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 static ssize_t cpu_capacity_show(struct device *dev,
58 struct device_attribute *attr,
61 struct cpu *cpu = container_of(dev, struct cpu, dev);
63 return sprintf(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id));
66 static void update_topology_flags_workfn(struct work_struct *work);
67 static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
69 static DEVICE_ATTR_RO(cpu_capacity);
71 static int register_cpu_capacity_sysctl(void)
76 for_each_possible_cpu(i) {
77 cpu = get_cpu_device(i);
79 pr_err("%s: too early to get CPU%d device!\n",
83 device_create_file(cpu, &dev_attr_cpu_capacity);
88 subsys_initcall(register_cpu_capacity_sysctl);
90 static int update_topology;
92 int topology_update_cpu_topology(void)
94 return update_topology;
98 * Updating the sched_domains can't be done directly from cpufreq callbacks
99 * due to locking, so queue the work for later.
101 static void update_topology_flags_workfn(struct work_struct *work)
104 rebuild_sched_domains();
105 pr_debug("sched_domain hierarchy rebuilt, flags updated\n");
109 static DEFINE_PER_CPU(u32, freq_factor) = 1;
110 static u32 *raw_capacity;
112 static int free_raw_capacity(void)
120 void topology_normalize_cpu_scale(void)
130 for_each_possible_cpu(cpu) {
131 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
132 capacity_scale = max(capacity, capacity_scale);
135 pr_debug("cpu_capacity: capacity_scale=%llu\n", capacity_scale);
136 for_each_possible_cpu(cpu) {
137 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
138 capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
140 topology_set_cpu_scale(cpu, capacity);
141 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
142 cpu, topology_get_cpu_scale(cpu));
146 bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu)
149 static bool cap_parsing_failed;
153 if (cap_parsing_failed)
156 ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz",
160 raw_capacity = kcalloc(num_possible_cpus(),
161 sizeof(*raw_capacity),
164 cap_parsing_failed = true;
168 raw_capacity[cpu] = cpu_capacity;
169 pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
170 cpu_node, raw_capacity[cpu]);
173 * Update freq_factor for calculating early boot cpu capacities.
174 * For non-clk CPU DVFS mechanism, there's no way to get the
175 * frequency value now, assuming they are running at the same
176 * frequency (by keeping the initial freq_factor value).
178 cpu_clk = of_clk_get(cpu_node, 0);
179 if (!PTR_ERR_OR_ZERO(cpu_clk)) {
180 per_cpu(freq_factor, cpu) =
181 clk_get_rate(cpu_clk) / 1000;
186 pr_err("cpu_capacity: missing %pOF raw capacity\n",
188 pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
190 cap_parsing_failed = true;
197 #ifdef CONFIG_CPU_FREQ
198 static cpumask_var_t cpus_to_visit;
199 static void parsing_done_workfn(struct work_struct *work);
200 static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
203 init_cpu_capacity_callback(struct notifier_block *nb,
207 struct cpufreq_policy *policy = data;
213 if (val != CPUFREQ_CREATE_POLICY)
216 pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
217 cpumask_pr_args(policy->related_cpus),
218 cpumask_pr_args(cpus_to_visit));
220 cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
222 for_each_cpu(cpu, policy->related_cpus)
223 per_cpu(freq_factor, cpu) = policy->cpuinfo.max_freq / 1000;
225 if (cpumask_empty(cpus_to_visit)) {
226 topology_normalize_cpu_scale();
227 schedule_work(&update_topology_flags_work);
229 pr_debug("cpu_capacity: parsing done\n");
230 schedule_work(&parsing_done_work);
236 static struct notifier_block init_cpu_capacity_notifier = {
237 .notifier_call = init_cpu_capacity_callback,
240 static int __init register_cpufreq_notifier(void)
245 * on ACPI-based systems we need to use the default cpu capacity
246 * until we have the necessary code to parse the cpu capacity, so
247 * skip registering cpufreq notifier.
249 if (!acpi_disabled || !raw_capacity)
252 if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL))
255 cpumask_copy(cpus_to_visit, cpu_possible_mask);
257 ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
258 CPUFREQ_POLICY_NOTIFIER);
261 free_cpumask_var(cpus_to_visit);
265 core_initcall(register_cpufreq_notifier);
267 static void parsing_done_workfn(struct work_struct *work)
269 cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
270 CPUFREQ_POLICY_NOTIFIER);
271 free_cpumask_var(cpus_to_visit);
275 core_initcall(free_raw_capacity);
278 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
280 * This function returns the logic cpu number of the node.
281 * There are basically three kinds of return values:
282 * (1) logic cpu number which is > 0.
283 * (2) -ENODEV when the device tree(DT) node is valid and found in the DT but
284 * there is no possible logical CPU in the kernel to match. This happens
285 * when CONFIG_NR_CPUS is configure to be smaller than the number of
286 * CPU nodes in DT. We need to just ignore this case.
287 * (3) -1 if the node does not exist in the device tree
289 static int __init get_cpu_for_node(struct device_node *node)
291 struct device_node *cpu_node;
294 cpu_node = of_parse_phandle(node, "cpu", 0);
298 cpu = of_cpu_node_to_id(cpu_node);
300 topology_parse_cpu_capacity(cpu_node, cpu);
302 pr_info("CPU node for %pOF exist but the possible cpu range is :%*pbl\n",
303 cpu_node, cpumask_pr_args(cpu_possible_mask));
305 of_node_put(cpu_node);
309 static int __init parse_core(struct device_node *core, int package_id,
316 struct device_node *t;
319 snprintf(name, sizeof(name), "thread%d", i);
320 t = of_get_child_by_name(core, name);
323 cpu = get_cpu_for_node(t);
325 cpu_topology[cpu].package_id = package_id;
326 cpu_topology[cpu].core_id = core_id;
327 cpu_topology[cpu].thread_id = i;
328 } else if (cpu != -ENODEV) {
329 pr_err("%pOF: Can't get CPU for thread\n", t);
338 cpu = get_cpu_for_node(core);
341 pr_err("%pOF: Core has both threads and CPU\n",
346 cpu_topology[cpu].package_id = package_id;
347 cpu_topology[cpu].core_id = core_id;
348 } else if (leaf && cpu != -ENODEV) {
349 pr_err("%pOF: Can't get CPU for leaf core\n", core);
356 static int __init parse_cluster(struct device_node *cluster, int depth)
360 bool has_cores = false;
361 struct device_node *c;
362 static int package_id __initdata;
367 * First check for child clusters; we currently ignore any
368 * information about the nesting of clusters and present the
369 * scheduler with a flat list of them.
373 snprintf(name, sizeof(name), "cluster%d", i);
374 c = of_get_child_by_name(cluster, name);
377 ret = parse_cluster(c, depth + 1);
385 /* Now check for cores */
388 snprintf(name, sizeof(name), "core%d", i);
389 c = of_get_child_by_name(cluster, name);
394 pr_err("%pOF: cpu-map children should be clusters\n",
401 ret = parse_core(c, package_id, core_id++);
403 pr_err("%pOF: Non-leaf cluster with core %s\n",
415 if (leaf && !has_cores)
416 pr_warn("%pOF: empty cluster\n", cluster);
424 static int __init parse_dt_topology(void)
426 struct device_node *cn, *map;
430 cn = of_find_node_by_path("/cpus");
432 pr_err("No CPU information found in DT\n");
437 * When topology is provided cpu-map is essentially a root
438 * cluster with restricted subnodes.
440 map = of_get_child_by_name(cn, "cpu-map");
444 ret = parse_cluster(map, 0);
448 topology_normalize_cpu_scale();
451 * Check that all cores are in the topology; the SMP code will
452 * only mark cores described in the DT as possible.
454 for_each_possible_cpu(cpu)
455 if (cpu_topology[cpu].package_id == -1)
469 struct cpu_topology cpu_topology[NR_CPUS];
470 EXPORT_SYMBOL_GPL(cpu_topology);
472 const struct cpumask *cpu_coregroup_mask(int cpu)
474 const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
476 /* Find the smaller of NUMA, core or LLC siblings */
477 if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
478 /* not numa in package, lets use the package siblings */
479 core_mask = &cpu_topology[cpu].core_sibling;
481 if (cpu_topology[cpu].llc_id != -1) {
482 if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
483 core_mask = &cpu_topology[cpu].llc_sibling;
489 void update_siblings_masks(unsigned int cpuid)
491 struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
494 /* update core and thread sibling masks */
495 for_each_online_cpu(cpu) {
496 cpu_topo = &cpu_topology[cpu];
498 if (cpuid_topo->llc_id == cpu_topo->llc_id) {
499 cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling);
500 cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling);
503 if (cpuid_topo->package_id != cpu_topo->package_id)
506 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
507 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
509 if (cpuid_topo->core_id != cpu_topo->core_id)
512 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
513 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
517 static void clear_cpu_topology(int cpu)
519 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
521 cpumask_clear(&cpu_topo->llc_sibling);
522 cpumask_set_cpu(cpu, &cpu_topo->llc_sibling);
524 cpumask_clear(&cpu_topo->core_sibling);
525 cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
526 cpumask_clear(&cpu_topo->thread_sibling);
527 cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
530 void __init reset_cpu_topology(void)
534 for_each_possible_cpu(cpu) {
535 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
537 cpu_topo->thread_id = -1;
538 cpu_topo->core_id = -1;
539 cpu_topo->package_id = -1;
540 cpu_topo->llc_id = -1;
542 clear_cpu_topology(cpu);
546 void remove_cpu_topology(unsigned int cpu)
550 for_each_cpu(sibling, topology_core_cpumask(cpu))
551 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
552 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
553 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
554 for_each_cpu(sibling, topology_llc_cpumask(cpu))
555 cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling));
557 clear_cpu_topology(cpu);
560 __weak int __init parse_acpi_topology(void)
565 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
566 void __init init_cpu_topology(void)
568 reset_cpu_topology();
571 * Discard anything that was parsed if we hit an error so we
572 * don't use partial information.
574 if (parse_acpi_topology())
575 reset_cpu_topology();
576 else if (of_have_populated_dt() && parse_dt_topology())
577 reset_cpu_topology();