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/arch_topology.h>
  16 #include <linux/cacheinfo.h>
  17 #include <linux/cpu.h>
  18 #include <linux/cpumask.h>
  19 #include <linux/init.h>
  20 #include <linux/percpu.h>
  21 #include <linux/node.h>
  22 #include <linux/nodemask.h>
  23 #include <linux/of.h>
  24 #include <linux/sched.h>
  25 #include <linux/sched/topology.h>
  26 #include <linux/slab.h>
  27 #include <linux/smp.h>
  28 #include <linux/string.h>
  29
  30 #include <asm/cpu.h>
  31 #include <asm/cputype.h>
  32 #include <asm/topology.h>
  33
  34 static int __init get_cpu_for_node(struct device_node *node)
  35 {
  36         struct device_node *cpu_node;
  37         int cpu;
  38
  39         cpu_node = of_parse_phandle(node, "cpu", 0);
  40         if (!cpu_node)
  41                 return -1;
  42
  43         cpu = of_cpu_node_to_id(cpu_node);
  44         if (cpu >= 0)
  45                 topology_parse_cpu_capacity(cpu_node, cpu);
  46         else
  47                 pr_crit("Unable to find CPU node for %pOF\n", cpu_node);
  48
  49         of_node_put(cpu_node);
  50         return cpu;
  51 }
  52
  53 static int __init parse_core(struct device_node *core, int package_id,
  54                              int core_id)
  55 {
  56         char name[10];
  57         bool leaf = true;
  58         int i = 0;
  59         int cpu;
  60         struct device_node *t;
  61
  62         do {
  63                 snprintf(name, sizeof(name), "thread%d", i);
  64                 t = of_get_child_by_name(core, name);
  65                 if (t) {
  66                         leaf = false;
  67                         cpu = get_cpu_for_node(t);
  68                         if (cpu >= 0) {
  69                                 cpu_topology[cpu].package_id = package_id;
  70                                 cpu_topology[cpu].core_id = core_id;
  71                                 cpu_topology[cpu].thread_id = i;
  72                         } else {
  73                                 pr_err("%pOF: Can't get CPU for thread\n",
  74                                        t);
  75                                 of_node_put(t);
  76                                 return -EINVAL;
  77                         }
  78                         of_node_put(t);
  79                 }
  80                 i++;
  81         } while (t);
  82
  83         cpu = get_cpu_for_node(core);
  84         if (cpu >= 0) {
  85                 if (!leaf) {
  86                         pr_err("%pOF: Core has both threads and CPU\n",
  87                                core);
  88                         return -EINVAL;
  89                 }
  90
  91                 cpu_topology[cpu].package_id = package_id;
  92                 cpu_topology[cpu].core_id = core_id;
  93         } else if (leaf) {
  94                 pr_err("%pOF: Can't get CPU for leaf core\n", core);
  95                 return -EINVAL;
  96         }
  97
  98         return 0;
  99 }
 100
 101 static int __init parse_cluster(struct device_node *cluster, int depth)
 102 {
 103         char name[10];
 104         bool leaf = true;
 105         bool has_cores = false;
 106         struct device_node *c;
 107         static int package_id __initdata;
 108         int core_id = 0;
 109         int i, ret;
 110
 111         /*
 112          * First check for child clusters; we currently ignore any
 113          * information about the nesting of clusters and present the
 114          * scheduler with a flat list of them.
 115          */
 116         i = 0;
 117         do {
 118                 snprintf(name, sizeof(name), "cluster%d", i);
 119                 c = of_get_child_by_name(cluster, name);
 120                 if (c) {
 121                         leaf = false;
 122                         ret = parse_cluster(c, depth + 1);
 123                         of_node_put(c);
 124                         if (ret != 0)
 125                                 return ret;
 126                 }
 127                 i++;
 128         } while (c);
 129
 130         /* Now check for cores */
 131         i = 0;
 132         do {
 133                 snprintf(name, sizeof(name), "core%d", i);
 134                 c = of_get_child_by_name(cluster, name);
 135                 if (c) {
 136                         has_cores = true;
 137
 138                         if (depth == 0) {
 139                                 pr_err("%pOF: cpu-map children should be clusters\n",
 140                                        c);
 141                                 of_node_put(c);
 142                                 return -EINVAL;
 143                         }
 144
 145                         if (leaf) {
 146                                 ret = parse_core(c, package_id, core_id++);
 147                         } else {
 148                                 pr_err("%pOF: Non-leaf cluster with core %s\n",
 149                                        cluster, name);
 150                                 ret = -EINVAL;
 151                         }
 152
 153                         of_node_put(c);
 154                         if (ret != 0)
 155                                 return ret;
 156                 }
 157                 i++;
 158         } while (c);
 159
 160         if (leaf && !has_cores)
 161                 pr_warn("%pOF: empty cluster\n", cluster);
 162
 163         if (leaf)
 164                 package_id++;
 165
 166         return 0;
 167 }
 168
 169 static int __init parse_dt_topology(void)
 170 {
 171         struct device_node *cn, *map;
 172         int ret = 0;
 173         int cpu;
 174
 175         cn = of_find_node_by_path("/cpus");
 176         if (!cn) {
 177                 pr_err("No CPU information found in DT\n");
 178                 return 0;
 179         }
 180
 181         /*
 182          * When topology is provided cpu-map is essentially a root
 183          * cluster with restricted subnodes.
 184          */
 185         map = of_get_child_by_name(cn, "cpu-map");
 186         if (!map)
 187                 goto out;
 188
 189         ret = parse_cluster(map, 0);
 190         if (ret != 0)
 191                 goto out_map;
 192
 193         topology_normalize_cpu_scale();
 194
 195         /*
 196          * Check that all cores are in the topology; the SMP code will
 197          * only mark cores described in the DT as possible.
 198          */
 199         for_each_possible_cpu(cpu)
 200                 if (cpu_topology[cpu].package_id == -1)
 201                         ret = -EINVAL;
 202
 203 out_map:
 204         of_node_put(map);
 205 out:
 206         of_node_put(cn);
 207         return ret;
 208 }
 209
 210 /*
 211  * cpu topology table
 212  */
 213 struct cpu_topology cpu_topology[NR_CPUS];
 214 EXPORT_SYMBOL_GPL(cpu_topology);
 215
 216 const struct cpumask *cpu_coregroup_mask(int cpu)
 217 {
 218         const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
 219
 220         /* Find the smaller of NUMA, core or LLC siblings */
 221         if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
 222                 /* not numa in package, lets use the package siblings */
 223                 core_mask = &cpu_topology[cpu].core_sibling;
 224         }
 225         if (cpu_topology[cpu].llc_id != -1) {
 226                 if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
 227                         core_mask = &cpu_topology[cpu].llc_sibling;
 228         }
 229
 230         return core_mask;
 231 }
 232
 233 static void update_siblings_masks(unsigned int cpuid)
 234 {
 235         struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
 236         int cpu;
 237
 238         /* update core and thread sibling masks */
 239         for_each_online_cpu(cpu) {
 240                 cpu_topo = &cpu_topology[cpu];
 241
 242                 if (cpuid_topo->llc_id == cpu_topo->llc_id) {
 243                         cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling);
 244                         cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling);
 245                 }
 246
 247                 if (cpuid_topo->package_id != cpu_topo->package_id)
 248                         continue;
 249
 250                 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
 251                 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
 252
 253                 if (cpuid_topo->core_id != cpu_topo->core_id)
 254                         continue;
 255
 256                 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
 257                 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
 258         }
 259 }
 260
 261 void store_cpu_topology(unsigned int cpuid)
 262 {
 263         struct cpu_topology *cpuid_topo = &cpu_topology[cpuid];
 264         u64 mpidr;
 265
 266         if (cpuid_topo->package_id != -1)
 267                 goto topology_populated;
 268
 269         mpidr = read_cpuid_mpidr();
 270
 271         /* Uniprocessor systems can rely on default topology values */
 272         if (mpidr & MPIDR_UP_BITMASK)
 273                 return;
 274
 275         /*
 276          * This would be the place to create cpu topology based on MPIDR.
 277          *
 278          * However, it cannot be trusted to depict the actual topology; some
 279          * pieces of the architecture enforce an artificial cap on Aff0 values
 280          * (e.g. GICv3's ICC_SGI1R_EL1 limits it to 15), leading to an
 281          * artificial cycling of Aff1, Aff2 and Aff3 values. IOW, these end up
 282          * having absolutely no relationship to the actual underlying system
 283          * topology, and cannot be reasonably used as core / package ID.
 284          *
 285          * If the MT bit is set, Aff0 *could* be used to define a thread ID, but
 286          * we still wouldn't be able to obtain a sane core ID. This means we
 287          * need to entirely ignore MPIDR for any topology deduction.
 288          */
 289         cpuid_topo->thread_id  = -1;
 290         cpuid_topo->core_id    = cpuid;
 291         cpuid_topo->package_id = cpu_to_node(cpuid);
 292
 293         pr_debug("CPU%u: cluster %d core %d thread %d mpidr %#016llx\n",
 294                  cpuid, cpuid_topo->package_id, cpuid_topo->core_id,
 295                  cpuid_topo->thread_id, mpidr);
 296
 297 topology_populated:
 298         update_siblings_masks(cpuid);
 299 }
 300
 301 static void clear_cpu_topology(int cpu)
 302 {
 303         struct cpu_topology *cpu_topo = &cpu_topology[cpu];
 304
 305         cpumask_clear(&cpu_topo->llc_sibling);
 306         cpumask_set_cpu(cpu, &cpu_topo->llc_sibling);
 307
 308         cpumask_clear(&cpu_topo->core_sibling);
 309         cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
 310         cpumask_clear(&cpu_topo->thread_sibling);
 311         cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
 312 }
 313
 314 static void __init reset_cpu_topology(void)
 315 {
 316         unsigned int cpu;
 317
 318         for_each_possible_cpu(cpu) {
 319                 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
 320
 321                 cpu_topo->thread_id = -1;
 322                 cpu_topo->core_id = 0;
 323                 cpu_topo->package_id = -1;
 324                 cpu_topo->llc_id = -1;
 325
 326                 clear_cpu_topology(cpu);
 327         }
 328 }
 329
 330 void remove_cpu_topology(unsigned int cpu)
 331 {
 332         int sibling;
 333
 334         for_each_cpu(sibling, topology_core_cpumask(cpu))
 335                 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
 336         for_each_cpu(sibling, topology_sibling_cpumask(cpu))
 337                 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
 338         for_each_cpu(sibling, topology_llc_cpumask(cpu))
 339                 cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling));
 340
 341         clear_cpu_topology(cpu);
 342 }
 343
 344 #ifdef CONFIG_ACPI
 345 static bool __init acpi_cpu_is_threaded(int cpu)
 346 {
 347         int is_threaded = acpi_pptt_cpu_is_thread(cpu);
 348
 349         /*
 350          * if the PPTT doesn't have thread information, assume a homogeneous
 351          * machine and return the current CPU's thread state.
 352          */
 353         if (is_threaded < 0)
 354                 is_threaded = read_cpuid_mpidr() & MPIDR_MT_BITMASK;
 355
 356         return !!is_threaded;
 357 }
 358
 359 /*
 360  * Propagate the topology information of the processor_topology_node tree to the
 361  * cpu_topology array.
 362  */
 363 static int __init parse_acpi_topology(void)
 364 {
 365         int cpu, topology_id;
 366
 367         for_each_possible_cpu(cpu) {
 368                 int i, cache_id;
 369
 370                 topology_id = find_acpi_cpu_topology(cpu, 0);
 371                 if (topology_id < 0)
 372                         return topology_id;
 373
 374                 if (acpi_cpu_is_threaded(cpu)) {
 375                         cpu_topology[cpu].thread_id = topology_id;
 376                         topology_id = find_acpi_cpu_topology(cpu, 1);
 377                         cpu_topology[cpu].core_id   = topology_id;
 378                 } else {
 379                         cpu_topology[cpu].thread_id  = -1;
 380                         cpu_topology[cpu].core_id    = topology_id;
 381                 }
 382                 topology_id = find_acpi_cpu_topology_package(cpu);
 383                 cpu_topology[cpu].package_id = topology_id;
 384
 385                 i = acpi_find_last_cache_level(cpu);
 386
 387                 if (i > 0) {
 388                         /*
 389                          * this is the only part of cpu_topology that has
 390                          * a direct relationship with the cache topology
 391                          */
 392                         cache_id = find_acpi_cpu_cache_topology(cpu, i);
 393                         if (cache_id > 0)
 394                                 cpu_topology[cpu].llc_id = cache_id;
 395                 }
 396         }
 397
 398         return 0;
 399 }
 400
 401 #else
 402 static inline int __init parse_acpi_topology(void)
 403 {
 404         return -EINVAL;
 405 }
 406 #endif
 407
 408 void __init init_cpu_topology(void)
 409 {
 410         reset_cpu_topology();
 411
 412         /*
 413          * Discard anything that was parsed if we hit an error so we
 414          * don't use partial information.
 415          */
 416         if (!acpi_disabled && parse_acpi_topology())
 417                 reset_cpu_topology();
 418         else if (of_have_populated_dt() && parse_dt_topology())
 419                 reset_cpu_topology();
 420 }