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 DEFINE_PER_CPU(unsigned long, freq_scale) = SCHED_CAPACITY_SCALE;
26 void arch_set_freq_scale(struct cpumask *cpus, unsigned long cur_freq,
27 unsigned long max_freq)
32 scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq;
35 per_cpu(freq_scale, i) = scale;
38 DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
40 void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
42 per_cpu(cpu_scale, cpu) = capacity;
45 static ssize_t cpu_capacity_show(struct device *dev,
46 struct device_attribute *attr,
49 struct cpu *cpu = container_of(dev, struct cpu, dev);
51 return sprintf(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id));
54 static void update_topology_flags_workfn(struct work_struct *work);
55 static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
57 static DEVICE_ATTR_RO(cpu_capacity);
59 static int register_cpu_capacity_sysctl(void)
64 for_each_possible_cpu(i) {
65 cpu = get_cpu_device(i);
67 pr_err("%s: too early to get CPU%d device!\n",
71 device_create_file(cpu, &dev_attr_cpu_capacity);
76 subsys_initcall(register_cpu_capacity_sysctl);
78 static int update_topology;
80 int topology_update_cpu_topology(void)
82 return update_topology;
86 * Updating the sched_domains can't be done directly from cpufreq callbacks
87 * due to locking, so queue the work for later.
89 static void update_topology_flags_workfn(struct work_struct *work)
92 rebuild_sched_domains();
93 pr_debug("sched_domain hierarchy rebuilt, flags updated\n");
97 static DEFINE_PER_CPU(u32, freq_factor) = 1;
98 static u32 *raw_capacity;
100 static int free_raw_capacity(void)
108 void topology_normalize_cpu_scale(void)
118 for_each_possible_cpu(cpu) {
119 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
120 capacity_scale = max(capacity, capacity_scale);
123 pr_debug("cpu_capacity: capacity_scale=%llu\n", capacity_scale);
124 for_each_possible_cpu(cpu) {
125 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
126 capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
128 topology_set_cpu_scale(cpu, capacity);
129 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
130 cpu, topology_get_cpu_scale(cpu));
134 bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu)
137 static bool cap_parsing_failed;
141 if (cap_parsing_failed)
144 ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz",
148 raw_capacity = kcalloc(num_possible_cpus(),
149 sizeof(*raw_capacity),
152 cap_parsing_failed = true;
156 raw_capacity[cpu] = cpu_capacity;
157 pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
158 cpu_node, raw_capacity[cpu]);
161 * Update freq_factor for calculating early boot cpu capacities.
162 * For non-clk CPU DVFS mechanism, there's no way to get the
163 * frequency value now, assuming they are running at the same
164 * frequency (by keeping the initial freq_factor value).
166 cpu_clk = of_clk_get(cpu_node, 0);
167 if (!PTR_ERR_OR_ZERO(cpu_clk)) {
168 per_cpu(freq_factor, cpu) =
169 clk_get_rate(cpu_clk) / 1000;
174 pr_err("cpu_capacity: missing %pOF raw capacity\n",
176 pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
178 cap_parsing_failed = true;
185 #ifdef CONFIG_CPU_FREQ
186 static cpumask_var_t cpus_to_visit;
187 static void parsing_done_workfn(struct work_struct *work);
188 static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
191 init_cpu_capacity_callback(struct notifier_block *nb,
195 struct cpufreq_policy *policy = data;
201 if (val != CPUFREQ_CREATE_POLICY)
204 pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
205 cpumask_pr_args(policy->related_cpus),
206 cpumask_pr_args(cpus_to_visit));
208 cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
210 for_each_cpu(cpu, policy->related_cpus)
211 per_cpu(freq_factor, cpu) = policy->cpuinfo.max_freq / 1000;
213 if (cpumask_empty(cpus_to_visit)) {
214 topology_normalize_cpu_scale();
215 schedule_work(&update_topology_flags_work);
217 pr_debug("cpu_capacity: parsing done\n");
218 schedule_work(&parsing_done_work);
224 static struct notifier_block init_cpu_capacity_notifier = {
225 .notifier_call = init_cpu_capacity_callback,
228 static int __init register_cpufreq_notifier(void)
233 * on ACPI-based systems we need to use the default cpu capacity
234 * until we have the necessary code to parse the cpu capacity, so
235 * skip registering cpufreq notifier.
237 if (!acpi_disabled || !raw_capacity)
240 if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL))
243 cpumask_copy(cpus_to_visit, cpu_possible_mask);
245 ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
246 CPUFREQ_POLICY_NOTIFIER);
249 free_cpumask_var(cpus_to_visit);
253 core_initcall(register_cpufreq_notifier);
255 static void parsing_done_workfn(struct work_struct *work)
257 cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
258 CPUFREQ_POLICY_NOTIFIER);
259 free_cpumask_var(cpus_to_visit);
263 core_initcall(free_raw_capacity);
266 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
268 * This function returns the logic cpu number of the node.
269 * There are basically three kinds of return values:
270 * (1) logic cpu number which is > 0.
271 * (2) -ENODEV when the device tree(DT) node is valid and found in the DT but
272 * there is no possible logical CPU in the kernel to match. This happens
273 * when CONFIG_NR_CPUS is configure to be smaller than the number of
274 * CPU nodes in DT. We need to just ignore this case.
275 * (3) -1 if the node does not exist in the device tree
277 static int __init get_cpu_for_node(struct device_node *node)
279 struct device_node *cpu_node;
282 cpu_node = of_parse_phandle(node, "cpu", 0);
286 cpu = of_cpu_node_to_id(cpu_node);
288 topology_parse_cpu_capacity(cpu_node, cpu);
290 pr_info("CPU node for %pOF exist but the possible cpu range is :%*pbl\n",
291 cpu_node, cpumask_pr_args(cpu_possible_mask));
293 of_node_put(cpu_node);
297 static int __init parse_core(struct device_node *core, int package_id,
304 struct device_node *t;
307 snprintf(name, sizeof(name), "thread%d", i);
308 t = of_get_child_by_name(core, name);
311 cpu = get_cpu_for_node(t);
313 cpu_topology[cpu].package_id = package_id;
314 cpu_topology[cpu].core_id = core_id;
315 cpu_topology[cpu].thread_id = i;
316 } else if (cpu != -ENODEV) {
317 pr_err("%pOF: Can't get CPU for thread\n", t);
326 cpu = get_cpu_for_node(core);
329 pr_err("%pOF: Core has both threads and CPU\n",
334 cpu_topology[cpu].package_id = package_id;
335 cpu_topology[cpu].core_id = core_id;
336 } else if (leaf && cpu != -ENODEV) {
337 pr_err("%pOF: Can't get CPU for leaf core\n", core);
344 static int __init parse_cluster(struct device_node *cluster, int depth)
348 bool has_cores = false;
349 struct device_node *c;
350 static int package_id __initdata;
355 * First check for child clusters; we currently ignore any
356 * information about the nesting of clusters and present the
357 * scheduler with a flat list of them.
361 snprintf(name, sizeof(name), "cluster%d", i);
362 c = of_get_child_by_name(cluster, name);
365 ret = parse_cluster(c, depth + 1);
373 /* Now check for cores */
376 snprintf(name, sizeof(name), "core%d", i);
377 c = of_get_child_by_name(cluster, name);
382 pr_err("%pOF: cpu-map children should be clusters\n",
389 ret = parse_core(c, package_id, core_id++);
391 pr_err("%pOF: Non-leaf cluster with core %s\n",
403 if (leaf && !has_cores)
404 pr_warn("%pOF: empty cluster\n", cluster);
412 static int __init parse_dt_topology(void)
414 struct device_node *cn, *map;
418 cn = of_find_node_by_path("/cpus");
420 pr_err("No CPU information found in DT\n");
425 * When topology is provided cpu-map is essentially a root
426 * cluster with restricted subnodes.
428 map = of_get_child_by_name(cn, "cpu-map");
432 ret = parse_cluster(map, 0);
436 topology_normalize_cpu_scale();
439 * Check that all cores are in the topology; the SMP code will
440 * only mark cores described in the DT as possible.
442 for_each_possible_cpu(cpu)
443 if (cpu_topology[cpu].package_id == -1)
457 struct cpu_topology cpu_topology[NR_CPUS];
458 EXPORT_SYMBOL_GPL(cpu_topology);
460 const struct cpumask *cpu_coregroup_mask(int cpu)
462 const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
464 /* Find the smaller of NUMA, core or LLC siblings */
465 if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
466 /* not numa in package, lets use the package siblings */
467 core_mask = &cpu_topology[cpu].core_sibling;
469 if (cpu_topology[cpu].llc_id != -1) {
470 if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
471 core_mask = &cpu_topology[cpu].llc_sibling;
477 void update_siblings_masks(unsigned int cpuid)
479 struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
482 /* update core and thread sibling masks */
483 for_each_online_cpu(cpu) {
484 cpu_topo = &cpu_topology[cpu];
486 if (cpuid_topo->llc_id == cpu_topo->llc_id) {
487 cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling);
488 cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling);
491 if (cpuid_topo->package_id != cpu_topo->package_id)
494 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
495 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
497 if (cpuid_topo->core_id != cpu_topo->core_id)
500 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
501 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
505 static void clear_cpu_topology(int cpu)
507 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
509 cpumask_clear(&cpu_topo->llc_sibling);
510 cpumask_set_cpu(cpu, &cpu_topo->llc_sibling);
512 cpumask_clear(&cpu_topo->core_sibling);
513 cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
514 cpumask_clear(&cpu_topo->thread_sibling);
515 cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
518 void __init reset_cpu_topology(void)
522 for_each_possible_cpu(cpu) {
523 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
525 cpu_topo->thread_id = -1;
526 cpu_topo->core_id = -1;
527 cpu_topo->package_id = -1;
528 cpu_topo->llc_id = -1;
530 clear_cpu_topology(cpu);
534 void remove_cpu_topology(unsigned int cpu)
538 for_each_cpu(sibling, topology_core_cpumask(cpu))
539 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
540 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
541 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
542 for_each_cpu(sibling, topology_llc_cpumask(cpu))
543 cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling));
545 clear_cpu_topology(cpu);
548 __weak int __init parse_acpi_topology(void)
553 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
554 void __init init_cpu_topology(void)
556 reset_cpu_topology();
559 * Discard anything that was parsed if we hit an error so we
560 * don't use partial information.
562 if (parse_acpi_topology())
563 reset_cpu_topology();
564 else if (of_have_populated_dt() && parse_dt_topology())
565 reset_cpu_topology();