arch/arm/kernel/topology.c

   1 /*
   2  * arch/arm/kernel/topology.c
   3  *
   4  * Copyright (C) 2011 Linaro Limited.
   5  * Written by: Vincent Guittot
   6  *
   7  * based on 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/arch_topology.h>
  15 #include <linux/cpu.h>
  16 #include <linux/cpufreq.h>
  17 #include <linux/cpumask.h>
  18 #include <linux/export.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/string.h>
  28
  29 #include <asm/cpu.h>
  30 #include <asm/cputype.h>
  31 #include <asm/topology.h>
  32
  33 /*
  34  * cpu capacity scale management
  35  */
  36
  37 /*
  38  * cpu capacity table
  39  * This per cpu data structure describes the relative capacity of each core.
  40  * On a heteregenous system, cores don't have the same computation capacity
  41  * and we reflect that difference in the cpu_capacity field so the scheduler
  42  * can take this difference into account during load balance. A per cpu
  43  * structure is preferred because each CPU updates its own cpu_capacity field
  44  * during the load balance except for idle cores. One idle core is selected
  45  * to run the rebalance_domains for all idle cores and the cpu_capacity can be
  46  * updated during this sequence.
  47  */
  48
  49 #ifdef CONFIG_OF
  50 struct cpu_efficiency {
  51         const char *compatible;
  52         unsigned long efficiency;
  53 };
  54
  55 /*
  56  * Table of relative efficiency of each processors
  57  * The efficiency value must fit in 20bit and the final
  58  * cpu_scale value must be in the range
  59  *   0 < cpu_scale < 3*SCHED_CAPACITY_SCALE/2
  60  * in order to return at most 1 when DIV_ROUND_CLOSEST
  61  * is used to compute the capacity of a CPU.
  62  * Processors that are not defined in the table,
  63  * use the default SCHED_CAPACITY_SCALE value for cpu_scale.
  64  */
  65 static const struct cpu_efficiency table_efficiency[] = {
  66         {"arm,cortex-a15", 3891},
  67         {"arm,cortex-a7",  2048},
  68         {NULL, },
  69 };
  70
  71 static unsigned long *__cpu_capacity;
  72 #define cpu_capacity(cpu)       __cpu_capacity[cpu]
  73
  74 static unsigned long middle_capacity = 1;
  75 static bool cap_from_dt = true;
  76
  77 /*
  78  * Iterate all CPUs' descriptor in DT and compute the efficiency
  79  * (as per table_efficiency). Also calculate a middle efficiency
  80  * as close as possible to  (max{eff_i} - min{eff_i}) / 2
  81  * This is later used to scale the cpu_capacity field such that an
  82  * 'average' CPU is of middle capacity. Also see the comments near
  83  * table_efficiency[] and update_cpu_capacity().
  84  */
  85 static void __init parse_dt_topology(void)
  86 {
  87         const struct cpu_efficiency *cpu_eff;
  88         struct device_node *cn = NULL;
  89         unsigned long min_capacity = ULONG_MAX;
  90         unsigned long max_capacity = 0;
  91         unsigned long capacity = 0;
  92         int cpu = 0;
  93
  94         __cpu_capacity = kcalloc(nr_cpu_ids, sizeof(*__cpu_capacity),
  95                                  GFP_NOWAIT);
  96
  97         cn = of_find_node_by_path("/cpus");
  98         if (!cn) {
  99                 pr_err("No CPU information found in DT\n");
 100                 return;
 101         }
 102
 103         for_each_possible_cpu(cpu) {
 104                 const u32 *rate;
 105                 int len;
 106
 107                 /* too early to use cpu->of_node */
 108                 cn = of_get_cpu_node(cpu, NULL);
 109                 if (!cn) {
 110                         pr_err("missing device node for CPU %d\n", cpu);
 111                         continue;
 112                 }
 113
 114                 if (topology_parse_cpu_capacity(cn, cpu)) {
 115                         of_node_put(cn);
 116                         continue;
 117                 }
 118
 119                 cap_from_dt = false;
 120
 121                 for (cpu_eff = table_efficiency; cpu_eff->compatible; cpu_eff++)
 122                         if (of_device_is_compatible(cn, cpu_eff->compatible))
 123                                 break;
 124
 125                 if (cpu_eff->compatible == NULL)
 126                         continue;
 127
 128                 rate = of_get_property(cn, "clock-frequency", &len);
 129                 if (!rate || len != 4) {
 130                         pr_err("%pOF missing clock-frequency property\n", cn);
 131                         continue;
 132                 }
 133
 134                 capacity = ((be32_to_cpup(rate)) >> 20) * cpu_eff->efficiency;
 135
 136                 /* Save min capacity of the system */
 137                 if (capacity < min_capacity)
 138                         min_capacity = capacity;
 139
 140                 /* Save max capacity of the system */
 141                 if (capacity > max_capacity)
 142                         max_capacity = capacity;
 143
 144                 cpu_capacity(cpu) = capacity;
 145         }
 146
 147         /* If min and max capacities are equals, we bypass the update of the
 148          * cpu_scale because all CPUs have the same capacity. Otherwise, we
 149          * compute a middle_capacity factor that will ensure that the capacity
 150          * of an 'average' CPU of the system will be as close as possible to
 151          * SCHED_CAPACITY_SCALE, which is the default value, but with the
 152          * constraint explained near table_efficiency[].
 153          */
 154         if (4*max_capacity < (3*(max_capacity + min_capacity)))
 155                 middle_capacity = (min_capacity + max_capacity)
 156                                 >> (SCHED_CAPACITY_SHIFT+1);
 157         else
 158                 middle_capacity = ((max_capacity / 3)
 159                                 >> (SCHED_CAPACITY_SHIFT-1)) + 1;
 160
 161         if (cap_from_dt)
 162                 topology_normalize_cpu_scale();
 163 }
 164
 165 /*
 166  * Look for a customed capacity of a CPU in the cpu_capacity table during the
 167  * boot. The update of all CPUs is in O(n^2) for heteregeneous system but the
 168  * function returns directly for SMP system.
 169  */
 170 static void update_cpu_capacity(unsigned int cpu)
 171 {
 172         if (!cpu_capacity(cpu) || cap_from_dt)
 173                 return;
 174
 175         topology_set_cpu_scale(cpu, cpu_capacity(cpu) / middle_capacity);
 176
 177         pr_info("CPU%u: update cpu_capacity %lu\n",
 178                 cpu, topology_get_cpu_scale(NULL, cpu));
 179 }
 180
 181 #else
 182 static inline void parse_dt_topology(void) {}
 183 static inline void update_cpu_capacity(unsigned int cpuid) {}
 184 #endif
 185
 186  /*
 187  * cpu topology table
 188  */
 189 struct cputopo_arm cpu_topology[NR_CPUS];
 190 EXPORT_SYMBOL_GPL(cpu_topology);
 191
 192 const struct cpumask *cpu_coregroup_mask(int cpu)
 193 {
 194         return &cpu_topology[cpu].core_sibling;
 195 }
 196
 197 /*
 198  * The current assumption is that we can power gate each core independently.
 199  * This will be superseded by DT binding once available.
 200  */
 201 const struct cpumask *cpu_corepower_mask(int cpu)
 202 {
 203         return &cpu_topology[cpu].thread_sibling;
 204 }
 205
 206 static void update_siblings_masks(unsigned int cpuid)
 207 {
 208         struct cputopo_arm *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
 209         int cpu;
 210
 211         /* update core and thread sibling masks */
 212         for_each_possible_cpu(cpu) {
 213                 cpu_topo = &cpu_topology[cpu];
 214
 215                 if (cpuid_topo->socket_id != cpu_topo->socket_id)
 216                         continue;
 217
 218                 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
 219                 if (cpu != cpuid)
 220                         cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
 221
 222                 if (cpuid_topo->core_id != cpu_topo->core_id)
 223                         continue;
 224
 225                 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
 226                 if (cpu != cpuid)
 227                         cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
 228         }
 229         smp_wmb();
 230 }
 231
 232 /*
 233  * store_cpu_topology is called at boot when only one cpu is running
 234  * and with the mutex cpu_hotplug.lock locked, when several cpus have booted,
 235  * which prevents simultaneous write access to cpu_topology array
 236  */
 237 void store_cpu_topology(unsigned int cpuid)
 238 {
 239         struct cputopo_arm *cpuid_topo = &cpu_topology[cpuid];
 240         unsigned int mpidr;
 241
 242         /* If the cpu topology has been already set, just return */
 243         if (cpuid_topo->core_id != -1)
 244                 return;
 245
 246         mpidr = read_cpuid_mpidr();
 247
 248         /* create cpu topology mapping */
 249         if ((mpidr & MPIDR_SMP_BITMASK) == MPIDR_SMP_VALUE) {
 250                 /*
 251                  * This is a multiprocessor system
 252                  * multiprocessor format & multiprocessor mode field are set
 253                  */
 254
 255                 if (mpidr & MPIDR_MT_BITMASK) {
 256                         /* core performance interdependency */
 257                         cpuid_topo->thread_id = MPIDR_AFFINITY_LEVEL(mpidr, 0);
 258                         cpuid_topo->core_id = MPIDR_AFFINITY_LEVEL(mpidr, 1);
 259                         cpuid_topo->socket_id = MPIDR_AFFINITY_LEVEL(mpidr, 2);
 260                 } else {
 261                         /* largely independent cores */
 262                         cpuid_topo->thread_id = -1;
 263                         cpuid_topo->core_id = MPIDR_AFFINITY_LEVEL(mpidr, 0);
 264                         cpuid_topo->socket_id = MPIDR_AFFINITY_LEVEL(mpidr, 1);
 265                 }
 266         } else {
 267                 /*
 268                  * This is an uniprocessor system
 269                  * we are in multiprocessor format but uniprocessor system
 270                  * or in the old uniprocessor format
 271                  */
 272                 cpuid_topo->thread_id = -1;
 273                 cpuid_topo->core_id = 0;
 274                 cpuid_topo->socket_id = -1;
 275         }
 276
 277         update_siblings_masks(cpuid);
 278
 279         update_cpu_capacity(cpuid);
 280
 281         pr_info("CPU%u: thread %d, cpu %d, socket %d, mpidr %x\n",
 282                 cpuid, cpu_topology[cpuid].thread_id,
 283                 cpu_topology[cpuid].core_id,
 284                 cpu_topology[cpuid].socket_id, mpidr);
 285 }
 286
 287 static inline int cpu_corepower_flags(void)
 288 {
 289         return SD_SHARE_PKG_RESOURCES  | SD_SHARE_POWERDOMAIN;
 290 }
 291
 292 static struct sched_domain_topology_level arm_topology[] = {
 293 #ifdef CONFIG_SCHED_MC
 294         { cpu_corepower_mask, cpu_corepower_flags, SD_INIT_NAME(GMC) },
 295         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
 296 #endif
 297         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
 298         { NULL, },
 299 };
 300
 301 /*
 302  * init_cpu_topology is called at boot when only one cpu is running
 303  * which prevent simultaneous write access to cpu_topology array
 304  */
 305 void __init init_cpu_topology(void)
 306 {
 307         unsigned int cpu;
 308
 309         /* init core mask and capacity */
 310         for_each_possible_cpu(cpu) {
 311                 struct cputopo_arm *cpu_topo = &(cpu_topology[cpu]);
 312
 313                 cpu_topo->thread_id = -1;
 314                 cpu_topo->core_id =  -1;
 315                 cpu_topo->socket_id = -1;
 316                 cpumask_clear(&cpu_topo->core_sibling);
 317                 cpumask_clear(&cpu_topo->thread_sibling);
 318         }
 319         smp_wmb();
 320
 321         parse_dt_topology();
 322
 323         /* Set scheduler topology descriptor */
 324         set_sched_topology(arm_topology);
 325 }