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/rcupdate.h>
  22 #include <linux/sched.h>
  23 #include <linux/smp.h>
  24
  25 static DEFINE_PER_CPU(struct scale_freq_data __rcu *, sft_data);
  26 static struct cpumask scale_freq_counters_mask;
  27 static bool scale_freq_invariant;
  28
  29 static bool supports_scale_freq_counters(const struct cpumask *cpus)
  30 {
  31         return cpumask_subset(cpus, &scale_freq_counters_mask);
  32 }
  33
  34 bool topology_scale_freq_invariant(void)
  35 {
  36         return cpufreq_supports_freq_invariance() ||
  37                supports_scale_freq_counters(cpu_online_mask);
  38 }
  39
  40 static void update_scale_freq_invariant(bool status)
  41 {
  42         if (scale_freq_invariant == status)
  43                 return;
  44
  45         /*
  46          * Task scheduler behavior depends on frequency invariance support,
  47          * either cpufreq or counter driven. If the support status changes as
  48          * a result of counter initialisation and use, retrigger the build of
  49          * scheduling domains to ensure the information is propagated properly.
  50          */
  51         if (topology_scale_freq_invariant() == status) {
  52                 scale_freq_invariant = status;
  53                 rebuild_sched_domains_energy();
  54         }
  55 }
  56
  57 void topology_set_scale_freq_source(struct scale_freq_data *data,
  58                                     const struct cpumask *cpus)
  59 {
  60         struct scale_freq_data *sfd;
  61         int cpu;
  62
  63         /*
  64          * Avoid calling rebuild_sched_domains() unnecessarily if FIE is
  65          * supported by cpufreq.
  66          */
  67         if (cpumask_empty(&scale_freq_counters_mask))
  68                 scale_freq_invariant = topology_scale_freq_invariant();
  69
  70         rcu_read_lock();
  71
  72         for_each_cpu(cpu, cpus) {
  73                 sfd = rcu_dereference(*per_cpu_ptr(&sft_data, cpu));
  74
  75                 /* Use ARCH provided counters whenever possible */
  76                 if (!sfd || sfd->source != SCALE_FREQ_SOURCE_ARCH) {
  77                         rcu_assign_pointer(per_cpu(sft_data, cpu), data);
  78                         cpumask_set_cpu(cpu, &scale_freq_counters_mask);
  79                 }
  80         }
  81
  82         rcu_read_unlock();
  83
  84         update_scale_freq_invariant(true);
  85 }
  86 EXPORT_SYMBOL_GPL(topology_set_scale_freq_source);
  87
  88 void topology_clear_scale_freq_source(enum scale_freq_source source,
  89                                       const struct cpumask *cpus)
  90 {
  91         struct scale_freq_data *sfd;
  92         int cpu;
  93
  94         rcu_read_lock();
  95
  96         for_each_cpu(cpu, cpus) {
  97                 sfd = rcu_dereference(*per_cpu_ptr(&sft_data, cpu));
  98
  99                 if (sfd && sfd->source == source) {
 100                         rcu_assign_pointer(per_cpu(sft_data, cpu), NULL);
 101                         cpumask_clear_cpu(cpu, &scale_freq_counters_mask);
 102                 }
 103         }
 104
 105         rcu_read_unlock();
 106
 107         /*
 108          * Make sure all references to previous sft_data are dropped to avoid
 109          * use-after-free races.
 110          */
 111         synchronize_rcu();
 112
 113         update_scale_freq_invariant(false);
 114 }
 115 EXPORT_SYMBOL_GPL(topology_clear_scale_freq_source);
 116
 117 void topology_scale_freq_tick(void)
 118 {
 119         struct scale_freq_data *sfd = rcu_dereference_sched(*this_cpu_ptr(&sft_data));
 120
 121         if (sfd)
 122                 sfd->set_freq_scale();
 123 }
 124
 125 DEFINE_PER_CPU(unsigned long, arch_freq_scale) = SCHED_CAPACITY_SCALE;
 126 EXPORT_PER_CPU_SYMBOL_GPL(arch_freq_scale);
 127
 128 void topology_set_freq_scale(const struct cpumask *cpus, unsigned long cur_freq,
 129                              unsigned long max_freq)
 130 {
 131         unsigned long scale;
 132         int i;
 133
 134         if (WARN_ON_ONCE(!cur_freq || !max_freq))
 135                 return;
 136
 137         /*
 138          * If the use of counters for FIE is enabled, just return as we don't
 139          * want to update the scale factor with information from CPUFREQ.
 140          * Instead the scale factor will be updated from arch_scale_freq_tick.
 141          */
 142         if (supports_scale_freq_counters(cpus))
 143                 return;
 144
 145         scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq;
 146
 147         for_each_cpu(i, cpus)
 148                 per_cpu(arch_freq_scale, i) = scale;
 149 }
 150
 151 DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
 152
 153 void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
 154 {
 155         per_cpu(cpu_scale, cpu) = capacity;
 156 }
 157
 158 DEFINE_PER_CPU(unsigned long, thermal_pressure);
 159
 160 void topology_set_thermal_pressure(const struct cpumask *cpus,
 161                                unsigned long th_pressure)
 162 {
 163         int cpu;
 164
 165         for_each_cpu(cpu, cpus)
 166                 WRITE_ONCE(per_cpu(thermal_pressure, cpu), th_pressure);
 167 }
 168
 169 static ssize_t cpu_capacity_show(struct device *dev,
 170                                  struct device_attribute *attr,
 171                                  char *buf)
 172 {
 173         struct cpu *cpu = container_of(dev, struct cpu, dev);
 174
 175         return sysfs_emit(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id));
 176 }
 177
 178 static void update_topology_flags_workfn(struct work_struct *work);
 179 static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
 180
 181 static DEVICE_ATTR_RO(cpu_capacity);
 182
 183 static int register_cpu_capacity_sysctl(void)
 184 {
 185         int i;
 186         struct device *cpu;
 187
 188         for_each_possible_cpu(i) {
 189                 cpu = get_cpu_device(i);
 190                 if (!cpu) {
 191                         pr_err("%s: too early to get CPU%d device!\n",
 192                                __func__, i);
 193                         continue;
 194                 }
 195                 device_create_file(cpu, &dev_attr_cpu_capacity);
 196         }
 197
 198         return 0;
 199 }
 200 subsys_initcall(register_cpu_capacity_sysctl);
 201
 202 static int update_topology;
 203
 204 int topology_update_cpu_topology(void)
 205 {
 206         return update_topology;
 207 }
 208
 209 /*
 210  * Updating the sched_domains can't be done directly from cpufreq callbacks
 211  * due to locking, so queue the work for later.
 212  */
 213 static void update_topology_flags_workfn(struct work_struct *work)
 214 {
 215         update_topology = 1;
 216         rebuild_sched_domains();
 217         pr_debug("sched_domain hierarchy rebuilt, flags updated\n");
 218         update_topology = 0;
 219 }
 220
 221 static DEFINE_PER_CPU(u32, freq_factor) = 1;
 222 static u32 *raw_capacity;
 223
 224 static int free_raw_capacity(void)
 225 {
 226         kfree(raw_capacity);
 227         raw_capacity = NULL;
 228
 229         return 0;
 230 }
 231
 232 void topology_normalize_cpu_scale(void)
 233 {
 234         u64 capacity;
 235         u64 capacity_scale;
 236         int cpu;
 237
 238         if (!raw_capacity)
 239                 return;
 240
 241         capacity_scale = 1;
 242         for_each_possible_cpu(cpu) {
 243                 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
 244                 capacity_scale = max(capacity, capacity_scale);
 245         }
 246
 247         pr_debug("cpu_capacity: capacity_scale=%llu\n", capacity_scale);
 248         for_each_possible_cpu(cpu) {
 249                 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
 250                 capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
 251                         capacity_scale);
 252                 topology_set_cpu_scale(cpu, capacity);
 253                 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
 254                         cpu, topology_get_cpu_scale(cpu));
 255         }
 256 }
 257
 258 bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu)
 259 {
 260         struct clk *cpu_clk;
 261         static bool cap_parsing_failed;
 262         int ret;
 263         u32 cpu_capacity;
 264
 265         if (cap_parsing_failed)
 266                 return false;
 267
 268         ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz",
 269                                    &cpu_capacity);
 270         if (!ret) {
 271                 if (!raw_capacity) {
 272                         raw_capacity = kcalloc(num_possible_cpus(),
 273                                                sizeof(*raw_capacity),
 274                                                GFP_KERNEL);
 275                         if (!raw_capacity) {
 276                                 cap_parsing_failed = true;
 277                                 return false;
 278                         }
 279                 }
 280                 raw_capacity[cpu] = cpu_capacity;
 281                 pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
 282                         cpu_node, raw_capacity[cpu]);
 283
 284                 /*
 285                  * Update freq_factor for calculating early boot cpu capacities.
 286                  * For non-clk CPU DVFS mechanism, there's no way to get the
 287                  * frequency value now, assuming they are running at the same
 288                  * frequency (by keeping the initial freq_factor value).
 289                  */
 290                 cpu_clk = of_clk_get(cpu_node, 0);
 291                 if (!PTR_ERR_OR_ZERO(cpu_clk)) {
 292                         per_cpu(freq_factor, cpu) =
 293                                 clk_get_rate(cpu_clk) / 1000;
 294                         clk_put(cpu_clk);
 295                 }
 296         } else {
 297                 if (raw_capacity) {
 298                         pr_err("cpu_capacity: missing %pOF raw capacity\n",
 299                                 cpu_node);
 300                         pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
 301                 }
 302                 cap_parsing_failed = true;
 303                 free_raw_capacity();
 304         }
 305
 306         return !ret;
 307 }
 308
 309 #ifdef CONFIG_CPU_FREQ
 310 static cpumask_var_t cpus_to_visit;
 311 static void parsing_done_workfn(struct work_struct *work);
 312 static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
 313
 314 static int
 315 init_cpu_capacity_callback(struct notifier_block *nb,
 316                            unsigned long val,
 317                            void *data)
 318 {
 319         struct cpufreq_policy *policy = data;
 320         int cpu;
 321
 322         if (!raw_capacity)
 323                 return 0;
 324
 325         if (val != CPUFREQ_CREATE_POLICY)
 326                 return 0;
 327
 328         pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
 329                  cpumask_pr_args(policy->related_cpus),
 330                  cpumask_pr_args(cpus_to_visit));
 331
 332         cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
 333
 334         for_each_cpu(cpu, policy->related_cpus)
 335                 per_cpu(freq_factor, cpu) = policy->cpuinfo.max_freq / 1000;
 336
 337         if (cpumask_empty(cpus_to_visit)) {
 338                 topology_normalize_cpu_scale();
 339                 schedule_work(&update_topology_flags_work);
 340                 free_raw_capacity();
 341                 pr_debug("cpu_capacity: parsing done\n");
 342                 schedule_work(&parsing_done_work);
 343         }
 344
 345         return 0;
 346 }
 347
 348 static struct notifier_block init_cpu_capacity_notifier = {
 349         .notifier_call = init_cpu_capacity_callback,
 350 };
 351
 352 static int __init register_cpufreq_notifier(void)
 353 {
 354         int ret;
 355
 356         /*
 357          * on ACPI-based systems we need to use the default cpu capacity
 358          * until we have the necessary code to parse the cpu capacity, so
 359          * skip registering cpufreq notifier.
 360          */
 361         if (!acpi_disabled || !raw_capacity)
 362                 return -EINVAL;
 363
 364         if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL))
 365                 return -ENOMEM;
 366
 367         cpumask_copy(cpus_to_visit, cpu_possible_mask);
 368
 369         ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
 370                                         CPUFREQ_POLICY_NOTIFIER);
 371
 372         if (ret)
 373                 free_cpumask_var(cpus_to_visit);
 374
 375         return ret;
 376 }
 377 core_initcall(register_cpufreq_notifier);
 378
 379 static void parsing_done_workfn(struct work_struct *work)
 380 {
 381         cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
 382                                          CPUFREQ_POLICY_NOTIFIER);
 383         free_cpumask_var(cpus_to_visit);
 384 }
 385
 386 #else
 387 core_initcall(free_raw_capacity);
 388 #endif
 389
 390 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
 391 /*
 392  * This function returns the logic cpu number of the node.
 393  * There are basically three kinds of return values:
 394  * (1) logic cpu number which is > 0.
 395  * (2) -ENODEV when the device tree(DT) node is valid and found in the DT but
 396  * there is no possible logical CPU in the kernel to match. This happens
 397  * when CONFIG_NR_CPUS is configure to be smaller than the number of
 398  * CPU nodes in DT. We need to just ignore this case.
 399  * (3) -1 if the node does not exist in the device tree
 400  */
 401 static int __init get_cpu_for_node(struct device_node *node)
 402 {
 403         struct device_node *cpu_node;
 404         int cpu;
 405
 406         cpu_node = of_parse_phandle(node, "cpu", 0);
 407         if (!cpu_node)
 408                 return -1;
 409
 410         cpu = of_cpu_node_to_id(cpu_node);
 411         if (cpu >= 0)
 412                 topology_parse_cpu_capacity(cpu_node, cpu);
 413         else
 414                 pr_info("CPU node for %pOF exist but the possible cpu range is :%*pbl\n",
 415                         cpu_node, cpumask_pr_args(cpu_possible_mask));
 416
 417         of_node_put(cpu_node);
 418         return cpu;
 419 }
 420
 421 static int __init parse_core(struct device_node *core, int package_id,
 422                              int core_id)
 423 {
 424         char name[20];
 425         bool leaf = true;
 426         int i = 0;
 427         int cpu;
 428         struct device_node *t;
 429
 430         do {
 431                 snprintf(name, sizeof(name), "thread%d", i);
 432                 t = of_get_child_by_name(core, name);
 433                 if (t) {
 434                         leaf = false;
 435                         cpu = get_cpu_for_node(t);
 436                         if (cpu >= 0) {
 437                                 cpu_topology[cpu].package_id = package_id;
 438                                 cpu_topology[cpu].core_id = core_id;
 439                                 cpu_topology[cpu].thread_id = i;
 440                         } else if (cpu != -ENODEV) {
 441                                 pr_err("%pOF: Can't get CPU for thread\n", t);
 442                                 of_node_put(t);
 443                                 return -EINVAL;
 444                         }
 445                         of_node_put(t);
 446                 }
 447                 i++;
 448         } while (t);
 449
 450         cpu = get_cpu_for_node(core);
 451         if (cpu >= 0) {
 452                 if (!leaf) {
 453                         pr_err("%pOF: Core has both threads and CPU\n",
 454                                core);
 455                         return -EINVAL;
 456                 }
 457
 458                 cpu_topology[cpu].package_id = package_id;
 459                 cpu_topology[cpu].core_id = core_id;
 460         } else if (leaf && cpu != -ENODEV) {
 461                 pr_err("%pOF: Can't get CPU for leaf core\n", core);
 462                 return -EINVAL;
 463         }
 464
 465         return 0;
 466 }
 467
 468 static int __init parse_cluster(struct device_node *cluster, int depth)
 469 {
 470         char name[20];
 471         bool leaf = true;
 472         bool has_cores = false;
 473         struct device_node *c;
 474         static int package_id __initdata;
 475         int core_id = 0;
 476         int i, ret;
 477
 478         /*
 479          * First check for child clusters; we currently ignore any
 480          * information about the nesting of clusters and present the
 481          * scheduler with a flat list of them.
 482          */
 483         i = 0;
 484         do {
 485                 snprintf(name, sizeof(name), "cluster%d", i);
 486                 c = of_get_child_by_name(cluster, name);
 487                 if (c) {
 488                         leaf = false;
 489                         ret = parse_cluster(c, depth + 1);
 490                         of_node_put(c);
 491                         if (ret != 0)
 492                                 return ret;
 493                 }
 494                 i++;
 495         } while (c);
 496
 497         /* Now check for cores */
 498         i = 0;
 499         do {
 500                 snprintf(name, sizeof(name), "core%d", i);
 501                 c = of_get_child_by_name(cluster, name);
 502                 if (c) {
 503                         has_cores = true;
 504
 505                         if (depth == 0) {
 506                                 pr_err("%pOF: cpu-map children should be clusters\n",
 507                                        c);
 508                                 of_node_put(c);
 509                                 return -EINVAL;
 510                         }
 511
 512                         if (leaf) {
 513                                 ret = parse_core(c, package_id, core_id++);
 514                         } else {
 515                                 pr_err("%pOF: Non-leaf cluster with core %s\n",
 516                                        cluster, name);
 517                                 ret = -EINVAL;
 518                         }
 519
 520                         of_node_put(c);
 521                         if (ret != 0)
 522                                 return ret;
 523                 }
 524                 i++;
 525         } while (c);
 526
 527         if (leaf && !has_cores)
 528                 pr_warn("%pOF: empty cluster\n", cluster);
 529
 530         if (leaf)
 531                 package_id++;
 532
 533         return 0;
 534 }
 535
 536 static int __init parse_dt_topology(void)
 537 {
 538         struct device_node *cn, *map;
 539         int ret = 0;
 540         int cpu;
 541
 542         cn = of_find_node_by_path("/cpus");
 543         if (!cn) {
 544                 pr_err("No CPU information found in DT\n");
 545                 return 0;
 546         }
 547
 548         /*
 549          * When topology is provided cpu-map is essentially a root
 550          * cluster with restricted subnodes.
 551          */
 552         map = of_get_child_by_name(cn, "cpu-map");
 553         if (!map)
 554                 goto out;
 555
 556         ret = parse_cluster(map, 0);
 557         if (ret != 0)
 558                 goto out_map;
 559
 560         topology_normalize_cpu_scale();
 561
 562         /*
 563          * Check that all cores are in the topology; the SMP code will
 564          * only mark cores described in the DT as possible.
 565          */
 566         for_each_possible_cpu(cpu)
 567                 if (cpu_topology[cpu].package_id == -1)
 568                         ret = -EINVAL;
 569
 570 out_map:
 571         of_node_put(map);
 572 out:
 573         of_node_put(cn);
 574         return ret;
 575 }
 576 #endif
 577
 578 /*
 579  * cpu topology table
 580  */
 581 struct cpu_topology cpu_topology[NR_CPUS];
 582 EXPORT_SYMBOL_GPL(cpu_topology);
 583
 584 const struct cpumask *cpu_coregroup_mask(int cpu)
 585 {
 586         const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
 587
 588         /* Find the smaller of NUMA, core or LLC siblings */
 589         if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
 590                 /* not numa in package, lets use the package siblings */
 591                 core_mask = &cpu_topology[cpu].core_sibling;
 592         }
 593         if (cpu_topology[cpu].llc_id != -1) {
 594                 if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
 595                         core_mask = &cpu_topology[cpu].llc_sibling;
 596         }
 597
 598         return core_mask;
 599 }
 600
 601 void update_siblings_masks(unsigned int cpuid)
 602 {
 603         struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
 604         int cpu;
 605
 606         /* update core and thread sibling masks */
 607         for_each_online_cpu(cpu) {
 608                 cpu_topo = &cpu_topology[cpu];
 609
 610                 if (cpuid_topo->llc_id == cpu_topo->llc_id) {
 611                         cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling);
 612                         cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling);
 613                 }
 614
 615                 if (cpuid_topo->package_id != cpu_topo->package_id)
 616                         continue;
 617
 618                 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
 619                 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
 620
 621                 if (cpuid_topo->core_id != cpu_topo->core_id)
 622                         continue;
 623
 624                 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
 625                 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
 626         }
 627 }
 628
 629 static void clear_cpu_topology(int cpu)
 630 {
 631         struct cpu_topology *cpu_topo = &cpu_topology[cpu];
 632
 633         cpumask_clear(&cpu_topo->llc_sibling);
 634         cpumask_set_cpu(cpu, &cpu_topo->llc_sibling);
 635
 636         cpumask_clear(&cpu_topo->core_sibling);
 637         cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
 638         cpumask_clear(&cpu_topo->thread_sibling);
 639         cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
 640 }
 641
 642 void __init reset_cpu_topology(void)
 643 {
 644         unsigned int cpu;
 645
 646         for_each_possible_cpu(cpu) {
 647                 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
 648
 649                 cpu_topo->thread_id = -1;
 650                 cpu_topo->core_id = -1;
 651                 cpu_topo->package_id = -1;
 652                 cpu_topo->llc_id = -1;
 653
 654                 clear_cpu_topology(cpu);
 655         }
 656 }
 657
 658 void remove_cpu_topology(unsigned int cpu)
 659 {
 660         int sibling;
 661
 662         for_each_cpu(sibling, topology_core_cpumask(cpu))
 663                 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
 664         for_each_cpu(sibling, topology_sibling_cpumask(cpu))
 665                 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
 666         for_each_cpu(sibling, topology_llc_cpumask(cpu))
 667                 cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling));
 668
 669         clear_cpu_topology(cpu);
 670 }
 671
 672 __weak int __init parse_acpi_topology(void)
 673 {
 674         return 0;
 675 }
 676
 677 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
 678 void __init init_cpu_topology(void)
 679 {
 680         reset_cpu_topology();
 681
 682         /*
 683          * Discard anything that was parsed if we hit an error so we
 684          * don't use partial information.
 685          */
 686         if (parse_acpi_topology())
 687                 reset_cpu_topology();
 688         else if (of_have_populated_dt() && parse_dt_topology())
 689                 reset_cpu_topology();
 690 }
 691 #endif