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/cpu.h>
  15 #include <linux/cpumask.h>
  16 #include <linux/export.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
  25 #include <asm/cputype.h>
  26 #include <asm/topology.h>
  27 #include <asm/smp_plat.h>
  28
  29
  30 /*
  31  * cpu power table
  32  * This per cpu data structure describes the relative capacity of each core.
  33  * On a heteregenous system, cores don't have the same computation capacity
  34  * and we reflect that difference in the cpu_power field so the scheduler can
  35  * take this difference into account during load balance. A per cpu structure
  36  * is preferred because each CPU updates its own cpu_power field during the
  37  * load balance except for idle cores. One idle core is selected to run the
  38  * rebalance_domains for all idle cores and the cpu_power can be updated
  39  * during this sequence.
  40  */
  41 static DEFINE_PER_CPU(unsigned long, cpu_scale);
  42
  43 unsigned long arch_scale_freq_power(struct sched_domain *sd, int cpu)
  44 {
  45         return per_cpu(cpu_scale, cpu);
  46 }
  47
  48 static void set_power_scale(unsigned int cpu, unsigned long power)
  49 {
  50         per_cpu(cpu_scale, cpu) = power;
  51 }
  52
  53 static int __init get_cpu_for_node(struct device_node *node)
  54 {
  55         struct device_node *cpu_node;
  56         int cpu;
  57
  58         cpu_node = of_parse_phandle(node, "cpu", 0);
  59         if (!cpu_node)
  60                 return -1;
  61
  62         for_each_possible_cpu(cpu) {
  63                 if (of_get_cpu_node(cpu, NULL) == cpu_node) {
  64                         of_node_put(cpu_node);
  65                         return cpu;
  66                 }
  67         }
  68
  69         pr_crit("Unable to find CPU node for %s\n", cpu_node->full_name);
  70
  71         of_node_put(cpu_node);
  72         return -1;
  73 }
  74
  75 static int __init parse_core(struct device_node *core, int cluster_id,
  76                              int core_id)
  77 {
  78         char name[10];
  79         bool leaf = true;
  80         int i = 0;
  81         int cpu;
  82         struct device_node *t;
  83
  84         do {
  85                 snprintf(name, sizeof(name), "thread%d", i);
  86                 t = of_get_child_by_name(core, name);
  87                 if (t) {
  88                         leaf = false;
  89                         cpu = get_cpu_for_node(t);
  90                         if (cpu >= 0) {
  91                                 cpu_topology[cpu].cluster_id = cluster_id;
  92                                 cpu_topology[cpu].core_id = core_id;
  93                                 cpu_topology[cpu].thread_id = i;
  94                         } else {
  95                                 pr_err("%s: Can't get CPU for thread\n",
  96                                        t->full_name);
  97                                 of_node_put(t);
  98                                 return -EINVAL;
  99                         }
 100                         of_node_put(t);
 101                 }
 102                 i++;
 103         } while (t);
 104
 105         cpu = get_cpu_for_node(core);
 106         if (cpu >= 0) {
 107                 if (!leaf) {
 108                         pr_err("%s: Core has both threads and CPU\n",
 109                                core->full_name);
 110                         return -EINVAL;
 111                 }
 112
 113                 cpu_topology[cpu].cluster_id = cluster_id;
 114                 cpu_topology[cpu].core_id = core_id;
 115         } else if (leaf) {
 116                 pr_err("%s: Can't get CPU for leaf core\n", core->full_name);
 117                 return -EINVAL;
 118         }
 119
 120         return 0;
 121 }
 122
 123 static int __init parse_cluster(struct device_node *cluster, int depth)
 124 {
 125         char name[10];
 126         bool leaf = true;
 127         bool has_cores = false;
 128         struct device_node *c;
 129         static int cluster_id __initdata;
 130         int core_id = 0;
 131         int i, ret;
 132
 133         /*
 134          * First check for child clusters; we currently ignore any
 135          * information about the nesting of clusters and present the
 136          * scheduler with a flat list of them.
 137          */
 138         i = 0;
 139         do {
 140                 snprintf(name, sizeof(name), "cluster%d", i);
 141                 c = of_get_child_by_name(cluster, name);
 142                 if (c) {
 143                         leaf = false;
 144                         ret = parse_cluster(c, depth + 1);
 145                         of_node_put(c);
 146                         if (ret != 0)
 147                                 return ret;
 148                 }
 149                 i++;
 150         } while (c);
 151
 152         /* Now check for cores */
 153         i = 0;
 154         do {
 155                 snprintf(name, sizeof(name), "core%d", i);
 156                 c = of_get_child_by_name(cluster, name);
 157                 if (c) {
 158                         has_cores = true;
 159
 160                         if (depth == 0) {
 161                                 pr_err("%s: cpu-map children should be clusters\n",
 162                                        c->full_name);
 163                                 of_node_put(c);
 164                                 return -EINVAL;
 165                         }
 166
 167                         if (leaf) {
 168                                 ret = parse_core(c, cluster_id, core_id++);
 169                         } else {
 170                                 pr_err("%s: Non-leaf cluster with core %s\n",
 171                                        cluster->full_name, name);
 172                                 ret = -EINVAL;
 173                         }
 174
 175                         of_node_put(c);
 176                         if (ret != 0)
 177                                 return ret;
 178                 }
 179                 i++;
 180         } while (c);
 181
 182         if (leaf && !has_cores)
 183                 pr_warn("%s: empty cluster\n", cluster->full_name);
 184
 185         if (leaf)
 186                 cluster_id++;
 187
 188         return 0;
 189 }
 190
 191 struct cpu_efficiency {
 192         const char *compatible;
 193         unsigned long efficiency;
 194 };
 195
 196 /*
 197  * Table of relative efficiency of each processors
 198  * The efficiency value must fit in 20bit and the final
 199  * cpu_scale value must be in the range
 200  *   0 < cpu_scale < 3*SCHED_POWER_SCALE/2
 201  * in order to return at most 1 when DIV_ROUND_CLOSEST
 202  * is used to compute the capacity of a CPU.
 203  * Processors that are not defined in the table,
 204  * use the default SCHED_POWER_SCALE value for cpu_scale.
 205  */
 206 static const struct cpu_efficiency table_efficiency[] = {
 207         { "arm,cortex-a57", 3891 },
 208         { "arm,cortex-a53", 2048 },
 209         { NULL, },
 210 };
 211
 212 static unsigned long *__cpu_capacity;
 213 #define cpu_capacity(cpu)       __cpu_capacity[cpu]
 214
 215 static unsigned long middle_capacity = 1;
 216
 217 /*
 218  * Iterate all CPUs' descriptor in DT and compute the efficiency
 219  * (as per table_efficiency). Also calculate a middle efficiency
 220  * as close as possible to  (max{eff_i} - min{eff_i}) / 2
 221  * This is later used to scale the cpu_power field such that an
 222  * 'average' CPU is of middle power. Also see the comments near
 223  * table_efficiency[] and update_cpu_power().
 224  */
 225 static int __init parse_dt_topology(void)
 226 {
 227         struct device_node *cn, *map;
 228         int ret = 0;
 229         int cpu;
 230
 231         cn = of_find_node_by_path("/cpus");
 232         if (!cn) {
 233                 pr_err("No CPU information found in DT\n");
 234                 return 0;
 235         }
 236
 237         /*
 238          * When topology is provided cpu-map is essentially a root
 239          * cluster with restricted subnodes.
 240          */
 241         map = of_get_child_by_name(cn, "cpu-map");
 242         if (!map)
 243                 goto out;
 244
 245         ret = parse_cluster(map, 0);
 246         if (ret != 0)
 247                 goto out_map;
 248
 249         /*
 250          * Check that all cores are in the topology; the SMP code will
 251          * only mark cores described in the DT as possible.
 252          */
 253         for_each_possible_cpu(cpu) {
 254                 if (cpu_topology[cpu].cluster_id == -1) {
 255                         pr_err("CPU%d: No topology information specified\n",
 256                                cpu);
 257                         ret = -EINVAL;
 258                 }
 259         }
 260
 261 out_map:
 262         of_node_put(map);
 263 out:
 264         of_node_put(cn);
 265         return ret;
 266 }
 267
 268 static void __init parse_dt_cpu_power(void)
 269 {
 270         const struct cpu_efficiency *cpu_eff;
 271         struct device_node *cn;
 272         unsigned long min_capacity = ULONG_MAX;
 273         unsigned long max_capacity = 0;
 274         unsigned long capacity = 0;
 275         int cpu;
 276
 277         __cpu_capacity = kcalloc(nr_cpu_ids, sizeof(*__cpu_capacity),
 278                                  GFP_NOWAIT);
 279
 280         for_each_possible_cpu(cpu) {
 281                 const u32 *rate;
 282                 int len;
 283
 284                 /* Too early to use cpu->of_node */
 285                 cn = of_get_cpu_node(cpu, NULL);
 286                 if (!cn) {
 287                         pr_err("Missing device node for CPU %d\n", cpu);
 288                         continue;
 289                 }
 290
 291                 for (cpu_eff = table_efficiency; cpu_eff->compatible; cpu_eff++)
 292                         if (of_device_is_compatible(cn, cpu_eff->compatible))
 293                                 break;
 294
 295                 if (cpu_eff->compatible == NULL) {
 296                         pr_warn("%s: Unknown CPU type\n", cn->full_name);
 297                         continue;
 298                 }
 299
 300                 rate = of_get_property(cn, "clock-frequency", &len);
 301                 if (!rate || len != 4) {
 302                         pr_err("%s: Missing clock-frequency property\n",
 303                                 cn->full_name);
 304                         continue;
 305                 }
 306
 307                 capacity = ((be32_to_cpup(rate)) >> 20) * cpu_eff->efficiency;
 308
 309                 /* Save min capacity of the system */
 310                 if (capacity < min_capacity)
 311                         min_capacity = capacity;
 312
 313                 /* Save max capacity of the system */
 314                 if (capacity > max_capacity)
 315                         max_capacity = capacity;
 316
 317                 cpu_capacity(cpu) = capacity;
 318         }
 319
 320         /* compute a middle_capacity factor that will ensure that the capacity
 321          * of an 'average' CPU of the system will be as close as possible to
 322          * SCHED_POWER_SCALE, which is the default value, but with the
 323          * constraint explained near table_efficiency[].
 324          */
 325         if (4 * max_capacity < (3 * (max_capacity + min_capacity)))
 326                 middle_capacity = (min_capacity + max_capacity)
 327                                 >> (SCHED_POWER_SHIFT+1);
 328         else
 329                 middle_capacity = ((max_capacity / 3)
 330                                 >> (SCHED_POWER_SHIFT-1)) + 1;
 331 }
 332
 333 /*
 334  * Look for a customed capacity of a CPU in the cpu_topo_data table during the
 335  * boot. The update of all CPUs is in O(n^2) for heteregeneous system but the
 336  * function returns directly for SMP system.
 337  */
 338 static void update_cpu_power(unsigned int cpu)
 339 {
 340         if (!cpu_capacity(cpu))
 341                 return;
 342
 343         set_power_scale(cpu, cpu_capacity(cpu) / middle_capacity);
 344
 345         pr_info("CPU%u: update cpu_power %lu\n",
 346                 cpu, arch_scale_freq_power(NULL, cpu));
 347 }
 348
 349 /*
 350  * cpu topology table
 351  */
 352 struct cpu_topology cpu_topology[NR_CPUS];
 353 EXPORT_SYMBOL_GPL(cpu_topology);
 354
 355 const struct cpumask *cpu_coregroup_mask(int cpu)
 356 {
 357         return &cpu_topology[cpu].core_sibling;
 358 }
 359
 360 static void update_siblings_masks(unsigned int cpuid)
 361 {
 362         struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
 363         int cpu;
 364
 365         if (cpuid_topo->cluster_id == -1) {
 366                 /*
 367                  * DT does not contain topology information for this cpu.
 368                  */
 369                 pr_debug("CPU%u: No topology information configured\n", cpuid);
 370                 return;
 371         }
 372
 373         /* update core and thread sibling masks */
 374         for_each_possible_cpu(cpu) {
 375                 cpu_topo = &cpu_topology[cpu];
 376
 377                 if (cpuid_topo->cluster_id != cpu_topo->cluster_id)
 378                         continue;
 379
 380                 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
 381                 if (cpu != cpuid)
 382                         cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
 383
 384                 if (cpuid_topo->core_id != cpu_topo->core_id)
 385                         continue;
 386
 387                 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
 388                 if (cpu != cpuid)
 389                         cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
 390         }
 391 }
 392
 393 #ifdef CONFIG_SCHED_HMP
 394
 395 /*
 396  * Retrieve logical cpu index corresponding to a given MPIDR[23:0]
 397  *  - mpidr: MPIDR[23:0] to be used for the look-up
 398  *
 399  * Returns the cpu logical index or -EINVAL on look-up error
 400  */
 401 static inline int get_logical_index(u32 mpidr)
 402 {
 403         int cpu;
 404         for (cpu = 0; cpu < nr_cpu_ids; cpu++)
 405                 if (cpu_logical_map(cpu) == mpidr)
 406                         return cpu;
 407         return -EINVAL;
 408 }
 409
 410 static const char * const little_cores[] = {
 411         "arm,cortex-a53",
 412         NULL,
 413 };
 414
 415 static bool is_little_cpu(struct device_node *cn)
 416 {
 417         const char * const *lc;
 418         for (lc = little_cores; *lc; lc++)
 419                 if (of_device_is_compatible(cn, *lc))
 420                         return true;
 421         return false;
 422 }
 423
 424 void __init arch_get_fast_and_slow_cpus(struct cpumask *fast,
 425                                         struct cpumask *slow)
 426 {
 427         struct device_node *cn = NULL;
 428         int cpu;
 429
 430         cpumask_clear(fast);
 431         cpumask_clear(slow);
 432
 433         /*
 434          * Use the config options if they are given. This helps testing
 435          * HMP scheduling on systems without a big.LITTLE architecture.
 436          */
 437         if (strlen(CONFIG_HMP_FAST_CPU_MASK) && strlen(CONFIG_HMP_SLOW_CPU_MASK)) {
 438                 if (cpulist_parse(CONFIG_HMP_FAST_CPU_MASK, fast))
 439                         WARN(1, "Failed to parse HMP fast cpu mask!\n");
 440                 if (cpulist_parse(CONFIG_HMP_SLOW_CPU_MASK, slow))
 441                         WARN(1, "Failed to parse HMP slow cpu mask!\n");
 442                 return;
 443         }
 444
 445         /*
 446          * Else, parse device tree for little cores.
 447          */
 448         while ((cn = of_find_node_by_type(cn, "cpu"))) {
 449
 450                 const u32 *mpidr;
 451                 int len;
 452
 453                 mpidr = of_get_property(cn, "reg", &len);
 454                 if (!mpidr || len != 8) {
 455                         pr_err("%s missing reg property\n", cn->full_name);
 456                         continue;
 457                 }
 458
 459                 cpu = get_logical_index(be32_to_cpup(mpidr+1));
 460                 if (cpu == -EINVAL) {
 461                         pr_err("couldn't get logical index for mpidr %x\n",
 462                                                         be32_to_cpup(mpidr+1));
 463                         break;
 464                 }
 465
 466                 if (is_little_cpu(cn))
 467                         cpumask_set_cpu(cpu, slow);
 468                 else
 469                         cpumask_set_cpu(cpu, fast);
 470         }
 471
 472         if (!cpumask_empty(fast) && !cpumask_empty(slow))
 473                 return;
 474
 475         /*
 476          * We didn't find both big and little cores so let's call all cores
 477          * fast as this will keep the system running, with all cores being
 478          * treated equal.
 479          */
 480         cpumask_setall(fast);
 481         cpumask_clear(slow);
 482 }
 483
 484 struct cpumask hmp_slow_cpu_mask;
 485
 486 void __init arch_get_hmp_domains(struct list_head *hmp_domains_list)
 487 {
 488         struct cpumask hmp_fast_cpu_mask;
 489         struct hmp_domain *domain;
 490
 491         arch_get_fast_and_slow_cpus(&hmp_fast_cpu_mask, &hmp_slow_cpu_mask);
 492
 493         /*
 494          * Initialize hmp_domains
 495          * Must be ordered with respect to compute capacity.
 496          * Fastest domain at head of list.
 497          */
 498         if(!cpumask_empty(&hmp_slow_cpu_mask)) {
 499                 domain = (struct hmp_domain *)
 500                         kmalloc(sizeof(struct hmp_domain), GFP_KERNEL);
 501                 cpumask_copy(&domain->possible_cpus, &hmp_slow_cpu_mask);
 502                 cpumask_and(&domain->cpus, cpu_online_mask, &domain->possible_cpus);
 503                 list_add(&domain->hmp_domains, hmp_domains_list);
 504         }
 505         domain = (struct hmp_domain *)
 506                 kmalloc(sizeof(struct hmp_domain), GFP_KERNEL);
 507         cpumask_copy(&domain->possible_cpus, &hmp_fast_cpu_mask);
 508         cpumask_and(&domain->cpus, cpu_online_mask, &domain->possible_cpus);
 509         list_add(&domain->hmp_domains, hmp_domains_list);
 510 }
 511 #endif /* CONFIG_SCHED_HMP */
 512
 513 /*
 514  * cluster_to_logical_mask - return cpu logical mask of CPUs in a cluster
 515  * @socket_id:          cluster HW identifier
 516  * @cluster_mask:       the cpumask location to be initialized, modified by the
 517  *                      function only if return value == 0
 518  *
 519  * Return:
 520  *
 521  * 0 on success
 522  * -EINVAL if cluster_mask is NULL or there is no record matching socket_id
 523  */
 524 int cluster_to_logical_mask(unsigned int socket_id, cpumask_t *cluster_mask)
 525 {
 526         int cpu;
 527
 528         if (!cluster_mask)
 529                 return -EINVAL;
 530
 531         for_each_online_cpu(cpu) {
 532                 if (socket_id == topology_physical_package_id(cpu)) {
 533                         cpumask_copy(cluster_mask, topology_core_cpumask(cpu));
 534                         return 0;
 535                 }
 536         }
 537
 538         return -EINVAL;
 539 }
 540
 541 void store_cpu_topology(unsigned int cpuid)
 542 {
 543         update_siblings_masks(cpuid);
 544         update_cpu_power(cpuid);
 545 }
 546
 547 static void __init reset_cpu_topology(void)
 548 {
 549         unsigned int cpu;
 550
 551         for_each_possible_cpu(cpu) {
 552                 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
 553
 554                 cpu_topo->thread_id = -1;
 555                 cpu_topo->core_id = 0;
 556                 cpu_topo->cluster_id = -1;
 557
 558                 cpumask_clear(&cpu_topo->core_sibling);
 559                 cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
 560                 cpumask_clear(&cpu_topo->thread_sibling);
 561                 cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
 562         }
 563 }
 564
 565 static void __init reset_cpu_power(void)
 566 {
 567         unsigned int cpu;
 568
 569         for_each_possible_cpu(cpu)
 570                 set_power_scale(cpu, SCHED_POWER_SCALE);
 571 }
 572
 573 void __init init_cpu_topology(void)
 574 {
 575         reset_cpu_topology();
 576
 577         /*
 578          * Discard anything that was parsed if we hit an error so we
 579          * don't use partial information.
 580          */
 581         if (parse_dt_topology())
 582                 reset_cpu_topology();
 583
 584         reset_cpu_power();
 585         parse_dt_cpu_power();
 586 }