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/cpufreq.h>
  18 #include <linux/init.h>
  19 #include <linux/percpu.h>
  20
  21 #include <asm/cpu.h>
  22 #include <asm/cputype.h>
  23 #include <asm/topology.h>
  24
  25 #ifdef CONFIG_ACPI
  26 static bool __init acpi_cpu_is_threaded(int cpu)
  27 {
  28         int is_threaded = acpi_pptt_cpu_is_thread(cpu);
  29
  30         /*
  31          * if the PPTT doesn't have thread information, assume a homogeneous
  32          * machine and return the current CPU's thread state.
  33          */
  34         if (is_threaded < 0)
  35                 is_threaded = read_cpuid_mpidr() & MPIDR_MT_BITMASK;
  36
  37         return !!is_threaded;
  38 }
  39
  40 /*
  41  * Propagate the topology information of the processor_topology_node tree to the
  42  * cpu_topology array.
  43  */
  44 int __init parse_acpi_topology(void)
  45 {
  46         int cpu, topology_id;
  47
  48         if (acpi_disabled)
  49                 return 0;
  50
  51         for_each_possible_cpu(cpu) {
  52                 int i, cache_id;
  53
  54                 topology_id = find_acpi_cpu_topology(cpu, 0);
  55                 if (topology_id < 0)
  56                         return topology_id;
  57
  58                 if (acpi_cpu_is_threaded(cpu)) {
  59                         cpu_topology[cpu].thread_id = topology_id;
  60                         topology_id = find_acpi_cpu_topology(cpu, 1);
  61                         cpu_topology[cpu].core_id   = topology_id;
  62                 } else {
  63                         cpu_topology[cpu].thread_id  = -1;
  64                         cpu_topology[cpu].core_id    = topology_id;
  65                 }
  66                 topology_id = find_acpi_cpu_topology_package(cpu);
  67                 cpu_topology[cpu].package_id = topology_id;
  68
  69                 i = acpi_find_last_cache_level(cpu);
  70
  71                 if (i > 0) {
  72                         /*
  73                          * this is the only part of cpu_topology that has
  74                          * a direct relationship with the cache topology
  75                          */
  76                         cache_id = find_acpi_cpu_cache_topology(cpu, i);
  77                         if (cache_id > 0)
  78                                 cpu_topology[cpu].llc_id = cache_id;
  79                 }
  80         }
  81
  82         return 0;
  83 }
  84 #endif
  85
  86 #ifdef CONFIG_ARM64_AMU_EXTN
  87
  88 #undef pr_fmt
  89 #define pr_fmt(fmt) "AMU: " fmt
  90
  91 static DEFINE_PER_CPU_READ_MOSTLY(unsigned long, arch_max_freq_scale);
  92 static DEFINE_PER_CPU(u64, arch_const_cycles_prev);
  93 static DEFINE_PER_CPU(u64, arch_core_cycles_prev);
  94 static cpumask_var_t amu_fie_cpus;
  95
  96 /* Initialize counter reference per-cpu variables for the current CPU */
  97 void init_cpu_freq_invariance_counters(void)
  98 {
  99         this_cpu_write(arch_core_cycles_prev,
 100                        read_sysreg_s(SYS_AMEVCNTR0_CORE_EL0));
 101         this_cpu_write(arch_const_cycles_prev,
 102                        read_sysreg_s(SYS_AMEVCNTR0_CONST_EL0));
 103 }
 104
 105 static int validate_cpu_freq_invariance_counters(int cpu)
 106 {
 107         u64 max_freq_hz, ratio;
 108
 109         if (!cpu_has_amu_feat(cpu)) {
 110                 pr_debug("CPU%d: counters are not supported.\n", cpu);
 111                 return -EINVAL;
 112         }
 113
 114         if (unlikely(!per_cpu(arch_const_cycles_prev, cpu) ||
 115                      !per_cpu(arch_core_cycles_prev, cpu))) {
 116                 pr_debug("CPU%d: cycle counters are not enabled.\n", cpu);
 117                 return -EINVAL;
 118         }
 119
 120         /* Convert maximum frequency from KHz to Hz and validate */
 121         max_freq_hz = cpufreq_get_hw_max_freq(cpu) * 1000ULL;
 122         if (unlikely(!max_freq_hz)) {
 123                 pr_debug("CPU%d: invalid maximum frequency.\n", cpu);
 124                 return -EINVAL;
 125         }
 126
 127         /*
 128          * Pre-compute the fixed ratio between the frequency of the constant
 129          * counter and the maximum frequency of the CPU.
 130          *
 131          *                            const_freq
 132          * arch_max_freq_scale =   ---------------- * SCHED_CAPACITY_SCALE²
 133          *                         cpuinfo_max_freq
 134          *
 135          * We use a factor of 2 * SCHED_CAPACITY_SHIFT -> SCHED_CAPACITY_SCALE²
 136          * in order to ensure a good resolution for arch_max_freq_scale for
 137          * very low arch timer frequencies (down to the KHz range which should
 138          * be unlikely).
 139          */
 140         ratio = (u64)arch_timer_get_rate() << (2 * SCHED_CAPACITY_SHIFT);
 141         ratio = div64_u64(ratio, max_freq_hz);
 142         if (!ratio) {
 143                 WARN_ONCE(1, "System timer frequency too low.\n");
 144                 return -EINVAL;
 145         }
 146
 147         per_cpu(arch_max_freq_scale, cpu) = (unsigned long)ratio;
 148
 149         return 0;
 150 }
 151
 152 static inline bool
 153 enable_policy_freq_counters(int cpu, cpumask_var_t valid_cpus)
 154 {
 155         struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
 156
 157         if (!policy) {
 158                 pr_debug("CPU%d: No cpufreq policy found.\n", cpu);
 159                 return false;
 160         }
 161
 162         if (cpumask_subset(policy->related_cpus, valid_cpus))
 163                 cpumask_or(amu_fie_cpus, policy->related_cpus,
 164                            amu_fie_cpus);
 165
 166         cpufreq_cpu_put(policy);
 167
 168         return true;
 169 }
 170
 171 static DEFINE_STATIC_KEY_FALSE(amu_fie_key);
 172 #define amu_freq_invariant() static_branch_unlikely(&amu_fie_key)
 173
 174 static int __init init_amu_fie(void)
 175 {
 176         cpumask_var_t valid_cpus;
 177         bool have_policy = false;
 178         int ret = 0;
 179         int cpu;
 180
 181         if (!zalloc_cpumask_var(&valid_cpus, GFP_KERNEL))
 182                 return -ENOMEM;
 183
 184         if (!zalloc_cpumask_var(&amu_fie_cpus, GFP_KERNEL)) {
 185                 ret = -ENOMEM;
 186                 goto free_valid_mask;
 187         }
 188
 189         for_each_present_cpu(cpu) {
 190                 if (validate_cpu_freq_invariance_counters(cpu))
 191                         continue;
 192                 cpumask_set_cpu(cpu, valid_cpus);
 193                 have_policy |= enable_policy_freq_counters(cpu, valid_cpus);
 194         }
 195
 196         /*
 197          * If we are not restricted by cpufreq policies, we only enable
 198          * the use of the AMU feature for FIE if all CPUs support AMU.
 199          * Otherwise, enable_policy_freq_counters has already enabled
 200          * policy cpus.
 201          */
 202         if (!have_policy && cpumask_equal(valid_cpus, cpu_present_mask))
 203                 cpumask_or(amu_fie_cpus, amu_fie_cpus, valid_cpus);
 204
 205         if (!cpumask_empty(amu_fie_cpus)) {
 206                 pr_info("CPUs[%*pbl]: counters will be used for FIE.",
 207                         cpumask_pr_args(amu_fie_cpus));
 208                 static_branch_enable(&amu_fie_key);
 209         }
 210
 211         /*
 212          * If the system is not fully invariant after AMU init, disable
 213          * partial use of counters for frequency invariance.
 214          */
 215         if (!topology_scale_freq_invariant())
 216                 static_branch_disable(&amu_fie_key);
 217
 218 free_valid_mask:
 219         free_cpumask_var(valid_cpus);
 220
 221         return ret;
 222 }
 223 late_initcall_sync(init_amu_fie);
 224
 225 bool arch_freq_counters_available(const struct cpumask *cpus)
 226 {
 227         return amu_freq_invariant() &&
 228                cpumask_subset(cpus, amu_fie_cpus);
 229 }
 230
 231 void topology_scale_freq_tick(void)
 232 {
 233         u64 prev_core_cnt, prev_const_cnt;
 234         u64 core_cnt, const_cnt, scale;
 235         int cpu = smp_processor_id();
 236
 237         if (!amu_freq_invariant())
 238                 return;
 239
 240         if (!cpumask_test_cpu(cpu, amu_fie_cpus))
 241                 return;
 242
 243         const_cnt = read_sysreg_s(SYS_AMEVCNTR0_CONST_EL0);
 244         core_cnt = read_sysreg_s(SYS_AMEVCNTR0_CORE_EL0);
 245         prev_const_cnt = this_cpu_read(arch_const_cycles_prev);
 246         prev_core_cnt = this_cpu_read(arch_core_cycles_prev);
 247
 248         if (unlikely(core_cnt <= prev_core_cnt ||
 249                      const_cnt <= prev_const_cnt))
 250                 goto store_and_exit;
 251
 252         /*
 253          *          /\core    arch_max_freq_scale
 254          * scale =  ------- * --------------------
 255          *          /\const   SCHED_CAPACITY_SCALE
 256          *
 257          * See validate_cpu_freq_invariance_counters() for details on
 258          * arch_max_freq_scale and the use of SCHED_CAPACITY_SHIFT.
 259          */
 260         scale = core_cnt - prev_core_cnt;
 261         scale *= this_cpu_read(arch_max_freq_scale);
 262         scale = div64_u64(scale >> SCHED_CAPACITY_SHIFT,
 263                           const_cnt - prev_const_cnt);
 264
 265         scale = min_t(unsigned long, scale, SCHED_CAPACITY_SCALE);
 266         this_cpu_write(freq_scale, (unsigned long)scale);
 267
 268 store_and_exit:
 269         this_cpu_write(arch_core_cycles_prev, core_cnt);
 270         this_cpu_write(arch_const_cycles_prev, const_cnt);
 271 }
 272 #endif /* CONFIG_ARM64_AMU_EXTN */