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/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 28 /* 29 * cpu power scale management 30 */ 31 32 /* 33 * cpu power table 34 * This per cpu data structure describes the relative capacity of each core. 35 * On a heteregenous system, cores don't have the same computation capacity 36 * and we reflect that difference in the cpu_power field so the scheduler can 37 * take this difference into account during load balance. A per cpu structure 38 * is preferred because each CPU updates its own cpu_power field during the 39 * load balance except for idle cores. One idle core is selected to run the 40 * rebalance_domains for all idle cores and the cpu_power can be updated 41 * during this sequence. 42 */ 43 static DEFINE_PER_CPU(unsigned long, cpu_scale); 44 45 unsigned long arch_scale_freq_power(struct sched_domain *sd, int cpu) 46 { 47 return per_cpu(cpu_scale, cpu); 48 } 49 50 static void set_power_scale(unsigned int cpu, unsigned long power) 51 { 52 per_cpu(cpu_scale, cpu) = power; 53 } 54 55 #ifdef CONFIG_OF 56 struct cpu_efficiency { 57 const char *compatible; 58 unsigned long efficiency; 59 }; 60 61 /* 62 * Table of relative efficiency of each processors 63 * The efficiency value must fit in 20bit and the final 64 * cpu_scale value must be in the range 65 * 0 < cpu_scale < 3*SCHED_POWER_SCALE/2 66 * in order to return at most 1 when DIV_ROUND_CLOSEST 67 * is used to compute the capacity of a CPU. 68 * Processors that are not defined in the table, 69 * use the default SCHED_POWER_SCALE value for cpu_scale. 70 */ 71 struct cpu_efficiency table_efficiency[] = { 72 {"arm,cortex-a15", 3891}, 73 {"arm,cortex-a7", 2048}, 74 {NULL, }, 75 }; 76 77 unsigned long *__cpu_capacity; 78 #define cpu_capacity(cpu) __cpu_capacity[cpu] 79 80 unsigned long middle_capacity = 1; 81 82 /* 83 * Iterate all CPUs' descriptor in DT and compute the efficiency 84 * (as per table_efficiency). Also calculate a middle efficiency 85 * as close as possible to (max{eff_i} - min{eff_i}) / 2 86 * This is later used to scale the cpu_power field such that an 87 * 'average' CPU is of middle power. Also see the comments near 88 * table_efficiency[] and update_cpu_power(). 89 */ 90 static void __init parse_dt_topology(void) 91 { 92 struct cpu_efficiency *cpu_eff; 93 struct device_node *cn = NULL; 94 unsigned long min_capacity = (unsigned long)(-1); 95 unsigned long max_capacity = 0; 96 unsigned long capacity = 0; 97 int alloc_size, cpu = 0; 98 99 alloc_size = nr_cpu_ids * sizeof(*__cpu_capacity); 100 __cpu_capacity = kzalloc(alloc_size, GFP_NOWAIT); 101 102 for_each_possible_cpu(cpu) { 103 const u32 *rate; 104 int len; 105 106 /* too early to use cpu->of_node */ 107 cn = of_get_cpu_node(cpu, NULL); 108 if (!cn) { 109 pr_err("missing device node for CPU %d\n", cpu); 110 continue; 111 } 112 113 for (cpu_eff = table_efficiency; cpu_eff->compatible; cpu_eff++) 114 if (of_device_is_compatible(cn, cpu_eff->compatible)) 115 break; 116 117 if (cpu_eff->compatible == NULL) 118 continue; 119 120 rate = of_get_property(cn, "clock-frequency", &len); 121 if (!rate || len != 4) { 122 pr_err("%s missing clock-frequency property\n", 123 cn->full_name); 124 continue; 125 } 126 127 capacity = ((be32_to_cpup(rate)) >> 20) * cpu_eff->efficiency; 128 129 /* Save min capacity of the system */ 130 if (capacity < min_capacity) 131 min_capacity = capacity; 132 133 /* Save max capacity of the system */ 134 if (capacity > max_capacity) 135 max_capacity = capacity; 136 137 cpu_capacity(cpu) = capacity; 138 } 139 140 /* If min and max capacities are equals, we bypass the update of the 141 * cpu_scale because all CPUs have the same capacity. Otherwise, we 142 * compute a middle_capacity factor that will ensure that the capacity 143 * of an 'average' CPU of the system will be as close as possible to 144 * SCHED_POWER_SCALE, which is the default value, but with the 145 * constraint explained near table_efficiency[]. 146 */ 147 if (4*max_capacity < (3*(max_capacity + min_capacity))) 148 middle_capacity = (min_capacity + max_capacity) 149 >> (SCHED_POWER_SHIFT+1); 150 else 151 middle_capacity = ((max_capacity / 3) 152 >> (SCHED_POWER_SHIFT-1)) + 1; 153 154 } 155 156 /* 157 * Look for a customed capacity of a CPU in the cpu_capacity table during the 158 * boot. The update of all CPUs is in O(n^2) for heteregeneous system but the 159 * function returns directly for SMP system. 160 */ 161 void update_cpu_power(unsigned int cpu) 162 { 163 if (!cpu_capacity(cpu)) 164 return; 165 166 set_power_scale(cpu, cpu_capacity(cpu) / middle_capacity); 167 168 printk(KERN_INFO "CPU%u: update cpu_power %lu\n", 169 cpu, arch_scale_freq_power(NULL, cpu)); 170 } 171 172 #else 173 static inline void parse_dt_topology(void) {} 174 static inline void update_cpu_power(unsigned int cpuid) {} 175 #endif 176 177 /* 178 * cpu topology table 179 */ 180 struct cputopo_arm cpu_topology[NR_CPUS]; 181 EXPORT_SYMBOL_GPL(cpu_topology); 182 183 const struct cpumask *cpu_coregroup_mask(int cpu) 184 { 185 return &cpu_topology[cpu].core_sibling; 186 } 187 188 void update_siblings_masks(unsigned int cpuid) 189 { 190 struct cputopo_arm *cpu_topo, *cpuid_topo = &cpu_topology[cpuid]; 191 int cpu; 192 193 /* update core and thread sibling masks */ 194 for_each_possible_cpu(cpu) { 195 cpu_topo = &cpu_topology[cpu]; 196 197 if (cpuid_topo->socket_id != cpu_topo->socket_id) 198 continue; 199 200 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling); 201 if (cpu != cpuid) 202 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling); 203 204 if (cpuid_topo->core_id != cpu_topo->core_id) 205 continue; 206 207 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling); 208 if (cpu != cpuid) 209 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling); 210 } 211 smp_wmb(); 212 } 213 214 /* 215 * store_cpu_topology is called at boot when only one cpu is running 216 * and with the mutex cpu_hotplug.lock locked, when several cpus have booted, 217 * which prevents simultaneous write access to cpu_topology array 218 */ 219 void store_cpu_topology(unsigned int cpuid) 220 { 221 struct cputopo_arm *cpuid_topo = &cpu_topology[cpuid]; 222 unsigned int mpidr; 223 224 /* If the cpu topology has been already set, just return */ 225 if (cpuid_topo->core_id != -1) 226 return; 227 228 mpidr = read_cpuid_mpidr(); 229 230 /* create cpu topology mapping */ 231 if ((mpidr & MPIDR_SMP_BITMASK) == MPIDR_SMP_VALUE) { 232 /* 233 * This is a multiprocessor system 234 * multiprocessor format & multiprocessor mode field are set 235 */ 236 237 if (mpidr & MPIDR_MT_BITMASK) { 238 /* core performance interdependency */ 239 cpuid_topo->thread_id = MPIDR_AFFINITY_LEVEL(mpidr, 0); 240 cpuid_topo->core_id = MPIDR_AFFINITY_LEVEL(mpidr, 1); 241 cpuid_topo->socket_id = MPIDR_AFFINITY_LEVEL(mpidr, 2); 242 } else { 243 /* largely independent cores */ 244 cpuid_topo->thread_id = -1; 245 cpuid_topo->core_id = MPIDR_AFFINITY_LEVEL(mpidr, 0); 246 cpuid_topo->socket_id = MPIDR_AFFINITY_LEVEL(mpidr, 1); 247 } 248 } else { 249 /* 250 * This is an uniprocessor system 251 * we are in multiprocessor format but uniprocessor system 252 * or in the old uniprocessor format 253 */ 254 cpuid_topo->thread_id = -1; 255 cpuid_topo->core_id = 0; 256 cpuid_topo->socket_id = -1; 257 } 258 259 update_siblings_masks(cpuid); 260 261 update_cpu_power(cpuid); 262 263 printk(KERN_INFO "CPU%u: thread %d, cpu %d, socket %d, mpidr %x\n", 264 cpuid, cpu_topology[cpuid].thread_id, 265 cpu_topology[cpuid].core_id, 266 cpu_topology[cpuid].socket_id, mpidr); 267 } 268 269 /* 270 * init_cpu_topology is called at boot when only one cpu is running 271 * which prevent simultaneous write access to cpu_topology array 272 */ 273 void __init init_cpu_topology(void) 274 { 275 unsigned int cpu; 276 277 /* init core mask and power*/ 278 for_each_possible_cpu(cpu) { 279 struct cputopo_arm *cpu_topo = &(cpu_topology[cpu]); 280 281 cpu_topo->thread_id = -1; 282 cpu_topo->core_id = -1; 283 cpu_topo->socket_id = -1; 284 cpumask_clear(&cpu_topo->core_sibling); 285 cpumask_clear(&cpu_topo->thread_sibling); 286 287 set_power_scale(cpu, SCHED_POWER_SCALE); 288 } 289 smp_wmb(); 290 291 parse_dt_topology(); 292 } 293