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 bool topology_scale_freq_invariant(void)
26 return cpufreq_supports_freq_invariance() ||
27 arch_freq_counters_available(cpu_online_mask);
30 __weak bool arch_freq_counters_available(const struct cpumask *cpus)
34 DEFINE_PER_CPU(unsigned long, freq_scale) = SCHED_CAPACITY_SCALE;
36 void topology_set_freq_scale(const struct cpumask *cpus, unsigned long cur_freq,
37 unsigned long max_freq)
42 if (WARN_ON_ONCE(!cur_freq || !max_freq))
46 * If the use of counters for FIE is enabled, just return as we don't
47 * want to update the scale factor with information from CPUFREQ.
48 * Instead the scale factor will be updated from arch_scale_freq_tick.
50 if (arch_freq_counters_available(cpus))
53 scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq;
56 per_cpu(freq_scale, i) = scale;
59 DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
61 void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
63 per_cpu(cpu_scale, cpu) = capacity;
66 DEFINE_PER_CPU(unsigned long, thermal_pressure);
68 void topology_set_thermal_pressure(const struct cpumask *cpus,
69 unsigned long th_pressure)
73 for_each_cpu(cpu, cpus)
74 WRITE_ONCE(per_cpu(thermal_pressure, cpu), th_pressure);
77 static ssize_t cpu_capacity_show(struct device *dev,
78 struct device_attribute *attr,
81 struct cpu *cpu = container_of(dev, struct cpu, dev);
83 return sysfs_emit(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id));
86 static void update_topology_flags_workfn(struct work_struct *work);
87 static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
89 static DEVICE_ATTR_RO(cpu_capacity);
91 static int register_cpu_capacity_sysctl(void)
96 for_each_possible_cpu(i) {
97 cpu = get_cpu_device(i);
99 pr_err("%s: too early to get CPU%d device!\n",
103 device_create_file(cpu, &dev_attr_cpu_capacity);
108 subsys_initcall(register_cpu_capacity_sysctl);
110 static int update_topology;
112 int topology_update_cpu_topology(void)
114 return update_topology;
118 * Updating the sched_domains can't be done directly from cpufreq callbacks
119 * due to locking, so queue the work for later.
121 static void update_topology_flags_workfn(struct work_struct *work)
124 rebuild_sched_domains();
125 pr_debug("sched_domain hierarchy rebuilt, flags updated\n");
129 static DEFINE_PER_CPU(u32, freq_factor) = 1;
130 static u32 *raw_capacity;
132 static int free_raw_capacity(void)
140 void topology_normalize_cpu_scale(void)
150 for_each_possible_cpu(cpu) {
151 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
152 capacity_scale = max(capacity, capacity_scale);
155 pr_debug("cpu_capacity: capacity_scale=%llu\n", capacity_scale);
156 for_each_possible_cpu(cpu) {
157 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
158 capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
160 topology_set_cpu_scale(cpu, capacity);
161 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
162 cpu, topology_get_cpu_scale(cpu));
166 bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu)
169 static bool cap_parsing_failed;
173 if (cap_parsing_failed)
176 ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz",
180 raw_capacity = kcalloc(num_possible_cpus(),
181 sizeof(*raw_capacity),
184 cap_parsing_failed = true;
188 raw_capacity[cpu] = cpu_capacity;
189 pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
190 cpu_node, raw_capacity[cpu]);
193 * Update freq_factor for calculating early boot cpu capacities.
194 * For non-clk CPU DVFS mechanism, there's no way to get the
195 * frequency value now, assuming they are running at the same
196 * frequency (by keeping the initial freq_factor value).
198 cpu_clk = of_clk_get(cpu_node, 0);
199 if (!PTR_ERR_OR_ZERO(cpu_clk)) {
200 per_cpu(freq_factor, cpu) =
201 clk_get_rate(cpu_clk) / 1000;
206 pr_err("cpu_capacity: missing %pOF raw capacity\n",
208 pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
210 cap_parsing_failed = true;
217 #ifdef CONFIG_CPU_FREQ
218 static cpumask_var_t cpus_to_visit;
219 static void parsing_done_workfn(struct work_struct *work);
220 static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
223 init_cpu_capacity_callback(struct notifier_block *nb,
227 struct cpufreq_policy *policy = data;
233 if (val != CPUFREQ_CREATE_POLICY)
236 pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
237 cpumask_pr_args(policy->related_cpus),
238 cpumask_pr_args(cpus_to_visit));
240 cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
242 for_each_cpu(cpu, policy->related_cpus)
243 per_cpu(freq_factor, cpu) = policy->cpuinfo.max_freq / 1000;
245 if (cpumask_empty(cpus_to_visit)) {
246 topology_normalize_cpu_scale();
247 schedule_work(&update_topology_flags_work);
249 pr_debug("cpu_capacity: parsing done\n");
250 schedule_work(&parsing_done_work);
256 static struct notifier_block init_cpu_capacity_notifier = {
257 .notifier_call = init_cpu_capacity_callback,
260 static int __init register_cpufreq_notifier(void)
265 * on ACPI-based systems we need to use the default cpu capacity
266 * until we have the necessary code to parse the cpu capacity, so
267 * skip registering cpufreq notifier.
269 if (!acpi_disabled || !raw_capacity)
272 if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL))
275 cpumask_copy(cpus_to_visit, cpu_possible_mask);
277 ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
278 CPUFREQ_POLICY_NOTIFIER);
281 free_cpumask_var(cpus_to_visit);
285 core_initcall(register_cpufreq_notifier);
287 static void parsing_done_workfn(struct work_struct *work)
289 cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
290 CPUFREQ_POLICY_NOTIFIER);
291 free_cpumask_var(cpus_to_visit);
295 core_initcall(free_raw_capacity);
298 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
300 * This function returns the logic cpu number of the node.
301 * There are basically three kinds of return values:
302 * (1) logic cpu number which is > 0.
303 * (2) -ENODEV when the device tree(DT) node is valid and found in the DT but
304 * there is no possible logical CPU in the kernel to match. This happens
305 * when CONFIG_NR_CPUS is configure to be smaller than the number of
306 * CPU nodes in DT. We need to just ignore this case.
307 * (3) -1 if the node does not exist in the device tree
309 static int __init get_cpu_for_node(struct device_node *node)
311 struct device_node *cpu_node;
314 cpu_node = of_parse_phandle(node, "cpu", 0);
318 cpu = of_cpu_node_to_id(cpu_node);
320 topology_parse_cpu_capacity(cpu_node, cpu);
322 pr_info("CPU node for %pOF exist but the possible cpu range is :%*pbl\n",
323 cpu_node, cpumask_pr_args(cpu_possible_mask));
325 of_node_put(cpu_node);
329 static int __init parse_core(struct device_node *core, int package_id,
336 struct device_node *t;
339 snprintf(name, sizeof(name), "thread%d", i);
340 t = of_get_child_by_name(core, name);
343 cpu = get_cpu_for_node(t);
345 cpu_topology[cpu].package_id = package_id;
346 cpu_topology[cpu].core_id = core_id;
347 cpu_topology[cpu].thread_id = i;
348 } else if (cpu != -ENODEV) {
349 pr_err("%pOF: Can't get CPU for thread\n", t);
358 cpu = get_cpu_for_node(core);
361 pr_err("%pOF: Core has both threads and CPU\n",
366 cpu_topology[cpu].package_id = package_id;
367 cpu_topology[cpu].core_id = core_id;
368 } else if (leaf && cpu != -ENODEV) {
369 pr_err("%pOF: Can't get CPU for leaf core\n", core);
376 static int __init parse_cluster(struct device_node *cluster, int depth)
380 bool has_cores = false;
381 struct device_node *c;
382 static int package_id __initdata;
387 * First check for child clusters; we currently ignore any
388 * information about the nesting of clusters and present the
389 * scheduler with a flat list of them.
393 snprintf(name, sizeof(name), "cluster%d", i);
394 c = of_get_child_by_name(cluster, name);
397 ret = parse_cluster(c, depth + 1);
405 /* Now check for cores */
408 snprintf(name, sizeof(name), "core%d", i);
409 c = of_get_child_by_name(cluster, name);
414 pr_err("%pOF: cpu-map children should be clusters\n",
421 ret = parse_core(c, package_id, core_id++);
423 pr_err("%pOF: Non-leaf cluster with core %s\n",
435 if (leaf && !has_cores)
436 pr_warn("%pOF: empty cluster\n", cluster);
444 static int __init parse_dt_topology(void)
446 struct device_node *cn, *map;
450 cn = of_find_node_by_path("/cpus");
452 pr_err("No CPU information found in DT\n");
457 * When topology is provided cpu-map is essentially a root
458 * cluster with restricted subnodes.
460 map = of_get_child_by_name(cn, "cpu-map");
464 ret = parse_cluster(map, 0);
468 topology_normalize_cpu_scale();
471 * Check that all cores are in the topology; the SMP code will
472 * only mark cores described in the DT as possible.
474 for_each_possible_cpu(cpu)
475 if (cpu_topology[cpu].package_id == -1)
489 struct cpu_topology cpu_topology[NR_CPUS];
490 EXPORT_SYMBOL_GPL(cpu_topology);
492 const struct cpumask *cpu_coregroup_mask(int cpu)
494 const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
496 /* Find the smaller of NUMA, core or LLC siblings */
497 if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
498 /* not numa in package, lets use the package siblings */
499 core_mask = &cpu_topology[cpu].core_sibling;
501 if (cpu_topology[cpu].llc_id != -1) {
502 if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
503 core_mask = &cpu_topology[cpu].llc_sibling;
509 void update_siblings_masks(unsigned int cpuid)
511 struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
514 /* update core and thread sibling masks */
515 for_each_online_cpu(cpu) {
516 cpu_topo = &cpu_topology[cpu];
518 if (cpu_topo->llc_id != -1 && cpuid_topo->llc_id == cpu_topo->llc_id) {
519 cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling);
520 cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling);
523 if (cpuid_topo->package_id != cpu_topo->package_id)
526 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
527 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
529 if (cpuid_topo->core_id != cpu_topo->core_id)
532 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
533 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
537 static void clear_cpu_topology(int cpu)
539 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
541 cpumask_clear(&cpu_topo->llc_sibling);
542 cpumask_set_cpu(cpu, &cpu_topo->llc_sibling);
544 cpumask_clear(&cpu_topo->core_sibling);
545 cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
546 cpumask_clear(&cpu_topo->thread_sibling);
547 cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
550 void __init reset_cpu_topology(void)
554 for_each_possible_cpu(cpu) {
555 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
557 cpu_topo->thread_id = -1;
558 cpu_topo->core_id = -1;
559 cpu_topo->package_id = -1;
560 cpu_topo->llc_id = -1;
562 clear_cpu_topology(cpu);
566 void remove_cpu_topology(unsigned int cpu)
570 for_each_cpu(sibling, topology_core_cpumask(cpu))
571 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
572 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
573 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
574 for_each_cpu(sibling, topology_llc_cpumask(cpu))
575 cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling));
577 clear_cpu_topology(cpu);
580 __weak int __init parse_acpi_topology(void)
585 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
586 void __init init_cpu_topology(void)
588 reset_cpu_topology();
591 * Discard anything that was parsed if we hit an error so we
592 * don't use partial information.
594 if (parse_acpi_topology())
595 reset_cpu_topology();
596 else if (of_have_populated_dt() && parse_dt_topology())
597 reset_cpu_topology();
600 void store_cpu_topology(unsigned int cpuid)
602 struct cpu_topology *cpuid_topo = &cpu_topology[cpuid];
604 if (cpuid_topo->package_id != -1)
605 goto topology_populated;
607 cpuid_topo->thread_id = -1;
608 cpuid_topo->core_id = cpuid;
609 cpuid_topo->package_id = cpu_to_node(cpuid);
611 pr_debug("CPU%u: package %d core %d thread %d\n",
612 cpuid, cpuid_topo->package_id, cpuid_topo->core_id,
613 cpuid_topo->thread_id);
616 update_siblings_masks(cpuid);