DVFS mechanism, there's no way to get the
 272                  * frequency value now, assuming they are running at the same
 273                  * frequency (by keeping the initial freq_factor value).
 274                  */
 275                 cpu_clk = of_clk_get(cpu_node, 0);
 276                 if (!PTR_ERR_OR_ZERO(cpu_clk)) {
 277                         per_cpu(freq_factor, cpu) =
 278                                 clk_get_rate(cpu_clk) / 1000;
 279                         clk_put(cpu_clk);
 280                 }
 281         } else {
 282                 if (raw_capacity) {
 283                         pr_err("cpu_capacity: missing %pOF raw capacity\n",
 284                                 cpu_node);
 285                         pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
 286                 }
 287                 cap_parsing_failed = true;
 288                 free_raw_capacity();
 289         }
 290
 291         return !ret;
 292 }
 293
 294 #ifdef CONFIG_CPU_FREQ
 295 static cpumask_var_t cpus_to_visit;
 296 static void parsing_done_workfn(struct work_struct *work);
 297 static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
 298
 299 static int
 300 init_cpu_capacity_callback(struct notifier_block *nb,
 301                            unsigned long val,
 302                            void *data)
 303 {
 304         struct cpufreq_policy *policy = data;
 305         int cpu;
 306
 307         if (!raw_capacity)
 308                 return 0;
 309
 310         if (val != CPUFREQ_CREATE_POLICY)
 311                 return 0;
 312
 313         pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
 314                  cpumask_pr_args(policy->related_cpus),
 315                  cpumask_pr_args(cpus_to_visit));
 316
 317         cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
 318
 319         for_each_cpu(cpu, policy->related_cpus)
 320                 per_cpu(freq_factor, cpu) = policy->cpuinfo.max_freq / 1000;
 321
 322         if (cpumask_empty(cpus_to_visit)) {
 323                 topology_normalize_cpu_scale();
 324                 schedule_work(&update_topology_flags_work);
 325                 free_raw_capacity();
 326                 pr_debug("cpu_capacity: parsing done\n");
 327                 schedule_work(&parsing_done_work);
 328         }
 329
 330         return 0;
 331 }
 332
 333 static struct notifier_block init_cpu_capacity_notifier = {
 334         .notifier_call = init_cpu_capacity_callback,
 335 };
 336
 337 static int __init register_cpufreq_notifier(void)
 338 {
 339         int ret;
 340
 341         /*
 342          * on ACPI-based systems we need to use the default cpu capacity
 343          * until we have the necessary code to parse the cpu capacity, so
 344          * skip registering cpufreq notifier.
 345          */
 346         if (!acpi_disabled || !raw_capacity)
 347                 return -EINVAL;
 348
 349         if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL))
 350                 return -ENOMEM;
 351
 352         cpumask_copy(cpus_to_visit, cpu_possible_mask);
 353
 354         ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
 355                                         CPUFREQ_POLICY_NOTIFIER);
 356
 357         if (ret)
 358                 free_cpumask_var(cpus_to_visit);
 359
 360         return ret;
 361 }
 362 core_initcall(register_cpufreq_notifier);
 363
 364 static void parsing_done_workfn(struct work_struct *work)
 365 {
 366         cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
 367                                          CPUFREQ_POLICY_NOTIFIER);
 368         free_cpumask_var(cpus_to_visit);
 369 }
 370
 371 #else
 372 core_initcall(free_raw_capacity);
 373 #endif
 374
 375 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
 376 /*
 377  * This function returns the logic cpu number of the node.
 378  * There are basically three kinds of return values:
 379  * (1) logic cpu number which is > 0.
 380  * (2) -ENODEV when the device tree(DT) node is valid and found in the DT but
 381  * there is no possible logical CPU in the kernel to match. This happens
 382  * when CONFIG_NR_CPUS is configure to be smaller than the number of
 383  * CPU nodes in DT. We need to just ignore this case.
 384  * (3) -1 if the node does not exist in the device tree
 385  */
 386 static int __init get_cpu_for_node(struct device_node *node)
 387 {
 388         struct device_node *cpu_node;
 389         int cpu;
 390
 391         cpu_node = of_parse_phandle(node, "cpu", 0);
 392         if (!cpu_node)
 393                 return -1;
 394
 395         cpu = of_cpu_node_to_id(cpu_node);
 396         if (cpu >= 0)
 397                 topology_parse_cpu_capacity(cpu_node, cpu);
 398         else
 399                 pr_info("CPU node for %pOF exist but the possible cpu range is :%*pbl\n",
 400                         cpu_node, cpumask_pr_args(cpu_possible_mask));
 401
 402         of_node_put(cpu_node);
 403         return cpu;
 404 }
 405
 406 static int __init parse_core(struct device_node *core, int package_id,
 407                              int core_id)
 408 {
 409         char name[20];
 410         bool leaf = true;
 411         int i = 0;
 412         int cpu;
 413         struct device_node *t;
 414
 415         do {
 416                 snprintf(name, sizeof(name), "thread%d", i);
 417                 t = of_get_child_by_name(core, name);
 418                 if (t) {
 419                         leaf = false;
 420                         cpu = get_cpu_for_node(t);
 421                         if (cpu >= 0) {
 422                                 cpu_topology[cpu].package_id = package_id;
 423                                 cpu_topology[cpu].core_id = core_id;
 424                                 cpu_topology[cpu].thread_id = i;
 425                         } else if (cpu != -ENODEV) {
 426                                 pr_err("%pOF: Can't get CPU for thread\n", t);
 427                                 of_node_put(t);
 428                                 return -EINVAL;
 429                         }
 430                         of_node_put(t);
 431                 }
 432                 i++;
 433         } while (t);
 434
 435         cpu = get_cpu_for_node(core);
 436         if (cpu >= 0) {
 437                 if (!leaf) {
 438                         pr_err("%pOF: Core has both threads and CPU\n",
 439                                core);
 440                         return -EINVAL;
 441                 }
 442
 443                 cpu_topology[cpu].package_id = package_id;
 444                 cpu_topology[cpu].core_id = core_id;
 445         } else if (leaf && cpu != -ENODEV) {
 446                 pr_err("%pOF: Can't get CPU for leaf core\n", core);
 447                 return -EINVAL;
 448         }
 449
 450         return 0;
 451 }
 452
 453 static int __init parse_cluster(struct device_node *cluster, int depth)
 454 {
 455         char name[20];
 456         bool leaf = true;
 457         bool has_cores = false;
 458         struct device_node *c;
 459         static int package_id __initdata;
 460         int core_id = 0;
 461         int i, ret;
 462
 463         /*
 464          * First check for child clusters; we currently ignore any
 465          * information about the nesting of clusters and present the
 466          * scheduler with a flat list of them.
 467          */
 468         i = 0;
 469         do {
 470                 snprintf(name, sizeof(name), "cluster%d", i);
 471                 c = of_get_child_by_name(cluster, name);
 472                 if (c) {
 473                         leaf = false;
 474                         ret = parse_cluster(c, depth + 1);
 475                         of_node_put(c);
 476                         if (ret != 0)
 477                                 return ret;
 478                 }
 479                 i++;
 480         } while (c);
 481
 482         /* Now check for cores */
 483         i = 0;
 484         do {
 485                 snprintf(name, sizeof(name), "core%d", i);
 486                 c = of_get_child_by_name(cluster, name);
 487                 if (c) {
 488                         has_cores = true;
 489
 490                         if (depth == 0) {
 491                                 pr_err("%pOF: cpu-map children should be clusters\n",
 492                                        c);
 493                                 of_node_put(c);
 494                                 return -EINVAL;
 495                         }
 496
 497                         if (leaf) {
 498                                 ret = parse_core(c, package_id, core_id++);
 499                         } else {
 500                                 pr_err("%pOF: Non-leaf cluster with core %s\n",
 501                                        cluster, name);
 502                                 ret = -EINVAL;
 503                         }
 504
 505                         of_node_put(c);
 506                         if (ret != 0)
 507                                 return ret;
 508                 }
 509                 i++;
 510         } while (c);
 511
 512         if (leaf && !has_cores)
 513                 pr_warn("%pOF: empty cluster\n", cluster);
 514
 515         if (leaf)
 516                 package_id++;
 517
 518         return 0;
 519 }
 520
 521 static int __init parse_dt_topology(void)
 522 {
 523         struct device_node *cn, *map;
 524         int ret = 0;
 525         int cpu;
 526
 527         cn = of_find_node_by_path("/cpus");
 528         if (!cn) {
 529                 pr_err("No CPU information found in DT\n");
 530                 return 0;
 531         }
 532
 533         /*
 534          * When topology is provided cpu-map is essentially a root
 535          * cluster with restricted subnodes.
 536          */
 537         map = of_get_child_by_name(cn, "cpu-map");
 538         if (!map)
 539                 goto out;
 540
 541         ret = parse_cluster(map, 0);
 542         if (ret != 0)
 543                 goto out_map;
 544
 545         topology_normalize_cpu_scale();
 546
 547         /*
 548          * Check that all cores are in the topology; the SMP code will
 549          * only mark cores described in the DT as possible.
 550          */
 551         for_each_possible_cpu(cpu)
 552                 if (cpu_topology[cpu].package_id == -1)
 553                         ret = -EINVAL;
 554
 555 out_map:
 556         of_node_put(map);
 557 out:
 558         of_node_put(cn);
 559         return ret;
 560 }
 561 #endif
 562
 563 /*
 564  * cpu topology table
 565  */
 566 struct cpu_topology cpu_topology[NR_CPUS];
 567 EXPORT_SYMBOL_GPL(cpu_topology);
 568
 569 const struct cpumask *cpu_coregroup_mask(int cpu)
 570 {
 571         const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
 572
 573         /* Find the smaller of NUMA, core or LLC siblings */
 574         if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
 575                 /* not numa in package, lets use the package siblings */
 576                 core_mask = &cpu_topology[cpu].core_sibling;
 577         }
 578         if (cpu_topology[cpu].llc_id != -1) {
 579                 if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
 580                         core_mask = &cpu_topology[cpu].llc_sibling;
 581         }
 582
 583         return core_mask;
 584 }
 585
 586 void update_siblings_masks(unsigned int cpuid)
 587 {
 588         struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
 589         int cpu;
 590
 591         /* update core and thread sibling masks */
 592         for_each_online_cpu(cpu) {
 593                 cpu_topo = &cpu_topology[cpu];
 594
 595                 if (cpuid_topo->llc_id == cpu_topo->llc_id) {
 596                         cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling);
 597                         cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling);
 598                 }
 599
 600                 if (cpuid_topo->package_id != cpu_topo->package_id)
 601                         continue;
 602
 603                 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
 604                 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
 605
 606                 if (cpuid_topo->core_id != cpu_topo->core_id)
 607                         continue;
 608
 609                 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
 610                 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
 611         }
 612 }
 613
 614 static void clear_cpu_topology(int cpu)
 615 {
 616         struct cpu_topology *cpu_topo = &cpu_topology[cpu];
 617
 618         cpumask_clear(&cpu_topo->llc_sibling);
 619         cpumask_set_cpu(cpu, &cpu_topo->llc_sibling);
 620
 621         cpumask_clear(&cpu_topo->core_sibling);
 622         cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
 623         cpumask_clear(&cpu_topo->thread_sibling);
 624         cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
 625 }
 626
 627 void __init reset_cpu_topology(void)
 628 {
 629         unsigned int cpu;
 630
 631         for_each_possible_cpu(cpu) {
 632                 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
 633
 634                 cpu_topo->thread_id = -1;
 635                 cpu_topo->core_id = -1;
 636                 cpu_topo->package_id = -1;
 637                 cpu_topo->llc_id = -1;
 638
 639                 clear_cpu_topology(cpu);
 640         }
 641 }
 642
 643 void remove_cpu_topology(unsigned int cpu)
 644 {
 645         int sibling;
 646
 647         for_each_cpu(sibling, topology_core_cpumask(cpu))
 648                 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
 649         for_each_cpu(sibling, topology_sibling_cpumask(cpu))
 650                 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
 651         for_each_cpu(sibling, topology_llc_cpumask(cpu))
 652                 cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling));
 653
 654         clear_cpu_topology(cpu);
 655 }
 656
 657 __weak int __init parse_acpi_topology(void)
 658 {
 659         return 0;
 660 }
 661
 662 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
 663 void __init init_cpu_topology(void)
 664 {
 665         reset_cpu_topology();
 666
 667         /*
 668          * Discard anything that was parsed if we hit an error so we
 669          * don't use partial information.
 670          */
 671         if (parse_acpi_topology())
 672                 reset_cpu_topology();
 673         else if (of_have_populated_dt() && parse_dt_topology())
 674                 reset_cpu_topology();
 675 }
 676 #endif