1 // SPDX-License-Identifier: GPL-2.0
3 * Arch specific cpu topology information
5 * Copyright (C) 2016, ARM Ltd.
6 * Written by: Juri Lelli, ARM Ltd.
9 #include <linux/acpi.h>
10 #include <linux/cpu.h>
11 #include <linux/cpufreq.h>
12 #include <linux/device.h>
14 #include <linux/slab.h>
15 #include <linux/string.h>
16 #include <linux/sched/topology.h>
17 #include <linux/cpuset.h>
18 #include <linux/cpumask.h>
19 #include <linux/init.h>
20 #include <linux/percpu.h>
21 #include <linux/sched.h>
22 #include <linux/smp.h>
24 __weak bool arch_freq_counters_available(struct cpumask *cpus)
28 DEFINE_PER_CPU(unsigned long, freq_scale) = SCHED_CAPACITY_SCALE;
30 void arch_set_freq_scale(struct cpumask *cpus, unsigned long cur_freq,
31 unsigned long max_freq)
37 * If the use of counters for FIE is enabled, just return as we don't
38 * want to update the scale factor with information from CPUFREQ.
39 * Instead the scale factor will be updated from arch_scale_freq_tick.
41 if (arch_freq_counters_available(cpus))
44 scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq;
47 per_cpu(freq_scale, i) = scale;
50 DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
52 void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
54 per_cpu(cpu_scale, cpu) = capacity;
57 DEFINE_PER_CPU(unsigned long, thermal_pressure);
59 void topology_set_thermal_pressure(const struct cpumask *cpus,
60 unsigned long th_pressure)
64 for_each_cpu(cpu, cpus)
65 WRITE_ONCE(per_cpu(thermal_pressure, cpu), th_pressure);
68 static ssize_t cpu_capacity_show(struct device *dev,
69 struct device_attribute *attr,
72 struct cpu *cpu = container_of(dev, struct cpu, dev);
74 return sprintf(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id));
77 static void update_topology_flags_workfn(struct work_struct *work);
78 static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
80 static DEVICE_ATTR_RO(cpu_capacity);
82 static int register_cpu_capacity_sysctl(void)
87 for_each_possible_cpu(i) {
88 cpu = get_cpu_device(i);
90 pr_err("%s: too early to get CPU%d device!\n",
94 device_create_file(cpu, &dev_attr_cpu_capacity);
99 subsys_initcall(register_cpu_capacity_sysctl);
101 static int update_topology;
103 int topology_update_cpu_topology(void)
105 return update_topology;
109 * Updating the sched_domains can't be done directly from cpufreq callbacks
110 * due to locking, so queue the work for later.
112 static void update_topology_flags_workfn(struct work_struct *work)
115 rebuild_sched_domains();
116 pr_debug("sched_domain hierarchy rebuilt, flags updated\n");
120 static DEFINE_PER_CPU(u32, freq_factor) = 1;
121 static u32 *raw_capacity;
123 static int free_raw_capacity(void)
131 void topology_normalize_cpu_scale(void)
141 for_each_possible_cpu(cpu) {
142 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
143 capacity_scale = max(capacity, capacity_scale);
146 pr_debug("cpu_capacity: capacity_scale=%llu\n", capacity_scale);
147 for_each_possible_cpu(cpu) {
148 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
149 capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
151 topology_set_cpu_scale(cpu, capacity);
152 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
153 cpu, topology_get_cpu_scale(cpu));
157 bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu)
160 static bool cap_parsing_failed;
164 if (cap_parsing_failed)
167 ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz",
171 raw_capacity = kcalloc(num_possible_cpus(),
172 sizeof(*raw_capacity),
175 cap_parsing_failed = true;
179 raw_capacity[cpu] = cpu_capacity;
180 pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
181 cpu_node, raw_capacity[cpu]);
184 * Update freq_factor for calculating early boot cpu capacities.
185 * For non-clk CPU DVFS mechanism, there's no way to get the
186 * frequency value now, assuming they are running at the same
187 * frequency (by keeping the initial freq_factor value).
189 cpu_clk = of_clk_get(cpu_node, 0);
190 if (!PTR_ERR_OR_ZERO(cpu_clk)) {
191 per_cpu(freq_factor, cpu) =
192 clk_get_rate(cpu_clk) / 1000;
197 pr_err("cpu_capacity: missing %pOF raw capacity\n",
199 pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
201 cap_parsing_failed = true;
208 #ifdef CONFIG_CPU_FREQ
209 static cpumask_var_t cpus_to_visit;
210 static void parsing_done_workfn(struct work_struct *work);
211 static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
214 init_cpu_capacity_callback(struct notifier_block *nb,
218 struct cpufreq_policy *policy = data;
224 if (val != CPUFREQ_CREATE_POLICY)
227 pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
228 cpumask_pr_args(policy->related_cpus),
229 cpumask_pr_args(cpus_to_visit));
231 cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
233 for_each_cpu(cpu, policy->related_cpus)
234 per_cpu(freq_factor, cpu) = policy->cpuinfo.max_freq / 1000;
236 if (cpumask_empty(cpus_to_visit)) {
237 topology_normalize_cpu_scale();
238 schedule_work(&update_topology_flags_work);
240 pr_debug("cpu_capacity: parsing done\n");
241 schedule_work(&parsing_done_work);
247 static struct notifier_block init_cpu_capacity_notifier = {
248 .notifier_call = init_cpu_capacity_callback,
251 static int __init register_cpufreq_notifier(void)
256 * on ACPI-based systems we need to use the default cpu capacity
257 * until we have the necessary code to parse the cpu capacity, so
258 * skip registering cpufreq notifier.
260 if (!acpi_disabled || !raw_capacity)
263 if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL))
266 cpumask_copy(cpus_to_visit, cpu_possible_mask);
268 ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
269 CPUFREQ_POLICY_NOTIFIER);
272 free_cpumask_var(cpus_to_visit);
276 core_initcall(register_cpufreq_notifier);
278 static void parsing_done_workfn(struct work_struct *work)
280 cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
281 CPUFREQ_POLICY_NOTIFIER);
282 free_cpumask_var(cpus_to_visit);
286 core_initcall(free_raw_capacity);
289 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
291 * This function returns the logic cpu number of the node.
292 * There are basically three kinds of return values:
293 * (1) logic cpu number which is > 0.
294 * (2) -ENODEV when the device tree(DT) node is valid and found in the DT but
295 * there is no possible logical CPU in the kernel to match. This happens
296 * when CONFIG_NR_CPUS is configure to be smaller than the number of
297 * CPU nodes in DT. We need to just ignore this case.
298 * (3) -1 if the node does not exist in the device tree
300 static int __init get_cpu_for_node(struct device_node *node)
302 struct device_node *cpu_node;
305 cpu_node = of_parse_phandle(node, "cpu", 0);
309 cpu = of_cpu_node_to_id(cpu_node);
311 topology_parse_cpu_capacity(cpu_node, cpu);
313 pr_info("CPU node for %pOF exist but the possible cpu range is :%*pbl\n",
314 cpu_node, cpumask_pr_args(cpu_possible_mask));
316 of_node_put(cpu_node);
320 static int __init parse_core(struct device_node *core, int package_id,
327 struct device_node *t;
330 snprintf(name, sizeof(name), "thread%d", i);
331 t = of_get_child_by_name(core, name);
334 cpu = get_cpu_for_node(t);
336 cpu_topology[cpu].package_id = package_id;
337 cpu_topology[cpu].core_id = core_id;
338 cpu_topology[cpu].thread_id = i;
339 } else if (cpu != -ENODEV) {
340 pr_err("%pOF: Can't get CPU for thread\n", t);
349 cpu = get_cpu_for_node(core);
352 pr_err("%pOF: Core has both threads and CPU\n",
357 cpu_topology[cpu].package_id = package_id;
358 cpu_topology[cpu].core_id = core_id;
359 } else if (leaf && cpu != -ENODEV) {
360 pr_err("%pOF: Can't get CPU for leaf core\n", core);
367 static int __init parse_cluster(struct device_node *cluster, int depth)
371 bool has_cores = false;
372 struct device_node *c;
373 static int package_id __initdata;
378 * First check for child clusters; we currently ignore any
379 * information about the nesting of clusters and present the
380 * scheduler with a flat list of them.
384 snprintf(name, sizeof(name), "cluster%d", i);
385 c = of_get_child_by_name(cluster, name);
388 ret = parse_cluster(c, depth + 1);
396 /* Now check for cores */
399 snprintf(name, sizeof(name), "core%d", i);
400 c = of_get_child_by_name(cluster, name);
405 pr_err("%pOF: cpu-map children should be clusters\n",
412 ret = parse_core(c, package_id, core_id++);
414 pr_err("%pOF: Non-leaf cluster with core %s\n",
426 if (leaf && !has_cores)
427 pr_warn("%pOF: empty cluster\n", cluster);
435 static int __init parse_dt_topology(void)
437 struct device_node *cn, *map;
441 cn = of_find_node_by_path("/cpus");
443 pr_err("No CPU information found in DT\n");
448 * When topology is provided cpu-map is essentially a root
449 * cluster with restricted subnodes.
451 map = of_get_child_by_name(cn, "cpu-map");
455 ret = parse_cluster(map, 0);
459 topology_normalize_cpu_scale();
462 * Check that all cores are in the topology; the SMP code will
463 * only mark cores described in the DT as possible.
465 for_each_possible_cpu(cpu)
466 if (cpu_topology[cpu].package_id == -1)
480 struct cpu_topology cpu_topology[NR_CPUS];
481 EXPORT_SYMBOL_GPL(cpu_topology);
483 const struct cpumask *cpu_coregroup_mask(int cpu)
485 const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
487 /* Find the smaller of NUMA, core or LLC siblings */
488 if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
489 /* not numa in package, lets use the package siblings */
490 core_mask = &cpu_topology[cpu].core_sibling;
492 if (cpu_topology[cpu].llc_id != -1) {
493 if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
494 core_mask = &cpu_topology[cpu].llc_sibling;
500 void update_siblings_masks(unsigned int cpuid)
502 struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
505 /* update core and thread sibling masks */
506 for_each_online_cpu(cpu) {
507 cpu_topo = &cpu_topology[cpu];
509 if (cpuid_topo->llc_id == cpu_topo->llc_id) {
510 cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling);
511 cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling);
514 if (cpuid_topo->package_id != cpu_topo->package_id)
517 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
518 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
520 if (cpuid_topo->core_id != cpu_topo->core_id)
523 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
524 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
528 static void clear_cpu_topology(int cpu)
530 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
532 cpumask_clear(&cpu_topo->llc_sibling);
533 cpumask_set_cpu(cpu, &cpu_topo->llc_sibling);
535 cpumask_clear(&cpu_topo->core_sibling);
536 cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
537 cpumask_clear(&cpu_topo->thread_sibling);
538 cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
541 void __init reset_cpu_topology(void)
545 for_each_possible_cpu(cpu) {
546 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
548 cpu_topo->thread_id = -1;
549 cpu_topo->core_id = -1;
550 cpu_topo->package_id = -1;
551 cpu_topo->llc_id = -1;
553 clear_cpu_topology(cpu);
557 void remove_cpu_topology(unsigned int cpu)
561 for_each_cpu(sibling, topology_core_cpumask(cpu))
562 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
563 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
564 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
565 for_each_cpu(sibling, topology_llc_cpumask(cpu))
566 cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling));
568 clear_cpu_topology(cpu);
571 __weak int __init parse_acpi_topology(void)
576 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
577 void __init init_cpu_topology(void)
579 reset_cpu_topology();
582 * Discard anything that was parsed if we hit an error so we
583 * don't use partial information.
585 if (parse_acpi_topology())
586 reset_cpu_topology();
587 else if (of_have_populated_dt() && parse_dt_topology())
588 reset_cpu_topology();